Social Media Framing and Racism

In social media networks, users tend to cluster around like-minded peers, which Himelboim et al. (Reference Himelboim, Smith and Shneiderman2013) describe as selective exposure. Selective exposure occurs when individuals actively seek information that matches their beliefs, connecting with content that is cognitively congruent with their preferences and prior beliefs. Within these social media bubbles, individuals are exposed to information consistent with their beliefs while deciding what content to accept and share and what content not to accept and consequently not to share. In other words, users are selectively exposed to information cognitively congruent or cognitively dissonant with their preferences and then decide whether to propagate this content across the network (Aruguete and Calvo Reference Aruguete and Calvo2018).

These social media bubbles are a marketplace of noncompeting frames. Given the vertical configuration of some social media networks, it is usually high-degree network authorities, users with a significant number of followers, who are interested in framing social media events to their advantage. Users are exposed to numerous frames and “vote” among the choices. Cognitively congruent frames—that is, frames accepted by users—will propagate; cognitively dissonant frames will go unshared and unseen (Aruguete and Calvo Reference Aruguete and Calvo2018).

Racist content is among the frames users are exposed to. Like other frames in social media bubbles, racist frames will sometimes be cognitively dissonant to users and therefore go unshared; other times, they will be cognitively congruent to users and will propagate. Unlike other frames, racist discourse is socially punished, particularly in public settings (Bonilla-Silva Reference Bonilla-Silva2015). In other words, racist discourse is a socially costly behavior among a menu of other less socially costly options. As such, people are unlikely to engage without a trigger. When will racist frames, a socially costly behavior, activate social media networks?

Homogenous communities formed in social media, particularly those in polarized environments, resemble in-groups with shared identities and social homophily (Keipi et al. Reference Keipi, Näsi, Oksanen and Räsänen2017).Footnote ² Social affiliations to gender, religious, and ethnic or racial groups promote in-group bias: greater attachments to and preference for in-group members (Tajfel Reference Tajfel1981). For example, in political parties, these affiliations motivate members to advance the party’s status (Huddy Reference Huddy2001).

However, in-group love is not reciprocal to out-group hate (Brewer Reference Brewer1999). Desires to benefit the in-group (or the in-party) do not necessarily drive biased behavior toward out-group members (or out-party members). Denigrating the out-group does not advance the in-group or the in-party status. Instead, in political contexts, threats to the in-party identity (Amira et al. Reference Amira, Cole Wright and Goya-Tocchetto2021) or status (Mason Reference Mason2016) are drivers of out-group hate and anger. Furthermore, research has shown that anger is a powerful political mobilizer (Groenendyk and Banks Reference Groenendyk and Banks2014), especially in strong partisans (Huddy et al. Reference Huddy, Mason and Aarøe2015).

I extend this logic to the role of racism in social media. I argue that in these homogenous communities of like-minded peers, threats to the in-group status will activate racist behavior. Racist frames are particularly salient because they appeal to the in-group’s identity. When the out-group threatens the status of the in-group, users will accept and propagate faster frames that serve as markers to separate the in-group identity more starkly from the identity of the out-group. Racialized frames serve as these markers, and the threat to the in-group status makes racist content cognitively congruent. The mobilization of in-group users in social media is carried out by engaging with and propagating content (e.g., racist content), which occurs with cognitively congruent frames. From the proposed theory, the formulation of the hypotheses follows:

The findings from Amira and colleagues (Reference Amira, Cole Wright and Goya-Tocchetto2021), Mason (Reference Mason2016), and Groenendyk and Banks (Reference Groenendyk and Banks2014) are relevant in explaining the activation of racist frames in social media, especially when users frame political events. Social media have become a battleground where political narratives are delivered and framed among voters (Barberá et al. Reference Barberá, Jost, Nagler, Tucker and Bonneau2015; Neumayer et al. Reference Neumayer, Rossi and Karlsson2016; Aruguete and Calvo Reference Aruguete and Calvo2018). The political network communities created by the selective exposure and dissemination of content often align with the political camps contending for political power in “real life.” Thus political psychology and political communication literature insights inform our understanding of users’ engagement with racist content in social media. Mainly, users will engage with racist content, one of many frames shared in social media bubbles, when the status of the in-group is perceived to be threatened. Consequently, racialized frames, especially those attacking the out-group, will be more readily accepted and shared faster by users.

Racist Discourse in Social Media

The literature on Latin America (Martínez-Echázabal Reference Martínez-Echazábal1998) and Ecuador (Roitman Reference Roitman2009) notes that race is a complex construction, due to mestizaje and the strong correlation between ethnic background, perceptions of race, and class. In other words, how issues of race are framed, and how race is socially constructed, have geographical caveats. In Ecuador, as in other parts of the world, discursive racism is usually framed as intolerance toward the out-group (e.g., indígena) by creating strict demarcations between the self and the “other.” However, individuals often dismiss racism and race as explanations for racist behavior, and state structures and government representatives often pay mere lip service to integration and ethnic identity. It is not surprising that racist language is normalized (Roitman and Oviedo Reference Roitman and Oviedo2017), even though this characteristic has different degrees and forms (De la Torre Reference De la Torre Espinosa1996). Everyday patterns of behavior and speech and the organization of the state are configured in a way that “indios” and “indígenas” are the subjects and objects of structural discrimination.

We would expect online social spaces, such as Twitter, to replicate public and private racial discourses, especially since online spaces amplify racist discourse and unmask where racist discourse is produced and how it is reproduced (Eschmann Reference Eschmann2020). The internet is a hybrid social space, at once public and private (Daniels Reference Daniels2013), where established and new forms of racism are facilitated (Daniels Reference Daniels2013; Nakamura Reference Nakamura2007). The user-generated communities on online platforms encourage intimate discursive interaction predicated on racial identity (Brock Reference Brock2009; Daniels Reference Daniels2013).

To a varying degree, social media platforms allow for technical anonymity and social anonymity. Some studies suggest that this can explain aggressive and hostile online behavior, given a user’s perceived freedom from social standards and sanctions (Christopherson 2007; Lapidot-Lefler and Barak Reference Lapidot-Lefler and Barak2012), even though there is also evidence to the contrary (Jaidka et al. Reference Jaidka, Zhou, Lelkes, Egelhofer and Lecheler2022). However, not all platforms grant or encourage the same levels of anonymity (Barlett et al. Reference Barlett, DeWitt, Maronna and Johnson2018). Given the importance of the relational component in less anonymous social media networks such as Twitter (Preussler and Kerres Reference Preussler and Kerres2010), users will be cognizant and cautious of the real-world ramifications of their actions (Barlett et al. Reference Barlett, DeWitt, Maronna and Johnson2018).

Racism has diverse discursive manifestations, both overt and covert.Footnote ³ From a practical standpoint, this study focuses on overt forms of racist content, which are easier to detect systematically than other forms of racism; in Ecuador, overt racist forms explicitly target someone because of their indigenous identity, using negative and hurtful comments. Most important, in a country such as Ecuador, many platforms and institutions, classes, and contexts reproduce a “blanco-mestizo” racist ideology in often subtle and normalized patterns that the user is not aware of or would not consider racist (Roitman and Oviedo Reference Roitman and Oviedo2017). Overt racism dispels any ambiguity.

The #ParoEcuador Context

On October 1, 2019, Lenín Moreno, then president of Ecuador, announced the elimination of gas subsidies. Two days later, the Unión de Transportistas (Transportation Union) announced a strike.Footnote ⁴ In addition to the transportation sector, eliminating gas subsidies had a dire effect on the indígena community. Increases in gas prices led to increases in the prices of goods, ultimately affecting the poorer sectors of Ecuadorian society, disproportionally represented by the indígena population. Two days after the Unión de Transportistas announced its strike, the Confederation of Indigenous Nations of Ecuador (CONAIE), the largest indígena organization in the country, followed suit, announced a national strike, and started mobilizing its base to move toward the capital, Quito. The indígena mobilization was a reaction to the consequences of eliminating the gas subsidies and a continuation and expansion of ongoing local protests about the general lack of government support for the community’s problems.

By the time the indígena movement reached the capital, the president had moved the seat of government to Guayaquil and declared a state of emergency. In a polarized environment, pro- and antigovernment media and protestors displayed contrasting sentiments toward the actions of the executive and the protesters and widely differing accounts of violent incidents. The political environment was complex, as the indígena-led protests incorporated an ample array of social and political actors, among them the Frente Unitario de Trabajadores and Frente Popular (two workers’ unions) and various student movements. Furthermore, followers of former president Rafael Correa joined the protest against the government. While indígena leaders consistently disassociated themselves from Correa and Correa supporters who also took to the streets, the government and some of its supporters blamed the strike on Correa. For example, some of the content reproduced in social media framed the indígena movement as being manipulated by “useful idiots” (tontos útiles).

