Wednesday, September 24, 2008

Too Close to Call

I've added a time decay factor to polls in the model so that older polls have less weight than more recent polls. National polls are assumed to decay in value over a 10-day period while state level polls are assumed to decay in value over a 14-day period. This change gives the daily tracking polls much more weight than before, as well they should. I also found a pile of mostly state-level polls done by the American Research Group that have not been included in the list of polls put on the RealClear Politics website and included these polls in my model. It appears there is some criticism of the accuracy of polling done by ARG, but a review of the polling results done by this organization suggest that their results seem in line with polling results from other organizations. All of this extra effort produced more or less the same results as if I had not made the effort. However, as every statistician knows, such efforts lower the uncertainty in the model and so I feel much, much better about the model, even if nothing else seems to have changed. It's like taking vitamin pills - they make you feel better if they actually don't make you any healthier.

The popular vote share remains the same as several days earlier:
While the movement of New Hampshire into McCain's column gives the Republican candidate an 18-vote margin in the Electoral College:
The reasons for the discrepency between the national popular vote and the electoral vote are extremely close races in five states (Colorado, Missouri, New Hampshire, Ohio and Virginia) all of which are in the McCain camp according to my model:

In reality, it is more reasonable to view these five states as toss-ups and the election as well.

No comments: