Weak ties, strong facts: how influencers amplify public health communication on X (formerly Twitter)

2.1 Weak ties and network centrality in social networks

In social networks, weak ties, defined as social contacts with looser connections (Kadushin 2012), serve as bridges connecting separate clusters of actors, enabling novel information to spread across communities (Granovetter 1973), and underpinning bridging social capital in social networks (Putnam 2000). Network centrality metrics formalize these roles: out-degree centrality (the number of outgoing connections a node has) reflects a user’s broadcast capacity, while betweenness centrality (how often a node lies on shortest paths between others) identifies its role as a bridge between groups (J. Zhang and Luo 2017). For example, a user with high out-degree (an “accelerator”) can share content to many contacts, and one with high betweenness (a “bridger” or “connector”) links otherwise disconnected audiences. Empirical studies confirm that high-degree nodes tend to be influential in information diffusion: one analysis of Twitter data found users with high degree centrality were more likely to spread information widely (Gupta et al. 2023; Hansen et al. 2020). In public health contexts, these weak-tie bridge roles can overcome echo chambers by carrying messages across social divides (Cava et al. 2023; Valenzuela et al. 2018).

During the COVID-19 pandemic, studies of vaccine discourse illustrate these network dynamics. For example, one multilayer Twitter network analysis found distinct pro- and anti-vaccine communities with very different structures: anti-vaccination users had much denser, more cohesive ego-networks than pro-vaccine users (Bonifazi et al. 2022). Similarly, a study of COVID-19 vaccine discussions on Weibo (a Twitter-like platform in China) found low overall echo-chamber effects: identified opinion leaders and “structural hole spanners” (bridging nodes) acted as bridgers connecting diverse vaccine topics and attitudes (Wang et al. 2022). These users linked disparate groups of users and vaccine viewpoints, exemplifying the Granovetter (1973) hypothesis that weak ties facilitate novel information flow. In line with this, research on Japanese Twitter data showed that a user’s community membership strongly predicted their opinion shift: users embedded in a pro-vaccine community were significantly more likely to adopt pro-vaccine stances over time, and similarly for anti-vaccine communities (Q. Wu et al. 2025).

This study integrates these frameworks to examine the role of weak-tie networks in shaping health communication dynamics on Twitter. Using public health discourse –vaccine-related discussions during the COVID-19 crises, this research explores how weak-tie power, which is measured through network centrality metrics, shapes the mechanisms of informational and normative social influence in health-related information dissemination.

Based on the literature review, it was hypothesized that weak ties (e.g., acquaintances, colleagues, or distant connections) enable more efficient information diffusion and bridging across communities compared to strong ties (e.g., family, close friends). Specifically, users with high out-degree centrality (accelerators or amplifiers) and high betweenness centrality (bridgers) are theorized to leverage weak-tie networks to accelerate the spread of information and connect fragmented subgroups, thereby overcoming echo chambers and expanding the reach of critical messaging (Rodrigues 2019). High out-degree centrality reflects a user’s capacity to amplify reach and distribute information widely, which aligns with the characteristics of weak ties as bridges for information (Granovetter 1973). This bridging function exemplifies weak-tie power, enabling novel information to traverse disconnected networks, which is consistent with Granovetter’s (1973) theory. Therefore, the following hypotheses were proposed:

2.2 Social influence and conformity on social media

Conformity theory, proposed by Solomon Asch based on a series of conformity experiments (Asch 1951, 1955), refers to individuals yielding their beliefs and behaviors to a group due to two different types of pressures (Crutchfield 1955). Normative social influence involves conforming to fit in and gain social acceptance (Sherif 1936; Sumner 1906) while informational social influence involves conforming because others are presumed to have accurate information (Festinger et al. 1950; Hardin and Higgins 1996).

On social media, these social pressures are made explicit through visible cues. In normative social influence, users retweet or like content to enhance their social image or avoid rejection by their peers (Ajzen and Fishbein 1980), aiming to mitigate risks tied to deviating from perceived majority behaviors (Latané and Wolf 1981). For instance, trending hashtags and high retweet counts signal majority norms, nudging users to adopt prevailing opinions or behaviors. This process amplifies specific narratives, reinforces (or challenges) existing social norms, as well as lays the foundation for the formation of new ones (Young and Jordan 2013). The visibility of these norms creates a feedback loop that further drives their adoption (Khasawneh et al. 2021; McGraw 2020; Samet 2020). Consequently, widely shared content, even if unverified, gains perceived validity through sheer exposure, as users interpret popularity as a proxy for credibility (H. Lee and Oh 2017). This dynamic extends to information-sharing behaviors (Yoo et al. 2014). In informational influence, users share what they perceive as credible or useful information, often relying on informational cues from perceived experts, opinion leaders, or popular figures (Cialdini and Goldstein 2004; Sherif 1936).

