EVERYONE: Please take the time to read this. VERY IMPORTANT

Actually, this is not important at all. I just wanted to point out that using a line like:

EVERYONE: Please take the time to read this. VERY IMPORTANT

as the subject line of a post is a terrible way to draw attention to your issue. It's like spam that has a subject line like "Make tons of money!!!1!!" People tend to ignore this stuff after a while; it's like crying "Wolf!" Tell us what the SUBJECT of the post is; that's why they call it a subject line.

(The above suject line was lifted from this thread.)

statistical data. So what does it all MEAN?

This is the way I see it, though I would feel more comfortable if the involved statisticians would confirm that I'm stating this properly:

First, I would go to my recent thread, "A non-statistician's view of the E-M exit poll controversy, and look at item #s 3, 4, and 5:
http://www.democraticunderground.com/discuss/duboard.ph...

Item # 3 briefly explains Mitofski's explanation of what he considers to be exit poll bias, and this explanation is referred to as the reluctant Bush responder (rBr) hypothesis, because it postulates that Bush voters were more reluctant to participate in the exit poll than Kerry. Item #s 4 and 5 describe some simple insights that USCV came up with which explain why the rBr hypothesis logically seems implausible. This is very important, because if the exit poll bias argument is implausible, that leaves election fraud as the only reasonalbe alternative.

Since USCV demonstrated in general terms the implausibility of the rBr hypothesis, they then decided to see if they could prove statistically whether or not the rbr hypothesis was plausible. What they came up with was a complicated statistical explanation which they said proves the implausibility of the rBr hypothesis (I believe that Ron Baiman was the primary author of this explanation, which is why he is now defending it). I don't understand this explantion well enough to offer an opinion on it, but as I note in item # 9 of my thread, I don't really think that it matters much, because USCV had already demonstrated the implausibility of the rBr hypothesis via a simple logical explanation -- certainly well enough to warrant a serious federal investigation into the integrity of the election.

Anyhow, Febble, who was a member of USCV at the time, at some point came to disgree with the USCV statistical explanation, and she broke away from the group and published her own paper on the subject. To make the point again that this whole argument between Febble and USCV doesn't matter much to the central issue of whether or not fraud occurred, here is a quote from Febble's response to Ron Baiman's thread: "My point was never to say that fraud did not occur, simply that it was not concentrated in high Bush precincts." So I say to that, who cares where the fraud was concentrated?

One more thing: Febble also said in her response (part g) that she "believes" that if major fraud occurred then there should be some correlation between the amount of bias in a precinct and the partisanship of the precinct (I'm translating her here, so I hope to be corrected if I'm mis-translating her), and she says that there was no correlation, therefore implying that major fraud did not occur. Whether or not such a correlation existed is beyond my understanding of this matter. But I certainly cannot see why the absence of such a correlation means that major fraud did not occur. In other words, what I'm saying is that if major fraud occurred, why couldn't it occur in different kinds of precincts, with no correlation to the partisanship of the precinct?

I have a few questions which, time permitting, I will try to answer for myself. It just takes so long to slog through all this stuff in a way that permits full understanding of it without just being a ditto-head. Given the importance of this subject, it's not enough for me to take someone else's word for it, and all the Q. Publics out there won't want to do that either. They will want to see the proof with their own eyes and be able to comprehend it. All we have now is "Mitofsky said this" and "USCV said that" (not to mention Febble) and the rest of us just have to decide for ourselves, as with so many other important issues of our time that are so poorly reported by our useless mainstream media whose job it is to explain them to us.

I KNOW the election could have been stolen! Anyone who reads a Diebold Users' Guide can figure that out and figure out HOW it could have been done without the need to HACK anything. But the question in my simple mind is whether these polls really prove it was done.

Here are my concerns:

1. First, why does rBr have to be a constant mean bias in the first place? The proof against rBr, as I understand it, is based on proving that the average bias wasn't constant (or uniform) across precinct partisanship. But why would it have to be? And why would fraud have to follow any particular pattern for that matter?

Like many crimes, election fraud is one of opportunity. Who is to say exactly where the opportunities would have presented themselves, or where those willing to commit the crime would have been positioned?

We can assume the swing states would have been targets, yet all this analysis is focused on the popular vote. Why?

An aggregate 20,000-vote swing to Kerry in NM, NV and IA would have tied the electoral college, even without Ohio. Who cares about the popular vote? (Yeah, I know, Bush thinks he has a "MANDATE." But like I said, WHO CARES?)

How many Republican BOE or vendor insiders would have had Democrats looking over their shoulders who actually knew what they were looking at? How many Democrats would have simply trusted their Republican counterparts NOT to stoop so low as to alter vote counts? And how many Democrats were Democrats in name only and participated in the theft?

All of this begs the question of what the fraud fingerprint would actually look like, based on the exit polls or otherwise.

2. There were only 40 Bush strongholds in the poll, equivalent to only 3.2% of all precincts. How anything can be inferred from such a small number of these precincts is beyond me. If we want to make the assumption that fraud is greater in Bush strongholds, we need more of them to test that hypothesis -- at least as many as the Kerry strongholds which numbered 90 (7.2%), not 40 (3.2%).

3. All this talk about rBr and constant mean bias may be muddying the waters. It relies on 2 assumptions that may be incorrect:
a) that poll bias would have a constant mean and
b) that fraud would not, right?

If not, please explain to me why I'm wrong?

The reason that it is important to refute constant mean bias is that that is the explanation that Mitofsky has stated as fact and the media has swallowed and repeated. If you can refute constant mean bias then you have proved that Mitofsky was just making it up and that rBr is just a convenient cover story, not an explanation based in fact.

The E/M report said (p. 31), "While we cannot measure the completion rate by Democratic and Republican voters, hypothetical completion rates of 56% among Kerry voters and 50% among Bush voters overall would account for the entire Within Precinct Error that we observed in 2004."

(That seems to be slightly wrong: apparently the gap has to be a bit bigger than 6 points in order to account for the 6.5 point WPE.)

Notice that "overall" doesn't imply constant mean with respect to any other variable. They go on to argue that the bias varies with respect to some variables, and apparently doesn't vary with respect to others.

