Media as Agents of Change in Post-genocide Societies

In 2004, a nongovernmental organization in Rwanda began broadcasting a radio soap opera (Musekeweya, or New Dawn) to teach listeners about the roots of violence, the importance of independent thought, and the dangers of excessive deference to authority. This media intervention, which enfolds educational messages in a dramatic fictional story, is part of a global phenomenon called “entertainment education,” which has become a popular technique among social change campaigns in developed and developing countries (Rosin Reference Rosin2006; Singhal et al. Reference Singhal, Cody, Rogers and Sabido2004). Such campaigns are especially popular in post-conflict countries where government agencies and civil society are often unreliable vehicles for promoting social or political change. Entertainment-education programs weave messages about health, conflict resolution, or development strategies into a story featuring realistic and entertaining characters. The theory, which draws on psychological social learning theory (Bandura Reference Bandura, Schmuch and Schultz2002), is that media can “both entertain and educate, in order to increase audience members’ knowledge about an educational issue, create favorable attitudes, shift social norms, and change overt behavior” (Singhal and Rogers Reference Singhal, Rogers, Singhal, Cody, Rogers and Sabido2004, 5).

The soap opera, launched in Rwanda one decade after the genocide, featured a fictional story of two Rwandan communities. Due to government restrictions against public discussion of ethnicity, the story of the communities serves as a transparent allegory for the history of cooperation and conflict between Rwandan Tutsis and Hutus. Intercommunity tensions created by a land shortage are set aflame by demagogic authorities who seek to accumulate power, and relations between the fictional communities disintegrate as the less prosperous community attacks its rival. Against this backdrop of communal violence, a romantic storyline unfolds between a young man and woman, each from a different community. Instead of falling victim to the violence and prejudices between their two communities, the Rwandan Romeo and Juliet form a coalition for peace with citizens from both communities. Their coalition defies the power-hungry authorities and seeks to mediate the conflict and help the victims.

We ask the following questions regarding the impact of this media program. First, can a drama written to reflect common experiences of the average Rwandan citizen encourage the adoption of norms and behaviors portrayed by the program's heroes? More broadly, can media interventions influence an aspect of a country's political culture? Questions about media influence on political behavior and culture abound, but rarely are investigators able to observe a media program unfold and deploy a methodological design that isolates its causal effects. We conducted a randomized experiment to assess whether one year after the radio program began changes could be observed in Rwandan radio listeners’ deference to authority, willingness to dissent, and collaborative participation in dispute resolution. By deference to authority, we mean submissive behavior toward another on the part of a subordinate actor or group (Shils Reference Shils1972), and by dissent, we mean the expression of opinions that are at variance with those previously, commonly, or officially expressed by peers or authorities (Oxford English Dictionary). In Kinyarwanda, the language in which the interviews were conducted, dissent was signified as ndabivuga, “I speak [when I disagree],” in opposition to ndicecekera, “I keep quiet.”

The project is part of a broader assessment of the impact of media on citizen beliefs, norms, and behavior. The radio program's social psychological effects have been described elsewhere. Paluck (Reference Paluck2009) shows that the radio program shifts listeners’ perceptions of social norms, or perceptions of typical and desirable behaviors, without affecting factual beliefs regarding the causes or consequences of those behaviors. In this article, we ask whether the media can influence Rwandans’ perceptions of the desirability of certain political behaviors, and whether these perceptions translate into political conduct.

Sampling of Participants

We selected 14 research sites to represent salient political, economic, and ethnic categories of present-day Rwanda: two genocide survivor communities (mostly Tutsi), two Twa communities (the Pygmy minority), two prisons, and eight general population communities from the four general provincial regions of Rwanda. The four provincial regions (excluding the capital) represent differing social and political circumstances. The northwest region was the base of Hutu extremism and support for former President Habyarimana, whose assassination marked the starting point of the genocide. The current (Tutsi-led) regime distrusts and monitors this region. The western region was the site of some of the most rapid genocidal killing in 1994 (in one prefecture, more than three fourths of its 59,050 victims were killed in less than two weeks; Straus Reference Straus2006, 56–8). The southern region is the historical seat of Tutsi royalty, and in modern times, has been the intellectual heart of the country, hosting the national university. The largest percentage of Rwandan Tutsis lived in the south prior to the genocide, and the region had one of the latest onsets of genocidal violence due to initial resistance by local authorities. The northeast region was in the Tutsi rebel (RPF) army's path as it invaded from Uganda in 1993 and 1994, eventually ending the genocide, but carrying out reprisal violence aimed at Hutu along their route to the central capital, Kigali.

