Do Terrorists Get the Attention They Want?

Abstract Terrorists aim at influencing audiences beyond their immediate victims, but can only achieve this if an attack receives sufficient public attention. Previous research shows that terrorism can affect public opinion, but these studies are mainly based on emblematic single cases and relate to varying measures of influence, which are difficult to compare. This research focuses on the first-order effect of terrorism: attention. To analyze whether terrorists get attention, we combine a quasi-experimental approach for causal identification with a comparative design. We compile data from Eurobarometer surveys and contrast responses of more than 80,000 individuals surveyed before and after five diverse Islamist attacks in Europe in 2013–2019. Attention to terrorism increases in all targeted countries, regardless of attack size. Yet, while all incidents raise attention to terrorism, only larger attacks exert a meaningful impact across Europe.


Do Terrorists Get the Attention They Want? Comparing Effects of Terrorism across Europe
This Appendix provides supplementary text and additional analyses that complement and further support the main article's argument and findings. These include: • A summary of the attacks included in the main analysis.
• We present further details on statistical analysis and tables comprising the results underlying the main text's figures.
• The SI comprises balance statistics of the control variables (Table S8).
• We have conducted an Oster Selection Test to assess the sensitivity of our results ( Figure S1).
• We assess whether there is a diffusion effect from terrorism abroad to attention to terrorism at home (Table S10).
• As a falsification test, we exchanged the outcome variable by trust in the European Central Bank and, as a placebo test, only assign the treatment to those interviewed two days before an attack (Table S11).
• We include individuals interviewed on the day of an attack (Table S12).
• In addition to fatalities, we also consider wounded people (Table S13).
• An overview of all Eurobarometer surveys coinciding with terrorist attacks until mid-2019 (Table   S14).

Information on Attacks
In the following, we provide information on each of the nine attacks included in our data set.
• London, May 22, 2013: Two assailants hit Lee Rigby, a British Army soldier, with their car and then proceeded to attack the man with knives and a cleaver in London, United Kingdom. Rigby, an off-duty British soldier, was killed in the incident. Michael Adebolajo and Michael Adebowale claimed responsibility in a video shot at the scene and stated that the attack was carried out in retaliation for British soldiers killing Muslims abroad. This attack is covered by Eurobarometer round 79.3.
• Paris/Saint Denis, November 13, 2015: : Bataclan attack comprising eight coordinated Islamic State of Iraq and the Levant (ISIL) attacks in Paris/St. Denis, killing 137 people. Suicide bombers struck outside the Stade de France (Saint-Denis), while several mass shootings and a suicide bombing occurred at coffee places and restaurants. This attack is covered by Eurobarometer round 84.3.
• Manchester, May 22, 2017: A suicide bomber, identified as Salman Abedi, detonated at the Manchester Arena following an Ariana Grande concert in Manchester, United Kingdom. In addition to Abedi, at least 22 people were killed and 119 people were injured in the blast. The Islamic State of Iraq and the Levant (ISIL) claimed responsibility and stated that the attack was carried out in response to "transgressions against the lands of the Muslims." This attack is covered by Eurobarometer round 87.3.
• Carcasonne, March 23, 2018: Redouane Lakdim stopped a car on the outskirts of Carcassonne and shot both of its occupants. Stealing the car, he then attacked four police officers by shooting at them and trying to run them over. In turn, he stormed a supermarket, killing two civilians and a police office, before the terrorist was killed by a police unit. In total, the attack saw five fatalities, including Lakdim, and 15 wounded. This attack is covered by Eurobarometer round 89.1.
• Utrecht, March 18, 2019: Three people were killed in a mass shooting on a tram in Utrecht, Netherlands (seven wounded people). Local police classified the incident as a terrorist attack. The suspect, a 37-yearold Turkish man, was arrested and he admitted carrying out the attack. The perpetrator, Gökmen Tanis, was charged with "murder with terrorist intent." This attack is covered by Eurobarometer round 91.2 Four additional attacks that occurred during the Eurobarometer survey period and were recorded by the GTD were not considered for analysis. We omitted three incidents classified as extreme right-wing terrorism ("National Socialist Underground") in Germany and the 2004 Madrid bombings. The German cases were not associated with terrorism at all, but rather crime, when committed; the Madrid bombings were excluded because the sample size for the analysis of the targeted country was too small to allow for robust inferences: only 107 people were interviewed in Spain before/after the incident, but only 9 of them within three days after the bombing.
Including random intercepts for NUTS units and survey rounds ensures that we carry out the pre and post comparisons within surveys and within units. Table S1 (the sample size is based on a 3 days before/after comparison) and used for most other analyses, we identified a series of relevant variables across several waves of the Eurobarometer that are operationalized in the same way, with the same question wordings. First, we control for respondents' age, which ranges between 15 and 99 in our sample (mean value of 49.82). Second, male respondents might differ from females in their views about terrorism. We created a variable, Female, which is coded 1 if a survey participant was female (0 otherwise). Third, when people indicated that they were not self-employed or employed, we code them as unemployed and include the corresponding binary variable in the analysis as well. Fourth, as urban interviewees could differ from individuals living in rural areas, we control for this with a dichotomous variable receiving a value of 1 if the latter scenario applies (0 otherwise). Finally, the variable Education captures individuals' overall degree of school/university education on a scale from 1 to 10. 0.106 0.000 0.047 0.000 0.029 0.000 * p < 0.10,** p < 0.05,***p < 0.01 Robust standard errors are used and p-values in parentheses (two-tailed). Constant included in all models, but omitted from presentation. All models include random intercepts for units (NUTS level) and survey rounds (years).

