Reply #21: OK

[Febble]

1. What I liked.

I will refer directly to Dan Tokaji's article, because I liked the same things. He was more measured than I was, because, I confess, I was put off by the fact that the first numbered item in the article was the section on exit polls, of which more later. But the fact that there was much to like was partly what made me so cross. I just think the article could have been so much more powerful had it not started by citing the most easily refutable part of the case.

2. What should he have emphasised.

I think he should have emphasised the stuff that I liked (duh...) But more importantly I think he should have emphasized that that that stuff about systemic disenfranchisement, of Blackwell's conflicts of interest, and all that Went Wrong In Ohio, is an outrage whether or not it made an iota of difference to the result If he wanted to quantify it, I think he should have been more circumspect - provided a range of possibilities perhaps, and I think he should certainly have checked his numbers, but mostly I think he should have emphasized that the point is not whether it affected the result but that it happened at all. Even if every disenfranchised voter was a Bush voter, it matters. It matters because democracy matters. Sheesh, I could almost rewrite the thing myself, if only my name was Kennedy.

3. Lou Harris and polls.

I can't comment directly as the cite isn't referenced, but I would like to know what he is referring to as "sticking out like a sore thumb", and what polls he is referring to. It may be that there are anomalies at county level, but this wouldn't show up in the exit polls, as fewer than one precinct per county on average is selected. So to address your more general comments about the exit poll: the sampling tries to be random (actually the longest selection interval was 10, the shortest was 1, i.e. every voter). However, even if strict random sampling was achieved, the sample would only be random if those who refused (or who were missed) were not "drawn from a different population" as we say in statistics from those who participated. If they were, the polls would show the well researched phenomenon of "non-response bias". Moreover, we know from the E-M report that redshift was greater where the selection interval was longer. The longer the interval, the easier it is to make an error, which won't matter if the errors are random. But the more likely a random error is, the easier it is for non-random selection to happen as well. If, for example, an interviewer is struggling for willing respondents, if the 10th voter looks surly, and the 11th looks friendly, and they are more or less abreast, then it is not difficult to see how you might end up with a sample more heavily weighted with friendly voters, so in that sense, it can induce another form of "non-response" bias, although "selection bias" is a better term.

As for the "mixing" with the vote count. There is simply no reason to infer preparation for fraud from this. If you look on the E-M website, and the FAQ, you will see exactly how the projections were always going to be made, which is the way they are made every time, and that page was up well before the election. It is simply not a secret. I think the misunderstanding here arises from confusion the meaning of "projections". All polls try to measure something that will tell them something else. By measuring how people think they will vote before an election, the pollsters, after a fair bit of data manipulation, will "project" the future election result. Therefore, as well as a margin of error due to sampling variance, there is also error due to the assumptions made by the pollsters about, for example, how likely each respondent is to vote. The exit polls are absolved from that assumption, but they still need to weight for other things likely to affect their "projections" including bias in the poll (non-respone bias, for example). This is done initially by observing the age, race and sex of those who were selected to be in the poll but who were not (they were missed because the interviewer was busy, or the voter refused), and weighting the sample for any mismatch in the age, race and sex proportions between the sample of voters interviewed, and the non-responders. However, while the pollsters might be aware that there were, for example, a smaller ratio of men to women in the respondent sample than in the refuser sample, and can therefore weight by the voting preferences of men in the sample, there is no way the pollster can no the voting preferences of the men not in the sample. It may be Bush-voting men are more likely to refuse than Kerry voting men. We know there is non-response bias by gender; we can't rule out non-response bias by vote. So, before the vote counts come in, they also weight according to whether the sample is in line with pre-election expectations. If it isn't, they suspect bias in their sample, and reweight accordingly. Then, once the vote returns come in, they also have further information as to bias in their sample, and they use this too.

The trouble, of course, is that this procedure assumes the vote returns are correct, although as the polls are designed to project the counted winner, not as a check on the count, this is only a problem if you want, as is of course reasonable, the poll to be an independent check on the count. But it isn't designed to be. The fact that it isn't designed to be, is not, in itself, cause for suspicion. It's been done this way for years, because the "projections" are designed to call the winner of the counted vote. Ironically this is probably why people think they are so uncannily accurate.

4. What do you think about Bill Chirolas' response to Manjoo?

I think he is less fair to Manjoo than Manjoo is to Kennedy. I've never bought into the "he who is not for us is against us" argument, and I think when your own side points out the inconsistencies and flaws in your argument, they are helping, not hindering the cause. But it may be a minority view around here ;)

5. RHP

I have a lot of respect for Phillips, and exchanged data and analyses on Ohio with him early in this whole saga. I don't think he's a statistician, but he is a diligent collector of data, and I think he's done an invaluable job. I think some of his inferential analyses fail to take into account some important factors, but I don't think inferential analysis is his strength. He prides himself on using nothing more than a handheld calculator, and I think that is an excellent approach - simply adding up the votes he thinks were lost. It's once we try to multiply them using inferential methods that we can up with wrong answers (although good inferential statistics are what we need).