Unpublished Supplement

Transcription

1 Unpublished Supplement This two-part supplement accompanies the paper Delegating Death: Military Intervention and Government Killing. Part One is a formal presentation of the theory developed informally therein. Part Two contains additional information as noted throughout the manuscript. Part One: A Strategic Logic of Government Killing Consider the following extensive form game of complete information: 1 Actors There are two players, a government (or regime; r) and a perpetrator (or military soldier; m). 2 Sequence of Play The game begins when the government assesses its situation and chooses an ideal percentage of civilian deaths - a proportion that would best allow it to pursue and achieve its goals. Next, the government orders some percentage of civilians killed. In response to the government s order, the perpetrator kills some percentage of civilians. More specifically, the sequence of moves is: 1. The government targets an ideal body count, 0 T 100, and then a percentage of civilians to be killed, t: 0 t T t. The ideal death toll T is the subset of the state s population whose death would allow the government to achieve its policy goal, while t is the percentage of civilians that perpetrators are asked to kill. 1

2 2. The perpetrator victimizes (kills) some percentage of civilians, v: 0 v Assumptions To specify the players preferences over outcomes, I employ the assumptions outlined in the body of this paper and made explicit below: Assumption 1: Humans value their physical person and their own life above the lives of others. Assumption 2: One-sided killing is a rational policy calculated to attain important political and/or military objectives. Assumption 3: The perpetrator is a unitary actor or a group with perfect discipline. Assumption 4: Governments desire a monopoly on the forces of coercion within the state. Thus, to the extent possible, they punish disobedience with p=1. 4 Utilities 4.1 The Government s Payoffs I assume that the government cares about the outcome that results from its interaction with perpetrators. Specifically, the government is interested in how closely the observed death toll (v) matches its ideal death toll (T ). I model this aspect of its preferences using a quadratic loss function, (T-v) 2. The term captures the benefits the government accrues from civilian deaths, assessing greater benefits as the government moves closer to its policy goal (as v moves toward T ). 2 2 Quadratic preferences capture the risk-averse nature of the government s payoff: the government has an ideal death toll in mind, and its utility declines as the true death toll moves farther from that ideal point in either direction (e.g., Bendor, Glazer and Hammond 2001). 2

3 In addition to its stake in the policy outcome of the interaction described above (e.g., the civilian death toll), the government is interested in the consequences of that interaction for its own future. Those consequences may influenced, first, by the potential international cost of killing. That concern is captured formally with bv 2, where b [0, 1] is the expected probability of international punishment for killing. Because of the international community s willingness to punish civilian killing, the government is concerned about whether or not it will be caught killing. As the bodies pile up (as v 2 increases), the international call for retribution gets louder and the punishment more severe. Therefore, increasing body counts become more costly as the government increasingly expects to be caught (as b increases from 0 to 1). International actors may increase the government s expectations about punishment for killing, but they may also attenuate those expectations. If an international ally supports, justifies, or even conceals killing from the result of the community, the expected probability of punishment for killing decreases. Thus increasing body counts become less costly as the government increasing expects to be protected (as b decreases from 1 to 0). The government s payoff also responds to the perpetrator receiving and subsequently disobeying the order to kill. I define disobedience as the squared difference between the percent ordered killed and the percent actually killed, (t v) 2. Disobedience may influence the government s utility for ordering t civilians killed in two ways. First, there are intrinsic costs to disobedience, captured formally as d(t v) 2. As discussed in the main text, disobedience undermines the government s monopoly on force and threatens its legitimacy. This dynamic is captured by imposing greater loss on the government as the final outcome diverges from its order. To capture the idea that intrinsic values differ across governments, the loss is weighted by d ɛ [0, 1], where d is the government s concern for its own legitimacy. When a government is unconcerned with the effects of disobedience on its own legitimacy (d = 0), this component reduces to zero. On the other 3

4 hand, when a government is highly concerned with its legitimacy (d = 1), the component assesses the maximum possible penalty. Second, whenever possible, a government will punish disobedience with p=1. The costs of punishing disobedience after observing v are expressed as ac(t v) 2, where a [0, 1] is the expected proportion of disobedience observed by the government and c 0 is the cost (to the government) of arresting, transporting, and punishing disobedient perpetrators. When a = 0 (when disobedience is not observed and therefore goes unpunished), the government does not pay the associated costs. When a = 1 (the government can observe disobedience and therefore punishment occurs), the costs have two components: c 0 is the cost of arresting, transporting and punishing disobedient perpetrators. When c = 0, punishing is costless. And, (t v) 2 is the severity of the deviation that must be punished. The more deviation occurs, the more punishment is metered out; the costs of punishment respond to the increase. Each of these considerations is captured in the following utility function, where u r (t) is the government s utility for ordering t civilians killed: u r (t) = (T v) 2 bv 2 d(t v) 2 ac(t v) The Perpetrator s Payoffs Like the government, the perpetrator s payoff responds to both international and domestic influences. Internationally, the perpetrator is concerned with the potential costs of punishment (for killing) from an international actor. These costs are captured by the term bv 2, and therefore depend on two factors. As above, b ɛ [0, 1] captures the perpetrator s beliefs about likelihood of punishment by a third party for killing civilians. When b = 0, the perpetrator doesn t worry about international costs. As b increases from 0 to 1 as she expects the likelihood of punishment for killing to increase killing becomes more costly for the perpetra- 4

