Pollster Peter Kellner argues the reverse Bradley effect could apply to the US election results as people who have never voted before turn out. 

In polling, as in so much else, the United States does not do things by halves, writes Peter Kellner. "Every day brings a slew of new polls, both national and at state level, on the race for the White House. As statistical theory would predict, they do not all agree on the numbers. 

Recent Obama polling leads range from two points to 14. What they have all agreed for the past six weeks is that Barack Obama leads not only nationally but in most of the battleground states that will determine the outcome.

However, some Democrats fear, and Republicans hope, that the polls may be wrong: that they may be overstating Obama’s support because some voters are telling pollsters that they will vote for him when, in fact, they will vote for John McCain or not vote at all. However, pollsters and pundits are divided on this, as I discovered last week on a visit to Palo Alto, California, home of YouGov-Polimetrix (whose clients for the US elections include the CBS television network) and Stanford University’s Hoover Institution (which houses some of the country’s best political scientists and which invited me over for the week).

Indeed, there is one school of thought that argues that the polls may be wrong for the opposite reason – they might be UNDERSTATING Obama’s support.

Here are the arguments put forward by the two opposing camps.

First, the reason why Obama’s lead might be overstated. It’s down to the ‘Bradley Effect’. This is named after Tom Bradley, who stood for Governor of California in 1982. He was ahead in the polls, but ended up losing. Some believe that he suffered because he was black – and some voters lied to pollsters, saying they would vote for him when because they were unwilling to admit their racism and their intention of voting for Bradley’s white opponent. (It’s important to note that the ‘Bradley effect’ is about polls, not general voting behaviour. There is no ‘Bradley effect’ where people vote against black candidates because of their race, and are happy to tell pollsters their true intentions.)

The ‘Bradley effect’ has been observed on a number of occasions since. Some observers attributed Obama’s defeat to Hillary Clinton in the Democratic primary nine months ago to the Bradley effect. If that is right, then the possibility that it might be distorting the polls in the current presidential race can’t be ruled out.

What, then, is the evidence? There are certainly some instances where pollsters have overstated the support for black candidates. There are, though two ‘buts’. The first is that Bradley himself may NOT have been a victim of the effect that bears his name. The daily tracking polls by Tarrance and Associates showed the race narrowing week by week during the campaign, from 12 points with a month to go, to one point in their final poll. Bradley’s narrow defeat could simply have been a continuation of that trend.

So why did the ‘Bradley effect’ gain such prominence? Because another pollster, Mervin Field, showed Bradley seven points ahead on the final weekend; an exit poll he conducted on election day also showed Bradley ahead. After the election he sought to explain his error. He said ‘race was a factor in the Bradley loss’. On the other hand, it could be that Bradley’s final campaign poll was not as good as Tarrance’s. As for the exit poll, Field failed to take account of postal votes, which went overwhelmingly to Bradley’s opponent, George Deukmejian. Bradley did, in fact, ‘win’ the votes cast on the day.

However, even if the ‘Bradley effect’ did not apply to Bradley’s own election, it does seem to have applied to a number of contests in the 1980s and early 1990s. A Harvard political scientist, Daniel Hopkins, has analysed 18 contests for the US Senate and state governorships where black candidates stood and polls were published. He finds that until 1996 the polls overstated support for black candidates by an average of three percentage points – but have not done so since then.

There are two possible reasons why things have changed. The first is that the notion of black Americans holding elective office is now more acceptable than it was. The second is that much of the ‘Bradley effect’ was linked to fears of crime, and the belief – conscious or subconscious – that black Americans were more likely than white Americans to be criminals. But since the early nineties, crime has been falling, especially in the big American cities, and the issue is far less of a concern than it was.

Against this, proponents of the Bradley effect point to the New Hampshire primary nine months ago, when both the pre-election polls and exit poll pointed, wrongly, to a clear Obama win. These polls can be explained in two ways. The first is that the campaign polls failed to detect a late swing to Hillary Clinton; the second is that there was a response bias in the exit polls, with well-educated (mainly Obama-supporting) voters willingly giving their views after they voted, while less well-educated (mainly Clinton-supporting) voters not taking part. Like campaign polls, exit polls correct for age and gender biases in response rates, as fieldworkers note the gender and guess the age of non-respondents; but, unlike campaign polls, exit polls DON’T correct for education (because it’s impossible to guess the level of education achieved by non-respondents).

Besides, apart from New Hampshire, there’s no evidence that the polls systematically overstated Obama’s support in the primaries. Indeed, according to an analysis by the website pollster.com, the polls actually UNDERSTATED Obama’s support in the primaries by an average of just over three points.

 
Interestingly, there seems to have been a regional pattern, with Obama’s support being systematically understated in states with a large black population, including the key swing states of Virginia and North Carolina, raising the suggestion that in some parts of America there may be a ‘reverse Bradley effect’, with polls failing fully to detect the enthusiasm for Obama among black Americans, especially younger electors – people who have generally not bothered to register or vote at all.

 This brings us to the arguments of those who think Obama might win by a bigger margin than the polls currently indicate. Because turnout in American elections has tended to be lower than in Europe, polls need to screen respondents, and count only those who they think will actually take the trouble to vote. It could be that the polls are screening out too many young and black Americans. This is partly because Obama’s well-resourced campaign has had some amazing successes with registration drives in the key battleground states. (Around one in four adults who are entitled to vote have traditionally not applied to join their state’s electoral register.) It’s hard for polls to keep up with the most up-to-date registration data, and some may be behind the curve in counting the proportion of pro-Obama voters in the registered electorate.

 In short, it is possible that the polls are underestimating Obama’s support, especially in the battleground states, with the result that Obama could win a comfortable majority of the national vote, and a large majority in the electoral college vote.

 And yet, and yet. All the evidence we have from past polls relate to state-level contests, for the US senate, or state governorships, of this year’s primaries. It’s possible that the polling dynamics of a national election are different. Some (though by no means all) of Obama’s supporters in the political science community remain edgy.

Personally, the poll I’ll be watching out for is the final state-by-state surveys conducted by our US company, YouGov-Polimetrix. They have already conducted two large-sample surveys for CBS’s network news. Their final figures will be broadcast by CBS and posted on the pollster.com website just before polling day. As online researchers with no interviewers for respondents to talk to, they should not be as prone as traditional telephone or face-to-face polls to any ‘Bradley effect’. And YouGov-Polimetrix is seeking to apply the latest registration data to their samples.

In just over two weeks we shall finally know the result of the election, and may have a better idea of whether the ‘Bradley effect’ can finally be laid to rest."

Peter Kellner is the president of YouGov.