An Eye for an Eye? An Integrated Model of Attitude Change Toward Protest Violence

ABSTRACT

How political violence emerges, why people support it, and how authorities can address it without escalating further radicalization remain an ongoing debate. In this study, we develop a quantitative model to predict violence endorsement as a function of repression severity, group identification, and individual emotions. We situate our investigation in the context of the Anti-Extradition Law Amendment Bill Movement in Hong Kong, during which the public acceptance of violence increased substantially, contrasting the city’s long history of peaceful protests. Results show that violence endorsement is associated with repression severity in a U-shaped fashion. While limited repression deters violence endorsement, excessive repression crossing a proportionality threshold escalates it. Group identification is a salient moderator that amplifies the backfire effects of repression. People who are more attached to protesting groups are more vigilant about repression and more supportive of protest violence. Furthermore, we also find that individual emotions exert more significant influences than repression and group identification. Anger, disgust, and fear can result in radicalized opinions. These findings unpack the complex and multifaceted communicative processes that shape the perceptions of protest violence. In contrast to the rational, organizational, and structural models of social movements, we argue that this analytical framework can offer more insights into protest dynamics amid the increasingly personalized political participation.

KEYWORDS:

• Attitudinal radicalization
• protest violence
• repression
• emotion
• computational methods
• anti-extradition law amendment bill movement
• Hong Kong

Introduction

Public order and nonviolence are social norms upheld by most societies during peacetime (Bond, 1997; Schwartz, 1992). Such norms can be overturned when institutionalized approaches to social demands are blocked and perceived legitimate calls for social changes are thwarted. From France, Hong Kong, Myanmar, Belarus to Chile, numerous violent clashes between state agencies and dissenters have broken out in recent years. As recorded in the Armed Conflict Location and Event Dataset,¹ 47.7% of social unrest events between 2018 and 2020 involved violence.

Despite moral presumptions against violence in general, protest violence has been accepted and justified by an increasing number of people. A recent study found that approximately 12% of Swiss adolescents agreed that violent acts, attacks, and kidnappings are sometimes necessary for a better world (Nivette et al., 2017). In the United States, 15% of respondents agreed that violence was justified for advancing political goals, and 40% agreed that violence could be justified if opponents acted violently first (Mason & Kalmoe, 2021). A debate has ignited on how violence emerges, how it is supported, and whether repression deters or provokes violence.

A range of seemingly rival theories have attempted to explain protest radicalization and violence endorsement. Some focus on macro-level approaches linking political violence to regime types and state responses (Della Porta, 2006; Lichbach, 1987; Tilly, 2003). Some examine the impacts of relational factors, such as partisanship and political polarization (Hsiao & Radnitz, 2006; P. G. Klandermans, 2014; Mason & Kalmoe, 2021). Others investigate individual psychological variation in emotions (Ayanian et al., 2020; Mackie et al., 2000) and aggressive traits (Kalmoe, 2014). Despite their insights, these accounts are grounded in disparate backgrounds, fragmented into different topic areas, and emphasize different aspects of the same process while neglecting the others (Della Porta & LaFree, 2011). Our accumulated knowledge remains inconsistent because of the epistemological fragmentation. As Tarrow (2022) points out, “how to combine structure, culture, and process is one of the next issues on the agenda of contentious politics” (p. 484) and there is substantial scope for synthesizing structural realities, relational dynamics, and psychological motivations in studying social movements.

To fill this gap, we aim to develop an integrated model of violence endorsement that factors in repression severity (situational factors), group identification (relational factors), and individual emotions (psychological factors). We assume that 1) repression that exceeds a proportionality threshold triggers backfire, provoking more positive attitudes toward protest violence; 2) group identification amplifies the backfire effects of repression; and 3) increased endorsements of violence are accompanied by emotional impulses of anger, disgust, and fear.

We choose the Anti-Extradition Law Amendment Bill (Anti-ELAB) Movement in Hong Kong as our research context. What makes the Anti-ELAB Movement special and relevant to our inquiry is that it has witnessed a drastic change in the public acceptance of violence in a city with a long history of peaceful protests. Over the last two decades, the protest events in Hong Kong have been overwhelmingly self-restrained. A moderate political culture that embraces public order and the rule of law is deeply rooted in society (Ku, 2009; Yuen, 2018). The peaceful protests achieved several landmark milestones by pressing authorities to make concessions in national security legislation and national education policy in 2003 and 2012, respectively. Later, in 2014, another remarkably peaceful movement, Occupy Central, mobilized sizable groups of people to occupy the central business district for 79 days. Despite scattered confrontations, the police-protester relationship remained largely congenial (Cheng & Chan, 2017; Stott et al., 2020). According to an opinion survey, before the Anti-ELAB protests, only 6.8% of Hong Kong citizens deemed violent protests acceptable (Lee et al., 2021). This nonviolent tradition, however, was overthrown amidst the widespread indignation and disillusion with the legal and political systems during the Anti-ELAB Movement. As the protest movement unfolded, the clashes between police and protesters became increasingly common. The percentage of citizens who deemed violent protests understandable surged to 55.7% and peaked at 68.4% at the height of the conflicts (Centre for Communication and Public Opinion Survey, 2020). This upsurge defies the commonly held belief that violent protests erode popular support (Chenoweth & Stephan, 2011) and provides a valuable opportunity to examine factors contributing to the rise of violence endorsement.

In practice, measuring attitude change requires longitudinal data. Panel surveys are very costly and subject to nontrivial nonresponse and social desirability bias, when it comes to sensitive issues like violent actions. Therefore, we opt to join a growing body of observational studies that have benefited from using online data to evaluate public opinions. We collected 390,841 posts about the Anti-ELAB protests from LIHKG, the central communication channel for protesters and protest sympathizers. Based on this data set, we trained supervised machine learning models to detect each post’s violence endorsement, group identification strength, and emotional intensity. Then, we conducted regression analyses using the first-difference estimation strategy to assess how offline political events, emotional impulses, and group identifications drive people to deviate from their original levels of violence endorsement. We hope our study can shed light on the nuances of attitude change in networked social movements and reconcile contrasting theories by providing systematic evidence that encapsulates multilevel factors.

Repression, Deterrence and Backfire

Police actions and protest behaviors reciprocally adapt to each other (Della Porta & Tarrow, 2011). Although civil disobedience strategies have been widely adopted, protesters tend to have divergent opinions on the efficiency and morality of the use of violence and adjust their opinions swiftly in response to on-site repression (Maguire et al., 2021; Steinert-Threlkeld et al., 2022).

