Author-card of document number 24266

Num
24266
Date
Saturday November 21, 2009
Ymd
Author
File
Size
1621019
Pages
51
Urlorg
Title
Propaganda and Conflict: Theory and Evidence From the Rwandan Genocide
Keyword
Keyword
Type
Article de revue
Language
EN
Citation
Propaganda and Con ict:
Theory and Evidence From the Rwandan Genocide
David Yanagizawa
IIES, Stockholm University
November 21, 2009

Abstract
This paper investigates the impact of propaganda on participation in violent conict. I examine the e ects of the infamous "hate radio" station Radio RTLM that
called for the extermination of the Tutsi ethnic minority population before and during the 1994 Rwanda Genocide. I develop a model of participation in ethnic violence
where radio broadcasts a noisy public signal about the value of violence. I then test
the model's predictions using a nation-wide village-level dataset on radio coverage and
prosecutions for genocide violence. To identify causal e ects, I exploit arguably exogenous variation in radio coverage generated by hills in the line-of-sight between radio
transmitters and villages. Consistent with the model under strategic complements in
violence, I nd that Radio RTLM increased participation in violence, that the e ects
were decreasing in ethnic polarization, highly non-linear in radio coverage, and decreasing in literacy rates. Finally, the estimated e ects are substantial. Complete village
radio coverage increased violence by 65 to 77 percent, and a simple counter-factual
calculation suggests that approximately 9 percent of the genocide, corresponding to at
least 45 000 Tutsi deaths, can be explained by the radio station.
[JOB MARKET PAPER] Contact: yana@iies.su.se. I thank Robert H. Bates, Ethan Kaplan, Masayuki
Kudamatsu, Rocco Macchiavello, Nancy Qian, Torsten Persson, David Stromberg and, especially Jakob
Svensson for their comments; the participants at Harvard Development Lunch, Econometric Society European Winter Meeting 09, the CIFAR IOG 09, and the EUDN Workshop at Oxford. I would also like to thank
Giovanni Zambotti at the Center for Geographic Analysis at Harvard University for ArcGIS assistance. All
mistakes are my own.

1

1

Introduction
"The radio encouraged people to participate because it said `the enemy is the
Tutsi'. If the radio had not declared things, people would not have gone into the
attacks."
-Genocide perpetrator, interviewed by Straus (2007)
Among all historical episodes of civil con ict the 1994 Rwanda Genocide is an extraor-

dinary event. During a period of only three months, a nation-wide extermination campaign
led by the Rwandan government against the Tutsi ethnic minority population resulted in at
least 500 000 Tutsi civilian deaths and a reduction by approximately 75% of the country's
Tutsi population (des Forges, 1999).1 In addition to the violence organized by the army
and militias, the high intensity killings were achieved by mass participation by hundreds of
thousand ethnic majority Hutu citizens using their machetes and clubs (des Forges, 1999;
Straus, 2004; Verwimp, 2006). Given the large-scale participation and the human lives lost,
understanding the determinants of participation in the genocide is of great importance. The
principal aim of this paper is to estimate the impact of one factor widely believed (BBC,
2003; Thompson, 2007) to have played a signi cant role in the genocide: propaganda spread
by the infamous "hate radio" station Radio Television Libre des Mille Collines (RTLM).
In order to understand the determinants of participation in violence, and the mechanisms
through which propaganda can a ect participation, the paper rst sets up a simple model
of propaganda and participation in ethnic violence. The model adopts the global games
framework (Carlsson and van Damme, 1993; Morris and Shin, 1998; 2005) and considers
a situation where individuals face some uncertainty about the value of con ict, but may
receive a noisy public signal about the value through the radio. The key insight of the
model is that propaganda, de ned as radio broadcasts signalling that the value of con ict
is high, will a ect participation through two mechanisms. First, by increasing the expected
value of con ict, independent of how many others that participate. Second, and potentially
more importantly, by changing the expectations individuals hold about whether others will
participate. If there are strategic complements in violence, the second mechanism implies
1

There was also a signi cant amount of moderate Hutus that were killed. For discussions on the
death tolls, see des Forges (1999), Verpoorten (2005), as well as Davenport and Stam's analysis at
www.genodynamics.com (Available 2009-11-05).

2

that propaganda will function as a coordination device and lead to large-scale increases in
participation if a su ciently large number of people receives the propaganda.
The predictions of the model are taken to a unique nation-wide village-level dataset that
combines data from several sources. First, as a proxy for participation rate, the paper uses
data on prosecution rates for violence during the genocide, provided by Rwanda's National
Service of Gacaca Jurisdictions. Second, it uses information on locations and technical
speci cations of Radio RTLM transmitters, and produces a nation-wide radio coverage map
at a 90 meter cell resolution. Using a digital map of village boundaries, the radio coverage
of each village is then calculated. Additional data on village characteristics is collected from
the 1991 Rwanda Census and the Africover database. The matched dataset contains data
on 1105 villages.2
The identi cation strategy exploits arguably exogenous variation generated by Rwanda's
highly varying topography consisting of hills and valleys. Using local within-commune village
variation in radio coverage, the variation exploited will be due to whether there happens to
be hills in the line-of-sight between radio transmitters and villages.3
Radio RTLM broadcasts had a substantial e ect on violence. The estimates imply that
going from no to full village radio coverage, increased civilian violence 65 percent and organized violence by 77 percent. Furthermore, the e ects are entirely driven by villages where
the Hutu ethnic majority was large relative to the Tutsi ethnic minority, and they are highly
nonlinear in the degree of radio coverage as there is a sharp increase in violence when the radio coverage is su ciently high. These results are consistent with the model under strategic
complements, and suggest that the broadcasts were most e ective when people knew that
many other village members were also listening to the same broadcasts. The propaganda,
therefore, appears to have functioned as a coordination device.
Moreover, and consistent with the model, the paper nds evidence that the ability to
access independent information can mitigate the propaganda e ects. In fact, there is no
e ect of radio coverage in villages in the upper literacy rate tertile, whereas the e ects
are large in villages in the lower literacy rate tertile. The results therefore suggest that the
2

The villages are formally called "administrative sectors". The term village is used for simplicity, highlighting that the units are relatively small. The median village area is 10.6 square kilometers.
3
The use of this method to examine media e ects in the social sciences is not new. Olken (2009) employs
a closely related but not identical approach in his study of the e ects of television and radio on social capital
in Indonesia.

3

propaganda caused more violence because there was a lack of alternative information sources
that could contest the content broadcasted by Radio RTLM.
To assess the extent to which the propaganda can explain the degree of violence in the
genocide, the paper presents a simple counter-factual calculation. The results suggest that
Radio RTLM caused approximately 9% of the genocidal violence, which corresponds to at
least 45 000 Tutsi deaths.4 Therefore, Radio RTLM was a quantitatively important causal
factor in the genocide.
This project is related and adds to several strands of literature. First, it contributes to
the literature on the determinants of the genocide (Verwimp, 2005, 2006; Straus, 2007), by
presenting novel evidence on the causal e ects of Radio RTLM.
Second, the Rwanda genocide may be extraordinarily grim, but it forms part of the wider
phenomenon of civil war and con ict. Since 1960, one third of all nations has experienced
civil war and one fth has seen episodes of more than 10 years of civil war (for an overview,
see Blattman and Miguel, 2009). Cross-country studies (Collier and Hoe er 1998, 2004;
Fearon and Laitin 2003; Miguel et al. 2004; Besley and Persson 2008) have focused on the
macro determinants of con ict onset, incidence and duration. There is also a small but
growing literature has used within-country regional data to identify factors determining the
intensity of civil violence (e.g., Murshed and Gates, 2005; Dube and Vargas ,2007; Do and
Iyer, 2007; Jha 2008). By presenting robust micro evidence on the role of information and
beliefs, this paper adds an important piece to the understanding of why people participate
in civil war and con ict, as well as how ethnic mobilization is achieved (e.g., Bates, 1986;
Fearon and Laitin, 1996). In their overview of the literature, Blattman and Miguel (2009)
conclude that the existing theory is incomplete. They argue that although the individual
participation choice should be a natural starting point for the analysis of civil con ict,
the literature lacks an understanding of the roots of individual participation. The workhorse model used to study determinants of group violence (including ethnic) is the contest
model (Haavelmo, 1954; Hirshleifer, 1988). By assuming unitary groups, the contest model
therefore typically ignores the participation problem at the individual level. In addition,
Blattman and Miguel argue that theories seldom specify the empirical predictions that can
test between competing accounts, and there is a lack of studies with convincing econometric
4

This is substantial considering that the radio signal was only receivable in about 19 percent of the
country.

4

identi cation. The model proposed in this paper analyzes the individual participation choice,
and delivers predictions that allow the data to disentangle whether participation in ethnic
violence is subject to strategic complements or strategic substitutes.5 A contribution of the
paper, in addition to estimating the causal e ects of Radio RTLM on participation in the
genocide, is therefore to shed light on the mechanisms driving ethnic violence. Speci cally,
the empirical results are consistent with strategic complements in violence, and inconsistent
with strategic substitutes. To the best of the author's knowledge, this is a novel nding.
Finally, the paper adds to the literature on media e ects (for an overview, see Della Vigna
and Gentzkow, 2009). Theoretically, self-interested politicians may supply biased mass media
in order to reduce the likelihood of regime change (Edmond, 2009) as well as to induce hatred
(Glaeser, 2005). The empirical e ects of mass media on political behavior have been studied
at least since Lazarfeld et al. (1954). A recent literature has found signi cant e ects.
This includes e ects on voting behavior (Gentzkow, 2006; Della Vigna and Kaplan, 2007;
Chang and Knight, 2008; Enikolopov et al., 2008; Gerber et al., 2009); accountability and
policy (Besley and Burgess, 2002; Str•omberg, 2004; Eisensee and Str•omberg, 2005); political
knowledge and beliefs (Gentzkow and Shapiro, 2004; Snyder and Str•omberg, 2008); and
social capital (Paluck, 2009; Olken, 2009). This paper adds to the literature by presenting
novel evidence showing that mass media can persuade individuals into what is arguably the
most extreme political acts of them all: killing members the political opposition.
Below, section 2 provides the background to the genocide and Radio RTLM; section 3
presents the model and derives empirical predictions; section 4 explains the data and the
empirical strategy; section 5 presents the results; and section 6 concludes the paper.

2

Background

This section provides a brief background in order to understand the pre-existing political
tensions leading up to the genocide, as well as the structure and content of Radio RTLM
broadcasts.
5

Under strategic interactions and complete information, multiple equilibria are typically present. However, under incomplete information (Carlsson and van Damme, 1993; Morris and Shin, 1998; 2005), there is
a unique equilibrium that allows one to derive testable predictions.

5

2.1

Political and ethnic tensions

After World War I, Belgium took control of Rwanda (previously a German colony) on a
mandate by the League of Nations. The Belgian rule reinforced pre-existing ethnic cleavages
by a range of policies favoring the ethnic minority Tutsi group (Prunier, 1995). However, with
the Hutu Revolution" and the independence from Belgium in 1962, there was a complete
reversal of power. After 1962, Rwanda became a Hutu-dominated one-party state.
In connection with the independence, there were several episodes of ethnic violence between the two ethnic groups that led to several hundreds of thousand ethnic Tutsi refugees
in neighboring countries (Prunier, 1995). A period of relative stability followed but in 1973,
there was more violence as ethnic clashes between Hutus and Tutsis in Burundi spilled over
into Rwanda. The unrest eventually led to the young Hutu military leader Juvenal Habyarimana seizing power in a coup in 1973.
In October 1990, a rebel army invaded Rwanda from Uganda. The rebels, of the Rwandan
Patriotic Front (RPF), represented the refugees that had ed during the Hutu Revolution
and demanded an end to the ethnically unbalanced policies. Internationally, they presented
themselves as a democratic multi-ethnic movement trying to overthrow a corrupt regime.6
In April 1992, a transitional multi-party government was formed. After periods of negotiations and unrest, a peace agreement was nally signed in Arusha in August 1993. With
sparse resources and a weak mandate, United Nations' peace-keeping forces were to facilitate
the installation of the transitional government. After periods of violence, unrest, and postponed installations, the Hutu president Habyarimana was assassinated when his jet was shot
down on April 6th 1994. Within days, extremists within Hutu-dominated political parties
managed to take over key positions of government, and an ethnic cleansing campaign spread
throughout the country shortly thereafter.
The branches of government took an active role in the killings, from Presidential Guards,
the regular army FAR, national gendarmes, via the civil administration down to the mobilization and supply of resources to the Interahamwe and Impuzamugambi militias (Prunier,
1995). In addition, there was large-scale civilian participation as several hundreds of thousands citizens participated in the attacks (Straus, 2004).
6

The rebel army of about four thousand well-trained troops mainly consisted of second-generation Rwandan refugees. They had gained military experience from Uganda's National Resistance Army which seized
power in Uganda in 1986.

6

The genocide ended in late July 1994 when the Tutsi RPF rebels defeated the Rwandan
army and militia groups, and managed to seize the capital Kigali. At that point, at least
500 000 Tutsis had been killed (des Forges, 1999).

2.2

Media and Radio RTLM

Radio RTLM started broadcasting in July 1993. The station was set up as a private company by a group of Hutu politicians, but with strong support from President Habyarimana
(Thompson, 2007). The broadcasts continued throughout the genocide, and did not end until
RPF rebels manage to take control of the country in mid-July.
Two radio transmitters were installed. One 100 watt transmitter was placed in Kigali,
the capital, and another 1000 watt transmitter was placed on Mount Muhe, one of the country's highest mountains. Compared to the only other national radio station in the country,
government owned Radio Rwanda, RTLM quickly became popular by airing western-style
talk shows and playing the latest music, especially popular Congolese songs.7
Importantly, the radio station called for the extermination of the Tutsi ethnic group and
claimed that preemptive violence against the Tutsi population was a necessary response of
"self-defense" (ICTR, 2003; Thompson, 2007).8 In her study of RTLM airtime content, Kimani (2007) reports that the most common in ammatory statements consisted of 1) Reports
of Tutsi RPF rebel atrocities (33%); 2) Allegations that Tutsis in the region were involved
in the war or a conspiracy (24%); and 3) Allegations that RPF wanted power and control
over Hutus (16%).
Although the radio station systematically called upon Hutus to be aware of Tutsi plots
and forthcoming attacks, it is still unclear to what extent Hutu citizens believed in the RTLM
broadcasts and viewed them as informative about the ongoing con ict between Hutus and
Tutsis, and to what extent citizens discredited the broadcasts as being biased. However, the
fact that there was a demand for the broadcasts suggests that citizens at least viewed the
broadcasts as bringing important information. For example, Des Forges described the high
demand of RTLM as "people listened to the radio all the time, and people who didn't have
7

There was also a station owned by the Tutsi RPF rebels, Radio Muhabura, that broadcast into Rwandan
territories from Uganda.
8
A common de nition of propaganda is "the spreading of ideas, information, or rumor for the purpose
of helping or injuring an institution, a cause, or a person".
(Available 2009-11-15)

7

radios went to someone else's house to listen to the radio. I remember one witness describing
how in part of Rwanda, it was di cult to receive RTLM, and so he had to climb up on the
roof of his house in order to get a clear signal, and he would stand up there on the roof of
his house with his radio to his ear listening to it".9
Furthermore, as alternative information sources were limited in the rural areas, it was
arguably di cult to verify the content in the broadcasts. Alternative media sources did
exist. In particular, the number of independent newspapers, including political opposition
newspapers, at the time of the genocide was between 30 to 60 (Alexis and Mpambara, 2003;
Higiro, 2005). However, the circulation and readership of these newspapers in the rural
areas, however, was naturally limited due to relatively low literacy rates in the country.10
Therefore, the radio became the sole source of news for most people (des Forges, 1999).

3

A Model of Ethnic Violence

Given their content, it is quite clear that one of the main motives for the RTLM broadcasts
was to a ect the beliefs among the Hutu population that a nondiscriminatory, preemptive,
attack against conspiring Tutsis was the appropriate course of action. We now turn to a
simple model that allows us to analyze how these broadcast might have a ected the beliefs
among the Hutu population, and how it could have in uenced the level of violence in Rwandan villages. Albeit relatively simple, the model sheds light on some interesting channels
through which propaganda might translate into violence. Most importantly, the model delivers a set of testable predictions that will be taken to the data in the subsequent sections
of the paper.
We proceed in several steps. First, we explain the basic setup and second, we nd the
equilibrium and show how it can be a ected by propaganda. Third, we present the empirical
predictions that will be taken to the data.

3.1

Basic setup

Consider a village with a continuum of individuals, where each individual is a member of
one of two ethnic groups, ethnic majority group H and ethnic minority group T . The
9
10

Interview with Alison des Forges, available (2009-11-16) at
The literacy rate was 66 percent (des Forges, 1999).

8

population size of group H is normalized to 1, and the size of group T in the village is t: The
analysis focuses on the discrete decision by group H members to participate in an attack
against minority group T in the village. Strategic behavior by minority group members is
not studied in order to keep things simple. Therefore, in what follows we exclusively focus
on the behavior of group H members.
The payo from participating in the attack depends on some fundamental value, , which
is possibly negative. We may consider

as the net bene t that depends on a range of factors

independently of how many other group H members participate in the attack, as well as the
size of group T . For example, factors determining

could be the amount of wealth of group

H, the opportunity cost of attacking group T , or the value associated with being the rst
side to attack the opposite group
In addition to the fundamental value, we allow the payo from participating in violence
to exhibit strategic complements or strategic substitutes. Under strategic complements, the
payo depends positively on how many other members of group H that participate in the
attack, h: Under strategic substitutes, the payo depends negatively on how many other
members that participate. On the one hand, violence is a dangerous and costly activity,
and there are good reasons to think that there exists strategic complements in violence. For
example, the larger is the group attacking, the smaller is the likelihood of being injured, or the
shorter is the duration of ghting required for success. On the other hand, if the appropriable
resources are limited and the participating members

ght over the same resources, then

there would be less appropriable resources per participating member the more members that
participate. Under such conditions, there could be strategic substitutes in violence.
Similarly, we allow the payo

from having more members participate in the attack h

to depend on the (relative) size of the defending group, t. Speci cally, to get a convenient
formalization, let the payo structure be the following

u =

8
<
:

+

h
t

0

if the member participates in the attack
if the member does not participate in the attack

If there are strategic complements (substitutes) in violence,
are no strategic interactions,

> 0 (

< 0). When there

= 0. We are interested in the equilibrium number of ethnic

majority members participating in the attack, h; and how h can be a ected by propaganda.

9

3.2

Information and beliefs

In reality, participating in con ict is a risky project. We formalize this by assuming that
members face uncertainty about the fundamental value of participating in violence, such
that there is incomplete information about : It is reasonable to believe that

cannot be

known with complete certainty in most cases of violent con ict. In this section, we describe
how members form their beliefs about .
Following the literature on global games, members do not observe

but receive informa-

tion about the value that allows them to form beliefs. We make the standard assumption
that members have a di use prior distribution of

on the real line. Each member i observes

an independent private signal xi = +"i ; where "i is independently and normally distributed
with mean zero and variance

2
x:

We can consider xi as all the independent private infor-

mation a member has from di erent sources that are relevant for the fundamental value of
con ict. Furthermore, we can consider a lower

x

representing having access to multiple

sources of information, or access to information sources of high quality.
Furthermore, the radio broadcasts a signal p about the value of . A fraction r of the
village population has radio coverage. Having radio coverage implies receiving the signal
p. For simplicity, we do not consider strategic behavior on behalf of whomever sends out
the radio signal. Instead, agents view the signal p as informative about the underlying
fundamental value of con ict, . The signal has the structure p = + b. To keep the analysis
simple, we assume that b is exogenous, unobservable, and distributed normally with mean
zero and variance

2 11
p:

Key to the model is that the radio signal is a public signal among

members with radio, i.e. there is common knowledge about the radio signal among majority
members with radio. Therefore, a member with radio will not only use the signal to update
his belief about ; he also knows that a fraction r of the other village members listens to the
radio and receives signal p; and everybody with radio knows that everybody else with radio
knows this, and everybody knows that everybody knows... ad in nitum. Individuals without
radio access do not receive the public signal. To focus on the choices of majority members
that receive the radio broadcasts and keep the analysis tractable, we make the simplifying
11

The key assumption about p is that 2p is nite, so that the broadcasts are informative. The zero mean
is not a binding assumption. If the radio signal is biased on average, individuals will adjust for this when
they form beliefs about : However, treating the signal as exogenous and without manipulation is clearly
a unrealistic simplication, made to keep the analysis simple. For a model with endogenous information
manipulation in a civil war context, see Edmond (2009).

10

assumption that members without radio are unaware of others receiving the radio signal.12
Individuals use Bayes' rule to update their beliefs about the fundamental value of violence.
Consider rst a member without radio. The private posterior distribution for member i that
N
i

receives private signal xi is normally distributed with mean
members with radio, the posterior expectation of
with mean13
R
i

3.3

=

= xi and variance

2
x:

For

given public information alone is normal
2
p xi
:
2
x

2
xp +
2+
p

Equilibrium

We are interested in the equilibrium level of participation, h: Consider a strategy where each
member follows a simple switching rule
j

a( i ) =

8
<

participate

if

: do not participate if

j
i
j
i

j

<

j

where j = N labels the strategy for members without radio and j = R for members with
radio. That is, members participate if and only if their beliefs about the fundamental value
of violence is su ciently high, above some threshold

j

: Following Morris and Shin (1998,

2005), this strategy is unique under some regularity conditions (see the web appendix for the
regularity conditions and the derivation of the equilibrium).14 For members without radio
coverage, the Bayes-Nash equilibrium threshold
N

=

N

2t

:

is
(1)

12

The key assumption is that a fraction 1 r of the members do not receive the signal p. One could in
principle allow 1 r members to not receive the signal p, but still be aware of the distribution of p, and that
some fraction r receives the signal p. This would complicate the analysis, but would most likely not change
the main results.
13

2

2

The posterior variance is 2x+ p2 :
x
p
14
The regularity conditions require that is bounded from above and below. The exact bounds are found
in the web appendix available at http://people.su.se/~daya0852/.

