The insurrection at the US Capitol on January 6, 2021, is widely considered one of the most remarkable examples of a violent attack on democratic institutions in a mature democracy in recent times (Bright Line Watch Reference Watch2021). Yet, even though many politicians and pundits condemned the insurrection and mass approval of President Trump decreased in its aftermath (Bump Reference Bump2021), we know little about the broader effects of political violence such as this episode on mass political behavior.
In this research letter, we take an important first step in uncovering the consequences of the Capitol insurrection for political affiliations by investigating changes in expressions of self-identification with the Republican Party and President Trump in the days immediately following the event. Specifically, drawing on recent work that uses social media self-descriptions as indicators of political identities (Rogers and Jones Reference Rogers and Jones2021), we study changes in identification with the Republican Party and then-President Donald Trump in the personal “bios” of around 117,000 users on the microblogging platform Twitter, all of whom express a partisan identity prior to the insurrection. Using panel data that track these users each day, we apply a flexible difference-in-differences model to estimate the causal effect of the insurrection on expressed partisanship. Our findings demonstrate that the insurrection caused an exceptionally clear immediate decrease in expressions of identification with the Republican Party and “Trumpism” that persists, at least in the short term, and is consistent across a wide series of robustness checks.
Our analysis contributes to a nascent literature concerning the consequences of violent protests for political behavior. Some prominent recent studies have brought this research agenda to the fore. One study finds that proximity to violent Black-led protests in the 1960s caused an increase in endorsements of “social control” and support for the Republican Party (Wasow Reference Wasow2020). Conversely, another study finds that proximity to the 1992 Los Angeles riots led to a liberal shift in policy support and an increase in Democratic Party voter registration (Enos, Kaufman, and Sands Reference Enos, Kaufman and Sands2019). This resonates with other findings showing that protests around liberal issues correlate positively with subsequent local vote share for Democratic candidates and vice versa for conservative issues and Republican candidates (Gillion and Soule Reference Gillion and Soule2018). Beyond these few but important studies, however, “we know little about the effect of these events on political behavior” (Enos, Kaufman, and Sands Reference Enos, Kaufman and Sands2019, 1012).
Our study advances our knowledge about the effects of violent protests in several important ways. First, in contrast to the aforementioned studies’ focus on historical cases, we investigate how a contemporary protest shapes political attachments. This is relevant because increased affective polarization and associated politically motivated reasoning (Iyengar et al. Reference Iyengar, Lelkes, Levendusky, Malhotra and Westwood2019) may have reduced the role of events—even extreme ones—in shaping political attachments. In other words, are violent political protests still consequential for voters’ political affiliations in a time of high political polarization? Second, we study whether reactions to violent protests extend beyond their immediate geographical locus. Given the increasingly nationalized nature of American political behavior (Hopkins Reference Hopkins2018), more widespread effects seem plausible. Third, in contrast to two of the studies highlighted above (Enos, Kaufman and Sands Reference Enos, Kaufman and Sands2019; Wasow Reference Wasow2020), which investigated protests associated with the political left, we examine the consequences of a right-led violent protest. This is important in light of known asymmetries between left- and right-wing movements (Grossmann and Hopkins Reference Grossmann and Hopkins2016; but see Gillion and Soule Reference Gillion and Soule2018, who find similar reactions to historical right- and left-led protests). Because the right-led insurrection arguably differs from typical left-led protests on many accounts (e.g., guiding motive and the racial and social composition of protesters) it is difficult to make any clear predictions from existing studies ex ante (Manekin and Mitts Reference Manekin and Mitts2022). Our study thus broadens our understanding of the immediate political behavioral consequences of political violence.