Posts related to the protests circulated extensively on Facebook and Twitter, the social media outlets with the largest user bases in Ecuador (AmericasBarometer 2018).Footnote ⁵ As state violence increased, many of the reports were initially broadcast by online media sources before being picked up by the more traditional outlets. The country was paralyzed for more than ten days, prompting mixed reactions from different groups. Labor unions joined the protests and various organizations, including universities, supported the indígena movement as government violence increased. However, the protests also paralyzed an already faltering economy. Business representatives, for example, denounced the indígena protest and praised the government for its position (Díaz and Mejía Artieda Reference Díaz and Mejía Artieda2020).

Beyond the role of social media in the Ecuadorian protest, there are structural characteristics worth discussing. Racism in Ecuador is a “system of ethnic-racial dominance” historically rooted in European colonialism (Van Dijk Reference Dijk and Adrianus2005), directed, in great part, toward the indígena population (Beck et al. Reference Beck, Mijeski and Stark2011). The state has marginalized the indígena population and done little to support its communities or grant them equal access to political spaces. Regardless, the indígena population has managed to organize around a unified banner (i.e., the “indígena” banner, despite the different and sometimes conflicting nationalities) to demand and achieve important political and social victories (see Becker Reference Becker2010). However, these victories have done little to change the racist ideology that permeates all levels of the state and society. Despite their political activism and mobilization, indígenas and the indígena community are still marginalized and remain the main target of the national racist ideology.

The #ParoEcuador Data

Between October 1 and October 24, 2019, I collected three waves of Twitter data using the strings paro and Ecuador, two politically and racially neutral terms used by the government and indígena supporters alike. To collect these data, I connected rtweet (Kearney Reference Kearney2019) to Twitter’s backward search application programming interface (API) and gathered tweets for the duration of the unrest in Ecuador. The data include 2,425,239 posts by 85,249 unique Twitter users for the Ecuadorian case. Of this sample, 93 percent were retweets of an original tweet.

I restricted data collection to various days before the beginning of the indígena strike and various days after the end of the strike. By limiting the data to October 2019, I could be more confident that the dialogue between users was about the actions of the indígena community during the strike. Throughout, the data show a latent level of perceived threat to the in-group status. Before the national conversation centered on the events surrounding the strike, little racist content was produced because little attention was paid to the indígena community.

The #ParoEcuador Network

Selective exposure in social media leads to homogenous and differentiated communities. To estimate these communities in the Twitter data, I first built a directed network in which each user was a node and an edge was created when a user H (hub) retweeted user A (authority).Footnote ⁶ To illustrate this interaction, figure 1 shows a diagram of how the Twitter network is created.

Figure 1. The Formation of a Network in Twitter

Notes: In the figure, @userA tweeted two tweets: “Tweet A” and “Tweet B.” @userH_1 and @userH_2 retweeted “Tweet A,” while @userH_3 retweeted “Tweet B.” Each user is a node in the network. When a user (node) retweets another user (node), a link (edge) is created between them. In the diagram, each arrow is an edge. Since @userH_1 retweeted a tweet from @userA, @userH_1 is called a hub, and @userA is the authority. Users (nodes) can be both hubs (when they retweet other users) and authorities (when other users retweet them).

Communities—clusters of nodes where the same information (tweets) is shared—are identified via a random-walk community detection algorithm (Pons and Latapy Reference Pons and Latapy2005).Footnote ⁷ This algorithm identified three primary communities in Ecuador: a progovernment network, which included 36,579 nodes; an indígena community network of 29,624 nodes; and a pro-Correa community network of 25,376 nodes. These communities are in line with the main groups disputing in the political battle during the protests (Díaz and Mejía Artieda Reference Díaz and Mejía Artieda2020). Figure 2 presents a basic Fruchterman-Reingold layout of the Ecuadorian network.Footnote ⁸ It describes the progovernment community with blue squares, the indígena community with red triangles, and the pro-Correa community with green circles. The size of the nodes is proportional to the nodes’ in-degree, with larger nodes indicating users retweeted by a larger number of followers.

Figure 2. Primary Connected Network During the Ecuadorian Protests

Notes: Ecuadorian protests between October 1 and October 24, 2019. Red triangles describe progovernment users. Blue squares describe users aligned with the indígena community. Green circles describe pro-Correa users.