11

R

is the solution

= 0:

(2)

For members with radio coverage, the equilibrium participation threshold
to the equilibrium condition
R

where

(2

+

2 2
x p

t
+

r

R

(p

1=2
4
( 2x
x)

+

2
2
x= y

)

2 1=2
.
p)

+ (1

r)

h

2t

+

R

i

The intuition behind equation 1 is relatively

straightforward. A member without radio coverage faces two forms of uncertainty. First,
there is uncertainty about

and second, there is also uncertainty about how many others

that will participate, h. This is because given the switching strategy, since the member is
uncertain about , he is also not certain about how many other members have expectations
of

above the threshold

N

: However, since he has independent information about , he

forms beliefs about the distribution of . In turn, this means that he holds beliefs about how
many other members are likely to hold expectations of
N

above the participation threshold,

. The higher expectation is the expectation of a member of the value of con ict,

N
i ;

the more other members he expects to participate. The equilibrium condition of equation 1
pins down the expectation

N
i

where a member is indi erent between participating and not

participating. Importantly, since members without radio do not receive the radio signal p and
are also unaware of the existence of the broadcasts, p and r do not change the participation
threshold whereby members are willing to participate.
The intuition behind equation 2 follows a similar logic. However, the important distinction between a member with radio coverage and a member without radio coverage is two-fold.
First, a member with radio receives the additional signal p about the value of con ict . This
will cause him to update his beliefs

R
i

by the same logic as in equation 1. Second, and most

importantly, due to the publicity of the signal he knows that everybody else with radio coverage also has received the same signal p:15 This is important because it will change his beliefs
about how likely it is that other members with radio will participate, h. For this reason,
the fraction r that has received the broadcasts is therefore a key variable in his decision of
whether to participate. When r is low, he knows that not too many have received p, so he
reasons similarly as someone without radio. When r is high, however, he knows that most
members have also received p too, which can dramatically change his expectations about
15

He also knows that everybody with radio knows that everybody with radio knows this, and that everybody... ad in nitum.

12

how others will behave, and can thus change his own willingness to participate. Therefore,
the fraction of the population with radio coverage, r; is a key variable for the equilibrium
participation in violence.

3.4

Participation

Having pinned down the equilibrium thresholds,

N

R

and

; we can investigate the equi-

librium participation, h. Given a fundamental value of violence , we can calculate the
N
i

proportion of non-radio members with beliefs
portion of radio members with beliefs

R
i

R

N

; given by equation 1; and the pro-

; given by equation 2. Using the distributions

for the private signal and the radio signal, conditional on , the total share of the majority
population participating that is a function of village radio coverage r
h = rhR + (1

r)hN ,

(3)

where hN is the proportion of members without radio coverage participating
hN =

2t

+

,

(4)

x

and hR is the proportion of members with radio coverage participating

hR =

2
4

2
x
2
p

2+ 2
x
p
2
p

p+
x

R

3

5 .

(5)

Lemma 1 The participation rate increases with radio coverage (@h=@r > 0; for all r) only
if radio broadcasts a signal that the fundamental value is su ciently high (p > p~

2t

).

De ning propaganda as a signal that the value of con ict is high (p > p~); increasing radio
coverage a ects participation through two propaganda e ects. First, through a direct "fundamentals e ect" that changes the share of the population with beliefs about the value of con ict
above the equilibrium participation threshold,

R

. Second, through an indirect "strategic ef-

fect" that a ects the expectations individuals hold about how many other individuals that will
participate, which changes the equilibrium participation threshold

13

R

:

Proof: see the web appendix. The equilibrium implies that members only participate if
their beliefs about the fundamental value of con ict is su ciently high. Given participation
thresholds for radio members and no-radio members, only if the radio broadcasts that the
fundamental value of con ict is su ciently high (above the participation thresholds) will
a larger fraction of the members with radio hold expectations of the fundamental value of
con ict above the participation threshold. This is the fundamentals e ect.
Furthermore, due the publicity of the radio signal, members with radio know that everybody with radio listens to the same broadcasts. When there is an increase in the radio
coverage, members with radio realize that more people now hold high expectations of the
fundamental value of con ict, which for each member with radio increases the expected
number of participants. This, in turn, changes the equilibrium participation threshold

R

whereby somebody with radio is willing to participate. This is the strategic e ect.
Importantly, the direction of the strategic e ect on participation crucially depends on
whether participation in con ict is subject to strategic complements or strategic substitutes.
Under strategic complements, the e ect is positive, whereas under strategic substitutes the
e ect is negative. Under strategic complements, the total payo of participation in con ict
is always higher the more people that participate. Therefore, when radio coverage increases,
each member with radio expects more people to participate, which makes each member with
radio more willing to participate by lowering the participation threshold. Individuals therefore participate at lower beliefs about the fundamental value of con ict when radio coverage
is high as compared to when it is low. Under strategic substitutes, on the other hand, the
total payo

of participation in con ict is always lower the more people that participate.

Therefore, when radio coverage increases, each member with radio expects more people to
participate, which makes each member with radio less willing to participate by increasing
the participation threshold,

R

.

Next, we derive the properties of participation in violence in the three possible cases:
no strategic interactions, strategic complements, and strategic substitutes. Since we are
interested in how propaganda may increase participation, from now on we assume that
p > p~.16
16

Since the focus of this paper is when p > p~, results are not presented for p < p~: It is worth noting that
in general the results go in the opposite directions when p < p~:

14

Benchmark case: a = 0
We rst state the properties for the benchmark case when

is zero and participation in

violence is free from any strategic interactions.
Proposition 1 ( = 0) : If there are no strategic interactions, then participation increases
linearly in radio coverage (@h=@r = c > 0) and the e ect is the same regardless of the size
of the ethnic minority (@h=@r@t = 0).
Proof: see the web appendix. The intuition behind this result is relatively straightforward. When there are no strategic components, the individual choice of participation does
not depend on how many others that participate. Instead, a member participates if his expectation of the fundamental value of participation is positive. Therefore, radio coverage
only a ects participation through the fundamentals e ect. As the fraction holding positive
expectations of the value of con ict is constant within the group of members with radio
coverage, the fundamentals e ect of radio coverage is linear and positive.
Strategic complements case: a > 0
Next, consider the case when

is positive and participation in violence is subject to

strategic complements.
Proposition 2 (

> 0) : If participation is subject to strategic complements, then radio

coverage exhibits increasing scale e ects (@ 2 h=@r2 > 0 for r 2 [0; r~]; and @ 2 h=@r2 < 0 for
r 2 (~
r; 1] ; where 0 < r~

1): Furthermore, the e ect of radio coverage is decreasing in the

size of the ethnic minority (@h=@r@t < 0 for r 2 [0; r^], where r^ = 1 as long as hR < 1=2: If

r^ < 1; the sign of @h=@r@t for r > r^ is ambiguous).
Proof: see the web appendix. The reason why radio coverage exhibits increasing scale
e ects under strategic complements is due to the combination of the fundamentals e ect and
the strategic e ect. In particular, both e ects are positive. As radio coverage increases, the
fundamentals e ect implies that more members with radio will hold beliefs about the fundamental value of con ict above the participation threshold, which increases participation.
In addition, when radio coverage increases, the strategic e ect implies that members with
radio expect more people to participate which, in turn, lowers the equilibrium participation
threshold, an e ect which further increases participation.
15

Figure 1A graphically shows equation 3 after solving equations 2, 4 and 5. The gure
shows how the participation rate changes as a function of radio coverage, for the benchmark
case and three di erent levels of strategic complements.17 To clearly see the importance
of the strategic e ects, the parameter values are set such that the fundamentals e ect of
radio coverage is essentially zero (i.e. very small and positive). We see that although
the fundamentals e ects are essentially zero (so that almost no members believe that the
fundamental value is su ciently high for participation), there are important positive strategic
e ects when radio coverage is su ciently high. The main insight is that the e ects of radio
coverage can be highly non-linear. The intuition behind this result is that at low levels of
radio coverage, most members with radio expect do not expect many others to participate
since only a small fraction of the population has received the radio broadcasts. At high
levels of radio coverage, however, members with radio know that many have received the
radio broadcasts and therefore, they expect many others to participate. Consequently, due
to these strategic e ects, increasing the radio coverage to high levels of radio coverage can
have dramatic e ects on participation.
Furthermore, the e ect of radio coverage on participation depends considerably on the
size of the ethnic minority group. Figure 1B graphically shows the e ect of radio coverage
for two di erent levels of ethnic minority size (keeping the other parameter values the same
as in Figure 1A). When the size of the ethnic minority is relatively small (t = 1=4), there is
a strong and positive strategic e ect of radio coverage. However, when the size of the ethnic
minority is relatively large (t = 2=5), the e ect of radio coverage almost completely goes
away as there is only a small increase in participation at very high levels of radio coverage.
The reason is relatively straightforward, since the marginal bene t of more participants is
lower when the ethnic minority is large. Therefore, even at high levels of radio coverage, most
members with radio coverage do not expect many others to participate and, consequently,
not many members are willing to participate.
Strategic substitutes case: a < 0
Finally, we treat the case when

is negative and participation in violence is subject to

strategic substitutes.
17

The other parameter values are: p = 0; t = 1=4; = 1; and the variances of private information
( x = 0:05) and radio information ( p = 0:1) are set such that the conditions for a unique equilibrium is
satis ed.

16

Proposition 3 ( < 0) : If participation is subject to strategic substitutes, then radio coverage exhibits decreasing scale e ects (@ 2 h=@r2 < 0 for r 2 [0; r~]; where 0 < r~

1): Further-

more, the e ect of radio coverage is increasing in the size of the ethnic minority (@h=@r@t > 0
for r 2 [0; r^], where r^ = 1 as long as hR < 1=2: If r^ < 1; the signs of @ 2 h=@r2 and @h=@r@t
for r > r^ are ambiguous).
Proof: see the web appendix. Under strategic substitutes, the strategic e ects are negative. Figure 2A graphically shows the importance of negative strategic e ects.18 When radio
coverage is low, the positive fundamentals e ect dominates the negative strategic e ect. The
participation rate therefore initially increases with radio coverage. When radio coverage is
high, however, members with the radio know that many people will have received the radio broadcasts and thus expect higher participation. Expecting many others to participate,
each member nds it less worthwhile to participate. Strategic substitutes therefore result in
decreasing scale e ects of radio coverage.
Furthermore, the e ect of radio coverage on participation depends importantly on the
size of the ethnic minority group. Figure 2B graphically shows the e ect of radio coverage
for two di erent levels of ethnic minority size (keeping the other parameter values the same
as in Figure 2A). We see that the e ect of radio coverage is larger when the size of the
ethnic minority is relatively large. The reason is relatively straightforward, as the marginal
payo of more participants is higher when the ethnic minority is relatively large. Therefore,
even though members with radio coverage expect a relatively large number of other people
to participate at high levels of radio coverage, since the ethnic minority is relatively large,
increases in radio coverage increase participation.
Independent information
In this section, we investigate how the e ects of radio coverage are related to the access
to independent information,

x.

