Likely Voter Filters
Kevin Drum makes an excellent find. NBC said something about their likely voter filter ! "Among likely voters — identified by their past voting history and their high level of interest in the November midterms."
As far as I know, until now, pollsters other than Gallup haven't explained their likely voter filter at all.
So hmm "past voting behavior" could be problematic. The last election was very unusual. Also, I think that past voting behavior will underweight early middle age people. To make my thought clear, consider a simplified imaginary USA where all young people don't vote and all old people vote. In that USA' there is some age when people vote for the first time. Those people will be voters not clasified as likely voters. My sense is that that's been a very good age group for republicans (see Atrios "Talking bout my generation") but isn't anymore. Oh no I don't have the link but about a week ago, someone at DailyKos noted that, looking at polls where both likely voter results and registered voter results were reported, the registered voter results were closer to the final outcome in more than half of cases in 2006 and 2008 (the only years checked). Was this because they were wave years ? Or was it because the usefulness of likely voter filters was based on generation X ? Likely voter filters haven't been around long enough to tell.
Now how about interest. Has anyone checked the relationship between stated interest in September and actual voting behavior ? Obviously interest typically rises during October -- many people only tune in then. I don't see how one can check (I recall a post by Kevin Drum which notes that according to polls 70% of eligible voters voted in the last election).
Something that is possible is checking the reliability of likely voter polls in September. Reliability is only checked using the very last poll. The idea is that people really change their mind so it isn't the pollster's fault if they don't vote in November the way they said they would vote if the election were held the day after some day in September. So ? Statistics are useful even when there is a disturbance term. But even Nate Silver himself just accepts likely voter filters. Why ?
I think it is clear that the Gallup filter is biased against the young until just before the election. One of their 7 questions is "do you know where your polling place is ?" They assess the filter using data on voting intentions of people who do and don't know in late October. They use the filter in late September. Clearly not knowing in late October means much more than not knowing in late September. Notably in the past two presidentials, Gallup gave better results for the Republican in September than other polls (search Gallup anomaly) then their last poll pretty much nailed it. What happens when you regress Dem share in the last Gallup poll on Dem Share in the first Gallup likely voter poll and a constant ? I must predict a positive constant which would be significant if one collected enough data.