The Impact of Terrorism on Social Capital: Evidence from the 2015 Charlie Hebdo Paris Shooting

We examine the impact of terrorism on social capital by exploiting variation in the 2014 European Social Survey administration dates coupled with the 2015 Charlie Hebdo attack in Paris, France. Using the difference-in-differences estimator, we find that the attack had a positive, causal impact on the overall level of social capital among French respondents. Further, the effect seems to be driven by an increase in institutional and interpersonal trust, as well as by engagement in social networks. This rise in social capital peaks in the immediate aftermath of the terrorist attack but subsequently decays to pre-attack levels within approximately one month.


Introduction
Social scientists frequently demonstrate the importance of social capital -or the "features of social organizations, such as trust, norms, and [social] networks" -in promoting efficient political institutions as well as a country's overall level of economic success (Putnam et al. 1993 p. 167). 1 Of late, and motivated by a sharp, ninefold increase in the number of terrorist-related events over the last two decades, new attempts have been made to quantify the role of terrorism in explaining a country's level of social capital (Chanley 2002;Putnam 2002;Collins 2004;Gross et al. 2004;Paxton 2005;Huddy and Feldman 2006;Bali 2007;Gassebner et al. 2008;Sander and Putnam 2010;Sinclair and LoCiero 2010;Blomberg et al. 2011;Arvanitidis et al. 2016).
Despite attracting the attention of social scientists at large, two notable limitations in the literature persist. First, existing studies tend to focus on narrowly construed definitions of social capital, namely trust in institutions and interpersonal trust. Second, the extant analyses utilize empirical methods that are insufficient for quantifying the causal effect of terrorism on a nation's level of social capital. Motivated by these shortcomings, we utilize the 2015 Charlie Hebdo shooting in France, coupled with the random and staggered nature that respondents were interviewed in the 2014 wave of the European Social Survey (ESS), to estimate the causal effect of terrorism on multiple dimensions of social capital using the difference-in-differences (DID) estimator.
As noted above, the number of terrorist attacks has increased dramatically over the last two decades. According to the Heritage Foundation (2016), there were 141,966 terrorist incidents between 1970 and 2014 worldwide resulting in 32,685 deaths in 2014 alone. 2 The geographic distribution of terrorist activity has also evolved. Before 2014, the vast majority of attacks were concentrated in developing countries. However, in recent years terrorism has become more prevalent in developed countries (Kim and Kim, 2017). This trend is especially true in Europe. According to a report by the Institute for Economics and Peace (2016), deaths from terrorism increased in OECD member countries to 577 in 2015 from 77 in 2014, with the majority of this increase occurring in European countries. Underscoring this finding, a recent article published in the New York Post (August 18, 2017) on the state of terrorism in Europe notes that, although terrorism is inherently abnormal, unfortunately "it is the new norm [for Europe],"and that "the situation is worsening, not improving." 3 As well, The Heritage Foundation's Robin Simcox states that, "Islamist terrorism poses a grave danger to Europe. Its face and its methods are always evolving, but it is not going anywhere" (Heritage Expert: Terror in Europe is Growing, Spreading, Changing, June 23, 2017).
Evidence to Simcox's claims are a series of major terrorist attacks in the West by Islamic extremist groups. One noteworthy attack, which is the focal point of this paper, was the Charlie Hebdo shooting in Paris, France. On January 7, 2015, two armed individuals shot and killed 11 people and injured 11 others at the office of the French weekly news magazine Charlie Hebdo. It was later found that the religious extremist group Al-Qaeda helped prepare the attack because the newspaper had published cartoons offensive to the central figure of Islam, the Prophet Muhammad. The attack was followed by intense global media coverage, social media exposure epitomized by YouTube footage of the Charlie Hebdo attackers shooting a French police officer in the streets of Paris, and massive public rallies following the attack. It is widely known that terrorist attacks have substantial economic costs. These costs include damages to physical and human capital, attenuated foreign direct investment (FDI) and tourism, misallocation of resources due to increased uncertainty, increased transaction costs due to heightened security measures, and increased adverse health effects among individuals who have directly or indirectly witnessed the terror attacks (Frey et al. 2007;Smith et al. 2009;Metcalfe et al. 2011;Kim and Kim 2017). Studies that investigate the macroeconomic impacts of terrorism find that the costs in terms of GDP, consumption, and economic growth are considerable, especially in countries with repeated and chronic exposure to terrorism (Abadie and Gardeazabal 2003;Blomberg et al. 2004;Eckstein and Tsiddon 2004).
In addition to these macroeconomic and health implications, it has been argued that terrorist attacks may have profound impacts on a country's level of social capital. Theoretically, however, the impact of terrorism on social capital is ambiguous. On the one hand, some have argued that terrorist incidents may strengthen bonds between citizens, increasing both trust among individuals and confidence in institutions (Chanley 2002;Putnam 2002;Collins 2004). This response can be interpreted as a "rally effect" according to which a large-scale terrorist attack leads to a collective loss of sense of security and, in the face of a national public threat, results in citizens placing more trust in political institutions and displaying a more unified sense of identity (Perrin and Smolek 2009;Woods 2011). On the other hand, others have argued that terrorism may foster feelings of insecurity, raising distrust in people of different ethnicities, cultures or religions, and/or in people and the government in general (Blomberg et al. 2011;Brehm and Rahn 1997;Huddy et al. 2002;Uslaner et al. 2004;Gould and Klor 2016). According to Tallman (2007), this potential negative effect of terrorism on social capital can be viewed as an "anti-rally effect" or a "jading effect." This ambiguity with respect to social capital stands in contrast to the impact of terrorism on physical and human capital, which is necessarily negative.
To the best of our knowledge, we provide the first quasi-experimental evidence indicating a causal impact of a terrorist attack on a nation's level of social capital. Specifically, we find that the 2015 Charlie Hebdo terrorist attack led to a significant increase in social capital immediately following the incident thus inducing a "rally effect" amongst the impacted populace. However, this rise in social capital peaks in the immediate aftermath of the terrorist attack and subsequently decays to pre-attack levels within approximately one month. Further, we show that this effect is driven primarily by increases in institutional and interpersonal trust, as well as an increase in engagement in social networks. Additionally, we explore various dimensions of heterogeneity with respect to the impact of the attack on social capital and further show that our results are robust to various falsification tests, alternative control groups, the inclusion of additional controls, and an alternative identification strategy. Finally, we provide suggestive evidence that our results for the Charlie Hebdo attack are externally valid relative to the overall pattern between terrorism and social capital when using the broader population of terrorist attacks in Europe.