First, even though each member does not know the exact

fundamental value of con ict, he uses his independent information to form expectations about
it.19 : Therefore, the e ect of radio coverage will crucially depend on how much independent
information members have.
18

Compared to Figures 1A and 1B, the value of the radio signal is now set higher (y = 4 instead of y = 0)
so that the benchmark case exhibits visible positive e ects.
19
Recall that the independent private information is equal to xi = + "i ; where "i is independently and
normally distributed with mean zero and variance 2x :

17

Proposition 4 When members have su ciently good access to independent information
(

x

! 0), the e ect of radio coverage disappears (@h=@r ! 0).
Proof: see the web appendix. Intuitively, the expectation a member holds about the

value of con ict, , will be a weighted average between independent information, xi ; and
the information broadcast on the radio, p. The better independent information about the
fundamental value of con ict that members have, the less weight will be put on the radio
broadcasts. Therefore, when members have very precise expectations about the fundamental
value of con ict through other information sources, they stop believing in the radio broadcast.
Consequently, propaganda will not a ect participation in the violence in that case.

3.5

Empirical predictions

We now summarize the results from the previous section into testable predictions.20 Lemma
1 and Propositions 1 to 4 imply the following predictions:
1. Main E ects: If radio coverage r increases the participation rate h; then radio broadcasts a signal that the fundamental value of con ict was high, p > p~: This prediction
follows from Lemma 1.
Moreover, if p > p~; then Propositions 1-4 imply:
2. Ethnic Polarization: The e ect of radio coverage r on the participation rate is
a) decreasing in ethnic polarization t, only with strategic complements in violence (Figure
1B).
b) increasing in ethnic polarization t, only with strategic substitutes in violence (Figure 2B).
3. Scale E ects of r: Radio coverage r exhibits
a) increasing scale e ects on participation h, only with strategic complements (Figure 1A).
b) decreasing scale e ects on participation h, only with strategic substitutes (Figure 2A).
20

We focus on the unambiguous e ects derived in the previous section. That is, we assume that the
additional conditions needed for the unambiguous e ects are ful lled. It is worth noting that the additional
condition h < 1=2 is ful lled in all the observations in the data.

18

4. Independent information: Radio coverage r does not a ect the participation rate h
when ethnic majority members have su ciently good access to independent information (

x

! 0).

Importantly, Predictions 2 and 3 imply to the extent we get consistent results, the data will
allow us to disentangle whether

4

is zero, positive, or negative.

Data and Empirical Strategy

This section describes the data, identi cation strategy, and econometric speci cations.

4.1

Measurement

The variables of interest are h; r; t; and

x.

Here, we present how they are measured. Several

sources of data are combined to construct a village-level cross-sectional dataset. Figure 3
shows a map of village boundaries in Rwanda. The nal dataset consists of 1105 matched
villages.21
Measuring the participation rate, h
Unfortunately, there is no dataset available that measures h directly. Instead, this paper
uses an indirect measure from a nation-wide village-level dataset on prosecutions for violent
crimes committed during the genocide. The data is provided from the government agency
National Service of Gacaca Jurisdictions. The proxy used for the participation rate h is
therefore the prosecution rate.22
The prosecution data for each village comes from local level Gacaca courts.23 The national
court system was set up in 2001 to process the hundreds of thousand of individuals accused
for crimes committed during the genocide.
There are two violent crime categories. Category 1 includes prosecutions for organized
violence, legally de ned as:
21

The term village is used for simplicity reasons, highlighting that the units are relatively small. The
correct term is "administrative sector". The median administrative sector in the dataset is 10.6 square
kilometers and has a population of 4336. There are some problems matching data across data sources, see
each section below.
22
The data used for village population and ethnicity is described below.
23
To see the laws governing the courts, see the National Service of Gacaca Jurisdictions homepage,
http://www.inkiko-gacaca.gov.rw/En/EnLaw.htm (Available 2009-11-05).

19

Planners, organizers, instigators, supervisors of the genocide.
Leaders at the national, provincial or district level, within political parties, army,
religious denominations or militia.
At the village level, these are typically prosecutions committed by local militias such as
the Interahamwe and Impuzamugambi. Category 2 prosecutions concern civilian violence,
de ned as:
Authors, coauthors, accomplices of deliberate homicides, or of serious attacks that
caused someone's death.
The person who - with the intention of killing - caused injuries or committed other
serious violence, but without actually causing death.
The person who committed criminal acts or became the accomplice of serious attacks,
without the intention of causing death.
The data speci es the number of prosecutions for each village in Rwanda. In the sample,
there are approximately 64 000 category 1 prosecution cases, and 362 000 category 2 cases.
Unfortunately, there is no data available on ethnicity at the village level (it is available only
at higher levels), only population numbers in 1991 (see below). The proxy used for the
participation rate h is therefore the prosecution rate, measured as prosecutions per capita.
Figures 6 and 7 show the prosecution rates in villages.24
Since we do not observe actual participation but prosecutions, and per capita rather than
per Hutu, we have some measurement error in the dependent variable. This will not lead to
any biased estimates unless the measurement error is correlated with the measured variation
in radio coverage.
Measuring radio coverage, r
The paper uses village-level data on predicted Radio RTLM coverage. The variable is
constructed in several steps. First, it uses data on Radio RTLM transmitter locations and
24

White areas on the map indicate no data. This is either because of national parks or Lake Kivu (to
the west), or because of matching problems. The data is matched on village names. There are two types of
matching problems. First, names have changed across data sources. Second, two villages within communes
sometimes have identical names.

20

technical speci cations, provided by the government agency O ce Rwandais d'Information.
Then, it predicts the radio coverage across the country by using digital topographic maps and
radio propagation software developed by engineers.25 The software (ArcGIS) uses an algoritm
called ITM/Longley-Rice, which is typically used by radio and TV engineers assessing the
signal strength of broadcasts. The software uses a digital topographic map of Rwanda,
provided by Shuttle Radar Topography Mission (SRTM), and it lets the software run the
ITM/Longley-Rice algoritm and predict the signal strength across the country. The software
produces a radio coverage map at a 90 meter cell resolution, indicating whether each cell has
radio coverage or not. Figure 5 shows predicted radio coverage.26
Using the digital map of village boundaries, the measure of r is calculated as the share
of the village area with coverage.27 As there is no available dataset on Radio RTLM listening rates, the paper will estimate the reduced form e ect of RTLM radio coverage on the
participation rate.28
Measuring ethnic polarization, t
Population and ethnic data is retrieved from the Rwanda 1991 population census, provided
by IPUMS International and GenoDynamics.29 The GenoDynamics data is used for the
population in each village. It does not contain any data on ethnicity. However, the 1991
census from IPUMS International reports the number of Tutsi and Hutu households in the
commune. The ethnicity of the household is de ned as the ethnicity of the household head.
The data is only available at the commune level, which is one administrative level above the
village (i.e., administrative sector). The measure used for t is therefore the number of Tutsi
25

The transmitter parameters are GPS position; transmitter height; transmission power; frequency;
polarization.
26
The software requires topography data in order to predict the radio signal. The digital map has complete
topography data of Rwanda. However, the software runs into a missing data problem for a small section
of villages in the very north and northeast, for signals radiation from the Mount Muhe antenna. This is
because the radio signal need to travel across Uganda in the north before reaching the northeastern Rwanda.
Therefore, the predicted radio signal is incorrect for those areas. The 205 villages a ected by this data
problem are dropped from the sample.
27
As the measure is predicted radio coverage rather than actual radio coverage, there could be some random
measurement error in the data. In that case, this will lead to attenuation bias and an underestimation of
the true e ects.
28
The commune average radio ownership rate in the sample is 34%, taken from the 1991 Census. Radio
ownership data is not available at the village level.
29
The data is available at https://international.ipums.org/international/, (Available 2008-06-08), and
http://www.genodynamics.com/, (Available 2009-05-11).

21

households divided by the number of Hutu households in the commune.
Since there are only two ethnic groups (98% of the population are either Hutu or Tutsi)
where the Tutsi population is all the villages (the maximum t in the data is 0.44), this
measure is equal to the commonly used measures of "ethnolinguistic fractionalization" and
"ethnic polarization", up to a scalar (see Montalvo and Reynal-Querol, 2005). Therefore,
we use t and ethnic polarization interchangeably.
Measuring access to independent information,

x

Ideally, we would want to test Prediction 4 directly through a measure of independent
information (

x ).

But this is naturally unobservable to the researcher. Instead, we proxy for

the access to independent information with the ability to access independent information,
by exploiting variation in literacy rates and education.
Independent information can, of course, come from a range of sources. Within the context of the Rwanda genocide, newspapers are particularly relevant. In the years preceding
the genocide, the independent press quickly expanded with multi-party politics and the legalization of opposition parties in June 1991. The number of independent newspapers that
not aligned with the government parties was between 30 to 60 during this period (Alexis and
Mpambara, 2003; Higiro, 2005). Arguably, a necessary requirement for access to newspapers
is literacy and basic primary education. In addition, Des Forges (1999) reports that, in
practice, not only the literate would read the newspapers, but those who knew how to read
were accustomed to reading newspapers to others.30 .
The data on literacy rates and primary education also comes from the 1991 Census
provided by IPUMS International. For the literacy rate, the fraction of Hutu household
heads that are literate is used. For primary education, the variable is the fraction of Hutu
household heads that have some primary education.31 Both variables are only available at
the commune level.
30

The model assumes that independent information is unbiased on average. However, since the newspapers
in Rwanda were typically aligned with political parties, each newspaper most likely supplied biased information. This does not necessarily mean that the newspapers were biased on average. In fact, Mullainathan and
Shleifer (2005) argues that with su cient political divisions the information will be unbiased on average.
31
The 1991 Census reports "last grade completed" for each household head. Since we would like to directly
measure x ; but use the proxy variables, there is measurement error. This will also lead to attenuation bias
if the error is classical.

22

Covariates
The SRTM topography data and ArcGIS software maps allow us to calculate the village
mean altitude, the village variance in altitude, and the min and max altitude of the village,
distance to the border, and village area. Using data from Africover, we can also measure the
village centroid distance to the nearest major town and the distance to the nearest major
road.
The summary statistics are presented in Table 1.

