The desensitization hypothesis

The concept of desensitization originates in research on repeated exposure to violent media, such as violent films or video games (e.g., Brockmyer Reference Brockmyer2015; Carnagey, Anderson, and Bushman Reference Carnagey, Anderson and Bushman2007; Funk et al. Reference Funk, Buchman, Jenks and Bechtoldt2003). According to that literature, desensitization refers to a psychological process involving two interrelated mechanisms: (1) a dampening of affective responses after repeated exposure to violence in media (e.g., reduced fear, sadness, or anger) and (2) cognitive normalization, whereby violent events are increasingly seen as routine, unremarkable, or to be expected. Consequently, repeated exposure to violence is argued to, among other things, diminish moral concern and increase tolerance for aggression (Bushman and Anderson Reference Bushman and Anderson2009; Funk et al. Reference Funk, Baldacci, Pasold and Baumgardner2004). In a broader sense, the concept of desensitization also resonates with habituation theory from neuroscience and behavioral psychology, which suggests that growing accustomization to a situation or stimulus can reduce responsiveness (Lader and Mathews Reference Lader and Mathews1968; Thompson and Spencer Reference Thompson and Spencer1966; Watts Reference Watts1979). Finally, there are parallels with the notion of compassion fatigue in healthcare, which posits that prolonged exposure to human suffering can trigger emotional disengagement (Joinson Reference Joinson1992; Sinclair et al. Reference Sinclair, Shelley Raffin-Bouchal, Mijovic-Kondejewski and Smith-MacDonald2017).

Previous literature suggests that desensitization may not be limited to video media, but extends to the case of Islamist terrorism (Amarasingam and Clarke Reference Amarasingam and Clarke2017; Castanho Silva Reference Castanho Silva2018; Markoulis and Katsikides Reference Markoulis and Katsikides2020; Nussio Reference Nussio2020). Specifically, the argument is that since Western citizens have been exposed to a growing number of attacks, they are likely to increasingly see Islamist terrorism as part of everyday life. As a result of increasing normalization, emotional responses become more muted, and citizens become less likely to update their perceptions of threat, mortality salience, and injustice.Footnote ¹ Furthermore, increasing normalization may lead to reduced and less emotionally charged media coverage, thus further reducing the salience of Islamist attacks. As Nussio (Reference Nussio2020, p. 1152) explains, this process eventually culminates in a situation where Islamist terrorism no longer produces detectable changes in political attitudes.

Limited empirical evidence and theoretical alternatives

However, despite the concept’s increasing popularity, it remains unclear whether desensitization extends to the case of Islamist terrorism in Western democracies. Most of the existing evidence for desensitization comes from media psychology, where exposure is very frequent and typically involves fictional depictions of nonpolitical violence (e.g., daily play of first-person shooter games). These studies rarely examine political attitudes, focusing instead on outcomes such as empathy for victims or tolerance for aggression (e.g., Brockmyer Reference Brockmyer2015; Bushman and Anderson Reference Bushman and Anderson2009; Ferguson Reference Ferguson2015; Funk et al. Reference Funk, Baldacci, Pasold and Baumgardner2004; Prescott, Sargent, and Hull Reference Prescott, Sargent and Hull2018). Hence, the conditions under which desensitization has been predominantly demonstrated differ markedly from those that characterize terrorism.

A key open question is whether Islamist terrorism is frequent enough in Western democracies to trigger desensitization. Islamist terrorism remains, at best, a semi-regular experience for most Western citizens. Furthermore, habituation theory suggests that emotional responses are only likely to decrease if stimuli remain similar in intensity and form (Lader and Mathews Reference Lader and Mathews1968; Thompson and Spencer Reference Thompson and Spencer1966; Watts Reference Watts1979). Yet, Islamist terrorist attacks tend to differ in terms of their locations, methods, targets, and lethality (etc.). In agreement with this, research suggests that most citizens of Western countries continue to see Islamist terrorist attacks as particularly transgressive events (Bove et al. Reference Bove, Efthyvoulou and Pickard2024; Nussio, Böhmelt, and Bove Reference Nussio, Böhmelt and Bove2021; Völker Reference Völker2024). In addition, Islamist terrorism in the West remains typically associated with perpetrators from ethno-religious outgroups, which has been shown to amplify perceived moral violations and blame (Bilali, Tropp, and Dasgupta Reference Bilali, Tropp and Dasgupta2012; Noor et al. Reference Noor, Kteily, Siem and Mazziotta2019). As a result, Islamist terrorism may not become normalized in the same way as depictions of high-frequency violence in fictional media. Despite repeated exposure, Islamist attacks may continue to activate the same cognitive and emotional mechanisms and therefore continue to have similar effects on political attitudes.