On Twitter, communities formed around political events and cleavages often have at their center political leaders or users strongly aligned to the leadership. Influential authorities in the 2016 US election communities included presidential candidates @HillaryClinton and @realDonaldTrump; for the 2018 #Tarifazo networks in Argentina, it was the opposition leader @CFKArgentina (Cristina Fernández de Kirchner) and then-president @mauriciomacr (Mauricio Macri). In the Ecuadorian network during the 2019 indígena protests, the progovernment community had at its center President @lenin, Vice President @ottosonnenh, and interior minister @mariapaularomo, as well as other prominent progovernment users. In the center of the pro-indígena community was the institutional account of @CONAIE_Ecuador and its president, @jaimevargasnae. And in the center of the pro-Correa community were former president @mashirafael and high-ranking members ofhis party. Beyond public officials or politicians, other influential users included media personalities, media outlets, and social media commentators.

The difference and sparsity of exchanges between each community give an account of the polarization of the Ecuadorian network. Of all edges in the indígena community, 78.0 percent are with members of the same community (i.e., indígena community → indígena community), and only 6.5 percent with members of the progovernment community.Footnote ⁹ Of all edges in the progovernment community, 91.6 percent are with members of the same community (i.e., progovernment community → progovernment community), and only 4.4 percent with members of the indígena community.

It is important to note that Twitter is not a representative sample of the Ecuadorian population. According to LAPOP’s AmericasBarometer project (2018), social media users in Ecuador are more likely to be young, urban, educated, and well-off. Despite these shortcomings, using Twitter data has several advantages. Unlike other, more popular social media sites, Twitter data are readily available. Twitter’s hierarchical nature resembles the political dynamics of real life, as a few important figures produce most of the information. At the same time, many peripheral users decide whether or not to replicate it (while not producing much information on their own). Furthermore, although they do not reflect the general population, Twitter networks, including the Ecuadorian network, are consistent and stable across time (Calvo and Aruguete Reference Calvo and Aruguete2020) and reflect the similar polarized interactions witnessed on the streets and in the press.

Detecting Racist Tweets

A large body of research is dedicated to detecting hate speech on social media (Schmidt and Weigand Reference Schmidt and Wiegand2017; ElSherief et al. Reference ElSherief, Shirin Nilizadeh, Vigna and Belding2018; Chatzakou et al. Reference Chatzakou, Nicolas Kourtellis, De Cristofaro, Stringhini and Vakali2017), but accurate and systematic hate speech detection is a challenging task (Davidson et al. Reference Davidson, Warmsley, Macy and Weber2017). Despite the many advances on the topic, there are still limitations to the automatic detection of racist discourse. Not the least of these is the contextual nature of racism (Van Dijk Reference Dijk and Adrianus2005). While racist discourse will work to maintain existing power structures and racist ideologies, the language will evolve and shift, depending on the particularities of society and time. Thus, even if we were to take pretrained models to detect hate speech, these would be useless in contexts different from the ones they were originally trained for.

To solve this problem, I adopted a multistep classification approach, like that employed by ElSherief et al. (Reference ElSherief, Shirin Nilizadeh, Vigna and Belding2018). I started by defining a racist attack toward a member of the indígena community or toward the indígena community in general as a “negative or hateful comment targeting someone because of their indigenous identity,” a variation of the definition used by the Google’s API Perspective.Footnote ¹⁰ I used Google’s API Perspective, a content-moderating tool that is the industry standard for automatic detection of toxic content in written comments. Perspective uses a convolutional neural network that scores the likelihood a text contains an identity attack.Footnote ¹¹ Perspective provides an identity attack score from 0 to 1, interpretable as the probability that a text will be perceived as an identity attack. I used a threshold score of 0.85 to create a dummy for whether a comment is an identity attack or not.Footnote ¹² The Perspective algorithm was trained to detect identity attacks on frequently attacked groups, focusing on sexual orientation, gender identity, and race.

Since this study is specifically interested in racist discourse directed toward the indígena community in Ecuador, I created a dictionary with key phrases that identify the indígena community or members of the indígena community (i.e., indígena, indio). I kept only identity attacks (tweets) that contained indígena, the term most used to refer to a member of the indígena community (e.g., el/la indígena) or the indígena community in general (e.g., los indígenas). An alternative, more charged term to refer to indígenas is indios. The blanco-mestizo population often uses indio (indian) as a derogative identifier. Despite the indígena community’s long history of reclaiming the term, it is still used to signal a racial attack. Therefore I followed a similar procedure as before (i.e., detect identity attacks and keep those that include the term indio) but lowered the threshold score to 0.75, given the charged nature of the term. Even though the term indio does not automatically reflect racism, it increases the probability of a text being racist (primarily when someone outside of the indígena community uses it). In addition, I created a second dictionary with local forms of racist discourse that the algorithm cannot detect. For example, during the protest in Ecuador, the phrase emplumados (feathered), was used mockingly to describe indígena leaders.Footnote ¹³

