To create a set of keywords that indicate opposition to vaccines, we used a snowballing sampling technique similar to that of DeVerna et al [11]. We started from a small set of manually curated keywords used exclusively in the context of strong vaccine hesitancy that appear on Twitter, such as #vaccineskill or #vaccinedamage. Using the Twitter Streaming API and the set of seed keywords, we collected the data for one day (October 18, 2020), after which we extracted other keywords that co-occurred with the seed keywords. We added the newly collected keywords to the list of seed keywords, checking them manually for relevance. We then repeated this step several times until we exhausted all the significant co-occurrences and narrowed our selection to approximately 60 keywords. The Twitter API can be queried with a substring of a longer keyword, and it will return the tweets that contain the substring. For example, the keyword novaccine will return the tweets that contain novaccineforme. We attempted to retain only the most informative and relevant stem words to capture most vaccine-related tweets and to avoid collecting less relevant tweets. The list of all keywords used to collect the streaming collection is listed in Table 1.

First, we identified a randomly sampled set of approximately 70,000 accounts that appeared in the streaming collection and that engaged in antivaccine rhetoric between October and December 2020, either by tweeting some of the tracked keywords or by retweeting tweets that contained some of the tracked keywords. Then, for those accounts, we collected their historical tweets using the Twitter API. By leveraging Twitter's Academic Research product track, we were able to access the full archival search and overcome the limit of 3200 historical tweets of the standard API. In this way, we collected almost all the historical tweets of the most queried accounts.

Our collection relies upon publicly available data in accordance with the Content Redistribution clause under Twitter’s Developer Agreement and Policy [12]. We released the data set with the stipulation that those who use it must comply with Twitter’s Terms and Conditions. The complete data set is publicly available on a GitHub repository and is accessible on the web [15].

We calculate the political leaning of each account by measuring the political bias of the media outlets it shared. We use a methodology proposed in prior work [16-18], and we identified a set of 90 prominent media outlets and accounts that appeared on Twitter. Each of these outlets and their associated Twitter accounts were placed on a political spectrum (left, lean left, center, lean right, right) per ratings provided by the nonpartisan service AllSides [19]. For each account in the data set, we maintained a record of all retweets and the original tweets that contained a domain name affiliated with the selected media outlets. The political bias of each account was calculated as the average political bias of all media outlets it shared content from.

We leveraged urllib, the Python URL handling module, to parse the URLs found in the data set. Each URL was broken into several components, including the addressing scheme, network location, and path. A third-party data set that contains the domains associated with websites that share misinformation was used as a ground truth to tag the domain names [20]. For URLs that were not in the data set, we queried the Media Bias/Fact Check website [21] for further identification. Because URL shortening services such as Bitly [22] are widely used on Twitter, shortened URLs appeared frequently. We used urlExpander [23] to expand the shortened URLs and retrieve the full URLs where possible. Domain names of popular news aggregators and social networks such as Twitter, Facebook, Instagram, Periscope, and YouTube were ignored in the analysis.

To infer a tweet’s geolocation, we used the information of the self-reported location of the account and matched it to a corresponding state in the United States. To calculate the average activity level per population, the absolute number of Tweets was normalized by the 2010 Census-reported population of that state as follows: I = N_i/P_i × 1,000,000, where N_i is the number of tweets originating in state i and P_i is that state’s population in 2010. This normalization provided information on the average number of collected tweets per million inhabitants. Note that we did not generate the geolocation map for the account collection, as it contains a relatively small number of accounts with self-reported locations.

A topic network was constructed to analyze the co-occurrence of hashtags in the streaming data set. Each node in the graph represented a hashtag, and an edge was added if two hashtags occurred in the same tweet. The node size was proportional to its degree of centrality, and the edge weight was the number of times two hashtags appeared together. For better visualization, nodes with fewer than 25 neighbors were ignored. To investigate the community structure of the network, we used the Louvain algorithm [24] on the topic network, which provided further insights about the links between antivaccine topics.

The streaming collection consists of 1.8 million tweets created by 719,000 unique accounts between October 18, 2020, and April 21, 2021. As shown in Figure 1, the number of relevant tweets in the streaming collection gradually increases from the start date. The chatter is relatively stable, with small spikes that do not often correspond to major announcements regarding vaccine research or vaccine authorization. We find this surprising, as the news usually drives the discussion on Twitter. Additionally, we observed a large spike in activity near the end of November 2020 that was not caused by any single event but rather by the increased activity of a small number of accounts.

The overwhelming majority of tweets originated from countries with predominantly English-speaking populations. Out of 1,832,333 tweets in the streaming collection, 1,245,986 (68%) originated in the United States, 229,041 (12.5%) in Great Britain, 100,778 (5.5%) in Canada, 21,987 (1.2%) in Ireland, and 20,155 (1.1%) in Australia; the rest of the tweets originated from other countries. In Figure 2, we show the geographical distribution of tweets in the United States. As expected, states with a large population, such as California, Texas, Florida, and New York, have more tweets in absolute terms (Figure 2, top). The number of tweets normalized by state population is depicted in Figure 2 (bottom), with the most tweets per capita originating from Hawaii, Alaska, and Maine, respectively.

