The anti-vaccination infodemic on social media: A behavioral analysis

Abstract

Vaccinations are without doubt one of the greatest achievements of modern medicine, and there is hope that they can constitute a solution to halt the ongoing COVID-19 pandemic. However, the anti-vaccination movement is currently on the rise, spreading online misinformation about vaccine safety and causing a worrying reduction in vaccination rates worldwide. In this historical time, it is imperative to understand the reasons of vaccine hesitancy, and to find effective strategies to dismantle the rhetoric of anti-vaccination supporters. For this reason, we analyzed the behavior of anti-vaccination supporters on the platform Twitter. Here we identify that anti-vaccination supporters, in comparison with pro-vaccination supporters, share conspiracy theories and make use of emotional language. We demonstrate that anti-vaccination supporters are more engaged in discussions on Twitter and share their contents from a pull of strong influencers. We show that the movement's success relies on a strong sense of community, based on the contents produced by a small fraction of profiles, with the community at large serving as a sounding board for anti-vaccination discourse to circulate online. Our data demonstrate that Donald Trump, before his profile was suspended, was the main driver of vaccine misinformation on Twitter. Based on these results, we welcome policies that aim at halting the circulation of false information about vaccines by targeting the anti-vaccination community on Twitter. Based on our data, we also propose solutions to improve the communication strategy of health organizations and build a community of engaged influencers that support the dissemination of scientific insights, including issues related to vaccines and their safety.

Fig 1. Anti-vaccination supporters are more engaged on Twitter.

We analyzed the behavior of three different groups: control (grey), anti-vaccination (red) and pro-vaccination (blue) (A). We calculated the number of tweets, replies and retweets per month (B). The anti-vaccination group scored the highest number of total Twitter actions (the sum of tweets, replies and retweets) per month (C). Anti-vaccination supporters tweeted less than control and pro-vaccination individuals (D), but they engaged in more discussion via an increased number of replies (E) and Retweets (F). Ordinary one-way ANOVA; **p<0.01; ****p<0.0001; Outliers were excluded with ROUT, Q = 0.1%; n = 50.

Fig 2. Anti-vaccination supporters are active science and vaccine communicators, share conspiracy theories and emotional content.

Both anti- (red) and pro-vaccination profiles (blue) share a larger number of science- and vaccine-related content per month, when compared with control profiles (grey) (A, B). We calculated the number of science- and vaccines-related content (tweets and retweets) published in the 24 hours before data analysis and normalized it for the total number of tweets published on average during a single day. 100 percent indicates that all generated contents are estimated to be science- or vaccines-related (A’, B’). Natural fluctuations above 100 percent are due to a variable Twitter activity during the 24 hours prior to data analysis compared to an average day. Anti-vaccination supporters publish conspiracy theories, whereas control and pro-vaccination individuals do not publish this type of material (C, C’). Anti-vaccination supporters share a larger number of tweets and retweets with emotional contents (and with emotional language) compared with the pro-vaccination and control groups (D, D’). Ordinary one-way ANOVA; ****p<0.0001; Outliers were excluded with ROUT, Q = 0.1%; n = 50.

Fig 3. The anti-vaccination group utilizes emotional language, but this does not determine the success of their tweets (engagement).

Most used words on Twitter by the anti- (red) and pro-vaccination groups (blue) normalized against the words predominantly used by the control-group (grey). Asterisks* indicate that words have been clustered (e.g. “vaccine” and “vaccines” are scored as a single word). n(profiles analyzed) = 42. Max = 18 indicates that particular word is used 18-times more in that specific group, when compared with the control. (A, A’). Most used words by anti- and pro-vaccination profiles normalized against each other. Asterisks* indicate clustered words. n(profiles analyzed) = 42 (B). Words are clustered for topic and normalized, with the value of 1 being assigned to the group utilizing the cluster of words the most. The most relevant clusters are shown. Words related to politics are greatly enriched in the anti-vaccination group; words related to health and medicine are predominantly used by pro- and anti-vaccination profiles, when compared with the control; phrasal words are underrepresented in the pro-vaccination group. Asterisks* indicate clustered words. (C). For the anti-vaccination group, the normalized number of emotional contents (relative to the total number of contents generated by a given profile) does not correlate with the number of engagements received on average for a single tweet (R² = 1.293*10⁻⁶; p = 0.99); n = 50 (D). Conversely, pro-vaccination profiles tweeting emotional content produce more engaging contents (R² = 0.2378; p = 0.003); n = 50 (D’).

Fig 4. Pro-vaccination profiles have more followers and produce more engaging content.

Pro-vaccination profiles (blue) generate more engagement in one day when compared with the control (grey) and anti-vaccination groups (red) (A), and normalization shows they produce more engaging content irrespectively of the number of contents generated in a given day (B). Pro-vaccination profiles have a larger number of followers when compared with the control and anti-vaccination groups (C). Hypothetical model to illustrate the results described so far. Anti-vaccination supporters are more engaged on Twitter, as they retweet contents more often than control and pro-vaccination profiles. They also share emotional content, although they generally produce less engaging content than their pro-vaccination counterparts. Despite the use of emotions as a tool to convey their message, given the lower engagement of anti-vaccination tweets, we hypothesized that a sense of community driven by common interest is key for the success of the anti-vaccination movement online (D). Ordinary one-way ANOVA; ***p<0.001; ****p<0.0001; Outliers were excluded with ROUT, Q = 0.1%; n = 50.

Fig 5. Anti-vaccination profiles establish a well-connected community sharing contents produced by a pull of influencers, whose most prominent exponent is Donald Trump.

The pro-vaccination Twitter web (A). Close up of the most relevant portion of the pro-vaccination web, which highlights the World Health Organization as the main influencer for the pro-vaccination group (A’). The anti-vaccination Twitter web (B). Close up of the most relevant portion of the anti-vaccination web, which highlights Donald Trump, his political entourage and public figures supporting his presidency as the main influencers for the anti-vaccination group (B’). The pro-vaccination and anti-vaccination Twitter webs are scaled 1:1 (A, B). For better readability, close up representations of the pro- and anti-vaccination webs are not equally scaled. Yellow color represents Twitter profiles (nodes) with 2 to 4 anti-vaccination profiles preferentially retweeting their contents within the top 10 most retweeted users (edges; 2≤ E ≤4; n = 42). Orange nodes represent profiles with 5 to 9 edges (5≤ E ≤9; n = 42), whereas red nodes indicate profiles with more than 10 connecting edges (E ≥10; n = 42). Size of the nodes is linearly scaled depending on the number of edges connecting the node (A-B’). The average number of neighbors in the web, the clustering coefficient, the density of the network and the characteristic path length of the anti-vaccination (red) web is greater than the pro-vaccination counterpart (blue) (C). Graphical representation and web parameters were generated with Cytoscape. Graphical representation of the main influencers in the pro- and anti-vaccination Twitter webs (threshold: E ≥5; n = 42). The size of the name tag assigned to the Twitter profile are linearly scaled for the number of edges. The Pro-vaccination influencer cloud only contains one profile (World Health Organization) (D), whereas the anti-Vaccination cloud contains 14 profiles, with former US President Donald Trump being the largest influencer (E).