Alan Reifman sends along a link to his webpage on weighting survey results to adjust for party ID.
The basic idea is, suppose 2 surveys are taken 1 week apart, and Bush has 40% support in the first survey and 45% in the second survey. But then you look at the stated party idenfications of the respondents, and you find that 5% more of the respondents are Republicans in the second survey. We know from experience that party ID almost never changes so dramatically; thus, it seems that what happened is that the second survey just happened to get a lot more Republicans (just from random sampling) than the first survey. Thus the evidence is that there was not a dramatic surge toward Bush in the general population.
Adjusting for party ID is a good idea, I think, which is why I wrote a paper on it (with Cavan Reilly and Jonathan Katz) a few years ago. I recommend reading that paper for more on the topic. Here’s the abstract:
We investigate the construction of more precise estimates of a collection of population means using information about a related variable in the context of repeated sample surveys. The method is illustrated using poll results concerning presidential approval rating (our related variable is political party identification). We use poststratification to construct these improved estimates, but because we do not have population level information on the poststratifying variable, we construct a model for the manner in which the poststratifier develops over time. In this manner, we obtain more precise estimates without making possibly untenable assumptions about the dynamics of our variable of interest, the presidential approval rating.
When it comes to using pre-election polls for forecasting elections, which is related but different than presidential approval, one should be mindful that weighting by party id might also be throwing away potentially useful information. Using the example above–while it is unlikely that there is that big a shift in the population in party id, it is conveivable that there is a shift that large in terms of party id of likely voters.
Most political polls for campaigns use questions to screen for likely voters. The results of the poll we see are for the population of likely voters–which changes from election to election.
So, an observed difference of 5% in party id could be attributable to three separate forces (in addition to chance error)–underlying change in party ID (presumably based on candidate messages), Republicans becoming more likely to turnout, and Democrats becoming more likely to NOT turn out.
How one would discern among these prior to an election is not immediately obvious. The only point is that there are potentially some contexts where weighting can make inferences more difficult.