There have been a lot of questions regarding the widely varying Gallup Polls. So, I thought I would take a minute to point out the major shortcoming of the poll methodology. As many have speculated, the major problem with the Gallup Poll is the Likely Voter Model they use. In short, Gallup determines whether someone is a likely voter on election day by asking a screen of questions such as how likely it is that you will vote and whether you voted in the last election. In 2000 they used a 7-point scale. They further assume that only 50% of those polled will turn out to vote. Usually the numbers run such that all of those who score a 7 are included in the LV sample while only a portion of those who are rated a 6 end up in the sample. If they have 40% from the 7-rated respondents then they need to get the additional 10% from the 6-rated respondents. So they do this by weighting the 6-rated respondent pool. Voila! The Likely Voter Pool!
Sound fishy? The Likely Voter Model does work rather well when it is used for what it was developed for ... to predict the election results on election day. However, b/c the likely voter sample changes over time, you can only use this methodology as the sample stabilizes. In other words, this method works great if the poll is done the night before the election. It works less well the farther from the election date you get. If you're using the method in September ... well, as you can see from the poll results ... you get junk. That's b/c the likely voter sample will depend on how each person feels that day. For example, suppose you are a Kerry supporter and you happen to be called on a day that you're feeling down on your candidate and wonder if it's even worth it to vote ... so you tell the interviewer that you are "likely" to vote instead of that you are "highly likely" to vote. Voila! You may have set yourself up to be "screened out" even though you end up voting for Kerry on election day. In practice this means that the likely voter sample will have a changing mix of Republicans and Democrats. If Republicans are all fired up just after or during the Republican convention then the Likely Voter Model for that period may have more Reps in it; likewise, for the Democratic convention. This is the fallacy of using the model this far out from the election. There are several other issues with this model but they do not have as large an impact.
To sum up, the Likely Voter Model was never developed to be used this far out from the election date. PEW has done some research in 1999 and found that the model works great for the last few days before the election and is acceptable as far out as two weeks from the election date ... but as you get farther out ... all you are measuring with this method is the change in the Likely Voter sample ... not the change in the voters' preference for a candidate. So, look at the Registered Voter samples until we get closer to the election. Also note that there are correction methods one can apply to the LV Model so it can be used this far out from the election, but in the end what you are essentially doing is guessing what the sample should look like. As always, it's buyer beware! I only wish those in the media understood enough about the polling methods so they could stop confusing and deluding the public with "pseduo-scientific" numbers.