Introduction

The Trump era has borne considerable political violence. From the deadly Unite the Right Rally in 2017 to the January 6th Capitol Insurrection, high-salience political violence has roiled American politics. As these examples suggest, the bulk of this violence has been perpetrated by the political right in America, but high-profile examples of left-wing violence – such as deadly Antifa protests in Portland, OR and isolated destructive Black Lives Matter protests in the summer of 2020 – have driven headlines as well. Additionally, during the 2024 election, there were two failed assassination attempts against President Trump by non-partisan, solitary actors. Political violence is a gross violation of democratic norms that risks destabilizing the democratic order, especially when encouraged by political leaders (Levitsky & Ziblatt 2019).

Motivated by this surge in violence, scholars have debated the extent to which Americans support political violence (Kalmoe & Mason 2022a, 2022b; Westwood et al. 2022a, b). Though most scholars agree that such support remains a dangerous but numerically small minority, domestic political violence represents a major national security threat and violation of democratic norms.

While most recent work on American attitudes toward political violence interrogates the correlates of support for violence as a legitimate means of grievance expression and/or policy advocacy, this article is concerned with a different question: given the occurrence of political violence, do citizens apply democratic standards of fairness and accountability (Howard 2013)? Are Americans generally willing to punish in-group members (of multiple groups) that engage in political violence for political ends that we might otherwise support? For democracies to sustain, average citizens must not only abstain from engaging in violence to achieve their aims, but also thoroughly reject others’ engagement in it – including from those who they might otherwise align with (Williams 2002). Doing so preserves democratic norms of resolving disputes using mainstream institutional channels. If citizens do not punish the few perpetrators, democratic norms deteriorate.

This proposition places great importance on understanding how Americans generally respond to instances of political violence. Existing work examining reactions to political violence typically focuses on specific events, such as the January 6th attacks (Armaly et al. 2022; Norman 2022). While these events are important, event-specific studies restrict the range of variation in the type of act and perpetrator. This variation could matter immensely for how people respond to political violence. Existing literature recognizes this and calls for a more context-specific approach to understanding responses to political violence (Kalmoe and Mason 2022b; Westwood et al. 2022b).

In this paper, we seek to broaden the conversation on public opinion about political violence by exploring what factors influence the punishment of political violence. In doing so, we ask two questions. First, how do Americans generally punish various instances of political violence? Second, how might we explain variation in punishment? While the specifics of the act could be a prime determinant of reactions to political violence (Clayton et al. 2021; Norman 2022), the social identities of perpetrators involved can affect reactions to their behavior (Iyer, Jetten and Haslam 2012; Tajfel and Turner 1979). In an affectively polarized polity, such as the United States, the public could display partisan and ideological double standards to punishment (Armaly et al. 2022; Mason 2018; Norman 2022; though see Lelkes and Westwood 2017).

To assess the roles of both act and identity in shaping reactions to political violence, we conducted – for the first time in this literature – two conjoint experiments as a novel measure of political violence evaluations. Importantly, this approach enabled us to present and vary several contextual details. In both studies, we presented participants with a scenario in which perpetrators of political violence were arrested for acts committed during a protest. Both experiments gave participants information on the target and severity of the act committed, as well as identity characteristics of the perpetrator. In the first experiment, respondents had to recommend sanctions for defendants. These sanctions provide a direct descriptive picture of how punitive Americans are of certain acts of political violence, and how much their punitiveness varies. In the second experiment, participants had to choose one of two perpetrators to walk free without charge. Requiring respondents to compare perpetrators in this way forces respondents to consider which acts and perpetrators are worse than others.

We find that, most Americans usually support significant sanctions for acts of political violence. When forced to let one person walk free, several participants volunteered that they did not wish to let any perpetrator go. The main factor Americans prioritize is the act itself. As the severity of an act increases, Americans are willing to apply more stringent sanctions and less likely to excuse these acts. The target matters as well. Americans treat acts against police as more reprehensible than acts against property or counter-protesters, holding severity constant. Surprisingly, there is clear bipartisan agreement in how to treat different acts of political violence. Some personal characteristics of the perpetrator also matter. While Americans are broadly willing to punish co-partisans and out-partisans alike, they are more lenient on co-partisans. Our results suggest that Americans’ reaction to political violence largely reflects the democratic ideal. However, some important double standards (e.g. partisanship) exist.

Conceptualizing Reactions to Political Violence

Support for political violence is fairly uncommon in democracies because it contradicts democratic norms (Kalmoe and Mason 2022a; Westwood et al. 2022a). While existing scholarship disputes whether support for political violence is increasing (Bartels 2020; Kalmoe and Mason 2022a; Westwood et al. 2022a), most scholars agree that supporters remain a small but dangerous minority.

