Research casts more doubt on the assumption that ‘hate speech’ causes violence

I’m pleased to announce that the political science journal Civil Wars has just published my article, “Did Radio RTLM Really Contribute Meaningfully to the Rwandan Genocide?: Using Qualitative Information to Improve Causal Inference from Measures of Media Availability.” The article, which is currently available online and will be published in print in an upcoming issue of the journal, questions whether, and if so to what extent, the broadcasts of the notorious “hate radio” station, Radio Télévision Libre des Mille Collines, caused listeners to participate in the Rwandan genocide in 1994.

The genocide, in which an estimated 800,000 to 1 million people were killed and 120,000 persons were subsequently charged with participating, took place under conditions which scholars have long recognized are conducive to genocide and other forms of mass violence. It took place during a civil war between a rebel army that was largely composed of members of the minority Tutsi ethnic group, and the government, which was largely composed of members of the majority Hutu ethnic group. There had been previous incidents of mass violence in the country. In addition, when the country became independent of Belgium in 1962, its leaders established a narrative that defined the nation as being Hutu, while defining the Tutsi as interlopers. Despite the fact that many of the conditions for genocide were clearly present in Rwanda, many politicians and media observers at the time placed particular blame for the genocide on the content of broadcasts by RTLM.

The RTLM broadcasts were odious — they routinely described Tutsis as “cockroaches” and explicitly encouraged Hutu civilians to join in their killing. Moreover, some broadcasts were, essentially, part of a conspiracy to commit murder, since on more than one occasion they instructed those engaged in killing on where to find victims, and even broadcast the license plate numbers of some of those who were fleeing the killing. (In America’s legal system, this would almost certainly meet the threshold for unlawful incitement that loses First Amendment protection.) It was for those broadcasts that RTLM executives were convicted of genocide by the International Criminal Tribunal for Rwanda. The popular conception of RTLM’s influence on the genocide goes much further, however, maintaining that the broadcasts caused people to hate, and then to kill, their neighbors. But most researchers have found little direct evidence to support that claim.

One exception is a study published in 2014 in the Quarterly Journal of Economics, which used estimates of radio reception and found that areas of Rwanda with better radio reception had higher levels of participation in the genocide, as measured by the number of prosecutions for genocide in local courts. The study estimated that ten percent of the violence was associated with RTLM broadcasts; however, the estimated effect was greater for those who were charged with being organizers of the violence than it was for those who were mere participants, and much of that effect was “spillover,” rather than direct persuasion (that is, high levels of radio reception in one village was associated with higher levels of violence in neighboring villages).   

My article casts some doubt on the findings of this study. I argue that radio reception is not really what the study is trying to analyze; rather, the study is trying to analyze the effect of radio consumption, and it uses radio reception as a proxy variable — i.e., as a way to indirectly estimate radio consumption. However, it is impossible to know how well radio reception represents radio consumption without deep knowledge of radio listening habits in the area. For example, if people are in the habit of gathering in places that have radio reception, most of the people in a village might listen to the radio, even if the radio signal reaches only a small part of the village. Indeed, one study 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.” If the militia member had to stand on the roof, radio reception was likely poor in that area, yet radio consumption (by him and his audience) was much higher than indicated by measures of radio reception.  

In the article, I review research on radio listening habits in Sub-Saharan Africa and find that there is strong evidence to doubt that a measure of radio reception is a good proxy for radio consumption. For example, people in Rwanda and other areas of Sub-Saharan Africa often listen to the radio in groups, gathering in areas of a village that have good radio reception. Hence, a village might have radio reception in only 20 percent of its area, yet the radio might be heard by much more than 20 percent of its residents. In addition, many crops in Rwanda are cultivated on slopes, which implies that Rwandans spend much of their time at relatively high altitudes, where radio reception is relatively good, rather than in valley bottoms, where radio reception is poor. Again, an area with deep valleys might have large geographical areas with no reception, yet most of its residents might nevertheless listen to the radio when working at higher elevations.

The original article used relatively standard multiple regression to find that a variable measuring radio reception is a statistically significant correlate of higher levels of participation in genocide. However, a 2010 paper by political scientists at Duke, the University of Washington, and University College London has shown that statistically significant variables often do not help predict the outcomes with which they are correlated. In my article, I used a technique recommended by those professors to test whether radio reception helps predict which areas of Rwanda had more participants in genocide.

That technique is relatively straightforward. Step one is to see how a well a model that includes all of the variables used in the original research (such as radio reception, poverty levels, education levels, etc.) predicts the amount of violence in a particular area. Step two is to repeat that process using a model with all of the original variables except radio reception. If the second model does worse at predicting violence than the first model did, then radio reception must be very important. On the other hand, if the second model does just as well as the first model did, then radio reception does not help the model to predict violence, and therefore radio reception is unlikely to cause violence. My results found that there was almost no difference between the predictive ability of the model with radio reception and the model without radio reception.   

That finding does not, of course, prove that the RTLM broadcasts had no effect on the genocide. As I mentioned earlier, there is strong evidence that some broadcasts were used explicitly to assist perpetrators who were already engaged in genocide. However, those who contend that “hate speech” causes violence usually make a different claim, which is that that “hate speech” inspires ordinary people to hate others and to engage in violence against them when they would not otherwise be inclined to do so. The Rwandan genocide is often used as “Exhibit A” in support of that argument. My paper casts serious doubt on that premise, and lends further support to the scholarship I cited above, which found little evidence that the RTLM broadcasts inspired violence in Rwanda.