Wednesday, October 10, 2012

Return of the Gallup Likely Voter Filter

Just when you thought it was safe to get back in the sample pool, Gallup switched on their likely voter filter.

The results and that the headline result isn't the registered voters subsample of the tracking poll showing Obama ahead by 3% but the likely voter subsample showing Romney ahead by 2% and my loyal readers (if any) get the latest entry of the continuing saga Robert dumps on the Gallup likely voter filter.

First note that in 2010 Gallup's likely voter poll forecast Republican share minus Democratic share 9 points greater than the outcome ?  Second note that Gallup knew something was wrong and broke with tradition by also reporting a broader likely voter sample using a filter based only on stated voting intentions.

Second note that I slipped and made a prediction in 2010 (I was wrong as usual).  My problem with the traditional Gallup likely voter filter (described below) is that I think that well before the election it introduces bias which goes away as election day approaches.  So I predicted that Gallup R-D minus other pollsters R-D would decline as election day 2010 approached. Oooops.  The opposite happened.  But the filter just didn't work.

I actually checked to see if Gallup is using their traditional filter -- they are.

The calculation of likely voters is based on registered voters' responses to a seven-question series that -- with some revisions along the way -- Gallup has used since 1952 to calculate voters' likelihood of voting. In some years, such as 2008, there was only a marginal difference between the vote choices of registered voters and likely voters. In others, such as 1996, there was a much more substantial difference.

At this point, Romney voters are somewhat more likely to respond that they will definitely vote, that they have thought a lot about the election, and that they are more familiar with where people in their local area vote. These attitudes indicate that Romney at this juncture will benefit from higher turnout on Election Day among his supporters than will Obama. These patterns could change closer to Election Day as more voters become engaged or if Republicans' or Democrats' enthusiasm for voting is altered by campaign events.

Note the amusingly unusual grammar in my italics , but the bold. is the part that matters.  One of Gallup's 7 questions (5 yes answers are required to be considered a likely voter) is (something like) do you know where to vote ?  Clearly a no answer means something different in early October and early November.  If the election is the day after tomorrow people who say they will go to, vote but don't know where, are not very likely to actually vote (has happened to me but I'm flaky).  The same ignorance in early October is less of a signal of unlikeliness to vote.

The problem is that Gallup has evaluated the filter for decades mostly based on the gap between the last poll and the election.  I don't know if they would notice a bias which fades (especially if covered by true drift in public opinions towards Republicans which, I think, was typical for many of those elections as conservative Democrats reluctantly voted Republican).

As predicted by @Austen_Goolsbee (great economist but if you cut people's taxes don't try to prevent them from noticing) this was reported as a shift to Romney.

Obama would have been 0.15% ahead.

No comments: