Reply #27: Well, I'd

[Febble]

hate you to go against your better judgment, but I will ignore the hostility and attempt to address substance.

OK. We know there was a large discrepancy between the raw exit poll responses and the official count. We know, therefore, that the crosstabulations were subjected to an unusually large degree of post-stratification reweighting, using the incoming vote returns.

There are two possible theories: one is that there was massive fraud. The other is that the sample was more than usually biased.

So what predictions arise from these theories?

In both cases we would expect the reweighted cross-tabulations to give a poor indication of who had voted for whom, where.

However, in the first case we would expect the degree of discrepancy to be correlated with the degree to which Bush increased his vote-share. In the second case we would expect no correlation.

The second hypothesis is therefore the disambiguating hypothesis, and it turned out that there was no correlation between increase in Bush's vote share and the degree of discrepancy. This suggests that the discrepancy was largely due to a biased sample. This would therefore also suggest itself as the most likely cause of peculiar looking cross-tabulations. This could arise from non-uniformity in the geographical distribution of the bias. However, it could also arise from the weighting process itself.

Before the vote returns come in, the pollsters can adjust the cross-tabs for age, race and sex differences between the respondents and the non-respondents (those selected but not included). However, if the incoming vote returns are still different from the reweighted responses, all the pollsters can do is to upweight all Bush voters across all demographic groups, as there is no further information about the composition of the apparently missing Bush voters. This means that if the Bush voters who should have been in the polls and weren't have a different demographic composition than Bush voters that should have been in the polls and were, then the demographics will be skewed. As it seems they were.

What the OP suggests is that either fraud was concentrated in the urban areas, or that the sample biasing factors may have been. This particular fraud hypothesis therefore suggests non-uniform fraud. If the fraud was non-uniform, then it is very hard to see how the exit poll discrepancies were totally uncorrelated with increase in Bush's vote share.

Can you suggest how?

http://journals.democraticunderground.com/Febble/11