In the American context, most studies examine the correlates of political violence support. American support for political violence is higher among those who endorse conspiracy theories (Baum et al. 2023), feel marginalized by society (Armaly et al. 2022; Munis et al. 2024), perceive low self-efficacy (Kalmoe and Mason 2022a) and display high trait aggression (Kalmoe 2013) or Dark Triad traits (Uscinski et al. 2021). Whether support for political violence is consistent across parties (Baum et al. 2023; Westwood et al. 2022a) or if conservatives are particularly prone to it (Piazza and Van Doren 2023) remains unresolved. Some suggest that party leaders’ adherence to democratic principles and whether one’s party is currently in power affects political violence support (Kalmoe and Mason 2022a).

By contrast, the context of the act on punishment decisions receives less attention (though see Westwood et al. 2022a, b). Democratic theory discusses how shared norms become binding law, and punishment in democratic society for violations of those laws and norms can fulfill multiple functions, including retribution, deterrence, rehabilitation, and restoration (Williams 2002). Principled, unbiased application of the law is essential in democracy, but especially in sanctioning acts, such as politically motivated violence, that are counter to democratic norms (Howard 2013). Failure to condemn political violence and dole out consequences in an even-handed way consistent with democratic standards of fairness undermines democracy along two fronts: first, by weakening deterrence of such violence, and second, by eroding public confidence in democratic institutions, including criminal justice processes.

To examine these norms, we investigate how the nature and context of acts affect reactions to political violence because some people who strongly oppose political violence in the abstract may display a more muted reaction to a specific instance of it. Alternatively, some who endorse such tactics in the abstract may strongly condemn their use when responding to a concrete example. Both produce divergence between abstract attitudes and concrete evaluations. Drawing an analogy to political tolerance, most Americans, if asked, want to guarantee First Amendment rights to all other Americans, regardless of how much they may disagree with their speech’s content (Stouffer 1955; Gibson 2013). When presented with specific groups in specific scenarios, what looks like strong support for a democratic norm begins to falter (Gibson and Bingham 1982).

Factors Affecting Reactions to Political Violence

The Act

One of the fundamental ways that instances of political violence vary regards the specifics of the act committed. In line with Norman (2022), we posit that there are two relevant subdimensions regarding acts of political violence: the type of act and its severity. Concerning the type of act, there are two important factors to consider: (1) whom is the violence directed toward, and (2) whether the act involves violence against persons and/or property. Politically motivated assaults of civilian subjects could be perceived as a greater violation of norms than acts against property, which may seem a less clear violation of legitimate conduct, because violence against humans elicits higher emotional distress and empathetic concern (Rosenthal-Von Der Pütten et al. 2014). Therefore, we expect greater condemnation of violence against people over acts against property (H1a). In the abstract, violence against the state, while still roundly condemned, receives greater justification than violence against non-state actors (e.g., Kalmoe and Mason 2022a; Munis et al. 2024). Thus, we expect Americans to especially condemn violence against non-state targets like small businesses over state targets like government property (H1b).

Furthermore, we argue that violence severity matters. One may perceive a verbal threat as disrespectful and a punch as alarming while considering inflicting serious injury with a weapon egregious. Existing work varying act severity offer supportive evidence of this argument (Kalmoe & Mason 2022a; Norman 2022; Westwood et al. 2022a). Therefore, we expect that people will prioritize the prosecution of more severe acts over lesser acts of violence (H1c).

Whether Democrats and Republicans react similarly to different targets of political violence remains an open question. For instance, Republicans are warmer towards police than Democrats (Reny and Newman 2021) and tend to be seen as the party of small businesses (Benoit 1999; Goggin et al. 2020).Footnote 1 Thus, we might expect Republicans to apply harsher sanctions to acts against police (H2a) and small businesses (H2b). In contrast, Democrats are traditionally more inclined to see government as a tool for positive action than Republicans (Baldassari and Park 2020). Similarly, they may condemn the targeting of government buildings more severely (H2c). However, if norms against political violence are sufficiently strong, partisan differences in sympathy toward targets might not affect reactions to violence against those targets. Alternatively, since conservatives tend to be more punitive than liberals (Redlawsk and Walter 2024; Yates and Fording 2005), Republicans may apply harsher sanctions overall.

The Perpetrator

Normatively speaking, democratic norms against political violence should be applied indiscriminately with no special affordances made based on who is committing violence (Clayton et al. 2021). However, expansive literature on group biases spanning multiple fields within the social sciences suggests that this is unlikely to be the case. Humans demonstrate a strong bias for favoring ingroup members (e.g., Balliet et al. 2014) and derogating outgroup members (e.g., Levin & Sidanius 1999), as discussed further in the next two paragraphs.

According to Social Identity Theory, people are apt to favor in-group members in their attitudes and behavior (Tajfel and Turner 1979). To enhance self-esteem, people make intergroup comparisons that involve selective attention or biased assimilation of information favoring groups they belong to (Kunda 1990). Therefore, moral and political transgressions committed by ingroup members may appear to be less severe, more understandable, or even justified (Brewer 1999). In contrast, outgroup-perpetrated violence reifies notions of ingroup superiority by showing that the outgroup “stoops lower” to achieve its goals. Thus, all else equal, people will likely punish acts of political violence with outgroup perpetrators more harshly.

Some intergroup distinctions, however, may matter more than others. In American society, one of the strongest sources of intergroup conflict is partisanship. Partisan animosity has increased in recent decades (Mason 2018). It can be acquired early in life (Phillips 2022) and has become more powerful through the overlapping association of ideological, racial, religious, educational, and urban-rural social identities with partisanship (Mason 2018). This overlap has reduced commonality between Democrats and Republicans, making out-partisans less relatable and easier to dislike. The rise of divergent partisan media ecosystems, which effectively act as echo chambers, reinforce perceptions of difference, distrust, and threat – further fueling polarization (Hobolt et al. 2024; Törnberg 2022). Indeed, the threat of the opposing political coalition is thought to be key to recent increased support for political violence (Kalmoe and Mason 2022a). Recent findings suggest that ideological ingroup members receive less condemnation for acts of political violence (Armaly et al. 2022; Kalmoe and Mason 2022a; Norman 2022; though see Lelkes and Westwood 2017). Therefore, all else equal, we expect people to punish out-partisans more than in-partisans (H3). However, our research design factors in the role of additional perpetrator characteristics beyond partisanship to capture other relevant ingroup and outgroup dynamics.

The Role of Perceptions of Motive

One important dimension in judgments of political violence is the perception of motive. Although Americans may not normally condone a given act, agreeing with the perpetrator’s cause might soften judgments. Inferring good intentions results in softer moral (Carlson et al. 2022) and criminal (Hessick 2006) penalties. Aspects of the act or the perpetrator could cue Americans that someone who committed political violence did so for a worthy cause. This might lead to less punitive judgements. However, if Americans strongly condemn the use of political violence, their sympathy to certain kinds of perpetrators or acts will not translate into judgments about the sanctions that the perpetrator deserves.

Method

To evaluate the extent to which Americans’ reactions toward political violence is conditional on contextual circumstances, we must simultaneously vary several dimensions of acts of violence and the perpetrator. Since conjoint experiments allow researchers to vary several attributes at once, we can vary the act committed and details of the perpetrators without sacrificing statistical power.

Conjoint experiments are well-suited to examine multidimensional evaluative contexts and increasingly common in political science (Hainmueller, Hopkins, and Yamamoto 2014). They have been used to examine candidate evaluations (e.g., Munis 2021) and immigration attitudes (e.g., Hainmueller et al. 2014), among other topics. Conjoint experiments present participants with several attributes of a profile they are asked to evaluate (Hainmueller et al. 2014). While they sacrifice some external validity, conjoints have enhanced realism compared to vignette experiments since, just as in the real world, many factors vary at once. Relatedly, conjoints are more predictive of real-world outcomes than vignette studies (Hainmueller et al. 2015). Additionally, because conjoints allow researchers to vary a multitude of dimensions, they can minimize “aliasing” where a number of attributes are correlated and it is difficult to discern their independent effects (Hainmueller, Hopkins, and Yamamoto 2014; Theodoridis et al. 2023).

Furthermore, conjoint experiments guard against social desirability bias and expressive responding (Horiuchi, Markovich, and Yamamoto, 2021; Theodoridis et al. 2023), which could greatly affect reported judgments of political violence (Kalmoe and Mason 2022a; Theodoridis et al. 2023). Because conjoints vary several attributes, it is harder to consciously avoid norm-violating bias towards certain social groups. Therefore, the effects derived from the conjoint are less likely to be distorted.

Our use of the conjoint method also allows us to bridge a central debate concerning American public opinion toward political violence. In their discussions over proper measurement of support for political violence, both Kalmoe and Mason (2022b) and Westwood et al. (2022b) call for examinations of political violence that take context into account. By leveraging rich contextual variation, our conjoint experiments provide clear evidence regarding which acts and actors people are more likely to excuse or condemn.

Samples

We conducted two separate experiments, both using samples recruited from Dynata (formerly known as Survey Sampling International). We included quotas for age, gender, state, race, and party to make the samples as descriptively representative of the U.S. population as possible (see Table 1 for sample demographics). Quota samples are commonplace in political science, with research showing that their results typically generalize to random/probability samples – especially for experimental research such as ours (Coppock et al. 2018). However, in both instances we needed to relax the use of quotas somewhat to recruit our target sample size (due to difficulty fully filling our quotas for some demographics). Therefore some demographic groups were slightly under/over-represented in our sample. Overall, divergences from national targets were rather negligible and both of our samples were highly similar in composition to the broader U.S. adult population (see Table 1). These experiments were approved by the Utah Valley University and Pennsylvania State University institutional review boards prior to fielding. The online pre-registrations can be found at https://osf.io/qw9ev (single-profile experiment) and https://osf.io/ez5tc (double-profile experiment).

Table 1 Survey experiment sample characteristics
Full size table

Procedure

Both experiments utilized a nearly identical hypothetical scenario. Participants were asked to imagine that there was a protest in their state’s capital city that descended into violence between protesters and counter-protesters, leading to numerous arrests. We purposefully left the basis of the protest vague in order to reduce the likelihood that results are driven by idiosyncratic responses to a particular kind of protest. In the single-profile experiment, participants were presented with a single perpetrator at a time (five perpetrators total). For each perpetrator, they had to apply a sanction to the perpetrator in the event the latter is found guilty and indicate, on a 5-point scale, how likely it was the perpetrator was performing the act for a good cause.Footnote 2 In the double-profile experiment, participants were presented with two perpetrators (seven pairs total) and had to decide which perpetrator they would let walk free without charge.

We gave respondents information on the type of act committed along with several perpetrator characteristics (see Table 2). Within act type, we varied act extremity, using parallel language for property destruction and violence against people. We chose different wording for people and property to reflect differences in the type of damage. Our selection of acts captures several realistic and common acts of political violence at protests, rather than the full range. We omitted the most horrific forms of political violence (e.g. mass murder, torture) and other targets (e.g. politicians, bystanders, congregations) because, while important to study, they are less common.

As shown in Table 2, our definition of political violence in this study is rather expansive and includes property damage. We acknowledge that it is debatable whether property damage constitutes violence. However, we chose to include property damage (against various targets and in varying degrees of severity) as act types in our study for two reasons. First, property damage within the context of political protest has long been treated as political violence by scholars (Mars 1975; Blum 2021). Second, property damage is considered violence in the media coverage of protests and under federal criminal code as well as, controversially, can constitute terrorism even where there is no danger of injury toward human bodies.Footnote 3 It is possible that some respondents consider property damage, especially minor property damage, neither violent nor political, which may add some noise to our estimates. However, we mitigate against this by clearly identifying hypothetical perpetrators as being present at a political protest and that the acts they committed occurred after the protest intensified, so political motivation for the act is strongly implied.

In addition to act characteristics, we informed respondents about several perpetrator characteristics: age, occupation, political party, gender, race/ethnicity,Footnote 4 marital status, and number of children. We also informed them whether the perpetrators were from the respondent’s state capital city, a rural area in the state, or out of state. Participants then had to decide the appropriate sentence if any (single-profile experiment) or which person should not be prosecuted between the two profiles (double-profile experiment). As with acts, these do not capture the full span of identity dimensions. For example, extremist group affiliated perpetrators are not included. This is because the extremist group(s) involved with any protest will be context-specific and naming them would inadvertently prime respondents to think the protest is about that cause. Sacrificing some range in identity guards the internal validity of the experiment. Table 2 displays the attributes and each of their levels.Footnote 5

Table 2 Levels of the conjoint attributes
Full size table

Each design has complementary strengths and weaknesses. The single-profile design allows us to answer two questions. How likely is a specific action or perpetrator characteristic to warrant incarceration? How many years in jail does a given action or perpetrator characteristic lead to? These metrics are intuitive and can be applied across perpetrator and act characteristics. Despite its contrived nature, this design mimics instances where people receive information about a specific act of political violence committed by a specific perpetrator and they evaluate that perpetrator on the spot. However, the design is vulnerable to some level of social desirability bias. While respondents can apply milder sanctions to sympathetic perpetrators, they may wish to communicate blanket opposition to political violence despite their actual opinions (whether sincerely or to signal social desirability). Thus, they may inflate the sanctions they give.

By contrast, forcing respondents to choose one defendant to excuse in a two-profile forced choice conjoint leaves little room for social desirability. Rather, respondents must weigh which factors make one perpetrator more worthy of prosecution than another. Therefore, this experiment is well-equipped to discern which factors matter most for judgments pertaining to political violence. Unfortunately, this design cannot recover raw levels of support for political violence. Thus, the single-profile experiment is best-positioned to gauge differences in the sanctions participants give single perpetrators, while the two-profile experiment is best-positioned to assess potential double standards.

