Mark Blumenthal | October 17, 2008
We logged 21 new statewide polls yesterday, including 9 of the controversial Zogby Interactive surveys conducted online using a non-random panel of Internet volunteers. The overall pattern of the new polls is roughly the same as we have seen in recent days. Most confirm the gains registered by the Democrats since September, although the net impact on our trend estimates over the last 24 hours is mostly negligible. The exception is Ohio where two new surveys, including one from Zogby, narrow our trend estimate enough to shift that state and its 20 electoral votes from lean Obama back to our toss-up category.
Most of these polls follow-up on previous surveys from the same pollster conducted in September or earlier, and as such, most confirm the recent gains by the Obama-Biden ticket. Only three of the the new surveys track results gathered earlier in October, two of those show a slight shift to Obama, one shows a slight shift to McCain.
Virtually all of the interviews in the national tracking surveys posted yesterday were conducted before the Wednesday night debate. Today's releases will be the first to indicate whether the debates made any noticeable dent in vote preference nationally. Please note that for the rolling-average tracking polls, the table above lists the previous non-overlapping sample for each pollster, not the release from the previous day.
The new surveys moved our trend estimates in both directions in the closer battleground states. In five states, Missouri, North Carolina, North Dakota and Virginia, the estimates shifted slightly in Obama's direction. In five states, Colorado, Florida Indiana, Nevada and Ohio, the estimates moved slightly in McCain's direction.
A new poll in North Dakota sponsored by a Democratic affliated union, only the second public poll released their in a month, confirms the very close result obtained by the The Forum there earlier this week and solidifies our surprising "toss-up" status for that state.
Two new surveys in Ohio were from Rasmussen (showing a 49% to 49% tie) and Zogby Interactive survey (showing McCain leading, 50% to 45). These narrow Obama's lead on our Ohio trend estimate by nearly two points, enough to shift Ohio back to the toss-up category. Our estimate shows Obama leading McCain by two and a half points (49.0% to 46.5%)
Ohio is the one state where -- for the moment at least -- our inclusion of the Zogby surveys affects the classification. If we use the "filter" tool on our charts to remove the Zogby surveys from the trend estimate (as in the modified chart above), Obama's lead widens by about a point (to 49.6% to 45.8%), which would have left Ohio in the lean Obama column.
All of which raises the question of how Internet panel surveys like Zogby's work and why we include them. The short version is that Zogby, like most companies doing survey research online draws, a sample from a non-random panel of volunteers that have agreed to complete surveys online. They then attempt to weight the completed interviews to match the demographics and partisanship of the electorate. (Interests disclosed: Pollster.com receives financial support from YouGov/Polimetrix, another company that conducts internet panel surveys).
Despite the hype on their website, the Zogby Interactive surveys have produced results of dubious accuracy. Following the 2006 elections, the Wall Street Journal's Carl Bialik found that Zogby's online surveys "missed by an average of 8.6 percentage points" in U.S. Senate races, "at least twice the average" of results from four other pollsters he examined, SurveyUSA, Rasmussen and Mason-Dixon. The Zogby Interactive surveys also rank second to last on Nate Silver's computations of "pollster introduced error," earning an error rate (5.73) more than double the average (1.97). Charles Franklin also noted some odd patterns in their national trial heat results in late 2007.
So why include them here? Our philosophy since launching Pollster.com two years ago has been to include all polls, good, bad and ugly. We do so partly to provide a reference and record of all polls, and partly because the loess regression trend lines usually resist the influence of "outlier" results from a single pollster (Charles Franklin discussed this issue at length here). And finally, we have worked hard to provide interactive charts that allow you to filter out individual pollsters to check for instances where one poll or one pollster may have a disproportionate impact on our perceptions of where the race stands.