The Early Warning Project aims to advance work on atrocities forecasting and prevention by bringing the best available methods and data to bear on these difficult problems and sharing the results with the public. As methods and data emerge and evolve, our project will continue to evolve, too. This Labs page spotlights our most active areas of research and development.
The Early Warning Project is using an innovative survey method called a pairwise wiki survey to elicit forecasts and opinions from experts and sometimes wider crowds on risks of mass atrocities and strategies for preventing or mitigating them.
Pairwise wiki surveys were developed by a Princeton-based project called All Our Ideas. The method involves a single question with many possible answers on which respondents vote in pairs that are selected automatically. Votes in a wiki survey ask respondents to make comparisons, so the method is particularly well suited to questions about relative risk ("Which event is more likely?") and policy or advocacy options ("Which action would be better?").
Wiki surveys differ from traditional survey in a couple of other important ways. First, a respondent never "finishes" a wiki survey; instead, the software keeps generating pairs and respondents may keep voting, as many times as they like in as many sessions as they like. This gives respondents who care more about an issue a chance to have their input weighted more heavily in the results. Second, in wiki surveys, respondents can add their own ideas to the list, and those added ideas will immediately become part of the mix on which everyone subsequently votes. This feature allows new ideas or concerns to "bubble up" from the crowd in a way that traditional surveys do not.
The Early Warning Project has already run a number of wiki surveys, including one on risks of state-led mass killing in 2014 (here, with results mapped below) and ones on policy options in Syria (here) and South Sudan (here). We plan to continue using this instrument routinely to ask about relative risks at the global level and occasionally to ask about risks and strategies for preventive action at the global, national, or subnational level.
The Early Warning Project is exploring ways to use new computer-generated data sets on political events around the world to monitor atrocities on a daily basis.
For decades, social scientists have produced event data with teams of trained readers who would read and summarize news stories on specific cases or issues. The data these projects produced have been vital to research on many topics, but the labor involved has made these data sets prohibitively expensive to build and sustain, and they have been limited by the small number of news sources available to them. Those constraints are finally falling away, however, thanks to the growth of the Internet and improvements in the computer hardware and software used to search, scrape, translate, and process its contents. The resulting machine-coded data sets stand on the analytical foundations of the projects that preceded them, but they are able to achieve much broader coverage at a tiny fraction of the cost. Updates that used to take large teams of trained researchers hours or days to produce can now be produced in seconds and at any time.
These machine-coded event data sets may include thousands or even tens of thousands of new records each day, but only a small portion of those relate to atrocities. Since 2013, the Early Warning Project has been experimenting with different strategies to cull and report information about atrocities from these torrents of data. Our initial research (see here, here, and here) focused on the Global Database of Events, Language, and Tone (GDELT), but we are now planning to focus our work on data sets being developed by the Open Event Data Alliance and, if they are made public, the U.S. government-funded ICEWS project.
As our project gains its footing, we hope to routinize and scale up our monitoring efforts to allow visitors to our web site to produce charts, maps, and animations for countries or groups of interest to them. Over time, we also plan to begin using these streams of machine-coded event data to develop new, more dynamic statistical models of atrocities risk at the national and subnational levels.