Here's what the report said about bias and partisanship (p. 36) --

"When the precincts were grouped based on their vote (high Kerry through high Bush), the high Bush precincts have the greatest statistical bias. The average signed WPE increases sharply with the increase in the Bush vote. A small Bush overstatement exists in the highest Kerry precincts. The analysis is more meaningful if the precincts where Kerry and Bush received more than 80% of the vote are ignored. In the highest Kerry precincts there is little room for overstatement of his vote. Similarly the highest Bush precincts have more freedom to only overstate Kerry rather than Bush. The three middle groups of precincts show a relatively consistent overstatement of Kerry."

So, they say that the high-Bush and high-Kerry precincts are out of line, but that the discrepancies probably don't matter much.

That may be right, it may be wrong, but it doesn't support the argument that "if you can refute constant mean bias then you have proved... that rBr is just a convenient cover story."

for WPE_Index with respect to partisanship had a slope of zero? That was Mitofsky's main point at AAPOR Miami. And, to answer my own question, his point was to imply no correlation between bias and partisanship. Just another way of saying constant mean bias.

For one thing, as Ron has pointed out, "no correlation" is not the same as "constant mean bias." If it were, he would have to concede "constant mean bias," and he hasn't. And he doesn't have to.

I have no way of knowing what Mitofsky thought he was proving with those slides. I think the zero slope, and the scatterplot distribution generally, are important for two or three reasons. They show that bias in Bush (and lack of bias in Kerry) strongholds doesn't stand out, visually or statistically. And they tend to infirm any vote shift model that assumes that some fraction of precincts was tampered with. (See my post in response to TfC and Bill.)

You said:



Now consider the following premise:



These are just two different ways of saying the same thing. They are saying that there is a constant mean bias. And you have been making exactly that argument over and over.

And the whole point of Mitofsky showing the straight line with zero slope through the scatterplot of WPE_Index was to say that there was a constant mean bias. He says, in effect, there is a strong linear relationship, just look at how the line matches the scatter. A strong linear relationship with slope of zero is the same thing as a constant mean.

If you (and Mitofsky) are finally admitting that the relationship is not linear then you are backing off what you've been saying all along and you are admitting the main point of USCV's position.

And to clear up a miscommunication we are having, when I said "no correlation", I meant those words in a broader sense than you took them. You inferred "no linear correlation" when I meant "no correlation of any type, linear or non-linear". If there is no correlation of any type, linear or non-linear, then there is a constant mean (or at least we can't say there is not one with any statistical significance).

In my field, when we throw around the term "correlation," we generally are referring to a linear correlation coefficient of some sort. So yes, that is a miscommunication.

(I don't think it is right to say that Mitofsky argues for a "strong linear relationship" with a slope of zero. Generally a "strong linear relationship" implies a high linear correlation, which would require at least one of relatively large slope or relatively small variance. What we actually have is very little slope and very large variance. It would never occur to me, or probably to Mitofsky, to call that a "strong linear relationship.")

On whether I am admitting a non-linear relationship, I first need to back up and describe how I defended the March USCV report on Daily Kos. Short version: Rick Brady (a man whom -- full disclosure -- I have come to like very much) argued that the report was wrong because it asserted that the reported response rate was higher in high-Bush precincts, and it hadn't demonstrated that the difference was statistically significant. I said that the statement was descriptively accurate whether or not the difference was statistically significant. It's a basic divide: description versus inference.

Descriptively, we have all started from the baseline that the mean WPE in the "Bush strongholds" is -10.0, which obviously does not equal the overall -6.5. Nothing to debate, nothing to discuss.

So, we seem to be debating whether the -10 is statistically significantly different from -6.5, or better, whether the bias in the high-Bush precincts (which I think we all agree is rather poorly measured by WPE) is significantly different from the bias elsewhere. To determine that, we need to choose statistical tests, and each test makes certain assumptions about possible distributions and causal mechanisms. And since we don't have direct access to the data, we can't readily test alternative assumptions. However....

In the O'Dell simulation (umm, I'm not sure whether this is variant 1 or 2, so take the results as heuristic), the distribution of ln alpha in the 40 high-Bush precincts has these summary statistics:

min -1.54
max 1.70
mean .2672
standard error of mean .1023
std. dev. .6473
skewness -.211, s.e. .374 (not significantly skewed)
kurtosis 1.622, s.e. .733 (distinctly, umm, platykurtic?)

I think the kurtosis may invalidate the standard error of the mean, but I am rusty on that. Regardless, you can see that these sample results are consistent with having been sampled from a population with mean ln(alpha) around 0.12 (I now think more like 0.15).

So, the burden remains on Ron and USCV to come up with any plausible inferential test by which the observed values are inconsistent with constant mean bias. Even if they do, they must further demonstrate that the inconsistency tends to falsify rBr generally, or else, I don't think they have demonstrated very much at all.

Does this help?

then they have refuted Mitofsky's claim that the regression line with slope of zero means anything. They will have proved that there is a non-linear relationship when Mitofsky and you have been implying there is no relationship.

Since the regression line with slope of zero (and hence constant mean bias) is what Mitofsky is putting forward as his main argument for rBr, if that argument is refuted then he is left with no argument, unless he comes up with a new one.

Since the E/M report expressly says that "the high Bush precincts have the greatest statistical bias," it seems implausible that constant mean bias is crucial to Mitofsky's argument.

Nor does it make sense to me that a zero slope and/or constant mean bias is an argument for rBr or other response bias hypotheses, per se. Either or both may be arguments against fraud. The E/M report does argue (or at least I read it to argue) that some of the other correlations in the report are consistent with response bias.

As I think ribofunk pointed out somewhere, the idea of response bias is rather a staple of survey analysis. In the minds of many (I would guess most) survey analysts, E/M doesn't have to meet a very high burden to demonstrate that response bias is at least a plausible explanation for the results. That's probably why AAPOR members have been slow to man the barricades.

What is Mitofsky's argument for concluding that rBr is the explanation?

First of all, as we all know, he doesn't call it rBr. (And actually, it's probably unfair to Lenski at Edison to keep talking about Mitofsky as if he does it all himself, but whatever.) But the January report does favor bias over fraud. Without close textual analysis, I think there are two main arguments.

* The observed WPEs by machine type infirm fraud because no one (as far as E/M knew, or as far as I know for that matter) postulated the observed pattern a priori. A lot of us expected high error rates in touchscreen precincts, and maybe in op scan precincts, but I didn't hear a lot about mechanical precincts. (Yes, the observed WPE is lower in paper ballot precincts, but there aren't many of those, and it's not obvious to me why you jigger op-scan figures and leave paper ballots alone.) Nevertheless, nothing in the January report comes anywhere near refuting fraud at the central tabulation stage.

I feel I have to give E/M a point for this argument, because in fact one of my colleagues was in my office the day after the election predicting higher error rates in touch-screen precincts. I thought that might be right. Turns out it wasn't. So, it's not as if E/M were setting up a straw man here.

* It's not obvious how fraud explains, say, the apparent interviewer influences on WPE and response rates (e.g., response rates 11 points higher among the oldest interviewers vs. the youngest). That looks more like an indicator of response bias. (Counter-argument: all the characteristics in the E/M report put together still don't come close to explaining all the bias. Counter-counter-argument: personally, I'm not sure why they should. If, for instance, some Bush voters shied away from mainstream media logos, we can't control for that factor after the fact.)

I'm somewhat making the case for E/M's arguments, but I'm not wedded to them. I just don't think they're obviously wrong.

One of the arguments is essentially that if you can't prove fraud then you are left with response bias. One could just as easily say that if you can't prove response bias then you are left with fraud. What if we can prove neither? Something still has to be the cause.

In the other argument (the one about response rates) the difference in response rate based on age of the interviewer could be because of some universal attitude that is based on age and has nothing to do with partisanship of the voter.

I don't see any direct arguments that response bias actually occurred.

BTW, I know you've said you'll be out of touch over the weekend. Just posting my thoughts and you can pick up again when you're back.

so I'll have a go!

You are absolutely right. It is notoriously difficult to prove causality with correlational analyses and that is what we have here.

And it is especially difficult when what you trying to estimate is not directly measurable.

We can't know why those who didn't take part in the poll didn't - because they didn't. Did they refuse because they didn't like the media logos? Did they manage to avoid being approached altogether? both are perfectly possible, but not directly provable.

And the fraud issue is not much better - we have to infer it, we can't measure it directly (and if we could, we wouldn't be messing about with these ruddy exit polls, scuse my french).

But if bias was greater where sampling was less likely to be random - it suggests non-random sampling had something to do with it.

If bias was greater in precincts with the most hackable machines, or in swing states, or in high Bush precincts, it suggests fraud had something to do with it.

But neither proves either. Sadly.

I think the most we can hope for from analysis of the exit poll is one or more suggestions of what may have happened. Those suggestions may help us focus our investigation but it is also possible they may lead us astray.

More importantly, I believe the exit poll discrepancy by itself (but especially in the absence of any convincing explanation) is enough to set off all the alarm bells. Whether we can drill into it and find some more specific pattern or not, it is enough all by itself.

Then you couple the exit poll discrepancies with all the known foul play and irregularities and you have to conclude that there should be a very serious investigation.

I believe you agree with these points since you signed a paper that had them (or something very similar) as its main premise (first USCV paper).

It's why I'm here on DU.

It's why I did that work on Florida, New Mexico and Ohio, still accessible at the USCV website.

And I signed that early USCV document when really all it was calling for was an investigation.

But the more detailed its case became, the weaker I felt it got, and I didn't - and still don't - think you strengthen a good case by supporting it with vulnerable evidence.

In fact I think you run the risk of wrecking it, as with the TANG memo.

The fact that the system can be hacked, the fact that Blackwell illegally obstructed the Ohio recount, the massive evidence of voter suppression, the differential spoilage rates, are all hard facts that mandate an investigation.

I think we shouldn't be misdirected. I think the exit polls are the proportionally spaced font of what is otherwise an irrefutable claim for electoral justice.

What we all want to know is WHAT produces bias.

We know it is not constant - it varies hugely. More than any of could have imagined. One hypothesis was that if fraud was the reason for increased bias in some precincts, it might be higher in high Bush precincts. It doesn't seem to, really, or only if you define high Bush very precisely at the 80% mark. Go much higher and you miss the high ones. Go much lower and you include a low one.

So it doesn't look as though bias varies systematically with the proportionate of votes cast for Bush. Which is a shame, because it would be definitely suggestive. But maybe it is only higher when those high Bush precincts are in a swing state? Or maybe whoever hacked the vote knew we would be looking for a slope, and concentrated his/her efforts on minimizing that effect?

So no - there is not constant mean bias. Just no linear relationship between bias and vote-count margin (and the only relationships we can realistically test statistically are linear - quadratic sometimes works, cubic should come out on a linear test, anything else has too many degrees of freedom to have any statistical power.)

And although I've said it before, it bears repeating - you can't prove a null. A slope would have suggested fraud (although even a slope might be explanable innocently, but fraud would have started to look like a goer). But the absence of a slope does not prove that fraud did not occur.

constant mean bias. Both you and OTOH jumped in immediately to say that is not true. Now both you and OTOH seem to be saying that there is not constant mean bias and that Mitofsky does not claim constant mean bias.

Just to be clear, Mitofsky, OTOH and Febble all say there is not constant mean bias. Is that right?

that "constant mean bias" is not a term that people generally use, and I'm not really sure what different people mean by it.

Here is what I am saying - and have always meant, even if I have not said it:

I said that it was POSSIBLE that bias would not be higher in high Bush precincts. Being a stats nerd, by that I meant that there might not be a significant linear correlation - in other words that bias might not systematically increase as Bush's share of the vote in a precinct increased. I suppose you could interpret that as "constant mean bias" but only in relation to Bush' share of the vote. It is certainly not constant. And, if you were to cut the plot into lots of vertical slices, I am quite sure the mean bias in any slice would not be the same as in any other slice. But where a variable is continous - like Bush's share of the vote - it doesn't really make sense to cut it into slices. What you really want to know is is there an overall slope.

To take a silly example - say there is a road between Townsville and Countryville. And you want to know whether Countryville is higher than Townsville. The road could be bumpy, or it could be smooth, but the only way you could really answer the question would be by measuring the straight line slope between the two towns. Now, if it happened that Countryville was on top of a cliff, the road could be fairly flat, then you might have to go up in funicular at the end. That would mean that there was a fairly absurdly non-linear relationship between the places on the way to Countryville. But actually, you could still measure it by measuring the linear slope between Countryville and Townsville.

This is not meant to be a kid's lesson, I'm just trying to be clear without resorting to technicalities. Probably failing, but whatever.

