So, in true nerd fashion, I wondered whether the 2004 presidential election polls actually had any accuracy in regards to who would win, on a state by state basis.
So, I downloaded all the non-partisan polls from 2004 from each of the 'swing states' in that election, from www.electoral-vote.com (a truly excellent polling site).
What did I learn? Well, my technique was to see what would happen if one averaged all the polls for a given state that were conducted in the 7 days prior to the election. I wanted to see if the average of these polls would accurately predict who would win that state.
I ran this analysis for all the 2004 swing states, and all those that ended up with a 5% or less spread between the candidates (CO, FL, IA, MI, MN, NV NH, NM, OH, OR, PA, WI). In 100% of cases, the averaged poll lead did go to the eventual winner of that state. I'm not saying the margin of win was correct, but the actual winner was accurately predicted by the average of all polls in the state.
What does this mean for the coming election? Well, with some caveats (which follow) it means we should be able to average all the results of polls available on Tuesday morning and predict the winner, state by state, and therefore the overall election.
The caveats here are: 1) this only works if the way polling is done has not changed substantively since the last election; 2) if voters are being as honest this year as 4 years ago about who they will vote for; 3) if there is not some other unaccounted for factor that has not been taken into account here (such as more young/first time voters this time around, or more cell phone users who weren't captured by the pollsters, etc). Still, overall, given that this method worked 100% for the last election, it seems probable to me that it will work this time as well.
We'll see.
