At roughly 9:20 am on 15 June 1996, members of the Irish Republican Army (IRA) drove a three-thousand-pound ammonium nitrate truck bomb into the center of Manchester, England. Eighty thousand people pack Manchester's department stores, restaurants, and cafés on a typical Saturday. IRA operatives parked the truck outside the glass display windows of Marks and Spencer, one of the largest department stores in the city. The stage was set for a bloodbath. But then the IRA did something peculiar: it placed phone calls to emergency hotlines and news media, giving the location of the bomb and the approximate time at which it would explode. Switchboard operators relayed the IRA's warnings to police, giving them ninety minutes to clear people from the city center. When the bomb went off, it tore through a ghost town. Flying rubble and glass laid waste to downtown Manchester, causing £1 billion in damage. But rescue crews combing the site afterward encountered no human bodies, only some shop mannequins blown from the store windows into the streets.Footnote 1
The IRA was eager to wreck the British economy, but it stopped short of mass murder. Why? Terrorism, by most definitions, involves attacking civilian targets.Footnote 2 Yet even as terrorists demolish businesses, transit systems, and other property, they sometimes choose to spare civilian lives. The terrorism and insurgency literature offers two broad theoretical frameworks to explain this choice. One framework emphasizes ideological variables (e.g., religion) that shape terrorists’ beliefs about the appropriateness of killing civilians.Footnote 3 A second framework emphasizes political and strategic factors that determine a militant group's need for civilian support—and the ease of obtaining that support. One variable relevant to the strategic/political framework is territory—whether militants possess strongholds or depend on the state's civilian population for shelter. Terrorists who shelter among civilians may restrain their violence to avoid offending their hosts.Footnote 4 Another relevant political factor is democracy: terrorists who fight democracies may eschew atrocities to avoid alienating civilians accustomed to political rights and the rule of law.Footnote 5
It is difficult to test ideological versus strategic/political explanations for casualty aversion, given the complex tactical settings where terrorism occurs. Body counts give an incomplete picture of whether terrorists meant to kill civilians because attacks produce unintended deaths. Target selection is also an imprecise measure. Groups like the IRA may attack crowded civilian areas without killing anyone. I consider a new measure of terrorists’ intent to avert civilian casualties: pre-attack warnings.
Terrorists who warn are explicitly announcing their intention to avert casualties, notwithstanding their decision to target civilian property. I present new data on 12,235 bombings by 131 terrorist groups during the years 1970 to 2016, coded by whether the perpetrators issued warnings beforehand. Fifty-nine of the groups in my sample gave warnings at least some of the time. Multilevel logit analyses, allowing random intercepts and random slopes for different terrorist groups, show that religion and other ideological factors have no significant effect on the incidence of warnings. Political and strategic considerations, specifically territory and democracy, do predict terrorists’ issuance of warnings. Groups fighting democracies are more likely to warn. Groups with territory are less likely to warn.
My findings cast doubt on purely ideological explanations for civilian victimization in terrorism, while supporting a political and strategic framework to explain terrorists’ use of casualty-aversion tactics. Additionally, my findings show the need to qualify claims from the civil war literature before applying them to the study of terrorism. For instance, my results appear to contradict civil war literature arguing that rebels with strongholds restrain their violence to win civilian support.Footnote 6 Yet it is important to consider the context and character of violence before importing rebel governance theories to terrorism. Rebels may eschew indiscriminate violence on territory they control, but terrorism is a mode of violence used against civilians on state-controlled territory.Footnote 7 The possession of strongholds enables militants to commit indiscriminate terrorism in state-governed areas without alienating their own supporters elsewhere. Terrorist groups without strongholds, who take shelter and draw support from the state's population, give more warnings to avert civilian casualties.
In what follows, I survey explanations for civilian victimization and humanitarian restraint in terrorism. I then present my warnings data set and my analysis showing that warnings are most common when terrorists lack territory and when they fight democracies. I discuss the implications for the literature on terrorism and civilian targeting. I also lay out potential reasons for my study's divergence from existing literature emphasizing religion, and from literature arguing that territory and democracy should have the opposite effects from what my analysis shows. These divergences demonstrate the importance of scope conditions and research designs tailored to the types of violence we observe in the specific context of terrorism.
