The (in)effectiveness of torture for combating insurgency

Abstract

It is commonly believed that torture is an effective tool for combating an insurgent threat. Yet while torture is practiced in nearly all counterinsurgency campaigns, the evidence documenting torture’s effects remains severely limited. This study provides the first micro-level statistical analysis of torture’s relation to subsequent killings committed by insurgent and counterinsurgent forces. The theoretical arguments contend that torture is ineffective for reducing killings perpetrated by insurgents both because it fails to reduce insurgent capacities for violence and because it can increase the incentives for insurgents to commit future killings. The theory also links torture to other forms of state violence. Specifically, engaging in torture is expected to be associated with increased killings perpetrated by counterinsurgents. Monthly municipal-level data on political violence are used to analyze torture committed by counterinsurgents during the Guatemalan civil war (1977–94). Using a matched-sample, difference-in-difference identification strategy and data compiled from 22 different press and NGO sources as well as thousands of interviews, the study estimates how torture is related to short-term changes in killings perpetrated by both insurgents and counterinsurgents. Killings by counterinsurgents are shown to increase significantly following torture. However, torture appears to have no robust correlation with subsequent killings by insurgents. Based on this evidence the study concludes that torture is ineffective for reducing insurgent perpetrated killings.

Keywords

civil war human rights insurgency repression torture

Nearly all states engaged in counterinsurgency between 1981 and 2010 committed some form of torture (Cingranelli & Richards, 2011). The pervasive application of torture in such settings suggests an enduring belief that torture is an effective tool for combatting insurgent threats. Yet while torture remains one of the most widely utilized forms of political repression employed by counterinsurgents, the practice is widely debated, in part because there is a dearth of evidence on the outcomes of torture (Greenberg, 2006; Hafner-Burton & Shapiro, 2010).

Emerging scientific research on torture identified several structural and behavioral characteristics associated with both the application of government torture and its restriction (e.g. Powell & Staton, 2009; Hill, 2010; Conrad & Moore, 2010; Conrad & Ritter, 2013). However, the evidence documenting torture’s effects remains severely limited (e.g. CIA, 1963; Lehner, 2006; Rejali, 2007; Walsh, 2009; Regan, 2009; Johnson & Ryan, 2012). As a result, contemporary decisions to engage in torture have not been grounded in systematic analysis, but have instead been made based largely on the presentation of anecdotal case evidence and the plausibility of various ‘what if’ scenarios (e.g. Benjamin, 2009; Finn, Warrick & Tate, 2009).

These limitations remain in spite of the extraordinary growth in efforts to understand insurgent and counterinsurgent violence (e.g. Kalyvas, 2012). Recent work has shown how civil war violence impacts a diverse set of variables, including reprisal killings, public health, territorial control, and political support (e.g. Condra & Shapiro, 2011; Chamarbagwala & Morán, 2011; Kocher, Pepinsky & Kalyvas, 2011; Lyall, Blair & Imai, 2013). However, the existing work has remained largely focused on a particular form of violence – killings – with limited attention paid to the effects of other violent acts. Thus, little is known about how other forms of violence, such as torture, impact counterinsurgent outcomes.

Using standard assumptions to model the dynamics of an insurgent–counterinsurgent conflict, this study identifies a series of theoretical mechanisms linking torture perpetrated by counterinsurgents to subsequent insurgent and counterinsurgent behavior.¹ As applied in counterinsurgent settings, torture is argued to be ineffective for reducing killings perpetrated by insurgents both because it fails to reduce insurgent capacities and because it can increase insurgent incentives for future killings. The theory also links torture to more pathological outcomes. Most discussions of torture explicitly consider the impact of this form of state coercion on the behavior of those challenging the government. However, torture represents only one part of the repressive repertoire and it is frequently connected to other government tactics. As argued, not only is torture ineffective for reducing killings by insurgents, but engaging in torture is also expected to be associated with an increase in killings perpetrated by counterinsurgents.

These arguments are evaluated using monthly municipal-level data on political violence from Guatemala’s counterinsurgency and a matched-sample, difference-in-difference identification strategy. The data were compiled from 22 distinct press and NGO sources as well as thousands of interviews. The analysis utilizes these data to estimate the relationship between torture and subsequent killings by insurgents and counterinsurgents. Results demonstrate that (1) torture is not related to any significant decrease in killings by insurgents and (2) torture is associated with increases in killings by counterinsurgents.

The article proceeds as follows. The next section details theoretical arguments connecting torture to subsequent patterns of killings perpetrated by insurgents and counterinsurgents. The third section reviews the data, case selection, and research design. The fourth section presents the empirical analysis. The results are summarized in the conclusion and implications are drawn connecting the findings to contemporary debates about torture as well as academic research on the topic.

The ineffectiveness of torture for combating insurgency

The United Nations Convention against Torture (UNCAT) defines torture as

[a]ny act by which severe pain or suffering, whether physical or mental, is intentionally inflicted on a person for such purposes as obtaining from him or a third person information or a confession, punishing him for an act he or a third person has committed or is suspected of having committed or intimidating or coercing him or a third person, or for any reason based on discrimination of any kind, when such pain or suffering is inflicted by […] a public official or person acting in an official capacity. (United Nations, 1984: 1)

The UNCAT definition suggests three possible motives for this behavior.

First, governments use torture to extract information. States torture captured enemies when they believe that those tortured will release secrets which might aid the state in its attempts to identify and eliminate domestic threats (Wantchekon & Healy, 1999). Second, states engage in torture in order to punish. The desired result is to create a link between disobedient behavior and pain, thereby reinforcing legal norms and associating transgression with negative sanctions (Foucault, 1979). Finally, states engage in torture in order to instill terror in the surrounding population (Walter, 1969).

By all three accounts, torture is a tool for managing dissent. From this perspective, torture can be viewed as effective when it leads to subsequent reductions in insurgent perpetrated killings of counterinsurgents or civilians. Justifications for torture do not rest on the contention that engaging in torture will reveal information, but on arguments that engaging in torture will allow state agents to somehow reduce insurgent perpetrated killings (e.g. Bagaric & Clarke, 2007; Derschowitz, 2003). If torture cannot produce discernible reductions in killings perpetrated by insurgents, then any immediate effects of torture, including the revelation (or non-revelation) of information, are of little consequence.

Understanding how torture impacts subsequent killings requires a basic model of insurgency that can account for when and where torture is applied in the context of counterinsurgent struggle, as well as how torture influences the incentives or constraints on insurgent and counterinsurgent perpetrated killings. In this setting, members of an armed insurgency aim to achieve concessions or revolution by combining direct violence against state institutions with efforts to limit the governability of particular geographic areas of the country (Galula, 1964). Insurgents win by exploiting various asymmetries inherent in the situation (Mack, 1975). Of greatest importance are asymmetries of information (Kalyvas, 2006). Counterinsurgents are generally uncertain who among the population belongs to the insurgency. Thus, while counterinsurgents seek to remove the insurgent organization by deterring or eliminating individuals willing to join these organizations, counterinsurgent forces are often uncertain as to whom they should target with such policies.

