I'll look again at the formula that I summarized on the methodology page. My description of weighting is correct, but I may not have summarized the formula properly.
What's the correct way of combining statistical data? I admit to being an amateur (Dammit, Jim, I'm a lawyer, not a statistician). I wanted to do more than a straight average because I believe Likley Voter polls should be weighted more heavily than Registered Voter polls, or all adults.
I'll post more later when I have time. When combining means, you need to weight inversely according to the variance of each mean. You may be ok with what you're doing. Roughly the variance of a poll is proportional the number sampled so you're doing that right. Modifying for likely-registered is ok too (as long as it's done consistently and noted in the results.) I think there's a minor modification I'll suggest later.
Also, if states are close (within a few percentage points) you should just call them a tossup.
BTW FWIW I agree with Dr. Stochastic's comments regarding the precision of the data as it's presented. The number of significant digits displayed does imply the data presented is accurate to that level.
The L.A. Times polls will not have much of an impact on your results because they don't do many of them, but the ARG and Rasmussen polls will. Your site can be seen as a sort of worst case scenario for Bush (or best case scenario for Kerry).
If I were smarter, I might be able to put together a site which had the polls normalized. That's a worthwhile project, which maybe I'll try if I can figure out how to log on to my computer. I also have to figure out make the pretty pictures with the EV's.
If you wanted to come up with a normalized map, that would be pretty cool. (One could choose which poll to use as the base line, and the map would display according to that).
You may also want to combine error estimates. The linked site gives methods for doing this.