Blood in the Village: A Local-Level Investigation of State Massacres

Definitions and Existing Research

In much the same way as the tactic of terror has been clearly differentiated from an espoused ideology of terrorism (e.g. Tilly, 2004), tactical understanding of the term massacres needs to be clearly distinguished from an alternative understanding of massacres as strategy. In the literature on political violence it is common to study massacres as part of larger strategies of violence, such as genocide or politicide (see reviews in Valentino et al., 2004; Straus, 2006; Kalyvas, 2006; Davenport, 2007). By strategy I refer to the broader organization of state forces and the ends towards which they are deployed. From a perspective that conceives of massacres as strategy, violence is the end goal. For example, in the literature on genocide, the strategic goal of the state is defined as the destruction of a collectivity (in whole or in part). To understand where and when states commit massacres, scholars in this tradition have pursued research strategies aimed at identifying when the strategic goals of the state are framed as the purging of ethnic, racial, political or religious groups (e.g. Valentino, 2004; Sémelin, 2007; Midlarsky, 2005).

While this perspective has identified settings in which strategies of group destruction are increasingly likely, research in this tradition has largely glossed over the dramatic variation in the application of massacres within states. Even within states committing campaigns of genocide, there are locations that appear to be relatively free from political violence (Straus, 2006; Davenport and Stam, 2008; Fujii, 2009). At the same time massacres have also been observed within states that have not committed themselves to pursuing strategies of genocide (Francisco, 2004). Working around the idea of massacres as a strategy has restricted the capacity of academic research to account for this dramatic variation and explain why massacres would be committed for reasons other than group destruction.

Alternatively, massacres can be conceived of as a tactic, entailing “the annihilation of a section of a group—men, women and children, as for example the wiping out of whole villages” (Kuper, 1983: 10; see discussion in Markusen and Kopf, 1995). Like other tactics of political repression, such as domestic spying (Davenport, 2005), torture (Sullivan, 2011), protest policing (Earl et al., 2003), and targeted killings (Kalyvas, 2006), massacres may be utilized to achieve a number of strategic goals. Ethnic group destruction may be one such strategic goal, but it is not the only one. Additional strategic goals that may motivate states to commit massacres include, among others, leadership survival, military victory, extending the territorial control of the state, social engineering, and regime consolidation.

Distinguishing between tactics and strategies opens up the theoretical box around which one might begin to identify the different strategic incentives that make the tactical selection of massacres more or less likely. The move allows for the development of theory that focuses less on temporally and spatially invariant (or slow moving) macro-structural variables associated with the political economy and towards more dynamic theorization about the conflict processes that may produce violence in some sub-national settings but not others.

But before one can generate a theory of when states are likely to deploy this particular tactic massacres must be distinguished from other repressive tactics commonly engaged in by state forces. One reliable metric for distinguishing among the repressive tactics available to the state is the ways in which different tactics target different subsets of the population. ¹ For instance, much attention has been paid to the production of selective violence, which targets individuals based on their presumed collaboration with an organized insurgency (e.g. Kalyvas, 2006; Kalyvas and Kocher, 2009). With selective violence, individuals who have aided the insurgents or provided information on state forces are targeted in the aim of deterring others from committing similar transgressions against the state. Tactics that might be associated with selective targeting include selective killings (e.g. Kalyvas, 2006), arrests (e.g. Davenport and Sullivan, 2011), torture (e.g. Sullivan, 2011), and protest policing (e.g. Earl et al., 2003). Selective violence is often distinguished from indiscriminate violence, which has been the subject of a smaller number of studies (e.g. Ziemke, 2008; Lyall, 2009; Kocher et al., 2011). With indiscriminate tactics, the state does not aim to limit its violence to those with presumed ties to insurgents and instead targets all civilians with an equal probability. Forms of indiscriminate violence include random shelling (Lyall, 2009) or pillaging tactics (Ferrara, 2003).

Between selective and indiscriminate forms of target selection there is another tactic—collective targeting (Steele, 2009). Massacres are a collective targeting tactic in which violence does not select victims based on individual characteristics, but instead selects on shared traits such as communal identity or residential location. Villages or groups are selected for repression based on the strategic aims of the state. Individuals within the targeted collectivity are victimized indiscriminately. Other forms of collective targeting include ethnic cleansing (e.g. Steele, 2009), the bombing of strategic hamlets (e.g. Kocher et al., 2011), genocidal killings (e.g. Straus, 2006; Fujii, 2009), and occupation and siege (e.g. Khawaja, 1993).

Having distinguished among different tactics available to those in charge of deploying repressive violence, the next task is to assemble a theory of strategic incentives and tactical selection that can specify when and where we should expect the state to engage in massacres. While there has been extensive work theorizing the production of selective violence and one-sided violence writ large, there has been significantly less work accounting for when we should expect a state to employ collective targeting tactics. And among the existing studies of state repression, there has been limited attention to the selection of particular repressive tactics over others (Sullivan, 2011; Conrad and Demeritt, 2011). The next section makes advances towards the development of such a theory, paying particular attention to the selection of massacres over alternative tactics of repression.

Massacres: Strategic Incentives and Tactical Selection

The challenges of domestic government produce numerous strategic incentives for states, including, but not limited to, military conquest, population control, wealth extraction, and the provision of public goods such as political order (Tilly, 1993; Olsen, 1993; Bueno de Mesquita et al., 2003). For states to remain in power they must match each strategic incentive with an appropriate tactical repertoire of repressive violence, selected to fulfill the aims and objectives of the state. Population control, for example, necessitates selective repression (Leites and Wolf, 1970; US Army/Marine Corps, 2007; Kalyvas, 2006). At the individual level repression only works to coerce civilians into cooperating with the state where states are able to utilize violence selectively in response to individual behavior (Kalyvas, 2006). When an individual defects and is violently punished for his or her actions, others make inferences about the potential consequences of their own actions and may be deterred from committing similar actions. Collective targeting, like indiscriminate violence, is unlikely to prove effective for population control. This is the case because violence loses its deterrent effect when civilians can no longer feel assured that not committing prohibited actions will protect them from state repression. If committing the prohibited act produces the same probability for being repressed as not committing that act, then there is no incentive to restrain from such activity (Kalyvas and Kocher, 2007; Mason and Krane, 1989).

But there are other strategic objectives of the state that can motivate the rational selection of collective targeting tactics, such as massacres. This research focuses on two—threat removal and the projection of state control over territory. Identifying the strategic incentives that lead to the selection of massacres helps to generate hypotheses about where and when we can expect to observe such strategies. It also helps highlight the contribution of this study. While the logic of selective violence has been well articulated by Kalyvas and his co-authors, the same cannot be said for the logic of collective targeting. ² And while the present study shares some of Kalyvas’s focus on territorial control, the theory below is able to generate positive predictions about where collective targeting tactics might be useful for the state, instead of relying on negative expectations about where the state cannot deploy other forms of violence. It moves away from a theory that privileges the agency of non-combatants and towards a theory of the decision making processes of state agents making calculated decisions about where to deploy different forms of political repression (see also Stanton, 2009). ³ By identifying the strategic settings that make the selection of massacres more likely than other tactics of political repression, the theory derives testable hypotheses about where and when we can expect to observe the deployment of this form of violence. ⁴

What strategic settings are likely to inspire the tactical selection and perpetration of massacres? The first is the removal of threats to political elites. Threat removal is a substantially different goal from population control. In Shelling’s (2008 [1966]) classic terminology, population control is a strategy of coercion, while threat removal is a strategy of brute force. With population control, the power to hurt is threatened and used in response to past behavior. The aim is to deter citizens from taking actions deemed undesirable by the state. With threat removal, violence is used to physically eliminate the source of the threat as quickly and efficiently as possible.

As Shelling (2008 [1966]: 150) notes, “Brute force works when it is used, whereas [coercion] is most successful when held in reserve.” Massacres can be more effective than other repressive tactics for threat removal because of the collective nature of most threats to the regime. To undercut the capacity of these threats to challenge the survival of regime, state elites may direct the deployment of violence at the collectivity without regard to whether or not each individual victim is associated with that threat. By contrast, selective targeting would require extensive information on who is or is not associated with the threat (Kalyvas, 2006), while indiscriminate violence would be insufficiently directed at the source of the threat and may do little to exterminate those who challenge the state (Kalyvas and Kocher, 2007).

But what types of actions and actors do states perceive as threatening? The perception and attribution of threat to actions by a challenging group are complicated social-psychological processes. However, by building models around a central state decision maker, we can begin to identify attributes likely to be deemed threatening to the regime. First among these, states appear threatened by insurgent violence (Poe and Tate, 1994; Davenport, 1995). States are theorized to feel threatened by insurgent violence because of the immediate consequences it has for their continued tenure. Where insurgent violence is targeted at state troops and supporters, violence imposes costs on the state and can threaten to overturn the political order. In the extreme case where the insurgents are able to deliver more violence than the state, state leaders may be concerned that insurgents could take the capital and oust the government.