Regarding the controls summarized in
Similar to the analyses in the first two tables, we also examined the effect of terrorism on public opinion in the focal country only (Tables S4-S5). That is, we omit all respondents not living in the country of an attack and re-estimate the main models. To provide a fully comprehensive overview of the effects, Tables S6 and S7 concentrate on the effects of an attack outside its borders only, i.e., we omit individuals from the target country.
The main analyses pools respondents from a country under attack and from the wider European audience. As  0.000 0.000 0.000 0.000 0.000 0.000 * p < 0.10,** p < 0.05,***p < 0.01 Robust standard errors are used and p-values in parentheses (two-tailed). Constant included in all models, but omitted from presentation. All models include random intercepts for units (NUTS level) and survey rounds (years). expected, the core finding is robust, but more strongly pronounced when focusing on the target state. Hence, proximity matters and terrorism more closely to a respondent has a bigger effect than more distant attacks (see also Böhmelt, Bove and Nussio, 2020;Nussio, Bove and Steele, 2019). And incidents in one's own state remain to be the most proximate events one could think of. 0.017 0.000 0.000 0.000 0.000 0.000 * p < 0.10,** p < 0.05,***p < 0.01 Robust standard errors are used and p-values in parentheses (two-tailed). Constant included in all models, but omitted from presentation. All models include random intercepts for units (NUTS level) and survey rounds (years).

Balance Statistics
The timing of attacks must be exogenous relative to that of the Eurobarometer interviews in order to meet a key identification requirement of our causal-estimation approach. It is plausible to assume that this is given in our setup. That said, there are other potential sources of bias. Specifically, individuals with specific features  0.000 0.000 0.000 0.000 0.000 0.000 * p < 0.10,** p < 0.05,***p < 0.01 Robust standard errors are used and p-values in parentheses (two-tailed). Constant included in all models, but omitted from presentation. All models include random intercepts for units (NUTS level) and survey rounds (years). 0.009 0.000 0.072 0.000 0.050 0.000 * p < 0.10,** p < 0.05,***p < 0.01 Robust standard errors are used and p-values in parentheses (two-tailed). Constant included in all models, but omitted from presentation. All models include random intercepts for units (NUTS level) and survey rounds (years). 0.000 0.000 0.000 0.000 0.000 0.000 * p < 0.10,** p < 0.05,***p < 0.01 Robust standard errors are used and p-values in parentheses (two-tailed). Constant included in all models, but omitted from presentation.

All models include random intercepts for units (NUTS level) and survey rounds (years).
may be more likely to be interviewed than others. If these characteristics correlate with the outcome variable, too, our results may be biased. In Table S6, we assess whether there are systematic differences in our control variables across individuals interviewed before and after an attack using normalized differences: Imbens and Wooldridge (2009) suggest as a rule of thumb that with a normalized difference exceeding one quarter, linear regression is likely to be sensitive to model specifications. However, as Table S8 shows, this is unlikely to be an issue in our case. When examining the patterns in missing data and non-responses, most of those cases do not pertain to a specific question, but all items of the Eurobarometer. There is also little support for the claim that missing data or non-responses in the outcome variable are more likely among those interviewed after an attack.

Oster Selection Test
In light of Balcells and Torrats-Espinosa (2018), we also assess the sensitivity of our results using one of Oster's selection tests (Oster, 2019). While we are confident in the validity of our research design's assumptions, a last test in this context examines the level of unobserved confounding that would make the results become invalid. Specifically, given a specific R 2 value (in essence, the one reported for our main, unconditional model that includes the control variables and is based on before/after comparisons of 3 days), we vary the degree of unobserved selection into treatment, which is defined as the ratio of selection on unobservables to selection on observables and calculate how the treatment effect would change. In substantive terms, a ratio of selection on unobservables to selection on observables of 1 implies that the unobservables are as important as the observables.
However, as Figure S1 highlights, when keeping the R 2 value constant, altering the ratio does not fundamentally change the size of the treatment effect.

