Bipolarity

Josh Marshall at Talking Points Memo makes a very good point about the disparity between the Zogby and SurveyUSA polls in California: one of them is going to be incredibly, hilariously wrong. But what does that mean about polling in general for today's contests?

The strangeness is here:

Consider this spread. Zogby has his final California number as Obama 49%, Clinton 36%. SurveyUSA has Obama 42%, Clinton 52%.

Both Zogby and SurveyUSA have been right and wrong, and usually right. Then again, they aren't usually this far apart in their predictive assertions. This isn't an internet poll against a phone poll, either. Zogby's was by phone and the survey size was plenty big. However, Zogby's sample doesn't appear to have been drawn from confirmed or likely voters except in the case of New York Republicans. SurveyUSA's sample looks a little more robust:

Filtering: 2,000 state of CA adults were interviewed 02/03/08 and 02/04/08. Of them, 1,762 were registered to vote. Of them, 872 had already voted or were determined by SurveyUSA to be likely to vote at the precinct on 02/05/08. Research conducted for KABC-TV Los Angeles, KPIX-TV San Francisco, KGTV-TV San Diego and KFSN-TV Fresno. Other polls in CA show Clinton with a smaller lead than SurveyUSA shows here. Two polls at this hour show Obama ahead in CA. For Obama to win in CA, he needs a larger turnout among males and African Americans than SurveyUSA's results indicate he will get.

SurveyUSA's sample size (872) at least seems to be comprised of a sample geared toward voting, and Zogby's sample is not necessarily comprised of voters. So Zogby is measuring Name ID and general population sample preference whereas SurveyUSA is measuring intent among a representative sample of likely voters.

So what does that indicate? For one thing, polls have margins of error for a reason — statistically, there's always some noise, some outliers, some restrictions on your sample that affect the outcome somewhat, and you try to minimize that noise but you can never wholly get rid of it. Secondly, I believe that in at least some cases, people are influenced in the last days leading up to an election by externalities that polls can't account for — a charismatic friend, a set of news stories that are particularly influencing, or even a particularly striking campaign ad. These things almost never swing an election (scandalous election day or day-1 news stories notwitshtanding), but it might disturb a population at a basic enough level so as to alter the character of a sample. This is all stats-nerd stuff, but it is important to consider.

That would all be well and good if the spread were within the margin of error, and some last minute shift of less than 5 points could swing it for one canddidate or another. But it isn't. The disparity between Zogby (Obama by 13) and SurveyUSA (Clinton by 10) is huge. Is it possible that the voting public is in flux enough for two random samples to produce results this far apart? The problem is that the samples are different even if the general methodology is similar, so it is impossible to compare them in a scientific way, unless you call looking at them and wondering what's going on scientific.

In the cases where two totally decent pollsters show two totally different results (and TPM Election Central points out that Missouri shows a less severe but similar disparity), things get tough. Does it come down to ground organization? Is there a mysterious constituency that wasn't polled? And if turnout continues to exceed expectations, what does that mean in terms of polls that measure "likely voters"?