Analytic Strategy

In both experiments, respondents evaluate multiple randomly generated profiles of alleged political violence perpetrators. As a result, there is a natural multilevel structure to the data of profiles nested within participants (single-profile: 17,010 profiles nested in 3,402 participants; double-profile: 38,682 profiles nested in 3,065 participants). We accommodate this structure using a multilevel modeling approach in the lme4 package in R, specifying random intercepts at the participant level. This approach is somewhat different to the use of pooled models for estimation with clustered standard errors at the individual-level (Leeper et al. 2020). While both methods adjust for non-independence between observations (Bliese and Hange 2004), multilevel modeling is more conservative, leading to fewer rejections of the null hypothesis when there are no true effects of a given attribute (Cheah 2009). Results with pooled models are substantively identical (see Figures A4-A5 and Table A7 in the Appendix).

We examine several dependent variables. In the single-profile experiment, the first is a binary indicator of whether the perpetrator was sentenced to jail, taking on a value of ‘1’ if the perpetrator was given a non-zero amount of incarceration, and ‘0’ otherwise. The second is an indicator of the number of recommended years in jail. Because the response categories are intervals of incarceration (e.g. `20–30 years in jail’), we coded this variable as the lowest number of years in the interval. The third is a 5-point scale of the perceived goodness of the perpetrator’s motivations (see footnote 1). In our double-profile experiment, leniency (taking on a value of ‘1’ if the profile is let go, ‘0’ if the profile is not let go) is the outcome of interest. Full results can be found in Tables A1-A6 of the Appendix.

Results

Americans Reliably Punish Political Violence

In the single-profile experiment, participants could choose how severely to sanction perpetrators. While not a direct measure of support for the act, harsher sanctions imply more negative reactions. Across choices, 75.7% of sanctions involved at least one day in jail. 29.9% of these sanctions involved at least one year in jail. The mean sentence was 2.1 years, with the median sentence lasting around 4 months. These figures indicate that most of the time participants did not want perpetrators to get away with political violence. Yet, participants still considered most acts unworthy of a lengthy prison term. As we detail in subsequent sections, these judgments vary considerably according to the action and attributes of the perpetrator.

In the double-profile experiment, participants had to let one of two perpetrators go with no punishment. Therefore, we could not directly gauge willingness to sanction any given perpetrator. However, after respondents indicated which perpetrator they would let off the hook if they had to, we asked them to indicate, in an optional open-ended response box, why they made the choice that they did. 55.6% of respondents gave responses, which were read and interpreted by this article’s authors. We determined that, of those who responded, 7.1% mentioned they did not want to free anyone generally. An additional 1.3% mentioned this for some choice sets but not others. By contrast, few (0.3%) wanted to free both profiles.

The Act is Incredibly Important for Judgments on Political Violence

Numerous scholars suggest that Americans’ reactions to political violence are contingent on the act committed. In line with these suggestions, 68.0% of participants in the double-profile experiment who commented on their decision-making placed great emphasis on the act. Many insisted only the act mattered or only mentioned the act (48.7%), while others mentioned it as a primary cause with (12.1%) or without (7.2%) mentioning other secondary concerns. However, these utterances could be attempts at responding in a socially desirable manner, or respondents failing to comprehend the reasons behind their decision-making.

Therefore, we turn to the results of the conjoint. Figure 1 depicts the likelihood of participants sending a given perpetrator to jail in the single-profile experiment (left panel), the length of a given sentence (middle panel) in single-profile experiment, and leniency (right panel) in the double-profile experiment. Overall, both the target of the act and severity mattered immensely to participants. In the single-profile experiment, 55.4% of the least severe acts across type earned incarceration. Across act type, participants were 25.1% points more likely to incarcerate to moderately severe acts than mild acts, and 36.8% points more likely for the severe acts than mild acts. Holding severity constant, participants were most likely to incarcerate perpetrators for acts against police. In the single-profile experiment, acts against government buildings received the next-most severe sanctions, followed by acts against small businesses, then acts against counter-protesters. In the double-profile experiment, acts against police receive less leniency than any other acts, which did not meaningfully differ from one another.

Fig. 1
Fig. 1The alternative text for this image may have been generated using AI.
Full size image

Likelihood of incarceration (left panel), average jail sentence (middle panel) in single-profile experiment and leniency (right panel) in double-profile experiment by perpetrator action. Derived from Models 1 and 3 of Table A3 and Model 1 of Table A4 in the Appendix

Findings are nearly identical for the length of sanctions in the single-profile experiment. Holding target and perpetrator characteristics constant, the least severe acts carried slightly less than 1 year in jail, moderately severe acts 1.4 years (roughly 200 more days), and the most severe acts carried 4.1 years. Over and above severity, acts against police yielded almost 1 year more in jail and acts against government buildings yielded 6 months more in jail than acts against counter-protesters or small businesses.

