There are two ways to forecast the election: polls and models. Polls are easy: go out and ask who will vote for whom, tally the results, and print ‘em. As long as the poll sample “looks like” the eventual voters, the poll will be somewhat accurate.
Models are just the same as any statistical model. They take data as input and spit out probability predictions. Bayesian models, just because they are Bayesian (which is of course not just the superior statistical philosophy, but the correct one) are not necessarily better than any classical model. Bayesian models can stink just as badly as classical ones, or worse because their users tends to be cockier (like Yours Truly).
Polls are inherently numerical; that is, a set of numbers is the result. 50% for Romney, 49% for Obama, as Gallup’s final has it, for example. Models are usually numerical, but needn’t be (mine in not). Polls are typically released to the nearest percentage. Modelers pretend that they have more insight and show us more digits. Nate Silver puts Mr Obama’s chances (as of today) at 91.6%. That’s point-six and not point-five.
Because our task will be to assess the goodness or badness of polls and models, let’s today document them. Corrections are welcome. I’ve tried to discover all the most relevant information but there are lacunae. If you can document these blanks, or suggest other sources that should be added, please do so in the comment boxes and I’ll put them in the main text. Be sure to include links.
|Source||Obama Projected||Romney Projected||Democrat Sample||Republican Sample||Independent Sample||Margin of Error|
|Pew; Oct 31-Nov 3||50||47||39||32||29||2|
|NBC/WSJ; Nov. 5||48||47||43||41||15||2.6|
|ABC; Oct. 31-Nov. 3||49||48||33||29||34||2.5|
|CNN; Nov. 2-4,||49||49||41||30||29||3|
|Gallup-USA Today Swing States; Oct 22-28, 27-31||48||48||?||?||?||3|
|Rasmussen; Nov 1-4||48||49||39||37||24||2.5|
|Gallup final Oct 31-Nov 3||49||50||?||?||?||2|
I’ve only found a few prominent academics and journalistic celebrities who are well known enough to have published predictions. Not to over-work an exhausted phrase, but garbage in, garbage out; and this is true no matter how sophisticated the apparatus or well credentialed the pundit (which could include me!). The top two modelers were so confident that they did not include explicit probabilities.
|Drew Linzer, Votamatic (R charts!); Nov 5||~100?|
|John Sides, The Monkey Cage; Nov 5.||~100?|
|Wang; Princeton Election Consortium; Nov 5.||99.8|
|Darryl Holman, Horses Asses; Nov 5.||98.8|
|Nate Silver; Nov 5.||91.6|
|Simon Jackman; Pollster; Nov 5.||> 50|
|Kenneth Bickers, Michael Berry, CU; article; Oct 4.||23|
|Briggs; Nov 5.||< 20 (less than twenty)|
Notes: Most models take as input the poll data, but they also use other data, like measures of the state of the economy, the year, etc., etc., etc. “Monte Carlo” doesn’t mean spit. It’s just one of many techniques with which to perform numerical integration. See also this journal with folks playing around. The WSJ looks at historical poll accuracy.
Update Clarified my prediction; less than 1 in 5 chance, but that’s as close to a number as I care to go.
Update 11:17 pm. So much for my model, which took too much account of poll error and supposed too much “Bradley effect.” Ah well. More: I had Obama down, and he was by 10 some million from 2008. But I had Romney up by about 3 million, while he turned out down about the same from McCain in 2008. Those Republicans who did vote were more “enthusiastic”, but what a dichotomy.
Update See this about the missing voters. “As of this writing, Barack Obama has received a bit more than 60 million votes. Mitt Romney has received 57 million votes. Although the gap between Republicans and Democrats has closed considerably since 2008, Romney is still running about 2.5 million votes behind John McCain; the gap has closed simply because Obama is running about 9 million votes behind his 2008 totals.” That difference was my blunder.
Update Well Real Climate folks! This is the space to discuss when to admit failure and abandon models which do not make skillful predictions. Like (sadly) my model. And climate models.