Explaining Casualty-Aversion Tactics
My focus here is casualty aversion in terrorist tactics. Terrorism is defined in US law as “premeditated, politically motivated violence perpetrated against noncombatant targets by subnational groups or clandestine agents.”Footnote 8 I seek to understand why militants who attack noncombatant targets such as businesses, infrastructure, and government buildings adopt tactics that spare civilian lives.
Previous studies advance different explanations for why terrorists might make this choice. One school of thought emphasizes ideology, which conditions militants’ beliefs about the appropriateness of killing civilians. For instance, religion narrows militants’ political constituencies to ingroups only.Footnote 9 Religious loyalty runs deep and outgroups are generally unpersuadable. Religious terrorists may therefore believe it necessary to kill, displace, or subjugate outgroups to achieve their political goals. Additionally, terrorists who believe they are remaking society on God's behalf may show less hesitancy about killing bystanders.Footnote 10 Divine sanction may even justify the killing of coreligionists, as when jihadists attack the governments of secular, majority-Muslim states. Moreover, transnational religious movements provide social networks, enabling the transmission of deadly tactics such as suicide bombing.Footnote 11 Individually and jointly, these arguments imply:
H1 Religious terrorist groups are less likely to employ casualty-aversion tactics than are secular groups.
Another school of thought emphasizes the political and strategic relationships between militants and civilians. Mao Zedong describes militants as “fish” and civilians as “the water” that sustains them.Footnote 12 Militants rely on civilians for shelter, intelligence, recruits, and materiel. The politics of obtaining these resources depend on who controls the territory. The civil war literature suggests that territorial control incentivizes softer treatment of civilians. Rebels with territory become “stationary bandits.”Footnote 13 They find it easier to levy resources if they treat civilians humanely.Footnote 14 Territorial control also enables militants to use force more discriminately because civilians provide intelligence to identify state agents.Footnote 15
The implications of territory are somewhat different in terrorism, however. Terrorism takes place on state-controlled territory, not in militants’ strongholds.Footnote 16 Militants with strongholds can afford to use indiscriminate terrorism in state-controlled areas because their own supporters, located in the strongholds, are not directly affected.Footnote 17 Groups without strongholds, who shelter among civilians in state-controlled areas, may be forced to restrain their violence. These groups may adopt a less bloody style of terrorism, destroying property but sparing civilian life to avoid “eviction” by their hosts.Footnote 18 Another peculiarity of terrorism, contrary to the logic of civil war more generally,Footnote 19 is that the lack of intelligence in state-controlled areas need not preclude the use of discriminate violence. Even if terrorists cannot distinguish civilians from state agents, pre-attack warnings enable civilians to flee the area and escape harm. Thus, casualty aversion is both feasible and politically useful for groups without territory. A testable hypothesis follows from this discussion:
H2 Terrorist groups with territory are less likely to employ casualty-aversion tactics than are groups without territory.
Another strategic factor affecting casualty aversion is the existence of a democratic regime. Democracy promises several benefits that reduce aggrieved citizens’ support for atrocities committed in their names. These include “commitment to civil liberties, enhanced political access for dispossessed minority groups, institutionalization of nonviolent conflict resolution mechanisms, [and] improved human rights practices.”Footnote 20 The same liberal norms amplify citizens’ negative reactions when terrorists’ atrocities directly or indirectly affect them. As a result, “terrorism tends to move democratic publics to the right, supporting far more hawkish and far less conciliatory attitudes toward terrorists and the populations that support them.”Footnote 21 To mitigate the political backlash from killing civilians, terrorists who target democracies may employ tactics such as warning. Another hypothesis emerges:
H3 Terrorist groups fighting democracies are more likely to employ casualty-aversion tactics than are groups fighting nondemocracies.
Two other strategic variables relevant to civilian targeting are state sponsorship and access to illicit resource trafficking revenue. These variables may reduce militants’ dependence on civilians, enabling them to commit atrocities with fewer consequences.Footnote 22 Although state sponsorship and resource revenues are not my main variables of interest, I account for their possible effects with appropriate controls.
Research Design
I test my hypotheses using data on 12,235 bombings from the START consortium's Global Terrorism Database (GTD).Footnote 23 I focus on bombings of civilian targets because this mode of violence corresponds with most definitions of terrorism.Footnote 24 With the bombing as my unit of analysis, I seek to determine what makes the perpetrator group choose the tactic of warning beforehand.
I employ warnings as a dependent variable because of their straightforward role in averting casualties. Previous studies use other, arguably rougher, measures to discern terrorists’ intent to kill or spare civilians. Casualty counts are a common choice.Footnote 25 However, in the context of terrorism, casualty counts may not always reflect the underlying intent of the attacker. One does not employ terrorism in areas one controls, where one can strike targets with surgical precision.Footnote 26 Militants use terrorism “behind enemy lines,” inducing fear, destabilizing the state, and sapping the public's will to resist.Footnote 27 Operating clandestinely in state-controlled areas, lacking freedom of movement, terrorists who prefer to avoid casualties may have difficulty doing so.Footnote 28 Rather than counting bodies, some studies focus on terrorists’ decision to attack “low-casualty” or “high-casualty” civilian targets.Footnote 29 This raises other concerns. As the IRA's Manchester bombing shows, “high-casualty” targets such as shopping districts can be devastated without causing deaths, if warnings are given. “Low-casualty” targets such as utilities and infrastructure may be scenes of carnage if no warnings are given. Consider an al-Qaeda cell's 2004 Madrid train station bombings, killing 191 people, or the Colombian ELN's 1998 attack on a Segovia oil pipeline—a fiery incident that killed forty-eight.Footnote 30 To avoid such atrocities, groups such as ETA and the Niger Delta Avengers have given warnings before attacks on railways and oil pipelines.Footnote 31 These warnings indicate a deliberate effort to spare civilians at the scene. As a dependent variable, warnings offer a chance to disentangle the correlates of casualty-aversion tactics among militant groups that use terrorism against civilian infrastructure and property.
I determine whether pre-attack warnings were given by consulting the LexisNexis database and other secondary sources. Research assistants coded a dichotomous warning variable, based on a detailed search and coding protocol. The protocol provides terms to enter into LexisNexis's searchable archive of print, radio, television, and news wire reportage. Research assistants specified a date range for each search: one day before the bombing through two days after the event, maximizing the chance of “observing” the bombing and any warnings given. The search terms included the common names of the perpetrator group and the boolean terms “bomb!” and “warn!” If these searches failed to generate hits, research assistants widened the search by removing the term “warn!” and using any hits to confirm that no warning was given.Footnote 32 If LexisNexis contained no information on a bombing, research assistants used other sources, including the US State Department's yearly Significant Incidents of Political Violence Against Americans reports and the GTD's “summary” field, which contains descriptions of some incidents.
I filter the GTD data to include only incidents where the attack type (attacktype1) is given as “Bombing/Explosion” and target type (targtype1) is a category other than “military,” “police,” “violent political party,” or “terrorists/nonstate militia.” I further restrict the analysis to include only attacks by militant groups that carried out at least twenty-five bombings in the years covered by my data (1970–2016). This restriction excludes groups whose limited notoriety decreases the coverage their attacks receive in news sources used to code the dependent variable. A lower level of press interest (fewer stories, fewer words per story) raises the prospect of underreporting bias that would understate the prevalence of warnings by certain groups. The twenty-five-bombing criterion admits 131 groups and 19,310 bombings, of which research assistants were able to find and code 12,235. (See Appendix Table A1 for a list of groups.)
Measures
Dependent Variable: warning
This is a dichotomous indicator of whether the perpetrator group gave a warning before the attack. Warnings take various forms: telephone calls to emergency services or targeted businesses, verbal warnings to people at the scene, and notes or graffiti indicating the presence of a bomb. Another tactic counted as warning is the forcible removal of people from the scene of an attack—for instance, disarming a security guard and leading her away before a bomb is detonated. Vague threats are not counted as warnings. Examples include general promises of violence if the government does not change a policy. Categorical warnings are counted, even if they are somewhat general in character. Examples include promises to bomb polling places during an election and promises to attack universities on a particular day. Appendix Table A1 shows the percentage of each group's attacks that were preceded by warnings.
Fifty-nine of the 131 groups gave warnings at least some of the time. Certain groups (e.g., the Corsican separatist group Resistenza) warn all of the time. Others (e.g., the Islamic State) warn as little as 0.5 percent of the time. Among groups that gave some warnings, the mean and median frequencies of warnings are 6.7 percent and 5.6 percent, respectively. The sample average frequency of warnings, including all groups, is 5.3 percent.
Independent Variables
I operationalize the independent variables from my hypotheses using data from several sources. These include the Big, Allied, and Dangerous data set (BAAD, versions 1 and 2),Footnote 33 the UCDP External Support in Armed Conflict data set,Footnote 34 reports from Stanford University's Mapping Militant Organizations project,Footnote 35 and US State Department reports. Most of my independent variables take BAAD variables as starting points. However, the BAAD data sets cover only the years 1998 to 2012 and some groups in my analysis do not appear in the data. I use the UCDP, Stanford, State Department, and other secondary source data to code these groups and years. (See the appendix and replication data set coding notes for the sources used for each coding.) The independent variables are:
• territory—a 0/1 variable indicating whether a militant group possessed territory in a given year. An alternative specification splits this variable into domestic territory and foreign territory, based on whether the militant group's strongholds were inside or outside the territory of the group's primary target state.
• polity2—a −10 to 10 scale developed by the Polity IV project, describing a government's authority characteristics in a given year.Footnote 36 Positive values indicate that the government was mostly democratic; negative values indicate the reverse. Each bombing's polity2 value is that of the perpetrator group's primary target state during the year in which the attack took place. I define “target state” as the state the militant group is attempting to coerce. For instance, the target state of Hamas is Israel. The target state is usually but not always the state in which the attack took place. For instance, when the Basque separatist group ETA carried out attacks in France, it did so as part of a campaign to coerce Spain. ETA attacks are always coded with the Polity2 score of Spain. Appendix Table A1 shows each militant group in association with its target state.
• state sponsored—a 0/1 variable indicating whether the militant group benefited from state sponsorship in a given year.
• resources—a 0/1 variable indicating whether the militant group drew funding from the trafficking of illicit drugs, precious metals, gemstones, or oil in a given year.
A set of dichotomous variables indicates the ideology and political goals of each militant group: religious, ethnic, leftist, reactionary, and separatist. These variables are not mutually exclusive. A group like Hamas could be motivated by religion, separatism, and Palestinian ethnic grievances simultaneously.
Some specifications include suicide, a 0/1 GTD variable. This control variable accounts for the possibility that suicide tactics are used to increase surprise and bloodshed, objectives potentially incompatible with warning.Footnote 37 Other dummy variables indicate the type of target attacked. These are derived from the GTD's targtype1, recoded to reduce the number of categories:
• government—where targtype1 is “Government (Diplomatic)” or “Government (General).”
• civil society—where targtype1 is “Educational Institution,” “Journalists & Media,” “Religious Figures/Institutions,” or “NGO.”
• infrastructure—where targtype1 is “Airports and Aircraft,” “Maritime,” “Transportation,” “Telecommunication,” or “Food or Water Supply.”
• private citizens—where targtype1 is “Private Citizens & Property.”
• tourists—where targtype1 is “Tourists.”
• business–where targtype1 is “Business.”
• utilities—where targtype1 is “Utilities.”
Specifications controlling for target type omit attacks on “Other” and “Unknown” targets. (Estimating coefficients on these target types would mean very little.) Models controlling for target type use utilities as the base category.
Table 1. Descriptive statistics
Analysis
I estimate multilevel logit regressions for my dichotomous dependent variable, warning. The multilevel specification is appropriate because my data consist of 12,235 observations (bombings) drawn from 131 terrorist groups.Footnote 38 The multilevel model allows a different regression constant (random intercept) for each terrorist group, assuming that an organization's baseline propensity to warn for any bombing is partly driven by unobservable group-level characteristics.Footnote 39 I also allow different slope coefficients (random slopes) on polity2 and suicide for each terrorist organization, granting that these variables’ effects may differ across groups.Footnote 40 (Groups may respond differently to variation in polity2 and think differently about the utility of warnings for suicide attacks.) The decision to allow random slopes is based on a variable's theoretical importance and whether the random slope explains more variance than a simpler random intercept specification.Footnote 41 Likelihood ratio tests (see replication files) show that of all the variables in models 1 to 4, only polity2 and suicide explain more variance with random slopes; the model with random slopes on both variables explains the most variance.
Table 2 shows the results of several model specifications. Model 1 contains the ideological variables religious, ethnic, leftist, reactionary, and separatist. Model 2 contains the strategic/political variables territory and polity2, plus resources and state sponsored. Model 3 combines the variables of models 1 and 2, testing the explanatory power of ideology against the explanatory power of strategic factors. Model 4 controls for suicide attacks. Model 5 controls for target types. Models 6 and 7 probe conjectures: a potential interaction between religious and suicide attacksFootnote 42 and the possibility (raised in the civil war literature) that foreign and domestic territorial strongholds exert different effects on terrorists’ treatment of civilians.Footnote 43
Table 2. Multilevel logit results
Notes: Obs./group (min., avg., max.) Model 1: 1, 93.4, 1,962; Models 2–4: 1, 87.2, 1,852; Models 5–7: 1, 80.9, 1,709.
Results
The results in Table 2 are consistent with the strategic theories explaining casualty-aversion behavior based on terrorists’ territorial control and the state's regime type.Footnote 44 Religion and the other ideological variables are not significant predictors of whether terrorists warn. The coefficients on religious are always negative, but they are statistically significant only in Model 1, which does not control for territory, the state's regime type, or tactical factors such as target type. Controlling for these variables renders the effect of religious statistically insignificant. The failure to uphold H1 defies the conventional wisdom that religion encourages indiscriminate terrorism.Footnote 45 Also surprising, given the extensive literature on religious suicide bombing, is that Model 6's interaction between religious and suicide is not statistically significant. In fact, suicide's main effect is insignificant in all models. The negative correlation between warning and suicide is modest: −0.0692. Terrorists occasionally pair the two tactics, and fourteen (1 percent) of the 1,403 suicide attacks in my data were preceded by warnings. For instance, ISIS has warned before suicide attacks on Iraqi polling places. The bombers were still able to access the polls before detonating themselves.Footnote 46
H2 and H3, concerning territorial control and democracy, are supported. The coefficients on territory are negative in all model specifications and significant at a 99 percent confidence level. Territorial control is associated with a lower probability of warnings. Model 7 shows that the direction of the effect is not dependent on whether the territory is held within the target state's borders. Nor are the coefficients on domestic territory and foreign territory statistically distinguishable. This makes sense, since terrorism is generally used on territory the state controls—territory that is by definition “foreign” to the militant group's stronghold.Footnote 47 (My bombings are not coded for whether they took place in state-controlled areas, but given the unsuitability of bombings to “policing” one's stronghold, it is reasonable to assume that a significant majority took place in areas controlled by the state).Footnote 48 Terrorists with strongholds anywhere have fewer incentives to avert civilian casualties in state-controlled areas. Confirming H3, the coefficients on polity2 are positive and statistically significant at a 99 percent level in all specifications. Democracy is associated with a higher probability of warnings.
Given the emphasis on religion throughout the terrorism literature, the failure of H1 deserves further investigation. The Pearson correlation coefficients in Table 3 offer some intuition for why the introduction of territory and polity2 in models 3 to 7 renders the effect of religious statistically insignificant. The pairwise correlations are moderately strong, with religious positively correlated with territory and negatively correlated with polity2. Substantively, religious terrorist groups are more likely to hold territory and less likely to fight democracies. territory is negatively correlated with polity2, suggesting that nondemocratic state opponents are also relatively poor at denying militants control of territory. Although religious groups are less likely to give warnings (the frequencies for religious and secular groups, respectively, are 1.7 and 9.5 percent) a regression model accounting for territory and polity2 attributes much of the apparent effect of religious to these latter variables. One can tell a plausible story that the association between religion and no-warning attacks is likely driven by the fact that many religious terrorist groups arise from (or come to fight) nondemocracies with tenuous control over their territory—countries like Afghanistan and Syria. Indiscriminate violence has fewer consequences for such groups, not because of their religion, but because their illiberal political environments and access to strongholds render them less accountable for victimizing the state's civilians.
Table 3. Pearson correlation matrix
Marginal Effects
To understand these results substantively, we can examine the predicted probabilities of warning implied by model 5. Setting all variables to 0 gives a predicted probability of 0.9 percent. Holding all variables at their means (see Table 1), the probability of warning is 2.6 percent. Substantively, this resembles some notable cases in the data. Turkey's PKK and Somalia's Shabaab have average territory and polity2 values close to the respective sample averages of 0.63 and 5.37. The PKK gave warnings for 3 percent of its 241 attacks; al-Shabaab gave warnings for 2 percent of its 394.
Figure 1 shows the average marginal effect of independent variables in model 5. (The effects of the target types are not displayed.) The significant binary predictor of warning, territory, decreases the probability of warning by 2.9 percent—from 6.7 percent for groups without territory to 3.8 percent for groups with territory. One group of interest, Boko Haram, effectively splits the difference, holding territory for 53 percent of its attacks, with an average polity2 score of 5.52, nearly equal to the sample average. Boko Haram gave warnings for 4.6 percent of its 459 attacks—2.1 percent when it held territory and 7.4 percent when it did not.
Figure 1. Average change in predicted probability (95 percent confidence intervals shown)
The average marginal effect of increasing polity2 by one category is to increase the probability of warning by 1.1 percent. Holding all other variables at their means, the probability of warning moves from roughly 0.5 percent at polity2=−10 to 10 percent at polity2 = 10. Greater changes are possible when polity2 and territory move simultaneously. At polity2=−10 and territory = 1, the probability of warning decreases to 0.4 percent. These realizations of the variables approximate those for the Free Syrian Army. With an average territory value of 1.0 and an average polity2 score of −8.95, the FSA gave warnings for none of its thirty-eight attacks. Looking at the opposite extreme, the predicted probability of warnings rises to 13 percent at territory = 0 and polity2 = 10. These territory and polity2 values approximate the average realizations for several organizations: Turkey's Dev Sol, which gave warnings for 13 percent of its thirty-two attacks; Greece's November 17 Revolutionary Organization, which gave warnings for 11 percent of its thirty-seven attacks; and the Puerto Rican FALN, which gave warnings for 9 percent of its ninety-two attacks. Of course, some groups with similar territory and polity2 profiles have higher frequencies of warning. ETA never held territory and fought a state with an average polity2 value of 9.9. The group's frequency of warnings was 45 percent in its 309 attacks. Likewise, the IRA never held territory and had an average polity2 value of 10. The group gave warnings for 44 percent of its 360 attacks.
Target selection partially accounts for these groups’ behavior. For example, fourteen of ETA's attacks (roughly 5 percent) targeted tourists. This target type is the most likely to receive warnings, with 48 percent of all tourism-related attacks preceded by warnings. ETA's “summer campaigns” against Spanish tourist beaches exemplify the group's relatively bloodless economic attrition strategy.Footnote 49 ETA buried bombs in the beaches, but it gave warnings 86 percent of the time and the bombs caused few injuries. However, an economic analysis finds that on average, each bombing dissuaded roughly 140,000 tourists from visiting Spain.Footnote 50
Even accounting for target selection, ETA is still something of an outlier. The group's 43 percent frequency of warning in nontourism attacks is also remarkably high, even for a group without territory, fighting a democracy. To some extent, ETA's behavior is driven by group-level idiosyncrasies. At the other end of the spectrum, groups like ISIS and Colombia's FARC warn less often than one might expect, given their average values of key independent variables. The groups’ respective averages of territory are 0.99 and 1.0, giving us a generally low expectation of warnings. But the groups’ polity2 averages are moderate, at 5.14 and 7.18. At these values of polity2, for groups with territory, we would expect warnings at rates of 2 and 4 percent. Yet ISIS and FARC give warnings at rates of only 0.5 and 0.7 percent. The heterogeneity among groups underscores the importance of a multilevel model accounting for differences in groups’ baseline propensities to warn and intergroup variation in the effects of key variables like polity2.
Robustness
Robustness checks in the appendix confirm that my findings are not altered by respecifications of the logit model. Table A5 controls for the possibility that some attacks take place in unoccupied spaces (e.g., retail businesses “after hours”) and for press freedom in the country where each attack occurred (because less press freedom may imply underreporting of warnings). Table A5 also includes a model with a quadratic polity2 termFootnote 51 and a model with an alternative polity2 measure based not on the terrorist group's target state but on the polity2 score of the country where the attack occurred. Table A6 replaces polity2 with alternative measures of democracy: the Political Terror Scale (PTS),Footnote 52 Freedom House civil liberties (CL) and political rights (PR) scores, and the democracy (democ) and autocracy (autoc) component scores of Polity2. Table A7 adjusts the threshold for inclusion in the analysis, omitting groups that carried out fewer than fifty, seventy-five, or one hundred attacks. Table A8 respecifies the multilevel model as a logit with fixed effects for each terrorist group. None of these respecifications substantially alters the effects or significance levels shown in Table 2, Model 5—nor is the quadratic polity2 term significant.
Table A9 tests for additional interactions associated with the variable religious. Perhaps H1 fails because religion's effect operates through an interaction with another variable such as territory or separatist. A religious terrorist group may become more brutal if it is also separatist, carving out a “pure” society from an impure world. Or perhaps a religious group kills more civilians if it achieves territorial control and begins to enact God's justice. Neither of these interactions is statistically significant. The results in model 5 are quite robust.
Discussion
Although many militant groups terrorize civilians indiscriminately, others take a different approach, targeting civilian property and employing tactics that reduce the likelihood of casualties. This study considers which variables incentivize the latter strategy. Rather than focusing on body counts or target types, I use pre-attack warnings as indicators of whether terrorists attempted to spare civilians in their attacks. My analysis suggests that terrorists employ casualty-aversion tactics for strategic reasons, particularly when they depend on civilians for shelter and when the state is a democracy. Religion and other ideological factors show no significant effect on the propensity to warn.
These findings have implications for how we understand terrorism and how we pursue terrorism research. The importance of scope conditions becomes especially clear. Terrorism is a distinctive variety of conflict behavior. Theoretical expectations derived from the broader civil war literature do not necessarily “travel” to the terrorist context. For example, I find that terrorists without territory are particularly likely to give life-sparing warnings before they attack. At first glance, this finding appears to contradict previous research arguing that militants with territory commit fewer abuses against civilians.Footnote 53 To collect funds, war materiel, and recruits, it is better to earn civilians’ loyalty by offering policy goods rather than using violence.Footnote 54 However, it is important to consider the context and character of violence before extrapolating this argument to terrorism. In fact, the civil war literature itself suggests the need for such scope conditions.
Research on civil war shows that militants use force discriminately in areas they dominate.Footnote 55 They prefer “direct” weapons such as small arms over “indirect” and potentially indiscriminate weapons such as bombs.Footnote 56 Yet my concern is specifically with terrorism and bombings. This type of violence is used almost exclusively in state-controlled areas, against state-governed civilians.Footnote 57 Thus, there is no reason to expect the logic of responsible rebel governance to apply.Footnote 58 Rather, we should expect militants with territory to be significantly more brutal. They are terrorizing the state's civilians, not the civilians in their strongholds. Militants without territory have stronger incentives to eschew indiscriminate terrorism. They depend on the state's civilians for shelter, and these civilians can “evict” them if they engage in indiscriminate no-warning attacks.
Context may also explain my study's divergence from previous research on religious terrorism. The religious terrorism literature is largely focused on Islam, and my sample of terrorist organizations includes the highest-profile jihadist groups, as measured by attack counts.Footnote 59 So why does the conventional wisdom about indiscriminate religious terrorism not bear out in my results? One reason appears to be that religious and secular terrorist groups come from very different political environments—Syria as opposed to Spain, Afghanistan as opposed to Ireland. On average, religious terrorist groups’ adversaries are less democratic and less effective at denying militants territory. Many religious terrorist groups can expect to be judged by low humanitarian standards, owing to the lack of liberal democratic norms. Many can take shelter and draw support in strongholds they control. The political costs of killing civilians in state-controlled areas are relatively low for these groups. It is no surprise, then, that after controlling for territory and democracy, religion shows no significant effect on terrorists’ propensities to warn.
However, religion may exert significant effects on modes of violence other than terrorism. For example, religious militants may govern with brutality. In the 1990s and early 2000s the Afghan Taliban were notorious for publicly mutilating and executing criminals.Footnote 60 Many of the most spectacular abuses we associate with ISIS are similar “policing” events—public executions of the caliphate's own subjects and state forces captured on the battlefield.Footnote 61 If anything, these incidents represent state terrorism or proto-state terrorism, phenomena outside of my scope. It is certainly worth examining the enabling role of religion in these cases, with research designs appropriate to the specific type of violence.
I also find that terrorists fighting democracies are more likely to give life-sparing warnings. This finding contributes empirical evidence to a long-running debate about whether democracy incentivizes or disincentivizes terrorism.Footnote 62 One body of literature argues that democracy creates a poor environment for terrorism because citizens come to value peaceful dispute resolution and the rule of law.Footnote 63 A contrasting body of literature argues that democracy encourages terrorism—because civil and political liberties enable terrorist organizing, because voters are perceived as guilty for the government's crimes, and because the public has the political means to induce policy concessions if terrorists inflict casualties on them.Footnote 64 Contradictory empirical findings may arise from the use of different research designs and dependent variables: the number of terrorist groups in a country,Footnote 65 the count of terrorist incidents,Footnote 66 fatality counts,Footnote 67 or the group's choice to attack “low-casualty” or “high-casualty” targets.Footnote 68 Each of these research designs addresses a different question about what democracy does: promoting terrorism's emergence, increasing terrorism's frequency, or increasing the level of bloodshed. My research question is nearest to the latter, examining terrorists’ adoption of casualty-aversion tactics, specifically warning. My findings support a more optimistic interpretation of democracy as restraining terrorism, in this case, incentivizing tactics that spare civilian lives while destroying property. The results are robust to controlling for target type. They are also consistent with the theoretical argument that militants restrain their violence when appealing to citizens accustomed to good governance and nonviolent political contestation.Footnote 69
This analysis contributes to several debates in the terrorism and insurgency literature. It elucidates the potential roles of territory and democracy in shaping terrorists’ tactics. In doing so, it highlights the complexities of empirical terrorism research. My hope is that this discussion reinforces the move toward creative data collection and research designs that directly address the particularities of terrorism as a subject.
Data Availability Statement
Replication files for this article may be found at <https://doi.org/10.7910/DVN/SZA9LU>.
Funding
This research was funded by the Horowitz Foundation for Social Policy and the Earth Institute at Columbia University.
Acknowledgments
Previous versions were presented at annual meetings of the International Studies Association (2017, 2018, and 2019). The author thanks Robert Jervis, Page Fortna, Shuai Jin, Tanisha Fazal, Victor Asal, Friederike Kelle, Costantino Pischedda, Mia Bloom, Paul Kowert, Luis Jimenez, Stacy VanDeveer, Samuel Barkin, Gary Ackerman, Peter Krause, Severine Autesserre, and Austin Long for their helpful comments. For their research assistance, the author thanks Nidale Zouhir, Nathan Akin, Uluç Karakaş, Vaibhav Sabharwal, Osama Mohamed, Mallory Rosso, Vivian Grossman, Gi Jae Han, Addy Sonaike, Antoine Sander, Katie Garcia, David Ray Anderson, Fatima Dar, and Kangdi Li.