Democrats and Republicans Agree on How to Punish Political Violence

It is unclear whether Democrats and Republicans in the electorate treat identical acts of violence differently. Republicans and Democrats may punish different targets due to their ideological coding and/or Republicans may give harsher sanctions than Democrats due to their higher punitiveness. Alternatively, norms against political violence could be strong enough to neutralize partisan differences. To discern between these possibilities, we assessed the relationship between partisanship and the likelihood and severity of sanction for different acts of political violence. We did this by including a series of respondent party x act interactions in each model.

Figure 2 depicts the results of these tests. Across acts, in the single-profile experiment, Republicans did not differ significantly from Democrats in likelihood of assigning incarceration or sentence severity (ps ≥ 0.083). A visual examination of response patterns from Democrats and Republicans in each experiment indicates they assign a comparable rank order of condemnation to different acts.

Fig. 2
Fig. 2The alternative text for this image may have been generated using AI.
Full size image

Likelihood of incarceration (top panel), average jail sentence (middle panel) in single-profile experiment and leniency (bottom panel) in double-profile experiment by perpetrator action and respondent political party. Derived from Models 2 and 4 of Table A3 and Model 2 of Table A4 in the Appendix

The act x respondent party interactions provide a more rigorous test of partisan differences. Only 5/24 party x act interactions in the single-profile experiment and 2/12 in the double-profile experiment were statistically significant. While occurring at a greater rate than chance, these differences were the exception rather than the rule. In the single-profile experiment, Republicans were more likely to incarcerate for graffitiing a government building or small business, and slightly more likely to incarcerate for burning a small business down. Republicans also gave longer sentences to people who punched police officers (+ 8.5 months) and burned down nearby small businesses (+ 8.3 months). In the double-profile experiment, Republicans were eight percentages points less likely to release someone who severely injured a police officer and 5% points less likely to allow someone who damaged a small business. These partisan differences are consistent with Republicans’ higher warmth towards businesses and police (Reny and Newman 2021, though see Hersh 2023) and greater baseline levels of punitiveness toward minor crimes, like graffitiing (Redlawsk & Walter 2024). However, most evidence indicates that different groups of partisans treat different acts similarly.

Americans Display In-Party Bias, but Perpetrator Identity Matters Little

By far, the actual act has the strongest effect on respondents’ sanctioning decisions. Meanwhile, most identity dimensions did not have a clear effect. However, one non-trivial pattern emerged in both the single and double-profile experiments: double standards based on the perpetrator’s partisan congruence. Figure 3 depicts the scope of these effects. When they had to choose who to let free, they were 6% points more likely to let go of an in-partisan than an out-partisan. When they could freely choose a sanction, out-partisans were ~ 2.4% points more likely to be incarcerated and received ~ 4 extra months in jail than in-partisans. While these effects pale in comparison to the effect of act, they indicate an unambiguous disparity.

Fig. 3
Fig. 3The alternative text for this image may have been generated using AI.
Full size image

Likelihood of incarceration (top panel) and average jail sentence (middle panel), single-profile experiment, and leniency (bottom panel) in double-profile experiment by perpetrator partisan congruence. Derived from Models 1 and 2 of Table A1 and Model 1 of Table A2 in the Appendix

One question that emerges from these findings is the extent to which affective polarization drives our findings. While we did not manipulate levels of affective polarization (Voelkel et al. 2023) experimentally, and therefore cannot make strong causal claims, we measured it in the double-profile experiment. Prior to the experimental task, participants indicated both their partisanship and 0-100pt scales of warmth toward their own party and the opposing party, allowing us to construct a measure of affective polarization capturing net in-party warmth. We then estimated its role as a moderator of the effects of perpetrator partisanship.

This interaction is depicted in Fig. 4. At minimum levels of affective polarization, participants do not make distinctions in who they excuse based on partisanship. However, starting at even moderate levels of affective polarization (a ~ 25 point difference in warmth), out-partisans (in-partisans) receive less (more) leniency than baseline. While not evidence of a causal role, the descriptive pattern is fairly clear. People make meaningful distinctions in how they punish according to partisanship, and affectively polarized partisans are particularly likely to do so. Given rising rates of affective polarization across time, driven by factors such as social sorting and the proliferation of echo chambers, this is a concerning finding (Mason 2018; Hobolt et al. 2024).