5 tor. Second, and also as above, the costs of international punishment depend on the severity of killing. These enter the utility function through the parameter v 2. As the body count increases, the international community is exponentially more likely to be involved and more likely to punish perpetrators. The second term in the utility function therefore responds to the following: How much bloodshed did I cause? and, Do I think I m going to get caught? Together, these parameters determine the cost of killing. Domestically, utility responds to the questions: Did I deviate from the killing order?, If so, do I think I m going to get caught?, and If I m caught, how much trouble am I in? These potential costs are captured with the term af(t v) 2. As above, a ɛ [0, 1] captures how easy it is for the government to identify disobedience. When a = 0, the perpetrator will not be caught or punished by the government for disobeying. As a increases to 1, she is increasingly likely to be caught and deviating becomes increasingly costly. Also as above, (t v) 2 captures the severity of the perpetrator s deviation. When this term equals 0 when there was no deviation there are no deviation costs. As the deviation becomes more severe, it also becomes more costly. Lastly, the parameter f 0 captures the expected punishment for disobedience. As f increases, the expected punishment is increasingly severe and the costs of deviating increase. Each of these considerations is captured in the following utility function, where u m (v) is the perpetrator s utility for killing v civilians: u m (v) = bv 2 af(t v) 2 5

6 5 Equilibrium 5.1 Perpetrator s Best Response Consider the perpetrator s best response to a given governmental order. That is, consider the perpetrator s utility for killing exactly v civilians: u m (v) = bv 2 af(t v) 2 = bv 2 aft 2 + 2aftv afv 2 Taking the partial derivative with respect to v, then maximizing, yields the perpetrator s optimal choice: u m (v) v = 0 = bv + aft afv v = aft b + af This choice has a number of intuitive properties. When disobedience is observed and punished with positive probabilities, the number of civilians perpetrators kill is in part a function of t, the number of civilians they were asked to kill. Specifically, when t = 0 (the perpetrators are not asked to kill), their best response is to kill exactly 0 civilians. As t increases from 0 (the perpetrators are asked to kill increasing numbers of civilians), the perpetrator kills more and the observed death toll increases as a result. In addition to capturing the killing order s implementation, the perpetrator s best response captures the effects of exogenous influences on civilian death tolls. If the government cannot observe or credibly threaten to punish disobedience, perpetrators cannot be punished for disobeying the killing order. In these cases, the perpetrator s best response is to similarly reduce the possibility of international punishment to 0 by avoiding killing altogether. Thus 6

7 when a = 0 and/or f = 0 (the expected punishment is absent or nonviolent), the model expects no civilian deaths. As a and f increase arbitrarily from 0 (the expected punishment becomes increasingly likely and/or increasingly violent), disobedience is less appealing and the model predicts mounting death tolls. In addition, the death toll responds to changes in the expected likelihood of international retribution (b). When perpetrators estimate this likelihood at 0, they will kill exactly g civilians. That is, when the international community is expected to do nothing but the possibility of punishment for disobedience exists, perpetrators will kill as many civilians as they are asked to kill. As the expected likelihood of international punishment for killing increases, the expected death toll decreases in response. And as the expected likelihood of international punishment for killing decreases towards 0, the expected death toll increases in response. Finally, when a 0, b 0, and f > 0, the perpetrator s optimal response takes its least specific form: aft b+af. 5.2 Government s Optimal Choice Next, consider the decision facing the government. To make the derivation and statement of the equilibrium easier, let z = af. Now, the government s utility for ordering t civilians b+af 7

8 killed is: u r (t) = (T v) 2 bv 2 d(t v) 2 ac(t v) 2 = (T zt) 2 bz 2 t 2 d(t zt) 2 ac(t zt) 2 = (T zt) 2 bz 2 t 2 (d + ac)(t zt) 2 = (T zt) 2 bz 2 t 2 (d + ac)(1 z) 2 t 2 = T 2 + 2T zt z 2 t 2 bz 2 t 2 (d + ac)(1 z) 2 t 2 Taking the partial derivative with respect to t, then maximizing, yields: u r (t) t = 2T z 2tz 2 2tbz 2 2t(d + ac)(1 z) 2 = 2T z 2t[z 2 + bz 2 + (d + ac)(1 z) 2 ] t[z 2 + bz 2 + (d + ac)(1 z) 2 ] = T z 0 = 2T z 2t[z 2 + bz 2 + (d + ac)(1 z) 2 ] t = T z z 2 (1 + b) + (d + ac)(1 z) Subgame Perfect Equilibrium The decision facing governments, together with the perpetrator s optimal response, define the game s unique pure-strategy, subgame-perfect equilibrium: {v = aft b + af, t = } T z. z 2 (1 + b) + (d + ac)(1 z) 2 8

9 6 Equilibrium Implications Above, I identified three possibilities with respect to the perpetrator s best response. The unique SGPE predicts that one of these three outcomes will obtain in every circumstance. Which outcome obtains depends on the values of the various parameters in equilibrium. Below, I walk through each situation in turn. 6.1 Self-Selection In two cases (a = 0 and f = 0), no civilians will die regardless of the government s choice. This provides an opportunity to ask, what happens when v = 0 regardless of the value of t? That is, what is the strategic logic of government killing when no civilians will die regardless of the government s choice? The answer reveals a previously-untheorized self-selection effect: A forward looking government will prefer not sending the killing order to sending it and having it disobeyed. Such disobedience is expected in 2 cases: When disobedience cannot be observed (a = 0), and when the expected punishment for disobedience fails to threaten perpetrators lives or liberties (f = 0). In both cases, anticipating disobedience, the government will not ask perpetrators to kill. Formally, when v = 0, g = 0 as well. Here is one answer to the puzzle, why do death tolls differ despite similiar conditions and incentives to kill? When the government knows its order will be ignored, it does not send the killing order regardless of the conditions and incentives that make killing an appealing policy choice. Proposition 1: When disobedient perpetrators cannot be observed, no civilians will die. Proposition 2: When the expected punishment for disobedience is nonviolent, no civilians will die. 9

