Violence and the State:

Evidence from Rwanda’s ‘Decade of Atrocities’∗

Leander Heldring

December 2016JOB MARKET PAPER

For the latest version, click here

Abstract

This paper shows that contemporary patterns of violence can be traced back to the establishment of

the precolonial state. Rwandan villages that were brought under centralized rule one century earlier

experience a doubling of violence during the state-organized 1994 genocide. Instrumental variable esti-

mates exploiting differences in proximity to Nyanza – an early capital – suggest these effects are causal.

In other periods, when the state pursued peace and rebuilding, with longer state presence, violence is

lower. Using data from several sources, including a lab-in-the-field experiment, I show that exposure

to state institutions impacted civil society, and in particular culturally transmitted norms of obedience

to political authority. In a lab setting today, individuals close to an abandoned border of the historical

state are more likely to follow an unenforced rule than individuals just across the border. The state’s im-

pact on rule following led to more violence when the Rwandan government mobilized for mass killing

and to less violence when the government pursued peace and rebuilding. These results suggest that the

interaction of government policy with deep-rooted aspects of civil society has the potential to reconcile

long-run persistence with rapid economic change.

Keywords: Violence, States, Rwanda. JEL classification: D73, D74, H70, N4.

∗Harvard University, Department of Economics, Littauer Center, 1805 Cambridge Street Cambridge, MA 02138, United Statesof America; e-mail: lheldring@iq.harvard.edu. Website: www.leanderheldring.com. This paper was previously titled“State Capacity and Violence: Evidence from the Rwandan genocide”. I would like to thank Melissa Dell, James Fenske, NathanNunn and James A. Robinson for helpful suggestions and conversations throughout this project. I would like to thank DaronAcemoglu, Martin Abel, Faisal Ahmed, Robert Allen, Robert Bates, Paul Collier, Simon Franklin, Ed Glaeser, Maximilian Kasy,Akos Lada, Horacio Larreguy, Karlijn Morsink, Suresh Naidu, Jean-Philippe Platteau, Kirsten Pontalti, Simon Quinn, Gautam Rao,Dan Rogger, Valeria Rueda, Philip Verwimp, Jonathan Weigel and David Yanagizawa-Drott for helpful conversations as well asseminar participants at Harvard, Oxford, Utrecht, Groningen, the 2013 CSAE conference, the 2013, 2015 NEUDC conference andthe 2013 HiCN Annual Workshop for helpful comments. I would like to thank Sophie Mukatizoni, Theogene Nkurunziza, ÖmerÖzak, Philip Verwimp and David Yanagizawa-Drott for kindly sharing their data. I would also like to thank Christian Iradukunda,Fidele Munezero, Christophe Ndahimana, Manasse Twagiramungu, Tim van der Maarel as well as several Rwandan enumeratorsfor excellent research assistance. Funding is gratefully acknowledged from the International Growth Centre, the Economic HistorySociety and the International Peace Research Association Foundation.

Several recent studies examine the long-run impact of precolonial states in Africa (Gennaioli and

Rainer, 2007; Michalopoulos and Papaioannou, 2013). The authors find that achieving greater bureau-

cratic complexity before colonization has a significant positive effect on contemporary economic devel-

opment. This effect has persisted to the present despite most precolonial states being absorbed into

larger countries during the Scramble for Africa (Michalopoulos and Papaioannou, 2016), despite Euro-

pean colonial powers radically changing political institutions (Acemoglu et al., 2001) and despite post-

independence governments frequently failing to represent the interests of inhabitants of these former

states (Easterly and Levine, 1997). Although the effect of precolonial states on development is arguably

causal, these studies do not focus on identifying exact causal mechanisms.

In this paper, I study civil society as one of the channels through which precolonial states impact con-

temporary economic development. Combining a village level data set and a lab-in-the-field experiment

with local variation in the establishment of Rwanda’s precolonial state, I test the hypothesis that exposure

to state institutions causes stronger norms of obedience to political authority to develop (Putnam et al.,

1994; Guiso et al., 2014). As the Rwandan state developed, it initially enforced demands, such as taxation

or conscription, by force. As the threat of violence became credible, individuals optimally responded by

obeying political authorities (Vansina, 2004; Des Forges, 2011). If over time obedience to authorities be-

comes a norm about the course of action which is most beneficial when faced with a demand from the

government – in other words, obedience becomes internalized –, it may persist even when the state weak-

ens or when governments change.1 Furthermore, historical differences in the strength of these norms may

be measurable today.

The setting of this paper is Rwanda, which is an attractive setting for two main reasons. First, I can

exploit the fact that the precolonial Rwandan state, called the Nyiginya kingdom, expanded slowly be-

tween its inception before 1700 and colonization in 1897, introducing local variation in exposure to state

institutions across precolonial districts.2 Second, Rwanda’s recent history of violence provides an attrac-

1I understand a ‘norm’ in this paper as a belief about the ‘right’ course of action. Norms may optimally arise as a heuristic fordecision making when information processing is costly (see the evidence summarized in Nunn (2012)).

2The focus in this paper is therefore on the intensive margin of the presence of the state, comparing villages in districts that wereincorporated earlier to those incorporated later. Because initial state formation was often resisted and bureaucratic hierarchies tookto time to form, I expect the effect of state on rule following behavior to be stronger with longer exposure to state institutions. Thisidea is consistent with notions of the accumulation of social, civic, and democratic capital discussed in Putnam et al. (1994); Guisoet al. (2014); Persson and Tabellini (2009) and Guiso et al. (2010). In a recent model of the incentives to cooperate by Tabellini (2008)the establishment of a government has a slowly diffusing effect on civil society due to parent-child transmission of cooperativevalues. Because the incentives for parents to inculcate cooperative values depend on the number of cooperators in society, it maytake several generations for stronger norms of cooperation to develop (Tabellini, 2008). Similar mechanisms predict that whenobedience is enforced by a community, enforcement norms are built up slowly (Acemoglu and Jackson, 2016). One reason forcommunity enforcement of obedience is the possibility of collective punishment for disobedience. Consistent with this logic, initialestablishment of the Rwandan state often involved the ransom of a chief (Des Forges, 2011).

1

tive, albeit tragic, setting to study the hypothesis of this paper because the motives of the government

changed dramatically within a short period of time. Specifically, between 1990 and 1993, the Rwandan

Hutu government pursued territorial control in the face of a rebellion. In 1994, it instead organized the

Rwandan genocide, in which it pursued mass mobilization of the Hutu majority population to extermi-

nate the Tutsi minority. After the genocide, a new – Tutsi – government sought to re-establish territorial

control while fighting former genocide perpetrators.3 In each of these episodes, observers have noted the

impact of Rwanda’s long tradition of centralized government. Studying the genocide, Gérard Prunier

writes: “Rwandese political tradition, going back to the Banyiginya Kingdom ... is one of systematic,

centralised and unconditional obedience to authority” (Prunier, 1995, p. 141). Studying post-genocide

rebuilding, Philip Reyntjens writes: “An ancient state tradition played an undeniable role here: a mere

two years after the extreme human and material destruction of 1994, the state was rebuilt. Rwanda was

again administered from top to bottom” (Reyntjens, 2013, p. 25).

Combining village level data on violence with my reconstruction of the expansion of the state, I find

that villages where the state was established earlier experience more violence during the genocide.4 This

result is consistent with individuals being more likely to comply with the state’s call to mobilize for vio-

lence in areas with longer state presence.5 In contrast, in the years just prior to and just after the genocide,

and again consistent with the hypothesis of this paper, I find that violence is lower.6 Figure 1 provides a

timeline of the main events studied in this paper and a summary of these results.

An alternative explanation for these results may be that ethnic polarization was stronger in places

where the state formed earlier, resulting in more violence against the minority Tutsi. I address this possi-

bility by controlling for the fraction of Tutsi in several regressions. I also control for differences in house-

hold income in 1990, as a summary measure of differences in development before the genocide.7 I then

3The Hutu and Tutsi are the main social groups in Rwanda, with the Hutu being the large majority. Although Nyiginya kingswere exclusively Tutsi, following a 1959 coup Rwanda had a Hutu-led government until the end of the genocide (Verwimp, 2013).

4I proxy for violence during the genocide by data on prosecutions for genocide violence, which I normalize by pre-genocidepopulation. The genocide violence data come from the Gacaca system, which was a network of 8000 traditional courts instituted tocombat the enormous backlog of cases against alleged genocide perpetrators. These data have been used before by e.g. Yanagizawa-Drott (2014). I verify that the effect of state presence on violence is consistent across alternative data sources. I measure violence inthe years surrounding as the count of violent events recorded and geocoded by the Uppsala Conflict Data Program (Sundberg et al.,2010; Sundberg and Melander, 2013).

5The data admit a more fine-grained study of types of violence. I separately study individuals involved in the organization ofthe violence, and participation of ordinary civilians. I find that the effect is stronger for mobilization of ordinary civilians.

6I focus on 1990-1993 and 1995-2000. 1990 coincides with the first rebel attacks that lead up to the genocide. By 2000, the Rwandangovernment had re-established a monopoly of violence and only sporadic violent events are recorded. Gerald Gahima, a formerrepresentative with the International Criminal Tribunal for Rwanda, referred to the period studied in this paper as a ‘decade ofatrocities’ (Gahima, 2013, Ch. 2).

7Ethnic composition and household income can be viewed as outcomes of the formation of the precolonial state. I thereforeseparately regress these measures on state presence, and find that there is no systematic correlation.

2

directly re-estimate the relationship between state presence and violence using distance to Nyanza, an

early capital village of the Nyiginya kingdom, as an instrumental variable. Specification checks using

data on density of archeological evidence for population density, terrain characteristics such as elevation,

and soil suitability for growing food show that Nyanza was not differentially densely settled or produc-

tive before the inception of the state. To account for a potential direct effect of proximity to Nyanza in

the sample period, I control for distance to Nyanza along the 1988 road network.8 The IV regressions

produce results that are qualitatively similar to the baseline OLS results. A one century increase in the

presence of the precolonial state at a local level is associated with an eight percentage point increase in

violence during the genocide (relative to a mean of 8%) and a 37% decrease in the number of violent

events before and after the genocide. Besides providing support for the hypothesis of this paper, the

results for the years surrounding the genocide support the hypothesis that precolonial states positively

impact long-run development in Africa (Gennaioli and Rainer, 2007; Michalopoulos and Papaioannou,

2013; Depetris-Chauvin, 2013). The result that violence during the genocide is positively affected by the

length of centralized rule supports the hypothesis of this paper and is consistent with a literature that

shows that factors usually considered conducive to development can have adverse consequences for de-

velopment as well (Satyanath et al., 2013; Acemoglu et al., 2014). Taken together, my findings show that

historical institutions can have rapidly changing effects on development outcomes depending on govern-

ment objectives.

Before turning to mechanisms, I test for heterogeneous effects by radio ownership. New rules need to

be communicated and legitimized (Glaeser, 2005), and because the Rwandan government used the radio

to formally sanction violence against Tutsi (Yanagizawa-Drott, 2014), I expect the effects of obedience to

authority to be stronger where the government can communicate its orders more effectively.9 I find sup-

port for this hypothesis using data on radio ownership in 1991. During the genocide the positive effect of

state presence is stronger in places with high radio ownership. Consistent with the changing objectives

of the government, the negative effect of state presence before and after the genocide is stronger in areas

with high radio ownership as well.

Having established the reduced form relationship between state presence and violence, I then turn

8I report the main results in this paper using the straight line distance to Nyanza as the instrument, and the driving distanceto Nyanza and Kigali – Rwanda’s current capital city – along the 1988 road network and the distance to the national border ascovariates. In addition, I compute several measures of walking time between Nyanza and the centroid of each precolonial district.I use these measures as alternative instruments. The results using each instrument, and without the driving distance covariates, arevery similar to the baseline strategy.

9There is a subtle distinction between following rules and following orders. However, because in the native language of Rwan-dans a law and an order are referred to by the same word (in Kinyarwanda both are igeteko), I do not pursue this distinction.

3

to distinguishing causal mechanisms. The first mechanism I consider is stronger norms of obedience to

political authority, which is the focus of this paper. I first directly measure rule following behavior in a

lab-in-the-field experiment with 420 participants in rural Rwanda, and then test for cultural transmission

of these norms over time. To separate rule following from the effects of the modern state, social networks

or other external influences, I ask participants to follow a rule in an anonymous lab setting.10 In the ex-

periment it is costly to follow a rule and beneficial to cheat. Anonymity implies that I do not observe

individual rule following, but I can compare rule following behavior over larger populations. I introduce

variation in historical state presence by comparing participants that live close to, but on either side of, an

abandoned outer border of the historical state in the eighteenth century. Across this border, the presence

of the state discretely changes by about one century. I find that individuals that are close to the boundary

but just happen to be on the side that has longer state presence exhibit significantly greater rule following

behavior.11 It may be the case that although there are differences in norms of rule following behavior, these

are sustained by differences in local government. To further distinguish between internalized norms of

rule following and factors that are external to the individual, such as socializing effects of education and

government institutions, I compute a second treatment measure based on where the respondent grew up,

rather than the location of the fieldwork. The idea behind this measure is that when participants move,

they take their norms and values with them into a new institutional environment. If the main source of

persistence is cultural transmission, then the place where participants grew up should be more important

than where they live. If local differences in institutions are the primary source of persistence their location

should be more important than the place of birth. My estimates show cultural transmission of norms of

compliance is the main source of persistence.

While supportive of the hypothesis of this paper, findings in a lab setting may not generalize to the

wider population. I therefore use data from household surveys to further test the hypothesis that exposure

to state institutions affects rule following, and to further distinguish between cultural versus institutional

sources of persistence. I first use data from the World Values Survey, and match respondents to precolo-

nial districts based on their place of residence. Individuals that live in places with longer state presence

are less likely to find rule breaking acceptable, and are more likely to think obedience to authorities is

central to democracy. I then turn to participation in village level meetings between citizens and govern-

ment representatives. Individuals who today live in areas with longer state presence are equally likely to

10I implement a modified version of the resource allocation game pioneered by Hruschka et al. (2014).11I show that trust, social capital, migration and altruism towards the government balance. In an accompanying paper (Heldring,

2016), I show that the presence of the local government, corruption, and several other measures that are likely candidates for omittedvariables also balance across the boundary.

4

attend community meetings, but are less likely to speak up in these same meetings and the government is

more likely to be the only party speaking. I finally assess a directly observable outcome of rule following.

The 2003 Rwandan constitution criminalized violence against women and I use data from the 2014 De-

mographical and Health Survey to study violence against women in the household. I find that women in

villages with longer state presence are less likely to experience verbal abuse, physical violence or sexual

violence at the hands of their partners. I then use government data on local provision of public goods

as well as local taxation and public expenditure to test whether state presence affects local government.

Across a wide range of public goods and revenue and expenditure categories I find precisely estimated

zero effects of state presence. Using data on several public goods and different years during the colonial

period, I verify that there are no differences in public good provision for this period either.12

The results in this paper provide support for its hypothesis: Exposure to state institutions historically

affects cultural norms of rule following today. Its ultimate effect on development, however, is determined

by the interaction of this channel with government policy. It is plausible that such effects are not con-

fined to the Rwandan context. For instance, rule following of the German population is often associated

with the successful execution of the Holocaust (Goldhagen, 1996). When the German government does

not pursue genocide, rule following behavior likely facilitates economic exchange and positively affects

prosperity (Tabellini, 2010). More broadly, the results in this paper speak to the literature on the persistent

effects of history (see Nunn (2009) for an overview). Persistent effects may be mediated by policy shocks

resulting in time-variant reduced form effects of historical institutions. This idea potentially explains why

we simultaneously observe persistence of economic development as well as rapid economic change.

The results in this paper furthermore complement a large literature that identifies a positive reduced

form effect of historical states on modern development outcomes (Acemoglu et al., 2015; Bockstette et al.,

2002; Dell et al., 2015; Depetris-Chauvin, 2013; Dincecco and Katz, 2014; Gennaioli and Rainer, 2007;

Michalopoulos and Papaioannou, 2013). I contribute to this literature by providing a mechanism that

drives these effects. By providing evidence for civil society as a mechanism, this paper contributes to the

literature that finds a positive relationship between historical states and civil society (Putnam et al., 1994;

Guiso et al., 2014) and to a growing empirical literature in economics that finds more mixed results. For

instance, Dell et al. (2015) show that when the state exercises its capacity through a strong civil society,

12I finally show that state presence does not affect population growth differentially, which has been hypothesized to lead toMalthusian pressures and violence (André and Platteau, 1998). There is also no effect of state presence on the coverage of thegovernment radio state whose propaganda affected violence during the genocide (Yanagizawa-Drott, 2014). In addition, controllingfor these variables does not affect the baseline effects of state presence on violence.

5

it is less open to foreigners, Acemoglu et al. (2014) show that chiefs in Sierra Leone use civil society in-

stitutions to cement their own authority and Satyanath et al. (2013) show that higher social capital led

to increased Nazi party membership in Germany.13 By showing that rule following is affected by the

historical Rwandan state, I contribute to a literature that tries to understand the effects of history on mod-

ern preferences, attitudes, and beliefs (Alesina and Fuchs-Schündeln, 2007; Di Tella et al., 2007; Nunn

and Wantchekon, 2011; Giuliano and Spilimbergo, 2014; Becker et al., 2015), and to a small but growing

literature that uses experimental methods to study long-run economic development (Lowes et al., 2014;

Blouin, 2014). By providing evidence that state presence and compliance are important determinants of

within-country violence, I contribute to the study of the determinants of violence and conflict (Blattman

and Miguel, 2010; Besley and Persson, 2009; Dube and Vargas, 2013; Collier and Rohner, 2008; Esteban

et al., 2012), as well as to the study of individual motivations to join conflict (Humphreys and Weinstein,

2008). Finally, this paper is related to a large and heterogeneous literature on socialization by nation states,

parents and modern governments, see e.g. Anderson (1983) and Cantoni et al. (2014). The papers that are

closest to this study are Yanagizawa-Drott (2014) who provides empirical evidence that the government

radio was instrumental in organizing the genocide and Lowes et al. (2014) who study the effects of the

historical Kuba kingdom on rule following today.

The rest of this paper proceeds as follows. Section 1 gives an overview of the relevant episodes of

Rwandan history. Sections 2 presents the estimation frameworks used in this paper. Section 3 presents

the main results of this paper, relating state presence to violence. Sections 4 provides direct evidence for

rule following as a mechanism, and section 5 concludes. An appendix provides additional results.

1 Historical background

This section presents the relevant historical and institutional background to the estimations in the rest of

this paper. Before discussing the recent history of conflict in Rwanda, I outline the expansion and the

organization of Rwanda’s precolonial state, the Nyiginya kingdom.

1.1 The expansion and organization of the Nyiginya kingdom

The Nyiginya kingdom was founded in the late sixteenth century and was initially confined to a small

area in central Rwanda called Nduga (Vansina, 2004). From 1700 to 1897, the small kingdom expanded

13In political science, there is a large literature on the interaction between the state and civil society, see for instance Mann (2005),Valentino (2013) and Bates (2008). See also Alesina and Giuliano (2015).

6

to Rwanda’s current borders (see figures A1 and A2 in the appendix). In 1897 Rwanda became part of

German East Africa and in 1916 it became a Belgian colony. The kingdom replaced lineage-based kin-

ship groups called umulyango. Although there were pressures towards consolidation of these groups,

there was no polity that managed to centralize power and project its authority beyond the limits of the

extended kinship group. The Nyiginya kingdom was the first polity to overcome these constraints by

forming a standing army and organizing a rudimentary bureaucracy.

These innovations allowed the Nyiginya army to expand from Nduga into one of the strongest pre-

colonial African states. The Murdock ethnographic atlas, which attempts to record systematic information

on precolonial African ethnic groups classifies Rwanda as having three levels of bureaucracy above the

village level (Murdock, 1967). The Nyiginya bureaucracy was therefore comparable in sophistication to

the Buganda state in Uganda and the Asante empire in Ghana.

Relative to these states, however, the Nyiginya kingdom was highly informal: It provided no public

goods, had no money or writing, and blocked all trade except for a few luxury items. It sustained its

army and its bureaucracy through patron-client networks in which protection was exchanged for labor

services, and in which chiefs raised taxes by appealing to pre-existing social norms prescribing the pay-

ment of tribute by their kinsmen (Vansina, 2004).14

Expansion generally involved a set of fixed steps (Des Forges, 2011). First, the army would move

into a territory and establish a monopoly of violence. The soldiers requisitioned food and support troops

through the chiefs of the local lineages (Umukuru w’umulyango). Second, over time, the court established a

formal bureaucracy that consisted of three classes of officials. Former army generals became administra-

tors charged with conscription and taxation of goods (these officials were called batware). Other officials

(banyabutaka) were responsible for land allocation, land grants and land taxes. A final class of officials

(banyamukenke) was responsible for pasture and cattle. Third, the local chiefs that used to be headmen of

their respective kinship groups were now coopted into the bureaucracy. The same norms that prescribed

paying tribute to the headman within the kinship group were now used to recruit conscripts for the army,

and raise taxes. Because every expansion required embedding the local elites into the patron-client net-

works that formed the bureaucracy, central authority was established slowly. Indeed, in areas that were

14Although only part of dense networks of mutual obligations between bureaucrats, cattle and land owners, and poorer individ-uals, the ubuhake contract is the most well known of these innovations (Vansina, 2004, p. 61). Under this arrangement, a patron gaveone or more head of cattle in usufruct to the client, in exchange for labor services. This system was the key to operating a bureau-cracy, since it personally bound each bureaucrat and military commander to the King, and established an enforceable commandstructure. The system was abolished in 1953 (Verwimp, 2013).

7

only recently incorporated, the German and Belgian colonial armies had to intervene regularly to stop tax

protests and outright rebellion against Nyiginya authority (see the case studies in Des Forges (1986, 2011),

Louis (1963) and Newbury (1987), as well as Botte (1985a,b) for a timeline of expeditions).

Aside from the theoretical reasons for longer exposure to state institutions being associated with

greater rule following mentioned in the introduction, the history of the kingdom provides further back-

ground to the hypothesis of this paper. Since the bureaucracy of the state built up slowly in the absence

of formal institutions, earlier establishment of the state plausibly results in more credible threats of pun-

ishment for disobedience and greater pressure for norms that prescribe rule following to develop.

Interwoven with the history of the state is the history of the Hutu and Tutsi identities. Although Hutu

and Tutsi as terms probably developed in the Nyiginya army, they quickly became social classifications of

rich, cattle-owning, elites (Tutsi) and poor, farming, masses (Hutu). The King was always from one of two

elite lineages which were Tutsi, the Abega and Abanyiginya, after whom the kingdom is named. The Bel-

gian colonizers turned this fluid economic distinction – upwardly mobile Hutu could become Tutsi and

Tutsi could become Hutu – into a rigid racial distinction by classifying every Rwandan as either Hutu

or Tutsi (or Twa, a marginal ethnic group accounting for about one percent of Rwanda’s population).

Although estimates vary, about 85% of Rwandans were Hutu and about 14% were Tutsi. The Belgians

favored Tutsi as their agents of indirect rule and restricted access to education and lucrative government

positions for Hutus.

In 1963 Rwanda gained its independence. The drive for independence resulted in part from a large

Hutu rebellion in 1959 which is seen as a culmination of the accumulated grievances held by Hutu

(Lemarchand, 1970).

1.2 Violence between 1990 and 2000

During the 1959 rebellion a large number of Tutsi fled to Uganda and many of their children joined the

Ugandan army. In 1990, they formed the Rwandan Patriotic Front (RPF) and invaded Rwanda. Between

1990 and 1993, the Rwandan army fought the RPF. It also used the fight against the RPF as a pretext to as-

sassinate political opponents and, as was later revealed, to organize practice killings for the genocide. In

1993, the invasion was halted and peace negotiations held in Arusha, Tanzania, resulted in the establish-

ment of a multiparty democracy. Several political parties were formed, some openly sympathetic to Tutsi

8

interests. Although multiparty democracy was supposed to lead to a peaceful integration of Hutu and

Tutsi in society and repatriation of exiled Tutsi, hardliners within the Rwandan government organized

the killing of all Tutsi which, they believed, would break the RPF’s power base and would keep them in

power (Mamdani, 2002).

On April 6th, 1994, an airplane carrying Rwanda’s president, Juvénal Habyarimana, was shot down

on its approach to Kigali. Within an hour, members of the political opposition were assassinated in their

homes and roadblocks were set up all across the country. After this initial wave of killings, which targeted

mostly senior government officials, the mass killing of all Tutsi began. In the hundred days that followed,

between 500,000 and 1,000,000 Tutsi and politically moderate Hutu were killed by the army, youth mili-

tias (the infamous interahamwe) or by machete blows from their next door neighbors, colleagues or even

family (Des Forges, 1999). It is estimated that about 75% of the number of Tutsi identified in the 1991

census were killed (Straus, 2006). Mobilization was extraordinary high, standing at 14-17 % of the adult

male population (Ibid.). The genocide ended when the RPF captured Rwanda’s capital, Kigali.15

As the RPF set out to rebuild, many Hutu genocide perpetrators fled to the Democratic Republic of the

Congo (DRC) and, to a lesser extent, to Tanzania. From a post-genocide population of about 6.9 million,

2.1 million fled to the DRC. Roughly 700,000 diaspora Tutsi emigrated back, almost fully replacing the

Tutsi population killed in the genocide. In total, about 3.6 million Rwandans, or about 45% of the pre-

genocide population, lived in their pre-genocide villages (Prunier, 2008). Between 1995 and 2000 former

members of the youth militias and the Rwandan army mounted raids and attacked government forces

from within Rwanda as well as across the border with the DRC. At the same time, the RPF established its

power by carrying out large numbers of covert killings.16 Around 2000, the RPF-dominated government

had re-established territorial control and had started a large-scale transitional justice effort, spearheaded

by a system of 8000 local courts, called Gacaca courts (see below).

15There is debate about the question when the government started planning for the genocide. It is clear that practice killings werebeing carried out in 1992 and 1993, which leads some commentators to point to the RPF invasion as the inception of the idea ofgenocide. For overviews, see Des Forges (1999) and Guichaoua and Degni-Segui (2010).

16These killings targeted families of former genocide perpetrators, educated citizens, members of the former Hutu political parties,and people who did not act ‘right’. According to Gérard Prunier: "What these people had in common was their constituting an actualor political elite” (Prunier, 2008, p. 20).

9

2 Estimation framework

The identification strategy of this paper exploits the fact that the Nyiginya expansion proceeded by adding

territory adjacent to its domain. Since armies and bureaucrats travelled from the administrative center of

the kingdom to conquer and administer new territories, places close to Nduga, its historical heartland are,

I hypothesize, more likely to be incorporated early. In order to take advantage of this idea I compute sev-

eral measures of distance to Nyanza, an early Nyiginya capital which was located in southern Nduga.17

After introducing these measures, the rest of this section shows that proximity to Nyanza is correlated

with state presence but uncorrelated with measures of pre-Nyiginya population density and measures of

agricultural productivity. Today, Nyanza is in a village called Nyabasindu and close to a modern town

also called Nyanza.

2.1 Constructing proximity to Nyanza

In this section I introduce four ways of measuring proximity to Nyanza, which I use as instrumental

variables in the rest of this paper. I compute these measures between Nyanza and the centroid of each

precolonial district. The reconstruction of these districts is described in the appendix.

Distance to Nyanza. The first instrument I use is the straight line distance in kilometers to Nyanza.

Distance is frequently used as a source of exogenous variation, such as in Becker et al. (2009) who use

distance to Wittenberg as an instrument for the spread of Protestantism, Dittmar (2011) who uses distance

to Augsburg as an instrument for the spread of the printing press and Nunn and Wantchekon (2011) who

use distance to the coast as an instrument for the intensity of the slave trade.

