Second, the networks almost never “call” truly competitive races on exit poll results alone. The decision desk analysts require very high statistical confidence (at least 99.5 percent) before they will consider calling a winner (the ordinary “margin of error” on pre-election polls typically uses a 95 percent confidence level). They usually only achieve that confidence for relatively close races after the exit pollsters obtain the actual vote results from the randomly selected precincts at which interviews were completed (and from other larger random samples of precincts) and combine all of the data into some very sophisticated statistical models.
Even then, if the models project that the leading candidates are separated by just a few percentage points, as pre-election polls suggest they will be in all of the key battleground states, the networks will usually wait until nearly all votes are counted to project a winner.
Third, the initial results of the exit poll interviews have had frequent problems with non-response bias, a consistent discrepancy favoring the Democrats that has appeared to some degree in every presidential election since 1988. Usually the bias is small, but in 2004 it was just big enough to convince millions of Americans who saw the leaked results on the Internet that John Kerry would defeat George W. Bush. It didn’t work out that way.