Two basic tenets of the model are of critical importance for understanding torture and its impacts. First, insurgents and counterinsurgents are strategic actors capable of using violence to manage information flows. Counterinsurgents can generate information by coercing or inducing civilians to inform on the insurgents, while insurgents simultaneously apply their own blend of force and incentives (e.g. selective violence, protection) to prevent defection. Second, because information is imperfect, insurgents and counterinsurgent forces may respond to flows of information with violence even when such information is inaccurate. The fog surrounding counterinsurgent struggles can motivate the warring parties to engage in strategic behaviors that often appear counterproductive in retrospect.

Within this context, torture is most likely to be applied when (1) counterinsurgents possess limited information on who is or is not a member of the insurgency, (2) insurgent attacks are ongoing, and (3) the population is not sufficiently cowed by fear to provide information or reduce insurgent support.² For counterinsurgents, these are situations in which alternative means for restoring order have either proven ineffective or are often closed off entirely. For example, when counterinsurgents control an area, civilians are more willing to provide information to state forces, typically in exchange for security (Kalyvas, 2006). In these settings, state forces possess greater information to target insurgents and killings perpetrated by insurgents are less common. Such situations present counterinsurgents with few incentives for committing torture, as other mechanisms for restraining insurgent perpetrated killings have proven effective.

However, when counterinsurgents are not firmly in control, the distribution of information is generally asymmetric, with insurgents holding greater local-level knowledge than counterinsurgents. Here, counterinsurgent forces can suspect an insurgent presence, but they do not possess sufficient information to target the insurgency more directly. Working from a position in which insurgent perpetrated killings are ongoing, but where other avenues for information collection and deterrence are either unavailable or ineffective, counterinsurgents commit torture to increase the flow of information, target subsequent violence at insurgents, and reduce insurgents’ ability to kill civilians or counterinsurgents.

Hypothesis 1: Torture will be associated with decreases in killings perpetrated by insurgents.

Yet despite these objectives, there are reasons to expect torture will fail to produce the results its proponents desire. To begin, torture by itself is unlikely to reduce killings perpetrated by insurgents. While it has been argued that torture can create a sense of fear that paralyzes potential insurgents (e.g. Walter, 1969), such arguments do not address the range of strategic options available to a repressed population seeking safety in the midst of shifting security conditions. Even among the truly fearful, torture often acts as a transformative event, inspiring a backlash among the repressed (Hess & Martin, 2006). Rather than being deterred from action against the regime, victim groups may turn to the insurgency both because joining increases their chances for survival (Kalyvas & Kocher, 2007) and because insurgent groups offer opportunities for organized opposition to future repressive action (Wood, 2003).

Any ensuing repression that torture generates can also be expected to be ineffective for reducing killings perpetrated by insurgents. Information-starved counterinsurgents often torture individuals who have little or no knowledge of the insurgency, which can lead to poorly targeted repression and an insurgent backlash.³ In order to stop their suffering, tortured individuals who do not know the answers to the torturer’s questions may decide to give up the names of innocent victims. They may even seek to turn subsequent repression to their advantage by identifying individuals with whom they share private grudges (Kalyvas, 2006).

Insurgencies, meanwhile, are constantly adapting. Built into the insurgents’ modus operandi are plans for minimizing the damage inflicted should one of their members be tortured. These include both proactive steps, such as compartmentalizing information and organizing in a cell structure, and reactive steps, such as shifting to new hideouts and altering their tactical repertoire in the event that one of their members is captured (see Rejali, 2007: Ch. 7).⁴ The result is to counter potential post-torture repressive actions taken to target the insurgents or limit their capacity to perpetrate attacks on the state or civilians.

Torture can even incite an increase in killings perpetrated by insurgents. Individuals who voluntarily join an insurgency enter the organization committed to its goals and aware of the possibility that they may experience physical violence (McAdam, 1990; Zwerman & Steinhoff, 2005). Such individuals are unlikely to be deterred by torture. Instead, to shore up their position, insurgent forces may respond to torture by exploiting sites where they can commit violence against counterinsurgents or by targeting civilians whom they suspect might be vulnerable to defection.

However, while torture may not be associated with decreased killings perpetrated by insurgents, there are reasons to expect that torture will be associated with an increase in killings by counterinsurgents.⁵ The mechanisms here are varied. First, logistically, counterinsurgent forces have to be present to commit torture. Indeed, the presence of such forces may be related to strategic decisions leading to a buildup of coercive capacity that simultaneously facilitates torture and subsequent killings. Here the relationship between torture and subsequent killings could be part of a broader escalation in repressive capacity. However, increasing the capacity for repression does not necessarily imply that counterinsurgent forces will employ it.⁶

Second, it might be the case that decisions to engage in torture are related to changes in the strategic value of a locality that also influence subsequent counterinsurgent decisions to engage in killings in the area. In this case, torture would be related to subsequent killings by counterinsurgents because both actions were related to some exogenous change in local strategy.⁷ But this account leaves no space for counterinsurgents to engage in torture and then adapt strategically to any information it provides.

Most immediately, torture can yield new information as victims attempt to limit their exposure to violence by appeasing torturers. In this way, torture opens up new opportunities and incentives for engaging in killings. As argued, counterinsurgents are generally at an informational disadvantage when committing torture. In the fog that surrounds such settings, counterinsurgents are aware of the insurgent presence, but are presented with few opportunities for controlling killings by insurgents. If counterinsurgents choose to commit torture in an effort to extract information and subdue the insurgency, we can infer that state forces will act on any information that is revealed.

While this information may be of questionable validity, counterinsurgent forces are often in a poor position to separate truth from falsehood.⁸ Verifying the claims made by victims of torture takes time and in situations in which time is at a premium, torturers may feel pressure to act without corroboration (Derschowitz, 2003). And while some steps taken following torture, such as rendition and detention, may be relatively less violent, counterinsurgents are highly specialized killers. In settings where insurgents are killing counterinsurgents and civilians, information revealed through torture is likely to motivate equally violent actions by counterinsurgents.⁹

Hypothesis 2: Torture will be associated with increases in killings perpetrated by counterinsurgents.

Research design

One explanation for the lack of success empirically evaluating the effects of torture is that estimating torture’s causal effects is an effort fraught with methodological challenges. Two threats to causal inference are particularly noteworthy and need to be discussed: the observation problem and the endogeneity problem. The observation problem presents challenges because (1) torture generally takes place under conditions of secrecy, and (2) systematically collecting data on insurgent and counterinsurgent behavior is extraordinarily difficult. The effect is to constrain observations of torture and its potential outcomes, as researchers must counter both the strategies employed by governments to hide the use of torture and the challenges associated with collecting data in a conflict-ridden or post-conflict society.

