"Regular readers will know that I generally refrain from making comments on the demographics within an individual poll. They will vary a bit from survey to survey based both on random variance and the different assumptions that pollsters make about just who will turn out. The random variance can be reduced by taking an average of surveys — and if pollsters have a persistent tendency to favor one candidate over another, we can account for that with our house effects adjustment, which is designed to detect and counteract these tendencies.
But once in a great while, a poll comes along with methodology that is so implausible that it deserves some further comment. The Foster McCollum White Baydoun poll of Florida is one such survey.
The poll was weighted to a demographic estimate that predicts that just 2 percent of Florida voters will be 30 or younger. It’s a decent bet that turnout will be down some among younger voters this year, but that isn’t a realistic estimate. In 2008, according to exit polls, 15 percent of voters in Florida were between 18 and 30.
The poll also assumed that 10 percent of voters will be between the ages of 31 and 50. In 2008, the actual percentage was 36 percent, according to the exit survey.
The poll projected Latinos to be 7 percent of the turnout in Florida, against 14 percent in 2008. And it has African-American turnout at 10 percent, down from 13 percent.
If the turnout numbers look something like that in November, then Mr. Obama will lose Florida badly. He’ll also lose almost every other state; his electoral map might look a lot like Walter Mondale’s.
But the share of voters 50 and younger in Florida is not going to drop all the way from half the electorate to roughly one-tenth of it, as the poll assumed. That is far beyond the range you can get from reasonable disagreement about methods, or from sampling error. It looks like the result from a from a badly-designed statistical model that never got a sanity check."
http://fivethirtyeight.blogs.nytimes.com/2012/08/20/aug-20-when-the-polling-gets-weird/