Citation
Did Radio RTLM Really Contribute Meaningfully to the Rwandan Genocide?:
Using Qualitative Information to Improve Causal Inference
From Measures of Media Availability
Gordon Danning
Foundation for Individual Rights in Education
510 Walnut St., Ste 1250
Philadelphia, PA 19106
1-215-717-3473 (phone)
1-215-717-3440 (fax)
gdanning@yahoo.com
Did Radio RTLM Really Contribute Meaningfully to the Rwandan Genocide?:
Using Qualitative Information to Improve Causal Inference
From Measures of Media Availability
Abstract
The scholarly literature on the relationship between media and communication
technology and social behaviour is as vast as it is fascinating, and is as
consequential as it is vast. In this article, I focus on a subset of that research, one
which examines the relationship between information which is disseminated
“from above” and political violence, and which employs estimates of media
exposure to explore that relationship. I argue that, while these methods hold
enormous potential for addressing some of the limitations that have long plagued
conflict research, they involve a potential pitfall, i.e., the possibility that the
variable that they measure, media availability, is an inadequate proxy for media
consumption, which is the actual variable of interest.
I further argue that
researchers often cannot be confident that that proxy is a valid one unless they
have a deep qualitative understanding of media consumption habits of the
population under study. I illustrate that concern by examining the findings of
Yanagizawa-Drott (2014), which estimated that roughly ten percent of the
violence which took place in the course of the Rwandan genocide can be
attributed to broadcast of the so-called “hate radio” station, Radio Télévision
Libre des Mille Collines (RTLM).
1
Introduction
The scholarly literature on the relationship between media and communication
technology and social behaviour is as vast as it is fascinating, and is as consequential as it is vast.
Researchers have argued that those technologies can, among other things, destroy social capital
(Olkean 2009), alter the political trajectories of fragile democracies (Adena et al. 2015;
Enikolopov et al. 2011), affect levels of political violence (Pierskalla and Hollenbach 2013;
Shapiro and Weidmann 2015; Warren 2015), and alter both patterns and levels of participation in
political protest (Little 2016; Reuter and Szakonyi 2015; Kern and Hainmueller 2009).
In this article, I focus on the subset of this research, one that examines the relationship
between information which is disseminated “from above” and political violence, and which
employs estimates of media exposure to explore that relationship. I argue that, while these
methods hold enormous potential for addressing some of the limitations that have long plagued
conflict research, they involve a potential pitfall, i.e., the possibility that the variable that they
measure, media availability, is an inadequate proxy for media consumption, which is the actual
variable of interest. I further argue that researchers often cannot be confident that that proxy is a
valid one unless they have a deep qualitative understanding of the media consumption habits of
the population under study. I illustrate that concern by examining the findings of YanagizawaDrott (2014), which estimated that roughly ten percent of the violence which took place in the
course of the Rwandan genocide can be attributed to broadcasts of the so-called “hate radio”
station, Radio Télévision Libre des Mille Collines (RTLM). I find that qualitative evidence
regarding radio consumption habits indeed undermines confidence in those findings, and employ
a cross-validation exercise to determine that inclusion of the radio reception variable in
Yanagizawa-Drott’s model does not improve its out-of-sample predictive performance.
2
Background
The international community has proclaimed that the prevention of mass atrocities is a
central responsibility of states, and has endorsed the concept of the Responsibility to Protect,
which mandates that states take steps to prevent atrocities, including intervening militarily as a
last resort (Petty 2013). However, some states have arguably used that responsibility as an
excuse for interfering in the internal affairs of their neighbors (Mohamed, 2012; Evans 2008). If
the international community is to navigate between the Scylla of under-intervention and the
Charybdis of over-intervention, understanding the causes of violent substate conflict, especially
mass atrocities, is exceptionally important.
Unfortunately, if there is any truism in the field of conflict studies, it is that precise and
accurate understandings of causal processes and other phenomena are exceedingly hard to come
by, in large part because of the difficulty of obtaining data that is sufficiently accurate and
complete (Weidmann 2016, Klayvas 2008; Davenport and Ball 2002). This issue has become
particularly problematic as scholars have come to realise that microanalysis, and hence microlevel data, is essential to understanding violent conflict (Kalyvas 2008).
The problem of causal inference in conflict studies is perhaps especially acute when
examining the causal role of persuasive media. Recent research indicates that propaganda and
other persuasive techniques are much less effective than policymaker and even some scholars
have long presumed (Mercier 2017). In particular, those techniques have often been found to be
effective in motivating behaviour only among those who are already predisposed to believe the
viewpoints being espoused. For example, Adena et al. (2015) found that the effect of Nazi antiSemitic propaganda varied according to the predisposition of listeners, in that it was most
3
effective in engendering denunciations of Jews to authorities and anti-Semitic letters to Der
Sturmer in areas where anti-Semitism was historically high, but had a negative effect in places
where anti-Semitism was historically low. Similarly, Chan et al. (2015) found that broadband
internet availability increases racial hate crimes in the United States only in those areas with
higher levels of racism; Weeraratne (2010) found that elite-orchestrated campaigns in Indonesia
which scapegoated minorities resulted in anti-Chinese riots only when the rhetoric employed
therein was rendered salient by local conditions; and Wilson (2011) employed process tracing to
determine that participants in violent ethno-religious conflicts in Indonesia tended to be acting
out of their own self-interest, rather than being incited by elite propaganda.
However, there are reasons to question whether these findings regarding the effect of
persuasive media hold as strongly in circumstances involving violent conflict or the risk thereof.
First, in those circumstances, propaganda or other persuasive media might operate to increase
violence through some mechanism other than changing listeners’ underlying beliefs, such as by
signaling to those inclined to violence that they will not be punished by authorities (Horowitz
2001).
Second, in conditions of ongoing or incipient political violence, the salience of certain
types of propagandistic appeals might be enhanced; for example, appeals to ethnic identity might
be more persuasive in times of conflict, because identity is often endogenous to ongoing conflict,
as violence often causes individuals to adopt oppositional identities by increasing the salience of
those identities. (Balcells and Steele, 2016; Kalyvas and Kocher 2007; Fearon and Laitin 2000).
Finally, during ongoing violence, even the most extreme appeal to “destroy them to save
us” (Straus 2012; De Forges 1999) might appear to be reasonable, and compliance therewith
might be perceived as rational, because the potential cost of being wrong — i.e., in refusing to
4
believe that the threat is real and therefore declining to engage in violence — is too high (De
Figueiredo and Weingast 1999).
That is one reason why the recent advances in the analysis of information communication
technology are so exciting. By analyzing variations in the exposure of subsets of the populations
under study to information communication technologies, scholars have been able to make
inferences about the relationship between exposure to information and outcomes of interest. For
example, Olkean (2009) employed the fact that Indonesia’s mountainous terrain creates
variations in radio and television signal strength in local villages to analyze the association
between access to signals and villagers’ participation in village governance and self-reported
levels of trust. Similarly, Enikolopov et al. (2011) used temporal and geographic variation in the
availability of the sole independent television station in Russia to estimate its influence on the
electoral fortunes of major opposition parties; Adena et al. (2015) found that variation in radio
exposure in pre-war Germany, as measured in part by radio signal strength, was associated with
variation both in electoral support for the NSDAP and in anti-Semitic acts, conditioned on
listeners’ predispositions toward anti-Semitism; Durante et al. (2017) found that early access to
Silvio Berlusconi’s commercial TV network was associated with greater electoral support for
Mr. Berlusconi's party; Yanagizawa-Drott (2014) found that variations in radio reception among
local Rwandan cells was associated with variations in levels of participation in the genocide of
1994; DellaVigna et al. (2014) found that residents of Croatian villages which had better
reception of Serbian radio broadcasts were more supportive of extreme nationalist parties; and
Gagliarducci et al. (2017) found that variation in reception of BBC broadcast, engendered by
sunspot activity, was associated with variations in levels of resistance activity in Italy in the late
stages of World War II. In contrast, Crabtree et al. (2015) found that there was no association
5
between the availability of West German television broadcasts and the incidence of anti-regime
protests in East Germany, and Warren (2015) found a negative association between radio
ownership and collective violence in Africa.
Scholars of political violence and collective action have also analyzed more
contemporary forms of information communication technology. For example, Bailard (2015),
Warren (2015), and Pierskalla and Hollenbach (2013) each found that greater availability of
cellular phone coverage is associated with a greater probability of violent collective action in
Africa, while Bergren and Bailard (2017) found that an increased availability of cellular phones
is associated with less violent contention in Myanmar, and Shapiro and Weidmann (2015) found
that cellular coverage in Iraq was negatively associated with attacks on government and
Coalition troops. Finally, there is a burgeoning literature on the role of the Internet on violent
contention (Gohdes 2018).