Cost distance to Nyanza (days). The second instrument I use considers elevation variability as an

impediment to travel. I compute optimal walking routes to Nyanza and I use the average travel time

along these routes to Nyanza (and back) as my second measure of proximity.18 I measure time in days

and assume that one day equals twelve hours of walking.19

17There were other capital villages such as Nyamagana and Nyundo. These may even have been established before Nyanza, but losttheir importance early on whereas Nyanza was prominent throughout the expansion of Nyiginya kingdom. By the early nineteenthcentury Nyanza had become the main capital (Newbury, 1991, p. 100), and Nyanza became Rwanda’s sole capital before colonization(Lugan, 1997). More pragmatically, Nyanza is also the only capital which can be precisely located. For Nyamagana and Nyundo,only approximate locations are known. Nyamagabe is said to have been in Southern Nduga (Vansina, 2004, p. 49), and Nyundo issaid to have been in Southern Nduga as well, close to a place called Bunyogombe (Vansina, 2004, p. 241). It is nevertheless possible touse modern village names to get a sense of where these capitals were. Using similar villages names near the approximate historicallocations suggests that Nyamagana and Nyundo were 13 and 15 kilometers away from Nyanza.

18For gentle slopes, walking downhill goes faster than going up. For steep slopes, both downhill and uphill travel is significantlyslower. Because of this asymmetry, I average over a return journey.

19I use Tobler’s function (Tobler, 1993) to convert elevation changes into travel speed. This function is calibrated on data collected

10

Figure 3 visualizes the process of generating this measure. Figure 3a maps elevation in Rwanda. Fig-

ure 3b plots contour lines of points that are six hours marching distance away from Nyanza. Figure 3c

maps the optimal paths from Nyanza to each district. Finding the optimal routes involves a straightfor-

ward application of Dijkstra’s algorithm (Dijkstra, 1959). Dell (2015) takes a similar approach.

Cost distance to Nyanza - Özak (days). Ömer Özak has proposed a Human Mobility Index, which

computes travel time taking into account not only elevation, but also weather patterns and soil conditions.

As a third instrument, I use travel time in days to Nyanza computed according to his procedure (Özak,

2010, 2013).

Cost distance to nearest Nyginya army location (days). I finally consider the expansion of the Nyiginya

kingdom using the location of Nyiginya armies as the origin of expansion, rather than Nyanza. Although

the Nyiginya bureaucracy was administered from the capital villages in the center of the kingdom, the

King positioned armies throughout the kingdom to defend against invasions and launch new expansions

(Vansina, 2004). I use data from Kagame (1963) who reports the location of each army battalion recruit-

ment as well as its year of establishment. I compute travel time to the nearest army that was established

before the annexation of a district as my fourth measure of proximity.

2.2 Estimating equations

I first estimate the relationship between state presence and violence. The unit of observation for the main

results is a sector, a small administrative unit which, at the time of the genocide, typically coincided with

a large village or several smaller villages. My baseline estimating equation is:

Ysdp = β0 + β1StatePresenced +X′dβ2 +Q′

sβ3 + rp + εsdp (1)

where s indexes sectors (n=1449 in most specifications), d indexes precolonial districts (n=50) and p

indexes modern provinces (n=5).20 Ysdp is an outcome of interest, such violence in the 1994 genocide.

StatePresenced is the number of years between the annexation of precolonial district d and colonization

in 1897.

by Imhof (1950). Travel time t(s) between two points that are one meter apart is a function of the increase in elevation over thatdistance, s, measured in degrees: t(s) = 18

5exp

(−3.5

∣∣(tan (s π180

))+ 0.05

∣∣). t(0) = 3.02205 km/h. To compute marching speedover paved terrain, multiply by 5/3 (Tobler, 1993).

20Sectors are the smallest relevant unit at the time of the genocide. In 2002 there was an administrative reform which reduced thenumber of sectors to 416. I use modern, post-reform, provinces since these provinces divide the country into four regions - north,east, south and west - plus Kigali and these are natural units for fixed effects. Eastern Rwanda has historically interacted with theDemocratic Republic of the Congo, for instance (Newbury, 1988).

11

Xd is a vector of travel distances along the 1988 road network. In addition to the distance to Nyanza,

this vector includes distance to Kigali, Rwanda’s current capital.21 Qs is a vector of sector level covariates,

including straight line distance to the national border. This vector also includes several other geograph-

ical and historical covariates. I discuss these as they are introduced in the empirics. This vector finally

also includes household income, as well as the fraction of the population that was recorded as Tutsi in the

1991 census.22 rp is a vector of fixed effects. εsdp is an error term.

When estimating the effect of state presence on violence in the genocide, I estimate equation (1) using

OLS. Because the number of violent incidents in other periods is measured as count data and exhibits

overdispersion, I estimate negative binomial regressions using maximum likelihood when using these

data.

IV estimates. I test whether proximity to Nyanza is correlated with the expansion of the Nyiginya

kingdom by estimating the following first stage relationship:

StatePresenced = γ0 + γ1Distanced +X′dγ2 +Q′

sγ3 + rp + νsdp (2)

where Distanced is either the straight line distance to Nyanza, or one of the cost distance measures

introduced in the previous section. StatePresenced and Distanced both vary at the precolonial district

level. I therefore report the first stage both at this level as well as the sector level.

I then use predicted state presence from the first stage in the second stage, which I estimate using two

stage least squares when using data on violence during the genocide, and using maximum likelihood

when using data on violence in other periods:

Ysdp = β0 + β1StatePresenced +X′dβ2 +Q′

sβ3 + rp + εsdp (3)

Ysdp is an outcome of interest, such as violence or mobilization during the genocide. β̂1 is the coeffi-

cient of interest. Because state presence varies at a higher level than my outcome data, I report clustered

standard errors for all specifications. I also report Conley (1999) standard errors that account for spatial

correlation. For most results, the Conley standard errors are smaller than the clustered standard errors.

21Distance from Kigali is important because militias, who were important killers in the genocide, travelled from Kigali (Rogall,2014).

22The census data vary at the commune level which is one administrative level above the sector but below the precolonial district.

12

For IV-maximum likelihood estimates, I report bootstrapped standard errors (Hilbe, 2011).

This instrumental variables strategy requires several assumptions to be met to be a valid approach for

causal inference. First, although proximity to Nyanza may be exogenous, Nyanza may have been cho-

sen because it was uniquely attractive for settlement. This would result in a violation of the exclusion

restriction to the extent that the factors that make Nyanza attractive correlate with violence. Second, the

correlation between proximity to Nyanza and state presence has to be sufficiently strong. To understand

the attractiveness of Nyanza, table 1 reports regressions of several characteristics of Rwandan villages on

the four instruments introduced above.

The first characteristic I consider is pre-Nyiginya population density. I use data on locations of arche-

ological finds indicating population settlement before the foundation of the Nyiginya kingdom from Pri-

oul and Sirven (1981). If prehistorical population density is correlated with proximity to Nyanza, the

exclusion restriction may be violated because population density may persist and correlate with εsdp. I

furthermore consider elevation and slope, as well as data on the suitability for growing banana, Rwanda’s

main staple crop, from the Food and Agricultural Organization. Each panel reports results for a different

instrument. All effect sizes are standardized (i.e. measured in standard deviations).

All instruments are uncorrelated with most pre-characteristics, which lends credence to the use of

proximity to Nyanza as a source of plausibly exogenous variation. Aside from the exclusion restriction

being met, the first stage correlation between the instruments and state presence needs to be sufficiently

strong. If this correlation is weak, the second stage results become difficult to interpret (Staiger and Stock,

1997). Table 2 reports estimations of equation 2 using state presence as the dependent variable and the

four measures of proximity to Nyanza as independent variables. Even-numbered columns report univari-

ate regressions using only the instruments as independent variables. Odd-numbered columns add travel

distances along the pre-genocide road network to Nyanza and Kigali, as well as straight line distance to

the national border and province fixed effects. The distance controls capture mechanical correlations be-

tween the instruments and state presence resulting from, for instance, accessibility for the Rwandan army,

the police or militias who were active killers in the genocide. These controls constitute a ‘baseline’ set of

controls and are included in all IV estimations in the remainder of this paper.

Columns (1) and (2) report results using the distance to Nyanza in kilometers as the variable of inter-

est. In line with the narrative in the introduction to this section, I find a negative correlation: Districts that

13

are further away from Nyanza were annexed to the Nyiginya kingdom later. This effect remains strong

and significant when introducing controls in column (2). I find similarly strong results using the cost

distance measures to Nyanza, in columns (3)-(4) and (5)-(6). Using cost distance to the nearest army as

the variable of interest does not produce robust results when introducing controls. The partial F-statistics

of the excluded instrument confirm these patterns, and show that, aside from cost distance to the nearest

army, the first stages are sufficiently strong.23

For the remainder of the analysis, I use the distance to Nyanza in kilometers as a baseline instrument.

This instrument is precisely measured and parsimonious, since it does not rely on specific assumptions re-

garding movement speed of soldiers nor on assumptions regarding the impediments posed by traversing

rugged terrain. In the appendix I report all IV results using the other instruments as well.24

2.3 Data

This section introduces the main variables used in the empirical analysis. Summary statistics for all vari-

ables used in this paper, as well as detailed data sources, are reported in the appendix.

Violence and mobilization in the genocide. Data on violence in the genocide is provided by the

National Service of Gacaca Jurisdictions court proceedings. A Gacaca court is a form of traditional local

justice, revolving around a village meeting in which individuals accused of genocidal crimes confess or

deny crimes and are sentenced or acquitted. Starting in 2001, over 8000 of these courts were instituted

in Rwanda to reduce the enormous backlog in the regular court procedures (for a detailed description of

the data as well as the Gacaca procedure, see Verpoorten (2011)). The data consist of three categories of

crimes:

1. Planners, organizers and supervisors of the genocide. This includes organizers at the local bureau-

cratic level and within political parties as well as the youth militias25

2. Murder, manslaughter, and non-lethal violence with the intention of killing

3. Property theft or damage, if not amicably settled

23In the appendix I report two robustness exercises. In table A2 I report estimates of equation (2) at the precolonial district level,the level at which state presence varies. The estimated effects are very similar. Following a recent contribution by Ashraf and Galor(2013) I then verify that the results in table 3 are not driven by the distance to Addis Abeba in table A3.

24I do not report results using the cost distance to the nearest army since the correlation in column (8) of table 2 is insignificant.25This category also includes individuals prosecuted for rape and sexual violence.

14

In my sample there are 816,325 prosecuted individuals; 76,572 in category 1, 431,265 in category 2 and

308,488 in category 3. In order to measure the intensity of genocidal violence and popular mobilization,

I construct three village level variables: The first is the sum of categories 1 and 2 divided by total popu-

lation from the 1991 census (Ministère du Plan, 1992). This measure captures total violence in a village.

Figure 4 contains a map of this variable. Second, I use the number of prosecuted individuals in category

1, normalized by the Hutu population, to understand the role Hutu organizers played. I finally use the

number of prosecuted individuals in category 2, normalized by the Hutu population to measure civilian

mobilization. It is important to note that the main results in this paper are robust to using alternative data

sources, such as a survey of genocide victims conducted by a Swiss NGO in 2004. I also show robustness

to normalizing by the number of Gacaca courts, rather than population, to account for the possibility that

longer state presence affects the presence of the Gacaca courts.

Violence before and after the genocide. I take pre- and post-genocide violence from the Uppsala

Conflict Data Program database (UCDP) (Sundberg et al., 2010; Sundberg and Melander, 2013), which

records the location, actors and death tolls of violent events since 1989. Its main sources are reports by

international news agencies (such as Reuters and the BBC) as well as reports from NGOs (such as Human

Rights Watch). In my sample there are 139 events in 1990-1993, 161 in 1994, and 147 in 1995-2000.26 The

UCDP data have been used extensively in economics, see e.g. Michalopoulos and Papaioannou (2016).

The UCDP database records three types of violence. First, state based armed conflict, which is defined

as “contested incompatibility that concerns government and/or territory where the use of armed force

between two parties, of which at least one is the government of a state, results in at least 25 battle-related

deaths in one calendar year”. Before the genocide, these conflicts occur mostly between the Rwandan

government and the invading RPF army. After the genocide, these conflicts occur between the Rwandan

government and Hutu rebels. Before 1994, 82 out of a 139 recorded conflicts are state based. After 1994,

59 out of 147 recorded events are state based. Second, one-sided violence, which is defined as the “use

of armed force by the government of a state or by a formally organized group against civilians which

results in at least 25 deaths in a year”. All remaining events are of this type. The final category is conflicts

"between two organized armed groups, neither of which is the government of a state, which results in at

least 25 battle-related deaths in a year."

26The UCDP data for 1994 are an underestimate. The total death toll for 1994 in this data is 155,217 which is far below the deathtoll estimated after the genocide as well as below the data introduced above (see Straus (2006) for a discussion). I therefore use theGacaca court data for 1994. Most results in the rest of this paper, however, are qualitatively similar when using the UCDP data,although I lose precision on some. These results are available upon request.

15

I use the total number of violent events between 1990-1993 and 1995-2000 as the main measure of vi-

olence in the years surrounding the genocide. This variable is mapped in figure 5. In addition, I use the

count of state based violent events and the count of one-sided violent events as variables of interest. In

the appendix, I report results using the count of violent events for different periods separately as well as

normalized by population.

State presence. I measure state presence as the number of years the Nyiginya state was present in a

district before colonization in 1897. Identifying when a district was annexed to the Nyiginya kingdom

requires reconstructing the administrative divisions of the precolonial kingdom and the identification of

the years of the annexation for each district. Full details of the procedure, as well as the raw data, are in

the appendix.

In addition to the data sources discussed in this paragraph, I use several historical and administrative

datasets. I introduce these in the text as they are used in an empirical estimation.

3 Results

This section reports the main results in this paper, the effect of state presence on violence before, during

and after the genocide. I first report results using OLS and negative binomial regressions, and then in-

troduce the IV estimations. I find a positive, significant and robust effect of state presence on violence in

the genocide. Using the UCDP data for 1990-1993 and 1995-2000, I find a negative effect of state presence

during these periods. Finally, the IV results are in agreement with the OLS results, providing evidence for

a causal interpretation of the effect of state presence on violence.

3.1 Results for violence and mobilization in the genocide

Table 3 reports estimates of equation (1) using measures of violence and mobilization from the Gacaca

records as dependent variables. Columns (1) and (2) use the fraction of population prosecuted for vi-

olence as the dependent variable, and columns (3) and (4) look at mobilization of the Hutu population

as organizers or as participants. The estimates in the first row show a consistent positive effect of state

presence on violence and mobilization. The mobilization effect is strongest in prosecutions for murder,

rather than organization, which is consistent with the case study literature which emphasizes the large

16

number of civilian perpetrators in the genocide (Straus, 2006). It is furthermore in line with the hypothe-

sis of this paper that the effect of the historical state is transmitted through rule following behavior. Note

that Conley (1999) standard errors are smaller in all regressions, suggesting that spatial correlation is not

very important for these results. Since the Tutsi were the targets of the genocide, a natural concern when

interpreting these effects is their spatial distribution. Column (2) includes the fraction Tutsi in the 1991

census as a covariate.27 Although the estimated effect of state presence falls somewhat, it remains positive

and highly significant. Overall, the results in this table suggest that state presence is positively correlated

with violence. These effects are driven by larger bottom up participation of the population, rather than

by more top down organizers.

Economic impact. The estimated coefficients in table 3 are not only significant, they are economically

meaningful as well. Consider the point estimate in column (1), 0.0307 (clustered s.e. 0.00953). A one

century increase in state presence (or about its mean) is associated with an increase in violence of about 3

percentage points. This increase is equal to about 40 percent of the mean incidence of violence. A different

way of benchmarking the effect size is to consider effect sizes moving through the interquartile range of

state presence. The interquartile range of state presence is 132 years. Taking the point estimate from

column (1) again, this range is associated with an 4 percent increase in violence, which is about 40 percent

of its interquartile range.

3.2 The impact of state presence on genocide intensity

The results in the previous section, although informative and precisely estimated, may be driven by omit-

ted variables. Below I consider several candidates for such variables. In this section I instead directly

estimate the causal effect of state presence using distance to Nyanza as an instrumental variable. The

main result of the paper is estimated in this section: An increase of one century in state presence doubles

violence in the genocide.

Table 4 reports estimates of equations (2) and (3), using distance to Nyanza as the excluded instrument.

Panel I reports two stage least squares estimates and panel II reports the corresponding first stages. All

regressions include travel distance to Nyanza and Kigali along the 1988 road network as well as distance

to the country border and province fixed effects. As before, columns (1) and (2) measure total violence

and columns (3) and (4) use mobilization outcomes. Panel II confirms the results in table 2: Distance to27Note that the fraction Tutsi may well be an outcome of state presence. The results in column (2) can therefore not be interpreted

as causal. I consider ethnic distribution as a channel below.

17

Nyanza is an informative instrument, and is strong enough (partial F-statistics are between 14 and 16)

to be confident about the second stage estimates. Column (1) presents the main result: A positive and

significant relationship between state presence and violence. As before, this effect is stronger for Hutu

civilian mobilization.

Economic impact. The estimated effect in column (1), 0.0807 (clustered s.e. 0.0270), implies that in-

creasing state presence by its mean increases violence by about 8 percent, or its mean. Similarly, increasing

state presence by its interquartile range increases violence by about its interquartile range. The estimated

effects of mobilization are similar. Note that these effects are large than the corresponding OLS estimates

in table 3. There are a number of reasons for why OLS estimates are typically smaller than IV estimates.

State presence could be measured with error, for instance. The IV estimates may pick up the Local Av-

erage Treatment Effect only for those sectors affected by the instrument, the OLS estimate may be biased

downwards due omitted variables or the exclusion restriction could be severely violated. Given the re-

sults in table 1, it is unlikely that the exclusion restriction is violated. It seems more likely that the OLS

estimates are biased downwards due to the presence of omitted variables. Section 3.4 below discusses

several candidates for such variables and finds that, while some covariates are correlated with violence,

the effect of state presence remains strong and significant.

3.3 Results for violence before and after the genocide

The previous section reports a strong and robust positive relationship between state presence and violence

during the genocide. This section estimates the same relationship in the years surrounding the genocide

when the Rwandan government was fighting rebels, rather than organizing mass killing. In this period,

it did not mobilize its population for violence. Rather, it urged the population not to support or join the

rebels and sought to unite the country against the invaders (Des Forges, 1999). The hypothesis of this

paper predicts that if the government does not mobilize for violence, violence should be lower in areas

where the precolonial state was established earlier. To understand if the effect of state presence is indeed

negative, this section reports results for violence in 1990-1993 and 1995-2000. I find a negative and robust

relationship between state presence and violence.

Table 5 reports estimates of negative binomial regressions of violence on state presence. In columns (1)

and (2), the dependent variable is a count variable of the number of violent events between 1990 and 2000,

excluding 1994. In columns (3) and (4) I break down this result by looking at state based and one-sided vi-

18

olence. As before, column (2) includes the fraction Tutsi as a covariate. The estimates in the first row show

a negative and significant effect of state presence on total violence and on both state based and one-sided

violence. In places with longer state presence, there are fewer violent events in the years surrounding the

genocide. There is no reason to hypothesize that the effects of state presence should be different for state

based or one-sided violence since both categories by and large involve the Rwandan government either

fighting rebels, or suppressing political opponents (for most of the sample period, Rwanda was a one

party state. Only between 1992 and the start of the genocide were multiple parties allowed (Guichaoua

and Degni-Segui, 2010). I explore results by the party affiliation of local bureaucrats below).

IV estimates. In order to understand whether these results admit a causal interpretation, I present

instrumental variable estimates in table 6. The columns present negative binomial regressions in panel

I and linear, first stage, regressions in panel II. The regressions in panel II are estimated using OLS. The

regressions in panel I are estimated using maximum likelihood and include the predicted residuals from

the first stage in a control function approach. This approach necessitates bootstrapping standard errors in

the second stage.28 The estimated effects in the first row show a negative effect of state presence on both

state based and one-sided violence and therefore provide a causal interpretation of the results in table 5.29

Economic impact. Because I estimate the relationship between state presence and violence in this

section using negative binomial regressions, I report marginal effects (at the mean). Consider the point

estimate in column (1) of table 6, -0.00371 (clustered s.e. 0.00116). This estimate implies that a one century

increase in state presence, at its mean, is associated with a 37.1% decrease in violence. Taking the point

estimates in columns (3) and (4), a one century increase in state presence results in a 14.2% reduction in

state based violence and a 18.6% reduction in one-sided violence.

The results in this section show that state presence has a large causal effect on contemporary violence.

This effect, however, is not constant. During the genocide I find a strong positive effect and in the years

immediately surrounding the genocide, this effect turns negative.30 These results are in line with the

hypothesis that the historical state affects rule following. The remainder of this paper is devoted to directly

28The regressions presented here are NB2 estimates. For background on this method see Hilbe (2011).29The control function approach admits an informal test of the endogeneity of state presence. The estimated effect of the predicted

residuals from the first stage is significantly different from zero which suggests that state presence is endogenous.30A natural next exercise is to exploit the time variation in violence data and estimate a continuous treatment differences in

differences model, using state presence as the continuous treatment and using a dummy for 1994 as the policy shift. Although thisappears attractive, creating the sample for this exercise involves combining the Gacaca data with the UCDP data. An alternativestrategy is to use the UCDP for 1994 as well, which has problems of its own (see above). Table A17 nevertheless implements thisidea. The interaction between a 1994 dummy and state presence is positive and significant and the magnitude is comparable to theresult in table 3.

19

testing this mechanism. Before moving on to these results, I briefly assess the robustness of the results in

this section.

3.4 Robustness

I undertake a series of robustness exercises, which I will only briefly discuss here. I assess two broad cat-

egories of robustness: Measurement and the impact of the colonial period. Overall, the main results are

highly robust to using alternative data sources, alternative ways of measuring the main outcomes and to

including a large number of historical controls. A sensitivity exercise shows that it is furthermore unlikely

that the main results are driven by unobservables. The appendix discusses each robustness check in more

detail.

Violence during the genocide. I first assess the robustness of the results for violence during the geno-

cide. I re-estimate equation (1) at the precolonial district level, the level at which state presence varies.

Second, I address the concern that the spatial distribution of violence during the genocide and in the sur-

rounding years reveals that most violence during the genocide happened in southern Rwanda whereas

most violence in surrounding years happened in northern Rwanda. This is potentially problematic since

the historical heartland of the state and the highest concentration of Tutsi are both in the South. I vary the

measurement of distance to Kigali and the national border by including polynomials in these distances

and interactions with province fixed effects, allowing for the effect of distance to the border to be different

for each province. I then show that normalizing by the number of Gacaca courts or using alternative data

sources for genocide violence results in qualitatively similar and statistically significant estimates. The

main objective of this exercise is to show that the main result is not driven by the idea that places with

longer state presence are better at prosecuting people, resulting in a mechanical correlation between state

presence and the number of prosecutions. Across all these exercises, the effect of state presence remains

positive and significant.

I then use equivalent household income in 1990 as a summary measure of differences in development

before the genocide. When using income as the dependent variable, I find no systematic relationship with

state presence. I then include income as a covariate. Because income data is only available for a subset

of villages, the estimated effects become noisier but the effect of state presence is stable. Ideally, I would

be able to measure income differences before the inception of the Nyiginya state. No such data exist but I

can use data from Prioul and Sirven (1981) that record archeological sites in Rwanda where remains have

20

been found that predate the Nyiginya kingdom. In addition to pre-existing differences, the German and

Belgian colonization may have had differential impacts across Rwanda. In order to understand the impact

of the colonial period, I use data from the Government of Belgium (1960) and the Parliament of Belgium

(1935) on the number of missionary stations in 1935 to account for missionary presence and the number of

cattle in 1960. In absence of income data, cattle is the best measures of prosperity in the colonial period. I

report results using these variables as covariates in appendix table A8. Including these variables reduces

the magnitude of the estimated effects from about 0.031 to about 0.0295.

This reduction in estimated effects can be further examined using a ‘selection on observables’ exercise

following Altonji et al. (2005) and Bellows and Miguel (2009). Let δ̂ be the reduction in the estimated

effect resulting from including observable covariates between two corresponding regressions in tables 3

and A8, and β̂ be the estimated effect in table 3. Then β̂

δ̂can be interpreted as a measure of how much

unobservables would have to reduce the estimated effects to explain them away, expressed in terms of

the reduction in the same effect resulting from including observables. This ratio, reported in every col-

umn in table A8, is around 25 for violence and mobilization of civilians, and equal to 12.45 for organizers

suggesting that, in order to explain away the effect, selection on unobservables would have to be at least

twelve times as strong as selection on unobservables.

Finally, I use the alternative cost distance instruments introduced above. The estimated effects of state

presence are virtually identical.

Violence before and after the genocide. In the appendix I show that the results in tables 5 and 6 are

robust to including the historical covariates introduced above: the number of archeological sites, mission-

ary stations and cattle. I also show that instead of using the number of violent events as the dependent

variable, I can normalize the count of events by 1991 population. I can also split violent events up by pre-

and post-genocide and by whether the Rwandan government in one of the parties involved in the vio-

lence (there is also violence of rebels against civilians, in which the government is not involved). I finally

verify that results estimated using two stage least squares rather than the IV-control function approach

give similar results. As before, using the alternative cost distance instruments gives very similar results.

21

3.5 Heterogeneous effects of state presence

The changing reduced form effect of states on violence during, before, and after support the hypothesis

tested in this paper because the motives of the government drastically changed during this period. In this

section I exploit the fact that changes in policy need to be communicated and legitimized (Glaeser, 2005).

Before the genocide the Rwandan government embared on a campaign that aimed at formally sanctioning

violence against Tutsi. The publication of the ‘Hutu Ten Commandments’ in the December 1990 edition

of Kangura, a government sponsored magazine, is seen as the start of this campaign (Des Forges, 1999).

The commandments forbade social interactions between Hutu and Tutsi and encouraged Hutu to ‘stop

having mercy’ on the Tutsi. Besides magazines, the government used radio to communicate its policy

changes (Yanagizawa-Drott, 2014). In this section I estimate heterogeneous effects of state presence by

radio ownership. I also test for heterogeneous effects by the political affiliation of the local burgomaster.

Yanagizawa-Drott (2014) has shown that the government radio station, the RTLM (‘radio television

libres des milles collines’), was instrumental in mobilizing the population. He finds that places that hap-

pened to have radio coverage due to varying geography see more violence. Should longer state presence

induce greater compliance with the demands of the government, I expect the effect of the radio to resonate

more with the local population in places with a deeper state history, since I hypothesize that individuals

in these districts are more likely to follow unenforced demands from the government. Although the radio

was not used to mobilize the population before and after the genocide, the government did use the radio

to instruct Rwandans to dissociate themselves from the rebels before the genocide and the radio remains

an important tool of communication. I therefore also use violence before the genocide as a dependent

variable. I test these ideas using data on radio ownership from the 1991 census.31

At a local level, the burgomasters, or mayors, of the 154 communes in Rwanda were instrumental

in mobilizing the population. Not all mayors, however, were members of the ruling MRND party. Fol-

lowing the introduction of multi-party government in Rwanda in 1992, a number of new political parties

attracted a significant following among local mayors.32 When the genocide started, many mayors recog-

31The data of Yanagizawa-Drott (2014) only provide coverage for the RTLM radio station which operated from July 8th, 1993 toJuly 31st, 1994. Using radio ownership allows me to study violence in the entire sample period.

