Readers of the USA Today were treated to the following headline splashed in huge type across the front page today: "Poll: Bush leads by 8 points". The headline referred to a 52-44 lead that the new Gallup poll found among likely voters. The accompanying story pointed out this was quite a turnaround compared to the Gallup poll of one week earlier, which had Kerry ahead by a point among likely voters (LVs).

A 9 point swing. That's pretty impressive. Of course, if you read the story closely, it does mention that Bush was ahead by just 3 points (49-46) in their registered voter (RV) sample. And, as it turns out--though this isn't mentioned in the story--that's a shift of only 3 points from a week ago, when Kerry and Bush were dead-even in the RV sample.

Much less impressive. Well, which is more believable? I think this a good time to review the basic case against Gallup's LV data:

Sampling likely voters is a technique Gallup developed to measure voter sentiment on the eve of an election and predict the outcome, not to track voter sentiment weeks and months before the actual election. There is simply no evidence, and no good reason to believe, that it works well for the latter purpose. In fact, the evidence and compelling arguments are on the other side: that the registered voters are the more reliable guage of voter sentiment during the course of the campaign.

Here’s why. Gallup decides who likely voters are based on 7 questions about their interest in voting, attention to the campaign and knowledge about how to vote (e.g., where their polling place is located). The interested/attentive/knowledgeable voters are designated “likely” and the rest are thrown out of the sample. But as a campaign progresses, the level of interest among voters tends to change, particularly among those with partisan inclinations whose interest level will rise when their party seems to be mobilized and doing well and fall when it is not. Because of this, partisans of the mobilized party (lately, Republicans) tend to be screened into the likely voter sample and partisans of the demobilized party (lately, Democrats) tend to get screened out. But tomorrow, of course, the Democrats could surge, in which case their partisans may be the ones over-represented in likely voter samples.

That suggests the uncomfortable possibility that observed changes in the sentiments of “likely voters” represent not actual changes in voter sentiment, but rather changes in the composition of likely voter samples as political enthusiasm waxes and wanes among the different parties’ supporters. And that is exactly what political scientists Robert Erikson, Costas Panagopoulos, and Christopher Wlezien find in their analysis of Gallup's 2000 RV/LV data in their forthcoming paper, “Likely (and Unlikely) Voters and the Assessment of Campaign Dynamics” in Public Opinion Quarterly: “shifts in voter classification as likely or unlikely account for more observed change in the preferences of likely voters than do actual changes in voters’ candidate preferences.”

That means that, instead of giving you a better picture of voter sentiment and how it is changing than conventional registered voter data, likely voter data give you a worse one since true changes in voter sentiment are swamped by changes in who is classified as a likely voter.

So: focus on the RV data, ignore the LV data. Indeed, in my view, it's time for Gallup to drop reporting these data altogether because they are highly likely to give an inaccurate picture of the state of the race and, by doing so--especially given the high profile of Gallup's polls and how they tend to drive media coverage--unfairly pump up one side of the race and demoralize the other. That doesn't seem acceptable to me. At a minimum, Gallup and other polling organzations that use similar approaches to defining likely voters should lead with their RV data and provide the LV data as a supplement, not the other way around.

That would make a difference in how the race is covered. Based on the Gallup LV result plus a couple of other recent LV results with fairly solid Bush leads (50-44 among Newsweek LVs and 50-46 among Washington Post (WP) LVs), other media outlets had stories today and yesterday on how Bush was surging and even breaking the critical 50 percent barrier in voter support (see, for example, this story by Ronald Brownstein in the Los Angeles Times which prominently features the three 50+ LV results).

But here are the same three polls, with RV results (all 3-way to match the data released by Gallup):

Gallup: 49-46
Newsweek: 48-46
WP: 48-47

So these three polls would then all have Bush under 50 (average: 48 percent) with only a 1-3 point lead. A very different picture and much dicier for Bush.

And probably much more accurate. Consider these other recent data, all of which paint a very different political picture than that implied by the Gallup LV data and the gaudy USA Today headline:

1. The Gallup RV data, while preferable to their LV data, may themselves be flawed. Their RV sample has a 2 point edge for the Republicans in party ID. If that sample is re-weighted to conform to the 2000 exit poll party ID distribution, Kerry leads Bush by 2 points, 49-47.

2. The latest Democracy Corps survey, conducted October 14-16, the exact same dates as the Gallup survey, has Kerry up by 3, 50-47. (And see this post for evidence on the superior accuracy of the DCorps survey, relative to Gallup, in the 2000 election.)

3. The Zogby and Rasmussen tracking polls have both closed by 4 points in the last two days, eliminating Bush's leads in these polls. Indeed, as Jerome Armstrong points out over at MyDD, if you include leaners in their trial heats (as most reported national results do), Kerry is now slightly ahead in both polls: 47.2-46.6 in Zogby and 49.5-47.7 in Rasmussen.

Doesn't sound like an 8 point Bush lead to me. Or much of a lead at all, for that matter.