Literature Review
Over the last decade, numerous works have explored the influence of terrorism on social capital. Drawing mainly on the experience of the United States, many of these studies find that terrorism is associated with an increase in the level of social capital. With respect to institutional trust, Chanley (2002), Collins (2004), Putnam (2002), and Sinclair and LoCicero (2010) conclude that terrorist attacks, or even fear of a terrorist attack, is associated with greater trust in government and other institutions. This boost in political trust can be interpreted as a rally effect; however, researchers note that this consensus starts to dissipate as people revert to prior political positions. Collins (2004), drawing on the 9/11 experience, reports that intense solidarity builds momentum 1-2 weeks after the event, remains at peak levels for roughly two months, and then gradually returns to prior levels.
With respect to interpersonal trust, Gross et al. (2004), Paxton (2005), and Putnam (2002) further find a positive association with terrorism. These studies assert that terrorismrelated events strengthen the bond between citizens and heighten social trust. For example, Putnam (2002) argues that the 9/11 terrorist attack led to Americans putting aside their differences and bonding across ethnic, racial, class, and partisan lines. As noted by Paxton (2005), the enhanced solidarity following 9/11 appeared to be at the group level, which generated its own process of conflict between groups (Collins 2004;Sander and Putnam 2010). However, these group-specific conflicts declined towards normalcy within 6-9 months following the attack (Paxton 2005;Sander and Putnam 2010).
In contrast to these studies, other work exploring the impact of terrorism in European countries finds that terrorist attacks are associated with a neutral, or even a negative effect, on social capital. For example, Bali (2007), Gassebner et al. (2008), and Montalvo (2011) find a negative or zero effect of terrorism on the electoral support of the incumbent government.
Likewise, Huddy and Feldman (2006) show that fear of terrorist attacks in the aftermath of 9/11 made people less likely to trust others. Blomberg et al. (2011) also report a negative influence of terrorism on trust, but acknowledge that the effects could, in principle, go in any direction, depending on the composition and characteristics of the affected population.
Among existing papers looking at the impact of terrorism on social capital, the study by Arvanitidis et al. (2016) is perhaps the closest to our work. Using data drawn from the European Social Survey (ESS) for the years 2004, 2012, and 2014, the authors find that terrorist incidents can trigger social dynamics that affect trust attitudes; however, the authors conclude that these effects are short-lived and dissipate rapidly. Specifically, the authors use four major terrorist events as case studies in their empirical examination: the 2004 Madrid train bombings, the 2011 Oslo and Utoya attacks, the 2014 Jewish Museum shooting in Belgium, and the 2015 Charlie Hebdo shooting in France. The outcome variables, in their study, include composite indices of institutional and interpersonal trust. The independent variable of interest is 'Terrorism elapsed days,' which counts the number of days elapsed between the time of the terrorist event and the interview date for each ESS participant in the corresponding country. Using an OLS framework, for the 2004 Madrid train bombings, they find evidence of an inverted U-shaped relationship between the terrorist incident and public trust in government and institutions; institutional trust increases in the first few days following the attack, but as time elapses, this initial strengthening of institutional trust levels off and subsequently exhibits signs of reverting towards pre-terrorist attack levels. However, for the 2015 Charlie Hebdo attack, they find evidence of a linear and negative effect of terrorism on trust attitudes. Although the authors argue that the timing of survey administration may be responsible for the contradictory results, one cannot rule out the possibility that unobserved country specific factors might be driving their results.
Like Arvanitidis et al. (2016), we also use the Charlie Hebdo attack to examine the impact of terrorism on social capital. However, our study differs in three main respects.
First, Arvanitidis et al. (2016) only explore how post-terror trust attitudes change over time, and therefore only use observations from the period after the Charlie Hebdo attack.
Our focus, instead, is on estimating the causal effect of a large scale terror attack on social capital, as well as understanding the temporal dynamics following the attack. As such, we use observations prior to and after the terrorist attack and implement a carefully designed empirical strategy comparing the difference in social capital between a country that is subject to the attack and countries that are not, both before and after the terrorist incident. Second, the set of elements of social capital that we examine is much broader compared to Arvanitidis et al. (2016). Specifically, while Arvanitidis et al. (2016) focus primarily on institutional and interpersonal trust, the set of outcomes we consider include these two elements of social capital as well as indices capturing engagement in social networks, norms of reciprocity, ethnic tolerance, and an overall index of social capital. As noted by the OECD (2001) and the Australian Bureau of Statistics (2002), a statistical framework for social capital should ideally contain multiple elements relevant to networks, social norms, and values that facilitate cooperation within and among groups. Lastly, we explore the impacts across various dimensions of heterogeneity including one's religiosity, political orientation, and exposure to television.

Data Description
The data for this study comes from the 2014 wave (round 7) of the European Social Survey (ESS) that was administered between April 2014 and September 2015. The 2014 ESS is a representative survey used to gather information on behaviors, beliefs, attitudes, and demographics of individuals across 22 European countries. Although the seventh round of the data is cross-sectional in nature, we exploit the fact that individual survey start times and survey end times were recorded, thus implicitly providing and additional dimension of time.
Further, the scheduled interview times were randomly set by ESS survey administrators.
This feature of the data, coupled with knowing the country of residence when the survey was taken, allows us to place the Charlie Hebdo shooting in a meaningful way with respect to individuals exposed to the attack (treatment vs. control) and individuals that completed the survey before or after the incident. As such, we are able to exploit the arguably random nature of the attack to estimate the causal effect of terrorism on social capital using a DID framework.
Motivated by the existing literature on social capital, our outcome of interest is a composite index of social capital comprised of measures on institutional trust, interpersonal trust, engagement in social networks, norms of reciprocity, and ethnic tolerance (Putnam et al. 1993;Fukuyama 1997;Durlauf and Fafchamps 2005). Each of these elements of social capital are constructed by standardizing, within each country, underlying survey responses to be mean zero with a standard deviation equal to one. 4 Once these measures are standardized, and similar to Kling et al. (2007), an equal weighted average of z-scores is taken to form component-level indices, and subsequently a broader index of social capital. Missing data related to a particular component of social capital is replaced using country specific means calculated over the non-missing data. Furthermore, the indices are constructed in such a way that for each element of social capital, higher ( Covariates used in our most robust model specifications include measures for age and age squared, years of education and years of education squared, an indicator for gender, and a full set of indicators classifying households within the distribution of income. Conditioning on these covariates, and using the broadest window of time possible, results in a sample size of 12,466 observations with 1,765 respondents residing in France (the treatment group) and 10,701 individuals residing in one of the control countries. 5 The 2014 wave of the ESS includes individuals sampled in: Austria, Belgium, Switzerland, Czech Republic, Germany, Denmark, Spain, Finland, United Kingdom, Republic of Ireland, Netherlands, Norway, Portugal, Sweden, and Slovenia. Among these countries, Belgium, 4 See Table A1 in the supplemental appendix for the list of survey questions that we have used to construct the different elements of social capital. 5 The sampling window of 68 days before the attack and 47 days after the attack is the maximum window of time we could use given the countries in our sample.
Switzerland, Germany, Spain, and Netherlands share a common border with France while Austria, Czech Republic, Denmark, Finland, United Kingdom, Republic of Ireland, Norway, Portugal, Sweden, and Slovenia do not. Henceforth, in our empirical work we consider three distinct set of control countries: all countries other than France; all countries other than France that share a border with France (we refer to this subset of controls as bordering countries); and all countries other than France that do not share a border with France (we refer to this subset of controls as non-bordering countries). To deal with the complex nature of how the data was collected across countries, sampling design weights were used throughout the analysis. Descriptive statistics broken down by treatment status and time can be found in supplemental appendix Tables A2 to A4.

Data Visualization and Pre-Trends
The difference-in-differences estimation strategy will identify the causal effect of the terror attack on social capital if the average change in social capital among individuals in France would have been proportional to the average change in social capital among individuals in the set of controls in the absence of the attack. However, this assumption cannot be explicitly tested since we do not observe the counterfactual outcome in France. Instead, we provide graphical evidence that the evolution of each social capital metric in the period of time leading up to the attack in France parallels the evolution of each outcome in the set of control countries.

