Double-Edged Bullets: The Conditional Effect of Terrorism on Vote for the Incumbent

Abstract

Terrorism often seeks to impact democratic politics. This article explores how it can influence the electoral fortunes of the incumbent. Existing research is contradictory. Models of retrospective voting predict a negative impact, as terrorism is detrimental to voters' welfare. However, the well-known ‘rally around the flag’ effect suggests otherwise: following a terrorist attack, voters often cling to the incumbent. We reconcile these arguments and argue that while both effects can coexist, the retrospective assessment is more durable than the rally around the flag. Using data on all deadly domestic terrorist attacks in Spain between 1977 and 2008, matched with municipal-level national election results, we show how exposure to strikes that occur during the last quarter of the term benefit the incumbent, while more temporally distant attacks are electorally harmful. In line with our theory, we find a more pronounced temporal heterogeneity for indiscriminate attacks and those that target civilians.

Terrorist activity in democracies often seeks to influence the political system in order to advance its goals. One key channel through which political violence may exert its influence can be its impact on elections. This is why it is important to understand how voters react to terrorist attacks. Will incumbent support increase or decrease as a consequence of terrorist activity? In standard retrospective voting models, the expectation tends to be the latter. Research in political behaviour consistently shows that voters generally hold the incumbent accountable for events that have a deleterious effect on their welfare. Moreover, this is often the case even with respect to events that are beyond the incumbents' direct control, such as terrorist attacks.

However, we also know that terrorism does not always hurt incumbents. In some circumstances, threatening events induce the so-called ‘rally around the flag’ effect: faced with external threats and uncertainty, citizens unite around the country's leadership and, as a result, incumbents get a popularity boost. The ‘rally around the flag’ effect was first identified for US presidents in the case of wars, international crises and external attacks (Mueller 1970), but it has also been found to occur elsewhere and in reaction to not only international crises or terrorism, but different types of threatening events more generally, such as wildfires (Ramos and Sanz 2020) or the COVID-19 pandemic (Bækgaard et al. 2020; Herrera et al. 2020; Yam et al. 2020).

In this article, we study the effects of terrorism on incumbent electoral support and address this fundamental theoretical ambiguity. Should we expect terrorist attacks to benefit or harm the incumbents' re-election prospects? While it is obvious that being targeted by terrorism is a negative shock to any society's welfare, terrorism has also been shown to be a favourable case for the ‘rally around the flag’ effect. In the aftermath of terrorist attacks, it is all but uncommon to witness pleas for unity and unconditional support of the incumbent from media, opinion leaders and even opposition parties (Chowanietz 2011).

There are hence two potentially competing expectations as to why and how terrorism might affect the electoral fortunes of incumbents. On the one hand, the expectation from a standard retrospective logic would be that exposure to terrorism will make voters more likely to vote against the incumbent. On the other hand, the rally hypothesis would posit that voters exposed to terror attacks will be more prone to support the incumbent in the interest of the country. Not surprisingly, empirical studies investigating the effect of terrorism on incumbent support have provided inconclusive results. Some authors find support for the rally around the flag, while others identify a negative effect of terrorism on incumbent support, and yet still other studies show null effects (see, for example, Baccini et al. 2021; Getmansky and Zeitzoff 2014; Kibris 2011).

However, given this theoretical ambiguity and seemingly contradictory findings, we are not interested in a horse race to adjudicate empirically between the two hypotheses. Both effects may coexist and, eventually, cancel each other out. Learning that in a given case, the average effect of terrorist attacks on incumbent support is either positive or negative (or indistinguishable from zero) would not provide much insight. It could simply be a weighted average of positive and negative effects that would remain obscured by the aggregation. Hence, we need to move beyond the empirical estimation of the average effect of terrorism on incumbent support and further develop the theory to derive some additional and more precise observable implications that are more informative of the actual electoral reactions to terrorism.

In this article, we discuss the conditions under which either the rally or the accountability effects should be expected to prevail, and then try to empirically assess them. We expect time to be the key conditioning factor. The ‘rally round the flag’ effect has often been found to be extremely short-lived. Furthermore, while retrospective voting models also show that voters are myopic and incumbents can escape punishment from temporally distant events, we expect the rally effect to dissipate even more rapidly than the retrospective voting effect.

This expectation is grounded on the nature of the political and psychological underpinnings of the rally around the flag. The key observable implication of this proposition, which we test, is that those attacks that occur close to the election should boost incumbents' electoral prospects, while those that occur at an earlier point in time should not benefit, and perhaps even harm, the incumbent.

In order to test this argument, we leverage fine-grained information on over 200 geolocated, deadly domestic terrorist attacks that occurred in Spain between 1977 and 2008 (De la Calle and Sánchez-Cuenca 2011). We merge the terrorism data with local-level results of Spanish general elections for the same period, generating a municipality–election-attack dyadic dataset. We use geographic proximity of a municipality to an attack as a measure of exposure.

Our empirical strategy leverages variation in the location, timing and characteristics of terrorist attacks (Bove, Efthyvoulou and Pickard 2021), together with the incumbency of various (left and right) parties during an extended period of time. This makes the Spanish case particularly suitable to examine our theory on the conditional effect of terrorism on incumbent electoral support.

In our empirical specifications, we use municipality and year fixed-effects models to account for various sources of heterogeneity. Our results indicate that, on average, exposure to terrorism tends to harm the incumbent. Electoral support for the party in office declines to a larger extent in municipalities that are more exposed to the attacks: one standard deviation increase in geographical proximity to one given attack decreases the incumbent's share of votes by around 0.3 percentage points, which is quite a substantial effect for a single attack, compared to the baseline electoral losses that government parties suffer. On average, therefore, we observe that the retrospective voting logic tends to prevail. However, our results also indicate that the effect of terrorism is highly conditional on the timing of the attacks with respect to the elections. Exposure to strikes that occur during the last year of a term is beneficial for the incumbent's electoral fate, as the rally theory would predict.

