Social movements, framing, values, identity, and ideology.

Social movements expend considerable time constructing “versions of reality, developing alternative visions of that reality, attempting to affect various audiences’ interpretations, and managing the impressions people form about their movement” ([8], pg. 678). They seek to build credibility, persuade audiences to join the cause, and set the terms of political discourse about issues that matter to them [23]. Movements persuade audiences through framing activities. A frame is an interpretive schema used to help people understand a situation [24] by selectively “punctuating and encoding objects, situations, events, and experiences within one’s present or past environment” ([25], pg. 137). Related, framing is the process through which reality is constructed by movement actors [25–27] and gives issues or events significance [28, 29].

Social movements frames can mobilize people, increase the likelihood of buy-in to the cause when grievances are presented in a way that resonates with people’s values [30], and change behavior and opinion [3, 31]. While there is a significant body of scholarship on social movement framing, scholars [2] have advanced the concern that many studies disregard ideology as a distinct but important mechanism for movement buy-in, advancing the framed values and issues of a movement.

In political campaigns, ideologies filter frames by increasing resistance to disconfirming information [32] or providing reasoned principles to make sense of public debates [9]. For example, the “US healthcare system is killing people” frame may appeal more to a far left flank that supports distributive justice while “The US healthcare system is unjust” may appeal more to a moderate base [33]. Both ideological perspectives on health inequity are united by an overarching goal of improving social conditions and the healthcare system, but how to resolve health inequities and why it “should” matter to a populace will be filtered and framed differently through differing ideologies.

Differences in framed values are therefore symptomatic of differences in ideology. This is critical, as framed values are what is seen in the messaging produced by movement actors, e.g. in their tweets. Put another way, tweets are frames that communicate values that are shaped by ideological perspectives. We therefore study variation in ideological perspectives (e.g. moderate, progressive, or the far left) that shape framed grievances and/or values and identities by examining tweets.

In addition to using Twitter to express and frame values, movement actors also convey identity through the content they share [34]. Identity on Twitter is also known to be expressed on Twitter user profile descriptions, or bios [35]. While ideology and identity are analytically distinct, scholarship on American politics often conflates political identities and ideologies [3]. Recent work on ideology has shown, however, that ideological perspectives vary considerably across individuals with shared political identities [17], and this may be particularly true of politically active individuals [36]. Here, we seek to identify collective identities separately from ideology, and to then assess how these two overlap.

Social movements, social media, and “Bernie Bros”.

Scholars have long explored how social movements use, and mobilize on, social media (see, e.g., early work on the Arab Spring [37–39] and more comprehensive book-length treatments from Castells [40] and Jackson et al. [41]). Most relevant to our work, however, is scholarship that uses social media data to understand movements, rather than understanding how movements use social media. Specifically, we use Twitter data as a window into ideological variation among Bernie supporters. Others that have analyzed activity patterns of movements on Twitter have focused on the structure [42] and dynamics [43] of networks, the evolution of content from movements and associated counter-movements [44], the spatial [45] and demographic [46] variation of movements, and their relationship to offline discourse and activity [47]. More closely related is work exploring the emergence of particular collective identities over time in movements [48] and how framing strategies are appropriated on Twitter by movements [10]. We complement this prior literature by focusing explicitly on identity and ideological variation and how we can find such variation within supporters of a single movement.

More specifically, our analysis examines the online behaviors of the movement to secure Bernie Sanders’ bid for the Democratic nomination during the 2020 presidential primaries. First popularizing the term Bernie Bros during the 2016 presidential election cycle, Atlantic writer Robinson Meyer [49] observed that the Bernie Bro is someone who “seems to have taken Sanders’ rhetoric that America is trapped in a number of deep, unprecedented crises to heart and always writes with an urgent, anxious seriousness when discussing national politics.” In jest, Meyer characterized Bernie supporters as an obstinate and earnest group. But the “Bernie Bros” phrase gained traction and transformed into a trope that portrayed all Bernie supporters as antagonistic misogynistic millennial racist males. In 2016, media pundits used the phrase “Bernie Bros” to describe a group of people who refused to accept that Hillary Clinton won the Democratic Party’s nomination because the primaries were rigged [13], as well as disbelief that she won stemmed from deeply held misogyny [14]. Similarly in 2020, Bernie supporters were homogenized as Bernie Bros who refused to accept that Joe Biden won the nomination, as Bernie held a considerable lead over other Democratic candidates until Super Tuesday, signaling another rigged election and conspiratorial thinking among so-called Bernie Bros [50, 51].

