Research questions

This paper is dedicated to the patterns of xenophobic attacks in Germany following the so-called refugee crisis in 2015/16. As mentioned, the high numbers of arriving refugees and asylum seekers were politicized far beyond right-wing extremist parties and led to unprecedented numbers of violent attacks against refugees and asylum seekers [for a historic overview, see 6]. We draw on many other studies that have investigated certain aspects of xenophobic violence against refugees in this time period [7–13]. However, our paper tries to include a wider picture of potential explanatory factors. Namely, we combine the substantive approaches of contagion effects, economic threat, contact theory, political opportunity structures, and look at the effect of media reports, fake news, and salient events that were perceived as threatening by the German population, namely Angela Merkel’s decision to keep the border open in September 2015 and New Year’s Eve in Cologne 2015/16. Summing up, we are interested in answering the following research questions:

• Do previous attacks increase the probability of further attacks in the area?
• Which roles do the media and fake news play?
• Do salient threatening events increase violence?
• Is violence a reaction to perceived economic threats?
• Do political opportunity structures of right-wing parties matter?
• Do opportunities for contact between ethnic groups diminish violence?

Our focus for this endeavor is on arson attacks (for the only other study that analyzes arson attacks separately, see [11]). Arson attacks have a long tradition in the history of xenophobia and racism [6], and are usually directed against the residential places where minority groups live or are designated to live. Arson attacks constitute a particularly visible and extreme form of violence with a clear signal: Offenders deny members of the minority groups the right to a new home. Due to the severity of this form of violence against refugees, such incidents should be less prone to underreporting than other common expressions of xenophobic violence, such as personal attacks or hostile demonstrations. In the following, we briefly introduce theoretical perspectives and related previous research on xenophobia.

Contagion effects: Transmitting violence through the social network

The idea that behavioral phenomena spread through social networks like infectious diseases is not new. Just like medical or industrial innovations, fashion trends, drug use and social movements of all kinds [14–16], researchers have conceptualized political violence (such as wars, coups, disorders, terrorism, and homicides) as phenomena that diffuse through social networks by contagion [14,17–21]. In more recent years, scientific studies have increasingly focused on–and provided evidence for–contagion effects related to the spread of xenophobic violence in time and space [8,9,11,22,23].

In principal, contagion effects can refer to ideas as well as behaviors. In a nutshell, they can occur whenever information is transmitted form one actor to another within a social network. As Burt [24] puts it, social contagion requires an actor who already has adopted an innovation or a certain behavior, and another who has not. As soon as information is transmitted from the first to the second actor, an idea or behavior can be imitated and adopted, that is, contagion can occur. Two mechanisms have been put forward to explain the mechanics of contagion effects. Firstly, they can happen by personal contact and communication, usually by word-of-mouth. Secondly, it has been argued that physical proximity is not a necessary precondition for social contagion. Rather, Burt assumed that people are more likely to exchange information and imitate behavior if they are socially proximate or similar (“structurally equivalent” or “socially homogeneous”).

Applied to the field of xenophobic attacks, we expect a contagion effect that diffuses violence through time and space: After one attack has happened in a certain district, we expect a heightened risk for further attacks, within the same county and in adjacent districts due to geographical proximity.

In addition, we will test a range of socio-political characteristics on the county level that might affect the number of xenophobic attacks (such as population size, share of immigrants and unemployment rate). These will be discussed later in the context of other hypotheses. However, finding the same patterns of xenophobic violence in counties that are similar in these socio-political characteristics could be a hint towards a contagion effect that goes beyond geographical proximity (as suggested by Braun and Koopmans [22]).

Media coverage

Social psychology offers an additional perspective on the spread of violence through contagion. If we assume that social norms are not fixed for all times but an agreement that we can renew or alter in our daily interactions (as symbolic interactionism would argue), different social norms can spread through a network just as any other idea and behavior. Some of these might be related to anti-immigrant sentiment, might downplay or trivialize violence and make xenophobic attacks more likely. This is why we now turn to the media, a powerful instrument of our time to disseminate ideas and change public opinion on a much bigger scale than personal communication.

