Iraq, the Surge, Regression to the Mean?

My int’l security class is discussing the causes of civil war this week, and one of the main case studies we’re looking at is Iraq. It’s well known that violence in Iraq has ebbed quite substantially since about 2007. If you ask most people why, the answer you’ll get is simple: the surge. George Bush’s strategy in 2007 worked. The strategy was basically to hire Robert Gates and David Petraus, and increase the troop levels in Baghdad and Anbar province by 20,000. Directly following the war there was a massive decline in violence which has continued in the post surge years. Not surprisingly, neoconservatives and Iraqi war supporters take this as proof that the surge worked, and that the U.S. has had great success in Iraq.

But can we necessarily infer cause and effect from this? I’ve long been convinced by Peter Galbraith’s claim that violence declined in Iraq for reasons that had really nothing to do with the surge: (a) Iraq’s Sunni/Shia ghettos which were the scene of so much violence had become ‘unmixed’ (b) Moqtad’ Al Sadr disarmed his forces for purely strategic reasons; the Sunni Ba’athists turned against their former Al Qaeda allies, who had grown to strong, and this led to a decrease in sectarian violence. In short, violence declined for other more particular and local reasons that had nothing to do with the surge.

In short, civil war ebbs and flows with a particular logic. We would not expect high levels of violence to be sustained for long periods of time, but a lull and violence is not an end of the war either. Having read Thinking: Fast, Slow by Daniel Kahneman it seems to me that this surge argument might fall victim to the regression fallacy.

Let me explain:

Regression to the mean is a statistical error caused by the reliance only of extreme observations. For example, consider the effect of acupuncture on migraine headaches. We’d like to know if it has an effect, and we’d like to know about the average. But many people, especially if they are self-reporting, misjudge cause and effect because of selection bias. For instance, migraine headache sufferers are most likely to seek out alternative medicine when pain is at its worst. The pain ebbs and flows: in other words, there is a random statistical distribution and there is a statistical “average level”. Overtime, we would expect extremes to return to the average. The unfortunate migraine suffer observes relief following treatment, but this is only because of selection bias, which they take as evidence of the effectiveness of the intervention. As interventions tend to occur when the patient is most desperate for relief, the mistake is compounded. As Kahneman says, as humans we like to tell stories, and we’re bad at statistics. Statistical reasoning would tell us this is just simple regression, but the causal story telling part of our brain, which usually dominates, likes to think that the intervention worked.

The variable “level of violence” also has a statistical distribution and an average. The intervention, namely the surge, occurred (not by accident) at the extreme point. There was a subsequent return to the mean, but this does not necessarily mean cause and effect. Thinking about causality in this instance is important, because the experience of the surge is likely to shape future nation-building efforts, affect thousands of lives, and eat up tons and tons of resources.