Russian Twitter Accounts and the Partisan Polarization of Vaccine Discourse, 2015–2017

Abstract

Objectives. To understand how Twitter accounts operated by the Russian Internet Research Agency (IRA) discussed vaccines to increase the credibility of their manufactured personas.

Methods. We analyzed 2.82 million tweets published by 2689 IRA accounts between 2015 and 2017. Combining unsupervised machine learning and network analysis to identify “thematic personas” (i.e., accounts that consistently share the same topics), we analyzed the ways in which each discussed vaccines.

Results. We found differences in volume and valence of vaccine-related tweets among 9 thematic personas. Pro-Trump personas were more likely to express antivaccine sentiment. Anti-Trump personas expressed support for vaccination. Others offered a balanced valence, talked about vaccines neutrally, or did not tweet about vaccines.

Conclusions. IRA-operated accounts discussed vaccines in manners consistent with fabricated US identities.

Public Health Implications. IRA accounts discussed vaccines online in ways that evoked political identities. This could exacerbate recently emerging partisan gaps relating to vaccine misinformation, as differently valenced messages were targeted at different segments of the US public. These sophisticated targeting efforts, if repeated and increased in reach, could reduce vaccination rates and magnify health disparities.

In this study, we examined the ways in which the personas underlying Twitter accounts operated by the Russian Internet Research Agency (IRA) discussed vaccines between 2015 and 2017. We build on research demonstrating the IRA’s efforts to sow public discord around vaccines,¹ as well as preliminary qualitative analysis of their general Twitter activity.² We demonstrate how anti- and provaccination IRA messages were used strategically to enhance the credibility of manufactured IRA “personas” in ways that could benefit the organization’s attempts to intervene in subsequent political discussions including those surrounding the 2016 US presidential election. As a result, such activity can increase the partisan polarization around vaccines.

Using unsupervised machine learning^3,4 and network analysis,^5,6 we modeled IRA Twitter personas as thematic communities,² groups of users that share the same thematic characteristics (discuss similar topics) and examined the ways in which their distinctive messaging about vaccination served each persona type, with a focus on pro-Trump and anti-Trump personas. We argue that vaccine-related content carried by specific Twitter personas has the potential to bolster partisan polarization around vaccines, resulting in a spiraling effect on vaccine hesitancy among specific US subpopulations.

Since mid-2014, the Russian government has operated a sophisticated online network of social media accounts to sow discord in the US political system and, before his election, to bolster the presidential candidacy of Donald J. Trump. This activity was conducted by the IRA, a Russian company based in St Petersburg and linked to the Russian government.² While engaging in political issues that could directly influence the elections,^7,8 IRA accounts have also used social media platforms such as Twitter to discuss a wide range of nonpolitical topics,² from popular culture⁹ to genetic engineering,¹⁰ fracking,¹¹ and, starting in early 2015, the safety and efficacy of vaccines.¹ A previous study by Broniatowski et al., which analyzed 889 tweets by IRA accounts that used the hashtag #VaccinateUS, found that IRA accounts posted information both for and against vaccines.¹ They suggested that the IRA’s aim was to sow discord and deepen disagreements among people in the United States.

A later study by Linvill et al.² observed that IRA discussions of political and nonpolitical topics were neither homogenous nor random. Using a qualitative analysis of almost 4000 tweets in the month leading to the elections, they identified 4 distinct types of accounts, differing in thematic content. They labeled these “right trolls,” “left trolls,” “newsfeed,” and “hashtag gamers.” The majority of activity that they isolated focused on nonpolitical content to camouflage their political intention. In other words, nonpolitical messages were used to bolster the credibility of their allegedly US accounts, by adhering to what Linvill et al.² dubbed “personas.”

Like Linvill et al.,² we used the term persona not in its conventional usage,¹² as a mask or performance aimed at projecting an improved image of the self.¹³ Instead, we focused on what we call “thematic personas,” types of social media accounts that are consistent in their use of specific topics and discourses. These thematic masks are built, in part, by the use of specific language patterns native to and, hence, impersonating, a target group. For example, accounts that attempt to attract likely Trump supporters will be more likely to discuss topics that are (at least stereotypically) associated with his followers (e.g., the Second Amendment) or use terms such as MAGA (Make America Great Again). Some of these topics are strictly political, in the sense that they relate to policy (e.g., gun control), while others are nonpolitical but nonetheless associated with the culture of political American subpopulations,¹⁴ embodied in preferences for food, art, music, sports, hashtags, and, as we demonstrate here, health topics such as vaccines.

