NEDA's Ohio Exit Poll Analysis Withdrawn

[sunshinekathy]

Release: National Election Data Archive

NEDA retracts Ohio Exit Poll Analsysis

November 03, 2005

Dear Fellow American Patriots,

The National Election Data Archive's paper analyzing the Ohio exit poll data has been temporarily withdrawn due to an error in the assumed definition of data given in the ESI report. In other words, NEDA’s Ron Baiman incorrectly assumed that ESI was giving the maximum and minimum of “vote” counts in its table of exit poll data, and ESI was actually giving the maximum and minimum of “exit poll results” if all non-responders to the exit poll in its sample had hypothetically completed the polls.

A solid independent analysis of Ohio's exit poll studies still must be done. On October 31st, the Election Science Institute's own analysis of its Ohio exit poll data was mathematically proven to be invalid.
http://uscountvotes.org/ucvAnalysis/US/exit-polls/ESI/ESI-hypothesis-illogical.pdf

NEDA’s analysis, using the corrected definition of “non-responders” will be redone and reposted - but our volunteer schedules do not permit us to do this soon.

A copy of ESI's data can be found here. NEDA welcomes any independent analysis of this Ohio exit poll data.
http://uscountvotes.org/ucvAnalysis/US/exit-polls/ESI/ESI-Mitofsky-files/ESI-OH-data-2004.pdf

If you would like to be notified when NEDA's paper is re-released, please email either:

[e-mail]
or
[e-mail]

We apologize for any inconvenience our error may have caused.

Sincerely,
Kathy Dopp, [e-mail]
National Election Data Archive, President

This public withdrawal of our Ohio exit poll analysis paper can be found at:
http://electionarchive.org/ucvAnalysis/OH/Ohio-Exit-Polls-2004.pdf

I want to thank Febble for pointing out this error at the DU. I had discussed the possibility of misinterpretion of ESI's definition of "nonresponders" with Ron earlier and had been verbally assured by Ron that he was interpreting it correctly. But I take full responsibility for not having carefully read for myself the definitions under the table of data provided by ESI.

The Ohio exit poll data is still highly suspicious, but is probably not "virtually irrefutable" evidence. Unfortunately we won't have time to complete this analysis now for several weeks, so perhaps other independent analysts may beat us to the punch on this particular job.

[Deep13]

1. Okay, I don't know much about this stuff.

Can you explain what this means in brass tacks?

[On the Road]

2. Does This Account for the Competing Claims

on whether certain precinct results were mathematically impossible according to the exit polls?

Some DUers were saying the vote count could not be reconciled with the exit polls under any circumstances, while others were saying that selection bias or unequal refusal rates could easily have accounted for the poll-vote discrepancy.

It's easy to read over unless you're drilling into this stuff, but it's actually quite a big error to make.

[OnTheOtherHand]

3. it's somewhat less stark than that

One can't say (I think this is now agreed) that "the vote count (can) not be reconciled with the exit polls under any circumstances."

I would not choose to say that "selection bias or unequal refusal rates could easily have accounted for the poll-vote discrepancy" (my emphasis). It's easy in the sense that one can often do it with hundreds of votes to spare.

Yes, I would say it was quite a big error, and that's why the paper was withdrawn. YMMV.

[On the Road]

27. I'm Glad That There's Starting to be Some Agreement

The size of the discrepancy between exit polls and votes should have set off alarm bells all over the place. At the same, once you start taking a hard look at how the exit polls were done and what can actually be proven, it's not enough to publish a claim of statistical proof of fraud. I actually think it's a good thing that those claims didn't get wider circulation lest a meme be built up that the fraud hypothesis has been debunked.

A lot of energy has been wasted on arguments among people who all suspect fraud and who all want Bush to be caught red-handed. Hopefully the focus can start to shift to some of the harder, more specific evidence that is more likely to uncover a smoking gun.

[Febble]

28. I hope you are right

I feared that precisely that meme did get built up, and that I was at least partially responsible for it (in that I tried to debunk the statistical evidence for a strong claim that the exit polls indicated fraud).

On the other hand, my POV has always been that modest claims backed up by good evidence have more power than strong claims backed up by weak evidence, precisely because they are less vulnerable to debunking - and I have tried to promote those modest claims.

If in doing so I have contributed to the "fraud hypothesis was debunked" meme, as opposed to the modest "the election was flagrantly unjust, unverifiable and may have been actually stolen" meme, I accept responsibility. It's the second I have been trying to put about.

The tragic split has been between the "exit polls are smoking gun for fraud" meme and the "exit poll evidence proves there was no fraud" meme, both of which are bunk IMO, and need debunking.

[Land Shark]

4. Different technologies, different procedures and properties

Technologies bring with them different properties. Note how electronic voting machines bring with them, just by signing a purchase contract and implementing them, the destruction of various checks and balances in elections.

The internet has its own effects. As opposed to conversation for example, one can go ON AND ON with a tangent that, if stated in normal conversation, would have been interrupted with a correction at an early point. Instead, the email form becomes a "flame" and the internet gets a reputation for being relatively rude.

The relevance here is that web-publishing these papers is a sort of peer review process (OF SORTS) that is relatively new, but is not nececssarily true peer review. yet, it has a function similar of quality review.

This new system WORKED here, catching an error. The paper will be re-written and will still have good evidence.

[Febble]

5. That's a very generous post, Kathy

And it has to be said that that ESI chart had my eyes crossed for quite a while before I figured out what it meant.

And I should add I didn't actually spot that Ron had got it wrong until OTOH pointed it out.

But thanks anyway.

Lizzie

[sunshinekathy]

6. Not generous - just truthful

Thanks for your kind words.

Now it would be nice if ESI and Mitofsky likewise would withdraw their analysis which they have wrongly let stand not just for 24 hours like NEDA did, but from June to November thus far.

NEDA's prior math proof of the invalid logic of the ESI and Mitofsky analysis still stands as 100% correct, yet I've not heard ESI or Mitofsky withdraw their prior analysis yet, as they rightly should.

Here is our press release on that:
http://electionarchive.org/ucvInfo/release/ESI-hypothesis-illogical-press-release.pdf

Perhaps you could pass it on to the folks at ESI and suggest that they withdraw their analysis.

I'm sure eventually Ron will get around to putting out a valid analysis on the Ohio exit poll data. I have no time to work on it for at least a few weeks myself.

It is likely that the Ohio data is indicative of vote fraud and highly suspicious, but that no "virtually irrefutable" evidence of vote fraud exists.

[sunshinekathy]

7. Clarification on our withdrawal

On November 2, 2005, NEDA released an analysis of the 2004 precinct level Ohio exit poll data entitled “The Gun is Smoking: Ohio Exit Poll Data Provides Virtually Irrefutable Evidence of Vote Miscount”. The analysis used data provided in the Election Sciences Institute (ESI) report of June 6, 2005.

On November 3, 2005 NEDA realized that its interpretation of the definition for the term “nonresponders” to the exit poll, as used by ESI, was incorrect.

What this means is most likely that the Ohio exit poll data is:

1. inconsistent with voter exit poll response explanations as put forth wrongly by Mitofsky in his Jan 19th paper

2. highly suspicious and very consistent with vote fraud explanations

but does "not" contain "virtually irrefutable" evidence of vote fraud.

In a positive sense, notice that NEDA withdrew its incorrect analysis within 24 hours, whereas ESI and Mitofsky, whose analysis has clearly been proven mathematically to be balderdash, has yet to show the integrity to withdraw its own analysis, which it has wrongfully let stand since June, and presented in October to the American Statistical Association fall conference in Philadelphia.

Anyone who has used our press release from yesterday, please substitute this press release instead:
http://electionarchive.org/ucvInfo/release/ESI-hypothesis-illogical-press-release.pdf

It is very important to get the word out that the Mitofsky/ESI analyses of Ohio exit poll data are bunk.

http://uscountvotes.org/ucvAnalysis/US/exit-polls/ESI/ESI-hypothesis-illogical.pdf

This above math proof still stands as 100% correct, has been reviewed by PhD mathematicians and physicists, and the "under construction" in its appendix should be removed from the appendix as we've checked it and found it correct.