Because prison, survivor, and Twa communities are relatively scarce and scattered across the country, we worked through the prison system and Twa and survivor advocacy groups to obtain demographic, socioeconomic, and detainee characteristics (for the prisons) for a range of sites, and we chose the two most similar sites of each community type, even if they were located in different provinces. We chose each general population community site on the basis of its accessibility by paved or dirt road (we eliminated sites unavailable by road for logistical reasons), and on the quality of its demographic and socioeconomic match with another site in the region, a site far enough away that the communities did not share markets (to prevent spillover) but close enough to be similar on a range of characteristics. We used Rwandan census data to match the sites, including gender ratio, quality of dwellings, religion, and education level. Figure 1 shows the exact locations of each research site within the four provincial boundaries.Footnote ⁵

FIGURE 1.

Research Sites by Location and Experimental Assignment

We randomly assigned one site in each pair to listen to the reconciliation program, and the other to the health program. Random assignment ensures that, in expectation, the treatment groups have the same observable and unobservable attributes. Block randomization of sites helps minimize observable differences between the communities in small samples. Even after blocking, there could be remaining observable differences, but these were controlled in multivariate analyses of the outcome data. Because the reconciliation and health sites have similar background attributes, these controls turned out to have little effect on the results.

Because radios and batteries are relatively expensive for Rwandans, they usually listen to the radio in groups. Thus, we used a group-randomized design in which adults from a community listened together either to the treatment (reconciliation) program or to the control program (another entertainment-education radio soap opera about health and HIV). Forty adults within each community were randomly sampled from official lists of the population after stratifying for sex and age (18–30 and 30 and older). The sampling technique invited at most one person per family.

The first author visited each community with four Rwandan research assistants who represented Hutu and Tutsi ethnic backgrounds. Research assistants sought out the selected individuals and explained the purpose and procedure of the study. The purpose—“to understand Rwandans’ opinions about radio programs and the issues addressed in those programs”—was defined broadly so as not to create expectations of any particular set of future questions. Nearly all individuals selected to participate accepted the invitation. On average, one or two selected individuals at each site turned down the invitation to participate due to poor health or indifference. We drew replacements from a randomized backup list created in advance.

Pretest

When an individual agreed to participate, the researcher conducted a brief pretest interview to collect information about the individual's background. Baseline pretest data for each community cluster confirmed that the matched randomization had divided the sample into two groups with similar background attributes. The pretest data also provided covariates that help minimize disturbance variance in the posttest analyses.

Table 1 presents statistics on the characteristics of participants at each research site. The total participant sample was 50% male, and the average age was 38.9 years. We did not achieve a male–female balance in every site because of the disproportionate loss of men during the genocide in some areas (in survivor, Twa, and in two general population villages). Ninety-three percent of men and 81% of women reported regular listening to the radio. Only 53% of the participants actually owned a radio; 83% of those who did not have a radio reported listening in groups with family or friends, which affirms the logic of our group listening study design. Seventy-nine percent of participants identified themselves as farmers, and 73% of men and 67% of women in the total sample had some schooling. The Twa and survivor communities were the poorest of all research sites.

TABLE 1.

Sample Characteristics by Research Site

R, reconciliation, H, health.

Note: Figures are percentages for each research site.

Ninety-nine percent of the sample was in Rwanda at the start of the genocide, and approximately 50% was displaced by the violence for a time of one week to a few years. Sixty-nine percent of the sample reported one or more of their relatives were killed in 1994; similar proportions of each community claimed material losses (property looted or burned) during the genocide. Twenty-eight percent of the general population participants had a relative in prison compared to 7% of survivors and 57% of Twa participants. As a randomization check, we regressed the binary treatment variable on background characteristics (including dummies for missing data in the pretest and accounting for clustered random assignment). An F test was insignificant (F = .73, p = .65), meaning that there was no unexpected association between characteristics of the participants and radio program to which they were assigned.

Experimental Design

The same Rwandan research assistants visited each community over the course of one year and played each month's four 20-minute episodes on a portable stereo for the listener groups. Although research assistants were aware of the program differences, they were blind to specific research hypotheses. Participants gathered in their respective community spaces as they do for nonresearch occasions to listen to the radio. Program content was the only difference between the two conditions—listening protocol, frequency, and outcome measurements were the same.

Control groups listened to the education-entertainment radio soap opera Urunana, which aims to change beliefs, norms, and behaviors about reproductive health and HIV. Comparing the outcome of groups listening to a radio soap opera with a different message isolates the impact of the content of the radio program from the social environment that accompanies media consumption. An alternative design would involve a no-listening control group. But the no-listening subjects would then differ along many dimensions (e.g., participation in groups, common knowledge about other listeners’ reactions, a monthly schedule) that potentially threaten the experiment's internal validity.

