Later this month, I hope to have the data I need to generate and post new statistical assessments of coup risk for 2015. Meanwhile, I thought it would be interesting and useful to experiment with applying a crowdsourcing tool to this task. So, if you think you know something about coup risk and want to help with this experiment, please cast as many votes as you like here:
For this exercise, let’s use Naunihal Singh’s (2014, p. 51) definition of a coup attempt: “An explicit action, involving some portion of the state military, police, or security forces, undertaken with intent to overthrow the government.” As Naunihal notes,
This definition retains most of the aspects commonly found in definitions of coup attempts [Ed.: including the ones I use in my statistical modeling] while excluding a wide range of similar activities, such as conspiracies, mercenary attacks, popular protests, revolutions, civil wars, actions by lone assassins, and mutinies whose goals explicitly excluded taking power (e.g., over unpaid wages). Unlike a civil war, there is no minimum casualty threshold necessary for an event to be considered a coup, and many coups take place bloodlessly.
By this definition, last week’s putsch in the Gambia and November’s power grab by a lieutenant colonel in Burkina Faso would qualify, but last February’s change of government by parliamentary action in Ukraine after President Yanukovich’s flight in the face of popular unrest would not. Nor would state collapses in Libya and Central African Republic, which occurred under pressure from rebels rather than state security forces. And, of course, Gen. Sisi’s seizure of power in Egypt in July 2013 clearly would qualify as a successful coup on these terms.
In a guest post here yesterday, Maggie Dwyer identified one factor—divisions and tensions within the military—that probably increases coup risk in some cases, but that we can’t fold into global statistical modeling because, as often happens, we don’t have the time-series cross-sectional data we would need to do that. Surely there are other such factors and forces. My hope is that this crowdsourcing approach will help spotlight some cases overlooked by the statistical forecasts because their fragility is being driven by things those models can’t consider.
Wiki surveys weren’t designed specifically for forecasting, but I have adapted them to this purpose on two other topics, and in both cases the results have been pretty good. As part of my work for the Early Warning Project, we have run wiki surveys on risks of state-led mass killing onset for 2014 and now 2015. That project’s data-makers didn’t see any such onsets in 2014, but the two countries that came closest—Iraq and Myanmar—ranked fifth and twelfth, respectively, in the wiki survey we ran in December 2013. On pro football, I’ve run surveys ahead of the 2013 and 2014 seasons. The results haven’t been clairvoyant, but they haven’t been shabby, either (see here and here for details).
I will summarize the results of this survey on coup risk in a blog post in mid-January and will make the country– and vote-level data freely available to other researchers when I do.
I don’t necessarily plan to close the survey at that point, though. In fact, I’m really hoping to get a chance to tinker with using it more dynamically. Ideally, we would leave the survey running throughout the year so that participants could factor new information—credible rumors of an impending coup, for example, or a successful post-election transfer of power without military intervention—into their voting decisions, and the survey results would update quickly in response to those more recent votes.
Doing that would require modifying the modeling process that converts the pairwise votes into scores, however, and I’m not sure that I’m up to the task. As developed, the wiki survey effectively weights all votes the same, regardless of when they were cast. To make the survey more sensitive to fresher information, we would need to tweak that process so that recent votes are weighted more heavily—maybe with a time-decaying weighting function, or just a sliding window that closes on older votes after some point. If we wanted to get really fancy, we might find a way to use the statistical forecasts as priors in this process, too, letting the time-sensitive survey results pull cases up or push them down as the year passes.
I can imagine these modifications, but I don’t think I can code them. If you’re reading this and you might like to collaborate on that work (or fund it!) or just have thoughts on how to do it, please drop me a line at ulfelder at gmail dot com.