The extant literature underlines that the relationship between repression and dissent must be examined in relation to the presence or absence of legitimacy (Jackson et al., 2013; Maguire et al., 2018, 2021; Opp & Roehl, 1990; Tyler, 1995). According to procedural justice theory (Tyler, 1995), if police use excessive violence, citizens’ sense of obligation to obey the law will decrease, and public tolerance of protest violence will increase. Greene (1974, as cited in Lichbach & Gurr, 1981) also noted that violent repression lowered state legitimacy and stimulated a spiral of radicalization that impelled apathetic individuals to become politicized, reformers to become radicalized, and revolutionaries to redouble their efforts. Drawing on a sample of male ethnic minority youth in London, Jackson et al. (2013) find that negative interactions with police are associated with diminished legitimacy judgments, which in turn lead to positive attitudes about protest violence. Using survey data on the Occupy Wall Street movement, Maguire et al. (2021) reveal that people who perceive police use of force as unjust are more likely to endorse protest violence. Similarly, Schock (1999) documented that violent and indiscriminate repression in the Philippines and Burma diminished state legitimacy and provoked antiregime activities. Although the acceptable limits of repression may vary from context to context, indiscriminate repression is typically more likely to instigate rather than prevent support for protest violence.

However, the studies cited above focus exclusively on violent repression while neglecting another prominent type of repression – legalistic repression. Legalistic repression involves mass arrests, high bail amounts, protest bans, and other forms of legal harassment (Barkan, 1984; Earl, 2003). In contrast to violent repression, legalistic repression is less confrontational and more targeted at radicals. It gives protesters more procedural guarantees under the law. We expect legalistic repression to claim more legitimacy and be more effective for deterring violence endorsement than violent repression. In line with this proposition, Barkan (1984) found that in cities where police used legal means and avoided violence, civil rights movements were more likely to be defeated.

Legalistic repression might face the same kind of backlash as violent repression when its legitimacy is no longer sustained. For instance, Shi (2019) finds that exposure to the threat of arrest can inhibit radical actions while promoting radical attitudes among Chinese peasant workers, heightening their distrust of and moral indignation toward authorities. In a study on German extreme right movements in the 1990s, Koopmans (1997) indicated that both arrests and police violence were counterproductive due to a lack of legitimacy.

Based on the existing literature, we propose that there exist invisible thresholds of proportionality for both types of repression, which constitute the normative boundaries of legitimate repression. Low levels of repression initially induce obedience, frighten potential opponents, and deter violence endorsement. Beyond the threshold of proportionality, repression fosters a strong sense of procedural injustice, motivating people to test alternative means to redress the oppressive status quo. Higher and higher levels of repression will become more and more counterproductive, stimulating more dissent than compliance (Lichbach, 1987). The relationship between repression severity and violence endorsement is expected to form a U-shaped curve resembling a boomerang, which can be formulated as follows:

H1 (Deterrence below Threshold): State repression can decrease violence endorsement when its severity is below acceptable limits.

H2 (Backfire above Threshold): State repression can increase violence endorsement when its severity exceeds acceptable limits.

Violence Endorsement from a Group Perspective

The social process of group identification through which individuals come to share similar opinions is also critical to understand individual attitudes toward protest violence (P. G. Klandermans, 2014; Kruglanski et al., 2014). The elaborated social identity model (ESIM; Drury & Reicher, 2016, 2009; Reicher, 1996; Stott et al., 2018) suggests that group identification forms the psychological basis for collective norms, motivating individuals’ moral reasoning of what is right or wrong, especially in an “us-them” intergroup context. We assume that group identification can increase endorsement of violence through both direct and indirect pathways.

Directly, politicized collective identity is arguably a powerful engine for radicalization (P. G. Klandermans, 2014; Simon & Kiel, 2001; Van Stekelenburg & Klandermans, 2013; Van Zomeren et al., 2008). On the one hand, Amira et al. (2019) and Mackie et al. (2000) indicate that people with strongly politicized collective identities are more committed to intergroup conflicts. When protesters consider themselves part of a group rather than solitary individuals, their collective ethos may overrun cool calculations of self-interest (Mackie et al., 2000; Turner et al., 1987) and impel them to embrace whatever means that is effective for advancing group goals (Ng & Fung, 2016). On the other hand, groups can also shield people from their fear of punishment (Le Bon, 2002). Based on a thematic analysis of 41 interviews, Stott et al. (2018) find that a sense of “us” can foster an increased capacity for collective violence and persuade people to believe that they will share risk with others. Hence, inhibition of violence will be mitigated if group identification is salient. Following this line of reasoning, we propose that group identification is positively associated with violence endorsement.

H3 (Direct Effect of Group Identification): Individuals’ social identifications with protesting groups are positively associated with their endorsements of violence.

In addition to its direct effects, group identification may also interact with repression severity to jointly enhance violence endorsement. Group identification engenders altruism and increases affection among peer group members. Such affection makes people more sensitive to repression and more inclined to resort to violence when confronted by opponents (Opp & Roehl, 1990; Stott et al., 2017). Additionally, repression can also fuel protest sympathizers’ motivated reasoning by offering a justification that affirms their preexisting stances. Motivated reasoning theory asserts that people are inclined to reason through their identities, but their ability to do so is constrained by their ability to construct justifications for their desired conclusions (Kunda, 1990; Redlawsk, 2002). It means protest sympathizers are predisposed to support protest violence only to the extent that evidence of repression permits. Even if a person wants to support violent protests, he or she may refrain from doing so when the evidence does not suffice to justify it. State repression thus offers a justification to the pro-violence motivated reasoning. Therefore, group identification can function as a sensitizing and justifying apparatus that amplifies the backfire effects of repression. Repression can trigger greater backlash, spurring endorsement of protest violence, as people become more attached to protesting groups.

H4 (Moderating Effect of Group Identification): Individuals’ social identifications with protesting groups can amplify the backfire effects of state repression.

Emotions as Affective Predictors of Violence Endorsement

Finally, conflicts bring up many, usually negative, emotions. As James Jasper put it, “emotions are present in every phase and every aspect of protest” (2011, p. 286) and it is emotions that “give ideas, ideologies, identities, and interests their power to motivate” (Jasper, 1998, p. 127). Emotions are even more important in the era of personalized politics where people are commonly seeking emotionalized experience from politics (Richards, 2010). Existing studies on emotions of protest have discussed the influence of anger extensively yet relatively little attention has been paid to other forms of emotions such as fear which can be equally important as anger during the social unrests in non-democracies (Asun et al., 2020; Jasper, 2011). In this section, we review the literature on anger, disgust, fear, and sadness to construct three research questions and one hypothesis about their potential influences on violence endorsement.