There are ongoing debates about how much the Bernie Bros trope accurately represents Bernie’s supporters. Altamura and Oliver [17] present evidence to the contrary, showing that during the 2020 election cycle, Bernie supporters were more likely to be young, Hispanic, Independent, or non-religious compared to all other Democratic primary voters (pg. 644). They also find that younger cohorts were more likely to vote for Sanders especially if they supported government-sponsored health insurance and were low in racial resentment. At the same time, a litany of lived experiences present evidence that the “Bernie Bro” catechism accurately described behaviors of at least some Bernie supporters.

Our work compliments the empirical work from Altamura and Oliver [17]—to our knowledge, the only existing academic article to quantitatively analyze Bernie supporters—by shifting away from voting behaviors and examining the ideologies and identities of Bernie supporters online. In particular, we look to frequently retweeted content as a potent signal, noting that, “political elites and journalists construct issues in value terms as a heuristic for the broader public, which often understand social issues not in finite detail but in terms of the values they present” ([31], pg. 513)

Analyzing collective identities in Twitter data.

The literature presents a number of different approaches to identify groups of users with particular characteristics or identities. The most straightforward is to use manually crafted heuristics based on user behavior, e.g. using whether or not users follow a pre-specified list of politicians to label those users as Democrats or Republicans [52], or using the sharing of specific hashtags to identify members of social movements [44].

One important question is thus what behaviors are used for identification. In the present work, we focus on 1) retweets and 2) profile descriptions. Specifically, we use the retweet network to identify groups of users who commonly retweet the same set of accounts, and profile descriptions to interpret collective identities shared within those clusters. Despite user claims that retweets “are not endorsements,” survey data suggests that retweets predominantly serve to promote or re-up tweets [53]. Because of this, prior work has shown that retweets can readily identify distinct social groups [21], and that analyzing profile descriptions of these users can uncover shared identities within the group [18].

Regardless of what behavior is used, however, the use of manual heuristics has known limitations. One such limitation is that manual heuristics provide a high precision, low recall identification strategy. For example, individuals who follow multiple Republican politicians are likely to be Republicans, but few Republicans follow right-leaning politicians [54]. One way to deal with this issue is to apply supervised machine learning methods, where manual heuristics serve as the training data [52]. Here, we adopt a machine learning approach to avoid this problem of low recall, but use a method that also seeks to address a second limitation to manual heuristics: relying on a predefined heuristic potentially limits us to a subset of the relevant users. In our setting, for example, predefined heuristics to identify Bernie supporters might lead us to miss ideological diversity if our preconceived notions of Bernie supporters limits us to a subset of the overall population.

Specifically, we adopt the recently proposed approach of Zhang et al. [18], who developed a mixed methods paradigm to identify what they call “flocks” based on patterns in who retweets whom. Flocks consist of sets of users and the accounts they commonly retweet. Flocks are first identified via a computational technique known as Vintage Sparse PCA, described in more detail below. Each flock represents a set of influential users and the users that commonly retweeted them. Social identities collectively shared by the group can then be identified, if they exist, via analysis of user profile descriptions and tweets.

Analyzing values and their frames in Twitter data.

Social media researchers have developed a number of strategies to identify the values and frames expressed by social movement actors. Qualitative approaches are generally restricted to a small subset of the data, but provide nuanced insights. In contrast, purely quantitative approaches can scale to tens of millions of tweets and thus identify, if only bluntly, rare behaviors that would escape study in a small sample.

Recent work has adopted a balance between these two approaches. These mixed methodological approaches include, for example, active learning methods that adopt models to a scholar’s qualitative coding as the codes evolve [55], and the use of unsupervised machine learning methods to filter data into clusters or groups which are then analyzed qualitatively [56]. The present work takes the latter approach. While some have advocated for iterative use of unsupervised machine learning methods and qualitative analysis [57], we here find that a single application of a machine learning model and subsequent qualitative analysis is sufficient for our research question.

Our work also differs from previous research in what data we apply a machine learning model to. Prior research focuses on the application of machine learning to cluster social media posts into themes, or more specifically to cluster posts based on their content (and potentially, associated metadata) [58]. In contrast, the present work creates clusters of tweets based on who retweets what content. That is, we identify groups of tweets that are retweeted by similar sets of users. We discuss in more detail below the motivations for doing so.

