«January 2014 Abstract This paper demonstrates that civilian participation in violence during civil conﬂicts can largely be explained by the ...»
The Escalation of Violence:
Armed Groups and Civilian Perpetrators∗
This paper demonstrates that civilian participation in violence during civil conﬂicts can largely
be explained by the presence of elite-level armed forces. Guided by a simple model, we empir-
ically examine how the central Rwandan Hutu government strategically allocated armed groups
like the Interahamwe militia or the National Hutu army to maximize civilian participation in the Rwandan Genocide. To establish causality, we exploit cross-sectional variation in transport costs
- the interaction of the shortest distance to the main road with rainfall along the way - at the village level. Our instrumental variables estimates reveal a huge multiplier effect: one additional militia man resulted in 7.2 more civilian perpetrators. High levels of Hutu radio ownership, a good proxy for information, do not mitigate the militia’s effects, indicating that it was the physical presence of armed groups which was crucial for mobilizing the population. Furthermore, building on the model predictions, we rule out that civilians were, at least the majority, forced into joining the killings. Instead, the data suggests that armed groups acted as catalysts and civilians followed orders.
JEL classiﬁcation: D74, N47 Keywords: Conﬂict, Rainfall, Transport Costs ∗ Thorsten is indebted to his advisors Torsten Persson, David Str¨ mberg, and Jakob Svensson for their invaluable o guidance, encouragement and support. We also thank Philippe Aghion, Konrad Burchardi, Masa Kudamatsu, Andreas Madestam, Laura Mayoral, Peter Nilsson, Rohini Pande, Debraj Ray, Jacob Shapiro, and Stergios Skaperdas as well as participants at the Summer Forum in Barcelona, SIPRI, ASWEDE Meeting, LICOS seminar, and HiCN workshop for many helpful comments. We also wish to thank Milda Jakulyte-Vasil for invaluable help with the Lithuanian Holocaust data.
† firstname.lastname@example.org, corresponding author.
‡ email@example.com 1 Introduction In many genocides and civil wars ordinary civilians with no military afﬁliation or military training whatsoever turn into perpetrators. To illustrate, during the Rwandan Genocide in 1994 Hutu perpetra- tors killed approximately 800,000 people of the Tutsi minority in just about 100 days (Prunier, 1995).
This high death rate could only be achieved because hundreds of thousands of civilians (about 85 percent of the total number of perpetrators) joined the militia and the army in the killings. In light of the immense human suffering, it is crucially important, especially for international policy makers contemplating an intervention, to understand the factors that trigger civilian participation. Two views coexist. On the one hand, civilian participation is often interpreted as an unstoppable outbreak of ancient hatred, usually fought along ethnic lines, thus ruling out a successful foreign intervention. One retired US admiral remarks on the subject, referring to the Bosnian War: ”Let them ﬁght. They’ve been ﬁghting for a thousand years.”2 There was no foreign intervention in Rwanda. On the other hand, anecdotal evidence suggests that strategically used elite-level armed groups trigger civilian participation. Brown (1996, p.23) writes, ”Elite decisions and actions are usually the catalysts that turn potentially volatile situations into violent confrontations.”3 Elite groups are naturally of much smaller size and thus potentially easier to stop. For example, General Romeo Dallaire - the Canadian commander of the UN force in Rwanda - insisted that with 5000 to 8000 well-equipped troops he could have stopped the genocide, hindering the various militia and army groups in Kigali and other big cities from spreading into the country.
This paper provides the ﬁrst empirical analysis of how important elite-level armed groups might be in inducing civilians to participate in killings. It answers three questions: First, how much do armed groups affect civilian participation? Second, what are the channels? Third, do armed-group leaders allocate their militiamen strategically? In answering these questions, we will focus on the Rwandan Genocide, to our knowledge the only conﬂict where data on civilian and armed-groups violence is separately available at a local village level.
