Anyone Got Freeman-Mitofsky Exit Poll Debate Results Yet?

[Bill Bored]

OK, so how was it? Was TIA there? I bet OTOH was.

Come on, all the true believers are waiting to find out who won the debate!

Was Ross Perot there? Any pretty graphs? Scatter plots? Pie charts? Just once I'd like to see a pie chart about the exit polls!

Has anyone figured out how they generated the WPE with the voting machines yet, or was it just about whether the polls were biased?

Come on, I'm interested; really I am! Post it here.

[Melissa G]

1. Well I have yet to hear but the format was that they each presented a

paper and then there was some time afterwards before lunch.. There was an entirely different presenter in the afternoon. Any DUers make it to fill in the details?

Kickin for answers...

[eomer]

2. Freeman won the debate by a 51/48 % margin...

according to those leaving the presentation, although many were reluctant to give an opinion. ;)

[demodonkey]

7. Oh-oh... RGRs!

(Reluctant GEEK Responders) LOL!

:P

[tiptoe]

11. Or...

vote RiGgeRs!

Clint Curtis (see below) talks about a prototype program he wrote for Jeb Bush's former running mate, Tom Feeney, that would "flip the vote" 51-49, whichever side one wanted to win. This 52-48 outcome looks suspicious. ;)

[autorank]

15. You left this out...tabulating was stopped when the race was tied 48-48%

There was a humidity crisis and the count ended for a brief period. Subsequent tabulation showed a last minute surge for Mitofsky, who was the declared winner 51-48%. This matched with the final exit poll of the debate, which incorporated the final count of debate ballots.

Mitofsky expressed surprise and said, "My methodology has obviously been accepted as an industry standard."

Freeman was unavailable for comment but was seen enjoying himself while eating a Cheese Steak in South Philly.

[Wilms]

16. It ain't over!

A third party (that wasn't allowed to join the debate) is asking for a recount.

[eomer]

20. In a breaking development...

a British data analyst¹ and a political science professor¹ jumped in to shield Mitofsky from a small but angry crowd throwing stones after the event. There was little the Brit could do to protect Mitofsky although she did suggest that the blows were finding their mark partly because of the way he was skewing his stance to one side. He appeared confounded at first but then applied her advice and was so pleased by the result that he gave her a handsome gratuity. The professor, on the other hand, seemed to be everywhere fending off potential damage. One mathematician¹ in the crowd seemed to have a lot of rocks and an uncanny knack for landing blows just where they smarted, if truth be told, but authorities eventually stepped in and chased him off to parts unknown.²

Footnotes:

1. To "the Brit", "the professor" and "the mathematician" (rest in peace): don't take offense, just kidding around you know.
   
2. To IndyOP: it's a joke.

Edit: fonts and wording

[Febble]

21. ...upon which

the "rocks" were discovered to be pies. These were distributed to the crowd, who swallowed them gratefully.



(another joke)

[OnTheOtherHand]

22. chocolate chip cookies would've hit the spot! n/t

[yowzayowzayowza]

23. I'm..... I'm ...

soooo confused. :cry:

[eomer]

25. Sorry, no cookies, but can I offer you...

:donut:

[OnTheOtherHand]

26. not right now, but maybe in the morning, thanks...

(These emoticons just blow my mind -- not to mention my cholesterol count.)

[Bill Bored]

37. Actually, I heard ESI had tried to do something like this for real. nt

[Febble]

3. Mmmm, pie.....

[Fly by night]

4. Would someone who attended the debate please post a summary. Thanks.

[OnTheOtherHand]

5. short answer--

mostly they talked past each other. In fact, even toward the end when they got to talk to each other, they were still largely talking past each other. However, I was stuck in Drive From Hell and missed Freeman's live presentation entirely (although I have the copy of slides he handed out), so I'm still trying to figure out what I think happened.

Freeman presented a slide that looked eerily like one of the projections from our "game" -- I almost ran screaming from the room when I saw it in the handout. One of his slides showed that WPE was somewhat higher in states with Republican gubernatorial control, the argument I know you were interested in. He also has a slide that shows rates of "MSNBC election day calls" per million voters, and apparently those rates are moderately correlated with WPE. I'm not sure I "buy" that result -- he doesn't control for anything else -- but it's an interesting idea.

Most of Mitofsky's presentation was close to what he did at AAPOR. He presented a cool new slide that shows the relationship between exit poll 'bias' (i.e., based on the difference between the exit poll result and the official count) and the "swing" from 2000 to 2004. If the 'bias' evinced fraud, you would expect Bush to do better, relative to 2000, in the precincts where the bias was largest (i.e., where Kerry did a lot better in the exit poll than in the official count). He didn't. I think that is a big, big problem for the exit poll argument. But since Freeman hadn't seen the slide before -- just as Mitofsky hadn't seen Freeman's results before -- the discussion wasn't very edifying. I'm trying to arrange for that slide to be posted somewhere, so people can stare at it and draw their own conclusions, ask their own questions, etc.

The presenters were well prepared and held their own; I doubt many minds were changed.

[rdmccur]

14. I wonder how much 'bias'

there may have been in the 2000 election,
referring to the new Mitofsky slide (which I suspect did not address that)? Also I wonder if Mitofsky examined this trend over different individual states? The data that he bases the slide on needs to be carefully examined as well.

[Febble]

18. Good questions

The plot, as with the ESI plot, correlates "swing" (change from 2000) with a measure of bias – and, as with the ESI plot, there is no significant correlation.

I think it is important to be really clear about what this is saying (and therefore what it is not saying). It is saying that redshift in the poll is not reflected in redshift in the swing. Bush didn’t do better than expected in precincts where the redshift was greatest. Nor did he do worse than expected in precincts where redshift was small, or where there was actually blueshift. So the straightforward interpretation of this is that whatever beast it was that red-shifted the vote relative to the poll, it was not the same beast that red-shifted the vote relative to 2000. If fraud was the beast responsible for both, the two effects should be correlated. It would seem they are not. They seem to be different beasts. That’s the straightforward interpretation.

But your two questions are good -

To take your first point: "bias" in the 2000 election:

1. The correlation between “swing” and “redshift” is based on what Freeman calls the PLD – precinct-level discrepancy, traditionally measured by the “WPE” (Within-Precinct Error). In 2000, the WPE was not substantially biased (nationally, the mean state WPE was not significantly different from zero, although it was slightly red-shifted). So while fraud could well have happened in 2000 (and we know, from Mebane’s work, that the butterfly ballot and overvotes in Florida were each alone enough to cost Gore the election) it cannot have been enough – or not of a form – to make a substantial impact on the WPE nationally. (Some forms of fraud will not show up in exit polls, so this is not to say that no fraud occurred in 2000). However, in 2004, the PLD (whether measured by the WPE or the measure used in Mitofsky’s slide) was massively significant – the famous red shift. So even if we postulate that fraud occurred in 2000, if fraud is going to account for the massively increased redshift in the poll in 2004 it has to be a massive increase in fraud - and, moreover, fraud of a type that will show up in exit polls. Which is the point of this entire debate of course. OK. The one scenario that would work is if the fraud in 2004 was completely uniform, across all states and all voting methods and tabulation systems (and bear in mind that in about 60% of precincts, the WPE was calculated from results obtained at precincts, not from county tabulations). However, if some, precincts, counties, systems, or states or were relatively clean – in other words if there was variance in the degree of fraud-increase, then we should find both less redshift in cleaner precincts, counties, systems or states, AND less of an increase in Bush’s share of the vote relative to 2000. In other words, redshift in the poll should would correlate with Bush’s swing. Clean states/voting methods/counties/precincts = less red shift AND less swing to Bush. Dirty states/voting methods/counties/ precincts = more red shift AND more swing to Bush = more red shift. But this does not seem to be the case.

Which brings me to your second point - trends in different states:

2. This is slightly more problematic because of the way statistics works. To demonstrate a null you need a lot of statistical power, which is possible when you have a large number of data points, as in Mitofsky's plot. However, the number of precincts in each state was of course much smaller, which means that within-state correlations will be fairly unstable, and any outliers could mask real effects. You also have to “correct” your criterion for statistical significance to compensate for the number of analyses you do. For example, the usual criterion in social sciences for “significance” is a 1 in 20 chance of something occurring by random processes. So, out of fifty states you’d actually expect 2 or 3 to have a “significant” correlation between swing and redshift, simply by chance - which means you have to make your significance criterion a lot more stringent, and this in turn means that there is a shortage of statistical power to do the analyses you suggest.

One way of avoiding this problem, is to devise an “a priori” hypotheses – in other words, before looking at the data, you decide which states are most likely to exhibit a relationship between swing and redshift – and test those first. Well, in this case, suspect numero uno was Ohio. Which was why ESI tested the correlation in Ohio. But in fact, despite a relatively large number of precincts and an a priori hypothesis, they found that it was not significant - no correlation between red-shift and swing in Ohio, not even a hint. And yet the net redshift in Ohio was substantial (not the highest nationally, but one of the highest, something that has frequently been point out as an argument FOR fraud). The mean WPE in 2004 in Ohio was -10.9 (negative means redshift); in 2000 it was -1. In other words, Ohio is a poster child for increased red-shift in 2004 – the PLD was low in 2000, indicating that if there was fraud in Ohio in 2000 it was a type not to impact heavily on the exit polls; it had a massive redshift in 2004; and yet that increase in red shift was not correlated with increase in Bush’s share of the vote.

This bring us back to my first point – perhaps there was uniform fraud in Ohio (although nothing about Ohio suggests uniform fraud – more like a Chinese restaurant menu of methodologies for boosting Bush’s vote by fair means or foul) – which is why it was of interest to see whether the non-relationship between swing and redshift would be observed at a national level as well as at the level of Ohio. If uniform corruption in Ohio accounted for ESI’s negative finding, then there still ought to be a positive correlation when done nationally. As long as there was variance SOMEWHERE in the level of fraud nationally, there should be a correlation at the national level between red-shift and swing to Bush. And we know there was between-state variance in red-shift because E-M give us the mean WPEs for each state, and in any case, the between-state variance in redshift has been the subject for countless analyses (TIA: the BEAST was in the EAST….)

So this is fairly strong evidence that fraud can’t account for the exit poll discrepancy nationally. That doesn’t mean there was no fraud – because some types of fraud won’t show up in the poll. And it doesn’t mean the election wasn’t stolen. Well-targetted small-scale fraud, and many of the kinds of disenfranchisement that are well documented in Ohio could, and possibly (probably?) did cost Kerry the presidency. But it does seem to me to mean that at a minimum, the exit poll discrepancy in itself can no longer be cited as evidence for fraud. And I would suggest that the new plot suggests rather more than this: that vote-switching cannot have happened at a massive scale – because otherwise it would have shown up in the polls as a red-shift that WAS correlated with a swing to Bush. Which rather means that the evidence is against the hypothesis that Kerry won the popular vote.

On the other hand the fact that it has taken nearly a year to achieve some kind of demonstration of this is in itself an argument for ensuring that the electoral system is secure and transparent, which it patently isn't. And even when it is, it will not solve the problem of voter suppression.

[eomer]

24. Febble, I am trying to get my head around this and not succeeding so far,

so maybe you can help me on a couple of points.

One issue is the following: are you saying that there is enough statistical power in the national case so that the apparent lack of correlation between those two factors in the sample allows us to conclude (with significance) that they really are not correlated in the full population? If in a hypothetical poll we start with a small sample size then an apparent lack of correlation could be due merely to the lack of statistical power and does not necessarily tell us there is no real correlation. If we jack the sample size up to almost all of the total population then an apparent lack of correlation is almost certainly a real lack of correlation. In other words, lack of enough significance to reach a conclusion does not automatically allow us to reach the converse. We still need to muster enough significance on behalf of the converse. Otherwise we are left not knowing whether there is a correlation or not. So where do we stand on this particular correlation? Can we say for sure there is no correlation or can we only make the lesser statement, that there may or may not be a correlation, in the full population?

And another issue: how do we know that there are not subsets of data with a negative correlation that would counteract other subsets with a positive correlation? Our national election system is a complicated patchwork of different technologies and different political contexts. Fraud in this system would not be smooth and predictable - it would be lumpy and jerky. Similarly, the true vote in 2000 and 2004 could have all sorts of weird patterns due to political and social patterns that are definitely not uniform throughout the country. These patterns in fraud and actual votes would come in big clumps based on the changes in technologies used, party control and legal climate as we cross jurisdiction boundaries and also on de facto patterns of voluntary segregation by ethnic, religious or other type of social affiliation. How do we know that such a hurky jerky system would result in a significant correlation between red shift and swing even if both were mostly caused by fraud? If actual swing (as opposed to swing caused by fraud) were zero then I can see a stronger case for your conclusion. But since actual swing is not zero and could be positive or negative in clumps, and those clumps could have a totally different pattern than the clumps of fraud-based swing, it feels to me like you've said more than we can really conclude with significance.

[Febble]

27. Well....

you never have enough statistical power to conclude a null. In fact the null is always false, in some sense, given enough decimal places. But with the power of the full sample, you have to conclude that the relationship between swing and redshift is extremely small - and that that we cannot be confident about whether it is positive or negative.

So we can make one lesser statement: that there may or may not be a positive correlation; and one stronger statement: that what relationship there is is very small. Tiny, in fact.

And your second point is a good one.

No, we can't tell from that plot that there are not subsets of data where there is a real relationship. There may well be. In fact I am sure there are - real data from huge populations is always heterogeneous - it's why I love it. So what I think you are saying you are saying is that in some sets of precincts real swing may be large and fraud may be small, whereas in other precincts, real swing may be small and fraud may be large - and variants on all these, which will tend to cancel out. And I agree this is possible.

But it still leaves us with no net correlation. And if fraud was the cause of the red shift, the effect size of the relationship between fraud and swing has to be substantial (because the redshift was substantial). If the effect of fraud on swing is cancelled out by swing that happened anyway where there was no fraud, and by fraud that happened anyway where there was no swing - then we don't have fraud as an explanation for the net discrepancy. Sorry if that's unclear. It's getting late here, and I may make a better job of this in the morning.

But I suppose, to return to my beasts metaphor - if fraud beasts are responsible for some of the swing, then either there are too few of them to show up in the noise (small effect) or their effect is counteracted by other fraud beasts pulling inthe other direction (pro Kerry fraud? There are some very blueshifted precincts in there).

It would certainly help if the plot were posted somewhere. If it is, I'll link.

Cheers, Lizzie

[eomer]

31. Let's try it this way...

Let's assume that actual swing turned out to a net zero in some hypothetical case. By actual swing I mean swing in actual intended votes of people who went to the poll and thought they voted. And by net zero I mean that the total at the national level didn't swing to the red side or the blue. Further assume that this actual swing has quite a bit of variation from one precinct to another and from one region of the country to another. The variation will need to have precincts that are red swing and precincts that are blue swing since the net result adds to zero. Also assume that the official count in 2000 was more or less correct.

Next assume an exit poll that would have been accurate at the national level if we could have compared it to the actual intended votes of people who cast a vote at the polls. The sum of all WPE for this poll (when measured against intended votes, not official votes) is zero but there is significant variation across precincts and this variation has its own unique distribution.

Now add on top of that base (still hypothetically) fraud that shifts the official count significantly toward the red. This hypothetical fraud is distributed across precincts in some pattern that is totally different than the pattern of distribution of actual swing in my first paragraph and also than the pattern of WPE in my second paragraph.

In this hypothetical, both the total swing at the national level and the total red shift at the national level are "caused" by fraud. They each would have been zero if no fraud were committed. The fraud then moved each of them significantly to the red.

We can break the total swing down as follows:
total swing = (actual swing) + (swing due to fraud)

And we can break the total WPE down as follows:
total WPE = (actual WPE) + (WPE due to fraud)

(Don't take the second formula too literally. I know that wouldn't be the actual equation but I mean that in general the effects of the actual component and the fraud component are additive.)

(Another clarification - I know that "actual WPE" isn't such a good name for the quantity it represents here but let's go with it for now. Actual WPE is defined here as the WPE we would calculate if we were able to know the true intended vote of each voter who voted at the polls. In other words, it is the WPE we would calculate if there had been no fraud.)

Now if we were able to calculate (swing due to fraud) and (WPE due to fraud) in each precinct, we would surely find a correlation between them because they are one and the same. If I fraudulently switch a vote in some precinct from blue to red then I add one vote to (swing due to fraud) and I add one vote to (WPE due to fraud).

The problem I'm getting at in this post is that we cannot calculate either (swing due to fraud) or (WPE due to fraud). We can only calculate swing and WPE after the fraud has been added on top of (actual swing) and (actual WPE). In other words, we can only calculate (total swing) and (total WPE).

I believe the above hypothetical is a counter example to your conclusion that you would have to see correlation between swing and red shift if they were both caused by fraud. In my hypothetical they are both caused by fraud and yet you cannot see correlation because the correlation is hidden by the different patterns in the base before fraud is added on top. If we could separate the fraud from the base then we would see the correlation in the fraud. But when we add the fraud to the base we get a total that shows no correlation. The base had zero swing and zero red shift. Adding fraud "caused" swing and red shift to be significantly positive. But we can't see correlation between swing and red shift because they had different patterns pre-fraud.

[Febble]

35. OK, I think I'm with you

up till your 9th graf.

And what I think you are getting at is that there will be variance in both WPE ("actual") and swing (also "actual") that has nothing to do with fraud, and without knowing the quantity of this we cannot subtract it from the observed WPE and swing. Let's not call it WPE for now - can we use Freeman's term, PLD (Precinct Level Discrepancy) as WPE has come to have a very precise meaning, not always equivalent to PLD

OK.

The regression equation for the observed data will be

Observed PLD = Beta x observed swing+ "error".

If in the unlikely event that there is zero actual swing in all precincts, and zero actual PLD in all precincts, but fraud was pervasive, then the error term will be zero, and Beta will be equal to 1. In other words if you know the PLD you will know the swing.

However, there will, of course, as you point out, variance in swing that is not due to fraud, and variance in PLD that is not due to fraud. In this case the value of any observed PLD will be given by a value of Beta (calculated to give the best fit to the data) times the observed swing in that precinct, plus an error term, the “residual”. The residual will tell us how far “out” the predicted value of the PLD will be if we use the Beta value to predict observed PLD from observed swing.

If most of the variance in PLD is due to something independent of whatever is causing variance in the observed swing, the calculated value for Beta will be small, and the residuals will be large. But the more variance in observed PLD is due to the same factor that is causing variance in observed swing, the larger Beta will be and the smaller the residuals. And the way we estimate whether variance in swing is accounting for a “significant” proportion of the variance in PLD is to look at the ratio of explained variance (i.e variance accounted for by swing) to unexplained variance (the residuals) i.e. an F test (alternatively, but it comes to the same thing, you can do a t test on Beta to see whether it is significantly different from zero).

This is why I tried to express the thing in terms of effect size rather than significance. What you can’t say, with a null finding, is that there is no relationship. What you can compute is the proportion of variance in variable A that is accounted for by variance in variable B (the “R squared” value), and estimate whether this proportion is large enough for us to conclude with confidence that it is not a chance finding. And according to Warren’s slide this amount is .009%. Which is extremely small – and not large enough to allow us to conclude with any confidence that it is actually different from zero.

And what it boils down to, which I should have said more clearly, is that we essentially have two unexplained phenomena. And they are both “swings” – a swing to Bush relative to the 2000 vote, and a swing to Bush relative to the poll. The suspicion has been (rightly) that the two swings are due to single cause – fraud. If this is the case, it clearly won’t be a one-to-one relationship but if fraud is a major – or even minor – cause of both, there will be a correlation. The smaller the correlation, the less of each can be attributed to a common cause. And it looks as though the shared variance is actually tiny. Which is why I say it leads to the pretty inevitable conclusion that the redshift in the poll wasn’t caused by fraud. There could still have been election theft –some voter suppression won’t show up as redshift anyway, and still could have cost Kerry Ohio, and possibly NM. But that R squared seems to say that the exit poll redshift is not a proxy for the magnitude of the fraud.

If you know all this, and I'm missing your point, forgive me.

(edited for consistency of terminology!)

[OnTheOtherHand]

36. I think eomer is also saying something else --

that just as a bivariate correlation between, say, use of hand-counted paper ballots and WPE could be spurious, a bivariate non-correlation between fraud and swing could also be spurious: some missing variable could confound the finding, or non-finding.

To which I would respond, yes, that is true -- the non-correlation doesn't actually singlehandedly sink all Massive Fraud theories. But it seems to me that any successful MF-salvage operation would have to acknowledge substantial sources of bias in the exit polls, and to offer a specific argument about a specific confound. I might even go so far as to say that the non-correlation sinks all 'really existing' MF theories (of which I am aware), although it doesn't doom all conceivable MF theories.

(EDIT: by "massive fraud" in this context, I was thinking of "fraud so extensive that Kerry won a majority of votes actually cast or attempted." At least 3 million votes nationwide. Freeman says that the exit poll data "indicated that Kerry defeated Bush by 7,000,000 votes nationally" -- so that implies 10 million net votes stolen, although it's not clear to me whether he really commits himself to that estimate.)

[eomer]

39. Yes, that is what I am saying.

And I agree with your follow on points.

I've never believed that a simple theory was going to explain the result we see. I've posted many times my opinion that our election system is messed up everywhere you look and every way you can imagine. A simple explanation would be a surprise to me.

[Febble]

41. Right.

I agree that there could be "suppressor" variables. I also agree there are very many ways the system is messed up. One point I keep banging on about is vote spoilage, particularly in predominantly African American precincts. This would tend to lead to red shift in the polls, and might account for Steve's finding regarding WPE and proportion of African Americans in the state (although there are collinearity issues to consider there) - and not only would it tend to lead to a red shift in the polls, but presumably it tends to happen in the same precincts year after year, so it wouldn't show up in that correlation.

So no, the answer is unlikely to be simple. But so far, any hypothesis that has postulated fraud as a primary factor in producing the red shift has failed to be supported by subsequent analysis - whereas hypotheses that postulate polling factors (various) as predictive factors HAVE been supported. Which is why I am actually being fairly circumspect in what I think this analysis tells us: that new, vote-switching fraud is not the cause of the exit poll discrepancy. I'm not saying that the analysis rules out fraud.

But perhaps I should be even more circumspect - perhaps I should say that if new, vote-switching fraud DID contribute to the exit poll discrepancy in some places (NOT Ohio, apparently) then it was counterbalanced by other effects that pushed the PLD in the opposite direction. And that still leaves us with my conclusion that the exit poll discrepancy is not a proxy for the magnitude of fraud.

[eomer]

46. I'm not ready to concede that any hypothesis has failed to be supported

by subsequent analysis if that subsequent analysis is not open to real peer review. In particular, Mitofsky's flat line is not convincing to me unless I can see it and play around with it (either personally or by proxy).

We've agreed (you, me and OTOH) that the election system is not a simple system and that there could be suppressor variables, confounds, currents, counter currents, all going on at once (okay I embellished a bit, sorry). In a complex system like this the analyst is forced to make numerous decisions on how to measure, slice and dice the data. If he fiddles long enough with different ways of slicing and dicing it is inevitable that a flat line will pop out at some point. If an outsider is presented that flat line all by itself and is not informed that there was a lot of fiddling with analysis decisions before it appeared then it looks very convincing, maybe even dispositive. But it's really not. It's just an artifact of fiddling.

There is an old joke among actuaries that goes like this:
A lawyer, an accountant and an actuary were each asked to render a professional opinion on the question "what is 2 + 2?". The lawyer said <boring lawyer stuff about why it is 4>. The accountant said <boring accountant stuff about why it is 4>. The actuary locked the door, pulled the blinds shut and said "what would you like it to be?".

The reason the fictitious actuary is able to make this offer to his client is that there are so many assumptions and interacting variables that it is often possible to come up with whatever answer you want if you are willing to back into it.

So to me Mitofsky's flat line is the statistician version of that joke. Would a flat line convince you there was no fraud? Give me a few days (or in this case quite a number of months) and I'll get back to you with a flat line.

I know that I am treating Mitofsky as guilty until proved innocent but that is what he gets for telling everyone that he is the keeper of the marbles and no one else can play with them except his bestest buddies.

[Febble]

48. Well that's entirely fair

Well, maybe not entirely:

what you say about fiddling is a bit ironic. In my experience, the fiddling usually goes the other way - trying to make lines slope, not make them flat. This one is actually pretty straightforward, although there would be more than one way of choosing the exact measure of swing and bias. The fiddling would start when you try to show that polling (or fraud) factors DID account for bias.

And no, a flat line does not convince me there was no fraud. It never would. As I've said all over this thread, you can't prove a null. But what you can prove is that one thing is not a measure of another. And I've read more times on DU than I've had hot dinners that the exit poll discrepancy indicates x gazillion votes were stolen. The plot, if it demonstrates nothing else, demonstrates that the polls do no such thing.

What demonstrates that votes were stolen is the kind of painstaking work done by RHP in Ohio (although I think his data collection is better than his inferential statistics), as well Conyers' catalogue of woes.

[anaxarchos]

96. Couldn't agree more...

The "simple" theory is attractive and it matches some of the evidence from 2000:

"2. Even after controlling for other factors, rates of ballot spoilage remain higher in predominantly black areas than in other areas of Florida. As the last model indicates, with all else being equal, for every 1-point increase in the percentage of registered voters who are black, there was a .07 percentage point increase in spoiled ballots. In addition, these rates were even higher where substantial numbers of blacks were found in counties with large margins for George W. Bush. All of this corresponds to and further reinforces the findings of the USCCR that there is evidence of racial disenfranchisement in the 2000 election in Florida. Consequently, it is important that federal authorities should investigate this matter more thoroughly."

Klinker, Philip. “Whose Votes Don’t Count?: An Analysis of Spoiled Ballots in the 2000 Florida Election,” Appendix XI in the USCCR Report (25 June 2001).

Available here:

http://www.hamilton.edu/news/florida/Klinkner%20Analysis.pdf

But 2000 doesn't really match the pattern of 2004... A simple explanation would be a surprise to me too.

[Febble]

97. Omigosh

I agree too.

All the evidence I have worked at (and I've worked at quite a bit) suggests that non-white Americans were overwhelmingly the victims of electoral injustice in 2004 as well as 2000. Any time I am tempted to forget it, I re-read this diary by Shanikka on DKos, written December 15th 2004:

http://www.dailykos.com/story/2004/12/15/9734/4841

And of her many points (some of which I agree with, some I don't) her most important point was that the fight for electoral justice matters REGARDLESS OF WHETHER KERRY WON OR LOST.

And securing the vote-counting software won't in itself fix that injustice.

[anaxarchos]

98. Shocking....

Let me push my luck and quote myself:

http://www.democraticunderground.com/discuss/duboard.php?az=show_topic&forum=203&topic_id=330828#333264

Posts #35, 40, 57....

[Febble]

114. And let me quote you

http://progressiveindependent.com/dc/dcboard.php?az=show_mesg&forum=120&topic_id=632&mesg_id=652

Is this what you believe?

Have you actually attempted to hear what I am saying?

Does it simply suit your argument to undermine my credibility on other boards, or do you honestly believe that shit?

Have you looked at any of my work done to try and expose electoral injustice in the 2004 polls?

Well, for the hell of it, here are some links:

http://uscountvotes.org/index.php?option=com_content&task=view&id=70&Itemid=63
http://uscountvotes.org/index.php?option=com_content&task=view&id=65&Itemid=63
http://uscountvotes.org/index.php?option=com_content&task=view&id=32&Itemid=63



Have you considered the possibility that I may not believe that the exit polls are an index of fraud because that is my reluctant but honest conclusion?

And that perhaps it is at least arguable that my contribution to the debate has been my attempt to direct investigation to where the case for electoral reform may be more fruitfully made?

OK, maybe you are right and I am wrong. Maybe there was massive fraud in 2004, and the exit polls were right and the vote-count was wrong. God knows how fervently I hoped that was true. Why in hell do you think I even touched the damn things?

Sorry, anax, but I am SO fed up of this shitty debate. OK the shitty stuff is now on another forum (from which I was banned after about 5 posts for stating, as a matter of honesty, that I had committed the heinous crime of voting for Tony Blair as leader of the Labour party about a decade ago, and for an anti-war member of his party in 2005). But when I see that bogus crap about me come up again, with the apparent sole purpose of undermining a perfectly decent argument, an argument that might actually help people look in the right direction for the rampant electoral injustice that occurred in 2004 instead at those BLOODY exit polls, well, let's say, it undermines my faith in human nature.

And certainly in the common decency of the Left.

But for your information, and this is not tortuous or difficult, the reason I have lost my last vestige of hope that the exit polls might have indexed substantial fraud is that the exit poll discrepancies at precinct level (the "raw data") were complete uncorrelated with increase in Bush's share of the counted vote.

I am still trying to find loopholes. I will continue. I think.

But frankly at this very moment I'm not sure I give a damn.

[eomer]

38. I'm not sure whether you agree on my essential point...

So let me try to capture just the essence of it.

Say that we have a set of variables that can all be sampled:

• D = Discrepancy
  
• D_e = Discrepancy due to error
  
• D_f = Discrepancy due to fraud
  
• S = Swing
  
• S_v = Swing due to change in voter preference
  
• S_f = Swing due to fraud

In this hypothetical the following relationships exist:

• D = D_e + D_f
  
• S = S_v + S_f

What I am saying is that there could be a significant correlation between D_f and S_f even if there is not a significant correlation between D and S (at whatever level of sampling we have chosen).

In fact, since I haven't assumed anything about any of the variables, there could be a strict linear relationship between D_f and S_f and still no significant correlation between D and S. This drastic result would require some fancy footwork in the form of just the right inverse relationship between D_e and S_v in order to cancel out all or almost all of the correlation between D_f and S_f but that is totally possible since we have not assumed anything in that regard.

Do you agree that these statements are true from a theoretical point of view?

[Febble]

47. OK

I am assuming that De and Sv are not correlated (although they could be, but I won't get into that right now*) and that Df and Sf are. You are saying that Df and Sf could still still be correlated in the presence of a non-significant correlation between D and S.

Well, yes, they could. But the straightforward interpretation says that the portion of the variance in D that is UNshared with S is much larger than the portion of the variance in D that IS shared with S, and for now we are assuming that all the variance in D shared with S is fraud. So there might be a real effect, but it is swamped by the error (error in the technical sense of unexplained variance).

The non-straightforward interpretation is that there could be suppressor effects. What follows is entirely hypothetical, but if fraud was more common in marginal states, but Bush voters were also keener to participate (make their voice heard?) in marginal states, you might get no significant zero-order correlation between swing and bias, but if you added "marginal state" (sorry I'm stuck here having used up the term "swing" for swing....) to the regression model, you might find the correlation between swing and bias became apparent. In other words, having controlled for the (hypothetical) tendency for Bush voters to be keener to participate in marginal states, the underlying relationship between swing and bias is revealed.