Research shows that on Twitter, many users primarily use the platform to see “what others are saying”, indicating the importance of observing social norms (McClain et al. 2021, 9). As a result, tweets with higher retweet volumes are perceived as more influential, reinforcing their propagation (Kapidzic et al. 2022). In light of previous literature on social conformity, and informational and social normative influence, the following hypothesis was proposed:

2.3 Influencers and networked social influence

Social media influencers, distinct from traditional opinion leaders who derive authority from institutional credentials (Rogers and Cartano 1962), leverage platform-specific strategies like curated self-presentation and parasocial engagement to shape public discourse (Archer et al. 2021; Kamiński et al. 2021). In the context of health topics, influencers range from medically trained communicators to celebrities discussing health. Those who incorporate factual, evidence-based messaging can exert strong informational influence, while those acting as “curators” of norms may reinforce social norms around behaviors by weaving health into personal narratives (Thorson and Wells 2016). Bridging scientific information and public audiences, influencers shape perceptions of both credibility and acceptability through the reinforcement of communal values (Pöyry et al. 2022).

Network centrality further amplifies influencers’ impact. They often occupy the high-centrality roles: an influencer with many followers (high out-degree) or one who bridges multiple interest groups (high betweenness) can disseminate content widely across the network (Rodrigues 2019). Because of these positions, influencers are likely to amplify both types of social influence, as they control information flow and decision-making in fragmented networks (Hanneman and Riddle 2005; J. Lee et al. 2013). Research has shown that users with many outgoing links indeed have more power to spread information (Mochalova and Nanopoulos 2013), and highly connected “third-party” accounts can connect disparate Twitter communities (Enke and Borchers 2019). Retweets, often interpreted as signals of credibility, reflect this dynamic, where users disproportionately share concise, factual content over opinion-based narratives (Saito et al. 2015).

Given the substantial body of previous literature highlighting the powerful impact of influencers in disseminating (or promoting) scientific information on social media, it was posited that influencers possess stronger weak-tie power capabilities, and consequently, exert a heightened degree of social influence upon other Twitter users. Considering their unique capacity to merge informational authority with networked reach, the following hypotheses were formulated:

Figure 1:

User, tweet type, and social influence on Twitter.

3.1 Data collection

A Python script utilizing the “Requests” library (Reitz et al. 2014) collected a sample of the Twitter stream posts around the COVID-19 vaccine via the Twitter search academic Application Programming Interfaces (API) v2. The Twitter API enabled the extraction of tweets that match search criteria such as keywords, hashtags, locations, and named places, among others. To gather relevant tweets on the public’s attitudes toward the COVID-19 vaccination, hashtags such as “pfizervaccine”, “modernavaccine”, “Johnson&Johnsonvaccine”, and “COVID-19 vaccine” were used as search terms in the crawler. A total of 5,554,372 tweets were collected from November 3, 2021, to May 5, 2022.

For the purpose of this study, the analysis focused exclusively on retweets rather than quote tweets, original tweets, or replies. Retweets provide a direct and quantitative indicator of social influence, as they reflect a user’s decision to endorse, propagate or conform to a message shared by another user. Prior research shows that retweet counts serve as a visible social cue that signals both informational and normative social influence (Kapidzic et al. 2022). Non-English tweets were removed using the Python package “Langdetect” (Rodrigues 2019), resulting in a final sample of 4,625,086 English retweets from users worldwide for analysis. Users’ locations were not filtered, as many did not provide this information.

3.2 Data analysis

First, Latent Dirichlet Allocation (LDA), a three-level hierarchical Bayesian unsupervised machine learning model, was utilized to uncover the hidden thematic structures – main topics – within the sampled tweets discussing COVID-19 vaccines. Each topic was modeled as an infinite mixture of underlying topic probabilities (Blei et al. 2003). The processing involved processing steps such as stop word removal, lemmatization, and tokenization. A dictionary was then created to catalog words frequencies, and text data were converted into a document-term matrix. The LdaModel from the Gensim Python package was selected for accuracy in identifying ten distinct topics (Tijare and Rani 2020).

