Conservatives, perhaps hoping to keep the base involved despite polls showing a growing Obama lead, are claiming that polls are systematically biased in how they choose their samples. They say the pollsters are counting too many Democratic voters, so of course their numbers will show an Obama lead. Party identification numbers which differ from similar numbers last time around are the source of the erzatz margin. A site called unskewedpolls.com removes the supposed bias by assuming a return to '08 party identification proportions.
This is not at all standard practice for polling. While they correct for things like age, gender and ethnicity, they make no effort to correct for party identification, allowing the interviewees themselves to determine the mix. Nate Silver at fivethirtyeight.com takes a look at the performance of polls in past elections and finds that this practice has historically resulted in predictions without partisan bias.