To check the internal validity of the measure, I hand-annotated a random subset of 1,500 tweets and compared the results to the ones obtained in the main procedure.Footnote ¹⁴ In a sample limited to tweets from the progovernment community, the multistep model obtained an F1 score of 0.89, with a recall score of 0.85 and a precision score of 0.94, suggesting that the model is accurate and particularly good at avoiding false positives (Type-I error).Footnote ¹⁵ No overt forms of racism were found in tweets produced by users from the indígena or pro-Correa communities in the hand-annotated sample. Nevertheless, the multistep process detected racist tweets in those communities, a limitation discussed below. After applying these filters, I identified 1,371 (2 percent) unique racist tweets in the progovernment community. This multistep process addresses both aspects of the definition of racist discourse: 1) the Perspective algorithm detects “negative or hateful comment targeting someone because of their identity,” and the key phrase dictionary identifies the indígena community or an individual indígena as the target of the toxic tweet.

This multistep process has some limitations. The first and most noteworthy is its reliance on the Perspective algorithm to identify racist tweets. Hosseini et al. (Reference Hosseini, Kannan, Zhang and Poovendran2017) show drawbacks to the Perspective toxic detection system, mainly underdetecting toxicity when keywords (e.g., words that signal toxicity) are misspelled and sending false alarms for benign phrases denouncing toxic behavior.Footnote ¹⁶ The latter was particularly problematic for tweets produced by users from the indígena or pro-Correa communities. While the rate of racist tweets detected by the model in these communities was low relative to those identified in the progovernment community, all of these were incorrectly identified as being racist.Footnote ¹⁷ These tweets usually denounce racist behavior and discourse from other users, political figures, media, and the police or point out racism in another user’s tweet. Therefore I focused solely on users from the progovernment community.

A second limitation is that the Perspective algorithm cannot capture subtler forms of racist discourse, such as those that infantilize indígenas (Guerrero Reference Guerrero1997) or use irony or wordplay to mock indígenas (Sue and Golash Boza Reference Sue and Golash-Boza2013). These forms of racism can be found in all the communities, including the indígena and opposition communities. However, because of their subtlety, I could not systematically detect them through the model, so I limited the analysis to overt forms of racism. Exploring subtler forms of racism is a promising avenue for further work.

In addition, Twitter has various mechanisms that detect and potentially eliminate tweets with racist content.Footnote ¹⁸ Unfortunately, it is impossible to know whether a tweet was deleted or if it was deleted before the collection of the corpus (see Timoneda Reference Timoneda2018).Footnote ¹⁹ Overall, most of the limitations led to underreporting of racist discourse in the corpus. Underreporting leads to Type II error, with a higher probability that the effects of racist discourse are underestimated. Thus, in a worst-case scenario, my findings are conservative estimates of the true effect.Footnote ²⁰

User Activation and Variables of Interest

The primary variable of interest is whether a tweet is racist or not. We want to know how these messages are accepted and spread across the network. In social media, “acceptance equals propagation” (Aruguete and Calvo Reference Aruguete and Calvo2018). Aruguete and Calvo (Reference Aruguete and Calvo2018) show that reaction time is associated with the diffusion of content (i.e., less time, more diffusion). They further show that time-to-retweet, a measure of reaction time, is a proxy for latency, which researchers frequently use in experimental settings to measure cognitive congruence or dissonance. Content that is accepted (i.e., cognitively congruent) will propagate. As Aruguete and Calvo (Reference Aruguete and Calvo2018, page 12) point out, “acceptance leads to propagation … as users share information, new users are exposed to the media content of their peers.”

To operationalize “acceptance” and the speed at which frames are spread in a network, I tested whether latency increased or decreased racist content and how latency for racist content varied under different conditions. Latency is measured as time-to-retweet, the number of seconds elapsed from when a user (authority) posts a tweet to when a second user (hub) retweets the same post. Users consuming messages on Twitter will propagate content from their community peers more quickly when these are accepted.Footnote ²¹

In addition to the tweet’s text, the tweet’s time, and the retweet time, I collected information on the users’ screen name, follower count, friend count, and the status of the users’ accounts (verified or not verified). From the network, I computed in-degree—the number of times a user had been retweeted—to identify high-authority users (i.e., users with a significant number of followers) and low-authority users. Users who are more central to the network usually behave differently from less prominent users (Calvo and Aruguete Reference Calvo and Aruguete2020), and this information allows us to control for that effect. I also controlled for the effect of bots by including the iratio of followers to friends.Footnote ²² In line with prior research, the Twitter data showed high degrees of concentration. In the #ParoEcuador network, less than 5.6 percent of accounts w responsible for 45 percent of the content circulated in the network.

