Contrary Findings on the Effects of Right-Wing Terror

A large body of research shows that associations of political actors with violence can trigger backlash – reducing identification with and support for parties and candidates deemed responsible for the violence (Eady et al. Reference Eady, Hjorth and Dinesen2023; Heger Reference Heger2015; Kadt et al. Reference Kadt, Johnson-Kanu and Sands2024; Lebas and Young Reference Lebas and Young2024; Rosenzweig Reference Rosenzweig2021). Drawing on this research, this study aims to clarify an important controversy in the literature on the effects of right-wing terror on voting choices for PRRPs.Footnote ³

Existing results regarding right-wing terrorism and political attitudes mostly point in the same direction: right-wing attacks tend to improve perceptions towards minority groups, reduce immigration skepticism, and make citizens shift away from nationalist attitudes (Jakobsson and Blom Reference Jakobsson and Blom2014; Solheim Reference Solheim2020; Pickard et al. Reference Pickard, Efthyvoulou and Bove2023; Wollebæk et al. Reference Wollebæk, Enjolras, Steen-Johnsen and Ødegård2012). It is only when it comes to studying how these attitudes translate into vote choice that two sets of opposing findings emerge.

The first group of studies suggests that there is a negative effect of right-wing terror on support for PRRPs. Based on survey experiments in Germany, the Netherlands, and Sweden, Jacobs and van Spanje (Reference Jacobs and van Spanje2021) show that exposure to news stories about a prevented right-wing terrorist attack can reduce self-reported preferences for PRRPs. Geys and Hernæs (Reference Geys and Hernæs2021) show that the 2011 terrorist attack in Norway led to a drop in support for party leaders of the right-wing Progress Party (Fremskrittspartiet) as well as a shift in voting intentions and actual vote shares away from the PRRP. Pickard et al. (Reference Pickard, Efthyvoulou and Bove2023) analyze the effects of the Jo Cox MP murder (2016) and the Finsbury Park attack (2017) in Great Britain. After the attacks, citizens express a lower intention to vote for the right-wing populist UK Independence Party.

Another group of studies points in another direction. Comparing municipalities with successful and failed terror attacks, Sabet et al. (Reference Sabet, Liebald and Friebel2022) show how the AfD benefits disproportionally from an increasing turnout in affected municipalities. Eger and Olzak (Reference Eger and Olzak2023) investigate the effects of violent anti-refugee incidents (2014–2019). They find differential effects of right-wing violence: it increases support for PRRPs only among citizens with anti-immigration attitudes. This finding is in line with Krause and Matsunaga (Reference Krause and Matsunaga2023), who find that right-wing violence increases support for the AfD – as former supporters of center-right parties turn to the AfD.

What explains these contrary findings? The discrepancy does not seem to be related to systematic differences in outcome measures. Most studies across both groups focus on party preferences (Jacobs and van Spanje Reference Jacobs and van Spanje2021 and Pickard et al. Reference Pickard, Efthyvoulou and Bove2023 in the first group; Eger and Olzak Reference Eger and Olzak2023 and Krause and Matsunaga Reference Krause and Matsunaga2023 in the second group). There are also studies in both groups that consider actual election vote shares of right-wing parties (Geys and Hernæs Reference Geys and Hernæs2021 in the first group and Sabet et al. Reference Sabet, Liebald and Friebel2022 in the second group).

Neither does the discrepancy seem to result solely from differences in the type of attacks under investigation. Research in the first group of studies has focused primarily on high-profile attacks killing and/or injuring multiple victims (Geys and Hernæs Reference Geys and Hernæs2021; Pickard et al. Reference Pickard, Efthyvoulou and Bove2023). However, similar negative associations for right-wing terrorism and PRRP support have also been found for pure threats of attacks (Jacobs and van Spanje Reference Jacobs and van Spanje2021). Studies in the second group have investigated a variety of localized, mostly non-fatal attacks (Eger and Olzak Reference Eger and Olzak2023) as well as datasets including the ‘most severe types’ of terrorist attacks (Krause and Matsunaga Reference Krause and Matsunaga2023).

Finally, opposing results may also be related to differences in study contexts – for example, in terms of characteristics of the respective PRRP, baseline PRRP support, or features of party systems. In fact, all studies finding positive correlations between right-wing attacks and PRRP support focus on the case of Germany. However, Jacobs and van Spanje (Reference Jacobs and van Spanje2021) find opposing results in a cross-country analysis that includes the case of Germany – casting doubts on the assumption that studies’ case selection alone explains the discrepancy in results.

Based on the relatively small number of previous studies, we cannot rule out that any of these factors contribute to explaining the opposing results of the two groups of studies. However, we believe that empirical patterns hint at a related but more specific moderator: the two groups of studies capture different types of public reactions to terrorist attacks. While studies across both strands highlight that it is citizens’ understanding of the causes of violence that determines their attitudinal reactions, the two groups differ in how they interpret the direction of public blame attribution.