At the same time, we know that the processes generating political violence are highly endogenous (Kalyvas, 2006). Generating counterfactual claims of what an observed torture site would look like had torture not taken place is difficult because the strategies of the two warring parties are influenced by earlier acts of violence. As a result, decisions to commit torture or to commit subsequent killings are both likely related to prior killings.

Countering the observation and endogeneity problems requires combining detailed records of torture and other forms of political violence with methodological techniques designed to address biases resulting from the non-random application of torture. The rest of this section explains the study’s case selection, data, operationalization, and identification strategy.

Case selection

This study employs micro-level data on political violence from Guatemala’s civil war (1977–94) to study the relationship between torture and subsequent killings by insurgents and counterinsurgents. The insurgency in Guatemala, largely organized under the banner of the Army of the Poor (Ejército Guerrillero de los Pobres; EGP), sought to overthrow the military government and impose a communist regime through a combination of popular mobilization and a hit-and-run strategy targeting government forces (Stoll, 1993). Throughout the war, the Guatemalan state engaged in high levels of political repression, which included the widespread application of torture (Ball, Kobrak & Spirere, 1999; CEH, 1999).

According to Guatemala’s Commission for Historical Clarification (Comisión para el Esclarecimiento Histórico; CEH, 1999: 489), torture was committed, ‘with the primary objective of obtaining information.… [A]s an additional purpose, it was used to punish the victim, and to generate terror in the population.’¹⁰ When it came to torture, the Guatemalan military was both highly skilled in the techniques of torture and unconstrained in its application (Schirmer, 1998). Using lessons learned in Vietnam, hundreds of Green Berets and CIA operatives were sent to Guatemala during the 1960s to restructure the military’s intelligence apparatus and to train the agents of the military intelligence unit, the G-2, in the application of different torture techniques.¹¹ At the same time, when civilian efforts were made to curtail the military’s operational independence, military power was able to subvert these efforts through a series of coups and states of emergency (Schirmer, 1998). The result was to produce a military force with both the technical knowhow to engage in torture and the ability to apply that knowledge wherever they saw fit. Among other tactics, agents of the Guatemalan military forced the victims to stand hooded for hours or days, forced them to eat excrement, forced them to stay awake for days at a time, refused to give them food or water, subjected them to electric shocks, stripped them naked, burned them with cigarettes, suspended them from chains, sexually abused them, submerged them in water, cut them, and broke their fingers (CEH, 1999: 472–489).

In the end, the overwhelming levels of repression imposed by the Guatemalan state worked to effectively eliminate insurgent activity within the country (Stoll, 1993). In this way, the country potentially serves as a most likely case for observing the effectiveness of torture as an instrument of counterinsurgency. Guatemala is a case where (1) torturers were expertly trained and unrestrained and (2) the repressive actions of the government ultimately undercut dissident mobilization and put an end to the insurgency. Alternatively, it may be the case that examining aggregate indicators of repression and dissent tell us little about the function of individual tactics. While we know that increases in overall levels of repression were related to decreases in overall levels of killings perpetrated by insurgents, without disaggregating the different repressive tactics engaged in by the state it is impossible to know whether torture operated to reduce killings by insurgents.

Data

The empirical analyses conducted below employ municipal-level events data on political violence in Guatemala collected from 17 different domestic press sources, the publications of five human rights organizations (Human Rights Watch, Amnesty International, and three Guatemalan human rights groups), and the transcriptions of more than 5,000 interviews with victims of the conflict (Ball, 1999). The dataset was constructed by the Center for Human Rights Research (CIIDH), a nongovernmental organization based in Guatemala City, during the truth and reconciliation processes that took place in Guatemala in the 1990s (Ball, Kobrak & Spirere, 1999). The project, which was directed by Patrick Ball and the Human Rights Data Analysis Group for the purpose of identifying incidents of human rights abuses committed by state forces, insurgents, and other militant actors, began with a thorough review of daily press reports and published human rights documents. The press sources were then supplemented with thousands of interviews, which began in 1994 among the survivors living near the Mexican border (Ball, Kobrak & Spirere, 1999). Using snowball techniques based on these interviews, the team identified areas of high reported violence throughout the country and spent two years interviewing victims and witnesses in these areas and other reported sites of violence (Ball, Kobrak & Spirere, 1999). Events from the different sources were then combined into a single database. Events reported in multiple sources were identified and duplicates were removed.

Taken together, the sources combine to produce more complete coverage of the variations in type, scope, intensity, and location of political violence than alternative events databases constructed around a single data source. The combination of multiple, partially overlapping sources cataloging political violence presents a more accurate portrayal of the distribution of violence across Guatemala than any one source would have on its own and helps makes the dataset one of the most complete sets of event data from a civil war available to date.¹² In total, the database records nearly 18,000 incidents of political violence, including massacres, selective killings, disappearances, and torture.¹³ More than 45,000 individual victims of political violence are identified, of whom more than 500 were victims of torture.¹⁴

Operationalization

Dependent variables

The units of analysis for the models below are municipality-months.¹⁵ Two sets of dependent variables are examined: killings perpetrated by insurgents and counterinsurgents. For both, the empirical analysis separately models the number of killings locally and regionally. For local killings, the models estimate counts of killings committed by the insurgents and counterinsurgents within a particular municipality-month.¹⁶ For regional killings, the models estimate the number of killings occurring across the country inversely weighted by the number of kilometers between the site of the killing and the municipality being examined (Ward & Gleditsch, 2008).¹⁷

Independent variable

The independent variable, torture, is identified by the CIIDH/AAAS dataset. The dataset used the following definition, which roughly matches the UN definition put forward above: ‘inflicting pain on a person in a premeditated and systematic manner … [where] the apparent motive is to obtain information or confessions, whether they are true or not’ (CIIDH, 1996: 1–9). Months in which the state engaged in at least one act of torture in a municipality are scored 1; they are scored 0 otherwise.

Identification strategy

To estimate torture’s association with subsequent killings by insurgents and counterinsurgents, this study employs a design in which assignment to treatment (i.e. torture) is based on self-selection, but where specific threats to causal inference are explicated and attended to through pre-test measurement and a control group over which the design has considerable control. In the study, I attempt to identify torture’s association with subsequent killings by focusing on the types of units that are treated and examining how changes in those units compare to changes in a sample of similar, untreated units (compare Lyall, 2009, 2010). In the absence of random assignment, it will be impossible to rule out all possible threats to causal inference. However, what the empirical analysis can do is analyze the expected outcomes of torture where torture occurred, while discrediting as many plausible rival hypotheses as possible (Campell & Stanley, 1963).