4.2

Identi cation strategy

To identify the causal e ects of radio coverage on the participation rate requires variation in
radio coverage to be uncorrelated with all other determinants of participation. In the model
radio coverage is exogenous, while in reality the placement of the two RTLM transmitters
was not random. One 100 watt transmitter was placed in the capital Kigali. The other
transmitter (1000 watt) was placed on Mount Muhe in the northwestern part of the country.32
The main endogeneity concern is that the transmitters could have been placed in areas with
high fundamental value of con ict ; little independent information

x;

or ethnic polarization

t. The simple correlation between radio coverage and participation rate would then violate
the identifying assumption. Importantly, since both

and

x

are unobservable, they cannot

be controlled for in a regression.
The following identi cation strategy addresses the problem in steps.33 Rwanda is a very
hilly country without any really at regions. Nick-named "The Land of the Thousand Hills",
Figure 2 shows a map with the topography of Rwanda. There are literally hilltops and valleys
everywhere in the country and the topographic variation shown in Figure 2 provides the basic
foundation for the identi cation strategy. In particular, the main idea is to exploit variation
in radio coverage due to hills in the line-of-sight between radio transmitters and villages in
between radio transmitters and villages.
Radio propagation follows the laws of physics for electromagnetic propagation. Given
32

The highest mountain, Mount Karisimbi, is right on the border to DR Congo and Uganda. Mount Muhe
is the second highest mountain in the country, but the highest one that is well within the country's border.
Together with the Kigali transmitter, the placement strongly suggest to have been driven by a maximizing
of listeners.
33
The strategy was pioneered by Olken (2009). The approach in this paper is similar but not identical to
Olken's.

23

transmitter height and power, the two main determinants of the signal strength are: distance
to the transmitter; and whether the receiver is in the line-of-sight of the transmitter34 In
free space, the power density of the radio signal decreases in the square distance from the
transmitter. Since the transmitter may have been placed strategically, the distance to the
transmitter most likely correlates with either

or

x.

The rst step is therefore to control

for a second-order polynomial in the distance to the transmitter.35 This will leave variation
in signal strength caused by variation in the line-of-sight between the transmitter and the
receiver.
Figure 6 shows graphically how radio coverage due to variation in the line-of-sight is
determined. Whether the receiver is in the line-of-sight of the transmitter will depend on two
factors: the topography where the receiver is located (the higher the altitude of the receiver,
the higher is the likelihood of its being in the line-of-sight) and the topography between the
transmitter and the receiver. Since the topography of a village may be correlated with the
other unobservable determinants of participation in con ict ( and

x ),

it will be controlled

for. The second step is therefore to control for the topography of the village. The control
variables consist of a second order polynomial in the mean altitude of the village and the
altitude variance. This will leave variation in radio coverage due to the topography between
the transmitter and the receiver.
Since the two Radio RTLM transmitters may have been strategically placed in parts of
the country with certain topography, the variation left (after controlling for the distance
to the transmitter and the topography of the village) may still be correlated with
x:

and

Therefore, in order to control for broad regional di erence in topography, the third and

last step is to include commune xed e ects.36 The variation in radio coverage exploited for
identi cation will therefore be highly local variation across villages within communes.37 This
variation is arguably uncorrelated with other determinants of con ict, as radio coverage is
determined by whether a hilltop randomly happens to be in the line-of-sight between the
34

If there are sharp edges that the electromagnetic signal encounters, there can also be some di raction.
The exact formula, and the Longley-Rice model, can be found at http:// attop.its.bldrdoc.gov/itm.html
(Available 2009-11-03).
35
The 2-order polynomial in the distance to the transmitter explains alone 44 percent of the variation in
radio coverage.
36
Commune xed e ects alone explain 82 percent of the variation in village mean altitude, and 72 percent
of the variation in radio coverage.
37
There are 129 communes in the sample and 8.6 villages per commune.

24

transmitter and the village.
Figure 7 shows graphically the topography and radio coverage variation within four communes in the northern part of the country. The radio signal in these communes comes from
the Mount Muhe transmitter located approximately 30 km west, outside the gure. The
gures show that within each commune, villages that happen to be situated to the east of
hilltops have low radio coverage, while villages that happen to be situated to the west of
hilltops have high radio coverage. This is because the signal comes in from the west, and
the hilltops are in line of sight to the transmitter. This arguably provides a credible identication strategy, as there is no plausible reason why other determinants of participation in
violence shoule be di erent across the eastern and western sides of the hilltops.38
Exogeneity check
If the identi cation strategy is valid and radio coverage is as good as randomly assigned,
there should be no correlation between the variation in radio coverage and the other determinants of participation in violence. In particular, there should be no correlation between
radio coverage and the fundamental value of participation in con ict , or the access to independent information,
test this assumption.

x.
39

Since these variables are unobservable, it it not feasible to directly

Instead, we test the validity of the exogeneity assumption by using

observable village characteristics that are likely correlated with

and

x,

namely 1991 pop-

ulation density; 1991 population levels; distance to the nearest major town; distance to the
nearest major road; distance to the nearest border point; and village area.40 . The regression
speci cation is
yc;i = rc;i + Xc;i +

c

+ "c;i ,

(6)

where yc;i is a characteristic of village i in commune c; rc;i is the radio coverage of village
i in commune c; Xc;i is the vector of village i controls and

c

is the commune xed e ects.

For completeness, we test using both levels and logs for each y.
The vector of standard village controls are: a second order polynomial in the kilometer
38

Note that in this particular case, the variation comes from the east-west relationship to the hilltops. In
other communes it will, of course, be in other directions.
39
Since there is no available data on ethnic polarization t at the village level, t is also an unobserved
determinant of participation.
40
The analogy used in randomized experiments is to check whether the treatment and control group is
balanced on observable pre-treatment characteristics.

25

distance to the nearest transmitter; a second order polynomial in the average village altitude
in kilometers; the variance in altitude within the village. If the exogeneity assumption is
correct, we expect

= 0:

Table 2 shows the results. None of the village characteristics are signi cant, and the
lowest p-value is 0.234. This lends credibility to the identi cation strategy. In the main
regressions, results will be presented both without and with village characteristics. The
results are similar with and without the inclusion of these characteristics.

4.3

Econometric speci cations

In this section, we present the econometric speci cations used to test each prediction.
Main E ects (Prediction 1): If radio coverage r increases the participation rate h; then
radio broadcast a signal that the fundamental value of con ict was high, p > p~.
That is, if we nd that radio coverage increased the participation rate, ethnic majority
members perceived the Radio RTLM broadcasts as information that the fundamental value
of con ict being high. To test this, we run the following regression41
log(hc;i ) = rc;i + Xc;i +

c

+ "c;i ,

(7)

where the dependent variable is the logged total number of prosecutions per capita, hc;i ; of
village i in commune c; rc;i is the RTLM radio coverage of village i in commune c; Xc;i is the
vector of village i controls; and

c

is the commune xed e ects.42 We will also run separate

regressions where hc;i is either civilian violence only or organized violence only. The vectors
of standard village controls are: a second-order polynomial in the kilometer distance to the
nearest transmitter; a second-order polynomial in the average village altitude in kilometers
and the variance in altitude within the village. In additional speci cations, we also add
controls for population density, distance to nearest major town, distance to nearest road,
and distance to the nearest border point. According to Prediction 1, if
41

> 0 then this is

Since the true conditional expectations function E[hi j ri ] depends on the unobservable parameters in
the model, it is unknown. We use a standard OLS regression model with a logged outcome variable. The
regression will provide a linear approximation of the true relationship.
42
Of the 1105 villages, 20 have zero prosecutions. Since the outcome variable is logged, we use
log[(prosecutions+1)/population] to deal with the problem of unde ned log function.

26

consistent with p > p~:
Ethnic Polarization (Prediction 2): The e ect of radio coverage r on the participation
rate is decreasing in ethnic polarization t, only if

> 0; and decreasing in t only if

< 0:

Therefore, testing for di erential e ects of radio coverage depending on ethnic polarization gives one method to separate whether there are strategic complements (
strategic substitutes (

> 0) or

< 0) in participation. We test for this using the following speci ca-

tion
log(hc;i ) = rc;i + rc;i

tc + Xc;i +

c

+ "c;i ,

(8)

where tc is a dummy variable indicating whether the size of the ethnic minority population
in commune c is large and the other variables are the same as previously. Speci cally, tc is
equal to one if the ethnic minority size is above the median (7.53%) commune. The main
parameter of interest is : According to Prediction 2, if
> 0: If

< 0; then this is only consistent with

< 0; this is only consistent with

> 0:

Scale E ects (Prediction 3): Radio coverage r exhibits increasing scale e ects, only if
> 0; and decreasing scale e ects, only if

< 0:

This provides an additional test that allows us to separate whether there are strategic
complements (

> 0) or strategic substitutes (

< 0) in participation. To investigate

Prediction 3, we use the following exible non-linear speci cation

log(hc;i ) =

1
X

s s
rc;i

+ Xc;i +

c

+ "c;i ,

(9)

s=0:1

where rs c;i is a dummy variable equal to one if s

0:1

other variables are the same as before. We estimate the

rc;i < s; and zero otherwise. The
s

in order to investigate the scale

e ects.
Independent Information (Prediction 4): Radio coverage r does not a ect the participation rate h when ethnic majority members have su ciently good access to independent
information (

x

! 0).

As described in section 4.4, we test this prediction using literacy rates and primary
education as proxy variables for access to independent information,
27

x.

We use the following

speci cation
log(hc;i ) =
where

j;c

1 rc;i

1;c

+

2 rc;i

2;c

+

3 rc;i

3;c

+ Xc;i +

c

+ "c;i ,

(10)

is a dummy variable indicating whether the Hutu literacy rate (or the Hutu

primary education level) commune c belongs to tertile j in the distribution of Hutu literacy
rates (or the Hutu primary education level): If
to zero; by Prediction 4 we expect

5

3

3;c

is a su ciently good proxy for

x

close

= 0:

Results

In the following sections, we present the results for each tested prediction.

5.1

Main e ects

The results for the test of Prediction 1 are presented in Table 3. Column 1 presents the
simple correlation between radio coverage and the participation rate, and shows a negative
correlation for total violence. However, this is unlikely to be a causal e ect of RTLM radio
coverage for a number of reasons mentioned in the empirical strategy section. Applying
the identi cation strategy by controlling for the main set of variables that determine radio propagation and commune xed e ects, Column 2 shows that radio coverage increased
participation in genocide violence. The e ect is signi cant at the 5 percent level. Column
3 shows that the point estimate is almost identical when additional village covariates are
added. Column 4 shows that RTLM reception has a positive and signi cant impact on civilian violence, and Column 6 shows signi cant e ects also on organized violence.43 Columns
5 and 7 show that adding covariates does little in the way of changing the point estimates,
which is not surprising given the identi cation strategy and the results in Table 2.44
The estimated e ects from the full speci cations in Table 3 are substantial. For overall
violence, Radio RTLM propaganda caused 71 percent (0.561 log points) more participation
43

Residual plots show that the results are not driven by outliers (not shown).
The estimates assume no spillover across villages, which might be unrealistic. If the violence increased
in villages with good radio coverage, which caused further violence in neighboring villages with low radio
coverage, this will lead to an underestimation of the true e ects. If this is the case, the estimates could be
interpreted as providing the lower bounds of the true e ects.
44

28

in violence for villages with full radio coverage (r = 1), as compared to villages unable to
receive the propaganda (r = 0). Looking at the two types of violence separately, civilian
violence increased by 65 percent (0.501 log points) and for organized violence, the increase
was 77 percent (0.572 log points).45
Interpreting these results within the framework of the model and Prediction 1, they imply
that Radio RTLM did indeed broadcast messages that the value of con ict was high, and
Rwandan citizens believed in them. Furthermore, the results are consistent with the model
under strategic interactions in violence, as well as without such interactions. That is, the
results presented in Table 3 are not informative about whether the participation increased
because Hutu citizens updated their beliefs about the fundamental value of violence, or
whether the broadcasts also changed the beliefs how many others were likely to participate
in the killings. Next, we present results that allow us to further understand the underlying
mechanisms that can explain why Radio RTLM caused more violence.

5.2

Ethnic polarization

The results for the test of Prediction 2 are presented in Table 4. Column 1 and 2 show
the estimated e ects for total violence. The interaction e ect between radio coverage and
ethnic polarization is negative with and without additional controls. Both coe cients are
signi cant at the 5 percent level. Columns 3 to 6 show that the interaction coe cients are
similar for civilian and organized violence. The coe cients for civilian violence are signi cant
at the 5 percent level, and insigni cant for organized violence.46 Interestingly, the estimated
coe cients imply that the broadcasts only had an e ect in areas with low ethnic polarization
(i.e., where the ethnic minority population is small), as the point estimate for the interaction
with high ethnic polarization is almost identical, but of the opposite sign, as the coe cient
when ethnic polarization is low.47
As stated in Prediction 2, the results are only consistent with the model under strategic
complements. Figure 1B graphically shows how the model, under strategic complements,
predicts the e ects of radio coverage depending on the relative size of the ethnic minority
45

Due to the speci cation, these are linear approximations of the causal e ects.
Strictly speaking, we cannot reject the null hypothesis for organized violence. Note, however, that this
is due to large standars errors. The coe cients for organized violence are very similar to those for civilian
violence.
47
The p-value for the test of e ects when ethnic polarization is high is 0.89.
46

29

group. The empirical results not only show that RTLM propaganda was ine ective when
the Tutsi population was relatively large, they also suggest that this was due to strategic
complements in ethnic violence. That is, Hutu citizens were more reluctant to participate in
the attacks against Tutsi citizens when the Hutu majority population was relatively small,
perhaps due to a fear that Tutsi villagers would be able to better defend themselves as a
group. Therefore, even though radio broadcast a message about the value of con ict was
high in general, the results show that the broadcasts were not su cient to persuade Hutu
citizens to participate in areas with high ethnic polarization.

5.3

Scale e ects

The results for the test of Prediction 3 are presented in Table 5. Column 1 shows that the
estimated coe cients are generally small and not signi cantly di erent from zero for low
levels of radio coverage, while for high levels of radio coverage, the coe cients are large and
statistically signi cant at the 1 or 5 percent level. Figure 10 graphically plots the coe cients
and the 95 percent con dence intervals. The gure shows that the e ects are highly nonlinear. For the range of up to 60-70 percent radio coverage, the point estimates are small but
not signi cantly di erent from zero. Most importantly, they are non-increasing in the range.
When radio coverage reaches approximately 70 percent, we see a sharp estimated increase
in the participation rate, however. The e ects are substantial. The increase in the point
estimates is almost three-fold. They imply that participation increased by approximately 70
percent when radio coverage reached above 70 percent. The coe cients are signi cant at
the ve-percent level.
Figure 10 suggests that the broadcasts were e ective only when people knew that many
other village members were also listening to the same broadcasts. The model allows us
to further interpret the results. By Prediction 3, under strategic substitutes there should
be decreasing scale e ects, whereas Figure 10 shows increasing scale e ects. This is only
consistent with the model under strategic complements in violence. Furthermore, the results
from the previous section showed that all of the e ects of radio coverage on participation
rates comes from villages where the Tutsi population was relatively small (i.e., low ethnic
polarization). This is also only consistent with the model under strategic complements.
Both results therefore suggest that Radio RTLM caused more violence due to strategic
coordination.
30

Figure 1B shows how the estimated e ects in Figure 10 can be interpreted. For low levels
of radio coverage, even though the ethnic minority is small, there is essentially no e ects
on participation in violence. When radio coverage reaches critically high levels, however,
there is a sharp increase in participation. In particular, when su ciently many receives the
broadcast, then everybody that listens to the radio knows that almost everybody else are
also listening to the same broadcasts. Under strategic complements in violence, individuals
are more willing to participate when they expect others to participate too. And this caused
a large-scale, 70 percent, increases in participation. The evidence therefore suggests that
there were important strategic complements in violence, and that Radio RTLM functioned
as a coordination device.

5.4

Access to independent information

The results for the test of Prediction 4 are presented in Table 6. Column 1 shows that
there is a signi cant e ect of radio coverage when the literacy rate is low. The coe cient is
large and signi cant at the 5 percent level. It implies that in villages with low literacy rates
(bottom tertile), complete radio coverage (r = 1) increased participation by 347 percent
(1.499 log points), compared to villages unable to receive the propaganda (r = 0). Column
1 also shows that in villages with medium literacy rates (middle tertile), radio coverage had
a signi cant e ect on participation. The coe cient is signi cant at the 10 percent level and
implies a 71 percent (0.535 log points) increase in participation when the radio coverage was
complete. Importantly, there is no e ect of radio coverage in the villages with the highest
literacy rates (upper tertile). The coe cient is negative and very close to zero. Column 2
shows that the e ects are similar when additional controls are included.
Columns 3 and 4 estimate the e ects of radio coverage for di erent levels of primary
education. The estimated coe cients show a similar pattern as literacy rates. Importantly,
there is no e ect in villages where the Hutu household heads have most primary education.
The coe cients in both columns 3 and 4 are very close to zero.
Interpreting relatively high literacy rates and a relatively high level of primary education
as better access to independent sources of information, the results con rm Prediction 4.
Moreover, the model allows us to interpret why literacy rates and primary education were
important and suggests why they mitigated the propaganda e ects. When people had better
access to independent information, for example through the 30-60 independent newspapers
31

available at the time, they did not put much weight on the RTLM broadcasts because, in
relative terms, RTLM did not contain much information. Therefore, they did not put much
belief in the messages and, consequently, they were not persuaded to participate in the
killings.

5.5

How much of the genocide is explained by Radio RTLM?

This section performs a simple counterfactual calculation to assess how much of the genocide
that can be explained by Radio RTLM. Speci cally, we use the estimated coe cients of Table
5 and calculate the participation in the absence of the radio station.
For each village i, we rst calculate the counterfactual (r = 0) participation
h
^ i;c (r = 0) = exp log (hc ;i )
h

^s

c;i

i

,

^ i;c is the counterfactual participation rate (prosecution rate) of village i in commune
where h
s
c, and ^ c;i is the coe cient estimate from Table 5, column 1, for the radio coverage indicator
variable equal to 1 for village i48 . Since participation is de ned as the number of village
prosecutions divided by the 1991 village population, we multiply with 1991 population in
order to the get counterfactual number of prosecutions. Summing over all villages, we nd
that Radio RTLM caused approximately 39 700 of the total 425 900 prosecution cases for
genocidal violence in the sample. The estimates therefore suggest that approximately 9%
of the genocide can be explained by Radio RTLM. This is non-trivial considering that only
about 20 percent of the population had radio coverage to receive the broadcasts.49
We can make the same calculations for civilian violence and organized violence, respectively. Using Table 5 column 2 for civilian violence and column 3 for organized violence, the
counterfactual calculation suggests that Radio RTLM caused approximately 32 000 more
civilian prosecution cases (the sample total is approximately 361 700 category 2 crimes) and
5 200 more prosecution cases for organized violence (category 1 crimes). Therefore, using
the separate estimates suggests that approximately 9% of the organized violence and 11%
48

We use the point estimates. Naturally, since there is uncertainty in the estimated coe cients, the
resulting numbers should be taken as approximate estimates.
49
We calculate the number by village radio coverage multiplied by the population number in each village,
given by the 1991 Census. Therefore, the number refers to the share of the population calculated to have
had radio coverage. Since only 34% of the households in the 1991 Census owned a radio (in the communes
in the sample), the number of listeners is most likely lower.

32

of the civilian violence can be explained by Radio RTLM.
Finally, we can use the numbers to assess how many in the Tutsi population were killed
due to Radio RTLM. According to des Forges (1999), at least 500 000 Tutsis were killed
in the genocide. Making the additional assumption that the number of Tutsi deaths is
proportional to the number of prosecutions, the estimated e ects therefore suggest that
Radio RTLM caused 45 000 Tutsi deaths.

6

Conclusion

This paper investigates the impact of propaganda on participation in civil con ict. Specifically, the paper examines the impact of the propaganda spread by the infamous "hate
radio" station Radio Television Libre des Mille Collines before and during the 1994 Rwanda
Genocide.
The paper

rst sets up a simple model of participation in ethnic violence. Then, it

derives a set of testable predictions that are consequently taken to the data. To identify the
causal e ects of the broadcasts, the empirical strategy exploits arguably exogenous variation
generated by Rwanda's highly varying topography consisting of hills and valleys.
The paper presents novel evidence on the e ects of propaganda. The results show substantial e ects of the Radio RTLM broadcasts on violence participation. The estimates
imply that when a village has full rather than zero radio coverage, civilian violence increased
by 65 percent and organized violence by 77 percent.
Furthermore, the paper presents evidence of strategic complements. First, the e ects are
entirely driven by villages where the Hutu ethnic group was large relative to the Tutsi ethnic
minority, which is only consistent with the model under strategic complements. Second,
as predicted by the model under strategic complements, the estimated e ects are highly
nonlinear in the degree of radio coverage as the is a sharp increase in violence when the
village radio coverage is su ciently high. This suggests that the broadcasts were e ective
only when people knew that many other village members were also listening to the same
broadcasts. Together, the evidence therefore suggests that the mechanism through which
the broadcasts increased violence was in part because it functioned as a coordination device.
The model also predicts that access to independent information can mitigate the propaganda e ects. It tests this prediction using variables associated with the ability to access
33

independent information, such as the 30-60 independent newspapers available in Rwanda at
the time of the genocide, by estimating whether the broadcasts had smaller e ects in villages
with higher levels of literacy and primary education. The empirical results show that more
education decreased the propaganda e ects, as there is no e ect of radio coverage in villages
in the tertile with the highest literacy rates and primary education.
To assess how much of the genocide that can be explained by the violence, the paper conducts a simple counter-factual calculation implying that Radio RTLM caused approximately
9.3% of the genocidal violence, corresponding to at least 45 000 deaths. The results therefore
suggest that Radio RTLM was a quantitatively important causal factor in the genocide.
Finding that the propaganda caused more violence, and was partly e ective because
of strategic complements in violence, opens up further questions. Why are there strategic
complements in violence? Is it because attacking in numbers is less risky? Or is it because
not participating is dangerous when many others participate? Are strategic complements
generally present in civil con icts? If so, what are the other devices used for coordination?
These are important questions left for future research.

