Terrorism as group threat

“It is the events seemingly loaded with great collective significance that are the focal points of the public discussion. […] When this public discussion takes the form of a denunciation of the subordinate racial group, signifying that it is unfit and a threat, the discussion becomes particularly potent in shaping the sense of social position” Blumer (1958): 6

Terrorism as group threat

• Terrorist attacks pose a threat to individual and collective safety
• Cause feelings of anxiety and anger
• And ultimately strengthen intolerance and prejudice
• Attacks can shape attitudes toward groups that are often associated with the attacks, immigrants or Muslims

Prior research: Survey data

• Natural experiments find mixed results
• More hostile attitudes after a terror attack
  • E. g.: Hopkins 2010; Legewie 2013, Böhmelt et al. 2019; Finseraas et al. 2011; Nussio, Bove, and Steele 2019
• Little or no effect
  • E. g.: Brouard et al. 2018; Finseraas et al. 2011; Larsen et al. 2019
• Methodological concerns and social desirability bias

Prior research: Social media data

• Tweets sympathetic yet nationalistic after attack (Fischer-Preßler et al. 2019)
• YouTube comments more hostile (Czymara et al. 2023)
• Tweets became more similar to AfD language (Giavazzi et al 2023)

Hypotheses

• H1: The Ethnic hostility increases after terror attacks
• H2: The effect of H1 levels off over time
• H3: Users post more ethnically hostile Tweets after attacks

Overview of attacks

Source: https://www.start.umd.edu/gtd/search/

Search string

• Goal: Identify Tweets that deal with immigration related topics
• No direct search for terrorism or attacks
• But search should include Muslims (much debated migrant group)
• No direct search for insults or derogatory terms
• Search language specific, English version:

 [1] "migration"   "immigration" "migrant"     "immigrant"   "foreigner"  
 [6] "migrants"    "immigrants"  "foreigners"  "refugee"     "refugees"   
[11] "asylum"      "islam"       "muslim"      "muslims"    

Distribution of Tweets before & after an attack

Measuring inter-ethnic hostility

• Relying on of Google Jigsaw’s Conversation AI’s Perspective API
• “Perspective is trained to recognize a variety of attributes (e.g. whether a comment is toxic, threatening, insulting, off-topic, etc.) using millions of examples gathered from several online platforms and reviewed by human annotators. Probability scores represent a probability, with a value between 0 and 1. A higher score indicates a greater likelihood that a reader would perceive the comment as containing the given attribute.” (Google)

Measuring inter-ethnic hostility

• To measure ethnic hostility, I use the API’s IDENTITY_ATTACK attribute, which should measure “Negative or hateful comments targeting someone because of their identity.” (Google)
• In this context, I assume identity proxies ethnicity or race (similar to Czymara et al. 2022)
• Based on convolutional neural networks (Dixon 2018), details of the algorithm are proprietary

Statistical model

• Dependent variable: Probability to be ethnically insulting (ranging from 0 to 100)
• Predictor: Posted either in the week before or after the attack (binary variable)
• Running variable that counts the time before and after the attack (in minutes) to measure trends
• Spline function: \(y=\beta_0+\beta_1attack+\beta_2time\_since\_attack +\) \(\beta_3attack*time\_since\_attack+\beta_4case+\epsilon\)

Results

Putting effect size into perspective

# trend by hour
round(tail(m3_full_trend_attack_only$coefficients, 1)*60, 5)

attack_diff_num:attack_befaft01 
                         -6e-04 

# trend by day
round(tail(m3_full_trend_attack_only$coefficients, 1)*60*24, 3)

attack_diff_num:attack_befaft01 
                         -0.014 

# after how many days are insult levels back to before attack?
round(-(m3_full_trend_attack_only$coefficients[3]/(tail(m3_full_trend_attack_only$coefficients, 1)*60*24)), 3)

attack_befaft01 
          6.845 

Variation across attacks

• Allow both the attack effect and the trends to vary across cases
• Three-way interaction model: \(y=\beta_0+\beta_1attack+\beta_2time\_since\_attack+\beta_3case +\) \([twoway\_interactions]+\beta_4attack*time\_since\_attack*case+\epsilon\)

Variation across attacks

Examining intra-individual change

• Analyzing a subset of accounts that Tweeted at least once before and after an attack
• Benefit: Testing differences within units over time
• Automatically controlling for unobserved heterogeneity
• But somewhat different sample
• Fixed effects model: \(\ddot{y}_{it}=\ddot{\beta}_1attack+\ddot{\beta}_2case+\ddot{\epsilon}_{it}\)

Examining intra-individual change

Robustness checks

• Binary outcomes (75% & 90% threshold) and logistic regression
• Removing highly active accounts (bots?)
• Removing Tweets on Muslims only
• Only before period, cut in half: no effect
• Random treatment timing: no effect

Conclusion

• Jihadist terrorism fuels online hate across cases and countries
• Higher levels of insulting after attacks
• Due to individuals becoming more hostile
• … and due to new people entering the debates
• Limitation: deletion of hate speech (results overly conservative)