Not surprisingly, Alan Abramowitz took exception to my comments earlier today. In the interest of fairness, I am posting a portion of his email reply. He reacted to my assertion that ABC now weights its likely voter numbers by party, but not its registered voter samples:
Do they? Who knows and who knows by how much? They don’t report either their party id distribution or vote by party id for either RVs or LVs. Until and unless they do that, I don’t think you or I or the man in the moon can tell what they’re really doing…What I do know is, they are projecting an 8.5 point turnout advantage for registered Bush supporters vs. registered Kerry supporters right now, and that’s just not plausible. That’s based on the Post’s RV and LV numbers…Unfortunately, with the Post poll, the LV vs. RV issue and the party id weighting issue are getting confounded but they are really separate issues. We need to see what the separate effect of each is on their results but we can’t
based on the way they’re presenting those results.
An update: Tonight’s press release from ABC (not yet posted online) reports a 48% to 48% “dead heat” among likely voters and also says the following: “The partisan makeup in this poll is about even between Democrats and Republicans, 37-36 percent, with 24 percent independents.”
You are incorrect regarding an equal number of democrats and republicans in today’s WashPost trackimg Poll Overall Result:
48%B
48%D
1% N
1% neither
2% Don’t know.
WaPo must have more democrats than republican respondents because…
WaPo internals reveal
Bush leads by 79% among republicans
Kerry leads by 77% among democrats
Bush leads by 2% among republicans.
the only way these internals can average to a tie is if more Democrats than Republicans responded.
using an equal weighting of 37%,37%,26% I
would yield 48.9%B 47.6%K 1.4%I 1.3% Don’t Know which doesn’t match the results
the weighting I used that matches the results is
D38% R36% I26%
Since 2000 exit poll was 39%D/35%R/26%I this weighting doesn’t seem too outrageous but not correct to say equal R, D in poll
The discussion about weighting by party overlooks the fact that different pollsters ask the Party ID question differently. Whereas the widely used University of Michigan version is prefaced by “Generally speaking, do you consider yourself to be . . . ” the Gallup Poll and many others preface the question by the phrase “As of today, do you consider yourself to be . . . .”
There was a rather intensive discussion of this difference in the American Political Science Review a few years ago. My sense of that debate is that the UMich version of the question (“Generally speaking”) is more stable — and less subject to short term change — than the Gallup version (“As of today”).
So is a survey organization is going to use Party IDS to weight or adjust the sample, they are on more solid ground if they are using the UMich version — it’s a more stable anchor for adjusting the current survey to a standard. And if that anchor or reference is based on a series of surveys conducted relatively recently in the same population, it can be a pretty reliable proxy for what the underlying Party ID is today on that same population.
There are in fact times when it is reasonable to use a Party ID adjustment, such as when the immediate circumstances of a given survey might be likely to mobilize respondents differentially by Party ID. But one then has to think carefully about whether one is using an appropriate base and anchor to do the weighting.
A further comment about how Party ID is measured. I just completed an on-line Zogby poll. The Party ID question there was “In which party are you either registered to vote or do you consider yourself to be a member. . .”.
If someone lives in a state in which registration by party, this is a reasonable way to find out someone’s “committed,” i.e., usual Party ID and is probably akin to the UMich version. But if someone lives in a state where there is no registration by party, then what does it mean when someone “considers himself to be a member”? I think it’s more akin to current rather than usual Party ID. That is, it’s closer in meaning to the Gallup than to the UMich concept. If so, then it’s more problematic to weight by Party ID based on the Zogby measure (in states where people do not register by party).