trumptman - you talk about "weights of the polls." Do you believe that pollsters pick how many Rs and Ds to survey? And that the polls are wrong because they're intentionally surveying too many Ds? And that this is part of the liberal media bias?
As I mentioned in the other thread, I feel bad for you guys if that's what you think - honestly, not even sarcastically (OK maybe a little sarcastically) - because many Ds convinced themselves of the same thing in 2004. I've been there, man. We have amazing defense mechanisms to protect ourselves from the truth.
They aren't wrong because they survey too many of one thing or another. However they can be wrong when they choose the wrong weights to apply to that poll to correct for what they declare to be under or over represented sub-groups. However if your sample consistently samples in a manner that tilts too far away from the known results in reality, shouldn't one question if they should adjust their methodology for obtaining the sample? In the other thread I noted that Silver had criticisms about the Rasmussen methodology. If your sample can't seem to find the group you are seeking the views of, you cant adjust the weights (which they are actually going in the wrong direction with and making it even more Democratic) but errors are going to be magnified as well.
This isn't about a defense mechanism.
All told, we see a statistically significant relationship between Obama's margin and the Democratic advantage in partisan identification. In other words, there appears to be a bimodal distribution of the polls. They are not converging around a single point. Instead, some (notably Rasmussen, Purple Strategies, Survey USA, and Mason-Dixon) see Obama ahead by just 1 to 3 points in the key swing states, while others (notably the Washington Post, Fox News, PPP, and NBC News/Marist) see an Obama lead that ranges between 4 and 8 points. And the difference looks to be built around how many Democrats are included in the polling samples.
I'm still unconvinced by the weighting argument when applied to party affiliation, since that is too closely related to the variable that the polls are trying to measure - voting intention. Weighting by independent demographic parameters that can be measured separately (age, gender, ethnicity, landline ownership etc.) is important when using relatively small samples that may not well represent the sampled set, but to weight results by party affiliation is to some degree begging the question in my opinion. If public opinion were swinging towards Republicans, for example, then it would be expected that a increasing fraction of those polled would indicate an intention to vote Republican, and an increasing fraction would self-identify as Republicans. It would be statistically incorrect then to weight the results towards Democrats on the basis that they were inappropriately under-represented in the sample - they would, in fact, be appropriately under-represented in the sample. I'd expect a positive linear correlation between those two increases (intent to vote Republican and self-identified as Republican), though with a non-unity constant of proportionality. And that appears to be what many unweighted polls are showing - in this case in favor of the Democrats.
In terms of the broad range of polls out there, I could well imagine a bimodal distribution of results with unweighted polls grouping around one peak and weighted polls grouping around the other - if the weighted polls all used the same weighting methodology. However - you are quoting polls that span the range from 1% - 8% in favor of Obama, and, even though that's a small number of data points, I don't see much evidence of two peaks in that distribution.
