Saturday, 26 April 2003

More John Lott

Glenn Reynolds today links to three articles in the Stanford Law Review that argue over John Lott’s More Guns, Less Crime thesis. Normally in these disputes it’s hard to say much of anything without having the data at hand; thankfully, Ian Ayres and John J. Donohue III have posted their data (warning—ZIP file) at Ayres’ website, so hopefully someone—who has more time on their hands than me and isn’t supposed to be writing a dissertation on a completely different topic at the moment, mind you—can make heads or tails of what’s going on.

My preliminary assessment (as a political scientist who plays an econometrician on TV and spends most of his time running limited dependent variable models): without knowing any of the authors’ statistical training, I’d be very reluctant to draw any conclusions from their writing alone, but Ayres and Donahue appear to be onto something. However, analyzing statistical models with fixed effects can be nasty business, particularly since theoretically the asymptotics that regression analysis relies on aren’t fulfilled (the number of independent variables in a fixed effects model increases as a function of the number of observations, rather than being constant) and throwing lots of atheoretical dummy variables into a model runs the risk of soaking up variance that really ought to be attributed to a substantive effect—but that applies equally to both sides in this debate.

I’m not quite sure why Ayres and Donahue use areg instead of xtreg in Stata to estimate fixed effects, but it shouldn’t make a substantive difference (it just makes the specification a bit harder); more generally, I’d be more comfortable if everyone involved used some sort of vaguely modern time-series analysis (ARIMA, VAR, cointegration—even Box-Jenkins!), but maybe I’m just weird that way. I assume the dependent variables are logged since economists log everything for some odd reason (perhaps just to make the regressions harder to interpret). It’s not entirely clear how any of the authors treat missing data.

Anyway, someone else will have to take over from here… this is far too much thought for a debate I have absolutely nothing invested in.

Julian Sanchez writes on this issue as well; I agree that this problem with his research (if borne out by the evidence) bothers me much more than the basically silly “Mary Rosh” business or other complaints (some of which seem to be based more on his political views on issues other than gun control).

Also, to clarify for the hardcore econometricians in the audience, by “time-series” I meant cross-sectional time-series or panel analysis. (Political scientists don’t get particularly worked up about the CSTS versus panel distinction in general, mainly because we tend to deal with much more of the latter; this data is closer to CSTS but has properties of both—large number of units, but also a fairly large number of time-points.)