Constructing the “who retweets whom” matrix.

We began by creating a matrix of who retweets whom in our sample, where ordinary users who retweeted others are the rows of the matrix, and influencers who were frequently retweeted are the columns. Overall in our data, 14,244,949 users (68% of all users) sent at least one retweet, and 2,308,610 users (11%) were retweeted at least once. The who retweeted whom network contains 166,021,415 total links.

Aligned with Zhang et al. [18], we performed a filtering step on this larger matrix. Critically, this biases our analysis towards heavier users of Twitter within our dataset. The core focus of our work, however, is on understanding whether or not there exists any variability. Consequently, finding variation in the sample of users we study is at worst an underestimate of the overall variability of Bernie supporters on Twitter. Our filtering step retains ordinary users who retweeted at least 12 different users, and influencers retweeted by at least 500 different users. These filter levels were selected by identifying the lowest levels at which 1) we observed interpretable clusters (as in Zhang et al. [18], we found that very low levels of filtering resulted in significant noise in cluster results), and 2) that were computationally feasible on the hardware available to us and likely available to other academic researchers (in our case, a single server with 16 CPU cores and 64 gigabytes of memory). This filtering step limits our analysis to 1,791,912 ordinary users, or 13% of those who sent at least one retweet, and 25,395 influencers, or 1.1% of all accounts that were retweeted at least once. Importantly, however, this sample retains 67% of the original links in the who retweets whom network, and accounts for over half (325M, or 55%) of all tweets in our dataset. Thus, while we sampled a small subset of users in our dataset, we retain the majority of content.

Finding clusters of users using vintage sparse PCA.

Following Zhang et al. [18], we then use Vintage Sparse PCA (VSP) to identify groups of ordinary users and the influencers that they commonly retweet. VSP performs the same computation as the traditional factor analysis in psychology: a principle component analysis followed by a varimax rotation. As such, both ordinary users and influencers can be said to “load onto” particular latent dimensions. We can then use the output of the factor analysis to place both ordinary users and influencers into groups by placing them into the factor they load the highest on [65], where these loadings onto each dimension are called factor loadings.

Our choice of VSP is motivated by a number of factors. First and foremost, it is feasible to run VSP on our data. Put another way, VSP is scalable; runs of VSP on matrices that involve millions of rows and tens of thousands of columns, as in the present work, can be carried out in under ten minutes. Relative to unimodal clustering algorithms (e.g. the Louvain methods [66]), VSP provides a principled approach to identifying the most influential accounts in each cluster, by simply looking at factor loadings of the matrix columns (influencers). Relative to other scalable, bi-modal clustering algorithms used in prior work analyzing Twitter data [67], VSP provides stronger theoretical (in a statistical theory sense) guarantees on robustness to non-normal inputs, and produces a sparse representation that in our initial analyses presented more salient differences across clusters.

Relative to other scalable, sparse bi-modal algorithms widely used in the literature, VSP has close links to other widely studied models. In particular, when data are not centered before applying the method, VSP is mathematically equivalent to the stochastic block model, a widely used hard clustering model for social networks [68]. Because we desire a hard clustering (i.e. every user is assigned to one and only one cluster), we use this approach to cluster the “who retweets whom” network. When data are centered, applying VSP is mathematically equivalent to latent Dirichlet allocation [69], a widely-used tool for soft clustering (i.e. Bayesian ad-mixture modeling) in the computational social sciences. We adopt this approach to cluster the “who retweets what” network, where we anticipate that users adopt multiple framed values instead of being fixed to just a single approach. Finally, VSP has been proven in prior work to be effective on the similar task of clustering the follower network on Twitter [18].

VSP thus presents an attractive tool for our work. However, as with any tool, there are notable limitations. First, VSP does not provide clear guidance on whether or not to scale inputs before use [65]. We opt not to scale in the present work, meaning our results may be more sensitive to outliers. Second, unlike some methods, VSP requires us to select the number of latent dimensions (or equivalently here, groups) to model. Recent work has argued that commonly used approaches to identifying the “correct” number of latent factors in a factor analysis may discard dimensions that hold potentially valuable data [70]. As such, we treated the problem of determining how many latent dimensions (clusters) to select as a task of finding a setting of k that is relatively consistent with other potential settings. To do so, we followed the method of Joseph et al. [67]. We first ran VSP with a variety of values for k, iterating from 5 to 60 in increments of 5. We then compared how consistent (in terms of the Adjusted Mutual Information [71]) the clustering produced by each value of k is with all other clusterings. We find that AMI is maximized when k = 35, and as such we selected that as the final k. See the Supporting information for additional details on the selection of k.

