The bar charts below show all countries in descending order by estimated risk. The length of the bar indicates the estimated risk, and the colors in the bars illustrate the contributions of those three models to that estimate. In the bottom two groups, some of the countries appear to have no data, but that's really just showing that their estimates are virtually indistinguishable from 0.
The Statistical Risk Assessments estimate the susceptibility of every country worldwide to an onset of state-led mass killing in the next year or so.
Our risk assessments come from a set of statistical models that represent a few different ideas about how best to anticipate when mass killing will occur. One model, labeled Bad Regime, emphasizes characteristics of a country’s national politics that hint at a potential for a regime to commit mass atrocities, especially in the context of political instability. The second model, labeled Elite Threat, emphasizes potential threats to a regime’s hold on power that might spur it to lash out against those threats in the form of a mass killing. The third model, labeled Random Forests, is actually a machine-learning algorithm that we apply to all of the factors identified by the other two. This model presumes that the subject-matter experts have identified the most important predictors but is more flexible in its search for patterns within the resulting data.
The Statistical Risk Assessment is produced by averaging the forecasts from these three models. We know from research that this model-averaging approach generally produces more accurate risk assessments than we could expect to get from any one model alone. Combining the forecasts allows us to learn from three different perspectives while hedging against the biases of any one of them. At the same time, we can also compare forecasts across the three models within each country to get a sense of why that country lands where it does in the global rankings.