In today's polling thread, I referred to the new SurveyUSA poll of Kansas, which showed a rather strong result for John McCain -- he was 23 points ahead, roughly double the margin he'd held in previous SurveyUSA polling of the state.
Bob Erikson of Columbia University, who was largely responsible for the trendline adjustment that we now apply on a daily basis, wrote to inform me that SurveyUSA had changed its procedure, now using a likely voter model as opposed to a registered voter model. The change actually isn't all that recent -- SurveyUSA had begin using likely voter models in most of its polling at some point in June -- but it had largely escaped my attention.
There is absolutely nothing wrong with SurveyUSA doing this -- many pollsters who begin with registered voter models switch to likely voter models at some point before the election.
It does mean, however, that we should be careful about trying to make apples-to-apples comparisons when a pollster switches from a registered to likely voter model, which have tended to produce fairly different results thus far in this election cycle. Therefore, for purposes of calculating our trendline adjustment, we will consider likely and registered voter surveys from the same pollster to constitute two separate data series, e.g. "SurveyUSA-LV" and "SurveyUSA-RV".
As a result of making this fix, Barack Obama picks up about 0.4 points in his popular vote margin, although John McCain remains the nominal favorite to win the electoral college, taking 50.1 percent of the simulations to Obama's 49.9 percent.
n.b. Some people had also argued that this same rule should be applied in another circumstance -- when Rasmussen began breaking out its results with "leaners" included several weeks ago. It is a judgment call, but I consider pushing leaners to be sort of a relatively trivial, organic variant of the pollster's existing model, rather than the more fundamental difference in philosophy that likely and registered voter polls represent. Besdies that, whereas pollsters almost always disclose whether they are using a likely or registered voter model, they frequently will not disclose how they handle leaners, which would make this rule hard to apply even if we wanted to.