Baseline Econometric Model
Given the ability to exploit the plausibly exogenous nature of the Charlie Hebdo terrorist attack combined with knowing the day, month, and year that individual surveys were started and completed, we are able to employ the DID estimator according to (1) as a means of examining the impact of terrorism on social capital. 6 Specifically, letting i denote an individual surveyed in country m at time t, our baseline DID models take the form: (1) Here, T reat m is an indicator variable set equal to one for any individual in our sample located in France, and zero otherwise. Likewise, P ost t is an indicator variable set equal to one for any individual in our sample that was surveyed after the Charlie Hebdo terror attack. C m is a set of country fixed effects, T t is a complete set of week fixed effects, and D i is a set of demographic controls including: second-order polynomials in age and education (in years); a binary variable indicating if the respondent is female; and a set of categorical variables indicating the decile of the respondent's household's total net income. Further, we estimate equation (1) multiple times using our broad measure of social capital, as well as 6 To the best of our knowledge there were no other major, non-terror related incidents that could be confounding the events of Charlie Hebdo (see a timeline of events here: https://www.bbc.com/news/world-europe-17299605; last accessed on November 25, 2018). With that said, there were multiple smaller incidents resulting in fatalities that occurred over the following two days after the attack at Charlie Hebdo. Most notably on January 8th an individual with ties to the Charlie Hebdo assailants shot and killed a police officer and injured a city employee in the Montrouge area, and on January 9th this same assailant took employees and customers hostage at a kosher supermarket in the Porte de Vincenne area. The standoff ended with elite forces storming the store and killing the assailant; four hostages were also killed with three others injured. As a quick check, we drop individuals in the post period but who were surveyed on the two days following the Charlie Hebdo shooting and get qualitatively unchanged results.
Of interest to us given equation (2) is the coefficient estimate of δ, which captures dif-ferences in the average trend in social capital in the treatment group relative to the average trend in social capital in the control group over the period of time leading up to the terror attack. If the prior trends assumption holds, then this coefficient estimate should be statistically indistinguishable from zero.

Temporal Dynamics
In the preceding analysis, we utilize the DID estimator to assess the average impact of the Charlie Hebdo terror attack on various measures of social capital. Implicit in the prior specification given by equation (1) Given equation (3), the parameter of interest is, evaluated over the domain of elapsed days in our sample, j = 1, ..., 47 days.

Results
We begin our formal analysis in section 5.1 by estimating equation (1), which allows us to estimate the average effect of the terror attack on each measure of social capital delineated in section (3). We then turn our attention to estimating the extent to which our estimated treatment effects captured in equation (1) vary over time in section 5.2.  Tables A5 to A10 coefficient estimates with and without the inclusion of these controls and also show that our main results are robust to the exclusion of countries immediately bordering France 11 .

Baseline Results
10 To aid in interpretation of the magnitudes of the effects of Charlie Hebdo, in supplemental appendix section A.2.4 we re-scale our difference-in-differences estimates of Charlie Hebdo in terms of the estimated effects of other model covariates. 11 This provides some evidence that the effects of Charlie Hebdo were confined to France. Please refer to the supplementary appendix section A.2.1 for a complete discussion of these results as well as additional tests for the presence of spillover effects which compare the evolution of social capital in bordering countries to non-bordering countries, before and after the attack.
Lastly, and as noted earlier in section 4.1 and visually inspected via Figure 1, a critical assumption required to identify the causal effect of the Charlie Hebdo attack is the existence of parallel trends in the outcome of interest prior to the attack. In effect, we are assuming that individuals in the set of control countries act as a valid counterfactual for our treated individuals residing in France. Figure 1 presents graphical evidence suggesting that the key underlying identifying assumption of our empirical models holds. Inspection of the raw data is useful for this purpose, but has its limitations. In our empirical setting, we do not utilize longitudinal data on a sample of individuals. Rather, we leverage the implicit time component in the ESS survey to estimate a DID model using a repeated cross-sections of respondents. In the traditional panel data setting, all unobserved determinants of social capital in the treatment and control group would be captured by country specific effects and ultimately be reflected in Figure 1 by level shifts in the overall trends between the treatment and the control group. In the repeated cross-sectional setting, we would want to be sure that once conditioning on the covariates included in our difference-in-differences specification, France and the controls exhibit similar pre-treatment trends in each outcome.
We provide evidence this is the case by estimating equation (2) using our broad social capital index, as well as the various individual components, as our outcome variable, Z itm .
We estimate various specifications of equation (2)  Broadly, we find that the coefficient estimates are statistically no different from zero.
There are exceptions, however, related to certain specifications associated with the institu-tional trust index. As illustrated in columns (1) and (2) Table A6 in the appendix, the estimated impact of the terrorist attack on institutional trust is stable across every model specification, thus alleviating some concern that the results are driven by divergent trends in the pre-attack period. This concern is further alleviated when the broad findings from pre-trend estimates are coupled with the pre-trend figures presented earlier. However, given the crucial nature of the pre-trends assumption and credible concerns regarding threats to identification, we provide a suite of robustness checks presented in the supplemental appendix. 12 Focusing attention to panels (A) -(C) in Figure 2, two things become immediately apparent. First, the impact of the attack on social capital, as well as on the individual measures for institutional and interpersonal trust, is greater than 0.5 standard deviations the day following the attack. Second, though this effect is seemingly large in the immediate aftermath, the impact decays quickly over the days following the event. In particular, the el-12 In particular, we provide a very comprehensive supplemental appendix exploring threats to identification and robustness of our main results (section A.2), re-estimating our main results varying the control set to be bordering and non-bordering countries and considering spillover effects (section A. Looking at the figure for engagement in social networks (panel (D) in Figure 2, a very similar pattern emerges. Specifically, on the day following the attack there is an increase in the normalized measure capturing engagement in social networks by roughly 0.5 standard deviations. However, the rate at which engagement in social networks returns to pre-attack levels is even faster than that documented for the broad measure of social capital, institutional trust, and interpersonal trust. In fact, after day 10 the impact is not statistically different from zero.

Temporal Dynamics
Next, we explore the temporal dynamics associated with the measures of reciprocity and tolerance. Although we do not find an average effect given equation (1), it could be the case that there was some impact very early on but then averaged out given the nature of the dynamics following the event. With that said, and as can be seen in panels (E) and (F) of Overall we find that the estimated causal effect of the Charlie Hebdo attack on social capital, and various factors comprising this measure, is non-trivial in the immediate aftermath of the terror incident, but tends to quickly revert to pre-attack levels. In particular, the factors driving the results associated with our broad measure of social capital are trust in institutions, interpersonal trust, and engagement in social networks. We find zero impact with respect to outcome measures of reciprocity and tolerance, both on average and in a temporal sense. Our results are thus in line with the findings of Chanley (2002), Putnam (2002), Collins (2004), Gross et al. (2004), Paxton (2005) and Sinclair and LoCicero (2010). As discussed in Section 2, these studies, like ours, document a positive influence of terror-related attacks on social capital that attenuates within a relatively short period of time following the attack. However, the aforementioned authors find social capital returning to normalcy over a 6-9 month time frame whereas we find effects decaying much more rapidly. Further, the findings of these previous studies are only capturing associations and are limited in making causal claims. With that said, and because our findings are much more fleeting in nature, the overall impact is arguably not as economically significant relative to the existing studies.
Thus, our work further contributes to the literature by tempering expectations regarding the persistent nature of the "rally effects" stemming from a terror-related attack.

Heterogeneity Analysis
We next evaluate how the effects on social capital stemming from the Charlie Hebdo terror attack vary across multiple dimensions of heterogeneity. Given the often political and religious nature associated with terrorist attacks, we first explore how the effect of the Charlie Hebdo shooting varies across the distribution of one's political orientation (right vs. left) and whether one identifies as being religious (religious vs. non-religious). 13 As well, given the amount of media exposure that these types of events get, we then look at how the impact of the attack varies across the distribution of television hours watched (high vs. low). 14 For each dimension of heterogeneity, we divide the sample into two groups, G1 i and G2 i , and test how our baseline difference-in-differences estimates vary across each group. In all cases the same set of demographic controls as outlined in section 4 and week fixed effects are included.
13 Hours Watched (High) is a binary variable equal to one for any respondent reporting an above median level of hours of television watched. The variable Right Wing was constructed from each respondents' response to a question in the ESS assessing their position on the left/right scale. Right Wing is set equal to one for any respondent reporting an above median level of this scale. 14 Other dimensions of heterogeneity that we explore are gender, education, employment status, and income.
These additional results can be found in supplemental appendix Tables A17 and A18.
To operationalize these tests, we estimate: 15 Of interest to us are coefficient estimates of β 1 and β 2 , which represent the average effect of the attack on individuals belonging to group 1 and group 2, respectively. Further, in each specification, we also report estimated differences in the impacts of the terror attack on each group (β 2 − β 1 ) along with the p-value associated with the test: H 0 : β 1 − β 2 = 0; All results can be found in Table 2.
First looking at the impacts of the attack along the dimension of political orientation (panel (a), social networks for the right wing is statistically significant at conventional levels. For both the right and the left, the impact of the terrorist attack on reciprocity and ethnic tolerance is statistically indistinguishable from zero. Lastly, differences in the estimated effects between the right and left wing are not statistically significant for the broad measure of social capital or any of the sub-components.
Next, we explore how the impact of the terrorist attack varies across the distribution of religiosity (panel (b), Table 2). In particular, we discretize the data into religious and nonreligious groups and again estimate (5 In panel (c) of Table 2 we look at high versus low television watchers and find that both groups are affected by the terrorist attack, in a positive and statistically relevant sense, with respect to the broadest measure of social capital. The same holds true for institutional trust and interpersonal trust. Interestingly, those in the high hours of television watched group experienced an impact of 0.343 standard deviations increase in reciprocity, with this result being highly statistically significant. This is the only point in all of the analysis where the impact of the terrorist attack results in a measurable, statistically significant effect on reciprocity. Further, the estimated difference in the impact of the terror attack on reciprocity between high and low television watchers is also statistically significant at conventional levels.
Overall in Table 2, institutional and interpersonal trust continue to be the primary components driving the results found when looking at social capital broadly. Two interesting points are worth noting. First, those who watch more television experience a statistically significant, sizable impact of the terrorist attack on reciprocity. Second, those who claim to be non-religious tend to increase their trust, economically and statistically, in both institutions and interpersonally relative to those who claim to be religious. However, one should exercise caution when interpreting the heterogeneity analysis presented here. Specifically, the results are purely descriptive in nature since the various dimensions of heterogeneity may themselves have been affected by the terrorist attack. 16

Conclusion
Acts of terrorism have increased dramatically over the previous two decades. The measurable costs of such events, related to physical capital and loss of life, are great in magnitude and relatively easy to quantify. However, the impact on a country's level of social capital is not so easily assessed. As such, we utilize the Charlie Hebdo attack, coupled with the staggered and random nature that respondents were interviewed for the 2014 wave of the European Social Survey, as a means of identifying the impact of terrorism on a country's level of social capital. Our results demonstrate that terrorism has a significant positive and causal effect on social capital. The effect seems to be driven by an increase in institutional trust, interpersonal trust, as well as engagement in social networks. Our results are consistent with the hypothesis of a "rally effect." In looking at the temporal dynamics, we find that the impact of terrorism on social capital is at its peak right after the terror attack, but quickly reverts to pre-attack levels.
16 Additional heterogeneity analyses across various demographic and socioeconomic measures can be found in supplemental appendix section A.1. In this section we also explore differential effects among French respondents (e.g. born in France vs. not born in France) as well as how the effect varies conditional on one's proximity to the epicenter of the attack in Paris.
Since the level of social capital is elevated in the immediate aftermath following a terrorist attack, the impacted country's government may consider addressing concerns about national security, homeland defense, and combating terrorism abroad by leveraging the increase in social capital as a means of winning support to carry out various prescriptive measures.
Further, the impacted government may consider using the augmented level of social capital as a policy tool to recover the direct economic losses due to terrorism or raise requisite revenues needed to implement programs aimed at mitigating radicalization both at home and abroad. As noted by Helliwell et al. (2017), the effect of positive interventions to restore the economic damages due to terrorism is likely to be amplified due to an increase in the stock of social capital. In addition, the higher stock of social capital may also be useful in diffusing threats of future terrorist attacks since it is found that "social capital is positively associated with increased [emergency] preparedness and with increased concern about future terrorism events" (Hausman et al., 2007(Hausman et al., , p. 1073. Lastly, the increase in social capital following an attack can, perhaps, be used to mitigate post-terrorism popular uprisings and/or inter-ethnic group conflicts. Durlauf, S.N., and Fafchamps, M. (2005). Social Capital. In: Aghion, P., Durlauf,   Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust (Huber-White) standard errors shown in parentheses. Columns (1) -(6) report coefficient estimates of the treatment group by post Charlie Hebdo interaction terms as specified in estimating equation (1). The social capital index of interest used on the left hand side of estimating equation (1) is reported in the column headings above the column numbers. The set of demographic controls include: Second-order polynomials in age and education (in years); a binary variable indicating if the respondent is female; and a set of categorical variables indicating the decile of the respondent's household's total net income. Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust (Huber-White) standard errors in parentheses. Columns (1) -(6) report coefficient estimates of the treatment group by post Charlie Hebdo interaction terms as specified in estimating equation (5) using the full sample of observations [-68 days, 47 days] and full sample of countries. Each model includes country fixed effects, week fixed effects, and demographic controls (second-order polynomials in age and education; a binary variable indicating if the respondent is female; and a set of categorical variables indicating the decile of the respondent's household's total net income). The social capital index of interest used on the left hand side of estimating equation (5) is reported in the column headings above the column numbers.The dimension of heterogeneity considered is given in each panel heading.   Difference-in-Differences Estimates (θj)

Appendix A Supplemental Appendix
Here we conduct additional heterogeneity analyses, rigorously explore threats to identification and robustness of our main results, discuss external validity, and reproduce the analysis using attitudes toward immigrants as the outcome of interest.

A.1 Additional Heterogeneity Analyses
Here we present additional dimensions of heterogeneity results that explore how the impact of the terrorist attack varies across multiple demographic and socioeconomic measures. For ease of exposition, we partition model estimates of equation 5 found in the main manuscript into two tables: Table A17 and Table A18. In Table A17 we consider gender (male vs. female), age (old vs. young), and education (high vs. low). In Table A18 we present results for marital status, employment status, and income. 1 Each column of Tables A17 and Table   A18 present model estimates of equation (5) from the main manuscript for social capital and each individual component of the index. As we did before, in each specification, we also report estimated differences in the impacts of the terror attack on each group (β 2 − β 1 ) along the the p-value associated with the test: We first present results pertaining to impacts across the dimensions of gender, age, and education (Table A17). Interestingly, we find statistically significant effects of the terrorist attack within groups for the broad measure of social capital and its various components, but find very few instances of statistically relevant results between groups. Specifically, we see that for both males and females (panel (a)), the impact of the terrorist attack on the broad measure of social capital is 0.255 and 0.280 standard deviations, respectively.
The impact on the individual components of social capital including institutional trust and interpersonal trust are 0.330 and 0.192 standard deviations for males and 0.291 and 0.170 standard deviations for females. These effects are statistically significant at conventional levels. However, the estimated impacts on social networks, reciprocity, and ethnic tolerance are statistically insignificant for both males and females.
Turning our attention to panel (b) of Table A17, we present results exploring the impacts of the terror attack across the dimension of age. In particular, we define "older" as an age at or above the median age and "younger" as an age below the median. Looking first at the impact of the terrorist attack on social capital broadly, we find that the terrorist attack resulted in a 0.200 and 0.341 standard deviation increase for older and younger individuals, respectively, with these results being statistically significant at conventional levels. For the various sub-components of social capital, the impact of the shooting on institutional trust is 0.256 standard deviations for older individuals and 0.368 standard deviations for younger individuals. Again these estimated impacts are statistically significant at conventional levels.
With respect to interpersonal trust, the impacts are 0.142 and 0.212 standard deviations for older and younger individuals, respectively. However, the estimated impact on interpersonal trust is only statistically significant for those classified as being younger. The same is true for social networks where the 0.227 standard deviation impact for younger individuals is statistically significant whereas the impact is near zero and statistically insignificant for individuals older than the median age. Lastly, the remaining impacts on reciprocity and ethnic tolerance are, for both age groups, approximately zero in magnitude and statistically insignificant at conventional levels.
In panel (c) of Looking along the dimension of income reveals findings similar to those presented earlier (panel (c), Table A18). As before, we use the median level of this measure to demarcate between "high" versus "low" levels of income. The results presented in Table 2 and Table A18 are similar to the results presented in with "Distance to Attack". Of interest to us are coefficient estimates on "Distance to Attack" interacted with the post Charlie Hebdo indicator variable; these results are shown in Table   A20. Inspection of Table A20 shows that across all model specifications and all dependent variables, coefficient estimates on "Distance to Attack × Post" are statistically insignificant.
If the impact of the attack in France varied with respect to proximity to the attack we would expect these coefficients to be statistically significant. However, the estimated effects are not providing evidence that Charlie Hebdo led to a uniform shift in social capital across France.