We will distinguish between the following types of media contents:

• Reports on attacks against immigrants fueling contagion effects (might be prone to imitation, that is, contagion effects, possibly because they normalize violence)
• Anti-immigrant sentiment in social media and fake news (seem particularly suited to evoke changes in social norms)
• Salient threatening events with omnipresent media coverage such as Merkel’s border opening and New Year’s 15/16 (might influence public opinion in a way that justifies violent behavior against foreigners)

Media reports on attacks against immigrants driving contagion effects

Media reports on arson attacks can be better understood from the perspective of agenda setting and framing theory [25–27]. Media outlets have the power to determine which aspects of the refugee crisis receive attention and how they are portrayed to the public. They can choose to highlight certain events, individuals, or perspectives while downplaying or ignoring others. These processes of agenda setting can be driven by different motivations including political affiliations, commercial interests, or the desire to cater to target audiences. They can shape the public’s perception of refugees and might potentially foster anti-immigrant sentiment.

On the one hand, agenda setting involves topic selection, in this case the decision whether to cover the refugee crisis as a significant news item at all, on the other hand it refers to the narratives that are conveyed with the media report. How is the topic framed: Does the report focus on an individual human story, on economic impacts, security concerns or political dimensions? Who is given a voice: politicians, NGOs, or refugees themselves? Which emotions are evoked by visual representations?

In the context of xenophobic violence, where the media is a powerful tool of transmitting violent behaviors through networks, many studies have focused on which events are reported and how prominently this is done. The mere number of studies on media portrayals of refugees in different European countries shows how important the media are and how controversial their role is [28–32].

Some of these analyses examine contents, some focus on the selection of contents. To name some non-representative examples, Georgiou and Zaborowski [33] conduct a content analysis of media reports related to the refugee crisis in eight European countries and conclude that immigrants were often either portrayed as vulnerable or as dangerous outsiders. Over time, empathetic reports were increasingly replaced by narratives of suspicion. Concerning media coverage, some researchers talk about selection bias, suggesting that attacks against immigrants that involve more victims and take place in closer proximity are more likely to be reported [3]. Others argue that key events trigger an overrepresentation in media coverage for following similar events [4] or point towards the disproportionate media coverage of dramatic, rare and violent content [34].

The worry is that such news reports trigger a temporary shift in social norms, making xenophobic violence more legitimate for others. This is in line with broken window theory, which argues that visible deviant behavior has a signaling function for other actors who are led to believe that such behaviors are accepted [35]. Dependent on the tone of the report, it seems plausible that media coverage of xenophobic events can have the power to challenge social norms of tolerance, human rights, and peaceful coexistence in favor of violence.‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬

Empirically, some studies have found links between media coverage and outbursts of violence [22,23,36]. For example, Braun & Koopmans [22] find that the rate of violent events against immigrants increases when the media previously covers such events in a more salient way (that is including a photo or reporting on the front page). Both, Braun and Koopmans [22] and Jäckle and König [9], found that the number of negative statements about immigrants in the media increased violence in the subsequent time period. Braun [23] does not replicate this effect but finds that any the mere salience of immigration as a topic on the news increases violence, irrespective of tone. In addition, media reports of previous attacks against immigrants made further attacks more likely.

There also is a scientific consensus that the effects of media coverage on xenophobic violence are rather short-lived, that is, they mark an only temporary change in social norms. In his study on racial rioting in the US, Myers [3] finds that severe riots increase the chances for further outbreaks for about one week before the effect diminishes. Braun [23] uses a time period of the previous 30 days to measure diffusion of xenophobic attacks in the Netherlands.