While important for our understanding of the IRA’s use of Twitter personas, the analysis by Linvill et al.² of 4 broad IRA personas is limited specifically in its ability to understand the discourse around topics such as vaccines, as the topic played a relatively small role in the general IRA activity (1968 tweets out of roughly 3 million) and therefore is not likely to be included in small samples used for qualitative analysis. Such analysis requires computer-assisted methods able to extract all mentions of vaccines and estimate their use among different personas. In addition, Linvill et al. only examined 1 month of IRA activity, out of more than 3 years online. Lastly, some groups of users might be too small to observe with a limited random sample of tweets used in qualitative analyses (e.g., our analysis identified an African American persona that was absent from previous work).

In this study, we harnessed the full data set of IRA activity over 3 years and used unsupervised machine learning and network analysis to identify 9 distinct thematic personas based on accounts’ language tendencies. We demonstrate that each persona discussed vaccines in ways that were congruent with the general purpose of the persona. Simply put, pro-Trump accounts talked about vaccines in ways that are different from anti-Trump ones, and the difference was congruent with recent tendencies of conservatives to oppose, and liberals to support, vaccines.¹⁵

This method allowed us to also extend the work of Broniatowski et al.¹ by contextualizing the IRA vaccine discourse they identified within the general IRA activity. We demonstrate how IRA accounts discussed vaccines not only to sow discord among people of the United States¹ but also to flesh out the personalities of their “American” accounts in a credible way. Based on our findings, we argue that, by selectively catering to people with different political orientations, IRA activity could result in an increase in partisan polarization around vaccines that, in turn, could affect vaccine hesitancy, especially among susceptible subpopulations of conservatives and African Americans.

We addressed 3 research questions. The first extends the work of Linvill et al.⁶ to the full data set of IRA activity while allowing the automated unsupervised algorithm to identify personas they did not find in their qualitative analysis. The second and third extend the work of Broniatowski et al.¹ by connecting the vaccine discourse to the IRA thematic personas:

• 1.What thematic personas did the IRA employ between 2015 and 2017?
• 2.Did these personas differ in the extent to which they engaged with the vaccine debate?
• 3.Did these personas differ in their attitudes regarding vaccines?

METHODS

Our data set included 2 827 928 English-language tweets from 2689 accounts identified by Twitter as operated by IRA,¹⁶ posted between January 1, 2015, and December 31, 2017, on Twitter.

We conducted analysis in 3 stages: (1) topic modeling of the textual data, (2) a novel thematic user network approach developed for this study to chart user thematic communities, and (3) manual coding of vaccine-related tweets for valence (neutral, provaccine, or antivaccine). Additional details regarding the method used such as model fit indicators and hyperparameter tuning can be found in the elaborated methodological report in Appendix A (available as a supplement to the online version of this article at http://www.ajph.org). For replication and transparency purposes, we provide the R script used for preprocessing, topic modeling, and networking via the researchers’ Github page (https://github.com/DrorWalt/AJPH2020). Data are publicly available at Twitter’s election integrity database.¹⁶

Topic Modeling

To address research question 1 and identify thematic personas among IRA accounts in an inductive way (i.e., not limited to the personas found by Linvill et al.²), we combined topic modeling and network analysis. Topic modeling is an unsupervised machine learning method for the analysis of textual data.³ Topic models are aimed at extracting a set of topics from which a corpus could be created.¹⁷ “Topics” are statistical entities, representing the probability that specific words will tend to belong to the same thematic unit based on the linguistic assumption of co-occurrence (i.e., words that tend to appear frequently in the same documents share thematic meaning). Every document is composed as a mixture of all topics.¹⁸

In this work, we used latent Dirichlet allocation and Gibbs sampling.¹⁷ Using 10-fold cross-validation and comparing the perplexity scores of multiple models, we chose to focus on an optimal model of 60 topics (α = 0.01). To interpret and label topics, we examined the words with the highest loading on each topic, the words that are both prevalent and exclusive to each topic (FREX words¹⁹), and the full tweets that were most representative of each topic. The top 20 unique words for each topic alongside their assigned labels can be found in Appendix B (available as a supplement to the online version of this article at http://www.ajph.org).