Insurgent violence also signals to the state that other repressive tactics such as selective targeting have proven ineffective (Valentino et al., 2004). Unable to end insurgent violence by other means, the state may turn to massacres as a “bloody but effective solution to the seemingly intractable problems of guerrilla warfare” (Valentino et al., 2004: 10). Massacres can destroy the mobilizing structures that support insurgents and contribute to the production of insurgent violence. By carrying out massacres where insurgent violence is most prevalent, the state aims to utilize such violence strategically to remove the insurgents and eliminate the perceived threat (Harff, 2003; Chalk and Jonassohn, 1990).

In addition to insurgent violence, massacres may also be employed to manage the threat posed by rival ethnic groups. Because the boundaries defining states and ethnic groups rarely overlap completely, political contestation in multi-ethnic states is frequently characterized by competing territorial claims. In many multi-ethnic settings one ethnic group is able to secure a dominant position by controlling the state apparatus and manipulating the state to discriminate against members of rival ethnic groups. Members of the dominant ethnic group are then able to utilize state resources to further the material and non-material goals of the group.

Yet the ability of dominant group members to maintain their position is inherently uncertain. The presence of a viable insurgency further increases the degree of uncertainty, especially when mobilization practices cut along ethnic cleavages (Barnes, 2005). ⁵ Massacres against ethnic group members can then be legitimated through logics of self-preservation which link the rise of the insurgency to a threatening increase in the power of the ethnic opposition (Sémelin, 2007; Midlarsky, 2005). The threatened state responds by labeling members of the rival ethnic group as targets for extermination. In order to counter the rise of the antagonistic ethnic group and restrain ethnic mobilization, the state can begin to direct its repressive energies towards areas with high concentration of rival group members. Massacres are employed to eliminate as many members of the ethnic group as possible.

Beyond threat removal, a second strategy that may motivate the tactical use of massacres is the projection of state control over territory. States have at least two goals when it comes to territorial control—consolidation and projection. States consolidate control by coercing individual members of the civilian population into cooperating with state authority and deterring them from defecting to the insurgency. Here selective violence is again the most useful tactic, and collective targeting is unlikely (Kalyvas, 2006). States project control by shifting the balance of power between state forces and their rivals. In settings where the state aims to project control into territory, massacres may be a preferred tactic.

Two mechanisms are posited as means by which massacres may function to project state control into uncontrolled regions. First, functionally and most horrifically, massacres are used to eliminate and disperse the populations living in areas uncontrolled by the state. Having destroyed the “community of resistance”, state forces may then move in to repopulate the area with military command and populations that are more willing to be sympathetic to state control (Barnes, 2005). Meanwhile, massacres create large numbers of internally displaced persons who may be resettled into areas of greater state control (Walter, 1969). Having redirected the population into areas it controls, state agents may more accurately monitor civilian behavior and selectively repress those who do not comply with state authority. Second, massacres signal to those living nearby that the state possesses overwhelming military force and that it is not afraid to use it (Lyall, 2009). Witnessing such violence can inspire a paralyzing sense of fear that prevents political action and provides space for state forces to impose order (Kalyvas, 2006: 143).

Because massacres may usefully be employed for projecting state control into areas not yet controlled by the state but not for consolidating state control where the state is dominant, it is possible to use the degree of state control held over an area to predict where states are likely to commit such violence. While massacres may still be useful for resettling local populations or for signaling state strength in areas where the state is dominant, such policies are largely unnecessary given the facts that the state may wield violence more selectively to consolidate control in these areas and that massacres may potentially be counter-productive for consolidating control (Kalyvas, 2006; Kalyvas and Kocher, 2007). In order to extend the reach of state power into the areas not under state control, state forces can turn to massacres as an extremely violent but strategic choice to project control into these regions.

Political Violence in Guatemala

Before reviewing the data, it is necessary to first provide a brief history of political violence during Guatemala’s civil war. The most recent period of violence began in 1960, with an urban insurrection led by the Guatemalan Labor Party. After staging a failed coup, the group fled to the jungles. With little public support and few arms, the movement made little headway in its initial campaign (REMHI, 1999). Then, after a period of relative calm in the late 1960s and early 1970s, political violence rapidly increased in the country. The Guatemalan Army of the Poor (EGP), who had formerly directed their principal mobilization efforts at the educated Ladino elite, began making claims in the indigenous jungle regions of Alta Verpaz and Quiche. By 1977–78, the insurgency had a strong hold over broad swaths of the jungle, and the Guatemalan government was forced to admit they were engaged in a “people’s war” (Schirmer, 2000). They recognized that as the insurgency came to control greater numbers of villages, the insurgents’ vision of the conflict became the dominant narrative in those villages. In response, the state’s deployment of repression changed dramatically. Where previous repression had been limited to the insurgents and their presumed supporters, the army began the first series of massacres in 1979.

During the first wave, massacres were targeted at villages close to the front lines of the conflict. Thousands fled to the jungle with the hopes that the insurgents would protect them (Stoll, 1993). The ranks of the insurgent army swelled. Within the army, Rios Montt’s coup in March of 1982 led to a further intensification of violence. Rios Montt chose to escalate the conflict by turning the army loose on villages in the highland jungle. Schirmer (2000: 45) contends that, “this meant literally emptying the local population from its socio-cultural and geographic habitat in order to create logistical and recruitment difficulties of every order”.

Anthropological studies confirm that once a village was targeted for extermination, no effort was made to separate combatants from non-combatants (e.g. Falla, 1994; Stoll, 1993; Manz, 2004). Because the overwhelming majority of victims in state massacres were indigenous peoples, scholars, NGO activists, and members of the international community began to levy charges of genocide. The UN sponsored Commission for Historical Clarification reached the same conclusion, asserting that, “agents of the State of Guatemala, within the framework of counterinsurgency operations carried out between 1981 and 1983, committed acts of genocide against groups of Mayan people” (CEH, 1999: 122).

However, unlike the preceding regime, the generals in charge of Rios Montt’s military campaign had a clear plan for eliminating the insurgents (Schirmer, 2000). The state had become aware that massacres were insufficient for prising peasant support away from the insurgency. As a result, massacres became the first part of the military’s three-pronged strategy to manipulate the support of the indigenous population. The second aspect of the plan dealt with the internally displaced persons (IDPs), of which there were between 300,000 and a million in 1982 (CIA estimates, Doyle, 1999). IDPs were forcibly resettled into “model villages”. There the government administered food and managed population migration. In the third part, the army established village level paramilitary units, known as civil patrols. The civil patrols were placed under direct military command and instructed to identify any civilians in their village who continued to support the insurgency (Stoll, 1993; Kobrak, 1997).

Eventually, the state’s wave of terror came to an end. The “beans and bullets campaign” was successful, though at great cost. Insurgent troop strength fell from a height of just over 3,000 in 1982 to fewer than 300 in 1984 (CIA estimates, Doyle, 1999). While the insurgent and counter-insurgent violence would continue for another decade before a peace agreement was signed, the majority of political violence had ended.

Data and Measurement

Data on political violence in Guatemala employed in this analysis were constructed by joining the records of the Commission for Historical Clarification (CEH), a UN sponsored truth commission, and the International Center for Human Rights Investigations (CIIDH), a human rights NGO charged with documenting the human rights abuses that took place during Guatemala’s civil war (Ball, 1999). These event databases contain information on both the perpetrators’ and victims’ identities, as well as the means of violence employed. The CIIDH database is based on over 5,000 interviews, as well as a comprehensive survey of 17 different domestic press sources and a review of the abuses recorded by four international and domestic human rights agencies (Ball, 2000; Davenport and Ball, 2002; Gulden, 2002). The database documents 17,423 violent events and records the deaths or disappearances of more than 45,000 victims (Ball, 1999, 2001). The CEH, meanwhile, conducted extensive research into the massacres committed by the Guatemalan government. The index of massacres was published in a seven-volume set cataloging the violence that occurred during the civil war. Massacres (defined below) were hand coded by the author and added to the CIIDH dataset for the analyses in this study.

For the analyses, recorded deaths between 1959 and 1996 were collapsed down to the month and municipality in which they occurred. These “municipality-months” serve as the units of analysis in the models below. ⁶

The CIIDH/CEH data represent one of the most complete quantitative surveys of civil war violence currently in use (compare Kalyvas, 2006; Cederman et al., 2009; Lyall, 2009). Combined, the three sources used to generate the CIIDH/CEH data (interviews, human rights reports, and newspaper listings) present a more accurate portrayal of the distribution of violence across Guatemala than any one source would have on its own. In an analysis of the data, Davenport and Ball (2002) show that that the non-state, non-media sources of information employed complement the newspaper data by capturing more rural incidents, capturing a greater number of smaller incidents and capturing alternative perspectives on the different incidents of violence. As a result, the data are superior to those used in existing studies based solely on convenience samples (e.g. Verwimp, 2006), data generated through newspaper clippings taken from the New York Times and newswires (e.g. Moore, 1998) or even data based on local press sources (e.g. Francisco, 1996).

Still, the non-random procedure by which the interviews were collected could lead to the dual challenges of selection bias and generalizability. However, concerns about selection bias are likely unfounded given the size and scope of the data collection effort. Patrick Ball, a principal consultant to the CIIDH and the CEH, as well as numerous other large-scale human rights data collection efforts, estimates that because of the volume of data collected it would require thousands of interviews with “fundamentally different stories” to alter the findings (Ball, 2001: 4). As to the question of generalizability, there is strong evidence to suggest that conclusions reached using the CEH data are applicable to the broader Guatemalan population. When processing the data, the collection team created a regionally stratified proportional sampling procedure to enter data into the database using random strata sampling (Ball, 2001). Statistical analyses were then conducted on both the random strata sample and the full sample. After half of the data had been collected, the findings from the randomized proportional sample approximated the full sample of statements nearly perfectly, suggesting that the correspondence between the collected data and the wider population is reasonably close (Ball, 2001).

However, the potential for sampling bias at the local level should not be underestimated. Given that neither the CIIDH nor the CEH attempted to sample the population randomly or to estimate the number of massacres at the municipal level, it would be imprudent to assume that the underreporting of political violence in the dataset is distributed randomly across space and time. To counter any potential bias imposed by systematically underreporting violence in particular municipalities or time periods, methods are developed below to estimate the probability of observing massacres conditional on the dataset’s recording patterns.

But before moving forward, it is necessary to first review the indicators used to identify the dependent (massacres) and independent variables (ethnic composition, insurgent violence, and state control), as well as the control variables.

Massacres

Consistent with the definition of massacres put forward earlier, the CEH defined massacres as “an indiscriminate attack” that involved “the execution of five or more people, in the same place, as part of the same operation and whose victims were in an indefensible state” (Mezquita, 2000: 6). Months in which a municipality experienced at least one massacre by the state are coded as experiencing massacres (1), while municipality-months without such violence are coded 0. ⁷ Figure 1 displays the spatial distribution of massacres and other forms of state violence across Guatemala’s municipalities during the height of the conflict.

OPEN IN VIEWER

Figure 1.

Variation in Political Violence by Municipality

As an example of what a massacre is and how massacres collectively target their victims, one can look at the massacre in the village of Dos Erres, Guatemala, which took place in December 1982. Following an insurgent attack on an army convoy elsewhere in the municipality, the Guatemalan Army’s elite killing squad, the Kabilles, received orders to enter Dos Erres and kill the inhabitants, who were all presumed to be insurgent supporters (CEH, 1999: §31). Despite the fact that a subsequent search of the town revealed no weapons or guerrilla propaganda, the Kabilles rounded up the town’s inhabitants and began killing them one by one. Beginning with the children, the army unit smashed each individual’s head against a tree until they were dead. Many women were also raped. No effort was taken to separate out individual victims. The entire village had been targeted for violence and everyone present was ultimately killed (CEH, 1999: §31).

Identification Strategy

The hypotheses put forward above are evaluated using a series of maximum-likelihood estimation procedures. To identify the effects of the independent variables on the dependent variable (i.e. the probability of experiencing massacres in a given municipality month), the models employ logistic regression with a rare-events correction. Because massacres are extremely rare (occurring on average in less than 1% of municipality months), rare-events logit is necessary to correct for potential biases that can lead to an under-estimation of such instances in the presence of a large proportion of zeroes (King and Zeng, 2001a,b). The rare-events logit models that also allow the study to re-estimate models including the civil patrol variables are using the full sample average rate of massacres instead of the subsample average (Models 2b and 4b below; see King and Zeng, 2001b).

The second series of models employs conditional logit as a means to guard against any potential biases imposed by the non-random data collection strategy employed by the CIIDH/CEH (Hosmer and Lemeshow, 1989). As noted, the dataset recorded less than a quarter of the total number of casualties estimated to have resulted from the civil war. The danger is that the underreporting of violence may not be random across municipalities or time periods, which could bias the results. The conditional logit models attempt to mitigate the potential for bias by first modeling the process by which the dataset recorded political violence in each municipality and then estimating the occurrence of massacres conditional on those estimates.

The method utilizes the dataset’s recording of violence in a municipality during the previous six months to estimate the probability that the CEH recorded political violence in a given municipality at time t. ¹³ Municipality-months that had similar reporting patterns over the prior six months are first unified in a stratum. The different strata are then incorporated into the models as a series of independent dummy variables and the models estimate the probability of observing massacres in a municipality controlling for the rate at which massacres are observed within their stratum and the sample population.

Municipalities that had experienced similar patterns of recorded violence in the recent past should have a similar probability of recording subsequent violence. As a result, this process should help mitigate the potential for selection bias to influence the results. The conditional-logit models estimate whether we observe significantly higher or lower rates of the dependent variable in the presence of the independent variable among municipalities that experienced similar patterns of recorded violence. The results should no longer be interpreted as estimates of the effects of the independent variables (X) on the dependent variable (Y) for all municipality-months, but instead as an estimate of the effects of X on Y conditional on the rate of massacres occurring in municipalities that experienced similar reporting patterns over the prior six months. Because these estimates are conditional on the recording of violence in the dataset, drawing valid inferences from their results requires less stringent assumptions about how the data were collected. These less restrictive models do not require that the reporting patterns of the data be homogenous across the entire dataset (i.e. all municipalities and all months), but only that the reporting patterns be consistent within reporting strata for short periods of time.

Empirical Analysis

The results of the analysis are reported in Tables 1 and 2. For each variable, the coefficient, standard errors, and level of significance are reported. Because the results of these models can be difficult to interpret, the percentage change in the predicted probability of massacres following a change in each independent variable (i.e. risk ratio, RR) are also reported. ¹⁴

OPEN IN VIEWER

Table 1.

Rare-Events Logit Analysis of Massacres in Guatemala, 1959–96

	Model 1	RR	Model 2	RR	Model 3	RR	Model 4a	RR	Model 4b	RR
Insurgent Violence	0.014***	+72.7%					0.017***	+95.6%	0.017***	+95.6%
	(0.000)						(0.000)		(0.000)
% Indigenous			2.564***	+695.3%			1.830*	+429.6%	1.830*	+438.0%
			(0.315)				(0.882)		(0.849)
State Control					−.950***	−63.5%	−1.006***	−63.5%	−.950***	−61.9%
					(0.251)		(0.252)		(0.252)
Spatial Lag	1.561***	+368.8%	2.051***	+693.0%	0.713	+104.4%	1.002**	+109.2%	1.017**	+109.5%
	(0.463)		(0.483)		(0.434)		(0.421)		(0.421)
ln Population	0.579***	+78.9%	0.782***	+118.4%	0.349	+41.2%	0.338	+75.6%	0.338	+80.1%
	(0.147)		(0.108)		(0.282)		(0.261)		(0.261)
Peace Months	−0.028***	−2.5%	−0.022***	−2.2%	−0.020***	−1.3%	−0.103**	−10.1%	−0.103**	−10.1%
	(0.002)		(0.001)		(0.002)		(0.044)		(0.044)
Constant	−7.310***		−11.633		−4.090		−5.558*		−7.722**
	(1.461)		(1.122)		(2.950)		(2.578)		(2.578)
N	116350		131625		1171		1171		1171

Municipality clustered Huber-White standard errors in parentheses.

†

p < .10, * p < .05, ** p < .01, *** p < .001; two-tailed test.

OPEN IN VIEWER

Table 2.

Conditional Rare-Events Logit Analysis of Massacres in Guatemala, 1959–96

	Model 5	RR	Model 6	RR	Model 7	RR	Model 8a	RR	Model 8b	RR
Insurgent Violence	0.025***	+106.0%					0.022***	+241.7%	0.022***	+241.6%
	(0.000)						(0.000)		(0.000)
% Indigenous			2.219***	+834.3			1.962*	+477.5%	1.962†	+515.5%
			(0.353)				(1.029)		(1.329)
State Control					−0.743†	−61.9%	−0.572*	−44.0%	−0.572*	−44.0%
					(0.401)		(0.282)		(0.282)
Spatial Lag	1.008*	+108.8%	1.493***	+354.8%	0.322	+102.9%	0.441	+103.8%	0.442	+104.1%
	(0.400)		(0.433)		(0.410)	(0.561)	(0.561)		(0.526)
ln Population	0.407***	+48.5%	0.551***	+73.2%	0.372	+40.8%	0.527*	+146.2%	0.537*	+153.3%
	(0.154)		(0.125)		(0.261)		(0.283)		(0.252)
Peace Months	−0.020***	−2.0%	−0.017***	−1.7%	−0.014***	−1.1%	−0.159***	−12.1%	−0.159***	−12.1%
	(0.002)		(0.001)		(0.003)		(0.048)		(0.048)
Constant	−6.840***		−10.152***		−6.065		0.022***		−0.159***
	(1.145)		(1.213)		(2.960)		(0.000)		(0.048)
N	116350		131625		1171		1156		1156

Municipality clustered Huber-White standard errors in parentheses.

†

p < .10, * p < .05, ** p < .01, *** p < .001; two-tailed test.

Table 1 presents the results from the first series of rare-events logit models estimating the probability of experiencing massacres in a given municipality month. Beginning the analysis, the first two models (Table 1, Models 1 and 2) review the two variables related to threats to the regime—Insurgent Violence and % Indigenous. The two hypotheses being tested, which specified that massacres should occur in response to state observations of the threats of insurgent violence and rival ethnicity, are strongly supported. Model 1 shows that increasing the number of individuals killed in a municipality by two standard deviations leads to a 72% increase in the predicted probability of a massacre being committed the next month. Model 2 shows that increasing the percentage of indigenous persons living in a municipality by two standard deviations increases the predicted probability of massacres more than seven-fold.

The analysis also evaluates the hypothesis put forward relating the projection of state control to the application of massacres. Model 3 evaluates this hypothesis without the other two independent variables. In the model, state controlled territory is predicted to be 63% less likely to experience massacres than territory not controlled by the state. This result holds for Models 4a and 4b, which include the two independent variables related to threats to the state. An astute observer will note that the number of observations drops off dramatically in these models due to the fact that data on civil patrols are only available for around 1,200 municipality-months. Still, these models present the opportunity to estimate the full set of independent variables and the results prove substantively interesting. For example, the municipalities for which data are available on civil patrols are located primarily in the jungle regions of northwest Guatemala, where the majority of the country’s violence took place and where the majority of the indigenous community lives. Rather than generalize to the wider population, the results for the models should be interpreted as being indicative of trends within this subsample. This is true for both Model 4a, which uses the subsample average rate of massacres in municipality months with civil patrol data, and Model 4b, which uses the full sample average rate of massacres. Control exerted by the state in a given municipality, as proxied by the control of the civil patrols, is both negatively and statistically significantly related to massacres across the models.

Where the state’s control was weakest, it was most prone to commit massacres. By contrast, massacres were significantly less likely to occur in municipalities controlled by state forces. Massacres were employed by counter-insurgent forces seeking to project control into contested regions, but were not employed to consolidate control where the state was dominant. These results indicate that accounting for when and where a state commits massacres requires that we examine not only those factors associated with ethnic rivalry, but also factors more commonly thought of as part of counter-insurgent strategy—insurgent violence and territorial control.

Addressing the marginal effects on the risk ratio of each, it appears at first blush as if the effects of the threat variables of ethnicity and insurgency are dramatically larger than the effects of the local-control counter-insurgency variable. But this is in part a function of the highly varied distribution of these variables. As noted above, a two standard deviation increase in insurgent violence signifies the killing of 40 additional people, while increasing the percentage of indigenous people by two standard deviations means increasing that figure by 72%. Looking at the effects of a one unit change from the mean in these variables on the risk ratio, we see that increasing the number of individuals killed by the insurgents from 0 to 1 leads to a 1.7% increase in the predicted probability of experiencing massacres, while increasing the percentage of indigenous people from 50% to 51% increases the risk ratio by less than 1%. Compared to these figures, the effect of state control is quite pronounced. State controlled territory was predicted to be less than half as likely to experience massacres compared to territory where state control was absent. ¹⁵

Table 2 examines the robustness of these results through the analysis of a series of conditional rare-events logit models estimating the probability of experiencing massacres conditional on the recording patterns documented in the dataset. Conditioning our estimates on the dataset’s recording practices does not appear to significantly alter many of the results. However, the results of the conditional analysis do present a few interesting challenges to the results presented above, which should be discussed.

Among municipalities that had similar patterns of recorded violence in the dataset, insurgent violence, the percentage of the rival ethnic group and state control are again significantly related to the use of massacres. However, comparing the results of Models 8a and 8b to Models 4a and 4b, which were identical but did not condition its estimates on the dataset’s recording practices, the estimated effects of each variable are dramatically different. The effect of a two standard deviation increase in insurgent violence is predicted in Model 8b to increase the predicted probability of experiencing genocidal massacres more than 240%. Meanwhile the causal effect associated with an increase in the municipality’s percentage of indigenous people, while still pronounced, drops by nearly half when we condition our estimates on the reporting patterns of the dataset. Examining the state control variable, we see that among municipalities experiencing similar recording patterns, state control is again negatively and statistically significantly related to the predicted probability of experiencing massacres. Municipalities controlled by the state are now predicted to have a 44% lower probability of experiencing massacres than those not controlled by the state. The sign and significance of the state control variable remain robust across model specifications, which lends greater support for the theories’ predictions.

In summation, massacres were much less likely to occur where the state controlled the local population and much more likely to occur where it did not. They were also much more likely to occur in settings where insurgents had recently perpetrated violence and where members of a rival ethnic group were highly concentrated. These results are robust across models that condition the estimates on the reporting patterns found in the dataset. To the extent that these conditional logit models help to control for sampling bias in the data, the consistent results between Table 1 and Table 2 suggest that the sampling bias in the data was either not severe or at least not severely correlated with the variables in these models. This should increase confidence that the results can be interpreted as representing underlying causal processes and not reporting patterns in the data.

When we look at the relative influence of the different variables, massacres appear to be shaped much more by the dynamics of insurgent violence and territorial control than the ethnic composition of a locality. This should come as something of a surprise to genocide scholars, who use ethnicity as part of the defining characteristic of their field and as one of the primary variables in their models. While the role of military strategy has been integrated into models of genocide for some time now (e.g. Sémelin, 2007; Midlarsky, 2005; Harff, 2003; Krain, 1997; Fein, 2000), the relative weight of counter-insurgency vis-à-vis ethnicity in producing massacres has yet to be identified. From the results produced in this analysis, it appears as though the timing and location of massacres are driven more by the objectives of eliminating insurgent violence and projecting state control over territory than by the objective of exterminating as many members of the rival ethnic population as possible.

Conclusion

This study investigated sub-national variation in the use of massacres by the state. The theory argued that massacres are best conceived of as a tactic of violence deployed by the state against collective targets. It identified two strategic settings that are likely to motivate the use of this tactic—eliminating threats to the state and projecting control over territory. The analysis showed robust correlations between massacres and the variables associated with (ethnic and insurgent) threats to the state and control over territory.

A model of local-level variation in the deployment of massacres has immediate practical implications. In the future, such models could be applied to formulate effective policies to end massacres. This would help intervening forces predict where massacres are likely to occur next and provide opportunities to prevent such violence. Intervening forces could target their operations in areas that are populated by rival ethnic groups, where insurgents are committing violence, and in areas poorly controlled by the state.

The results of this study also hold significant implications for future research. To begin with, it would be useful for future work to test the robustness of the results across different cases (e.g. Burundi, Burma, Rwanda). The theory presented above is presented in general terms and thought to be widely applicable. By focusing on the local-level variation in the deployment of massacres, it was able to hold constant many of the factors examined in the cross-national literature on mass violence. But the results may still be sensitive to charges that they are uniquely Guatemalan. Only by branching out and examining the deployment of massacres across a broader range of cases can we be certain that these findings are truly robust across different conflict structures.

The cross-national literature on mass violence could be strengthened through a deeper investigation of the strategic arguments presented in this article. Following the conventions of international law, genocide scholars have too commonly assumed that the deployment of violence during genocide was categorically distinct from the deployment of repressive violence in other forms of conflict. This need not be the case. As was shown in this study, the same tactic used to eliminate members of rival ethnic groups was used to collectively target settings of insurgent activity and to extend the state’s control over its territory. Additional work should be done to explicate the relationship between ethnicity, counter-insurgent strategy, and the use of massacres. While ethnic targeting certainly occurred and played a significant role in determining when and where the state engaged in massacres, results from this study show that massacres were also inspired by desires to reduce insurgent violence and project state control over territory. But these factors are certainly related. More research will be needed to better understand the relationship between ethnic divisions, counter-insurgent strategy, and collective targeting tactics, such as massacres.