Friday, September 17, 2004
To model the likely electorate, researchers used a weighting method based on a model developed after the 1996 and 1998 elections, and tested on all general elections since 1996. It weights each respondent on a combination of demographic, attitudinal and behavioral questions correlated with voting. The individual questions measure income, certainty of voting, having voted in the last presidential election, and registration status.
I wonder why they've not updated the model with data post-1998? It would seem to me that issue saliency vis-a-vis income, for example, would be different between 2004 (when one candidate is proposing big tax cuts and the other big spending increases) would be different than it was in 1996 or 1998. Voting behavior in time of military conflict also would be different. Out-of-sample prediction is hard enough without biasing the sample with elections that look nothing like the current one. This could be a source of the poor forecast performance Trunk is talking about.