Anyway - what I postulated, way back in April, was that just because the road looked bumpy from the mean and median WPEs we had then, it didn't mean that Countryville was actually higher than Townsville, especially as there was something wrong with the map. And we'd have to hope Mitosfky would do the math and tell us the answer with the right map. And he did.

And the answer is that although the road is bumpier than our worst nighmares, Countryville is not higher than Townsville. Or most of it isn't. Or most of Townsville isn't lower than Countryville. Every single damn bump on the road could represent fraud, but you can't say its happening in Countryville more than Townsville.

Which is a shame, because if Countryville was masquerading as Townsville (i.e. it only has lots of Bush votes because the votes were stuffed) you'd hope it might be a bit higher.

But it DOESN'T PROVE THERE WAS NO FRAUD. It just means we have to look more closely at why some precincts are higher than others. Because they certainly are.

Updated to correct typo

Not true. There are various ways to look at it and you can't know until you look which one will be better at revealing the truth. There are many sets of data where looking at slices will reveal what's really going on while looking at a regression line will be totally misleading.



There you go again with the slope of the line. Your analogy is not apt. We are not trying to see whether one town is higher than another. A huge hill in our curve (of WPE_Index) would represent extra votes and the fact that you come back down off the hill before getting to the "end of the road" is irrelevant. The votes are still extra.



I believe it means that, if you could get a large enough sample in each band of partisanship, you would come out with the same mean as that of each of the other bands.

then you test that with a linear correlation, if you've got a continous variable, which we have, or with an ANOVA if you've got discrete variables, which you haven't. If you slice in percentiles, you don't get those categories because the boundaries will come in from the edges.

And I agree, a quadratic test would be interesting - it's as far as I'd go though, and you still lose precious degrees of freedom. A cubic you would test with a linear.

However - we are talking about inferential statistics here, which have strict rules. There's nothing wrong with staring at the plot and trying to figure out what's going on.

I'm tackling your post backwards here, sorry.

Re the cliff - I take your point, up to a point. If there really were an obvious cliff, I'd be happy (although an obvious cliff at one end usually gives you a slope, leverage in stats, as in teeter-totters, being greater at the extremes). However, there isn't really a cliff. There are four high data points in a sparse category. There's a very low data point at 79% pulling the teeter-totter down. And there are loads of very high ones at the other end pulling the other end up.

That's why I say, forget the stats - Bush's vote count clearly has no "explanatory power" as we say in stats. We need other explanations. Fraudulent ones included.

Can I say yet ONE MORE TIME - THE FLAT LINEAR FIT DOES NOT PROVE RBR!!!

All it does is refute Bvmcc ("Bush strongholds have more vote corruption") which was the hypothesis advanced by Ron. It was a good one. I just don't think it happens to be true.

Not true. There are many cases of non-linear correlation that you will miss entirely if the only tool in your toolbox is the linear regression line.

But those tools are expensive in terms of degrees of freedom, particularly if you don't have an a priori hypothesis.

If you know what fraud might look like in terms of a non-linear pattern you can test for it. But trying to find a function that will fit the pattern is a bit like looking for that face of a madonna in a grilled cheese sandwich. It might really be there, and it might really be miraculous, but you'd have to write a pretty good a priori specification to convince anyone that it was divine intervention. Or have some backup.

Which is not to say it's not worth looking. But the reason the General Linear Model is such a powerful tool is that it gives you a lot of bang for your buck. And it will usually pull out a cubic. In fact what it usually pulls out is an ogive, seeing as most apparently linear relationships actually flatten off at the extremes. So a Weibull would be another thing to fit. But if linear doesn't work, I don't think a Weibull will either.

And if you just carry on piling on the polynomials, yes you'll get a fit. But what will it tell you? Not as much as just looking at the plot. Or doing a multiple regression, which is what Ron is calling for, and I am with him all the way. But that uses the General Linear Model too. Usually.

Well, it's not necessarily terrible to run some slice tests -- although I hope you've noticed the force of Febble's point that based on the scatterplots, these particular tests might be very sensitive to where we cut the slices.

That said, go back to #13, where I use the O'Dell simulator to run a test on the high-Bush slice. And now I'll add some O'Dell results for the high-Kerry slice: mean ln(alpha) is 0.069, standard error of the mean is 0.052. Again not statistically significantly different from a postulated constant mean alpha (i.e., alpha uncorrelated with partisanship in every respect) of 0.12 or 0.15. And again the distribution has unusually high kurtosis -- long tails -- which probably only adds to the noise in the mean.

Obviously it would be more satisfying to do these things with the actual data. And it would be much better to have proper multivariate analysis. Failing that, some theory would be nice. I think our analysis of the high-Bush precincts has more bearing on "Bush stronghold vote count corruption" than it does on fraud vs. response bias generally. And I don't think it refutes vote count corruption -- but I sure don't see how it supports the big hoo-hah that the USCV May working paper still claims it does.

has anyone thought of blowing up the scatter plot, applying some gridlines and estimating the individual WPEs in an equal number of "strong Kerry" and "strong Bush" precincts, instead of this 90/40 stuff, and doing t-tailed paired t-test to determine p-values?

Would this add anything to the debate?

And it's why I keep encouraging everyone to look at those plots! There's a lot data in them thar plots.

And it would have been much more sensible for E-M to have given the means for quintile splits than that arbitrary cutting of the tails.

You might need new glasses though!

And I still think, if you are going to carve up a continous variable, you are better off with a good hypothetical reason for your divisions rather than just where the slicer falls.

Is that not the data from the scatter plots, or at least the first one?

their recorded vote - and thus can not calculate the WPE for the NEP precincts or determine their % partisanship.

As far as I know, only 15 NEP precincts have been matched to the actual precinct data. I'm sure we can locate those 15 precincts on the scatterplot. :-)


Cheers
kiwi

Well, you may and you may not. That plot gets a bit busy around the middle. But it does seems as if the scatter plot could provide some useful data.

You are right. None of the 15 matched NEP precincts stand out on the WPE scatter plot.

I thought Cincinnati 4-M, with its 28% Kerry decrease ("Kerry overestimate") at Bush Vote 60%, would be sure to stand out. But it is just part of the central mass.
http://www.mysterypollster.com/main/2005/05/aapor_exit_poll.html

from the fraud issue.

The issue is to show that there could have been fraud -- not just to show that Mitofsky's explanation for the discrepancy is wrong. Even if rBr is true, it doesn't rule out fraud, does it? rBr and fraud are not mutually exclusive, are they?

So why spend so much time debunking a hypothesis that doesn't rule out fraud in the first place (even if it was stated as fact with nothing to back it up)?

I'm not saying not to debunk it -- just saying that we should state at the outset that rBr doesn't rule out fraud even if it's true. (Unless USCV believes that it does.)

rBr is specifically a hypothesis that postulates bias in the exit polls towards Kerry. So if the bias that rBr postulates is large enough it could explain the whole discrepancy between the exit polls and the official vote count, in which case fraud would almost be a dead issue.

However, as you know if you've been reading my posts, I believe rBr to be highly implausible. In other words, I believe that the exit polls were not biased enough to account for the discrepancy between them and the official vote count.

It may take Febble a while to find this thread -- I only read it by accident. I'm time-limited and going on the road for the long weekend later, so I can't do extensive back-and-forth, but please let me know (in fact, best to PM me, too) if there are more-or-less factual questions I can help with, or if I've blundered in something below.

(1) As I document in a response to eomer, E/M actually don't claim constant mean bias, so the answer to Bill's first question in #1 is, it doesn't. (EDIT: To clarify, I'm referring to his message number 4, the point numbered 1, which contains several questions!)

(2) However, USCV argued (quite reasonably) that the high bias in the high-Bush precincts could be evidence of "Bush stronghold vote count corruption," and can't be explained by any plausible manifestation of reluctant Bush response alone. (USCV has also argued that the bias in high-Kerry precincts is too low to be explained by a constant mean bias, which IMHO is a less compelling argument since E/M didn't claim c.m.b. So a big question is what is "plausible" in the high-Bush precincts.) I don't agree with the USCV analysis for high-Bush precincts, but I won't belabor that here.

(3) TfC, I agree, it's important not to confuse the argument about high-Bush precincts with the larger argument about fraud.

(4) TfC, I think what you are seeing in Febble's response part (g) is a development of thinking. Several of us have been trying to sort out, during May, whether a vote shift/theft would necessarily or probably induce a positive correlation between partisanship and bias. Basic idea: if votes are shifted/stolen in a precinct, the precinct will move up and right in the scatterplot. So if votes are shifted in some precincts but not others, you should get an upward-sloped line that connects the cloud of unshifted precincts and the cloud of shifted precincts. (In fact, that sentence assumes vote shifting for Bush, but vote-shifting for Kerry would tend to do the same thing, only the points move down and left. I didn't figure that out, by the way.)

Of course, if votes are shifted in almost all precincts, this doesn't work. It's hard to think of a test to distinguish more-or-less uniform vote shift from more-or-less uniform response bias. But, for instance, if 1/3 of precincts regardless of true partisanship had votes stolen, probably this positive correlation would emerge. (The general idea of stealing votes in some fraction of precincts comes, I think, from another USCV statistician.)

(Everyone chew on that for a moment. Bill, this speaks to the question of "fraud fingerprint," but it certainly isn't the only possible answer.)

But there are so many other things that could be going on that it's hard for cautious people like Febble and myself (grin) to bang the table and say that there should be a positive correlation if massive vote fraud occurred. I'm closer to believing that than I was three weeks ago.

(5) Bill, I guess the literal answer to your question about the focus on the popular vote is that people are trying to account for the wide-ranging gaps between exit polls and official returns, many of the biggest of which were in safe-Bush and safe-Kerry states. (That point is obscured because most of those states are also small, and also were not heavily sampled because they were considered safe anyway -- so they have larger margins of error.)

But if we boldly stop arguing about the exit polls for a moment, then you are absolutely right: if Bush stole Ohio through any combination of vote suppression and fraud, then that's the election. And many forms of vote suppression won't register in the exit poll.

(6) Briefly to Bill's point #2, I basically agree, but obviously not everyone does. Certainly it's hard to see how the election could have been stolen in "Bush strongholds" alone.

(7) Briefly to Bill's point #3: very well put. Assumption (a) isn't a core E/M tenet, it's more something we fell into while hashing our what we thought about the high-Bush precincts. Assumption (b) is addressed in my points 2 and 4 above, but it's at least possible that massive fraud is compatible with constant mean bias vis-a-vis precinct partisanship.

But it doesn't match the original post of this thread, which was intended to call attention to the inappropriateness of the subject title of another post, but apparently has had unintended effects.

I posted here originally because I wanted to respond to Amaryllus' question and be sure he saw my response. He had apparently mistakenly re-posted his question on this thread because he confused it with the other thread of the same subject title.

So does anybody have an idea of how we can transfer this discussion to another thread?

I'll get back on this later -- I have to go to work now.

Thank you OTOH for an illuminating discussion. Your explanation of the significance of the correlation (or lack of) between vote count margin and WPE certainly helps clarify this issue for me.

I would like to note four issues related to this, the first three being possible explanations for fraud in the presence of a zero slope, and I would be very interested in hearing what you have to say particularly about 3 and 4. The first two I don't feel are that important, but I'll say them first to get them out of the way:

1) Perhaps you could have a slope that is not significantly different from zero, but still be enough to represent enough fraud to affect the outcome of the election. Well, maybe and maybe not. I believe that both you and Febble have made statements somewhat similar to this.

2) As you note, if the fraud was widespread throughout different levels of partisanship, the slope would not deviate from zero. I don't think that this is a terribly important issue because I do believe that the fraud was probably concentrated in a relatively small number of precincts.

3) If the fraud was perpetrated mainly in Kerry strongholds, then I don't believe that that would create a positive slope (is that correct?) I think that there are some reasons to believe that this may be the case. For one thing, in the Kerry strongholds there would be more room for deviation between the exit polls and the official vote count, and it seems to me that this would be where the gold is for those conspiring to commit fraud.

Secondly, I believe that there is some empirical evidence to the effect that this is what happened in Florida, specifically in the three southeastern counties that are strongly Democratic: See my recent post that looked at electronic "vote switching" incidents:
http://www.democraticunderground.com/discuss/duboard.php?az=show_mesg&forum=203&topic_id=371211&mesg_id=371211
42 of 87 reports of electronic vote switching from Kerry to Bush reported by the national Electronic Incident Reporting System were reported from those three counties (My assumption is, obviously, that the reports represent only a very small tip of the iceberg), which were the main hope of giving Kerry an electoral victory in Florida.

