ZurichOpenRepositoryand
Archive
UniversityofZurich
UniversityLibrary
Strickhofstrasse39
CH-8057Zurich
www.zora.uzh.ch
Year:2014
Propagandaandconict:evidencefromtheRwandangenocide
Yanagizawa-Drott,David
Abstract: Thisarticleinvestigatestheroleofmassmediaintimesofconictandstate-sponsoredmass
violenceagainstcivilians.Weuseauniquevillage-leveldatasetfromtheRwandangenocidetoestimate
theimpactofapopularradiostationthatencouragedviolenceagainsttheTutsiminoritypopulation.
Theresultsshowthatthebroadcastshadasignicanteectonparticipationinkillingsbybothmilitia
groupsandordinarycivilians.Anestimated51,000perpetrators,orapproximately10%oftheoverall
violence,canbeattributedtothestation.Thebroadcastsincreasedmilitiaviolencenotonlydirectly
byinuencingbehaviorinvillageswithradioreceptionbutalsoindirectlybyincreasingparticipationin
neighboringvillages.Infact,spilloversareestimatedtohavecausedmoremilitiaviolencethanthedirect
eects. Thus,thearticleprovidesevidencethatmassmediacanaectparticipationinviolencedirectly
duetoexposureandindirectlyduetosocialinteractions.
DOI:https://doi.org/10.1093/qje/qju020
PostedattheZurichOpenRepositoryandArchive,UniversityofZurich
ZORAURL:https://doi.org/10.5167/uzh-137646
JournalArticle
AcceptedVersion
Originallypublishedat:
Yanagizawa-Drott,David(2014).Propagandaandconict:evidencefromtheRwandangenocide.Quar-
terlyJournalofEconomics,129(4):1947-1994.
DOI:https://doi.org/10.1093/qje/qju020
PROPAGANDA AND CONFLICT: EVIDENCE FROM
THE RWANDAN GENOCIDE
David Yanagizawa-Drott
∗
Harva rd Uni versity
August 2014
Abstract
This paper investigates the role of mass m edia in times of conflict and state-sponsored
mass violence against civilians. We use a uniqu e village-level dataset from the Rwandan
Genocide to estimate the impact of a popular radio station that encouraged violence
against the Tutsi minority population. The results show that the broadcasts had a s ig-
nificant impact on participation in killings by both militia groups and ordinary civilians.
An estimated 51,000 perpetrators, or approximately 10 percent of the overall violence,
can be attributed to th e station. The br oadcasts increased militia violence not only
directly by influencing behavior in villages with radio reception, but also indirectly by
increasing participation in neighb oring villages. In fact, spillovers are estimated to have
caused more militia violence than the direct effects. Thus, the paper provides evidence
that mass media can affect participation in violence directly due to exposure, and indi-
rectly due to social interactions.
JEL codes: D7, N4
Keywords: Conflict, Genocide, Mass M edia, Propaganda, Social Interactions
∗
79 JFK St, John F. Kennedy School of Government, Harvard University, Cambridge, Massachusetts
02138. E-mail: david
yanagizawa-drott@harvard.edu. This paper was previously titled “Propaganda and
Conflict: Theory and Evidence from the Rwandan Genocide.” I thank Robert H. Bates, Tim Besley, Raquel
Fe rnandez, Ethan Kaplan, Asim Khwaja, Masayuki Kudamatsu, Rocco Macchiavello, Vestal McIntyre, Nancy
Qian, Rohini Pande, Tors ten Persson, David Stromberg, and Jakob Svensson for their comments, and seminar
and conference participants at Harvard; Stanford SIEPR; LSE; Dartmo uth; U Pompeu Fabra; Warwick; U
Namur, Center for Global Development; EUDN Oxford; ESEWM, CIFAR, and BREAD 2011. I would also
like to thank Giovanni Zambotti for ArcGIS assistance, a nd Annalise Blum and Aletheia Donald for excellent
resear ch assistance. All mistakes are my own.