Our study also has implications beyond the developing literature on the political behavioral consequences of violent protests. First, it connects to the related literature on political violence and more specifically to the costs and benefits of violence to political actors in the comparatively rare setting of a developed democracy (Rosenzweig Reference Rosenzweig2021). In essence, public reactions to the Capitol insurrection indicate whether political violence is an attractive strategy for political elites to appeal to US voters. Second, our study addresses the limits (or lack thereof) of partisanship in the US, at least for the Republican Party. Often attributed to increases in political polarization (Iyengar et al. Reference Iyengar, Lelkes, Levendusky, Malhotra and Westwood2019; Mason Reference Mason2018), the strength of partisanship is now so socially and politically consequential in the US that, in the words of a recent study, “it is difficult to overstate the importance of party loyalty” (Barber and Pope Reference Barber and Pope2019, 39). Indeed, partisanship has been shown to drive economic behavior (McConnell et al. Reference McConnell, Margalit, Malhotra and Levendusky2018) and is itself linked to violence (Kalmoe and Mason Reference Kalmoe and Mason2022). Thus, by analyzing whether (expressed) partisans are willing to forego identifying with their party in the face of exceptional political violence, we examine the scope conditions of the “unmovable” character of partisanship in the United States. Finally, by investigating the immediate consequences of the insurrection, our study addresses subsequent struggles within the Republican Party over how to respond to the insurrection (Cheney Reference Cheney2021) and thus contextualizes later efforts and reversals by Republican politicians to mitigate the insurrection’s political costs.
Research Design
Data and Sample
We collected data daily from the Twitter bios of 3.4 million geolocated US users starting seven months prior to the Capitol insurrection until approximately two months afterward (June 1, 2020–March 15, 2021). This sample was drawn from a population defined as active US social media users who are at least minimally politically engaged. We defined this as any user who followed at least one of an ideologically diverse set of major US news media accounts including MSNBC, Huffington Post, New York Times, Washington Post, CNN, Wall Street Journal, FOX News, and Breitbart News. We first collected the profiles of followers of each media account. Then, to identify active users (and reduce the likelihood of collecting bot accounts), we included only users who sent at least one tweet in the past year, sent at least 25 tweets ever, and had at least 10 followers. Finally, we included only US-based users based on geocoordinates and text location information.
In total, this yields an initial sample consisting of 3.4 million users. Because our group of interest is partisans, we reduce this sample to the subset of users with expressed partisanship (see the next section) at any point during our study period. This yields a final sample of around 117,000 users—around 3.5% of the initial sample—roughly equally divided between Republican and Democratic partisans. We show each step of our sampling process in Appendix C, where we also report descriptive statistics on Twitter metadata demonstrating that users in our final sample are relatively more active and connected across a variety of metrics (e.g., number of Tweets and likes). This implies that our final sample of expressed partisans plausibly consists of users considerably more engaged than Twitter users overall, who are themselves more engaged than the average American (Blank Reference Blank2017). Although we cannot infer to the population more broadly, this should make for a comparatively hard test of the malleability of expressed partisanship, as party identification is generally more stable among more politically engaged individuals (Green, Palmquist, and Schickler Reference Green, Palmquist and Schickler2004). Due to the General Data Protection Regulation adopted by the European Union, we are unable to publicly share individual-level data, but we make aggregate-level data available as part of our replication materials (Eady, Hjorth, and Dinesen Reference Eady, Hjorth and Dinesen2022). For more on data availability, including how to potentially access the individual-level data, see the Data Availability Statement below.
Measuring Expressed Partisanship
We measure expressed partisanship based on partisan terms in users’ bios. To identify these terms, we apply a keyword expansion algorithm, which is shown to be superior to ad hoc keyword selection for social media data (King, Lam, and Roberts Reference King, Lam and Roberts2017). Beginning with a minimal set of seed words (“Democrat” and “Republican”), we identify relevant terms that users would include in their profiles to explicitly indicate their partisanship (or remove to deidentify from it). We detail this procedure in Appendix A. For robustness, we also run analyses using only the terms “Democrat” and “Republican”.