A third theoretical possibility is response escalation. Recurrent terrorist attacks may be interpreted as evidence of institutional failure and a lack of state capacity (Montalvo Reference Montalvo2011), which could successively increase threat perceptions and mortality salience. Similarly, repeated norm violations could reinforce moral outrage rather than diminish it and therefore increasingly harden attitudes toward perpetrators and associated outgroups. Accordingly, repeated exposure may not diminish but intensify attitudinal responses to Islamist terrorism.

Ultimately, whether public responses to Islamist terrorism have weakened, remained stable, or intensified is an empirical question. Yet the evidence for change in public responsiveness to Islamist terrorism remains limited. Few existing studies explicitly test for changes in responses to Islamist terrorism over time; instead, desensitization is often invoked as a post-hoc explanation for null findings in recent studies. Furthermore, many claims of desensitization are based on single-case studies, which cannot establish broader temporal trends. Meanwhile, the best available evidence stems from a narrative review, which is able to provide a broader overview (Nussio Reference Nussio2020); yet, narrative reviews are susceptible to incomplete coverage, selective emphasis, and heterogeneity between studies.

Reliably establishing trends in reported effect sizes is far from straightforward. In any larger body of literature, results are bound to be mixed. The literature on the effects of Islamist terrorism on public opinion is no exception. As Nussio has observed, many studies of older Islamist attacks report sizable effects on public opinion (e.g., Hobfoll, Canetti-Nisim, and Johnson Reference Hobfoll, Canetti-Nisim and Johnson2006; Lerner et al. Reference Lerner, Gonzalez, Small and Fischhoff2003). However, there are also notable exceptions reporting much smaller or even null effects (e.g.,  Dinesen and Jager Reference Dinesen and Meier Jæger2013; Legewie Reference Legewie2013; Vlandas and Halikiopoulou Reference Vlandas and Halikiopoulou2025). Studies of more recent Islamist attacks similarly report a variety of different results (e.g., Castanho Silva Reference Castanho Silva2018; Muñoz, Falcó-Gimeno, and Hernández Reference Muñoz, Falcó-Gimeno and Hernández2020; Larsen, Cutts, and Goodwin Reference Larsen, Cutts and Goodwin2020; Schmidt-Catran and Czymara Reference Schmidt-Catran and Czymara2020; Vlandas and Halikiopoulou Reference Vlandas and Halikiopoulou2025). The high variability in reported effects makes it difficult to reliably aggregate results by hand. Furthermore, it suggests that effect sizes are shaped by a variety of factors, potentially including repeated exposure but also broader context factors (e.g., country of study, attack severity, and the target) and a study’s research design (e.g., the precise outcome variable, the sample, or the quality of causal inferences).Footnote ² It follows that apparent trends in reported effect sizes may not reflect the consequences of repeated exposure, but other factors. Overall, the significant heterogeneity in terms of reported findings, the methods used, and the contexts studied renders narrative reviews an unreliable tool for the establishment of trends in reported effects. Testing whether public responses to Islamist terrorism have changed requires a more systematic approach.

Data

Study 1 builds on a previously published meta-analytical dataset that compiles estimates of the effects of terrorism on political attitudes from quantitative studies (Godefroidt Reference Godefroidt2023). The dataset adopts a broad definition of political attitudes, encompassing (a) orientations toward security policy, the protection of civil liberties, right-wing authoritarianism, and related ideological positions; (b) attitudes toward outgroups, including prejudice, stereotyping, and support for (anti-)immigration policies; and (c) rally-around-the-flag responses, such as national identification, political trust, and political participation. To be included in the dataset, studies had to provide individual-level effect estimates; be written in English, French, or Dutch; and report sufficient information to compute standardized effect sizes.

Godefroidt (Reference Godefroidt2023) used this dataset to examine general patterns across all forms of terrorism worldwide. Our study narrows this scope in two ways. First, we restrict the sample to studies of Islamist terrorism, excluding those on unspecified or other forms of terrorism (e.g.,  far-right or far-left terrorism). Second, we focus exclusively on Western democracies, including countries in Europe (excluding Belarus and Russia), North America, Australia, and New Zealand.Footnote ³

Analyzing change in effect sizes requires good coverage of recent attacks. Hence, we extended the temporal coverage of Godefroidt’s original dataset. Whereas the original data included studies available up to December 2019, we added all relevant studies formally published or otherwise made available (e.g., pre-prints) between January 2020 and March 2023. To ensure consistency and comparability, we applied the same inclusion and exclusion criteria and used the same four-part search strategy for the identification of relevant manuscripts:

1. (1) Searches in Web of Science, ProQuest, and EBSCO with the same search terms;

2. (2) Screening of recent narrative reviews;

3. (3) Public calls for unpublished work via social media and mailing lists;

4. (4) Backward and forward citation tracking.

SI Appendix § 1.1 provides further details on the data collection process. With the temporal update included, our meta-analytical dataset comprises 837 estimates of the effects of Islamist terrorism on political attitudes. These estimates are drawn from 130 manuscripts (15 added as part of our update), which include 170 unique studies conducted between 2001 and 2019. Importantly, effect estimates are based on diverse methodologies: 165 (20%) stem from randomized experiments (e.g., vignette or priming experiments), 351 (42%) from natural experiments (e.g., pre/post comparisons of cross-sectional or panel data), and 321 (38%) from correlational designs (e.g.,  associations between threat perceptions and political attitudes). The most frequently studied outcome category is attitudes toward outgroups (42% of estimates), followed by preferences for tougher security policies and other conservative shifts (36%), and finally, rally-around-the-flag responses (22%). The 837 effect estimates span a total of 24 Western countries, including Australia, Austria, Belgium, Canada, Czechia, Denmark, Estonia, Finland, France, Germany, Ireland, Italy, Luxembourg, the Netherlands, Norway, Poland, Romania, Slovakia, Slovenia, Spain, Sweden, Switzerland, the UK, and the USA.

The original manuscripts report effect estimates in various ways, including bivariate correlations, multiple regression coefficients, and mean differences. To facilitate cross-study comparisons, we converted all estimates into Cohen’s d coefficients and calculated corresponding sampling variances. Cohen’s d is a standardized effect size measure that reflects the difference between two group means as a proportion of the outcome’s standard deviation. Where necessary, we reversed the direction of the coefficients so that positive values consistently indicate a shift to the political right (e.g., greater outgroup hostility) and stronger rally-around-the-flag tendencies (e.g.,  increased national identification).

Results

We examine changes in attitudinal responses to Islamist terrorism in three ways. First, we test for an overall time trend by regressing reported effect sizes on the study year. Second, we assess whether reported effects correlate with the number of previous attacks in the studied country and across all Western countries. Third, we compare estimates of the effects of specific Islamist attacks. We also conduct several robustness checks, including tests for publication bias and sensitivity analyses. As a preliminary first step, however, we establish the baseline effect of Islamist terrorism on political attitudes across all previous studies.

We use a three-level random-intercept model to estimate the average effect of Islamist terrorism across studies, which accounts for both the sampling variance of reported effects and the hierarchical structure of the data (i.e., statistical dependencies within and between manuscripts) (Cheung Reference Cheung2015; Noortgate et al. Reference Noortgate, López-López, Marín-Martínez and Sánchez-Meca2013). The full model specification is provided in SI Appendix § 1.2. We find that Islamist terrorism is generally associated with a small-to-moderate rightward shift in political attitudes in the existing literature ( $d = 0.22$ , SE $ = 0.03$ , $p \lt .001$ , 95% CI $\left[ {0.16,{\rm{\;}}0.27} \right]$ ). Note that this represents the average effect across studies and not the average effect of an Islamist attack, which is smaller since selected high-profile attacks associated with comparatively large effect sizes have been studied more frequently (see below).Footnote ⁴

Over-time variation

If repeated exposure to Islamist terrorism leads to smaller (larger) effects on political attitudes, then we should find that more recent studies tend to report smaller (larger) effects. We test for the existence of a time trend in reported effects by adding the study year to our baseline specification (see above). For natural experiments, the study year generally corresponds to the year of the attack.Footnote ⁵ For correlational studies and randomized experiments, we use the year in which the survey or experiment was fielded. For better interpretability, we recoded the study year so that the earliest year in our dataset (2001) is coded as 0. Since the most recent study year is 2019, the variable ranges from 0 to 18. We were unable to retrieve the study year for 113 effect sizes (13.5% of the data). This leaves us with 109 manuscripts and 724 effect estimates for this analysis.