Network Analysis

For each IRA account, we calculated the amount of language associated with each topic. This allowed us to calculate the similarity between accounts in terms of discussed topics, using cosine similarity over the topic-user matrix. We used the topical similarity scores between accounts to create a thematic network, where each node represents a single account. Importantly, the network does not describe social connections (e.g., liking, retweeting, following) but rather a thematic similarity relationship, in which accounts that tend to post about similar topics are more strongly connected, and are spatially closer to each other. We used a community detection algorithm (Louvain²⁰) to segment our network to distinct thematic communities. Our algorithm identified 9 communities.

To understand the content posted by accounts employing different personas, 2 independent coders qualitatively content analyzed a sample of 100 tweets representing each community or persona (n = 900 tweets) and labeled them. These tweets can be found in Appendix C (available as a supplement to the online version of this article at http://www.ajph.org). In addition, the full code we provide can be combined with the publicly available data set provided by Twitter.

Volume and Valence

To address research question 2, for each community, we counted the share of vaccine-related content out of the total community activity. We used the string “vaccin” and identified 1968 tweets from 270 accounts in our whole sample. In addition, we also calculated the share of users that mentioned vaccines at least once, out of the community’s total number of users.

Lastly, to address research question 3 and assess the valence of tweets mentioning vaccines in each community, 2 independent coders manually analyzed whether the reference was antivaccine, provaccine, or neutral or irrelevant (Krippendorff’s α = 0.85) for a sample of 50 tweets per each of the 7 communities (n = 350) that mentioned vaccines at least once (as detailed in Results, 2 of the 9 communities never discussed vaccines).

RESULTS

We identified 60 topics and 9 thematic personas. Figure 1 presents the top words for each of the thematic personas.

FIGURE 1—

Frequency of 15 Top Words Found in the Analysis of the Russian Internet Research Agency Activity Related to Vaccines Between 2015 and 2017, for the 9 Thematic Communities: (a) Hard News, (b) Anti-Trump, (c) Pro-Trump, (d) Youth Talk and Celebrities, (e) African Americans and Black Lives Matter (BLM), (f) Mixed International Topics, (g) Ukraine, (h) Soft News, and (i) Retweets of Various Topics and Hashtag Games

Based on the qualitative open-ended analysis of tweets from each community, we labeled the 9 thematic communities:

• 1.hard news,
• 2.anti-Trump,
• 3.pro-Trump,
• 4.youth talk and celebrities,
• 5.African Americans and Black Lives Matter,
• 6.mixed international topics,
• 7.Ukraine,
• 8.soft news, and
• 9.retweets of various topics and hashtag games.

Figure 2a shows the results of this analysis, with each node representing an IRA account; the edges between nodes representing how often different accounts talked about similar topics; the colors representing membership in a thematic community, as identified by a Louvain algorithm; and the size of nodes representing reach in terms of total retweets of all their communication between 2015 and 2017.

FIGURE 2—

A Network Representation of 2681 Russian Internet Research Agency Accounts, 2015–2017

Note. BLM = Black Lives Matter; IRA = Internet Research Agency. Nodes represent the various IRA accounts. Edges represent topical similarity between accounts. Size of nodes indicates accounts’ reach as the sum of all retweets from that account during the research time frame. Color in Figure 2a indicates Louvain community membership. Color in Figure 2b indicates discussion of vaccines by the account (red indicates the account mentioned vaccines at least once between 2015 and 2017; black indicates no vaccine mentions).

After thematically mapping the different IRA personas (research question 1), we assessed whether accounts in each community ever talked about vaccines and, if so, how often (research question 2). Each node in Figure 2b represents an IRA account; the edges between nodes indicate how often different accounts talked about the same topics; red nodes are accounts that tweeted about vaccines at least once; and black nodes accounts that never did so in the period analyzed. As can be seen in Figure 2, vaccine discourse was limited to specific areas of the graph and was more common (in terms of number of accounts that ever talked about vaccines) in some personas than others.

Lastly, we analyzed a sample of 50 vaccine-related tweets for each community to determine the valence toward vaccines (pro-, anti-, or neutral). The volume and valence per persona can be seen in Table 1. We elaborate more on our findings in the following paragraphs.

TABLE 1—

Vaccine Activity and Label of Each Thematic Community Found in the Analysis of the Russian Internet Research Agency Activity Related to Vaccines Between 2015 and 2017

		Fractional Share Mentioning Vaccines
Community No.	Main Discourse	Users	Tweets	Vaccine Discourse Sentiment Distribution (% Against/% Neutral/% Pro)
1	Hard news	0.017	0.0004	24/57/19
2	Anti-Trump	0.017	0.006	10/52/38
3	Pro-Trump	0.17	0.0004	54/22/24
4	Youth talk and celebrities	NA	NA	NA
5	African American and BLM	0.11	0.0003	40/24/36
6	Mixed international	0.10	0.0009	42/14/44
7	Ukraine	NA	NA	NA
8	Soft news	0.22	0.001	6/78/16
9	Retweets and hashtag games	0.13	0.0002	32/22/46
All		0.08	0.0005	30/38/32

Note. BLM = Black Lives Matter; NA = not applicable.

The first persona type engaged mostly with links to hard news online updates (e.g., war, terrorism). Only 2% of these accounts ever talked about vaccines, and only 0.04% of all their tweets were about them. When talking about vaccines, these accounts tended to provide links to news articles without expressing their own opinion (60%).

We found a similar user ratio for the second persona, which focused on liberal and anti-Trump content. Here the rate of vaccine tweets was slightly higher (about 0.5%). Importantly, tweets seemed to be mostly provaccine (38%) or neutral (52%), reporting on developments in regulation, legislation, and life-saving operations globally or locally. Only 10% were coded antivaccine.

For the third persona, representing pro-Trump discourse (with accounts having relatively higher reach), many of the accounts using the persona talked about vaccines—17% of these accounts mentioned vaccines at least once. However, while many accounts under the persona talked about vaccines, vaccines accounted for only 0.04% of their tweets. On average, they expressed mostly antivaccine sentiment (54% of vaccine tweets). Thus, the anti-Trump persona was dedicated to supporting vaccines, while the pro-Trump persona opposed them.

Accounts belonging to the fourth persona immersed themselves in topics often associated with youths, such as celebrities (many of which are Russian-related such as Maria Sharapova or Alexander Nevsky). Tellingly, these alleged young users never talked about vaccines.

The fifth persona imitated African American users, both in topics (e.g., Black Lives Matter activism and African American celebrities such as Talib Kweli) and language (e.g., stereotypical use of the n-word). While less negative toward vaccines then the “pro-Trump” persona, this persona exhibited a balance between antivaccine messages, directed mostly against corporations and the government (40%), and provaccine ones (36%). Like the pro-Trump persona, many accounts using this persona tweeted about vaccines at least once (11%), but this content accounted for only 0.03% of their total tweets. Another similarity to the pro-Trump persona was the existence of a relative high number of high-reach users—that is, users with large number of total retweets for their overall communication between 2015 and 2017.

The sixth persona focused on a mixture of international topics, including some conversation about the Russian–Ukrainian conflict and the 2011 Fukushima nuclear disaster. Ten percent of the accounts using this persona tweeted about vaccines at least once, accounting for 0.09% of their tweets. As was the case in the persona imitating the African American community, sentiment toward vaccines was balanced with 42% of tweets coded as antivaccine, 44% as provaccine, and 14% as neutral.

As the figures indicate, the seventh persona exhibited a unique thematic pattern. These accounts focused almost exclusively on Ukraine, highlighting the alleged injustice of the US stance on the conflict between Russia and Ukraine, the immorality of its people (e.g., Ukraine aiding Nazi brigades), and the failures of the Ukrainian government. This persona did not tweet about vaccines at all.

The eighth persona focused on soft news updates (e.g., health, sports, and local topics). Twenty-two percent of these accounts mentioned vaccines at least once, and roughly 1 out of every 1000 tweets created by this persona type was about vaccines. However, their vaccine tweets tended to be neutral (78%), with only 16% pro- and 6% antivaccine.

Finally, accounts using the ninth persona did not focus on 1 specific topic, instead focusing their activity on retweeting others and on engagement with trendy hashtag games (viral online prompts encouraging users to create content around a specific premise or topic—for example, prompts promoted by the popular Comedy Central show @midnight). Their vaccine-related content (13% of users but only 2 in every 10 000 tweets) was more pro- (46%), than anti- (32%) vaccine, and extensively used the #VaccinateUS hashtag, a finding consistent with that of Broniatowski et al.¹

DISCUSSION

Using the full data set of more than 3 million tweets published by IRA accounts on Twitter between 2015 and 2017, we show that the vaccine discourse that was first identified by Broniatowski et al.¹ was part of a larger effort to solidify and bolster identifiable thematic personas, probably for a later political use including during the 2016 presidential elections. Our automated analysis identified 9 personas, compared with the 4 found in qualitative research.² We were able to identify the additional personas because we were working with a higher resolution and larger sample.

