page 3
http://www.electionscience.org/Members/stevenhertzberg/report.2005-07-19.7420722886/report_contents_file/Regarding your question: sort of.
It is actually quite tricky to state a statistical hypothesis in words, although it is actually stated in the terms of the analysis.
You wrote:
"If fraud is responsible for the exit poll discrepancy, then in precincts where, in 2004, Bush improved the most on his share of the 2000 vote, the exit poll discrepcancy should be the greatest. In precinct where he his share of the vote declined vs. 2000, the redshift should be lowest, or even of negative value."
Which is more or less right.
But the point about Analysis of Variance is that you cannot disprove it by a single counter-example. There will be other factors affecting both shift and discrepancy, and sometimes Bush will do badly because he is deeply unpopular - but fraud will rescue him from doing quite so badly. In others Bush will do well for other reasons - whether there is fraud or not.
This is why we partition the variance. The idea is to determine the proportion of variance in Swing that is
shared with Redshift. If a lot is shared, it will look as though there is a common factor - fraud. If most of the variance is unshared, then it looks as though there is not a common factor. The statistical test (one of them) is the F ratio - between the variance shared and the variance unshared. In both analyses, the variance unshared was far greater than the variance shared, rendering the correlation between the two "insignificant". In the presence of an "insignifant" correlation we cannot say whether there was a common factor or not.
In the ESI study the number of precincts was small - so the upper confidence limit on the correlation might allow for a bit of a relationship - there might be a "true" relationship of some size, that might be due to fraud.
However in Mitofsky's analysis the full 1250 precincts were included, making the confidence limits on the correlation pretty tight. It really looks as though there was very little shared variance.
Thus far, I maintain both studies are a perfectly logical test of the hypothesis that fraud was responsible for both variance in swing and variance in redshift.
However, it rests on certain assumptions, which Kathy correctly challenges, and which I have also challenged.
If fraud was targetted in particular precincts, it might affect the Bush's vote share but not show up as a correlation. However, this would also limit its explanatory power regarding red-shift. I'm trying to figure out how much.