Logistic Regression Models
The hierarchical models we employ are linear models, rendering all estimations essentially linear-probability models that control for the hierarchical nature of the data. A key advantage of this setup is that coefficients can be directly interpreted. However, a potential shortcoming is that we induce heteroskedasticity when employing a linear model on a binary dependent variable. To address any concerns in this regard, we re-estimated the main (unconditional) models with regular logistic regression models that comprise fixed effects for survey rounds and NUTS units. In this setup, only the signs of the coefficients and their significance levels can be interpreted directly. However, as Table S9 emphasizes, our findings are robust to the change in the estimator.

Diffusion Effect
The main finding in Böhmelt, Bove and Nussio (2020) suggests that terrorism abroad can shape migration attitudes at home. That is, not only attacks in the focal country shape public opinion, but also incidents in other countries. By implication, one may also expect a relatively strong impact, e.g., of terrorist attacks in France on public opinion in Germany. Similar patterns seem plausible for the UK and Ireland, or the Netherlands and Belgium. To explore this possibility in more detail, Tables S6 and S7 above examine a cross-European only effect, but we created here a variable that captures whether the country of a terrorist attack and the state where an individual has been interviewed are neighbors as defined in Stinnett et al. (2002). To code country neighbors, we focus on direct land borders and separation of up to 24 miles of water. We interact this contiguity variable with the treatment and re-estimate our core model with the control covariates.
According to Table S10 in this appendix, the main result is also robust to considering a diffusion effect of terrorism on public opinion, but we do indeed find evidence for the latter as well (also similar to Tables S6 and   S7). On one hand, the treatment's coefficient now pertains to those cases where Contiguity is set to 0, i.e., we omit neighboring countries here. The coefficient estimate remains statistically significant and positively signed in the first column, providing further evidence for a causal impact of terrorism at home on public opinion. On the other hand, the interactive term's coefficient captures those cases where public opinion is influenced by terrorism in the neighborhood -and the estimate is positive and significant as well at least in Model 39. This finding may add to the "European sphere" literature (e.g., Trenz, 2005).

Falsification and Placebo Tests
To further increase the confidence in our findings, we conducted one falsification test and, secondly, assigned a placebo treatment at a different date than the actual terrorist attacks (see also Muñoz, Falcó-Gimeno and Hernández, 2020). On one hand, we replaced the outcome variable by an item, which should substantively not Interviewees were asked what institutions they generally trust, and the ECB is among the possible answers. We used this information to create a binary variable receiving a value of 1 if individuals answered that they trust in the ECB institution and 0 otherwise. In turn, we use this item as the new dependent variable and re-run the main, unconditional model based on before/after comparisons of 3 days). On the other hand, we altered the treatment by assigning it only to people interviewed two days before an attack took place. To this end, we follow Muñoz, Falcó-Gimeno and Hernández (2020, p.12) as we "address the possibility that preexisting time trends, unrelated to the event of interest, could bias the findings," and thus "test for the existence of such a trend before the event took place. " Muñoz, Falcó-Gimeno and Hernández (2020, p.12) suggest "placebo treatments constructed at arbitrary points at the left of the cutoff-point," and two days before an attack meets this criterion. It seems implausible that those interviewees would know about an incident to come and, hence, we should not detect a significant treatment effect either. Table S11 summarizes the corresponding results.

Inclusion of People Interviewed on the Day of an Attack
In all previous models and the ones presented in the main text, we exclude individuals who were interviewed on the day of an attack. On one hand, we cannot fully ensure that an interview then took place indeed after an incident and not before. On the other hand, even if an interview, say, was conducted about two hours after a terrorism attack, it may not have made it to the news and is covered by the media so that individuals can be affected by this. However, as a robustness check, we do include all people interviewed on the day of an attack and, as demonstrated, this does not affect the substance of our main result (Table S12). 0.000 0.000 * p < 0.10,** p < 0.05,***p < 0.01 Robust standard errors are used and p-values in parentheses (two-tailed). Constant included in all models, but omitted from presentation. Models include unit and survey fixed effects.

Inclusion of Wounded People
While we have considered the influence of fatalities, our model estimations did not include the number of wounded people. However, a large number of wounded may also affect media attention and, eventually, public opinion. Against this background, we created the variable Casualties (ln), which is the logarithm of the number of dead and wounded casualties of an attach. Table S13 summarizes our findings for models with controls for before/after comparisons of 1, 3, and 5 days for the full sample covering all countries. When additionally considering wounded victims, however, the results are virtually identical to those merely focusing on fatalities only.

All Eurobarometer Surveys Coinciding with Terrorism
Although we focus on Islamic terrorism, we also identified a series of other incidents that coincide with Eurobarometer surveys. Table S14 summarizes the necessary information. The information on attacks was retrieved from the Global Terrorism Database. 1 Most of these surveys were not used for our analysis as they did not include the relevant survey questions (e.g., "what do you think are the two most important issues facing (OUR COUNTRY) at the moment?").