Nate Silver used to run his own show, then went to work for The New York Times, then for ABC News, and is now back to running his own show, Silver Bulletin. He just wrote an interesting column about herding, complete with an illustrative photo of it in one of these two domains:

Silver analyzed the polling data for October in the seven swing states, ran some math on them, and concluded that there should be more outliers. Basically, in a sample with a mean (average) M and a standard deviation S, 95% of the data should fall between M-2S and M+2S. Five percent should be outliers, half below M-2S and half above M+2S. He didn't observe this in the data. When pollsters quote a margin of error of 4, this is (by convention) two standard deviations. In polling terms, 2.5% of the time the true mean will be outside the margin of error on the high side and 2.5% of the time it will be outside the margin of error on the low side. This has nothing to do with bad polling methodology or shy voters. It's just math and also applies when estimating the number of pages in a library book by sampling 1,000 books in a library. Sometimes you get a freaky sample and the mean is way off. This is normal and expected.

How could herding happen? If a polling team gets a result that is way different from other published results, the team may begin to question its methodology or sample and begin to wonder if it did something wrong. All polls are corrected to make sure the gender, age, partisan, race, education, income, and other distributions match the expected electorate. The team may then begin fiddling with the corrections to make their results look like everybody else's. It may not be malicious. The team members may just be afraid they did something wrong and want to fix it.