But what about the other questions where exit pollsters cannot simply make visual assessments of non-respodnents? For example, “For whom would you vote in the presidential election if it were held today?” We have no idea if the non-respondents are somehow different than actual respondents of a similar gender, race, or age.

To understand why this might be a problem, consider the following (very exaggerated) hypothetical. Assume that the initial exit polls find that the race is tied, and that Obama trails Romney by 30 points among Walker voters.

As the results come in, it is apparent that Walker leads by 10 points. So these Walker voters are weighted more heavily, and the president’s lead over Romney shrinks.

But what if the non-respondents were also disproportionately the most conservative voters, who don’t trust media folk? In other words, what if the president trails by 50 points within this category? Weighting these non-respondents the same as actual respondents on the presidential horse-race question will actually overstate the president’s standing — it won’t shrink enough.