Table 3 lists the top 15 most tweeted hashtags in the streaming collection. The count column represents the total number of times a hashtag appears, and the proportion column quantifies the proportion of tweets that contain a specific hashtag out of all tweets with any hashtag. Note that many tweets contain no hashtags, and many tweets with a hashtag contain more than one hashtag. In addition to the most common general hashtags that we expected to find, such as #vaccine and #covid19, we observed a high proportion of hashtags that carry strong antivaccine sentiment, such as #novaccineforme, #vaxxed and #vaccineinjury. For example, #novaccineforme can be found in more than 25,000 tweets, accounting for 6.6% of all tweets in the streaming collection that contain any hashtags. A large set of common hashtags is related to some debunked conspiracy theories that claim there is a global plot by rich individuals to reduce the world population, often expressed through hashtags such as #depopulation, #billgatesbioterrorist and #arrestbillgates. Another set of very frequent hashtags appears benign on the surface. Hashtags such as #learntherisk and #informedconsent appear to communicate genuine concerns about the safety of the vaccines; however, those hashtags are usually decoys and are very often used by the same accounts that strongly oppose vaccination and that otherwise often use more explicit antivaccine hashtags.

The account collection differs from the streaming collection, as it is focused on historical tweets from a set of accounts. The process of collecting the historical tweets is explained more in detail in the Methods section. The current account collection consists of more than 135 million tweets published by over 78,000 unique accounts, and it spans the period from March 3, 2007, to February 8, 2021. In Figure 3, we illustrate some of the most important statistics from this data collection. The left panel in Figure 3 shows the distribution of the number of tweets per account. Out of 78,954 accounts, 39,350 (49.8%) published fewer than 1500 tweets, 31,581 (40%) of the accounts have more than 2000 tweets, and 1184 (1.5%) have more than 5000 tweets. The right panel in Figure 3 shows the number of tweets over time. Most of the tweets originate in the year 2020, with the oldest tweet dating back to 2007. For 55,267 (70%) of the 78,954 accounts, the oldest collected tweet dates from 2020. There is a significant portion of accounts whose historical tweets date much earlier; for 14,211 (18%) of the 78,954 accounts, the earliest tweet was dated before 2018, and for 5368 (6.8%) of the accounts, the earliest tweet was dated before 2014. This relatively long-spanning collection of historical tweets at the account level may allow for a comprehensive temporal analysis of vaccine hesitancy development on Twitter over several years.

The 15 most common hashtags appearing in the account collection are displayed in Table 4. In addition to the common COVID-19–related hashtags, we observe many hashtags referring to US politics. During the period of the US 2020 presidential election and the political campaign, the accounts that appear in our collection were particularly active. Hence, we can see that many politically motivated narratives in the data originated during that period.

Vaccine hesitancy is usually fueled by misinformation originating from websites with questionable credibility. In Figure 4, we list the top 10 URLs that can be found in the streaming collection, and we illustrate the number of times each appears. The vast majority of those websites can be found in the Iffy+ database of low credibility sites [20]. One of the most commonly shared sources is the website of an American antivaccine group called Learn The Risk; it is known for its campaigns against vaccination, which assert that vaccines are responsible for a large number of deaths of young children. It is followed by Vaccine Impact, a well-known news and information website that promotes pseudoscience; this website often shares antivaccination propaganda and promotes alternative medicine, holism, and alternative nutrition. The only website on the list with high credibility is the website of the National Center for Biotechnology Information (NCBI), a PubMed parent company.

In Figure 5, we list the top 10 URLs that can be found in the account collection, and we illustrate the number of times each appears. Figure 5 shows that many far-right news media sites appear frequently in the account collection. The Gateway Pundit [25], which is known for publishing falsehoods, hoaxes, and conspiracy theories, occurs more than 400,000 times. Other far-right media outlets, such as Breitbart News [26] and the Epoch Times [27], also appear very often. Considering the sources that usually fall in the group of mainstream news media sites, such as Fox News [28] and the New York Post [29], conspiracy spreaders selectively quote reports from these sources to increase the credibility of often false claims.

In Figure 6, we show the distribution of political leanings of the accounts. The political leaning of an account was estimated based on its media diet (see the Methods section). The x-axis represents the account’s political leaning and can take any value between “far left” and “far right.” The y-axis is the normalized number of accounts with a corresponding political leaning. The political leaning of the accounts engaged in the antivaccination narratives is shown in orange. We observed a bimodal distribution with a significantly higher right peak. The blue bars illustrate the distribution of the political leanings for random Twitter accounts. The random Twitter accounts are a random sample of approximately 6000 accounts from the previously published Twitter data set related to the US 2020 Presidential election by Chen et al [30]. It has been previously shown that the Twitter users are younger on average and more likely to vote Democrat than the general public [31,32]. These results are not surprising, as they align with earlier studies showing that political orientation is a strong predictor of vaccine hesitancy in the United States [33,34].

