Search the Entire SiteSearch
Genocide and mass atrocities are devastating crimes in their scale and scope, in their enduring scars for survivors, and in the long-term trauma they cause in societies where they occur.
Such crimes also pose serious threats to US interests, particularly for national security and the economy. Murders on a mass scale can destabilize entire regions because of the resulting mass displacement of people and the propensity of violence to spill across borders. Responding to the resulting humanitarian catastrophe—with food, medical, and refugee aid—becomes an international obligation that often continues decades after the violence ends.
Despite past efforts to address systematic killing, and a body of law formed after the Holocaust to prevent and punish perpetrators, such crimes persist.
In studying these tragedies, we have learned that genocides are never spontaneous. They are always preceded by a range of early warning signs. If these signs are detected, their causes can be addressed, preventing the potential for catastrophic progression.
The United States Holocaust Memorial Museum’s founding charter, written by Holocaust survivor Elie Wiesel, mandates that our institution strive to make preventive action a routine response when the signs of genocide appear. Wiesel wrote that “only a conscious, concerted attempt to learn from past errors can prevent recurrence to any racial, religious, ethnic or national group. A memorial unresponsive to the future would also violate the memory of the past.”
The Museum’s Early Warning Project is part of the Simon-Skjodt Center’s efforts to fulfill this aspect of the Museum’s mandate: to prevent genocide by doing for today’s victims what was not done for the Jews of Europe.
In partnership with Dartmouth College, we studied past situations where governments and non-state groups systematically targeted and killed thousands of civilians. In doing so, we identified a range of patterns and circumstances that often precede such violence. On the basis of this research, we developed a “Statistical Risk Assessment”—published since 2014—which uses a range of publicly available data to identify countries where similar conditions exist and, as a consequence, the risk of mass atrocities is elevated.
The assessment identifies the risk—the possibility—that a mass killing may take place. On average, one or two countries experience a new episode of mass killing each year. But relative infrequency does not make the brutality less devastating for victims: a mass killing, by our definition, is 1,000 or more civilians deliberately killed by armed forces (whether government or non-state), over the period of a year or less, because of their membership in a particular group. Virtually all cases of genocide include mass killings that meet this definition.
In many places, such violence is ongoing—in countries such as Burma, Syria, and South Sudan. These cases are well known and we address them in this report. But our risk assessment’s primary focus—and the gap we seek to fill—is to highlight those cases where systematic mass atrocities have not yet begun: the places where people live under the shadow of violent persecution.
Preventing genocide is of course difficult. In deciding how to respond, policy makers face an array of constraints and competing concerns. The choice to prevent one potential tragedy often takes a back seat when policy makers are confronted by multiple ongoing conflicts. We aim for this risk assessment to serve as a tool and a resource for policy makers and others interested in prevention. We hope this helps them better establish priorities and undertake the discussion and deeper analysis that can help reveal where preventive action can make the greatest impact in saving lives.
Simon-Skjodt Center for the Prevention of Genocide
The Early Warning Project’s Statistical Risk Assessment uses publicly available data and statistical modeling to produce a list of countries ranked by their likelihood of experiencing an onset, or new episode, of mass killing.
By our definition, a mass killing occurs when the deliberate actions of armed groups in a particular country (including but not limited to state security forces, rebel armies, and other militias) result in the deaths of at least 1,000 noncombatant civilians in that country over a period of one year or less. The civilians must also have been targeted for being part of a specific group. Mass killing is a subset of “mass atrocities,” which we define more generally as “large-scale, systematic violence against civilian populations.”
This report highlights findings from our Statistical Risk Assessment for 2018–19, focusing on:
We recognize that this assessment is just one tool. It is meant to be a starting point for discussion and further research, not a definitive conclusion. We hope it will help governments, international organizations, and nongovernmental organizations determine where to devote resources for additional analysis, policy attention, and, ultimately, preventive action.
Our 2018–19 assessment is the result of new and refined data and statistical methods that have been incorporated to maximize the accuracy and practical utility of the results. The changes reflect our commitment to continually learn from developments in statistical forecasting practice and our experience working to translate early warning analysis into effective preventive action. As a result of these changes, risk estimates and rankings from 2014 through 2016 should not be compared directly with results from 2017 onward.