Sharing Racist Tweets in the #ParoEcuador Network

Before analyzing the (racist) reaction of progovernment users to threats to the in-group status, let us look at a more general snapshot of users interacting with racist content. Figure 3 describes user activity of the progovernment #ParoEcuador Network across time, labeled by critical dates The red line represents overall tweets, and the black line racist tweets. The network activity increased right after the Unión de Transportistas announced its strike. However, it was only after the CONAIE announced its strike that the racist content began and continued for some time after the end of the strike, eventually waning, as did the attention to the issue. The timeline shows how users became more active in the network as the strike intensified and how racist activity did not necessarily follow the same pattern. In general, racist content was a latent frame in the progovernment community.

Figure 3. Timeline of #ParoEcuador Activity, October 1–October 24, 2019

Hypothesis 1 argues that users will reject racist frames and decrease the speed with which they share racist content. To systematically explore how racist language increases (or decreases) latency in Twitter user engagement, the determinants for time-to-retweet were analyzed. Higher values for time-to-retweet mean longer times between the original post of the tweet by the authority and the retweet of the post by the hub. I estimated a Cox proportional-hazards model, with unstandardized coefficients describing changes in the hazard rate of time-to-retweet. The Cox proportional-hazards model estimates the survival rate of an object of study; in this case, how long before a tweet is retweeted. The less a tweet survives before being retweeted (i.e., the faster it is retweeted), the more the user accepts the content. In terms of the results given by the model, positive coefficients indicate an increase in the hazard rate (faster time-to-retweet) and acceptance of the frames. In contrast, negative coefficients indicate slower times and greater rejection of the frames.

One important aspect is that verified users neither posted nor retweeted racist messages.Footnote ²³ This behavior is probably a direct result of the costs public figures incur when engaging with racist content, something that less-public users can get away with. Also note that many verified users are news outlets and government officials who often have staff managing their social media accounts and more to lose if they were to engage publicly in overt forms of racism.

Table 1 shows the results from the Cox proportional-hazards model. For an intuitive interpretation of the magnitude of the effect, consider the effect of our covariate of interest, racist attack, among users of the progovernment community, which is negative and takes the value of –0.148 (p < 0.01) in model 1. The exponentiated value of the coefficient (0.86) is the incidence rate and can be interpreted as the (instantaneous) change in time-to-retweet when a tweet has racist content. One minus the incidence rate is 0.14, showing that the time-to-retweet for racist posts is about 14 percent slower. Thus, users are not eager to retweet racist tweets posted by their own authorities, suggesting that the content of racist messages produces rejection among users, or that sharing racist messages increases the latency of the decision to bear the (social) cost of sharing racist message content. The survival probability curves of time-to-retweet for racist and nonracist content are plotted in figure 4. The results are in line with the expectations from hypothesis 1.

Table 1. Racist Content and Time-to-Retweet in the Ecuadorian Progovernment Community

Standard errors are reported in parentheses, with confidence levels reported as follows: *p < .1; **p < .05; ***p < .01.

Note: Hazard estimates of time-to-retweet. Positive numbers indicate positive increases in hazard rate and shorter time to retweet.

Figure 4. Survival Probability of Cox Proportional-hazards Modes

Note: From table 1, model 1, with time-to-retweet as the dependent variable. Survival curves with 95 percent confidence intervals.

There are other particularities about the network worth mentioning. The progovernment community was not eager to retweet messages from its authorities. Results show that users more central to the network (authorities with an in-degree above the log median) were retweeted 20 percent more slowly than users less central to the network (authorities with an in-degree below the log median). However, users more central to the network retweeted posts 9 percent faster than users less central to the network. The difference in behavior shows that the “soldiers” of the network—i.e., users less central to the network—were not being activated by the content posted by the central users but rather by content created by other, less prominent users.

As anticipated by hypothesis 1, racist content had a higher time-to-retweet than nonracist content. The larger latency in time-to-retweet suggests that, in general, racist content induced rejection from users and a slower diffusion of the frame. Furthermore, the results suggest that public figures avoid producing racist content and avoid reproducing it. The social cost of overt racism is a likely deterrent for that behavior, especially damaging to the image of more prominent political personalities. Alternatively, more-public figures might also be less racist. While we cannot discard this possibility, it seems less plausible, considering the racist structure embedded in this network.