The main difﬁculty in estimating the effects of armed groups on civilian participation arises from joint determination and reverse causality. Furthermore the direction of the bias is a priori unclear. On the one hand, village-speciﬁc unobservable characteristics that affect both civilian and armed-group violence, for instance local leader quality, could produce a spurious positive correlation between the two, biasing the estimate upwards. On the other hand, if army and militia were strategically sent into areas where civilian participation was unobservably low, the estimate would be downward biased.4 To overcome these endogeneity issues, we use an exogenous measure of transport costs for estimating the effect of armed groups on civilian participation in civil conﬂict. More speciﬁcally, we exploit two sources of variation. First, we exploit variation in distance to the main road. There is abundant anecdotal evidence that army and militia troops were sent around the entire country to promote the killings. Because the few main roads crossing the country in 1994 were the only ones in 2 Rear Admiral James W. Nance (ret.) is quoted in Tom Ashbrook (1995).
3 Of the 34 major internal conﬂicts he considers 26 were elite-triggered (Brown, 1996).
4 In addition, measurement error might bias the OLS estimate downwards.
1 reasonable condition, we expect areas further away from these main roads to be harder to reach by the militia. However, because distance to the road is certainly correlated with other, possibly unobservable, determinants of civilian violence such as education, health or income, we further exploit variation in rainfall during the period of the genocide, introducing a novel, high resolution rainfall data set.5 Speciﬁcally, the instrument is the distance to the main road interacted with rainfall along the way between each village and the closest point on the main road during the period of the genocide. The idea is simple: We expect the movements of army and militia, mostly performed by motor vehicles, to be limited by the heavy rains that characterize the ﬁrst rainy season, which partly overlaps with the genocide.
Using this interaction has several advantages; ﬁrst, it allows us to control for the main effects of our instrument, in particular distance to the road. Second by controlling for distance to the road interacted with rainfall between village and road during the 100 genocide days of an average year we only exploit the seasonal weather variation in the year of the genocide. Finally by controlling for rainfall during the 100 days in 1994 and its long-term average in each village, that is at the armed group’s destination, we ensure that identiﬁcation only stems from short-term variation in rainfall along the distance measure, which is arguably exogenous and should only affect the militia’s transport costs.6 We ﬁnd that higher transport costs are in fact strongly negatively related with the number of militiamen arriving in a village (ﬁrst stage). Following a one standard-deviation increase in rainfall, a village with an average distance to the road receives 18 fewer militiamen (35 percent of the mean).
There is one concern regarding the excludability of our instrument: villages that were difﬁcult to reach by armed groups might have also been difﬁcult to reach by civilian killers or informants. However, civilian violence was very localized and we will devote an ample amount of care to corroborate that this concern is unwarranted.
We proxy for armed-group and civilian violence by the number of people prosecuted for armedgroup genocide violence and civilian violence in the Gacaca courts, normalized by village Hutu population. There were about 10,000 of these local (grassroots) courts set up all over the country to prosecute the crimes committed during the genocide.7 Using prosecution instead of actual participation rates might introduce some bias. However, ﬁrst we show that the Gacaca data is strongly correlated with other measures of violence from various different sources. Second, we also directly take potential bias into account in the empirical analysis.
The OLS results indicate a positive relationship between armed group and civilian violence: a 1 percent increase in the number of militiamen per Hutu is associated with a 0.631 percent increase in the civilian participation rate. In contrast, the instrumental variable estimates are about twice as large. The numbers imply that, on average, one additional militia man resulted in 7.2 more civilian perpetrators. Put differently, the 50,000 militia and army men roaming the country, about 10 percent 5 The genocide lasting only 100 days is another advantage for our identiﬁcation strategy as this limits the presence of time confounding factors.
6 Rainfall in each village might be correlated with malaria prevalence or civilian’s transport costs within the village, both of which are likely to directly affect civilian participation.