10 6.2 Perfect Implementation When there is no expected international retribution for killing (when b = 0), the perpetrator s ideal point will be the same as the government s, and therefore the perpetrator will kill as many civilians as the government asks her to kill. What is the strategic logic of government killing when the perpetrator will kill as many civilians as the government likes? When the government knows its order will be implemented perfectly, it asks for its ideal body count. Formally, when v = t, t = T. Proposition 3: When international punishment is not expected, the observed death toll equals the government s ideal death toll. 6.3 Comparative Statics & Conditionalities Comparative Static Analysis When an international actor observes disobedience with a positive probability (a > 0) and the expected punishment for disobedience threatens human liberty (f > 0), the predicted outcome is the least specific statement of the equilibrium: {v = aft b+af, t = T z z 2 (1+b)+(d+ac)(1 z) 2 }. What is the strategic logic of government killing when the death toll is a function of the government s order as well as a set of exogenous parameters? The SGPE equilibrium permits a ceteris paribus examination of the effects of each influence on the outcome, civilian body counts. To determine the instantaneous effects of changing the various model parameters on the equilibrium outcome, I take the partial first derivatives of each equation with respect to each parameter of interest. To extract relationship predictions from this analysis, I next take the partial second derivative of each equation with respect to each parameter of interest. 3 3 Differentiation of v was conducted by hand and confirmed using the computer algebra system Maxima. Differentiation of t was conducted using Maxima, which is available online at maxima.sourceforge.net. I am grateful to Justin Esarey, Kris Ramsay, and David Siegel for introducing me to the system. 10

11 The sign on the second derivative represents the directional effect of increasing the parameter of interest on the dependent variable under investigation. So for analyses with respect to t, the sign on each second derivative is the effect of increasing the parameter of interest on the government s ordered death toll. And for analyses with respect to v, the sign on each second derivative is the effect of increasing the parameter of interest with respect to the civilian death toll. I begin with a comparative static analysis of the number of civilians the government orders perpetrators to kill: t = T z z 2 (1 + b) + (d + ac)(1 z) 2 To determine the effect on the ordered death toll of increasing the cost (to the government) of punishing disobedience, I calculate t c and 2 t 2 c : t c 2 t 2 c = = ( 1 af a 2 f ( ( (af + b) (d + ac) 1 af af+b ( 1 af af+b) 2 T 2a 3 f ( ( (af + b) (d + ac) 1 af af+b af+b) 4 T ) ) a 2 (b+1)f 2 (af+b) 2 ) ) a 2 (b+1)f 2 (af+b) 2 2 t 2 c is strictly negative. Therefore, as punishing disobedience becomes more costly, the government orders a decreasing number of civilians killed. To determine the effect on the ordered death toll of increasing the government s con- 11

12 cern for its own legitimacy, I calculate t d and 2 t 2 d : t d 2 t 2 d = = ( 1 af af ( ( (af + b) (d + ac) 1 af af+b ( 1 af af+b) 2 T 2af ( ( (af + b) (d + ac) 1 af af+b af+b) 4 T ) ) a 2 (b+1)f 2 (af+b) 2 ) ) a 2 (b+1)f 2 (af+b) 2 2 t 2 d is strictly negative. Therefore, as the government becomes increasingly concerned with its own legitimacy, it orders a decreasing number of civilians killed. Next, I conduct a comparative static analysis of the number of civilians the perpetrator actually kills: v = aft b + af To determine the effect on the observed death toll of increasing the expected proportion of disobedience the government can observe, I calculate v a and 2 v 2 a : v a 2 v 2 a = = ft af + b af 2 t (af + b) 2 2af 3 t (af + b) 2f 2 t 3 (af + b) 2 2 v 2 a is strictly positive. Therefore, as the government observes an increasing proportion of disobedient perpetrators, the civilian death toll increases. Next, to determine the effect on the observed death toll of changing the expected 12

13 likelihood of international punishment for killing, I calculate v b and 2 v 2 b : v b 2 v 2 b = aft (af + b) 2 = 2aft (af + b) 3 2 v 2 b is strictly negative. Therefore, as the expected likelihood of international punishment for killing increases, the civilian death toll decreases. And as the expected likelihood of international punishment for killing decreases, the civilian death toll increases. Finally, to determine the effect on the observed death toll of increasing how violent the punishment for disobedience is expected to be, I calculate v f and 2 v 2 f : v f 2 v 2 f = = ag af + b a2 ft (af + b) 2 2a 3 ft (af + b) 2a2 t 3 (af + b) 2 2 v 2 f is strictly positive. Therefore, as beliefs about the punishment for disobedience become more violent, the civilian death toll increases Conditional Effects Finally, examining the actors equilibrium choices reveals a series of conditionalities. These conditionalities yield insight into how changes in one parameter affect the relationship between another parameter and each actor s choice in equilibrium. The interplay of the domestic environment and international system reveal hypotheses that are not only untheorized (until now), but have real policy implications for international actors interested in limiting noncombatant deaths. I focus on one such implication. 13