Fig. 4
Fig. 4The alternative text for this image may have been generated using AI.
Full size image

Affective polarization and leniency by party of perpetrator, double-profile experiment. Derived from Model 3 of Table A6 in the Appendix

By contrast, other identity dimensions had more muted and inconsistent effects. The only other clear tendency in both experiments related to perpetrator race. Black perpetrators were less likely than white perpetrators to face sanction and faced shorter sanction lengths (see Figure A1 in the Appendix). This was primarily driven by Black respondents, whereas white respondents tended not to distinguish based on perpetrator race – findings that are consistent with other conjoint studies measuring the effects of race on punitiveness (e.g., Doherty et al. 2022). We lacked the statistical power to precisely estimate the preferences of other racial and ethnic groups. Given that we used text-based indicators of race, and this method of communicating information on race in conjoints tends to have stronger, not weaker effects (Abrajano et al. 2018), we are fairly confident in these muted effects.

Identity vs. Act – A Harder Test

In general, perpetrator actions had a strong and consistent effect on Americans’ reactions to political violence, while perpetrator identity, with some exceptions, had less consistent effects. Informally, the size of effects illustrates this point. Increasing act severity extends the likelihood of sanction by over 10% points in each experiment, and the length of sanction by years in the single-profile experiment. By contrast, effects of perpetrator identity tended to affect jail likelihood by only a few percentage points, and length of sentences by months at most.

More formally, a series of pairwise comparisons derived from our models underscores this point. In the single-profile experiment, 55/66 act comparisons (83.3%) for likelihood of jail and 50/66 (75.8%) for length of sentence were statistically significant. Likewise, in the double-profile experiment, 56/66 (84.4%) of act comparisons were statistically significant. By contrast, only 5/37 identity comparisons for likelihood of jail (13.5%) and 4/37 (10.8%) for length of sentence in the single-profile experiment, and 10/42 comparisons (23.8%) in the double-profile experiment were statistically significant.

Another way to test this is to examine whether, holding a given action constant, a given identity dimension has a significant effect on Americans’ judgments of sanction (and vice versa). We find similar results. Holding each possible identity on each dimension constant yielded 1,782 pairwise comparisons each for likelihood of jail and length of sentence in the single-profile experiment and 1,848 in the double-profile experiment. Of these, 68.9% of pairwise comparisons for likelihood of jail and 55.9% for length of sentence were statistically significant in the single-profile experiment, as were 73.4% of comparisons for leniency in the double-profile experiment. By contrast, holding each action constant, only 10/444 identity comparisons for likelihood of jail (2.3%) and 13/444 (2.9%) for length of jail term were statistically significant in the single-profile experiment, and 21/504 (4.2%) were for leniency in the double-profile experiment.

However, it is possible that in these comparisons, we are stacking the deck against finding significant effects for identity. Because there are 12 potential actions perpetrators could take, there are fewer observations for detecting effects of identity than there are for detecting effects of action. To resolve this disparity, we performed post-hoc pairwise comparisons for identity holding constant separately the target and the severity of the act. While more effects are statistically significant, the changes are not drastic. Holding target constant, 4/148 (2.7%) of comparisons for likelihood of jail and 6/148 (4.1%) for length of sentence in the single-profile experiment and 15/168 (8.9%) of comparisons in the double-profile experiment were statistically significant. Holding severity constant, 3/111 (2.7%) of comparisons for likelihood of jail and 7/111 (6.3%) for length of sentence in the single-profile experiment and 17/126 (13.5%) in the double-profile experiment were statistically significant. In some cases, the rate of finding significant effects is higher than chance. Given the consistency of some findings, it is unlikely that identity is irrelevant to Americans’ judgments of political violence. Overall, while the act being the most consistent and powerful factor influencing evaluations of political violence is our most prominent finding, the fact that some identity characteristics consistently emerged as influential – especially partisanship – should not be discounted.

Identity Matters Somewhat More for Perceived Motive, Motive Does Not Explain Sanction

So far, our discussion of Americans’ reactions to political violence has been about actual sanctions. However, in the single-profile experiment, we also asked respondents to indicate on a 5-point scale how likely it was that the perpetrator committed the act for a good cause (question wording in footnote 1). This measure, albeit crude, is a gauge of sympathy for the perpetrator. Two distinct possibilities exist. First, sympathy for the perpetrator may explain the effects of the act and identity of the perpetrator on sanctions. Second, there may be a disjunction between how Americans decide to allocate sympathy as opposed to sanctions.

Figure A2 in the Appendix depicts the effects of each action and identity dimension on the perceived goodness of the perpetrator’s motives. These effects are juxtaposed with that of likelihood of jail and years in jail. In general, findings are similar. Act matters strongly and some identity characteristics matter too. However, two differences emerge. First, while Americans display some in-party favoritism in sanctioning, it is magnified with perceived motive. Second, Americans, as a whole, were more likely to think perpetrators of color had good motives compared to white perpetrators.

These differences prove prescient for understanding how perceived motive relates to sanction. In general, there is only a weakly negative repeated-measures correlation between perceived goodness of perpetrator motives and ultimate sanction (r=-.155, p <.001). Furthermore, including motivation in models determining sanction does not generally change the sign or significance of effects (see Figure A3 in the Appendix). This indicates that while Americans might display differences ins sympathy based on some identity dimensions, these sympathies largely do not translate into sanctions.

Discussion and Conclusion

Overall, our findings paint a clear picture of how Americans judge acts of political violence. Descriptively speaking, Americans are averse to excusing political violence. Given the option, most respondents incarcerate political violence perpetrators. Lacking the option to punish both perpetrators, they frequently express a desire to do so.

Far outweighing any other piece of information, the target and severity of the act itself factored most prominently in Americans’ evaluations. Acts that threaten lives or irrevocably destroy property receive near-unanimous sanction. By contrast, acts that can easily be repaired (such as spraying graffiti on a building) or acts that result in no physical harm (threats) receive weaker sanction. Across experiments, Americans treat acts against police with the most severity. There is also broad bipartisan agreement on which acts are more worthy of sanction.

Meanwhile, the partisanship of the perpetrator plays a clear role in shaping the sanctions people seek to give. Partisans, particularly those higher in affective polarization, are more likely to condemn and recommend harsher sanctions for their partisan opponents compared with non-partisan or co-partisan perpetrators. In practical terms, carrying an out-partisan identity comes with an extra 4 months of recommended incarceration, which is evidence of appreciable partisan bias.

By comparison, perpetrator race played a more muted role. While there was evidence that Black respondents were less likely to sanction racial ingroup members, other findings did not consistently emerge. One can argue, given reactions to real-world events such as Black Lives Matter, that this finding lacks external validity and is driven by social desirability bias. However, the fact that we delivered information about perpetrator race via text normally amplifies rather than mutes the effect of profile race in conjoints (Abrajano et al. 2018). Furthermore, conjoint experiments tend to guard against social desirability bias (Horiuchi et al. 2021). Therefore, while it is possible the artificial nature of each scenario affects the role of perpetrator race, other aspects of the experiments do not display that issue.

Thus, results of our study present somewhat of a normative “mixed bag” regarding Americans’ evaluations of political violence. The legitimacy of punishment in a democracy depends on the existence of a moral community where citizens view the law as embodying shared norms (Howard 2013). Moreover, these laws, and the norms undergirding, must be applied fairly. We find that, overall, shared norms against political violence are strong. More severe departures from democratic norms receive harsher sanctions, indicating agreement that certain acts are beyond the pale. Yet, there are some issues concerning the fairness standard – there is a non-trivial amount of differential treatment based on the personal characteristics of the perpetrators, particularly partisanship. This is especially the case when we examine citizens’ sympathy for a given perpetrator. However, double standards factor much more into judgments of sympathy than of actual leniency. This suggests that despite their sympathies, Americans believe that acts of political violence need to be sanctioned.

There are some key limitations to the paper as a result of research design choices made. The most central limitation stems from our decision not to include the reason for the protest in order to reduce the odds that responses were protest-specific. In doing so, we expose ourselves to the risk that responses are less comparable because respondents likely have different kinds of protests in mind while completing the conjoint tasks. A potential consequence is that we may systematically underestimate identity-based perpetrator effects. It also reduces the realism of the task. Further work would do well to supply different scenarios to examine how situation-contingent judgments of political violence are. A related limitation is that our experiment neither varied delivery medium nor source cues, which is notable as these features of political communication likely shape how people process and evaluate information regarding political violence. We urge future researchers to explore these communication dynamics. Furthermore, we did not use the full spectrum of acts and identities. This was a deliberate choice to maintain some experimental realism and avoid indicating what the protest was about. However, it means that we do not know how much extremist identities affect Americans’ judgments or how extreme actions might shape the effects of perpetrator identity. Future work should expand the range of acts and identities. A final limitation worth noting is that it is possible that a non-trivial number of respondents did not consider some of the acts we presented as constituting political violence, which, if true, could add some noise to our estimates. Future work should continue to gauge citizen understandings of political violence and the range of acts that constitute it (including both what they consider violent and, separately, political).

Nevertheless, our paper advances our knowledge of how citizens evaluate political violence in the United States. Our conjoint design allowed us to examine the potentially multidimensional nature of evaluations that more closely approximates the world than a single survey question (or two or three) that asks about political violence in the abstract. Our work can also inform future work examining the abstract overall level of support for violence through illuminating what contextual features should be specified/held constant in survey questions.

That said, there is still much to unpack regarding how citizens think of acts of political violence. In particular, our paper usefully illuminates which considerations Americans find most significant when they are asked to judge political violence but does not shed light on which factors influence perceptions of legitimacy or illegitimacy surrounding political violence, nor the extent to which a given act’s legitimacy shapes willingness to sanction the act. These are important questions which we hope future research will immediately turn to.