YouGov's John Sides pours over the numbers and says he doesn't trust those voters who say they still haven't made up their minds. He believes those "undecided" voters will end up making a choice that reflects their pre-existing biases.
His model predicts those undecideds will break almost down the middle, meaning they will likely be inconsequential in this election.
I will assume that undecided voters will make a decision that reflects three things: their party identification, their approval of Obama (as a “referendum” model would suggest), and how favorably they feel toward Obama and Romney (the difference between how they feel about each candidate, as a “choice” model would suggest). I estimate this model on all decided voters and then predict the choices of the undecided voters based on the model’s results. The model predicts the choices of decided voters correctly 99% of the time, which is no surprise given the factors in the model.
Here’s the first finding. The model predicts that these undecided voters will split almost exactly evenly: 50.1% for Obama and 49.9% for Romney. There is substantial uncertainty in this estimate, naturally. The 95% confidence interval for Obama’s predicted vote share is 44% to 56%.
If we focus on those undecided voters who either have voted or say they “definitely” will vote—this is about 58% of undecided voters— the balance tilts more toward Romney: 44% Obama vs. 56% Romney. As in polls more generally, Romney tends to do a bit better among self-described likely voters than among voters as a whole. (Again, insert caveat about uncertainty.)
However, consider the math here. Assume that only these likely undecided voters (the 58%) actually make it to the polls. You’ve got 56% of 58% of 6.5% voting for Romney and the rest for Obama. This would add at most 0.4 points—that is, less than half of 1 percentage point—to Romney’s margin over Obama nationwide. That is, if you assume the model is correct, that we can safely ignore the underlying uncertainty, that only these undecided voters will vote, and so on and on.