[Febble]

8. Mitofsky/ESI hypothesis NOT bunk.

Kathy's paper correctly demonstrates three things:

1. that fraud need not result in Bush doing better than he did in 2000.

2. that the the Mitofsky/ESI hypothesis is not a test of whether there was fraud.

3. that if fraud were concentrated in very few large precincts it could do help Bush a lot, without producing a correlation between redshift and swing.

However, these are all straw men.

The hypothesis Kathy attempts to dismiss as BUNK simply states that Bush will do better in precincts where there is pro-Bush fraud. This is self-evident. It does not state that he will do better than in 2000 where there is fraud. This is self-evidently false, as Kathy demonstrates.

And what the hypothesis tests is whether fraud was the cause of the exit poll discrepancy, NOT whether there was fraud. Many forms of fraud would not produce much or any discrepancy in the exit poll, for example fraud concentrated, as Kathy suggests, in very large precincts.

To my knowledge, neither Mitofsky nor ESI have claimed that the hypothesis is a test of fraud. They have claimed that the results strongly suggest that fraud was not the cause of the exit poll discrepancy. This is a quite different claim from the one alleged by Kathy.

In fact, both the ESI and the Mitosfky analyses are extremely informative, and, if read rightly, could help us delineate where fraud is most likely to have occurred, and what form it might have taken. Ironically, it has already led Kathy to the interesting hypothesis that fraud might have been concentrated in large precincts.

Inviting people to discard perfectly good analyses because of the way some have interpreted them (and NOT, to my knowledge, their authors) seems to be a great way of shooting oneself in the foot.


(edited for grammar)

[eomer]

10. Could you make it a bit less simple?

What exactly is the hypothesis again? You said:

The hypothesis Kathy attempts to dismiss as BUNK simply states that Bush will do better in precincts where there is pro-Bush fraud.

More specifically, will do better than what? Better than he otherwise would have done? Obviously that would be true but there is no way to ever prove or disprove that one so it is not a very useful hypothesis.

Not trying to be cute, I really would like to know and for some reason can't open the ESI paper right now.

[Febble]

11. Well that's the broad hypothesis

What we need to do next, and what both ESI and Mitofsky did, was to account for some of the variance in "better". Clearly Bush will do better anyway where he has traditional support. ESI therefore baselined his performance relative to his performance in 2000; Mitofsky did essentially the same in a slightly different way.

So on ESI's Y axis we have the difference between Bush's vote share in 2000 and his vote share in 2004. A value of zero means he his vote share stayed the same. A positive value means he did better; a negative value means he did worse.

On Mitofsky's Y axis we have the residuals from the Bush's vote share in 2004 regressed on his vote share in 2000. In plain English this means that a positive value means Bush did better than his average gain on 2000 (because it was a net gain, in counted votes) and a negative value means he did worse than his average gain, which could include an actual loss.

But the point of "covarying" for vote in 2000 is that it gives us a way of accounting for variance due to variance in his underlying base, leaving residual variance that might be due to fraud. But note that the hypothesis would work just as well if in fact Bush's vote-share had dropped (imagine if Bush had dropped his vote share despite widespread voteswitching - "swing to Bush" would still be positively correlated with redshift, even though most of Bush's swing would be in the negative range (i.e. a drop in vote share).

Now it is possible that 2000 is not a good baseline, as there may have been fraud in 2000 - but the precinct-level discrepancies in 2000 were fairly low, so as a measure of the kind of fraud that induces redshift (and there are many that wouldn't) it is not a bad baseline for the hypothesis that fraud caused the exit poll discrepancy in 2004.

The general point about the General Linear Model, which is a lovely tool, is that you are essentially accounting for variance. There is huge variance in Bush's voteshare - but a lot of it, as is obvious from both ESI's first plot, and Mitofsky's first plot - can be "accounted for" by his voteshare in 2000. So we then test to see how much of the "residual" variance - that not accounted for by vote-share in 2000, can be accounted for by redshift. And the simple answer is: zilch.

Now there are loopholes. You found one. Kathy found one. I thought I'd found one, then it didn't quite work, and now OTOH is on to another one. Anax promised one, but I think he's still clearing up after Wilma. The thing certainly isn't bulletproof - but at least it tells us what we are up against. Which is why I would really rather it was not dismissed as BUNK!

On edit to try and make this clearer:

The hypothesis, in essence, says that if fraud results in extra phantom votes for Bush, two things will happen: Bush will do better (than otherwise) and the exit polls will show redshift. Redshift therefore ought to correlate with Bush doing better. But that is difficult to measure, as a lot of other things will determine how well Bush does in a precinct, including whether he has a base of support. So controlling (covarying) for his base of support will soak up some of the extra variance in Bush's performance, allowing us to see more clearly whether redshift also correlates with his performance.

[eomer]

12. Thanks, but that doesn't seem like a hypothesis to me.

Shouldn't a hypothesis be a provisional explanation that fits all the facts we've observed so far? Then we look for more facts. If all the new facts are consistent with the hypothesis then we provisionally keep it. As soon as we find one fact that is inconsistent with the hypothesis then we are forced to abandon it.

If that is what a hypothesis is then here are some possible hypotheses:

• The exit poll discrepancy is caused by fraud in the official count.
  
• The exit poll discrepancy is caused by polling error.
  
• The exit poll discrepancy is caused by innocent error in the official count.
  
• The exit poll discrepancy is caused by fraud in the exit poll.
  
• The exit poll discrepancy is caused by some combination of the above causes.
  

From what I can tell, all of those hypotheses can still be provisionally put forward since no one has come up with a fact that disproves any of them.

The ESI paper claims a fact and then says that their fact rules out the first hypothesis in the list. The fact they claim is that there is no correlation between redshift in the exit poll and precincts in which Bush did better in the official count. The problem with their logic is that, while their fact does seem to be in tension with the first hypothesis, it is not demonstrably inconsistent with it. Or at least ESI hasn't provided such a demonstration with enough specificity to warrant discarding the hypothesis. Until they do, the hypothesis stands.

On edit: I used a phrase that I would like to amend. ESI does not claim to have "ruled out" the first hypothesis. In fact they specifically say that further analysis is needed in order "(t)o rule out important errors or irregularities within the election system". Rather than "rule out", they use words like "strongly indicate" and "is much more likely". Thanks to OTOH for pointing this out in another thread. That doesn't, however, change my position that the first hypothesis still stands and in fact strengthens that position since ESI's statement against it is a bit softer than I thought originally.

[Febble]

13. The hypothesis certainly leave open

others. In fact that's what hypotheses are for.

Statistically, the null hypothesis was:

Swing is not higher where shift is redder.

If the null was rejected, then fraud would be a fairly justifiable inference.

If the null is retained, as it was, then we need to put confidence limits on the size of the relationship. For ESI's study, the confidence limits are fairly wide (bigger loophole for the fraud explanation). For Mitofsky's it is pretty narrow.

However, alternative hypotheses might be: fraud was targetted at, only large precincts; fraud was absolutely uniform in distribution and magnitude (yours, I believe); fraud was targetted only on precincts where Bush was known to be doing badly.

If these can be translated into testable statistical hypotheses, I think we can test them. If not, then we can certainly look for on-the-ground evidence that supports them.

Regarding your hypotheses: ESI tested a prediction flowing from your first. What we need are testable predictions flowing from the rest.


(edited for grammar!)

[eomer]

14. A minor quibble and then back to the main point.

First the minor quibble - the hypothesis that you credit to me should be slightly modified. It should be: fraud was absolutely relatively uniform in distribution and magnitude. Fraud doesn't need to be absolutely uniform to sneak in under the radar, just relatively so (relative to the variance of the data onto which it is applied).