Moreover, leveraging various characteristics of the attacks, we show how those strikes that pose a more serious threat for the population at large exacerbate this conditional relationship. Indiscriminate attacks and those that target civilians are both more harmful for the incumbent if they occur early on during the term and more beneficial if they are perpetrated during the last months before an election. This result provides additional evidence in support of our argument of the two effects coexisting sequentially.

Our findings support the idea that terrorism elicits both rallies around and blaming of the incumbent parties; the relative prevalence of one or the other may just depend on the circumstances surrounding the attacks. Rallies around the incumbent to facilitate the effective management of the aftermath of a terror attack by the sitting leadership could yield an electoral premium for government parties when terrorist violence occurs near election dates. However, as this short-term effect vanishes, voters exposed to terrorism may rather prefer replacing the incumbent.

In the next section, we discuss the relevant literature and develop our argument in detail. Then, we contextualize the case and present the data and empirical strategy used to test our theory. The fourth section describes and discusses the results, explores the mechanisms, and rules out some potential alternative explanations. We also briefly discuss a number of additional robustness checks that we present in the Online Supplementary Material. We conclude with a general discussion of our findings and their implications for our understanding of the interplay between terrorism and democratic politics.

Theory

Related Literature

It is well established that terrorist attacks are salient events that capture the attention of the public (Nussio, Bohmelt and Bove 2021) and have the potential to shape citizens' political attitudes on different domains, from law and order (Brouard, Vasilopoulos and Foucault 2018) to foreign policy (Gadarian 2010) or immigration (Böhmelt, Bove and Nussio 2020; Legewie 2013; Nussio, Bove and Steele 2019). However, terrorist violence also affects the cornerstone of democratic politics: elections. Research on the electoral consequences of terrorism has tended to examine whether attacks benefit conservative (‘hawkish’) or progressive (‘dovish’) parties, rather than focusing on the incumbent-versus-opposition logic. In general, these studies conclude that terror attacks appear to benefit conservative and authoritarian parties (Berrebi and Klor 2006; Berrebi and Klor 2008; Getmansky and Zeitzoff 2014; Gould and Klor 2010; Helbling and Meierrieks 2022; Kibris 2011). This is often attributed to the salience of law and order, a type of issue generally owned by the Right, and the sense of threat leading to seek protection from hard-line leaders.

However, studies specifically aimed at investigating the impact of terror attacks on incumbent electoral support are scarce and, if anything, mostly inconclusive. One reason is that in most designs, incumbency is collinear with the partisan composition of the government. In single-case studies, it is not possible to disentangle the progressive–conservative dimension from the incumbent–opposition effect. In any case, some works on the subject find a negative effect of terrorism on incumbent support, others find a positive effect, and yet still other recently published works point to a null effect.

On the one hand, in the literature, there is evidence that, in some cases, incumbents lose electoral support following attacks and casualties, as expected by the retrospective voting theory. Karol and Miguel (2007), for instance, find that Iraq casualties significantly decreased Bush's vote share in the 2004 US presidential election. Similarly, for the Spanish case, Bali (2007) and Montalvo (2011; Montalvo 2012) show that the 2004 Madrid bombings had a remarkably negative impact on electoral support for the incumbent party, even though this effect could be due to the extremely controversial handling of the information on the authorship of the attack by government officials. More informative is the study in which Gassebner, Jong-A-Pin and Mierau (2008) analyse over 800 elections in about 115 countries over the period 1968–2002. They find that terror attacks increase the likelihood that the incumbent government is replaced and that the magnitude of the effect increases with the severity of the attack. Finally, a negative effect of attacks on trust in government has also been identified in contexts of weaker institutions (Gates and Justesen 2020) or larger-scale violence (De Juan and Pierskalla 2016).

The negative effects of terrorist attacks on incumbent support, whenever they are found, square well with the standard models of accountability. A standard expectation from political science retrospective voting models is that whenever a negative shock to citizens' welfare occurs, this will negatively affect incumbent support (Healy and Malhotra 2013). This is also often the case for those ‘exogenous’ events that are deleterious for citizens' welfare while not being under the incumbents' direct responsibility, such as shark attacks (Achen and Bartels 2012), the weather (Healy and Malhotra 2010) or even sports events (Healy, Malhotra and Mo 2010).

Some authors interpret these effects as instances of the malfunctioning of the democratic accountability mechanism, calling it ‘blind accountability’ (Achen and Bartels 2017). Other scholars have pointed to the fact that even for negative shocks that are not under the responsibility of the incumbent, such as natural disasters, voters might reasonably expect the incumbent to prevent, respond to and mitigate their consequences (Healy and Malhotra 2010). It has also been shown how such events that are outside incumbents' control provide opportunities for voters to learn about the incumbents' qualities that would have otherwise remained unknown (Ashworth, de Mesquita and Friedenberg 2018). In other words, facing a security crisis may expose certain qualities of the incumbent, which are rewarded or sanctioned by voters (Ramos and Sanz 2020). In the case of terrorism, given that governments are in charge of national security, voters may attribute them responsibility for the failure in protecting the country from terrorism, or they may judge the incumbents' ability to respond effectively in the aftermath of an attack.

However, there is also abundant research that points to the opposite prediction: terrorism increases incumbent support through the so-called ‘rally around the flag’ effect (Mueller 1970). In the aftermath of a serious attack, it is commonplace to find spikes in government popularity and trust. In fact, several empirical studies have documented that these rallies do generate a pro-incumbent reaction among the public, whereby citizens respond to the terrorist threat by coming together in support for the incumbent government and institutions.