A mixed methods approach to identifying Bernie supporting clusters.

Given the output of a particular VSP model, we propose a multi-step, qualitative approach to identifying Bernie supporting accounts, as well as identity variations within those accounts. As a first step, we generated a sample of the top 25 influencers and top 25 ordinary accounts for each cluster. We defined “top 25” in each case using the factor loadings, which are, informally, a numeric value that represent the VSP model’s estimate for how representative a given influencer (ordinary account) is of a given cluster. We then conducted a first round of open coding [72] with information about these users, including their profile descriptions, usernames, locations, etc. Three paper authors then independently provided a label for the cluster and a decision as to whether or not the cluster appeared to represent Bernie supporters. In a second step of focused coding [72], for any cluster that at least one of the authors believed was a group of Bernie supporters, all authors additionally looked at the top 50 retweets for that cluster in terms of overall shares by ordinary users within the cluster. The three authors then came to a consensus decision on whether or not each of the 35 clusters represented Bernie supporters or not.

Once final decisions were made, we conducted a pair of quantitative evaluations to assess evidence from our qualitative findings, which were restricted to only top users, across all of the users within the clusters. Our first assessment focused on the extent to which ordinary users in the cluster were likely to retweet Bernie Sanders relative to the other candidates in the 2020 Democratic primaries, or Donald Trump. We measure the likelihood of retweeting one candidate relative to the others by computing the (weighted) log odds of a user within the group retweeting each candidate relative to all others, using the formula in Equation 21 of [73] and as implemented in the R package tidylo. For our second quantitative evaluation, we first identified a set of character strings we found to be representative of Bernie supporting accounts in our qualitative work. We then scanned profile descriptions of all accounts and determined the percentage of influencer and ordinary accounts in each cluster that contained the character strings “bernie,” “sanders,” “notmeus,” “socialist,” “Chapo,” or the red rose icon, which implies a socialist political orientation.

Finally, we use insights from both our qualitative and quantitative analysis to explore how collective identities varied across the five social groups we found to contain users that were likely to support Bernie. In addition to these existing methods, we additionally computed the weighted log odds of users expressing particular phrases in their profile descriptions, following the approach of Pathak et al. [35].

Moderate ideology clusters.

Tweets with framed values shaped by a moderate ideology were rare relative to those drawing from a far-left or progressive ideology, but still appeared to support Bernie during the 2020 election season. Drawing on Carmines and Berkman [75] we define a moderate ideology as one containing centrist views, support for the status quo, and instrumentally oriented. To this end, clusters of framed values we identified as emblematic of a moderate ideology were non-confrontational, passive conveyances of support, and friendly reminders to get out and vote.

Cluster 13: Pro-Biden. As mentioned above, a cluster of Bernie supporters signal boosted Joe Biden’s bid to become president in the fall of 2020. This translated to re-tweeting Biden’s campaign message that asked people to consider what would happen if Trump was re-elected. For some, the stakes were too high, even if Bernie’s supporters were not convinced that Biden was the best Democratic candidate. As one of Biden’s tweets rhetorically asked, “When Donald Trump says tonight you won’t be safe in Joe Biden’s America, look around and ask yourself: How safe do you feel in Donald Trump’s America?” Almost all tweets presented here have been retweeted at least 100 times, often substantially more, are from public figures, and are still available via the API. Tweets following this one are thus quoted verbatim. In the few cases where the tweet is no longer available, we do slightly modify content to preserve privacy and do not provide a citation.

Cluster 14: Hashtag heavy. Tweets in this cluster relied heavily upon hashtags to demonstrate their support for Bernie and his platform of signature issues; signaling support by writing, “#MedicareForAll,” “#BernieBeatsTrump.” “#Bernie2020,” “#NotMeUs” which was followed or preceded by the message of “Vote for Bernie.” Hashtags symbolically convey a person’s stance and are a mechanism for advocating for one’s position [76]. We coded this cluster as ‘moderate because most of the hashtags passively named issues or endorsement of Bernie, thereby upholding civility in discussions of politics [77].