Thus, it is clear that analysis of media and communication technology has the potential to
greatly expand our understanding of the role that communication plays in conflict, a potential
which can only expand as communication technology grows more ubiquitous, and the data it
generates becomes more voluminous. However, a possible pitfall exists in this literature. The
primary independent variable employed in most of these studies is some measure of the
availability of the technology at issue. However, media availability is merely a proxy for the true
variable of interest, which is actual consumption of the medium under study, whether that
medium takes the form of radio broadcasts, television broadcasts, cellular phone usage, or social
media. If media availability is an inadequate proxy for media consumption, then our confidence
in the results of analyses of media and communication technology must perforce be reduced.
6
That problem is, of course, not unique to this particular set of studies; Hegre et al. (2017,
117) note that “many of the ‘usual suspect’ variables [in peace and conflict research] are proxies
that on their own cannot exert causal effects[,]” and Kalyvas (2008, 397) identifies the
“considerable distance between theoretical constructs and proxies” as a “well known” problem of
econometric studies of civil war. Nevertheless, it is a problem that must be addressed, and most
of the above-cited studies attempt to do so by including some direct measures of technology
usage and demonstrating that those direct measures correlate with the measures of availability
which constitute each study’s primary independent variable. For example, Olkean (2009)
included a survey which found that respondents with better television reception tended to spend
more time listening to radio and and watching television; Enikolopov et al. (2011) constructed a
dummy variable from a survey of media consumption habits and determined that increased
availability of television was indeed associated with greater consumption thereof; Adena et al.
(2015) used local radio subscription rates to determine the levels at which predicted signal
strength, which was their primary main explanatory variable, actually resulted in greater levels of
listenership; DellaVigna et al. (2014) conducted a survey and determined that large numbers of
Croats in areas where Serbian radio signals are strong in fact listen to Serbian radio with some
regularity; and Durante et al. (2017) examined survey data on television consumption and
political attitudes, and found that support for Mr. Berlusconi’s party was higher among age
cohorts who tend to watch the most television.
In each of these cases, the author or authors were able to present evidence indicating that
their primary proxy variables, measures of availability of the media in question, were correlated
with more direct measures of their actual dependent variable, consumption of that medium.
7
Hence, in these cases, we have high confidence that the studies are really examining what they
purport to examine.
However, there are many cases in which it is not possible to employ that strategy. It is
the central claim of this article that, in such cases, even the most sophisticated analysis of media
availability is of limited utility, unless the author can draw upon a deep qualitative understanding
of the culture of media use of the population under study in order to increase confidence in the
validity of the proxy variable. I test that claim by examining Yanagizawa-Drott’s (2014) study
of the role of radio in the Rwandan genocide, which found that greater availability of RTLM
“hate radio” broadcasts was associated with greater levels of participation in the 1994 genocide.
Case Study - Yanagizawa-Drott (2014)’s Findings Regarding RTLM and Rwanda
Yanagizawa-Drott (2014) examined the role played by Radio Télévision Libre des Mille
Collines (RTLM) in the Rwandan genocide. That issue has been of particular interest to both
scholars and policymakers because, as Nyseth Brehm (2017) observes, “[a]lthough the genocide
was orchestrated by the state, hundreds of thousands of civilians implemented the killing, and
viewing the violence strictly as a top-down endeavor ignores the other factors that influenced
their participation” (23). Hence, understanding who participated in the Rwandan genocide, and
why, is clearly a significant question.
Background – Radio and the Rwandan genocide
Until the Summer of 1993, Rwanda had only one radio station, the government-operated
Radio Rwanda (Des Forges 1999, 58). RTLM, which began broadcasting to the entire country in
8
July of 1993 (Mironko 2007, 126), was created by Hutu hard-liners; although the station was
officially distinct from the government, it used many of the same staff and much of the same
equipment as Radio Rwanda, and many of its founders had close connections with high
government officials (Des Forges 1999, 59). Unlike Radio Rwanda, RTLM was designed to be
highly entertaining, featuring popular music, call-in talk shows, and gossip, and it quickly
became extremely popular (Des Forges 1999, 60; Petrova and Yanagizawa-Drott 2016).
Prior to the 6 April 1994, downing of the aircraft carrying Rwandan President Juvénal
Habyarimana, RTLM broadcasts included statements by radio personalities and political figures
which, among other things, framed Rwandan history as the oppression of the Hutu majority by
the Tutsi minority; often used the terms “Inyenzi” (cockroach) to refer to members of the
ongoing Tutsi insurgency and even to Tutsis in general; conflated Tutsi rebels with Tutsis in
general; depicted the Tutsi insurgency and the Tutsi in general as existential threats to the Hutu;
engaged in ethnic stereotyping; and accused specific Tutsi civilians and their families as being in
league with the insurgents (The Prosecutor v. Ferdinand Nahimana 2003, 118-133). After the
death of President Habyarimana and the subsequent initiation of the genocide, RTLM broadcasts
began to explicitly define as enemies the Tutsis as a whole, rather than just the rebels;
increasingly referred to Tutsis as Inyenzi; began calling for the killing or “extermination” of
Tutsi civilians; and specifically named as enemies particular persons, many of whom were
subsequently killed (The Prosecutor v. Ferdinand Nahimana 2003, 133-158). In some instances
after the killing began, RTLM gave specific directions regarding where to find particular targets
(Des Forges 1999, 158), and even broadcast the license plate numbers of those attempting to
escape (Kirschke 1996).
9
Yanagizawa-Drott’s Findings
As noted by Wilson (2015), in comparison to most studies of RTLM’s role in the
genocide, Yanagizawa-Drott’s (2014) is a relative outlier. Yanagizawa-Drott tested the
hypothesis that higher levels of consumption of RTLM broadcasts were associated with higher
levels of participation in the genocide. He measured participation in genocide by the number of
prosecutions in each of the local cell-level Gacaca courts, which were established by the
Rwandan government in 1997 to prosecute genocide participants (Schabas 2005). The Gacaca
courts divided suspects into three categories: Category 1 suspects were those accused of being
planners or organizers, government officials, leaders who participated or who urged, coerced, or
incited others to participate, and those accused of rape or sexual torture; Category 2 suspects
were “notorious murderers,” individuals accused of murder, torture or the defilement of bodies,
and their accomplices; and Category 3 suspects were those accused of property crimes (Nyseth
Brehm et al. 2014). Although most killers acted in groups (Fujii 2009, 7), Yanagizawa-Drott
maintains that most Category 1 offenders at the cell level were members of local militia
members (Yanagizawa-Drott 2014, 1959), and hence for convenience he denominates Category
1 killings as “militia” killings, and Category 2 killings as “individual” killings (YanagizawaDrott 2010, 1960).
Yanagizawa-Drott’s data included a total of roughly 77,000 persons who were prosecuted
for Category 1 (“milita”) violence, and approximately 432,000 persons who were prosecuted for
Category 2 (“individual”) violence. In order to measure cell-level consumption of RTLM
broadcasts, Yanagizawa-Drott employed a proxy variable, a measure of radio reception in each
of the cells. Because of the very hilly topography of Rwanda, which is known colloquially as the
10
“land of a thousand hills” (Carr and Halsey 2000), levels of reception varied greatly from village
to cell. According to Yanagizawa-Drott’s data, in the median cell1 only ten percent of the cell
received radio signals; the mean was .19 and the standard deviation was .23. Just over twenty
percent of the cells had no reception at all, and in only 104 of 1065 cells did at least fifty percent
of the cell have reception. Yanagizawa-Drott employed a fixed-effects model with controls for
various potential determinants of violence in order to test his hypothesis that higher levels of
radio reception were associated with higher levels of participation in the genocide.
Yanagizawa-Drott found that levels of participation were indeed higher in cells with
greater radio reception. He estimated that RTLM broadcasts increased overall participation by
approximately ten percent; that it increased the incidence of Category 1 (“militia”) violence
participation by almost a third; and that there was a “spillover” effect for militia violence, but not
individual violence, in that the likelihood of a person in a given cell engaging in militia violence
was significantly higher when radio reception in neighboring cells was higher.