There are multiple instances of rally effects following terrorist attacks. Research has mostly focused on a subset of extremely salient attacks, such as the 11 September 2001 (9/11) attacks in the United States (Bonanno and Jost 2006; Chanley 2002; Hetherington and Nelson 2003; Landau et al. 2004; Woods 2011), the 2004 Madrid bombings in Spain (Dinesen and Jæger 2013), the 2011 Utoya shootings in Norway (Wollebæk et al. 2012), the November 2015 series of attacks in Paris (Coupe 2017) or the 2015 Charlie Hebdo attack also in Paris (Muñoz, Falcó-Gimeno and Hernández 2020). Similar rally effects are also found in contexts of more widespread violence, such as Afghanistan (Deglow and Sundberg 2021).

We can identify two distinct mechanisms through which this rally effect may occur: one is political and the other is psychological. The political mechanism is related to the reaction of media and opposition parties. External threats have been said to create incentives for elites to cooperate across party lines, which can bring ‘politics as usual’ to a standstill (Lijphart 1969). Specifically, in the aftermath of terrorist attacks, opposition parties and the media are often the first to ‘rally around the flag’. Mainstream parties, including the opposition, often launch pleas for unity and unconditional support of the incumbent as a sign of responsibility and common effort against an external enemy. Examples of this include reaction after 9/11 in the United States, as well as to other attacks in France, the UK or Germany (Chowanietz 2016), and also, as a matter of fact, Spain (Chowanietz 2011).

There are reasons to believe that these pleas for unity around the incumbent could work against accountability. By calling to support the incumbent against an external threat, the attribution of responsibility that is necessary if voters are to hold the government accountable becomes more difficult. Rallies around the flag, in a certain sense, could be seen as diluting the clarity of responsibility (Powell and Whitten 1993) because the main political actors rhetorically take collective responsibility and refrain from blaming the incumbent for security failures leading to the attacks. Moreover, voters will not be able to punish the incumbent if the opposition fails to provide a clear alternative (Maeda 2010).

The psychological mechanism that underpins the rally around the flag operates through fear and sense of threat. Terrorist attacks are often conceived to induce feelings of peril and insecurity among the population at large, well beyond those directly affected. Often, media coverage and framing of terrorist attacks, as well as its seeming unpredictability, heighten the subjective perceptions of threat. According to terror management theory, one of the basic functions of leaders is precisely to help citizens to manage this fear of death (Solomon, Greenberg and Pyszczynski 1991). Consequently, when voters perceive themselves as vulnerable or at risk, they have an incentive to support the sitting government, which is in the position to protect them from the threat. In the face of an external menace, it is reasonable to expect that citizens will not question the country's leadership, but, on the contrary, come together to reinforce it (Huddy et al. 2005). Consistently with these premises, the experiments of Landau et al. (2004) showed that when reminded about 9/11, US voters increased their support for the incumbent president (George W. Bush), and when primed to think about their own death, they also decreased their support for the challenger presidential candidate (John Kerry).

Finally, some recent articles fail to identify any causal effect of terrorism on voting for the incumbent. Berrebi and Klor (2008) find that in Israel, suicide attacks increase the vote share of the right-wing bloc regardless of the incumbent's political affiliation. Leveraging the unexpected occurrence of eight Euskadi Ta Askatasuna (‘Basque Country and Freedom’) (ETA) strikes during the fieldwork of various public opinion surveys, Balcells and Torrats-Espinosa (2018) find a positive effect on Spanish voters' intent to participate in elections but no effect on their intention to vote for the incumbent party. Similarly, Baccini et al. (2021), with a refined identification strategy based on the alleged randomness of the success and failure of attempted attacks, also fail to find any electoral effects of terrorist attacks in the United States. Finally, Randahl (2018) shows that when excluding the 9/11 attack, terrorism has no short-term effect on support for the US president.

Argument

There are, therefore, two mechanisms that lead to opposing expectations of the effect of terrorism on incumbent support. The accountability mechanism implies that terrorism should harm incumbents, while the rally mechanism leads to the opposite prediction. Available evidence points in all directions. In this context, an additional study seeking to adjudicate between the two mechanisms in a sort of horse race would not solve the theoretical and empirical puzzle.

Here, we propose a different argument. We argue that retrospective incumbent punishment and the ‘rally around the incumbent’ effect may coexist after a terrorist attack. Both are reasonable reactions to these events. Either of them can predominate in a given case or moment, and they could even cancel each other out. Therefore, the effect of any terrorist attack on incumbent support may represent a sort of weighted average between a positive and a negative effect, which could add up to a positive, negative or even null overall effect. Even if this average was credibly identified, it would still be a black box from which it would be difficult to adjudicate between the different alternative theories.

This is why we need to move beyond the horse race. In order to do so, we propose that time is the crucial moderator that can help us in the identification of the two effects. There are reasons to expect the rally effect to be more short-lived than the accountability effect. Therefore, we expect a short-term positive effect of terrorism on incumbent support, which should fade away relatively quickly, but in the longer term, a retrospective voting logic would lead to a sanction of the incumbent.

A basic assumption of the literature, which we incorporate into our theory, states that the electoral effect of any event should decay over time. This is something that the literature on retrospective voting has long been aware of. In the economic voting research, this is labelled as the ‘myopic voter hypothesis’: voters tend to give more weight to more recent events than to more distant ones. The psychological underpinning of voters' myopia is the recency bias, according to which humans tend to weigh more recent events over previous ones, or the more extreme end heuristic, which is a tendency to substitute the end for the whole in retrospective assessments (Healy and Lenz 2014). However, despite the results from the myopic voter literature, in some cases, the accountability mechanism has also been shown to have the potential to last for a long time. Analysing the case of the 2002 Elbe flooding in Germany, Bechtel and Hainmueller (2011) show that voters in the affected areas rewarded the incumbent party for its policy response in both the 2002 and the 2005 elections.