This study employs a matching procedure to model the assignment of treatment (i.e. torture) and generate a control group in which the latent probability of observing torture is comparable to the treatment group. While all municipality-months could be considered ‘at risk’ of experiencing torture within the context of the Guatemalan counterinsurgency campaign, the probability that torture was assigned at any given time or place varied significantly. Matching pre-processes the data to pair each municipality-month experiencing torture with a sample from the group of municipality-months not experiencing torture to reduce imbalances in the observable covariates thought to influence assignment to treatment (Ho et al., 2007). The result is to minimize residual imbalance across the distribution of covariates, which ensures common support while also matching the underlying propensity for torture between the treatment and control units (Imbens, 2009).

The assignment model used to identify the latent probability of treatment is based on a series of variables thought to influence state strategy. The variables can be broadly clustered into five categories. First, there are deep structural variables that broadly influenced insurgent and counterinsurgent strategy. These include the log of the municipality population and the percentage of the municipality that was indigenous as well as the department in which a municipality was located, and the year in which the assignment month took place. The first two variables were selected because population and ethnic divisions have previously been demonstrated to impact the dynamics of an insurgency. The latter two were included to incorporate any confounding factors potentially operating at the departmental or annual level. Second, there are variables associated with regional strategy. These include spatial lags of torture, killings perpetrated by counterinsurgents, and killings perpetrated by insurgents. The spatial lags measure ongoing counts of different forms of violence inversely weighted by how far they occurred from a given municipality (Ward & Gleditsch, 2008). They were included to attend to the spatial clustering of insurgent and counterinsurgent strategies. Third, there are variables associated with other counterinsurgent tactics occurring during the assignment month, the use of which could potentially influence decisions to use torture and subsequent killings. Specifically, two measures are used to identify other forms of ongoing state repression: selective state killings and state massacres, both measured at time t. Fourth, there is a variable to account for temporal dependence in the use of torture: a count of the number of torture events occurring in the municipality during the previous months. Finally, the matching process includes dynamic measures of the killings perpetrated by insurgents and counterinsurgents in a municipality in the recent past (i.e. T–6 through T–1) to account for the tactical setting occurring locally as state forces make decisions whether to engage in torture in a given municipality-month. Including these pre-treatment levels of killings is crucial both for modeling the assignment of treatment and for balancing time trends for the difference-in-difference estimates.

Treatment and control municipality-months were matched 1:1 with replacement using genetic matching to maximize balance across the covariates (Sekhon & Mebane, 1998; Sekhon, 2011). To ensure that members of the treatment group were not smuggled into the control group at later points in time, months without torture in treated municipalities were dropped during the matching procedure if they occurred within the six months prior to or following an instance of torture.

Table I displays the mean counts of the assignment variables for municipality-months experiencing and not experiencing torture. Among other differences, the first 12 rows of Table I show that for the unmatched sample, municipality-months experiencing torture witnessed much higher rates of killings perpetrated by both counterinsurgents and insurgents in the preceding months. For example, during the previous month, counterinsurgents committed killings in 44% of all municipalities in which counterinsurgents engaged in torture. By comparison, less than 1% of municipalities where torture did not occur experienced killings by counterinsurgents during the previous month. If the higher rates of killings in municipality-months experiencing torture not only helped provoke torture but also inspired increases in other forms of violence, the models could severely overestimate torture’s association with subsequent killings. Alternatively, if killings are generally mean-stationary, such that levels of killings fluctuate and then revert back to a mean, comparing the more deadly periods in which torture was applied to less deadly periods where it was not could lead to an underestimate of the association between torture and subsequent killings.

The table also displays those same rates for the treatment and control group in the post-matching sample. For both pre-treatment levels of killings and the other assignment covariates, the remaining imbalances between municipalities experiencing and not experiencing torture now appear quite small. The table includes two balance statistics for both the full sample and the matched sample: the raw mean eQQ difference and the scaled maximum eCDF difference. Both statistics compare the empirical distributions of the different sets of data and help to identify imbalance across covariates. They display similar but slightly different information, with the raw mean eQQ difference indicating the mean difference on the eQQ plot of the treatment and control groups and the scaled maximum eCDF difference providing the maximum difference in the eQQ plot presented on the scale of the variable in order to improve comparability across the different variables.¹⁸ The two statistics were selected because such tests accurately characterize the sample populations and because they are less sensitive to sample size than P-tests or the Kolmogorov-Smirnov test (Imai, King & Stuart, 2008; Diamond & Sekhon, 2013). As can be seen in the table, across the distributions of the covariates, the level of imbalance is strongly reduced in the matched sample. This is particularly true for the prior levels of killings perpetrated by counterinsurgents and insurgents, which proves crucial for estimating the difference-in-difference models.

The difference-in-difference models employ the matched sample to compare changes in rates of killings by insurgents and counterinsurgents in municipalities experiencing and not experiencing torture. The idea is to use rates of change in sites unaffected by torture as an estimate for what killings would occur had torture not taken place. One key assumption underlying these models is that pre-intervention temporal trends in the treatment and control groups are roughly comparable (Angrist & Pischke, 2009). If the two groups have inconsistent temporal trends prior to assignment, estimates of the treatment effect may remain biased (Ashenfelter, 1978). Combining the matched sample with difference-in-difference estimation ensures that the observable differences in pre-intervention histories of killings are comparable between the treatment and control groups (e.g. Card & Sullivan, 1988).

Formally the difference-in-difference model is constructed as:

Y i, s, t = α + γ D s + λ T t + δ (D s * T t) + ξ s + μ t

The difference-in-difference model estimates the number of killings (Y) occurring both before and after the intervention (i.e. T–1 and T+1) as a function of two dummy controls (D and T) and their interaction (Angrist & Pischke, 2009). The subscript i refers to unit level variation, t refers to time period, and s refers to treatment group. The first dummy variable (γD) is measured 0 for the control group and 1 for the treatment group across all time periods. It identifies any systematic, time-invariant and unobservable differences that exist between the two groups.¹⁹ The second (λT) is a time-period fixed effect, coded 0 for periods preceding the intervention and 1 for periods following intervention. It controls for any temporal trends that are common to both the treatment and control groups.

By interacting these two variables (δ(D*T)), the analysis identifies the average treatment effect on the treated units (ATET), which represents the amount of change in municipalities experiencing torture compared to change in municipalities where torture did not occur, controlling for systematic and unobservable

Table I.

