I measure onset of genocidal violence as a nominal variable of early, middle, and late onset. This measure is obtained from two groups of scholars—Scott Straus (2006) and the Genodynamics Project (2013). Straus drew upon six key sources to determine the onset date for each commune. These include transcripts from the International Criminal Tribunal for Rwanda (ICTR) and reports from a commission appointed by the Rwandan Ministry of Higher Education, Human Rights Watch, African Rights, Ibuka (a Rwandan NGO), and Straus’ interviews with 230 perpetrators of violence (Straus 2006:249). In his book, Straus provides each of the onset estimates given in the reports and a final onset estimate he created after weighing the potential credibility of each report.
Christian Davenport and Allan Stam’s Genodynamics Project (2013) also sought to ascertain when violence started in each commune. They coded five sources—including the reports from the Ministry of Higher Education, Human Rights Watch, African Rights, and Ibuka—as well as a report from the Ministry of Youth, Culture, and Sport. Then, using Bayesian methodology, they analyzed these data to ascertain the best estimate for the onset date in each commune by allocating various amounts of credibility to the sources. They also assessed the validity of the dates by conducting interviews at the ICTR.
I obtained access to both of these datasets. As the specific date of onset varied widely across sources and when comparing the two datasets, here I rely on a measure of early, middle, and late onset, sacrificing precision in favor of validity. I define early onset as April 6-8th. While this is a small range of time, almost a third of the communes saw violence within these days. Middle onset is defined as April 9-14th, while late onset is
defined as April 15th and later. This matches Straus’ assessment of early, middle, and late onset, and it breaks the communes into three relatively even groups.
To assign the value of early, middle, or late to a commune, I followed three steps. First, if Straus’ estimate and Davenport and Stam’s fell within the same category (early, middle, or late onset), I assigned that category to the commune. Second, if one of the datasets was missing the onset date, the onset data from the other dataset was assigned, resulting in a total of 92 communes (onset agreement).110 Third, for the remaining communes, I analyzed the sources used by the two different research teams and drew a conclusion based on the agreement of the sources, was always in line with either Straus or Stam and Davenport (onset expanded). I use the more complete measure, onset expanded, for analysis, though I also restrict all analyses to the 92 communes (onset agreement) to ensure that the results remain similar across the two case sets (not shown). In addition, even with the more complete measure, reliable data on the onset of violence do not exist for 26 communes, and they are excluded.
Table 3.4: Commune Onset Dates
Data Source Early Middle Late Total
Onset (agreement) 31 (33.7%) 27 (29.4%) 34 (37.0%) 92 (100%) Onset (expanded) 43 (36.1%) 40 (33.6%) 36 (30.3%) 119 (100%)
These dependent variables are summarized in Table 3.4, while Table 3.5 includes descriptive statistics by onset date. All independent variables match those above with the exception of the spatial autocorrelation weight, which now is a measure of the onset in
Table 3.5 Mean Values of Independent Variables by Onset (Expanded)
Independent Variables Early Middle Late
Targeted Violence and Ideology
Tutsi 9.99% 8.62% 13.51% Catholic 56.63% 63.83% 67.83% Radio Ownership 36.76% 32.28% 34.22% Education 2.84 2.86 3.2 Community Organization Always Lived 67.08% 73.10% 69.16% Divorce 2.99% 2.96% 3.66% Marriage 59.77% 57.44% 52.41% Unemployment 0.22% 0.27% 0.15% Formal Employment 11.64% 10.75% 11.30% Young Men 12.22% 12.57% 13.50% Intermarriage 4.65% 4.59% 8.62% Resource Competition Population 56,854 47,937 43,960 Population Density 399.04 455.81 411.68 Population Growth 3.47% 2.66% 2.25% Kilocalorie Production 206.18 219.68 193.63 Hutu Formal Employment 10.99% 10.38% 10.55% Tutsi Formal Employment 17.97% 18.56% 21.50%
Organized Actors (April)
RPF Troops 30.23% 25.00% 5.56%
FAR Troops 88.37% 90.00% 94.44%
RPF/FAR Frontline 46.51% 40.00% 2.78%
Broader Spatial and Temporal Factors
Elevation 1,768.85 1,778.09 1,668.99
Capital Region 100% 0% 0%
Urban Region 16.28% 15.00% 2.78%
Distance from Roads 4.78 4.66 4.25
Distance from Cities 17.80 15.03 14.69
Borders Burundi 14.00% 5.00% 16.67%
International Border 41.86% 20.00% 19.44%
Past Violence 11.63% 2.50% 0%
Distance from Kigali 65.22 59.57 56.14
Surrounding Onset 1.50 1.80 2.64
Note: While these values are transformed in the analyses and in the other descriptive tables, values presented here are in original units.
surrounding communes.111 In addition, troop location is restricted to troop presence in April, rather than throughout the entire episode of violence, since all onset dates were in April. Because of this, the presence of French troops is excluded. Population, which is included in the dependent variable (a rate) above, is now tested as an independent variable. Lastly, I include a measure of distance from Kigali to test the theory that violence radiated outward.
Descriptive statistics were also analyzed for the 26 communes missing an onset date. These communes had significantly lower percentages of Tutsi (three percent compared to eight percent in the non-missing communes). In addition, they had much higher average elevation (averaging over 1880 km2) and much lower numbers of people killed (averaging 2,000 deaths per commune compared to the total average of over 7,000 deaths per commune). The comparatively lower number of people killed likely explains why the onset date is unknown and, overall, illustrates that communes excluded are those with much lower levels of killing.112
Analysis
To analyze the factors associated with the onset of violence within a commune, I employ multinomial logistic regression, an appropriate model when the dependent variable is categorical and there are more than two outcomes. In this case, as discussed above, there are three outcomes (early, middle, or late), and early onset is excluded as the
111 I created the onset variable manually, as GeoDa is unable to calculate onset when
missing values are included. Thus, it is the average onset (1, 2, or 3, corresponding to early, middle, and late, respectively) of surrounding communes excluding those with missing values.
112 Currently, there is no way to remedy this. However, on a future trip to Rwanda, I will
conduct interviews to better understand the onset in these regions (specifically, before turning this dissertation into a book).
comparison outcome. As with models of magnitude presented previously, I begin by building bivariate models, shown in Table 3.6. I review these briefly and then turn to multivariate models, illustrated in Table 3.7. Note that results are presented in odds ratio, so coefficients larger than one suggest that the variable is associated with higher odds of middle or late onset in comparison to early onset, while coefficients less than one suggest that the variable is associated with lower odds of middle or late onset in comparison to early onset. Note also that I cluster standard errors by prefecture due to the nested nature of the data.113