On 'Likely Voters': Did screening for likely voters lead to accurate polling?

President Obama, campaigning in Colorado on Nov. 1, 2012.

In the weeks prior to the presidential election, I wrote of the rather formidable gap between how President Obama fared in polls where respondents were classified as "likely voters" and how he performed in polls where all registered voters were included.

At the time, I went back to data from the 2004 and 2008 presidential elections, and I discovered the following:

Of the 50 state presidential polls conducting during the final month of the 2004 and 2008 presidential campaigns, the RV result was closer to the final outcome than the LV result in fully half of them. In just 38 percent of them was the LV screen closer to the final outcome than the RV screen. In six of the polls, incidentally, there was no difference between the RV/LV results in a poll.

While there are still hundreds of thousands (if not millions) of ballots remaining to be counted in the 2012 presidential election, we have enough data to draw some pretty firm conclusions on the merits of screening for likely voters.

Exactly a half dozen pollsters, looking at the Obama versus Romney national numbers, bothered to differentiate their "likely voter" screens from their results among all registered voters. Here is how they broke down:

• Gallup Tracking, 11/5/2012: Romney +1 (LV); Obama +3 (RV)
• Ipsos/Reuters Tracking, 11/5/2012: Obama +2 (LV); Obama +4 (RV)
• CNN/Opinion Research, 11/4/2012: Obama +0 (LV); Obama +2 (RV)
• Fox News, 10/31/2012: Obama +0 (LV); Obama +1 (RV)
• YouGov, 10/31/2012: Obama +1 (LV); Obama +2 (RV)
• CBS/New York Times, 10/30/2012: Obama +1 (LV); Obama +5 (RV)

In other words, given the current national margin (Obama +3.2, which is likely to expand with additional provisional and absentee ballots from Obama-friendly states like California and New York), the "likely voter" screen was further removed from the final result than the registered voters screen in all six cases.

Now, in fairness, virtually every national poll in the final week underestimated the eventual Obama margin of victory. Therefore, if likely voter screens create more Republican-leaning results (which has long been a given in polling analysis), if everyone underestimated Obama, it stands to reason that the RV screens would perform better.

That said, some of those likely voter screens were quite a ways off. Most notably, the crew at Gallup has had quite a bit of criticism heaped upon them for missing the fairway, especially since this is the second cycle in a row that their likely voter screen has erred badly in favor of the GOP.

As Steven Shepard, the resident polling analyst over at National Journal, pointed out in an excellent campaign post-mortem:

On Oct. 26, Gallup released a demographic analysis of those respondents classified as likely voters in its daily tracking poll between Oct. 1 and Oct. 24. Of those voters, 78 percent were classified as non-Hispanic white, significantly more than the percentage of white voters measured by exit pollsters, 72 percent.

Four years ago, Gallup also found an electorate that was 78 percent white, an overestimation from the 74 or 75 percent recorded by exit polls. But this year's disparity is of a greater magnitude.

Gallup has also underrepresented younger respondents in its measures of likely voters. Over the first 24 days of October, 13 percent of Gallup's likely-voter sample was younger than age 30. Exit polls show these younger voters made up 19 percent of the national electorate.

Democratic pollster Mark Mellman believes this kind of likely-voter screen is counterproductive. Speaking at the event at the Gallup building, Mellman said that pollsters and the media have "overfetishized this whole notion of likely voters."

"We should not be concerned about finding likely voters," said Mellman. "We should be concerned about simulating the likely electorate."

Mellman has a salient point. Modeling a potential electorate is always an act of guesswork, but a quick examination of presidential exit poll data should've exposed the folly of the electorate that Gallup was expecting to see fill the polls on November 6th.

In 1992, the electorate was 87 percent white. In 1996, the electorate was 83 percent white. In 2000, the electorate was 81 percent white. In 2004, the electorate was 77 percent white. And, finally, in 2008, the electorate was 74 percent white.

You don't need a Master's degree in statistics to see a certain trajectory in those numbers. And, yet, somehow, Gallup was presuming not merely a stoppage of that trajectory, but a reversal of it.

(In the name of complete disclosure, there are alternate theories on the issue of nonwhite populations in exit polling. Read Harry Enten's excellent post-election article in the Guardian on the matter.)

To say that a marked increase in the white share of the electorate seemed unlikely, especially when between 2008 and 2012, there were four million new registered Latino voters, is perhaps the understatement of the year. And it is why Gallup is taking a heap of criticism, quite a bit of which would appear to be justified.