Among the emotions, anger is widely considered the prototypical aggressive emotion (Izard, 2013; Van Stekelenburg & Klandermans, 2017). Anger, by definition, is a high-arousal others-condemning emotion (Brady et al., 2017; Haidt, 2001). Many scholars argue that anger can make people more attentive to provocative content (Cohen et al., 1998) and incite aggressive actions that offend, attack or humiliate agents who pose a threat to individuals (Haidt, 2001; Mackie et al., 2000). However, other researchers contend that anger is a high-efficacy emotion with reconciliatory intentions. Angry people tend to believe they are capable of coercing changes through less destructive means. In this view, anger is more related to peaceful actions than to offensive actions (Tausch et al., 2011; Van Stekelenburg & Klandermans, 2017). That being said, research on the relationship between anger and violence endorsement is inconclusive. Therefore, we frame this relationship as a research question that warrants further investigation.

RQ1: Does anger increase endorsement of protest violence?

Disgust² encapsulates both other-condemning and other-belittling sentiments. Disgust typically arises when anger-driven actions fail to solve a problem (Fischer & Roseman, 2007). In contrast to anger, which wanes quickly once anger-inducing behaviors cease (Ohbuchi et al., 1989), disgust is longer lasting and based on more stable and fundamental assessments of others (Fischer & Roseman, 2007; Hutcherson & Gross, 2011; Matsumoto et al., 2012). Mackie et al. (2000) and Pond et al. (2011) argue that disgust deters violence by turning people away from offensive entities. Similarly, Fischer and Roseman (2007) assert that disgust might entail social exclusions and avoidances but not confrontational actions.

Disputing deterrence theory, Matsumoto et al. (2012) insist that disgust escalates violence tendencies much more than anger. According to their theory, disgust renders its source morally inferior with no chance of rehabilitation. Anger impels people to seek changes in “bad” acts, while disgust impels people to seek the elimination of “bad” persons. Hence, they posit that it is disgust, not anger, that contributes to the breakdown of relationships and the emergence of intergroup violence. To support this proposition, they draw on speeches by ideological leaders, such as Hitler, to show how an escalation of disgust leads to violent conflicts. Furthermore, Tausch et al. (2011) and Van Stekelenburg and Klandermans (2017) affirm that disgust³ consistently predicts support for nonnormative violent actions. As the arguments and evidence have been offered for both deterrence and backfire effects of disgust, we opt to formulate disgust’s impact on protest violence as a research question:

RQ2: Does disgust increase endorsement of protest violence?

Fear is widely regarded as the major source of law-abiding and norm-respecting behaviors that reduces the probability of violent actions or radical thoughts (Ayanian et al., 2020; B. Klandermans et al., 2008; Mackie et al., 2000; Van Stekelenburg & Klandermans, 2013). Experimental evidence demonstrates that fear leads to more pessimistic perceptions of risks and reduces dissents in both actions and expressions (Young, 2019). Rozenas and Zhukov (2019) find that in fear of retribution, Ukrainians exposed to the “terror by hunger” became more loyal to Moscow. Taylor and Whittier (2020) also noted that in women’s movements, fear made unpleasant memories of victimization easy to recall and undermined commitments to protest. Therefore, successful mobilization usually requires protest organizers to strategically downplay fear (Goodwin & Pfaff, 2001; Johnston & Carnesecca, 2014; Taylor & Whittier, 2020). These studies consistently describe fear as an incapacitating emotion that deters opposition, reduces dissent, and induces compliance.

However, others inferred that fear might preemptively radicalize opinions and cause self-defensive aggressions (Simunovic et al., 1992). As Nina Simone put it, “Freedom means no fear.” People can support violent protests because they are tired of living in fear and want to pursue the freedom to act fearlessly, with violent protests as a last resort. Thus, people with fears may have more grievances and more incentives to support violent protests than unafraid people (Bar-Tal, 2001; Jarymowicz & Bar-Tal, 2006). Moreover, Jarymowicz and Bar-Tal (2006) argue that fear as a primary emotion could flood consciousness, limit self-regulation, and activate human instincts to embrace primitive ways to resolve conflicts.

RQ3: Does fear increase endorsement of protest violence?

Sadness, one of the universal moral sentiments endorsed by Adam Smith (2010), usually arises from one’s own or others’ miserable experiences. Violent protests are typically laden with sadness because protesters are hurt, beaten, intimidated, and even killed in violent protests. Many scholars argue that sadness is an emotion of withdrawal and avoidance (Roseman, 2011). As Berkowitz (1993) vividly put it, sadness can throw people into a “blue funk where they don’t want to do anything and certainly don’t want to assault others openly.” The civil right movement leader Malcolm X asserted that “when people are sad, they don’t do anything. They just cry over their condition.” Therapeutic wisdom in traditional Chinese medicine also holds that sadness counteracts anger, and patients who have illnesses caused by anger can be cured by inducing sadness (Jiang & Wei, 1996). Leveraging experimental evidence, Lutz and Krahé (2018) assert that sadness can offset impulses of anger and prevent anger-driven aggressive behaviors. In addition, Zhan et al. (2017) find that their participants acted less aggressively when exposed to sadness-inducing comments and movie clips. Therefore, we propose that sadness can mitigate people’s support for violence.

H5: Sadness mitigates endorsement of protest violence.

Methodology

Data

The Anti-ELAB protests are widely characterized as a decentralized, or as some may say, “leaderless,” social movement, which involves a loosely knit alliance of participants. Elite actors did not claim the legitimacy and authority to represent, command, or broker the masses (Cheng et al., 2022; Liang & Lee, 2021). In place of them, digital communication platforms like LIHKG emerged as the hub for mobilization and cooperation (2022 Lee, Liang, et al., 20212021; Urman et al., 2021). Numerous protest events, including the siege of the LegCo complex on June 12 and the citywide general strike on August 5, were first initiated by a few ordinary netizens and gained unexpected momentum with the aid of digital platforms. LIHKG is thus strategically critical for the development of the Anti-ELAB protests and it accommodates a large number of tactical negotiations and emotional expressions which can effectively inform our model and illuminate the nuances of the attitudinal radicalization process.

However, LIHKG data are admittedly unrepresentative of the entire population. LIHKG users are more inclined to support the Anti-ELAB protests and to be more attentive to protest events than the general public. Nevertheless, given that approximately 60–70% of Hong Kong citizens support the protests (Centre for Communication and Public Opinion Survey, 2020; Hong Kong Public Opinion Research Institute, 1999) and an estimated 2 million, or two in seven of the city’s population, marched to show their support (BBC, 17 June 2019), this sample can still offer critical insights into the protests as well as the evolution of public opinions in the city. We used an automatic web scraper to collect 390,841 posts from LIHKG between June 1 and December 31, 2019.⁴

Violence Endorsement, Emotions, and Group Identification

We divided posts into sentences, extracted a random sample of 20,251 sentences,⁵ and recruited three coders to manually identify the violence endorsement, emotions, and group identification underlying each sentence (Supplementary Materials, Note 5). To establish intercoder reliability, we used a common set of 700 sentences to test the three coders and obtained Krippendorff’s alpha score of 0.79, suggesting good intercoder reliability. With guaranteed reliability, the three coders continued to code the remaining sentences in the sample.

