Seasonality Revisited

SEASONALITY REVISITED....A few days ago I argued that much of the recent decrease in violence in Iraq was probably due to normal seasonal factors: violence tends to peak in fall and spring, and drop in winter and summer. I showed this seasonality pictorially via a chart of U.S. troop fatalities, one of the few data series that's both consistent and available for the entire course of the war.

But troop fatalities are merely a proxy for violence, and an anonymous fellow who goes by the handle Engram argues that if you look at civilian death statistics there's no seasonality at all. Thus, the recent drop in deaths is almost certainly due to the surge.

Now, Engram is pretty seriously invested in a highly distinctive view of the sources of violence in Iraq that I won't get into here (i.e., insurgency vs. terrorism vs. civil war), and is therefore also pretty invested in demonstrating that the surge is succeeding. Still, numbers are numbers. So what does he have?

The answer is on the right: a chart of civilian casualties since March 2005 that's taken from ICCC data and modified slightly to remove some artifacts. Engram then fits a 6th order polynomial through the data to demonstrate that there's no evident seasonality.

Unfortunately, there are several problems with this. The reason I didn't use this data in the first place is because it's available for only two and a half years. That's just not long enough to show seasonality, especially with a noisy data set. Even my original chart, with 4+ years of data, was barely long enough to pick out seasonal differences from the long-term noise of Iraqi violence, and the ICCC data is nearly useless in this regard. It's simply not going to show much seasonality whether it's there or not.

There are other problems too. First, there was an enormous secular increase in violence throughout 2006. This increase is so pronounced that it drowns out any local variability in the data when 2006 makes up nearly half your series.

Second, fitting a polynomial to a small data set, even a 6th order polynomial, is going to smooth the data. If there is any seasonality, it's going to make it harder to see, not easier.

In fact, if you look at the data hidden beneath Engram's curve, you can see glimmers of seasonality that the curve is hiding. The local peaks are all in spring and fall: May and September of 2005, March and November of 2006, and May of 2007. If I wanted to, I could draw a curve through the data that shows off this seasonality — and in fact I did draw just such a curve. But it's not worth putting up. It's just fighting one ad-libbed curve with another ad-libbed curve, and there's not much point in that. There's simply not enough data here to draw a conclusion one way or the other.

The seasonality of Iraqi violence has been well known and much discussed for a long time (especially the fall peak), and the 4+ years of troop fatality data bear that out. After all, why would troop fatalities show seasonal variation unless there were also some seasonal variation in the broader violence levels? In the end, I think the data suggests that (a) violence in Iraq follows a modest seasonal pattern, (b) violence rose dramatically throughout 2006 and then began to drop before the surge began, and (c) fatality levels this summer were down only mildly compared to earlier in the year — and probably not down at all when you take seasonality into account. Bottom line: Violence is as bad as it's ever been, sectarian cleansing is proeeding at a murderous pace, Kirkuk and Basra are both timebombs, refugees are fleeing the country at staggering rates, the Iraqi infrastructure is in ruins, and the surge just doesn't seem to be having much of an effect outside of a small number of handpicked neighborhoods in Baghdad. What's more, it's having no effect on the political situation, which continues to be a sectarian disaster.

I'll end with two related points. First, no matter how many different ways you look at the civilian casualty numbers, what you're going to see is a huge increase in violence during 2006 followed by a slow decrease. That's just all there is. If you want to argue that the surge is responsible, you can do that, but since the decrease began in December of last year that requires some pretty creative torturing of the data. Or maybe the decrease is due to seasonality. Or the Anbar Awakening. Or something else. But no matter what you do, you've basically got one observation (violence rose through November 2006 and then started to decrease) and that's it.

Second, and ironically, the ICCC data gives the lie to the primary pro-surge talking point these days: namely that violence is down 75% from its peak. It's actually down by only a fraction of that amount, and the data is even more discouraging if you pick a more appropriate starting point. In February, just before the surge began, there were 1,529 civilian fatalities. In the past couple of months we've averaged....1,528 fatalities per month.

And what about the latest talking point, namely that what we really ought to be paying attention to is the single month of August, which was awesome? Aside from the fact that this doesn't really seem to be true (and the military data to the contrary seems to be very heavily massaged), you're in pretty desperate straits when you have to hang your entire case for continuing a failed policy on equivocal data from a single month. It's no way to run a war.

POSTSCRIPT: I know, I know, I know. I keep saying I'm not going to obsess over casualty stats anymore. I feel like Michael Corleone in Godfather III. What can I say?