I'll take it in bits. This bit is boring, but it's wrong (in my view) so I'll put it on record:
Flaws in Liddle's Analysis included:
a) By looking at a simulation of an exaggerated ratio of Kerry-to-Bush voters' exit poll response (Liddle assumed a 2:1, K/B ratio giving an "alpha" of 2) 64 Liddle produced an “inverted U” shaped WPE/WPD graph that seemed asymmetric enough, so that it appeared that it could
approximate the E/M reported WPD outcomes. The asymmetry of the “inverted U” WPD curve – which gives a slightly larger WPD in high Republican precincts seemed to be consistent with the (much higher) WPD of high Bush precincts in the E/M data. The asymmetry is a mathematical result of linking an absolute difference (WPD) measure to a ratio measure (alpha=K/B). This “mathematical nit” cannot possibly explain the dramatic asymmetry in the E/M data. If an absolute difference “differential partisan response” measure is used (Bush - Kerry voter response rates), this small asymmetry disappears altogether. Only with highly magnified levels of alpha (such as a 2: 1 ratio of K:B representing alpha=2) will this small effect look significant.
The authors claim that a "mathematical nit" (whatever that means) cannot explain the "dramatic assymmetry in the E/M data". The aggregate mean WPEs for five arbitrarily drawn "partisanship" categories do indeed look "dramatic". They certainly alerted me when I saw them in the E-M report. However, when the full set of data points is seen in the scatterplot displayed by Mitofsky at AAPOR, a) the assymmetry is seen to be very much less dramatic, and b) also disappears when my proposed measure is substituted for WPE. My measure may not be the best measure of Precinct Level Discrepancy - indeed I think it is not, although I think it is very much better than WPE, which has two major problems. I am not sure what measure is being proposed in the USCV document, but "Bush minus Kerry response rates" begs a huge question - you can only compute Bush and Kerry response rates if you know the numbers of responses for each candidate in each precinct and the number of votes counted for each candidate in each precinct. Clearly USCV do not know this - it is why they advocate for access to the data. Moreover I would argue (though not here) that it is not a better measure than my measure, even if computed using the actual data.
The measure I proposed (and would still defend) was not the difference between the two response rates, but the ratio. There may be an argument for using a difference measure, and indeed Mark Lindeman and myself have spent an inordinate amount of time attempting to derive a measure that is unconfounded by vote-share. It has been an extraordinarily difficult (but fascinating, to geeks) task. However, I think Mark will support me when I say that our conclusion is that a ratio measure is probably the way to go. But I digress. The point is that neither can be produced without access to the actual data, and what Mitofsky did, using the actual data, was to compute a ratio measure. (And if by a difference between "Bush and Kerry response rates", USCV mean the difference between "K" and "B" values computed using their formula, a function of completion rates, then I would maintain that their difference measure is simply wrong.)
b) Liddle’s analysis at this point was based entirely on hypothetical simulations. In response, NEDA did some simulations – trying to match E/M WPD and response outcomes with constant alpha. NEDA's simulations showed that matching E/M reported mean and median WPD levels, and over-all response rates with constant alpha was highly improbable to impossible. In other words, NEDA’s simulations showed that even for the K:B ratio that worked best, the rBr hypothesis was not a likely explanation for the discrepancy between official results and exit poll results.
Well, before Mitofsky did his analysis on the actual data, all we had was simulations. I maintained that the aggregate category data was compatible with a flat linear regression line; NEDA maintained (as I understood it) that it was not. In fact, it turned out it was. There may indeed be non-linear relationships between bias and Bush's vote share, but there does not appear to be a linear relationship. Thus the hypothesis that fraud had a tendency to be greater in precincts in which Bush's share of the vote was greater, is not supported. This does nothing, however, to disprove fraud - it just says there wasn't a generalizable tendency for it to be more prevalent in precincts with high Bush counts. It could have been concentrated in Kerry strongholds.
c) Defenders of Liddle's analysis claimed that the only unusual thing about the scatter plot data are four high Bush outliers in the precincts that voted over 80% for Bush, that are not offset by any high Kerry outliers, and that these Bush outliers should be dropped. However, four outliers represents 10% of a sample of 40. If 10% of all of the high Bush precincts in the country were corrupted, this could represent a very serious problem. Moreover, even if the four outliers are removed, the other ways in which the E/M data are not consistent with constant mean bias hypothesis need to be addressed. The NEDA reports have shown, for example, that the E/M hypothesis is not consistent with the high median WPE in precincts with 80% or greater Bush vote, and with the very small mean and median WPE in precincts with an 80% of greater Kerry official vote, that would not be affected by removing a small number of outliers in high Bush vote precincts.
Well, I can't speak for my defenders (and a major problem with this paper is that this, like many similar statements, are not referenced) but I have certainly never suggested that ANY outliers should be dropped. Outliers are important. All I have suggested is that the four highly discrepant precincts in the high Bush category are no more nor less discrepant than many similar precincts in other categories. In other words there is nothing particularly generalizable about high Bush precincts - no reason to look in those rather than in any others for fraud. What I HAVE pointed out, and the paper also points out, is that there are a fair number of similarly discrepant precincts in which the shift is BLUE. And moreover that one of these just misses the arbitrary category boundary between mod and high Bush. In other words, had the Bush's vote count in that precinct been a single percentage point higher, the mean of the "high Bush" category would have been in line with the rest (except for the low Kerry category). Moreover, the mean of zero in the low Kerry category can be seen to arise not because there is are no discrepant precincts in that category, but because there are highly discrepant precincts in both directions. So my point (and Mitofsky's) is a very simple one: there is no generalizable tendency for red-shift to be greater where Bush's share of the vote is greater. That does not mean (sigh) that there is no fraud. It just means there is no special reason to look for fraud in high Bush precincts. I have NO idea why this finding has become so ridiculously controversial.
I'll post more stuff as I find it.
On edit: missed a bit. I'd just point out that a 2:1 ratio is not particularly exaggerated. Mitofsky's plot indicates that many, many precincts were discrepant at that level or greater.
Link to Mitofsky's plots are given in this DKos diary:
http://www.dailykos.com/story/2005/5/24/213011/565