The three most important changes from our past assessments are:
1. The new results incorporate the risk of both state-led and non-state-led mass killings, addressing a significant gap in our past approach, which focused exclusively on state-led mass killings. In the last few years, numerous non-state armed groups—including the Islamic State (IS), Boko Haram, and militias in South Sudan and Central African Republic—have perpetrated widespread atrocities. As data on such events have become available, we have incorporated those risks into our annual assessment. Note: Our model estimates risk at the country level and does not identify the potential perpetrators of the violence. Additional analysis would be required to assess whether potential perpetrators in a specific country are state or non-state actors, and if the latter, which specific groups.
2. The new results are estimates of risk over a two-year period (January 2018–December 2019); past results were estimates of risk over one year. Two-year forecasts allow more time for results to be addressed—for the necessary analysis and planning and to implement preventive actions. In addition, considering that the exact timing of a mass killing onset is difficult to forecast, the two-year horizon allows the model to perform with higher accuracy.
Both of these changes in our methodology have an impact on the risk calculation for each country. Followers of our project will note that the percentage risk estimated for most high-risk countries in the 2018–19 assessment is significantly higher than in previous published assessments. This is the case due to the two changes cited above: 1) Our estimates now reflect the likelihood of state-led and non-state-led mass killing and 2) our new model now assesses risk for two years—2018 and 2019. As stated above, because of these changes, risk estimates and rankings from 2014 through 2016 should not be compared to results from 2017 onward.
3. Our risk assessment uses newly-available data, most notably several measures from the Varieties of Democracy (V-Dem) dataset on governance and human rights characteristics, which could improve the accuracy of our assessment. Even with these additions, however, the availability and quality of cross-national quantitative data are significant constraints: We lack systematic data on some hypothesized warning signs, such as hate speech. And in situations where governments deliberately restrict access to international observers, such as in Burma’s Rakhine state or China’s Xinjiang region, existing data might not fully reflect conditions on the ground.
Figure 1: Risk assessment heat map of estimated likelihood of onset of mass killing, 2018–19 (Data: Early Warning Project)
Before discussing the results, we underscore four points about interpreting this Statistical Risk Assessment:
First, as a statistical matter, mass killings are rare. On average, just over one percent of countries see a new mass killing in any given year—that means one or two countries. Our risk model predicts a similar number of onsets, so the average two-year risk score produced by our model is between two and three percent. Just eight out of 162 countries have a two-year risk score greater than 10 percent, and the highest-risk country we assessed has about a one in four chance of experiencing a new mass killing in 2018 or 2019.
Second, our model is designed to assess the risk of an onset of new mass killing, not of the continuation or escalation of ongoing episodes. This feature is especially important to bear in mind when interpreting results for countries that are currently experiencing mass killing events, such as Burma/Myanmar and Syria (see Figure 1 and our website for a full list of these countries). For these countries, our assessment should be understood as an estimate of the likelihood that a mass killing event would be launched by a different perpetrator or targeting a different civilian group in 2018 or 2019. Our model estimates that having a mass killing currently in progress is associated with lower risk of another such conflict beginning, as it is rare for a country to have two distinct mass killing episodes concurrently.
Third, for practical reasons, we only forecast mass killings within countries (i.e., in which the perpetrator group and the targeted civilian group reside in the same country. This risk assessment does not forecast interstate conflict). Situations in which large numbers of civilians are killed deliberately by an armed group from another country are not captured in our historical data or current forecasts. This decision does not involve a value judgment about the moral or practical significance of such atrocities, only a pragmatic judgment about what we are able to forecast reliably.
Fourth, readers should keep in mind that our model is not causal: The variables identified as predicting higher or lower risk of mass killings in a country are not necessarily the factors that drive or trigger atrocities. For example, large population size does not directly cause mass atrocities; however, countries with large populations have been more likely to experience mass killing episodes in the past, so this factor helps us identify countries at greater risk going forward. We make no effort to explain these kinds of relationships in the data; we only use them for their predictive value. An important consequence of the non-causal nature of these forecasts is that actions aimed at addressing risk factors are not necessarily effective ways of mitigating the risk of mass atrocities. For example, although countries that have banned opposition parties are at greater risk of experiencing mass killings than are other countries, our analysis does not imply that action to encourage or pressure governments to allow participation by opposition parties would necessarily reduce the risk of mass killings.