By relying on users’ Twitter profiles, our measure of partisanship departs from traditional survey-based measures. To highlight its expressive quality, we refer to it as “expressed partisanship.” However, there are reasons to expect the measure to track users’ partisan loyalties. First, our measure has high face validity: it stands to reason that publicly expressing support for a party or associated movement is an indication of identification. Second, in Appendix B we present two analyses validating our measure against the content of users’ tweets, another behavioral manifestation of partisanship. We first collect a total of 16.5 million tweets from a random subsample of Republican- and Democrat-identifying users. We then demonstrate a strong correlation between expressed partisanship and tweet sentiment: users identifying as Republican tweet less negatively about their own party than about the Democratic Party and vice versa. Then, in a more exploratory approach using supervised machine learning, we demonstrate that the terms in users’ tweets that are most predictive of Republican and Democratic expressed partisanship are highly politically loaded (e.g., among the terms most predictive of identifying as Republican are “msm” (“mainstream media”), “communist,” and “swamp”). These checks indicate that our measure of expressed partisanship picks up meaningful variation in users’ partisan loyalties.
Finally, the removal of Republican partisan terms reflects a distancing from the Republican Party but may be animated by different motives. Changes in these expressions may reflect a weakening identity, but they may also reflect that the social costs of associating with the Republican Party has increased (i.e., an act of “preference falsification”). Although we cannot address these motives directly, we do provide some tentative evidence in auxiliary analyses.
Difference-in-Differences Model
To estimate the effect of the Capitol insurrection on Republican partisan deidentification, we apply a flexible difference-in-differences (event study) model to data collected within a 10-day window around the event from users whose profiles include a Republican or Democratic keyword on at least one day within this period. This allows us to capture the dynamics of the effect of the insurrection on Republican partisan identification relative to Democrats, the natural counterfactual group, in the immediate aftermath of the event. Event study estimates also allow us to visually assess the parallel trends assumption for causal identification. More formally, we use the following model:
$$ {y}_{it}\hskip0.35em =\hskip0.35em {\alpha}_i+{\lambda}_t+\sum \limits_{t=-10}^{10}{\beta}_t{\mathrm{Republican}}_i\times {\mathrm{Day}}_t+{\varepsilon}_{it}, $$
where the outcome
$ {y}_{it} $
is a binary variable indicating whether user
$ i $
’s profile contains a partisan-identity keyword on day
$ t $
, and
$ {\mathrm{Republican}}_i $
is a binary variable indicating whether user
$ i $
’s partisan identity as measured by keyword use during the period is Republican (Republican = 1) or Democratic (Republican = 0). The parameters of interest,
$ {\beta}_t $
, are day-specific interaction coefficients that capture the difference in differences between partisan-identifying Democrats and Republicans on a given day
$ t $
relative to the day before the insurrection (
$ t=0 $
), which we set as a baseline. User and day fixed effects are α
i and λ
t, respectively. Finally, it is possible that any observed effect is driven by an increase in Democratic identity. However, as we show in Appendix D, we observe no major discontinuity in Democratic identification. The results below thus appear wholly driven by changes in expressions of Republican identification.
Results
To begin, Figure 1 shows descriptively the daily net changes in the number of users who indicate a Republican identification across the entire data collection period. Immediately after the insurrection, we see a dramatic net decrease in users identifying with the Republican Party and President Trump, with a continued net deidentification over the following two months. In the three weeks immediately following the insurrection, a substantial 1 in 15 users (7%) remove Republican-identifying terms. This compares with just one in 108 users (1%) with Democratic terms. For comparison, during the three weeks before the insurrection, deidentification for Republican and Democratic terms was essentially equivalent (~0.5%). The postinsurrection drop is also far more pronounced than that following the 2020 presidential election, which is roughly equal among Republicans (2%) and Democrats (2%).
Figure 1. Daily Net Change in Republican Party Identification from June 2020 to March 2021
Note: Values below zero indicate a net decrease in users with Republican identity terms compared with the previous day. LOESS regression included for reference.
Because Figure 1 considers only Republicans, observed changes may reflect cross-partisan alienation from politics rather than reactions specific to Republicans. We thus fit our difference-in-differences model to examine changes in expressions of party identification among Republican- versus Democrat-identifying users.Footnote 1
Figure 2 presents the model estimates, using data from a 10-day window around the insurrection. Each point represents the difference between Republican and Democratic users in the predicted probability of party identification relative to the preinsurrection baseline. Negative values indicate that Republican users deidentify more (i.e., drop all party-related terms) on a given day relative to Democrats.
Figure 2. Event Study Estimates (with 95% CIs)
Note: Data were collected each morning, and thus observations on January 6 (before vertical line) are preinsurrection. Standard errors are clustered at the user level.
Figure 2 clearly shows that before the insurrection, Republican and Democratic users changed their expressed identification to a similar extent. This implies an absence of preinsurrection differential deidentification, which substantiates the parallel trends assumption. After the insurrection, the change is clear and dramatic. Within a few days, Republican users were on average 2 percentage points less likely to express a partisan identity relative to Democrats than they were before the insurrection. This relative difference increases to around 4 percentage points within a week and a half. These estimates, based on a narrower time frame than shown in Figure 1, imply that within this short period, roughly one in 25 Republican identifiers removed markers of partisan identification from their biographies. This result is substantively equivalent when examining the count of partisan terms in users’ bios (Appendix E), and it is robust to alternative choices of partisan keywords (Appendices F and H).
Partisan terms indicating identification with Trumpism (e.g., “Trump,” “MAGA”) are more frequent than those referencing the Republican Party itself, which raises the question of whether the effect is wholly driven by deidentification with Trump rather than the party. Unsurprisingly, users overlap in their use of these terms (see Appendix F). However, by considering only the terms “Republican” and “Democrat”, we can gauge whether the effect is driven solely by deidentification with Trumpism. As we report in Appendix F, the effect is similar, although diminished, when using only party labels.
A natural follow-up question to this set of findings is whether those who deidentified as a result of the insurrection reidentified shortly afterward. Figure 3 demonstrates that this is not the case by presenting the frequency of reidentification among Republican users who deidentified in the first week after the insurrection. By the end of our data collection period—almost two months after the insurrection—only 6% of deidentifiers had reidentified. The observed Republican deidentification was thus not merely an ephemeral shift but appears to be persistent, at least in the short term. Last, as we show in Appendix M, users identifying by more moderate terms (using only the term “Republican”) before deidentifying were relatively less likely to reidentify, although the difference is not statistically significant.
Figure 3. Percentage of Republican Deidentifiers who Reidentified by the End of the Data Collection Period
Note: Data include any Republican-identifying user who deidentified within a week after the Capitol insurrection.
We conducted a number of robustness tests and auxiliary analyses, most of which are reported in the appendix. First, we consider whether the certification of the presidential election may have driven the observed deidentification, perhaps out of aversion to being on the losing side. We cannot entirely rule out this alternative mechanism, but two observations speak against it. First, when restricting the data to those collected before the official certification (early morning on January 7), we find that deidentification begins before the certification, and the observed effect precertification on January 7 is essentially identical to that of the full sample used in Figure 2. This suggests that the trend toward deidentification started prior to the certification. Second, as Figure 1 shows, Republican deidentification in the wake of the election, where Trump’s defeat was widely announced, was much smaller than was the subsequent deidentification after the insurrection. Although some supporters may still not have accepted the outcome at this point, we find it implausible that the certification could lead to Republican deidentification on a much larger scale than the election, especially because the most committed partisans are less likely to change their partisan identity in general (Green, Palmquist, and Schickler Reference Green, Palmquist and Schickler2004).
Second, another alternative explanation is that users changed their bios out of fear of legal prosecution (including but not limited to actual rioters). Finding a significant drop when considering only the party label speaks prima facie against this alternative explanation. However, we can address this more directly by excluding users who scrubbed their timelines of potentially incriminating tweets, here proxied by removing tweets on the same day that they dropped Republican partisan terms from their bio. In Appendix I we show that the findings are robust to excluding these users.
