Here are a couple interesting stories from two well-respected poll watchers.
Jay Cost looks at the polls in Ohio and thinks that he sees a bimodal distribution of their results. Some polls (Rasmussen, Gravis, ARG) consistently show a very tight race. Other polls (CBS, NBC, PPP) continue to show a sizable Obama lead. In a normal distribution of polling results we would expect to see polls clustered about the mean, with each polling group sitting on boths sides of the average. Here there are virtually no polls showing an Obama lead of two points, which is the approximate arithmetic mean.
The aptly named Sean Trende explores the divergence between the state polls and the national polls and concludes that they can’t both be right. National polls show a Romney lead of approximately one point, but the state polls continue to show Obama leads in many of the swing states. Trende constructed a model using the average of state polls to try and come up with the national result. He couldn’t make the numbers work.
Both Cost and Trende are observing something I have been pointing out for weeks. How you build your sample predetermines your result. If you assume that 2012 turnout is going to closely match 2008, then you probably have a sample showing a slight Obama lead. But if you assume a return to a more historically normal turnout, then your poll will like have Mitt Romney with a comfortable margin.
This just goes to show what I like to say about polling: “It is an educated guess heaped upon conjecture piled atop assumptions filtered through subjectivity and complicated by lies.” It would be a heresy to mathematics to treat polling as gospel truth.