4) Lastly, I don't understand the argument that there is a zero slope. As you know, by far the largest negative WPE was reported from the category of precincts that were >= 80% Bush. At the other end of the spectrum, by far the lowest negative WPE was reported from the Kerry strongholds. The middle looks to have a zero slope, although rather bumpy.

In addition, the mean negative WPE in the Bush strongholds is far more negative than the median in these precincts, thus suggesting that a large portion of this negativity was concentrated in a small number of precincts. This would, it seems to me, be exactly what you would predict, in your explanation to me, if fraud was prevalent in these precincts. So could you please explain to me why that isn't suggestive of fraud?

I understand that Febble feels that WPE is not the best measure of bias, because it is confounded by partisanship, and that she has proposed what she considers to be a better measure of bias. I'm not familiar enough with this issue to have an opinion on whether her measure is a better indication of bias than WPE. But even if it was, would that negate what I say above?

This may be my last post until the end of the weekend, so let me know if I leave you hanging!

I see you caught Febble's own explanation of why slope matters.

(1) I don't know whether fraud inducing a statistically insignificant slope could nonetheless be large enough to affect -- umm, let's say the popular vote, since we've been talking about the national exits. It's a separate issue what might have happened in Ohio and how that might have affected the Ohio exits. Ditto Florida (a bit on that below).

(2) Actually -- well, it depends on what you mean by widespread. Fraud in 50%, 60%, 70% of precincts across the partisanship range would probably induce a slope. You would need to have fraud in the vast majority of precincts (but how vast depends on the amount of fraud) before the slope disappeared. Or so I think at present.

(3) I'm not sure, but you may be right. Off-hand that doesn't seem to fit the observed results very well, although I don't much trust my intuitions on that front any more. It would probably be worth testing on the state level (although just off the top of my head, I'm not sure what the right test is).

(4) This is several related issues, and I'm afraid I will answer them badly in a hurry. Maybe one good thing is to look at the scatters via
http://www.dailykos.com/story/2005/5/24/213011/565
and see what you think. You can see the four high outliers in high-Bush precincts -- which don't really stand out amongst all the points. But could there be fraud in them? Sure.

In a nutshell, four precincts out of 1250 doesn't imply enough fraud to matter. It doesn't rule out finely targeted fraud that did matter. It just isn't anything like a smoking gun on its own, IMO.

I hope this is somewhat helpful. Best wishes for the Memorial Day weekend!

So are you saying that the -10.0 WPE in the >80% Bush strongholds weren't enough enough to produce a statistically significant positive slope?

OR

Are you saying that when one uses Febble's measure as a substitute for WPE one doesn't get a positive slope?

I'll await your response upon your return. You have a good weekend too.

I must admit the numbers are swirling but I found something I understand very well, politics and intellectual debate.

tommccintyre asked you a question in the link below, namely will you support the appearance of UCSV before the Carter-Baker committee at the upcoming meeting. You answer is of great interest.



http://www.democraticunderground.com/discuss/duboard.php?az=view_all&address=203x372464#373191

Depends on what you mean by "support" -- I may dislike the arguments they would make -- but I think all the arguments should be aired.

I become maximum leader, in which case, it's highly advisable) haha. Thanks.

"If the fraud was perpetrated mainly in Kerry strongholds, then I don't believe that that would create a positive slope (is that correct?". -TfC

That was my reading of Febble's explanation(#14), also. But I am struggling to follow most of the statistical discussion here.

If OTOH and Febble do not disagree with your statement, TfC, then it might make sense for our intrepid Ohio ballot recounter(s) to focus on the Kerry strongholds.


However, the Ohio counties most suspected of fraud - Butler, Clermont and Warren - hardly feature on the scatterplots.

NEP might have missed the BIG fraud completely.


Cheers

I actually think you might be right. There might have been massive fraud, but not caught by the NEP.

Now that would be ironic.

As I keep saying, the exit polls are a lousy way to monitor elections.

But I can tell you one thing that supports that possibility:

Here is an analysis of "vote switching" incidents reported to the national Election Incident Reporting system that I recently posted: http://www.democraticunderground.com/discuss/duboard.php?az=show_mesg&forum=203&topic_id=371211&mesg_id=371211

This analysis shows that of the 87 reports of vote switching from Kerry to Bush in the whole U.S., 42 of them occurred in the Kerry strongholds of southeastern Florida -- Palm Beach, Broward, and Miami-Dade Counties. These counties were absolutely critical for any chance Kerry might have of winning Florida.

If there had been a slope, it wouldn't have meant that there was more fraud in Bush strongholds - it would have mean that fraud was everywhere, and pushing precincts Bushwise.

I don't want to take credit for this idea - it was Josh Mittledorf of the USCV team who suggested it, and I think he's right. I just modelled it, because the way my brain works, I need to model things to see them.

Now, if the hacker was a smart as Josh (and smarter than me) s/he might have realised that a slope would be a give-away - and concentrated the fraud in Kerry precincts. That, as you say, would have hidden the slope.

So the suspect precincts should be the high bias precincts at the Kerry end - and there are quite a few (more, in fact, than at the high Bush end)!


But to be devil's advocate again (seems to me my lot on life, or at least on DU):

a) it is hard to prove, especially when the really high data points are not the ones that are really pulling the mean bias upwards. The whole mass of points much nearer the zero line are still, on balance, higher than zero rather than lower.

b) those high points are matched by not a few very low points. How does one justify, theoretically, claiming that the high points are Bush fraud without also allowing the possibility that the low points are Kerry fraud?

c) finally - there is a real problem with the expression "rBr", even if you accept that "differential non-response" was a factor in producing bias - nothing in that data set suggests that in all precincts, Bush voters were less likely to respond that Kerry voters. There is plenty of "bias" around that is well outside sampling error, in both directions. It looks to me as though (if you are assuming non-response bias as an explanation) there were many precincts with "rKr" as well. Just rather more with "rBr".

None of which is to dispute your excellent point - just to point out where the tigers are crouching!

And it still means that high Kerry precincts, whether you have exit poll data on them or not, are the ones worth looking more closely at.

How does one justify, theoretically, claiming that the high points are Bush fraud without also allowing the possibility that the low points are Kerry fraud?