However - this hypothesis depends on the assumption that something other than fraud contributes to the bias. In this hypothetical case, it depends on the assumption that Bush voters were less likely to respond to pollsters in non-swing states. Which is of course perfectly plausible if you believe in rBr. But if you believe in rBr, then aren't you a heretic?

And I'm not suggesting this happened - just that that's how suppressor effects can work. But it means, AFAICT, that the exit poll discrepancy is not a proxy for the magnitude of fraud. Or, if you will, that fraud is neither necessary nor sufficient to account for the exit poll discrepancy.


*For example, you might imagine that Bush GOTV efforts might encourage shy Bush voters to vote - but that they might still baulk at responding to a poll. In that case the better the GOTV in a precinct the bigger the swing - and the smaller the response rate from Bush voters. So a correlation De and Sv might be correlated, and result in a correlation between D and S - except that it doesn't seem to have done.

[eomer]

49. "you never have enough statistical power to conclude a null"

Right.

And not having enough significance to reject a null does not mean that you have proven it.

You seem to be saying we can conclude that fraud was not the cause.

I think the most you can really say is that we haven't been able to muster enough significance to conclude that it is not true that fraud was not the cause.

A tricky difference but one that means you are asserting something that is not supported if you follow the rules of statistical inference.

[Febble]

50. No, that is why

I am steering clear of the word "significance". We cannot say there was NO relationship. What we can do is put confidence limits on the extent and direction of that relationship.

There is a similar case with regard to the MMR vaccine. No research can prove that there is no relationship between MMR and autism. But research can tell you that IF there is a relationship, it is a very small one. Which, practically, means that autism can occur without MMR and MMR can be given without producing autism. And that MMR has no predictive power with regard to the development of autism.

The flat line in Mitofsky's plot does not allow us to conclude there is NO relationship between swing and bias. It does allow us to conclude that the amount of variance in bias accounted for by swing is a fraction of a percentage point - and we don't even know in which direction it goes.

But your point about complexity is also good - and requires me to modify the above statement. The flat line also allows for the possibility that in some cases fraud may have accounted for both swing and bias in the same precinct - but that if so, something correlated with fraud was pushing the bias in the opposite direction, leaving no net effect.

So to repeat: the flat line does not allow us to say there was no fraud. It does allow us to say that fraud cannot have been a net contributor to the bias. And a direct corollary of that is that the magnitude of the bias cannot be invoked as a proxy for the magnitude of fraud.

Fraud is compatible with that plot. The plot does not rule out fraud. But the plot indicates that the exit poll does not measure it.

[rdmccur]

99. Well and succinctly put

your 3rd sentence!

[rdmccur]

28. I think your 2nd paragraph

makes a good point. I will hold off any mental gymnastics on this until I see Freeman's response (which I reall would like to see); not the response he gave during the debate, but a more deliberate response.
Although I'm an amateur mathematician, I'm not a statistician by any stretch of the definition.

[OnTheOtherHand]

19. just a few thoughts

Mitofsky of course has no way of measuring possible vote miscounting in the 2000 election. However, the 2000 exit polls were fairly close on average to the official counts (for instance, the average WPE was -1.8, compared with -6.5 in 2004). So if one postulates that there may have been almost as much "bias" in the 2000 election as in the 2004 election, then one has a problem with the 2000 exit polls. (Also, if one thinks DREs are a big part of the story, then one would have to see if that picture holds up going back to 2000.)

Dunno whether Mitofsky looked at individual states. It's certainly possible that individual states could be dirty as all get-out, and the exit polls might or might not pick that up reliably.

[rdmccur]

29. On the other hand

(no pun intended) consider (in contrast to Mitofsky's new slide's trends) the plethora of statistics presented by du'er TruthIsAll and the arguments Freeman makes (some are essentially equivalent to TruthIsAll's) that point toward massive fraud. I'm withholding serious
consideration of the purported implication of Mitofsky's new slide until
I know more about it, including the data he used to produce it.
A conclusion can be elicited by manipulating or massaging data intentionally to give that conclusion. I've seen it done in my work (which is in the physical sciences, not the social sciences).

[OnTheOtherHand]

33. sorry, I hold the copyright on that phrase (grin)

If you care to cite any argument of TIA's that you have found persuasive, chances are that I have a rebuttal ready at hand. Not that TIA is always wrong -- in fact, his math is generally pretty careful. But in my opinion he is generally wrong about the most important stuff.

Since we haven't talked before (that I can recall), I should say that I'm not committed to arguing against "massive fraud" -- although I think that fairly interpreted, the pre-election polls put some limits on how massive it was likely to have been. In any case, I try to assess arguments individually, not to argue The Case For Fraud or The Case Against Fraud.

I don't find any of Freeman's arguments especially persuasive either, once we acknowledge that non-response bias in the exit poll is possible.

It's reasonable to postpone serious consideration of evidence you haven't seen, Speculations of deliberate "massaging" may be less reasonable under the circumstances; it seems to me that people should believe either that the exit polls prove fraud or that Mitofsky cooks the numbers, but not both. (EDIT: actually, I believe neither.)

[rdmccur]

43. A few issues:

The non-response bias you refer to may be possible but I am convinced (now anyway) that it is not very probable. I am much more suspicious of the people who have access the voting machinery and to tabulating the vote and therefore consider fraud as much more probable (not absolutely 100% likely though). This is based on a multitude of evidence and statistics I have read but have not itemized (I won't cite specifics; there's way too much out there that I've seen). The statistical analyses that I saw (many of them) from TIA were very thorough (my opinion). His premises may sometimes be called into question, but many of them cannot be resolved as true or not with the current information available.

I will add the caveat that massaging of data may not be 'consciously' intended (the person or persons responsible may not premeditate the
massaging). It's done to fit within some paradigm that they have adopted, if you will.

[OnTheOtherHand]

44. OK, but I have to say...

Non-response bias is very, very common, so if TIA gave you a different impression, you have been misinformed. Misinformation on this point has run rampant on DU and among election reform activists. I am trying to choose my words carefully here, and I am _not_ saying that I am at all certain that non-response bias explains the exit poll discrepancies (although I presently believe that it generally does). But a lot of people here seem to think that Mitofsky somehow made up non-response bias as a desperate maneuver to explain away his own damning results; that just isn't true.

I need to caution you that indeed, I am not aware of any of TIA's arguments that has much support among analysts. It's true that many of his premises can't be definitively refuted, but that isn't the usual scientific standard of argument. You should regard his arguments with the same caution with which I would regard, say, articulate and impressively detailed arguments for "scientific creationism."

(EDIT: I don't know what else to say. The view that an election outcome could be predicted with 99+% probability from polls that show the race practically tied is way, way outside the mainstream.)

There is considerable evidence for vote suppression and fraud, some of it good, some of it less good; in my experience it generally doesn't dovetail very well with the putative exit poll evidence, but it needs to be evaluated on its own terms.

[rdmccur]

52. I do not have any

knowledge of the pedigree of TIA's analyses in the view of other professionals (where has he gone anyway, haven't seen any recent posts by him?). I only considered the mathematics I saw in his work (and I admit that I am not a statistician, although I do have an MS in mathematics and work in a technical field). I am ignorant with regard to social sciences and political sciences' phenomenon of 'response bias'. I buy into Freeman's arguments about it. They are the best I've heard. But I would sincerely like to hear from you and others of other citations about 'response bias' that support what you are claiming.

But getting back to original issue of the new slides of Mitofsky,
I really want to hear from Freeman after he has had a chance to analyze how Mitofsky got his graphical result.

[Febble]

53. Well, as I said

the findings haven't been publicly quantified, but the E-M report

http://www.exit-poll.net/election-night/EvaluationJan192005.pdf

gives a number of factors that were associated with greater apparent bias, including things like the interviewing interval (how many voters had to be counted off before the next one was approached); distance of interviewers from polling place (further = more bias) and other interviewer characteristics, suggesting that factors that increased the chance of departure from strict random sampling protocol were associated with increased bias. They are only proxy measures of course, but all these hypotheses can only be tested through proxy measures. But in support of them, there is evidence for bias in UK exit polls despite the fact that we have pretty well assured clean counts (PBHC). Answers to your other questions can be found here:

http://www.democraticunderground.com/discuss/duboard.php?az=user_profiles&u_id=106042

http://www.progressiveindependent.com/dc/dcboard.php?az=show_topics&forum=120

[rdmccur]

56. Thank you

[OnTheOtherHand]

54. fair enough

Again, the problem generally isn't with TIA's math, so you're on solid ground there.

This link isn't exactly "social science" at all, but it has the benefit of being reasonably short and sweet:
http://www.stats.org/record.jsp?type=oped&ID=127
I can dig up some more detailed ones (or others may have links).

Yes, I too will be interested to see what Freeman thinks of Mitofsky's slide.

[Melissa G]

57. I don't know what Freeman thinks of Mitofsky's slide

But I would would be very curious to know what Mitofsky says about the 20 points that Freeman brought up that Mitofsky did not answer.

[Febble]

59. Well I know what I think of some of them

If you'd like to list, I'll have a go. I've got a copy of the slides, but didn't hear the actual points. One or two I simply don't understand (the paper ballot one, for instance). But some of the others were things I've thought about quite a bit (I even contributed to a couple....)

[Melissa G]

66. I sent Freeman an email to ask if he has posted it anywhere yet

Is there not commentary on the list serve you and OTOH are on?
Has the presentation been posted there?

[Melissa G]

75. Febble, Steve F said he sent it to you earlier today,..

Check your email and you should have the report in detail so to better opine..
mg

[Febble]

78. Well I got a report

and I've got the slides. I thought there might be 20 points I've missed. I'll look at what I've got.

I thought some of the points were interesting.

[OnTheOtherHand]

83. if you mean AAPORNET...

(the American Association for Public Opinion Research listserve), no, there has been no reference to the debate there whatsoever. I haven't even bothered to post there myself. As far as I can tell, there aren't many public opinion researchers who are very interested in the exit poll arguments; statisticians are more curious. (Presentations can't be posted on the listserve because it doesn't accept attachments.)

[Febble]

45. I'd like to know

(this is NOT snark) why you are convinced that non-response bias is not very probable. It is a question that genuinely intrigues me, and is probably at the root of division between people who are essentially on the same side. People who have done surveys are only too aware of the perils of non-response bias - it is almost impossible to avoid. In my own field we bypass it completely and describe our samples as "volunteer samples" - and just hope our results are still fairly generalizable.