The studies that find a negative effect of terrorism on right-wing voting suggest that people associate the right-wing violence with the PRRP itself. Jacobs and van Spanje (Reference Jacobs and van Spanje2021, 740) highlight that ‘right-wing extremist terror threat could activate or reinforce a cognitive link’ between PRRPs’ political programs and terrorism. Consequently, people blame the ideology of the PRRP for the attack and distance themselves from the party (Solheim Reference Solheim2020; Bove et al. Reference Bove, Efthyvoulou and Pickard2024b).

The studies that find a positive effect of right-wing attacks on PRRP voting propose that people associate the violence not with the PRRP but with immigration. Right-wing attacks increase the media salience of immigration, while PRRPs frame right-wing violence as a consequence of immigration. Thus, right-wing terrorist attacks increase immigration concerns among citizens and thereby increase – rather than dampen – support for PRRPs (Krause and Matsunaga Reference Krause and Matsunaga2023; Eger and Olzak Reference Eger and Olzak2023).

Consequently, a closer look at the public reactions to right-wing terror attacks may help to reconcile the previous opposing findings. Several recent studies demonstrate that media framing of events can affect political opinions and behavior (Grossman et al. Reference Grossman, Margalit and Mitts2022; Foos and Bischof Reference Foos and Bischof2022; Devine and Murphy Reference Devine and Murphy2020). More specifically, previous research indicates that the type of media framing of terrorist attacks can shape people’s attitudinal reactions to these attacks (Matthes et al. Reference Matthes, Schmuck and von Sikorski2019; Solheim Reference Solheim2021; Bove et al. Reference Bove, Di Leo, Efthyvoulou and Pickard2024a). Building on these findings, we argue that the effects of right-wing terrorist attacks on PRRP support (positive or negative) depend on the way that the media reports on these attacks.

How Media Backlash Conditions the Effects of Terrorism

We propose a simple mechanism that links right-wing terrorist attacks to a decrease in support for PRRPs through anti-PRRP ‘media backlash’. Figure 1 illustrates the main elements of this argument.

Figure 1.

Proposed mechanism linking terrorism to PRRP voting through media backlash.

The term ‘backlash’ describes the reaction of dominant social groups to high-intensity events that threaten the status quo (Bishin et al. Reference Bishin, Hayes, Incantalupo and Smith2016). We conceptualize ‘media backlash’ as consisting of a sudden and substantive media reaction to right-wing terrorist attacks highlighting causal links between PRRPs’ political programs and political violence (Solheim Reference Solheim2020).

Our argument focuses on backlash in ‘traditional media’ rather than on social media for two main reasons. First, newspapers, radio, or television news still constitute people’s most important sources of information across most OECD countries (see Figure A.1 in the Appendix). Thus, we attribute the highest opinion-forming potential to this type of media. Second, research shows that news reports in these media often shape discussions in social media (King et al. Reference King, Schneer and White2017).

We highlight two processes through which media backlash can shape the effects of terrorist attacks. First, media reports ‘frame’ events (Goffman Reference Goffman1974). Gamson and Modigliani (Reference Gamson and Modigliani1989) describe media frames as ‘stories’ providing meaning to real-world developments. These ‘storylines’ can shape people’s attitudes by activating or altering the weight of considerations and by adding previously unavailable beliefs in individuals (Lecheler and De Vreese Reference Lecheler and De Vreese2019). Thus, by framing right-wing terrorist attacks as consequences of PRRP rhetoric, the media can activate cognitive links between PRRPs and extremism and violence. Previous studies show that these links can undermine public support for PRRPs (van Heerden and van der Brug Reference van Heerden and van der Brug2017; van Spanje and Azrout Reference van Spanje and Azrout2019).

Second, media reports inform their audience about the distribution of opinions in society by presenting opinion polls, interviews with politicians, and reports on public expressions of opinion such as demonstrations (Gunther Reference Gunther1998). According to public opinion research, the perceptions of dominant opinions can then produce attitude change at the individual level, as people feel an intrinsic need to be in agreement with others (Joslyn Reference Joslyn1997). Thus, an anti-PRRP media backlash communicates to citizens that a majority of the population believes that PRRPs are to blame for the right-wing terrorist attack, making citizens align their interpretations to these dominant opinions and inducing a distancing process from PRRPs.

If strong media backlash can shape political opinions, then what determines the strength of the backlash? Building on previous research on moderators of media framing effects (Lecheler and De Vreese Reference Lecheler and De Vreese2019), we argue that two main attack properties shape the likelihood of strong anti-PRRP media frames.

The first property is the intensity of the attack. Violent attacks do not always receive a high level of coverage in the media – depending on how journalists and editors evaluate the news value of attacks relative to other events. Previous research demonstrates that attacks with more ‘dramatic’ consequences tend to be associated with a higher likelihood that they are widely covered in the news (Kearns et al. Reference Kearns, Betus and Lemieux2019; Jacobs and van Spanje Reference Jacobs and van Spanje2023; Hellmueller et al. Reference Hellmueller, Hase and Lindner2022; Sui et al. Reference Sui, Dunaway, Sobek, Abad, Goodman and Saha2017). Highlighting the role of variation in media coverage, Nussio et al. (Reference Nussio, Böhmelt and Bove2021) show how more intensive attacks tend to generate higher public attention to terrorism. This leads us to the cynical expectation that a high intensity of attacks (that is, above-average numbers of dead/wounded victims) constitutes a general prerequisite for substantive media reactions. Previous research suggests that this heightened media attention may also lead to more pronounced and more long-lasting attitudinal and emotional reactions of the population (Bove et al. Reference Bove, Efthyvoulou and Pickard2024b).