Other than that we seem to be on the same page. I would just bring up a point we have discussed before - that there is likely more complexity than we are accounting for. You say that ESI tested a prediction based on my first hypothesis. I would change that a little bit and say instead that they tested a prediction based on a simplistic version of the first hypothesis. And I would add that any simplistic model of any hypothesis can probably be shot down. When it is we are not ready to abandon that hypothesis. Instead, the next step is to try a more complex version of it. Otherwise we will quickly run out of hypotheses and have to conclude that the election can't have occurred the way we know it did.

Edit: fonts

[OnTheOtherHand]

15. I think ESI would say that those are theories, not hypotheses

page 2 of their brief: "Two explanations for the discrepancy remained. The fraud accusation theory assumes that there was substantial and systematic fraud in the counting of the votes, and that such vote-shifting favored President Bush; in other words, the official results were wrong. The non-response rate theory assumes that Bush voters were less likely to respond to the interviewers than Kerry voters; in other words, the exit polls were wrong. It is also possible that irregularities or errors exist within the election system and that the discrepancy resulted partially from this, or a combination of factors."

So, a theory is a causal account, and a hypothesis is a testable prediction that gives some leverage on assessing the plausibility of the theory. But very rarely, in social science at least, does one find a killer hypothesis such that the outcome of the test ends the debate over theory.

We agree that what they call the fraud accusation theory is still tenable, or at least that some elaboration of it might be tenable. Where I decisively part company with Dopp is over the notion that a single counterexample can render the correlational analysis simply irrelevant -- no, worse than that, "bunk." I think very few social scientists will (or should) take seriously an article that gussies up "lack of correlation doesn't prove non-causation" in a bunch of high school logic symbology and presents it as a definitive refutation.

[eomer]

16. "some elaboration of it might be tenable"

That phrase sums up my position.

I mispoke in a reply to Febble above. I said we would quickly run out of hypotheses if we only test simplistic versions of them. When you think about it, we have used up about a year and have now seriously discredited two overly simplistic hypotheses, uhm, sorry, theories - first simple rBr as a sole cause and second simple fraud as a sole cause. Not exactly a blazing pace. I'm hunkering down because I know BillBored is going to come over and give us all a good flogging any minute now for frittering away our time on the great exit poll debates.

[Febble]

17. I'm not sure "we" have discredited

"rBr" - much though I dislike the term - I'd rather call it "polling bias" myself, which would encompass non-response bias (Bush voters refusing at a higher rate that Kerry voters) and sampling bias (Kerry voters being selected at a higher rate than Bush voters, either because Kerry voters "volunteered" and were not rejected, or because Bush voters were more likely to escape selection). There is anecdotal evidence to support both these.

Bias in the poll may well not be the sole cause, but what* is your basis for considering it "discredited"?. I know it is often stated on DU that it has been discredited, but I do not actually know why. The E-M evaluation, contrary to assertion by some, does actually provide evidence in support of it as an explanation.

But I hasten to add: even if could be shown that all the discrepancy could be attributed to polling bias, it would not rule out fraud. And it would make not one jot of difference to what I consider statistical fact that voting machine rationing in Franklin County, Ohio, resulted in a substantial net loss of votes for Kerry in a key swing state.



*This is not a "cute" question either - I'd really like to know why you consider it has been (maybe you could convince me - we seem to be in a season of retractions).

[eomer]

18. The words "simple" and "sole" are key to what I am saying.

"simple rBr as a sole cause"

What I am saying (and let me change it to "I" rather than "we") is that simple X as the sole cause of the exit poll discrepancy is most likely not the case. Fill in anything for X you want.

If we construct a testable hypothesis for the rBr theory (thanks OTOH), we have to give it some form that we can express mathematically. We can't allow ourselves to say that it will take any form it needs to because that is effectively begging the question. Once we put it into the form of some simple mathematical formula then it fails to reproduce the data all by itself. So if rBr is going to be the explanation then it needs to have a complex formula, not a simple one, or else it needs some simultaneous help from some other effect.