In any case, even if the negative electoral effects of a terrorist attack are expected to decay over time, there are empirical and theoretical reasons to believe that the ‘rally around the flag’ mechanism will be more intense in the short term but will decay at a faster pace and, hence, be more short-lived. Theoretically, the mechanisms that underpin the rally effect are likely to be very circumstantial; both the psychological and the political mechanisms would suggest a short-term rally.

First, the sense of fear spurred by a terrorist attack is probably very transient (unless the attacks scale up). For instance, in his study of the Berlin Christmas market attack in 2016, Nussio (2020) finds that the negative emotions related to the event (sadness, anger and fear) were elevated for up to three days after the attack. Secondly, regarding the political drivers of the rally around the flag, after the initial moment of shock, the opposition briefly has strong incentives to distance itself from the incumbent and take electoral advantage of the crisis. Kam and Ramos (2008) explain the short duration of the rally effect with a social identity framework: the rally arises because the threat makes the nation state the salient in-group referent at the expense of the ‘lower’ groups (that is, political parties), bolstering allegiance to national symbols (for example, the flag, the president and so on). However, as the sense of threat dissipates, the position of the nation state as the dominant in-group also recedes.

Empirically, most research on the rally effect focuses on the short term. Whenever its duration has been evaluated, results tend to point in the same direction: the rally effect, when present, is very short-lived (Hetherington and Nelson 2003; Mueller 1970; Norrander and Wilcox 1993; Oneal and Bryan 1995; Perrin and Smolek 2009). The same short-lived rally effect has also been identified for the case of the COVID-19 crisis (Herrera et al. 2020). While the question of the duration of the rally effect may not be very important for the studies that explore attitudinal outcomes, it becomes central if we care about the electoral consequences of terrorism because elections are generally fixed or, at least, not determined by the attack.

If both the ‘rally around the flag’ and the accountability mechanisms are activated by terrorist attacks, and if the rally is more intense in the short run but more short-lived, then the observable implication that we can derive states that the time between the attack and the election will be the key moderator. Our prediction is that for those attacks that occur closer to the election, the rally effect will prevail and the average effect on incumbent support should be positive. However, conversely, for those attacks that occur at a more distant point in time, the retrospective voting mechanism will be more salient vis-à-vis the rally and an overall negative effect is to be expected. Some related empirical findings point to this important role of time in conditioning voters' response: the study of Kuijpers (2019) in ten countries shows that an increase in military casualties boosts the popularity of incumbent parties during the first year of a conflict but the effect turns negative for military interventions that started more than four years before.

It must be underlined that our argument is not contradictory with the well-established myopic voter hypothesis. Even if voters are near-sighted and, hence, punish the incumbent more intensely for recent than temporally distant attacks, an intense but short-lived rally effect can compensate this trend. The actual electoral toll of terrorism is, according to our argument, the result of both the retrospective voting logic and the rally effect. If the incumbent benefits from the rally around the flag shortly after an attack, then we may observe less or no overall punishment when the attacks occur close to the election, but we may see more punishment for attacks that occurred earlier, for which the emotional and political reactions that underlie the rally effect may have faded away.

An additional observable implication of our theory, which we test in the following, states that this temporal heterogeneity should be more pronounced the more threatening or detrimental for the social welfare of the overall population the attack is. That is to say, more consequential and threatening attacks should induce both a more intense short-term rally and a more intense punishment in the longer term. Next, we test these key expectations.

Empirics

The Case: Domestic Terrorism in Spain

Spain has a long history of domestic terrorism. Since the restoration of democracy in the late 1970s, several armed groups have been active in the country, making it one of the Western European countries that has experienced more intense political violence over the last four decades.

The main driver of this level of political violence was the Basque pro-independence leftist group ETA. From 1968 to 2008, ETA engaged in a long-lasting and often intense terrorist campaign, killing over 820 people. Most of the killings took place during the transition to democracy and the first years of the new democracy in Spain, but its activity extended over the years (Sánchez-Cuenca 2007). About 40 per cent of their victims were civilians, killed either in indiscriminate attacks or in strikes that specifically targeted politicians and other salient public figures. The share of civilian causalities increased over time.

ETA's tactics combined car bombings with direct shootings. Its most deadly attack occurred in 1987, with a car bomb in a shopping mall in Barcelona that killed twenty-one people and injured forty-five. Other deadly actions included attacks on military police headquarters in Madrid (twelve dead), Zaragoza (eleven dead) or the Catalan town of Vic (nine dead). Politicians were also a frequent target of ETA. In 1973, it famously killed the appointed successor of Franco, and in 1995, it attempted to kill the opposition leader and future Prime Minister José María Aznar. Lower-rank politicians (such as local councillors), public officials, police officers, businesspeople and even journalists or civilians were also targets of the terrorist campaign. Some kidnappings of politicians or prison officers also attracted massive attention and spurred mass protests.

The terrorism of ETA was a highly salient issue in Spanish politics for decades. Over the years, the state used many different tactics to fight the terrorist group, including the organization of paramilitary death squads (the GAL: Grupos Antiterroristas de Liberación). However, there were several attempts at finding negotiated solutions as well. In 1989, 1998 and 2006, different governments negotiated with ETA, coinciding with periods of ceasefires. None of these negotiations ended successfully, and the dominant political discourse about ETA tended to harden over time. Especially among the Right, the idea of any dialogue whatsoever with a terrorist group became increasingly stigmatized as morally unacceptable. The legislative and judicial strategies also hardened over time. ETA's alleged political branch (Batasuna) was outlawed in 2003, and two Basque nationalist newspapers were shut down under accusations of cooperation with ETA. Severely weakened by a changing political context and increasingly effective police action, in 2011, ETA unilaterally announced the end of its activities, with disarmament coming in 2017 and formal dissolution in May 2018.

ETA was by far the most active, deadly and politically salient group. However, it was not the only group operating in Spain since the transition to democracy. The second most deadly organization was the far-left GRAPO (Grupos de Resistencia Antifascista Primero de Octubre). The GRAPO was a Maoist-inspired group that followed a wave of revolutionary terrorism in Europe (for a comparative study, see Sánchez-Cuenca 2019). It was formed in 1975 and killed eighty-four people (twenty civilians and sixty-four members of the security forces) in over 1,000 strikes. It was mostly active during the years of the transition to democracy, but its last deadly attack was as late as in 2006.

Together with ETA and the GRAPO, which jointly account for over 86 per cent of all victims of domestic terrorism in Spain, various right-wing extremists and neo-Nazi groups have also been active in Spain during and after the transition to democracy. These groups account for almost seventy victims (De la Calle and Sánchez-Cuenca 2011).

Data

In this study, we present observational evidence on the electoral consequences of exposure to terrorist attacks. We estimate the effect of the degree of exposure to terrorism of each municipality in each term on the local vote share for the national incumbent in the next general election, using data on the nine general elections that were held between 1979 and 2008.

We combine these electoral data with the Domestic Terrorist Victims (DTV) dataset, which includes information on all victims caused by domestic terrorist groups in Western Europe for the 1965–2005 period (De la Calle and Sánchez-Cuenca 2011) and that we updated to 2008 for Spain. The DTV dataset is built at the victim level, and we aggregated it at the attack level. Therefore, our data include deadly attacks only. The data provide information on the location and date of the attack, the identity of the perpetrators and victims, and various other characteristics of the attack. Figure 1 shows the geographic distribution and authorship of the attacks included in our analysis, as well as their temporal distribution with respect to each election.

We proxy local exposure to terrorism using geographic distance to the attacks. In order to measure exposure, we geolocated the attacks and calculated the geographical distance between each strike and each municipality. The intuition is that for any given municipality, an attack occurring closer is more consequential than one that occurs at a more distant location (Braithwaite 2013; Fischhoff et al. 2003). This strategy allows us to capture variation across municipalities in the level of exposure to terrorism. The use of geographic distance to proxy exposure to terrorism is a standard practice in the literature (Bove, Efthyvoulou and Pickard 2021; Getmansky and Zeitzoff 2014; Huddy, Khatib and Capelos 2002; Kibris 2011; cf. Agerberg and Sohlberg 2021). We believe that in the context of a sustained terrorist campaign, such as the one we are analysing here, geographic proximity is especially likely to condition the political reactions to the attacks. Crucially, rather than an embedded assumption in our model, this is something that is subject to empirical scrutiny in our design: if the effects were constant across the whole country, we should not observe a significant effect of distance.

We work with a dataset at the municipality/term-attack level. For each municipality-term, the dataset has one row per every attack that occurred during that term. Specifically, as a measure of exposure, we use the geodetic distance between every municipality and attack. In addition, we have information on the date of the attack, the number of victims and whether it killed members from the military or police forces, politicians, and/or civilians.¹

As shown in Figure 1, a large proportion of the attacks were concentrated in the Basque territories for which ETA claimed secession (the Basque Autonomous Community and Navarre). This may be an important source of bias if we want to exploit distance to the attacks, as our measure would end up being, in large part, a measure of proximity to the Basque Country, which may correlate with other unobservable factors. This is the reason why in our preferred specification, we exclude these attacks and analyse only the effect of the attacks that occurred outside the Basque Country. However, our results do not depend on this decision, as discussed in Online Appendix B.

Figure 1. Geographical and temporal distribution of terrorist attacks in Spain, 1977–2008.

Source: Own elaboration based on the DTV dataset (De la Calle and Sánchez-Cuenca 2011).

Moreover, although we analyse all terrorist attacks, the lion's share of the phenomenon in Spain is linked to ETA, and regarding ETA activity, the political dynamics and reactions within the Basque Country are expected to be different. In the Basque provinces, there was always an electorally relevant pro-independence party that somehow supported ETA's activity, together with other pro-independence but pacifist parties. This is why we decide to exclude the Basque municipalities from our analysis, though in Online Appendix B, we show that our results are robust to their inclusion.²

Likewise, we exclude the municipalities from the Canary Islands, as their average distance with respect to the attacks is more than four standard deviations larger than the average distance with respect to municipalities in the Spanish mainland. The resulting dataset comprises a total of 1,660,383 dyads, with information on 223 attacks, 7,500 municipalities and nine elections. In these elections, three different incumbents ran for re-election ((the centrist Unión de Centro Democrático, the socialdemocrat Partido Socialista Obrero Español and the conservative Partido Popular) the centrist UCD, the social-democrat PSOE and the conservative PP). These changes in the incumbent party allow us to circumvent the collinearity between incumbency and partisanship that was often present in previous studies.

It is important to underline that our empirical approach and case selection make it possible to leverage variation in four relevant dimensions: exposure (as measured by geographical distance), timing, the characteristics of the attacks and the identity of the incumbent. This is one crucial difference between our article and most other published studies on terrorism and elections that focus on a single attack or a reduced and temporally concentrated set of attacks.

Empirical Strategy

If we regress local incumbent support on exposure to terrorism, there are two main threats to identification that we need to address if we are to provide a causal interpretation of the coefficients. First, there is the unobserved heterogeneity across municipalities: those localities that are more exposed to terrorist attacks will probably differ systematically from those less exposed in several relevant dimensions. This heterogeneity is likely to be expressed in the electoral results at t − 1, so we include this lagged variable in our models. More specifically, we include the results in t − 1 of the party that was incumbent at t as a key control for this heterogeneity.³ In some of our specifications, we also include municipality fixed effects, so we can account for any time-invariant heterogeneity across municipalities.

Another relevant concern may come from the strategic, and hence potentially endogenous, selection of locations to be attacked. If terrorist groups decided where to attack by taking into consideration the electoral evolution of the area, then our estimates would not have a causal interpretation. However, we believe that this is a minor concern in our setting and design because: first, our model specifications already control for most of the heterogeneity across municipalities that might have been taken into account by perpetrators in the selective targeting of localities; secondly, we use election results from the general election at the local level and exploit differential variation in incumbent support, so it is not obvious how attacking one area as opposed to another would impact the national result, which comes from the aggregation at the provincial and national level of party support in the Spanish proportional representation system; and, thirdly, if electoral calculus played a role in the selection of localities to attack, it was most likely a minor one, as operational concerns were arguably predominant.

With this strategy in mind, the baseline specification we use is the following:

(1)$$Y_{mti} = \beta _1Exp_{mi} + \beta _2Y_{mt-1} + \gamma _m + \lambda _t + \zeta _i + u_{mti},$$

where: Y _mti is the share of votes for the Spanish government incumbent party in municipality m in election t after attack i was perpetrated; β ₁ captures the effect of exposure (as measured by geographical proximity) of municipality m to attack i (the standardized reverse of the log of distance in kilometres)⁴; and γ _m, λ _t and ζ _i represent vectors of municipality, election and attack fixed effects, respectively, to account for any remaining heterogeneity across localities, terms and attacks that might affect the relationship between exposure to the terrorist strike and election results. Hence, we exploit variation over time within each municipality, net of potential municipal time-invariant and other temporal and attack-specific confounders. Finally, u _mti refers to the error term.⁵

In order to test our argument about the interaction between exposure and timing of the attack relative to the election, we interact exposure with proximity to the next election, as shown in the following equation, where T _i refers to the share of the term elapsed when attack i was perpetrated and β ₄ is, therefore, our main quantity of interest:

(2)$$Y_{mti} = \beta _1Exp_{mi} + \beta _2Y_{mt-1} + \beta _3T_i + \beta _4Exp_{mi} \times T_i + \gamma _m + \lambda _t + \zeta _i + u_{mjti}.$$

Finally, as we discuss later, in the mechanisms section, we include an additional term that forms a triple interaction between exposure, time and various characteristics of the attack, such as the type of victims or the method employed.

Results

In Table 1, we present the results for models based on Equations 1 and 2. The average effects of exposure to a terrorist attack are presented in Columns 1 to 3. Models 4 to 6 present the interaction with temporal proximity to the next election, measured as the share of the term elapsed when the attack occurred, varying between 0 if the attack was perpetrated the very same day of the previous election, and 1 if the attack occurred on the election day.

Table 1. Effect of exposure to attack on incumbent support (regression estimates)

Notes: Standard errors clustered by municipality in parentheses. * p < .05; ** p < .01; *** p < .001.

Models 1 to 3 show a consistent and strong negative average effect of exposure to terrorism on incumbents' vote share. Being closer to an attack by one standard deviation reduces the predicted incumbent vote share by about 0.3 percentage points. This result is robust to controlling for incumbent vote in t − 1, even after including election, attack and municipality fixed effects.

In order to facilitate interpretation of these coefficients, in Table 2, we print the average predicted incumbent vote share at different distances from an (average) attack. While an attack right on the municipality is associated with an average predicted vote share of 33.9 per cent, one that occurs 100 km away leads to a predicted incumbent vote of about 3 percentage points more, up to 36.5 per cent. Incumbents tend to do better in the less exposed parts of the country. The least exposed municipality in the table, located 1,000 km from the attack, has an expected vote share of 37.8 per cent.

Table 2. Effect of exposure to attack on incumbent support (predicted values)

Note: Estimates from Model 2 in Table 1.

However, the relevant test of our core argument lies in the interaction between exposure to terrorism and time. As we argued, rallying around the incumbent should be more intense in the very short run, while retrospective punishment should last longer. In Table 1, we show how the interaction between exposure to an attack and the temporal proximity to the election is clearly positive. This coefficient indicates that the negative effect of exposure to an attack decreases as elections approach, up to a point that it turns positive when the term is coming to an end.

Figure 2 represents the estimates of Model 5 graphically. It shows how the more temporally distant an attack is from the next election, the more harmful it is for the incumbent. However, conversely, exposure to attacks that happen in the last part of the term tend to be much less detrimental for the incumbent, and the effect is even positive in the temporal vicinity of the election.

Figure 2. Marginal effect of exposure to terrorism on incumbent support conditional on timing of elections.

Note: Estimates from Model 5 in Table 1.

In the right panel of Figure 2 we represent the interaction effect without the linearity assumption imposed in the left panel, following the recommendations in Hainmueller, Mummolo and Xu (2019). We cut the terms into four quarters, which correspond roughly to one year. As stands out quite clearly, although the pattern of decreasing sanction extends to the whole term, the main driver of the interaction is the attacks that occurred in the last 25 per cent of the tenure of the incumbent in office (typically, the last year). Exposure to those late attacks is clearly beneficial for the incumbent's electoral prospects. The figures also include a histogram showing the distribution of the observations across time, as suggested by Hainmueller, Mummolo and Xu (2019), to show common support for the estimates. There is no evidence that a handful of extreme cases is driving our results.

All in all, what we see is that there is a clear negative average effect of exposure to terrorism, which may be interpreted as supporting the retrospective voting argument. However, the temporal heterogeneity suggests otherwise. When the attacks occur when elections are close, they are beneficial for the incumbent. This result, we argue, can only be explained by a short-lived ‘rally around the incumbent’ effect that is able to offset and even reverse the negative effect of terrorism on incumbent vote in the short term.

Mechanisms: Types of Attacks

To provide further evidence, we also explore how the characteristics of the attacks moderate this conditional relationship. Not all terrorist attacks are equal; they vary in selectivity and type of victim. It is reasonable to expect more indiscriminate attacks and those that target civilians to be electorally more consequential. However, if our argument is correct, these attacks should be more consequential in both the positive and negative sense: they shall produce not only a more intense rally in the short term, as citizens may feel more vulnerable and emotionally impacted, but also a more intense punishment in the long run, as they represent a more significant negative shock to societal welfare. We therefore expect these attacks to weigh more in citizens' voting decisions.