Fake news fostering anti-immigrant sentiment

Apart from media reports on xenophobic attacks, we argue that also fake news play a crucial role in spreading anti-immigrant sentiment. Previous research suggests that right-wing organizations typically use social media to deliberately spread misinformation and hatred [37]. Wahlström and Törnberg [38] identify three mechanisms by which social media can generally lead to violence: a) users look for information that confirms their beliefs, b) they express mutual recognition and legitimization of extreme beliefs, and c) they share practical information (e.g., on how to burn down a house) and co-ordinate violent activities.

Müller and Schwarz [39] empirically show that anti-refugee sentiment on Facebook predicts crime against refugees on a municipal level in Germany by investigating users that engage on the AfD (Alternative für Deutschland, right-wing populist party) facebook page. An additional characteristic of social media that seems crucial is its lack of content management or supervision. Since anyone can post and share contents without any serious quality control, extreme perspectives can easily be expressed. These circumstances make it particularly easy for more and less organized right-wing and populist entities (for an overview see e.g. [40]) to actively engage in spreading their ideology and influencing public opinion, while fact-checkers are facing difficulties in disseminating their corrections [41].

To the best of our knowledge, no other study has examined the correlation between fake news and xenophobic violence during the refugee crisis. However, there is some descriptive research on fake news in other contexts, such as the Myanmar genocide [42], the mob lynchings in India [43], the storming of the US capitol [44], and the Brazilian Presidential Election 2018 [45]. A study on the presidential election 2016 in the US [46] shows that social media users are much more likely to believe fake news that corroborate their own ideological and political beliefs. From a psychological perspective, this finding relates to concepts of motivated reasoning and confirmation bias, suggesting that people are more likely to consume and accept information that reinforces their already existing beliefs, while perceiving information that contradicts their views as less convincing [47]. For violence against refugees, this means that exposure to fake news is unlikely to convince people with opposed opinions but might well contribute to the radicalization of people with pre-existing anti-immigrant sentiments. Such processes of radicalization seem particularly concerning, since we know from previous research that fake news spread faster through social media than facts [48] and that negative emotions can easily become contagious to other people in an online setting [49].

Salient ‘threatening events’ changing public opinion

Slightly different from Jäckle and König [9], who focus on terrorist attacks and their impact on xenophobic violence, we use the term ‘threatening event’ for the rare incidents when media reports that portrayed refugees as a threat became omnipresent throughout the whole country. The most obvious example of such an incident surely is New Year’s Eve in Cologne 2015/16, during which women were insulted, sexually harassed and occasionally robbed, mainly by refugees from Algeria, Morocco, Iraq and Syria. The event was omnipresent in the German media for a while, particularly because local police seemed to be unable to cope with the situation. It is a wide-spread belief that the tone of news reports and the public opinion towards immigration changed as a consequence of these events, particularly towards male Muslim refugees [32,50]. Köttig and Sigl [51] argue that the night "marked a turning point in perceptions of the need to protect refugees, especially male refugees". This result is somewhat confirmed in a factorial survey that finds lower acceptance towards immigrants after New Year’s Eve 2015 [52], although the experimental design of the study is not flawless [53]. In addition, some studies find that the incidents at New Year’s Eve 2015/16 were indeed followed by a rise in xenophobic violence against refugees [10], possibly as an act of revenge against a perceived out-group. A similar shift seems to have happened after Angela Merkel decided to keep the borders open for refugees on the Balkan route in September 2015 (Merkel’s ‘border opening’). The period that followed has been coined the ‘refugee crisis’.

Unlike xenophobic attacks, threatening events do not offer a clear behavior that could be imitated. Instead, they are connected to a more or less vague compilation of anti-immigrant sentiments, which can contribute to a shift in public opinion and a temporary or permanent change in the acceptance of violence against immigrants. We expect direct nation-wide effects on xenophobic violence, going beyond attacks in geographical or social proximity to previous attacks.