Monthly field visits ensured that participants listened to the program but preserved the most natural environment possible. Research assistants who visited every month sat and listened with the group as part of what was designed to be a casual community gathering. The group shared customary local drinks (purchased by the research team) and often lingered afterward to chat. Research assistants never initiated or guided discussion. They filled out standardized observation sheets after they left each site to record attendance and to rate the group's levels of observable enthusiasm, attentiveness, confusion, emotional expressions, and the amount and subjects of discussion during and after the program.

A content analysis of the radio program's scripts confirms the essential similarity of the two soap operas.Footnote ⁶ As Table 2 makes plain, the principal topics of each program are distinct. Health issues (e.g., AIDS testing, health insurance, pregnancy precautions) appear in 83% of Urunana episodes compared to 2% of Musekeweya episodes. Expressions of dissent, coded when opposition to authorities or peers was voiced (e.g., disagreeing with plans to attack the neighbors; denouncing looters; proclaiming friendship with “the enemy”; excluding minor interpersonal disagreement, such as the best way to make banana beer), never appear in Urunana episodes but arise in 56% of the episodes in the Musekeweya reconciliation program. That the programs’ messages do not overlap boosts confidence that any observed change pertaining to the reconciliation program content can be attributed to listening to the reconciliation program. In other respects, the programs are similar in format and present the same or homologous topics, providing study participants with relatively equivalent listening experiences. Like all soap operas, the radio dramas were filled with romance, humor, villains, deceit, and emotion. In the reconciliation program, the lovers were the Romeo and Juliet from opposing communities (27% of episodes); the health program also featured two young lovers, and a comedic and affectionate married couple expecting their first baby (25% of episodes). The reconciliation program villains were local authorities and elders who fomented hatred and violence between the two communities (29% of episodes); health program villains included elders who covered up the rape of a young girl, an ignorant work supervisor who stigmatized people with HIV, and “quack” doctors who charged high prices for unhygienic and retrograde medicine (17% of episodes). When violence was depicted, it was primarily interpersonal (rape) in the health program (10%), and mostly intergroup (attacks and looting) in the reconciliation program (8%). Equivalent amounts of conflict spurred positive but different resolutions in each program: interpersonal apology and economic cooperation in the health program (23%), and empathy and cooperation for peace in the reconciliation program (21%). As indicated by Table 2, both programs portrayed rural Rwanda communities in which some people lived comfortably but others struggled in poverty.

TABLE 2.

Topics of Reconciliation and Health Radio Programs as Percentage of Total Episodes Played to Participants

Note: Percentages are calculated by dividing the number of episodes in which each theme appears by the 48 total episodes. Each episode contained more than one theme; thus, percentages do not total 100.

Overall, the reconciliation program focused on social and political conflict and citizens’ responsibility for conflict reduction, whereas the health program focused on social and familial conflict and citizens’ responsibility for their health. Both programs were penned by Rwandan scriptwriters relying on traditional Rwandan storytelling styles (one reconciliation scriptwriter originally worked for the health program) and were financed by nongovernmental organizations.

Random assignment of listeners to one of the radio programs addresses the chronic problem of self-selection—listeners’ preexisting tastes may guide their choice of media programs and their attitudes and behaviors. To ensure the health group remained “untreated” by the reconciliation program, which was broadcast nationally during the evaluation period, we asked the health program listeners to refrain from listening to the reconciliation program for one year. As an incentive to comply, we promised 14 cassette tapes containing the year's worth of reconciliation program episodes at the end of the year. Thus, health participants understood their promise as a postponement and not a sacrifice. To maintain symmetry across experimental groups, these cassettes were also promised to the reconciliation group. (See Appendix I for a detailed discussion of the experimental procedure and the ethical considerations that guided the design.)

Health participants do not appear to have listened to broadcasts of the reconciliation program. We instructed research assistants to make casual comments to health participants about the reconciliation program halfway through the year, but participants indicated that they were not listening and seemed unaware of its main characters. The high rate of compliance may reflect the fact that the new reconciliation drama had yet to build up a following, and three alternative programs were aired on other stations in its time slot. If, however, some participants “crossed over” to the treatment group without our knowledge, our statistical findings would underestimate the true effect of the reconciliation program (Freedman Reference Freedman2006). Our results therefore provide a conservative estimate of the reconciliation program's impact.