Threats to the In-Group and Racist Content

The main argument of this article is that racism is activated when the status of the in-group is threatened (H2). To provide empirical evidence, we can look at two events in which the indígena community (out-group) threatened the status of the government, either by calling on forces to challenge the government or by calling the government to acquiesce to its demands. Specifically, I test the hypothesis by examining the moments when Jaime Vargas, the president of the CONAIE, called the police and military forces to disobey government orders and join the protest, and when the Moreno government announced the end of the strike after agreeing to the demands of the indígena leadership.

To determine whether these two events affected users’ time-to-retweet to racist content, I followed Calvo et al. (Reference Calvo and Aruguete2020) and employed an interrupted time series analysis, a variety of regression discontinuity design (RDD) in which the running variable is time (Morgan and Winship Reference Morgan and Winship2015; Mummolo Reference Mummolo2018). The key assumption of this approach is that any immediate change in the time-to-retweet can be attributed to the events and not to any other factor affecting time-to-retweet that also changed systematically at the same point in time.Footnote ²⁴ The granularity and abundance of Twitter data, as well as the public nature of these events, allows me to isolate the change in time-to-retweet on the specific second the event took place and thus make a more plausible assumption that other omitted variables (e.g., other events not attacking the status of the in-group) are not also changing suddenly at the time of the events. To further isolate the effect of the event, I estimated the models within a six-hour window around the cutoff (i.e., time of the events).Footnote ²⁵

As noted, the primary quantity of interest is the immediate change in time-to-retweet after the cutoff. I estimated the models using a nonparametric local linear regression (LLR) to approximate the treatment effect at the cutoff points (Gelman and Imbens Reference Gelman and Imbens2019). In other words, I fitted two separate regression functions, one before and one after the cutoff events. The treatment effect (i.e., the effect of the event) is the difference of both functions at the limits of the cutoff.Footnote ²⁶

A possible threat to the model’s validity is the potential of sorting by users right before the event. In this case, users might anticipate the outcome of the events (i.e., a perceived threat to the status quo) and change their behavior accordingly. Given that we expect the event to decrease time-to-retweet, any anticipation to the treatment is likely to go in the same direction. Therefore, the model would underestimate the effects of the treatment, suggesting that the true effect of the event would be stronger. However, I ran a McCrary test and found no evidence of sorting around the cutoff for racist users engaging with racist content (McCrary Reference McCrary2008). Appendix B provides additional tests to verify the continuity assumption, including placebo checks with the running variable. Overall, the results ensure the internal validity of the RD design. We now turn to the events and how they constitute a threat to the status of the progovernment group.

Calls to Arms and Acquiescing

Historically, the indígena uprisings have had three political outcomes: subjugation through violence, policy reform and expansion of rights (e.g., constitutional recognition of Ecuador as a plurinational state), and coups. The 2019 protests began as a series of demands (i.e., policy reform) made by the indígena community in reaction to government action that culminated in eliminating gas subsidies. Indígenas began marching from their communities, often located in rural Ecuador or Amazonia, and headed, as is customary, to the capital, Quito, where they would occupy city spaces of symbolic importance and negotiate with the government.

On October 10, the indígena assembled in a large forum in Quito, where eight police officers, among them a colonel, were detained for more than two hours, as well as various journalists and their crews.Footnote ²⁷ Jaime Vargas spoke to the crowd on live television (the detained broadcast journalists and their crews came in handy), with the detained police officers in the background. At 11:51 a.m., Vargas asked the detained colonel, the police force, and the military to join the protests and disobey the orders of the government.

The moment of Vargas’s sedition created a reaction in the progovernment community. Given recent Ecuadorian political history, Vargas’s call was not an empty threat but rather a clear challenge to the power and the status of the government.Footnote ²⁸ Through Vargas’s call, the out-group conspicuously challenged the in-group. We would expect this action to be perceived as a threat to the in-group and to trigger the progovernment community’s shared identity, particularly the racial identity.

Similarly, on October 13, high-ranking officials from the Moreno government sat down with indígena leaders to negotiate the end of the strike. The indígena leaders demanded, among other things, that Moreno rescind the order that prompted the mobilizations in the first place. After three hours of negotiations, news stations broadcast the meeting live as each party made public statements. At 9:45 p.m., a United Nations mediator announced the conditions under which the strike would be terminated. Lenín Moreno agreed to reinstate the gas subsidies, and the indígena leaders agreed to suspend the mobilization. In the eyes of progovernment users, the indígenas had won. The adjudication of the strike materialized the challenges to the in-group status. As in the previous event, the immediate actions that led to the end of the strike were broadcast and garnered collective attention.