Model 1 in Table 1 shows the raw time trend. We find that the reported effects of Islamist terrorism on political attitudes decreased by an average of around 0.01 Cohen’s $d$ every year since 2001 ( $p = 0.02$ ). This initial result is consistent with the desensitization hypothesis. However, the raw time trend does not account for heterogeneity across manuscripts in the attacks they study and their study designs. For example, bias could emerge since more recent attacks tend to have fewer casualties, and attack severity may independently shape public reactions. Furthermore, methodological improvements or other changes could have led to smaller reported effects over time.

Table 1.

Annual change in reported effects of Islamist terrorism on political attitudes in Western democracies

Note: * $p \lt $ 0.05; ** $p\lt$ 0.01; *** $p \lt $ 0.001.

To mitigate omitted-variable bias, we add controls for attack severity (fatalities) and several aspects of study design: design type (correlational, natural experiment, randomized experiment), sample quality (probability vs. non-probability), and outcome domain (conservative shifts, outgroup attitudes, rally tendencies). We also include a U.S. dummy to capture systematic contextual differences, as prior work suggests that terrorism studies conducted in the U.S. yield larger effect sizes (Godefroidt Reference Godefroidt2023).

Model 2 shows the results: with relevant controls included, the time trend is close to zero ( $b = - 0.003$ ) and clearly not statistically significant ( $p = 0.44$ ). This suggests that while reported effects have decreased over time, this is likely due to factors other than desensitization. In particular, our results show that studies with a U.S. focus tend to yield significantly larger effects ( $b = 0.14,p = 0.007$ ), and earlier studies are disproportionately more often focused on the U.S. (63% of studies conducted before 2005 were conducted in the U.S. compared to only 22% thereafter). Put another way, the raw negative time trend is statistically explained by the overrepresentation of U.S. studies in earlier years, many of which were focused on 9/11 and yielded comparatively large effects.Footnote ⁶ We also find some evidence that attacks with higher numbers of fatalities, as well as less well-identified correlational studies, tend to yield slightly larger effects. However, these effects are not statistically significant at the conventional 5% level.

Correlation with the number of previous attacks

The results thus far suggest that, ceteris paribus, the effects of Islamist terrorism on political attitudes have not changed over time. However, time is only an indirect proxy for repeated exposure to Islamist terrorism. In a second step, we therefore consider a more direct measure of exposure: the cumulative number of Islamist terrorist attacks prior to the study year. We draw data on the number of Islamist attacks from Fondapol (2019). Because it is unclear whether desensitization (or escalation) is primarily a function of recent or all prior attacks, we test both the number of attacks in the preceding five years and the total number of prior attacks. Analogously, we consider both the number of prior attacks in the same country and across all Western democracies. All models contain the full set of controls, including attack severity, study design, sample quality, outcome domain, and country context. Table 2 reports the results. We find little evidence for changes in reported effect sizes due to repeated exposure, irrespective of our measurement approach.

Table 2.

Association between past Islamist terrorist attacks and reported effects of Islamist terrorism on political attitudes in Western democracies

Note: All models include the control variables as specified in Table 1. * $p\lt $ 0.05; ** $p\lt $ 0.01; *** $p\lt $ 0.001.

Specific attacks

Finally, we examine whether there is an observable trend in the effect estimates for specific Islamist attacks. For this analysis, we restrict the sample to estimates tied to concrete Islamist terrorist events, dropping estimates from correlational and experimental studies that focus on generic Islamist terrorism without reference to a particular attack. This yields a sample of 580 effect estimates from 91 manuscripts, covering 15 Islamist attacks perpetrated between 2001 and 2018. Analogously to our analysis of the overall effect across studies, we estimate the average effect of each attack using three-level random-intercept models that account for dependencies within and between manuscripts. For reference, we also establish the average effect of an Islamist terrorist attack using a four-level meta-analytical model that, in addition, accounts for clustering of effect sizes within attacks. This takes into account that each attack has specific features, such as context, scale, or method, which may make it more or less likely to affect attitudes.

