Abstract
This article offers a theory of rebel targeting in civil war. Rebels face two logics of targeting: the logic of compliance in areas where they are fighting for territory, and the logic of legitimacy in areas where they remain clandestine. The first boosts civilian targeting because rebels use it as a way to discipline local populations. The second moderates it because unqualified civilian killings could prompt rebels’ supporters to turn against the group and easily help bring down its clandestine structures. In order to avoid this backlash, rebels focus on attacking ‘legitimate’ targets such as security forces and authorities. This supporters’ constraint is mitigated when the state heavily represses the rebels’ constituency – which increases supporters’ appetite for indiscriminate retaliation – and when the rebels can rely on independent funding – which reduces rebels’ dependence on supporters’ help. I test the argument with detailed information about rebel presence and political violence in Peru from 1980 to 1995. The results support that civilian victimization is driven by rebel strength, since the more control the rebels hold, the more civilians are targeted. Thus, when Shining Path was forced to operate clandestinely on a permanent basis, it carried out more actions against hard targets, so as to remain a legitimate force within its urban support base.
Introduction
Two of the most infamous episodes of violence perpetrated by Shining Path (Sendero Luminoso, SP henceforth) during the Peruvian conflict involved civilians. On 3 April 1983, SP cadres executed 69 civilians in Santiago de Lucanamarca, a backwater district in the Andean region of Ayacucho. Apparently, the massacre was a punitive raid in the face of decreasing support for the rebels in the area. Nine years later, on 16 July 1992, an SP urban cell planted a no-warning car bomb in Tarata Street, inside the well-off Lima district of Miraflores, which took 25 civilian lives. This time, the action aimed to raise the cost for the wealthy of collaborating with the newborn regime led by President Fujimori. Despite being a highly centralized organization where the long-serving leader, Abimael Guzmán, tightly controlled every dimension of the military strategy, SP leaders largely condoned the first action, while censuring the second. In the words of Abimael Guzmán, ‘Lucanamarca was an excess, whereas Tarata was a mistake’ (Guzmán, 2002).
Why should two similarly vicious attacks against civilians receive such different interpretations from the insurgent leadership? Most irregular wars are fought on two fronts – the rough terrain of the countryside, where rebels aspire to gain territory, and the cities, where state power is strongest and rebels usually remain clandestine – and they do not necessarily involve the same warfare dynamics (Kalyvas & Balcells, 2010; Lockyer, 2010). Most models and theories on civilian targeting either make no distinction or simply interpret the urban fight as epiphenomenal.
In this article I offer a novel theory of target selection in civil war. I claim each theater of conflict has its own logic of targeting: the fight in the countryside is usually driven by the search for territory and civilian compliance; the fight in the cities is usually driven by the search for legitimacy and civilian support. The former puts a premium on territorial control, and civilian loyalty can be seen as much a consequence of control as a cause. The latter focuses on gaining popularity in the areas under safe state authority, because in these areas rebels depend upon supporters to collect information on targets, offer hideouts, and contribute resources.
Rebels are therefore constrained by their military force in areas where they fight for territory, whereas they are constrained by supporters’ preferences in areas where they fight for legitimacy. Civilian targeting will ultimately vary depending on this dual nature of conflict, with attacks against civilians being more probable in areas where rebels rely mainly on their military strength than in areas where rebels rely on ‘legal’ supporters to operate. As long as supporters have more moderate preferences on civilian targeting than foot soldiers, these rebels will have to select their victims more carefully in urban areas than in the countryside. As a consequence, civilian targeting should be more pervasive on the rural front than on the urban front.
The moderating effect of supporters on rebels’ targeting patterns may be mitigated when the state heavily represses the potential support constituency for the rebellion or when the rebels can rely on autonomous sources of funding, such as lootable resources. In the first case, supporters call for revenge against state raids, which allows the rebels to prey on civilians supporting the other side. In the second scenario, rebels do not rely on supporters to gain resources to fuel the fight and therefore supporters’ constraints are less binding.
I test these propositions with district-year data from the Peruvian civil war, adding up to around 30,000 observations (1,800 districts in 16 years). This conflict spanned from 1980 to 1995 and centered on confrontations mainly between the Maoist SP and the Peruvian military. All datasets code the Peruvian case as a civil war due to its lethality and the existence of rebels liberating some areas of the country. Still, SP also ranks high in datasets on terrorist events due to its operations in the urban areas – mainly in the Lima province. I argue here that the dynamics of the Peruvian conflict illustrates quite well some trends of violence that are visible in most contemporary civil wars.
Because my theory speaks to the scale of violence as well as to the distribution of targets, I use two different statistical methods. I first explore the causes behind some districts having more attacks against civilians (measured in absolute numbers) than others. Second, I compare for every district the share of attacks against civilians with the share of attacks against authorities and security forces, as a way to investigate whether civilians are more commonly hit than other targets.
The results by and large support that civilian victimization is driven by rebel strength, since the more control the rebels hold, the more civilians are targeted – in absolute as well as relative terms. Remarkably, when rebels are forced to operate clandestinely on a permanent basis, they carry out more actions against ‘legitimate’ targets, so as to not jeopardize their support base. These constraints seem to be less binding when the state represses heavily and the rebels have access to lootable resources to finance the insurgency, although evidence in this regard is less compelling.
The rest of this article is organized as follows. In the next section, I lay out the theory. Then, I suggest a new way to operationalize territorial control by looking at the capacity of the rebels to boycott local elections, and describe the dataset. After that, I run several statistical models to account for target choice. A discussion on the generalizability of the main findings closes the article.
A theory of rebel targeting
This article is about rebel targeting in conflicts where the balance of power between the rebels and the state clearly favors the latter. ‘Balance of power’ usually captures the ratio between the resources rebels arrange and those in the hands of the government. Typical resources are troop strength, weaponry, and legitimacy (Lockyer, 2010).
Warfare in government-advantaged conflicts has been identified as ‘irregular’ (Kalyvas & Balcells, 2010) and ‘guerrilla’ (Lockyer, 2010). In guerrilla warfare, rebels aim to conquer power through harassing state forces and seizing hitherto state-controlled territory wherein they become the rulers – typically imparting justice and raising taxes. Usual tactics are ambushes, town raids, and improvised explosive devices (IEDs). Guerrilla fighting has consistently been the most common warfare strategy since World War II (56% of all insurgencies), and it is on the rise – 70% of all insurgencies during the last decade (Jones & Johnston, 2013).
The main theoretical contributions to the study of targeting in guerrilla wars have relied on a balance-of-power approach, with two apparently contradictory hypotheses. On the one hand, Kalyvas’s influential theory places most civilian targeting in the areas where rebels are stronger (but not the strongest), since they can seize and hold territory there and rule over local dwellers. Violence against civilians is used as a mechanism to impose compliance (Kalyvas, 2006; Kalyvas & Kocher, 2009). In the same vein, some authors have found that rebels whose capabilities increase because of foreign transfers also rely tactically on one-sided civilian violence (Hovil & Werker, 2005; Wood, 2014).
Conversely, it has been emphasized that civilian victimization is driven by the decreasing capabilities of rebels, since they substitute lower armed strength with more intimidation (Eck & Hultman, 2007; Wood, Kathman & Gent, 2012). These two approaches can be aligned if we assume that shifts in the balance of power between the rebels and the state prompt the former to update their targeting scheme. Thus, when rebels increase their territorial sway, they force civilians to comply with the new rulers; and when rebels are on the losing side, they turn against civilians as a way to coerce them into obeying the group and supporting its goals.
This research generally focuses on the fight between insurgents and the state in the countryside. Thus, its predictions are about the areas where the rebels are most of the time fighting for gaining a territorial foothold – usually in sparsely populated, border provinces. The balance-of-power approach, however, remains largely silent about the patterns of victimization we should observe in areas remaining under safe state control – quintessentially the capital cities. According to Kalyvas’s theory, rebel violence in the state strongholds should be indiscriminate, because the militants do not have the ability to gather good evidence against potential targets and successfully vet them. By the same token, the weakening hypothesis should predict high levels of civilian killings in areas where the rebels cannot but resort to the last available tactic. Still, it makes little sense, from a strategic standpoint, to expect that clandestine groups will alienate supporters through unqualified attacks against civilians.
Groups forced to operate most of the time in areas under safe state control heavily depend on supporters to gather resources – funds, intelligence, and legitimacy. In this type of conflict, supporters always have the exit option – they can turn to the state at any time – and they usually embrace a more restrictive view on potential targets than recruits (Blair et al., 2013). Using Zhukov’s terminology, civilians may balance in urban contexts – that is, supporting the side with a more careful approach to civilians – whereas they tend to bandwagon in rural settings – that is, joining the ranks of the winning side, regardless of its behavior toward civilians (Zhukov, 2013). As a consequence, militants must be very careful about picking out targets, since unqualified attacks could drain urban support for the militants’ goals (Condra & Shapiro, 2012; Jaeger et al., 2015).
Given this legitimacy constraint, rebels choose targets that are palatable for their constituency base, with security forces and authorities being the most typical candidates. Civilians may also be targeted, but only when the victims have taken a public stance against the terrorists or belong to groups siding openly with the state.
Anecdotal evidence endorses this legitimacy constraint, since several terrorists have written about the critical role of supporters to sustain the fight (Alcedo, 1996; Collins, 1997; Mouesca, 2006). Al-Qaeda’s current mastermind, Ayman al-Zawahiri, illustrates nicely this point. In a 2005 letter to Abu Masab al-Zarqawi, al-Qaeda’s then chief of operations in Iraq and open perpetrator of civilian massacres, al-Zawahiri complained to him that indiscriminate attacks in Iraq were alienating the population. The edict admitted no ambiguity: ‘the mujahedeen movement must avoid any action that the masses do not understand or approve of’ (quoted in Bergen, 2006: 366).
In a more systematic fashion, recent work by Heger (2015) and De la Calle & Sánchez-Cuenca (2013) has found that terrorists who participate in the political process avoid civilian targeting in order not to jeopardize public support for their cause. As a matter of fact, research on the target profiles of clandestine groups in Western Europe also shows that civilians rarely make up the largest proportion of victims of terrorism, the exception being extreme right-wing groups largely unconcerned with gathering support (De la Calle & Sánchez-Cuenca, 2011).
Guerrilla conflicts traditionally see most of the rebel violence concentrated in areas far from the capital city, which is usually the government’s powerhouse (Buhaug, 2010; Butcher, 2015). And yet, most insurgencies in conflicts where state forces are overwhelmingly dominant try to maintain a presence in the two theaters of war – the countryside and the capital city. They fight for territory in the rural areas of the country, but they also seek to rally supporters and expose the vulnerability of the state in its city strongholds. In the countryside, targeting is driven by compliance; in the cities, it is driven by legitimacy. It is not that rebels do not care for legitimacy in the countryside – rebellions sometimes break out in areas where insurgents could anticipate some level of support for their cause (Holtermann, 2016). The issue is that, regardless of the distribution of popular support, rebels anticipate that legitimacy usually goes for the actor ruling the area, rather than the other way around (Cohen, 2014; Kalyvas & Kocher, 2007). In this sense, targeting is driven by territorial concerns, rather than by precautionary considerations.
It follows from this that civilians should be more often killed, and in a larger proportion compared to other targets, in areas where the rebels are fighting for territory rather than in areas where they are fighting for hearts and minds. I derive two testable implications from this discussion: Implication 1: Rebel violence against civilians should be concentrated in zones where rebels are stronger than state forces (but not the strongest). Implication 2: Rebel violence against civilians should be lower in cities, where the state is usually strongest and rebels have to operate most of the time under clandestine conditions.
Supporters’ constraints on civilian targeting could be overrun when recruits have autonomous sources of income (Hovil & Werker, 2005; Wood, 2014). According to Weinstein, armed groups with a larger share of bad recruits produce more civilian victims. The rationale is that bad recruits are driven into the insurgency by greedy considerations, which facilitate more predatory behavior against civilians (Azam & Hoeffler, 2002). Good recruits, to the contrary, are driven by ideology, restraining themselves from attacking civilians unless the leadership demands so (Weinstein, 2007). We should then observe that rebels target civilians more systematically when they do not depend on supporters to fund the fight. Implication 3: Rebel violence should be larger when rebels draw resources from non-supporters.
Finally, a theory of rebel targeting must consider the behavior of state security forces. Because the supporting constituency of the insurgency tends to manifest less radical preferences on the use of violence than rebel cadres, the latter must refrain themselves from boundless targeting. However, rebel groups sometimes fight on behalf of a constituency that is so hardened by the conflict that it endorses any type of civilian retaliation. Hardcore supporters are usually those with experiences of repression (Irvin, 1999). When the state circumvents the laws and triggers aggressive tactics against potential supporters of the insurgency, these groups could call for retribution against civilians siding with the state (Bhavnani, Miodownik & Choi, 2011; Fielding & Shortland, 2012; Lyall, Blair & Imai, 2013; Ricolfi, 2006). Implication 4: State repression feeds civilian victimization by the insurgents.
Before ending the section, I would like to briefly discuss the distinction between selective and indiscriminate violence. I think this distinction is unquestionably difficult, on at least three bases. First, the purely indiscriminate attack rarely takes place, since there is always a measure of selection in targeting. For instance, collective massacres, a quintessential example of ‘indiscriminate’ violence, may make sense when ethnic militias aspire to weaken rival ethnic groups’ sources of funding (Fjelde & Hultman, 2014). Car bombs planted in wealthy neighborhoods usually aim at threatening key groups supporting the state. The scant evidence on urban attacks against civilians show that most instances are either selective, driven by anti-rebel behavior, or collective, carried out as a way to punish communities (De la Calle & Sánchez-Cuenca, 2013; O’Duffy, 1995; Ron, 2001). Second, evidence on the counter-effects of non-selective violence is mixed (Lyall, 2009). Finally, there is always a subjective dimension when adjudicating between rebels’ motivations to select targets: some groups commit mistakes, whereas others pretend to have hit the right person, regardless of the public information about the victim. For all these reasons, I henceforth assume that most rebel groups do not select targets randomly.
Background
The civil war in Peru started on 17 May 1980, when several SP cadres broke the ballot boxes in Chuschi, a small town in Central Ayacucho. Driven by a Maoist ideology, SP took advantage of decades of failed sociopolitical mobilization to redress the political and economic grievances of the indigenous peoples of the Andes, with the region of Ayacucho as its stronghold (Comisión de la Verdad y Reconciliación [CVR henceforth], 2003: volume VIII; Degregori, 2011; Koc-Menard, 2007).
The initial reluctance of Fernando Belaúnde’s administration, the first democratic government since 1968, to grant exceptional powers to the military contributed to quickening the growth of SP. When the army was called in, in 1982, its tactics of collective repression further reduced state legitimacy in the area (McClintock, 1998). After consolidating its strongholds in the countryside, SP quickly moved toward the Lima-Callao conurbation, site of one-third of the country’s population. SP was able to set up urban commandos there that carried out a permanent terrorist campaign against politicians and security forces in the safest areas of the country (Burt, 1998).
During Alan García’s term (1985–90), the army received full power to deal with the rebellion, declaring the government numerous provinces under emergency law (Palmer, 1995). Although damaged by the spread of self-defense militias in the rural areas (Degregori, 1998; Starn, 1998), SP launched the ‘final offensive’, which pretended to encircle Lima to force its downfall. Despite the campaign, SP started to lose ground in the rural areas where its increasing reliance on massacres to counter the self-defense squads revealed its shrinking armed capabilities.
On 5 April 1992, Alberto Fujimori, president since July 1990, dissolved Congress and empowered the military with legal backing to smash SP. Five months later, the arrest in Lima of the longtime leader of the group, Abimael Guzmán (Comrade Gonzalo), gave Fujimori a tremendous success and large popularity. By the end of 1993, only one year after the fall of Guzmán, SP was largely defeated. Guzmán’s arrest had dramatic effects on the morale of the SP militants, and many gave up weapons and had recourse to the repentance law passed in 1992 (Bermúdez, 1995). Guzmán’s later call for surrender simply hastened the process.
The war took 69,000 victims, between SP, state repression, and the actions of other minor groups. Most of the violence was concentrated around the central Sierra region, with Ayacucho, Huancavelica, and Apurímac as the deadliest departments, plus the Huallaga corridor in Huánuco. Not by chance, these areas experienced the largest presence of the insurgency. Only Lima showed a different pattern, with many attacks but no noticeable open presence of the rebels.
Operationalization and data sources
Far from being an outlier in the universe of conflicts, SP’s rebellion may be portrayed as the median insurgency nowadays. Firstly, SP relied on a combination of guerrilla tactics in the countryside and IEDs and selective assassinations in the urban centers. This tactic profile has become overwhelmingly dominant over the last decade (Jones & Johnston, 2013) and is apparently very successful in forcing states to negotiate (Thomas, 2014). Second, the Peruvian conflict took place under a democratic façade, and rebels strategically took advantage of regular elections to show military strength and popular support. As most dictatorships are transitioning into electoral authoritarian regimes (Svolik, 2012), rebels will have more chances of using votes as another armed tactic in their toolkit. Recent conflicts in Afghanistan, Iraq, and Ukraine attest to this. Finally, SP was led by a tight-knit leadership ring, overseeing a highly hierarchical organization. As this rebel structure promises better outcomes (Johnston, 2008), it is no wonder that recent successful rebel groups, such as the ISIS, draw on a pyramid-like organizational structure.
Following Mao’s lessons from the Chinese revolution, the SP’s leadership concentrated its proselytizing efforts on the countryside. For a decade prior to the outbreak of hostilities, the organization built networks of rural cadres and combatants with the intention of moving the conflict as quickly as possible to the ‘guerrilla’ stage. However, SP leaders were soon forced to come to terms with the reality that rural insurgency on its own is doomed to fail in largely urbanized countries, where government personnel and state forces entrench themselves in the main city centers (Marks, 2003).
Ingeniously, Abimael Guzmán theorized the war as a conflict on two fronts: the countryside, with Ayacucho as its epicenter, and Lima, the capital city – neither below the other in order of priority. SP went on to raze the rural areas to seize territory, whereas its urban plan consisted of a more prudent strategy of mobilizing supporters and undermining the regime (Gorriti, 1990: 188; McCormick, 1992: 25). SP did care about public opinion. A public survey released in the late 1980s showed that 33% of adolescents in urban low-class schools considered militants as ‘social fighters’ rather than criminals or fanatics (Portocarrero, 1998: 205). Similarly, a countrywide study commissioned by the Peruvian Senate in 1989 found that 22% of Peruvians did not identify SP members as ‘terrorists’ (Senado de la República, 1989).
The Lima branch of SP was particularly active during most of the conflict, with a key presence in universities, associations of the young, 1 and the so-called young towns surrounding the capital city, but was unable to seize any district there (Degregori, 2011: Ch. 5). To cater to its base, SP seemed to select targets in Lima that had sided with the government in the fight against the subversives (Jiménez, 2000: 553–554). Given its clandestine network structure, it was important to avoid desertions in the cities. As one leading senderologo put it: ‘Unlike the rural members of SP, who were peasants recruited out of the strength of the rebels in their locality, the urban cells were largely made up of convinced militants, who blindly believed in the group’s discourse’ (Tapia, 1997: 137, my own translation).
My goal is to systematically test whether SP adjusted its strategy to the two logics of target selection. To test this argument, I have created a dataset with around 30,000 district-year observations, including 1,835 current districts and 16 years (from 1980 to 1995), and information collected from several sources. For data on violence, I have relied on the CVR dataset. For the main independent variable, territorial control, I have relied on electoral data from municipal elections in Peru since 1980.
The most important source about the Peruvian civil war is the reports collected by the Truth and Reconciliation Commission (Comisión de la Verdad y Reconciliación/CVR), a non-party group commissioned by the Peruvian Congress to investigate the origins, dynamics, and consequences of the civil war. The CVR dataset includes information on 16,383 deaths coded by location, date, status of the victim (target), and organization responsible. I collapse target type into three categories: civilians, security forces, and authorities. According to the CVR, 8,631 victims were killed by SP, and the army was responsible for 5,690 deaths. I use these deaths produced by the military to test I4, conveniently geocoded at the district level and lagged to avoid reverse causality.
The CVR collected its information through victims’ relatives’ testimonies, and this may explain why the dataset underreports the number of security personnel and self-defense members (so-called ronderos) killed by SP. In the first case, the security forces did not collaborate with the commission in releasing information about their deaths during the war. However, the Defense and Home ministries later provided the CVR with a list of 1,600 military and police victims. I coded this list and merged it with the CVR dataset, ending with a total of 8,018 victims killed by SP that were timely and spatially identified – almost 80% of the victims included in the dataset.
The second source of underreporting is more debatable. Rondas were poorly armed volunteer groups that emerged out of a mix of local concerns against rebel oppression and military imposition (Fumerton, 2001). It is a puzzle why ronderos’ relatives failed to recognize that their kin were killed because of militia involvement, as the self-defense groups were on the winning side, and this usually diminishes the culpability associated with armed membership. It may be the case that rebel retaliation against self-defense groups concentrated on part-time volunteers or collective targeting at the village level, which would explain the low number of ronderos identified in the dataset.
The most critical variable in this work is territorial control. By rebel control I understand the capacity of the rebels to seize and rule geographical areas of the country, displacing previous state authorities. Rebel control has been rarely used in empirical analysis, because measures of rebel presence are too difficult to gather. Leaving aside ethnographic work that collects local data, most large-N analyses draw on measures of rebels’ activity to map their presence. This is problematic because fully liberated areas could experience less violence than contested districts.
I propose a new measure of rebel control drawn from the degree of success of electoral boycotts promoted by the insurgents. Authors analyzing SP’s fortunes always looked at its boycott campaigns as a way to gauge SP’s territorial spread (McClintock, 1998; Pareja & Torres, 1989). Electoral outcomes have also been used as informational input for displacement campaigns (Steele, 2011) and targeting in conventional civil wars (Balcells, 2010). In addition, rebellions in the shadow of elections are becoming very common, giving insurgents the possibility of showing strength through either electoral boycotts or participation with a political mouthpiece. Pro-Russian rebels in Ukraine, ISIS fighters in Iraq, and the Taliban in Afghanistan have already employed this electoral strategy.
I exploit the variation in electoral outcomes at the district level to operationalize SP’s strength. Voting in Peru is mandatory, and the government used its full power to force people to vote (see Flores, 2004: 218 for an illustration). As the Senderista leadership always called for its supporters to boycott the elections in order to denounce the ‘democratic farce’, local elections in many districts were either never held or annulled afterwards because of low turnout or failure of state authorities to oversee the election. Besides, the number of spoiled votes is extraordinarily high in many districts. Given that during the civil war, seven local elections (1980, 1983, 1986, 1989, 1993, and 1995) and many intermediate recall elections were called, we have enough data points to generate an operational proxy for territorial control. 2
To translate votes into war zones, I lay out a slightly adjusted version of Kalyvas’s five-zone continuum of territorial control. His continuum has the following categories: zone 1 (safe state control); zone 2 (under state control, but with rebels operating clandestinely); zone 3 is the prototypical area under dispute, with contested authority over the district (usually, the army rules during the day, and the rebels during the night); zone 4 (rebel control, but with the army penetrating the district regularly); and finally, zone 5 (safe rebel control).
Instead of a five-zone continuum, I suggest a three-zone continuum applies better to the Peruvian case. It has been argued that SP never had a safe rearguard, where the state could not penetrate (CVR, 2003), and my data prove it, since there was not even a town where the rebels were in charge during the whole cycle of violence. For the sake of symmetry, I use a three-zone scale, where 1 is a safe/slightly contested state area, 2 is a contested area, and 3 is a slightly contested SP area. 3
I consider a district as under SP’s control (zone 3 in my scale) when the election was annulled (or never held). I consider a district as contested (zone 2 in my scale) when the election was held, but the number of spoiled votes was larger than 50% of the votes cast. If the state is sufficiently strong to run the election, but not strong enough to prevent SP from forcing (or encouraging) local citizens to cast a spoiled ballot, the situation resembles one of contested power. Finally, if the election was run smoothly and few spoiled votes were cast, then the district belongs to zone 1 and is considered under safe state control. For interelection years, the district maintains the value of the immediately previous election. This instrument is used to test I1 and I2. I also create a proxy for the districts belonging to the province of Lima to run a more demanding test of I2.
On average, 20% of the districts were either contested or under rebel control, with a peak in the three most affected war areas: Ayacucho, Apurímac, and Huancavelica (56%, 52%, and 41%, respectively). The temporal variation also fits the narrative about the war, since SP triggered its last offensive in the late 1980s, bringing the democracy to the brink of breakdown in 1992. The successful arrest of Abimael Guzmán, SP’s founder and single leader, in September 1992 ended the campaign and caused the quick defeat and collapse of the insurgency. 4
The third implication is about the effect of alternative sources of funding on civilian violence. SP did not count on major foreign sponsors, but it extracted unaccountable resources from the coca-growing area of Alto Huallaga. 5 The coca business boomed in the area during the 1980s, attracting SP’s cadres since the middle years of the decade. Initially welcomed by coca dealers and local growers alike, because SP aspired to pacify a highly conflictive area and to impose some provisions on the deals, its presence was soon resented (CVR, 2003: Volume V). Although SP levied millions of dollars from the coca trade, the SP’s leader declared that the money never flowed into the city coffers, a statement that is controversial (Guzmán, 2002).
I include a dummy for Alto Huallaga to test this implication. This is certainly an imperfect proxy, since it does not vary with local fluctuations in coca cultivation. Unfortunately, there are no disaggregated data on coca production for the years under analysis in this study.
The fourth implication proposes that civilian targeting goes up when the state is heavily repressive. I look at the number of locals killed by the Peruvian security forces in each district, as reported by the CVR. I additionally investigate whether state killings have more impact on lifting supporters’ constraints in the cities by interacting state repression with the conflict zone to which the district belongs.
Finally, I include three controls. First, the logarithm of the district population (taken from the 1981 Census Handbook); second, the three presidency terms, as a way to capture different ideological approaches to the conflict; and third, zone switching. Against my theory, this last variable identifies situations where rebels do not quickly update their targeting patterns in reaction to sudden changes in control. For instance, districts with increasing rebel presence could have less violence than expected – because rebels still want to satisfy their supporters’ preferences on targeting – whereas districts with decreasing rebel presence could suffer more violence than expected – because rebels resort to civilian victimization to cover gaps in military power.
I code a district as ‘decreasing control’ when it moved for the rebels from zone 3 to zone 2, from zone 2 to zone 1, or from zone 3 to zone 1. Symmetrically, a district is coded as ‘increasing control’ when it moved from zone1 to zone 2, zone 2 to zone 3, or zone 1 to zone 3.
Descriptive results
My theory speaks to the scale of rebel violence and its breakdown into targeting profiles. I start with a look at the absolute number of civilian victims killed by the state and SP. I distinguish between areas where the state was in control, areas of contested territory, and areas of rebel control. Figure 1 suggests that the geography of rebel victimization is concentrated in the polar conflict zones, with less violence in the contested areas. The same pattern applies to state repression, with a lower reliance on civilian targeting in areas under fragmented authority.
Total number of civilian deaths, by group and zone
By showing the average share of targets in every war zone instead of absolute numbers, Figure 2 uncovers large variation in civilian deaths by control zone. It is worth noting that security personnel are more targeted in safe state zones (34% of all victims), whereas authorities are proportionally killed more in contested areas (24% of deaths).
Control zones and share of targets by SP
Why should this be the case? If underground rebels would have trouble in finding targets, they could just simply carry out low-risk collective attacks that require little evidence-gathering work. But if they take the costly effort of monitoring security forces and authorities, the explanation could lie in the search for legitimacy. On the other hand, the fact that authorities are most killed in contested areas suggests that publicly known authorities could serve as scapegoats of rebel violence when militants lack good information about potential enemies in those zones.
Results
I turn now to two more rigorous tests of the implications. First, I investigate the absolute number of civilian deaths provoked by SP in every district-year. Because there is overdispersion and an excessive number of zeros in the count data, I use zero-inflated negative binomial models with district-clustered standard errors. 6 Zero-inflated models involve calculating two processes: why some units experience more positive counts than others, and why some units have no counts at all. To model the second process, I introduce a measure of SP activity in every district, the number of years in which SP killed at least one person, as a way to instrument the probability of observing civilian victimization there. 7
The three models included in Table I report incidence rate ratios, which allow for a direct comparison of the effects of the different variables: values of 1 for absence of effect, values above 1 for positive effects, and below 1 for negative effects over the dependent variable. The first column is the baseline model. The second column checks whether violence against civilians was substantially lower in the Lima province. The third model incorporates interactions between state repression and rebel control and between the Huallaga region and rebel control, in both cases to test if the effects of repression and resources are larger in areas where rebels are weaker, as a more stringent version of my theory would imply. All models include three controls: sudden changes in control, presidential term (García and Fujimori, Belaúnde being the baseline) and (logarithm of) district population. 8
ZINB models of the number of civilians killed by SP (irr coefficients)
Exponentiated coefficients; t statistics in parentheses; * p < 0.05, ** p < 0.01, *** p < 0.001.
To test the four implications, I introduce measures of control zones (contested and rebel, state control being the baseline) (I1), a dummy for districts belonging to the province of Lima (I2), a dummy for Alto Huallaga (I3), and a yearly count of victims killed by the state in the district (I4). The proxies for control and state repression are lagged one year. 9
Models 1 and 2 show that rebel control has a strong and positive impact on civilian victimization. Not only did SP kill more civilians in rebel areas, compared to safe state territory, but also in contested zones. The effect of Lima province is very robust, with a significant reduction in civilian victimization in the main conurbation of the country. 10 In accordance with Weinstein’s micro evidence from Alto Huallaga, this region suffered higher levels of civilian killings than the country in general. Still, more civilian targeting may be driven by the presence of more competitors to grab the rents afforded by the illicit trade (drug-funded private militias, security forces, and insurgents) or by the worse quality of recruits – or even both motives. Finally, state repression triggers rebel retribution.
Model 3 tests whether the effect of repression and independent sources of funding have a larger impact in areas where rebels operate clandestinely. Both interactions show the expected signs, although the repression coefficients are not significant. The effect of funding is more robust, however. Civilian victimization was larger in the Alto Huallaga districts where SP did not hold or fight for territory. This could be an indirect sign that targeting in the area was more affected by the quality of recruits than by group competition. If the latter were the driving force, then violence against civilians should have been concentrated in contested districts, where all actors tried to reign over armed rivals (Metelits, 2010).
Figure 3, derived from Model 3, plots the expected number of civilians killed by SP depending on levels of repression, control zones, and the production of coca (Alto Huallaga vs. rest of the country), with the remaining variables kept at their means.
Expected number of SP killings
The expected value of civilian deaths is relatively low, since many districts were spared the conflict. But there is still an interesting variation. The contested zones seem to be those that captured most of the tit-for-tat effect: regardless of the region, contested zones experienced exponential increases in civilian targeting in reaction to higher state repression. This qualifies the expectation that repression should have a larger impact on safe state areas. In contested areas, the local dynamics of retribution could drive up civilian targeting.
There is an alternative way to understand hypotheses about targeting. Instead of accounting for absolute numbers, they could be interpreted as predicting violence against civilians in comparison to violence against other targets. War-ridden areas suffer more violence in general, regardless of the target. The point is, however, whether the proportion of civilian victims varies along control zones. In this case, we would be dealing with ‘compositional’ data. The proper statistical technique to analyze this type of data is ‘seemingly unrelated regressions’ (SUREG), because the proportion of attacks an insurgency carries out against one specific type of target is related to the proportion perpetrated against other targets. 11
Table II includes the three types of targets (civilians as compared to security forces and authorities, respectively). Only district-years with at least one victim are included – otherwise, they would have zeros in the three proportions and the log-ratios could not be calculated. The table includes two models, each of which contains two columns – the comparison between the proportion of security forces killed in the district and the proportion of civilians, and the corresponding comparison between authorities and civilians. Positive coefficients denote that in the district the share of civilians compared to security forces (or authorities) went down when the independent variable increased its value. To the contrary, negative values denote a larger proportion of civilian victims compared to the other two targets.
Patterns of targeting by Shining Path (sureg models)
Standard errors in parentheses; * p < 0.05, ** p < 0.01, *** p < 0.001.
Model 1 further proves that rebel and contested zones in Peru had on average a larger share of civilian victims than state areas. The only exception is the comparison between authorities and civilians (column 2), which is not statistically significant in contested areas – perhaps an indication that for rebels the authorities may be a convenient target in areas where they cannot gather information about civilians collaborating with the enemy. Both repression and funding (Alto Huallaga) contribute to raising the proportion of civilian victims. Finally, in a remarkable contrast with the previous results, switches in territorial control have a low impact on the targeting profile of SP. It would seem that changes in the logic of rebel fighting affect the levels of civilian victimization, but not the patterns of target distribution. This result casts doubt on the expectation that rebels substitute intimidation for military strength when their capabilities are eroding.
Model 2 offers an additional test of the main argument of this article – that rebels are more cautious when selecting targets in the safest state areas – in this case, the Lima conurbation. The findings indicate that Lima underwent a different type of conflict, more affected by attacks against security forces and authorities than other areas of the country. Far from being highly indiscriminate, SP’s clandestine units in the capital city carefully focused on selecting ‘legitimate’ targets, in open contrast to its disproportionate behavior in the rural areas.
Table III includes the predicted share of attacks of each target type depending on conflict zones, regions, and levels of repression. I distinguish between state zones and the Lima province, because the latter is the most paradigmatic example of an area where rebels have no chance of fighting for territory. I single out three scenarios: (i) districts outside Alto Huallaga that were unaffected by state repression; (ii) districts outside Alto Huallaga affected by high levels of repression; and (iii) Alto Huallaga districts with low repression. 12
Predicted shares of targets by region and control zone
Several comments are in order. First, an increase in rebel control leads to an increase in the targeting of civilians, regardless of the scenario. The tendency is apparently linear, with areas under state control suffering less violence against civilians. This result is more striking for the province of Lima, where security forces and authorities make up a spectacular 83% of the victims (80%, in the districts with high repression). Second, the expected effect of state repression on the proportion of civilian targets is larger in state areas. Third, the largest proportion of attacks against authorities takes place in contested zones, where rebels may try to remove state officials from the district. Finally, the coca-growing county of Alto Huallaga experienced the highest shares of civilian targeting.
In all, these results point to a war fought in two different terrains, with two different targeting logics: when SP fought for territory, civilian victimization became a way to ensure collaboration and compliance; when it fought for support and legitimacy in the cities, civilian victimization became a burden, and it selected instead more cautious target choices.
Conclusions
I have proposed in this article a new theory of rebel targeting in guerrilla conflicts. Drawing on previous insights from literatures on civil wars and terrorism, I have claimed that most insurgencies operate under two logics – the logic of compliance and the logic of legitimacy. The first applies when rebels fight for territory, usually in the rural areas; the second applies when rebels fight for support, usually in the cities, where they are forced to operate clandestinely and rely on supporters’ networks of collaboration.
The fight for territory is driven by the imposition of compliance on local dwellers. The fight for legitimacy is driven by the battle for hearts and minds. To govern liberated territory and clear contested districts, rebels use civilian victimization as a way to discipline local populations and punish potential enemies. As a consequence, civilians suffer heavily in these districts. To gather support in cities, the safest state areas in most conflicts, rebels carefully select their targets, since unqualified killings could backfire. As a consequence, they usually pick out security forces, authorities, and outspoken civilians who campaign against the rebels.
Supporters’ constraints are nonetheless lifted when insurgents’ support constituencies have suffered heavy losses inflicted by the state or when the rebels can rely on autonomous sources of funding – such as illegal commodities. In the first case, supporters may call for retribution; in the second, rebels no longer depend on supporters to raise funds and continue the fight.
I tested this argument with data from the Peruvian internal conflict during the 1980s. This prototypical asymmetric war had SP fighting on two fronts: a guerrilla war in the countryside and a terrorist campaign in the cities. The results indicate that not only did rebels kill fewer civilians in urban areas than in the areas where they fought for territory, but also the proportion of civilian victims compared to security forces and authorities was significantly lower. Although rebels might maximize harm by placing no-warning bombs in the cities, this tactic could jeopardize supporters. In Lima, the safest state area in Peru, SP implemented a very cautious approach to target selection, with a low focus on civilians.
The results offer mixed evidence on the effect of state repression over supporters’ preferences for civilian victimization. On the one hand, repression increases rebel violence against civilians. Retribution is a deep source of civilian victimization. On the other, contested areas suffer more violence against civilians, but this effect is proportionally stronger in safe state areas.
Finally, the existence of coca-based receipts also contributed to increasing the number of civilian deaths. Still, as the Alto Huallaga region did not encompass any urban center, it is difficult to tell whether the funding effect would also apply in areas where the rebels have to hide permanently.
The findings endorse the idea that rebels fight with different logics in mind: compliance and legitimacy. Although intuitive, this argument goes against the current literature on violence in three ways. First, it claims that rebel violence against civilians in areas under state control is not necessarily indiscriminate and large-scale. To the contrary, violence should be even more selective, just to avoid jeopardizing potential supporters who could easily switch sides and root for the government. Second, contested areas are far from being safe havens for civilians. Although it can be difficult for rebels to screen rivals in these areas, they may still carry out collective attacks against uncooperative districts. And third, I suggest that urban political violence is not the last resort of highly alienated rebels but a complementary and yet necessary wing of modern insurgencies.
The logic of war unfolded in this article seems on the rise. Recent experiences of groups such as ISIS make the point. ISIS, for instance, moved swiftly from an urban network of underground cells to a group capable of seizing Anbar province and later large swaths of Northern Iraq and Eastern Syria. Although we lack detailed data on the target profile of ISIS, its own annual report points to three interesting metrics: (a) the absolute number of ISIS attacks skyrocketed from 2012 to 2013 as the group started to seize and hold territory; (b) the most violent provinces in Iraq were those where the group was fighting for control, with much lower levels of violence in Baghdad and Kirkuk – the safest state areas; and (c) car bombs hit the last two cities disproportionately, showing a different tactical approach to state areas (Bilger, 2014).
Compared to ISIS’s no-prisoners policy that is dramatically increasing civilian victimization in the areas now under its control, its activities in Baghdad have remained low profile, with bombs targeting goal-inimical groups and government personnel. It is still too early to tell who will prevail in this conflict, but one could claim that, before starting to retreat territorially, the Iraqi government had already lost the fight for hearts and minds within the Sunni constituency. In the end, it is no doubt better for a state to wage a self-restrained fight against clandestine rebels than to let them move into large-scale rebellion, which inevitably mostly affects civilians.
Footnotes
Replication data
Acknowledgements
Earlier drafts of this article were presented at Yale, Uppsala, CIDE, IC3JM, and LSE. I would like to thank those present for their comments and suggestions. I am also grateful to three anonymous reviewers and the JPR editors.