32In 1993, 96 out of 145 communes had a mayor that was affiliated with the ruling MRND party. 24 mayors were affiliatedwith the Hutu MDR party, and 16 were affiliated with either the PL or the PSD, parties that were sympathetic to Tutsi interests. 6positions were vacant, one mayor was unaffiliated, and for two communes the affiliation of the mayor is not known (Guichaouaand Degni-Segui, 2010). The MRND, the Mouvement Revolutionnaire National pour le Developpement, was the party of the president,Habyarimana, and was a Hutu party. The radical arm of this party, the Akazu, has been identified as the organizing force behindthe genocide. The MDR, the Mouvement Democratique Republicain, was the main Hutu opposition party and consisted mainly ofsupporters of the previous president, Gregoire Kayibanda. Geographically, the MRND had its support among Hutu in North-WestRwanda, Habyarimana’s home region, and the MDR amoung Hutu living in central Rwanda. The PSD, Parti Social Democrate, and

22

nized that the orders that were given were coming from the MRND party, and some opposed the genocide

(although we do not know the exact number of mayors who opposed, see Straus (2006, p. 79) for a con-

crete example and Des Forges (1999, ch. 4) for detailed description of the interaction between local and

central government, and the role of mayors in Gikongoro and Butare provinces). These case studies sug-

gest that the effect of state presence may be less strong in communes with opposition mayors. I test this

hypothesis by investigating heterogeneous effects by party affiliation of the local mayor using data from

Andre Guichaoua (Guichaoua and Degni-Segui, 2010), coding an indicator variable equal to one if the

local mayor was member of an opposition party and estimating heterogeneous effects by splitting the

sample by this indicator variable.

Table 7 explores these hypotheses using the count of violent events in 1990-1993 and 1995-2000 as the

dependent variable in columns (1) and (2). In columns (3) and (4) I use the fraction of Hutu prosecuted

for murder as a measure of civilian mobilization. I report heterogeneous effects by splitting the sample at

the median of radio ownership (the median of radio ownership is equal to 33% of the 1991 population). In

line with the hypothesis tested in this paper, I find that that effect of state presence on mobilization during

the genocide is positive and significant in places with above median radio ownership positive but smaller

and insignificant in places with below median ownership (Chow tests show that the difference between

coefficients in these subsamples is statistically significant). The effect of state presence on the number

of violent events surrounding the genocide is negative in both subsamples but significantly lower in the

subsample with higher radio ownership rates. This evidence supports the idea that the effect of state

presence is concentrated in places where the government can reach individuals more easily. Columns (5)

and (6) use mobilization as their dependent variables as well, but split the sample by political affiliation

of the mayor. As hypothesized the effect of state presence is concentrated in areas with mayors that are

aligned with the government. Note that due to the small sample in column (5) the sub sample effects are

not statistically distinguishable.

4 Mechanisms

The previous sections have quantified the causal effect of state presence on genocide intensity and mobi-

lization of the Hutu population, as well as on state based and one-sided violence in the years surrounding

the genocide. In this section, I explore the mechanisms through which a longer history of centralized

government may affect violence. Although there are many potential mechanisms, I hypothesize in this

the PL, Parti Liberal were opposition parties based that were sympathetic to the Tutsi (PSD) or had many Tutsi as members (PL).

23

paper that the expansion of the Nyiginya kingdom had an effect on civil society. In this section I provide

evidence that state presence results in greater rule following in a fieldwork experiment. I corroborate this

result using survey evidence. I find that individuals that live today in areas with longer state presence

are less likely to find rule breaking acceptable, are less likely to challenge the government in community

meetings, and are less likely to abuse their wife at home.

4.1 Rule following as a mechanism

This section provides motivating case study evidence for obedience to political authority, which I’ll call

‘rule following’ in the empirical part of this section, as a channel of transmission of the effect of the

historical state.33 The idea that the extraordinary high degree of obedience of ordinary Rwandans to

the state was central to successful mobilization is widespread in case studies of the Rwandan genocide,

(Prunier, 1995; Hintjens, 1999; Des Forges, 1999; Newbury and Newbury, 2000; Straus, 2006).34 Obedience

is, in turn, linked to the state: Gérard Prunier writes: "Rwandese political tradition, going back to the

Banyiginya Kingdom through the German and Belgian colonial period, is one of systematic, centralised

and unconditional obedience to authority" (Prunier, 1995, p. 141). Despite the prevalence of these ideas,

there is little systematic evidence on their empirical validity. One exception is a survey of 209 incarcerated

genocide perpetrators, in which Scott Straus finds that 45% of perpetrators was motivated by ‘Intra-Hutu

coercion and/or obedience’. He describes (2006, p. 137):

“Here respondents said that they joined attacks because doing so was ‘the law’ (igeteko).

Others said that they went with murderous groups or killed because they were ‘obeying’ what

they had been told to do. Still other said that they participated because they had been ‘autho-

rized’ to kill Tutsis. In these accounts, respondents stressed that ‘the state’ or ‘the authorities’

had mandated participation for all able-bodied Hutu men. Killing was ‘the law’.”

Scholars have mostly investigated obedience in the context of the genocide. However, one telling

quote comes from a refugee in the Democratic Republic of the Congo. Although he or she was now re-

moved from the coercive power of the government that organized the genocide and likely forced him/her

33The authorities were usually local government officials, militia, or so-called intimidators. Before the genocide, the governmenthad drawn up of lists of people who could be used as leaders in the genocide. These ‘intimidators’, as the civilian perpetrators calledthem, were in charge of morale and order. Usually, intimidators would be shopkeepers, teachers and other local notables. When theintimidators could not motivate the civilians or when they encountered resistance, the authorities would call on the national policeor the youth militias, to break the resistance or lead the killings. Before the genocide, however, training programs were set up ineach district to train the youth to fight with machetes. One person in ten would be trained while living at home, under the pretenceof self defence. Coordination of these activities was left to local bureaucrats and the local police. See Rogall (2014) for empiricalevidence on the role of militias.

34The government managed to mobilize about between 14 and 17% of adult Hutu males Straus (2006). This rate of mobilizationis comparable to the highest rates in modern European wars (Onorato et al., 2014).

24

to become a refugee, he/she says: “.. we do not know what to think because our leaders are not around

just now. We are waiting for a new burgomaster to give us our orders.” (Prunier, 2008, p. 25). It is of

course hard to disentangle the effect of obedience from expected punishment for non-compliance, social

pressure and the prospect of material gain (this is the conclusion Straus reaches, see p. 140).35 Further-

more, arguments about obedience are always made comparing Rwanda to other countries. There is no

empirical evidence that provides guidance on whether there are systematic differences in rule following

behavior within Rwanda. To further isolate rule following that is driven by the historical state, I report

results from a fieldwork exercise which, as part of an experiment, holds environmental effects constant.

I find a greater propensity to follow rules in villages with a longer state presence. After presenting these

results, I present supporting evidence from several household surveys.

4.2 Mechanisms: experimental evidence

The motivation for individuals to engage in violence is determined by their environment, social situation

and personal factors, such as the expected gain from participation (Humphreys and Weinstein, 2008).

When the government orders individuals to participate, these same motivations plausibly apply: Some

perpetrators valued the loot that they would receive, others said that when the authorities tell you to

do something, you have to comply (see the case studies in Hatzfeld (2005)). To isolate the voluntary

component of this compliance from the coercive part by local intimidators, police or militias and the

prospect of material gain, I present results from a lab-in-the-field experiment which I implemented in fall

2014 in Southern Rwanda.

4.2.1 An experimental approach to measuring rule following

Experimentally measuring rule following faces several challenges. First, one has to isolate rule following

from other factors, such as trust, altruism and fear of punishment. The setup of the experiment is aimed at

keeping these factors balanced through random selection of participants, and I directly show balance on

many potential confounders. Second, one has to relate choices faced in the experiment to the government.

I do so by choosing taxation as a specific setting of the experiment. Finally, one has to link experimental

outcomes to differences in exposure to the historical state. I do so by sampling individuals on either side

of a historical boundary of the expanding Nyiginya kingdom.

The experiment proceeds as follows: I provide an endowment equal to 6000 Rwandan Francs (about

35See also the debate on compliance of ordinary Germans with Nazi demands (especially Browning (1993) and Goldhagen (1996)),and Stanley Milgram’s experiments (Milgram, 1963).

25

equal to wage compensation for one week of work). The participants are led into the experimental lab,

which is prepared with thirty stacks of two coins of 100 Rwandan Francs (RWF). Participants are also

provided with two unmarked envelopes which they are asked to associate in their head with the govern-

ment, and with themselves. The other player in the game is therefore the Rwandan government. Finally,

the participants are provided with an extra coin. They are asked to flip this coin, which determines the

split of each pile of two coins between the participant and the government. The rule that participants

need to follow is: If the coin comes up heads, the participants pays one coin, or 100RWF, into the envelope

associated with the government. The other coin is deposited in the envelope associated with the partici-

pant. If the coin comes up tails, both coins are to be deposited in the envelope for the participant. When

the participant has completed thirty flips for each pile, the game ends. The participant then keeps the

money that he/she allocated to him/herself and the envelope for the government is handed back to an

enumerator, who has left the lab for the duration of the game.36

There are three ways this game is associated with the government. First, participants are explicitly

asked to share between themselves and the government. Second, the contributions to the envelope asso-

ciated with the government are framed as a tax, so that when the coin comes up heads, the rule prescribes

paying taxes.37 Third, I transfer the money to the Rwandan government using their sovereign wealth

fund, the Agaciro fund, which takes credit card payments. Over half of the participants in my sample

report to have contributed to this fund themselves.38 This process is then repeated for every pile. The rep-

etition ensures that I know what, on average, should have been paid if the rule was perfectly followed.39

Importantly, the participants are unsupervised throughout this process. After observing the coin flip,

being aware of the rule that decides whether this money goes to the government or not, the participant

chooses to follow this rule or deviate. Because the experiment provides an anonymous environment with-

out any penalty for deviations, but with a financial incentive to deviate from the prescribed behavior, I

interpret the amount of money allocated to the government as a measure of rule following. I express the

amount allocated to the government as a fraction of the amount that should have been paid (on average)

if the fair coin was strictly followed in all thirty rounds of the game. In a recent contribution, Lowes et al.

36These instructions are based on the resource allocation game pioneered by Hruschka et al. (2014). A close variant of this gamewas recently used by Lowes et al. (2014) to measure rule following in the Democratic Republic of the Congo.

37Recently, Dwenger et al. (2016) provide evidence ‘duty-to-comply’ preferences are important for explaining voluntary tax com-pliance in Germany.

38In addition to explaining that the tax would go to the government, the exact wording of the instructions is: “I will provide RWF200 to you, as well as a coin. You will flip the coin. If you get heads, you have to pay RWF 100 in tax, which leaves RWF 100 for youto take home. If you get tails, you keep the RWF 200. Let me repeat that, if you get heads, you have to pay tax. If you get tails, youdo not have to pay tax.” Virtually everyone in Rwanda pays either income tax, VAT (18%) or one of a variety of other taxes.

39Since the coin dictates paying 50RWF with probability 0.5 in every round, this amount is 1500RWF or 25% of the endowment.

26

(2014) use a variant of the same game to understand the impact of the Kuba kingdom in the Democratic

Republic of the Congo on rule following behavior.

To understand how longer state presence affects rule following in this experiment, I compare partici-

pants close to the an abandoned boundary of the expanding Nyiginya kingdom to those just on the other

side of the boundary. It may be the case that individuals living in areas with a longer tradition of central-

ized rule are just more trusting, or feel more altruistic towards the government. I address these concerns

below.

4.2.2 Historical background and sample selection

Figure 6 highlights the boundary between two precolonial districts, Mayaga and Bugusera (see figure A1

in the appendix for a map with precolonial district names). Mayaga became part of the Nyiginya kingdom

around 1700 and Bugusera in 1799. The boundary was crossed by farmers in 1799 after the polity occu-

pying Bugusera, called Gisaka, was attacked by Burundi in the south. After Bugusera had been settled

by the farmers, the King formally annexed it. Jan Vansina notes about this annexation: “The .. conquest ..

illustrates the role of serendipity and private initiative in the process" (Vansina, 2004, p. 155). This history,

which is more fully detailed in Heldring (2016), suggests that villages on either side of the boundary are

similarly attractive for incorporation. I formally test balance across the boundary on elevation and agri-

cultural productivity below.

The study boundary coincides with a river, the Akanyaru, which is a small tributary river to Rwanda’s

main river, the Nyabarongo. Up to the administrative reorganization of 2002, the study boundary also

coincided with the border between three communes in Bugusera and four communes in Mayaga. After

the 2002 redrawing of boundaries, Bugusera became one district and Mayaga is subdivided into three

districts.40 These districts are mapped in figure 6 as well. Within these districts, I randomly sampled

villages to participate in a fieldwork exercise and within these villages, I randomly sampled participants.

Each village was visited twice, once to identify participants, and once for to carry out the fieldwork. In

total, there are 21 villages and 420 participants in the sample. An important identifying assumption for

the empirical part of this section is that these participants are randomly selected. I provide balance tests

below to shed light on the extent to which this assumption is met.

40This means that the effect that I observe may be driven by post-2002 district level confounders. This is however unlikely sincedistrict are very similar across Rwandan in terms of duties, and employ the same number of employees everywhere. The mainadvantage of focusing on a river is that I measure the historical boundary without error. The results are robust to including fixedeffects for the pre-2002 communes.

27

The game is implemented in an experimental ‘lab’ for the day, usually a small room in a local hotel

or bar, which are common in Rwandan villages. This location had to meet two criteria: 1) it can not be

owned by the government to assuage suspicions that the government was monitoring the experiment and

2) it should guarantee anonymity of choices the participant makes as part of the experiment.41

4.2.3 Balance

Table A18 in the appendix provides balance tests for several individual-level characteristics in panel I

and several village-level characteristics in panel II. Columns (1) and (2) report the effect of longer state

presence, now measured as an indicator variable equal to one for those individuals living on the side

of the Akanyaru river with longer state presence, on migrant status, accounting for differential sorting

across the study boundary. Columns (3) and (4) report results for two measures from a dictator game,

measuring altruism towards employees of the local and central government.42 Columns (5) and (6) assess

effects on trust and social capital. Columns (7) and (8) use survey questions on the expectation of gains

from taxation. Specifically, I asked individuals whether they thought taxation benefited themselves or the

country and coded two indicators equal to one if participants agreed with this statement. In panel II, I

report results using two dependent variables that measures underlying differences in geography between

places across the study boundary; elevation and terrain suitability for growing banana, Rwanda’s main

food crop. Pairs of columns restrict the sample to villages closer to the study boundary and columns (7)

and (8) report effects within the experimental sample.

Table A18 reports results for individual level characteristics in the experimental sample, as well as

results for sample village level characteristics. In panel I, all individual observables balance. In panel II,

elevation and banana suitability start balancing as I restrict to villages closer to the study boundary. These

results suggest that villages and participants look similar in terms of observables before the start of the

fieldwork.

4.2.4 Experimental results on rule following

This section uses OLS to estimate the relationship between state presence and rule following. I estimate

the following relationship:

41All results are robust to dropping those participants that indicated in post-experiment interviews that they felt observed.42After an initial survey and before any games were played, each participant was given a RWF1500 sitting fee. Right after being

handed the money, the participant was given the opportunity to share this money with an anonymous member of the local (sector)or the central government (a government official from the capital).

28

yiv = α+ βD ·Div + x′iv · βX + f(location) + εiv (4)

yiv is compliance with the demand to pay taxes for individual i in village v. Dv is an indicator variable

equal to one if village v is in Mayaga, rather than in Bugusera such that β̂D is the coefficient of interest,

the measured effect of being on the side of the river with a deeper history on rule following. f(location)

is a linear function of distance to the study boundary interacted with Dv to allow for a differential effects

of proximity to the Akanyaru. Xiv includes age, age squared and a gender dummy, as well as equivalent

daily income in Rwandan Francs (RWF) and education in years. Following Young (2016) I report p-values

from a permutation test to assess statistical significance.

Table 8 reports results. Column (1) reports the effect of state presence using just interacted distance

to the study boundary as a covariate. Column (2) includes demographic covariates, and column (3) adds

equivalent income and education as covariates to account for the effect these variables may have on rule

following. In all specifications, I find a positive and significant effect of state presence on rule following.

To take the estimated effect in column (1), a hundred years longer state presence (which is about the mean

of state presence), captured by the state presence dummy, leads to an increase in compliance with a taxa-

tion demand in the experimental fieldwork of 8.916% (permutation p-value 0.040). This effect is equal to

about 13% of the control group mean of compliance.43

Cultural norms as the main transmission mechanism. I naturally face the question what transmits

the effect of the precolonial state, which ceased to exist over a century ago, to the present. The litera-

ture has identified institutions – i.e. government enforced rules – and culture – norms that are internal

to the individual – as two important sources of persistence (see e.g. Acemoglu et al. (2001) and Nunn

and Wantchekon (2011)). In the context of the experimental results, I can use detailed information on the

history of the migration of participants to disentangle the effect of external – institutional – factors from

cultural norms. I do so by constructing a second treatment dummy which is based on the presence of the

state in the place where a respondent grew up, rather than where he or she lives today. For those individ-

uals who were born near to where they live today, the two measures will be equal. However, for those

who were born in a part of Rwanda where the state expanded earlier (later), and subsequently moved to

the side of the boundary where the state established later (earlier), the two measures will differ. Since in-

43Because participants observe the coin flip privately and follow the rule based on this coin flip, it may be the case that individualsare equally compliant with a rule but individuals in places with a longer state presence are less likely to lie about the result of thecoin flip. Since participants are not required to disclose the result of the coin flip to anyone but themselves, the only motivation tofalsely report coin flip is the anticipation of wanting to break the rule.

29

ternalized values and norms move with a participant while institutions are tied to the place where he/she

lives, comparing a location treatment (the treatment used before) to a birthplace treatment (the new treat-

ment based on place of birth) is informative about the transmission mechanism of the long-run effect of

the state. In particular, if the effect of the state works primarily through internalized norms of rule fol-

lowing, the effect of the birthplace treatment should overwhelm the effect of the location-based treatment.

Following this logic, 10.5% of my sample is assigned a new treatment status in the birthplace treat-

ment.44 Since the two treatment measures are the same for the rest of the sample, this subsample – those

who moved across the treatment boundary – is the source of identification. Since this is a smaller group

of people than the full sample, estimates are likely to become more noisy.45

Column (4) of table 8 includes both treatment measures in equation (4). As expected, both location

treatment and the birthplace treatment positively affect compliance, suggesting that the effect of the his-

torical state works both through the location of a respondent – through for instance enforcement of com-

pliance by peers – as well as through cultural norms. The birthplace treatment, however, is larger and

more precisely estimated. This result suggests that cultural norms are the main transmission mechanism

of the effect of the state.

Robustness. The balance tests reported in this section show that there are no systematic differences in

trust, social capital, altruism towards the government or beliefs about the usefulness of paying taxes. It is

also important to check for the presence and accountability of the modern government, social capital, and

the the intensity of genocidal violence across the study boundary. Aside from robustness, these tests sup-

plement the study of the transmission mechanism of the long-run effect of the Nyiginya kingdom in the

previous paragraph. In an accompanying paper I test for the effects of the presence on these variables and

find that the effect of the historical state is stable and strong throughout (Heldring, 2016). Importantly,

villages in the sample are balanced with respect to observable measures of modern local government.

The results in this section support the main hypothesis of this paper: Longer exposure state institutions

affects rule following behavior. The attractiveness of measuring rule following experimentally is that

the lab setting keeps the environment constant and random selection, the results in this section suggest,

44Overall migration in my sample is higher. Almost half of all participants have moved, but frequently they moved very locally,often to marry someone in a nearby village.

45Due to the smaller sample, it is difficult to assess the reasons for moving. One likely explanation however is marriage. Thosewho moved are significantly more likely to be married, and focus groups reveal frequent relocation for family reasons.

30

induced balance on individual observables. Nevertheless, results from a lab setting may not generalize.

The next section therefore turns to evidence from household surveys as a next test of the hypothesis of this

paper. In this section, too, I distinguish between norms of compliance, using data on beliefs for instance,

and government institutions as sources of persistence.

4.3 Mechanisms: Survey evidence

In this section, I present further evidence for the hypothesis that longer exposure to centralized govern-

ment is systematically linked to rule following today. Whereas the previous section directly measures

rule following, in this section I focus on beliefs about rule following and obedience from the World Values

Survey waves for 2007 and 2012. I also reports on civic engagement from the Rwanda Threshold survey

and evidence on domestic violence from the Demographic and Health Survey for 2014. I find effects con-

sistent with the hypothesis advanced in the previous section: individuals who live in parts of Rwanda

with longer state presence are less likely to find rule breaking acceptable, are less likely to speak up in

community meetings and are less likely to engage in domestic violence. These results should, however,

be interpreted with caution. Aside from being confounded by the genocide, these surveys do not provide

village level geographical coverage. I do not have enough spatial variation to report IV estimates or re-

duced form estimates. I therefore report OLS estimates.

I start with the World Values Survey (WVS). In 2007 and 2012, participants were asked a series of

questions about rule following. Specifically, questions V198-V202 asked participants to indicate on a scale

from 1 to 10 how justifiable it is to: ‘avoid a fare on public transport’, ‘steal property’, ‘cheat on taxes if

you have a chance’ and how justifiable it is for ‘someone to accept a bribe in the course of their duties’. A

score of 1 is ‘never justifiable’ and 10 ‘always justifiable’. Results for these questions are reported in Table

9A. Since the World Values Survey has limited geographical coverage, resulting in 22 clusters, I report

p-values from a permutation test in addition to heteroskedasticity robust standard errors.

I also consider question v138 from the 2012 wave which asks about essential aspects of democracy. It

asks: “Many things may be desirable, but not all of them are essential characteristics of democracy. Please

tell me for each of the following things how essential you think it is as a characteristic of democracy. Use

this scale where 1 means "not at all an essential characteristic of democracy" and 10 means it definitely

is "an essential characteristic of democracy". I consider the statement: “People obey their rulers”. I then

consider several variables from the Rwanda Threshold survey which asks question about participation in

31

civil society. I focus on community meetings in which the government announces new plans and citizens

have an opportunity to speak up to challenge the government. The survey asks whether individuals re-

cently went to a ‘budget meeting’ in which the budget of the local government was discussed or to any

other government meeting. I code an indicator equal to one if the respondent says that he/she went to

either meeting. The survey then asks whether other people spoke up in the meeting and whether the

participant spoke up. From the first question I code an indicator if the government was the only party

speaking at the meeting, and from the second question I code an indicator equal to one if the respondent

speaks up. The results from these questions are in table 9B.

I then consider several questions from the 2014 wave of the Demographic and Health Survey (DHS),

asking about violence against women. Before the genocide, women could not formally enter into business

relations. Since the genocide a series of laws gave women the right to own property, start business and

inherit property (Burnet, 2008). The 2003 Rwandan constitution formally codifies the equality of men and

women and affirms its commitment to this principle by ‘ensuring that women are granted at least thirty

per cent of posts in decision making organs’ (Republic of Rwanda, 2003, p. 3). Because men and women

are equal before the law, violence against women is a crime. I take absence of violence against women

as auxiliary evidence of rule following behavior. The DHS asks women whether they have experienced

verbal violence, physical violence or sexual violence from their husband. I code indicators for each. I also

assess how justified men find it to beat their wives using question V208 in the 2012 wave of the World

Values Survey. I report results in table 9C.

When estimating the effect of state presence on these outcomes, I match the respondents based on their

place of residence to precolonial districts, and control for age, age squared, a gender dummy as well as

survey year fixed effects. When reporting WVS results, I report heteroskedasticity robust standard errors

and p-values from a permutation test because the WVS has limited geographical coverage which results

in a small number of clusters. The other surveys have wider geographical coverage, and I report clustered

standard errors for these. To ensure comparability of results across regressions I report standardized co-

efficients.

Table 9A-C reports results. Consider table 9A. In places with longer state presence, respondents are

less likely to consider rule breaking justifiable across all categories. The results in table 9B suggest that

respondents in areas with longer state presence are more likely to think that obedience to authorities is a

central aspect of democracy. Although they are equally likely to attend community meetings, they are less

32

likely to speak up in these meetings and it is more likely that the government is the only party speaking.

Finally, the results in table 9C suggest that men find it marginally less acceptable to beat their wives,46

and violence against women is lower in areas with longer state presence. Although the estimated effects

in table 9 are not large, they are supportive of the hypothesis that the historical state in Rwanda has a

persistent effect on rule following today.47

4.4 The local state

Because the Nyiginya state did not innovate any top-down institutions, I do not expect state presence

to be associated with measures of the presence or capacity of the local government before the genocide.

The colonial and post-colonial governments did build government offices, promulgated laws and more

generally set up a institutionalized bureaucracy. It may be the case that these efforts are more successful

in areas with longer state presence. State capacity building and public investment during the colonial

period, however, were directed at unifying state capacity throughout the colony to facilitate indirect rule

(see also the history of suppression of rebellion in section 1). The Belgian minister of colonies remarked

in 1921:

"We will practice in Rwanda and Urundi a colonial protectorate policy. The basis of this pol-

icy is the maintenance of indigenous institutions....this is perfectly realizeable in these coun-

tries where the organization is ancient and remarkable and whose ruling class shows evident

political talents...Our administration will maintain royal authority and reinforce it."

Minister Franck, 1921, cited in Straus, 2006, p. 209.

Post-colonial public investment was instead aimed at the home region of the president (Verwimp, 2013).

In this section I nevertheless directly tests the effect of state presence on public good provision and fiscal

capacity.

Table 10 uses data from the Rwandan government on a number of public investments and other as-

pects of local government to test for a relationship between state presence and local government presence

or capacity. In column (1) I use the number of primary schools in 1980 as the dependent variable, in col-

umn 2 I use the number of hospital beds. Column (3) uses the number of post offices, and column (4)

46The p-value from the permutation test suggests that there is a marginally significant negative effect of state presence.47I also use data from the World Values Survey to verify that there are no systematic differences in several psychological outcomes

variables. Using survey questions, I check that risk taking, individualism, locus of control (the extent to which respondents thinkthey are in control of their lives), trust and confidence in the national government are not systematically correlated with statepresence. I describe the variable definitions in the appendix. The results in Table A19 suggest that areas of Rwanda that became partof the Nyiginya state earlier look similar in terms of psychological outcomes.

33

uses the number of government controlled markets. Column (5) uses total local government revenue and

column (6) uses total government expenditure. Row 1 reports IV estimates of the effect of state presence

on these variables. Consistent with the absence of formal institutions in the Nyiginya kingdom, I find a

consistently small and insignificant effect of state presence on these outcomes.48

Additionally, I look at a subset of these outcomes over time. Results are reported in table A21. I use

the number of hospitals in 1935, 1960 and the number of hospital beds in 1980. I also use the number

of missionary stations in 1924 and 1935 as well as schools in 1960 to measure the supply of education

(before the second world war, missionaries were the primary educators in Rwanda). As before, I estimate

precise zeroes on all these outcomes, but for the number of schools in 1960, for which I estimate a small

and marginally significant effect.

4.5 Other mechanisms

André and Platteau (1998) argue that Rwanda’s rapid growing population caused agricultural plots to be

subdivided up to the point that farming an average plot did not meet subsistence needs. The genocide,

they argue, was a struggle over land as a scarce resource, or a Malthusian ‘positive check’. I use data on

population growth between the 1978 and 1991 census as well as a measure of the extent to which Rwandan

communes could provide the calories to feed the local population in 1980 from the Rwandan government

to test this idea within my sample. Another theory of genocide intensity views the genocide as an outburst

of ethnic hatred between Hutu and Tutsi (see Prunier (1995) for a statement of this view). Hatred towards

other groups is hard to measure, and I will show results on the composition and fractionalization of the

local population.49 In places where Tutsi are more numerous, the Hutu population may, for instance, feel

more competition over resources or propaganda initiatives by the government may be more effective.

Finally, Yanagizawa-Drott (2014) shows how coverage of the RTLM radio station affected violence during

48In the appendix I test for an effect of state presence using a wider set of outcomes, as well as principal components of groups ofoutcomes. These results are reported in table A20. I first look at local public good provision. In column (1) I use the first principalcomponent of number of energy centers, energy transformers, water installations and water pumps in 1980. In column (2) I use thefirst principal component of the number of primary, secondary and secondary technical schools in 1980. In column (3) I use the firstprincipal component of the number of hospitals, maternity clinics, health clinics and other clinics in 1980. I then look at semi-publicinstitutions. In column (4) I use the first principal component of the number of social centers, reading centers, nutritional centers,and orphanages in 1980. In column (5) I use the first principal component of the number of post offices, grain silos, gas stationsand banks in 1980. In column (6) I use the first principal component of the number of agricultural cooperatives, markets, tradecenters and commercial centers in 1980. I finally look at public revenue and expenditure. In column (7) I use the first principalcomponent of total tax received, trade tax received and other tax received, all in per capita terms, in 1980. In column (8) I use thefirst principal component of total public expenditure, administrative expenditure, social expenditure and economic expenditure, inper capita terms, in 1980. The results suggest that there is no systematic relationship between state presence and the presence orcapacity of the local government: All standardized estimated effects are smaller than 0.02 and statistically insignificant.

49For Hutu, H and Tutsi, T , let πg be the in the fraction of the local population belonging to group g, g ∈ H,T . I disregardthe marginal Twa ethnic group. Fractionalization is then defined as πH(1 − πH) + πT (1 − πT ). This measure is maximized atπH = πT = 0.5 and can be interpreted as the probability that two randomly selected individuals will belong to different groups.

34

the genocide. I use his data on fraction of each village that had radio coverage to test if state presence

affects radio coverage. I assess the relationship between state presence and these hypotheses in appendix

table A22. Column (1) uses the percent increase in population between the 1978 and 1991 censuses as

the dependent variables. Column (2) uses as the dependent variable an indicator variable equal to one

if a sector was unable to meet the caloric needs of the local population in 1980. I find a negative and

insignificant effect of state presence on population growth and a negative and significant effect of state

presence on the inability to provide sufficient calories. These results find no support for the Malthusian

hypothesis. I find no systematic relationship between state presence and the presence of Tutsi and on the

fractionalization of Hutu-Tutsi, nor on radio coverage. Table A23 in the appendix assesses the robustness

of the correlation between state presence and violence and mobilization to including these variables as

covariates. The effect of state presence remains positive, stable and significant.50

5 Conclusion

A growing literature has tested the idea that more successful states historically affect prosperity even

when these states have disappeared or after several leadership transitions. In order to leverage these re-

sults for policy purposes, it is important to understand what drives these effects. This paper has tested the

hypothesis that longer state presence affects civil society, and in particular obedience to political authority,

in the context of violence in Rwanda.

The results in this paper show that a longer history of centralized rule impacts violence today. Sup-

porting the hypothesis of this paper, this effect is different depending on the rules that government agents

enforce. When the Rwandan government pursued mass mobilization and violence during the 1994 Rwan-

dan genocide, violence is higher in those parts of Rwanda where the state formed earlier. When the

Rwandan government pursued territorial control, violence is lower in these areas. Instrumental vari-

able estimates exploiting the proximity to Nyanza – an early capital – provide evidence that these results

have a causal interpretation. To further test the hypothesis that state presence affects rule following, I

report heterogeneous effects by radio ownership. In places where radio ownership was higher, and rules

were more easily communicated, the positive effect of state presence on violence during the genocide is

stronger than in places with low radio ownership. Before and after the genocide, I observe the opposite.

The effect of state presence is more strongly negative in areas where radio ownership is higher.

50Radio coverage is only available for two thirds of the village in my sample. To preserve power, I therefore test its effect byincluding it separately of the other covariates in column (3).

35

I then turn to directly testing the effect of state presence on rule following. I show, using data from

a lab in the field experiment, that in villages with longer state presence, individuals display a greater

propensity to comply with unenforced government demands. While this experiment directly measures

rule following, lab results may not generalize. I provide supporting evidence from household surveys.

I find that individuals that live in areas of Rwanda where the state formed earlier are less likely to find

rule breaking acceptable, are more likely to think that obedience to rulers is central to democracy, are less

likely to speak up in community meetings and are less likely to experience household violence.

The results in this paper speak more broadly to the persistence of historical institutions. While there

is now a large body of evidence showing that history matters (Nunn, 2009), there is less evidence on

whether persistent effects are constant over time. This paper has documented that a deep historical factor,

the length of exposure to centralized rule can have rapidly changing effects on economic development

because the channel of transmission interacts with public policy.

The results in this paper furthermore suggest that policies aimed at promoting state capacity and poli-

cies aimed at promoting civil society need to be cautious. Policies enacted by a strong state – such as the

genocide – interact with historically determined norms to potentially produce socially undesirable out-

comes. In anticipation of the abuse of state capacity, capacity building may be delayed (see also Besley and

Persson (2011) on this point). If civil society reform increases ‘civic mindedness’ or otherwise increases

rule following, individuals are easier to mobilize for socially unproductive ends as well. Policies aimed

at building ‘civic’ or ‘social’ capital should therefore be similarly cautious.

36

References

Acemoglu, D., C. García-Jimeno, and J. A. Robinson (2015). State capacity and economic development: A

network approach. The American Economic Review 105(8), 2364–2409.

Acemoglu, D. and M. O. Jackson (2016). Social Norms and the Enforcement of Laws. Journal of the European

Economic Association, forthcoming.

Acemoglu, D., S. Johnson, and J. A. Robinson (2001). The Colonial Origins of Comparative Development:

An Empirical Investigation. The American Economic Review 91(5), 1369–1401.

Acemoglu, D., T. Reed, and J. A. Robinson (2014). Chiefs: Economic Development and Elite Control of

Civil Society in Sierra Leone. Journal of Political Economy 122(2), 319–368.

Alesina, A. and N. Fuchs-Schündeln (2007). Good-Bye Lenin (or Not?): The Effect of Communism on

People’s Preferences. The American Economic Review 97(4), 1507–1528.

Alesina, A. and P. Giuliano (2015). Culture and institutions. Journal of Economic Literature 53(4), 898–944.

Altonji, J. G., T. E. Elder, and C. R. Taber (2005). Selection on Observed and Unobserved Variables: As-

sessing the Effectiveness of Catholic Schools. Journal of Political Economy 113(1), 151–184.

Anderson, B. (1983). Imagined communities: Reflections on the origin and spread of nationalism. London: Verso.

André, C. and J.-P. Platteau (1998). Land Relations under Unbearable Stress: Rwanda caught in the

Malthusian Trap. Journal of Economic Behavior & Organization 34(1), 1–47.

Ashraf, Q. and O. Galor (2013). The “Out of Africa” hypothesis, human genetic diversity, and comparative

economic development. The American Economic Review 103(1), 1–46.

Bates, R. H. (2008). When things fell apart. Cambridge: Cambridge University Press.

Becker, S. O., K. Boeckh, C. Hainz, and L. Woessmann (2015). The empire is dead, long live the empire!

Long-run persistence of trust and corruption in the bureaucracy. The Economic Journal, forthcoming.

Becker, S. O., L. Woessmann, et al. (2009). Was Weber Wrong? A Human Capital Theory of Protestant

Economic History. The Quarterly Journal of Economics 124(2), 531–596.

Bellows, J. and E. Miguel (2009). War and Local Collective Action in Sierra Leone. Journal of Public

Economics 93(11), 1144–1157.

Besley, T. and T. Persson (2009). Repression or Civil War? The American Economic Review 99(2), 292–297.

37

Besley, T. and T. Persson (2011). Pillars of Prosperity: The Political Economics of Development Clusters. Prince-

ton: Princeton University Press.

Bisin, A. and T. Verdier (2000). " Beyond The Melting Pot": Cultural Transmission, Marriage, And The

Evolution Of Ethnic And Religious Traits. Quarterly Journal of Economics, 955–988.

Blattman, C. and E. Miguel (2010). Civil War. Journal of Economic Literature 48(1), 3–57.

Blouin, A. (2014). Culture and Contracts: The Historical Legacy of Forced Labour. Working paper, Univer-

sity of Toronto.

Bockstette, V., A. Chanda, and L. Putterman (2002). States and Markets: The Advantage of an Early Start.

Journal of Economic Growth 7(4), 347–369.

Botte, R. (1985a). Rwanda and Burundi, 1889-1930: Chronology of a Slow Assassination, Part 1. Interna-

tional Journal of African Historical Studies, 53–91.

Botte, R. (1985b). Rwanda and Burundi, 1889-1930: Chronology of a Slow Assassination, Part 2. Interna-

tional Journal of African Historical Studies, 289–314.

Browning, C. R. (1993). Ordinary Men: Reserve Police Battalion 101 and the Final Solution in Poland. New

York: Harper Collins.

Burnet, J. E. (2008). Gender balance and the meanings of women in governance in post-genocide rwanda.

African Affairs 107(428), 361–386.

Cantoni, D., Y. Chen, D. Y. Yang, N. Yuchtman, and Y. J. Zhang (2014). Curriculum and Ideology. Journal

of Political Economy, forthcoming.

Collier, P. and D. Rohner (2008). Democracy, Development, and Conflict. Journal of the European Economic

Association 6(2-3), 531–540.

Conley, T. G. (1999). GMM Estimation with Cross Sectional Dependence. Journal of Econometrics 92(1),

1–45.

Dell, M. (2015). Trafficking networks and the Mexican drug war. The American Economic Review 105(6),

1738–1779.

Dell, M., N. Lane, and P. Querubin (2015). State Capacity, Local Governance, and Economic Development

in Vietnam. Working Paper, Harvard University.

38

Depetris-Chauvin, E. (2013). Local State History and Contemporary Conflict: Evidence from Sub-Saharan

Africa. Working paper, Brown University.

Des Forges, A. (1986). “The Drum Is Greater than the Shout”: The 1912 Rebellion in Northern Rwanda.

In Banditry, Rebellion and Social Protest in Africa, pp. 311–331. London: James Currey.

Des Forges, A. (1999). Leave none to tell the story. New York: Human Rights Watch.

Des Forges, A. (2011). Defeat is the Only Bad News: Rwanda Under Musinga, 1896–1931. Madison: University

of Wisconsin Press.

Di Tella, R., S. Galiani, and E. Schargrodsky (2007). The formation of beliefs: Evidence from the allocation

of land titles to squatters. The Quarterly Journal of Economics, 209–241.

Dijkstra, E. W. (1959). A note on two problems in connexion with graphs. Numerische mathematik 1(1),

269–271.

Dincecco, M. and G. Katz (2014). State Capacity and Long-run Economic Performance. The Economic

Journal. forthcoming.

Dittmar, J. E. (2011). Information technology and economic change: The impact of the printing press. The

Quarterly Journal of Economics 126(3), 1133–1172.

Dube, O. and J. F. Vargas (2013). Commodity Price Shocks and Civil Conflict: Evidence from Colombia.

The Review of Economic Studies 80(4), 1384–1421.

Dwenger, N., H. Kleven, I. Rasul, and J. Rincke (2016). Extrinsic and intrinsic motivations for tax compli-

ance: Evidence from a field experiment in Germany. American Economic Journal: Economic Policy 8(3),

202–232.

Easterly, W. and R. Levine (1997). Africa’s Growth Tragedy: Policies and Ethnic Divisions. The Quarterly

Journal of Economics 112(4), 1203–1250.

Esteban, J., L. Mayoral, and D. Ray (2012). Ethnicity and Conflict: An Empirical Study. The American

Economic Review 102(4), 1310–1342.

Gahima, G. (2013). Transitional justice in Rwanda: Accountability for atrocity. New York: Routledge.

Gennaioli, N. and I. Rainer (2007). The Modern Impact of Precolonial Centralization in Africa. Journal of

Economic Growth 12(3), 185–234.

39

Giuliano, P. and A. Spilimbergo (2014). Growing up in a Recession. The Review of Economic Studies 81(2),

787–817.

Glaeser, E. L. (2005). The political economy of hatred. The Quarterly Journal of Economics 120(1), 45–86.

Goldhagen, D. J. (1996). Hitler’s Willing Executioners: Ordinary Germans and the Holocaust. New York:

Alfred A. Knopf.

Government of Belgium (1960). Ruanda-Urundi: Economy 1. Brussels: Office de l’information et des

relations publiques pour le Congo Belge et le Ruanda-Urundi.

Guichaoua, A. and R. Degni-Segui (2010). Rwanda, de la Guerre au Génocide: Les Politiques Criminelles au

Rwanda (1990-1994). la Découverte.

Guiso, L., P. Sapienza, and L. Zingales (2010). Civic Capital as the Missing Link. In A. B. Jess Benhabib

and M. Jackson (Eds.), Handbook of Social Economics.

Guiso, L., P. Sapienza, and L. Zingales (2014). Long Term Persistence. Working paper, University of Chicago.

Hatzfeld, J. (2005). Machete Season: The Killers in Rwanda speak. Macmillan.

Heldring, L. (2016). The Persistence of State Capacity in Rwanda. Working paper, Oxford University.

Hilbe, J. M. (2011). Negative binomial regression. Cambridge: Cambridge University Press.

Hintjens, H. M. (1999). Explaining the 1994 Genocide in Rwanda. The Journal of Modern African Stud-

ies 37(2), 241–286.

Hruschka, D., C. Efferson, T. Jiang, A. Falletta-Cowden, S. Sigurdsson, R. McNamara, M. Sands, S. Munira,

E. Slingerland, and J. Henrich (2014). Impartial Institutions, Pathogen Stress and the Expanding Social

Network. Human Nature 25(4), 567–579.

Humphreys, M. and J. M. Weinstein (2008). Who fights? The determinants of participation in civil war.

American Journal of Political Science 52(2), 436–455.

Imhof, E. (1950). Gëlande und karte. Zürich: Rentsch.

Kagame, A. (1963). Les milices du Rwanda précolonial. Brussels: Académie royale des sciences coloniales.

Lemarchand, R. (1970). Rwanda and Burundi. New York: Praeger Publishers.

Louis, W. R. (1963). Ruanda-Urundi, 1884-1919. Oxford: Clarendon Press.

40

Lowes, S., N. Nunn, J. A. Robinson, and J. Weigel (2014). The Evolution of Culture and Institutions:

Evidence from the Kuba Kingdom. Working Paper, Harvard University.

Lugan, B. (1997). Histoire du Rwanda: De la préhistoire à nos jours. Bartillat.

Mamdani, M. (2002). When Victims Become Killers: Colonialism, Nativism, and the Genocide in Rwanda. Prince-

ton: Princeton University Press.

Mann, M. (2005). The Dark Side of democracy: Explaining Ethnic Cleansing. Cambridge: Cambridge Univer-

sity Press.

Michalopoulos, S. and E. Papaioannou (2013). Pre-Colonial Ethnic Institutions and Contemporary African

Development. Econometrica 81(1), 113–152.

Michalopoulos, S. and E. Papaioannou (2016). The Long-Run Effects of the Scramble for Africa. The

American Economic Review 106(7), 1802–1848.

Milgram, S. (1963). Behavioral study of obedience. The Journal of abnormal and social psychology 67(4),

371–378.

Ministère du Plan (1992). Recensement Général de la Population et de l’Habitat 1991. Kigali: Rwanda:

République Rwandaise.

Murdock, G. P. (1967). Ethnographic atlas: a summary. Ethnology 6(2), 109–236.

Newbury, C. (1988). The Cohesion of Oppression: Clientship and ethnicity in Rwanda 1860-1960. New York:

Columbia University Press.

Newbury, D. (1987). Bunyabungo: The Western Frontier in Rwanda. In The African Frontier: The Reproduc-

tion of Traditional African Societies, pp. 162–192. Bloomington: Indiana University Press.

Newbury, D. and C. Newbury (2000). Bringing the peasants back in: Agrarian themes in the construction

and corrosion of statist historiography in Rwanda. American Historical Review, 832–877.

Newbury, D. S. (1991). Kings and Clans: Ijwi Island and the Lake Kivu Rift, 1780-1840. Madison: University

of Wisconsin Press.

Nunn, N. (2009). The Importance of History for Economic Development. Annual Review of Economics 1(1),

65–92.

Nunn, N. (2012). Culture and the historical process. Economic History of Developing Regions 27(1), 108–126.

41

Nunn, N. and L. Wantchekon (2011). The Slave Trade and the Origins of Mistrust in Africa. The American

Economic Review 101, 3221–3252.

Onorato, M. G., K. Scheve, and D. Stasavage (2014). Technology and the Era of the Mass Army. The Journal

of Economic History 74(02), 449–481.

Özak, Ö. (2010). The Voyage of Homo-œconomicus: some economic measures of distance. Working paper,

Southern Methodist University.

Özak, Ö. (2013). Distance to the technological frontier and economic development. Working paper, Southern

Methodist University.

Parliament of Belgium (1935). Rapport sur du l’administration Belge du Ruanda-Urundi. Bruxelles.

Persson, T. and G. Tabellini (2009). Democratic Capital The Nexus of Political and Economic Change.

American Economic Journal: Macroeconomics, 88–126.

Prioul, C. and P. Sirven (1981). Atlas du Rwanda. Ministère de la coopération de la République Française

pour le compte de l’Université de Kigali-Rwanda.

Prunier, G. (1995). The Rwanda crisis: History of a Genocide. New York: Columbia University Press.

Prunier, G. (2008). Africa’s world war: Congo, the Rwandan genocide, and the making of a continental catastrophe.

Oxford: Oxford University Press.

Putnam, R. D., R. Leonardi, and R. Y. Nanetti (1994). Making democracy work: Civic traditions in modern

Italy. Princeton: Princeton University Press.

Republic of Rwanda (2003). The constitution of the republic of Rwanda.

Reyntjens, F. (2013). Political governance in post-genocide Rwanda. Cambridge: Cambridge University Press.

Rogall, T. (2014). The Escalation of Violence: Armed Groups and Civilian Perpetrators. Working paper,

IIES.

Satyanath, S., N. Voigtländer, and H.-J. Voth (2013). Bowling for Fascism: Social capital and the Rise of

the Nazi Party in Weimar Germany, 1919-33. National Bureau of Economic Research working paper w19201.

Staiger, D. and J. Stock (1997). Instrumental variables regression with weak instruments. Economet-

rica 65(3), 557–586.

Straus, S. (2006). The Order of Genocide: Race, Power, and War in Rwanda. Ithaca: Cornell University Press.

42

Sundberg, R., M. Lindgren, and A. Padskocimaite (2010). UCDP GED Codebook version 1.0-2011. Depart-

ment of Peace and Conflict Research, Uppsala University.

Sundberg, R. and E. Melander (2013). Introducing the UCDP georeferenced event dataset. Journal of Peace

Research 50(4), 523–532.

Tabellini, G. (2008). The Scope of Cooperation: Values and Incentives. The Quarterly Journal of Eco-

nomics 123(3), 905–950.

Tabellini, G. (2010). Culture and institutions: Economic development in the regions of Europe. Journal of

the European Economic Association 8(4), 677–716.

Tobler, W. (1993). Non-isotropic geographic modeling. Three presentations on geographic analysis and model-

ing. Santa Barbara: National Center for Geographic Information and Analysis, University of California.

Valentino, B. A. (2013). Final Solutions: Mass killing and genocide in the 20th century. Ithaca: Cornell Univer-

sity Press.

Vansina, J. (2004). Antecedents to modern Rwanda: the Nyiginya kingdom. Madison: University of Wisconsin

Press.

Verpoorten, M. (2011). The intensity of the Rwandan genocide: Measures from the Gacaca Records. Peace

Economics, Peace Science and Public Policy 18(1).

Verwimp, P. (2013). Peasants in Power, The Political Economy of Development and Genocide in Rwanda. Dor-

drecht: Springer.

Yanagizawa-Drott, D. (2014). Propaganda and Conflict: Evidence from the Rwandan Genocide. The

Quarterly Journal of Economics 129(4), 1947–1994.

Young, A. (2016). Channeling Fisher: Randomization Tests and the Statistical Insignificance of Seemingly

Significant Experimental Results. Working paper, London School of Economics and Political Science.

Young, H. P. (2001). Individual strategy and social structure: An evolutionary theory of institutions. Princeton:

Princeton University Press.

43

Figure 1: TIMELINE OF MAIN EVENTS IN RWANDA, 1990-2000, AND SCATTER PLOTS OF MAIN ESTIMATEDRELATIONSHIPS

Panel I: Timeline of main events in Rwanda, 1990-2000

1991 1992 1993 1994 1995 1996 1997 1998 1999 2000

RPF rebels

activ

e

Genocid

e

(1) (2) Hutureb

elsac

tive

Panel II: Binned scatterplots of the relationship between state presence and violence in different periods

(a) 1990-1993

0.00

0.05

0.10

0.15

0.20

0.25

50 100 150E( State presence | X)

E(

Vio

lenc

e, 1

990−

1993

| X

)

(b) April - July 1994

6

8

10

50 100 150E( State presence | X)

E(

Vio

lenc

e in

the

geno

cide

| X

)

(c) 1995-2000

0.00

0.05

0.10

0.15

0.20

0.25

50 100 150E( State presence | X)

E(

Vio

lenc

e, 1

995−

2000

| X

)

Notes: Panel I depicts a timeline of the events that are studied in this paper. On October 1st, 1990, the RPF rebels invade Rwandafrom Uganda, starting a period of rebel activity in Rwanda. In December 1990, the Hutu 10 commandments are published, whichstarted a campaign that placed Tutsi outside the law (indicated as point (1) on the timeline). Point (2) indicates the signing of theArusha accords, on August 4th, 1993. The genocide starts on April 8th, 1994 and ends mid-July 1994. From 1995 and 2000, theRwandan government faced rebel activity from Hutu rebels, many of whom were genocide perpetrators, and operated from withinRwanda as well as from the Democratic Republic of the Congo. Subfigures (a)-(c) are binned scatterplots of the relationship betweenviolence and the length of the presence of the centralized state across sectors in Rwanda (n=1449). The depicted relationships areconditioned on the distance to Kigali, Rwanda’s capital, the national border, and province fixed effects. Subfigure (a) contains ascatter of the count of violent events recorded in the UCDP violence data (Sundberg and Melander, 2013). I restrict the sample toevents that occurred before the Rwandan genocide, between January 1990 and December 1993. This relationship is estimated in tableA13 in the appendix. Subfigure (b) contains a scatter of the fraction of individuals prosecuted for genocide violence as a fractionof 1991 population in the Gacaca data (see text for details on this data source), against state presence in years. This relationship isestimated in table 3. These data only cover the Rwandan genocide. Subfigure (c) contains a scatter of the count of violent eventsrecorded in the UCDP violence data (Sundberg and Melander, 2013). I restrict the sample to events that occurred after the Rwandangenocide, between January 1995 and December 2000. This relationship is estimated in table A13 in the appendix.

44

Figure 2: MAP OF THE NYIGINYA EXPANSION

1700175018001850

Legend: Year of annexation

PlacesNyanza

Map of districts in precolonial Rwanda. A darker shade indicates districts that have a longer history of centralized rule as part ofthe Nyiginya kingdom. The raw data are reported in the Appendix. Nyanza is also indicated.

45

Figure 3: THE CONSTRUCTION OF COST DISTANCE INSTRUMENTS

(a) Step 1: Terrain elevation

1000200030004000

Legend: Elevation

PlacesNyanza

(b) Step 2: Terrain slope and distancedetermine travel time. Contours indi-cate six hour marches from Nyanza

Legend: Places

Nyanza

(c) Step 3: Travel time grid definesoptimal paths and minimum traveltime to precolonial districts

Legend: Places

Nyanza

Notes: Subfigures (a)-(c) describe different parts of the construction of the cost distance instruments. The full procedure is describedin the main text. This process starts with a grid that defines the cost of traversing a grid cell, such as elevation in (a). Using Tobler’shiking function (Tobler, 1993) I convert the elevation into a travel time grid. Subfigure (b) visualizes this step by plotting travel timecontour lines onto the elevation surface in subfigure (a). Using Dijkstra’s algorithm, I then find the least time-consuming path totravel from Nyanza to each precolonial district’s centroid (Dijkstra, 1959). These paths are visualized in subfigure (c). The HumanMobility Index proposed by by Özak (2010, 2013) proceeds similarly using a coarser grid. For details, see Özak (2010, 2013).

46

Figure 4: MAP OF VIOLENCE DURING THE GENOCIDE

Notes: Map of administrative sectors. Sectors are shaded by the number of individuals prosecuted for being an organizer of genocideviolence or for murder in the Gacaca courts divided by total population in the 1991 census.

47

Figure 5: MAP OF VIOLENT EVENTS BEFORE AND AFTER THE GENOCIDE

●

●●

●

●●

●

●

●

●

●

●●

●

●

●●

●

●

●

●

●

●

●

●

●●

●

●●

●

●

●●●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●●

●

●●

●

●

●

●

● ●

●●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●●

●

●

●

●

●

●

●

●

●

●

●

Legend: Type of violence

●

●

State based violenceOne−sided violence

Count●

●4812

16

Notes: Map of Rwanda. Violent events are indicated by dots, which are scaled by the number of violent events occuring in thatsector between 1990-1993 and 1995-2000. Colors indicate types of violence, state based or one-sided.

48

Figure 6: THE LOCATION OF PARTICIPANTS IN THE RULE FOLLOWING EXPERIMENT

Legend: Places

Kigali

(a)

Legend: Places

Kigali

(b)

Notes: Subfigures (a) and (b) map the villages that are in the fieldwork sample. In subfigure (a), the black line demarcates Rwanda’snational boundaries and the districts within which participating villages are randomly selected are depicted in grey. Villages aredepicted in black. The boundary separating the villages by their exposure to the Nyiginya kingdom is indicated in blue. Subfigure(b) depicts districts in grey, villages in black, the boundary separating the villages by their exposure to the Nyiginya kingdom isindicated in blue. Kigali, Rwanda’s capital city, is also indicated.

49

Table 1: CORRELATION OF PREDETERMINED CHARACTERISTICS AND DISTANCE TO NYANZA

Pre-Nyiginyasettlement Elevation Slope

BananaSuitability

(1) (2) (3) (4)

Panel I

Distance to Nyanza 0.425 0.0754 0.118 -0.112White s.e. (0.405) (0.163) (0.145) (0.230)

Conley s.e. [0.225] [0.124] [0.101] [0.117]

R2 0.171 0.547 0.364 0.455

Panel II

Cost distance to Nyanza 0.125 0.191 0.0522 -0.0835White s.e. (0.200) (0.197) (0.190) (0.171)

Conley s.e. [0.104] [0.151] [0.118] [0.100]

R2 0.122 0.560 0.361 0.454

Panel III

Cost distance to Nyanza - Özak 0.315 0.222 -0.0513 -0.113White s.e. (0.338) (0.177) (0.219) (0.218)

Conley s.e. [0.192] [0.104] [0.148] [0.095]

R2 0.144 0.559 0.361 0.455

Panel IV

Cost distance to nearest Nyiginya army location 0.122 -0.108 0.287* -0.0845White s.e. (0.0962) (0.137) (0.165) (0.144)

Conley s.e. [0.042] [0.079] [0.053] [0.084] ]

R2 0.126 0.553 0.416 0.456

Observations 50 50 50 50

Notes: All regressions are estimated using OLS. The unit of observation is a precolonial district. All point estimates are standardized.All regressions include fixed effects at the modern province level. Parentheses report heteroskedasticity robust (White) standarderrors. Square brackets report Conley (1999) standard errors, correcting for two-dimensional spatial correlation (assuming a 4decimal degree cutoff). * indicates significance at the 10 percent level, ** at the 5 percent level, *** at the 1 percent level, computedusing White standard errors.