A.2 Threats to Identification and Robustness
We proceed by discussing and assessing threats to the identifying assumptions of our model. provides evidence from a regression discontinuity design.

A.2.1 Border vs. Non-Border Countries and Spillovers
In the preceding sections, we report our main results using all countries other than France as controls. However, one might be concerned that the amount of publicity of the attack may have induced changes in the level of social capital in neighboring countries. Along these lines, in this section we we investigate the extent to which spatial spillovers exist.
First, in Tables A5 to A10, for each social capital index, we investigate how our model estimates change in response to refining the definition of the control group. We report estimates for each social capital index separately in each table with column (3) reporting our main results summarized in Table 1 which use all countries other than France as the control group. For the sake of completeness, columns (1) and (2) test the sensitivity of our main results to the exclusion of demographic controls and week fixed effects. Column (4) replicates column (3) after excluding the set of non-bordering countries from the sample.
Likewise, column (5) Table A7). All results pertaining to the impact on interpersonal trust are statistically significant at conventional levels.
In panel (a) of Table A8 which reports our results for engagement in informal social networks, the average impact of the attack ranged from 0.211 to 0.122 standard deviations (panel (a), Table A8). However, only the results found in columns (1) and (2)  To operationalize these tests, we first omit France from the sample. Second, we reestimate equation (1) from the main manuscript after replacing the variable T reat m with a binary variable set equal to one for the set of countries bordering France and zero otherwise.
The results of this exercise, which are reported in Table A21, show that there does not exist any statistically significant change in social capital in bordering countries relative to nonbordering countries before and after Charlie Hebdo irrespective of the social capital index or sub-sample we consider. With that said, even in the event of spillovers occurring from France to the set of control countries, we would be, in effect, estimating a lower bound.

A.2.2 Additional Robustness Checks
Throughout the paper we estimate each model using all countries in the ESS; however, for three countries (Spain, Norway and Portugal) there exists data on respondents surveyed before the attack but not after. In Table A26, for each outcome of interest we re-estimate the most robust version of our empirical model (e.g. equation (1) with the full set of covariates specified in the sub-panel titled Baseline covariates and sample definitions of Table A26) after excluding data from Spain, Norway and Portugal from consideration. In addition, we report p-values for two-tailed tests of differences between each coefficient estimate reported in Table A26 and the corresponding set coefficient estimate obtained from estimates of equation (1) that includes Spain, Norway and Portugal. As shown in Table A26, model estimates are qualitatively unaffected to refining the set of control countries.
Next, in Table A27 we test the robustness of our empirical models to the inclusion of additional control variables. For each social capital index of interest, Table A27 reports coefficient estimates of the treatment group by post Charlie Hebdo interaction terms as specified in estimating equation (1)

A.2.3 Cluster-Robust Standard Errors
Throughout the paper, statistical inference is conducted with robust, heteroskedasticityconsistent (e.g. Huber-White) standard errors. In Tables A29 and A30 we consider two alternative approaches to obtaining standard errors. Specifically, in Table A29 for our most robust empirical specification and for each measure of social capital, statistical inference is performed with standard errors clustered at the country level. In Table A30, standard errors are clustered at the level of country by pre/post. And finally, in Table A31, standard errors are clustered at the country by week level. At all levels of clustering, standard errors are generally smaller than the corresponding set of Huber-White robust standard errors used throughout the paper.

A.2.4 Re-scaled Coefficient Estimates of Charlie Hebdo
Here we re-scale model estimates of T reat m × P ost t from estimating equation (1)  Turning attention first to the effects of Charlie Hebdo re-scaled in terms of education, in the first row, column (1) of Table A32 we find an estimate of 6.624. What this value suggests is that in order to reproduce the same effect on social capital that was induced by the Charlie Hebdo attack, education among individuals would have to be increased, on average, by roughly 6 years. Re-scaled effects of Charlie Hebdo in terms of changes in education for institutional and interpersonal trust are roughly 13 years and 5 years, respectively. All other effects re-scaled in terms of education can be found in Table A32. almost two times the effect. In essence, the higher one climbs the income distribution, the magnitude of the disparities in average levels of social capital between these higher income groups and the bottom 10% of income earners is more than the induced rally effect achieved by the terrorist attack. The complete set of re-scaled effects across all deciles of income and across all components of social capital, relative to the bottom 10% of income earners, can be found in Table A32.

A.2.5 Placebo Test -Timing of Event
Next, in Table A33 we report the results of a placebo test whereby we modify the timing of the attack. To perform these tests, we first restrict attention to the pre-treatment time-period.
Next, we split the sample in half with respect to time and test for any differential change in social capital in France relative to the set of controls. For example, restricting attention to the broadest pre-treatment window of -68 to -1 days, we generate an indicator variable, I(−35 Days, −1 Day) t , that is set equal to one for any respondent surveyed between -35 and -1 days of Charlie Hebdo. We then re-estimate equation (1) replacing the P ost t with I(−35 Days, −1 Day) t . We perform the same exercise restricting the pre-treatment window from -47 to -1 days and -30 to -1 days and report all estimated coefficients on the interaction terms of interest in Table A33 Inspecting Table A33 reveals that of the eighteen coefficient estimates reported, only one is statistically significant. Specifically, the coefficient estimate reported in column (2) of panel (a) is positive and marginally significant. This result implies is that relative to the set of controls, France experienced a small change in institutional trust in the 35 day window of time before Charlie Hebdo. However, in the smallest pre-treatment window of -30 to -1 days, we fail to detect any statistically significant effects. Recall that our formal estimates for differences in trends reported in Table A12 also fail to indicate any statistically significant difference in the overall trend in institutional trust among French respondents relative to respondents in all other countries, bordering countries, and/or non-bordering countries. These results, coupled with the placebo results associated with the -30 to -1 day window of time, provides additional evidence suggesting that the underlying parallel trends assumption holds within the 30 day window of time around the terrorist attack.

A.2.6 Evidence from a Regression Discontinuity Approach
In section A.2.1 we test for spatial spillovers due to Charlie Hebdo by contrasting bordering countries to non-bordering countries the results of which indicate that in relative terms, there does not exist any differential change in social capital between bordering and nonbordering countries. These findings lend credence to the validity of using countries other than France as a control group but do not necessarily rule out the possibility that our model estimates may be biased due to spillover effects stemming from both bordering and nonbordering countries responding to Charlie Hebdo. This might be particularly problematic for the interpretation of our estimates if, for instance, there was a decrease in social capital in all countries, but a little less in France. Before proceeding any further, we highlight here that our visual inspection of the data seem to suggest this is not the case. For every measure of social capital, Figure 1 found in the main manuscript fails to indicate any structural shift in social capital among the set of control countries immediately following the attack.
In contrast, Figure 1 shows an immediate spike in social capital following the onset of the terrorist attack in France.
These descriptive results provide evidence that the effects of the terrorist attack on social capital were localized in France. We proceed by quantifying this immediate jump in social capital in France more formally using a regression discontinuity framework. To do this, we restrict attention to French respondents and estimate: Of interest to us are coefficient estimates of β 1 , which capture the immediate shift in the average level of social capital among French respondents. Estimates obtained from equation (6), which are are reported in Table A22, indicate that immediate jumps in our broadest measure of social capital, institutional trust, and interpersonal trust, as shown in Figure 1 of the main manuscript, are also statistically relevant.

