So the pollsters try to minimize these confounding factors by making a weighting adjustment for party affiliation. The problem is they have to guess what ratio of Democrats, Independents, and Republicans to use. Usually they base this on exit polling from previous elections. Some pollsters do not make such an adjustment. Still others decide whether or not they think they need the adjustment after the poll is taken.
Polls, such as Newsweek, that do not weight tend to be rather erratic in these ratios, and in the case of most of the polls that I have looked at that had a debate bounce for Kerry, they have shifted from a high Republican mix to a high Democrat mix. This suggest either they have fallen prey to confounding factors, or there has been a sudden national shift in party affiliation.
Polls, such as Fox/Opinion Dynamics that decide whether to weight after seeing their results tend to show very little change in general, and don't tend to be helpful. Essentially they end up being tweaked to show what the pollster already thinks.
Polls that weight by party affiliation are probably the best for tracking trends, but they are still of questionable use in determining the real results. I have not seen much of a Kerry debate bounce in these polls.
Those are all good points, but you must remember that there are still good reasons to let partisan identification vary in the polls.
The presumption that partisan identification is an independent variable, i.e. that it is a unchanging cause rather than a variable effect, freezes the poll from picking up potential changes in it, which would make it a dependent variable. That partisan identification can change, that it can act as a dependent variable, would mean that polls that treat it as an independent variable have a poor methodology.
These polls with varying numbers are not just commenting on the horserace. They are also commenting on how many partisans they pick up in the real world. These numbers can vary depending upon events. Thus, Bush had a poor night on Thursday, so the next day when Newsweek or Gallup calls, weak Republicans identify themselves as Independents, Democrat-leaning Independents identify themselves as Democrats. The change in party ID is caused by events on the ground; thus, it is best understood (in this case) as a dependent variable. Meanwhile, Zogby, Fox News or the others that freeze party ID fail to pick up that change because they treat it as independent.
I am not saying here that Gallup and the other polls that let party ID vary are right, but I am also not saying that Fox News is right, either. Both sides are taking a guess on this issue. Both guesses have some justification and neither is a decisively clear choice. You have made a strong case for considering party ID as an independent variable, but there are strong reasons to consider it a dependent variable. What is the reality? We shall not know the answer until 11-2, where we will see if partisan ID is different than it has been in years past.
More than that, thoughI think you are making a mountain out of what is a statistical molehill. I have a lot of problems with the Newsweek poll, but both that poll and the Gallup poll have partisan samples that are within the margin of error from their past partisan sample. Statistically speaking, you cannot say that anything is really happening in that poll. The change in partisan ID is entirely explicable by the kind of random statistical variation that occurs anytime you take a sample of a population to draw inferences about the latter.
This is a point that people (not you in this post) usually fail to take from polling. When, for instance, the CBS News poll show the race tied at 47-47, what they are saying is that they are 95% confidant that Bush has somewhere between 44% and 50%, and that they are 95% confidant that Kerry has somewhere between 44% and 50%. So, all they are saying is that the race is somewhere between Bush +6 and Kerry +6.
Statistical sampling of any population is, at best, a rough estimate. Usually, as in the case of most social science questions, that roughness is not a problem. But in a presidential contest, one where the difference in 2000 was 537 votes, there is a precision required that polling simply cannot provide. Pollsters call their work an "art" in recognition of this fact. They are being a tad disingenuous when they say that. It is not an "art." Rather, it is a science that is poorly suited for presidential politics.