The second property is the combination of the motive of the perpetrator and the identity of the victims. According to Mudde (Reference Mudde2007), ‘nativism’ is at the core of PRRPs’ narratives.Footnote ⁴ Right-wing terrorist attacks reflect this narrative if (a) they are motivated by the goal to create a homogeneous nation state and (b) if they are directed against non-native people (that is, ‘immigrants’). After such attacks, the media will be particularly likely to articulate frames that link these attacks to PRRPs – because this link is obvious to journalists and because it induces outside actors such as politicians and experts to push this interpretation into the media (see Wlezien Reference Wlezien2024; Lecheler and De Vreese Reference Lecheler and De Vreese2019).

We do not expect media coverage to be uniform. Some newspapers may be more likely to associate attacks with PRRPs than others. In particular, more right-leaning ones may be more likely to refrain from making such associations explicit in their reports on terrorist attacks. However, in cases in which the combination of the motive of the perpetrator and the identity of the victims makes ideological links to PRRP platforms obvious, we expect to see reports on these links in newspapers across the political spectrum of the mainstream media.

The media framing of right-wing terrorism may not influence the preferences of the most radical right-wing voters. First, those voters may self-select news reports, avoiding media content that associates PRRPs with violence and extremism (Burghartswieser and Rothmund Reference Burghartswieser and Rothmund2021). Second, even if these voters are exposed to the media backlash, their political convictions and related cognitive bias may make them more likely to dismiss arguments blaming PRRPs for right-wing terrorism (Arceneaux Reference Arceneaux2012).

Thus, we assume that strong and credible associations of PRRPs with extremist violence in the media are most likely to sway those voters that agree with PRRPs’ anti-immigration policies, but disapprove of extremist ideas (Jacobs and van Spanje Reference Jacobs and van Spanje2021; Collier and Vicente Reference Collier and Vicente2012). The effect of the backlash on this segment likely reduces the average public support for PRRPs. Thus, our main hypothesis is that right-wing terrorist attacks dampen public support for PRRPs, if the properties of the attacks generate a strong anti-PRRP media backlash.

Right-Wing Terrorism and PRRP Support in Germany

Since World War II, radical right parties have had limited success in German federal politics. The upper panel of Figure 2 displays self-reported voting intentions for right-wing parties over time. In the 1990s, far-right splinter parties like the DVU (German People’s Union), the NPD (National Democrats), and the Republikaner/REPs (Republicans) had some success in state elections (most notably in the states of Bavaria and Baden-Württemberg) or European elections, but none of them was ever represented in the German Federal Parliament, the Bundestag.

Figure 2.

Right-wing voting and violence in Germany, 1990–2020.

Note: panel A displays monthly share of respondents who indicate they would vote for either AfD/DVU/NPD/Republicans, based on Forsa-Bus surveys. Panel B displays all ninety-eight right-wing attacks with at least one fatality as recorded by the RTV dataset (Ravndal Reference Ravndal2016). We code attacks as ‘high intensity’ if they resulted in either two or more fatalities or two or more injured.

The ‘Alternative for Germany’ (AfD) was established in 2013 as a single-issue party focusing on the Euro crisis and fiscal policy. Since then, the party has seen a constant reorientation to the radical right and an unprecedented rise in German politics. With a leadership change in 2015, the party began to focus on immigration as its new core issue. It capitalized on grievances related to the European ‘refugee crisis’, attracting voters from radical right and centre-right parties. Over the years, the AfD’s agenda has grown increasingly populist and radical. In 2023, the party was represented in the Bundestag, and almost all state legislatures (see Arzheimer Reference Arzheimer2015, Reference Arzheimer2019 for detailed accounts of the rise of the AfD).

While the electoral success of the far right is a relatively new phenomenon in post-World War II Germany, right-wing violence has been frequent since at least the 1970s. The Right-Wing Terrorism and Violence (RTV) dataset (Ravndal Reference Ravndal2016) records a total of ninety-eight fatal terrorist attacks in Germany for the period from 1990 to 2021. No other European country has experienced a similar level of violence. The lower panel of Figure 2 displays the temporal distribution and the number of victims of the fatal ninety-eight attacks. In fact, Germany alone accounts for around 40 per cent of all fatal right-wing attacks that occurred in the sixteen countries included in the dataset.

Particularly in the early 1990s, Germany experienced a wave of intense xenophobic attacks. These attacks clearly mirrored the nativist ingroup versus non-nativist outgroup rhetoric of the right-wing parties of the time: they were mostly perpetrated by groups of German skinheads against migrants and asylum shelters. While this violence had clear political objectives and was aimed at instilling fear among migrants and political opponents, it was not discussed under the label of right-wing ‘terrorism’. The public debate changed in 2011, however, when German security services discovered the right-wing ‘National Socialist Underground’ (NSU), which allegedly assassinated at least ten persons over a period of almost fourteen years.

These and several other terrorist attacks in the 2000s differed from the violence in the 1990s in that the political motives and the identity of the perpetrator were not known immediately after the attacks, preventing any associations of the violence with the party platforms of PRRPs. In other cases, the violence was directed against the political left or against representatives of the state. Lacking obvious parallels between the victims of these attacks and the anti-migrant rhetoric of the PRRPs, they did not trigger any substantive debate on the role of PRRPs in instigating right-wing violence.

This changed, in particular, with the attack that took place on 19 February 2020 in the small city of Hanau, close to Frankfurt in Western Germany. At around 10 p.m., a 43-year-old German citizen opened fire in two shisha bars. The shooting killed nine people and injured five others. All of the victims had migration backgrounds. The attacker then fled the scene and later killed his mother before taking his own life.

The far-right background of the attack was evident. The gunman had posted a racist video and manifesto on his personal website, expressing hatred for migrants and for German citizens who had allowed immigrants into the country. The attack prompted massive public outrage. Tens of thousands of people attended demonstrations of solidarity and protests against racism (BBC 2020). The attack also triggered an intense public debate on the role of the AfD. Leading politicians and other prominent figures accused the AfD of creating the breeding ground for the racist attack (Guardian 2020).

Figure 3 illustrates how German news heavily associated the AfD with violence and right-wing extremism. An analysis of the content of more than 400 of these newspaper articles published after the attack shows that more than 75 per cent imply that the AfD shares some responsibility for right-wing violence; only around 30 per cent include defenses of the AfD against such accusations. While we find that this backlash has been somewhat more pronounced among left-leaning newspapers, we find reports associating the AfD with the attack across the entire political spectrum of mainstream media (see more detailed information on this media backlash in the Appendix Section C.2). The subsequent sections investigate if such media reactions shape the effects of right-wing terrorism on citizens’ voting intentions.

Figure 3.

Media backlash against the AfD following the Hanau terror attack.

Note: the figure displays the share of German print and online articles mentioning ‘AfD’ in conjunction with right-wing extremism (search terms: ‘AfD’ AND ‘rechtsextrem*’ (right-wing)) and violence (search terms: ‘AfD’ and ‘gewalt’ (violence)) among all articles that mention the AfD in the days surrounding the Hanau attack. Newspaper data are taken from the Genios newspaper database (http://www.genios.de).

The Effects of the Hanau Terrorist Attack

We start our empirical analysis with a single-case study of the Hanau terrorist attack. According to our theoretical argument, we expect particularly strong negative effects of the attack on AfD support due to its high intensity, its obvious nativist motives, and its non-native targets, as well as the ensuing media backlash. We draw on three different data sources and employ two types of inferential strategies to identify the causal effects of the Hanau attack on PRRP support. First, we use daily survey data on voting intentions in an ‘Unexpected Event during Survey Design’ (UESD, Muñoz et al. Reference Muñoz, Falcó-Gimeno and Hernández2020). Second, we replicate this analysis using a behavioral measure of AfD support: followers of AfD accounts on social media. Third, we investigate the results of an online survey experiment.

Effects on Voting Intentions in Daily Surveys

We use individual-level data from the Forsa-Bus survey to investigate the effect of the Hanau terror attack on voting intentions for the AfD (Forsa 2021). The Forsa-Bus samples and asks around 500 new respondents per day (as a repeated cross-section) about their electoral preferences and socio-demographic attributes. The survey is based on computer-assisted phone interviews and is representative of the German population of voting age.

In a UESD, the identifying assumption is that the timing of an event is exogenous to the timing of the interviews due to its unexpected occurrence (we discuss and investigate this assumption in Appendix Section B.1.1). In the context where the unexpected event is a terrorist attack, the attack divides the interview respondents randomly into a ‘treatment’ group (those interviewed after the attack) and a ‘control’ group (those interviewed before the attack). We define as ‘treated’ all respondents who were interviewed on or after 20 February 2020, and all those interviewed before 20 February 2020 as the ‘control’ group.

Our main specification takes the following form:

(1) $$AfD\ voting\ intentio{n_{id}} = \beta\ Post - attac{k_{id}} + \delta {X_{id}} + {\epsilon _{id}}$$

where $AfD\ voting\ intentio{n_{id}}$ is a dummy variable of the voting intention of individual $i$ in survey day $d$ for the AfD in the next state/federal elections. $Post - attac{k_{id}}$ is a dummy variable that takes the value 1 for all individuals interviewed on 20 February or later and 0 for those interviewed on 19 February or earlier. ${X_{id}}$ is a vector of the following socio-demographic covariates taken from the Forsa-Bus: gender, state, birth decade, income level, education level, occupation status, religion, children, married, and mobile versus landline sample. All covariates are categorical and enter the models that include covariates as flexible dummies (see Appendix Section B.1.2 for covariate categories). ${\varepsilon _{id}}$ is the error term. We compute heteroskedasticity-robust standard errors. All models are estimated with OLS. The coefficient of interest $\beta $ captures the direct effect of the Hanau attack on voting intentions for the AfD in the next state/federal elections. A positive sign of $\beta $ indicates that an increased share of voters prefer the AfD in the next elections whereas a negative sign represents reduced support for the AfD.

The upper panel of Figure 4 displays trends for AfD voting intentions. They show a visible drop in voters’ preferences for the AfD directly after the Hanau attack. The lower panel of Figure 4 shows the result of estimating Equation (1) with different bandwidths around the treatment date. The first specification uses the broadest treatment window of −15 to +15 days around the attack. We then narrow the bandwidth to five, three, and, most conservatively, one day before and after the attack. All models are estimated with and without a full set of covariates. Across models, we report a negative effect of the Hanau attack on voting preferences for the AfD. Similar to the visual patterns in the upper panels, the negative effect of the attack is most clearly discernible for AfD voting intentions in state elections – which is what we would expect given the AfD’s higher popularity in state elections than on the federal level (Weisskircher Reference Weisskircher2020).

Figure 4.

The Hanau attack and AfD voting intentions.

Note: the upper panel displays daily averages of voting intentions for the AfD in the next federal (left panel) and state (right panel) elections before/after the Hanau attack. The lower panel displays coefficients from OLS models with 95 per cent (thin) and 90 per cent (thick) confidence intervals based on heteroskedasticity-robust standard errors. The dependent variable is a dummy for voting intentions for the AfD in the next federal (left panel) and state (right panel) elections. Coefficients can be interpreted as percentage points. Covariates include dummies for: gender, state, birth decade, income level, education level, occupation status, religion, children, married, and mobile versus landline sample.

For both election types, the effect size ranges from about −1 to −3.8 percentage points. Effect sizes seem particularly large in models relying on a one-day bandwidth. We attribute this pattern to a single-day spike in AfD support on the day of the attack itself (interviews took place before the attack occurred) – mirroring similar upward and downward outliers pre- and post-treatment (see upper panel of Figure 4). Our preferred specification is the model that relies on a five-day bandwidth around the attack. It balances the trade-off between sample size and random noise on the one hand and exogeneity/narrow treatment window on the other hand. In this OLS model, the effect size is between −1.8 and 2.0 percentage points for the state elections, and around −1.7 percentage points for federal elections.Footnote ⁵ Given the AfD’s overall share of about 11 per cent of vote intentions in the Forsa-Bus of 2020, this amounts to a loss of up to a sixth of the party’s potential supporters as a consequence of the attack.

In Appendix B.1, we present a battery of robustness tests of our main results, following the advice for UESD robustness tests proposed by Muñoz et al. (Reference Muñoz, Falcó-Gimeno and Hernández2020), including balance tests, placebo tests, multiple bandwidths, falsification tests, analysis of non-responses, differently estimated standard errors, and consideration of survey weights. None of the tests significantly challenges our main finding.

In Appendix B.2 we probe a series of alternative explanations. First, we assess the possibility that COVID-related events drive our main findings. We compare the salience of the pandemic and the Hanau attack in Twitter activity over time. While there was some increase in the public salience of the pandemic, the Hanau attack was the dominant issue during the first few days after the incident. Thus, it seems highly implausible that our main findings – especially the ones using small time windows – are driven by the pandemic.

Second, we assess the role of social desirability bias. Mirroring the approach of Singh and Tir (Reference Singh and Tir2023) we investigate the treatment effect on alternative outcomes that (1) gauge socially desirable ‘society-supporting’ behavior and that (2) should not be affected by the treatment: reported turnout in prior (pre-treatment) elections, reported AfD voting in prior elections, and non-responses. We find null effects of the attack on these placebo outcomes.

Third, we draw on monthly surveys on perceptions towards refugees to assess the alternative explanation that the attack changed citizens’ attitudes towards refugees. We observe a small reduction in negative attitudes towards refugees only more than twenty days after the attack – after the announcement of COVID-related restrictions. This time gap makes it implausible that this attitudinal change reflects an effect of the Hanau attack.

Finally, we estimate treatment effects on support for other parties beyond the AfD. We do not find evidence that the Hanau attack triggered a swing in public opinion away from the smaller parties towards the governing parties (rally-around-the-flag). More generally, this analysis indicates that our main effect reflects a shift away from the nativist AfD to mainstream parties to the left of the political spectrum (that is, the Social Democrats and the Green Party).

Effects on AfD Support in Social Media

One potential weakness of our previous analyses is that they rely on self-reported intended voting behavior. It is unclear, however, to what extent respondents would really be willing to act in line with their survey responses. In the absence of actual voting data around the Hanau attack, we rely on an alternative behavioral measure: we investigate to what extent the attack prompted people to ‘unfollow’ AfD party accounts on Facebook. Contrary to the survey-based outcome investigated above, individuals must become active to signal a change in political preferences – a signal that is also unaffected by potential interviewer bias.

We consider the number of Facebook followers for accounts created by both state-level as well as federal-level representatives and entities. For every state we include the account of the party’s regional association and, if existent, the official page of the AfD’s faction in the state parliament. At the federal level, we include the accounts of the party’s national association, the faction in the parliament, and every AfD representative who has been a member of the party’s federal board during our period of investigation. Our final dataset includes information on fifty accounts in total. We collect information regarding daily numbers of followers for each account through the Intelligence Tool of the CrowdTangle project (CrowdTangle 2022).

