First, there is only one good reason to purposely over sample a certain group. I stated that reason, in my first post.
You raise a good question. You can base the 'control number' to weight back to through known data (voter registration information, census data, etc.). You could use a truly random survey of the poputlation. That's not easy, due to refusal bias. Or, you could use a combination of the above.
Known data and census information are dated sources, so a shift in partisan allegience, wouldn't be accounted for.
In my humble opinion, a random sample of the population would be the best way to account for such a shift. Surveys done for the government, usually by colleges, really do a good job of getting the refusal bias out of a survey. They use methods that are cost prohibitive to the private sector, but the government doesn't seem to mind :-)
Polling is not a perfect science. This is especially true with political polls, because likely voters can fluctuate significanly, as voting day approaches. Plus, they are self selected. More people tend to say they are likely to vote, than actually do vote.
Any time you have to weight, you reduce the accuracy of the poll. If they is a major change in the information you are using to weight the data back to, that isn't accounted for, you're in trouble.
In political polls, I look for trends, instead of 'so and so is up by X percent'. I'm liking the trend, right now.
Thanks. I wonder if polls in the months after 9/11 showed any swings in party allegiance...do you know?