Data and Statistical Considerations

At the end of one year, a team of seventeen Rwandan researchers (including those who had visited the communities throughout the year) accompanied the first author to each research site for three days. Researchers conducted individual structured interviews, semistructured focus groups, “role plays,” and unobtrusive observations of community negotiations. Before analyzing the results, we first comment on some statistical issues.

Analytic Issues of Clustered Random Assignment. With fourteen clusters and approximately 40 individuals per community, the total sample is 556 individuals (fewer than 10 people left the study over the course of the year; there was no differential attrition for reconciliation and health). However, the “effective N” of this study is lower because individuals within each cluster are not independent observations. Participants living together in each cluster may share unobservable characteristics. Their shared characteristics violate the assumption of independent disturbances, an assumption on which conventional standard errors are calculated. We therefore present robust cluster standard errors, which account for within-cluster covariance (Arceneaux Reference Arceneaux2005). We also increase the power of the experiment by including covariates from the pretest, which potentially reduce the intracluster correlation among the disturbances and decrease the disturbance variance.

Evidence of Equivalent Listening Experiences. A mean of 35.6 participants attended each month's listening sessions at each site, with no meaningful difference in attendance between health (m = 35.8, sd = 3.5) and reconciliation (m = 35.4, sd = 3.7) treatment groups. There were no differences between reconciliation and health groups’ observed interest in the program, using the research assistants’ ratings combined across all visits (using a scale from least to most, 1–5; m = 4.0, sd = .75; m = 4.2, sd = .83, respectively), enthusiasm (m = 3.3, sd = .96; m = 3.6, sd = .90), distraction (m = 1.9, sd = .90; m = 1.7, sd = .78), confusion (m = 1.5, sd = .60; m = 1.5, sd = .78), and discussion during the broadcast (m = 2.9, sd = .90, m = 3.3, sd = 1.1). After the broadcast ended, researchers observed any further spontaneous discussion and reported that participants spent, on average, 63% of their time discussing the program (as opposed to other topics) before leaving, a rate that was almost identical in the two experimental conditions. These results indicate that the listening experience was similar for reconciliation and health, with the key difference being the content of the media program.

Individual Interviews and Focus Groups. In the structured individual interviews, researchers read each participant a series of statements. Participants specified how much they agreed or disagreed with each statement by pointing to one of four progressively larger circles printed on a large index card; the smallest circle represented “disagree strongly” and the largest “agree strongly.” We focus here on the item used to measure perceptions of a prescriptive (“that is the way people should act”) norm about dissent: When I disagree with something that someone says, I should dissent. Because the response options are a series of ordered categories, we used ordered probit regression to analyze the results.

Table 3 shows the effect of the reconciliation program on responses to this item across three specifications. The first specification includes only the random treatment (“reconciliation program”) as a predictor. The second includes dummy variables for each community pair—from the northwest sites to the survivor community sites. The final specification includes the community dummies as well as measures from the pretest survey. These pretest measures included whether the respondent was ever displaced by the violence in Rwanda, respondent sex, and amount of radio listening.

TABLE 3.

Ordered Probit Estimates of Reconciliation Program's Effects on Willingness to Express Dissent

Notes: *p <. 05; **p <. 001, two tailed. The response scale for dissent is 1 (I should stay quiet) to 4 (I should dissent); pairs of communities are entered 1-0 dummies for each general population (NW–S), and Twa, prison, and survivor pair. The omitted pair of communities is the western pair. Sex is a dummy variable with males equal to one. Age is measured in years. Displacement and radio listening are coded from 1 (never displaced, do not listen often) to 3 (displaced a long time, listen often).

The results show that reconciliation listeners were .26 to .29 probits more likely than health listeners to indicate that they should speak up. These estimates are statistically significant (p < .001) regardless of model specification, including the most conservative specification, a linear model using the research site as the unit of analysis (ß = 1.70, se = .63, p < .05, N = 14). They are also substantively large. A shift of .26 probits implies, for example, that a health group respondent with a 30% chance of strongly agreeing to dissent would move to a 40% chance if assigned to the reconciliation program group. This is a large, but not implausibly large, shift in opinion.

Focus Group Results

Does a perception that one should express an unpopular viewpoint translate into actual behavior? To test this, we raised the question of trust in both the private interview and the public focus group, to determine if participants would express similar views in the two contexts. From knowledge of the research site and previous research (e.g., Gabisirege and Babalola Reference Gabisirege and Babalola2001), we expected trust to be low in the wake of the genocide. However, we also knew that the expression of mistrust would go against the official government line of unity and reconciliation.

In privately conducted individual interviews, participants were asked to rate their agreement with the statement: There is mistrust in my community. The vast majority (80%) of participants agreed with this statement. There was no difference in privately reported mistrust comparing reconciliation and health listeners. This finding is consistent with other evidence demonstrating that social desirability did not guide the responses of reconciliation listeners. Paluck (Reference Paluck2009) reports that reconciliation listeners did not become more sanguine about the effects of intermarriage and other types of cross-group social interaction.

Half of all participants were organized into single-sex focus groups of 10 before their private interview, and the other half took part in a focus group after the interview.Footnote ⁷ Researchers posed open-ended questions and encouraged participants to discuss them with each other. With little moderation by researchers, the focus groups were more like community forums than-question-and-answer sessions; on average, there were 11 contributions for each question, and there was no significant difference in the average amount of talking between reconciliation and health focus groups (ß = −4.96, se = 3.59, p = .22).

Among other topics, researchers asked focus groups to discuss whether there was mistrust in their community. Figure 2 shows that in these public settings denials of mistrust became much more frequent—but only among health groups.Footnote ⁸ With the exception of the survivor communities (described later), more than 60% of the health group comments were denials of mistrust, compared to less than 20% of reconciliation group comments (ß = .96, se = .37, p < .01). In light of the uniformly high levels of mistrust reported in the individual interviews, the difference in focus group responses seems to reveal more about the reconciliation group's willingness to speak out on difficult subjects than about actual levels of community mistrust.

FIGURE 2.

Denials of Mistrust Are More Frequent When Made in Public among Health Listeners

Focus groups were only conducted in prisons, survivor communities, and south and northwest communities (combined here into “general population” communities).

The one exception to this finding was the survivor groups—none of whom modified their strong stance that there is mistrust in their community. A plausible explanation for this break in the pattern, which we discuss further in this section, is that their identity as genocide survivors and as coethnics with the current regime grants them greater license to speak out about the social effects of the genocide. Although the current governmental regime pressures Rwandans to adhere to their official line of “unity and reconciliation,” genocide survivors (backed by strong political interest groups within Rwanda and featured in government appeals for aid) are to a certain extent granted exception.

Social Distance Attitudes

It should be emphasized that the behavioral changes just described were not accompanied by a more general willingness to affiliate with members of other groups. Our questionnaire asked whether participants were interested in greeting, sharing beer with, working with, and intermarrying a child with a member of a different group (the implication was a different ethnic group, although the prohibition on using ethnic terms prevented us from specifying this directly).Footnote ⁹ We found no reduction in social distance when comparing the health and reconciliation groups. The null results reported in Table 4 are important insofar as they run counter to the hypothesis that demand characteristics of the experiment caused respondents in the reconciliation group to express more tolerant opinions.Footnote ¹⁰

TABLE 4.

Linear Regression Estimates of Reconciliation Program's Effects on Social Distance

Notes: *p < .05; **p < .001, two tailed. Social distance is an index of willingness to greet, share a beer, work together, or marry child to another group. The response scale for each social distance item is 1 (not willing) to 4 (very willing). Pairs of communities are entered as 1-0 dummies for each general population (NW–S), and Twa, prison, and survivor pair. The omitted pair of communities is the western pair. Sex is a dummy variable with males equal to one. Age is measured in years; displacement and radio listening are coded from 1 (never displaced, do not listen often) to 3 (displaced a long time, listen often).

Role Playing

We were also interested in measuring group-generated outcomes because dissent and deference to authority is expressed socially and conditioned by one's social surroundings. In particular, we want to test participants’ approach to a collective dispute. To single-sex groups of participants at each site, we played recordings of unfinished radio scenes written by the nongovernmental organization's (NGO's), scriptwriters and the first author. Each unfinished scene centered on a common problem in Rwandan society that had also been addressed in the reconciliation radio program. In the scene relevant to the present analysis, the characters argue about how to handle a group of refugees fleeing a famine.

This refugee scene (reprinted in the Appendix III) presented three main characters. One alarmed and angry man comes running to tell two neighbors that he has learned about the fleeing neighboring community. He exclaims that their own community is too poor to help the refugees and asks the other two to help him block the border and, if necessary, use force to repulse the refugees. The two neighbors hearing the news are divided: one agrees, but the other argues that they should welcome the refugees. The three men debate, raise a few different points about why they should or should not help the refugees (i.e., violence is wrong, these neighbors may help us in the future, we have no money, I suspect them of stealing money from me in the past), when the scene suddenly ends. The narrator who introduces the scene announces that she would like the participants to act out the end of the scene: “What do you think should happen now?”

This exercise was performative by definition—our aim was not to approximate what group members would actually do in such a situation, but rather to gauge their prescriptive ideas about what they should do. Educators, therapists, and activists use participatory theater in other settings (e.g., Boal Reference Boal and Jackson1992; Breed Reference Breed, Anheier and Isar2007), but rarely have social scientists measured group dynamics or social prescriptions with role playing, making this an exploratory measure. After listening to a scene about a community meeting that familiarized them with the exercise, both male and female participants seemed to enjoy and even “lose themselves” in the role plays. One group of women staged an argument in their practice scene with such conviction that the group attracted the attention of a military guard who was patrolling the general area (he was persuaded to leave by the first author).

Seventeen groups were formed across eight research sites. Each group, on average, performed approximately two unique new endings (allowing different members of the group to assume the primary roles) for a total of twenty-one scenes. Typically, the entire group participated in each scene; those who did not assume one of the three primary roles acted as bystanders or invented new characters who joined the action (e.g., a local authority or a family relation of a main character). Two researchers hand-recorded their speech and action in script format.Footnote ¹¹

Content Analysis of Role-Play Scripts

The first author coded each speaking turn in the “scripts” that resulted.Footnote ¹² We were interested in how each group solved the problem (i.e., by welcoming them or mobilizing in various ways to provide for them). We noted patterns in the vocabulary participants used to argue for and against various ideas.

Problem Resolution Codes. We coded whether the script ended by helping the refugees; if the refugees were offered help, we coded how participants mobilized to help them. It turned out that the means of helping, and not whether they did help, was the distinguishing factor among the scenes. Specifically, we found, as illustrated in the following two excerpts, that actors either helped the refugees themselves or called on the Rwandan government or on NGOs or international organizations (IOs) for help.

Scene resolution 1: Helping refugees with own resources

[Action: The refugees, dragging their feet and looking hungry and tired, approach the two men at the border who want to fight them (Mukama and Cacana).]

[Action: Mukama and Cacana walk away from the refugees to talk, where others from their community stand watching.]

Scene resolution 2: Helping refugees with state or NGO resources

[Action: Important local authority calls a meeting with the ten smaller authorities (Nyumbakumi), each in charge of a block of ten households.]

[Action: Approval of the group, arrival of donors.]

Language. We also content analyzed the vocabulary used in each scene. A remarkable aspect of these transcripts is the way the “public transcript” of the Rwandan government appeared to influence the language used in each research site. Following Scott (Reference Scott1990), we use public transcript to mean government discourse, with which Rwandan peasants are familiarized through speeches broadcast on the radio and in community “sensitization” meetings in which local authorities teach citizens about new policies.

Some of the most important policies publicized and enacted by the Rwandan government in recent years draw on language of forgiveness (e.g., community court processes that emphasize public apologies and acts of forgiveness), reeducation (e.g., ingando or reeducation camps to which newly released prisoners convicted of crimes of genocide visit before returning home), and decentralization of state power. This discourse appeared frequently in the role plays, for example: “I publicly ask for forgiveness for trying to block the refugees”; “These people have bad ideology and need to change!”; and “As the decentralization system stipulates, it's up to you to take a final decision.” We coded each of these speaking turns to note their themes and prevalence.

Role-Play Results

We coded 359 “lines” from the scripts, which include turns of speech and notes about action. Each scene featured, on average, seventeen unique speech or action lines, and each line could receive more than one code.Footnote ¹³ We calculated the relative frequency of each code in terms of the total number of codes using two methods: weighting all sites equally, and weighting according to scene length. The results do not change appreciably in either method (see Appendix IV for calculations), and we present results weighting sites equally in Figure 3.

FIGURE 3.

Content Analysis of Reconciliation and Health Listeners’ Resolutions of Role-Play Dilemma

Percentages were calculated by dividing the number of scenes that were resolved in a given way by the total number of scenes enacted, across all communities in the Health and Reconciliation conditions.

Problem Resolution Results. In every role play, participants decided to welcome the refugees and to shame and sometimes punish those who wanted to keep them out. The striking difference between reconciliation and health scenes lies in how participants arranged to care for the refugees. The two dominant motifs were to handle the problem within the community, by collectively organizing shelter and gathering resources from each family, or to request assistance from the government or NGOs/IOs.

Figure 3 shows that reconciliation groups never asked for help from the government or from an NGO/IO—each time they decided to welcome the refugees, they agreed that they would share with them “the little they had.” By comparison, 16 of the health group's 21 “resolution” comments about how they would provide for the refugees involved asking the government or an NGO to help out (p < .001). Requesting assistance from the state or aid organizations was not a dominant theme of the health program (see Table A1 in Appendix V for precise ratios of speaking turns focusing on community vs. government/NGO/IO help). Some reconciliation scenes did not reach one particular conclusion about how to provide for the refugees (e.g., in three scenes played at the reconciliation prison group, one role-play group self-organized to help the refugees, and the two others were resolved by punishing those who wanted to fight off the refugees). In the health condition, some scenes chose both community and government/NGO/IO help (e.g., deciding to feed the refugees until the Red Cross came).

