In 1982, California gubernatorial candidate Tom Bradley went into Election Day leading in the polls. He lost in a squeaker. In 1989, New York City mayoral candidate David Dinkins had as much as an 18-point lead in the polls over his Republican opponent, Rudy Giuliani. On Election Day, he eked out a narrow victory. That same year, Virginia gubernatorial candidate Douglas Wilder enjoyed as much as a 15-point lead, only to win by a few thousand votes out of nearly 2 million cast.

All three candidates were black. These peculiar results have induced political observers to ask whether there is a systematic “Bradley Effect” — i.e., whether black candidates can expect to poll better among white voters than they perform at the ballot box. As Harvard University post-doctoral fellow Daniel Hopkins recently put it, “During a survey, respondents may be less likely to indicate voting intentions that are socially stigmatized, such as their unwillingness to support a black candidate.”

Hopkins’s analysis indicates that while there was a Bradley Effect through the 1980s, it had dissipated by the mid-’90s. Hopkins speculates that this was due to a decline in the “salience of racialized issues” like crime and welfare.

But that is not the end of the story. University of Washington professor Anthony Greenwald found evidence suggesting that the Bradley Effect was in play during the 2008 Democratic-primary battle between Hillary Clinton and Barack Obama. He noted that the variation in polling margins depended on the number of blacks living in a state: Polls in states with small black populations overestimated his performance, as the Bradley Effect predicts; but polls in states with large black populations actually underestimated Obama’s final performance, meaning that more people voted for Obama than the polls predicted.

So Greenwald found both a Bradley Effect and a Reverse Bradley Effect. His results are preliminary, but they highlight the new salience of this topic. Might the Bradley Effect be affecting Barack Obama, and might it manifest itself on Election Day 2008? Greenwald’s results suggest this is possible.

I designed a statistical model that builds upon Greenwald’s work. The model at­tempts to account for the difference be­tween Obama’s predicted vote margin and his actual vote margin.

Following Hopkins, the expectation is that if the Bradley Effect exists, it has to do with white respondents’ understanding of what is socially desirable. This could be related to a number of factors, and I included some in my analysis. First was the one Greenwald measured, the percentage of each state’s population that is black.

Socioeconomic status is another possibility; white respondents of a higher socioeconomic status might be more attuned to social cues — and therefore more likely to feel pressured to say they’re voting for the black candidate. I measured socioeconomic status in each state as the interaction of two variables: the percentage of whites with college degrees and the median white income in the state (both as of 2000).

The existence of a Bradley Effect might also be related to age. Older respondents might not be as attuned to social cues as younger respondents, and might be more likely to give an authentic answer even if it’s socially unacceptable. To measure this, I included as a variable the median age among whites in each state.

I added a variable indicating whether each state’s primary contest took place before, during, or after Super Tuesday. I also measured the effect of other factors: the geographical region of the contest (West, Midwest, Northeast, or South, as defined by the Census Bureau), the pollster who conducted the poll, and Obama’s final vote share (because polls often overestimate the winner’s share of the vote).

My results confirmed Greenwald’s. As the black proportion of a state’s population increases, the polls become more likely to underestimate Obama’s vote share. This is consistent with Greenwald’s Reverse Bradley Effect.

I also find a tight relationship between socioeconomic status and polling error. As socioeconomic status among whites increases, the polls are more likely to overestimate Obama’s share of the vote. This is consistent with the Bradley Effect.

It is unclear whether age has an effect; my analysis produces only inconclusive evidence. The evidence does suggest that the Bradley Effect might have been en­hanced after Super Tuesday, however.

These results suggest that the Bradley Effect (and the Reverse Bradley Effect) might have been in play during the primaries. Unfortunately, moving beyond this rather slim statement pushes us into a serious problem: I analyzed only this race, instead of determining how black candidates fared against whites in many races.

Therefore, while my results are consistent with the Bradley Effect, they could also result from any number of non-racial differences between this race’s candidates. They are consistent with an “Experience Effect”; it might be that white voters consistently shifted their opinions at the end of the campaign because, on reflection, they were more comfortable with Hillary Clin­ton and her experience than Barack Obama and his newness.

This could easily be related to age and socioeconomic status, given that Obama was more competitive with voters of a younger age and a higher socioeconomic status. Perhaps Clinton was able to peel off some of these voters at the last moment. This would be consistent with the typical finding in the exit polls that a significant number of respondents made up their minds in the last three days before the primary.

In theory, good polling data might enable us to decide between these two hypotheses (i.e., a Bradley Effect or an Experience Effect). But for now, I must conclude that we cannot know.

Both the Bradley Effect and the Ex­perience Effect, however, rest on the fact that people don’t always vote the way they tell pollsters they will. We will not know the country’s true opinion until Election Day.