34

References
[1] Bates, Robert H. (1986), "Modernization, Ethnic Competition, and the Rationality of
Politics in Contemporary Africa," Governing in Black Africa, 6(4):457-84.
[2] BBC

(2003),

"The

Impact

of

Hate

Media

in

Rwanda,"

BBC

article,

, Available (2009-11-06)
[3] Besley, Timothy and Burgess, Robin (2002), "The Political Economy of Government
Responsiveness: Theory and Evidence from India," Quarterly Journal of Economics,
117(4):1415-51
[4] Besley, Timothy and Persson, Torsten (2008), "The Incidence of Civil War: Theory and
Evidence," mimeo
[5] Bethany, Lacina and Gleditsch, Nils Petter (2005), "Monitoring Trends in Global Combat: A New Dataset of Battle Deaths," European Journal of Population 21(2{3):145{66
[6] Blattman, Chris and Miguel, Edward (2009), "Civil War," NBER Working Paper,
14801, March
[7] Boesche, Roger (2003), "Kautilya's Arthasastra on War and Diplomacy in Ancient
India,". The Journal of Military History 67(1):9{37
[8] Carlsson, Hans and van Damme, Eric (1993), "Global Games and Equilibrium Selection," Econometrica, vol. 61(5):989-1018, Sept
[9] Chang, Chun-Fang and Knight, Brian (2008), "Media Bias and In uence: Evidence
from Newspaper Endorsements," NBER Working Paper, 14445
[10] Chretien, J., Dupaquier, J., Kabanda, M., Ngarambe, J., and Reporters Sans Frontieres
(1995), Rwanda: Les Medias du Genocide [Rwanda: The media of the genocide]. Paris:
Karthala.
[11] Collier, Paul and Hoe er, Anke (1998), "On Economic Causes of Civil War," Oxford
Economic Papers, 50(4):563-73
[12] Collier, Paul and Hoe er, Anke (2004), "Greed and Grievance in Civil War," Oxford
Economic Papers, 56(4):563-95
35

[13] Della Vigna, Stefano and Kaplan, Ethan (2007), "The Fox News E ect: Media Bias
and Voting," Quarterly Journal of Economics, 122(3):187-234
[14] Della Vigna, Stefano and Gentzkow, Matthew (2009), "Persuasion: Empirical Evidence"
Annual Review of Economics, forthcoming
[15] Des Forges, Alison (1999), Leave None to Tell the Story: Genocide in Rwanda. Human
Rights Watch and the International Federation of Human Rights Leagues, New York,
NY, USA. , Available (2009-11-06).
[16] Diamond, Jared (2005), Collapse: How Societies Choose to Fail or Succeed, New York:
Penguin.
[17] Do, Quy-Toan and Iyer, Lakshmi (2007), "Poverty, Social Divisions and Con ict in
Nepal," Unpublished working paper, Harvard Business School.
[18] Dube, Oeindrila and Vargas, Juan F. (2007), "Commodity Price Shocks and Civil Conict: Evidence From Colombia," Unpublished working paper, Harvard University and
UCLA
[19] Edmond, Chris (2008), "Information Manipulation, Coordination and Regime Change,"
Stern School of Business, mimeo
[20] Enikolopov, R., Petrova, M. and Zhuravskaya, E. (2009), "Media and Political Persuasion: Evidence from Russia", mimeo
[21] Eisensee, Tomas and Str•omberg, David (2007), "News Droughts, News Floods, and U.S.
Disaster Relief," Quarterly Journal of Economics, 122(2):693-728
[22] Fearon, James D. and Laitin, David (1996), "Ethnicity, Insurgency and Civil War,"
American Political Science Review, 97(1):75-90
[23] Fearon, James D. and Laitin, David (1996), "Explaining Interethnic Cooperation,"
American Political Science Review, 90(4):715-35
[24] Gentzkow, Matthew and Shapiro, Jesse (2004), "Media, Education and Antiamericanism in the Muslim World," The Journal of Economic Perspectives, 18(3):117-33

36

[25] Gentzkow, Matthew (2006), "Television and Voter Turnout," Quarterly Journal of Economics, 121(3):931-972
[26] Gerber A, Karlan, D and Bergan, D. (2009), "Does the Media Matter? A Field Experiment Measuring the E ect of Newspapers on Voting Behavior and Political Opinions,"
American Economic Journal: Applied Economics, 1(2):35-52
[27] Glaeser, Edward (2005), "The Political Economy of Hatred," Quarterly Journal of Economics, 120(1):45-86.
[28] International Criminal Tribunal for Rwanda, ICTR (2003), Case No. ICTR-99-52-T
Judgement. http://www.ictr.org/ENGLISH/cases/Barayagwiza/judgement/SummaryMedia.pdf (Available 2009-09-05)
[29] Jha, Saumitra (2008), "Trade, Institutions and Religious Tolerance: Evidence from
India," Stanford GSB Working Paper No. 2004, January
[30] Knight, Brian and Chiang, CF (2008), "Media Bias and In uence: Evidence from newspaper endorsements," mimeo
[31] Berelson, Bernard, Lazarsfeld, Paul and William M. McPhee (1954), Voting: A Study
of Opinion Formation in a Presidential Campaign. University of Chicago Press.
[32] Miguel, Edward, Satyanath, Shanker and Sergenti, Ernest (2004), "Economic Shocks
and Civil Con ict: An Instrumental Variables Approach," Journal of Political Economy,
112(4):725-53
[33] Montalvo, J.G. and Reynal-Querol, Marta (2005), "Ethnic Polarization, Potential Conict and Civil War," American Economic Review, 95(3),:796-816.
[34] Morris, Stephen and Shin, Hyun Song (1998), "Unique Equilibrium in a Model of SelfFul lling Currency Attacks," American Economic Review, 88(3):587-97.
[35] Morris, Stephen and Shin, Hyun Song (2005), "Notes on Strategic Substitutes and
Complements in Global Games", mimeo
[36] Mullainathan, Sendhil and Shleifer, Andrei (2005), "The Market for News," American
Economic Review, 95(4): 1031{53.
37

[37] Murshed, S. Mansoob and Gates, Scott (2005), Spatial-Horizontal Inequality and the
Maoist Insurgency in Nepal," Review of Development Economics 9, 1:121-34.
[38] Olken, Ben (2009), "Do Television and Radio Destroy Social Capital? Evidence from
Indonesian Villages," American Economic Journal: Applied Economics, 1(4):1-33
[39] Paluck, Elizabeth L. (2009), Reducing Intergroup Prejudice and Con ict Using the
Media: A Field Experiment in Rwanda," Journal of Personality and Social Psychology
96 (March): 574-587
[40] Prat, Andrea and Str•omberg, David (2005), Commercial Television and Voter Information,", CEPR Discussion Papers 4989.
[41] Prunier, Gerard (1995), The Rwanda Crisis: History of a Genocide. New York:
Columbia University Press.
[42] Verpoorten, Marijke (2005), "The Death Toll of the Rwandan Genocide: a Detailed
Analysis for Gikongoro Province," Population 60(4):331-68.
[43] Verwimp, Philip (2005), "An Economic Pro le of Peasant Perpetrators of Genocide:
Micro-level Evidence from Rwanda," Journal of Development Economics, 77(2):297323.
[44] Verwimp, Philip (2006), "Machetes and Firearms: the Organisation of Massacres in
Rwanda," Journal of Peace Research, 43(1):5-22
[45] Straus, Scott (2004), "How Many Perpetrators Were There in the Rwandan Genocide?
An Estimate," Journal of Genocide Research, 6(1):85-98
[46] Straus, Scott (2007), "What Is the Relationship between Hate Radio and Violence?
Rethinking Rwanda's "Radio Machete"," Politics & Society, 35:609
[47] Str•omberg, David (2004), "Radio's Impact on Public Spending," Quarterly Journal of
Economics, 119(1):189{221.
[48] Str•omberg, David and Snyder, Jim (2008), "Press Coverage and Political Accountability,", mimeo
[49] Thompson, Allan (2007), The Media and the Rwandan Genocide. London: Pluto Press.
38

Figure 1A. The figure plots the participation rate as a function of the
radio coverage. It shows the importance of strategic effects due to
strategic complements. Under the benchmark case (fundamentals effects
only), there is essentially no effect. The higher are the degrees of
complements, the larger are the strategic effects.

Figure 1B. The figure plots the participation rate as a function of the
radio coverage for two levels of ethnic minority size. It shows that the
effect of radio coverage is smaller, and can disappear, when the ethnic
minority is relatively large.

Figure 2A. The figure plots the participation rate as a function of the
radio coverage. It shows the importance of strategic effects due to
strategic substitutes. Under the benchmark case (fundamentals effects
only), the effect is linear. The effects of radio coverage decrease with
higher degrees of substitutes.

Figure 2B. The figure shows the participation rate as a function of the
radio coverage for two levels of ethnic minority size. It shows that the
effect of radio coverage is larger when the ethnic minority is large.

Figure 3. The Topography of Rwanda
Source: Shuttle Radar Topography Mission

Figure 4. Rwandan Village Boundaries
Source: Analog map by Organisation Administrative du territorie de la Republic
Rwandaise, digitized by the author.

Figure 5. RTLM Radio Coverage
The figure shows the predicted radio coverage based on SRTM 90 meter digital topography maps and
ArcGIS radio propagation software. The two red dots mark the transmitters. The north-western 1000 watt
transmitter is on Mount Muhe. The central 100 watt transmitter is in the capital Kigali. Yellow indicates
radio coverage. The map also shows a software calculation error in the north due to missing topography
data (see the data section for the details). Villages affected by this error are excluded from the sample.
Source: Author’s calculations in ArcGIS using the Longley-Rice Propagation Model.

Figure 6. Theoretical Radio Coverage
Dotted space marks low signal strength, and striped space marks even lower signal strength. The figure shows
that the signal strength for a point on the ground is lower when there is a hilltop in the line-of-sight between
the transmitter and a point on the ground. The red bars mark hypothetical village boundaries.

Figure 7. Predicted Radio Coverage, 4 communes example
This left picture shows the height of ground, where brighter marks higher altitude. The right picture shows the
empirical radio coverage, where grey marks radio coverage. The signal comes from the Mount Muhe
transmitter located 30 km to the west (outside the figure). The figures show that within each commune
(boundaries in thick white lines), villages (boundaries in thin white lines) to the east of hill tops have low
radio coverage due the hilltops in the line-of-sight to the transmitter.
Source: SRTM topography data, Author’s own calculations of radio coverage in ArcGIS software.

Figure 8. Civilian Violence. White areas are no data areas, either because of
Lake Kivu, Natural Reserves, or villages that are missing due to unmatchable
data issues.

Figure 9. Organized Violence. White areas are no data areas, either because
of Lake Kivu, Natural Reserves, or villages that are missing due to
unmatchable data issues.

0.98

0.78

Coefficient

0.58

0.38

0.18

-0.02

0.1-0.2

0.2-0.3

0.3-0.4

0.4-0.5

0.5-0.6

0.6-0.7

-0.22

Radio Coverage

Figure 10. Scale Effects, Total Violence

0.7-0.8

0.8-0.9

0.9-1

Table 1. Summary Statistics
Variable

Observations

Mean

Std. Dev.