[Febble]

19. Why could it not be the sole cause?

If we define "rBr" fairly simply: that Bush voters were less likely to participate in the polls than Kerry voters (which I think is how the E-M report phrases it) - what undermines your belief that this cannot be the sole cause?

Again, this is a serious question.

[eomer]

20. Because...

that is not a simple definition, it is a loose one.

A simple definition, in the way I mean the word, is one that can be expressed as a simple linear equation or some such.

The essence of the discussions earlier this year about rBr is that USCV constructed a hypothesis using a simple formula for rBr and found that that simple definition did not reproduce the data.

Your definition is neither simple nor complex, it is amorphous.

[OnTheOtherHand]

21. I would agree that no "simple" theory accounts for the data

For better or for worse, that is almost irrelevant. Most analysts aren't up nights worrying about how complex a theory is needed to account for the data. At most, they worry about big questions, such as: do the exit poll data 'point to' (deliberately woolly) fraud, and if so, where? or, do the exit poll data point to big methodological problems, and if so, can we fix them? and, to what extent (if any) do exit polls need to be improved in order to serve their purposes? (That last depends, of course, on the purposes.)

USCV refuted a "simple" theory that AFAIK no one held, and was left to wonder why the refutation had so little impact. I don't think that work's Straw Man Problems were as bad as the more recent work's, but the basic outcome was the same.

[Febble]

23. Ha!

We are simultaneously disagreeing again!


On edit:

I get it. You are a social scientist, I am a psychologist. I'm focussed on the postulated underlying phenomenon (Bush voters being more reluctant to participate); you are focussed on the methodology that might have allowed it to affect the survey.

The first is essentially simple (a single behavioural phenomenon, however complex the cause) and the second is complex (aspects of survey methodology).

I agree the second is more useful for generating testable hypotheses. But I'd maintain the first is essentially simple.

[eomer]

24. Some analysts are indeed up nights ...

worrying about how complex a theory is needed to account for the data. I'm an actuary and I would say that is, in a nutshell, what actuaries do (and unfortunately they are often up nights doing it).

For example, actuaries model the simultaneous effect of multiple decrements to explain the behavior of a closed population of lives. It wouldn't be very useful to have at hand a population we know to be subject to multiple decrements and then try to explain its behavior by modeling just one of those decrements. Our result would never approximate very closely the actual behavior of the population.

But that is what we seem to be doing once and again in this discussion.

So we should either abandon modeling all together or we should move past the simple models and into the complex ones.

[OnTheOtherHand]

29. I meant the exit poll data in particular

Of course, I have no way of knowing how many people may be up at night thinking about the exit poll data, but not a whole lot of interesting ideas have made it into the open.

[Febble]

22. Actually it IS quite simple -

what is amorphous are the ways of measuring it. Like so many things, what we are really interested in (whether non-participants were a "different population" of voters from participants) is impossible to measure directly. So we have to devise a series of proxies.

So to use OTOH's nice distinction between theory and hypothesis (Brits tend to use these more inter-changeably than you logical Americans), our "theory" is that Bush voters are more likely to want to avoid participating than Kerry voters. OK. Given that our theory is that they will "want" to avoid participating, we might hypothesise that the greater the chance they are given to do so, the more they will. This can then be turned into a testable linear prediction. We consider some attribute of a precinct that will allow "reluctant" voters to escape more easily (say the pollster is more than 25 feet from the precinct). If Bush voters are more reluctant than Kerry voters, we would expect greater redshift in these precincts than in precincts where the pollster was within 25 feet of the precinct or actually inside the polling place.

And indeed, this is what E-M found. They also found greater red-shift in precincts in which other factors likely to make "escape" easier for "reluctant" participants - low interviewing rate (e.g. interviewer selects every 10th voter, rather than every voter or every second voter). And again, interviewing rate correlated with redshift.