14 In the figure below, each graph is a plot of the model s equilibrium predictions given theoretically-relevant values of the parameters. In other words, each line is produced by plugging numeric values into the equilibrium outcome, and arithmetically calculating the model s expectations with respect to the government s ordered death toll (t) and the number of civilians who perpetrators kill (v). For example, the top graph presents predictions from the formal model about how many deaths are ordered, and how many civilians executed, when the expected probability of international punishment for killing is 10% (so b =.1) and the government s ability to observe disobedience (a) varies from 0 to 1. All other parameters are held to the middle values of their theoretical ranges. The model suggests a positive relationship between the government s ability to observe disobedience (monitoring capacity) and civilian death tolls. It also suggests that this relationship may be conditioned by the expected probability of international punishment for killing. That conditional relationship is presented in Figure S.1. The top panel captures the relationship between domestic monitoring capacity and body counts when the expected likelihood of international punishment is low (10%), while the bottom panel captures the same relationship when the expected likelihood of international punishment is high (90%). As Figure S.1 shows, the strength of the effect of domestic monitoring capacity is conditional on actors expectations about the likelihood of international punishment for killing: Perpetrators must weigh international consequences for killing against domestic consequences for ignoring the order to kill. Proposition 4: As the expected likelihood of international punishment for killing increases, the positive relationship between the government s ability to monitor disobedience and body counts weakens (becomes less positive). 14

15 Figure S.1: Moderating Effect of International Punishment on the Relationship Between Domestic Monitoring and Civilian Death Tolls 15

16 Part Two: Supplementary Information Table S.1: List of Countries in Estimation Sample Albania ( ) Egypt ( ) Kyrgyzstan ( ) Singapore ( ) Algeria ( ) Fiji ( ) Latvia ( ) Slovenia ( ) Australia (1999) Finland ( ) Lesotho ( ) South Korea ( ) Austria ( ) France ( ) Malaysia ( ) Spain ( ) Azerbaijan ( ) Georgia ( ) Mauritius ( ) Sri Lanka ( ) Bahrain ( ) Germany ( ) Mexico ( ) Sudan ( ) Belarus ( ) Greece ( ) Moldova ( ) Sweden ( ) Belgium ( ) Guatemala ( ) Namibia ( ) Switzerland ( ) Bhutan ( ) Guinea ( ) Nepal ( ) Syria ( ) Botswana ( ) Hungary ( ) Netherlands ( ) Tajikistan ( ) Brazil ( ) India ( ) Nicaragua ( ) Trinidad (1995) Bulgaria ( ) Indonesia ( ) Oman ( ) Tunisia ( ) Burundi ( ) Iran ( ) Pakistan ( ) Uganda ( ) Cameroon (1995, ) Ireland ( ) Panama ( ) Ukraine (1999) Canada ( ) Italy ( ) Papua New Guinea ( ) United Kingdom ( ) China ( ) Ivory Coast ( ) Peru ( ) Uruguay ( ) Costa Rica ( ) Jamaica ( ) Philippines ( ) Venezuela ( ) Croatia ( ) Jordan ( ) Portugal ( ) Zambia ( ) Cyprus ( ) Kazakhstan ( ) Russia (1999) Zimbabwe ( ) Czech Republic ( ) Kenya ( ) Senegal ( ) Denmark ( ) Kuwait ( ) Sierra Leone ( ) 16

17 Defining Government Killing I define one-sided government killing as the intentional, extralegal killing of civilians by governments. This is similar to the definition of mass killing proposed by Valentino (2004,10-12), although his requires a minimum of 50,000 intentional deaths over no more than five years. It differs from most other definitions (e.g., Harff 2003, Midlarsky 2005) in that it does not specify that the perpetrators must intend to destroy a particular group, and in that because I do not limit the focus to killing during civil or international war I focus on killing civilians rather than noncombatants. Dependent Variable: Benefits and Potential Issues My dependent variables, presence of death tolls and number of civilians killed, come from the Political Instability Task Force s (PITF) Worldwide Atrocities Dataset (Ulfelder and Schrodt 2009). The full dataset currently records civilian casualties between January 1994 and December 2009 in all political and military conflicts, and contain 6,000 killing events across a sample of 113 countries. These events represent the population of cases reported in major international press agencies (Agence France Presse, Associated Press, New York Times, and Reuters) in which at least five civilians were killed. Below I discuss some issues associated with the data. To my knowledge, most existing datasets on government killing either cast the macro event of a larger conflict process as the unit of observation (e.g., Fein 1993, Harff 2003, Valentino, Huth and Croco 2006, Downes 2007) or focus on micro events within the context of a single conflict process (e.g., Ball, Kobrak and Spirer 1999). These decisions limit researchers to analyzing either the onset of conflicts involving atrocities or the dynamics of violence in the context of genocide. As the Atrocities Dataset Codebook notes, one consequence is the inability to draw inferences about the likelihood of any deliberate violence 17

