« Midwest Battleground State Polls from Chicago Tribune | Main | Kerry Wins Third Debate »

Kerry Still Ahead in the WP/ABC News Tracking Poll

Data on the WP website show that Kerry still leads by 2 among RVs in the WP/ABC tracking poll. Moreover, he now also leads by 1 among LVs, the first time this has happened in their tracking poll.

Speaking of LVs and RVs, it has come to my attention (thanks, Mark Blumenthal) that ABC News does now do some sort of party-weighting with their likely voter samples--but apparently not their RV samples. Just to confuse things even more the WP half of this partnership draws a different likely voter sample which may or may not be party-weighted and even reports different registered voter results than those reported by ABC--which may or may not reflect a different weighting approach on the part of the WP.

Clear? I thought so.

I feel somewhat less charitable toward Blumenthal for implying (well, saying) that my and Alan Abramowitz’ failure to embrace the ABC LV results shows deplorable hypocrisy on our parts. After all, he reasons, they’re party-weighting their LV results so shouldn’t we be delighted with these data and prefer it to their RV data?

But this doesn’t follow at all. I couldn’t have made my views clearer on how useless I think these sort of LV samples are for tracking changes in voter sentiment during a campaign and applying some sort of party-weight to them does not change my viewpoint. Indeed, RVs vs. LVs and party-weighting vs. not party-weighting are two separate issues and should not be collapsed, as Blumenthal, for some reason, insists on doing.

Now if ABC (or WP) were party-weighting their RV samples as well, I would be more sympathetic. And I suspect if they were, that would amplify, rather than reduce, the horse race differences we currently see between their weighted LV and unweighted RV samples. The reason for this is simple: the way ABC weights their likely voter sample is by averaging the unweighted party ID distribution in their LV sample with the average exit poll distribution from the last three elections (which is about +4D). Therefore, if they applied the same methodology to their RV sample it would likely increase the Democrats' party ID advantage in that sample, unless it was already over +4D, and therefore move the RV sample even farther toward Kerry.

Of course, I don't really know how much weighting would affect the RV results, since neither ABC nor WP release the party ID distributions or vote by party ID for their RV samples. (For that matter, these organizations don't release the analagous information about the LV samples either, with the exception of vote by party ID in the WP LV sample.)

In short, we lack enough information to sort out the the LV vs. RV issue and the party-weighting vs. not party-weighting issue--which, as I stressed, are two separate issues--and how they affect these data. But I welcome the release of this information, should these organizations see fit to do so.

Presumably, Blumenthal would agree with myself and Alan Abramowitz that this information should be released. I might add that we're both pleased to note that he has come out of the closet as a "third way" party-weighter, who has parted company with the "purists" who think party-weighting is a thought-crime of the highest order.

Welcome to the club, Mark. Now we can argue about how to do it, not whether to do it. And--who knows?--perhaps avoid casting aspersions on each other's intellectual integrity.

Comments

There should be more emphasis on raw data in polls, and less on selected questions by particular pollsters. A percentile conclusion is seized upon, without reference to obvious underlying defects.

If a story began "we polled the crowd at brunch at the country club Sunday and got these results" we would know it had faulty conclusions. A poll that shows 8% more Repubs than Dems is such a poll.

What a wonky post. It just appalls me that I'm wonkified enough myself to understand what you mean about weighting partisan affiliation in these polls.

During the summers I work as a data analyst for a market research firm and when I design field studies I have to struggle a lot with issues like whether I should weight the findings by demographic categories, which intuitively I feel will help our clients get a better picture of what's going on in the market.

But it always seems to turn out that I get better and more reliable results when I don't weight them. It's weird, it's counterintuitive, but mostly it's the laws of random selection in action. Random polling works better than weighted sampling. I do not want to admit this and I'm sure neither do the researchers who are selling these numbers to the candidates for their careers. But when you slant the numbers to fit a pre-fabbed demographic model, you end up slanting your findings away from what a truly random sampling will yield.

If I had a PhD I could probably tell you why this is. But I don't, so I'm putting it down to the usual trickery from Loki and the fire spirits.

Bucky, it probably has to do with the degree to which the variables used for weighting are correlated with the variables you are surveying and are offset by other countervailing but unrecognized factors.

For Party ID in a presidential election (especially this year), it's hard to find variables that are both more strongly correlated with support for a candidate than party ID and yet less strongly correlated with Party ID. In general, the reason to avoid weighting samples is because there may be several unknown variables that are important to the results and yet whose effect will be suppressed as a result of the weighting. I cannot begin to conceive of what such a demographic trait would be in this electorate.

this convoluted examination of LV, RV, and weighting is just about incomprehensible. i don't know what LV or RV mean, and you could at least have said what they stand for before launching into your analysis.