This mystery could be resolved if we could get them to release the raw data from the exit polls so we could see responses on a county or precinct level. But here's how a 52-48 Kerry lead in Ohio could have accurately forecast a Bush win 50-49.
Assume you have 3 precincts in Ohio, and over the last n elections, they have typically voted 52-48 Democrat, 50-50, and 52-48 Republican. Say that on the whole the state of Ohio votes 52-47 Republican. Say that 300 people are sampled from each precinct. Say that the results from the first precinct show 54-46 Democrat, second shows 52-48 Democrat, and third shows 50-50. That's 900 people so what you have is a 52-48 Democrat sample with a MOE of +/- 3.3%, right?
Wrong. More below.
