Reply #49: OK, to put the record straight:

[Febble]

I referred to TIA on DKos:

http://www.dailykos.com/story/2005/6/22/183230/208

so it is only fair that I comment here.

Here is where I agree with TIA:

I agree that his table showing the "12.22" exit poll responses shows a) Kerry winning and b) proportions of respondents who say they voted for Bush and Gore respectively that match the proportions that actually voted for Bush and Gore in 2000.

One possible interpretation therefore, is that the poll responses accurately reflected both the proportions of votes in 2000 and the proportions of votes in 2004.

I also agree with TIA that his table showing the "1.25" responses shows a) Kerry losing and b) proportions of respondents who say they voted for Bush and Gore that do not match the proportions who actually voted for Bush and Gore in 2000. Moreover, I agree with TIA that not only are the proportions wrong for the 2000 vote in the second table, the proportions are impossible, as it that implies more repeat Bush voters in 2004 than actually voted for Bush in 2000.

One possible interpretation of this, therefore, is that the responses have been reweighted to match the results (we know this was done), but that the fact that this reweighting makes the plausible number of Bush repeat voters implausible (and in fact, impossible) suggests that the vote count used for the reweighting must have been fraudulent.

This is a good argument. It makes mathematical sense. And I agree that it is suggestive that the Gore/Bush proportions in the early table actually match the vote-count Gore/Bush proportions, a point in favour of their veracity. So far so good.

Here is where I diverge from TIA's thinking:

Suppose there is a discrepancy between the count and the poll, as there was. And that this was either due to the fact that Kerry voters had been polled at, say 1.12 times the rate of Bush voters, or that a proportion of Kerry votes had been switched to Bush (I haven't worked out the proportion, but I know TIA has).

When the precinct results start to come in, the E-M computer program will start to discover this, and reweight the projections in line with the precinct results. I do not know how this is done, only that it is done, and I know this from the E-M FAQ:

http://www.exit-poll.net/faq.html

How are projections made?
Projections are based on models that use votes from three (3) different sources -- exit poll interviews with voters, vote returns as reported by election officials from the sample precincts, and tabulations of votes by county. The models make estimates from all these vote reports.

My guess is that the way it is done is by identifying particularly categories of responses (from a particular demographic group, say) that seem to be most divergent, and adjusting those. I imagine it works rather like your optimizer, except that the constraints on the output start being very loose and only become tighter as the results come in.

And we know that when the final adjustment had been made, the projection made a nonsense of the Gore/Bush responses.

One reason for this nonsense, could, I agree, be fraud. However, I see two alternative explanations that I find plausible, though I respect your right to disagree.

One is that no-one knows whether the weightings were correctly applied - my guess is that they would be optimized to constraints more binding (like gender) than past vote, as past vote is a notoriously unreliable measure. However, there are almost certainly multiple solutions to the possible weights that could be applied to match the poll data to the count data (and I agree - who could not? - that they were matched: they were designed to be matched) and some may give a more veridical match to the Gore/Bush proportions. So the final Gore/Bush proportions may simply be wrong, even though the estimate of the vote count now matches the vote count. Weighting solutions may exist that do a better job of approximating the actual Gore/Bush proportion.

But secondly, I do not believe the Gore/Bush raw responses are necessarily accurate. I find it plausible that the initial proportions of Gore/Bush respondents only coincidentally matched reality. I am not saying this is so, just that it could be so. In other words it is possible that the final table tells you something real about what people who voted for Bush in 2004 wanted to think about they way that they had voted in 2004. And that some Bush voters in 2004 wanted to think they had voted for Bush in 2000, whether they had in fact voted for him, or whether in fact they had voted at all. We know, from good research, that people do this.

However, I do not think this is necessarily the case. I think it may simply be an artefact of the way the weighting was done.

But it means that although I follow your mathematical reasoning, I cannot agree that Gore/Bush proportions in the final table are incontrovertible proof of fraud, particularly in the light of evidence that bias was greater where random sampling protocol was most likely to be compromised, for example, where interviewing rate - the rate at which voters are selected for interview - was low.

And I think to assert that it is incontrovertible proof is misleading. It is contingent on a notoriously unreliable measure of past voting behaviour from a minority subsample of the participants in the poll, and moreover, the final Bush/Gore proportions diverge from what is possible in exactly the way that one would expect - that people like to report have previously voting for the person they are currently voting for. I think it is known as the "consistency bias".

However, these are behavioural arguments, not mathematical. I do not dispute your math.

Please join the debate with me on DKos if you would like. I won't respond on DU again, for reasons I gave in my last post.

(edited for typo)