The labeled dataset was then split into training, validation, and test sets by 60%, 20%, and 20%, respectively. Based on the training set, we developed six independent binary classifiers using recurrent neural network structures. All models yielded good performance metrics in the validation set (F1 scores: mean = 0.74, SD = 0.08, ROC AUC scores: mean = 0.91, SD = 0.03). Then, we applied the models to the entire dataset to predict every sentence’s scores for violence endorsement (VEndorsement), group identification (Identity), and emotions (Anger, Disgust, Fear, and Sadness). The predicted scores were dichotomized into 0 and 1 using 0.5 as the cutoff point. Finally, we calculated what percentage of sentences in each post were coded 1 and used the returned values as post scores. For example, if a post consists of 10 sentences and 7 out of the 10 sentences show significant support for violent protests, the magnitude of the post’s endorsement of violence would be 0.7.

To analyze attitude change toward protest violence, we needed to compare an individual’s current level of violence endorsement with the past level. The time interval between the current and past posts should not be arbitrarily long, since an overly long interval might include too many unwanted confounders beyond the scope of the investigation. Nor should it be too short, as an overly short interval would exclude too many users who do not post more than once within it and would cause data sparsity problems. Therefore, we focused on attitude changes that occur within 32 hours. We also tried out alternatives, namely 16-hour and 48-hour intervals, and affirmed that the results were not swayed by changes in time intervals (see Supplementary Materials, Note 1).

To measure attitude changes within 32 hours, we first divided each day into three 8-hour fractions⁶ according to sociotemporal regularities. Then, we ran a sliding window (Figure 1) of four fractions to scan through the entire timeline. The former two fractions (df1 & df2) constitute the past time frame where people’s past violence endorsement levels were measured (VEndorsement_t-1). If more than one post was created by the same person in the same time frame, their average would be used. The latter two fractions (df3 & df4) constitute the current time frame where people’s current violence endorsement levels were measured (VEndorsement_t). We calculated the difference between VEndorsement_t-1 and VEndorsement_t to obtain individuals’ attitude change ∆VEndorsement_t, which was the dependent variable in this study. Users who did not post within any two continuous time frames, such as users A and E in Figure 1, were excluded from further analysis. In addition to ∆VEndorsement_t, we also evaluated people’s emotional changes ($\mathrm{\Delta }Ange{r}_t$,$\mathrm{\Delta }Disgus{t}_t$,$\mathrm{\Delta }Fea{r}_{it}$, $\mathrm{\Delta }Sadnes{s}_t$) and group identification shifts ($\mathrm{\Delta }Identit{y}_t$) by following the same approach.

Figure 1. Examples of the sliding window. The window is composed of two adjacent time frames, one representing the current time frame (yellow) and the other representing the past time frame (green). Each time frame is composed of two fractions and spans 16 hours. The window was moved across the timeline at a pace of one fraction per step.

To test H1 and H2, we set the third fraction df3 as the focal fraction; evaluated the magnitudes of violent repression, legalistic repression, and protests in df3 (the method of measurement will be explained in the next section); and examined whether attitude change was associated with political events in df3. As the window moved across the timeline, all day fractions except the first and last two alternately played the role of df3, and all political events within this range were exhaustively tested.

This sliding window approach has two significant merits. First, it is equivalent to taking first-order differences in time series data. Given that some people might inherently be more inclined to violence than others, the first differencing design that exploits temporal variation within individuals can eliminate the endogeneity problem that arises from autocorrelation and individual-invariant omitted variables (Wooldridge, 2010). Second, people might not immediately respond when certain political events take place. In the current approach, the time frame spans 16 hours, and the window spans 32 hours. Response latency can be effectively incorporated to capture individual responses more fully.

Political Events

To measure repression severity, we referred to the official numbers of live rounds, tear gas, rubber bullets, bean bags, and arrests listed by the Hong Kong Police Force (HKPF) in their regular press conferences. We summed the first four numbers to obtain a general proxy for violent repression severity⁷ and used the numbers of arrests as the proxy for legalistic repression severity.

In addition to repressive events, we are also interested in examining and controlling for the influence of protest events. Therefore, we manually compiled a dataset pertaining to the number of participants in each protest and entered it into the model in parallel with the two types of repression events.

Control Variables

User Fame was gauged by the average number of replies (in thousands) that users received.

User Engagement was gauged by the total number of posts and replies that users created. As this value was positively skewed, we applied the logarithm transformation to conform it to normality.

Post Topics were gauged by a 30-dimensional vector returned from the Latent Dirichlet Allocation model trained on the complete dataset. Each post was represented by a 30-dimensional vector, where each dimension represented the probability that the post was related to a specific topic area.

Data Analysis: First-Difference Estimators

A regression model (Model 1) was implemented using a first-difference estimation strategy. The model can be represented as:

$\begin{array}{rl}& \mathrm{\Delta }VEndorsemen{t}_{it}=VEndorsemen{t}_{it}-VEndorsemen{t}_{it-1}\\ & ={\alpha }_1+{\beta }_{A1}\mathrm{\Delta }Ange{r}_{it}+{\beta }_{D1}\mathrm{\Delta }Disgus{t}_{it}+{\beta }_{F1}\mathrm{\Delta }Fea{r}_{it}+{\beta }_{S1}\mathrm{\Delta }Sadnes{s}_{it}\\ & +{\gamma }_{1}ProtestSiz{e}_t+{\delta }_1\mathrm{\#}Arrest{s}_t+{\mu }_1\mathrm{\#}Round{s}_t\\ & +{\theta }_1\mathrm{\Delta }Identit{y}_{it}+{\theta }_{P1}\mathrm{\Delta }Identit{y}_{it}\ast ProtestSiz{e}_t\\ & +{\theta }_{A1}\mathrm{\Delta }Identit{y}_{it}\ast \mathrm{\#}Arrest{s}_t+{\theta }_{R1}\mathrm{\Delta }Identit{y}_{it}\ast \mathrm{\#}Round{s}_t\end{array}$

where ∆VEndorsement_it is the difference between individual i’s current and past violence endorsement levels; $\mathrm{\Delta }Ange{r}_{it}$, $\mathrm{\Delta }Disgus{t}_{it}$, $\mathrm{\Delta }Fea{r}_{it}$, and $\mathrm{\Delta }Sadnes{s}_{it}$ are individual i’s emotional changes from time frame t-1 to t; and $\mathrm{\Delta }Identit{y}_{it}$ is the change in the strength of individual i’s identification with the protesting group.