The substantial discrepancy between the exit polls and the official count, along with the related WPE computations makes it obvious that either there was substantial bias towards Kerry in the exit polls and/or fraud favoring Bush in the election. Obviously a small amount of vice versa can't be ruled out, but certainly any major trends would have to be bias for Kerry and/or fraud for Bush -- right?

But I'm just saying the bulk of the bias that is shifting the mean isn't coming from those outliers. The whole sausage is floating above the line.

Sorry, I look at scatterplots for a living, so I must find them fascinating! Fascination may not be shared. But in some ways they tell you more than parametric statistics can ever tell you.

If we are going to account for the substantial discrepancy between the poll and count by a fraud theory, it needs to account for all those small amounts of bias (relatively small)clustered close to zero, on both sides of the line, but more above than below.

But I take your point.

I am trying, at the moment, believe it or not, to figure out a scenario by which fraud could have occurred and NOT produced a slope.

It's actually tricky. I'm not saying it's impossible, and I'm not saying that the absence of a slope proves that fraud didn't occur (I'm never going to say that - absence of proof is not proof that something didn't happen). But a slope would have been nice.

If there had been a slope I would definitely have said fraud is suggested by the evidence.

So what we should now be thinking - I think - is: what could be hiding the slope? (I can think of lots).

And why do I think we should be looking for a slope?

Because the sample of precincts has a bell curve. Imagine that plot, if you can (or get it on screen). The one that plots WPE_index against Bush's share of the vote. The precincts have a roughly normal distribution - there are more precincts bunched in the middle than at the extremes. Right. Now imagine that, say 25% of those dots (precincts) had vote fraud. That means that their "true" position is to the left of where it currently is. In other words a precinct that recorded "80% Bush votes" should really be at, say, the "75%" mark. So imagine 25% of the dots moving leftwards. Aha - but if they were in their "true" position, they would also move downwards because they wouldn't have so much bias. So the plot for the same election without vote fraud should have a quarter of the dots further left and further down. This would reduce the mean bias to something like, say, zero. Fine, we've got rid of the exit poll discrepancy.

But what we have done is we have said the high dots should move down and left. In other words we are saying that the reason they are high is that they are also too far right. If this is true, we should have more high dots on the right, and fewer high dots on the left, because the high right dots are really a whole bunch of middle dots that have moved up and right. And this would give a slope - might not be actually linear, but it should give a linear fit. I've set up a model here that you can play with. By my estimates, any amount of vote shift required to get the mean bias down to zero produces a significant slope.

HOWEVER - and this is a big however - there is LOTS of noise in that plot. There may be lots of things masking the slope. A zero order correlation is a very crude measure. But I'm increasingly convinced that the slope is what we are looking for.

My model is here if you are interested;

http://uscountvotes.org/ucvAnalysis/US/exit-polls/simulators/liddle

You can select the amount of fraud in either percentage of precincts or size of fraud, and see what it does to the slope and the mean. You can also alter "alpha" - an alpha of 1 means no "rBr" - an alpha of more than one means rBr - an alpha of between 0 and 1 means rKr.

Knit your own election!

I believe that you are explaining much the same issue that OTOH discusses in post # 9 of this thread. So please see my post # 36 for a response to both of you.

I do have one question for you on this post, however: Your scatterplots show a slightly positive relationship between Kerry vote count margin and WPE, and with your log function. Are those relationships statistically significant?

Log function isn't. Which nicely demonstrates my point that the WPE introduces an artefactual slope.

On the other hand, seeing as I wanted Kerry to win I'd have been delighted to find it still their after removing the artefact!

Why does this show that WPE introduces an artefactual slope?

showed that even where the underlying bias was not dependent on vote-count margin (or one candidate's share of the vote), the WPE would tend to show a slope when you plotted it against vote-count margin (because it is actually a function of vote-count margin). In other words it would look as though a) greater bias was occurring in the middle of the plot and b) more at one end than the other. So that if you ran a regression line through the lot, you'd get a bit of a tilt.

But if you apply my function to the WPE it removes this artefact (the tilt). If there is still a slope after applying my function, then it is a real slope - there really was more bias at one end. But it may simply go flat. So it looks to me as though the significant tilt you get with the WPE fit line, is just an artefact of this WPE problem. It disappears after you apply the transform.

If you look at Kathy and Ron's simulators (or probably better, mine, as it has scatters) you can show that a net bias across all precincts produces a slope in the WPE plot where no slope should be. But it disappears when you apply the transform. In my simulator you can also simulate random fraud. That produces a slope in both plots - i.e. the slope doesn't disappear when you apply the transform.

http://uscountvotes.org/ucvAnalysis/US/exit-polls/simulators/liddle

So if there is fraud we should see a slope - and there might be one, but it might be hiding! That's what I'm thinking about.

Lizzie

I may be wrong, but what you're saying to me sounds like cyclic reasoning. You say that even when bias was not dependent on vote count margin there was still a slope. Well, if there was still a slope, then what reasoning are you using to say that bias was not dependent on vote count margin? And if your answer is that applying your function got rid of the slope, then how do we know that the original WPE wasn't right and your function was the artefact?

It isn't actually circular, but probably sounded like it.

There is an unfortunate and purely mathematical dependency of WPE on vote count margin - because vote-count margin actually goes into the calculation of the WPE. So to get at any underlying real relationship between bias and vote-count margin, we have to get rid of the "confound" i.e. algebraically remove vote-count margin from our measure of bias.

Then we can see if real bias (i.e. biased counting or biased polling) bears some kind of relationship with vote-count margin. For example, is it greater where the margin is in Bush's favour, i.e. at the high Bush end of the plot?

The reason I know that the original WPE is an artefact is just algebra. It's rather boring, but it is in my paper, and if you send me your email address I could even send you the spreadsheet that lays it all out.

But in essence, for a given ratio (not equal to 1) between two response rates, the WPE will be larger in even precincts than at the extremes, and larger at one end than the other. So to use the dreaded term "constant bias" the WPE would not be constant.

However I am not suggesting, and have never suggested, that there is "constant bias" in that plot. There clearly isn't. It's just for any given WPE, you have to take into account the vote-margin in the precinct to calculate the underlying bias. Once you've got that, you've got the real bias, and you can see how it's distributed. If there's a slope, there's a real slope.

