Not to harp too much on this, but today Eugene Robinson makes a dubious argument that I've seen elsewhere:
Was it that voters told pollsters they intended to vote for African American candidates and then, in the privacy of the voting booth, chose white candidates instead? Not really. In each of these instances, preelection polls were quite accurate in predicting the black candidate's vote. What happened was that the polls greatly underestimated the vote for the white candidates. Unusually large numbers of self-described undecided voters ended up making the same decision.
Well, maybe. But, to point out the obvious, votes are fungible: given a certain number of votes for a candidate, it's impossible to tell whether those votes came from previously undecided voters or not. It's possible that 90 percent of undecided voters broke for Hillary, but it's not very likely. Robinson writes, "The exit poll...indicated that those last-minute deciders broke equally for Clinton and Obama--which pretty clearly was not the case." In reality, though, there's simply no way of telling whether it was the case or not: a wholesale breaking of undecided voters toward Clinton is indistinguishable from a splitting of undecided voters combined with a defection of some Obama voters to Clinton.
I tend to think the Bradley effect, while perhaps present in some measure, wasn't primarily responsible for the inaccuracy of New Hampshire polls. The nonresponse-bias explanation (blue-collar voters, who prefer Clinton for reasons unrelated to race, are more likely to refuse to answer pollsters' inquiries) strikes me as fairly plausible. But we can't rule out the possibility of voters lying to pollsters simply because Obama's final vote tally matched what the polls predicted. There's just no good scientific way of ascertaining what impact the Bradley effect might have had in any given race.