50

Table 2: OLS ESTIMATES OF THE DETERMINANTS OF PRESENCE OF THE NYIGINYA STATE

Dependent variable:State Presence (years) Distance to Nyanza

Cost distance toNyanza (days)

Cost distance toNyanza - Özak (days)

Cost distance tonearest Nyiginya army

location (days)

(1) (2) (3) (4) (5) (6) (7) (8)

estimated coefficient -1.356*** -1.675*** -26.41*** -14.80** -108.3*** -134.7*** -12.06*** -5.718Clustered s.e. (0.194) (0.435) (4.234) (7.157) (15.52) (34.50) (3.798) (4.302)

Conley s.e. [0.062] [0.220] [1.731] [3.038] [7.085] [14.781] [2.739] [3.113]

Travel distance to Nyanza along 1988 road network N Y N Y N Y N NTravel distance to Kigali along 1988 road network N Y N Y N Y N YDistance to country border N Y N Y N Y N YProvince fixed effects N Y N Y N Y N Y

F-stat excluded instrument 48.81 14.83 38.90 4.28 48.70 15.25 10.08 1.77Number of clusters 50 50 50 50 50 50 50 50Observations 1553 1553 1553 1553 1553 1553 1553 1553R2 0.404 0.622 0.443 0.594 0.356 0.629 0.116 0.564

Notes: All regressions are estimated using OLS. The unit of observation is an administrative sector in 1994. State presence (years)is the number of years a sector was under centralized rule before colonization in 1897. Distance to Nyanza is the distance (inkilometers) between the centroid of a precolonial district and Nyanza. Cost distance to Nyanza (days) is the travel distance betweenthe centroid of a precolonial district and Nyanza for a person on foot, measured in hours. Cost distance to Nyanza - Özak (days) isthe travel distance between the centroid of a precolonial district and Nyanza for a person on foot, measured by the Human MobilityIndex proposed by Özak (2010, 2013). Cost distance to nearest Nyiginya army location is the travel distance (in days) between thecentroid of a precolonial district and the nearest Nyiginya army battalion, measured in hours. Travel distance is defined as thedistance (in kilometers) between the centroid of a precolonial district and an endpoint along the Rwandan road network in 1988.Regressions include fixed effects as indicated in the table. Parentheses report clustered standard errors at the level of the precolonialdistrict. Square brackets report Conley (1999) standard errors, correcting for two-dimensional spatial correlation (assuming a 4decimal degree cutoff).* indicates significance at the 10 percent level, ** at the 5 percent level, *** at the 1 percent level, computedusing standard errors clustered at the precolonial district level.

51

Table 3: DETERMINANTS OF VIOLENCE AND MOBILIZATION IN THE RWANDAN GENOCIDE

Dependent variable: Fraction of population prosecuted for: Fraction of Hutu prosecuted for:

Violence Violence Organization Murder(1) (2) (3) (4)

State Presence (Years) 0.0307*** 0.0287*** 0.00314 0.0287***Clustered s.e. (0.00953) (0.00763) (0.00243) (0.00972)

Conley s.e. [0.00508] [0.00692] [0.00148] [0.00364]

Fraction of population Tutsi 1991 0.250***(0.0611)

Travel distance to Kigali along 1988 road network Y Y Y YDistance to country border Y Y Y Y

Mean of the dependent variable 7.95 7.95 1.40 7.75

Number of clusters 50 50 50 50Observations 1449 1449 1449 1449R2 0.191 0.234 0.113 0.184

Notes: All regressions are estimated using OLS. The unit of observation is an administrative sector in 1994. Fraction of populationprosecuted for violence is the number of individuals prosecuted for being an organizer of genocide violence or for murder in theGacaca courts divided by total population in the 1991 census. Fraction of Hutu population prosecuted for organization (murder)is the number of individuals prosecuted for organizing genocide violence (murder) in the Gacaca courts divided by total Hutupopulation in the 1991 census. State presence (years) is the number of years a sector was under centralized rule before colonizationin 1897. Fraction of population Tutsi 1991 is the total number of Tutsi divided by total population in the 1991 census. Travel distanceis defined as the distance (in kilometers) between the centroid of a precolonial district and an endpoint along the Rwandan roadnetwork in 1988. All regressions include fixed effects at the modern province level. Parentheses report clustered standard errorsat the level of the precolonial district. Square brackets report Conley (1999) standard errors, correcting for two-dimensional spatialcorrelation (assuming a 4 decimal degree cutoff). * indicates significance at the 10 percent level, ** at the 5 percent level, *** at the 1percent level, computed using standard errors clustered at the precolonial district level.

52

Table 4: IV ESTIMATES OF THE EFFECT OF STATE PRESENCE ON VIOLENCE AND MOBILIZATION IN THERWANDAN GENOCIDE

Dependent variable: Fraction of population prosecuted for: Fraction of Hutu prosecuted for:

Violence Violence Organization Murder(1) (2) (3) (4)

Panel I: Second stage estimates

State Presence (Years) 0.0807*** 0.0664*** 0.00957* 0.0799***(0.0270) (0.0211) (0.00527) (0.0268)

Mean of the dependent variable 7.95 7.95 1.40 7.75

Panel II: First stage estimates

Distance to Nyanza -1.675*** -1.720*** -1.675*** -1.675***(0.435) (0.430) (0.435) (0.435)

F-stat of excluded instrument 14.59 15.98 14.59 14.59R2 0.622 0.623 0.622 0.622

Fraction of population Tutsi 1991 N Y N NTravel distance to Nyanza along 1988 road Y Y Y YTravel distance to Kigali along 1988 road Y Y Y YDistance to country border Y Y Y Y

Number of clusters 50 50 50 50Observations 1449 1449 1449 1449

Notes: All regressions in Panel I are estimated using two stage least squares using distance to Nyanza as the instrument. Allregressions in panel II are estimated using OLS. The unit of observation is an administrative sector in 1994. Fraction of populationprosecuted for violence is the number of individuals prosecuted for being an organizer of genocide violence or for murder in theGacaca courts divided by total population in the 1991 census. Fraction of Hutu population prosecuted for organization (murder)is the number of individuals prosecuted for organizing genocide violence (murder) in the Gacaca courts divided by total Hutupopulation in the 1991 census. State presence (years) is the number of years a sector was under centralized rule before colonizationin 1897. Distance to Nyanza is the distance (in kilometers) between the centroid of a precolonial district and Nyanza. Fraction ofpopulation Tutsi 1991 is the total number of Tutsi divided by total population in the 1991 census. Travel distance is defined as thedistance (in kilometers) between the centroid of a precolonial district and an endpoint along the Rwandan road network in 1988.All regressions include fixed effects at the modern province level. Parentheses report clustered standard errors at the level of theprecolonial district. * indicates significance at the 10 percent level, ** at the 5 percent level, *** at the 1 percent level.

53

Table 5: DETERMINANTS OF VIOLENCE BEFORE AND AFTER THE RWANDAN GENOCIDE

Dependent variable:

Number of violentevents

1990-1993, 1995-2000(count)

State based violence(against rebels)

1990-1993, 1995-2000(count)

One-sided violence(against civilians)

1990-1993, 1995-2000(count)

(1) (2) (3) (4)

State Presence (Years) -0.00152*** -0.00132*** -0.00059*** -0.00046*(0.00045) (0.00049) (0.00018) (0.00026)

Fraction of population Tutsi 1991 -0.00362(0.00339)

Travel distance to Kigali along 1988 road network Y Y Y YDistance to country border Y Y Y YPopulation 1991 Y Y Y Y

Mean of the dependent variable 0.18 0.18 0.09 0.09

Log-likelihood -507.08 -505.79 -309.33 -354.92Number of clusters 50 50 50 50Observations 1553 1553 1553 1553

Notes: All regressions are negative binomial regressions, estimated using maximum likelihood. Estimated coefficients are marginaleffects (at the mean). The unit of observation is an administrative sector in 1994. Number of violent events 1990-1993, 1995-2000(count) is a count variable measuring the number of violent events resulting in more than twenty five casualties that are classifiedas either state based violence or one-sided violence in the UCDP dataset occurring between January 1990 and December 1993, andbetween January 1995 and December 2000. State based violence (against rebels) 1990-1993, 1995-2000 (count) is a count variablemeasuring the number of violent events resulting in more than twenty five casualties that are classified as state based violence inthe UCDP dataset occurring between January 1990 and December 1993, and between January 1995 and December 2000. One-sidedviolence (against civilians) 1990-1993, 1995-2000 (count) is a count variable measuring the number of violent events resulting inmore than twenty five casualties that are classified as one-sided violence in the UCDP dataset occurring between January 1990 andDecember 1993, and between January 1995 and December 2000. State presence (years) is the number of years a sector was undercentralized rule before colonization in 1897. Fraction of population Tutsi 1991 is the total number of Tutsi divided by total populationin the 1991 census. Travel distance is defined as the distance (in kilometers) between the centroid of a precolonial district and anendpoint along the Rwandan road network in 1988. 1991 Population is total population in the 1991 census. All regressions includefixed effects at the modern province level. Parentheses report clustered standard errors at the level of the precolonial district. *indicates significance at the 10 percent level, ** at the 5 percent level, *** at the 1 percent level.

54

Table 6: IV ESTIMATES OF THE EFFECT OF STATE PRESENCE ON VIOLENCE BEFORE AND AFTER THERWANDAN GENOCIDE

Dependent variable:

Number of violentevents

1990-1993, 1995-2000(count)

State based violence(against rebels)

1990-1993, 1995-2000(count)

One-sided violence(against civilians)

1990-1993, 1995-2000(count)

(1) (2) (3) (4)

Panel I: Second stage estimates

State Presence (Years) -0.00371*** -0.00357*** -0.00142** -0.00186***(0.00116) (0.00112) (0.000456) (0.000710)

Mean of the dependent variable 0.18 0.18 0.09 0.09

Panel II: First stage estimates

Distance to Nyanza -1.767*** -1.822*** -1.767*** -1.767***(0.463) (0.457) (0.463) (0.463)

R2 0.622 0.623 0.622 0.622

Fraction of population Tutsi 1991 N Y N YTravel distance to Nyanza along 1988 road Y Y Y YTravel distance to Kigali along 1988 road Y Y Y YDistance to country border Y Y Y YPopulation 1991 Y Y Y Y

Number of clusters 50 50 50 50Observations 1553 1553 1553 1553

Notes: All regressions in Panel I are negative binomial regressions, reporting marginal effects (at the mean) estimated using a controlfunction - maximum likelihood approach with distance to Nyanza as the instrument. All regressions in panel II are estimatedusing OLS. The unit of observation is an administrative sector in 1994. Number of violent events 1990-1993, 1995-2000 (count) isa count variable measuring the number of violent events resulting in more than twenty five casualties that are classified as eitherstate based violence or one-sided violence in the UCDP dataset occurring between January 1990 and December 1993, and betweenJanuary 1995 and December 2000. State based violence (against rebels) 1990-1993, 1995-2000 (count) is a count variable measuringthe number of violent events resulting in more than twenty five casualties that are classified as state based violence in the UCDPdataset occurring between January 1990 and December 1993, and between January 1995 and December 2000. One-sided violence(against civilians) 1990-1993, 1995-2000 (count) is a count variable measuring the number of violent events resulting in more thantwenty five casualties that are classified as one-sided violence in the UCDP dataset occurring between January 1990 and December1993, and between January 1995 and December 2000. State presence (years) is the number of years a sector was under centralizedrule before colonization in 1897. Distance to Nyanza is the distance (in kilometers) between the centroid of a precolonial districtand Nyanza. Fraction of population Tutsi 1991 is the total number of Tutsi divided by total population in the 1991 census. Traveldistance is defined as the distance (in kilometers) between the centroid of a precolonial district and an endpoint along the Rwandanroad network in 1988. Population 1991 is total population in the 1991 census. All regressions include fixed effects at the modernprovince level. In panel I, parentheses report bootstrapped standard errors. In panel II, parentheses report clustered standard errorsat the level of the precolonial district. * indicates significance at the 10 percent level, ** at the 5 percent level, *** at the 1 percent level.

55

Table 7: ESTIMATES OF HETEROGENEOUS TREATMENT EFFECTS OF STATE PRESENCE ON VIOLENCE ANDMOBILIZATION

Dependent variable:

Number of violent events1990-1993, 1995-2000

(count) Fraction of Hutu population prosecuted for murder

Below medianradio ownership

1991

Above medianradio ownership

1991

Below medianradio ownership

1991

Above medianradio ownership

1991Opposition

mayorGovernment

mayor

(1) (2) (3) (4) (5) (6)

State Presence (Years)) -0.0262 -0.158*** 0.137 0.309*** 0.160 0.234**(0.0466) (0.0498) (0.106) (0.0500) (0.137) (0.0985)

Travel distance to Kigali along 1988 road Y Y Y Y Y YDistance to country border Y Y Y Y Y Y

Chow test of coeffecient equality (p-value) 0.0275 0.0275 0.0832 0.0832 0.636 0.636Number of clusters 50 50 50 50 50 50Observations 769 784 709 740 221 1146R2 0.016 0.027 0.239 0.196 0.029 0.176

Notes: All regressions are estimated using OLS. The unit of observation is an administrative sector in 1994. All point estimates arestandardized. Number of violent events 1990-1993, 1995-2000 (count) is a count variable measuring the number of violent eventsresulting in more than twenty five casualties that are classified as either state based violence or one-sided violence in the UCDPdataset occurring between January 1990 and December 1993, and between January 1995 and December 2000. Fraction of Hutupopulation prosecuted for murder is the number of individuals prosecuted for murder in the Gacaca courts divided by total Hutupopulation in the 1991 census. Travel distance is defined as the distance (in kilometers) between the centroid of a precolonial districtand an endpoint along the Rwandan road network in 1988. All regressions include fixed effects at the modern province level.Parentheses report clustered standard errors at the level of the precolonial district. * indicates significance at the 10 percent level, **at the 5 percent level, *** at the 1 percent level.

56

Table 8: THE EFFECT OF STATE PRESENCE ON RULE FOLLOWING

Dependent variable: Compliance with tax demand (% of base)(1) (2) (3) (4)

State Presence (100 years) - Place of residence 8.916** 8.663** 8.687** 4.586Permutation test p-value 0.040 0.041 0.041 0.368

State Presence (100 years) - Place of birth 6.862*Permutation test p-value 0.0896

Demographic controls N Y Y YIncome N N Y NEducation N N Y NDistance to boundary * State Presence Y Y Y Y

Control group mean 68 68 68 68

Clusters 21 21 21 21Observations 415 413 413 413R2 0.010 0.027 0.027 0.031

Notes: All regressions are estimated using OLS. The unit of observation is an individual respondent.Compliance with tax demand (% of base) is the tax paid by an individual in the random taxation game,divided by the theoretical tax base, 1500 Rwandan Francs. State Presence (100 years) is an indicator equalto one if a respondent lived to the East of the Akanyaru river in October 2014. Demographic controlsinclude age, age squared and a dummy for gender. Income is equivalent daily income in Rwandan francs.Education is education in years. P-values of the permutation test give the fraction of t-statistics greaterthan the t-statistic of the estimated effect in the reported regression across 5000 permutations of the statepresence dummy. * indicates significance at the 10 percent level, ** at the 5 percent level, *** at the 1percent level.

57

Table 9: OBERVATIONAL EVIDENCE ON RULE FOLLOWING

TABLE 9A: ATTITUDES

Is it justifiable to:

Dependent variable:Avoid a fare onpublic transport Steal property Cheat on taxes

For someone toaccept a bribe

(1) (2) (3) (4)

State Presence (Years) -0.0547** -0.0534* -0.0601** -0.0481*(0.0272) (0.0277) (0.0282) (0.0269)

Permutation test p-value 0.045 0.052 0.033 0.080Clusters 22 22 22 22Observations 2593 2589 2593 2582R2 0.017 0.058 0.011 0.059

Notes: All regressions are estimated using OLS. The unit of observation is the individual. All point estimates are standardized.Dependent variables are several questions from the World Values survey waves of 2007 and 2012 that ask whether. State presence(years) is the number of years a sector was under centralized rule before colonization in 1897. All regressions include demographiccontrols and survey year fixed effects. Demographic controls include age, age squared and a gender dummy. Survey year fixedeffects are indicator variables for whether an individual respondent was interviewed as part of the 2007 or the 2012 World Valuessurvey wave. P-values of the permutation test give the fraction of t-statistics greater than the t-statistic of the estimated effect in thereported regression across 5000 permutations of the state presence dummy. Parentheses report heteroskedasticity robust standarderrors. * indicates significance at the 10 percent level, ** at the 5 percent level, *** at the 1 percent level.

TABLE 9B: OBEDIENCE AND PROTEST AGAINST GOVERNMENT

Dependent variable:

Obedience to rulersimportant for

democracy

Participation incommunity

meetings

Governmentonly speaking at

meetings

Respondent speaksup in community

meeting

(1) (2) (3) (4)

State Presence (Years) 0.103*** 0.0187 0.0794** -0.0540**(0.0398) (0.0350) (0.0387) (0.0202)

Permutation test p-value 0.010Clusters 22 50 50 50Observations 1203 4550 3328 3638R2 0.015 0.020 0.013 0.033

Notes: All regressions are estimated using OLS. The unit of observation is the individual. All point estimates are standardized. Obe-dience to rules important for democracy is a question asking about how central to democracy obedience to rulers is. Participationin community meetings in an indicator variable equal to one if respondent indicated participation in community meetings. Govern-ment only speaking is an indicator equal to one if the government is the only party speaking at these meetings. Respondent speaksup in community meetings is an indicator variable equal to one if respondent indicated speaking up in these community meetings.State presence (years) is the number of years a sector was under centralized rule before colonization in 1897. All regressions includedemographic controls. Demographic controls include age, age squared and a gender dummy. P-values of the permutation test givethe fraction of t-statistics greater than the t-statistic of the estimated effect in the reported regression across 5000 permutations of thestate presence dummy. Parentheses report heteroskedasticity robust standard errors in column (1) and clustered standard errors atthe level of the precolonial district in columns (2)-(4). * indicates significance at the 10 percent level, ** at the 5 percent level, *** atthe 1 percent level.

58

TABLE 9C: VIOLENCE AGAINST WOMEN

Dependent variable:Justifiable to beat

wife Verbal violence Physical violence Sexual violence

(1) (2) (3) (4)

State Presence (Years) -0.0472 -0.0664* -0.0652** -0.0669**(0.0349) (0.0333) (0.0318) (0.0271)

Permutation test p-value 0.062Clusters 22 50 50 50Observations 1280 1859 1858 1859R2 0.16 0.009 0.010 0.006

Notes: All regressions are estimated using OLS. The unit of observation is the individual. All point estimates are standardized.Justifiable to beat wife is question that asks how justifiable it is to beat your wife. Verbal violence is an indicator equal to one if arespondent experienced verbal abuse from his/her partner. Physical violence is a variable ranging between zero and two which isthe sum of two indicators, one equal to one if a respondent experienced pushing, shaking, throwing of objects, slapping, punching,hitting with objects, arm twisting, hair pulling and one equal to one if a respondent experienced kicking, dragging, strangling,burning or threatening with a weapon from her partner. Sexual violence is an indicator equal to one if a respondent experiencedforced sex or sexual acts from her partner. State presence (years) is the number of years a sector was under centralized rule beforecolonization in 1897. All regressions include demographic controls. Demographic controls include age, age squared and a genderdummy. P-values of the permutation test give the fraction of t-statistics greater than the t-statistic of the estimated effect using statepresence across 5000 permutations of state presence. Parentheses report heteroskedasticity robust standard errors in column (1) andclustered standard errors at the level of the precolonial district in columns (2)-(4) * indicates significance at the 10 percent level, ** atthe 5 percent level, *** at the 1 percent level.

Table 10: ESTIMATES OF THE EFFECT OF STATE PRESENCE ON LOCAL GOVERNMENT

Dependent variable:Primary

schools 1980Hospital beds

1980Post offices

1980 Markets 1980

Governmentrevenue (Francsper capita) 1980

Governmentexpenditure(Francs percapita) 1980

(1) (2) (3) (4) (5) (6)

State Presence (Years)) 0.00124 -0.173 -0.00231 0.00219 0.101 2.720(0.0200) (0.162) (0.00282) (0.00527) (5.815) (5.880)

Travel distance to Nyanza along 1988 road Y Y Y Y Y YTravel distance to Kigali along 1988 road Y Y Y Y Y YDistance to country border Y Y Y Y Y Y

Mean of the dependent variable 11.3 21.2 0.45 1.33 4014 3923

Number of clusters 50 50 50 50 50 50Observations 1553 1553 1553 1553 1297 1468R2 0.187 0.053 0.036 0.076 0.270 0.229

Notes: All regressions in Panel I are estimated using two stage least squares using distance to Nyanza as the instrument. The unitof observation is an administrative sector in 1994. Primary schools 1980 is the number of primary schools in 1980. Hospital beds1980 is the number of hospitals beds available in local hospitals in 1980. Post offices is the number of post offices in 1980. Marketsis the number of government-sanctioned local markets in 1980. Government revenues (Francs per capita) 1980 is total governmentrevenue from taxes divided by 1978 population in 1980 Rwandan Francs. Government expenditure (Francs per capita) 1980 is totalgovernment expenditure from taxes divided by 1978 population in 1980 Rwandan Francs. State presence (years) is the numberof years a sector was under centralized rule before colonization in 1897. Travel distance is defined as the distance (in kilometers)between the centroid of a precolonial district and an endpoint along the Rwandan road network in 1988. All regressions includefixed effects at the modern province level. Parentheses report clustered standard errors at the level of the precolonial district. *indicates significance at the 10 percent level, ** at the 5 percent level, *** at the 1 percent level.

59

Appendix for:

Violence and the State:

Evidence from Rwanda’s ‘Decade of Atrocities’

Leander Heldring∗

December 2016

Contents

1 Further results 2

2 Reconstructing the expansion of the Nyiginya kingdom 27

3 Data Sources 33

∗e-mail: lheldring@iq.harvard.edu; website: http://www.leanderheldring.com/.

1

1 Further results

Table A1 reports summary statistics for all variables used in the paper and in this appendix.

OLS estimates - first stage. Table A2 reports estimates of the first stage relationship reported in table

2, using precolonial districts as the unit of observation, rather than administrative sectors at the time of

the genocide. The estimated effects in this table are virtually identical to those in table 2.

Table A3 repeats the regressions in table A2, adding distance to Addis Abeba (in kilometers) as a co-

variate, accounting for the effect distance to Addis Abeba has on comparative development (Ashraf and

Galor, 2013). Estimated effects are slightly larger, as are standard errors, but the relationship between the

measures of distance to Nyanza and state presence is stable and robust.

OLS estimates. Tables A4-A8 report a series of empirical exercises assessing the robustness of the

results reported in table 3 in the paper, in which the determinants of violence and mobilization in the

Rwandan genocide were explored using OLS.

First, table A4 uses the precolonial district level as the unit of observation, rather than an administra-

tive sector at the time of the genocide. The measures of violence and mobilization used in table 1 vary at

this finer (n=1449) whereas state presence varies at the level of the precolonial district (n=50). Both the

estimated effect sizes and the precision of the estimates are virtually identical to those in table 3.

Second, table A5 assesses the robustness to including polynomials in distance to Kigali and the na-

tional border, and interactions between distance to the border and province fixed effects. Using the frac-

tion of population prosecuted for violence as the dependent variable, this table shows that the main OLS

result is robust to varying these location controls.

Third, table A6 reports results using four different measures of violence in the genocide. To under-

stand the effect of state presence on the ability of the Rwandan government to identify and convict geno-

cide perpetrators in the Gacaca courts, columns (1) and (2) reports the number of individuals prosecuted

for murder and violence, respectively, normalized by the number of Gacaca courts, rather than by popula-

tion. Columns (3) and (4) use data, collected at the commune level (n=135 and n=107, respectively) taken

2

from surveys of the number of genocide victims in Rwandan commissioned in 2004 and 1996 (see data

appendix below for sources). For each of these alternative violence measures, the estimated relationship

between state presence and violence is positive and significant.

Fourth, table A7 includes the log of equivalent household income as a covariate (see Clay et al. (1997)

and Verwimp (2005) for background on the income data). The sample size falls because the income data

is available for a subset of Rwandan sectors. Estimated effects of state presence are virtually unchanged.

Fifth, table A8 adds additional controls that capture aspects of the colonial period in Rwanda. The

rationale behind adding these controls is described in the paper. Each regression includes the number of

archeological remains found in a sector as a measure of pre-Nyigingya population density, the number

of cattle in 1960, and the number of missionary stations (of any denomination) in 1935. Point estimates

fall somewhat, indicating that there may be selection on one or more of these controls. In the main text I

report a sensitivity test based on these estimates.

IV estimates - genocide violence. Table A9 reports IV regressions that are similar to the main IV re-

gressions for violence and mobilization in the genocide reported in table 4, using cost distance to Nyanza

(days) as the instrument. Because the first stage partial F-statistics indicate that this instrument may be

weak, I report Anderson-Rubin confidence intervals with the second stage results. The estimated second

stage effects, and the fact that the Anderson-Rubin confidence intervals do not include zero, suggest a

positive causal effect of state presence on violence and mobilization using this instrument. Note that the

estimated effect sizes are larger than in table 4, although the difference is not statistically significant (re-

sults available upon request).

Table A10 repeats the regressions in table 4 as well, using cost distance to Nyanza - Özak (days) as

the instrument. The estimated second stage effects suggest a positive causal effect of state presence on

violence and mobilization using this instrument. Note that the estimated effect sizes are larger than in

table 4, although the difference is not statistically significant (results available upon request).

Negative binomial - maximum likelihood estimates. Table A11 repeats the regressions in table 5,

including the same covariates as Table A6, the number of archeological remains found in a sector as a

measure of pre-Nyigingya population density, the number of cattle in 1960, and the number of mission-

3

ary stations (of any denomination) in 1935. Results are very similar.

Table A12 repeats the regressions in table 5, normalizing the number of violent events by 1991 popu-

lation. These regressions are estimated using OLS, and show that the negative correlation between state

presence and violence is robust to this normalization

Table A13 reports results using alternative measures of violence from the UCDP data. Whereas in ta-

ble 5 I report results for different types of violence, here I split the violence up by period and whether the

Rwandan government was involved. In column (1), the dependent variable is the count of all events (state

based and one-sided) in the period 1990-1993. Column (2) reports the number of events in the same pe-

riod, restricting to events for which the Rwandan government is recorded as one of the parties involved.

The incidents excluded by this measure are mostly violence of rebel groups against civilians. Columns (3)

and (4) repeat these measures for the period 1995-2000. For each of these alternative violence measures,

the estimated relationship between state presence and violence is positive and significant.

Table A14 repeats the regressions in table 6, estimating using two stage least squares. Results are sim-

ilar.

Table A15 repeats the regressions in table 6, using cost distance to Nyanza (days) as the instrument.

The second stage estimates show a negative causal effect of state presence on violence before and after the

genocide.

Table A16 repeats the regressions in table 6 as well, using cost distance to Nyanza - Özak (days) as the

instrument. The second stage estimates show a negative causal effect of state presence on violence before

and after the genocide as well.