The upper panel in Figure 5 shows average follower trends for federal- and state-level Facebook accounts of the AfD. We observe a discontinuous decrease in the average number of followers per Facebook page around four to five days after the attack for both federal- and state-level accounts.

Figure 5.

The Hanau attack and AfD Facebook followers.

Note: the upper panel displays daily average page likes for AfD accounts of federal-level (left panel) and state-level (right-panel) entities and representatives. The dotted vertical line indicates the cut-off used in our specification. The lower panel displays coefficients from OLS models with 95 per cent (thin) and 90 per cent (thick) confidence intervals based on heteroskedasticity-robust standard errors. The dependent variable is the daily growth rate in the number of users who follow an account through a page like.

As unfollowing requires a user to actively detach his or her name from an individual account, we think it is plausible that any observable effects of the attack will likely operate with a time lag. We validate this assumption by looking at another instance of user reactions to macro-political discourses: In January 2020, the head of the Bavarian state government publicly accused the AfD for the rise in antisemitic sentiment across Germany (Zeit Reference Zeit2020). While this led to a sharp decrease in Facebook users following the accounts of the AfD’s branch in Bavaria, the reaction only materialized between three and four days after the beginning of the public debate. We therefore shift the cut-off in our estimation by four days. We replicate our main analysis using this alternative outcome and report the results in the lower panel of Figure 5. To account for time trends in the data, we use the daily follower growth rate as the outcome. Overall, the results mirror the findings regarding changes in self-reported voting intentions.

Effects in a Survey Experiment

We implement a (pre-registered) survey experiment to provide further robustness of our findings so far and to probe the media backlash mechanism for the Hanau case. We administered the survey among 3,000 members of the online panel of the German provider Bilendi, focusing on the voting population of 18- to 74-year-old German citizens.Footnote ⁶ The sample is representative of that target population in terms of age, gender, and region.Footnote ⁷ Data collection took place between 13 April and 9 May 2023.

The experiment randomly assigned respondents to a control group and two treatment groups (see Appendix B.3.1 for the vignettes). All three versions of the survey confronted respondents with media content related to fatal violent attacks that occurred in Germany in 2020: newspaper headlines, press photos, and a bar chart of public opinion polls related to the respective attacks. We designed all three control and treatment conditions to be as similar as possible in terms of content and visualization.

The control vignette presents media information about an actual, but apolitical attack: in December 2020, a man ran over pedestrians with an SUV at high speed in the city of Trier, killing six people. Police investigations associated the attack with the perpetrator’s mental health problems and alcohol abuse. We selected media content to reflect this apolitical nature of the attack.

The first treatment vignette presents headlines, press photos, and public opinion results on the Hanau attack. We designed the vignette to explicitly prime the radical, right-wing nature of the attack. The headlines emphasize the racist motive of the perpetrator and the migrant background of the victims.

The second treatment aims to explicitly prime respondents on the media backlash against the AfD following the Hanau attack. It differs from the first treatment condition in only two respects: first, while both treatment vignettes include pictures of public protest after the attack, the picture of the second treatment condition features a large sign with the slogan ‘The AfD also fired shots’, mimicking the way news media portrayed the connection between attacker identity and the AfD. Second, while both vignettes include figures reporting the results of opinion polls, the figure of the second treatment condition adds the information that 60 per cent of Germans agree with the statement ‘The AfD shares responsibility for right-wing violence’.

Thus, contrary to the first treatment, the second treatment combines both suggested channels through which media backlash may reduce public support for PRRPs: (1) information on the motives of the perpetrator and the identity of the victims, activating cognitive links between PRRPs and terrorism, and (2) information on the distribution of opinions in society, highlighting that the dominant interpretation of the attack emphasizes the responsibility of the PRRP for the attack.

Our main outcome of interest is AfD party support. Our pre-registered expectation is that respondents exposed to media reports on the right-wing Hanau attack will display lower support for the AfD than respondents exposed to the apolitical Trier attack. In addition, if it is particularly the media backlash to the terror that drives respondents’ alienation from the AfD, we should see a stronger negative effect of treatment two on AfD support than of treatment one.

We use two survey items to measure the dependent variable, AfD support. We asked respondents (1) how likely it is that they would ever vote for the AfD and (2) to what extent the political objectives of the AfD match with the respondents’ political interests. Both items are measured on a 7-point Likert scale, ranging from −3 to +3, allowing us to capture more nuanced changes in AfD support as opposed to a simple voting intention.Footnote ⁸ We create an ‘AfD support’ index by taking the mean of both items.Footnote ⁹

Experimental assignment across the control and two treatment groups worked well (see Appendix B.3.2). Nevertheless, slight imbalances occur with respect to gender and political interest in treatment group 2. We therefore report results with and without adjustment for a set of pre-treatment covariates, as specified in the pre-analysis plan. Covariates include measures of gender, age, political interest, education, and six items measuring party preference for all German parties in the Bundestag.Footnote ¹⁰

Our survey also included an attention test that allowed us to assess treatment compliance – that is, the extent to which respondents really noticed the specific content of the vignettes presented to them.Footnote ¹¹ Following our pre-analysis plan we only keep respondents in the control and treatment groups that correctly identified at least one of the three attention tests. Importantly, as we administered attention tests to both the control and the treatment conditions, we mitigate typical differential attrition problems (and resulting biases) that arise when researchers drop only treated respondents who fail an attention test (Aronow et al. Reference Aronow, Baron and Pinson2019).Footnote ¹²

The main results of the survey experiment are reported in Figure 6. The plot displays coefficients from OLS models that regress AfD support on differently specified treatment dummies with the control group being the reference category.

Figure 6.

Replicating the effects of the Hanau attack in a survey experiment.

Note: the plot displays coefficients and 95 per cent (thin) and 90 per cent (thick) confidence intervals from OLS regressions of the AfD support index on differently specified treatment dummies (control condition is the reference group). The dependent variable is an index variable for AfD support, ranging from 1 (low) to 7 (high) and coefficients are on the scale of this index. Covariates specified in the pre-analysis plan: pre-treatment measures of gender, age, political interest, party preference for all six German parties in the Bundestag, including pre-treatment AfD preference.

We report two main findings. First, we replicate the negative effect of the Hanau attack on AfD support from the UESD in the survey experimental set-up. Across specifications and treatment groups, we observe a negative point estimate of being exposed to the right-wing Hanau terror attack on AfD support compared to being exposed to the apolitical Trier attack. While estimates of models without covariates are noisy with large confidence intervals (grey coefficients), estimates become more precise once we included the pre-registered set of covariates. Additional analyses indicate that it is particularly supporters from the politically conservative/liberal spectrum, including AfD supporters, who are most affected by the treatment (see Appendix B.3.6).

We compare the effect size that we find in the survey experiment to the effect size estimates from the natural experiment reported above (see Appendix Figure B.3.5). Effect sizes are very similar between the two designs and data types: we estimate that the treated groups in the survey experiment are about 1.6 percentage points less likely to ‘ever vote for the AfD’, compared to −1.8 to −2 percentage points in UESD analyses with the Forsa-Bus data. The similar effect sizes reinforce our confidence of our main findings. In particular, they also suggest that the results of the UESD analysis are unlikely to be purely driven by compositional effects in the Forsa sampling before and after the attack, since sample composition is constant in the survey experiment.

Second, and contrary to our expectations, we do not find stronger effects for the more pronounced media backlash exposure (treatment group 2). Point estimates for the two treatment conditions do not differ in a substantive and statistically significant way. We view the presence of ‘pre-treatment effects’ as the most likely explanation for this result (Kane Reference Kane2025): because most respondents have been exposed to the actual media backlash after the Hanau attack, both of our treatments may reactivate individuals’ association of the party with right-wing extremism and violence. In line with this post hoc explanation, we find that respondents in both treatment groups agree more with the statements ‘Many people in Germany blame the AfD for right-wing extremist violence’, ‘Extremist thought is widespread in the AfD’, and ‘The AfD increases the risk of conflict and violence in Germany’.

An alternative explanation for the similarity of the effects of the two treatments could be that the information on the motives of the perpetrator and the identity of the victims has been enough to create a link between the attack and the party platform of the AfD. Additional information on public reactions and interpretations of the attack provided in the second treatment vignette may not have reinforced this link in any substantive way.

In Appendix B.3.4 we assess other potential mechanisms. First, we probe a potential impact of our treatment on attitudes towards refugees. However, we do not observe a precisely estimated effect of any of the two treatment variants on attitudes towards refugees.

Second, we test the possibility that treatment effects represent a pure social desirability bias. Our results suggest that at least parts of the main effects of the treatment without explicit media backlash prime may reflect social desirability bias. We do not see a similar pattern for our main backlash treatment.

Analysis of the Full Population of Fatal Right-Wing Attacks in Germany, 1990–2020

To what extent does our theoretical argument generalize out beyond the case of Hanau? And what is the precise role of media backlash in linking right-wing terror attacks to PRRP voting preferences?

We have selected the case of Hanau because it fits the two attack properties that are likely to generate media backlash – high attack intensity as well as a match between victim identity and a PRRP anti-immigrant narrative. But for our theoretical argument to hold more broadly, we should observe two additional empirical patterns (cf. Figure 1). First, attacks that lack these properties should not generate an observable media backlash against PRRPs. And second, the negative effect of right-wing attacks on PRRP voting intentions should be strongest in cases in which we observe a strong media backlash against the PRRP.

We turn to an analysis of the full population of fatal right-wing attacks in Germany between 1990 and 2020 to investigate these two observable implications – we draw on the RTV dataset to identify all ninety-eight attacks that resulted in at least one fatality (Ravndal Reference Ravndal2016).

In order to probe our first observable implication, we classify the ninety-eight attacks along two dimensions. First, we distinguish between low-intensity and high-intensity attacks. A high-intensity attack is one in which at least two or more people died or in which two or more people were wounded; low-intensity attacks are all the others. Second, we differentiate between different types of high-intensity attacks. We use RTV data on target groups to code attacks that were directed against individuals or groups that fall into the AfD’s nativist outgroup definition: (1) Muslims and (2) immigrants/foreigners/asylum seekers/refugees. We also hand-coded information on perpetrator motives from news reports. We code attacks as reflecting PRRP narratives if news reports indicate that nativist motives of the perpetrators were obvious (for example, when groups of Nazis publicly attacked refugee shelters to create a homogeneous homeland) and/or formally reported by the police or other state authorities in the immediate aftermath of the respective attacks (on the day of the attack or the day after the attack). This classification allows us to compare three types of attacks: (1) low-intensity attacks, (2) high-intensity attacks that do not reflect PRRP narratives, and (3) high-intensity attacks that reflect PRRP narratives.Footnote ¹³

We compare the extent of the media backlash across these three different types of attacks. We draw on a database of German newspapers to quantify the media backlash. We identify the daily number of articles including references to right-wing violence and PRRPs within three days before and after each violent attack. Crucially, since we extend the analysis all the way back to 1990, we do not only capture newspaper reactions to the AfD, but also to other PRRPs that were active before the rise of the AfD, such as the NPD or the Republikaner. We then calculate the share of PRRP articles referring to right-wing violence and PRRPs within these three-day windows and use the difference in this variable before and after an attack as a measure of the strength of media backlash.

Figure 7 displays the average level of media backlash across different types of attacks. The figure supports two core assumptions of our theoretical argument. First, we find that right-wing attacks of relatively low intensity generate only weak responses by media outlets on average. Second, while the share of articles discussing PRRPs and right-wing violence generally increases after high-intensity attacks, it is particularly those attacks targeting outgroups defined by PRRPs’ nativist ideology and in which a perpetrator acts with a clear right-wing ideology that generate a high level of media backlash.

Figure 7.

Properties of right-wing terrorist attacks determine the strength of media backlash against PRRPs.

Note: the plot displays the level of media backlash across different types of attacks. We measure the level of backlash by identifying all German newspaper articles mentioning a PRRP (search terms: ‘AfD’ or ‘Republikaner’ or ‘NPD’ or ‘DVU’) as well as articles referring to violence (search terms: ‘Terror’ or ‘Anschlag’ (attack) or ‘Gewalt’ (violence)) within windows of three days around attacks. Newspaper data are taken from the Genios newspaper database (http://www.genios.de).

Are negative effects of right-wing attacks on PRRP support strongest in cases with high anti-PRRP media backlash? In order to probe this second implication of our argument, we extend the same UESD as used in the analysis of the Hanau case to the ninety-eight other fatal attacks. To measure support for PRRPs before the establishment of the AfD in 2013, our main outcome variable dummy codes the vote intention for any of the following three different parties (as well as the AfD). Die Republikaner (The Republicans), founded in 1983, entered the European and a German state parliament in the early 1990s. Contrary to The Republicans, the two other right-wing parties were both classified as ‘right-wing extremist’ by the German domestic intelligence service: The Nationaldemokratische Partei Deutschlands (National Democratic Party of Germany, NPD), established in 1964, and The Deutsche Volksunion (German People’s Union, DVU), founded in 1987 and later merged into the NPD.

Figure 8 shows the results of UESD analyses comparing right-wing voting intentions in five-day windows before and after attacks at different levels of media backlash in all fatal attacks (left panel) and in high-intensity cases where the motive of the perpetrator and the identity of the victim group coincide with the ideology of PRRP parties (right panel).

Figure 8.

The effect of right-wing terrorism on PRRP voting intentions across different types of right-wing attacks, 1990–2020.

Note: the plot displays coefficients of OLS models that predict voting intentions from a post-attack dummy for any of the following PRRPs: AfD, DVU, NPD, Republikaner (REP) using daily Forsa-Bus surveys with 95 per cent (thin) and 90 per cent (thick) confidence intervals, based on robust standard errors. Dependent variable is a binary indicator of vote intention and coefficients can be interpreted as probabilities. We use a +/− five-day window around each attack date and remove all survey days that overlap across attacks. Models include attack ID fixed effects as well as the following covariates: state (Bundesland), gender, education, employment status, and birth year. We estimate separate models for (1) each of the two categories of possible attacker/victim constellations and (2) the full sample and a sample that keeps only respondents from the states of Bavaria and Baden-Württemberg before 1995. We use the median of all cases with non-zero values in our backlash measure to identify high backlash cases.

We report two results. First, we look at all fatal attacks and observe substantive and statistically significant effects only for attacks that were accompanied by substantive media backlash – in line with our theoretical expectations (left panel). Second, in order to rule out that the characteristics of the attacks confound the moderating effects of the media backlash, we hold the type of attack constant and focus only on attacks that are clearly motivated by far-right ideology and directed against outgroups matching PRRPs’ nativist ideology (right panel). Even when conditioning on the type of attack, we still find evidence that only those attacks that are accompanied by high levels of media coverage linking PRRPs and right-wing violence elicit anti-PRRP effects (effect estimates are statistically significant at the 10 per cent level).

Taken together, the patterns presented in Figure 7 and Figure 8 therefore suggest that it is indeed the media backlash – driven by attack properties – that shapes the effect of right-wing violence on PRRP support.