A person with a Master's degree in philosophy and an associate professor of biology please themselves in thinking they have refuted Ray Fair's econometric model of presidential elections.

The philosopher assumes what is to be proven, that public opinion polls are accurate.

And the biologist proudly administers the coup de grace: "Correlation is not causation."

I'm sure Professor Fair will be resigning in disgrace any minute now.

(By the way, all the regression-haters should check out this forecast of how many medals each country will win in the Olympics. Even if a biologist needed to warn a world-renowned econometrician about correlation not being causation, this simple regression model for medals might do pretty well.)

Comments

[Patrick R. Sullivan]

Oh, c'mon, we all remember President Dukakis who had an 18 point lead in the polls over George H W Bush in 1988.

Actually, Fair's model makes a lot of sense--and I'd say Hillary Clinton believes it since she didn't bother to run for the Dem nomination. The only thing you can criticize Fair for is not having enough data points; there have been 22 elections in the period he covered. I'm told by a friend that econometricians usually want a minimum of 30.

[arthegall]

"Correlation is not causation" is a coup de grace, precisely because of predictive models that come with statements like this:

"Bernard and Busse show that over the last 40 years, national Olympic medal totals have been driven by four distinct factors: population, per capita income, past performance, and a host effect."

Hmmm... "driven?" Try, "share a common cause with."

This is exactly the point: you can make all the predictive models you want, but ultimately you can't rule out hidden causes, commone effects, etc. Now maybe Fair's model isn't so susceptible to this; perhaps it's hard to think of what could be a hidden cause for both positive economic performance and winning an election. But Bernard and Busse, or at least the people who write their copy, should be more careful.

The problems with Fair's model are different anyway: a small number of datapoints, a (relatively) large number of variables that were selected post-hoc for their performance (and without discussion of the complexity of the total model set from which they were selected) and the updating of parameters to re-fit the model after every election.

I mean, his results are fine... probably quite adequate for the (limited) use of examining voter behavior. It's a little more problematic if you're trying to fit his model into the mold of "miraculous predictor of election results."