Although the first differencing can effectively mitigate within-individual autocorrelation, it did not address the endogeneity caused by the hierarchical structure of data. That is, individuals’ radicalization shifts within the same time period are presumably correlated, and endogeneity can arise as error terms are clustered. To adjust for the cross-sectional dependence of observations, we ran another model (Model 2) with time fixed effects (Angrist & Pischke, 2008):

$\begin{array}{rl}& \mathrm{\Delta }VEndorsemen{t}_{it}={\alpha }_2+{\beta }_{A2}\mathrm{\Delta }Ange{r}_{it}+{\beta }_{D2}\mathrm{\Delta }Disgus{t}_{it}+{\beta }_{F2}\mathrm{\Delta }Fea{r}_{it}+{\beta }_{S2}\mathrm{\Delta }Sadnes{s}_{it}\\ & +{\gamma }_{2}ProtestSiz{e}_t+{\delta }_2\mathrm{\#}Arrest{s}_t+{\mu }_2\mathrm{\#}Round{s}_t\\ & +{\theta }_2\mathrm{\Delta }Identit{y}_{it}+{\theta }_{P2}\mathrm{\Delta }Identit{y}_{it}\ast ProtestSiz{e}_t\\ & +{\theta }_{A2}\mathrm{\Delta }Identit{y}_{it}\ast \mathrm{\#}Arrest{s}_t+{\theta }_{R2}\mathrm{\Delta }Identit{y}_{it}\ast \mathrm{\#}Round{s}_t+{\pi }_t\end{array}$

where ${\pi }_t$ is the fixed effect dummy of time frame t.

The last model (Model 3) includes control variables to account for post topics and user characteristics:

$\begin{array}{rl}& \mathrm{\Delta }VEndorsemen{t}_{it}={\alpha }_3+{\beta }_{A3}\mathrm{\Delta }Ange{r}_{it}+{\beta }_{D3}\mathrm{\Delta }Disgus{t}_{it}+{\beta }_{F3}\mathrm{\Delta }Fea{r}_{it}+{\beta }_{S3}\mathrm{\Delta }Sadnes{s}_{it}\\ & +{\gamma }_{3}ProtestSiz{e}_t+{\delta }_3\mathrm{\#}Arrest{s}_t+{\mu }_3\mathrm{\#}Round{s}_t\\ & +{\theta }_3\mathrm{\Delta }Identit{y}_{it}+{\theta }_{P3}\mathrm{\Delta }Identit{y}_{it}\ast ProtestSiz{e}_t\\ & +{\theta }_{A3}\mathrm{\Delta }Identit{y}_{it}\ast \mathrm{\#}Arrest{s}_t+{\theta }_{R3}\mathrm{\Delta }Identit{y}_{it}\ast \mathrm{\#}Round{s}_t\\ & +\sigma UserFam{e}_i+\tau UserActivit{y}_i+\sum _{n=1}^{29}{\omega }_{n}Topi{c}_{nit}\end{array}$

where $UserFam{e}_i$ and $UserEngagemen{t}_i$ represent the popularity and engagement level (or outspokenness) of user i and $Topi{c}_n$ represents the average possibility of user i’s posts in time frame t belonging to n^th topic. Since the sum of possibilities of 30 topics equals 1, to avoid collinearity problems, we drop one topic and enter only $Topi{c}_1$ to $Topi{c}_{29}$ into the model.

To assure the robustness, we also tested alternative models on different dependent variables. For details, please refer to Supplemental Materials Note 2 and 3.

Results

Descriptive Statistics

Figure 2 illustrates the numbers of daily posts on LIHKG. Discussion volume reached its peak on August 12, the day after the citywide violent clashes between police and protesters. Late on the night of August 11, photos and videos emerged of a female protester lying on the ground with blood streaming from her right eye. A bean bag round was seen lodged in her goggles. In the following days, tens of thousands of people were mobilized to join “an eye for an eye” protests (Cheung, 2019; Ho & Kilpatrick, 1996). As noted by many scholars, this incident was a milestone in the Anti-ELAB protests; after it, the protests’ central focus shifted from blocking controversial legislation to opposing police brutality (Stott et al., 2022). Mirroring the slogan “an eye for an eye,” many protesters started to act with escalated violence in retaliation for excessive repression by the government (Hui, 2020). In addition, other violent events, such as the Yuen Long Attack and Mong Kok Attack, also provoked fierce discussions on LIHKG. Legal and electoral events can also stimulate significant discussions, yet these discussions were lower in volume than those triggered by violent events. Generally, discussion on LIHKG has been highly reactive to offline events, especially violent incidents. This reaffirms our selection of LIHKG as a valid research setting for the current study.

Figure 2. Numbers of daily posts on LIHKG. Authors manually checked the days with extreme volumes of discussions, identified their main themes and annotated the peaks accordingly. The background colors of annotations indicate different types of political events (red: violent clashes between protesters and police/countermovement activists; gray: legal or electoral events; blue: mass protests).

For the mean values of key variables (Table 1), anger and disgust were the two most prevalent themes, accounting for 18% of online communication. Moreover, the bivariate correlation between them was the strongest (r_{anger_disgust} =.48). Thus, anger and disgust were closely correlated and frequently co-occurrent in posts. In addition, 17% of the content expressed solidarity with the protesting group, implying that Anti-ELAB has successfully commanded a high degree of solidarity from its supporters. Fear, sadness, and violence endorsement were highly right-skewed, with a few observations of extremely large values and an excessive number of small values. Nevertheless, their first-differenced counterparts are all distributed approximately normally with moderate skewness scores ranging between −1 and 1. In other words, although the expressed feelings are heterogeneous, the changes in the expressed feelings are homogenous. This somewhat justifies the efficacy of the first difference estimation strategy, which can place the analysis on more solid ground with normally distributed variables. Sadness, the least common theme at only 5%, might be a socially undesirable emotion that many people wittingly avoided disclosing.

Table 1. Descriptive statistics.