1
1 Introduction
Since 1945 as many as 22 million noncombatants have been killed in nearly fifty genocides
and politicides (Harff, 2003). These are political mass killings that are typically sp onsored or
initiated by elites in control of the government, where those elites have agendas to reduce or
eliminate certain groups (ethnic or religious) that are thought to constitute political threats.
1
The tremendous costs of political mass killings in terms of human life wa r rants a full investi-
gation of how to prevent them; isolating the mechanisms that enable elites to carry them o ut
is key to a fuller understanding.
Joseph Goebbels, Hitler’s propaganda minister, called radio “the most importa nt instru-
ment of mass influence that exists anywhere” (Welch, 1993). Elites in control of autocratic
states have repeatedly used mass media – often under their direct control – with the intent ion
to induce citizen support of, and participation in, violence against certain g roups (Lee, 19 45;
Lasswell, 1971). Cross-country evidence indicates that when persecution of certain g r oups in
society is made the official ideolo gy of the elite in power, the likelihood of a conflict transition-
ing into po litical mass killings is significantly higher (Harff, 2003). Yet, it is an open question
whether and how propaganda that explicitly encourages violence against a certain group can,
in fact, directly induce violence against that group.
This paper investigates the role of mass media in the spread of violence by estimating
the effects of propaganda disseminated via radio during the 1994 Rwandan Genocide. This
planned campaig n was led by key ethnic Hutu members of the government against the Tutsi
ethnic minority. In addition to violence by the milita ry, att acks and massacres conducted by
local militias groups and ordinary civilians contributed to a death toll of 0.5–1.0 million deaths
(Des Forges, 1999; Straus, 2004; Verwimp, 2006) . In a country with low newspaper circulation
and few television sets, radio was the dominant medium for the government to deliver messages
to the population. The radio station Radio T´el´evision Libre des Mille Collines (RTLM) led the
propaganda effort s by broadcasting inflammatory messages calling for the extermination of
the Tutsi minority. Although qualitative evidence suggests this “hat e radio” station catalyzed
violence (Hatzfeld, 2005; Straus, 2007), and its cofounders were found guilty of instigating
genocide by the International Criminal Tribunal for Rwanda , there is no quantitative evidence
establishing whether, how, and to what extent the broadcasts caused more violence.
1
Political mass k illings, i.e. genocides and politicides, are considered to be distinctly different phenomena
from civil war a nd re volutions, primarily because of the intent of state authorities to destroy certain groups in
society, but also because the violence is large scale and one-sided. That said, multiple definitions exis t (e.g.,
Harff and Gurr, 1988; Krain, 1997; Harff, 2003).
2
We hypothesize that mass media could have fueled participation in the violence via two
broad mechanisms. First, in line with the literature on persuasive communication, the broad-
casts could have had a direct persuasion effect by convincing some listeners t hat participation
in the attacks on Tutsis was preferable to no n-participation.
2
This mechanism is plausible
given t hat the broadcasts contained not only strong anti-Tutsi rhetoric that may have in-
creased pro-violence preferences, but also information about relevant tradeoffs: they made it
clear that the government would not punish participation in t he killing of Tutsi citizens or
the appropria tion of their property, but instead encouraged or even mandated such behav-
ior. Second, following a long traditio n in the social sciences on the role of social interactions
in general, and their importance in intermediating mass media effects in particular, a direct
persuasion effect could influence the spatia l diffusion of violence, even beyond the immediate
areas of radio reception. One would expect this to be the case if violence begets violence,
leading to contagion, or if information and beliefs spread via social networ ks.
3
Put simply,
the broadcasts may have affected overall violence via local spillover effects, in addition to
direct effects from exposure.
We build a unique village- level dataset fro m Rwanda to examine these hypotheses. We
use information on RTLM tr ansmitters a nd radio propagation software to produce a dataset
on radio coverage at a high spatial resolution, allowing us to calculate the area with reception
within each village. To identify causal effects, our empirical strategy exploits variation in
radio reception generated by Rwanda’s highly varying topography, which is practically ran-
dom and, therefore, arguably uncorrelated with other determinants of violence.
4
To measure
participation in the violence, we use data on the number of persons prosecuted for violent
crimes committed during the genocide in each village. The prosecution data contain two dis-
tinct legal categories of crime: the first fo r members and accomplices of organized forms of
violence, primarily from local militias (77,0 00 persons in total); the second for less organized
individual violence carried out by perpetrato r s who are not members or accomplices of any of
2
For an overview of the persuasion literature, see DellaVigna and Gentzkow (2010), and Glaeser (2005) for
a po litical economy mo de l of propaganda and hatred towards minority groups.
3
The hypothesis that social interactions provide an intermediate channel for persuasion effects on behavior
dates back to Lazarfeld et al. (194 4) and Katz and Lazarfeld (1955). Formal models of social interactions
under complementarities in violence production go back to at least Granovetter (1978), and early information-
based models o f herd behavior and contagion include Banerjee (1992) and Bikchandani, Hirshleifer and Welch
(1992); see Chamley (2004) for an overview. Early theoretical work on the diffusion of crime include Sa h
(1991), and on the empirical side papers by Case and Katz (1991) and Glaeser, Sacerdote and Scheinkman
(1996). For overviews of the social interactions literature, see Manski (2000), Durlauf (2004 ), and Jackson
and Yariv (2011).
4
Olken (2009 ) was the first to use a similar, but not identical, strategy to identify media effects.
3
the organized groups in the first category (432,000 persons in to t al). For simplicity, hereafter
we refer to the first category as militia violence, the second category as individ ual viol ence,
and the sum of the two as total violen ce .
The results show that the broadcasts led to more violence during the genocide. First,
there is a direct effect on participation, with violence increasing in radio coverage in the
village. A one standard deviation increase in ra dio coverage is associated with a 12–13 percent
increase in participation in total violence. The effect is similar for militia violence (13–14
percent) and individual violence (10–11 percent). A battery of robustness tests show that the
effects are unlikely to be spurious due to omitted variables, outliers, or measurement error
in violence. Moreover, placebo tests show that another radio station t hat did not broadcast
propaganda instigat ing geno cide had no effects on violence, indicating that radio reception
irrespective of content did not influence par t icipation. We also present suggestive evidence
that the RTLM broadcasts were mo st effective in inducing violence in villages where the
population was relatively uneducated and illiterate, and where Tutsis made up a relatively
small minority.
Second, we find evidence that the broa dcasts exhibited positive spillover effects in militia
violence. The number o f persons engaged in militia violence in a given village was significantly
higher when a larger share of the population in neighboring villages had radio coverage, con-
sistent with the hypothesis that social interactions determine the spatial diffusion of violence.
There are no spillover effects on individual violence, suggesting that local complementar ities
or information diffusion among ordinary citizens were weak or nonexistent, at least relative
to that among members of organized militias.
Third, we use the regression estimates on the direct effects and indirect effects and perform
a simple counterfactual calculation of t he countrywide effects on participation, enabling us to
also quantify the relative importance of the spillovers. This analysis suggests that 1 0 percent
of the total participation in the genocide, or approximately 51,000 prosecuted persons, was
caused by the radio station. Spillovers ha d a greater overall impact on militia violence (16,000
additional persons) than did the direct effects (6,000 additional persons). This is consistent
with existing qualitative evidence from perpetrator interviews by Hatzfeld (2005) and Straus
(2007), which suggest that the broadcasts persuaded a limited number of key agents of t he
local elite in villages, and these agents in turn recruited individuals in neighboring villages by
engaging in face-to-face mobilization.
The empirical literature on the determinant s of conflict has mostly focused o n two-sided
4