Potential weaknesses in Yanagizawa-Drott’s proxy for radio consumption
Yanagizawa-Drott’s findings are broadly consistent with what we know about RTLM
broadcasts during the genocide, in that on several occasions the station broadcast specific
instructions regarding exactly where to find particular victims (Des Forges 1999, 158; Kirschke
1996). Specifically, the fact that RTLM sometimes broadcast such specific instructions is very
consistent with Yanagizawa-Drott’s finding that spillover effects were confined to Category 1
violence, if Yanagizawa-Drott is correct in stating that most Category 1 participants were militia
Yanagizawa-Drott uses the term, “village,” but the administrative level at which Gacaca courts
were established is denominated the “cell” in Rwanda (Government of Rwanda 2001, Article 3;
Government of Rwanda 2004, Article 3) .
1
11
members, because militia members might have gone to neighboring cells in search of particular
named victims, and committed the crimes there, which would have subjected them to Gacaca
prosecution in that cell.2 In addition, militia members who traveled to neighboring cells in
search of named victims might have recruited or coerced residents there to participate.
However, there are some reasons to be wary of Yanagizawa-Drott’s findings.
First, as noted previously, his findings are a bit of an outlier, as other studies have
struggled to find a significant connection between RTLM broadcasts and participation in the
genocide.
Second, Yanagizawa-Drott uses cell-level data on radio reception. As Warren (2015)
notes, analyses of the same phenomenon that are based on different geographical units of
analysis can sometimes yield different empirical inferences (299). And, indeed, Nyseth Brehm
(2017) analyzed Yanagizawa-Drott’s data at the commune level3, and found no significant
association with the level of killing.
Third, data on radio ownership in Rwanda indicates that radio consumption might have
been greater than implied by Yanagizawa-Drott’s data on radio reception. The 1991 Rwandan
census found that 34% of households owned radios (IPUMS International 2012), yet
Yanagizawa-Drott’s data indicates that only 19.8% percent of the population had radio reception.
While radio ownership is not a perfect proxy for radio consumption — Spitulinik (2002) found
that eight percent of radios owned by Zambians are broken at any given time, and that an
additional four percent are inoperative because they lack batteries — nevertheless a wide
discrepancy between ownership and reception raises a suspicion that radio reception data does
2
The jurisdiction of the Gacaca courts was based on where the crimes were committed, rather
than where the defendants resided. (Organic Law 2004, Art. 44-45).
3
The commune was a now-defunct administrative unit, each of which was subdivided into
several sectors; each sector was divided into several cells.
12
not fully capture actual levels of radio consumption. The fact that Warren (2015) found a
negative association between radio ownership and collective violence in Africa at the country
level of course reinforces that suspicion to some degree.
Finally, unlike many of the studies discussed herein, Yanagizawa-Drott was unable to
investigate whether his measure of radio reception is correlated with measures of radio
consumption.
Qualitative studies of radio consumption habits in Sub-Saharan Africa cast additional doubt
on the validity of Yanagizawa-Drott’s proxy for radio consumption
Therefore, there is some doubt whether radio reception is a sufficiently valid proxy for
Yanagizawa-Drott’s variable of interest, radio consumption. That doubt might well be dispelled
if there is qualitative evidence regarding Rwandan media consumption patterns indicating that
radio reception is likely to be closely correlated with radio consumption; however, if anything,
the evidence trends in the opposite direction. There was little specific research on Rwandan
radio listening habits in the pre-genocide period, but research conducted in analogous regions
indicates that radio reception data might seriously underestimate levels of radio consumption,
because it appears that that radio consumption in Sub-Saharan Africa is very much both a social
and a mobile activity.
The leading exploration of radio consumption habits in Sub-Saharan Africa is
Spitulinik’s (2002) study of 1990s Zambia. Spitulinik found, among other things, that Zambian
radio consumption practices were shaped by “the sociality of domestic space. … In situations
where there are portable radios and batteries people listen outside during daylight. Such
listening occurs in an outdoor workspace, under a shade tree, on the stoop of the home, at the
13
marketplace, or on the road” (313). She also noted that “... the tendency is for most daytime
social activity and a great deal of domestic labor to take place outdoors; thus, the portable radio
moves as people move” (313), and that “[p]eople carry radios to the office when presidential
press conferences or football matches are to be broadcast, and they are sometimes brought into
the field to listen to while farming” (314). In addition, “[r]adios are on in public spaces such as
bus stops, minibuses, shops, foodstands, bars and markets” (314), and when a visitor brings a
radio to an area where radios are scarce, it “temporarily becomes a collective asset within his
wider kin and friendship networks” (315). Finally, Spitulinik found that “... rural men were 1.5
times more likely to listen to radio than rural women …” (312).
Spitulinik’s findings are echoed by those of Ambler (2002), who notes that, again in
Zambia, radio listening became “less of a conscious leisure activity and more an accompaniment
to work or socializing” (131), and also that radio listening became popular during World War
Two, when “concern about global warfare once again impinging on Northern Rhodesia created a
substantial audience for world news” (133).
Gathigi (2009) studied radio listening habits in rural Kenya and also found that “[m]ost
men in Kieni West used radio as a companion in their places of work” (142) and that
men were more likely to carry the radio when going to work in the fields than were women
(142). Indeed, he quotes one 68-year-old male respondent as claiming that his portable radio
“has become a part of me” (142).
These findings tend to further undermine confidence in the utility of a measure of radio
reception as a proxy for radio consumption in Rwanda. If radio owners in Rwanda, like radio
owners in Zambia and Kenya, carried their radios to work in the fields in the period during which
RTLM was broadcasting, then variation in consumption of those broadcasts from cell to cell
14
might be much less than is implied by levels of radio reception. For example, YanagizawaDrott’s data indicates that only about five percent of the cell of Vumbi in the Runyinya commune
had radio reception; in contrast, radio reached 13% of the cell of Kibingo, which was in the same
commune. Yet, if residents of both cells spent their days tending fields at relatively high
altitude areas, where radio reception is at its best, then they might in fact have had roughly
equivalent levels of consumption RTLM broadcasts. The evidence that men are more likely than
women to listen to radio at work in rural occupations is significant as well, since the majority of
the participants in the genocide were male (Adler et al. 2007). Hence, radio consumption by the
actual participants in the genocide might vary less than is implied by measures of radio
reception.
In addition, because radio consumption is often a social activity in Sub-Saharan Africa,
residents might have tended to gather in the evenings in those few locations with good radio
reception, as has historically often been the case in rural areas worldwide when media becomes
available (Opt 1992; Roberts 1995; Leguy 2007). If so, then this again implies that residents’
consumption of radio broadcasts might vary less among cells than is indicated by data on
reception.
This evidence regarding media consumption patterns in other sub-Saharan African
countries is of course not dispositive of those habits in Rwanda. Because the primary crops in
Rwanda during the period before the genocide were not necessarily the same as those in Zambia
and Kenya – for example, in Rwanda the primary cash crops were coffee and bananas (Verwimp
2003, 173-174), while in Zambia in the 1990s and early 2000s they were cotton and groundnuts
(FAO/WPP 2002, 3), and in Kenya major crops included maize, sugarcane, tea and also coffee
(Kibaara et al. 2008) –work patterns and hence media consumption habits during working hours
15
in Rwanda might differ from those in Zambia and Kenya. Moreover, unlike Zambia and Kenya
during the years covered by the studies thereof cited herein, Rwanda was experiencing serious
internal armed conflict in the years before the genocide, which of course is likely to have been
disruptive of all manner of settled patterns of residents’ social and economic life, including
media consumption habits related thereto. However, such disruption would not have been either
uniform nor universal, as noted by Sperling (1997), who reports that “[t]ime spent away from the
homestead . . . was on average, four months, although this varied greatly by region (from 3-4
weeks in the southwest to 54.6 in the northwest) . . . [and] 30% of those still farming in Rwanda
at the end 1995 had not been displaced at all, not even for a single day” (21).
However, there is evidence from Rwanda which indicates that the general patterns of
media consumption observed in Zambia and Kenya were common in Rwanda as well. First, as
was the case elsewhere in Africa, Rwandans listened to radio “in bars and at work, and . . . you
could hear it in taxis and at the market” (The Prosecutor v. Ferdinand Nahimana 2003, 117).
Indeed, as elsewhere, in Rwanda “[r]adio bookend[ed] and punctuate[d] the daily routine of
many ordinary Rwandans” (Li 2004, 20), and “‘people who did not have radios went to someone
else’s house to listen to the radio’” (Bonnier et al. 2016, 25 n. 33).
