Today, Gallup released its tracking data, which show three different results, all of which reflect tightening from yesterday:
Today's Result (Yesterday's Result)
RV O50, M43 (O51, M42)
LVE O52, M44 (O53, M43)
LVT O49, M46 (O51, M45)
RV means within the sample of 2785 registered voters (for today's results). LVE means an "expanded" likely voter sample of 2319. LVT is a "traditional" likely voter sample of 2160.
As Gallup has explained:
The "expanded" model determines likely voters based only on their current voting intention. This estimate would take into account higher turnout among groups of voters traditionally less likely to vote, but who may be inspired to vote this year. That model has generally produced results that closely match the registered voter figures . . . .
The "traditional" likely voter model, which Gallup has employed for past elections, factors in prior voting behavior as well as current voting intention. This has generally shown a closer contest, though with Obama still ahead.
Unfortunately, no public pollsters except the much-maligned Zogby provide their results to the decimal place. So when Gallup says Obama has a 50-43 lead among registered voters, that actually could be as small as 6.01 points (49.50 to 43.49) or as large as 7.99 points (50.49 to 42.50). So this makes it a little difficult to estimate exactly how many voters Gallup is weeding out with its "likely voter" screen. But they tell us their sample sizes, so we can make some estimates.
They start with "no fewer than 1,000 U.S. adults nationwide each day." Since it's a three-day tracking poll, we know the total sample of adults is at least 3000, and probably some number over that.
For today's three-day rolling sample, they weeded the 3000+ adults they reached down to 2785 who are actually registered to vote. Among that sample, as mentioned above, Obama leads 50-43. Ergo, between 1379 and 1406 of the registered voters in the sample must have picked Obama. Any figure below 1379 would round to 49%; any figure above 1406 would round to 51%. By the same calculation, between 1184 and 1211 supported McCain in the RV sample.
Looking at the "expanded" likely voter sample of 2319 (that is, those who informed Gallup that they are likely to vote on Nov. 4th), we can calculate that between 1195 and 1217 supported Obama and between 1009 and 1031 supported McCain.
Gallup further weeds the "expanded" likely voter sample down based on "past voting behavior." They don't say how they do this, but presumably they ask whether the person voted in the 2004 presidential election, also in 2000, etc. This "traditional" methodology cuts the sample down to 2160. Of those 2160, we know that between 1048 and 1069 picked Obama and between 983 and 1004 picked McCain.
I apologize if this is too math heavy, but I am building up to a real point here. In cutting the RV sample down to the LVE sample, Obama loses between 162 and 211 supporters; McCain loses between 153 and 202. In other words, of the total respondents cut out of the RV sample to get to the LVE sample, between 45 and 58% are Obama supporters. This means that Obama supporters are cut by this first likely voter screen in roughly the same proportion that they were in the RV sample to begin with (about 50%).
To make the second cut, however, from LVE to LVT, Gallup weeds out between 126 and 169 Obama supporters while weeding out no more than 48 and as few as 5 McCain supporters. Put another way, of the total respondents cut out of the LVE sample to get to the LVT sample, at least 72% and as many as 97% are Obama supporters. At the median, this second likely voter screen is falling very heavily (85%) on Obama supporters.
To Gallup's credit, they are at least employing the two models. My impression is that most of the other pollsters that have switched from RV to LV in the last few weeks (such as Diageo Hotline or Democracy Corps) only use the LVE methodology. Others like Research 2000 and Rasmussen have used likely voters from the start, which I suspect is based on their weightings to assess what they think the final electorate will look like (in terms of party and demographic ID), rather than specific questions designed to assess whether particular respondents are likely to turn out.
However, it seems to me that Gallup's LVT methodology is (to borrow a phrase from Barack) employing a hatchet rather than a scalpel.