In the next step, to identify influencers and measure users’ capacity for social influence, network analysis was executed using the igraph library in Python (Hermida 2010; Himelboim 2017; Kolaczyk and Csárdi 2014). Network analysis is a valuable tool for identifying influencers in large, loosely organized groups of users and verifying truthful and reliable information in social network communication (Hermida 2010; Himelboim 2017). For this study, network analysis was executed to measure users’ index of influence (degree of centrality and betweenness centrality), which are widely recognized measurements for social networks (J. Zhang and Luo 2017). Degree centrality counts a node’s neighbors in a network, to measure individuals’ capacity to connect with a network (Gupta et al. 2023; Hansen et al. 2020). No weights were added in the index of influence computation.

The top 1,000 Twitter users with the top in-degree, out-degree, and betweenness centrality were identified as influencers and categorized into three categories: “authoritarians”, who received the most attention and gained popularity; “accelerators”, who provided accelerating weak-tie power through likes, retweets, and comments (based on the degree of centrality); and “connectors”, who acted as connective communicative tissues and provided bridging weak-tie power (based on betweenness centrality) (Nisbet and Kotcher 2009; Saito et al. 2015). The descriptive analysis of the 94 sample prominent influencers revealed that half of the sample are news media users, while public health and government accounts made up 36.17 %, and personal accounts 8.51 %. These influencers were highly connected, averaging 2,506,147 followers and following 3,103 others, which suggests their central role in vaccine-related discussions (Appendix: Tables 1 and 2).

To evaluate the weak-tie power, two centrality metrics, out-degree centrality, and betweenness centrality, were calculated for the 1,728,543 Twitter users, with average values of 3.88 and 641,557.36 respectively. Out-degree centrality captures the extent to which a user actively initiates connections and disseminates information, indicating greater accelerating weak-tie power because such users can rapidly broadcast information to diverse audiences. Betweenness centrality, on the other hand, identifies users who act as bridges between communities, facilitating communication and information exchange across different groups. This bridging role enables novel information to flow across network clusters and overcome echo chambers (Granovetter 1973). Together, these metrics provide a robust measure of weak-tie power in social networks and these values served as criteria for evaluating ties or connections between users. Among these users, 320,374 (19 %) had above-average out-degree centrality, indicating significant accelerating weak-tie power, while 1,408,169 (81 %) had below-average values. Additionally, 54,254 (3 %) users had above-average betweenness centrality, signifying strong bridging weak-tie power, whereas 1,674,289 (97 %) had below-average values.

Finally, sentiment analysis using Natural Language Processing (NLP) was executed through crowd coding, a method shown to be as reliable as trained coders but more cost-effective, transparent, and replicable (van Atteveldt et al. 2021). This analysis determined the prevalence of informational versus opinion-based content in each tweet by evaluating the text of all tweets.

Sentiment analysis of all tweets was performed using the TextBlob Python library, which offers a standardized API for various NLP tasks, including tagging, classification, and sentiment analysis (Loria 2021). TextBlob utilizes a Naïve Bayes classifier trained with unigram features (Araque et al. 2017; Loria 2021), to calculate sentiment scores, specifically the subjectivity score, using the pattern library (Giatsoglou et al. 2017). Subjectivity quantifies the amount of personal opinion and factual information contained in the text. The subjectivity score ranges from 0 to 1, where 0 denotes high objectivity and 1 denotes high subjectivity. The higher subjectivity means that the text contains personal opinion rather than factual information. For example, both ‘great’ and ‘not great’ have a subjectivity score of 0.75. This score reflects the balance of personal opinion (expressing of opinions, feelings, or subjective judgments) versus factual information in a tweet, with higher values indicating greater subjectivity (opinions) and lower values indicating greater objectivity (factual information) (Shah 2020).

To evaluate the reliability of TextBlob’s sentiment scoring, an intercoder reliability test was conducted using the Pattern sentiment library (Rodríguez López and de Jesús Hoyos Rivera 2019) as an independent automated coder. A random 10 % sample of the dataset (n = 462,288 tweets) was analyzed separately with Pattern, and the sentiment outputs (polarity and subjectivity) from the two tools were compared. The agreement between TextBlob and Pattern was assessed using Cronbach’s alpha, yielding α = 0.70 for both polarity and subjectivity. These values demonstrate a high level of consistency between the two sentiment analysis tools, supporting the robustness of TextBlob’s measures for large-scale analysis.

5.1 Theoretical implications

This study reinforces Granovetter’s (1973) theory of weak ties as conduits for rapid information diffusion. Weak ties, characterized by a loosely connected network structure with minimal social overlap, prioritize content that is factual, concise, and broadly applicable over nuanced debates. This structure inherently favors updates such as vaccine distribution statistics, highlighting the tension between reach (broad dissemination) and depth (in-depth discussion) in digital ecosystems. Contextual trust-building, for which social conformity is relied on, is inhibited within Twitter’s virtual communities. This underscores Twitter’s role as a megaphone for institutional messaging, amplifying authoritative sources while marginalizing subjective perspectives. These findings extend Granovetter’s (1973) framework to health communication on social media, illustrating how weak-tie networks, particularly bridging weak-tie power, facilitate public health messaging during crises while limiting opportunities for norm curation.