We find that the average effect of an Islamist attack is $d = 0.16$ (SE $ = 0.04$ , $p \lt .001$ , 95% CI $\left[ {0.09,{\rm{\;}}0.23} \right]$ ). However, we observe substantial variability in effect sizes (see Figure 1). Some attacks, such as 9/11, the 2007 Fort Dix plot, or the November 2015 Paris attacks, are estimated to have produced comparatively large effects, with Cohen’s $d$ values exceeding 0.20 and in some cases even 0.30. Other attacks, such as the 2004 van Gogh murder, the 2017 Manchester Arena bombing, or the 2015 Copenhagen shootings, are associated with much smaller effects ( $d \approx 0.05$ – $0.10$ ). Finally, in nearly half of the cases, the confidence intervals include zero (e.g., the 2005 London bombings, 2013 Boston Bombing, or 2016 Berlin truck attack). Most importantly, though, we do not detect any systematic trend of decline (or increase) in public responsiveness over time. Instead, the observed heterogeneity appears to vary with attack characteristics and research design features.

Figure 1.

Average reported effects of different Islamist terrorist attacks.

Note: ${N_i}$ = number of effect estimates; ${N_j}$ = number of manuscripts.

Case selection

Study 2 analyzes the effects of the 2017 and 2019 London Bridge attacks. The 2017 London Bridge attack involved a group of terrorists deliberately ramming a van into pedestrians on the London Bridge and subsequently stabbing people in a nearby food market. A total of 11 people died, including the three perpetrators, and 48 more were wounded. The 2019 London Bridge attack involved a single perpetrator stabbing several people in the same area. Three people died, including the perpetrator, and three more were injured.

We argue that the two London Bridge attacks represent a strong testing ground for the desensitization hypothesis, for three main reasons. First, while no two terrorist attacks are ever identical, the two London Bridge attacks share several important characteristics that could plausibly shape public responses. Both attacks occurred in practically the same location in the heart of London; in both, the perpetrators wore fake suicide vests and were ultimately shot dead by police on the scene; both were claimed by the same terrorist organization, the Islamic State; both occurred during election campaigns; and, although the 2017 attack led to a higher number of casualties, both were comparatively small in scale. In the absence of desensitization (and escalation), we would therefore expect that the two London Bridge attacks generated broadly similar responses.

Second, the two London Bridge attacks were by no means the first Islamist attacks in the UK (see Table 3). In the two years before the 2017 London Bridge attack alone, there were four Islamist terrorist attacks, including one (i.e., the Manchester Arena bombing) only two weeks prior and another (i.e., the Westminster attack) just three months prior.Footnote ⁷ There were two more attacks between the 2017 and 2019 London Bridge attacks (i.e., the 2017 Parsons Green Train bombing and the 2018 Manchester Victoria Station stabbing). If there was desensitization, it should therefore be visible in the London Bridge attacks. Specifically, we should observe comparatively small effects after both attacks, especially after the 2019 attack. Analogously, if there was response escalation, we would expect comparatively large effects, especially after the 2019 attack.

Table 3.

Islamist terrorist attacks in the United Kingdom, 2005–2019

Source: Fondapol (2019).

Finally, since both London Bridge attacks occurred during election campaigns, we are able to estimate their effects using a constant data source including similar or even identical outcome measures (see below). This increases our ability to attribute changes in effect sizes to desensitization or escalation.

Research design

We follow several recent studies and leverage the exogenous timing of the London Bridge attacks during the collection of cross-sectional survey data (e.g., Balcells and Torrats-Espinosa Reference Balcells and Torrats-Espinosa2018; Bove, Efthyvoulou, and Pickard Reference Bove, Efthyvoulou and Pickard2024; Breton and Eady Reference Breton and Eady2022; Castanho Silva Reference Castanho Silva2018; Giani Reference Giani2021; Muñoz, Falcó-Gimeno, and Hernández Reference Muñoz, Falcó-Gimeno and Hernández2020; Nussio Reference Nussio2020). However, unlike previous studies, we do not rely on commercial or academic surveys, but on data from two Voting Advice Applications (VAAs). VAAs are online information tools that are made available to voters before elections (Germann and Gemenis Reference Germann and Gemenis2019). VAA data has a key advantage for our purposes since VAAs tend to be used by large numbers of people over a short period of time. By contrast, the daily number of interviews in more commonly used surveys in this field of study tends to be more limited. Therefore, previous studies often had to rely on relatively long temporal bandwidths spanning weeks or even months, which increase the risk of bias due to other, unrelated events (for exceptions, see Balcells and Torrats-Espinosa Reference Balcells and Torrats-Espinosa2018; Breton and Eady Reference Breton and Eady2022; Giani Reference Giani2021). Our reliance on VAA data allows us to significantly reduce the temporal bandwidth while maintaining high statistical power. Another notable advantage of VAA data is that VAA users have a clear incentive to reveal their true policy preferences, which reduces the risk of social desirability bias. Of course, VAA data is no panacea. Pre/post comparisons hinge on several strong identification assumptions (Muñoz, Falcó-Gimeno, and Hernández Reference Muñoz, Falcó-Gimeno and Hernández2020), and additional concerns arise due to the self-selected nature of VAA data. We perform a large number of tests to assess the plausibility of our identifying assumptions, the comparability of effect estimates across years, and their generalizability to the population of British voters.

Results

We proceed to discuss the effects of the 2017 and 2019 London Bridge attacks on political attitudes. We report all effects as Cohen’s $d$ values (i.e., effects divided by the standard deviation of the outcome variable) to improve comparability between outcome variables and with Study 1. Higher values indicate stronger support for restrictive security and immigration policies as well as heightened in-group identification. The spikes represent 95% confidence intervals.

Figure 2 shows the results. We find that even though the UK was subject to several Islamist terrorist attacks in the years, months, and even weeks prior, the 2017 London Bridge attack led to changes in political attitudes that are broadly comparable to previous attacks in the UK and elsewhere.Footnote ¹⁰ The largest effect emerges in the case of support for tough security policies ( $d = 0.15-0.20$ ). We also observe substantial increases in anti-immigrant sentiment ( $d = 0.08-0.11$ ) and a rally effect, although, interestingly, we find a stronger effect on English identification ( $d = 0.07-0.10$ ) than on British identification ( $d = 0.02-0.04$ ). A possible reason is that British identity tends to be seen as more ethno-religiously inclusive, and Islamist terrorism may be especially likely to activate more narrow, White, and Christian identities.

Figure 2.

Short-term effects of the 2017 and 2019 London Bridge attacks on political attitudes.

Islamist terrorism continued in the months after the 2017 London Bridge attack. Despite this, the 2019 London Bridge attack had broadly comparable effects on political attitudes. This applies, in particular, to the effects on English ( $d = 0.07-0.09$ ) and British identification ( $d = 0.03-0.07$ ), which had the exact same wordings in both years. Meanwhile, the effects on security ( $d = 0.12-0.13$ ) and immigration policy ( $d = 0.06-0.07$ ) are somewhat smaller in 2019, but the differences are moderately sized and most do not reach statistical significance (see Table 4). Notably, the wordings of the policy statements related to security and immigration policy differed somewhat between the years (see SI Appendix § 2.5).

Table 4.

Formal comparisons of the short-term effects of the 2017 and 2019 London Bridge attacks

Note: The table provides the standardized impacts of the 2017 and 2019 London Bridge attacks (Cohen’s $d$ ), the differences between them, and the 95% confidence intervals of the differences. We estimate the statistical significance of differences between 2017 and 2019 using seemingly unrelated estimation. CI = confidence interval; LB = lower bound; UB = upper bound. * $p \lt$ 0.05; ** $p\lt$ 0.01; *** $p\lt$ 0.001.

Overall, the results of Study 2 point to response stability. Despite many prior attacks, several of which were in the recent past, both London Bridge attacks had comparable effects on political attitudes. The effect sizes we report are also comparable to those found for other attacks in previous studies, including two attacks in the UK: the 2005 London bombings ( $d = 0.13$ ) and the 2017 Manchester Arena bombing ( $d = 0.07$ ).

Additional robustness checks and analyses

We report several additional robustness checks in SI Appendix § 2.8. First, we provide the full-sample results including all observations that were removed as part of the data cleaning. Second, we report the results when we match exactly on all covariates instead of using linear regression. Exact matching fully removes any measured imbalances between treatment and control groups and thus increases confidence in the ignorability assumption. Third, we report naive pre/post comparisons that do not include any control variables. The results are always similar.

In SI Appendix § 2.9, we consider the effects of the terrorist attacks on each of the individual policy questions making up our summated rating scales measuring security and immigration attitudes. We find that the effects are generally similar across items, with the most notable difference being that we find particularly large increases in people’s opposition to accepting refugees from war-torn countries.

Finally, in SI Appendix § 2.10 we investigate the durability of the effects of the London Bridge attacks beyond the three-day time window. We find that the effects remained similar in size over a period of up to two weeks after the attacks. Notably, the effects also remained visible on election day in both 2017 and 2019. Unfortunately, our data does not allow us to establish effects beyond the two-week mark (VAAs tend to stop data collection after election day). We note that effects become less well-identified the longer the time frame.