Reply #9: Probably [View All]

[Febble]

What is required, as Ron Baiman has frequently pointed out, is a multiple regression analysis, where some measure of the discrepancy (traditionally WPE, but better, I suggest, something like my function which is the log of the ratio between the sampling rates for each candidate) is regressed on factors likely to be associated with sampling bias. If this accounts for all the bias, then fraud is not ruled out, but can probably ruled out as a major causal factor in the discrepancy.

If it doesn't then it leaves some question.

One problem is that we are short of good fraud hypotheses - there are plenty of factors we know might affect sampling bias, and we know in which precincts these factors were present (from the E-M report) - what we don't know is whether they added up to a sufficient explanation for the total bias, because no statistical details were given in the report.

But by definition, we don't know where fraud occurred. If we could find a "fingerprint" to look for, that would help, and that is where these simulations (Bruce's; mine, I would argue) help. Where I don't personally find the Baiman Dopp analysis helpful is that it doesn't have an a priori fraud hypothesis - it seems rather to look at a rather meagre set of data points (the ones in the E-M report) and infer a fraudulent pattern from them. Which was sort of fair enough, when it was all we had. But we know have more, and it doesn't really support their inference.

Bruce's simulation is terrific. It doesn't prove fraud, nor does it prove that the explanation for the discrepancy is differential response rates. But it does suggest the kinds of pattern we might want to look for as a finger print of fraud, given what we know about the way that such a fingerprint may be hidden. And it does look as though it is hidden, if it is there at all.

Whether the fingerprint is really there or not, we may never know. Just because the finger-print is missing doesn't mean the crime didn't happen. It just means that fingerprints aren't going to be very useful for the prosecution case.

But it may also be worth saying that the better the non-fraud variables account for the data, the less chance there is of finding wiggle room for major fraud in the exit polls.