The irony is that there is actually good evidence in favour of non-response bias, given in the E-M report, although the effects weren't quantified, whereas the evidence for fraud as a contributor has been thin, and I think actually contra-indicated by the ESI study and the analysis reported by Mitofsky in Philadelphia.

I often read on DU that "no evidence" has been given for non-response bias as a factor in the exit poll discrepancy, whereas actually there has been quite a lot. Certainly more than the evidence for fraud as a contributor (although not necessarily more than the evidence for electoral corruption per se).

[rdmccur]

77. My opinion is based

to a great degree on what Freeman has stated in his latest paper on the exit poll discrepancy. There are a few other analyses that I have seen (e.g. TIAs) that add more weight to Freeman's arguments. That's it, succinctly speaking.

I do not recall the details anymore, unfortunately, but I vaguely recall a study that showed that the hypothetical response bias (proposed by Mitofsky) did not behave in a way that would be expected (the trends across precincts did not make sense). This is almost too vague to bring up but I'm hoping perhaps someone else here has a better recollection of what I refer to. Perhaps Freeman has discussed this in his paper, I need to re-read it. It's late and I'm crashing mentally. Goodnight.

[Febble]

90. I presume you are thinking

of the USCV reports here:

http://electionarchive.org/ucvAnalysis/US/Exit_Polls_2004_Edison-Mitofsky.pdf

And here:

http://uscountvotes.org/ucvAnalysis/US/exit-polls/USCV_exit_poll_simulations.pdf

In which case you might like (!) also to read my own work cited here;

http://www.mysterypollster.com/main/2005/04/the_liddle_mode.html and here:

http://www.mysterypollster.com/main/2005/04/the_liddle_mode.html

Some people (including TIA) maintain the that original USCV position, and its current update are sound; others have adopted something closer to my position. But it's a pretty geeky battleline.

One summary of the debate is here:

http://www.mysterypollster.com/main/2005/06/uscv_vs_uscv.html

but the currently last word literally belongs to USCV though I still don't agree with it (see second link).

I've given links to Mystery Pollster as he has plenty of links to other relevant sources.

[OnTheOtherHand]

92. I think the second MP link was intended to be to

http://www.mysterypollster.com/main/2005/05/aapor_exit_poll.html

[Febble]

93. Yes, thanks

Sorry too late to edit.

[Febble]

42. Well I can tell you a little

It was based on the 1250 precincts that went into the section on WPE in the E-M report. This was a subset of the original 1480.

Since election day we have examined information from all 1,480 exit poll precincts in our samples, including all of the exit poll data from election day. This includes presidential vote tallies; questionnaires by demographics; refusals and misses by demographic, etc. There were only 20 precincts where we were unable to get the vote returns from the precinct in our exit poll sample.

The WPE values do not measure just the error of the exit poll in precincts that contain significant absentee vote. When absentees were greater than 15 % statewide, we removed precincts from this study that had the absentees merged with the precinct vote. In these precincts we cannot obtain counts of the election day vote separate from the absentee vote. Also, not included in this study are any precincts with fewer than 20 interviews as well as three additional precincts with large absolute WPE (112, -111, -80) indicating that the precincts or candidate vote were recorded incorrectly. Out of the 1,480 exit poll precincts, 1,250 were included in the analysis that follow.

(page 34)

Residuals were computed for each precinct by regressing Bush's share of the counted Bush + Gore vote in 2000 for that precinct, and Bush's share of the counted Bush + Kerry vote (mostly counted at the precinct at election night)in 2004.

So the residuals give you the degree of "swing", relative to the mean "swing". Precincts where Bush increased his vote share more than the average would have positive residuals, and precincts in which his increase was less than his average (or a decrease) would have negative residuals.

The plot referred to had the residuals on the Y axis and a measure of precinct-level exit poll discrepancy on the X axis. The regression line was flat, and the R squared was correspondingly tiny. In other words, the degree to which Bush increased his share of the vote was not correlated with the degree to which the exit poll apparently over-estimated Kerry's share.

[Time for change]

71. How about this

I think that this is similar to the points that Bill and eomoer have been making, but the wording is different, and perhaps simpler.

Consider a situation where there is no exit poll bias and no fraud, and Kerry has a narrow victory in Ohio.

Then, on top of that, you fraudulently switch 1.1% of the vote from Kerry to Bush, and the swithes are somewhat randomly distributed in various precincts throughout Ohio. That gives Bush a 2.2% official victory and a red shift of about 2.2%, all due to fraud. With that 1.1% switch of the vote, will that give you a statistically significant correlation between WPE (or bias index) and swing?

And if the answer to that is yes, what if, rather than the switches being randomly distributed they were targeted more towards precincts where the swing was small to begin with?

[Febble]

80. Well the better targetted

the fraud, the smaller it has to be, and the smaller it is, the less likely it is to show up as a correlation.

And the margin was tight-ish in Ohio, and tighter in NM, so I agree it wouldn't take much to swing the election, and that much (especially if some of it was voter suppression) might not show up.

But I do think that the plot argues against massive vote-switching fraud. I simply cannot see how massive vote-switching fraud could fail to show up as a correlation.

[OnTheOtherHand]

85. we need a better word than "massive," perhaps -- or a glossary

Freeman's presentation suggests that the popular vote margin was miscounted by over 8 million votes (one table implies perhaps 9-10 million); that instead of Bush winning the popular vote by about 2.5 points, Kerry originally won by about 4.6 points. I think "affecting millions and millions of votes all over the country" is more or less what you and I both mean by "massive" (and then there is the further qualifier "vote-switching").

[Time for change]

88. Then let me see if I understand what you're saying

You're saying that the data argues against "massive" vote switching

But not necessarily against enough vote switching to fraudulently switch Ohio to Bush and therefore the Presidency.

If so, I think that we're pretty much in complete agreement. :-)

[Febble]

94. Yup

I think Ohio could easily get under the wire (with enough of a shove).

[Time for change]

95. And also, it's important to remember that

the vote switching that was addressed by the analysis that correlated swing with exit poll discrepancy is only one of several mechanisms of fraud that could have been used in Ohio.

[texpatriot2004]

6. Anyone else have any info about this? nm

[Melissa G]

8. Just received an email from Freeman

He said he could put it up on his web page and maybe he can get a link on the ASA web page. He may also see about a video.

[IndyOp]

9. Please keep us posted! I am encouraged about the 52/48 vote -

from a bunch of academic stuff-shirts - that is good!

[FogerRox]

10. thanks OTOH

[Fly by night]

12. That was a joke, I think.

[IndyOp]

13. Thanks ---

I thought they had actually asked for immediate audience feedback.

:blush:

[eomer]

17. Yes, definitely a joke.

My bad, sorry if it wasn't clear. :spank:

[tiptoe]

30. That damned Census Bureau: Now look what TIA's done!

CLINCHER IV: CENSUS MATCHES NAT EXIT POLL (EDUCATION, RACE, GENDER, INCOME) Mon Oct 17th 2005, 06:04 AM (-4 GMT)

Of course, I mean the near-pristine National Exit Poll at the 12:22am time line (13047 respondents), NOT the Final at 1:25pm, which was highly contaminated when it was forced to match the (cough) recorded DRE-enabled vote count.

When we determined that the National Exit Poll (1:25pm, 13660 respondents) demographic "HOW VOTED IN 2000" Bush/Gore 43/37% weightings were totally impossible, simple logic dictated that all the other demographic weights and/or percentages had to be fiction as well.

If only ONE demographic which has been matched to the vote is proven to be impossible, then all other demographics, likewise matched to the vote, must also be impossible. A further bonus: we get to eat our cake and keep it, too, because that also proves that the final vote must be bogus...

(emphases mine)

[kiwi_expat]

32. Why assume that ALL 1:25pm demographics are "impossible"?

"If only ONE demographic which has been matched to the vote is proven to be impossible, then all other demographics, likewise matched to the vote, must also be impossible." -TIA

Everyone (Mitofsky, included) agrees that the final exit poll demographics are made-up ("fiction", if you like). But that doesn't automatically make them incorrect. Some might be correct and some might not be. Mitofsky probably focused on making some demographics plausible (e.g., gender) and neglected others.

[kiwi_expat]

60. A simple technique to make any count false.

Simply assign an incorrect demographic to a count and PRESTO the count, itself, becomes false. For example:

LET A= "WHAT I ATE FOR DINNER" DEMOGRAPHIC
LET B = VOTE COUNT

IF A = FALSE THEN
IF A = B
THEN B = FALSE


(I invalidated all of the 1988-2004 exit poll totals and vote counts, before breakfast.)


Please see illustrations of other clever techniques at
http://www.datanation.com/fallacies/index.htm

[foo_bar]

62. specifically denial of the antecedent

IF A = FALSE THEN
IF A = B
THEN B = FALSE

If A then B
Not A
Therefore, Not B

http://www.datanation.com/fallacies/deny.htm

[kiwi_expat]

65. I would say it is an example of "false equivalency"

(not on the list).

[foo_bar]

67. visual basic isn't the best way of expressing a syllogism

TIA was thinking of "If A = B, and B = C, then A = C". Instead, it evolved into the classic fallacy "If Not A, and A = B, then Not B":

If I am in Calgary then I am in Alberta.
I am not in Calgary, thus, I am not in Alberta.

http://www.datanation.com/fallacies/deny.htm

If A, then B
Not A; therefore not B
Fallacy of Denying the Antecedent (DA)

http://www.virtualsalt.com/think/deduhypo.htm

If the demo data is correct, then the vote count is correct (already a false premise, but consider it a->b)
If the demo data is incorrect, then the vote count is incorrect (~a->~b, "denying the antecedent", aka the logical inverse)

The correct form of the second statement is the contrapositive, "if the vote count is incorrect, then the demo data is incorrect" (~b->~a). But since the initial antecedent is flawed, the consequences are also meaningless.

[kiwi_expat]

68. Thanks!

I was missing (for the obvious reason) the implied assumption "If the demo data is correct, then the vote count is correct".

But now that you mention it, I have seen that formulation before.

I now understand the "clincher" arguments. ;-)

[foo_bar]

72. perhaps I was rationalizing it

"A = B" was supposed to speak for itself, "A= HOW VOTED DEMOGRAPHIC" and "B = VOTE COUNT". I assumed it was a syllogism, since there's no way to parse "HOW VOTED DEMOGRAPHIC = VOTE COUNT" in human or machine language (maybe COBOL).

