The Corner

The one and only.

‘78.4 Percent’


That’s what Nate Silver’s model says are the odds that President Obama will win the Electoral College (as of 1:55:05 P.M. EST.) Not 78.5 percent. Not 78.39 percent. Not 80 percent or 60 percent. 78.4 percent.

However you come down on the soundness of Silver’s model. — FWIW though I’m a layman, I was a fan of Silver the baseball statistician, and think his model, from what he’s written about it publicly, is plausibly built, even if I suspect that he’s following in sabermetrics godfather Bill James’ footsteps by assigning weights to inputs based on what produces outcomes that feel right — Where was I? Oh yes, however you come down on the soundness of Silver’s model, I’m irked by how both supporters and detractors seem to be interpreting it.

Let’s be clear. When Nate Silver says there’s a 78.4 percent chance Obama will win the Electoral College, he’s expressing a degree of confidence in a prediction. He’s not inventing a new verb and saying “President Obama will be seventy-eight-point-four-reelected” or creating a hundred parallel universes wherein “78.4 percent of Barack Obamas will be reelected.” He’s making a prediction about a binary outcome, based on his reading of the data, and then stapling on a measure of confidence to that prediction.

So one can think that Silver is probably right about the Electoral College, and simultaneously think that the 78.4 percent number is basically meaningless. Or rather, that it is impossible to formulate the epistemological difference between “There is a 78.4 percent chance that Obama will win” and “There is a pretty good chance Obama will win.” A lot of Silver’s credibility counts on somehow elucidating the distance between those two propositions, in somehow explaining how his model adds value to what amounts to an expression of the conventional wisdom. As I understand it, Silver runs a very large number of simulations and the 78.4 percent number reports the distribution of outcomes in those simulations. If this is a way of saying “Obama has more paths to victory than Romney”, that’s, again, probably true, but it’s also the same thing that every two-bit pundit with a humanities degree (myself included) has been saying for a while.

UPDATE: Generalizing hugely, Silver’s electoral model works like PECOTA — the career-modeling algorithm developed for Baseball Prospectus. It incorporates past election outcomes, and most crucially the correlation between actual outcomes and what the polls predicted. The issue there, for me, is that PECOTA had a pretty good N of playeer careers to work with. If you want to know what kind of year Robinson Cano is likely to have in 2013, PECOTA has a lot of actual prior career trajectories, weighted for era, home ballpark, and the like, to compare with Cano’s, and the resultant predictions, and especially “similarity scores” are intuitive — they usually “make sense” to baseball fans. But as one of the commenters pointed out, presidential electoral dynamics (and demographics) are fluid enough and quality, extensive polling so recent a phenomenon, that the number of comps is maybe four or five — or maybe zero.

Anyway, that’s spitballing. I hope no one takes me to think myself qualified to appraise the validity of Silver’s model. What the original post was about is how we should characterize the relationship between Silver’s technical work and the vaguer, layman-speak we use to talk about the election.