On Election Day, though, Clinton won the state by more than two points.
On Super Tuesday, pre-election polls in Massachusetts and California similarly overestimated Obama's strength. But polls in Alabama and Georgia underestimated his eventual winning margins.
These results suggest that the "Bradley effect" -- a term used to describe the fact that pre-election polls often overestimate support for black candidates -- may still play a role in elections, according to a new analysis by University of Washington researchers Anthony Greenwald and Bethany Albertson. But, the researchers say, the Bradley effect is complicated, and there may be evidence for a "reverse Bradley effect" in which pre-election polls underestimate support for black candidates in states with large black populations, such as in Alabama and Georgia.
The Bradley effect is named for former Los Angeles Mayor Tom Bradley, who narrowly lost the 1982 race for California governor to his white Republican opponent, George Deukmejian. Polls had suggested he had a comfortable lead, and researchers theorized afterwards that while many white voters told pollsters that they would vote for Bradley, they behaved differently once they entered the voting booth.
Analysts found the same pattern in Virginia's gubernatorial 1989 race, between Douglas Wilder, who is black, and Marshall Coleman, who is white, and in many other races throughout the 1980s and 90s.
By the 2006 mid-term election, though, the effect seemed to have dissipated. In five statewide gubernatorial or senate races that featured black candidates competing against white ones, pre-election polls accurately mirrored election results, according to Scott Keeter, director of survey research for the Pew Research Center.
"So on the basis of this, I concluded that this phenomenon that had bedeviled us in the 1980s and 90s was finished," Keeter says.
But that might not be the case after all this election year. Greenwald, a social psychologist interested in people's implicit and unstated biases, and Albertson, a political scientist, are analyzing the results of the Democratic primaries with an eye on the Bradley effect.
Last week, they compared pre-election polls to actual results in the 13 states for which they could find multiple late pre-election polls, analyzing data in 11 Super Tuesday states plus New Hampshire and South Carolina. They found that in states where less than 10 percent of the population is black, including California, Arizona, Massachusetts and New Hampshire, polls over-predicted the Obama vote by as much as 12 percent. But in states where more than 25 percent of the population is black --Alabama, South Carolina and Georgia -- polls under-predicted Obama's lead by 15 percent or more.
Those numbers are intriguing, but they raise many more questions than they answer, the researchers and others say.
For starters, there are many possible explanations for the Bradley effect. The easiest, perhaps, is that people simply lie to pollsters -- but Greenwald says that he suspects that that's not the case.
Another, which he finds more plausible, is that the factors that influence someone responding to a pollster are not necessarily the same ones that influence them in the voting booth.
"This is different than lying, it's something [psychologists] call impression management," Greenwald says.
"Let's say you're white, and you pick the white candidate -- you understand that you're giving a response that could be construed as racist," he explains. "If you say that you're supporting the black candidate, you couldn't possibly be."
So a genuinely undecided voter -- of which there could be many in a primary election -- might be more likely, if white, to tell a pollster he planned to vote for Obama. And the same influences might work in reverse on a black voter, prompting him to say he planned to vote for the white candidate in order not to appear racist -- and thus create a "reverse Bradley effect."
Finally, Greenwald offers a third possibility. Much of his work focuses on people's subconscious biases, which he measures through a test called the implicit association test. The test ferrets out these biases by asking people to respond to positive and negative words and to other stimuli -- such as black or white faces, or particular politician's faces. Biases are revealed when people find that it takes longer to respond when positive words are linked with a certain stimuli (such as a black person's face, or a political candidate they dislike) than when negative words are paired with that stimuli.
The test has garnered attention over the past few years for revealing that many people who consider themselves entirely unbiased actually hold these biased associations.
Greenwald is using the Implicit Association Test to examine whether people's self-reported political opinions reflect their implicit opinions of the candidates. But, he cautions "we're not now prepared to say that the Bradley Effect -- in either direction -- is a manifestation of implicit prejudice."
Finally, poll analyst Scott Keeter believes that the most probable explanation for the apparent Bradley effect in this season's primaries may be based not on psychology at all, but on polling methods. It's notoriously more difficult for pollsters to predict primaries than general elections -- both turnout and results. People are in general less certain of whether they'll be voting, and who they'll be voting for.
With Obama capturing up to 80 percent of the black vote in some states, under-predicting black voter turnout by only a little could significantly skew the results, Keeter says: "A primary is just a more uncertain process."
Several states have yet to hold their primaries, and Greenwald and Albertson plan to continue their analysis throughout primary season. Tuesday's primaries in Washington, D.C., Maryland and Virginia will provide their next data points.