Further additional results. Table A17 estimates the effect of state presence on violence in a differences-

in-differences framework. The sample for this regression includes two observations of violence for each

unit of observation, one for violence in 1994 and one for violence in 1990-1993 and 1995-2000. For violence

in 1994, I use either the number of individuals prosecuted for violence in the Gacaca courts or the number

of violent events recorded in 1994 in the UCDP data. I then define an indicator variable equal to one for

the observations with the 1994 violence data, zero otherwise. I interact this dummy with state presence.

4

I then run regressions using violence as the outcome variable. Column (1) uses the Gacaca data for 1994,

and column (2) uses the UCDP data. Row 1 contains estimates for the interaction between state presence

and the 1994 dummy, and both regressions include the main effects. The 1994 dummy can be interpreted

as capturing the policy shift from pursuing peace and territorial control to the organization of mass vi-

olence. Interpreted this way, the results in this table suggest that violence is relatively higher in areas

with longer state presence when the government aims at organizing violence, relative to the difference in

average violence in the period in which the government pursued peace.

Table A18 reports balance tests for several individual level and village level covariates, in either the

experimental sample, or a sample of village in modern Rwanda. The interpretation of these tests is dis-

cussed in section 4 of this paper.

Table A19 uses questions from the World Values Survey waves of 2007 and 2012 to assess balance on

a series of psychological variables. Questions V86 in the 2007 wave and V76 in the 2012 wave ask how

important it is to the respondent to experience adventure and take risks. The answer scale runs from 1

to 10. I interpret the responses to this question as a measure of risk aversion. It is important to assess

the effect of risk aversion since risk takers may be more likely to cheat in the experiment. I similarly look

at individualism and locus of control, which captures the idea that some people think they are in control

of their lives whereas others feel that fate determines outcomes (Rotter, 1966). Questions V65 in the 2007

wave and V216 in the 2012 wave ask individuals to what extent they agree with the statement that ‘I seek

to be myself rather than to follow others’, which I interpret as a measure of individualism.1 Question V122

in the 2007 wave asks individuals to judge on a scale from 1 to 10 how much they think outcomes are de-

termined by fate versus control. 1 stands for ‘Everything is determined by fate’ and 10 stands for ‘People

shape their fate themselves’. I finally assess trust and confidence in the national government. Questions

V23 in the 2007 and V24 in the 2012 wave ask whether most people can be trusted or whether ‘you can’t

be too careful’. Question V115 in the 2012 survey asks how much confidence respondents have in the

central government in Kigali (on a scale of 1-4, from ‘A great deal’ to ‘None at all’). Across all outcomes,

there is no systematic effect of state presence.

Table A20 reports IV estimates for different measures of local public good provision and state capacity

1The wording of the 2012 question is slightly different. Respondents are now asked to agree with the statement ‘I see myself asan autonomous individual’.

5

before the genocide, using distance to Nyanza as the instrument. First principal components are reported

in column (1). These principal components are also presented in table 7. All estimated effects are small

and insignificant.

Table A21 reports IV and reduced form estimates for different measures of local public good provision

at different points in time before the genocide, using distance to Nyanza as the instrument. Columns

(1)-(3) use the number of hospitals in 1935 and 1960 as well as the number of hospital beds in 1980 as

dependent variables. Columns (4)-(6) use the number of missionary stations in 1924 and 1935 as well as

the number of schools in 1960 as the dependent variables. Both the IV and the reduced form estimates

show no systematic relationship between state presence and these outcomes.

Table A22 explores alternative hypotheses for local differences in genocidal violence from the litera-

ture. Columns (1) and (2) explore the hypothesis advanced in André and Platteau (1998) that Rwanda was

caught in a Malthusian trap. The genocide, in this view, was a Malthusian ‘positive check’. Column (1)

uses the population increase (in percent) between the 1978 and 1991 censuses as a measure of increased

demands placed on the food supply (Ministère du Plan, 1982, 1992). Column (2) uses a measure of the

degree to which a local community could supply the necessary calories to its members in 1980, collected

by the Rwandan government (Lamoure, 1980). The estimated effects are discussed in the text. Columns

(3) and (4) explore the idea that the genocide was the culmination of deep-rooted ethnic hatred between

Hutu and Tutsi (Prunier, 1995). I use the fraction of Tutsi, the group targeted in the genocide, in 1991 as

the dependent variable in column (3). In column (4) I use the fractionalization between Hutu and Tutsi.

Although state presence is positively correlated with both measures of ethnic composition, the effects are

insignificant (column (3)) or marginally significant (column (4)).

Table A23 implements further robustness tests of the main OLS presented in table 3, adding a series of

geographical controls, as well as adding the variables used as dependent variables in table A22 as covari-

ates. Columns (1) and (3) add a number of geographical covariates. To the extent that, for instance, ele-

vation affects agricultural productivity, elevation may correlate with the presence of the Nyiginya state.

The geographical covariates included in this table are terrain elevation, terrain slope, the geographical

suitability for growing banana and an index capturing the local prevalence of the savannah Tsetse fly.

Columns (2) and (4) then add these covariates together with the measures of alternative hypotheses from

6

table A22. Throughout, the estimated effect of state presence remains positive and significant.2

2The estimated effect size decreases somewhat, which is to be expected given that there was a correlation between state presenceand, for instance, fractionalization in table A22.

7

TABLE A1: SUMMARY STATISTICS

N mean sd 10th percentile 90th percentile

Unit of observation: Sector. Sample: Main sampleFraction of population prosecuted for violence 1460 7.916858 7.809828 .3157611 17.24832Fraction of Hutu population prosecuted for organization 1460 1.403843 2.045621 0 3.518452Fraction of Hutu population prosecuted for murder 1460 7.736806 7.886511 .2740837 17.44656Number of violent events 1990-1993, 1995-2000 (count) 1576 .1814721 1.074997 0 0State based violence (against rebels) 1990-1993, 1995-2000 (count) 1576 .089467 .6929968 0 0State based violence (against civilians) 1990-1993, 1995-2000 (count) 1576 .0920051 .5609851 0 0State Presence (Years) 1553 100.9472 65.71094 2 197

Number of individuals prosecuted for genocide participation per Gacaca court 1369 1628.053 1968.748 62.5 3641.379Number of individuals prosecuted for murder per Gacaca court 1369 1379.381 1690.874 47.82609 3116.667Genocide victims per capita 2004 estimate 139 .1298848 .1293098 .0074702 .3536004Genocide victims per capita 1996 estimate 110 .2209414 .304267 .0024985 .6473615Violence 1990-1993 (count) 1576 .088198 .7931214 0 0Violence involving the Rwandan government 1990-1993 (count) 1576 .0805838 .6986453 0 0Violence 1995-2000 (count) 1576 .0932741 .577509 0 0Violence involving the Rwandan government 1995-2000 (count) 1576 .0501269 .3446093 0 0

Distance to Nyanza 1576 65.27946 30.62782 23.12311 101.3946Cost distance to Nyanza (days) 1553 2.90364 1.656719 .86905 5.50025Cost distance to Nyanza - Özak (days) 1553 .7586982 .3620835 .3017139 1.228012Cost distance to nearest Nyiginya army location (days) 1553 2.780276 1.859381 .3982925 6.0696Distance to Addis Abeba 1576 1588381 49647.07 1530541 1656573

Travel distance to Nyanza along 1988 road 1553 105528.5 49801.4 43244.2 169805.6Travel distance to Kigali along 1988 road 1553 99027.58 53825.47 34194.9 168643.6Travel cost distance to Kigali 1553 -.1792182 1.076034 -1.77203 1.447757Distance to country border 1576 22463.18 13972.85 4477.82 42812.34Terrain elevation 1576 1765.349 281.4835 1445 2148Terrain slope 1576 392.1288 298.7535 67 821Distance to nearest river 1576 6.222947 4.908907 1.319759 13.07343Tsetse prevalence 1573 .0790543 .2132588 9.85e-06 .1894174Banana suitability 1576 6.998553 13.51527 0 31.27

Population 2002 (sector) 1565 13601.21 1904.695 11055 16197Population 1991 (commune) 1576 51730.88 18156.11 33315 73115Population 1978 (commune) 1567 35282.79 11216.6 24491 46458Total number of Hutu 1991 1576 4699.1 1706.841 2844 6775Total number of Tutsi 1991 1576 415.6263 421.2897 9 939Fraction of population Tutsi 1991 1576 8.108401 7.934117 .2380953 18.86503Population growth 1978-1991 1567 .4773635 .29467 .238652 .7177965Fractionalization Hutu Tutsi 1991 1576 .1483521 .1209627 .0066854 .3247043log equivalent household income 1990 829 9.367825 .4428594 8.788908 10.06533Unable to meet caloric need of population 1980 1576 .2956853 .456495 0 1Local leader opposition 1993 1483 .163857 .3702707 0 1Radio ownership 1991 1576 .3277789 .0864121 .2386889 .4300992

Principal component energy and water 1980 1576 1.19e-09 1.072162 -.8778389 1.75577Energy centers 1980 1576 .1884518 .4400792 0 1Energy transformers 1980 1576 .2392132 .6708519 0 1Water installations 1980 1576 .4149746 .6573668 0 1Water pumps 1980 1576 .1465736 .3847419 0 1Principal component schools 1980 1576 -1.18e-09 1.14416 -1.352822 1.464512Primary schools 1980 1576 11.32614 4.013814 7 17Secondary schools 1980 1576 .8794416 .8215718 0 2Secondary technical schools 1980 1576 .0513959 .2208741 0 0Principal component health services 1980 1576 -8.57e-10 1.274519 -.9346766 1.50034Hospitals 1980 1576 21.27919 61.51257 0 80Maternity clinics 1980 1576 10.02157 14.55981 0 32Health clinics 1980 1576 16.02602 17.86237 0 40Other clinics 1980 1576 2.637056 21.24841 0 0Principal component social provisions 1980 1576 1.86e-09 1.173693 -1.100518 1.612993Social centers 1980 1576 3.501269 1.436043 2 5Reading centers 1980 1576 1.97335 2.059878 0 5Nutritional centers 1980 1576 .6732234 .694061 0 1Orphanages 1980 1576 .2392132 .6207096 0 1Principal component commercial institutions 1980 1576 -6.13e-09 1.109594 -.832226 .5988494Banks 1980 1576 .3927665 .4885206 0 1Post offices 1980 1576 .4543147 .511897 0 1Grain silos 1980 1576 .090736 .2873245 0 0Gas stations 1980 1576 .0513959 .2208741 0 0Principal component markets 1980 1576 -4.38e-09 1.320582 -1.401663 1.855265Agricultural cooperatives 1980 1576 .1345178 .3413161 0 1Markets 1980 1576 1.331218 1.006993 0 3Trade centers 1980 1576 1.107868 .9083956 0 2Commercial centers 1980 1576 .1091371 .3119101 0 1Principal component fiscal capacity 1980 1318 2.19e-09 1.554533 -1.041444 1.257922Tax received per capita 1980 1318 4014.304 1670.217 2634.2 5873.9Trade tax received per capita 1980 1318 999.2687 1170.166 339 1821.3Other taxes received per capita 1980 1318 1115.142 1323.262 130.1 2158.6Principal component public expenditure 1980 1489 1.40e-09 1.667038 -1.522031 2.537059Public expenditure per capita 1980 1489 3923.835 1577.298 2711.8 5471.4Administrative expenditure per capita 1980 1489 1951.896 460.0686 1499.4 2483.4Social expenditure per capita 1980 1489 574.2789 281.4922 324 892.4Economic expenditure per capita 1980 1489 875.6055 505.1217 377.3 1536.5

Data sources are given in section 3 of this appendix.

8

TABLE A1: SUMMARY STATISTICS CONTINUED

N mean sd 10th percentile 90th percentile

Unit of observation: Sector. Sample: Main sampleNumber of Hospitals 1935 1576 .0164975 .127419 0 0Number of Hospitals 1960 1576 .0590102 .2357185 0 0Number of Hospital beds 1980 1576 21.27919 61.51257 0 80Number of Missionary Stations 1924 1576 .0101523 .106421 0 0Number of Missionary Stations 1935 1576 .0253807 .1573284 0 0Number of Schools 1960 1576 1.046954 1.119673 0 2Number of coffeepulping centers 1960 1576 .1656091 .3852651 0 1

Unit of observation: Individual. Sample: World Values Survey for RwandaJustifiable: Avoiding a fare on public transport 2593 1.909371 1.422605 1 3Justifiable: Stealing property 2589 1.877945 1.595907 1 3Justifiable: Cheating on taxes if you have a chance 2593 2.148477 1.565671 1 4Justifiable: Someone accepting a bribe in the course of their duties 2582 1.801317 1.354184 1 3Justifiable: For a man to beat his wife 2593 3.262245 2.018402 1 6Democracy: People obey their rulers 1203 7.283458 2.657101 3 10Risk taking 2515 2.708946 1.578615 1 5I see myself as an autonomous individual 2578 2.140419 .8224556 1 3Fate versus control 1388 5.201009 2.630162 2 9Most people can be trusted 2590 .1057915 .30763 0 1Confidence: The government (in your nation’s capital) 1203 1.758936 .8280098 1 3Age 2597 34.39353 12.92395 21 52Sex 2597 .4936465 .5000559 0 1

Unit of observation: Individual. Sample: World Values Survey globalJustifiable: avoiding a fare on public transport 285366 2.586752 2.435045 1 6Justifiable: Stealing property 76598 1.805674 1.782325 1 4Justifiable: cheating on taxes 289689 2.255564 2.206697 1 5Justifiable: someone accepting a bribe 303788 1.814966 1.804048 1 4Age 318687 40.03661 16.78253 21 64Sex 318687 .4177045 .7265308 0 1

Unit of observation: Individual. Sample: DHS 2014Emotional violence 1906 .2686254 .4433609 0 1Physical violence 1905 .4440945 .7131331 0 2Sexual violence 1906 .1159496 .3202485 0 1Age 2679 29.59388 8.830199 18 42

Unit of observation: Individual. Sample: Threshold civil society surveyParticipation in community meetings 4937 .8537573 .3361984 0 1Government only speaking at meetings 3615 .0070539 .0824536 0 0Respondents speaks up in community meetings 3939 .5591521 .4930233 0 1Age 9295 35.06595 14.30507 19 56Gender 9319 .4482241 .4973387 0 1

Unit of observation: Individual. Sample: Experiment participantsCompliance with tax demand (% of base) 416 71.34615 27.29979 40 100State presence (dummy) 422 .6184834 .4863354 0 1Age 420 39.9119 13.53734 24 58Male dummy 422 .450237 .498108 0 1Education in years 420 5.454762 3.462772 0 11Daily income 420 1261.986 5116.758 0 2555.556Immigrant 420 .4761905 .5000284 0 1Immigrant from outside sample area 422 .7535545 .4314525 0 1Amount shared with sector employee 422 56.87204 166.243 0 500Amount shared with national gov’t employee 422 82.93839 204.4467 0 500Trust 420 .102381 .3035102 0 1Social Capital 420 2.2 1.248579 1 4Thinks taxes benefit country 404 .9876238 .1106951 1 1Thinks taxes benefit individual 404 .9579208 .2010188 1 1

Unit of observation: Village. Sample: Modern villagesTerrain elevation, village within 10KM from study boundary 300 1450 44.24619 1382 1495Banana suitability, village within 10KM from study boundary 300 42.58467 3.996373 38.12 48.59Terrain elevation, village within 5KM from study boundary 160 1445.875 50.18065 1360 1510Banana suitability, village within 5KM from study boundary 160 41.0825 4.438938 38.12 48.59Terrain elevation, village within 3KM from study boundary 80 1405.75 37.73576 1360 1462Banana suitability, village within 3KM from study boundary 80 40.7375 4.562247 38.12 48.59Terrain elevation, villages in experiment 21 1446.476 47.42217 1388 1495Banana suitability, villages in experiment 21 39.76571 7.847312 25.46 48.59

Data sources are given in section 3 of this appendix.

9

TABLE A2: OLS ESTIMATES OF THE DETERMINANTS OF PRESENCE OF THE NYIGINYA STATE, USINGPRECOLONIAL DISTRICTS AS UNIT OF OBSERVATION

Dependent variable:State Presence (years) Distance to Nyanza

Cost distance toNyanza (days)

Cost distance toNyanza - Özak (days)

Cost distance tonearest Nyiginya army

location (days)

(1) (2) (3) (4) (5) (6) (7) (8)

estimated coefficient -1.470*** -1.504*** -26.33*** -15.33** -114.8*** -121.5*** -9.117** -2.596(0.144) (0.371) (3.513) (7.516) (14.22) (33.18) (4.401) (4.970)

Travel distance to Nyanza along 1988 road network N Y N Y N Y N NTravel distance to Kigali along 1988 road network N Y N Y N Y N YDistance to country border N Y N Y N Y N YProvince fixed effects N Y N Y N Y N Y

Observations 50 50 50 50 50 50 50 50R2 0.466 0.683 0.432 0.657 0.389 0.682 0.063 0.623

Notes: All regressions are estimated using OLS. The unit of observation is a precolonial district. State presence (years) is the numberof years a sector was under centralized rule before colonization in 1897. Distance to Nyanza is the distance (in kilometers) betweenthe centroid of a precolonial district and Nyanza. Cost distance to Nyanza (days) is the travel distance between the centroid ofa precolonial district and Nyanza for a person on foot, measured in hours. Cost distance to Nyanza - Özak (days) is the traveldistance between the centroid of a precolonial district and Nyanza for a person on foot, measured by the Human Mobility Indexproposed by Özak (2010, 2013). Cost distance to nearest Nyiginya army location is the travel distance (in days) between the centroidof a precolonial district and the nearest Nyiginya army battalion, measured in hours. Travel distance is defined as the distance (inkilometers) between the centroid of a precolonial district and an endpoint along the Rwandan road network in 1988. Regressionsinclude fixed effects as indicated in the table. Parentheses report heteroskedasticity robust (White) standard errors. * indicatessignificance at the 10 percent level, ** at the 5 percent level, *** at the 1 percent level.

TABLE A3: OLS ESTIMATES OF THE DETERMINANTS OF PRESENCE OF THE NYIGINYA STATE, WITHDISTANCE TO ADDIS ABBEBA AS A COVARIATE

Dependent variable:State Presence (years) Distance to Nyanza

Cost distance toNyanza (days)

Cost distance toNyanza - Özak (days)

Cost distance tonearest Nyiginya army

location (days)

(1) (2) (3) (4) (5) (6) (7) (8)

estimated coefficient -2.057*** -1.963*** -26.71*** -13.30* -193.3*** -169.0*** -17.40*** -6.704(0.230) (0.544) (4.340) (6.787) (23.41) (45.33) (4.577) (4.298)

Distance to Addis Abbeba Y Y Y Y Y Y Y YTravel distance to Nyanza along 1988 road network N Y N Y N Y N NTravel distance to Kigali along 1988 road network N Y N Y N Y N YDistance to country border N Y N Y N Y N YProvince fixed effects N Y N Y N Y N Y

Observations 50 50 50 50 50 50 50 50R2 0.466 0.683 0.432 0.657 0.389 0.682 0.063 0.623

Notes: All regressions are estimated using OLS. The unit of observation is a precolonial district. State presence (years) is the numberof years a sector was under centralized rule before colonization in 1897. Distance to Nyanza is the distance (in kilometers) betweenthe centroid of a precolonial district and Nyanza. Cost distance to Nyanza (days) is the travel distance between the centroid ofa precolonial district and Nyanza for a person on foot, measured in hours. Cost distance to Nyanza - Özak (days) is the traveldistance between the centroid of a precolonial district and Nyanza for a person on foot, measured by the Human Mobility Indexproposed by Özak (2010, 2013). Cost distance to nearest Nyiginya army location is the travel distance (in days) between the centroidof a precolonial district and the nearest Nyiginya army battalion, measured in hours. Travel distance is defined as the distance (inkilometers) between the centroid of a precolonial district and an endpoint along the Rwandan road network in 1988. Regressionsinclude fixed effects as indicated in the table. Parentheses report heteroskedasticity robust (White) standard errors. * indicatessignificance at the 10 percent level, ** at the 5 percent level, *** at the 1 percent level.

10

TABLE A4: DETERMINANTS OF VIOLENCE AND MOBILIZATION IN THE RWANDAN GENOCIDE USINGPRECOLONIAL DISTRICTS AS UNIT OF OBSERVATION

Dependent variable: Fraction of population prosecuted for: Fraction of Hutu prosecuted for:

Violence Violence Organization Murder(1) (2) (3) (4)

State Presence (Years) 0.0367*** 0.0286*** 0.00369 0.0366***(0.0110) (0.00979) (0.00252) (0.0114)

Fraction of population Tutsi 1991 0.365**(0.140)

Travel distance to Kigali along 1988 road network Y Y Y YDistance to country border Y Y Y Y

Mean of the dependent variable 7.36 7.36 1.28 7.19

Observations 50 50 50 50R2 0.603 0.731 0.558 0.583

Notes: All regressions are estimated using OLS. The unit of observation is a precolonial district. Fraction of population prosecutedfor violence is the number of individuals prosecuted for being an organizer of genocide violence or for murder in the Gacaca courtsdivided by total population in the 1991 census. Fraction of Hutu population prosecuted for organization (murder) is the numberof individuals prosecuted for organizing genocide violence (murder) in the Gacaca courts divided by total Hutu population in the1991 census. State presence (years) is the number of years a sector was under centralized rule before colonization in 1897. Fractionof population Tutsi 1991 is the total number of Tutsi divided by total population in the 1991 census. Travel distance is defined as thedistance (in kilometers) between the centroid of a precolonial district and an endpoint along the Rwandan road network in 1988. Allregressions include fixed effects at the modern province level. Parentheses report heteroskedasticity robust (White) standard errors.* indicates significance at the 10 percent level, ** at the 5 percent level, *** at the 1 percent level.

TABLE A5: DETERMINANTS OF VIOLENCE AND MOBILIZATION IN THE RWANDAN GENOCIDE, USINGDIFFERENT LOCATION COVARIATES

Dependent variable: Fraction of population prosecuted for violence

(1) (2) (3) (4)

State Presence (Years) 0.0294*** 0.0250** 0.0242*** 0.0266***(0.00973) (0.00968) (0.00870) (0.00805)

Distance to Kigali Third order poly. Linear Linear Third order poly.Distance to country border Linear Third order poly. Linear × province f.e. Third order poly. + f.e. interactions

Observations 1449 1449 1449 1449R2 0.185 0.195 0.194 0.212

Notes: All regressions are estimated using OLS. The unit of observation is an administrative sector in 1994. Fraction of populationprosecuted for violence is the number of individuals prosecuted for being an organizer of genocide violence or for murder in theGacaca courts divided by total population in the 1991 census. State presence (years) is the number of years a sector was undercentralized rule before colonization in 1897. All regressions include fixed effects at the modern province level. Parentheses reportclustered standard errors at the level of the precolonial district. * indicates significance at the 10 percent level, ** at the 5 percentlevel, *** at the 1 percent level, computed using standard errors clustered at the precolonial district level.

11

TABLE A6: DETERMINANTS OF VIOLENCE IN THE RWANDAN GENOCIDE USING ALTERNATIVEVIOLENCE MEASURES

Dependent variable:

Number of individualsprosecuted for genocide participation

per Gacaca courtNumber of individuals

prosecuted for murder per Gacaca courtGenocide victims per capita

2004 estimateGenocide victims per capita

1996 estimate

(1) (2) (3) (4)

State Presence (Years) 5.125** 4.531** 0.00101*** 0.00181***(2.127) (1.790) (0.000208) (0.000639)

2002 Population Y Y N NTravel distance to Kigali along 1988 road network Y Y Y YDistance to country border Y Y Y YProvince fixed effects Y Y N N

Mean of the dependent variable 1633.416 1384.178 13.08 22.17

Number of clusters 50 50 NA NAObservations 1359 1359 135 107R2 0.202 0.188 0.188 0.086

Notes: All regressions are estimated using OLS. The unit of observation is an administrative sector in 1994 in columns (1) and (2) andthe administrative commune in columns (3) and (4). Number of individuals prosecuted for genocide participation per Gacaca court isthe number of individuals prosecuted for being an organizer of genocide violence or for murder in the Gacaca courts divided by totalnumber of Gacaca courts. Number of individuals prosecuted for murder per Gacaca court is the number of individuals prosecutedfor murder in the Gacaca courts divided by total number of Gacaca courts. Genocide victims per capita, 2004 estimate is an estimateof genocide victims as identified by the Rwandan government in 2004 (see appendix for exact source) divided by total populationin the 1991 census. Genocide victims per capita, 1996 estimate is an estimate of genocide victims as identified by the Rwandangovernment in 2004 (see appendix for exact source) divided by total population in the 1991 census. 2002 Population is the totalpopulation in the 2002 census. Travel distance is defined as the distance (in kilometers) between the centroid of a precolonial districtand an endpoint along the Rwandan road network in 1988. Regressions include fixed effects as indicated in the table. Parenthesesreport clustered standard errors at the level of the precolonial district. * indicates significance at the 10 percent level, ** at the 5percent level, *** at the 1 percent level.

TABLE A7: DETERMINANTS OF VIOLENCE AND MOBILIZATION IN THE RWANDAN GENOCIDE,WITH INCOME AS A CONTROL

Dependent variable:

Fraction ofpopulation

prosecuted for: Fraction of Hutu prosecuted for:

Log equivalentadult income

1990Violence Organization Murder

(1) (2) (3) (4)

State Presence (Years) 0.00109 0.0340** 0.00353 0.0321**(0.000880) (0.0136) (0.00314) (0.0140)

Log equivalent adult income 1990 N Y Y YTravel distance to Kigali along 1988 road network Y Y Y YDistance to country border Y Y Y Y

Number of clusters 50 50 50 50Observations 814 763 763 763R2 0.373 0.177 0.103 0.173

Notes: All regressions are estimated using OLS. The unit of observation is an administrative sector in 1994. Log equivalent adultincome is the natural logarithm of the equivalent adult income in Rwandan Francs in 1990. Fraction of population prosecuted forviolence is the number of individuals prosecuted for being an organizer of genocide violence or for murder in the Gacaca courtsdivided by total population in the 1991 census. Fraction of Hutu population prosecuted for organization (murder) is the number ofindividuals prosecuted for organizing genocide violence (murder) in the Gacaca courts divided by total Hutu population in the 1991census. State presence (years) is the number of years a sector was under centralized rule before colonization in 1897. Travel distanceis defined as the distance (in kilometers) between the centroid of a precolonial district and an endpoint along the Rwandan roadnetwork in 1988. All regressions include fixed effects at the modern province level. Parentheses report clustered standard errors atthe level of the precolonial district. * indicates significance at the 10 percent level, ** at the 5 percent level, *** at the 1 percent level.