Furthermore, both episodes were highly salient events. High-salience events focus the public’s attention and redefine the situation (Pride Reference Pride1995. Lin et al. (Reference Lin, Keegan, Margolin and Lazer2014, pg. 2) argue that highly salient media events “create a social condition called ‘shared attention,’ … where the larger potential audiences, altered norms, and the high level of shared understanding can all contribute to shifts in how users … attend to content” (emphasis added). Lin et al. (Reference Lin, Keegan, Margolin and Lazer2014) explain that in media events, users can converse without needing to explain the context, allowing for the creation of a more disciplined message. In both of these events, progovernment users reacted to these perceived threats to the in-group status by engaging with racist frames.

The Effects of Perceived Threats on Racist Content

Table 2 reports the robust bias-corrected treatment effects and 95 percent confidence intervals as suggested by Calonico et al. (Reference Calonico, Cattaneo and Titiunik2014). The dependent variable is time-to-retweet for models 2.A and 2.B. In model 2.A, the cutoff time of the discontinuity is centered on October 10, 2019, at 11:51 a.m. local time, when Jaime Vargas, the president of CONAIE, asked the police force and the military to join the protests and disobey the orders of the government. In model 2.B, the time of discontinuity is centered on October 11, 2019, at 9:45 p.m. local time, when the United Nations mediator announced the terms agreed upon for the end of the strike. The treatment effect for racist tweets in both events is negative—time-to-retweet is reduced—and statistically significant at the 95 percent level.

Table 2. Regression Discontinuity Estimation for Racist and Nonracist Tweets

Notes: 95 percent confidence intervals reported in brackets, with confidence levels reported as follows: *p < .1; **p < .05; ***p < .01.

Model 2.A centers treatment on October 10, 2019, at 11:51 a.m. local time, when Jaime Vargas, the president of the CONAIE, asked the police force and the military to join the protests and disobey the orders of the government. Model 2.B centers treatment on October 11, 2019, at 9:45 p.m. local time, when the United Nations mediator announced the terms agreed on for the end of the strike.

To better illustrate the effect, panels a and b in figure 5 plot the regression discontinuity results for the event. For both panels, the vertical axis reports time-to-retweet, and it is interpreted as usual: lower values mean less time-to-retweet and less user latency. The sample is divided into racist and nonracist tweets. The horizontal axis has a range of 12 hours, 6 hours before and after the event. A LOESS smoother was used to fit the underlying regression function separately before and after the event. The discontinuity shows that there is no change in latency for users sharing nonracist content at the time of Vargas’s speech. However, there is a statistically and substantively significant difference in latency for users sharing racist messages (p < 0.01). Immediately after Vargas’s call, racist content increases engagement and reduces latency among progovernment users. Notice that before the cutoff, racist content is spread, on average, slower than nonracist content. However, at the time of Vargas’s speech, users engage with racist content faster than with nonracist content. The reaction to the perceived threat to the status of the in-group points to progovernment users’ accepting racist frames and increasing the speed of their diffusion.Footnote ²⁹

Figure 5. Time-to-Retweet During the Ecuadorian Protests

Top panels: Centering on October 10, 2019, at 11:51 AM, local time, when Jaime Vargas, the president of the CONAIE, asks the police force and the military to join the protests and disobey the orders of the government. Bottom panels: Centering on October 11, 2019, at 9:45 PM, local time, when the United Nations mediator announced the terms agreed on for the end of the strike.

Panels c and d in figure 5 show a similar treatment effect when the end of the strike was announced. In this case, the discontinuity shows that the end of the strike decreased the latency for both nonracist and racist content (p < 0.05), suggesting that adjudication (i.e., the act of granting or denying ownership of an outcome to groups in social media) also affects the reaction of users to content, a result in line with. Thus, some of the effects of the event on latency might be a product of adjudication. Nevertheless, the treatment effect for racist tweets (–2.17) is larger than the treatment effect for nonracist tweets (–0.66), suggesting that some of the change is a reaction to the threat to the in-group status.

Analogous to Vargas’s call, racist content went from having a longer time to retweet than nonracist content to roughly the same time when the end of the strike was announced. In other words, both types of frames created similar acceptance from progovernment users after the strike adjudication.

Overall, I find that users of the in-group, the progovernment community, engaged with racist frames when the status of the in-group was perceived to be threatened. The analysis of two high-salience events shows reduced latency of engagement and greater acceptance from progovernment users of racist content. I also find that, absent these threats, racist content creates rejection from users.