Who Can Really Say?

Likely Voters and the Incumbent Rule

Continuing a time honored tradition here at WCRS (a few weeks is a period of time after all), I link again to the fascinating stuff put up by the Mystery Pollster. Today he blogs about two of the hallmark polling issues in the recent campaign -- the likely voter model and the incumbent rule.

As for the latter, he says it didn't happen this time around:

I’ll concede that the incumbent rule (under which undecided voters on the final survey typically break to challengers) did not hold in this year’s presidential race, but I’m dubious of a break towards Bush. In the national exit poll, those who made their final vote choice in the last four days went for Kerry 53% to 44%, while those deciding earlier went for Bush (52% to 47%). The same pattern held in Florida and Ohio. The average of the national tracking surveys also showed no trend to Bush over the final weekend; if anything, they had Kerry running slightly closer.

I may be missing something here but two things bother me about this. First, the incumbent rule isn't about what polls are doing over the weekend -- it's about what happens when comparing the last poll before the election with the election itself. If Bush has 48.6 and Kerry 46.9 after the average of the final polls and they end up 51 and 48 respectively, it seems to me that Bush got the better of the deal. As well, those making "their final vote choice" over the weekend may well have made a different choice earlier in the campaign. If the incumbent rule is supposed to assist in predicting how the undecideds in the last poll will vote, then I'm not sure I get this.

Next, he turns to the likely voter model, something he blogged about frequently before the election. His unsurprising take -- some of them worked better than others. The juice is in which ones were better.

When I presented the details on likely voter models, I noticed that the pollsters that used a variant of the Gallup likely voter model showed Bush doing consistently better than other surveys. That difference now looks prescient. The following table shows the results of those using the Gallup likely voter model either in the final week (Gallup, Pew, Newsweek) or in the final two weeks (adds Time and the LA Times). In both cases, the Gallup-model showed a Bush margin closer to the actual result (3.2%+) than the average of the other surveys (0.9%). There were three surveys in the “other” category that correctly forecast Bush’s final three-point margin (notably, Pew, TIPP, ICR), but the other 10 showed Kerry doing slightly better.

In part Gallup eschewed weighting by party ID in their model, which other polls embraced. MP argued (also in part) that weighing by party ID is problematic since responses to those questions are more volatile than say, responses by age or sex. But two polls which weighed by party ID were still pretty close, and he says he'll have more on that in a bit.

The MP is one of my favorite sites and if I got any traffic (aside from his in the first place) I'd apologize for being such a lazy SOB in not adding him to my blogroll, which suffers a form of blogoshpehric rigor moritis.

Posted by Peter at November 12, 2004 09:21 PM