12

TABLE A8: DETERMINANTS OF VIOLENCE AND MOBILIZATION IN THE RWANDAN GENOCIDE,WITH ADDITIONAL CONTROLS

Dependent variable: Fraction of population prosecuted for: Fraction of Hutu prosecuted for:

Violence Violence Organization Murder(1) (2) (3) (4)

State Presence (Years) 0.0295*** 0.0277*** 0.00291 0.0276***(0.00950) (0.00783) (0.00242) (0.00972)

Fraction of population Tutsi 1991 0.247***(0.0641)

Pre-Nyiginya population density Y Y Y YNr. of cattle 1960 Y Y Y YNr. of missionary stations 1935 Y Y Y Y

Travel distance to Kigali along 1988 road network Y Y Y YDistance to country border Y Y Y Y

Altonji et al. (2005) selection statisticusing corresponding table 3 point estimate 25.53 26.24 12.45 25.89

Mean of the dependent variable 7.95 7.95 7.75 7.75

Number of clusters 50 50/142 50 50Observations 1449 1449 1449 1449R2 0.210 0.252 0.201 0.289

Notes: All regressions are estimated using OLS. The unit of observation is an administrative sector in 1994. Fraction of populationprosecuted for violence is the number of individuals prosecuted for being an organizer of genocide violence or for murder in theGacaca courts divided by total population in the 1991 census. Fraction of Hutu population prosecuted for organization (murder)is the number of individuals prosecuted for organizing genocide violence (murder) in the Gacaca courts divided by total Hutupopulation in the 1991 census. State presence (years) is the number of years a sector was under centralized rule before colonizationin 1897. Fraction of population Tutsi 1991 is the total number of Tutsi divided by total population in the 1991 census. Pre-Nyiginyapopulation density is the number of archeological sites indicating population presence before the establishment of the Nyiginyakingdom. Nr. of cattle 1960 is the number of cattle in 1960. Nr. of missionary stations 1935 is the number of missionary stations(any denomination) in 1935. Travel distance is defined as the distance (in kilometers) between the centroid of a precolonial districtand an endpoint along the Rwandan road network in 1988. Parentheses report (double) clustered standard errors at the level of theprecolonial district, or at the level of commune and the precolonial district. * indicates significance at the 10 percent level, ** at the 5percent level, *** at the 1 percent level.

13

TABLE A9: IV ESTIMATES OF THE EFFECT OF STATE PRESENCE ON VIOLENCE AND MOBILIZATION IN THERWANDAN GENOCIDE USING COST DISTANCE TO NYANZA (DAYS) AS INSTRUMENT

Dependent variable: Fraction of population prosecuted for: Fraction of Hutu prosecuted for:

Violence Violence Organization Murder(1) (2) (3) (4)

Panel I: Second stage estimates

State Presence (Years) 0.117** 0.0871** 0.0189* 0.0730**Clustered s.e. (0.0586) (0.0419) (0.0110) (0.0365)

Anderson-Rubin c.i. [0.050, 0.347] [0.032, 0.251] [0.006, 0.061] [0.051, 0.351]

Mean of the dependent variable 7.95 7.95 1.40 7.75

Panel II: First stage estimates

Cost distance to Nyanza (days) -14.80** -15.60** -14.80** -14.80**(7.157) (7.139) (7.157) (7.157)

F-stat of excluded instrument 4.43 5.19 4.43 4.43R2 0.594 0.595 0.594 0.594

Fraction of population Tutsi 1991 N Y N YTravel distance to Nyanza along 1988 road Y Y Y YTravel distance to Kigali along 1988 road Y Y Y YDistance to country border Y Y Y Y

Number of clusters 50 50/142 50 50Observations 1449 1449 1449 1449

Notes: All regressions in Panel I are estimated using two stage least squares. All regressions in panel II are estimated using OLS. Theunit of observation is an administrative sector in 1994. Fraction of population prosecuted for violence is the number of individualsprosecuted for being an organizer of genocide violence or for murder in the Gacaca courts divided by total population in the 1991census. Fraction of Hutu population prosecuted for organization (murder) is the number of individuals prosecuted for organizinggenocide violence (murder) in the Gacaca courts divided by total Hutu population in the 1991 census. State presence (years) is thenumber of years a sector was under centralized rule before colonization in 1897. Cost distance to Nyanza (days) is the travel distancebetween the centroid of a precolonial district and Nyanza for a person on foot, measured in hours. Fraction of population Tutsi 1991is the total number of Tutsi divided by total population in the 1991 census. Travel distance is defined as the distance (in kilometers)between the centroid of a precolonial district and an endpoint along the Rwandan road network in 1988. All regressions includefixed effects at the modern province level. Parentheses report (double) clustered standard errors at the level of the precolonialdistrict, or at the level of commune and the precolonial district. * indicates significance at the 10 percent level, ** at the 5 percentlevel, *** at the 1 percent level.

14

TABLE A10: IV ESTIMATES OF THE EFFECT OF STATE PRESENCE ON VIOLENCE AND MOBILIZATION INTHE RWANDAN GENOCIDE USING COST DISTANCE TO NYANZA - ÖZAK (DAYS) AS INSTRUMENT

Dependent variable: Fraction of population prosecuted for: Fraction of Hutu prosecuted for:

Violence Violence Organization Murder(1) (2) (3) (4)

Panel I: Second stage estimates

State Presence (Years) 0.102*** 0.0824*** 0.0159** 0.0997***(0.0312) (0.0223) (0.00663) (0.0309)

Mean of the dependent variable 7.95 7.95 1.40 7.75

Panel II: First stage estimates

Cost distance to Nyanza - Özak (days) -134.7*** -142.2*** -134.7*** -134.7***(34.50) (31.85) (34.50) (34.50)

F-stat of excluded instrument 15.01 20.70 15.01 15.01R2 0.629 0.632 0.629 0.629

Fraction of population Tutsi 1991 N Y N YTravel distance to Nyanza along 1988 road Y Y Y YTravel distance to Kigali along 1988 road Y Y Y YDistance to country border Y Y Y Y

Number of clusters 50 50 50 50Observations 1449 1449 1449 1449

Notes: All regressions in Panel I are estimated using two stage least squares. All regressions in panel II are estimated using OLS. Theunit of observation is an administrative sector in 1994. Fraction of population prosecuted for violence is the number of individualsprosecuted for being an organizer of genocide violence or for murder in the Gacaca courts divided by total population in the 1991census. Fraction of Hutu population prosecuted for organization (murder) is the number of individuals prosecuted for organizinggenocide violence (murder) in the Gacaca courts divided by total Hutu population in the 1991 census. State presence (years) is thenumber of years a sector was under centralized rule before colonization in 1897. Cost distance to Nyanza - Özak (days) is the traveldistance between the centroid of a precolonial district and Nyanza for a person on foot, measured by the Human Mobility Indexproposed by Özak (2010, 2013). Fraction of population Tutsi 1991 is the total number of Tutsi divided by total population in the 1991census. Travel distance is defined as the distance (in kilometers) between the centroid of a precolonial district and an endpoint alongthe Rwandan road network in 1988. All regressions include fixed effects at the modern province level. Parentheses report (double)clustered standard errors at the level of the precolonial district, or at the level of commune and the precolonial district. * indicatessignificance at the 10 percent level, ** at the 5 percent level, *** at the 1 percent level.

15

TABLE A11: DETERMINANTS OF VIOLENCE BEFORE AND AFTER THE RWANDAN GENOCIDE,WITH ADDITIONAL CONTROLS

Dependent variable:

Number of violentevents

1990-1993, 1995-2000(count)

State based violence(against rebels)

1990-1993, 1995-2000(count)

One-sided violence(against civilians)

1990-1993, 1995-2000(count)

(1) (2) (3) (4)

State Presence (Years) -0.00150*** -0.00130*** -0.00036*** -0.00045*(0.00047) (0.00048) (0.00012) (0.00026)

Fraction of population Tutsi 1991 -0.00387(0.00310)

Pre-Nyiginya population density Y Y Y YNr. of cattle 1960 Y Y Y YNr. of missionary stations 1935 Y Y Y Y

Travel distance to Kigali along 1988 road network Y Y Y YDistance to country border Y Y Y Y

Mean of the dependent variable 0.18 0.18 0.09 0.09

Log-likelihood -507.08 -505.61 -309.88 -354.77Number of clusters 50 50 50 50Observations 1553 1553 1553 1553

Notes: All regressions are negative binomial regressions, estimated using maximum likelihood. Estimated coefficients are marginaleffects (at the mean). The unit of observation is an administrative sector in 1994. Number of violent events 1990-1993, 1995-2000(count) is a count variable measuring the number of violent events resulting in more than twenty five casualties that are classifiedas either state based violence or one-sided violence in the UCDP dataset occurring between January 1990 and December 1993, andbetween January 1995 and December 2000. State based violence (against rebels) 1990-1993, 1995-2000 (count) is a count variablemeasuring the number of violent events resulting in more than twenty five casualties that are classified as state based violence inthe UCDP dataset occurring between January 1990 and December 1993, and between January 1995 and December 2000. One-sidedviolence (against civilians) 1990-1993, 1995-2000 (count) is a count variable measuring the number of violent events resulting inmore than twenty five casualties that are classified as one-sided violence in the UCDP dataset occurring between January 1990 andDecember 1993, and between January 1995 and December 2000. State presence (years) is the number of years a sector was undercentralized rule before colonization in 1897. Fraction of population Tutsi 1991 is the total number of Tutsi divided by total populationin the 1991 census. Pre-Nyiginya population density is the number of archeological sites indicating population presence before theestablishment of the Nyiginya kingdom. Nr. of cattle 1960 is the number of cattle in 1960. Nr. of missionary stations 1935 is thenumber of missionary stations (any denomination) in 1935. Travel distance is defined as the distance (in kilometers) between thecentroid of a precolonial district and an endpoint along the Rwandan road network in 1988. All regressions include fixed effectsat the modern province level. Parentheses report (double) clustered standard errors at the level of the precolonial district, or at thelevel of commune and the precolonial district. * indicates significance at the 10 percent level, ** at the 5 percent level, *** at the 1percent level.

16

TABLE A12: DETERMINANTS OF VIOLENCE BEFORE AND AFTER THE RWANDAN GENOCIDE,NORMALIZING BY POPULATION

Dependent variable:

Number of violentevents

1990-1993, 1995-2000 (%of population)

State based violence(against rebels)

1990-1993, 1995-2000 (%of population)

One-sided violence(against civilians)

1990-1993, 1995-2000 (%of population)

(1) (2) (3) (4)

State Presence (Years) -0.0000328** -0.0000310** -0.0000271** -0.00000568(0.0000162) (0.0000153) (0.0000112) (0.00000650)

Fraction of population Tutsi 1991 -0.000176**(0.0000780)

Travel distance to Kigali along 1988 road network Y Y Y YDistance to country border Y Y Y Y

Mean of the dependent variable 0.18 0.18 0.09 0.09

Number of clusters 50 50 50 50Observations 1553 1553 1553 1553R2 0.014 0.016 0.018 0.014

Notes: All regressions are estimated using OLS. The unit of observation is an administrative sector in 1994. The unit of observationis an administrative sector in 1994. Number of violent events 1990-1993, 1995-2000 (% of population) is the number of violent eventsresulting in more than twenty five casualties that are classified as either state based violence or one-sided violence in the UCDPdataset occurring between January 1990 and December 1993, and between January 1995 and December 2000, normalized by 1991population. State based violence (against rebels) 1990-1993, 1995-2000 (% of population) is the number of violent events resultingin more than twenty five casualties that are classified as state based violence in the UCDP dataset occurring between January 1990and December 1993, and between January 1995 and December 2000, normalized by 1991 population. One-sided violence (againstcivilians) 1990-1993, 1995-2000 (% of population) is the number of violent events resulting in more than twenty five casualties that areclassified as one-sided violence in the UCDP dataset occurring between January 1990 and December 1993, and between January 1995and December 2000, normalized by 1991 population. State presence (years) is the number of years a sector was under centralizedrule before colonization in 1897. Fraction of population Tutsi 1991 is the total number of Tutsi divided by total population in the1991 census. Travel distance is defined as the distance (in kilometers) between the centroid of a precolonial district and an endpointalong the Rwandan road network in 1988. All regressions include fixed effects at the modern province level. Parentheses reportclustered standard errors at the level of the precolonial district. * indicates significance at the 10 percent level, ** at the 5 percentlevel, *** at the 1 percent level.

17

TABLE A13: DETERMINANTS OF VIOLENCE BEFORE AND AFTER THE RWANDAN GENOCIDE, USINGALTERNATIVE VIOLENCE MEASURES

Dependent variable:Violence 1990-1993

(count)

Violence involvingthe Rwandangovernment

1990-1993 (count)Violence 1995-2000

(count)

Violence involvingthe Rwandangovernment

1995-2000 (count)

(1) (2) (3) (4)

State Presence (Years) -0.00075*** -0.00071*** -0.00064 -0.00047***(0.00024) (0.00024) (0.00037) (0.00016)

Travel distance to Kigali along 1988 road network Y Y Y YDistance to country border Y Y Y Y

Mean of the dependent variable 0.089 0.081 0.09 0.048

Log-likelihood -283.25 -279.46 -360.16 -242.22Number of clusters 50 50 50 50Observations 1553 1553 1553 1553

Notes: Negative binomial regressions, estimated using maximum likelihood. Estimated coefficients are marginal effects (at themean). The unit of observation is an administrative sector in 1994. Violence 1990-1993 (count) is a count variable measuring thenumber of violent events that resulted in more than twenty five casualties in the UCDP dataset occurring between January 1990and December 1993. Violence involving the Rwandan government 1990-1993 (count) is a count variable measuring the number ofviolent events involving the Rwandan government that resulted in more than twenty five casualties in the UCDP dataset occurringbetween January 1990 and December 1993. Violence 1995-2000 (count) is a count variable measuring the number of violent eventsthat resulted in more than twenty five casualties in the UCDP dataset occurring between January 1995 and December 2000. Violenceinvolving the Rwandan government 1995-2000 (count) is a count variable measuring the number of violent events involving theRwandan government that resulted in more than twenty five casualties in the UCDP dataset occurring between January 1995 andDecember 2000. Travel distance is defined as the distance (in kilometers) between the centroid of a precolonial district and anendpoint along the Rwandan road network in 1988. All regressions include fixed effects at the modern province level. Parenthesesreport clustered standard errors at the level of the precolonial district. * indicates significance at the 10 percent level, ** at the 5percent level, *** at the 1 percent level, computed using standard errors clustered at the precolonial district level.

18

TABLE A14: ESTIMATES OF THE EFFECT OF STATE PRESENCE ON VIOLENCE BEFORE AND AFTER THERWANDAN GENOCIDE USING TWO STAGE LEAST SQUARES

Dependent variable:

Number of violentevents

1990-1993, 1995-2000(count)

State based violence(against rebels)

1990-1993, 1995-2000(count)

One-sided violence(against civilians)

1990-1993, 1995-2000(count)

(1) (2) (3) (4)

Panel I: Second stage estimates

State Presence (Years) -0.00754** -0.00689*** -0.00501** -0.00254**(0.00296) (0.00264) (0.00208) (0.00119)

Mean of the dependent variable 0.18 0.18 0.09 0.09

Panel II: First stage estimates

Distance to Nyanza -1.776*** -1.848*** -1.776*** -1.776***(0.472) (0.467) (0.472) (0.472)

F-stat of excluded instrument 14.38 16.02 14.38 14.38R2 0.627 0.629 0.627 0.627

Travel distance to Nyanza along 1988 road Y Y Y YTravel distance to Kigali along 1988 road Y Y Y YDistance to country border Y Y Y Y

Number of clusters 50 50 50 50Observations 1553 1553 1553 1553

Notes: All regressions in Panel I are estimated using two stage least squares. All regressions in panel II are estimated using OLS.The unit of observation is an administrative sector in 1994. Number of violent events 1990-1993, 1995-2000 (count) is a countvariable measuring the number of violent events resulting in more than twenty five casualties that are classified as either state basedviolence or one-sided violence in the UCDP dataset occurring between January 1990 and December 1993, and between January1995 and December 2000. State based violence (against rebels) 1990-1993, 1995-2000 (count) is a count variable measuring thenumber of violent events resulting in more than twenty five casualties that are classified as state based violence in the UCDP datasetoccurring between January 1990 and December 1993, and between January 1995 and December 2000. One-sided violence (againstcivilians) 1990-1993, 1995-2000 (count) is a count variable measuring the number of violent events resulting in more than twentyfive casualties that are classified as one-sided violence in the UCDP dataset occurring between January 1990 and December 1993,and between January 1995 and December 2000. State presence (years) is the number of years a sector was under centralized rulebefore colonization in 1897. Cost distance to Nyanza (days) is the travel distance between the centroid of a precolonial district andNyanza for a person on foot, measured in hours. Travel distance is defined as the distance (in kilometers) between the centroid ofa precolonial district and an endpoint along the Rwandan road network in 1988. All regressions include fixed effects at the modernprovince level. In panel I, parentheses report bootstrapped standard errors. In panel II, parentheses report clustered standard errorsat the level of the precolonial district. * indicates significance at the 10 percent level, ** at the 5 percent level, *** at the 1 percent level.

19

TABLE A15: ESTIMATES OF THE EFFECT OF STATE PRESENCE ON VIOLENCE BEFORE AND AFTER THERWANDAN GENOCIDE USING COST DISTANCE TO NYANZA (DAYS) AS INSTRUMENT

Dependent variable:

Number of violentevents

1990-1993, 1995-2000(count)

State based violence(against rebels)

1990-1993, 1995-2000(count)

One-sided violence(against civilians)

1990-1993, 1995-2000(count)

(1) (2) (3) (4)

Panel I: Second stage estimates

State Presence (Years) -0.00702** -0.00633** -0.00436** -0.00266**(0.00279) (0.00248) (0.00183) (0.00120)

Mean of the dependent variable 0.18 0.18 0.09 0.09

Panel II: First stage estimates

Cost distance to Nyanza (days) -15.41** -16.62** -15.41** -15.41**(7.343) (7.291) (7.343) (7.343)

R2 0.597 0.599 0.597 0.597

Travel distance to Nyanza along 1988 road Y Y Y YTravel distance to Kigali along 1988 road Y Y Y YDistance to country border Y Y Y Y

Number of clusters 50 50 50 50Observations 1553 1553 1553 1553

Notes: All regressions in Panel I are negative binomial regressions, reporting marginal effects (at the mean) estimated using acontrol function - maximum likelihood approach with Cost distance to Nyanza (days) as the instrument. All regressions in panel IIare estimated using OLS. The unit of observation is an administrative sector in 1994. Number of violent events 1990-1993, 1995-2000(count) is a count variable measuring the number of violent events resulting in more than twenty five casualties that are classifiedas either state based violence or one-sided violence in the UCDP dataset occurring between January 1990 and December 1993, andbetween January 1995 and December 2000. State based violence (against rebels) 1990-1993, 1995-2000 (count) is a count variablemeasuring the number of violent events resulting in more than twenty five casualties that are classified as state based violence inthe UCDP dataset occurring between January 1990 and December 1993, and between January 1995 and December 2000. One-sidedviolence (against civilians) 1990-1993, 1995-2000 (count) is a count variable measuring the number of violent events resulting inmore than twenty five casualties that are classified as one-sided violence in the UCDP dataset occurring between January 1990and December 1993, and between January 1995 and December 2000. State presence (years) is the number of years a sector wasunder centralized rule before colonization in 1897. Cost distance to Nyanza (days) is the travel distance between the centroid of aprecolonial district and Nyanza for a person on foot, measured in hours. Travel distance is defined as the distance (in kilometers)between the centroid of a precolonial district and an endpoint along the Rwandan road network in 1988. All regressions includefixed effects at the modern province level. In panel I, parentheses report bootstrapped standard errors. In panel II, parenthesesreport clustered standard errors at the level of the precolonial district. * indicates significance at the 10 percent level, ** at the 5percent level, *** at the 1 percent level.

20

TABLE A16: ESTIMATES OF THE EFFECT OF STATE PRESENCE ON VIOLENCE BEFORE AND AFTER THERWANDAN GENOCIDE USING COST DISTANCE TO NYANZA - ÖZAK (DAYS) AS INSTRUMENT

Dependent variable:

Number of violentevents

1990-1993, 1995-2000(count)

State based violence(against rebels)

1990-1993, 1995-2000(count)

One-sided violence(against civilians)

1990-1993, 1995-2000(count)

(1) (2) (3) (4)

Panel I: Second stage estimates

State Presence (Years) -0.00392*** -0.00376*** -0.00138*** -0.00200***(0.00109) (0.000951) (0.000452) (0.000639)

Mean of the dependent variable 0.18 0.18 0.09 0.09

Panel II: First stage estimates

Cost distance to Nyanza - Özak (days) -139.8*** -148.4*** -139.8*** -139.8***(36.12) (33.53) (36.12) (36.12)

R2 0.633 0.638 0.633 0.633

Travel distance to Nyanza along 1988 road Y Y Y YTravel distance to Kigali along 1988 road Y Y Y YDistance to country border Y Y Y Y

Number of clusters 50 50 50 50Observations 1553 1553 1553 1553

Notes: All regressions in Panel I are negative binomial regressions, reporting marginal effects (at the mean) estimated using a controlfunction - maximum likelihood approach with Cost distance to Nyanza - Özak (days) as the instrument. All regressions in panel IIare estimated using OLS. The unit of observation is an administrative sector in 1994. Number of violent events 1990-1993, 1995-2000(count) is a count variable measuring the number of violent events resulting in more than twenty five casualties that are classifiedas either state based violence or one-sided violence in the UCDP dataset occurring between January 1990 and December 1993, andbetween January 1995 and December 2000. State based violence (against rebels) 1990-1993, 1995-2000 (count) is a count variablemeasuring the number of violent events resulting in more than twenty five casualties that are classified as state based violence inthe UCDP dataset occurring between January 1990 and December 1993, and between January 1995 and December 2000. One-sidedviolence (against civilians) 1990-1993, 1995-2000 (count) is a count variable measuring the number of violent events resulting inmore than twenty five casualties that are classified as one-sided violence in the UCDP dataset occurring between January 1990 andDecember 1993, and between January 1995 and December 2000. State presence (years) is the number of years a sector was undercentralized rule before colonization in 1897. Cost distance to Nyanza - Özak (days) is the travel distance between the centroid ofa precolonial district and Nyanza for a person on foot, measured by the Human Mobility Index proposed by Özak (2010, 2013).Travel distance is defined as the distance (in kilometers) between the centroid of a precolonial district and an endpoint along theRwandan road network in 1988. All regressions include fixed effects at the modern province level. In panel I, parentheses reportbootstrapped standard errors. In panel II, parentheses report clustered standard errors at the level of the precolonial district. *indicates significance at the 10 percent level, ** at the 5 percent level, *** at the 1 percent level.

21

TABLE A17: DETERMINANTS OF VIOLENCE AND MOBILIZATION, 1990-2000

Dependent variable:Violence, Gacaca (1994) and UCDP

(1990-1993, 1995-2000)Violence, UCDP (1994 and 1990-1993,

1995-2000)

(1) (2)

State Presence (years) * 1994 dummy 0.00539*** 0.00163**(0.00138) (0.000640)

Main effects: State Presence (years) and 1994 dummy Y YTravel distance to Kigali along 1988 road network Y YDistance to country border Y Y

Number of clusters 50 50Observations 3002 3106R2 0.066 0.012

Notes: All regressions are estimated using OLS. The unit of observation is an administrative sector in 1994. The sample containseach sector twice, once for 1994 and once for the period before and after the genocide. Violence, Gacaca (1994) and UCDP (1990-1993, 1995-2000) is a measure of violence that uses the Gacaca violence data for 1994 and the UCDP data for the surrounding period.Violence, UCDP (1990-1993, 1995-2000) is a measure of violence that uses the UCDP data for both 1994 and the surrounding years.State presence (years) is the number of years a sector was under centralized rule before colonization in 1897. State Presence (years)* 1994 dummy is the interaction of an indicator variable that equals one for 1994, and zero for the surrounding years, with statepresence in years. Travel distance is defined as the distance (in kilometers) between the centroid of a precolonial district and anendpoint along the Rwandan road network in 1988. All regressions include fixed effects at the modern province level. Parenthesesreport clustered standard errors at the level of the precolonial district. * indicates significance at the 10 percent level, ** at the 5percent level, *** at the 1 percent level, computed using standard errors clustered at the precolonial district level.

TABLE A18: CORRELATION OF PREDETERMINED CHARACTERISTICS AND STATE PRESENCE

(1) (2) (3) (4) (5) (6) (7) (8)

Dependent variable: ImmigrantImmigrant from

outside sample areaAmount shared

with sector employee

Amount sharedwith national gov’t

employee Trust Social capitalThinks taxes

benefit countryThinks taxes

benefit individual

Panel I: individual covariates

State Presence (100 years) -0.126 0.119 0.0478 -0.162 -0.0214 0.0477 0.114 0.0692(0.0982) (0.101) (0.101) (0.106) (0.0989) (0.0974) (0.119) (0.108)

Observations 420 420 420 420 420 420 404 404R2 0.046 0.056 0.006 0.015 0.050 0.043 0.013 0.010

Dependent variable:Elevation

<10KM from AkanyaruBanana Suitability

<10KM from AkanyaruElevation

<5KM from AkanyaruBanana Suitability

<5KM from AkanyaruElevation

<3KM from AkanyaruBanana Suitability

<3KM from AkanyaruElevation

experimental sampleBanana Suitability

experimental sample

Panel II: Geography

State Presence (100 years) 30.55*** -5.415*** 15.77** 0.114 4.775 0.379 32.88 -3.565(10.99) (0.909) (6.792) (1.113) (8.720) (1.307) (20.37) (3.519)

Observations 300 300 164 164 82 82 21 21R2 0.149 0.260 0.033 0.000 0.004 0.001 0.119 0.051

Notes: All regressions are estimated using OLS. The unit of observation is an individual respondent in panel I and a village inpanel II. Education in years is the total number of years a respondent has received fulltime education. Eq. daily income (RWF) isthe equivalent daily income earned by a respondent. Immigrant is an indicator variable equal to one if the respondent immigratedinto the sample village. Immigrant from outside area is an indicator variable equal to one if the respondent immigrated into thesample village from outside the sample districts Bugesera, Kamonyi, Ruhango and Nyanza. Amount shared with sector employeeis the amount shared (out of RWF 1500) with an anonymous employee of the local administrative sector in which the samplevillage was located in a dictator game. Amount shared with national government employee is the amount shared (out of RWF1500) with an anonymous employee of the national government in a dictator game. Trust is an indicator variable equal to oneif the respondent answered that most people can be trusted in the post-fieldwork. Social capital is the number of civil societyorganizations a respondent is a member of. Thinks taxes benefit country is an indicator variable equal to one if a respondentanswers that he/she thinks taxes benefit Rwanda in a fieldwork survey. Thinks taxes benefit individual is an indicator variableequal to one if a respondent answers that he/she thinks taxes benefit him/herself in a fieldwork survey. Terrain elevation is theelevation in meters under the centroid of the village the respondent lives in. Banana suitability is the geographical suitability forgrowing bananas in the respondent’s village. State Presence (100 years) is an indicator equal to one if a village is situated to the Eastof the Akanyaru river in October 2014. Parentheses give heteroskedasticity robust standard errors. * indicates significance at the 10percent level, ** at the 5 percent level, *** at the 1 percent level.

22

TABLE A19: ALTERNATIVE PSYCHOLOGICAL MECHANISMS FOR RULE FOLLOWING

Dependent variable: Risk taking Individualism Locus of control TrustConfidence ingovernment

(1) (2) (3) (4) (5)

State Presence (Years) -0.0222 -0.0397 -0.0488 -0.00722 0.0123(0.0277) (0.0275) (0.0444) (0.0274) (0.0386)

Demographic controls Y Y Y Y YSurvey year fixed effect Y Y Y Y Y

Permutation test p-value 0.42 0.15 0.27 0.79 0.75Observations 2515 2578 1388 2590 1203R2 0.007 0.028 0.003 0.046 0.017

Notes: All regressions are estimated using OLS. The unit of observation is the individual. All point estimates are standardized. Eachdependent variable is taken from questions in the World Values survey waves of 2007 and 2012. Risk taking is the answer to thequestion whether ‘Adventure and taking risks are important to this person’. Individualism is the extent to which a respondent seeshim/herself as an autonomous individual. Locus of control is the extent to which the respondent feels he/she is in control of his/herlife. Trust is a question asking whether most people can be trusted or that you can’t be too careful. Confidence in government askshow confidence the respondent has in the government in Kigali. State presence (years) is the number of years a sector was undercentralized rule before colonization in 1897. Demographic controls include age, age squared and a gender dummy. Survey yearfixed effects are indicator variables for whether an individual respondent was interviewed as part of the 2007 or the 2012 WorldValues survey wave. Heteroskedasticity robust standard errors are in parentheses. * indicates significance at the 10 percent level, **at the 5 percent level, *** at the 1 percent level.

