Earlier this year, I participated in a workshop at the Council on Foreign Relations on predicting political instability. Not surprisingly, the decision to hold the workshop was partially inspired by events in the Middle East and North Africa, and one of the questions on our minds was, “Can we do a better job anticipating these kinds of uprisings?”
To see how statistical modeling might answer that question, I applied a technique called Bayesian model averaging (BMA) to data from 1972 through 2009 to develop an algorithm that could be used to estimate the probability of nonviolent uprisings in 2011 (and future years) in countries worldwide. For this analysis, I used historical data on nonviolent rebellions produced by Professors Erica Chenoweth and Maria Stephan, whose work on the subject is described in this journal article and this op-ed. In their research, a nonviolent rebellion is defined as “a campaign of purposive, nonviolent mass events in pursuit of a political objective.” As I understand their definition and data, occasional demonstrations or strikes do not qualify; for protest activity to get tagged as a nonviolent rebellion, it must be directed at central state authority, involve large numbers of participants, and be sustained. For example, I’d say that popular uprisings in Egypt, Syria, and Tunisia in 2011 would meet their definition, while smaller and more sporadic protests so far this year in Georgia and Sudan would not. My statistical analysis looked at the onsets of these protest, whether or not one or more campaign was already occurring in the same country. Chenoweth and Stephan’s data set ends in 2006; I made informed guesses to extend it through 2009, and I changed a few historical observations with which I disagreed.
I’ll provide more information about the results of the statistical analysis toward the end of this post, but, to avoid losing readers who are less interested in those technical details, let me cut to the chase: Where have popular uprisings occurred so far this year, and how good was the statistical “model” at distinguishing those countries from ones that have not seen protest campaigns yet this year?
Let’s start with the first question. I have two lists of countries that have seen onsets of nonviolent rebellion in 2011 as of early June: one I generated, and one Erica Chenoweth shared with me on request (thank you, Erica!). Those lists overlap but are not identical. To reflect that uncertainty about where these campaigns are occurring, I will discuss both versions.
- My list: Albania, Bahrain, Burkina Faso, Egypt, Syria, Tunisia, Uganda, Yemen
- Erica’s list: Bahrain, Egypt, Iran, Libya, Oman, Saudi Arabia, Syria, Yemen
Erica and I both identify eight onsets of nonviolent rebellion so far in 2011, but there are only four countries that appear on both lists: Bahrain, Egypt, Syria, and Yemen. One of the differences is just a matter of timing; I put the start of the Tunisian uprising in 2011, but Erica sees it as having started in December 2010. The other points of divergence reflect differences in judgment and, probably, available information. I identify new nonviolent uprisings this year in Albania, Burkina Faso, and Uganda, where Erica sees them in Iran, Libya, Oman, and Saudi Arabia.
Those two lists give us alternate versions of a “ground truth” for the first half of 2011 against which we can compare our statistical forecasts. So how are those forecasts faring? The quickest answer comes in the form of the area under the receiver operating characteristic curve, or AUC, a quantity statisticians often use to summarize the accuracy of a model that’s meant to classify cases on either/or outcome. The AUC represents the probability that a given classifier–here, the forecasting algorithm produced by Bayesian model averaging–will rank a randomly selected positive case (onset) higher than a randomly selected negative case (non-onset). An AUC of 0.5 is what you’d expect to get from coin-flipping. A score in the 0.70s is good; a score in the 0.80s is very good; and a score in the 0.90s is excellent. The mid-year AUC scores for the two lists (with 95% confidence intervals) are as follows.
- Out-of-sample AUC using my 2011 list: 0.75 (0.57, 0.94)
- Out-of-sample AUC using Erica’s 2011 list: o.87 (0.79, 0.95)
The samples are very small (8 or 9 events in 163 countries with populations larger than 500,000), and the year is only half over, but those preliminary results are solid out-of-sample scores for a rare-events analysis, comparable to the accuracy rates achieved by state-of-the-art efforts to predict political instability. Based on those scores, I think it’s fair to say (tentatively) that this statistical tool is doing a pretty good job so far this year at assessing risks of popular uprisings that some people have argued are impossible to forecast well.
AUC scores are a useful summary statistic, but they don’t give us very specific information about how the forecasts would perform in ways that people might actually use them. One of the challenges when using statistical analysis to forecast rare events is that the distribution of the forecasts inevitably skews toward the base rate, which is close to zero (on average, only a few onsets of nonviolent rebellion occur each year). Because of that skewing, we sometimes focus on relative rather than absolute risk when using statistical estimates to try to forecast rare events. Instead of looking at the predicted probabilities and saying we don’t think any of these events are going to happen anywhere in 2011, we start by assuming that a few of them are going happen and then using our rank-ordered list to identify which countries are the most likely candidates. One way to compare the relative risks is to group the ordered data by quantile. Here, I’m going to use quintiles (i.e., fifths), referring to the first quintile as the “most likely” group, the second quintile as the “moderately likely” group, and the bottom three quintiles as the “least likely” group. It might seem more natural to use terciles (thirds) if we’re going to break the forecasts into three groups, but I think the 1:1:3 breakdown of quintiles is more useful with rare-events forecasts because it more accurately reflects the long tail in the distribution of the predicted probabilities (i.e., the fact that most of the forecasts are approximately 0).
So, using that categorization scheme, how have my BMA-based forecasts done so far in 2011? Once again, the results look pretty good. The bar plot below shows the number of onsets so far this year that have occurred in each quintile, using the two different lists. Each quintile comprises 32 or 33 countries. Of the eight cases I have called onsets, five were in the most-likely group (top quintile), two were in the less-likely group (2nd quintile), and one (Albania) was in the least-likely group (5th quintile). Of the eight cases where Erica sees onsets so far this year, six are in the most-likely group, and the other two are in the less-likely group (2nd quintile). Based on my experience trying to forecast rare forms of political instability, I would say that those are very good out-of-sample results.
Some of you are probably wondering what other countries this statistical tool identifies as being among the most likely to see onsets of nonviolent rebellion in 2011. The following dot plot shows, in descending order, 2011 forecasts for the 40 countries with the greatest likelihood that a nonviolent uprising would begin there at some point this year, according to the algorithm I got from Bayesian model averaging.
Some of you are probably also wondering what the statistical analysis underlying these forecasts tells us about the correlates of nonviolent rebellion. What follows is a list of the variables I included in the Bayesian model averaging exercise, based on my reading of prior theory and research and the availability of reasonably reliable time-series cross-sectional data at no cost.
- Population Size. Relative to annual global median, logged.
- Poverty. Infant mortality rate relative to annual global median, logged.
- Urbanization. Percent of population living in urban areas.
- Literacy. Adult literacy rate.
- Mobile Telephony. Subscribers per 1,000 population.
- Internet Usage. Users per 1,000 population.
- Economic Growth. Year-to-year percent change in real GDP per capita.
- Civil Liberties. Freedom House’s seven-point scale.
- Democracy. Polity’s 21-point scale, rescaled to 0-10.
- Regime Stability. Years since last change of 3+ points in Polity score, logged.
- Recent Uprising. Indicator of any nonviolent rebellion in same country in previous year.
- Recent Civil War. Indicator of any violent rebellion in same country in previous year.
- Uprisings in Region. Count of countries in same region with nonviolent rebellion in previous year, logged.
- ICCPR 1st Optional Protocol. Indicator of whether or not country has signed International Covenant on Civil and Political Rights’ 1st Optional Protocol, which gives citizens the right to petition the UN for alleged violations.
- GATT/WTO Member. Indicator of whether or not country is a signatory to GATT or (after 1994) a member of the WTO.
- Post-Cold War. Indicator for post-cold war period, identified as 1989 or later.
- Colonial Legacies. Series of binary variables identifying country’s last colonizer (or lack thereof).
Those are the variables that went into the analysis, but only nine of them are actually influencing the forecasts. According to my statistical analysis, the rest are not particularly useful for predicting the onset of nonviolent rebellion. Here is a list of the nine that do, along with the posterior means and probabilities from Bayesian model averaging. [For those of you accustomed to reading results from a single regression model, the posterior mean is akin to an estimated coefficient, and the posterior probability is akin to 1 minus the p-value (so higher values indicate more confidence that the variable is a useful predictor of the outcome in question). Most of the coefficients are not on a common scale, so they shouldn’t be compared directly to each other.]
- Population Size: 0.379 (100%)
- Democracy Score: -0.033 (100%)
- Literacy: 0.012 (67%)
- Uprisings in Region: 0.132 (22%)
- Civil Liberties: -0.072 (22%)
- Post-Cold War: 0.029 (5%)
- ICCPR Ist Optional Protocol: 0.025 (4%)
- GATT/WTO Member: 0.021 (4%)
- Economic Growth: -0.003 (3%)
I will not succumb to the temptation to draw causal inferences from these associations. Even without making that heroic leap, though, we can talk about the associations we see in these results. Other things being equal, nonviolent rebellions are more likely to occur…
- In countries with the least democratic institutions;
- In countries with more expansive civil liberties;
- In countries with higher literacy rates;
- When more uprisings are already occurring in regional neighbors;
- In the post-cold war period;
- In countries that belong to the WTO;
- In countries that have signed the 1st Optional Protocol of the ICCPR: and
- When economic growth is slower.
There are also some interesting negative findings here. According to my analysis, variables that are not particularly useful for forecasting nonviolent rebellion when the measures listed above are also in the mix include:
- Poverty (as measured by infant mortality);
- Cellular phone penetration (as measured by mobile phone subscribers per 1,000 population); and
- Internet access (as measured by users per 1,000 population).
These three negative findings contradict many of the on-the-fly explanations I’ve read for the protests that are occurring this year in the Middle East and North Africa. It’s also worth pointing out that the association identified between economic growth and nonviolent uprisings is pretty tiny. It’s not quite zero, but it’s awfully close to it, and that result contradicts the prevailing belief that economic slowdowns are one of the, if not the, most important triggers to popular unrest.
On the whole, I think this exercise reaffirms the claim that we can get useful forecasts of rare forms of political instability, including popular uprisings, from statistical analysis of widely available country-level data. That doesn’t mean we can’t do even better. What’s needed to do this particular analysis better is higher-resolution data on dynamics of nonviolent rebellion. That kind of data would allow us to differentiate more subtly between situations like Egypt’s and, say, Sudan’s. Some scholars are doing excellent work right now using software to turn news reports into event data that should enable kind of analysis, but to the best of my knowledge, we’re not there yet.