Reply #25: Because [View All]

it is not necessarily constant from election to election, although it tends to go in the same direction each year (and in 1992 was nearly as great as in 2004). What seems clear is that while there is an underlying trend from year to year for Democratic voters to be better represented in the exit poll than Republican voters (I say this knowing the risk of being flamed, but contrary to frequent claims, there is actually good evidence for this), where polling protocol is good, the effect is minimized. In other words, given good conditions, training, close adherence to random sampling protocol etc, Republicans have less opportunity to evade the pollsters, and/or Democrats have less opportunity to increase their chances of selection. The trouble is that the factors likely to provide opportunities for the underlying phenomenon to affect the sampling are only known after the election.

In fact, of course, you have put your finger on what actually happens, which is that the pollsters expect bias, and use the vote-count itself as the "overweight". As the vote-count comes in, they spot the what they assume is bias, and weight accordingly. The trouble here is that the procedure assumes the vote-count is correct! But this is at the root of a lot of misunderstanding of the exit poll. It is not designed as a check on the count - it is designed to project the counted result. So the pollsters, while they try to get the interviewing right, are not too worried about bias (unlike pre-election telephone interviewers who are trained interviewers and collect a much smaller sample over a longer period, the election day interviewers are casual hires with a short telephone training) because the intention is always to weight any observed bias by the vote-count itself. What they are really worried about is projecting a result that results in a "call" for a state that turns out to be "incorrect" (i.e. not in accord with the final counted result).

If you wanted to design an exit poll as a check the count (which would be prohibitively expensive, which is why they don't do it) you would have to have teams of trained interviewers rigidly sticking to Nth voter protocol throughout election day, and even then your results would be vulnerable to non-response bias (more refusals from voters for one candidate than the other). Ironically, ploys to improve response rate (e.g. giving away free folders) don't necessarily eliminate bias and may even increase it.


ETA: results are also weighted for the observed age/race/sex of non-responders, but this won't help with selection bias, only with non-response bias. Pre-election polls are also used in the weightings - if exit poll tallies are out of whack with pre-election polls, that is another indication that bias may be present in the exit poll sample, and they can be weighted accordingly.