Economic threat

Although the impacts of immigration on wages and employment seem to be rather small (for a meta-analysis see [54]), a popular argument when talking about anti-immigrant sentiments and xenophobic violence is the concept of economic threat or deprivation [55]. The idea is that competition for scarce goods, such as employment, is what triggers xenophobic attitudes and actions. In other words, xenophobic attacks should be more likely where unemployment is high and numbers of immigrants are large.

Generally speaking, empirical evidence on the topic is mixed [56,57]. Studies range from analyses of survey data on attitudes about immigration [58–61] or ethnic prejudice [62] to macro-level analyses of the correlation between economic threat and interracial crimes in the US [63,64] or economic threat and right-wing violence in Germany [65,66].

Out of the studies that have investigated xenophobic violence during the refugee crisis in Germany, none has found convincing empirical evidence for this idea. Generally, studies do not find an effect of unemployment rate or other related measures [8,9,11,13]. In their study on anti-immigrant attitudes, Entorf and Lange [7] report a small effect that disappears when controlling for East- versus West-Germany.

One obvious reason for this apparent lack of economic threat is that refugees face serious disadvantages on the German labor market. Their right to work depends on their residence status, and their capacity to compete is limited by potential problems to get their degrees recognized. Nonetheless, some researchers have argued that xenophobic attitudes might be triggered by subjective perceptions of economic threat rather than objective situations of greater competition. This idea is somewhat supported empirically by Semyonov et al. [67] who find that the perceived size of the foreign population is associated with perceived threat while the actual size is not.

Contact theory

A somewhat contradictory effect of immigration would be expected in the light of contact theory [68]. In a nutshell, it states that contacts between different ethnic groups reduce prejudice. Higher numbers of immigrants yield more opportunities for contact between the majority and minority groups within a society and should prevent or reduce prejudice on both sides.

Empirical evidence is somewhat mixed. Among the studies that have analyzed xenophobic violence in the context of the refugee crisis in Germany, some have found support [7–9] or partial support [11,13] for contact theory. Schaub et al. [12], however, who examine refugee-related attitudes and behaviors around the same time, come to a more skeptical conclusion. The study uses a matching approach to compare survey data from 236 German municipalities, out of which half received refugees, while the other half did not. Interestingly, the authors find that local exposure to refugees did not have any effects on right-wing, populist or anti-refugee attitudes. Similarly, the following natural experiment on attitudes in the context of the refugee crisis comes to skeptical conclusions: Using survey data from different islands in the Aegean Sea that have partly experienced considerable refugee arrivals while others have not, Hangartner et al. [69] find sizable increases in hostile attitudes against refugees and immigrants among the local population.

More generally speaking, the idea that contact reduces prejudice has been corroborated in some studies [70,71], while others have found the opposite effect [22, see 72 for an experimental trial]. Two meta-analyses, one of them only on studies with random assignment to treatment groups, come to the conclusion that contact “typically reduces prejudice” [73,74].

Wagner points out a reason why some studies might fail to find empirical support for contact theory. He argues that mere contacts are not enough: “It is friendship, that is, intimate and personal contact, which has an effect on prejudice.” [27,75]

Piatkowska [76] offers an alternative interpretation for the negative effect of immigration on xenophobic violence that she finds: As the size of the minority population increases, their potential power and the possibility for retaliation increases and therefore makes crimes against them less likely (“power-differential hypothesis”).

Prevalence of right-wing/populist parties and ideology

As previous research has argued, right-wing parties portray immigrants as enemies, evoke feelings of hatred and, in this way, contribute to a political climate that is more conducive to xenophobic violence [77]. Particularly PEGIDA and later the right-populist AfD became popular actors during the refugee crisis for fueling the fear that Germany would be swamped by uncontrolled mass immigration, bringing chaos and terrorism to the country [78,79]. In addition, right-wing and populist parties are usually well connected to other groups that are susceptible to xenophobic violence and easy to mobilize [80].