Cluster 19: Get out the vote. Similar to the other moderate re-tweet clusters, this cluster’s main concerns were instrumentally oriented in reminding people to get out the vote in the primaries, reach out to friends and family members to vote, and contribute money to Bernies’ campaign which offered hope that a better future was possible if he was elected as the Democratic nominee, “Despite what the political establishment wants you to think, running a campaign of, by, and for the people isn’t a weakness—it’s our greatest strength. Together, we can prove it tomorrow”

Progressive ideology clusters.

Whereas clusters of framed values representing a moderate ideology were likely to uphold the status quo, retweet election strategies, and signal passive support for Bernie, progressives were more likely to retweet grievances around social issues, inequities, or disdain for other candidates in the race. We draw on McCright and Dunlap [7] to define a progressive ideology in the most generalized sense as one oriented toward improving social conditions for various social groups or issues.

Cluster 2: Anti-Biden. In the days leading up to Super Tuesday, Bernie supporters became more vocal in stating their opposition to Biden and calling attention to Biden’s historical record of conservative policy-making such as supporting the Iraq War, cutting social security, and the Defense of Marriage Act (DOMA). As one prominent retweet stated, “The American people are sick and tired of endless wars in the Middle East which have cost hundreds of thousands of lives and trillions of dollars. We will not defeat Trump with a candidate like Joe Biden who supported and voted for the Iraq War alongside Republicans.”Tweets in this cluster were concerned that Biden’s centrist policies and policy record appeared to be aligned with Republican values.

Cluster 6: Young and/or Black voices. While Bernie was often presented in the media as not being popular among Black voters, there was a cluster of Black users who vocalized support of Bernie or discontent with Biden. This was often, though not always, vocalized using humor. Humor in political mobilization can be an effective way to frame grievances or offer support [78] and this cluster used it in both ways to promote Bernie or to call out Biden and Harris for not being responsive to progressive Black voters. For example, as one retweet jokingly referenced Biden and Harris walking out of a building with a pole in between them, “They split the pole, we done for.”And another reflected, “This poor guy is gonna die tryna save the U.S. and you are not paying him any mind.” Voices representing “young people’’ also emerged here, with several tweets appealing to young adults to vote or to think specifically about youth and their life situations.

Cluster 7: The Lavender Left. Bernie had a consistent record supporting queer rights which this cluster emphasized, while calling out other candidates, such as Warren and Biden, who had previously voted in favor of anti-LGBTQ+ issues. As one tweet emphatically stated, “elizabeth warren used to be anti-lgbt and a republican. biden also was until recently. Here’s our guy Bernie: [link]” Undermining the credibility of other candidates by positioning not only their policies but identities as anti-LGBTQ+, these frames discredited Biden, Warren, and their supporters for not upholding progressive values. Moreover, this cluster advocated for an interdisciplinary platform of gender justice and anti-racist progressive feminist platform while rallying against Biden’s spotty and moderate neoliberal track record on abortion, women’s equality, and the rights of sexual assault survivors. This frame extended to a critique of those who claimed to be progressive or liberal while supporting candidates who did not uphold these values, especially in the context of rape allegations against Joe Biden and promoting nationalism over progressive values. As one tweet suggested, “…I smile serenely at the lack of ideology.”

Cluster 9: Covid-19. Faced with the Covid-19 pandemic during the height of the 2020 primary season, Bernie supporters began to make connections between one of Sanders’ signature issues—Medicare For All—and the unfolding global health crisis. From a video link tweeted out on April 3, 2020, Bernie opined, “If there’s anything the American people are learning during this horrific crisis is how weak and how dysfunctional our current healthcare system is…”

As one of the first candidates to begin speaking openly about the COVID-19 pandemic and potential economic hardship facing the country, Bernie framed his Progressive healthcare policies and hope for universal health coverage as a viable option for ensuring that all people had access to healthcare, even during a global pandemic.

Cluster 15: Fact-checkers. Scholars have consistently found that merely fact-checking potential mis- or disinformation does little to change people’s perspectives (Hoffman 2015). Despite the lack of efficacy, one cluster used fact-checking tactics to draw attention to the inconsistencies in Bernie’s opponents’ voting record on issues that progressives typically support. Tweets in this cluster were similar to those in the Anti-Biden one (Cluster 2), but appealed to progressives by using ‘fact-checking’-like language to position opponents as too centrist, or to frame Bernie as clearly on the right side of political debates in putting people before profits.