And to make it 100% clear (I hope) - the ratio between the apparent "response rates" computed in this way can depend on either how the voters responded, or on how their votes were mangled. The math itself is fraud-neutral. It's the patterns that might be interesting.

I guess my problem is that I don't understand the necessary relationship between WPE and vote count margin. Maybe if you send me the material you spoke of I will be able to understand it. I will pm you with my e-mail address. Thank you.

moderately Dem precincts shifted to the middle quintile and the same from the middle to the moderately Repub quintile. (Ignore the 2 extreme quintiles for now; they are not enough to affect the electoral outcome anyway.)

Here's what I'm suggesting:

165 moderate Dem (not shifted, Med WPE = -5.5 due to actual WPE and bias index)
540 middle (some shifted from previous quintile, Med WPE = -8.3 due to actual WPE, vote shift and bias index)
415 moderate Repub (some shifted from middle quintile, Med WPE = -6.1 due to WPE, vote shift, and bias index)

So here's the question: If this were not fraud, why would there be a larger number of precincts with a higher Med (and Mean) WPE in the last quintile than in the first one?

Any randomly distributed fraud model produces a slope, as long as the precincts have a normal-ish distribution (i.e. more precincts in the mid-range of partisanship than at the extremes. This is because fraudulent data points from throughout the plot will move up and right, and so you will start to colonise the upper right corner of the plot with relative large numbers of fraudulent migrants from the middle. The middle will be re-colonised by the rather fewer fraudulent migrants from the Kerry end. And the Kerry end will be denuded of precincts altogether.

So you will get a slope, even if the distribution is a bit skewed to start with. All it needs is for it it be more populous in the middle than the ends.

So: Fraud of a magnitude to shift the mean that doesn't produce a slope must be more prevalent at the high Kerry end

(I should put that facility into the model, but I haven't done it yet).

It sounds paradoxical, but it's true - random fraud will produce a slope. Concentrated fraud at the Kerry end would kill the slope.

There isn't a slope, so one fraud hypthesis that still has legs is that it was concentrated at the Kerry end. Which, as someone or other on some thread has pointed, makes some sense.

I assumed that the average Kerry % in each quintile was as follows:
90
70
50
30
10
Using that assumption and the above switching, Kerry's overall % would have been 52.56 pre-fraud and 47.6 post fraud. Not too much different from the exit poll discrepancy -- a little more actually.

So has this been considered? I don't think hacking < 25% of the precincts or counties was so impossible.

It does not seem to limit the partisanship of the precincts in which the fraud occurred, does it? In other words, what is the distribution of the fraud assumed in this model? It appears to be uniform, but it's a pretty busy spreadsheet! I'm thinking the fraud may have been more gaussian.

There seems to be an unlimited number of high Bushes in the model as you keep adding more fraud. If the fraud occurred primarily in moderate Kerry and middle precincts, they would be converted to middle and moderate Bush precincts respectively. But would this contradict any other findings?

It selects a different bunch of precincts each time, regardless of where. That was the interesting thing - to put it the other way round, it would have meant that had there been a slope in the data, it wouldn't necessarily have meant "more vote corruption in Bush strongholds" but random fraud anywhere.

Now, when you say a Gaussian distribution - what do you mean exactly? The fraudulent precincts have same Gaussian as the precinct sample - ie the fraudulent precincts are always numerous in the middle of the range, because the model selects a random sample of all precincts, and there are more precincts in the middle of the range.

The way it works is that the model generates a Gaussian distribution of "true" precincts. Fraud shifts a proportion of them in the Bush direction, and thereby also increases the bias. So in effect you have two Gaussian populations of precincts superimposed, one set shifted upwards and rightwards - and because they are both Gaussian, the centroids of each distribution move diagonally apart, creating the sloped regression line.

It would be good (had I worlds enough, and time) to tweak the model to distribute fraud on one side of the plot rather than the other. Feel free to do so if you like! The spreadsheet may be locked, but it doesn't need a password to unlock it - it was just locked to prevent people accidentally adjusting the wrong cell.

be important, and it was. He's one of our best writers and analysts. No one better for writing a plain English sentence that everybody can understand (translating the esoterica of statistics, for instance). His plain English post on the exit poll stuff was a godsend.

His "everyone please read" post takes you to a link, with a very difficult letter on statistics by the Ron Baiman, which tommcintyre then translates (in a post, below). I was quite glad to be apprised that it was "very important." Baiman bursts forth with outrage at the end of his long, difficult treatise on the Edison-Mitofsky twisted scam, and gives all election fraud activists a great compliment.

You should go there and review what is said. It's NOT like crying, "Wolf!" if tommcintyre cries it. It's IMPORTANT! And the post bears it out.

Enough of the internal criticism around here.

Peace Patriot. Coming from TomMcintyre you know it is important he does not even need a subject line but not everyone knows who everyone is here.

I was grateful for it, and for the subject heading which alerted me to its importance.

title on a thread. But I guess that's a difference of opinion between tom and myself. I'm sorry that several posts intended for tom's thread ended up here, but that just goes to show why having generic thread titles isn't such a hot idea.

And for those of us who have not been following tom's contributions, a descriptive title is a better way to clue us into a thread. The name says it all for those of you who know tom, so why use a generic title that says the same thing?

At this point, tom's PM'd me to flame me, and I flamed him back. I hope that's that. I know tom does great work, and I don't want to start a feud. And I'd like to apologize to him for creating this distracting parallel thread.

how to pick a title that will give a post a lot of attention.

I guess so.

http://www.independentmediasource.com/voteintegrity.htm I love this media blaster, I had it on my blog, also there is a nice one from Impeach.org for congress. this one is such a great resource!

I also am encouraging folks to go a step beyond an electronic blip or 2 (or hundreds) (Don't forget the Conyer's one of course).

Please join the BEYOND virtual Orange Post card capaign!
Support verified voting,and election fraud investigations; without which reforms will be meaningless! Investigate, prosecute IMPEACH punish to restore our democracy! Also If you have time; Please send some to the media. Cards of course don't have to be orange-progressive ones, index cards etc.
<http://jarnocan.blogspot.com> petitions and free graphics-print on orange or white
Also at care2.com- <http://www.photochains.com/view/B.S.> here are some anti-Bu**SH** photos along with ideas for actions and links. You can also contribute pictures to this and make comments-wide progressive but not all up on this stuff-audience. Might be a good way to get the word out about issues you care about!