The study advances the theoretical understanding of social influence by demonstrating that accelerating and bridging weak-tie power are stronger predictors of retweetability than follower count or tweet volume. This directly challenges Suh et al.’s (2010) study, which emphasizes follower metrics, indicating a positive association between a user’s follower and following counts and retweetability. Instead, this current study resonated with Cha et al. (2010) assertion that network structure and tie strength, not the number of followers, drive engagement.

In addition, the findings’ alignment with Liu and Zheng’s (2024) conclusion that the informative value of content critically shapes parasocial relationships between influencers and followers, highlights a key mechanism driving user engagement on Twitter. Influencers who disseminate evidence-based updates enhance trust and credibility, driving retweetability. This dynamic further distinguishes social media influencers from traditional opinion leaders, as their influence stems not from institutional authority but from digital attributes such as platform-specific social ties and curated relatability (Cheung et al. 2022; Kim and Kim 2022). For instance, influencers leveraging weak-tie connections can transcend traditional hierarchies, embracing Katz’s (1957) assertion that influence in fragmented networks depends more on “whom one knows” (weak-tie reach) than “who one is” (institutional authority).

The limited retweetability of opinion-based posts suggests that normative influence (conforming to shared group values) is constrained by the platform’s structure. Unlike long-form platforms (e.g., YouTube), where community interaction fosters norm curation, Twitter’s design favors central-route processing of factual information (Cho et al. 2024). For instance, connectors drove over half of retweets by disseminating factual updates, while authoritarians (e.g., institutional accounts) had limited reach. This supports Zhang et al.’s (2017) finding that ordinary active users, not traditional elites, often wield the greatest influence in digital networks. As a result, Twitter’s network design constraints limit the potential for subjective narratives to shape collective behavior (McClain et al. 2021) and heighten engagement during the COVID-19 pandemic (Samet 2020). The minimal impact of opinion-based content further underscores Twitter’s role as a real-time information hub rather than a space for norm formation. This reinforces the idea that follower counts signal popularity but not credibility, reinforcing that conformity on Twitter is driven more by informational influence (adherence to credible sources) than by normative influence (adherence to group values).

5.2 Practical implications

The findings of this study offer actionable strategies for public health communicators. Social conformity models must account for platform-driven dissociation of informational and normative influence. Health campaigns on Twitter should prioritize concise, evidence-based data messaging (but not opinions) tailored to the platform’s rapid diffusion dynamics. Moreover, the non-significant role of follower count suggests that investing in accounts solely for their large audiences may be inefficient. Instead, strategies should focus on identifying users with high weak-tie bridging or accelerating power, even if they have modest followings. Accelerators, often science communicators or local healthcare workers, can broadcast information to broad audiences, while connectors, such as community leaders, help bridge polarized subgroups, ensuring messages reach ideologically diverse communities. Authoritarians like institutional accounts may rely on their credibility to propagate accurate information, but require collaboration with connectors to penetrate fragmented audiences. Furthermore, public health campaigns should adopt a cross-platform strategy, reserving nuanced, emotion-driven narratives for platforms like YouTube, where long-form content fosters communal validation and norm curation.

5.3 Limitations

While this study provides critical insights into Twitter’s role in health communication, several limitations must be acknowledged. First, the exclusion of retweets from suspended or deleted accounts may introduce bias, potentially overrepresenting persistent institutional voices while undercounting transient or dissenting users. Second, reliance on automated natural language processing (NLP) for sentiment analysis risks misclassifying nuanced content. Finally, the dynamics observed may not fully generalize to non-crisis health topics or less polarized issues.

5.4 Future research

Future studies should expand on these findings by exploring how platform-specific affordances shape health communication across diverse contexts. Comparative analyses of weak-tie dynamics on video-centric platforms (e.g., TikTok) versus closed networks (e.g., WhatsApp groups) could reveal how content format and network structure interact to drive engagement. Additionally, investigating the mediating role of issue valence such as contentious topics versus neutral ones, could clarify how controversy amplifies or diminishes weak-tie influence. Longitudinal research tracking shifts in influencer roles during public health crises versus routine advocacy periods would further illuminate the stability of network-driven leadership. Finally, integrating mixed-method approaches (e.g., NLP with qualitative interviews) would deepen understanding of how users interpret and act on health information in fragmented digital ecosystems.