Cluster 20: We are all underdogs. The underdog is a powerful trope used in social movement mobilization to garner sympathy for one’s position or grievance [79]. Progressives used the underdog rhetoric to position Bernie as a hero and champion for the masses. As one tweet read, “Thank you Senator @BernieSanders for your compassion in these trying times, and for your solidarity with millions of laid-off workers, including tens of thousands @unitehere members in the hospitality industry. #ThankYouBernie.” Bernie’s policies and consistent concern for worker’s rights, affordable housing, and other progressive policies not only offered hope for people experiencing inequalities, but enabled them to frame their concerns from an underdog perspective and Bernie as the person who was well-poised to become a powerful advocate and champion.

Far left ideology clusters.

In contrast to clusters representing moderate or progressive ideologies, there were several clusters shaped by a far left ideology, which we define as one that upholds anti-establishment views and eschews institutional norms [51]. While the issues discussed were similar to those we labeled progressive, how these issues were framed was more emphatically conveyed, antagonistic, and hostile. At times, some of the clusters in the far left offered conspiratorial rhetoric to frame their grievances.

Cluster 10: Character attacks on other candidates. Engaging in character attacks against other candidates can be a potent strategy for discrediting an opponent and framing grievances [80]. One cluster engaged in such attacks, “If @JoeBiden was so good at building a coalition why did he need Klobuchar, Buttigieg, and Bloomberg to drop out and endorse him in order to get into a statistical tie with @BernieSanders? Because Biden is a bad candidate and Bernie’s message speaks to more voters.”

Using conspiracy-adjacent rhetoric, this cluster framed Biden’s weakness as not only a failure of the candidate, but of the DNC in gatekeeping Bernie from proceeding to the top of the Democratic ticket. While clusters from moderate and progressive ideologies framed opposition to Biden and support for Bernie, a far left ideology attacked Biden’s character by calling into question his integrity as a nominee and ability to win an election, which was further supported by the perceived coup by the DNC.

Cluster 11: Anti-Trump/police state. A distinct group of Bernie supporters were gravely concerned that Trump was attempting to suppress democracy, “Make no mistake about it. This is a blatant attempt by Trump’s handpicked Postmaster General (and campaign contributor) to sabotage the Postal Service, suppress the vote and undermine democracy” thereby making it difficult for mail-in ballots to reach their intended destination while turning cities into a police state, “Against the wishes of state and local officials, unmarked federal agents in combat gear have invaded parts of Portland, Oregon. This is what a police state looks like. These troops must be withdrawn immediately. We must defend the Constitution. We must defeat Trump.” Noting that Trump was acting against the wishes of voters, elected officials, and the public, Trump was characterized as deliberately upending democracy and working against the interests of the everyday person.

Cluster 16: Anti-establishment. Anti-establishment values framed this cluster’s retweeting patterns. Bernie often campaigned on the message of promoting progressive policies to end the wealth gap in the US, but the grievances suggested by this cluster often promoted the complete redistribution of wealth, “I have a message for the top 1% and the large profitable corporations in this country: Enjoy the massive tax breaks and loopholes you have right now. Under a Bernie Sanders administration, we’re going to end them and invest in working people.”

Cluster 17: Biden is a villain. Villainizing one’s opponent can also be a potent mobilization strategy if the source is perceived as credible [80] as character attacks rely upon well-worn tropes of heroes and villains [81]. This cluster tweeted with the conviction that Biden was, “not a decent person, a plagiarist, and pathological liar.” Framing Biden as a villain offered this cluster the rhetorical maneuver to promote a Vote for Bernie. Unlike the progressive clusters, however, those from the Far Left deduced instead that if the Presidential election was a choice between Trump and Biden they would, “…never under any circumstances vote for Joe Biden. Ever.”

Cluster 18: M4A. Across the ideological spectrum, many heralded Bernie’s signature Medicare for All platform. However, instead of mirroring Bernie’s call to action through unity, this cluster was shaped by anti-establishment thinking and conspiracy-adjacent perspectives by offering that even with a Democratic supermajority, the status quo democratic politician will, “…say no to it. What then? Keep waiting for a time that never arrives? You create your openings through force of will and public pressure. 300k Americans dead. Fight now.” Similar to cluster 16, there is a collective “we” invoked that positioned Bernie supporters in opposition to the status quo and in conflict with the average Democrat. As they surmised, achieving healthcare for all may mean imposing one’s will upon an unnamed entity.