Balance statistics

Variable	Full sample					Matched sample
	Sample mean	Treatment mean	Control mean	Raw mean eQQ difference	Scaled maximum eCDF difference	Control mean	Raw mean eQQ difference	Scaled maximum eCDF difference
Killings perpetrated by counterinsurgents
T–1	0.02	0.44	0.01	0.43	0.21	0.43	0.05	0.03
T–2	0.02	0.39	0.01	0.39	0.19	0.36	0.03	0.01
T–3	0.02	0.40	0.01	0.39	0.20	0.36	0.03	0.02
T–4	0.02	0.38	0.01	0.37	0.19	0.38	0.00	0.00
T–5	0.02	0.34	0.01	0.33	0.17	0.33	0.01	0.01
T–6	0.02	0.40	0.01	0.40	0.20	0.37	0.03	0.01
Killings perpetrated by insurgents
T–1	0.01	0.03	0.01	0.02	0.01	0.03	0.00	0.00
T–2	0.01	0.05	0.01	0.05	0.03	0.05	0.00	0.00
T–3	0.01	0.03	0.01	0.03	0.02	0.02	0.01	0.00
T–4	0.01	0.03	0.01	0.03	0.02	0.02	0.01	0.00
T–5	0.01	0.02	0.01	0.01	0.01	0.02	0.03	0.01
T–6	0.01	0.03	0.01	0.03	0.02	0.03	0.00	0.00
Prior torture	0.30	3.24	0.23	2.94	0.14	1.68	1.34	0.06
Spatial lag of torture	0.03	0.09	0.03	0.09	0.25	0.08	0.05	0.13
Spatial lag of killings perpetrated by counterinsurgents	3.16	18.61	3.12	19.86	0.31	20.12	4.69	0.05
Spatial lag of killings perpetrated by insurgents	0.47	6.93	0.48	14.39	0.19	21.54	12.98	0.05
Massacres by counterinsurgents T	0.01	0.28	0.01	0.28	0.14	0.27	0.01	0.01
Killings by counterinsurgents T	0.01	0.75	0.01	0.74	0.37	0.70	0.05	0.02
% Indigenous	0.52	0.74	0.52	0.23	0.19	0.70	0.06	0.07
Log population	9.17	9.89	9.13	0.77	0.27	9.86	0.13	0.06
Year	1985.56	1983.50	1985.60	2.65	0.16	1983.20	0.54	0.04
Department	11.97	7.62	12.09	4.51	0.21	7.48	1.08	0.05

differences between the treatment and control groups, as well as for any temporal trends in the data.

Analysis

Tables II and III present the results of matched-sample difference-in-difference estimates for killings perpetrated by insurgents and counterinsurgents.²⁰ Table II presents a two-month difference-in-difference design (i.e. one month pre-and post-), while Table III presents a

Table II.

Difference-in-difference estimates of killings perpetrated by insurgents and counterinsurgents: 1 month pre- and post-

	Killings perpetrated by insurgents				Killings perpetrated by counterinsurgents
	1A		1B		2A		2B
	Local		Regional		Local		Regional
	Point estimate (Standard error)	95% CI	Point estimate (Standard error)	95% CI	Point estimate (Standard error)	95% CI	Point estimate (Standard error)	95% CI
Δ ATET (T+1)	0.21 (0.22)	–0.22, 0.64	–5.21 (8.80)	–22.52, 12.10	86.62* (36.77	14.29, 158.96	4.23 (4.35)	–4.33, 12.79
γ Treatment group	0.02 (0.02)	–0.02, 0.06	–6.92 (6.32)	–19.37, 5.52	–8.43 (34.80)	–76.91, 60.05	–4.36 (3.89)	–5.22, 3.28
λ T+1	–0.01 (0.01)	–0.03, 0.04	2.95 (8.13)	–13.05, 18.95	–75.31* (27.23)	–128.89, −21.73	0.99 (3.16)	–5.22, 7.21
Intercept	0.01 (0.01)	–0.01, 0.02	12.81* (5.68)	1.63, 24.02	99.82*** (26.92)	46.84, 152.80	21.54*** (2.59)	16.44, 26.64
N	660		660		660		660

Unit clustered robust standard errors in parentheses. *p < 0.05, **p < 0.01, ***p < 0.001.

Table III.

Difference-in-difference estimates of killings perpetrated by insurgents and counterinsurgents: 6 months pre- and post-

	Killings perpetrated by insurgents				Killings perpetrated by counterinsurgents
	3a		3b		4a		4b
	Local		Regional		Local		Regional
	point estimate (Standard error)	95% CI	Point estimate (Standard error)	95% CI	Point estimate (Standard error)	95% CI	Point estimate (Standard error)	95% CI
δ ATET (T+1)	0.08 (0.23)	–0.37, 0.53	–2.78 (6.56)	–15.70, 10.13	86.91*** (25.41)	36.92, 136.90	12.11** (4.39)	3.46, 20.75
δ ATET (T+2)	31.88 (31.76)	–30.60, 94.36	9.07* (3.51)	2.08, 15.92	63.61* (25.85)	12.75, 114.47	20.45*** (4.83)	10.93, 29.97
δ ATET (T+3)	31.82 (31.76)	–30.65, 94.31	6.83 (3.69)	–0.44, 14.11	33.99 (22.76)	–10.78, 78.76	29.92*** (5.40)	19.28, 40.55
δ ATET (T+4)	31.85 (31.76)	–30.63, 94.32	9.91*** (2.71)	4.57, 15.25	23.05 (24.20)	–24.56, 70.66	28.72*** (5.67)	17.55, 38.89
δ ATET (T+5)	31.86 (31.76)	–30.61, 94.34	10.13*** (2.88)	4.45, 15.82	50.08* (24.84)	1.20, 98.58	28.03*** (5.45)	17.29, 38.76
δ ATET (T+6)	31.84 (31.76)	–30.64, 94.32	13.24*** (3.58)	6.18, 20.30	46.60* (21.18)	4.92, 88.27	26.59*** (5.37)	16.01, 37.17
γ Treatment Group	0.14 (0.07)	–0.01, 0.30	–9.34*** (2.47)	–14.22, –4.47	–8.72 (17.92)	–43.99, 26.55	–12.24*** (3.69)	–19.52, –4.96
λ T+1	–0.02 (0.02)	–0.05, 0.01	4.41 (6.27)	–7.93, 16.76	–67.11*** (11.29)	–89.33, –44.89	–10.41*** (3.21)	–16.73, –4.07
λ T+2	–0.03* (0.01)	–0.06, –0.01	–9.22*** (3.24)	–15.59, –2.84	–62.67*** (11.86)	–85.99, –39.36	–14.62*** (3.48)	–21.48, 7.79
λ T+3	0.01 (0.03)	–0.04, –0.07	–7.41* (3.52)	–14.34, –0.48	–52.77*** (12.02)	–76.45, –29.13	–21.03*** (3.19)	–27.31, –14.75
λ T+4	–0.02 (0.02)	–0.05, 0.01	–10.55*** (2.49)	–15.44, –5.62	–37.89*** (14.55)	–66.56, –9.62	–20.88*** (3.41)	–27.58, –14.17
λ T+5	–0.03* 0.01	–0.06, –0.01	–10.15*** (2.58)	–15.22, –5.06	–42.24*** (13.26)	–68.26, –16.24	–23.01*** (2.91)	–28.75, –17.28
λ T+6	–0.02 (0.02)	–0.01, 0.30	–9.78*** (2.47)	–14.22 –4.47	–43.82*** (9.71)	–62.93, –24.71	–23.86*** (3.03)	–29.84, –17.89
Intercept	0.03* (0.01)	0.01, 0.06	11.35*** (2.41)	6.61, 16.10	91.62*** (12.21)	67.62, 115.63	32.94*** (2.85)	27.52, 38.36
N	3,960		3,960		3,960		3,960

Unit clustered robust standard errors in parentheses. *p < 0.05, **p < 0.01, ***p < 0.001.

