No way undecideds break entirely for one candidate. 60-40 at best. Basically at this point late in the game, the polls need to be wrong en masse, or McCain is done.
When they all use the same basic model that is wrong, then they will be incorrect en masse.
If they're "undecided" at this point, they ain't voting for Obama. 67-33 to McCain is the most conservative case I'd consider, with as high as 80-20 to McCain.
You heard me, Mister! Or Miss! Or Mrs.! Or Ms.
Here's the question though: do you really believe that those 8% are really "undecided," this late in the game?
For those who are truly "undecided," it means that their hearts are for Obama, but their brains are for McCain. Come crunch time, it's a fair bet that most of those folks will go with their brains. But they're a small number.
But for the majority, it's almost certain they're not really undecided at all, even if that's what they told the pollsters. The question is, why would they lie to a pollster? To echo Dick Morris, that bodes ill for Obama: the only reason to lie is that they don't want to be seen bucking the trend of the polls, which invariably slant toward Obama. Those folks will either go toward McCain, or not vote at all.
60-40 is not ‘at best’. Also there is a margin of error that must be accounted for.
IBD was accurate in 2004 but they were not in 2000. They may not be accurate in 2008, or they may be. No one knows, no one.
Daily polls are never based on a quality random sample. There is not enough money/time. Thus, each polling organization has built-in biases caused by their sampling plan and execution of the plan. Those biases can be considered a random variable among polling organizations. Thus again, some polling organizations will be ‘lucky’ to be accurate on election day. IBD was in 2004, and likely someone else this go-round.
The averages and error bands are not important because they are unreliable when some polling results show estimates within the MOE. What is reliable from a field of mixed results is ‘trend’, as it contrasts out base levels leaving uncorrelated differences. But then again each polling org’s built-in biases will either accurately detect trend or confound actual trend signal.
Treating each polling organzation as a cluster of random variable generation, an aggregate analysis of their trend signals and the strength of their trend signals via smaller error margins of trend estimates can be reliable (less biased) and useful.
So the fact that the McCain campaign is touting momentum detection of trend and the Obama campaign is worried is most telling information.
If none of this makes sense to you, then the take-away is that PhD statisticians know these polls are dismal examples of status and forecasting. Any statistician worth their salt would be reluctant to certify any of these polling results unless there was a large separation of the levels (averages) by a number of standard deviations and the sample was sufficiently random enough for accurate inference with respect to the number of standard deviations used to specify real separation.
I disagree. Huge moves of “undecideds” to one candidate happen frequently (Reagan in 1980, for example). In this case, the unknown is Hussein. He is finally being scrutinized and will be found wanting. I think McCain will get 75% - 80% of the undecideds this time.
The cool, hip thing would be to simply go with the crowd and the media. (Hope I’m right.)