Friday, August 20, 2004

Complicate until reaching desired conclusion 

I've done a few posts on how models of economics predicts a rather sure electoral victory for President Bush, and one of the models most people point to is that of Ray Fair, a professor at Yale. I've met Prof. Fair a few times; one of his students was one of my professors long ago, and since I wrote my dissertation on political business cycles I had to read many of Fair's papers. To have seen a presentation by Fair is to have seen someone use a rather old-fashioned form of econometric analysis (anachronistic to many younger macroeconomists) in a very methodical fashion to whatever solution it takes him. If the method is executed correctly and if the answers given are not absurd, he takes them as valid. As he should.

Unless, that is, if you are a New York Times interviewer.

As a professor of economics at Yale, you are known for creating an econometric equation that has predicted presidential elections with relative accuracy.

My latest prediction shows that Bush will receive 57.5 percent of the two-party votes.

The polls are suggesting a much closer race.

Polls are notoriously flaky this far ahead of the election, and there is a limit to how much you want to trust polls.

Why should we trust your equation, which seems unusually reductive?

It has done well historically. The average mistake of the equation is about 2.5 percentage points.

In your book ''Predicting Presidential Elections and Other Things,'' you claim that economic growth and inflation are the only variables that matter in a presidential race. Are you saying that the war in Iraq will have no influence on the election?

Historically, issues like war haven't swamped the economics. If the equation is correctly specified, then the chances that Bush loses are very small.

But the country hasn't been this polarized since the 60's, and voters seem genuinely engaged by social issues like gay marriage and the overall question of a more just society.

We throw all those into what we call the error term. In the past, all that stuff that you think should count averages about 2.5 percent, and that is pretty small.

It saddens me that you teach this to students at Yale, who
could be thinking about society in complex and meaningful ways.
(Emphasis mine.)

Eugene Volokh has the rest. T-jic puts it well:

So this guy does research, and says "there are two or three main variables that predict phenomena X", and someone - effectively speaking in the editorial voice of the NYT - is "saddened" that he'd dare teach this to students who should be thinking about society in "complex and meaningful" ways.

This is an amazing clash of rationalism, on the one hand, with irrational totem-worshipping, on the other. The NYT writer is effectively saying "your 'facts' may prove one thing, but I think that the world is complex, and therefore needs complex models, and any model that merely explains actual phenomena, but does not reflect my world view is illegitimate, and it saddens me that you would explain this predictive tool to others, instead of ideological models that I like but aren't as efficient".


Yeah, and then the reporter doesn't even have the decency to follow up with Fair's own note to try to explain this in a manner that is very much Fair -- he offers up every reason his model could be wrong, and arrives at the following.
If you experiment on the site with alternative vote predictions, you will see that no realistic economic values can bring the predicted vote share to even
about 53 percent. (Remember that there is only one quarter, 2004:3, for which actual economic data are not available.) This means, given the standard error of 2.4, that if the equation is correctly specified, the probability that Bush loses is very small. The bottom line is that the equation has to be misspecified in order for Bush to lose. And this is where the pitfalls come in. Regression analysis can only take us so far; possible pitfalls are always lurking.
So yes, he could be wrong, but that would mean that all the other correct predictions of the model were just luck, something which is possible but again with low probability. That the model predicts victory for the candidate he does NOT prefer matters little to Professor Fair.
I am not attempting to be an advocate for one party or another. I am attempting to be a social scientist trying to explain voting behavior.
And to this, the interviewer is utterly clueless.
But in the process you are shaping opinion. Predictions can be self-confirming, because wishy-washy voters might go with the candidate who is perceived to be more successful. ...