Reply #65: Ron, there is a serious problem here

[Febble]

Check your email, but I'll post it here too:

Your case seems to be that after subtracting from each precinct some value equivalent to the mean postulated "bias" for the nation (though why it should be uniform across the nation is something I've never understood, even though it has been repeatedly ascribed to me) you are left with some measure of bias that must be attributable to fraud because it is positively correlated with Kerry's vote share (WPE more redshifted where Bush's vote-share is greater) and negatively correlated with his exit poll share (WPE more redshifted where Kerry's exit poll share is greater).

Let's take this in two parts, and I'll take the second one first, as it's novel.

It should be pretty clear, from simple algebra, or even from considering the axes of the second plot, that any error in the poll - including sampling error - will induce a negative correlation between WPE and Kerry's vote share: because there is shared variance - error in the poll. Where, even simply by sampling error, you sample a greater proportion of Bush voters than the the proportion who vote, the WPE will be more negative, and Bush's exit poll share will be higher. Ditto for Kerry voters - if, purely by chance, you sample a greater proportion of Kerry voters than the proportion in the total, then your WPE will be more positive (or less negative) and Kerry's exit poll proportion will be higher.

So your second plot is entirely meaningless - it just says that where there were too many Bush voters in the exit poll, there tended to be more Bush voters in the exit poll than when there were too few Bush voters in the exit poll. See what I mean?

OK, first plot.

Well, as can be ascertained from the ESI data you use, the correlation is not significant. It is a trend only. Not only that, but, as we argued out long long months ago, WPE as a measure, tends to have a positive linear correlation with Kerry's vote share in the presence of any bias, even if the mean bias is zero. It was what that geeky paper was about. Yes, there is an U shaped relationship between any WPE for any given value of alpha, but the U is not symmetrical. Where alpha is greater than 1 (pro Kerry bias inthe poll, or pro Bush bias in the count), WPE will be more less negative at the high Kerry end of the plot than at the high Bush end. Where alpha is less than 1 (pro Bush bias in the poll or pro Kerry fraud in the count) WPE will be more positive at the high Kerry end than at the high Bush end. In both cases you will tend to get a positive correlation between WPE and Kerry's share of the vote. Only precincts with an alpha of 1 will have a flat correlation. There is not a single "U" but a distribution of assymmetrical Us, corresponding to given distribution of alpha. And that distribution looks like a tilted football. The tilt is what gives it the linear correlation.

Sorry if this is a bit geeky for anyone new to the series, but Ron knows what I mean, or if he doesn't he'd better re-read my paper.

In other words, both patterns are exactly what the math will produce in the presence of some kind of bias, whether in poll or count. They are NOT a fingerprint of one or the other. They are a fingerprint of WPE.

Well, one is the fingerprint of WPE, the other is the fingerprint of a tautology.

If that's the bottom line, I think we're in trouble.