Third, one concern might be that the results could be driven by Twitter’s deletion of accounts related to QAnon—a loosely knit group of political conspiracy theorists, some of whose profiles may overlap with the set of Republican-identifying users—in the weeks after the insurrection (Singh Reference Singh2021). This is not the case because deleted accounts do not affect our event study estimates on any given day, as we only exploit within-user variation and deleted users who drop out of the sample are not coded as deidentifying. Another possibility is that users may have preemptively scrubbed their timelines and profiles to potentially prevent detection. In Appendix J, we show that our findings are substantively unaffected when excluding any user-day observations for users whose accounts were deleted, suspended, or made private following the insurrection. Moreover, excluding users whose bios on the eve of the insurrection include terms “qanon,” “wwg1wga” (a QAnon acronym), and the related “#StopTheSteal,” produces effectively identical results to those in Figure 2.
Fourth, as noted in the introduction, we interpret the effect as a national-level shock rather than one driven by users geographically close to the insurrection. In Appendix K, we substantiate this by demonstrating that the effect is unchanged when excluding users geolocated to Washington, DC, and neighboring states.
Fifth, to gauge whether the observed deidentification is driven primarily by increased social costs of affiliating with the Republican Party (i.e., an act of preference falsification) as opposed to a weakened party identity, we compare event study models among users whose user names match and do not match a first name in US Social Security Administration records (as a proxy for potentially being identifiable, see Appendix L). If increased social costs were the primary animating motive, one would expect users who use a real name—and therefore bear higher costs due to potentially being identifiable—to be more likely to deidentify. We do not find this to be the case. This suggests that the observed effect is at least partly driven by a weakening of identification with the Republican Party.
Sixth, in Appendix G we consider mentions of political parties in users’ tweets as an alternative outcome. Consistent with our main finding, we show that Republican deidentifiers make fewer references to parties in their postinsurrection tweets, a decrease driven by fewer references to Republicans.
Seventh and finally, in Table C1 in Appendix C we show that Republican deidentifiers are more active (e.g., tweet more) and more connected (e.g., have more followers) on Twitter than do the average Republican identifiers. This implies that this subset of users may have a broader influence than is reflected by their raw numbers.
Conclusion
Studying the effect of the US Capitol insurrection on expressed partisanship, we find that the insurrection caused a substantial number of Republican partisans to actively remove expressions of identification with the Republican Party and Donald Trump in its immediate aftermath. Our findings add to our understanding of the effects of violent protests on mass political behavior in several ways. Complementing studies of historical cases of left-wing protest, we provide evidence of the effects of violent protests in a contemporary setting and on the political right. Furthermore, we show that deidentification in response to the insurrection is nationalized—that is, it extends beyond its immediate geographical context. Last, we document that this immediate effect persists in the short term, with only a small minority of deidentifiers reidentifying during the following two months.
More broadly, our findings suggest that extreme events, such as those that violate democratic norms, can drive even some avowed partisans to distance themselves from their party. In the context of the ongoing debate about the negative consequences of polarization in the United States (Finkel et al. Reference Finkel, Bail, Cikara, Ditto, Iyengar, Klar and Mason2020; Iyengar et al. Reference Iyengar, Lelkes, Levendusky, Malhotra and Westwood2019), this finding is encouraging, as it suggests that there are limits to partisan loyalty. Our results thus complement recent work finding that exposure to incivility in same-party media leads partisans to distance themselves from their party (Druckman et al. Reference Druckman, Gubitz, Lloyd and Levendusky2019). Furthermore, this carries the positive democratic implication that political violence potentially deflects and demobilizes at least some copartisans, raising the political cost of using such tactics.
Nevertheless, the potentially positive conclusions from our study should not be overstated. Expressed deidentification is a potential indicator of distancing from the party and its leader, but it does not imply that partisan identities are no longer salient or consequential. Republican politicians’ responses to the insurrection, for example, resulted in a struggle over the meaning of the party’s identity, rather than its abandonment. A minority of radical partisans, as in the Capitol insurrection, may also use violence for their own ends despite its potential costs to a political party. Therefore, our results should be seen as contributing to an evolving understanding of the conditions under which partisanship may be curbed or amplified in an age of polarization (Druckman et al. Reference Druckman, Gubitz, Lloyd and Levendusky2019; Finkel et al. Reference Finkel, Bail, Cikara, Ditto, Iyengar, Klar and Mason2020; Iyengar et al. Reference Iyengar, Lelkes, Levendusky, Malhotra and Westwood2019; Kalmoe and Mason Reference Kalmoe and Mason2022).