Despite this striking difference between local responsibility and external aid, it is important to point out that both reconciliation and health participants invented authority characters for each scene. Twenty-six percent and 15% of all “process” comments in scenes in health and reconciliation groups, respectively, concerned the local government. Characters in scenes across all research sites called on important local authorities or on the police for help and advice.

Results for Language Use. The percentage of comments that used government language shows that, on average, 8% of all comments made in each scene drew from the “public transcript.” There were no differences between the reconciliation and health listeners in the percentage of comments using government language. More important, the use of government language was not reliably associated with any particular decision regarding the refugee dilemma. Participants used government language to justify whatever actions they took. (“This is the time of decentralization” was used to argue for providing for the refugees themselves and also for applying to international organizations such as the Red Cross for help.) This finding reveals participants’ opportunistic use of elite parlance, rather than a passive acceptance of media-borne messages “coming from the top.”

Behavior: Unobtrusive Measurement of Community Negotiations

Thus far, we have measured behavior within artificial settings—in focus group discussions and in role-playing activities. We also sought to trigger a more natural dilemma for our participants that would not feel simulated or monitored so that we could observe behavior that is closer to “actual” behavior. At the end of our time in each community, the first author and the research team gathered community members together to thank them for their participation and to share refreshments. During this goodbye party, we presented the group with one battery-operated portable stereo and a set of 14 cassette tapes of the radio program. We acknowledged that there was only one, and suggested that because they were all together they could use this time to decide how to share the stereo, when the group members would listen, and how to keep it supplied with batteries. After this casual suggestion, the research team thanked the participants and sat back down.

Given the monetary value of a portable stereo, this discussion was of great significance to the participants. The measure also captured spontaneous behavior that participants believed to be “off record”—their discussions took place immediately after the research team's presentation. Two researchers sitting discreetly in the back of the group recorded these ensuing discussions by hand, with notebooks in front of them to appear as if they were finishing notes from earlier in the day.

Did the reconciliation radio program's messages have an effect on actual participant behaviors regarding community issues of consequence? Of interest was whether the participants allowed for multiple viewpoints and whether they expressed optimism about cooperating to share this community property. We were also interested in the groups’ ultimate decisions about how to share the communal stereos.

During the health groups’ deliberations about the portable stereo and cassettes, we frequently observed the following pattern: the first member of the group to speak would propose handing over the stereo and cassettes to the village's local authority, who could regulate usage and financial contributions for the batteries. Following this proposal, group members would overwhelmingly support the motion and close the discussion.

In the reconciliation groups, deliberations typically followed a different pattern. After the same initial proposal to entrust the stereo to the authorities, one or more of the participants would challenge this suggestion, for example, claiming that the group should be collectively responsible or should elect one of their members to manage it. Comments about the group's ability to cooperate came up more frequently, such as “We've been coming together to listen all of this time, why can't we come together to listen to this stereo together, just as we did before?”

These different patterns were borne out by statistical analyses of the coded transcripts. Table 5 presents the number of dissenting propositions that followed the initial proposal to assess the extent and openness of the deliberation session. This indicator reveals that reconciliation groups proposed and debated a significantly greater number of views on how to share the communal stereo compared to health groups (using a Mann-Whitney test, z = −2.6, p < .01; using Wilcoxon signed-rank test, which controls for the site strata, z = −2.2, p < .05). The number of positive comments made about group cooperation was also more frequent in reconciliation groups (using a Mann-Whitney test, z = −2.6, p < .01; using Wilcoxon signed-rank test, which controls for the site strata, z = −2.2, p < .05). Importantly, these results were the same for ethnically homogenous Twa and survivor communities and for heterogeneous communities in the general population.

TABLE 5.

Behaviors Observed During Discussion of Community Resource

Notes: Letters a–g each represent a pair of communities; in each pair, one community was randomly assigned to listen to the health program and the other, the reconciliation program. Transcripts from community discussions were assigned one point for offering a dissenting opinion about the initial position taken by one or more members of the group and for any comment made about the group's ability to cooperate.

