One of the most contentious debates in this election cycle has been over party ID samples in polling.
Poll skeptics (mostly conservative) argue that the party ID sample of a poll should roughly reflect the likely party ID breakdown of the electorate on Election Day. If there are likely to be 33% Republicans, 33% Democrats, and 33% independents voting, then polls should reflect that in order to get an accurate picture of how they plan to vote. This is of course a very difficult thing to estimate in the first place. Many factors affect turnout, and it varies from year to year.
Poll purists (mostly liberals, media figures, and pollsters themselves) instead argue that the party ID sample of their polls is an organic finding of the electorate in a state or nationally at any given time, mostly reflective of voter enthusiasm (the more the electorate identifies with a party, the more they will identify with it in a poll). Purists also argue that they should leave the sample distribution as they find it, since estimating what the party ID sample should be and adjusting their numbers as necessary is presumptive at best, and completely distorts the data at worst.
The purist argument makes sense from a statistical perspective, but the pollsters’ own data argues against it. In the last week, PPP, a private polling firm mostly hired by Democrat groups, conducted two different polls of Iowa voters. This turns out to be a perfect opportunity to test their own theory of party ID. PPP’s 10/19/12 poll, conducted 10/17-19, found Romney up 1 point overall, with a party breakdown of R+2. Their 10/21/12 poll, conducted 10/18-19, found Obama up 1 point overall, with a party breakdown of D+7. Both cannot be true. If party ID sample in polls represents organic electorate composition and enthusiasm in a state or nationally, then two polls taken by the same firm at the same time in the same state should show the same party ID. But they don’t.
So where does that leave us? Party ID is important on Election Day. Turnout wins. But it’s not predictive in polling, it’s determinant. As PPP’s two Iowa polls show, shift the sample (intentionally or not), and you shift the result. That has been the biggest statistical story in this election cycle, I think. The polls haven’t changed all that much in the actual inter-party result; Republicans are voting for Romney, Democrats are voting for Obama, and independents favor Romney by an average of 10ish points. The only thing that determines who wins is how many of each group answers the phone or turns out to vote.