Second, many crops in Rwanda are cultivated on slopes — as of 2007, 77 percent of
cultivated land in Rwanda had slopes between 13% and 55% (Republic of Rwanda 2007) —
which implies that agricultural labor took place not in valley bottoms but rather higher up, where
radio reception is generally better. Hence, a cell that has low radio reception by YanagizawaDrott’s measure might actually have a relatively large percentage of its populace living in the
small areas where reception is good.
16
Third, Rwanda experienced a crash in commodity prices and a serious drought in the
early 1990s (Newbury 1998). Rural residents in such circumstances are likely to be driven to
travel further afield in search of land to cultivate or other income activities. This implies that
radio reception in participants’ home cells might not accurately estimate their actual exposure to
radio broadcasts. Moreover, Verwimp (2005) found that Hutu households with at least one
participant in genocide averaged twice the number of days worked off the farm than did Hutu
households with no participants. It is of course possible that the factors which caused Hutu
individuals to seek work off the farm also caused such individuals to be more likely to participate
in genocide. But, it is also possible that the experiences of those who worked off the farm had an
independent effect on genocide participation rates. If exposure to RTLM broadcasts was one
such factor, then the effects of that exposure cannot be estimated by a measure of home cell
reception.
Fourth, in the two decades before the initiation of the genocide, the Rwandan government
had implemented a resettlement policy, the Payasannat program, in which rural residents were
encouraged to move either to hillsides or to valley bottoms; by the early 1990s, an estimated
76% of rural households had done so (Niazi 2002). Valley bottoms, of course, have poorer radio
reception than hillsides. That is not problematic if the percentage of households located in valley
bottoms in each cell is randomly distributed vis-a-vis estimates of radio reception. In other
words, variation in cell-wide radio reception is a good proxy for variation in household
consumption only if the proportion of the households living in valley bottom is unrelated to radio
reception patterns. However, the Longley-Rice propagation model employed by YanagizawaDrott includes local terrain profiles, including terrain slope (Kasampalis et al. 2013; Weiner
2005). Hence, if residents’ decisions on whether to locate their households in valley bottoms
17
rather than hillsides is also related to the nature of the local hillside terrain, which seems
plausible, then there might be systematic biases in the variation between the propagation model’s
estimate of reception and actual reception, and hence consumption, of broadcasts.
Finally, the years before the genocide had of course been marked by an ongoing
insurgency; in such periods of instability and potential threat, it is likely that those without radio
reception at home would make a special effort to listen to news by finding neighbors or local
gathering places that had reception (Bratić 2005, 4).
For all of these reasons it is entirely possible that variation in radio consumption among
Rwandan cells is less than the estimated variation in radio reception which is generated by the
propagation model employed by Yanagizawa-Drott. That raises the very real possibility that
Yanagizawa-Drott’s findings are somewhat less convincing than they appear. That possibility is
all the more likely because previous studies of the Rwandan genocide tend not to support
Yanagizawa-Drott’s findings.
Previous studies of the Rwandan genocide have generally downplayed the role of RTLM
Although the popular media has blamed RTLM for causing or at least greatly
exacerbating the genocide by generating hatred against the Tutsi (Smith 2003; Dalliare 2003,
272; Melvern 2000, 71), most researchers have found little direct evidence to support that claim.
First, qualitative studies of participation in the Rwandan genocide generally minimize the
role of RTLM in causing the violence, and find that face-to-face interactions with peers and
community members were key to catalyzing participation. For example, Straus (2007)
interviewed 210 sentenced and self-confessed perpetrators, and found that most were mobilized
by local actors (626). Similarly, Fujii (2009) found that social ties were the primary
18
determinants of how individuals responded to the genocide, which social ties “were the product
of talking, gossiping, greetings, visiting, sharing beers, and participating in umuganda [the
obligatory communal labor system]” (128); and McDoom’s (2013) study of Category 2
participants in one sector of Rwanda found that participants often resided in neighbourhoods, and
were members of households, with high concentrations of other participants. He concluded that
“horizontal pressure” from peers was the primary determinant of participation; in contrast, he
found that the evidence for pressure from above was merely “tentative” (462). Straus (2007) also
notes that the most extreme RTLM broadcasts did not take place until the later stage of the
genocide, after the bulk of the killing had already taken place (622)
McDoom (2013) also discusses the onset of violence in the Tare section, which at the
time of the genocide was part of the Maraba commune (458). He notes that the violence in the
Maraba commune, which was the focus of his study, did not begin until nearly two weeks after
the assassination of President Habyarimana, but only the day after Rwanda’s new president
visited the commune; within four days, most of the local Tutsi population had been eliminated
(459). That implies that the RTLM broadcasts had little effect in Maraba, and that instead it was
in-person organizing that was the driving force of participation.
Finally, Nyseth Brehm (2017) examined variation in killing at the commune level and
found that neither RTLM coverage nor an interaction between radio coverage and radio
ownership was significantly associated therewith, but instead that higher levels of violence were
associated with lower levels of social control and cohesion. However, Nyseth Brehm did note
that an earlier study (Nyseth Brehm 2014) had found that greater Radio RTLM coverage was
associated with earlier onsets of violence at the commune level.
19
Qualitative studies of the genocide also cast more direct doubt on the assumption that
reception of RTLM is a valid proxy for consumption thereof. For example, Mironko (2007)
interviewed 100 perpetrators of genocide and found that although many did not listen to RTLM
broadcasts, nevertheless oftentimes “[t]hey heard the messages from others” (134), and Li (2004)
notes that “[b]roadcasts were often reincarnated elsewhere as rumour,” and relates the story of a
militia member who “used to spend mornings on the roof of his shop with a radio clutched to his
ear, listening to RTLM” and who would subsequently “climb down and gather people to tell
them what he had heard” (19). That final anecdote is a particular telling example of the dangers
of assuming that radio reception is a valid proxy for radio consumption; the individual in
question presumably spent hours on his roof because radio reception in his area was poor, yet he
and his neighbors were nevertheless able to consume the broadcasts and the contents thereof.
However, it is possible to overstate the extent to which these studies are inconsistent with
Yanagizawa-Drott. For example, McDoom’s (2013) finding that participants tended to be
concentrated in certain neighborhoods and households is perfectly consistent with YanagizawaDrott’s finding that variation in radio reception helps explain variation in rates of participation,
since the neighborhoods and households with many participants could easily be those with better
radio reception. In addition, according to Yanagizawa-Drott’s data, the commune in which
McDoom conducted his research had rather poor radio reception. Hence, the fact that residents
thereof did not engage in killing until prompted by a visit by the President does little to
undermine Yanagizawa-Drott’s argument. Finally, Yanagizawa-Drott’s broad findings are
consistent with Mironko’s (2007) finding that any impact of RTLM was likely far greater in
Kigali than in rural areas, in that Kigali had more broad radio reception than most rural areas.
20
Most importantly, Yanagizawa-Drott’s findings are relatively modest – he estimates that
ten percent of the overall violence can be attributed to the RTLM broadcasts – and most of the
qualitative work on the genocide concludes that radio probably had at least some effect on
participation. For example, although Straus (2007) “conclude[s] that radio alone cannot account
for either the onset of most genocidal violence or the participation of most perpetrators[,]” he
nevertheless “find[s] some evidence of conditional media effects” (611), especially among more
hard-core participants, although he notes that “[t]he causal arrows remain unclear” (628).
Similarly, Mironko (2007) concludes only that RTLM and other government-sponsored
messages “alone did not cause them to kill” (134, emphasis in original), and Li (2004) maintains
that RTLM broadcasts probably increased listeners’ tendencies toward participation by
“appropriate[ing] and transform[ing] elements of three of the dominant public discourses of postcolonial Rwandan modernity” (12-13), and by “implicat[ing] ordinary listeners in the activities
of the genocide” (18).
Hence, Yanagizawa-Drott’s findings are not entirely inconsistent with prior research.
Quantitative analysis of the predictive ability of Yanagizawa-Drott’s proxy for radio
consumption
It appears that, while Yanagizawa-Drott’s findings are not necessarily inconsistent with
previous research, nevertheless there is reason to question the validity of his proxy for radio
consumption. As a result, his data might misstate the variation in radio consumption among
cells, and hence his estimates of the effect of RTLM broadcasts on participation might be
inaccurate. Had Yanagizawa-Drott been able to gather qualitative evidence regarding radio
consumption habits in Rwanda, we could be more confident that his proxy captures the true
21
variation in RTLM consumption among cells. In the absence of that qualitative evidence, it is
useful to conduct an additional quantitative analysis of his data, in order to determine whether
higher levels of radio reception predict higher levels of participation in the genocide.