Heterogeneity of Experimental Effects

Table 6 summarizes individual- and group-level results for each separate research site, underscoring the remarkable homogeneity of the reconciliation program effects across a range of Rwandan communities. As described in this section, we do find exceptions to this homogeneity in two ethnic subcommunities: the northwest (predominantly Hutu) communities, and the genocide survivor (Tutsi) communities. We test but find little evidence of significant interactions between individual characteristics and exposure to the reconciliation program.

TABLE 6.

Overview of Results

R, reconciliation; H, health.

Notes: *p < .05; **p < .01. Pluses and minuses represent the prediction for each variable, where a plus indicates that we expect higher scores for the reconciliation listeners.

Entries for columns 1 & 2 are averages of the ratings reported by each individual to statements about dissent and social distance; entries for columns 3–7 are counts of group-generated behavior observed in focus groups (3), role plays (4 & 5), and community negotiation (6 & 7).

^a Columns 1 & 2 present ordered probit estimates of treatment effects at the individual level, columns 3–5 present ordered probit estimates of treatment effects at the group level, and the Z statistics in columns 6 & 7 are Mann-Whitney test results of treatment effects at the group level. Focus groups and role plays were not conducted in Twa, western, and northeastern pairs.

Community Characteristics. At the community level, Hutu living in the northwestern research sites and Tutsi living in the genocide survivor communities deviated from the general pattern of experimental findings. Although our small sample of each type of community does not allow us to rule out the possibility of chance fluctuations, these interactions between community identity and reaction to the radio program seem to accord with the communities’ political standings in Rwanda.

As described previously, the predominantly Hutu communities in the northwest region (the seat of Hutu extremism prior to the genocide) are closely monitored by the current Tutsi regime. During data collection, citizens in this region were keenly aware of government surveillance as a result of attacks launched by ex-Habyarimana military from the nearby border, which were reportedly supported by members of the local population (in focus groups, many participants discussed the importance of trusting that your neighbor will not harbor “infiltrators” and thus invite repercussions on the whole community). Table 6 shows that unlike the rest of the sample, the northwest reconciliation listeners were not more likely than the northwest health listeners to perceive a norm of dissent or actually to dissent in the group deliberation (in fact, neither group dissented at all). However, we observe an experimental effect on northwest reconciliation listeners’ open discussions of mistrust, cooperation in group deliberation, and reliance on local aid in the role plays. These findings tell a story consistent with the region's political standing: the reconciliation program successfully facilitated better local interaction and cooperation, but had no effect on willingness to speak out in the context of government surveillance and censure.

Results for the genocide survivor communities stand out in a different way. Genocide survivors are the least likely of all communities in the sample to be censured for politically incorrect speech, as coethnics of the current government and as the government's living justification for speech restrictions and international aid. At odds with government rhetoric about Rwandan unity, and contrary to the rest of the communities’ patterns, no genocide survivor publicly denied the existence of mistrust in their communities. In addition, Table 6 shows that reconciliation listeners in the survivor community called on the government for aid almost as frequently as they offered their own aid in role plays. In line with the rest of the sample, however, survivor reconciliation listeners were more likely to perceive a norm of dissent and to express dissent in the group deliberation compared to survivor health listeners.

Although the two communities react differently to the radio program with respect to certain outcomes, it is important to remember that these communities are similar to other communities with respect to other outcomes—dissent among survivors, and collective action among the northwest communities. In other words, the communities’ politicized identities did not push them entirely out of the radio's range of influence.

Individual Characteristics. Our data do not reveal differences in experimental effects according to individual characteristics such as age, experience of violence, and access to media (see Appendix V, Tables A2 and A3). Although some theories of political cultural change predict that younger generations are more likely to adopt new cultural values (Inglehart Reference Inglehart1997, 15; c.f. Prewitt Reference Prewitt1975), like Rogowski (Reference Rogowski1974), we find no real differences for older and younger cohorts.Footnote ¹⁴ Models of self-interest (Sears and Funk Reference Sears and Funk1991) or grievance and blame (Gurr Reference Gurr1970; Iyengar Reference Iyengar1991) might predict the radio impact would vary according to the magnitude of participants’ personal losses in the 1994 violence—but we do not find this to be the case. Treatment effects do not appear to vary by a pretest question asking how many family members were killed by war, genocide, or related violence in 1994. Finally, theories of media sophistication and socialization (Gerbner et al. Reference Gerbner, Gross, Morgan, Signorielli, Shanahan, Bryant and Zillmann2002; Heath and Bryant Reference Heath and Bryant1992) would predict that radio ownership might affect participants’ responses; radio ownership also does not interact with experimental media exposure. This finding also addresses concerns that participants with radios surreptitiously listened to radio programs not assigned to their group in their own time. If contamination had been a serious problem, we might have found different effects for those who own radios.