Vaccines as Part of Personas’ Discourse

Our method identified 9 thematic personas, operationalized through the consistent preference of specific topics and language. This linguistic behavior, we believe, was an attempt to project coherent and reliable personas that do not focus exclusively on the elections and politics (thus avoiding being tagged as suspicious by their Twitter followers).

None of the personas dedicated itself exclusively to vaccines, nor to health topics in general. Two were focused on soft and hard news, 2 focused on supporting or opposing presidential candidate Trump, 1 mimicked African Americans with a focus on Black Lives Matter activism, and 2 were mixed in topics, focusing on retweets and the language and interest of younger users. Two personas focused on international issues, 1 focusing solely on the conflict between Russia and Ukraine, and 1 focusing on a larger number of international issues including the Russian–Ukrainian conflict and the 2011 Fukushima nuclear disaster.

Importantly, in line with our argument that different accounts disguised themselves as different personas with different opinions and interests, some personas tweeted about vaccines, while others did not. Among those who did, we identified persona-tied differences in intensity and centrality of the vaccine issue, as well as differences in the valence of opinions about vaccines. This finding contextualizes previous findings offered by Broniatowski et al.¹ Of particular importance for public health, the pro-Trump personas tended to oppose vaccines, while the anti-Trump ones did not. For example, an account associated with the pro-Trump persona, the supposed conservative Christian @ameliebaldwin, wrote on November 12, 2016, that “Holistic doctors found #autism-causing carcinogens in #vaccines before being murdered.” On the contrary, the allegedly African American user @imissobama wrote on January 13, 2017, that “The anti-vax movement can only exist bc few living Americans can recall what polio actually did to ppl. I fear the same is true of fascism.”

The personas focused on sharing soft and hard news reports tweeted about vaccines relatively often, but their tweets did not take a position on the issue. Rather, they tended to merely retweet news stories on vaccines, just as they did for other non–vaccine-related issues. Finally, personas focusing on youth issues and Ukraine did not post any tweets about vaccines. Thus, different accounts, aimed at different subpopulations of the US public, shared dramatically different information regarding vaccines.

While beyond the scope of this study, attention should be directed in the future to the dynamics of persona use over time. Our data indicate that a higher level of vaccine-related IRA tweets appeared around the vote on a California bill on mandatory school vaccinations (SB277). Future studies are needed to better understand the strategic use of vaccines and other health topics over time, with a focus on whether it changes around events tied to citizen voting.

In summary, our findings indicate that, in addition to sowing discord, vaccine-related IRA Twitter activity functioned to fashion believable personas. Such behavior could have detrimental effects on public health if targeted to vulnerable vaccine-hesitant communities.

Public Health Implications

Regardless of the motivation behind the Russians’ involvement with vaccine discourse online, this behavior could threaten US public health, especially when targeted at specific users. Historically, the antivaccine movement and the individuals supporting it were not associated with one political party or the other.¹⁶ However, like other topics, political polarization can affect attitudes toward vaccination as well. It is worrisome that recent polls and surveys indicated that individuals who identify as conservative have begun shifting against vaccination.¹⁶

The partisan polarization of public health issues is cause for concern. Attitudes toward and intentions to follow public health recommendations should be formed as the result of a consideration of risks and benefits, based on scientific, nonpartisan inquiry. The increase in partisan polarization could result in an increase in the number of conservatives who make their judgment of vaccines’ safety and necessity not based on the science but based on their political disposition and their perception of how the topic aligns with their support for a political leader. Such a phenomenon could be further fueled by misinformation that distorts credible scientific evidence,²¹ especially in social media and online sources where antivaccination messages are more common than in mainstream media. This, in turn, could increase vaccination hesitancy.

Even if small in magnitude, the intentional Russian spread of antivaccine discourse targeted at specific subpopulations that are susceptible to it (e.g., pro-Trump users and African Americans on Twitter) could be the beginning of a new front in the ongoing informational cyberwar.