On Thu, 23 Sep 2004, Seth Ackerman wrote:
> I understand why a bad LV screen would give a skewed finding.
Actually the argument is that any LV screen used at any other time other than the eve of the election will give not only a skewed finding, but a wildly gyrating one that positively reinforces and exaggerates every shift.
> But why would Gallup consistently get so many Republicans in their RV
> sample? Is there some reason why Republicans are more likely to answer
> the phone when Gallup calls?
I have no idea. And it's not just Gallup, it's CBS too. Standard random sampling is supposed to give you a representative sample 19 times out of 20. And the re-weighting technique Teixeira is using here is usually very frowned upon because party id varies (just like interest and enthusiasm). In fact, this is exactly the process that most pollsters mock Zogby for.
The difference is that in Zogby's case, he's forced to do this because he's already violating so many other rules for getting a clean sample that he never gets a normal party breakdown. His argument that the solution to breaking rules A-D is to break rule E too makes kind of a mockery of the whole process. In Gallup's and CBS's case, we are assuming they are following the standard procedures. So why they would both end up with such skewed samples is a complete mystery that indicates that something is completely out of whack. Either they aren't following the standard process, or some new development has thrown a spoke in the standard process, or something is wrong with our assumptions.
As for assumptions angle (the only one we can really analyze without more information) our prime assumption is that party ID varies in smooth continuous trends and doesn't gyrate wildly. Something may have changed and maybe it is. But offhand, it doesn't make sense why that would suddenly happen this of all elections, because this is the most partisan election since 1972, with the highest, earliest interest level and the smallest amount of independents and persuadables. Those seem like the conditions least conducive to frantic party switching.
Bottom line: no idea. But if our assumptions about the basic distribution of party ID are right -- and I think they are -- then these samples are off. How they got that way is an interesting question.
Michael