18 against noncombatants in situations where none has yet occurred, or the dynamics of such violence in situations that may or may not rise to the level of genocide (Ulfelder and Schrodt 2009,2). The PITF data record information from press accounts of killings occurring in all countries of the world during their period of observation, allowing researchers to study the risk of killing that may or may not accumulate to the level of mass killing. 4 A final benefit of the PITF data is that it is updated approximately every six months. As a result, replications, extensions, and challenges of this work should be relatively straightforward. Sample Selection: Details and Results Recall that the conditions that make government killing an appealing policy choice are distinct from actors expectations about punishment. To this point, scholars of repression have implicitly assumed that killing is a decision governments make: It occurs when the government decides it will help them achieve a desired goal. If one omitted these influences on killing from the empirical analysis one would produce a nonrandom sample: Since unobserved forces systematically affect whether and how much killing occurs, the country-years that produce killing are not drawn randomly from the population of country-years in the study. As Lemke and Reed (2001,126) note, whenever cases are drawn by any rule other than 4 In an analysis of alternative sources of information about state repression in Guatemala, Ball and Davenport (2002) find that different sources of information focus on different aspects of the behavior under investigation. They find that newspapers are useful in documenting obvious behavior or that which is deemed politically salient within a specified political-geographic context...[but] may be relatively weaker at identifying events in more remote areas that occur during periods of relatively less state repressiveness and that are relatively smaller in scale. (Ball and Davenport 2002,445) One consequence of using journalistic sources, then, is that low-level government killings may be underreported in my dependent variable. Resultantly, the implications drawn from the quantitative analyses below may be less applicable to low-level killing events in rural areas when government repression is low than to other government killing events. 18

19 randomization, selection bias is a potential problem. Thus, were I to test my hypotheses using only cases where government killing campaigns had begun, my estimates would likely be biased. To limit this inferential threat and conduct unbiased hypothesis tests, I model the process by which governments decide to kill civilians with sample-selection estimation. Existing work provides ample guidance on this issue. As noted in-text, that work suggests that leaders are influenced to repress (including kill) by a set of conditions (e.g., democracy, violent dissent, ethnic conflict; see, for example, Lichbach 1987, Moore 2000, Davenport and Armstrong 2004, Bueno De Mesquita et al. 2005, Davenport 2007) and a set of incentives (wars of attrition or territorial annexation, guerrilla tactics, civilian support for guerrillas; see, for example, Valentino et al. 2004, Kalyvas 2006, Downes 2008). To incorporate these forces and produce unbiased second-stage estimates, I conduct a maximum likelihood estimation of these influences on the presence/absence of civilian death tolls. The estimation produces a parameter (the inverse Mills ratio) which I use in the subsequent estimation to proxy the desirability of killing civilians. 5,6 I use Polity s [0,10] index of democratic institutions to measure democracy (Marshall and Jaggers N.d.) and the PRIO/Uppsala Conflict Data Project s measures of the presence of civil and international war (Gleditsch et al. 2002). Measures of attrition and territorial annexation come from Downes (2007), and all remaining variables come from Valentino et al. (2004). Table S.2 presents the results of the sample-selection estimation. As the literature would expect, both international war and civil war, as well as elements of civil war (particularly civilian support for rebels), increase the likelihood that government killing occurs (e.g., Valentino et al. 2004, Kalyvas 2006). Other elements of civil war (guerrilla tactics and 5 The IMR is calculated as λ(x) = φ(x) 1 Φ(x), where φ(x) is the standard normal pdf, and Φ(x) is the standard normal cdf. 6 This is the same procedure detailed by Heckman (1979). For a longer discussion of its use as the first of three equations, see Moore and Shellman (2007, ). 19

20 Table S.2: Probit (Selection) Estimation DV=Gov t Killing it ɛ [0, 1] β r.s.e.(β) domestic influences Democracy it Government Killing it Civil War it Guerrilla Tactics it Guerrilla Threat it Civilian Support for Guerrillas it Identity Conflict it international influences International War it War of Attrition it Constant *** p 0.01; ** p 0.05; * p 0.1, (two-tailed). N=960. LPL= , Wald χ 2 (9df)=144.28, Pseudo-R 2 =.2838 Robust standard errors are clustered by country. Wars of annexation were specified, but dropped due to collinearity. identity conflict), as well as elements of international war (wars of attrition) are positively signed as expected but statistically insignificant. The effect of the level of threat posed by guerrillas is also statistically insignificant. Finally, the estimated effect of democracy is negatively-signed and statistically significant. These results capture the conditions and incentives that make killing civilians an attractive government policy, independent of the delegation process modeled in the article s main ZINB estimation. By including them in subsequent analysis, I can test hypotheses relating to the process of government killing while controlling for the decision to kill civilians. Sensitivity Analysis: Measuring Democracy Above, I operationalize democracy using Polity s [0,10] measure of democratic institutionalization (Marshall and Jaggers N.d.). This is an appropriate operationalization because as a state becomes more democratic, its institutions make repression decreasingly appealing to leaders. Because there is some debate in the literature about the appropriate way to con- 20

21 ceptualize and measure democracy (Keith 2002, Davenport and Armstrong 2004, Bueno de Mesquita, Koch and Siverson 2004, Bueno De Mesquita et al. 2005), it is helpful to confirm that my results are also robust to other measurement strategies. I begin by respecifying Model S.2 using the following measures of democracy and democratization: (1) the Polity [-10,10] measure of regime type; (2) a binary specification that codes states democratic if Polity [-10,10] 6; (3) a binary specification that codes states democratic if Polity [-10,10] 7; (4) A binary decomposition of democracy; (5) a linear (collapsed) version of democracy based on the binary decomposition; (6) the Polity [-10,10] measure of regime type along with its squared term; and (7) a measure of winning coalition divided by selectorate size as discussed in Bueno de Mesquita et al. (2003). Then, I show and discuss the correlations among IMRs for each specification. Table S.3: Probit Estimation where Democracy is Polity [-10,10] DV=Gov t Killing it ɛ [0, 1] β r.s.e.(β) domestic influences Democracy it Government Killing it Civil War it Guerrilla Tactics it Guerrilla Threat it Civilian Support for Guerrillas it Identity Conflict it international influences International War it War of Attrition it Constant *** p 0.01; ** p 0.05; * p 0.1, (two-tailed). N=948. LPL= , Wald χ 2 (9df)=152.28, Pseudo-R 2 =