In order to test whether the intensity of the threat moderates the relationship, we estimate models with a triple interaction between exposure, time and different measures of the severity of the attack: the type of victims (military/police, politicians/public officials or civilians) and the type of attack (selective versus indiscriminate). In Figure 3, we study how this conditional effect varies by type of attack. The left panel presents the estimates of the triple interaction with the type of victim, distinguishing between attacks killing civilians and attacks that killed politicians/public officials and members of the military/police forces. The right panel separates by type of attack: selective versus indiscriminate, as defined by the DTV dataset (De la Calle and Sánchez-Cuenca 2011).⁶

Figure 3. Marginal effect of exposure to terrorism on incumbent support conditional on timing of elections, by type of victim and type of target.

The results of these analyses suggest that those attacks that pose a stronger threat to the population at large, either because they target civilians or because they are indiscriminate, show a clearer pattern of temporal heterogeneity. They are both more harmful for the incumbent in the long run and more beneficial in the short term. This is especially true for exposure to indiscriminate attacks: if they happen during the last third of the term, they have a strong positive effect on the incumbent's vote.

We interpret this result as indicating that terrorist attacks that are more threatening for the population at large are more consequential in both directions. They generate both more negative retrospective voting and more rally around the incumbent in the short term. This is an additional piece of evidence in support of our argument and contributes to our understanding of the mechanisms: the sense of threat caused by terrorists in the population at large produces both a short-term reaction along the lines of the rally around the incumbent and a more durable negative update on the incumbents' ability.

Attacks on politicians follow the opposite pattern: they seem to harm the incumbent more if they occur late in the term. We interpret this result as suggesting that the killing of politicians does not activate the ‘rally around the flag’ mechanism; hence, the pattern we observe is produced by the standard, myopic retrospective voter. We can only speculate as to why this may be the case. There are two potential reasons: first, voters may identify less with the victim, especially in a context of very low political trust, such as Spain; and, secondly, politicians were typically killed in selective, not indiscriminate, attacks, which moderates the rally effect. These two factors may reduce the sense of threat spurred by an attack and, therefore, prevent the activation of the rally around the flag.

Ruling out Alternative Explanations

There are two alternative explanations that could explain our results. The first one is a story of differential memory: perhaps terrorism does only generate negative effects for the incumbent but we observe a stronger effect of exposure, as measured by geographical proximity, to attacks that are more temporally distant to the election because those attacks are only recalled by voters living in their vicinity. Conversely, if we observe a smaller effect of exposure to attacks that occur closer to the election, it may be due to the fact that those recent attacks are in the memory of all voters and, therefore, we cannot identify an effect of differential exposure as operationalized by geographical distance.

One first piece of evidence that would run counter to this alternative explanation is the positive marginal effect for recent attacks shown in Figure 2. Such a reversal of the effect is only compatible with some kind of rally around the incumbent. However, in order to rule out the alternative explanation of differential memory more convincingly, we explore one additional observable implication that allows us to adjudicate between our argument and the alternative explanation: the effect of time across different levels of exposure, as measured by geographical distance.

If the memory story was driving our results, then the effect of time should be more pronounced in those municipalities that are far from the place in which the attacks were perpetrated. As time goes by, memory of the attacks would be more enduring in the more exposed areas and more evanescent in the less exposed ones. Conversely, if the results in Figure 2 are explained, as we contend, by a combination of a more durable retrospective voting and a short-term rally around the incumbent, then we should expect the temporal heterogeneity to be stronger, not weaker, in the geographical vicinity of the attacks.

In Figure 4, we explore this possibility by analysing the marginal effect of temporal distance to the election across different geographical areas. Results provide support for our interpretation. In more exposed municipalities, the timing of the attacks relative to the election appears to be more important in determining levels of incumbent support. Put differently, the differential effect of the attacks that occur early in the term and those that occur closer to the election is mostly driven by the most, and not by the least, exposed municipalities.

Figure 4. Marginal effect of proximity to next election on incumbent support over exposure to attack.

Note: Estimates from Model 4 in Table 1.

A second alternative explanation that would be compatible with our results concerns the effect of the accumulation of attacks on their electoral consequences. By definition, attacks that occur later in the term come after a larger number of previous attacks, so timing and accumulation may confound each other.

This accumulation of attacks may reduce the accountability effect. The occurrence of terrorist attacks gives information to voters about the government's ability to prevent and protect the population. Therefore, each new attack is an opportunity for voters to learn about the incumbents’ quality. This would drive the accountability mechanism. However, as attacks accumulate, the marginal effect of an additional attack on voters' beliefs about the incumbent should be decreasing. That could explain why the negative effect of attacks closer to the election tends to dissipate: rather than time, the relevant factor would be the accumulation of attacks.

Secondly, one could think that the accumulation of attacks could have the opposite effect on the rally around the flag and exacerbate it. If the support for the incumbent stems from feelings of fear and uncertainty, the occurrence of several attacks could boost the rally around the incumbent. It is likely that citizens feel more threatened as the number of attacks increases given that a situation with frequent strikes is more dangerous than one with rare strikes.

If this was the case, the reason for a stronger rally effect for attacks that are closer to the elections would be, again, not the timing of the attacks, but their accumulation (that is, strikes close to the elections carry the cumulative effect of the previous ones).⁷

In order to test this alternative explanation, in Table 3, we present the marginal effect of exposure to an attack at the beginning and at the end of the term, controlling for an interaction between exposure and the order of the attack.⁸ This specification allows us to control for the confounding effect of the accumulation of attacks. We operationalize the order of the attacks in two different ways: first, as the order during the whole period of analysis, 1977–2008 (see Columns 1–3); and, secondly, as the count during the tenure of the sitting incumbent (see Columns 4–6). As Table 3 shows, the main results remain unchanged: despite controlling for the accumulation of attacks, we still see a clearly negative effect on the incumbent vote of exposure at the beginning of the term and a clearly positive effect if the attack occurs at the end of the term. The evidence, therefore, allows us to rule out the possibility that our results are driven by accumulation.

Table 3. Marginal effects of exposure to attack on incumbent support, controlling for accumulation of attacks

Notes: Standard errors clustered by municipality in parentheses. * p < .05; ** p < .01; *** p < .001.

Robustness Checks

In the Online Supplementary Materials, we present a number of robustness checks in order to provide additional support for our empirical findings. In Table B1, we replicate the main analyses with an extended sample of attacks, one that also includes all the attacks that took place in the Basque Country. This exercise increases the number of attacks from 223 up to 701. Despite the substantial increase in the number of observations, the results remain basically unchanged: the effect of exposure to terrorism on local incumbent vote is negative when the attacks occur early in the term and positive when they occur close to the election. Results are also similar if we add election results from Basque municipalities, increasing the number of municipalities from 7,500 to 8,022 (see Table B2).

In Table B3, we replicate the main analysis with a linear measure of exposure instead of the log distance used in the main specification. Again, this decision does not affect the substantive findings. In Table B4, we include a control for those municipalities that were directly hit in the attack. This allows us to further validate our measure of exposure based on geographic distance: it is not only capturing the influence of attacked municipalities, as exposure matters beyond direct affectation.

In Tables B5 and B6, we replicate our estimation controlling for the proximity of the municipality to Madrid and its interaction with the timing of the attack. Since a large number of the attacks outside the Basque Country occurred in Madrid or its surroundings, the distance from Spain's capital city could confound the effect of our measure of exposure. These tables show that the results remain unaltered.

In Table B7, we use as a dependent variable the share of votes for the incumbent as a percentage of the electoral census, instead of the percentage of the valid votes. This alternative operationalization allows us to account for the potential variation in turnout caused by exposure to terrorism. The results are also substantially unchanged, so our findings are not driven by changes in participation. In Figure B3, we also show that if we use an alternative non-linear measure of proximity to elections divided into quintiles, which is more aligned with the suggestion by Hainmueller, Mummolo and Xu (2019), our core conditional relationship holds: incumbents reap electoral benefits from attacks that occur close to the election day but are sanctioned for acts of violence that are committed earlier during the term.

In Table C1, we explore whether our results are consistent across the three different incumbents that Spain had during the period under analysis. The interaction between exposure and proximity to the election is always positive and significant. Hence, we can rule out that our findings are driven by a specific ideological reaction when the Left or the Right are in government. This test allows us to rule out that our results are merely a by-product of a conservative or progressive shift. Finally, in Figure C2, we explore potential heterogeneity by perpetrators and find that our main findings hold for both the attacks by ETA and those perpetrated by other organizations.

Concluding Remarks

Retrospective voting theory assumes that citizens will punish the incumbent in the ballot box if they experience a negative shock in their welfare. The standard expectation is that voters would hold the governing party accountable for deleterious events, even for ones that are beyond its actual responsibilities. However, even if this is the case for many political issues, events like terrorist attacks are likely to generate a harder case for accountability to work. When faced with an ‘external’ threat, political elites tend to cooperate across party lines and to call for national unity and unconditional support for the incumbent. This phenomenon, described as ‘rally around the flag’, is supposed to blur lines of responsibility and is likely to counteract the electoral effects of the accountability mechanism for the incumbent. Moreover, a sense of external, uncontrolled threat may induce a psychological reaction that would reinforce the ‘rally around the incumbent’. The literature on the electoral consequences of terrorism has so far found inconclusive results, and we have argued that this may be explained by this theoretical ambiguity.

In this article, we have proposed an argument to reconcile these seemingly contradictory findings. Terrorist attacks may activate both the retrospective logic and the ‘rally around the incumbent’ mechanisms. However, they have different rates of temporal decay: the rally around the incumbent, spurred by fear and temporal displays of unity by the opposition, tends to prevail in the short term, while the retrospective voting logic will be more dominant in the longer term given that the rally fades away quicker.

We have tested this argument using the case of Spain. This is a good test bed because we can have data for over 200 deadly attacks over nine elections and 7,500 municipalities. Therefore, we can leverage variation in exposure, timing, the type of attack and the partisan identity of the incumbent. Using a number of models with municipality, year and attack fixed effects, we have estimated the effect of exposure to terrorist attacks on incumbent vote share.

Our results show a negative average effect, in line with the retrospective voting argument. However, crucially, we also observe a strong interaction with time: those attacks that occur closer to the election not only are less detrimental, but can also positively impact incumbents' electoral prospects. This is especially true for indiscriminate attacks, which are expected to generate a more widespread sense of threat and fear among the general population. These results suggest that even if ‘politics as usual’ can be temporally suspended in favour of a rally around the incumbent, this is not sustainable in the long run. If enough time passes between the attack and the elections, voters will eventually hold the incumbent accountable for the violence they have been exposed to.

Overall, we have seen that the effects of terrorist violence on incumbent support are largely conditional on the context surrounding the terror attacks. In general, voters seem to sanction the incumbent when they are more directly exposed to terrorism, but our evidence also suggests that in the short term, a ‘rally around the incumbent’ effect can be found. Moreover, this is especially true for indiscriminate attacks and those whose victims are civilians. These findings have, we believe, important implications for our understanding of the political impact of terrorism in democracies. In the short term, terrorist attacks may unify the attacked country around its sitting incumbent, but in the longer term, terrorism has the potential to destabilize the ruling majority and erode its social support.