	Mean	SD	Skewness	VE*	Anger	Disgust	Fear	Sadness	Identity
VE*	.06	.18	4.04	1.00
Anger	.18	.31	1.76	.10	1.00
Disgust	.18	.32	1.80	.14	.48	1.00
Fear	.09	.22	3.19	.05	.13	.07	1.00
Sadness	.05	.17	4.45	.03	.09	.07	.19	1.00
Identity	.17	.30	1.83	.08	−.06	−.09	−.05	.01	1.00
	Mean	SD	Skewness	∆VE*	∆Anger	∆Disgust	∆Fear	∆Sadness	∆Identity
∆VE*	−.0007	.23	−1.00	1.00
∆Anger	−.002	.39	−.37	.07	1.00
∆Disgust	−.0003	.38	−.38	.10	.45	1.00
∆Fear	−.0007	.27	−.80	.04	.12	.06	1.00
∆Sadness	.0003	.21	−1.11	.02	.08	.05	.19	1.00
∆Identity	−.0006	.35	−.49	−.06	−.06	−.08	−.05	.01	1.00
#Users	63,256		#Posts per user		6.11
#Posts	386,534		#Posts per day		1,806

*VE stands for Violence Endorsement.

First-Difference Estimation Models

To better understand the nuances of the radicalization process, we developed three models, one with time-fixed effects, one without them, and one with additional control variables. The results are displayed in Table 2. Positive coefficients suggest backfire effects, while negative coefficients suggest deterrence effects. As shown in the table, the inclusion of time-fixed effects does not substantively affect the model results. Most coefficients in Models 1 and 2, when rounded to one significant figure, are the same and are close to those in Model 3. For the sake of clarity, we will mainly report the results of Model 1.

Table 2. First-Difference estimation models predicting violence endorsement.

	Model 1			Model 2			Model 3
Variables	b (S.E.)	t		b (S.E.)	t		b (S.E.)	t
Political Events
Log(Protest Size’)	.0008(.001)	.57		.0013(.001)	.87		−.0008(.001)	.57
Log(Protest Size’)^2	.0003(0)	1.17		.0002(0)	.69		.0005(0)	1.94
Log(#Arrests’)	−.0055(.003)	−1.86		−.0059(.003)	−1.84		−.0083**(.003)	−2.84
Log(#Arrests’)^2	.0020(.001)	1.47		.0014(.002)	.96		.0036**(.001)	2.70
Log(#Rounds’)	−.0090***(.002)	−4.21		−.0086**(.003)	−3.20		−.0084***(.002)	−3.96
Log(#Rounds’)^2	.0028***(.001)	3.52		.0029**(.001)	2.97		.003***(.001)	3.78
Group Identification
∆Identity	.049***(.002)	24.83		.048***(.002)	24.23		.043***(.002)	21.36
∆Identity * Log(Protest Size’)	.0004(.001)	0.29		.0004(.001)	.30		.00007(.001)	.06
∆Identity * Log(#Arrests’)	.0037(.003)	1.09		.0033(.003)	.97		.0044(.003)	1.30
∆Identity * Log(#Rounds’)	.0039**(.001)	2.69		.0042**(.001)	2.90		.0038**(.001)	2.66
Emotions
∆Anger	.021***(.002)	14.27		.021***(.002)	14.22		.015***(.002)	9.92
∆Disgust	.054***(.002)	35.74		.054***(.002)	35.25		.045***(.002)	29.41
∆Fear	.026***(.002)	13.64		.026***(.002)	13.41		.023***(.002)	12.00
∆Sadness	.0024(.003)	.93		.0017(.003)	.67		−.001(.003)	−.39
Time Fixed Effects	No			Yes			No
Control Variables
User Fame	−			−			−.020**(.008)	−2.61
User Engagement	−			−			−.0075***(.001)	−8.95
Topics (N = 30)	−			−			Yes
Constant	−.0005(.001)	−.72		−.0012(.003)	−1.14		.084***(.006)	14.26
N	192,611			192,611			192,611

* p < .05; ** p < .01; *** p < .001. Since the logarithmic transformation can only be applied to positive values, we add constant 1 to all observations of #Rounds, #Arrests, and Protest Size to escape zeros. The resulting values are re-named as #Rounds’, #Arrests’, Protest Size’.

H1-2: Political Events

Among all situational variables in the “Political Events” block, only Log(#Rounds) and its quadratic term Log(#Rounds)² have## significant coefficients. These two variables together form a U-shaped parabola (Panel 1, Figure 3) that opens upward with the axis of symmetry located at Log(#Rounds’) ≈ 1.61 (or #Rounds = 10^1.61 −1 ≈ 40) and crosses the horizontal zero line at Log(#Rounds’) ≈ 3.17 (or #Rounds = 10^3.17 −1 ≈ 1,478). Hence, as #Rounds increases but does not exceed 40, violent repression is increasingly effective in deterring violence endorsement. When #Rounds > 40, the deterrence effect decreases. Increasing levels of repression will generate diminishing returns for the government. When #Rounds > 1,478, the deterrence effect diminishes to zero, and the radicalization effect gains the upper hand. Violent repression beyond this level is counterproductive, risking public outrage and inciting more radical thoughts in protesters and protest sympathizers. We want to underline that the interaction terms between #Rounds and ∆Identity are statistically significant and therefore the specific thresholds for individuals are varying according to their group identification strength. The numbers mentioned above are inferred from the main effects of #Rounds, which are valid for people with ∆Identity = 0 only. For people whose ∆Identity > 0, even though the U-shaped relationship between #Rounds and ∆VEndorsement holds true, the thresholds for them are lower. We will elaborate on this later when we discuss interaction effects.

Figure 3. Marginal effects of #Rounds, #Arrests and protest size in Model 1, leaving out their interactions with ∆Identity to a later section. The horizontal axes represent repression severity or protest size. The solid lines show the marginal effects of political events on radicalization. The light gray areas represent the 90% confidence bands of the marginal effects, and the dark gray areas represent the 95% confidence bands.

Although not statistically significant, Log(#Arrests’) also shows a similar U-shaped curve (Panel 2, Figure 3). Its parabola is symmetric around Log(#Arrests’) = 1.375 (or #Arrests = 23) and intersects the x-axis at Log(#Arrests’) = 2.78 (or #Arrests = 602). Arrests, in general, can also deter violence endorsement as long as #Arrests < 602. However, the magnitude of the deterrence effect diminishes as #Arrests > 23. Moreover, if more than 602 people are arrested, the deterrence effect reverses to become a backfire effect, impelling protesters to endorse a more radical approach in defiance of the legal system.

H1 and H2 are therefore supported in the case of violent repression and partially supported in the case of legalistic repression.

Pertaining to the protests, the third panel of Figure 3 indicates that the marginal effects of protest sizes are monotonically increasing and have always been positive (Panel 3, Figure 3). In other words, offline protests can increase violence endorsement, and the radicalization effect increases as protest size expands.

H3-4: Group Identification

The statistical significance of ∆Identity ranks second only to that of ∆Disgust in Model 1 (b_∆Identity = 0.049, S.E. = 0.002, t = 24.83). If one’s identity strength increases from 0 to 1, i.e., ∆Identity = 1, his or her expressed support for violence would be augmented by 4.9%. H3 is supported.