12-month difference-in-difference design (i.e. six months pre- and post-). For each independent variable included, the coefficient estimates are presented along with the standard error, level of statistical significance, and 95% confidence intervals.²¹

Models 1a and 1b in Table II estimate torture’s association with killings perpetrated by insurgents. The first thing to note is that the ATET does not appear statistically significant in either model. The point estimate is positive, indicating that where applied, torture is on average associated with increased killings by insurgents during the month following torture. However, the 95% confidence intervals for local and regional killings perpetrated by insurgents overlap with zero.

The results from these first two models cast serious doubt on the belief that torture is an effective means for reducing killings perpetrated by insurgents. If torture is an effective instrument of counterinsurgency, torture would be expected to be associated with a decrease in the number of individuals killed by the insurgents. However, there is no evidence from the first two models to suggest that this is the case.

Models 3a and 3b in Table III replicate these analyses using a six-month pre-and post- difference-in-difference design. When examining insurgent behavior over a longer time frame, the evidence again refutes the contention that torture is associated with reductions in killings perpetrated by insurgents. Examining killings locally in Model 3a, torture appears unrelated to patterns of killings by insurgents in a municipality over the next six months. Looking at regional patterns in Model 3b, torture has a positive and statistically significant association with subsequent killings perpetrated by insurgents. This evidence suggests that torture may actually have been counterproductive, increasing the number of killings committed by insurgents in the surrounding areas over the subsequent six months. While the results do not provide strong, consistent evidence supporting the conclusion that torture correlates with significant increases in killings perpetrated by insurgents, there is no evidence that torture is related to a decrease in killings by insurgents. I return to this finding in the conclusion.

Turning to the relationship between torture and killings perpetrated by counterinsurgents, Models 2a and 2b in Table II examine rates of killings by counterinsurgents in the month preceding and following torture. Model 2a shows that when compared to the matched sample, municipalities where torture occurred experienced a significant increase in killings committed by counterinsurgents during the month following torture. Substantively, the predicted number of killings perpetrated by counterinsurgents in a municipality is expected to increase by between 15 and 159 victims during the month following torture.

Models 4a and 4b in Table III examine changes over the longer 12-month period. When looking at the models, torture is again associated with increases in killings perpetrated by counterinsurgents in the municipalities experiencing torture. The most robust increase occurs in the month immediately following torture, in which the counterinsurgent forces are expected to kill between 37 and 137 additional victims in the municipality where torture occurred. Looking regionally, Model 4b estimates that following torture, killings by counterinsurgents increased in the surrounding region as well. The results indicate that increases in regional killings peaked during months two through four, suggesting a lag not found in local killings by counterinsurgents.

In summation, the results support the expectations that torture is ineffective for reducing killings perpetrated by insurgents. Across the models, torture was either unrelated to killings by insurgents or associated with positive increases in such killings. The results also provide evidence confirming the expectation that torture is positively associated with increases in killings perpetrated by counterinsurgents. Counterinsurgent forces killed greater numbers of individuals following torture than they were predicted to have killed had torture not taken place.

Remaining concerns and additional analyses

The models include two important controls, one for systematic unobservable differences between the treatment and control groups and one for systematic differences that exist in both groups before and after the intervention. Interestingly, across the models, the Treatment Group variable is only significant in Models 3b and 4b (both of which deal with regional patterns of killings), which suggests that there are few significant, systematic, unobservable differences between the matched samples of municipality-months experiencing and not experiencing torture. In both cases, the treatment group was systematically less violent than the control group, effectively biasing the results against the findings that torture is related to increases in killings committed regionally by insurgents and counterinsurgents. At the same time, the post-intervention period control is significant across four of the six models. This suggests that while the matching process has identified comparable samples of treatment and control municipality-months, both groups experience common temporal dynamics that would bias the results if not controlled for.

One remaining concern is that the patterns are being driven by some other time-variant difference between the treatment and control groups that made the treatment groups both more likely to receive torture and more likely to witness state killings later on. An example would be if the state somehow found greater strategic interest in the municipalities experiencing torture just before or just following torture.²² Given the observational nature of the data, this possibility cannot be ruled out entirely. What can be done is to narrow down on the relationship between torture and killings committed by counterinsurgents and attempt to demonstrate that the identified increase in killings by counterinsurgents happened immediately following torture, rather than in the preceding or subsequent months. An empirical strategy for such an investigation would be examining ex ante changes in state strategy through placebo tests (e.g. Ladd & Lenz, 2009). Another would be Granger causality tests examining torture’s association with killings perpetrated by counterinsurgents compared to torture’s association with any potential ex post strategic changes (Angrist & Pischke, 2009: 237–238). The results of three such tests are presented in Table IV.

The first two models presented in Table IV are placebo tests examining rates of change in local killings by counterinsurgents that follow a placebo treatment occurring either two months before (Model 5a) or one month before (Model 5b) torture was actually observed. Both models replicate the analysis in Model 3a above with a recoded pre-treatment placebo. If there was some strategic change that took place prior to the use of torture that might have motivated both the use of torture and subsequent killings perpetrated by counterinsurgents, these two placebo tests should record statistically significant results. Looking at the results, both models demonstrate insignificant coefficients, suggesting that there was no significant change in state strategy immediately prior to torture. However, a close look at Model 5b reveals a substantively large, positive coefficient for the placebo test that approaches conventional standards of significance. The finding does not nullify the identified positive and significant association between torture and subsequent killings perpetrated by counterinsurgents, but it does complicate a strict causal interpretation of these results.

Another alternative is that there was some strategic change occurring in the months after torture that was unrelated to the use of torture but that produced the observed changes in killings perpetrated by counterinsurgents. The Granger causality tests presented in Model 5c examine the effect of treatment compared to a post-treatment placebo effect. The model estimates the association between both variables and killings perpetrated by counterinsurgents during the next six months (T+1 to T+6) for the treatment, and the next five months (T+2 to T+6) for the post-treatment placebo. In the results, only the treatment variable is significantly related to subsequent changes in killings perpetrated by counterinsurgents, which indicates that the changes occurred during the month when torture took place and not during the following month.

It must be noted that the results of these tests cannot rule out the possibility that some other change occurred in the treated months at the time of treatment and inspired both the use of torture and subsequent killings committed by counterinsurgents. This fact, along with the results of Model 5b, suggests some caution must be taken with granting a strictly causal interpretation to the analyses. Given these data, one can say descriptively that torture is associated with increased killings by counterinsurgents and that this relationship is consistent with the causal arguments above.

