Do Soldiers Returning From Combat Increase the Murder Rate – Results of a Quick and Dirty Regression, or Modeling on the Fly

Here’s Megan McArdle:

In the comments to the previous post, Brooksfoe, who I respect a great deal, points out that the narrative “Soldiers exposed to combat come home to kill” does not flunk the gut check test.

No, indeed, it doesn’t. I can build a quite plausible story where combat makes people into crazy killers.

The problem is, it’s just a story. History is full of those stories that turned out not to be true. As my commenter points out, I can also build a plausible story where combat makes you realize the sanctity of human life and makes you less likely to kill. Or where the amazing human capacity for compartmentalization makes it have no overall effect. A look at one piece of gross evidence–the massive return of combat troops post World War II does not seem to show evidence of a killing spree; homicide rates fall during the war (not surprising; we shipped our prime homicide age overseas by the millions, plus national emergencies tend to surpress both crime and suicide); rates return to their 1940 level in 1946, then fall rapidly, which is not what I would expect if combat is really so brutalizing. But lots of things changed, yadda yadda. The point is, there is usually more than one plausible story. To check whether your story is true, you need data.

She’s right… you need data. But it helps to have some idea how to look at data when you have it. See, this story seems pretty plausible to me. The drop-off after 1946 she finds so puzzling to me indicates not a problem, but adds to the plausibility of the story. After all, combat is traumatic, but I would imagine, like most traumas, for most people it fades with time. It may never go away, but for most people the pressing effect of it is going to be felt most deeply in the year or so after combat ends, not thirty years later.

Now, I like data, and I like testing things, so I figured… let’s see what I could do in a few minutes to test both whether demobilized troops returning from combat can raise the murder rate, and my little hypothesis that this is likely a short term (one year or so) effect. So I decided to check out the data to which she links… and which, when you follow the links, comes originally from this table at the Justice Dep’t.

Well, if we have data, we can build a model. I’m going to build a very simple model, looking at how the change in the murder rate from one year to the next is affected by troops coming home, and the population in the prime homicide age (which I assume is males 18 – 25). The data on population through 1979 comes from here and the data on population after that comes from here. (This is data I pulled a couple of months ago… the second link doesn’t seem to be active right now. Sorry about that. Makes me really hope I didn’t screw up the data since it makes hard to check it!) Since the pre-1979 is such a pain in the butt to put together, I’m only going to go back to 1940. (I’m curious, but I’m not that curious, and this is just a hobby and I only have a bit of time. Someone give me a grant or some money and I’ll do something a lot more rigorous.) I also construct a dummy variable that is equal to zero in most years, and is equal to 1 in the year of a demobilization of combat troops, in the year after a demobilization of combat troops, and in the year after that.

So here are the years for which the dummy is not equal to zero:
– I assumed that for WW2, there wasn’t much of a demobilization until 1945… so the dummy is equal to 1 for 1945 through 1946.
– Similarly, I assumed three wasn’t much of a demobilization for the Korean War until 1953, so the dummy is equal to 1 for 1953 through 1954.
– For Vietnam, it seems combat troops were in and out, usually after a year of combat… I assumed there wasn’t much of an American presence there until 1964 – so there wouldn’t be any major demobilizations until 1965. I also assumed most US combat troops were out by 1973… so the dummy is equal to 1 from 1964 to 1974.
– For Gulf War 1 – the War ended in 1991, and the dummy was set to 1 for 1991 – 1992.
– The Justice Dep’t data to which McArdle indirectly linked ends in 2002… I assume there were some troops returning from Afghanistan then, so 2002 is set to 1.
– Other wars… Grenada, Panama, Kosovo, etc., I figured were too small for many soldiers to have suffered the effects of shell-shock or other events that lead to excessive trauma.

The dummy is quick and dirty… if I had the time, maybe I’d try to figure out how many troops were coming home (or what percent of the population it amounted to) and multiply that by the dummy. But like I said, this is just a hobby and I’m doing this at speed.

A quick note to other data jockeys reading this… the correlation between the change in the murder rate and its lag is about 0.37. I’d prefer lower, but its low enough that I’m willing to assume not enough autocorrelation to be a problem in a very quick and extremely dirty analysis. Differencing the dependent variable means I don’t think I don’t have to worry about unit root issues.

Anyway, I’ve put up a file that has the data, and the regression results here. (Congrats to me for finally figuring out how to use Google Docs!!!)

A summary…

1. The model has an adjusted R2 is only 0.11. Not great, but bear in mind… its explaining the variation in the change in the murder rate, not the murder rate itself, which is much harder to do. So for its purposes, and given its a quick and dirty model, I’m not complaining too much. Sure, there’s a lot of room for improvement. So what?
2. Males 18 – 24 as a % of the population has a positive sign… but is not significant – the p-value is .36
3. The demobilization dummy is positive… and has a P-value of 0.004. Put another way, the model does seem to indicate that within 1 years of soldiers coming home from combat, there is an uptick in the murder rate, and its significant, even at an alpha of 1%. (Back to the numbers nerds for a moment – even my assumption of autocorrelation was generous, autocorrelation will raise the t Stat, but I wouldn’t be on it taking a variable that isn’t significant and giving it this kind of a P-value if there isn’t something to it given how long the t to t+1 correlation was.)

This is not to speak ill of soldiers. Most of them do not come home and commit murder. And those that do… are probably not entirely themselves. Combat does things to people. So any failure these numbers might suggest belong exclusively to three parties:

1. The DoD for not doing enough to help these people who gave so much
2. The politicians for not doing something about it
3. Us for not making our politicians do something about. Slapping a yellow ribbon magnet on your car doesn’t do diddly, quite frankly.

Anyway, that’s my quick and dirty model. Have at it.