In addition to direct effects, group identification can also interact with situational variables to exert influence. All interaction terms have positive coefficients. However, only the interaction between ∆Identity and Log(#Rounds’) is statistically significant (b_{∆Identity*Log(#Rounds’)} = 0.039, S.E. = 0.001, t = 2.69). When people grow a stronger group identity, they become more sensitive to political events, especially violent repression events, and more inclined to support protest violence as repression severity or protest size increases.

To obtain a more heuristic grasp of such complicated interactions, we visualized the marginal effects of political events at two extreme levels of ∆Identity in Figure 4. By comparing the lower panel with the upper panel, we can see the “boomerangs” of legalistic and violent repression move upward in the lower panel when ∆Identity = 1. Consequently, the axes of symmetry (dashed lines) and the x–intercepts move leftward. The probability that repression will provoke rather than deter radicalization emphatically rises when individuals become more attached to the protesting group. Therefore, group identification can amplify the backfire effects of state repression, and H4 is supported.

Figure 4. Marginal effects of political events on attitudinal radicalization when ∆Identity = 0 and ∆Identity = 1. The upper panel is identical to Figure 3 and is retained here as a reference for the lower panel. The vertical dashed lines represent the turning points where the marginal effects turn positive, and the deterrence effects are reversed to become radicalization effects.

RQ1-3 & H5: Emotions

All emotions except sadness have a significantly positive influence on the radicalization process. RQ1-3 are therefore answered, and H5 is rejected. Notably, emotions, collectively, have contributed the most to model predictions. ∆Disgust was the most significant predictor, with a positive coefficient of 0.054. This implies that if a person’s intensity of disgust increases from 0 to 1, i.e., ∆Disgust = 1, his or her support for violence will increase by 5.4%. Disgust could thus be considered an anti-establishment and anti-doctrine emotion that makes people particularly likely to defy social norms, reducing their inhibition of aggression. In addition, feelings of anger and fear can also significantly impel people to support violence, and their effect sizes are second only to ∆Disgust and ∆Identity. Given that the Anti-ELAB protests were initiated to block controversial legislation and defend the status quo of Hong Kong as a free city and international financial hub, the fear of change is a salient mobilizing factor, making people more likely to sympathize with the grievances of protesters and support protest violence. Finally, sadness, the least common emotion expressed on LIHKG, is an emotion of inaction that does not substantively influence individual opinions about the use of violence.

Control Variables

Model 3 suggests that both user fame (b_{user_fame} = -.020, s.e. = .008, t = -2.61, p = .009) and the past engagement level (b_{user_engagement} = -.0075, s.e. = .001, t = -8.95, p = .000) are negatively correlated with violence endorsement. Users who are more accepted and more active on the platform are less likely to support violence. One possible interpretation is that popular and outspoken users might be more confident in their ability to influence political decisions and thus less inclined to support non-normative violence.

To shed light on the relationship between attitudinal radicalization and future engagement, we aggregated the post data by week and ran a bivariate linear model using a user’s current engagement level User Engagement_w and radicalization level ∆VEndorsement_w to predict their future engagement level User Engagement_w+1 (w in subscript stands for week). Contrasting the common belief that radicalization can catalyze sustained engagements, the regression Model 3 and 4 in Table 3 indicate attitudinal radicalization is associated with declined engagements when the past engagements are controlled for. We argue that the common belief might arise from a spurious bivariate correlation pattern. As can be seen in Model 2, the coefficient of correlation between current radicalization and future engagement is positive when no control is added. If we enter the auto-correlation term, the coefficient of radicalization falls from positive to negative and its statistical significance diminishes. In other words, people’s future engagements are more forcefully shaped by their past posting habits and the attitudinal radicalization as a configural force would drive them to post less than they would do otherwise. Taken together with earlier findings, attitudinal radicalization does not enhance engagements and enhanced engagements do not lead to further radicalization. The assumption of a reinforcing spiral between engagement and radicalization is not substantiated in this case.

Table 3. Linear regression models predicting future user engagement (N = 88,964).

	Model 1		Model 2		Model 3		Model 4
Variables	b (S.E.)	t	b (S.E.)	t	b (S.E.)	t	b (S.E.)	t
Constant	.456***(.011)	41.8	2.715***(.017)	157.0	.458***(.011)	41.9	0(.000)	−.2
User Engagementw-1	.822***(.002)	443.7	−		.823***(.002)	442.3	.829***(.002)	453.0
∆VEndorsementw-1	−		.782***(.004)	20.0	−.102***(.022)	−4.7	−.099***(.022)	−4.6
Week Fixed Effects	−		−		−		Yes

Conclusions

In a time of political polarization, conflicts between state and society are increasingly inevitable. As for liberal democracies, enhancing the knowledge of political violence and conflict resolution has become an essential prerequisite for sustaining peace (Deutsch & Coleman, 2016). Whereas as for regions undergoing democratic backsliding like Hong Kong, such knowledge is critical for understanding how and which fraction of the population might insist on resisting, even in violent forms, in the face of state repression and coercion.

In this study, we clarify how psychological, relational, and situational factors shape people’s perceptions and reactions during conflictual episodes. In contrast to the rational, organizational, and structural models of social movements, we argue this analytical framework can offer more insights into protest dynamics amid the increasingly personalized political participation. The results reveal that increased support for protest violence occurs when repression severity exceeds the relevant proportionality threshold and is more salient for people who experience stronger disgust, anger, or fear or who are more attached to protesting groups. Moreover, we also find that group identification significantly amplifies the backfire effects of violent repressions. Group identification is like a spark, and violent repression is the gasoline whose interactions drive the Anti-ELAB protests into a more radical direction. As we frequently heard people on the frontlines chanting “if we burn, you burn with us,” a strong sense of “us” can make people more vigilant of repression.

This study has broad theoretical and practical implications. Theoretically, we make a pioneering attempt to synthesize multilevel factors into a holistic model to quantitatively determine the impacts of repressions, identities, and a combination of different emotions on violence endorsement. This integrated model can complement, if not replace, conventional rational choice models by putting psychological motivation into equation. Moreover, it helps elucidate the repression-dissent nexus. We find that the relationship between repression severity and violence endorsement is curvilinear. Repression deters violence endorsement if its severity is below relevant proportionality thresholds and provokes violence endorsement when these thresholds are exceeded. Ignoring this nonlinearity might introduce errors and obscure reality. We also find that the effect sizes of repression-related terms are weaker than those of emotions and group identification. This partly explains why protest dynamics nowadays seem to be less predictable than in the past. Although repression is still salient, people’s situational encounters with repression – with what emotions and group identification strength – are more important in determining their political choices. As participants of networked social movements do not need to conform to authoritative action frames, their responses are conditioned more by their own emotions and identities. These findings underline the importance of emotion management in political communication strategies when a crisis surfaces. Effective coping with political contentions requires appropriate responses to not only political demands but also affective unrests.

