Jowei Chen sent along this paper:

In the aftermath of the summer 2004 Florida hurricane season, the Federal Emergency Management Agency (FEMA) distributed $1.2 billion in disaster aid to Florida residents. This research presents two empirical findings that collectively suggest the Bush administration engaged in vote buying behavior. First, by tracking the geographic location of each aid recipient, the data reveal that FEMA treated applicants from Republican neighborhoods much more favorably than those from Democratic or moderate neighborhoods, even conditioning on hurricane severity, home value, and demographic factors. Second, I compare precinct-level vote counts from the post-hurricane (November 2004) and pre-hurricane (November 2002) elections to measure the effect of FEMA aid on Bush’s vote share. Using a two-stage least squares estimator, this analysis reveals that core Republican voters are easily swayed by FEMA aid – $16,800 buys one additional vote for Bush – while Democrats and moderates are not. Collectively, these results suggest the Bush administration maximized its 2004 vote share by concentrating FEMA disaster aid among core Republicans.

This is interesting. In many aspects of politics, it seems clear that politicians reward their supporters, but political scientists sometimes really resist this idea, arguing on logical grounds that candidates should be focusing their efforts on the median voter. It’s interesting to see some clear evidence where supporters are getting rewarded–and it’s also good to see someone getting down and dirty with the data, rather than just reanalyzing the same old datasets over and over (which is what I usually do…). Sure, it’s ultimately an n=1 study, but I imagine it will add something useful to the literature on government spending

Also, my little thoughts:

1. Chen has a good description on page 2 of why the term “vote buying” might be appropriate here, but I don’t think the word “bribery” is appropriate. Giving someone federal aid to motivate them to vote for you could be called “vote buying” but I don’t see how it’s a “bribe” in the usual sense of the word.

2. In the abstract, Chen writes, “core Republican voters are easily swayed by FEMA aid –
$16,800 buys one additional vote for Bush.” Is that really “easily swayed”? $17,000 is a lot of money, no? To sway a million votes would take $17 billion, which can’t quite be buried in the federal budget. I’ve heard it said that in typical election campaigns, it costs something like $40 to change a vote.

3. The usual comments about rounding, tabular displays, etc. In the abstract, $16,800 should be $17,000 (or even $20,000). The sort of precision implied by “$16,800” just isn’t there, and can really never be there, given that conditions are always changing. And then on page 4 it says “$15,989”! I mean, really! Why not give the cents, too?

The tables should be graphs, and also the predictors should be rescaled so you’re not in the awkward position of having to interpret a coefficient of 0.054 for wind speed in miles per hour. (One more mile per hour corresponds to a change of 0.054 . . . hmmm, what’s that again?) Table 2 has meaningless numbers like 107.16 and house values to the nearest dollar . . . (Yeah, yeah, I’ve been an offender too; see Table 1 here. But that won’t stop me from trying to get others to clean up their acts.) Other tables have the no-no of including interactions without first centering the predictors (see Chapter 3 for discussion of that point). Finally, Figure 1 has some nice features, but that business of adding 1 so you can take the log . . . that’s ugly, man. I mean, why add $1, why not $100 or $1000 or whatever (maybe that’s what was actually done). Also better to use log10 or, better still, to display dollar amounts.

4. Some of the labeling is confusing. Models (1), (2), (3), (4) on page 24 don’t seem to be the same as models (1), (2), (3), (4) on page 27. This is a problem for readers like me who like to jump to the results right away when reading a paper.

5. I’m a little worried by the analysis associated with Figure 1. You shouldn’t be taking 2004 vote minus 2002 vote; you should be regressing 2004 on 2002 [typo fixed] and looking at the residuals. Otherwise you can get the usual regression-to-the-mean artifacts. It’s also sort of weird that, in the regressions, some variables are logged (household income) but others aren’t (for example, house value). Probably no big deal but it looks funny somehow.

6. I noticed that a citation to a paper by someone named Sam Houston. Well, I suppose that sort of thing has to happen sometime.

7. On the bottom of page 1 there’s a copright notice. I don’t recall seeing that sort of thing before on a working paper. Is this a new trend?

P.S. John Sides has some comments here.