To obtain further insights into the provided data set, we explored the clusters of antivaccine narratives by identifying the antivaccine topics that usually co-occurred. We ran the Louvain community detection algorithm on the topic co-occurrence network, as described in the Methods section. The topic network is illustrated in Figure 7. We identified 3 distinct communities; all of them contained antivaccine keywords, but with different focuses on topics. The largest topic community, colored purple, focuses on debunked claims around the conspiracy narrative that the vaccine is a plot by rich people to reduce the world population. The second topic community, colored orange, mostly focuses on vaccine safety, as hashtags such as #doctorsspeakup, #vaccinesafety, and #vaccineinjury appear often. The smallest topic community, in green, contains a mixture of various hashtags that range from strongly antivaccine, such as #informedconsent, #learntherisk, and #vaxxed, to some neutral hashtags, such as #vaccine, to some provaccine hashtags, such as #vaccineswork.

In this paper, we present a comprehensive data set consisting of tweets related to antivaccination narratives, organized in streaming and account collections. We characterized the data in several ways, including frequencies of prominent keywords, news sources, geographical location of the accounts, and political leaning of the accounts. The streaming collection consists of a random sample of tweets that contain any of the specific keywords promoting strong antivaccination sentiments. This is a common method used to collect Twitter data on vaccination hesitancy and other similar topics [35-42]. It is well understood by academics and is often used to provide useful insights about the chatter on the web about a particular topic in a specific period. The account collection was gathered using a relatively new method of collecting Twitter data by querying the historical activities from a set of tracked accounts. This collection was made possible after Twitter introduced the Academic Research product track API. In this way, by gathering massive amounts of historical tweets, researchers can characterize individual accounts rather than populations on average. This data set will be useful for scientists interested in the demographic and psychographic characteristics of Twitter users who are prone to spreading antivaccination narratives.

The news sources shared by the users in the streaming collection are predominantly websites with low credibility. However, the most shared URL is the website of the NCBI [25], which is part of the United States National Library of Medicine, a branch of the National Institutes of Health. NCBI houses PubMed, the largest bibliographic database for biomedical literature. This finding can create a false impression that the tweets from the streaming collection contain information from legitimate scientific sources. When we examined the context in which those papers were shared, we discovered that most of the papers from PubMed were cited with false and misleading conclusions. Sometimes, antivaccine advocates would share legitimate scientific papers documenting rare side effects of the vaccines, while overemphasizing the observed adverse effects and calling for vaccine boycotts. Sharing a scientific study in a tweet provides an illusion of credibility. Cherry-picking desirable sentences and relying on the fact that most of the audience will not make an effort to read a scientific paper in detail is a very effective strategy for manipulation.

It is often valuable to know the political affiliation of users who share antivaccine narratives. Knowing users’ position on a political spectrum can be useful in identifying their most likely moral values and possible stances toward specific societal issues. This knowledge can be used to design appropriate future messaging and campaigns. We were able to identify the political affiliation for the accounts collection, as we had enough tweets for each account. Accounts that share common misinformation related to vaccines often share other conspiracy narratives, usually politically charged ones. The population susceptible to such narratives strongly skews conservative [18]; therefore, we expected that a large number of accounts in the account collection would be right leaning.

Although the data sets give an overview of vaccine hesitancy on Twitter, potential limitations warrant some considerations. First, our streaming collection relies on a defined set of keywords. The antivaccine lingo is constantly evolving as the COVID-19 pandemic unfolds. Although we have made our best efforts to find the most representative keywords, they may not fully cover all antivaccine topics. The set of keywords we used was designed to capture the strongest antivaccine sentiments and may have missed various nuances in the multifaceted nature of vaccine hesitancy. Second, this data set should not be used to draw conclusions for the general population, as the Twitter user population is younger and more politically engaged than the general public [31]; this means that our data may be biased in various ways. Additionally, the keywords used for the collection were derived from the English vocabulary, highly biasing the geographical distribution of the tweets toward the English-speaking regions of the world. Finally, to prevent the spread of misleading COVID-19 information, Twitter has enacted specific rules and policies. The accounts violating these rules and policies may be banned by Twitter, making their tweets unreachable. At the time of writing, our estimate is that more than 40% of the accounts in the streaming collection and 30% of accounts in the accounts collection had been either banned or deleted. With each update of the streaming data set, we expect this proportion to change.

In addition to the streaming collection, which tracks tweets as they appear in real time, perhaps the most important contribution of this study is the account collection, a data set consisting of almost all historical tweets for a sample of users who were actively sharing antivaccination narratives. This data set can be used to provide further insights into the accounts that engage in antivaccine propaganda. Our intention in publishing this paper and data sets is to provide researchers with assets to enable further exploration of issues revolving around vaccine hesitancy and to study them through the lens of social media. The data sets collected and provided here could be useful for researchers interested in tracking the longitudinal characteristics of accounts engaging with antivaccine narratives. It can help provide better insights into the socioeconomic, political, and cultural determinants of vaccine hesitancy.

The data set is released in compliance with the Twitter Terms and Conditions and the Developer’s Agreement and Policies [12]. Researchers who wish to use this data set must agree to abide by the stipulations stated in the associated license and conform to Twitter’s policies and regulations.

The data are available at GitHub [15].