Figure 2: Top 30 countries by estimated likelihood of onset of mass killing, 2018–19 (Note: * indicates ongoing state-led mass killings; ° indicates ongoing non-state-led mass killings. Some countries have multiple ongoing episodes. The probabilities displayed here are associated with the onset of an additional mass killing episode. See the full list of ongoing mass killings.)
Figure 2 displays the estimated likelihood of a new onset of mass killing (state-led or non-state-led) in 2018 or 2019 for the 30 highest-ranked countries. For every country in the top 30, we recommend that policy makers consider whether they are devoting sufficient attention to addressing the risks of mass atrocities occurring within that country. As noted earlier, our model generates a single risk estimate for each country; additional analysis is required to assess whether potential perpetrators in a specific country are state or non-state actors, and if the latter, which specific groups. Here we highlight the three countries that topped our risk list in both the 2017–18 and 2018–19 assessments and the top factors accounting for their risk estimates (see “About the modeling process” below for more detail on the risk factors in the model):
The remaining seven countries in the top ten are South Sudan, Pakistan, Yemen, Angola, Turkey, Sudan, and Somalia. Analysis of Angola, Yemen, and Turkey is included later in this report. To learn more about what accounts for the high risk score of any of these countries, visit our website.
One way in which global statistical risk assessments are helpful is in identifying countries whose relatively high (or low) risk scores may surprise regional experts. In cases where our statistical results differ substantially from expectations, we recommend conducting deeper analysis and revisiting assumptions. We highlight three countries that, in our informal judgment, fall into this category.
Several countries have appeared near the top of our rankings for several years but have yet to experience a mass killing onset.
In addition to Côte d’Ivoire and Haiti, discussed previously, two countries moved up in our rankings substantially between the 2017–18 and 2018–19 assessments, and one declined significantly.
On the basis of prior empirical work and theory, we selected more than 30 variables, or risk factors, as input for our statistical model (see the discussion of our modeling approach, below). All data used in our model are publicly available, regularly updated, and available without excessive delay. They also have, in our estimation, minimal risk of being retrospectively coded in ways that could depend on observed mass killings or their absence. We include variables reflecting countries’ basic characteristics (e.g., the number of years a country has existed, geographic region, population); socioeconomic measures (e.g., changes in gross domestic product per capita); measures of governance (e.g., regime type); levels of human rights (e.g., freedom of movement); and records of violent conflict (e.g., battle-related deaths, ongoing mass killings).
Our modeling approach is described in detail on our website. In summary, based on the variables detailed above and data on mass killing going back to 1945, the algorithm identifies predictive relationships in the data, resulting in an estimated model. We then apply this model to recent data (from 2016 for the 2017–18 assessment and from 2017 for the 2018–19 assessment) to generate forecasts. Although expert knowledge is inherently involved in our choice of which variables to collect, the model determines the contribution of each variable to the risk score without input from us. The particular modeling approach we use is also able to “drop” factors from the model if they are not informative regarding risk.
According to our forecasting model, the factors determined to be informative of the risk of mass killing include large population size; lack of freedom of movement for men; anocratic regime type (i.e., neither full democracy nor full autocracy); uneven respect for civil liberties; high ethnic fractionalization; high infant mortality; geographic region; history of mass killings; existence of politically motivated killings; duration of current regime; high numbers of battle-related deaths; banning of opposition parties; not being a state signatory of Optional Protocol to the International Covenant on Civil and Political Rights; coup attempts within the past five years; country age; lack of ongoing mass killing, repression of civil society; and recent significant alteration of judicial powers. We emphasize again that these risk factors should not be interpreted as causes or “drivers” of risk but simply as correlates of risk that have proven useful in forecasting. For a complete description of the variables used in our model, see our data sources.
We assessed the accuracy of this model in ways that mimicked how we use its results: We built our model on data from a period of years and then tested its accuracy on data for later years (i.e., we conducted out-of-sample testing). Our results indicate that eight out of every ten countries that later experienced a new onset of mass killing had risk estimates of greater than 4 percent (which usually meant they were among the 30 top-ranked countries in a given year). We are preparing a technical paper in which we assess our model and others according to multiple performance measures.
from the Early Warning Project and the Simon-Skjodt Center for the Prevention of Genocide