So their theory is "supported" by the results from testing specific hypotheses concerning the relationship between precinct characteristics and redshift. We do not know whether the hypotheses accounted for all the redshift, because that information is not provided in the report. But as these measures are only proxies for the hypothesized underlying "cause" (greater reluctance of Bushies to participate in the poll) there is no particular reason to expect that they would, although they might. But they do support the theory itself.

In contrast, we have a hypothesis (ESI's and Mitofsky's) flowing from the theory that the exit poll discrepancy was due to fraud - the hypothesis is that redshift will be greater where Bush does better relative to his 2000 performance. The result from both these tests is negative. Neither study finds an association between the two measures. Thus the theory that the exit poll discrepancy is due to fraud not supported. We may make a new theory - that the exit poll discrepancy was caused by fraud targetted at, say, precincts where Bush was anticipated to do badly. And from this we can, perhaps devise a testable prediction - a new operational hypothesis.

And testable hypothesese like these are what I personally have been trying to solicit for the last few months!


(edited to replace accidentally deleted phrase and a bit of grammar....)

[eomer]

25. Another testable hypothesis is...

that voters will be more reluctant when they feel they are in the minority (in hostile territory) and more willing when they feel they are in the majority (home turf).

This prediction was tested by USCV and the prediction failed.

So we have dueling predictions from the theory. One fails, the other succeeds. That leaves us uncertain whether the theory is the true one or not.

Edit: wording

[Febble]

26. Yes, there seems no evidence for that

but I'm not quite sure why it was advanced. I'd maintain that Mitofsky's formulation, if not proven, is supported, and not challenged by anything within the exit poll data itself*. If "rBr" accounts for the net redshift, then it would appear to be distributed orthogonally with regard to Bush's precinct vote-share.

USCV argued that bias was higher in precincts with heavy Bush support, and some (not all) of the original authors still maintain this, but my own view is that this was not supported by subsequent analysis.

This was why my notorious "Fancy Function" was widely regarded as "supporting rBr", whereas it merely demonstrated (IMO) that an alternative hypothesis (that redshift was greater in Bush Strongholds, thus suggesting fraud) was not supported.

And, to keep repeating: even if "rBr" was responsible for the entire redshift, it would not rule out fraud. Plenty of types fraud won't result in redshift, and nor will many of things that we know deprived Kerry of votes in Ohio, whether or not they can strictly be described as "fraud".


*on edit: one piece of exit poll data that presents a slight challenge to hypothesis of ubiquitous Bush reluctance was the finding that completion rates were not lower in high Bush precincts (nor were they statistically higher). This should give us pause for thought; however, the discussion on an earlier thread is relevant here: there may not be a simple linear correlationship between completion rate and reluctance, if extreme reluctance in a precinct (by any kind of voter) tends to lead to non-random sampling in the cause of achieving an adequate sample size. In such cases, if only apparently willing voters were "selected", completion rates might actually go up. So it remains a suggestive, but in my view, far from conclusive, piece of the puzzle.

[OnTheOtherHand]

30. a brief word on simple/complex down here

I wasn't so much thinking about methodology as about data fit. No one has tried to show that some constant "rBr" parameter, plus random noise, could account for the observed data. I guess that USCV satisfied its collective self that no constant "rBr" parameter could account for the observed data. Unfortunately for USCV, most survey folks Just Didn't Care. We can imagine some of the variables that might influence the extent of non-response bias, but we wouldn't expect to be able to measure them all -- so the project of building a model that would either match the observed data, or generate simulated data 'like' the observed data, would not enthrall us.

Besides, it doesn't really seem that the proponents of fraud have even tried to meaningfully account for the variance, unless someone really thinks that Bush was trying to cut into Kerry's (putative) 30-point margin in Vermont in order to make better TV in the early evening.

So, given that most of us think that non-response bias could account for the general magnitude of the exit poll discrepancy, the burden shifts to the proponents of fraud to change our minds, whether that is fair or not. We won't burn the midnight oil trying to model all the variance. (Well, some of us especially sick puppies might do it now and then.)

[Febble]

31. OK, fair enough

I'm certainly with you on data fit, speaking as an especially sick puppy.

And on the rest, in fact. Oh well, the disagreement was refreshing while it lasted.

[sunshinekathy]

32. ESI's entire analysis is based on an invalid illogical premise

Quote from ESI's paper:

"We conclude that the data do not support accusations of election fraud in the Ohio presidential election of 2004."

end quote.

Please read this proof that the ESI hypothesis is invalid.

Vote fraud can have ANY relationship with the variables that ESI analyzes, so they can conclude nothing at all about vote fraud or the causes of exit poll disparities from their analysis. i.e. ESI's analytical methods are useless and meaningless when it comes to determining anything.

http://electionarchive.org/ucvAnalysis/US/exit-polls/ESI/ESI-hypothesis-illogical.pdf

As an actuary, you will probably understand this proof. I was an actuary myself for a while.

[eomer]

33. Thanks. I was just about to copy/paste from another thread into this one

part of what you just said: "Vote fraud can have ANY relationship with the variables that ESI analyzes,..."

We've been poking around with various counterexample candidates and I think your phrase is a helpful generalization of some of the ideas we've looked at.

One somewhat successful counterexample was one where fraud was defined as exactly the amount to negate swing in a subset (50%) of the precincts where swing was otherwise to the blue. (I say somewhat because it looks like we can generate some portion of the PLD redness but not all or even most of it before we start showing correlation between red shift and swing.)

The reason, I think, that this counterexample works is because of the relationship between fraud and one of the other variables being analyzed.

So your generalization may point us toward other counterexamples by thinking about other ways fraud might be involved in relationships.

And your point is an important one to remember when pondering the meaning of the lack of correlation between red shift and swing - it may mean what it looks at first glance that it means, or, it may mean something totally different since there may be unknown relationships, variables that are systematically or algorithmically distributed rather than randomly distributed and plenty of other pitfalls.

[Febble]

34. Not "any" Kathy:

You write: "Vote fraud can have ANY relationship with the variables that ESI analyzes"

The number of relationships are, indeed, "countably infinite", but they are nonetheless fairly heavily constrained. If the relationship between Bush's true swing and fraud is effectively random (i.e. not systematically designed to interact with it) then we can estimate a probability of a result such as that in your counter-example occurring (which in fact is what, in effect, Mitofsky and ESI in their different analyses did). Unless, as eomer I think suggested, the Bushies have complete control over the entire election system, and switch the same* proportion of votes in virtually every precinct. However, that would infirm some of your other hypotheses (more or greater fraud in precincts with higher Bush vote counts, for example).

However, if the relationship between Bush's swing and fraud is NOT random - in other words, if it is systematic - then we can shift the bounds of the possible ways considerably. Nonetheless they remain remarkably firmly bounded.

This is precisely why I think both the ESI and the Mitofsky analyses should not be dismissed as BUNK. They both, as eomer, perhaps inadvertently, points out, allow us to make clear hypotheses about the kinds of systematic fraud that are compatible with both findings - and potentially lead us to hypotheses that could be tested against the data.

Unfortunately so far, I'm not finding it easy to construct an algorithm for calibrating the fraud in such a way that it will tend to result in non-correlation between swing and redshift (there is a bit of leeway in Ohio, but not much leeway in the 1250 precinct plot).

OTOH may be faring better.


(edited for typos, grammar, and precision).

* or at least similar

[OnTheOtherHand]

9. fact check

"ESI and Mitofsky... has yet to show the integrity to withdraw its own analysis, which it has wrongfully let stand since June, and presented in October to the American Statistical Association fall conference in Philadelphia."

The poor grammar here betrays poor meaning.

"ESI and Mitofsky" are not "it."

ESI presented an analysis of Ohio at AAPOR in May, and 'published' some results from it in early June. Mitofsky presented an analysis of nationwide data in October. The analyses are somewhat similar (although they use different measures), but not interchangeable. Still less are ESI and Mitofsky interchangeable.

As for the rest, I associate myself with the remarks of the estimable Febble.