Thursday, September 23, 2004
Yet More Adventures in Polling
by Tom Bozzo
The poll was almost immediately labeled an outlier by both campaigns.
The internals (note: PDF), which trailed the Cap Times story a bit, suggest more of the funkiness seen in the Gallup and NY Times polls, noted elsewhere. Party ID was strongly GOP-tilted (29D/36R/25I). The 2000 election question responses were also fun: 61% Bush, 35% Gore, actual result 47.8% Gore, 47.6% Bush.
At least the UW survey folks seem to recognize these as potential anomalies to be explained, though the explanations are reliant on a moderate degree of cognitive dissonance among the surveyed. The Bush/Republican ID bounce is explained as a sympathy effect of the CBS 'memogate' flap ("response to short term campaign developments" when the survey was in the field). If this were true, though, one wonders why Kerry wouldn't see a sympathetic bounce from the Swift Boat Vets for "Truth." The 2000 election result is described as "typically show[ing] a clear advantage for incumbents unless they are wildly unpopular." I've certainly heard of this (with respect to Kennedy, for instance), but it seems like a big effect for an incumbent who isn't especially popular. (Particularly given that the effect is higher than that measured in the March '02 Badger Poll, when Bush was still riding the post-9/11 approval surge.) An amusing detail is that Democrats and independents appear to retrospectively over-report their support for Nader.
Update 9/24/04: Previous Badger Polls have yielded party identifications not (statistically) significantly different from the Harris results cited at Left Coaster (and two posts ago). This is Badger Poll XVII, and XVII independent trials gives around a 58% cumulative Type I error probability. Outlier result is not improbable. This is arguably bad news for Tim Michels, who fared poorly against Russ Feingold in a Republican-friendly sample. I would be surprised, however, if both races weren't close.
Curse words were muttered in the Marginal Utility master bedroom this morning when Wisconsin Public Radio reported that the latest Badger Poll showed a 14 point lead (RVs, leaners pushed) for Bush.
The poll was almost immediately labeled an outlier by both campaigns.
The internals (note: PDF), which trailed the Cap Times story a bit, suggest more of the funkiness seen in the Gallup and NY Times polls, noted elsewhere. Party ID was strongly GOP-tilted (29D/36R/25I). The 2000 election question responses were also fun: 61% Bush, 35% Gore, actual result 47.8% Gore, 47.6% Bush.
At least the UW survey folks seem to recognize these as potential anomalies to be explained, though the explanations are reliant on a moderate degree of cognitive dissonance among the surveyed. The Bush/Republican ID bounce is explained as a sympathy effect of the CBS 'memogate' flap ("response to short term campaign developments" when the survey was in the field). If this were true, though, one wonders why Kerry wouldn't see a sympathetic bounce from the Swift Boat Vets for "Truth." The 2000 election result is described as "typically show[ing] a clear advantage for incumbents unless they are wildly unpopular." I've certainly heard of this (with respect to Kennedy, for instance), but it seems like a big effect for an incumbent who isn't especially popular. (Particularly given that the effect is higher than that measured in the March '02 Badger Poll, when Bush was still riding the post-9/11 approval surge.) An amusing detail is that Democrats and independents appear to retrospectively over-report their support for Nader.
Update 9/24/04: Previous Badger Polls have yielded party identifications not (statistically) significantly different from the Harris results cited at Left Coaster (and two posts ago). This is Badger Poll XVII, and XVII independent trials gives around a 58% cumulative Type I error probability. Outlier result is not improbable. This is arguably bad news for Tim Michels, who fared poorly against Russ Feingold in a Republican-friendly sample. I would be surprised, however, if both races weren't close.