If radio reception is, in fact, a valid proxy for radio consumption, and if consumption or
RTLM broadcasts did, indeed, cause increased participation in the genocide, then higher levels
of radio reception should predict increased genocide participation. If, on the other hand,
measures of radio reception are poor proxies for radio consumption, as I hypothesise, or if radio
consumption had no causal effect on genocide participation, then higher levels of radio reception
should not predict higher levels of genocide participation. This is because, although there is an
important distinction between prediction and causal inference (Cederman and Weidmann 2017;
Shmueli 2010), and of course some predictive variables are not causal (Gohdes and Carey 2017),
a variable’s failure to increase the predictive power of a model does reduce confidence in its role
as a causal factor (Chenoweth and Ulfelder 2017; Hegre et al. 2017; Muchlinski et al. 2015;
Wischnath and Buhaug 2014; Schneider et al. 2011).
An analysis of the predictive power of RTLM reception in Yanagizawa-Drott’s model is
also important because Ward et al. (2010) found that statistically significant variables in models
of conflict onset are often poor predictors of onset. Hence, it cannot be assumed that RTLM
reception in fact has predictive power, merely because Yanagizawa-Drott found it to be a
statistically significant independent variable.
Therefore, following Ward et al. (2010), I assessed the predictive power of radio
reception on genocide participation by employing Yanagizawa-Drott’s fixed effects model. The
dependent variable (Gacaca court prosecutions), the primary independent variable (radio
22
reception) and all control variables are those employed by Yanagizawa-Drott, who graciously
provided his replication data and code.
Ward et al. (2010) measured the predictive power of conflict models using a calculation
of the area under the Receiver Operating Characteristic (ROC) curve for the test set. The ROC
curve was developed during World War II to assess the effectiveness of radar, and was
subsequently applied to the field of medical diagnostics (Streiner and Cairney 2007). In both
radar detection and medical diagnostics, there is a tradeoff between false positives (for example,
a medical test which indicates the presence of disease when the patient is healthy) and false
negatives (for example, a medical test that fails to detect disease in a patient who in fact has the
disease). The ROC curve is a graph of the predictor’s true positive rate (the number of correctly
predicted events divided by the total number of cases where the event in fact occurred) against
the false positive rate (the number of incorrect predictions divided by the total number of cases
where the event did not occur) (Ward et. al 2010, 366). The area under the ROC curve is
frequently used to assess the overall predictive accuracy of a model. The area’s value ranges
from 0.5 to 1.0, and it is a calculation of how much better the model does at predicting outcomes
than does random chance; for example, an area of 0.5 indicates that the model only outperforms
a random guess fifty percent of the time, and hence that it performs no better than a random
guess (Ward et al. 2010, 266-267).
Because ROC analysis requires a binary dependent variable, I instead employed the
method developed by Obuchowski (2005, 2006) for conducting ROC-style analysis on models
with a continuous dependent variable, using the NonbinROC package in R (Nguyen 2007).
Following Ward et al., I first assessed the predictive power of Yanagizawa-Drott’s full model,
and then dropped the radio reception variable from the model and reran the analysis. If the
23
predictive power of the truncated model dropped substantially, that would imply that radio
reception improves the model’s predictive ability, and hence that high levels of radio reception in
a given cell (in Yanagizawa-Drott’s parlance, village) predicts high levels of Gacaca
prosecutions in that cell; if the truncated model performed just as well as the new model, then
that implies that high levels of radio reception does not predict high levels of prosecution, which
thus implies either that radio reception is a poor proxy for radio consumption, or that radio
consumption did not increase genocide participation.
A common challenge in assessing the predictive accuracy of models is that, often, all of
the available data is used to generate the model, leaving no data which can be used to test the
ability of the model to predict outcomes based on out-of-sample data. The solution to this
dilemma is to use K-fold cross-validation, in which the data is randomly divided into k equal
sized subsamples. One subset is withheld, and the remaining k-1 subsets are combined into a
single training set. The training set is used to train the model, and then the model’s predictive
ability is then assessed using the remaining data, i.e., the test set. The procedure is then
repeated k times, with a different subset being used each time as test set. As Ward et al. (2010)
note, the results of that process can sometimes be dependent on the initial random partition;
hence, the best practice is to repeat the entire process numerous times, and to then average the
results (370).
Following Ward et al. (2010), I employed 4-fold cross-validation;4 however, while Ward
et al. repeated the process ten times, I found that repeated iterations of ten repeats yielded
slightly different outcomes, so I instead repeated the process on one hundred different random
partitions of the Yanagizawa-Drott data, and averaged the results.
4
Increasing the number of folds to ten did not meaningfully change the results, which were the
same within two decimal places.
24
I compared the results of the continuous ROC analysis on models with and without the
radio reception variable by dividing the four hundred outcomes from the continuous ROC
analysis into ten sets of forty each, and employing the paired t-test to calculate a p-value for each
set. I repeated that analysis on ten different random sorts of the continuous ROC outcomes, and
averaged the results. I chose the paired t-test because it is the most conservative of many
options; that is, it has high power but also high rates of Type I error (Dietterich 1998), and hence
is unlikely to underestimate any improvement in prediction from including the RTLM variable in
the model.
The results of that analysis are set forth in Table 1.
Table 1
Results of Continuous ROC analysis
Y-D Model
Model w/o RTLM
Paired t-test p-value
All violence
0.644
0.645
0.202
Category 1 violence
0.637
0.637
0.456
Category 2 violence
0.641
0.635
0.143
These results are interpreted as follows (Nguyen 2007, 5):
Regarding total violence, of two randomly chosen cells, a) there is a 64.4% chance that
the cell with greater total participation in violence will have a larger predicted participation from
the full model than the cell with less total participation in violence; b) there is a 64.5% chance
that the cell with greater total participation in violence will have a larger predicted participation
from the model that omits radio reception than the cell with less total participation in violence;
and c) the difference between the two results is not statistically significant.
25
Regarding Category 1 violence, of two randomly chosen cells, a) there is a 63.7% chance
that the cell with greater participation in Category 1 violence will have a larger predicted
participation in Category 1 violence from the full model than the cell with less participation in
Category 1 violence; b) there is a 63.7% chance that the cell with greater participation in
Category 1 violence will have a larger predicted participation in Category 1 violence from the
model that omits radio reception than the cell with less participation in Category 1 violence; and
c) the difference between the two results is not statistically significant.
Regarding Category 2 violence, of two randomly chosen cells, a) there is a 64.1% chance
that the cell with greater participation in Category 2 violence will have a larger predicted
participation in Category 2 violence from the full model than the cell with less participation in
Category 2 violence; b) there is a 63.5% chance that the cell with greater participation in
Category 2 violence will have a larger predicted participation in Category 2 violence from the
model that omits radio reception than the cell with less participation in Category 2 violence; and
c) the difference between the two results is not statistically significant.
The results indicate that the out-of-sample predictive power of the model without radio
reception is virtually the same as the the out-of-sample predictive power of the full model with
radio reception included.
As Ward et al. (2010) note, the fact that statistically significant variables do not improve
the model’s ability to make correct predictions indicates that “something is amiss” (Ward et al.
2010, 372). Given the qualitative evidence discussed herein regarding the culture of radio
consumption in Sub-Saharan Africa in general and in Rwanda in particular, it seems likely that
what is “amiss” is that radio reception is an insufficiently precise proxy for radio consumption in
period before and during the Rwandan genocide.
26
Conclusion
The findings herein do not, of course, mean that the RTLM broadcasts had no effect on
the genocide. Indeed, there is good reason to think that they had some effect on specific
instances of killing, given that there is evidence that the broadcasts were used explicitly to
coordinate the actions of some of the genocidaires (Des Forges 1999; Kirschke 1996). In
addition, because, as Straus (2015) notes, elites do not need large scale popular participation for
genocide to take place but rather only need “popular compliance — they need citizens not to
mobilise to resist violence” (66), it is certainly plausible that exposure to anti-Tutsi propaganda
might have made implementation of the genocide easier, such as by reminding listeners of the
founding, exclusionary ideology of the country (Straus 2015), by causing civilians to internalise
norms that legitimised the killing (Smeulers and Hoex 2010), or by spurring the adoption of
oppositional identities (Balcells and Steele, 2016; Kalyvas and Kocher 2007; Mandami 2001;
Fearon and Laitin 2000).
What the findings do imply, however, is that quantifying the effect of media and
communications technologies is even more challenging than previously thought. New
methodologies such as the use of radio propagation models and the analysis of “big data”
generated by social media platforms hold great promise, but their use must be coupled with a
qualitative understanding of the society and the medium in question.
That understanding can help researchers choose the best tools and methods for their task.
For example, because televisions are less mobile than radios, an analysis of television
broadcasting reception is likely to be a better proxy for television program consumption than is
an analysis of radio reception. Similarly, because people obviously have greater ability to move
27
through space than through time, an analysis that relies on temporal variation in media
availability is likely to yield a more precise measure of media consumption than is an analysis
that relies on geographic variation in media availability.
Moreover, scholars who employ measures of media availability should make every effort
to also employ surveys or other instruments which can help to verify that media availability is
indeed closely correlated with media consumption; the ability of Adena et al. (2015), DellaVigna
et al. (2014), Durante et al. (2017), Enikolopov et al. (2011), and Olkean (2009) to do so greatly
enhances DellaVigna et al. (2014) conducted a survey and determined that large numbers of
Croats in areas where Serbian radio signals are strong in fact listen to Serbian radio with some
regularity; and Durante et al. (2017) the reader’s confidence in their findings.
Attention to the need for a qualitative understanding of media consumption habits is also
likely to enhance the ability of future studies which employ analysis of information
communication technology to contribute to the burgeoning literature on the role in political
violence of such factors as social ties, McDoom (2014); communication networks, Larson and
Lewis (2018); norm diffusion, Fujii (2004); and framing, Luft (2015), Shaikh (2016).
Finally, understanding the predictive power of models of political violence and of the
variables included therein has important implications for public policy, including policies
concerning prevention of mass atrocities, the limits of humanitarian intervention, and in the case
of the role of the media in engendering violence, freedom of speech. (Nanlohy et al. 2017;
Petrova and Yanagizawa-Drott. 2016; Dafoe and Lyall 2015; Muchlinski et al. 2015; Buhaug et
al. 2014; Ward et al. 2010).
Moreover, the media, including social media, has been hypothesized to play important
roles both in exacerbating violence and in advancing respect democracy and human rights. As
28
an example of the former, rumour has long been recognized as playing a role in sparking
violence (Bhavnani et al. 2009; Horowitz 2001), and authorities have sometimes claimed that
media can exacerbate that tendency (Polianskaya 2018; Ngdona 2017). On the other hand, many
activists argue that social media can be a valuable tool for atrocity response, mitigation, and
documentation (Tuckwood 2014), and that it can enhance citizens’ political engagement (Gil de
Zúñiga and Valenzuela 2011).
Therefore, understanding the precise role of communication technologies has the
potential to improve policy, including policies designed to intervene in order to forestall violence
(Kogen 2013). Research on communication technologies which is attentive to local media
consumption habits therefore both improve our understanding of past events, and assist
policymakers to design more effective policies.
Acknowledgments
I am indebted to Cyril Ghosh, Vincent La, and two anonymous reviewers for their
insightful comments and recommendations on earlier drafts of this article.
Declaration of Interests
The author(s) declared no potential conflicts of interest with respect to the research,
authorship, and/or publication of this article.
Funding
This article was researched and written in the course and scope of the author’s
employment with the Foundation for Individual Rights in Education. The author’s position at
29
the Foundation for Individual Rights in Education was funded by a grant from the John
Templeton Foundation (grant no. 59218).
Replication data
This analysis was conducted using the statistical programming language R, and the
statistical program Stata. Full replication data are available at
https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/W1REJL.
References
Adena, M., R. Enikolopov, M. Petrova, V. Santarosa, and E. Zhuravskaya. 2015. “Radio and the
Rise of the Nazis in Prewar Germany.” The Quarterly Journal of Economics 130, no.4:
1885-1939.
Adler, R. N., C. E. Loyle, and J. Globerman. 2007. “A Calamity in the Neighborhood: Women's
Participation in the Rwandan Genocide.” Genocide Studies and Prevention 2(3): 209233.
Ambler, C. 2002. “Mass Media and Leisure in Africa.” The International Journal of African
Historical Studies 35(1): 119-136.
Balcells, L., and A. Steele. 2016. “Warfare, Political Identities, and Displacement in Spain and
Colombia.” Political Geography 51: 15-29.
Bailard, C. S. 2015. “Ethnic Conflict Goes Mobile: Mobile Technology’s Effect on the
Opportunities and Motivations for Violent Collective Action.” Journal of Peace
Research 52, no 3: 323–337.
30
Baisley, E., 2014. “Genocide and Constructions of Hutu and Tutsi in Radio Propaganda. Race &
Class 55, no. 3: 38-59.
Bergren, A., and C. S. Bailard. 2017. "Information and Communication Technology and Ethnic
Conflict in Myanmar: Organizing for Violence or Peace?." Social Science Quarterly 98,
no.3: 894-913.
Bhavnani, R., Findley, M.G. and Kuklinski, J.H., 2009. “Rumor Dynamics in Ethnic
Violence.” The Journal of Politics 71, no. 3: 876-892.
Bonnier, E., J. Poulsen, T. Rogall and M. Stryjan. 2016. Preparing for Genocide: QuasiExperimental Evidence from Rwanda. Stockholm Institute of Transition Economics
Working Paper No. 31.
Bratić, V. 2005. Media Effects During Violent Conflict: Evaluating Media Contributions to
Peace Building.” Conflict & Communication Online 4(2).
Buhaug, H., L. Cederman, and K. S. Gleditsch. 2014. “Square Pegs in Round Holes: Inequalities,
Grievances, and Civil War." International Studies Quarterly 58(2): 418-431.
Carr, R. H., and A. H. Halsey. 2000. Land of a Thousand Hills: My Life in Rwanda. New York:
Penguin.
Carver, Richard. 2000. “Broadcasting and Political Transition: Rwanda and Beyond." African
Broadcast Cultures: Radio in Transition, edited by Richard Fardon and Graham Furniss,
188-197. Oxford: James Currey.
Cederman, L., and N. B. Weidmann. 2017. “Predicting Armed Conflict: Time to Adjust Our
Expectations?.” Science 355 (6324): 474-476.
Chan, J., A. Ghose, and R. Seamans. 2016. “The Internet and Racial Hate Crime: Offline
Spillovers from Online Access.” MIS Quarterly 40, no. 2: 381-403.
31
Chenoweth, E., and J. Ulfelder. 2017. “Can Structural Conditions Explain the Onset of
Nonviolent Uprisings?” Journal of Conflict Resolution 61(2): 298-324.
Crabtree, C., D. Darmofal, and H. L. Kern. 2015. “A Spatial Analysis of the Impact of West
German Television on Protest Mobilization During the East German Revolution.”
Journal of Peace Research 52, no. 3: 269–284.
Dafoe, A., and J. Lyall. 2015. “From Cell Phones to Conflict? Reflections on the Emerging ICT–
Political Conflict Research Agenda." Journal of Peace Research 52(3): 401-413.
Dallaire, R. with B. Beardsley. 2003. Shake Hands with the Devil: The Failure of Humanity in
Rwanda. Toronto: Random House.
Davenport, C., and P. Ball. 2002. "Views to a Kill: Exploring the Implications of Source
Selection in the Case of Guatemalan State Terror, 1977-1995." Journal of Conflict
Resolution 46, no. 3: 427-450.
De Figuerido, R., and B. R. Weingast. 1999. “The Rationality of Fear: Political Opportunism and
Ethnic Conflict.” Civil Wars, Insecurity and Intervention, edited by Barbara Walter and
Jack Snyder, 261-302. New York: Columbia University Press.
DellaVigna, S., R. Enikolopov, V. Mironova, M. Petrova, and E. Zhuravskaya. 2014. "Crossborder Effects of Foreign media: Serbian Radio and Nationalism in Croatia." American
Economic Journal: Applied Economics 6, no. 3: 103-132.
Des Forges, A. 1999. Leave None to the Tell the Story: Genocide in Rwanda. New York: Human
Rights Watch.
Dietterich, T. G. 1998. “Approximate statistical tests for comparing supervised
classification learning algorithms.” Neural Computation 10, no. 7: 1895–1923.
32
Durante, R., P. Pinotti, and A. Tesei. 2017. The Political Legacy of Entertainment TV. CEPR
Discussion Papers 1475.
Enikolopov, R., A. Makarin and M. Petrova. 2017. “Social Media and Protest Participation:
Evidence from Russia.” Available at SSRN: https://ssrn.com/abstract=2696236 or
http://dx.doi.org/10.2139/ssrn.2696236
Enikolopov, R., M. Petrova, and E. Zhuravskaya. 2011. “Media and Political Persuasion:
Evidence from Russia. The American Economic Review 101, no. 7: 3253-3285.
Evans, G. 2008. “Russia and the 'Responsibility to Protect.'” Los Angeles Times. August 31.
Available at http://www.latimes.com/la-oe-evans31-2008aug31-story.html.
Fearon, J. D., and D. D. Laitin. 2000. “Violence and the Social Construction of Ethnic Identity.”
International Organization 54, no. 4: 845-877.
FAO/WFP Crop and Food Supply Assessment Mission to Zambia. 2002. Special Report. Rome:
Food and Agriculture Organization of the United Nations. Available at
https://reliefweb.int/sites/reliefweb.int/files/resources/0E2D56211E6A3945C1256BDC00
3546F9-fao-zam-18jun.pdf (accessed July 27, 2018).
Fujii, L.A., 2004. “Transforming the Moral Landscape: The Diffusion of a Genocidal Norm in
Rwanda.” Journal of Genocide Research, 6, no. 1: 99-114.
---. 2009. Killing Neighbors: Webs of Violence in Rwanda. Ithaca, NY: Cornell University Press.
Gagliarducci, S., M. Gaetano Onorato, F. Sobbrio, and G. Tabellini. 2017. War of the Waves:
Radio and Resistance During World War II. IZA Institute of Labor Economics
Discussion Paper No. 11244. Available at SSRN: https://ssrn.com/abstract=3092226.
33
Gathigi, G. W. 2009. Radio Listening Habits Among Rural Audiences: An Ethnographic Study
of Kieni West Division in Central Kenya. Diss. Ohio University.
https://etd.ohiolink.edu/!etd.send_file?accession=ohiou1249668973&disposition=inline
Gil de Zúñiga, H. and Valenzuela, S., 2011. “The Mediating Path to a Stronger Citizenship:
Online and Offline Networks, Weak Ties, and Civic Engagement. Communication
Research 38, no. 3: 397-421.
Gohdes, A. R. 2018. “Studying the Internet and Violent Conflict.” Conflict Management and
Peace Science 35(1): 89-106.
Gohdes, A. R., and S. C. Carey. 2017. "Canaries in a Coal-mine? What the Killings of
Journalists Tell Us About Future Repression." Journal of Peace Research 54(2): 157174.
Government of Rwanda. 2001. Organic Law N°40/2000 of 26/01/2001 Governing the Creation
of Gacaca Courts.
Government of Rwanda. 2004. Organic Law N°16/2004 Establishing the Organisation,
Competence and Functioning of Gacaca Courts Charged with Prosecuting and Trying
the Perpetrators of the Crime of Genocide and Other Crimes Against Humanity,
Committed between October 1, 1990 and December 31, 1994.
Hegre, H., N. W. Metternich, and H. M. Nygård. 2017. “Introduction: Forecasting in Peace
Research.” Journal of Peace Research 54 (2): 113-124.
Horowitz, D. L. 2001. The Deadly Ethnic Riot. Berkeley and Los Angeles: University of
California Press.
Ilboudo, J. 2000. “Prospects for Rural Radio in Africa.” African Broadcast Cultures: Radio in
Transition, edited by Richard Fardon and Graham Furniss, 42-71. Oxford: J. Currey.
34
IPUMS International. 2012. Integrated Public Use Microdata Series, International: Version 6.1
[Machine-readable database]. Minneapolis: University of Minnesota.
Kalyvas, S. N. 2008. “Promises and Pitfalls of an Emerging Research Program: The
Microdynamics of Civil War.” Order, Conflict, and Violence, edited by S. Kalyvas, I.
Shapiro, & T. Masoud, 397-421. Cambridge: Cambridge University Press.
Kalyvas, S. N., and M. A. Kocher. 2007. “Ethnic Cleavages and Irregular War: Iraq and
Vietnam.” Politics & Society 35, no. 2: 183-223.
Kasampalis, S., P. I. Lazaridis, Z. D. Zaharis, A. Bizopoulos, S. Zetlas, and J. Cosmas. 2013.
"Comparison of Longley-Rice, ITM and ITWOM Propagation Models for DTV and FM
Broadcasting." In Wireless Personal Multimedia Communications (WPMC), 2013 16th
International Symposium, 1-6. IEEE.
Kern, H. L., and J. Hainmueller. 2009. “Opium for the Masses: How Foreign Media can Stabilize
Authoritarian Regimes.” Political Analysis 17, no. 4: 377-399.
Kibaara, B., J. Ariga, J. Olwande, & T.S. Jayne. 2008. Trends in Kenyan Agricultural
Productivity: 1997–2007. Nairobi: Tegemeo Institute of Agricultural Policy and
Development.
Kimani, Mary. 2007. “RTLM: The Medium that Became a Tool for Mass Murder.” The Media
and the Rwandan Genocide, edited by Allan Thompson 110-124. London and Ann
Arbor: Pluto Press.
Kirschke, L. 1996. Broadcasting Genocide: Censorship, Propaganda & State-sponsored
Violence in Rwanda 1990-1994. London: Article 19.
35
Kogen, L. 2013. “Testing a Media Intervention in Kenya: Vioja Mahakamani, Dangerous
Speech, and the Benesch Guidelines.” The Center for Global Communication Studies.
http://www.global.asc.upenn.edu/app/uploads/2016/04/Vioja-Final-Report-Dec-2013.pdf.
Larson, Jennifer M., and Janet I. Lewis. 2018. "Rumors, Kinship Networks, and Rebel Group
Formation." International Organization: 1-33.
Leguy, C. 2007. “Revitalizing the oral tradition: Stories broadcast by Radio Parana (San, Mali).”
Research in African Literatures 33, no. 3: 136-147.
Li, D. 2004. "Echoes of Violence: Considerations on Radio and Genocide in Rwanda." Journal
of Genocide Research 6, no 1: 9-27.
Little, A. T. 2016. "Communication Technology and Protest." The Journal of Politics 78, no.1:
152-166.
---. 2017. "Propaganda and Credulity." Games and Economic Behavior 102: 224-232.
Luft, A. 2015. “Genocide as Contentious Politics. Sociology Compass, 9, no 10: 897-909.
Mandani, M. 2001. When Victims become Killers. Princeton, NJ: Princeton University Press.
McDoom, O. S. 2013. “Who Killed in Rwanda’s Genocide? Micro-space, Social Influence and
Individual Participation in Intergroup Violence." Journal of Peace Research 50, no. 4:
453-467.
---. 2014. “Antisocial Capital: A Profile of Rwandan Genocide Perpetrators’ Social
Networks.” Journal of Conflict Resolution, 58(5), 865-893.
Melvern, L. A People Betrayed: The Role of the West in Rwanda’s Genocide. London: Zed
Books.
Mercier, H. 2017. "How Gullible are We? A Review of the Evidence from Psychology and
Social Science." Review of General Psychology 21, no. 2: 103-122.
36
Mironko, C. 2007. “The Effect of RTLM’s Rhetoric of Ethnic Hatred in Rural
Rwanda,” The Media and the Rwandan Genocide, edited by Allan Thompson 125-135.
London and Ann Arbor: Pluto Press.
Mohamed, S. 2012. “Taking Stock of the Responsibility to Protect” Stanford Journal of
International Law 48(2): 319-339.
Muchlinski, D., D. Siroky, J. He, and M. Kocher. 2015. “Comparing Random Forest with
Logistic Regression for Predicting Class-imbalanced Civil War Onset Data.” Political
Analysis 24(1): 87-103.
Nanlohy, S., C. Butcher, and B. E. Goldsmith. 2017. “The Policy Value of Quantitative Atrocity
Forecasting Models.” The RUSI Journal 162(2): 1-9.
Newbury, D. 1998. “Understanding Genocide.” African Studies Review 41(1): 73-97.
Ngonda, S. 2017. “Kenya May Shut Off Social Media During August Elections.” Capital FM. 17
July. https://allafrica.com/stories/201707170910.html.
Nguyen, P. 2007. “NonbinROC: Software for Evaluating Diagnostic Accuracies with Nonbinary Gold Standards.” Journal of Statistical Software 21(10): 1-10.