7 Gacaca roughly translated means short, clean cut grass.
2 of the total number of perpetrators, were directly and indirectly responsible for at least 82 percent of the Tutsi deaths.8 The results are robust to the inclusion of various geographic controls, they pass several indirect tests concerning the exclusion restriction and are also relevant for other cases of state-sponsored murder, in particular the killings of the Jews in Lithuania in the 1940s.
In the second main part of the paper we examine different channels through which armed groups might have spurred civilian participation. First, we show that high levels of Hutu radio ownership in a village, a good proxy for information, do not mitigate the militia’s effect on civilians. This result is consistent with the militia providing more than simple information about the ongoing genocide or put differently that their physical presence was crucial for mobilizing the population. A natural next question is whether the militia needed to force opposing civilians to join in the killings or whether they rather organized the killings and taught civilians how to kill? Unfortunately we do not have data to directly distinguish between these two possibilities. So instead, we test the theoretical implications of these two scenarios.
Our model suggests that the militia’s effects are increasing if force is needed and decreasing otherwise, a result based on the reasonable assumption that forcing Hutu civilians to participate gets more effective the more militia members arrive but that only a few militiamen are needed to provide an example. Furthermore the model predicts different signs for the interaction effect with the Tutsi minority share. In the data, we ﬁnd decreasing effects of the militia as well as a positive interaction effect with the Tutsi minority share. Both these ﬁndings suggest that the militia did not force people to join in the killings but rather functioned as a role model.
In the last main part of the paper, we ﬁnd that the central planners in Kigali allocated their armed groups strategically. We model a genocide planner who wants to maximize civilian participation (under the role model) but faces a transport constraint. Both predictions of the model are conﬁrmed in the data: villages which are more costly to reach receive fewer militiamen (this is essentially the ﬁrst-stage result; however, since that result might be driven by villages which were, due to very bad weather, simply impossible to reach, we also drop those villages with high rainfall and the negative relationship remains). Furthermore, transport costs matter less for villages with large Tutsi group shares.
Our paper contributes to the literature in several ways. First of all, it adds to the vast conﬂict literature, of which Blattman and Miguel (2010) give an excellent review, vehemently calling for well-identiﬁed studies on the roots of individual participation in violent conﬂict and the strategic use of violence. This paper starts ﬁlling the gap, adding to the literature on the determinants of conﬂict by providing novel evidence on the strong effects that armed groups have on civilian participation and on the strategic use of armed groups. Recent studies consider government policy, foreign aid, propaganda, income, and institutions (Dell, 2012; Nunn and Qian, forthcoming; Yanagizawa-Drott, 2012; Dube and Vargas, 2012; Besley and Persson, 2011, respectively). Furthermore, our paper complements the literature on the causes of the Rwandan Genocide (Verwimp, 2003, 2005, 2006;
Straus, 2004; Friedman, 2010; Verpoorten, 2012b).
8 See the results section for the necessary assumptions for this back-of-the-envelope calculation.
3 Our ﬁndings also speak to the discussion on the effects of rainfall on conﬂict other than through the income channel (Sarsons, 2011). Our results suggest that especially in areas with poor infrastructure, such as Africa, rainfall might have negative direct effects on conﬂict through transport costs.
Regarding the importance of transport costs our paper contrasts recent contributions by Donaldson (forthcoming), Faber (2013), and Banerjee et al. (2012) that highlight the positive economic effects of low transport costs. Our ﬁndings loosely echo the ones in Nunn and Puga (2012) who show that high transport costs in Africa, i.e. rugged terrain, have positive effects on todays GDP because they hindered slave traders.
The remainder of the paper is organized as follows. Section 2 provides some background information on the Rwandan Genocide. Section 3 sets up a simple model to highlight three channels through which armed groups might have affected civilian killings and to distinguish between these channels in the data. It also models the strategic use of armed groups from a central genocide planner perspective.
Section 4 presents the data used for the analysis and section 5 lays out our empirical strategy. Section 6 presents the results and accesses their robustness. Section 7 discusses the external validity of our results. Section 8 concludes with possible policy implications.