[OnTheOtherHand]

34. if you're going to make TIA's argument, be prepared to make it!

There actually isn't much difference in most of the demographics between the 12:22am release and the 1:25pm release. The numbers that shift the most are the ones most strongly correlated with vote, of which recalled 2000 vote is of course an excellent example. (Another is party ID, which shifted from 38% Dem 35% Rep in the 12:22am to 37-37 in the final weightings.)

As kiwi points out, TIA's "simple logic" is simply wrong. His premise is misleading as well. I and others have presented ample evidence that people often misreport how they voted years earlier. Therefore, the "impossibility" of the reported vote figures -- in the sense that they don't jibe with the actual 2000 vote -- doesn't prove that the final 2004 vote must be bogus, _or_ that any of the other demographics are wrong.

Maybe the reason this 43/37 argument has failed to catch on, even though it was made within days of the election, is that most experts don't take it the least bit seriously -- and maybe they are right not to. Just a thought.

[Bill Bored]

40. How many votes PER PRECINCT on average would it take to swing

the popular vote enough to change the outcome?

And how many in each of the 50 states + DC?

And how many in just the swing states of FL, OH, NM, IA and NV?

Answer: NOT MANY!

Forgive my rounding errors but we have a 6% poll discrepancy in the popular vote which could be completely reversed by a 3% vote swing.

To change only the outcome of the popular vote would require only a (1.5% + 1) vote swing.

So how many votes per precinct is that?

NOT MANY.

If this number is large enough to be detected in these exit polls, then you might be able to argue that they don't prove fraud. But you'd also have to know that the exit poll precincts weren't known by the hackers so they could avoid them and thus avoid detection. I think this is unlikely, unless they were blind (no offense).

Now if the fraud were more concentrated, say tens or hundreds of votes per precinct, so few precincts would be required that the small sample polled in the exit polls would certainly not detect enough of those to change the outcome. This is why we have to use a hypergeometric distribution to select precincts to audit EVEN WITH A MUCH LARGER UNBIASED RANDOM SAMPLE THAT COUNTS EVERY VOTE to be able to prove who won a close election.

If the number of votes per precinct (or the number of precincts) to change the outcome of the election is small enough NOT to be detected in the exit polls, we're probably wasting our time to some extent by saying the polls do or do not prove fraud.

If TIA were here, I'd ask him to graph this one by state to see just how many states could have been hacked without being detected in these polls. But it might interfere with his belief of the polls' infallibility. Still, it would show that the election could have stolen without a trace and I'd think that would be an important finding for our side.

Peace.

[Febble]

51. Bill, I'm not sure about this

If a large vote-switch hack happened in exit poll precincts, I think it would show up in that correlation. And if it happened in unpolled precincts it would show up as redshift in the precinct selection - and precinct selection was good i.e. the polled precincts on the whole were representative of their state.

I'd like to know an answer to your question about how much vote-switching fraud could occur without detectable in the polls. My own efforts say not much, but I possibly have not been ingenious enough.

I still think the best way of stealing an election is through quasi legal means, as yowza points out, of which the cleverest (if it was clever not stupid) was allocating voting machines on the basis of "past turnout" and thus selectively capping turnout in low-turnout (i.e. Democratic) precincts.

[Bill Bored]

55. OK, let's take these one at a time:

"If a large vote-switch hack happened in exit poll precincts, I think it would show up in that correlation."

Sure, if it were correlated. But I believe there were 2 unexplained deviations out of 43 precincts in the ESI Ohio data. That's about 5% of the precincts. If this happened statewide, it comes out to about 500 precincts in which the discrepancy was not explainable by response bias alone.

Since we have no idea what happened in the unpolled ~10,000 precincts, who knows how many other discrepancies not attributable to response bias there would have been, or what they would actually be attributable to?

"And if it happened in unpolled precincts it would show up as redshift in the precinct selection - and precinct selection was good i.e. the polled precincts on the whole were representative of their state."

So we have roughly 43 out of 10,000 precincts with a vote margin representative of the entire state of Ohio. That's like auditing 0.4% of the precincts. Except it's not a true audit because the exit polls don't count every vote in these precincts, and the precincts weren't selected randomly. But even taking this into account, it still leaves 99.6% of all the precincts in the state pretty much unexamined, except for their official vote count. And that still means on average that only a few votes per precinct could change the outcome.

I may have some of these numbers wrong, but you get the drift. Just because the exit poll precincts correctly represented some demographics that mirrored those of the entire state doesn't prove that in the other 99.6% of the precincts there wasn't vote count corruption, does it?

And the same could apply at the national level.

And on edit:

"I'd like to know an answer to your question about how much vote-switching fraud could occur without detectable in the polls. My own efforts say not much, but I possibly have not been ingenious enough."

Well, to avoid an audit, you'd want to switch relatively large numbers of votes in relatively few precincts. Same if you're going to avoid exit poll detection, which might even be easier because the exit polls do not even select their precincts randomly! You'd know in advance that the polls were trying to mirror overall demographics of the state, but since votes are votes, you could switch them in precincts that deliberately would NOT mirror statewide demographics as the exit polls do, thereby evading detection -- especially if Mitofsky's precinct selection is as impeccable as everyone seems to think it is!

[Febble]

58. Keep going, Bill

I'd like to be convinced. I do think there is room for targetted fraud. But as I said, if it mostly happened in unpolled precincts, the precinct selection in the exit polls would be off. And if it happened in the polled precincts the WPE would be off. Which of course it was. But the WPE would also then be correlated with swing from 2000.

Unless there was fraud in that same precinct in 2000. But the WPEs were pretty close in 2000. And we are talking about NEW technology.

You see where I'm going? It's like a Rubik's cube (except I bet you can do those, I can't). You get one side the right colour and the other side gets in a muddle. I don't see how it can be vote-Switching AND new AND massive. It could be massive and not vote-switching. Or it could be vote switching and targetted under radar (i.e. small but effective). Or it could be plausibly deniable voter suppression. Or it could be electronic spoilage (new). But I don't see how you can have them all.

Tell me.

[Bill Bored]

69. A little of this, a little of that.

First of all, this whole discussion presumes that the exit polls are sensitive enough to detect fraud, and as you know, I don't necessarily subscribe to that view. So this is largely an academic exercise.

Next, you have not addressed the issue of the 2 precincts (5%) in the ESI report whose discrepancies were beyond the range of possible response bias. If rBr has been partially ruled out in 5% of precincts, and we do not know the true cause of this discrepancy, this could be significant, couldn't it? So refresh my memory on this if you would. Why couldn't there have been fraud in that 5% of the polled precincts in Ohio?

That said, you said:

"I'd like to be convinced. I do think there is room for targeted fraud. But as I said, if it mostly happened in unpolled precincts, the precinct selection in the exit polls would be off."

Or the results in the polled precincts were made to mirror the rest of the state after the fact. Mitofsky would certainly not argue with such flattery; makes him look brilliant!

"And if it happened in the polled precincts the WPE would be off. Which of course it was. But the WPE would also then be correlated with swing from 2000."

But would it? I would think that the Bias Index would be a better indicator since a shift from 2000 would change the precinct partisanship and the WPE would no longer be a particularly reliable metric.

And there is still the matter of the 5% of the polled precincts in Ohio that were so far out of whack that their discrepancies could not be explained by response bias alone. What are the alternate explanations for those again and exactly how do they rule out fraud?

"Unless there was fraud in that same precinct in 2000. But the WPEs were pretty close in 2000. And we are talking about NEW technology."

Well in Ohio there was not a lot of new technology. And furthermore, I think it's just as easy to steal votes with the new stuff as the old stuff.

"You see where I'm going? It's like a Rubik's cube (except I bet you can do those, I can't)."

Really? I'm surprised. Neither can I actually and I refuse to brake down and buy that book that shows how to do it.

"You get one side the right colour and the other side gets in a muddle. I don't see how it can be vote-Switching AND new AND massive."

Switching AND new is not a problem at least not technology-wise. Massive isn't necessary, but of course massive is a relative term.

"It could be massive and not vote-switching. Or it could be vote switching and targeted under radar (i.e. small but effective). Or it could be plausibly deniable voter suppression. Or it could be electronic spoilage (new). But I don't see how you can have them all."

"Tell me."

I "vote" for targeted vote switching and electronic spoilage. I'm not sure what you mean by spoilage though; I mean undervotes. Suppression would not show up in the exit polls, but of course it occurred as well.

Please respond to my questions about Bias Index vs. WPE re 2000/2004 correlations, and the 2 precincts in Ohio that were so far off. Did anything similar happen nationally (precincts with WPEs so large that they could not have been caused by response bias alone?)

[Febble]

81. Sorry Bill, missed this post

A quick response, maybe more later.

ESI limited their interpretation of "response bias" to literal "non response bias". Mitofsky has shown that interviewing interval is an important predictor of bias. This suggests that one form of "response bias" is actually "sampling bias" - one group of people selected for interview at a higher rate than the other. Once interviewers start to become selective (non random) about who they approach, completion rates may go up, but bias remains in the sample. You could have identical response rates from those selected (Bush and Kerry voters) but still have bias if Bush voters were less likely to be selected (if they looked less enthusiastic, perhaps). On the other hand, as you say, those precincts might have had fraud. Statistically we can't tell from Ohio alone.

Which is why the analysis on the whole dataset is important. 5% fraudulent precincts in Ohio might not show up if the underlying swing happened, for other reasons, to be a bit low in those two precincts. But nationally, it should show up as a correlation. If 5% of precincts had fraud then that 5% should also, on average, have higher swing. If the fraud was any good. (Aha! here's a loophole: if fraud was targetted at precincts where Bush was predicted to do relatively poorly, then maybe that might beat the system. Which presupposes Rove anticipated those scatterplots....)

Re Bias index - that was what was used - actually a better version, namely artan(alpha)-arctan(1), which Mark Lindeman calls tau. Tau was not correlated with swing, as defined as the residuals from Bush's vote share (Gore + Bush) in 2000 regressed on his vote share in 2004 (Kerry + Bush). Counted vote that is.

[eomer]

84. What about this scenario?

What if fraud was selectively applied to precincts that, before the fraud, were showing the most swing to the blue.

I remember there being some suspicion that certain precincts that were swinging to the blue compared with 2000 were tampered with to make the percentages come out the same as they did in 2000.

If you applied this approach to an otherwise clean slate and had no other competing effects to worry about, how would this show up in terms of correlation between swing and red shift?

It seems like swing in those fraudulent precincts would be zero since the fraud was designed to make it so and red shift in those precincts would be positive.

Would this kind of fraud slip through undetected in your correlation analysis?

[Febble]

89. Well I think it might

Let me think a bit more. That was sort of where I found myself going in response to Bill upthread (downthread?).

[Febble]

117. On further thought

one prediction flowing from your hypothesis does not seem to be supported, namely that on inspection of the plot, swing to Bush seems, if anything, to be negatively rather than positively skewed. If the algorithm was designed to stop Bush's vote-share dropping a certain level relative to 2000, there should be a positive skew - a shortage of precincts where he did really badly.

TfC is also working on some hypotheses that could fit.

Edit: forgot to add link to plots

[kiwi_expat]

61. Blackwell-initiated fraud could have avoided the NEP precincts.

I'm pretty sure Blackwell would have known which precincts were NEP precincts. The BOEs know which they are.

(And aren't the NEP precincts the same each election?)

[Bill Bored]

63. Even if he didn't know, all he'd have to know is that NEP precincts

would mirror the state as a whole. So he picks a bunch of precincts that would NOT do so. Call them the anti-NEP precincts.

Then he'd pick precincts whose vote count would be more than 3% of the county, to avoid the "random" recount.

So you end up with a bunch of un-polled, un-recounted precincts and you switch the votes there with impunity. How hard is that?

[kiwi_expat]

64. I didn't realize there were precincts that big!

"Then he'd pick precincts whose vote count would be more than 3% of the county" -BB

[Bill Bored]

70. Sorry, might be several precincts. Should I have said significant

fraction of 3%?

[kiwi_expat]

74. I seem to recall that one county picked the biggest precincts to recount

....and then topped them off with a few little precincts.

If Blackwell were trying to select precincts for fraud, he'd have to couple precinct selection with a "plausible deniability" cover, in case he got caught.


Cheers!

[Bill Bored]

76. Cover was that he was trying to save money by not recounting.

See, the $100,000 cost was never adjusted for inflation, so they had to avoid actual hand counts which would have cost a lot more than just the 3%. And they wanted to come as close to the 3% as possible without exceeding it.

[OnTheOtherHand]

86. "anti-NEP precincts"

I think this is a muddle, but maybe I'm just confused by the words. The NEP precincts collectively (and apparently) "mirror the state," more or less, because they are intended to represent all different kinds of precincts. It's not as if E/M identifies two kinds of precincts, the Bellwethers and the Oddities, and samples only Bellwethers -- that would actually be a real mistake.

I would follow kiwi's line instead -- how long in advance would Blackwell know which were the NEP precincts? No tengo la menor idea.

Bill, the other thing to keep in mind is that it isn't good enough to avoid quirky exit poll results; if you concentrate fraud in relatively few precincts, it will (or may) still show up in a Mebane-like historical analysis (depending on the magnitude and concentration)...

-- unless (following eomer's line) you can adjust the distribution so that fraud is focused in precincts where Kerry otherwise would have done better than expected. Best done on election night, assuming that someone has the ability to manipulate precinct results all over the state (but not necessarily in every county) without being detected; possibly done in advance by guesstimate.

This line of argument seems pretty labored to me, but maybe it can be articulated more sympathetically. If all we can say is that no one can refute vote-switching fraud in Ohio 2004, then maybe there isn't much point in talking about it -- except, of course, where we can say more about specific anomalies, vulnerabilities, etc.

[eomer]

87. "possibly done in advance by guesstimate"

or... wait a second.... :tinfoilhat: ... some algorithm programmed in advance that had target results embedded in it and would kick in only if necessary to achieve those targets... put hat away... :)

[Bill Bored]

100. My point was that ANY precinct distribution that didn't mirror the entire

state could be used confidently by vote switchers to avoid detection by exit polling. So that makes exit polling even less reliable than a random audit.

Just hacking Dem strongholds for example would work if a state is split 50/50 because NEP would be forced to select at least half of their precincts that were not Dem strongholds to gather their results.

A random audit by contrast would have no such restriction.

So an anti-NEP precinct fraud would be one in which even if the exit poll precincts were not known in advance, they could be avoided at least in part by concentrating the fraud in strongholds that do NOT reflect the entire state.

So the exit poll is really a very small (and incomplete) sample with a non-random bias that would not even qualify as an audit to detect election fraud. And yet we have people on both sides claiming it does and does not prove fraud.

After discussing this with Febble for a couple of days now, I think our whole problem here is that those of us who believe there are exploitable weaknesses in the election system, and with good reason, are tired of guys like Mitofsky and Scheuren appearing to defend that system, whether that is their intent or not.

If they are not educated about how the e-voting junk that's being put there works, they should admit that and defer to some others who are, or they should work with guys like Freeman, instead of against him, to bring this problem to light, regardless of what the exit polls say. As I said in the other thread, if they're not with us, they're against us. (And I'm not even an exit poll true believer!)

[OnTheOtherHand]

101. I don't yet follow your reasoning

If we assume that half of precincts are "Dem strongholds," it's true that a stratified sample would probably try to end up with exactly half such precincts, whereas a random audit might end up with 45% or 55% such precincts (or whatever, depending on the number of precincts in the audit). That wouldn't be decisive. What would be Really Bad, from the standpoint of trying to use the exit polls as an audit, would be to use only the same old precincts every time.

I think Mitofsky is pretty much damned if he does, damned if he doesn't. Freeman was throwing everything plus the sink at him as evidence of fraud. Mitofsky was basically sticking to the exit polls, which was after all the topic he had been asked to address. I don't remember exactly what he said about the last slide, but having been on the receiving end of that "with us or against us" business, I would cut him some slack. He did say very clearly at the outset that his argument wasn't about fraud, it was about using the exit polls to prove fraud -- or not.

That said, it's probably true that he doesn't know much about voting systems.

[Bill Bored]

105. He reportedly said, "This kills the fraud argument."

If that's being taken out of context, so be it.

But then why NOT advocate for further investigation?

I like the "with us or against us" stuff in this context because we are talking about the possibility of the loss of our Democracy.

I'm not asking anyone to compromise their scientific principles -- just advocate for free, fair, and verifiable elections. If you're in the exit poll biz, maybe that means devising a REAL exit poll that can be used to verify the outcome. If you think that can't be done, figure out another way to do it. I hear none of this from Mitofsky. Have I missed something?

[OnTheOtherHand]

107. nope, I didn't hear either debater advocate for verifiable elections

Maybe they took the debate topic too literally.

(I do think that what Mitofsky meant by "the fraud argument" was "the argument that the exit polls prove that Kerry won by millions of votes" -- or something along those lines -- not all arguments for fraud.)

[Bill Bored]

108. Were you there for the whole thing?

On pages 2 and 3 of his remarks, he states the problems of the elections systems:

<http://www.appliedresearch.us/sf/Documents/ASAP-Improbabilities%20&%20Neglected%20Correlations%20051014%20text.pdf>

Did he not present this?

Isn't it fair to say that by doing so, he is advocating change?

[OnTheOtherHand]

110. I'm sure he presented that, but no, I really don't agree --

He could have closed his remarks by hammering on the election system; he chose not to. I agree with MelissaG that he is coming from a different place, but then we should cut Mitofsky some slack on the same score. The speakers were invited to come debate the exit polls, and they did. I bristle at the idea that one side of the exit poll debate is inherently the progressive one. (If Mitofsky had spent time trying to rebut Freeman's comments about electronic voting, then we would have an issue.)

Several folks in the Q&A did draw implications for how the system should change.

My own opinion is that in the entire field of investigating 2004 and working toward better elections in the future, exit poll data are one of the least important aspects. (That's too bad for me, since I have spent a lot of time trying to understand them -- and I think Mitofsky still ought to do more to make the data available for indirect analysis, although I don't think there is a "clincher" in there.) Right now, I think EIRS data are one of the most underutilized data sources; there's a lot more to be done with precinct-level returns and, well, direct investigation -- there are plenty of anomalies to look into. You know a lot of the things that have to be done.

[Bill Bored]

111. EIRS

I wonder if there are EIRS data that have been misclassified. I.e., stuff that could be machine related that is somehow stated as something else.

And the per-capita problems in each state are interesting too. NY's were quite low as I recall. Lots of voters, not too many problems -- at least not machine related problems. But in Foo's neck of the woods (Brooklyn), Bush does seem to have gotten too many votes, doesn't he?

How about correlations between number of complaints per capita and vote margin, election outcome, etc?

Trouble with EIRS is that some people will always say it's biased, even more than the exit polls.

Did you have a chance to watch this video yet?:
<http://www.iwtnews.com/videoplayer/pamela_de_maigret/broadband>
I have always said that we need to work with people like this from the other side of the aisle, if she's still on that side.

I'll look at EIRS some more if you promise not to bring up the fact that not everyone had the 800 number!

Another idea would be to actively solicit voters who actually experienced machine glitches, instead of just relying on EIRS. For every one that called, there are probably 10 others who didn't. Were there any :puke: EXIT POLL questions about this? I know a fair amount of respondents thought there votes would NOT be counted as cast. But neither side in the EP debate talks about that very much which is a shame.

[OnTheOtherHand]

113. hey, nothing's perfect...

The fact that the EIRS and MSNBC data diverge so substantially indicates that they are far from perfect indicators -- but the fact that we can construct an index that is correlated with each of them at 0.7 to 0.8 indicates that they probably measure something.

My MSNBC incident index approaches correct-signed significance as a predictor of "surprise" (the gap between the pre-election polls and the official returns), which is a lot better than I can say for WPE. The EIRS index does somewhat worse but again at least has the right sign. So, my guess is that one can't drop in these DBs and expect them to "prove" anything, but they might have some interesting relationships with vote returns at e.g. the county level (and in some cases even the precinct level).

No, New York's reports are substantial overall. Eyeballing, it looks like it ranks about 6th in the MSNBC index (per capita), maybe 12th in the EIRS -- but this includes all incidents, so it may be that NY's are disproportionately polling-location questions that don't really signal trouble. I spot-checked Brooklyn -- it doesn't have many long-line reports, but it has a bunch of broken machines. In my armchair judgment as a former Brooklyn voter, the broken machines are about par for the course, but whether the emergency ballots got tabulated correctly is something I cannot know. (The state results were close to pre-election polls.)

