In 2010 I made the prediction that the United States will experience a period of heightened social and political instability during the 2020s. Recently, several people challenged me to make this prediction more quantitative. There are all kinds of caveats, and I will get to them eventually.
But first, the TL;DR version.
David Andolfatto suggested on Twitter (@dandolfa) that I should simply use the quantitative metric that was established in my 2012 JPR article. This is a sensible suggestion because the methods are clearly described in the article and anybody will be able to check that there is no funny business once we get to comparing the predictions with data. Here’s the relevant figure from the article:
Structural-demographic theory (SDT) suggests that the violence spike of the 2020s will be worse than the one around 1970, and perhaps as bad as the last big spike during the 1920s. Thus, the expectation is that there will be more than 100 events per 5 years (see the upper panel in the figure). In terms of the second metric (the lower panel) we should expect more than 5 fatalities per 1 million of population per 5 years, if the theory is correct.
And there you have it. If violence doesn’t exceed these thresholds by 2025, then SDT is wrong.
And now for the caveats.
First, as I said on many occasions, it’s not a prophecy but a scientific prediction. So if you see me testifying before the Congress in the next few years, all bets are off. However, and realistically, political violence would probably have to reach the predicted levels before people start listening and politicians introducing reforms (if then).
Second, a critical defining feature of a political violence event, in order for it to be counted in the database, is that someone is killed (see the article why). Because modern medicine is much better at treating gunshot wounds, victims of political violence today are more likely to survive being shot, compared to 100 years ago, and that introduces a certain statistical bias.
Third, it’s important to understand that the nature of the dynamical processes generating collective violence imposes severe limits on our ability to predict them. Wars, whether between states, or internal to states, are like earthquakes. Small variations in the magnitude of the initial rupture can result in it either dissipating without much effect, or amplifying to truly catastrophic consequences. In technical language, magnitude of collective violence is governed by a fat-tailed distribution. Large-scale events (measured by the number of people killed), while relatively uncommon, are much more probable than we think. This idea has been popularized by Nassim Nicholas Taleb in The Black Swan. While I disagree with Taleb on his pessimistic view on the prospects of a science of history (not everything in social dynamics is governed by fat-tail distributions), his point is perfectly valid when we consider the possibility of predicting major ruptures, like the American Civil War, or who knows what awaits us after 2020.
The comparison between political violence and earthquakes is not a poetic one. Statistical analysis of the frequency-severity distribution showed that the number of deaths per instability event in the United States between 1780 and 2010 (again, severity is simply the number of people who are killed) is characterized by approximate scale invariance in which the frequency scales as an inverse power of the severity:
What does it mean for making predictions? In this blog post, following my 2012 article in the Journal of Peace Research, I use two metrics in quantifying political violence: the number of such events per 5-year interval, and the number of people killed in such incidents, divided by the total US population. Both measures have problems, as is discussed in the article. In particular, the second measure is quite unstable, because of its sensitivity to uncommon, but extreme events. A single large event, such as Oklahoma City bombing, can generate a spike all on its own.
This is why I use these historical data to frame my prediction about the 2020s in both metrics.
Finally, what we really need is not ability to predict the future, but ability to predict the consequences (including unintended ones) of our interventions.