In line with typical right-wing ideology, frequent motives among perpetrators are competition for jobs, cultural or value-related threat, exploitation of the German security system ([81], although the author also points out that other non-ideological motives such as group dynamics might drive xenophobic attacks). In light of these considerations, one would expect that greater strength of right-wing parties and hence visibility of right-wing ideology should shape public opinion in a way that fosters xenophobic violence.

Interestingly, also the exactly contradictory argument has been put forward in previous research on xenophobic violence [22,82]. The underlying idea is that representation within the established circles of politics channels anti-immigrant sentiments into legal boundaries and makes extra-institutional mobilization superfluous. In the context of xenophobic violence, this means that a strong representation of right-wing parties on the municipal level should reduce the need for extra-institutional violence against immigrants. Looking at previous research, empirical evidence that the strength of right-wing parties inhibits extra-institutional outbursts of xenophobic violence is provided by some studies [22,83], but not by others [23,82].

East vs. West Germany

Many studies on xenophobic violence in Germany test and find differences between the West and the former Socialist East, with the East yielding consistently more xenophobic violence despite far lower numbers of immigrants (see e.g. [8]). Sometimes these differences can be partly or entirely explained by unemployment levels or other socioeconomic characteristics, sometimes they remain significant despite controlling for such variables [13,22,65,84]. Adam [85] sees the reason for the remaining socio-cultural differences in the “official denial of the fascist legacy” in East Germany, while the West engaged in an “effective re-education policy” (p.1), for an explanatory attempt on the basis of survey data, see [86]). Whatever the exact reasons might be, we will distinguish between East and West Germany to capture such socio-cultural differences.

Summing up, we hypothesize that the following indicators correlate with the likelihood of xenophobic attacks. If not stated otherwise, we expect positive correlations.

1. (a) the number of previous attacks in spatial neighborhoods
2. (b) local media reports on attacks nearby
3. (c) anti-immigrant fake news originating nearby
4. (d) salient threatening events with nation-wide media-coverage

(indicated by a dichotomous variable that marks the time point of these events)

1. (e) unemployment in the county
2. (f) share of immigrants in the county

(contact theory predicts positive, economic threat literature negative effect)

1. (g) right-wing voters in the county
2. (h) East (vs. West) Germany

Time-variant county-level data (lagged variables)

To test hypothesis a, we generate a lag variable indicating (previous) attacks in the time period before the current time period for all counties and their (direct) neighboring counties. We use this next neighbor approach (all neighbor counties with a common border to the county of focus) for the main analysis.

As mentioned earlier, we categorize the time from 2015 (first quarter) to 2017 (first quarter) into 59 intervals of two-weeks each. As a consequence, the actual time interval between two different attacks measured by the indicator variable in one county could be between one day and almost four weeks. The number of (previous) attacks ranges from 0 (in 93.6% of all counties in the space-time matrix) to 6 (for only 10 counties in the space-time matrix). For 1,482 counties, at least one attack occurred in this or a neighboring county in the previous time period. Note that using the lag variable reduces the number of cases by 402 since there is no lagged information available for the first time interval under study. This leaves us with a total of 23,316 county-periods for analysis.

Local media

As outlined above, we additionally consider the local media coverage of (previous) attacks (hypothesis b) as well as fake news about refugees (hypothesis c) with a geographic reference. Equivalent to the specification used for previous attacks, we model their delayed and their spatial impact.

We approximated the media coverage of the event data with the help of the DIGAS press database provided by the Axel Springer Syndication. We decided to use the regional editions of BILD. BILD is a very popular daily tabloid newspaper that is read all over the country. It is particularly suited for our purposes because it does not consist of just one nation-wide edition but instead comes with a few pages on regional news with a more targeted geographical focus, which is where we would expect coverage of local arson attacks. These regional editions are comparable in scope and provide consistent coverage across the country. A list of the included regional editions of BILD can be found in S2 Appendix. For each arson attack, we restricted the text corpus to reports published on the incident date up to one week after the incident. We searched with the following (German) search query:

(Brandanschlag OR Brandsatz) AND (Asylbewerberheim OR Flüchtlingsheim OR Asylbewerberunterkunft OR Flüchtlingsunterkunft OR Asylunterkunft OR Asylheim)AND(`Incident Location´)

It combines the keyword “arson” with a range of terms referring to facilities for refugees. `Incident Location´ reflects the location string of the current incident.