Methodologically, we renovate time series analysis by devising a sliding window approach to capturing within-subject variation from digital trace data. Digital trace data, which consist of hundreds of thousands of individual timelines, could be regarded as a kind of panel data that are easier to obtain. However, digital trace data as a panel are highly porous, with excessive “holes” in the timelines, because we cannot require users to create posts at specific times. The proposed sliding window approach allows us to escape missing data and convert porous digital traces into structured formats. This methodological renovation could be helpful for scholars who want to replicate panel techniques in observational research environments.

With that being said, this study is subject to some limitations. First, given that LIHKG has been almost exclusively used by protest supporters and sympathizers, it would be difficult to generalize our results to the entire population of Hong Kong. Second, our research design is capable of elaborating short-term changes evoked by immediate situations, but it is less capable of reflecting long-term dynamics. Nevertheless, since protest events are usually bursty and conflict events usually occur in episodes where actions and reactions all affect one another in the short run (Lichbach & Gurr, 1981), short-term effects are still a vital part of the conflict process. Third, we rely on netizens’ spontaneous expressions to infer their attitudes and feelings. This text-as-data approach might suffer from the disparity between expressions and experiences. People who express support for violence might not genuinely agree with violence it but only say so in order to appeal to others or to show loyalty to an increasingly radicalized group. However, we assert that social desirability bias is presumably limited on an anonymous and decentralized platform like LIHKG. LIHKG users are not required to reveal their real identities. The platform does not support the following/friending function and not provide any badges to showcase users’ popularity or engagement level. It means people won’t be rewarded for behaving socially and won’t be punished for behaving unsocially. Therefore, we argue that when embedded in such a highly anonymous and decentralized network, people will perform more freely and the disparity between expression and experience is arguably minimized.

Open Scholarship

This article has earned the Center for Open Science badges for Open Data and Open Materials through Open Practices Disclosure. The data and materials are openly accessible at https://doi.org/10.7910/DVN/NPRBTR.

Disclosure Statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The data described in this article are openly available in the Open Science Framework at https://doi.org/10.7910/DVN/NPRBTR.

Supplementary Material

Supplemental data for this article can be accessed on the publisher’s website at https://doi.org/10.1080/10584609.2022.2053915

Additional information

Funding

This work was supported by the Hong Kong Institute for Data Science (grant no. 9360163) and the Hong Kong Research Grants Council (grant no. 11605820).

Notes on contributors

Yuner Zhu

Dr. Yuner Zhu is a postdoc researcher in the Department of Public Policy at the City University of Hong Kong. Her research interests primarily lie in applying computational methods to understand public opinion formation, political polarization, and information operations on social media in Greater China.

Edmund W. Cheng

Edmund W. Cheng is an associate professor in the Department of Public Policy and the Director of Political Analysis Lab at the City University of Hong Kong. His research interests include contentious politics, political sociology, political communication, and research methods. He is a co-editor of Social Movement Studies.

Fei Shen

Fei Shen is an associate professor in the Department of Media and Communication at the City University of Hong Kong. His research interests include political communication, public opinion, and computational methods.

Richard M. Walker

Richard M. Walker is the Chan Hon Pun Professor in Behavioral and Policy Sciences, Chair Professor of Public Management in the Department of Public Policy and Dean of the College of Liberal Arts and Social Sciences at City University of Hong Kong, where he is the Director of the Laboratory for Public Management and Policy and Co-Director of the Center for Public Affairs and Law. Prof. Walker is a Fellow of the Academy of Social Sciences.

Notes

1. Data were retrieved from https://acleddata.com/ on November 24, 2020. For details on the event classification and archival methods, please refer to Raleigh, Linke, Hegre, and Karlsen (2010).

2. We used the term “disgust” to refer to a cluster of “other-condemning and other-belittling” emotions, such as disgust, contempt, disdain and detestation. Disgust and contempt are two closely related emotions. They both involve a sense of disapproval, disrespect, and superiority. They are often viewed as conceptually similar (Harris & Fiske, 2006; Matsumoto et al., 2012), linguistically interchangeable, and empirically inseparable (Haidt & Keltner, 1999). Furthermore, contempt is arguably less prominent than disgust and is often recognized as a variant of disgust (Biehl et al., 1997; Haidt & Keltner, 1999). Therefore, when we coded LIHKG posts (see supplementary materials, Note S1), we considered both disgust and contempt but labeled the results only as disgust for the sake of simplicity.

3. The way Tausch et al. (2011) and Fischer and Roseman (2007) measure contempt is compatible with our measurement of disgust. Tausch et al. (2011) identify contempt as the mean values of self-reported disdain and detestation levels, which correspond to our other-belittling and other-condemning dimensions, respectively. Fischer and Roseman (2007) measure contempt based on individuals’ autobiographical accounts of contempt-inducing experiences, which is also within the scope of our disgust construct. Therefore, given the substantive overlaps in conceptualization and operationalization designs, their findings are applicable to our study’s research context.

4. The project was approved by the Human Subjects Ethics Sub-Committee, City University of Hong Kong (Ref. No. H002346). Due to the observational and aggregate nature of our investigation (no private or identifiable information is collected, and user identities cannot be identified from the data), the ethics committee has waived the requirement for informed consent for this study.

5. Model performance will be impeded if the training set is not large enough to fully capture the critical connections between inputs and outcomes (under-fitting problem). Therefore, based on the rule of parsimony, we continued coding on more sentences if the model was underfitted and stopped when the model performance was satisfactory or reached a saturation point where obtaining more data would not help improve the model.

6. The first fraction is 2:00–10:00, which is the late night and early morning period before the working hours (typically quiet hours). The second fraction is 10:00–18:00, which is the work time on weekdays and the prime time for large-scale demonstrations on weekends. The last fraction is 18:00–2:00 (next day), which is the post-work time when most violent acts, clashes, and arrests occur.

7. To fit the repression measurement into the 8-hour sliding window (Figure 1), we consulted a data archive created by researchers from the University of Hong Kong (Teo & Fu, 2021), which recorded when and where tear gas incidents were reported on 92 telegram channels. Compared with the official data release, this observational dataset can provide more fine-grained information on the spaces and times of repression events. We assumed that the use of tear gas was proportionate to the use of other crowd-control agents and estimated the severity of violent repression in relation to the pace of tear gas events.