Over the last two weeks, we looked at the potential impact of "" voters missing from many poll samples and concerns expressed by some that the "likely voter" models used by pollsters might miss a flood of new younger voters. Today I want to take one more look at the so-called "Bradley Effect" and similar theories suggesting that polls may missing a hidden vote for among those who say they are undecided.
Much has been written about "The Effect" in recent weeks (including an op-ed by Lance Tarrance, the pollster for Tom Bradley's opponent, who argues that the effect did not occur even in the 1982 California governor's race for which it is named), but unfortunately, much of the recent speculation mischaracterizes the pattern pollsters often saw.
To be clear, the Bradley effect that pollsters discuss is not about the role that the race of the candidates may play in voters' choices. Rather, it is about a tendency for polls to underestimate the support for the white candidates in biracial contests.
As I reviewed in this space back in June, the effect was best described in a 1993 report [PDF] by Larry Hugick of Princeton Survey Research. He examined the final polls in 10 biracial elections in the 1980s and early 1990s and found that they accurately forecast the percentage of the vote received by the black candidate, but typically understated the share of the vote won by the white candidate.
"Only black candidates who broke the 50% level in the final poll were victorious," Hugick wrote 15 years ago. To "improve the accuracy of pre-election polls in biracial elections," he suggested a new method: "[Assign] all the black undecided vote to the black candidate and all the remaining undecided vote to the white candidate."
Today some -- including Republican consultant Bill Greener -- argue that we should apply Hugick's method to the current polls. "If you're a black candidate running against a white candidate, what you see is what you get," Greener wrote recently for Salon.com. "If you're not polling above 50 percent, you should be worried."
McCain pollster Bill McInturff appears to apply that logic in a memo released Wednesday. He argues that "the campaign" in the battleground states "is functionally tied" (emphasis added), in part since 's support is "dropping below 50 percent." Echoing Greener, he added, "I am becoming more and more convinced Senator Obama 'gets what he gets in the tracking.'"
So will we see a Bradley Effect next week? Much of the recent debate centers on whether the effect ever really existed (see the skeptical take by ABC's Gary Langer) or whether it existed and then disappeared 10 or 15 years ago (see the exhaustive report [PDF] by Harvard political scientist Daniel Hopkins).
Rather than trying to resolve this debate, I would prefer to focus on the evidence pollsters have collected this year. Is there any current empirical evidence suggesting that the effect observed 20 years ago might reappear next week?
Let's consider two important categories of evidence:
• Race of Interviewer. In those famous biracial contests 20 or more years ago, some pollsters reported seeing a "race of interviewer" effect. White respondents who talked to white interviewers were more supportive of white candidates than white respondents who talked to black interviewers. The best-documented example involved a survey of the 1989 Virginia governor's race described in a 1989 article [PDF] in Public Opinion Quarterly.
Are pollsters seeing any such differences now? Not that I could find.
The pollsters at ABC News examined more than 7,000 interviews conducted this year among white respondents and found that white interviewers obtained nearly identical vote preference results as black interviewers. Similarly, the pollsters at CBS News regularly check, and have seen no race-of-interviewer effect this year (though they saw a small effect involving surveys on the Obama-Clinton contest in late 2007).
At my request, Christopher Borick, director of the Muhlenberg College Institute of Public Opinion, checked the nearly 4,000 interviews he has conducted since Labor Day for the Morning Call Pennsylvania tracking survey. Here too, he finds nearly identical results when tabulating by race of interviewer.
• A Hidden "Undecided" Vote?: Typically, pollsters can say little about the undecided voters on their final surveys because the single-digit percentages yield (at best) only a few dozen respondents for analysis. However, the massive rolling-average national tracking surveys offer a unique opportunity to put larger-than-average samples of undecided voters under an analytical microscope.
The pollsters at Financial Dynamics were kind enough to share with Charles Franklin and me the raw, respondent-level data from more than 3,449 interviews conducted from Oct. 1 to Oct. 22 for the Diageo/Hotline poll.
We can learn two things from this data. First, roughly 6 percent of the respondents were initially undecided, but split almost evenly (47 percent for Obama, 53 percent for McCain, n=193) when pushed for how they "lean."
Second, Franklin constructed a statistical model to predict the vote choice among those who expressed a preference, then ran the model among the 267 respondents who were completely undecided. This process allows us to draw on every variable that seems predictive of vote preference -- including party identification, age, race, gender, education, frequency of church attendance and geographic region -- and use it to predict how the currently undecided voters will ultimately "break."
Franklin's finding? The model predicts that the totally undecided voters in this sample will split 54 percent for Obama and 46 percent for McCain (more details on Franklin's model here).
Of course, this result is far from the final word on this subject. The pollsters at Gallup, ABC/Washington Post, Research2000, Rasmussen Reports and three other organizations have been collecting similarly large pools of data that should enable comparable analyses. Hopefully, we will hear more from all of them soon.
As always, we need to watch the final surveys for any evidence of late shifts in opinion and voter preference. However, if we focus on the data before us, we find little or no evidence of a hidden vote for John McCain.