Before querying, we removed over-identifying words from the location string, meaning that e.g. Berlin-Kreuzberg was reduced to Berlin. After removing false-positive hits from the result lists, we generated a binary variable indicating whether an arson attack was covered by the newspaper or not. Transferred to a panel lag format and using the neighbor concept for counties, the media coverage variable reflects the number of arson attacks reported on in a specific county-period c_t-1 for the county and all direct neighbours.

In total, this strategy leads to a total number of 265 media coverage units in BILD. In 98.9% of all counties in the space-time matrix, we record no media reports. As expected according to hypothesis b, local media coverage correlates with the dependent variable of arson attacks (.377; p < .001). In addition, there was no media coverage without a registered event in the chronicle, which highlights the reliability of the data.

Fake news

We use spatial event data from the hoax map project (https://hoaxmap.org/) to calculate the number of fake news spread in a county-period c_t-1. The website lists discrediting information about refugees that proved to be erroneous. In addition to rumors about refugees, rumors about people with southern complexions or non-German accents are included in the data base because such characterizations are typically associated with refugees. The most frequent accusations concern refugees robbing, assaulting other men and sexually harassing women.

For each piece of fake news, the database contains a location string of the supposed incident, which we assign to its respective county with Google’s geocoding API. The total number of fake news gathered is 416 (132 in East Germany, 31.7%). After transferring the original information to the time-space matrix (considering neighbor counties and the previous time period), there is again huge variation over time and space: 92.1% of units do not record any fake news. The maximum for one unit of analysis was 29 pieces of fake news (Kreis Lippe, West Germany). In sum, we counted 2.759 pieces of fake news in the time-space matrix (neighbor counties included). This refers to 1,837 county-periods with values greater than or equal to one.

For the computation of neighbor variables, we used a county-level geography of Germany provided by the Bundesamt für Kartographie und Geodäsie (territorial state: 01.01.2016). As mentioned briefly above, for each county, we identified all counties that have a common border with the county by means of the R-package spdep. Then, we computed the sum of arson attacks, whether an arson attack was covered by Bild, and the sum of fake news assigned to the county and the neighboring counties for a county-period c_t-1.

Time-variant nation data

Statistical model

Our analysis strategy is inspired by Steele’s application of a two-level binary random intercept model to discrete-time event history data [2]. In contrast to traditional event data analysis, this means that a county in our data is not only analyzed until a xenophobic attack happens, but several attacks might happen over time. At the same time, we acknowledge the multilevel structure of the data, where multiple time periods (level 1) are observed for each county (level 2). Using the xtlogit command in stata, we estimate the log-odds of an arson attack in a given county c at time t as follows: where β₀ is the mean intercept, u_c represents the random effect on the county level and e_ct the idiosyncratic error, which are both assumed to follow a normal distribution. As predictors, the formula includes W_c, a vector of time-constant county-level variables (e.g. unemployment rate 2015), X_t for time-variant nation data (e.g. New Year’s Eve with decay function), and Z_c1(t−1), lagged time-variant county-level variables (e.g. arson attacks in same or neighboring counties for the previous time period, an indicator that measures if the risk of an attack increases if the same county or an adjacent one has experienced an attack in the previous time period).

To facilitate interpretation, we will report average marginal effects (AME) instead of logit coefficients throughout this paper. This means that the coefficients indicate the average changes in the probability of an arson attack if the predictor variable increases by one unit.