Super Tuesday Results thread

[Measure for Measure]

Originally Posted by Martin Hyde

What I object to is your analysis that because of some number from intrade you have factual support for the claim that 2012 was a fine year to run for President as a Republican and there was no reason for any prominent Republicans not to throw their hat into the ring. I'm not against talking about intrade, but you cannot use it that way. There is more to being a professional politician than looking at polls 18 months out and deciding if the President is vulnerable or not. A lot of people won't even consider a run against a sitting President, regardless of the polling numbers or how weak he might seem. Incumbency is incredibly powerful in our system, and your reliance on intrade makes you blind to the fact that some politicians will simply choose not to participate in a fight against an incumbent when they have the option to marshal their resources and wait four years and get another run at it.

I guess I have 2 comments:
1. The odds of a Romney November victory are somewhere between 30 and 60 percent. Anywhere in that span are not bad odds. If you don't toss your hat in the ring your chances are zero. I use intrade as a rough way of placing current odds: again, anybody who thinks Romney's chances are 5-10% can make a killing at Intrade. They only need to put their money where their mouth is.

Less mathematically, it's a long time till November. GWBush's odds looked excellent in Jan 1991. In Nov 1992, not so much.

2. Incumbency is worth about 2.75 points in the popular vote. That's a lot. But it's not insurmountable as Clinton and Reagan can attest. Especially when there's a decent chance of economic collapse during the election year.

I have absolutely no idea how you are addressing the actual part of my post you quoted when you driveled off with these numbers. You need to do a better job of linking numbers you pull from election forecaster's models and the actual things you are saying. I didn't make any claims about Obama's performance, for one. You were quoting me saying McCain wasn't a dreg of a candidate and respond with something that doesn't really address that at all.

Martin. We're here to fight ignorance. That means if you argue "A", I'm not obliged to argue "Not A". You said McCain wasn't a dreg. I agree: if he was a dreg (or if Obama's oratory was electorally hypnotic as some conservatives like to claim) Obama would beat the Fair model by something like 4 points, rather than a merely respectable 1.5. The model only tells you relative performances in various matchups though. My take is that Obama is a solid but not particularly strong campaigner and that choosing Palin should have disqualified McCain electorally. It did not.

Dukkakis: now that man was a dreg.

From what I've seen election forecasting is mostly a sham, and it's obvious you think it is really cool and interesting but actually looking at the history of it, it is simply weak and uninteresting to me.

For one, Ray Fair has been working on his model since 1976. With any of these models, the people making it have an immediate advantage on all of their retrocasts because they develop a model that can predict the very outcomes they use to develop their model in the first place. (That's perhaps strange to parse linguistically, but essentially the retrocasting isn't impressive at all.)

It's impressive when you modify your variables (modestly) once in a 25 year period. That helps combat concerns with data-dredging. You are correct to distinguish between within sample and out of sample forecasts. I wasn't aware that they are all that different though, and I can't locate the 1992 and 1996 out of sample forecasts. I can say though that 1992 was a bad year, due to Perot's intervention. Third party runs mess up the model because we don't have enough data on such experiences.

1. The models are tail loaded to place more importance on data later in the election cycle. This makes sense, but it also means it's a much less impressive feat to say they successfully projected an election based on October data than on June data. Since they can revise their predictions throughout the cycle and their models place more important on tail end data this decreases substantially my interest in these models anytime prior to September or October.

Hm. Well, Nate Silver certainly doesn't take that approach. And Fair's point is primarily about noting that the economy matters much more than other factors which a babbled about by our nation's pundits. If you want to predict electoral outcomes, I understand that approval ratings are an invaluable component, which Fair doesn't use for endogeneity reasons. [Elaboration on request.]

2. Most of them actually produce a series of results out of a few different versions of their models. Some of the forecasters of the 1996 election then significantly publicized the results that very closely mirrored the results, and basically de-emphasized all the other iterations of their model that missed the mark.

3. Most of the models have "special exceptions" or "special factors" added in to elections that they can't really explain. It's been a long time since I've paid attention to election forecasting, but the last time I read about Fair he had a special variable for elections where he felt war "improperly affected the results" (1920, 1944, 1948.) However he doesn't use this variable in many other war elections, so it almost feels like these econometric forecasters simply throw special variables and such in whenever it fixes the predictive power of their model. That's fine, but it also makes me trust them far less when it comes to being able to provide meaningful basis for discussing an election months before the fact. In one of his models Thomas Holbrook added a special variable for "extremism" to explain why Barry Goldwater and George McGovern lost in landslides that his model failed to accurately explain.

So what it really comes down to is you have a bunch of college professors that release a bunch of predictions over and over, most of them placing increased weight to more recent data (which means the models tend to be more accurate in September and October because the professors know studies have shown late polling data is more meaningful than early polling data.) The predictive ability isn't impressive much in such a scenario because it is like me predicting who will win the AL East after 145 games have been played and the team in first place is up 5 games. That's a much easier prediction to make than one made in June, and the value of a model that could predict the outcome of the AL East race in June would be immensely higher than one that only really gets very accurate around August.

Additionally, the college professors follow the practice of cold readers, in that they actually release tons of predictions and then highlight the successful ones. That's precisely what happened in 1996 (the first election I really knew about election forecasters), guys like Holbrook and Wlezien released tons of predictions some which showed Clinton winning close and some showing him winning big, then they emphasized their prediction that showed him winning big and basically ignored the slight variations in their model that showed other things.

Numbers 2 and 3 don't apply to the Fair model.

As I think I've mentioned in another thread, even if we take the models at face value all they do is use econometrics to get "kinda close." In a Presidential election things are often decided by very small overall percentage differences in the popular vote, so the value in being able to predict the vote within a few percentage points isn't that exciting to me at all because depending on how things shift in certain states that wobble room can be the all the difference between who is President or not and whether the election is considered a landslide or not in the EC.

The average spread in the popular vote is 5.6 points. The median spread is 4.9 points. 1916-2008. Admittedly, the spreads have declined in recent years: the median spread from 1972-now was 3.7. That's not small: a typical error in the model of 2.1 or less won't make a difference usually. A bigger error might.