A.3 External Validity and Extensions
We highlight here that the 2015 Charlie Hebdo terrorist attack was not an attack on random people, but rather an attack against politically engaged journalists. As such, one might raise the concern is that our findings are not externally valid with respect to the impact of terrorist attacks on social capital more generally. Additionally, understanding how terrorism impacts attitudes towards immigrants is an interesting extension worth investigating given the available data. As such, we perform three additional analyses using data on the  (1) from the main manuscript where T reat m is now an indicator variable now set equal to one for any individual in our sample located in Germany, and zero otherwise. 4 As before, P ost t is an indicator variable set equal to one for any individual in our sample that was surveyed after the terror attack, C m is a set of country fixed effects, T t is a complete set of week fixed effects, and D i is the same set of demographic controls used in the Charlie Hebdo analysis. We again estimate estimate (1) multiple times using our broad measure of social capital, as well as the various components of social capital, as our outcome of interest. We further estimate the model multiple times varying the window of time around the Christmas attack with the full sample ranging from -112 days before the attack to 88 days after the attack. All results can be found in Table A24.
Inspecting It is worth noting that the December Berlin Christmas attack was the fourth major attack over the preceding six months, with these first three attacks resulting in 10 fatalities and 25 major injuries. 5 Unlike the Charlie Hebdo attack in France, which resulted in the first terrorism-related casualty since 2013, attacks in Germany were arguably not surprising in nature if individuals were augmenting expectations using the history of recent terrorist attacks as reference-points. Using a similar line of reasoning as in Backus et al. (2017), but adapted to the current context, the longer a resident of a particular country goes without encountering an "abrupt" terrorist attack leads to greater responses on the margin with respect to social capital. As such, it is reasonable to think that the string of terror-related events in Germany prior to the Christmas attack could possibly induce individuals to make minimal, if any, adjustments on the margin given what is, perhaps, perceived as the new normal. We address this further and attempt to reconcile our findings in the section below.

A.3.2 Global Terrorism Analysis
The unique circumstances Germany faced in the period of time leading up to the Berlin Christmas Attack may lead one to doubt whether the empirical results we report for Germany could be interpreted as causal, or not. As a result, it is difficult to contrast our empirical findings for the Berlin Christmas attack with our empirical findings for the Charlie Hebdo attack. However, our study of Germany does motivate us to think more deeply regarding the degree to which the results we report for France are externally valid. Along these lines, we proceed by discussing the principal advantages, as well as the potential disadvantages, of using Charlie Hebdo as a quasi-experiment for drawing inferences regarding the broader socioeconomic impacts of terrorism at large.
There are many aspects of the Charlie Hebdo attack that motivate the appropriateness of the difference-in-differences methodology utilized in this study. Most notably, Charlie Hebdo was the first terrorist attack in the West by Islamic extremists only to then be followed by the November 2015 Paris attacks, the Berlin 2016 attack, and the three 2017 London attacks, among others. Rare is the circumstance where we can visualize how social capital evolves over the period of time immediately preceding and following an attack. Further, given the magnitude by which terrorism impacts countries around the world, it is even more challenging to identify experimental settings where concerns regarding bias from other contemporaneous treatments are relatively minimal. Collectively, these observations inspire confidence in the capacity of the difference-in-differences framework to identify a causal relationship.
On the other hand, we also note that there are certain aspects of the Charlie Hebdo attack that may or may not be representative of the average terrorist attack. As noted earlier, Charlie Hebdo was not an attack against random people, but against a French weekly news magazine that made strong political statements. The reaction to this event was both massive in terms of media coverage and public rallies. One could also argue that Charlie Hebdo was highly symbolic in terms of the people who were killed. Further, following the attack the French government and president worked to both unify the nation and display a common sense of identity. For these reasons, even with the climate in France leading up to the "abrupt" attack being an ideal setting to identify the causal effect of terrorism on social capital, the effects that we measure might, nonetheless, work through idiosyncratic components of the Charlie Hebdo attack. Motivated by these observations, we advance an alternative model relating terrorism to social capital on a broader geographic and temporal scale. To draw inferences regarding the extent to which one may interpret our estimates of the Charlie Hebdo attack as externally valid, we subsequently contrast the results of this exercise with our estimates for Charlie Hebdo across each measure of social capital.
First, we create a country-level index of terrorism. To do this, we acquire data from the GTD, which includes information for over 180,000 terrorist incidents around the world.
For each act of terrorism, the GTD records both the affected country as well as the year of each event. Using the GTD, we construct the variable T errorism mt which equals the total number of fatalities due to terrorism in country m in year t standardized mean zero.
Next, after stacking the seventh and the eighth waves 6 of the European Social Survey, we assess the correlation between terrorism and social capital in the broader economy with the estimating equation, Here, Z imt represents our broad measure of social capital, as well as the various components of social capital. C m is a set of country fixed effects. D i is the suite of demographic control specified in section 4.1. T t is a vector of time-controls. In our baseline specification, T t represents a complete set of year fixed effects. In our robust specification, in addition to year fixed effects, we also includes a complete set of linear country-specific time trends to help control for other unobserved determinants of social capital varying through time within a country.
We present model estimates of estimating equation (7) in Table A25. Each column and panel of Table A25 represents the results of a variant estimating equation (7). The dependent variable used is indicated in each sub-panel heading. To gain a sense of the raw correlation between social capital and terrorism in the data, column (1) presents model estimates of equation (7) excluding all controls other than country specific effects. Column (2) adds year fixed effects while column (3) adds demographic controls.
Our robust specification is reported in column (4) which further adds country specific time trends.
Turning attention to column (1), with the exception of reciprocity, overall we observe that social capital and terrorism are positively correlated. Specifically, panel (a) suggests that a one standard deviation increase in the number of fatalities due to terrorism leads to a corresponding 0.018 standard deviation increase in our broadest measure of social capital.

However, inspection of the results reported in panels (b) to (d) show that this effect is largely
driven changes in institutional trust and interpersonal trust. As we move from column (1) to column (4) we observe that the raw correlations in the data as indicated by the model results reported in column (1) persist. In fact, point estimates for the impact of terrorism on social capital and institutional trust become larger in magnitude once we include a robust set of covariates. In no specification do we detect any statistically significant or economically meaningful relationship between terrorism and reciprocity.
Interpreting the results in column (4) as causal would require one to assume that once conditioned on country specific effects, demographics and time, terrorism is random. While we caution the reader in terms of interpreting the results reported in column (4)

A.4 Attitudes towards Immigrants as an Outcome
While not the primary focus of our paper, here we assess how the terrorist attack affected attitudes towards immigrants. In particular, the outcome of interest is now a standardized index capturing attitudes towards immigrants with higher index values corresponding to a more favorable view of immigrants. Specifically, the index is constructed from item responses related to the perceived role that immigrants have on culture, religion, crime, and the economy. The set of demographic controls are the same as in the analysis with social capital as the outcome of interest. We again estimate the model using the full sample, the restricted sample ranging from -47 to 47 days around the attack, and the time frame of -30 to 30 days around the attack. All results can be found in Table A23.
When looking at the results in Table A23  For the sake of completeness, we perform a similar exercise as we did with social capital and evaluate the impact of the attack in a regression discontinuity framework using time as the running variable. Results for this exercise are qualitatively similar to those presented above. However, a severe loss in precision is experienced given the loss of data when only using data associated with French respondents. As such no results presented are statistically significant at conventional levels. All results from this exercise can be found in appendix Table A34.
Overall it appears there is no clear effect of a terror-related attack on attitudes towards immigrants as measured by our index. However, in light of the imprecisely estimated coefficients, these null results should be interpreted with caution. It may very well be that terrorism does affect society's attitudes toward immigrants in a statistically meaningful way, and we are simply unable to draw such conclusions using this particular wave of the European Social Survey. Tables   Table A1:

ESS Survey Questions Used to Construct Indices
Variable Name Survey Question trstprl How much you personally trust each the following institutions : the country's parliament? (0 ̶ 10; 0 = no trust at all; 10 = complete trust).
trstlgl How much you personally trust each the following institutions : the country's legal system? (0 ̶ 10; 0 = no trust at all; 10 = complete trust) trstplc How much you personally trust each the following institutions : country's police? (0 ̶ 10; 0 = no trust at all; 10 = complete trust).
trstprt How much you personally trust each the following institutions : country's political parties? (0 ̶ 10; 0 = no trust at all; 10 = complete trust). trstep How much you personally trust each the following institutions : the European parliament? (0 ̶ 10; 0 = no trust at all; 10 = complete trust).
trstun How much you personally trust each the following institutions : the United Nations? (0 ̶ 10; 0 = no trust at all; 10 = complete trust).
ppltrst Generally speaking, would you say that most people can be trusted, or that you can't be too careful in dealing with people? (0 ̶ 10; 0 = You can't be too careful; 10 = Most people can be trusted).
pplfair Do you think that most people would try to take advantage of you if they got the chance , or would they try to be fair? (0 ̶ 10; 0 = Most people would try to take advantage of me; 10 = Most people would try to be fair).
pplhlp Would you say that most of the time people try to be helpful or that they are mostly looking out for themselves ? (0 ̶ 10; 0 = People mostly look out for themselves ; 10 = People mostly try to be helpful).
sclact Compared to other people of your age, how often would you say you take part in social activities ? (0 ̶ 5; 1=Much less than most; 5 = Much more than most).
ipeqopt Now I will briefly describe some people. Please listen to the description and tell me how much she/he person is or is not like you. She/he thinks it is important that every person in the world should be treated equally .        Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust (Huber-White) standard errors in parentheses. Columns (1) -(5) report coefficient estimates of the treatment group by post Charlie Hebdo interaction terms as specified in estimating equation (1) with the Engagement in (Informal) Social Networks index as the dependent variable. The set of demographic controls include: Second-order polynomials in age and education (in years); a binary variable indicating if the respondent is female; and a set of categorical variables indicating the decile of the respondent's household's total net income. Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust (Huber-White) standard errors in parentheses. Columns (1) -(5) report coefficient estimates of the treatment group by post Charlie Hebdo interaction terms as specified in estimating equation (1) with the Values/Norms of Reciprocity index as the dependent variable. The set of demographic controls include: Second-order polynomials in age and education (in years); a binary variable indicating if the respondent is female; and a set of categorical variables indicating the decile of the respondent's household's total net income. Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust (Huber-White) standard errors in parentheses. Columns (1) -(5) report coefficient estimates of the treatment group by post Charlie Hebdo interaction terms as specified in estimating equation (1) with the Ethnic Tolerance/Tolerance of Diversity index as the dependent variable. The set of demographic controls include: Second-order polynomials in age and education (in years); a binary variable indicating if the respondent is female; and a set of categorical variables indicating the decile of the respondent's household's total net income. Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust (Huber-White) standard errors in parentheses. Columns (1) -(5) report coefficient estimates of the treatment group by days since Charlie Hebdo interaction terms as specified in estimating equation (2) with the Social Capital index as the dependent variable. The set of demographic controls include: Second-order polynomials in age and education (in years); a binary variable indicating if the respondent is female; and a set of categorical variables indicating the decile of the respondent's household's total net income. Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust (Huber-White) standard errors in parentheses. Columns (1) -(5) report coefficient estimates of the treatment group by days since Charlie Hebdo interaction terms as specified in estimating equation (2) with the Trust in Institutions index as the dependent variable. The set of demographic controls include: Second-order polynomials in age and education (in years); a binary variable indicating if the respondent is female; and a set of categorical variables indicating the decile of the respondent's household's total net income. Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust (Huber-White) standard errors in parentheses. Columns (1) -(5) report coefficient estimates of the treatment group by days since Charlie Hebdo interaction terms as specified in estimating equation (2) with the Interpersonal Trust index as the dependent variable. The set of demographic controls include: Second-order polynomials in age and education (in years); a binary variable indicating if the respondent is female; and a set of categorical variables indicating the decile of the respondent's household's total net income. Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust (Huber-White) standard errors in parentheses. Columns (1) -(5) report coefficient estimates of the treatment group by days since Charlie Hebdo interaction terms as specified in estimating equation (2) with the Engagement in (Informal) Social Networks index as the dependent variable. The set of demographic controls include: Second-order polynomials in age and education (in years); a binary variable indicating if the respondent is female; and a set of categorical variables indicating the decile of the respondent's household's total net income. Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust (Huber-White) standard errors in parentheses. Columns (1) -(5) report coefficient estimates of the treatment group by days since Charlie Hebdo interaction terms as specified in estimating equation (2) with the Values/Norms of Reciprocity index as the dependent variable. The set of demographic controls include: Second-order polynomials in age and education (in years); a binary variable indicating if the respondent is female; and a set of categorical variables indicating the decile of the respondent's household's total net income. Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust (Huber-White) standard errors in parentheses. Columns (1) -(5) report coefficient estimates of the treatment group by days since Charlie Hebdo interaction terms as specified in estimating equation (2) with the Ethnic Tolerance/Tolerance of Diversity index as the dependent variable. The set of demographic controls include: Second-order polynomials in age and education (in years); a binary variable indicating if the respondent is female; and a set of categorical variables indicating the decile of the respondent's household's total net income. Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust (Huber-White) standard errors in parentheses. Columns (1) -(6) report coefficient estimates of the treatment group by post Charlie Hebdo interaction terms as specified in estimating equation (5) using the full sample of observations [-68 days, 47 days] and full sample of countries. Each model includes country fixed effects, week fixed effects, and demographic controls (second-order polynomials in age and education; a binary variable indicating if the respondent is female; and a set of categorical variables indicating the decile of the respondent's household's total net income). The social capital index of interest used on the left hand side of estimating equation (1) is reported in the column headings above the column numbers. The dimension of heterogeneity considered is given in each panel heading. Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust (Huber-White) standard errors in parentheses. Columns (1) -(6) report coefficient estimates of the treatment group by post Charlie Hebdo interaction terms as specified in estimating equation (5) using the full sample of observations [-68 days, 47 days] and full sample of countries. Each model includes country fixed effects, week fixed effects, and demographic controls (second-order polynomials in age and education; a binary variable indicating if the respondent is female; and a set of categorical variables indicating the decile of the respondent's household's total net income). The social capital index of interest used on the left hand side of estimating equation (5) is reported in the column headings above the column numbers.The dimension of heterogeneity considered is given in each panel heading. Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust (Huber-White) standard errors in parentheses. Columns (1) -(6) report coefficient estimates of the treatment group by post Charlie Hebdo interaction term as specified in estimating equation (1) defining the set of French respondents born in France as the treatment group (e.g. Born in France = 1) and the set of French respondents not born in France as the control group (e.g. Born in France = 0). Columns (1)-(6) exclude all countries other than France from the estimation. Finally, the social capital index of interest used on the left hand side of estimating equation (1) is reported in the column headings above the column numbers. The set of demographic controls include: Second-order polynomials in age and education (in years); a binary variable indicating if the respondent is female; and a set of categorical variables indicating the decile of the respondent's household's total net income. Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust (Eicker-Huber-White) standard errors shown in parentheses. Columns (1) -(6) report coefficient estimates of the treatment group by post Charlie Hebdo interaction terms as specified in estimating equation (1) replacing the treatment indicator with Distance to Attack (which measures the distance from the centroid of each region in France to the terrorist attack), restricting attention to France, and including region fixed effects. The social capital index of interest used on the left hand side of estimating equation (1) is reported in the column headings above the column numbers. The set of demographic controls include: Second-order polynomials in age and education (in years); a binary variable indicating if the respondent is female; and a set of categorical variables indicating the decile of the respondent's household's total net income. Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust (Huber-White) standard errors in parentheses. Columns (1) -(6) report coefficient estimates of the treatment group by post Charlie Hebdo interaction term as specified in estimating equation (1) defining the set of countries which border France as the treatment group (e.g. I(Bordering Country) = 1) and the set of nonbordering countries as the control group (e.g. I(Bordering Country) = 0). Columns (1)-(6) exclude France from the estimation. Finally, the social capital index of interest used on the left hand side of estimating equation (1) is reported in the column headings above the column numbers. The set of demographic controls include: Second-order polynomials in age and education (in years); a binary variable indicating if the respondent is female; and a set of categorical variables indicating the decile of the respondent's household's total net income. Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust (Huber-White) standard errors in parentheses. Columns (1) -(6) report coefficient estimates of estimating equation (6). The social capital index of interest used on the left hand side of estimating equation (6) is reported in the column headings above the column numbers. The set of demographic controls include: Second-order polynomials in age and education (in years); a binary variable indicating if the respondent is female; and a set of categorical variables indicating the decile of the respondent's household's total net income. Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust (Huber-White) standard errors in parentheses. Columns (1) -(5) report coefficient estimates of the treatment group by post Charlie Hebdo interaction terms as specified in estimating equation (1) with the Attitudes Towards Immigrants index as the dependent variable. The set of demographic controls include: Second-order polynomials in age and education (in years); a binary variable indicating if the respondent is female; and a set of categorical variables indicating the decile of the respondent's household's total net income. Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust (Huber-White) standard errors in parentheses. Columns (1) -(5) report coefficient estimates of the treatment group by post event interaction terms as specified in estimating equation (1). The variable Treat is set equal to one for Germany and zero otherwise. Likewise, the variable Post is set equal to one for any period time after the Berlin Christmas Market Attack. The social capital index of interest used on the left hand side of estimating equation (1) is reported in the column headings. The set of demographic controls include: Second-order polynomials in age and education (in years); a binary variable indicating if the respondent is female; and a set of categorical variables indicating the decile of the respondent's household's total net income. Panel (A) reports estimates using the full sample of observations which span -112 to 88 days since the event.   (1) using the baseline covariates and sample definitions highlighted above but excluding the additional covariates and sample restrictions highlighted above).  (1) using the baseline covariates and sample definitions highlighted above but excluding the additional covariates and sample restrictions highlighted above). Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust (Huber-White) standard errors in parentheses. Columns (1) -(6) report coefficient estimates of the treatment group by post Charlie Hebdo interaction terms as specified in estimating equation (1). The social capital index of interest used on the left hand side of estimating equation (1) is reported in the column headings above the column numbers. The set of demographic controls include: Second-order polynomials in age and education (in years); a binary variable indicating if the respondent is female; and a set of categorical variables indicating the decile of the respondent's household's total net income. Balanced country restriction = y if the estimation sample excludes countries that do not have data on respondents both before and after Charlie Hebdo. Marital status = y if the model includes a binary variable indicating if the respondent is married. Political orientation = y if the model includes categorical variables indicating the respondents placement on the left/right scale. Citizenship = y if the model includes a binary variable indicating if the respondent is a citizen. Employment = y if the model includes a binary variable indicating if the respondent is employed. Religiosity = y if the model includes a binary variable indicating if the respondent is religious. Hours spent watching television = y if the model includes categorical variables indicating the number of hours of television watched. Lastly, we report the p-value associated with the two-tailed test for differences in the coefficient estimates of Treat x Post reported above and coefficient estimates of Treat x Post obtained from our most robust baseline specification (e.g. coefficient estimates of Treat x Post obtained from estimating equation (1) using the baseline covariates and sample definitions highlighted above but excluding the additional covariates and sample restrictions highlighted above). Notes: *** p<0.01, ** p<0.05, * p<0.1. Standard errors shown in parentheses are clustered at the country level. Columns (1) -(6) report coefficient estimates of the treatment group by post Charlie Hebdo interaction terms as specified in estimating equation (1). The social capital index of interest used on the left hand side of estimating equation (1) is reported in the column headings above the column numbers. The set of demographic controls include: Second-order polynomials in age and education (in years); a binary variable indicating if the respondent is female; and a set of categorical variables indicating the decile of the respondent's household's total net income. Notes: *** p<0.01, ** p<0.05, * p<0.1. Standard errors shown in parentheses are clustered at the country by pre/post level. Columns (1) -(6) report coefficient estimates of the treatment group by post Charlie Hebdo interaction terms as specified in estimating equation (1). The social capital index of interest used on the left hand side of estimating equation (1) is reported in the column headings above the column numbers. The set of demographic controls include: Second-order polynomials in age and education (in years); a binary variable indicating if the respondent is female; and a set of categorical variables indicating the decile of the respondent's household's total net income. Notes: *** p<0.01, ** p<0.05, * p<0.1. Standard errors shown in parentheses are clustered at the country by year by week level. Columns (1) -(6) report coefficient estimates of the treatment group by post Charlie Hebdo interaction terms as specified in estimating equation (1). The social capital index of interest used on the left hand side of estimating equation (1) is reported in the column headings above the column numbers. The set of demographic controls include: Second-order polynomials in age and education (in years); a binary variable indicating if the respondent is female; and a set of categorical variables indicating the decile of the respondent's household's total net income.  Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust (Huber-White) standard errors in parentheses. Columns (1) -(6) report coefficient estimates of estimating equation (1) restricting attention to observations before the onset of Charlie Hebdo (e.g. the number of days elapsed since Charlie Hebdo is between -68 and -1). The indicator variable Treat equals one for France and zero otherwise. In panel (A), the indicator variable I(-35 Days, -1 Day) is set equal to one for any period of time such that the number of days elapsed since Charlie Hebdo is between -35 and -1; models restrict attention to -68 days to -1 day. In panel (B), the indicator variable I(-24 Days, -1 Day) is set equal to one for any period of time such that the number of days elapsed since Charlie Hebdo is between -24 and -1; models restrict attention to -47 days to -1 day. In panel (C), the indicator variable I(-15 Days, -1 Day) is set equal to one for any period of time such that the number of days elapsed since Charlie Hebdo is between -15 and -1; models restrict attention to -30 days to -1 day. Finally, the social capital index of interest used on the left hand side of estimating equation (1) is reported in the column headings above the column numbers. The set of demographic controls include: Second-order polynomials in age and education (in years); a binary variable indicating if the respondent is female; and a set of categorical variables indicating the decile of the respondent's household's total net income. Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust (Huber-White) standard errors in parentheses. Columns (1) reports coefficient estimates of estimating equation (6). The set of demographic controls include: Second-order polynomials in age and education (in years); a binary variable indicating if the respondent is female; and a set of categorical variables indicating the decile of the respondent's household's total net income.