Relatedly, because this study examines the short-term consequences of the insurrection, it helps contextualize the longer-term efforts by Republicans to minimize the insurrection’s political costs. In its immediate aftermath, for example, Senate majority leader Mitch McConnell was highly critical of the insurrection and Donald Trump’s role in fomenting it. Five months later—citing the insurrection’s political costs—he sought to block a bipartisan commission designed to investigate it (Fandos Reference Fandos2021). Many rank-and-file Republican politicians, furthermore, sought to deemphasize the violence of the insurrection and obscure its partisan origins. A pressing question for future research thus concerns whether and how political elites are able to minimize the political costs of violence by strengthening the partisan identities that may have been weakened as a result of antidemocratic violence.
It is also worth considering the scope conditions of our findings. For one, they reflect the affordances and user base of a popular social media platform, Twitter. Understanding expressed partisanship and other behaviors on emerging fringe platforms (e.g., Parler), where some users may have moved their political communication after Twitter’s deletion of QAnon-related accounts in the aftermath of the insurrection, is an important task for future research.
Moreover, as we have employed a novel behavioral measure of expressed partisanship, future research might examine the effect of the insurrection on other (behavioral) measures of partisanship (e.g., party registration) or investigate the downstream consequences of changes in expressed partisanship for other political behaviors and attitudes (e.g., voting behavior or policy support).
Given the short aftermath of the insurrection studied herein, it is also relevant to further investigate whether deidentification proved long-lasting, especially in light of changes in high-profile Republican politicians’ approach to the event.
Last, the generalizability of our findings across the political spectrum is also a pertinent question. Put more substantively, would we expect to see a parallel deidentification among Democrats in response to the Black Lives Matter protests following the May 2020 police killing of George Floyd? Although our data collection began too late to test this directly, we conjecture, based on recent work, that this is not the case. For one, some recent work suggests that political responses to protest are contingent on protesters’ group status (Manekin and Mitts Reference Manekin and Mitts2022). Moreover, the George Floyd protests specifically appear to have prompted strong Democrats to become more liberal in their evaluations after the protests (Reny and Newman Reference Reny and Newman2021), suggesting that, if anything, we may even expect an uptick in identification with Democrats after the Black Lives Matter protests. Future research would do well to further scrutinize these and other contemporary violent protests to provide an understanding of the political behavioral consequences of such significant events.
Supplementary Materials
To view supplementary material for this article, please visit http://doi.org/10.1017/S0003055422001058.
DATA AVAILABILITY STATEMENT
Documentation and data for replicating results from the paper are available at the American Political Science Review Dataverse: https://doi.org/10.7910/DVN/EHH9XT.
The posted replication code and data reproduce the aggregate-level results in the article. As a consequence of the EU-wide General Data Protection Regulation, we are unable to share individual-level data in the replication materials as they are personal data, which enable identification of individuals. Thus, although we provide the code necessary to replicate all figures and tables in the article, we cannot make public the data needed to replicate the individual-level results. For more details, please see the documentation at the Dataverse. The individual-level data can potentially be accessed by other researchers for replication if (1) permission is obtained from the Danish Data Protection Agency and (2) a data transfer agreement is signed with the University of Copenhagen.
ACKNOWLEDGMENTS
The authors would like to thank participants at the Annual Conference of the NYU Center for Social Media and Politics, members of the Aarhus Research on Online Political Hostility project, members of the Political Communication group at Vrije Universiteit Amsterdam, and Charles Breton and Evelyne Brie for their thoughtful comments and suggestions on early versions of this manuscript.
FUNDING STATEMENT
This research was funded by the Independent Research Fund Denmark (grant number: 9038-00123B).
CONFLICT OF INTEREST
The authors declare no ethical issues or conflicts of interest in this research.
ETHICAL STANDARDS
The authors affirm this research did not involve human subjects.
Open access