Dependent Variables
Participation Rate, Total
Participation Rate, Civilian
Participation Rate, Organized

1105
1105
1105

.084
.072
.013

.070
.060
.016

Independent Variables
Radio Coverage
Altitude, Mean
Altitude, Variance
Distance to Transmitter
Distance to Major Town
Distance to Major Road
Distance to the Border
Village Area
Hutu Literacy Rate
Hutu Primary Education
Tutsi Minority Size
Population
Population Density

1105
1105
1105
1105
1067
1071
1074
1105
1105
1105
1105
1105
1105

.189
1.713
9208.3
5.171
.200
.058
.217
15.07
.503
.579
.098

.226
.229
10531.6
2.841
.120
.052
.127
44.6
.056
.060
.085

4846.7
.528

2456.5
.868

The dependent variables are violent crimes prosecutions divided by the village population in 1991; Organized Violence is crime category 1 prosecutions
against organizers, leaders, army and militia; Civilian Violence is crime category 2 prosecutions for homicides, attempted homicides and serious violence.
Total is the combined Civilian and Organized. Radio Coverage is the share of the village area that has RTLM reception. Altitude, Mean is the mean altitude
in the village in kilometers. Altitude, Variance is the village variance in altitude in meters, Distance to Transmitter is the distance in kilometers to the
nearest RTLM transmitter. The other distance variables are measured in decimal degrees. Hutu Literacy Rate is the fraction of Hutu household heads in the
commune that are literate. Hutu Primary Education is the fraction of Hutu household heads in the commune that have at least some primary education.
Education and literacy data are taken from the 1991 Census, available only at the commune level. There are 129 communes in the sample, and
approximately 8.6 villages per commune. Population is the population number in the village and Population Density is 1000 people per square kilometers,
also from the 1991 Census.

Table 2. Exogeneity Check
Dependent Variable

Population
Density, 1991

Population, 1991

Village Area, km2

Distance to Major
Town

Distance to Major
Road

Distance to the
Border

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

(11)

(12)

level

log

level

log

level

log

level

log

level

log

level

log

0.240
[0.352]

0.177
[0.205]

-557.32
[766.21]

-0.047
[0.094]

-28.484
[31.305]

-0.224
[0.191]

0.006
[0.010]

0.096
[0.112]

-0.012
[0.010]

-0.233
[0.212]

0.001
[0.011]

0.091
[0.146]

Controls

Y

Y

Y

Y

Y

Y

Y

Y

Y

Y

Y

Y

Commune FE

Y

Y

Y

Y

Y

Y

Y

Y

Y

Y

Y

Y

N

1105

1105

1105

1105

1105

1105

1067

1067

1071

1071

1074

1074

R-squared

0.44

0.42

0.42

0.45

0.18

0.56

0.95

0.90

0.81

0.70

0.96

0.92

P-value of Radio Coverage

0.496

0.389

0.468

0.618

0.365

0.243

0.528

0.390

0.234

0.275

0.957

0.535

Radio Coverage

Radio Coverage is the share of the village area that has RTLM radio coverage. The controls are: A second-order polynomial in village mean altitude, altitude variance, and a
second-order polynomial in the distance to the nearest transmitter.
Robust standard errors in parentheses, clustered at the commune level. There are 129 communes in the sample. * significant at 10%; ** significant at 5%; *** significant at 1%

Table 3. Main Effects
Dependent Variable
(1)
Radio Coverage

Log(Participation Rate)
Total Violence
(2)
(3)

-0.717
[0.260]***

0.571
[0.229]**

0.561
[0.244]**

Log(Participation Rate)
Civilian Violence
(4)
(5)

Log(Participation Rate)
Organized Violence
(6)
(7)

0.520
[0.229]**

0.559
[0.294]*

0.501
[0.246]**

0.572
[0.288]**

Log(Population Density)

-0.127
[0.071]*

-0.120
[0.071]*

-0.101
[0.082]

Distance to Major Town

1.019
[1.534]

1.224
[1.526]

-0.518
[1.768]

Distance to Major Road

-2.791
[1.548]*

-2.646
[1.554]*

-4.527
[1.810]**

Distance to the Border

1.910
[1.317]

2.150
[1.366]

0.198
[1.625]

Controls

N

Y

Y

Y

Y

Y

Y

Commune FE

N

Y

Y

Y

Y

Y

Y

N

1105

1105

1066

1105

1066

1105

1066

R-squared
0.02
0.62
0.63
0.61
0.62
0.51
0.52
Participation Rate is the number of violent crimes prosecutions per capita; Total Violence is the sum of Civilian and Organized Violence, Organized Violence
is crime category 1 prosecutions against organizers, leaders, army and militia; Civilian Violence is crime category 2 prosecutions for homicides, attempted
homicides and serious violence. Radio Coverage is the share of the village area that has RTLM radio coverage. The radio propagation controls are: A secondorder polynomial in village mean altitude, village altitude variance, and a second-order polynomial in the distance to the nearest transmitter. Robust standard
errors in parentheses, clustered at the commune level. There are 129 communes in the sample.
* significant at 10%; ** significant at 5%; *** significant at 1%

Table 4. Ethnic Polarization
Log(Participation Rate)
Total Violence
(1)
(2)

Log(Participation Rate)
Civilian Violence
(4)
(5)

Log(Participation Rate)
Organized Violence
(6)
(7)

Radio Coverage

0.932
[0.303]***

0.936
[0.325]***

0.849
[0.301]***

0.834
[0.325]**

0.870
[0.379]**

0.922
[0.369]**

Radio Coverage x High Ethnic Polarization

-0.972
[0.411]**

-0.972
[0.427]**

-0.884
[0.412]**

-0.864
[0.430]**

-0.839
[0.619]

-0.907
[0.614]

Dependent Variable

Log(Population Density)

-0.126
[0.070]*

-0.118
[0.070]*

-0.100
[0.081]

Distance to Major Town

0.845
[1.509]

1.069
[1.504]

-0.681
[1.738]

Distance to Major Road

-2.736
[1.528]*

-2.598
[1.537]*

-4.475
[1.789]**

Distance to the Border

1.824
[1.296]

2.073
[1.346]

0.118
[1.613]

Controls

Y

Y

Y

Y

Y

Y

Commune FE

Y

Y

Y

Y

Y

Y

N

1105

1066

1105

1066

1105

1066

R-squared
0.62
0.63
0.61
0.62
0.51
0.52
Participation Rate is the number of violent crimes prosecutions per capita; Total Violence is the sum of Civilian and Organized Violence, Organized
Violence is crime category 1 prosecutions against organizers, leaders, army and militia; Civilian Violence is crime category 2 prosecutions for
homicides, attempted homicides and serious violence. Radio Coverage is the share of the village area that has RTLM radio coverage. The radio
propagation controls are: A second-order polynomial in village mean altitude, the village altitude variance, and a second-order polynomial in the
distance to the nearest transmitter. Robust standard errors in parentheses, clustered at the commune level. There are 129 communes in the sample.
* significant at 10%; ** significant at 5%; *** significant at 1%

Table 5. Scale Effects
Dependent Variable
Radio Coverage, 0.1 - 0.2
Radio Coverage, 0.2 - 0.3
Radio Coverage, 0.3 - 0.4
Radio Coverage, 0.4 - 0.5
Radio Coverage, 0.5 - 0.6
Radio Coverage, 0.6 - 0.7
Radio Coverage, 0.7 - 0.8
Radio Coverage, 0.8 - 0.9
Radio Coverage, 0.9 – 1

Log(Participation Rate)
Total Violence
(1)
0.181
[0.119]
0.178
[0.148]
0.278
[0.153]*
0.194
[0.138]
0.232
[0.191]
0.154
[0.187]
0.602
[0.201]***
0.518
[0.228]**
0.498
[0.211]**

Log(Participation Rate)
Civilian Violence
(2)
0.163
[0.119]
0.180
[0.143]
0.281
[0.158]*
0.196
[0.137]
0.227
[0.199]
0.169
[0.178]
0.559
[0.205]***
0.429
[0.239]*
0.381
[0.189]**

Log(Participation Rate)
Organized Violence
(3)
0.346
[0.121]***
-0.004
[0.145]
0.147
[0.149]
0.115
[0.216]
-0.005
[0.199]
0.171
[0.320]
0.594
[0.285]**
0.855
[0.288]***
0.810
[0.390]**

Controls
Y
Y
Y
Commune FE
Y
Y
Y
N
1105
1105
1105
R-squared
0.62
0.61
0.52
Participation Rate is the number of violent crimes prosecutions per capita; Total Violence is the sum of Civilian and Organized
violence, Organized Violence is crime category 1 prosecutions against organizers, leaders, army and militia; Civilian Violence is
crime category 2 prosecutions for homicides, attempted homicides and serious violence. Radio Coverage is the share of the
village area that has RTLM radio coverage. The radio propagation controls are: A second-order polynomial in village mean
altitude, village altitude variance, and a second-order polynomial in the distance to the nearest transmitter.
Robust standard errors in parentheses, clustered at the commune level. There are 129 communes in the sample.
* significant at 10%; ** significant at 5%; *** significant at 1%

Table 6. Ability to Access Independent Information
Dependent Variable
Radio Coverage x Low Hutu Literacy
Radio Coverage x Medium Hutu Literacy
Radio Coverage x High Hutu Literacy

Log(Participation Rate)
Total Violence
(3)

(1)

(2)

1.499
[0.582]**
0.535
[0.323]*
-0.013
[0.308]

1.549
[0.602]**
0.484
[0.355]
-0.042
[0.321]

Radio Coverage x Low Hutu Education

0.855
[0.473]*
0.824
[0.329]**
0.015
[0.337]

Radio Coverage x Medium Hutu Education
Radio Coverage x High Hutu Education
Log(Population Density)
Distance to Major Town
Distance to Major Road
Distance to the Border

-0.122
[0.071]*
0.963
[1.513]
-2.592
[1.526]*
2.178
[1.299]*

(4)

0.811
[0.480]*
0.980
[0.366]***
-0.139
[0.364]
-0.128
[0.070]*
0.971
[1.536]
-2.821
[1.541]*
2.011
[1.304]

Y
Y
Y
Y
Controls
Commune FE
Y
Y
Y
Y
N
1105
1066
1105
1066
R-squared
0.62
0.63
0.62
0.63
Participation Rate is the number of violent crimes prosecutions per capita; Total Violence is the sum of Civilian and Organized
violence, Organized Violence is crime category 1 prosecutions against organizers, leaders, army and militia; Civilian Violence is
crime category 2 prosecutions for homicides, attempted homicides and serious violence. Radio Coverage is the share of the village
area that has RTLM radio coverage. The radio propagation controls are: A second-order polynomial in village mean altitude, village
altitude variance, and a second-order polynomial in the distance to the nearest transmitter. The other variables are described in the
data section. Robust standard errors in parentheses, clustered at the commune level. There are 129 communes in the sample.
* significant at 10%; ** significant at 5%; *** significant at 1%

Top

fgtquery v.1.9, February 9, 2024