22 Table S.4: Probit Estimation where Democracy is Polity 6 DV=Gov t Killing it ɛ [0, 1] β r.s.e.(β) domestic influences Democracy it Government Killing it Civil War it Guerrilla Tactics it Guerrilla Threat it Civilian Support for Guerrillas it Identity Conflict it international influences International War it War of Attrition it Constant *** p 0.01; ** p 0.05; * p 0.1, (two-tailed). N=948. LPL= , Wald χ 2 (9df)=158.77, Pseudo-R 2 =.2766 Table S.5: Probit Estimation where Democracy is Polity 7 DV=Gov t Killing it ɛ [0, 1] β r.s.e.(β) domestic influences Democracy it Government Killing it Civil War it Guerrilla Tactics it Guerrilla Threat it Civilian Support for Guerrillas it Identity Conflict it international influences International War it War of Attrition it Constant *** p 0.01; ** p 0.05; * p 0.1, (two-tailed). N=948. LPL= , Wald χ 2 (9df)=161.94, Pseudo-R 2 =

23 Table S.6: Probit Estimation with Binary Decomposition of Democracy [0,10] DV=Gov t Killing it ɛ [0, 1] β r.s.e.(β) domestic influences Democracy=1 it Democracy=2 it Democracy=3 it Democracy=4 it Democracy=5 it Democracy=6 it Democracy=7 it Democracy=8 it Democracy=9 it Democracy=10 it Government Killing it Civil War it Guerrilla Tactics it Guerrilla Threat it Civilian Support for Guerrillas it Identity Conflict it international influences International War it War of Attrition it Constant *** p 0.01; ** p 0.05; * p 0.1, (two-tailed). N=913. LPL= , Wald χ 2 (18df)=183.56, Pseudo-R 2 =.3238 Reference category is Democracy=0. Table S.7: Relationship between Democracy [0,10] and Collapsed (Linear) Scale Polity (Democracy) Linear Scale

24 Table S.8: Probit Estimation with Collapsed (Linear) Scale of Democracy DV=Gov t Killing it ɛ [0, 1] β r.s.e.(β) domestic influences Democracy it Government Killing it Civil War it Guerrilla Tactics it Guerrilla Threat it Civilian Support for Guerrillas it Identity Conflict it international influences International War it War of Attrition it Constant *** p 0.01; ** p 0.05; * p 0.1, (two-tailed). N=913. LPL= , Wald χ 2 (9df)=151.23, Pseudo-R 2 =.2920 Table S.9: Probit Estimation where Democracy is Polity [-10,10] and Polity 2 DV=Gov t Killing it ɛ [0, 1] β r.s.e.(β) domestic influences Democracy it Democracy 2 it Government Killing it Civil War it Guerrilla Tactics it Guerrilla Threat it Civilian Support for Guerrillas it Identity Conflict it international influences International War it War of Attrition it Constant *** p 0.01; ** p 0.05; * p 0.1, (two-tailed). N=948. LPL= , Wald χ 2 (10df)=173.59, Pseudo-R 2 =

25 Table S.10: Probit Estimation where Democracy is Winning Coalition/Selectorate Size DV=Gov t Killing it ɛ [0, 1] β r.s.e.(β) domestic influences Democracy it Government Killing it Civil War it Guerrilla Tactics it Guerrilla Threat it Civilian Support for Guerrillas it Identity Conflict it international influences International War it War of Attrition it Constant *** p 0.01; ** p 0.05; * p 0.1, (two-tailed). N=732. LPL= , Wald χ 2 (9df)=106.21, Pseudo-R 2 =.2582 The results in Table S.2 model the likelihood that killing is a desired policy for state leaders. This selection equation enters the outcome equations in the article via the Inverse Mills Ratio as described above and by Heckman (1979). As shown in Tables S.3 through S.10, the selection equation is robust to a series of specifications of democracy. To ensure that the outcome equations are similarly robust to the respecifications, Table S.11 presents Pearson s ρ correlations among IMRs across all selection equations. The extremely high correlations suggest that the ZINB results reported and discussed in the main article are similarly robust to these changes. 25

26 Table S.11: Correlations (Pearson s ρ) Among IMRs Across Measures of Democracy Table S.2 Table S.3 Table S.4 Table S.5 Table S.6 Table S.8 Table S.9 Table S.10 Table S Table S Table S Table S Table S Table S Table S Table S

27 Table S.12: In-Sample Interventions Target State Intervener Years Direction Type Guatemala Belize 1995 Hostile External Peru Ecuador 1995 Hostile Ground, Air Peru Ecuador 1997 Hostile Ground Peru Ecuador 1998 Hostile Ground Croatia Yugoslavia Hostile Air Croatia Yugoslavia 1999 Hostile Ground Moldova Russia Hostile unknown Azerbaijan Armenia Hostile External Uganda Sudan 1998 Hostile Air Sudan Eritrea 1998 Hostile External South Korea North Korea 1999 Hostile Naval India Pakistan 1999 Hostile Ground Pakistan India 1998 Hostile External Philippines China Hostile Naval Philippines Malaysia 1999 Hostile Naval Albania Italian-led multinational force 1997 Supportive Ground Croatia U.N Supportive Ground, Air Sierra Leone Guinea Supportive Ground Sierra Leone Nigeria Supportive Ground Cameroon Nigeria Supportive Ground, Naval Namibia Angola 1999 Supportive Air Kuwait United States 1996 Supportive Ground, Naval, Air Kyrgyzstan Kazakhstan 1999 Supportive Air Kyrgyzstan Tajikistan 1999 Supportive Air China United States 1998 Supportive Air Papua New Guinea New Zealand-led truce Supportive Ground, Air monitoring group Papua New Guinea Australian-led peace Supportive Ground monitoring group Papua New Guinea Australia 1998 Supportive Ground, Air Papua New Guinea New Zealand 1998 Supportive Ground, Air Papua New Guinea United States 1998 Supportive Ground, Air Source: Pickering and Kisangani External interventions involve firing by the intervener from outside the target. 27