No, the exit polls didn't ask anything about voting experience or confidence.

(And no, I haven't watched the video yet, bear with me.)

[Bill Bored]

115. They DID ask about confidence. (Do I have to reopen those dreadful PDFs?)

I think EIRS NY machine-only per-capita numbers were pretty low, but could be wrong.

No one else talks about the MSNBC stuff on this forum. How come? Did it have a lot of problems that favored Kerry or something? :)

[OnTheOtherHand]

116. oops...

I was thinking that the confidence q was only on one version of the national questionnaire, which would be pretty useless for most purposes -- but apparently it was on six state questionnaires: FL, GA, LA, MD, OH, and WA. Any questions about that? I'm happy to atone by running crosstabs if you are curious.

I see a total of 211 machine-only problems in NY compared with 280 in OH, 367 in FL... only 94 in TX.... NY is not super high overall, but I don't have a per capita ranking at the moment.

If the MSNBC stuff is linked to a database one tenth as informative as the EIRS one, I don't know where it is. The report I looked at from a link foo_bar sent me was pretty darn cryptic -- one couldn't tell who was favored. Does anyone know of other sources?

[Melissa G]

109. There was very little time to even make one's case at least for Freeman

I have yet to see Mitofsky's text and I would like to. OTOH, If you see it somewhere please post a link.

Freeman is Not an activist.. that is not his background. I see him as coming from a practical but very academic place. His writing has gotten a lot more accessible over the course of the year though....

[Bill Bored]

112. Might be interesting to do a "background" check on him.

What other projects has he done prior to this?
I ask just out of curiosity -- not with any agenda in mind.
(But then I've spent enough time on this stuff the last few days.)

[kiwi_expat]

102. re. "...even if the exit poll precincts were not known in advance"

As I indicated in post #61, there is reason to believe that the same NEP precincts are used each election.

A person in one of the county BoE offices told liam_laddie that the same precincts were used in 2000.

[Bill Bored]

106. Fine. Then it's even easier to avoid them.

But the counter argument is that the vote count in these precincts mirrored that of the entire state and the country as a whole. So unless this was coincidence, it means there was some hacking in the NEP precincts and some hacking in the non-NEP precincts. None of which is inconceivable, is it?

[kiwi_expat]

103. Exit poll precincts were definitely known to Ohio BoEs in Sept 2004

"Per a Sept 23, 2004 letter on an Edison/Mitofsky letterhead to BoE Director of Elections, J.M. Williams......The following seven locations are on the M/S exit-polled list..." -liam_laddie

http://www.democraticunderground.com/discuss/duboard.php?az=view_all&address=203x337293#340875

[Melissa G]

104. nice work Kiwi! n/t

[garybeck]

73. here is a report

Subject: Freeman/Mitofsky debate
Date: Mon, 17 Oct 2005 14:20:54 -0400
From: XXXXX
(name removed because I don't have permission to post his/her name. I'm sure it would be fine but just being safe....)



Here's the 5-minute summary of Friday's debate in Philadelphia between Mitofsky and Freeman:


See http://electionarchive.org/index.php?option=com_content&task=view&id=120&Itemid=66
for graphs


It was poorly attended. We tried in vain to get TV coverage - ended
up with 1 Washington Post and 1 BBC reporter, and 2 documentary filming
crews.
I urged Steve to take it seriously, and perhaps we overdid it in that
department - he was up past 3:00 the previous night, and his delivery
started off energetic but then got flustered, I think for sleeplessness.
------------------------------------------------------------------------
Steve started by running quickly through many reasons to be suspicious
of the election, making the case that the threshold of belief for a
corrupted election should not be high, that indeed the government should
have the responsibility of assuring the public by transparency and
openness that the election is clean beyond doubt.


He spent the bulk of his talk on what he knows best: the state-level
data from the 19Jan report, and the 5-division summaries of the data by
partisanship. He made the case from our report last March, and added to
it several 50-state distributions which are his unique contribution:
Republican and Democratic governors; proportions of blacks in the state,
complaints about election problems - all are positively correlated with
WPE, which we re-christened "PLD" for precinct-level disparity, on the
theory that we need not presume there's any "error". Steve put up a
histogram of PLD, which is not the one from our March report



but instead one that looked like this:

(not sure what's missing here....)

Gives you a very different feeling for the data - doesn't it? There are
2 differences between this new plot and the old one: (1) The WPE/PLD
numbers are based on what NEP had labeled "IM-WPE" in the 19Jan report.
Steve said these are the same averages but outliers not filtered out. I
didn't corroborate this, though I preparted the graph. (2) Steve and
Stephanie and Elizabeth and I decided it was justified to use the
binomial formula, which gives a fraction of a percent, compared with
2-3% used in the 19Jan report, derived from the std dev of the mean of
the sampled precincts in each state. I've argued that this is a large
overstatement, since the variance of these precincts is not statistical
scatter, but rather represents the diversity of the state, by design.


Steve ended by putting up the CNN screen shot from election night,
derived from the same NEP poll, showing how voters voted in 2004
compared to 2000. He made two points from this:


(1) a- Bush retained 90% of his 2000 base, but Kerry retained 91% of
Gore's 2000 base.
b- Bush got a minority of new voters and of Nader voters.
So - given that he lost the popular vote in 2000, where could he
have picked up enough votes to prevail in 2004?
(2) Despite the fact that the exit poll showed a Kerry victory, the
majority of pollees said they voted for Bush in 2000. This is another
indication that the poll was oversampling R's, not D's as in the rBr
hypothesis.


Throughout, Steve made an appeal for making the precinct-level data
public, which seemed to go over well in a room full of statiticians.
------------------------------------------------------------------------
Warren's talk was elegant and professional, but slower-paced, not at all
tightly reasoned, with much less data and most of his points rhetorical
and unsubstantiated. He wasted a lot of time talking about how there
was data released before the poll was complete, and how unreliable that
was. I thought that was utterly irrelevant. He went through the
factors that make for unreliability: inexperienced pollsters, and large
distances from the polling place. He said the response rate was low
because the form was too complicated and time-consuming, and that the
networks had insisted on a 2-page form when he thought an instant "Who
did you vote for?" would yield a much higher response rate, and
correspondingly greater accuracy.


At one point in the talk, he put up the scatterplot by partisanship
from last May and said that this analysis was done by Elizabeth Liddle
and Mark Lindemann, who formerly had believed that the election was
stolen, before they looked carefully at the data, and then changed their
minds. I thought that was an unfair characterization of Elizabeth's
views, and struggled internally whether I should bring this up in a
question - asking myself what she would have wanted. Fortunately, I
didn't have to make the point because Mark had driven down from
Kingston, arrived late, but heard his views mischaracterized, and was
able to speak for himself (and for Elizabeth), saying that Warren had
mischaracterized his view.


There was a new plot at the end of Warren's talk, which I haven't seen
before. It has a cluster of blue dots on
the left and red dots on the right. I couldn't see what the axes were,
and Warren didn't explain anything about how the plot was derived or how
he interpreted it, but stated baldly that "This plot kills the fraud
argument." (Later, in an email exchange with Steve, Warren told us what
the axes were in this plot, and how it was derived.*)


He reiterated his refusal to make data public on grounds of
confidentiality. (I have trouble with that reasoning, and apparently
others do, too.)


Finally, Warren said that Steve's WPE data was just wrong. He looked at
the numbers, compared them with the data from the 19Jan report, and they
were way off.
------------------------------------------------------------------------
In questions after the debate, Steve explained that his data came from
data labeled IM-WPE in the 19Jan report, and Warren continued to say
that the numbers were just wrong. The most important thing that got
debated in the question period, I thought, was that Steve said raw
PLD/WPE by itself is a telling statistic, being the difference between
what people said they voted for and how those votes were recorded. In
this sense, it's a pure measure of fraud, whether or not you can predict
the election with these data. Warren took the position that you can't
tell anything from the precinct-level data until you run it through the
algorithms and weightings that he has spent so many years developing.
(My opinion is that this is a blind spot for him. It's a rather subtle
point that the PLD/WPE is itself an indicator of fraud, whether or not
the exit polls or the precinct results are of any use in predicting the
election. It may be that Warren has spent so many years learning to
predict election outcomes that it's hard for him to conceptualize a
statistical test that bypasses prediction entirely.)


* Here's the story on the plot that "kills the fraud argument": It was
prepared by Elizabeth as a national version of a similar scatterplot
done for Ohio by ESI. On one axis is the ratio of votes, by precinct,
2004/2000. On the other is arctan(alpha) - which is Elizabeth's
rationalized measure of PLD. There is no significant correlation, and a
regression line through the "blue-shift" precincts is indistinguishable
from the regression line through the "red-shift" precincts. Since WPE
was modest in 2000 (not distinguishable from zero, by some measures),
this is an argument that whatever the exit polls are saying to us about
2004, they can't be taken as a demonstration of fraud.


I've already begun to think about this argument, how it might be
reconciled with some of the other things we know that indicate Democrats
were not oversampled in the poll, and therefore that the 2004 exit poll
/is/ a reliable indicator of fraud. Were outliers removed before the
data even got to Elizabeth? Was the 2000 data "pure" or mixed in with
pre-election polls? Were outliers removed from the 2000 data?



forwarded to 'garybeck' by another election enthusiast by email.

[Febble]

79. Gary

The description of the Mitofsky plot is not quite right. There were two. One plotted vote share in 2000 against vote share in 2004, and identified precincts as either redshifted or blue shifted. This is the one where the regression lines were not significantly different. The other showed the residuals from that regression on the vertical axis and bias (a version of my measure, devised together with Mark Lindeman) on the horizontal axis. In this one the regression line was flat.

The precincts that went into the analysis is the 1250 subset used in the E-M report. More info in this post:

http://www.democraticunderground.com/discuss/duboard.php?az=show_mesg&forum=203&topic_id=397074&mesg_id=397399

It was based on the actual response data, not projections. No outliers were removed from the 1250 dataset.

(edited to remove question regarding provenance of email)

[Febble]

82. Oh, and

Mark and I worked on the measure (and continue to do so). Mark didn't do the actual analysis.

[Melissa G]

91. And Here is the real thing from Freeman's site...

Starting a new thread here...http://www.democraticunderground.com/discuss/duboard.php?az=view_all&address=203x397502

Here is the direct link
http://www.appliedresearch.us/sf/epdiscrep.htm