23

TABLE A20: ESTIMATES OF THE EFFECT OF STATE PRESENCE ON LOCAL GOVERNMENT

(1) (2) (3) (4) (5)

Panel I: Dependent variable:Principal component

energy and water 1980Energy

centers 1980Energy

transformers 1980Water

installations 1980Water

pumps 1980

State Presence (Years) 0.00517 -0.00228 0.00329 -0.00601 0.00441*(0.00681) (0.00280) (0.00371) (0.00655) (0.00241)

Observations 1553 1553 1553 1553 1553

Panel II: Dependent variable:Principal component

schools 1980Primary

schools 1980Secondary

schools 1980Secondary

technical schools 1980

State Presence (Years) 0.000398 -0.0199 0.000345 0.00226(0.00751) (0.0252) (0.00510) (0.00161)

Observations 1553 1553 1553 1553

Panel III: Dependent variable:Principal componenthealth services 1980 Hospital beds 1980

Maternityclinics 1980

Healthclinics 1980

Otherclinics 1980

State Presence (Years 0.00507 0.0559 0.0400 0.133 -0.111(0.00697) (0.294) (0.0847) (0.115) (0.102)

Observations 1553 1553 1553 1553 1553

Panel IV: Dependent variable:Principal componentsocial provisions 1980

Socialcenters 1980

Readingcenters 1980

Nutritionalcenters 1980 Orphanages 1980

State Presence (Years) 0.0106 -0.00711 0.00710 0.0117* 0.00362(0.0103) (0.0108) (0.0125) (0.00649) (0.00435)

Observations 1553 1553 1553 1553 1553

Panel V: Dependent variable:

Principal componentcommercial institutions

1980 Banks 1980Post

offices 1980Grain

silos 1980Gas

station 1980

State Presence (Years) 0.00435 -0.00536 -0.00320 0.00313 -0.00140(0.00704) (0.00396) (0.00375) (0.00217) (0.00156)

Observations 1553 1553 1553 1553 1553

Panel VI: Dependent variable:Principal component

markets 1980Agricultural

cooperatives 1980 Markets 1980Trade

centers 1980Commercialcenters 1980

State Presence (Years)) 0.0109 -0.00342 0.00122 -0.00522 -0.000296(0.0120) (0.00260) (0.00794) (0.00821) (0.00252)

Observations 1553 1553 1553 1553 1553

Panel VII: Dependent variable:Principal componentfiscal capacity 1980

Tax receivedper capita 1980

Trade tax receivedper capita 1980

Other taxes receivedper capita 1980

State Presence (Years) 0.00353 -0.0000777 0.0000415 -0.0000857(0.00678) (0.000226) (0.000130) (0.000193)

Observations 1288 1288 1288 1288

Panel VIII: Dependent variable:Principal component

public expenditure 1980public expenditure

per capita 1980

Administrativeexpenditure

per capita 1980Social expenditure

per capita 1980Economic expenditure

per capita 1980

State Presence (Years) 0.00316 -0.0000643 -0.0000823 0.0000556 -0.0000891(0.0120) (0.000243) (0.0000948) (0.0000460) (0.0000943)

Observations 1468 1459 1459 1459 1459

Travel distance to Nyanza along 1988 road Y Y Y Y YTravel distance to Kigali along 1988 road Y Y Y Y YDistance to country border Y Y Y Y Y

Number of clusters 50 50 50 50 50

Notes: All regressions are estimated using two stage least squares. The unit of observation is an administrative sector in 1994.Principal components in column (1) are the first principal component of columns (2)-(5). State presence (years) is the number ofyears a sector was under centralized rule before colonization in 1897. All regressions include fixed effects at the modern provincelevel. Parentheses report clustered standard errors at the level of the precolonial district. * indicates significance at the 10 percentlevel, ** at the 5 percent level, *** at the 1 percent level.

24

TABLE A21: ESTIMATES OF THE EFFECT OF STATE PRESENCE ON LOCAL GOVERNMENT

Dependent variable: Hospitals 1935 Hospitals 1960Hospital beds

1980

Number ofMissionary

Stations 1924

Number ofMissionary

Stations 1935Number of

Schools 1960

(1) (2) (3) (4) (5) (6)

Panel I: 2SLS estimates

State Presence (Years) 0.000435 0.000492 -0.00281 0.000628 0.000622 0.00439*(0.00161) (0.00131) (0.00263) (0.00254) (0.00149) (0.00255)

Panel II: Reduced form estimates

Distance to Nyanza 0.000561 0.000914 -0.00166 0.000664 0.00160 0.000502(0.00140) (0.00151) (0.00260) (0.00160) (0.00129) (0.00217)

R2 0.004 0.001 0.063 0.001 0.006 0.105

Travel distance to Nyanza along 1988 road Y Y Y Y Y YTravel distance to Kigali along 1988 road Y Y Y Y Y YDistance to country border Y Y Y Y Y Y

Number of clusters 50 50 50 50 50 50Observations 1553 1553 1553 1553 1553 1553

Notes: All regressions in Panel I are estimated using two stage least squares and distance to Nyanza as the instrument, all regressionsin panel II are estimated using OLS. The unit of observation is an administrative sector in 1994. All point estimates are standardized.State presence (years) is the number of years a sector was under centralized rule before colonization in 1897. Travel distance isdefined as the distance (in kilometers) between the centroid of a precolonial district and an endpoint along the Rwandan roadnetwork in 1988. All regressions include fixed effects at the modern province level. Parentheses report clustered standard errors atthe level of the precolonial district. * indicates significance at the 10 percent level, ** at the 5 percent level, *** at the 1 percent level.

TABLE A22: ALTERNATIVE EXPLANATIONS FOR VIOLENCE IN THE RWANDAN GENOCIDE

Dependent variable:Population growth

1978-1991

Unable to meetcaloric need of

population 1980Fraction of population

Tutsi 1991FractionalizationHutu Tutsi 1991

Radio Coverage inSector

Share with Reception

(1) (2) (3) (4) (5)

State Presence (Years) -0.0706 -0.0918** 0.0858 0.134 0.0642(0.0844) (0.0429) (0.101) (0.105) (0.0743)

Travel distance to Kigali along 1988 road Y Y Y Y YDistance to country border Y Y Y Y Y

Number of clusters 50 50 50 50 50Observations 1544 1553 1553 1553 1056R2 0.410 0.179 0.330 0.401 0.425

Notes: All regressions are estimated using OLS. The unit of observation is an administrative sector in 1994. All point estimates arestandardized. Population growth 1978-1991 is the percentage increase in population between the 1978 census and the 1991 census.Unable to meet caloric need of population is an indicator variable equal to one if a village was unable to supply enough food toprovide its inhabitants with a minimum number of calories in 1980. Fraction of population Tutsi 1991 is the total number of Tutsidivided by total population in the 1991 census. fractionalization Hutu Tutsi 1991 is the index of local Hutu Tutsi fractionalizationdescribed in the text. Radio Coverage in Sector, Share with Reception is the fraction of a sector that has RTLM radio coverage(Yanagizawa-Drott, 2014). State presence (years) is the number of years a sector was under centralized rule before colonization in1897. Travel distance is defined as the distance (in kilometers) between the centroid of a precolonial district and an endpoint alongthe Rwandan road network in 1988. All regressions include fixed effects at the modern province level. Parentheses report clusteredstandard errors at the level of the precolonial district. * indicates significance at the 10 percent level, ** at the 5 percent level, *** atthe 1 percent level.

25

TABLE A23: DETERMINANTS OF VIOLENCE AND MOBILIZATION IN THE RWANDAN GENOCIDE,WITH FURTHER ADDITIONAL CONTROLS

Dependent variable: Fraction of population prosecuted for: Fraction of Hutu prosecuted for:

Violence Violence Violence Organization Murder(1) (2) (3) (4) (5)

State Presence (Years) 0.176** 0.159*** 0.162** 0.0890 1.253**(0.0661) (0.0540) (0.0656) (0.138) (0.539)

Population growth 1978-1991 N Y N Y YUnable to meet caloric need of population 1980 N Y N Y YFraction of population Tutsi 1991 N Y N Y YFractionalization Hutu Tutsi 1991 N Y N Y YRadio Coverage in Sector, Share with Reception N N Y N N

Travel distance to Kigali along 1988 road network Y Y Y Y YDistance to country border Y Y Y Y YGeographical controls Y Y N Y Y

Mean of the dependent variable 7.95 7.95 7.95 1.40 7.75

Number of clusters 50 50 50 50 50Observations 1449 1441 1042 1449 1449R2 0.240 0.272 0.145 0.148 0.234

Notes: All regressions are estimated using OLS. The unit of observation is an administrative sector in 1994. Fraction of populationprosecuted for violence is the number of individuals prosecuted for being an organizer of genocide violence or for murder in theGacaca courts divided by total population in the 1991 census. Fraction of Hutu population prosecuted for organization (murder)is the number of individuals prosecuted for organizing genocide violence (murder) in the Gacaca courts divided by total Hutupopulation in the 1991 census. State presence (years) is the number of years a sector was under centralized rule before colonizationin 1897. Fraction of population Tutsi 1991 is the total number of Tutsi divided by total population in the 1991 census. Populationgrowth 1978-1991 is the percentage increase in population between the 1978 census and the 1991 census. Unable to meet caloricneed of population is an indicator variable equal to one if a village was unable to supply enough food to provide its inhabitants witha minimum number of calories in 1980. Fraction of population Tutsi 1991 is the total number of Tutsi divided by total populationin the 1991 census. Fractionalization Hutu Tutsi 1991 is the index of local Hutu Tutsi fractionalization described in the text. RadioCoverage in Sector, Share with Reception is the fraction of a sector that has RTLM radio coverage (Yanagizawa-Drott, 2014). Traveldistance is defined as the distance (in kilometers) between the centroid of a precolonial district and an endpoint along the Rwandanroad network in 1988. Geographical controls include terrain elevation, terrain slope, the geographical suitability for growing bananaand the local prevalence of the savannah tsetse fly. All regressions include fixed effects at the modern province level. Parenthesesreport clustered standard errors at the level of the precolonial district. * indicates significance at the 10 percent level, ** at the 5percent level, *** at the 1 percent level.

26

2 Reconstructing the expansion of the Nyiginya kingdom

This section contains further background to the reconstruction of the history of the expansion of the

Nyiginya kingdom. I compute the variable of interest in this paper, state presence in years, for each

precolonial district by subtracting its year of annexation to the Nyiginya kingdom from 1897, the year

Rwanda was colonized and became part of German East Africa. This paragraph describes the procedure

to reconstruct the expansion.

Step 1: Mapping precolonial districts. Precolonial districts are reported by Vansina (2004). He reports

names and approximate locations for districts in precolonial Rwanda. He also provides maps of the ex-

tent of the Nyiginya kingdom in 1700, 1796, 1867 and 1897. He does not, however, report boundaries for

the historical districts. To assign boundaries to each district, I use the fact that Belgian colonial chiefdoms

(‘chefferies’ in French), were based on precolonial administrative units (Paternostre de La Mairieu, 1972).

37 of 50 district directly match with the chiefdoms and I use the chiefdom boundaries as the precolonial

district boundaries. For all other precolonial districts I geographically merge the approximate locations to

colonial chiefdoms using standard georeferencing procedures available in ArcGIS. Figure A1 reports the

resulting map.

Step 2: Identifying the year of annexation for each precolonial district. In terms of the availability of data

on the year of annexation, there are three categories of districts. First, districts whose year of annexation

into the Nyiginya kingdom is explicitly mentioned by Vansina (2004). Second, districts for which the in-

corporation is linked to a specific year of the reign of a Rwandan king. Third, those districts not explicitly

mentioned by Vansina. For the first group I directly use the year of annexation Vansina reports. For the

second group, I use the list of royal succession in table A24 to assign a year to the annexation of these

districts. If a district was annexed ‘early’ in the reign of a King, I code this as five years after his accession

to the throne. For the third group, I use Vansina’s cross sectional maps of the expansion of the kingdom

up to 1700, 1796, 1867 and 1897. I consider a district annexed in the year of the earliest cross sectional map

it appears on which it appears.

Table A25 reports results of this coding scheme. It reports colonial chiefdom names (for which I know

geographical boundaries) in column (1). Column (2) gives the corresponding district from Vansina. Col-

umn (3) gives the latest year at which the district was incorporated from the cross sectional maps. Column

27

(4) contains a reference to the page number in Vansina (2004) that I used to identify the year of annexa-

tion in column (6). Column (5) has the name of the reigning king at the time of the annexation of the

district. Column (6) gives the year of annexation used to compute the length of state presence. This year

is obtained using the procedure in the previous paragraph, and is either taken from the page number in

column (4) or from column (3). Column (7), finally, reports the map in Vansina from which I take the

precolonial district name. The numbers refer to the maps printed before the introduction that display the

kingdom as a central, a western and an eastern part. Map 1 refers to the map of central Rwanda, map 2

refers to the map of eastern Rwanda and map 3 refers to the map of western Rwanda.

28

TABLE A24: LIST OF KINGS

Name of king Reign

RUGANZA Ndori around 1650Some unknown kingsMIBAMBWE Gisanura after 1700-1735YUHI Mazimpaka 1735-1766KAREMERA Rwaka 1766-1770CYIRIMA Rujugira 1770-1786KIGERI Ndabarasa 1786-1796MIBAMBWE Sentaboyo 1796-1801YUHI Gahindiro 1801-1845MUTARA Rwogera 1845-1867KIGERI Rwabugiri 1867-1895MIBABMWE Rutarundwa 1895-1897YUHI Musinga 1897-1931

Notes: This table contains a list of royal succession in the Nyiginyakingdom. The names printed in lower case are the actual namesof the kings. The names printed in upper case are the sym-bolic names, which are often used interchangeably with the actualnames. Source: Vansina (2004).

29

Table A25: Precolonial Districts

Colonial District Precolonial District Cross section Page King Year of annexation Map(1) (2) (3) (4) (5) (6) (7)

Biru Biru 1867 153 1867 3Buberuka Buberuka 1897 178 1897 2

Rukiga Buberuka 1897 178 1897 2Budaha Budaha 1796 114 Mazimpaka 1766 1

Bufundu Bufundu 1796 116 Rujugira 1775 1Buganza Buganza 1796 118 Rujugira 1771 2Bugoyi Bugoyi 1867 159 Rujugira 1786 3

Bugusera Bugusera 1867 124 Sentabyo 1799 1Buhanga/Ndara Buhanga and Ndara 1720 50 Gisanura 1720 1

Buhema Buhoma 1867 153 1867 3Bukonya Bukonya 1796 125 Sentabyo 1796 1Cyesha Bukunza 1897

Bumbogo Bumbogo 1700 46 Ndori 1700 1Nyaruguru Bungwe 1796 124 1796 1

Bunyambiriri Bunyambiriri 1796 116 Rujugira 1766 1Ndiza Burembo 1796 114 Gisanura 1735 1

Bumbogo Buriza 1796 46 Ndori 1700 2Busanza Busanza 1720 50 Gisanura 1720 1Bushiru Bushiru 1867 153 1867 3Buliza Busigi 1700 46 Ndori 1700 1

Bukunzi Busozo 1897 178 1897 3Busozo Busozo 1897 178 1897 3Buyaga Buyaga 1897 178 1897 2Buyenzi Buyenzi 1897 157 Rwabugiri 1872 1

Bwanacyambe Bwanacyambe 1796 50 Gisanura 1720 1Bashumba-Nyakare Bwanamukari 1796 124 1796 1

Mwejuru Bwanamukari 1796 124 1796 1Rusenyi Bwishaza 1867 160 Rwogera 1845 3Kingogo Cyingogo 1867 124 1796 3Gihunya Gisaka 1867 155 Rwogera 1850 3Impara Impara 1867 153 1867 3

Kabagari Kabagari 1700 49 Ndori 1700 1Kanage Kanage 1867 153 Rwogera 1867 1Kibari Kibari 1796 125 Sentabyo 1796 1Cyesha Kinyaga 1867 160 Sentabyo 1801 3

Marangara Marangara 1796 124 1796 1Mayaga Mayaga 1700 49 Ndori 1700 1Migongo Migongo 1867 153 1867 2Mubari Mubari 1897 178 1897 2

Bugarura Murera 1867 159 Gahindiro 1845 1Mulera Murera 1867 159 Gahindiro 1845 1

Buganza North Mutara 1897 1895 Rwabugiri 1895Mutara Mutara 1897 1895 Rwabugiri 1895Ndiza Ndiza 1796 114 Gisanura 1735 1

Ndorwa Ndorwa 1897 167 Rwabugiri 1868Nduga Nduga 1700 49 Ndori 1700 1

Nyantango Nyantango 1796 114 Mazimpaka 1766 1Rukaryi Rukaryi 1796 124 1796 2Rukoma Rukoma 1700 49 Ndori 1700 1Bwishaza Rusenyi 1867 153 1867 3Rwankeri Rwankeri 1867 153 1867 3Bukamba Yomba/Jomba 1867 159 Gahindiro 1845 3

30

FIGURE A1: MAP OF PRECOLONIAL DISTRICTS IN RWANDA, SHADED BY THE LENGTH OF STATE PRES-ENCE

Bugoyi Bushiru

Kanage

Kingogo

Budaha

Nyantango

Rusenyi

Cyesha

Bwishaza

Bunyambiriri

Biru

Impara

BukunziBusozo

Buyenzi

Bufundu

Nyaruguru

Mwejuru

Bashumba−Nyakare

Buhanga/Ndara

Busanza

Mayaga

Kabagari

Marangara

Nduga

Rukoma

Ndiza

Bukonya

Rwankeri

Buhema

Bukamba

Bugarura

Kibari

Bumbogo Buliza

Rukiga

Buberuka

NdorwaMutara

Buyaga

Buganza

Bwanacyambe

Rukaryi

Mirenge

Gihunya

Migongo

Bugusera

Mubari

Mulera

Buganza North

1700175018001850

Legend: Year of annexation

Map of precolonial districts in Rwanda, shaded by the length of state presence. The legend indicates the year of annexation to theNyiginya kingdom. District names correspond to in table A26.

31

FIGURE A2: MAPS OF THE EXPANSION OF THE NYIGINYA KINGDOM

1700 1720 1735

1740 1766 1771

1786 1796 1799

1801 1845 1850

1867 1868 1872

1895 1897

Maps of the expansion of the Nyiginya kingdom, by year of expansion. Districts names are given in figure A1.

32

3 Data Sources

Variable Source Comment

Dependent Variables

Genocide violence data,number of Gacaca courts

Gacaca courts report athttp://www.inkiko-gacaca.gov.rw/En/EnLaw.htm

Website no longer operational as of March 1st,2013.

Other violence data UCDP UCDP project (Sundbergand Melander, 2013)

I use the following dataset: UCDP GED ConflictPolygons dataset version 1.1-2011 available athttp://ucdp.uu.se/downloads/, accessedFebruary 1st, 2016.

Victims 2004 Guichaoua and Degni-Segui(2010)

These data vary at the commune level (n=142)

Victims 1996 Rwanda: Ministère del’Enseignement Supérieur,de la Recherche Scientifiqueet de la Culture

Available at http://rwanda.free.fr/docs1_c.htm,accessed May 6th, 2016. These data vary at thecommune level (n=142)

State Presence

State presence Vansina (2004) Procedure described above

Distance variables

Location of Nyanza Historical Nyanza is in Nyabisindu village, closeto modern Nyanza

Distance to Nyanza, Kigali Computed in ArcGIS

Cost distance to Nyanza(days)

Computed in ArcGIS using an elevation grid, seebelow, and Tobler’s hiking formula (Tobler, 1993).I rescaled Tobler’s function by 3/5 to account forunpaved terrain. A day is assumed to last twelvehours.

Cost distance to Nyanza –Özak (days)

Provided by Ömer Özak, using the methodologydeveloped in Özak (2010, 2013). A day is assumedto last twelve hours.

Cost distance to nearestNyiginya army location(days)

Army locations fromKagame (1963)

Computed in ArcGIS using an elevation grid, seebelow, and Tobler’s hiking formula (Tobler, 1993).I rescaled Tobler’s function by 3/5 to account forunpaved terrain. A day is assumed to last twelvehours.

Travel distance to Nyanza,Kigali along 1988 road

1988 road network fromRhineland-Palatinate (1988)

Distance along road network computed inArcGIS, assuming route starts at closest point todestination point on the road network.

Distance to country border,nearest river, Addis Abeba

Computed in ArcGIS

33

Variable Source Comment

Other variables

Population 2002 The 2002 Rwandan Census(Rwanda. National Census Service,1992)

Available at www.statistics.gov.rw andIPUMS international athttps://international.ipums.org/international,accessed February 2015

Population 1991,number of Hutu andTutsi 1991

The 1991 Rwandan Census(Ministère du Plan, 1992)

Available through at www.statistics.gov.rwand IPUMS international athttps://international.ipums.org/international,accessed November 2013.Fractionalization between Hutu and Tutsiis computed using this data. Populationgrowth 1978-1991 is also computed usingthis data. These data vary at the communelevel (n=142)

Population 1978 The 1978 Rwandan Census(Ministère du Plan, 1982)

Population growth 1978-1991 is computedusing this data. These data vary at thecommune level (n=142)

Unable to meetcaloric need ofpopulation 1980

Lamoure (1980) These data vary at the commune level(n=142)

Public goods andfiscal capacity in1980

Lamoure (1980) These data vary at the commune level(n=142)

Local leaderopposition

Guichaoua and Degni-Segui (2010) These data vary at the commune level(n=142)

Radio ownership1991

The 1991 Rwandan Census(Ministère du Plan, 1992)

Available through at www.statistics.gov.rwand IPUMS international athttps://international.ipums.org/international,accessed November 2013. These data varyat the commune level (n=142)

RTLM radiocoverage

Made available by DavidYanagizawa-Drott (Yanagizawa-Drott,2014)

Word Values Survey2007 and 2012 wavesfor Rwanda

Available at www.worldvaluessurvey.org,accessed May 2014. Full citation for 2012:WORLD VALUES SURVEY Wave 62010-2014 OFFICIAL AGGREGATEv.20150418. World Values SurveyAssociation(www.worldvaluessurvey.org). AggregateFile Producer: Asep/JDS, Madrid SPAIN.Full citation for 2007: WORLD VALUESSURVEY Wave 5 2005-2008 OFFICIALAGGREGATE v.20140429. World ValuesSurvey Association(www.worldvaluessurvey.org). AggregateFile Producer: Asep/JDS, Madrid SPAIN.

34

Variable Source Comment

Other variables continued

Civil participationvariables

Rwanda threshold programdatabase

Available at https://www.mathematica-mpr.com/our-publications-and-findings/projects/rwanda-democracy-and-governancel, accessed October2016

Household violencevariables

Demographic and Health Survey2014

Available at www.dhsprogram.com,accessed August 2016

Colonial period variables

1924 colonialvariables

1924 missionary map available athttp://scholar.harvard.edu/nunn

1935 colonialvariables

Parliament of Belgium (1935)

1960 colonialvariables

Government of Belgium (1960)

Geographical variables

Elevation, slope CGIAR consortium athttp://srtm.csi.cgiar.org/

Slope computed using elevation grid fromCGIAR.

Inland rivers andwater bodies

Digital Chart of the World availablethrough www.diva-gis.org

Distances computed in ArcGIS

Suitability forBananas

FAO athttp://webarchive.iiasa.ac.at/Research/LUC/GAEZv3.0/

I used the rain-fed, low intensity, baselineperiod settings

Suitability for Tsetse FAO athttp://www.fao.org/ag/againfo/programmes/en/paat/home.html

I use the Morsitans (Savannah) Tsetsesubspecies suitability data, since this is theonly species that is endemic to Rwanda.

Experimental data

Data generated infieldwork

All data generated in the fieldwork, suchas demographic observables, orcompliance with government demandswere observed in fieldwork in Summerand Fall 2014. The main text describes thisfieldwork and the procedures used.

35

References

André, C. and J.-P. Platteau (1998). Land Relations under Unbearable Stress: Rwanda caught in the

Malthusian Trap. Journal of Economic Behavior & Organization 34(1), 1–47.

Ashraf, Q. and O. Galor (2013). The “Out of Africa” hypothesis, human genetic diversity, and comparative

economic development. The American Economic Review 103(1), 1–46.

Clay, D. C., T. Kampayana, and J. Kayitsinga (1997). Inequality and the Emergence of Non-farm Employ-

ment in Rwanda. In N. E. Johnson and C.-l. Wang (Eds.), Changing rural social systems: adaptation and

survival, pp. 93–130. East Lansing: Michigan State University Press.

Government of Belgium (1960). Ruanda-Urundi: Economy 1. Brussels: Office de l’information et des

relations publiques pour le Congo Belge et le Ruanda-Urundi.

Guichaoua, A. and R. Degni-Segui (2010). Rwanda, de la Guerre au Génocide: Les Politiques Criminelles au

Rwanda (1990-1994). la Découverte.

Kagame, A. (1963). Les milices du Rwanda précolonial. Brussels: Académie royale des sciences coloniales.

Lamoure, G. (1980). Rural atlas of Rwanda. République Rwandaise: Ministère des travaux publics et de

l’équipement.

Ministère du Plan (1982). Recensement Général de la Population et de l’Habitat 1978. Kigali: Rwanda:

République Rwandaise.

Ministère du Plan (1992). Recensement Général de la Population et de l’Habitat 1991. Kigali: Rwanda:

République Rwandaise.

Özak, Ö. (2010). The Voyage of Homo-œconomicus: some economic measures of distance. Working paper,

Southern Methodist University.

Özak, Ö. (2013). Distance to the technological frontier and economic development. Working paper, Southern

Methodist University.

Parliament of Belgium (1935). Rapport sur du l’administration Belge du Ruanda-Urundi. Bruxelles.

Paternostre de La Mairieu, B. (1972). Le Rwanda: Son effort de développement, antécédents historiques

et conquêtes de la révolution Rwandaise.

36

Prunier, G. (1995). The Rwanda crisis: History of a Genocide. New York: Columbia University Press.

Rhineland-Palatinate (1988). République rwandaise, réseau de transport interurbain 1:500 000. Ministerium

des Innern und fuer Sport. Abteilung Vermessungs- und Katasterwesen.

Rotter, J. B. (1966). Generalized expectancies for internal versus external control of reinforcement. Psycho-

logical monographs: General and applied 80(1), 1.

Rwanda. National Census Service (1992). Recensement Général de la Population et de l’Habitat 2002. Kigali :

République du Rwanda, Ministère des finances et de la planification économique.

Sundberg, R. and E. Melander (2013). Introducing the UCDP georeferenced event dataset. Journal of Peace

Research 50(4), 523–532.

Tobler, W. (1993). Non-isotropic geographic modeling. Three presentations on geographic analysis and model-

ing. Santa Barbara: National Center for Geographic Information and Analysis, University of California.

Vansina, J. (2004). Antecedents to modern Rwanda: the Nyiginya kingdom. Madison: University of Wisconsin

Press.

Verwimp, P. (2005). An Economic Profile of Peasant Perpetrators of Genocide: Micro-level Evidence from

Rwanda. Journal of Development Economics 77(2), 297–323.

Yanagizawa-Drott, D. (2014). Propaganda and Conflict: Evidence from the Rwandan Genocide. The

Quarterly Journal of Economics 129(4), 1947–1994.

37