What the results do demonstrate is that the increase in killings committed by counterinsurgents occurred in municipalities experiencing torture during the month when torture took place, and not during the months before or after. In the Appendix to this article, further evidence is provided to support these claims, including tests examining the causal mechanisms as well as tests focusing on the use of torture in different strategic settings. All of the evidence is consistent with the expectation that torture is related to significant increases in killings by counterinsurgents. At the same time, there is no evidence that torture is related to any significant reduction in killings perpetrated by insurgents.

Conclusion

Using micro-level data on political violence from Guatemala and a research design that combined matching procedures with difference-in-difference estimation, this study estimated the relationship between torture and

Table IV.

Robustness tests of killings perpetrated by counterinsurgents

	Killings perpetrated by counterinsurgents
	5A		5B		5C
	Local		Local		Local
	Placebo test: Treatment lagged 2 months		Placebo test: Treatment lagged 1 month		Granger causality test
	Point estimate (Standard error)	95% CI	Point estimate (Standard error)	95% CI	Point estimate (Standard error)	95% CI
δ Pretreatment placebo effect T+1	29.84 (39.60)	–48.07, 107.76	60.84 (34.78)	–7.58, 129.28
δ Granger ATET T+1 to T+6					70.89** (24.84)	22.01, 119.77
δ Post-treatment placebo effect T+2 to T+6					–24.21 (17.65)	–58.95, 10.51
γ Treatment group	–38.27 (36.06)	–109.22, 32.67	–7.27 (34.59)	–75.67, 61.13	–8.72 (17.90)	–43.93, 26.49
λ T+1	–26.15 (29.30)	–83.80, 31.49	–63.02* (27.69)	–117.51, -8.54
λ T+1 to T+6					–51.09*** (10.05)	–70.87, -31.31
Intercept	125.98* (27.78)	71.31, 180.63	99.22 (26.77)	46.55, 151.89	91.61*** (12.19)	67.64, 115.61
N	660		660		3,960

Unit clustered robust standard errors in parentheses. *p < 0.05, **p < 0.01, ***p < 0.001.

subsequent killings by insurgents and counterinsurgents. The results suggest that torture is an ineffective tool for combatting insurgency. Torture was not associated with any significant decrease in killings by insurgents, but was associated with significant increases in killings by counterinsurgents.

In addition to shining light on this controversial policy, these results have important implications for research on insurgent and counterinsurgent violence more generally. Existing research on the topic has focused almost exclusively on deadly acts of violence, while disregarding other forms such as torture. Expanding the topic to include other counterinsurgent strategies may contribute to the refinement of important theories on the topic, such as the control-collaboration model (Kalyvas, 2012). It may also yield contrasting conclusions, for example regarding the presumed relationships between counterinsurgent violence and information.

Extensions could investigate the use of torture in other contexts, such as transnational torture or torture applied domestically by governments not engaged in counterinsurgency. Future work might also probe the relationship between torture and other methods of political repression (Davenport, 2007). Integrating the findings of policy-level studies, such as this one, with a broader literature on repression would provide clues as to how different repressive policies might interact, substitute for one another, or have varying effects depending on the context in which they are applied. States are victorious in an overwhelming majority of domestic conflicts. This leads to the conclusion that some policies of repression must be effective, but at present we know little about what form these policies take or what contexts make various types of repressive activity more or less effective.

Footnotes

Replication data

The Appendix, dataset, and do-files for the empirical analysis in this article can be found at http://www.prio.no/jpr/datasets and at .

Acknowledgements

The author wishes to thank Benjamin Appel, Jacob Aronson, Courtenay Conrad, Christian Davenport, Robert Franzese, Will Moore, David Nickerson, Darius Rejali, Rocio Titiunik, the anonymous reviewers, and the editor, Sabine Carey, for reading and commenting on earlier drafts.

Notes

References

Angrist

Joshua D

Pischke

Jörn-Steffen

(2009) Mostly Harmless Econometrics. Princeton, NJ: Princeton University Press.

Ashenfelter

Orley

(1978) Estimating the effect of training programs on earnings. Review of Economics and Statistics 60(1): 47–57.

Bagaric

Mirko

Clarke

Julie

(2007) Torture: When the Unthinkable Is Morally Permissible. Albany, NY: State University of New York Press.

Ball

Patrick

(1999) CIIDH/AAAS Database of Human Rights Violations in Guatemala ATV20.1 (http://hrdata.aaas.org/page/ciidh-data).

Ball

Patrick

Kobrak

Paul

Spirere

Herbert

(1999) State Violence in Guatemala, 1960–1996: A Quantitative Reflection. Washington, DC: AAAS.

Benjamin

Mark

(2009) Soufan: CIA torture actually hindered our intelligence gathering (http://www.salon.com/news/feature/2009/05/14/torture/index.html).

Campbell

Donald

Stanley

Julian

(1963) Experimental and Quasi-Experimental Designs for Research. Boston, MA: Houghton Mifflin.

Card

David

Sullivan

Daniel

(1988) Measuring the effect of subsidized training programs on movements in and out of employment. Econometrica 56(3): 497–530.

Chamarbagwala

Rubiana

Morán

Hilcías

(2011) The human capital consequences of civil war: Evidence from Guatemala. Journal of Development Economics 94(1): 41–61.

10.

CIA (1963) KUBARK Manual. Washington, DC: Booksurge.

11.

CIIDH (1996) Draining the sea: An analysis of terror in three rural communities in Guatemala (1980–1984). Unpublished manuscript.

12.

Cingranelli

David

Richards

David

(1999) Respect for human rights after the end of the Cold War. Journal of Peace Research 36(5): 511–534.

13.

Cingranelli

David

Richards

David

(2011) The Cingranelli-Richards (CIRI) Human Rights Database (http://humanrightsdata.blogspot.com/).

14.

Commission for Historical Clarification (1999) Guatemala, Memory of Silence. Guatemala City: CEH.

15.

Condra

Luke

Shapiro

Jacob

(2011) Who takes the blame? The strategic effects of collateral damage. American Journal of Political Science 56(1): 167–187.

16.

Conrad

Courtenay

Moore

Will H

(2010) What stops torture? American Journal of Political Science 54(2): 459–476.

17.

Conrad

Courtenay

Ritter

Emily

(2013) Treaties, tenure, and torture: The conflicting domestic effects of international law. Journal of Politics 75(2): 397–409.

18.

Davenport

Christian

(2007) State repression and political order. Annual Review of Political Science 10: 1–27.

19.

Davenport

Christian

Ball

Patrick

(2002) Views to a kill: Exploring the implications of source selection in the case of Guatemalan state terror, 1977–1995. Journal of Conflict Resolution 46(3): 427–450.

20.

Derschowitz

Alan

(2003) Why Terrorism Works: Understanding the Threat, Responding to the Challenge. New Haven, CT: Yale University Press.

21.

Diamond

Alexis

Sekhon

Jasjeet

(2013) Genetic matching for estimating causal effects: A general multivariate matching method for achieving balance in observational studies. Review of Economics and Statistics 95 (3): 932–945.

22.

Fariss

Christopher

Schnakenberg

Keith

(2013) Measuring mutual dependence between state repressive actions. Journal of Conflict Resolution doi: 10.1177/0022002713487314.

23.

Finn

Peter

Warrick

Joby

Tate

Julie

(2009) How a detainee became an asset. Washington Post, 29 August: A1, A6 (http://www.washingtonpost.com/wp-dyn/content/article/2009/08/28/AR2009082803874.html).

24.

Foucault

Michel

(1979) Discipline and Punish. New York: Vintage.

25.

Galula

David

(1964) Counterinsurgency Warfare: Theory and Practice. Westport, CT: Praeger Security International.

26.

Greenberg

Karen

(2006) The Torture Debate in America. Cambridge: Cambridge University Press.

27.

Hafner-Burton

Emilie

Shapiro

Jacob

(2010) Tortured relations: Human rights abuses and counter-terrorism cooperation. PS: Political Science and Politics 43(3): 415–419.

28.

Hess

David

Martin

Brian

(2006) Repression, backfire and the theory of transformative events. Mobilization 11(2): 249–267.

29.

Hill

Daniel

(2010) Estimating the effects of human rights treaties on state behavior. Journal of Politics 72(4): 1161–1174.

30.

Daniel

Imai

Kosuke

King

Gary

Stuart

Elizabeth

(2007) Matching as nonparametric preprocessing for reducing model dependence in parametric causal inference. Political Analysis 15(3): 199–236.

31.

Imai

Kosuke

King

Gary

Stuart

Elizabeth

(2008) Misunderstandings between experimentalists and observationalists about causal inference. Journal of the Royal Statistical Society: Series A (Statistics in Society) 171(2): 481–502.

32.

Imbens

Guido W

(2009) Better LATE than nothing: Some comments on Deaton (2009) and Heckman and Urzua (2009). NBER Working Paper no. 14896.

33.

Johnson

David

Ryan

John

(2012) The interrogation game: Using coercion and rewards to elicit information from groups. Unpublished manuscript (http://econ.ucalgary.ca/sites/econ.ucalgary.ca/files/unitis/publications/1-4658443/Interrogation_Game_Draft.pdf).

34.

Jonas

Susanne

(1996) Dangerous liaisons: The U.S. in Guatemala. Foreign Policy 103(Summer): 144–160.

35.

Kalyvas

Stathis

(2006) The Logic of Violence in Civil War. Cambridge: Cambridge University Press.

36.

Kalyvas

Stathis

(2012) Micro-level studies of violence in civil war: Refining and extending the control-collaboration model. Terrorism and Political Violence 24(4): 658–668.

37.

Kalyvas

Stathis

Kocher

Matthew

(2007) How free is ‘free riding’ in civil wars? Violence, insurgency, and the collective action problem. World Politics 59(2): 177–216.

38.

Kocher

Matthew

Pepinsky

Thomas

Kalyvas

Stathis

(2011) Bombing as an instrument of counterinsurgency in the Vietnam War. American Journal of Political Science 55(2): 201–218.

39.

Ladd

Jonathan

Lenz

Gabriel S

(2009) Exploiting a rare communication shift to document the persuasive power of the news media. American Journal of Political Science 53(2): 394–410.

40.

Lehner

Paul

(2006) Options for scientific research on education practices. In: Swenson

Russell

(ed.) Educing Information: Intelligence Science Board Phase 1 Report. Washington, DC: National Defense Intelligence College, 303–310.

41.

Lyall

Jason

(2009) Does indiscriminate violence incite insurgent attacks? Evidence from a natural experiment. Journal of Conflict Resolution 53(3): 331–362.

42.

Lyall

Jason

(2010) Are co-ethnics more effective counterinsurgents? Evidence from the second Chechen war. American Political Science Review 104(1): 1–20.

43.

Lyall

Jason

Blair

Graeme

Imai

Kosuke

(2013) Explaining support for combatants during wartime: A survey experiment in Afghanistan. American Political Science Review 107(4): 679–705.

44.

Mack

Andrew

(1975) Why big nations lose small wars: The politics of asymmetric conflict. World Politics 27(2): 175–200.

45.

McAdam

Doug

(1990) Freedom Summer. Oxford: Oxford University Press.

46.

McClintock

Michael

(1985) The American Connection Volume II: State Terror and Popular Resistance in Guatemala. New York: Zed.

47.

Otterman

Michael

(2007) American Torture: From the Cold War to Abu Ghraib and Beyond. New York: Pluto.

48.

Powell

Emilia

Staton

Jeffrey

(2009) Domestic judicial institutions and human rights treaty violation. International Studies Quarterly 53(1): 149–174.

49.

Regan

Patrick

(2009) Political repression as a public good. Unpublished manuscript.

50.

Rejali

Darius

(2007) Torture and Democracy. Princeton, NJ: Princeton University Press.

51.

Schirmer

Jennifer

(1998) The Guatemalan Military Project. Philadelphia, PA: University of Pennsylvania Press.

52.

Sekhon

Jasjeet

(2011) Multivariate and propensity score matching software with automated balance optimization: The matching package for R. Journal of Statistical Software 42(7): 1–52.

53.

Sekhon

Jasjeet

Mebane

Walter R

(1998) Genetic optimization using derivatives: Theory and application to non-linear models. Political Analysis 7(1): 189–213.

54.

Stoll

David

(1993) Between Two Armies in the Ixil Towns of Guatemala. New York: Columbia University Press.

55.

United Nations (1984) Convention against torture and other cruel, inhuman or degrading treatment or punishment. United Nations Treaty Collection 1465(85): 1.

56.

Vrij

Aldert

Mann

Samantha

Kristen

Susanne

Fisher

Ronald P

(2007) Cues to deception and ability to detect lies as a function of police interview styles. Law and Human Behavior 31(5): 499–518.

57.

Walsh

James

(2009) Does torture stop terror? Nope. International Rehabilitation Council for Torture Victims 3(3): 2–3.

58.

Walter

Eugene

(1969) Terror and Resistance. Oxford: Oxford University Press.

59.

Wantchekon

Leonard

Healey

Andrew

(1999) The game of torture. Journal of Conflict Resolution 43(5): 596–609.

60.

Ward

Michael

Gleditsch

Kristian Skrede

(2008) Spatial Regression Models. London: Sage.

61.

Wood

Elizabeth

(2003) Insurgent Collective Action and Civil War in El Salvador. Cambridge: Cambridge University Press.

62.

Zwerman

Gilda

Steinhoff

Patricia

(2005) When activists ask for trouble. In: Davenport

Christian

Johnston

Hank

Mueller

Carol

(eds) Repression and Mobilization. Minneapolis, MN: University of Minnesota Press, 85–107.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB