To me the most likely scenario is that a number of undecideds have shifted to Obama, while an equal number of previously-decided McCain voters have now become unsure. Doesn’t this seem more intuitively right than a bunch of decided people changing their minds completely, while the undecideds in between fail to make a choice? Scold me if this has already been put forth.

Imagine a sort of upside-down bell curve, with “candidate preference” as the x-axis and “decidedness” as the y-axis. The more strongly you prefer a certain candidate, the more likely you are to consider yourself “decided”, with a sort of valley of undecideds in the middle. In such a situation, if one pushes the preference of the nation as a whole in one direction, then some of the undecideds will now be pushed over the threshold into decided territory, and on the other end, some who previously considered themselves decided fall into the valley.

If you take this as a model, it brings up all sorts of interesting stuff, because we don’t know how the actual population is distributed along the x-axis. For example, if there are a tremendous number of McCain supporters who identify themselves as “decided”, but are very close to the valley, and there are relatively small numbers of people in the undecided group, then a shift in mass public opinion in favor of Obama would create a big point shift, not because he gains a lot of newly decided voters, but because previously decided McCain voters suddenly become uncertain. You would be able to detect such a situation if the poll margin changed and the number of undecideds expanded. That hasn’t happened, which would suggest under this model that the distribution of the voting public along the “preference” axis is generally flat.

Josie, likely voter screens include questions such as whether the person voted in the last election, the degree of intent to vote, and so on. Evidently it is possible to be a near-certain voter yet be undecided.

Look at the change in decided-voter margin over time. At a minimum, this swing represents either the number of undecideds forming a preference that they can state, half that number of decideds changing their mind, or a mixture.

To pick the largest example, in the last 4 weeks the swing has been 9 points. At the same time, the number of undecideds has stayed the same. The simplest interpretation is that at least some voters had an expressed preference, but switched.

One caveat has to do with likely voter screens. Obama and McCain supporters might have maintained their preference, but their likelihood of voting has changed. However, this should still cause fluctuations in the undecided fraction, which aren’t observed. So for now, I lean against this hypothesis.

Adriel, your argument is appealing, but “common sense” is a speculation, not a conclusion. I am not aware of historical data that support the idea that undecideds will break massively in one direction.

Josie – Do people say they are undecided because they don’t have access to their true opinion? Are they truly undecided? Are they mulishly stubborn about answering some stranger’s questions? I suspect the first two are dominant factors, but I don’t know.

Nic, there appear to be fewer “undecideds” in swing states. For example, since Sept. 1 they are 5.5 +/- 0.6% in OH (N=30) and 5.0+/-0.5% in FL (N=28). In regard to the Bradley effect, based on the last 12 years of gubernatorial and Senate races, it is on average zero. Any other assumption would not be a suitable default assumption. I have posted on this topic before.

Everyone else: individual pollsters can guess wrong in likely voter screens and other factors such as weighting. But on average, state polls are quite accurate in predicting the winner. Keep in mind that we are talking about effects here that are likely to add up to no more than a few percentage points. They could matter, though not if the election were held today.

Thank you for this marvellous website and its interesting analyses.

Are you aware of any data which shows what proportion of “undecided” people eventually decide just to stay home on election day, and thus don’t end up affecting things one way or the other?

On a more general note, I’m guessing that most of the polls you use are of “likely” voters. Are you aware of any data which shows how closely the polling firm’s designation corresponds to reality?

I ask these questions because it strikes me that the difference in voter turnout from election to election (as shown in the table here: http://www.infoplease.com/ipa/A0763629.html ) seems to exceed the current meta-margin and I do not understand completely how turn out interacts with and affects your analyses.

Again, thanks for your work.

It doesn’t look like any of this will matter this year, but it is something that could come up again.

Has anyone seen a worthwhile discussion of statistical and/or meta-analysis use in detecting fraud, say in those Diebold voting machines or other classic fixed voting scams? Will we have U.N. election monitors in the U.S.? Should we?