Friday, August 5, 2011

Data Revisions and the Guns and Butter Model

Karl Smith has a request:
Has anyone run the guns a butter model on the last election with the new disposal income data? Supposedly the Dems lost an extra 20 seats in the House or so above what could be explained structurally. However, now that we know the structure of the economy was worse than the frontline data does that estimate still hold?

I haven’t run the numbers but I am guessing what looked to be policy backlash will vanish in the structural void with new estimates.
I decided to check this out. First, I went with Douglas Hibbs site, the original source of the "Guns and Butter" model. He had predicted that the Dems would win 211 seats, roughly an overestimate of 20 seats (relative to Democrat actual wins of 193). However, this estimate came with caveats:
In fact there is uncertainty about income growth during last quarter - the 2010q2. The personal income data for q2 posted by the Commerce Department's Bureau of Economic Analysis on 30 August 2010 are second estimates and they are subject to potentially large revisions later.
Of course, this is exactly what happened.

I decided to download his data and update the consumption statistics. I took the latest disposable per capita income, which have been revised going back to 2008Q1. I couldn't get the CPI data to match exactly, but this page seems to do well enough to deflate the numbers.

At this point, I decided to run the full model with the 2010 election as an additional data point. This is the relevant point to use in evaluating all of the data to use for future modeling; but it may overstate the fit exactly for 2010 slightly. Here's what I have:

The new prediction for 2010 is 202.5 seats. This overstates Dem gains about about 10 seats, but does cut the overstatement.

I wondered how much adding 2010 did on its own, so next I threw out the 2010 data, and fit the 2010 election based on previous election data, but current economic data: now, I get a Democrat prediction of 206 seats. Roughly, a fourth of the Democrat "underperformance" can be accounted for by economic conditions that were worse than thought at the time.

I'm not sold on this model -- with so few elections to go through, it seems likely that many elections will be "anomalies" ex ante and then rationalized ex post through the model (you can sort of see that here -- throwing in the new data lowers the rate of misfit for 2010). Plus there are the various structural reasons to mistrust any model like this - Andrew Gelman offers some comments here, and then there is the Lucas Critique.

But if you're looking for ways in which worse economic data should change your priors, here's one of them -- the Democrats faced a worse economic climate in 2010 than commonly realized, and their performance is more understandable as a result.

No comments: