Reply #3: From Elizabeth Liddle

[Febble]

As you've re-posted Kathy's comment, I'll re-post my response below.

But can I say first that I absolutely back Kathy's appeal for funds. The fact that we are even debating whether the exit polls or the vote count were wrong is a scandal. Elections must be both fair and seen to be fair. Also the voter suppression in Ohio must never happen again. When is Blackwell going to jail?

And Kathy is working tirelessly in Utah to prevent DREs being bought there. You must be able to audit your elections directly. Exit polls are far too crude, especially when response rates are so low.

Which takes me to: response rates:

The E-M report states that there was no significant difference in response rates between precinct partisanship categories, although they do not give details of the statistical test. However, it is true that the "differential response" hypothesis predicts that response rates should actually be lower in Bush strongholds than in Kerry strongholds - because allegedly, Bush voters don't like responding.

This is an important argument. However, statistically, a small difference is only detectable if you have enough statistical power. The predicted difference is small (in theory you should get 56% response from an all-Kerry precinct, and 50% in an all Bush district, so the maximum difference is 6 percentage points). Statistical power depends on two things - plenty of data and good signal-to-to noise ratio.

Here the numbers are small where it matters - in the extreme precincts where the differences should be maximised. There were only 90 "high Kerry" precincts and only 40 "high Bush precincts". So the data would have to be pretty clean. We know it was not, as the E-M report gives as a value called the "absolute" error, which, without going into too much detail, tells us how much noise there was in the data.

When I did my model, I compared the amount of noise in my model data with the amount of noise in the real data. I had less, and even so, statistical significance of the lower responses in my "high Bush" precincts was only borderline. I estimated that if I increased the noise so that it matched the real data, the difference would be lost in the noise. It only takes a couple of stray data points to lift the response rates at the "high Bush" end, or lower them at the "high Kerry end" to destroy the effect.

So I agree that the lack of evidence (and it is no more than that)of lower response rates in "high Bush" precincts than in "high Kerry" precincts is an argument worth considering. However, the way statistics works that what you test, statistically is a "null hypothesis". If you cannot show that an effect is unlikely to have occurred by chance you "retain the null". If you "retain the null", have not proved that the effect was not there. You simply have not proved that it was. It is the statistical equivalent of "presumption of innocence".

Unless E-M provide us with better data, I do not think we can reject the null hypothesis that over-estimates of Kerry's vote were randomly distributed amongst precincts of all degrees of partisanship.

This in itself, however, does not rule out fraud as a cause of the over-estimate. The same effect would result from under-counting of Kerry votes.

Bruce has done a great job with his simulation, and he is clearly at the top of his field. However, the point here is not the skill of the modeller, but the parameters of the model. With the greatest respect to Bruce (and we have been exchanging information very co-operatively over the modelling, for which I am extremely grateful to him), I simply disagree that his model disposes of widespread rBr, or for that matter, of widespread fraud. (Other forms of analysis may do the latter, but not this one).

This may be a matter of different philosophical approaches to data. Bruce is an engineer, my background is in the social sciences (psychology) although I also have done a bit of engineering. When data points are people, noise behaves badly. I think what we are seeing is noise, and I would argue that there is a sound statistical basis for my claim.

Lizzie