Today NPR has a poll out showing Obama leading Romney by seven points. Here's how they managed to do it:
1. They over-sampled women, using a pool of 53% women.
2. They over-sampled union households, using a pool with 17% union households (when the actual figure is something like 12%).
3. They over-sampled people with advanced degrees (urban elites who skew Democratic), using a pool with 14% when the real total is closer to 10%.
4. And, of course, they oversampled Democrats, using a sample that is D+7.
Party affiliation can change over time, and I understand the argument that you shouldn't "correct" your pool to fit an assumption about party identification. But the number of people with advanced degrees or the number of households with union members or the number of women don't change much over time, and to the extent that those are indicators of party affiliation, if you take too many female teacher's union members with master's degrees or female AFSCME members with master's degrees in social engineering or human resource management or whatever it is that government bureaucrats do... well, pretty obviously you are going to end up with more Dems.