Niazi, T. 2002. “The Ecology of Genocide in Rwanda.” International Journal of Contemporary
Sociology 39(2): 223-48.
Nyseth Brehm, H. 2014. Conditions and Courses of Genocide. PhD Dissertation. University of
Minnesota.
---. 2017. “Subnational Determinants of Killing in Rwanda.” Criminology 55, no. 1: 5-31.
Nyseth Brehm, H., Uggen, C., & Gasanabo, J. D. 2014. Genocide, Justice, and Rwanda’s Gacaca
Courts. Journal of Contemporary Criminal Justice, 30(3), 333-352.
37
Obuchowski N. A. 2005. “Estimating and Comparing Diagnostic Tests’ Accuracy When the
Gold Standard Is Not Binary.” Academic Radiology 12(9): 1198–1204.
---. 2006. “An ROC-Type Measure of Diagnostic Accuracy When the Gold Standard is
Continuous-Scale.” Statistics in Medicine 25(3): 481–493.
Olkean, B. A. 2009. “Do Television and Radio Destroy Social Capital? Evidence from
Indonesian Villages.” American Economic Journal: Applied Economics 1, no. 4: 1-33.
Opt, S. K. 1992. “The Development of Rural Wired Radio Systems in Upstate South Carolina.”
Journal of Radio Studies 1(1-2): 71-82.
“Organic Law Establishing the Organisation, Competence and Functioning of Gacaca Courts
Charged with Prosecuting and Trying the Perpetrators of the Crime of Genocide and
Other Crimes Against Humanity, Committed between October 1st, 1990 and December
31, 1994.” 2004. Official Gazette of the Republic of Rwanda, no. 16/2004 of 19/6/2004.
Available at
https://repositories.lib.utexas.edu/bitstream/handle/2152/4582/3677.pdf?sequence=1.
Petrova M., and D. Yanagizawa-Drott. 2016. “Media Persuasion, Ethnic Hatred, and Mass
Violence: A Brief Overview of Recent Research Advances.” Economic Aspects of
Genocides, Other Mass Atrocities, and their Prevention, edited by Charles H. Anderton
and Jurgen Brauer 274-286. Oxford and New York: Oxford University Press.
Petty, K. A. 2013. “Humanity and National Security: The Law of Mass Atrocity Response
Operations.” Michigan Journal of International Law 34(4): 745-827.
Pierskalla, J. H., and F. M. Hollenbach. 2013. “Technology and Collective Action: The Effect of
Cell Phone Coverage on Political Violence in Africa.” American Political Science Review
107, no. 2: 207-224.
38
Polianskaya, A. 2018. “Villagers Lynch Five Men in India over Child Kidnap Gang Claims
Spread on WhatsApp.” The Independent. 2 July.
Republic of Rwanda; Ministry of Lands, Environment, Forestry, Water and Mining. 2007.
Building Capacity for Sustainable Land Use and Management in Rwanda. UNDP/GEFMSP Project on Land Degradation in Rwanda.
http://www.rw.undp.org/content/dam/rwanda/docs/operations/Projects/Environment%20a
nd%20energy/RW_EE_Sustainable_land_use_and_management.pdf
Reuter, O. J., and D. Szakonyi. 2015. “Online Social Media and Political Awareness in
Authoritarian Regimes.” British Journal of Political Science 45, no. 1: 29–51.
Roberts, J. T. 1995. “Expansion of Television in Eastern Amazonia." Geographical Review
85(1): 41-49.
Schabas, W. A. 2005. “Genocide Trials and Gacaca Courts." Journal of International Criminal
Justice 3, no. 4: 879-895.
Schneider, G., N. P. Gleditsch, and S. Carey. 2011. “Forecasting in International Relations: One
Quest, Three Approaches.” Conflict Management and Peace Science 28(1): 5-14.
Shaikh, S.J., 2016. “Elite Communication and Legitimization of Violence during Inter-Group
Conflicts.” Impact of Communication and the Media on Ethnic Conflict, edited by Steven
Gibson and Agnes Lucy Lando, 107-125. Hershey, PA: IGI Global.
Shapiro, J., and N. Weidmann. 2015. “Is the Phone Mightier than the Sword? Cell Phones and
Insurgent Violence in Iraq.” International Organization 69, no. 2: 247–74.
Shmueli, G. 2010. “To Explain or to Predict?.” Statistical Science. 25(3): 289-310.
Smeulers, A., and L. Hoex. 2010. “Studying the Microdynamics of the Rwandan Genocide." The
British Journal of Criminology 50(3): 435-454.
39
Smith, R. 2003. “The Impact of Hate Media in Rwanda.” BBC News Online. 3 December.
http://news.bbc.co.uk/2/hi/africa/3257748.stm.
Sperling, L. 1997. “The Effects of the Rwandan War on Crop Production and Varietal Diversity:
A Comparison of Two Crops.” ODI Agricultural Research and Extension Network Paper
no. 75. London, UK: Overseas Development Institute.
Spitulinik, D. 2002. “Mobile Machines and Fluid Audiences: Rethinking Reception through
Zambian Radio Culture.” Media Worlds: Anthropology on New Terrain, edited by Faye
Ginsburg, Lila Abu-Lughod, and Brian Larkin, 337-354. Berkeley and Los Angeles:
University of California Press.
Straus, S., 2007. “What is the Relationship between Hate Radio and Violence? Rethinking
Rwanda's ‘Radio Machete’”. Politics & Society 35, no. 4: 609-637.
---. 2012. “‘Destroy Them to Save Us’: Theories of Genocide and the Logics of Political
Violence.” Terrorism and Political Violence 24, no. 4, 544-560.
---. 2015. Making and Unmaking Nations: War, Leadership, and Genocide in Modern Africa.
Ithaca, NY: Cornell University Press.
Streiner, D. L., and J. Cairney. 2007. "What's Under the ROC? An Introduction to Receiver
Operating Characteristics Curves." The Canadian Journal of Psychiatry 52, no. 2: 121128.
The Prosecutor v. Ferdinand Nahimana, et al. 2003. Trial Judgement), ICTR‐99‐51‐T,
International Criminal Tribunal for Rwanda (ICTR), 3 December, available at:
http://unictr.unmict.org/sites/unictr.org/files/case-documents/ictr-99-52/trialjudgements/en/031203.pdf.
40
Tuckwood, C., 2014. “The State of the Field: Technology for Atrocity Response.” Genocide
Studies and Prevention: An International Journal, 8, no. 3: 81-86.
Verwimp, P. 2003. “The Political Economy of Coffee, Dictatorship, and Genocide.” European
Journal of Political Economy 19, no. 2: 161-181.
--- . 2005. “An Economic Profile of Peasant Perpetrators of Genocide: Micro-level Evidence
from Rwanda.” Journal of Development Economics 77(2): 297-323.
Ward, M. D., Greenhill, B. D., & Bakke, K. M. (2010). The perils of policy by p-value:
Predicting civil conflicts. Journal of Peace Research, 47(4), 363-375.
Wallenstein, J., 2001. “Punishing Words: An Analysis of the Necessity of the Element of
Causation in Prosecutions for Incitement to Genocide.” Stanford Law Review 54: 351398.
Warren, T. C. 2015. “Explosive connections? Mass Media, Social Media, and the Geography of
Collective Violence in African States.” Journal of Peace Research 52, no. 3: 297–311.
Weeraratne, S. 2010. Ethnic Entrepreneurs and Collective Violence: Assessing Spatial
Variations in Anti-Chinese Rioting Within Jakarta During the May 1998 Riots (World
Institute for Development Economics Research, Working Paper No. 2010.55).
Weidmann, N. B. 2016. “A Closer Look at Reporting Bias in Conflict Event Data.” American
Journal of Political Science 60, no. 1: 206-18.
Weiner, M. M. 2005. Monopole Antennas. New York: Marcel Dekker, Inc.
Wilson, C. 2011. “Provocation or Excuse?: Process-Tracing the Impact of Elite Propaganda in a
Violent Conflict in Indonesia.” Nationalism and Ethnic Politics 17(4): 339-360.
Wilson, R. A. 2015. “Inciting Genocide with Words.” Michigan Journal of International Law
36(2): 277-320.
41
Wischnath, G., and H. Buhaug. 2014. "On Climate Variability and Civil War in Asia." Climatic
Change 122(4): 709-721.
Yanagizawa-Drott, D. 2014. "Propaganda and Conflict: Evidence from the Rwandan Genocide."
The Quarterly Journal of Economics 129, no. 4: 1947-1994.
42