28 Table S.13: In-Sample Collinearity Among Independent Variables tax cap intrvn(against) cap*int against intrvn (support) recent rep aid(wtd) openness (wtd) ln(pop) war IMR tax capacity intvn (against) cap*int against intvn (support) recent repression aid (wtd) openness (wtd) ln(pop) war IMR

29 Additive Substantive Impacts on Killing Figure S.2 presents the predicted effects of increasing state capacity. The top panel of shows capacity s effects on the likelihood that the government can order killing given that it is a desirable policy option. 7 The x-axis encompasses the full in-sample range of capacity, from.06 (Moldova, 1998) to.80 (Oman, 1997), and the results reveal that increasing domestic monitoring capacity has a monotonic and significant positive effect on the probability that government killing can occur. The bottom panel of Figure S.2 predicts the effects of increasing capacity on the expected civilian death toll. These results, too, show that increasing capacity has a monotonic and significant positive effect. Next, Figure S.3 presents the predicted impacts of hostile intervention. The top panel shows intervention s effects on the likelihood that killing can occur when it is appealing to the government. These results echo the findings from Table 1: Intervening against the government increases the likelihood that government killing can occur (thereby undermining hypothesis one), but depresses expected death tolls when they emerge (thereby supporting hypothesis two). The predicted effects of supportive intervention are shown in Figure S.4. Again, the top panel shows intervention s effects on the likelihood that government killing can occur when desired. And again as above, these findings echo statistical results: Supportive intervention decreases the probability that killing can occur (thereby supporting hypothesis one), but increases expected death tolls when they emerge (thereby undermining hypothesis two). 7 To ease interpretation, results are inverted from the first-stage predictions, which estimate substantive effects on the likelihood that killing cannot occur regardless of the government s desire to kill civilians. 29

30 Figure S.2: Additive Impact of Domestic Monitoring Capacity on Government Killing Source: Table 1, columns one (top panel) and two (bottom panel). 30

31 Figure S.3: Additive Impact of Hostile Intervention on Government Killing Source: Table 1, columns one (top panel) and two (bottom panel). 31

32 Figure S.4: Additive Impact of Supportive Intervention on Government Killing Source: Table 1, columns one (top panel) and two (bottom panel). 32

33 Sensitivity Analysis: Domestic Monitoring Capacity In the paper, I use income tax capacity as a proxy for the state s ability to monitor potential perpetrators. As is often true in studies seeking to quantitatively test unobservable forces, this is not a perfect measure. For example, in 1994 the Hutu-led Rwandan government was not especially adept at collecting income tax, but was quite capable of organizing genocide and monitoring its murderous agents. If my tax-based proxy is systematically limited, I expect it will be in states like Rwanda states with weak economic monitoring capacity but stronger military monitoring capabilities. To see whether my results hold when this limitation is addressed, I conduct a sensitivity analysis using a second proxy for the state s monitoring ability. I assume that monitoring ability increases with military expenditures. Specifically, increased military spending increases the presence and density of other potential perpetrators and improves their ability to report disobedience. Since perpetrators know how many other perpetrators are around, and how easy it is for the others to report his disobedience, increasing military spending corresponds to an increasing expectation that disobedience would be observed and punished (by the state). I measure these forces via military expenditures in thousands of current-year US dollars. The data come from the Correlates of War project s National Material Capabilities Data (Singer, Bremer and Stuckey 1972, Singer 1987). Results show little statistical change from the specification in the main table, and the changes that do occur relate to significance (as opposed to sign). Hostile intervention remains negatively signed in both stages of both models, and statistically significant in stage two. Supportive intervention remains positively signed in both stages of the additive model and stage two of the interactive version, and statistically significant in both stage two estimations. The product of monitoring capacity and hostile intervention remains statistically significant in both stages of the estimation, and the joint significance test suggests significant combined effects of the constituent and product terms in stage two. While my main measure of 33

34 Table S.14: ZINB Analysis of Influences on Government Killing (Military Spending Proxies Monitoring Capacity) Model I (Additive) Model II (Interactive) Pr(all zeroes) E(death toll) Pr(all zeroes) E(death toll) international influences Intervention Against (.791) (.737) (.651) (.473) Intervention Supporting (.830) (1.363) (.570) (.738) Foreign Aid (weighted) (.001) (.0007) (.00004) (.0007) Trade Openness (weighted) (.006) (.005) (.004) (.008) domestic influences Monitoring Capacity -1.89e e e e 8 (8.26e 9 ) (1.97e 7 ) (4.83e 9 ) (1.18e 8 ) Recent Repression (.346) (.348) (.188) (.324) ln(population) -4.27e e 6 (1.18e 6 ) (9.76e 7 ) interactive influences Intervene Against*Capacity 2.97e e 7 (1.60e 7 ) (2.83e 7 ) controls Inverse Mills Ratio (2.152) (.973) (1.934) (.829) Civil and/or Int l War (1.522) (1.308) (1.030) (.870) Constant (4.307) (1.599) (2.750) (1.468) Interaction terms joint test N (non-zero N) 691(61) 691(61) ln(α) Log pseudo-likelihood Wald χ 2 (df) (9) (10) *** p 0.01; ** p 0.05; * p 0.1, (two-tailed). Robust standard errors, in parentheses, are clustered by country. : χ 2 test (3df) of H 0: Intervention, capacity, and their interaction are jointly insignificant. 34

