With the release of the Kos/R2000 VA poll today, we now have three sets of data to work with. Unfortunately this poll didn't have the McCain/Obama voting question on it that the others did, but now we've got two sets of ages (SUSA used different age splits) and three sets of party ID breakdowns to work with.
Research2000's poll, as a good many of you know by now, showed Bob McDonnell (R) leading Creigh Deeds (D) by 51-43, with 6 undecided. This was as opposed to the other two recent polls, which gave Bob McDonnell leads of 15% (SUSA) and 14% (PPP).
My inclination is that the R2000 poll better controlled for younger voters and party ID. I'm not sure how, but they did. Controlling for younger voters may have led to knock-on effects in party ID or vice-versa (I think the age problem was a part of what was wrong in the other two polls), but at the very least, both age and party ID were different in R2000 vis-a-vis the others, likely causing the differing results. R2000's is probably more reliable for this reason.
For reference, here are the splits on party ID and age (with the CNN exit polls from 2008 thrown in for good measure):
R D I
CNN/06 39 36 26
CNN/08 33 39 27
SUSA 38 32 29
PPP 35 32 33
R2000 33 39 28
18-29 30-44 45-59 60+
CNN/06 12 27 37 24
CNN/08* 21 30 [38] [11]
SUSA* [17] [34] [29] [19]
PPP 9 21 49 21
R2000 17 32 37 14
*SUSA's age splits were: 18-34, 35-49, 50-64, and 65+. CNN also used 45-64 and 65+ in 2008, though they used the listed categories in 2006.
SUSA and PPP both had Republican-leaning samples (GOP edges of 6 and 3) vs. R2000's sample, which gave Dems a 6-point lead. Both had smaller samples of younger voters: SUSA gets 17% by adding in 5 more years (or 1/3 of the next age bracket up), which likely translates into about 6-7 of the 17% (this is a ballpark guess, as I do not have access to the underlying data).
Age variation, however, doesn't cover everything. In the 18-29 age group, PPP showed McDonnell beating Deeds by 10% (33-43). R2000 shows Deeds up 51-44. The McDonnell figure therefore would require a cross-section margin of error that one could probably fit a small European country in. Clearly something is still up with that part of the sample.
On party ID, the three polls offer us this:
SUSA
Party R D I
McD. 88 11 60
Deeds 7 86 35
Undec. 5 3 5
PPP
Party R D I
McD. 94 4 52
Deeds 2 80 33
Undec. 4 16 16
R2000
Party R D I
McD. 88 17 55
Deeds 6 77 40
Undec. 6 6 5
All three tell the same general story: McDonnell wins the vast majority of Republicans, more crossovers than Deeds does, and wins Independents by some margin above 15 points. However, SUSA and PPP give Independents to McDonnell by 19 and 25 points while R2000 gives them to him by 15. Also, interestingly enough R2000 also gives McDonnell more Democrats than does anyone else.
My guess is that R2000 is in the right neighborhood. The general narratives are all about the same...Deeds is losing independents and is behind because of this, and he's losing a bit more of his party than McDonnell is as well. The question is on margins, and I think that R2000 is more on the ball both because (as I've said before and I can't say enough):
A) They have more younger voters (at 17% vs. 9% and probably about 11%)
and
B) They have party ID more in line with what we saw in 2008, and age more in line.
In sum, McDonnell's lead in the R2000 poll is about what I figured on it being before. Their splits are more in line with what I felt was reasonable, cleaning up the "mess" in the other polls. I won't run the figures now, but if one altered all three polls' party IDs to be in line with what R2000 generated, those big illusory leads would vanish in a big hurry.
Update: In response to concerns, I've added the 2006 exits as well, for comparison. Age-wise, I do expect a younger electorate than 2006 but older than 2008 (both because the new registrants will have aged a year and because I do expect some of them to fall off). I will also say that PPP heavily oversampled the 45-59 demographic while probably overshooting slightly the 60+ demo as well. One thing that is not explained in PPP's data is why younger voters lean towards McDonnell by such a margin (10 points). My guess is some sort of sampling bias.