Poll Potheads

Pollsters--along with nearly everyone else on earth--failed to predict the result of the New Hampshire Democratic primary. According to Real Clear Politics, they estimated that Barack Obama would defeat Hillary Clinton by an average of eight percent. She won by three, and eleven percent is an awful lot for pollsters to be wrong by--well beyond the margin of error. In the scramble to explain how this could have happened, several writers, including Andrew Sullivan and Washington Post columnist Eugene Robinson, have suggested that the discrepancy might be the result of what is called the “Bradley effect.”

In the 1982 California gubernatorial election, Los Angeles Mayor Tom Bradley, an African-American, enjoyed a comfortable lead in the polls against his white Republican opponent George Deukmejian going into the election--but Deukmejian won. It turned out that a large number of white voters had either lied to the pollsters about their willingness to back Bradley, or had changed their mind on polling day and decided to vote for Deukmejian. Could it be that voters lied to pollsters this time, too?

Andrew Kohut, a pollster for the Pew Research Center, thinks so. Kohut is the eminence grise among pollsters. His interpretation was published in The New York Times. Suffice it to say, it carried a lot of weight. Kohut’s argument goes as follows: Clinton did much better in the final count than Obama among poorer, less educated voters. These voters “have more unfavorable views of blacks” than wealthier, more educated voters. Kohut doesn’t accuse these voters of lying. Instead, he argues that the voters who have unfavorable views of blacks tend to be underrepresented in polling samples, because they “do not respond” to pollsters--thus accounting for the inaccurate readings of support for Clinton and Obama.

This is an incendiary argument. Not only does it purport to explain why the pollsters got the results wrong, but it also implies that Clinton’s success in New Hampshire can largely be attributed to the racism of low-income, less educated whites. But Kohut’s evidence seems flimsy at best.

Kohut provides no data--none at all--to back up his contention that New Hampshire’s lower-income, less educated whites have a more unfavorable view of blacks than their wealthier, more educated counterparts. I think he is simply inferring from national studies or studies that were conducted elsewhere, but he doesn’t say. Yet New Hampshire is not Georgia or Mississippi, states with long histories of racial problems, nor is it the polarized New York City of 1989, where Kohut claims he encountered the Bradley effect. This kind of explosive claim deserves to have been backed up by some kind of evidence. I certainly don’t know of any.

Still, let’s assume for the sake of argument that Kohut is right about New Hampshire’s downscale white Democrats. Were they really underrepresented in the polls? If they were, then one would expect that the pre-election polls would overstate Obama’s support and understate Clinton’s among downscale white Democrats--and that the more extensive post-election polls would show a dramatic rise in Clinton’s support, and a fall in Obama’s among these voters.

But nothing of the kind took place. The pre-primary polls sponsored by CNN and WMUR and conducted by the University of New Hampshire Survey Center--polls, I might add, that are considered among the most reliable in the state--published a breakdown of Democratic primary voters by educational level. One can compare that survey, taken on the Sunday before the election, with the exit polls to determine whether a dramatic change took place in the white working class vote. Here are the numbers:

As you can see, Obama’s support among New Hampshire Democrats without college degrees slightly increased from the pre-election poll to the exit poll. There was an increase in Clinton’s support among voters with only a high school degree or less, but there was also a slightly smaller increase among Obama voters from this group. Where Clinton dramatically picked up support from the pre-election poll to the final poll was among voters with college degrees and higher. That’s exactly the opposite of what Kohut’s version of the Bradley effect would predict.

Here we come to a far more likely explanation (albeit one of several) for why the pre-election polls were wrong: Women--and college-educated women in particular-- shifted to Clinton. The polls show that while four percent of men switched over to Clinton, 12 percent of women did. Since there was very little change among voters without college degrees, one must infer that the bulk of the change came from women voters with college degrees. And it seems unlikely that racism can help to explain that, since well-educated women may be the least racist sub-group in American society.

Robinson seems to think that the exit polls themselves failed to account for Clinton’s margin over Obama. Robinson says, “The exit poll done for the television networks indicated that nearly four out of 10 Democratic voters made their decisions in the last three days before the primary. But the exit poll also indicated that those last-minute deciders broke equally for Clinton and Obama--which pretty clearly was not the case.” Robinson is, however, looking at the wrong data. Three days is not a good figure.

The change would have come in the last two days, after the debate and Clinton’s crying episode. There is no exit poll figure for these days precisely, but the exit polls did ask voters whether they made up their mind during the last day--17 percent of voters did, and they went for Clinton by a margin of 39 to 36 percent. That still probably understates the change, but if the networks had asked for the last two days, they might have gotten a figure that fully explained the final result.

More evidence for a last-minute shift to Clinton--and for voter volatility rather than deceit as the culprit--comes from the Rasmussen and Zogby polls, both of which reported seeing a turn toward Clinton on the Monday before the vote. The pollsters didn’t poll late enough to record these changes fully, and as Zogby admitted afterwards, they folded the Monday results into the results from the previous two days, which obscured the final shift.

But an additional factor may have been at work. Some of the polls seem to have significantly underrepresented the women’s vote. For instance, the Suffolk University/WHDH poll, which surveyed voters on Sunday and Monday and came the closest to predicting the final result, estimated a 39 to 34 percent Obama win by working off the assumption that 53 percent of primary voters would be women. According to exit polls, though, the breakdown was 57 percent women to 43 percent men. If you rejigger the Suffolk/WHDH poll to take into account the real mix of women to men, what you get is something like 38 percent Obama and 35 percent Clinton--which is within the 4.38 percent margin of error for the final results.

That may not be the reason why other polls got the result so wrong, but the under-representation of woman voters, coupled with the volatility of the electorate (as evidenced by the last minute shift of college-educated women voters), is a far more plausible hypothesis than the one that Kohut, Sullivan, and Robinson provide. This is not to say that there weren’t people who did not vote for Obama because he is black. But, clearly, a hidden racist vote is neither an explanation for Clinton’s victory nor the pollsters’ error in predicting it. A closer reading of the evidence also has the benefit of not accusing half of New Hampshire Democratic primary voters of being racists.

John B. Judis is a senior editor at The New Republic and a visiting fellow at the Carnegie Endowment for International Peace.