35 monitoring capacity is insignificant in both stages of both models, the new proxy (military expenditures) is positive and significant in both stage two estimations. The substantive results are even more robust, supporting virtually all substantive insights generated by the main model. Figure S.5 shows the positive additive relationship between state monitoring capacity (proxied through military spending) and the likelihood and severity of state killing. As shown in Figure S.6, hostile intervention increases the likelihood that killing can occur, and decreases its severity when it does. Figure S.7 shows that supportive intervention decreases the likelihood of killing while increasing civilian death tolls. Finally, Figure S.8 presents the moderating impacts of hostile intervention on the positive relationship between monitoring capacity (i.e., military spending) and killing. The top panel shows that in the absence of such intervention, increased capacity leads to an increased likelihood and severity of killing. When hostile intervention is present, the likelihood of killing quickly declines to zero. The bottom panel of Figure S.8 shows that, in the absence of hostile intervention, increased capacity leads to increasingly high death tolls. At low levels of military spending, hostile intervention increases this number; as military spending increases, however, the relationship inverts dramatically and intervention lowers the expected death toll. In sum, like the results reported in the paper, these results suggest that the likelihood of killing and resultant death tolls covary positively with military spending, that hostile intervention makes killing more likely but less severe, that supportive intervention makes killing less likely but more severe, and that the relationship between hostile intervention and domestic monitoring capacity (i.e., military spending) is interactive such that intervention against the government decreases both the likelihood and in most cases the severity of killing. 35

36 Figure S.5: Sensitivity Analysis: Additive Impact of Domestic Monitoring Capacity on Government Killing Source: Table S.14, columns one (top panel) and two (bottom panel). 36

37 Figure S.6: Sensitivity Analysis: Additive Impact of Hostile Intervention on Government Killing Source: Table S.14, columns one (top panel) and two (bottom panel). 37

38 Figure S.7: Sensitivity Analysis: Additive Impact of Supportive Intervention on Government Killing Source: Table S.14, columns one (top panel) and two (bottom panel). 38

39 Figure S.8: Sensitivity Analysis: Moderating Impact of Hostile Intervention on the Relationship Between Monitoring Capacity and Killing Source: Table S.14, columns three (top panel) and four (bottom panel). 39

40 References Ball, Patrick and Christian Davenport Views to a Kill: Exploring the Implications of Source Selection in the Case of Guatemalan State Terror, Journal of Conflict Resolution 46(3): Ball, Patrick, Paul Kobrak and Herbert Spirer State Violence in Guatemala, : A Quantitative Reflection. Washington, D.C.: American Association for the Advancement of Science. Bendor, Jonathan, Amihai Glazer and Thomas H. Hammond Theories of Delegation. Annual Review of Political Science 4: Bueno De Mesquita, Bruce, Feryal M. Cherif, George W. Downs and Alastair Smith Thinking Inside the Box: A Closer Look at Democracy and Human Rights. International Studies Quarterly 49(3): Bueno de Mesquita, Bruce, James D. Morrow, Randolph M. Siverson and Alastair Smith The Logic of Political Survival. Massachussetts: MIT Press. Bueno de Mesquita, Bruce, Michael T. Koch and Randolph M. Siverson Testing Competing Institutional Explanations of the Democratic Peace: The Case of Dispute Duration. Conflict Management and Peace Science 21(4): Davenport, Christian State Repression and the Domestic Democratic Peace. New York: Cambridge University Press. Davenport, Christian and David A. Armstrong Democracy and the Violation of Human Rights: A Statistical Analysis from American Journal of Political Science 48(3):

41 Downes, Alexander B Restraint or Propellant? Democracy and Civilian Fatalities in Interstate Wars. Journal of Conflict Resolution 51(6): Downes, Alexander B Targeting Civilians in War. Ithaca, New York: Cornell University Press. Fein, Helen Accounting for Genocide after 1945: Theories and Some Findings. International Journal on Group Rights 1: Gleditsch, Nils Petter, Peter Wallensteen, Mikael Eriksson, Margareta Sollenberg and Havard Strand Armed Conflict : A New Dataset. Journal of Peace Research 39(5): Harff, Barbara No Lessons Learned from the Holocaust? Assessing Risks of Genocide and Political Mass Murder Since American Political Science Review 97(1): Heckman, J.J Sample Selection Bias as a Specification Error. Econometrica 47(1): Kalyvas, Stathis N The Logic of Violence in Civil War. Cambridge, UK: Cambridge University Press. Keith, Linda Camp Constitutional Provisions for Individual Human Rights ( ): Are They More Than Mere Window Dressing?. Political Research Quarterly 55(1): Lemke, Douglas and William Reed The Relevance of Political Relevant Dyads. Journal of Conflict Resolution 45(1): Lichbach, Mark Irving Deterrence or Escalation? The Puzzle of Aggregate Studies of Repression and Dissent. Journal of Conflict Resolution 31: