For the record: the new exit poll evidence

I wanted to know who won the 2004 US election.

I developed a better tool with which to find out.

It led me to believe that USCV's hypothesis that "Bush strongholds have more vote corruption" was flawed, not because they made an error, but because the information on which it was based, in the Edison-Mitofsky report, was flawed.

It was flawed because E-M used a poor measure of bias (the WPE). I proposed a better measure. Like the WPE, it does not distinguish between bias in the count and bias in the poll. Unlike the WPE it is not confounded with the actual vote-count margin.

The conclusion I drew from my analysis was that whatever happened in the Bush strongholds was probably happening everywhere.

The evidence released by Mitofsky at AAPOR, using my bias measure, confirms this, and this evidence is now in the public domain.

This is important information, whether you interpret it as evidence for or against fraud.

And whatever you think of my motives.

The new evidence does absolutely nothing to prove or disprove either.

Lizzie

http://www.democraticunderground.com/discuss/duboard.php?az=show_mesg&forum=203&topic_id=369886&mesg_id=369886

Mitofsky gave a talk at the AAPOR conference. Mystery Pollster reported the conference and has some slides on his blog here:

http://www.democraticunderground.com/discuss/duboard.php?az=show_mesg&forum=203&topic_id=369886&mesg_id=369886

The comments on the blog include comments from Ron Baiman who is a USCV author of both their recent reports, and myself.

All know about the flawed inaccuracies in Mitofsky's original data. But you seem to be backpedaling, changing your story alot and saying only now that Mitofsky's original data is inaccurate.

Which is it Febble? Did Mitofsky have incorrect data until you magically came along? Or did you suggest to him the WPE_Index that just happened to fit his hypothesis?

Either way if you recognize where people are coming from, it begs the clear question. Of course there was "Response Bias" and Fraud, there was just no such thing as a "Uniform Response Bias" which is what Mitofsky has outlayed in bold print for the viewership.

It's not the problem Elizabeth that you got it wrong, or made any mistakes no the reason people have actually had contention with you is your inability to disclose the secret relationships.

First it came with not telling anyone you worked with Mitofsky. Then it came with not telling anyone how you were deploying information for him, then it came with the close relationship observed and documented with Mystery Pollster. It suddenly comes with the changing of your story, which is that now what Mitofsky showed was flawed to begin with.

See, it becomes a problem to trust that kind of logic. Now if you don't mind me asking, where does the actual WPE_Index formula come from? Where is the number algorithm that reflects this precisely? I think it would put the experts especially USCV at ease, to know what the BASIS is for those said formulas.

Finally it would be prudent to know if W-M uses those formulas or not, as it seems Mystery Pollster is the source of the formulas. Elizabeth if you got the formula from Mark, why the cover story or deception? Why say it was your work? I've seen that similar formula on Mark's website....Do you actually work independantly? Are you at all recognizing how many parts of your story you have altered?

The data is the data.

From the data, E-M produced a measure, called the WPE for each precinct. It is calculated by taking the predicted margin (from the polled responses) and subtracting that from the vote margin (from the precinct count). So if your polled responses say that 70% of voters voted for Kerry and 30% for Bush, but the votes say that 60% of votes were for Kerry and 40% for Bush, your predicted margin is 40% and your vote count margin is 30%. Your "WPE" is -10%. It's a measure of "bias" and does not distinguish between bias in the count or bias in the poll.

It was the information on the WPEs that were in the E-M report (amongst other things). They seemed to be higher in "Bush strongholds" (precincts with over 80% of votes for Bush). USCV interpreted this as evidence (in combination with some other evidence) that vote corruption was greater in these precincts. I agreed with them. Until I realised that the WPE is a flawed measure, because for a given amount of true bias, it varies according to the proportion of votes counted for each candidate.

I figured out a way of measuring bias more accurately. USCV used a similar formula. Unfortunately it needs to be applied to precinct level data.

Mitofsky did this.

I know what you think of me, you have told me. I understand that.

But I am certainly not back-pedalling. I am simply saying, that when a function that is now widely agreed to be valid (by USCV as well as other pollsters and statisticians) is applied to the precinct level data a different pattern emerges.

It could easily make the case for fraud stronger. It shows it was not a feature of the small number of "Bush strongholds" in the poll (40) but of the entire sample.

I am not making that case. But I am pointing out that there is absolutely nothing that I have done, or that Mitofsky has presented, that undermines that case. It is simply that the measure of bias is now more accurate, and can bias can thus be seen to be widespread.

And your characterization of my own history is wrong. You may not believe it, and I take responsibility for the fact that you do not believe it, but it is.

But the truth is simple. I want to know who won that election.

" I am simply saying, that when a function that is now widely agreed to be valid (by USCV as well as other pollsters and statisticians) is applied to the precinct level data a different pattern emerges. "

Again, how are we to distinguish this is the official WPE_index? USCountVotes has run a new paper, I see nothing mentioned about your own new "WPE formula" index in the graphs. Instead using your methodology, they reapply a new formula to the mean data and it shows up as a spike.

It becomes very contentious when the source of these mathematical calculations is not revealed. It appears that USCV has not yet outlined what exactly was at the APPOR, as I have been in contact with them and they are not aware of what you are saying.

However, they do have what is known as a model simulation. It refutes the data. You can't claim one formula is more correct over another, it just doesn't work that way I'm afraid. It appears this is what Mitofsky has attempted to do, and made sure it was supported by others including yourself. He has not released the full chart for how these calculations were made. There is no precinct map of which to map it from. It is highly suspect.

Although the underlying map of your hypothesis is correct, and clearly shows a median to mean problem, the prospect W-M and his staffers are proposing is not and we seem to have not met up on why this is being stressed.

Ergo, there clearly is bias and fraud but it is not going to be measured using the calculation demonstrated here. It can not be as the updated paper shows the problems. If it were to get the actual result, and record a proper measurement than they will need to re-run the real simulation in an actual precinct using the updated data and hopefully precinct questions Kathy and Ron should have in possession(if the news is accurate).

Then we will all be able to tell how much bias and fraud there is according to that one area, and I would think that one of them should have no trouble getting W-M to release the rest of the WPE if he is co-operative.

I called it "bias index"

USCV call it "ln(alpha). They are all the same thing. I can give you the formula if you like.

The USCV data is simulated data. I also simulated data.

Mitofsky's data is the actual data.

The important thing, which is why I posted today, is to say that now that you can see both WPEs (the original measure) and bias, WPE_index, ln(alpha) or whatever you like to call it, plotted against vote-count margin you can see exactly where the "bias" occurred. All over the place.

This would be interesting information, I would have thought, for anyone wishing to make the case for fraud.

Your following statement is not honest:



USCV is using both actual data and simulated data. They are comparing the two. That is the point of their paper. In fact that is almost always the point of simulation. What is your point?

Your following statement is not honest:



Your statement is not honest because it exaggerates what you have shown. You have only shown that if you force the data (WPE Index) into a best-fit line, then that line is roughly horizontal (no slope). There are plenty of non-linear shapes that could be present in the data that would resolve to a horizontal line if you force them into a best-fit line.

USCV most definitely takes issue with your use of a linear model and they have done so for quite some time. The point they make has always been about what happens in the data at the two ends of the range. You can't tell what is happening at the ends compared with what is happening in the middle if you force the data into a single straight line.

The USCV paper is crystal clear on these points. And the points are very, very simple - not points you could possibly have missed or misunderstood. These arguments - the actual arguments being proffered by USCV - are the ones you should be addressing.

I thought you were talking about their simulations.

Both my work and USCV's work was based on the data given in the E-M report and simulated data.

Yes, there are plenty of non-linear shapes that can be fitted. There are in fact an infinite number. Perhaps one of them is the fingerprint of fraud.

All I am saying is that the linear fit is with bias is zero. I think Ron Baiman agrees. The linear fit with WPE was not zero, as Ron correctly inferred.

USCV are entirely entitled to take issue with my use of a linear model. I only ever said that I thought the data was consistent with a linear fit of zero.

USCV and I are probably now in agreement, and the good news is that we now have a better way of measuring bias. If anyone can see fraudulent patterns in the bias plot, they are more likely to be real patterns.

Unfortunately non-linear statistics are much less powerful than linear statistics, so non-linear hypotheses are much harder to test.

It does not mean they are not true.

That fairly sums up the current state of things.

I think it would advance the conversation if you would address some of that nuance on a routine basis, especially whenever the topic has anything to do with how you and USCV agree or disagree on the question.

But both the USCV ideas and mine have been in flux, and it hasn't always been easy to see when they have diverged. Ironically, I have often thought we have been in agreement, but they have insisted they disagree!

Language, even mathematical language, can be misunderstood, and perhaps I simply have not been clear.

But I've tried. And to be fair on myself, I think we now do have a much better measure of the thing we are all interested in - the bias. Even if we don't know what was causing it.

But being able to measure a thing properly is a start.

A couple of posts above you say: "...you can see exactly where the "bias" occurred. All over the place."

Your statement would make the reader think it has been shown that the bias is uniform and that there is no shape or pattern to it. I don't believe that much has been shown yet.

"all over the place" isn't the definition of uniform.

It means "all over the place". If someone can find a pattern that means fraud, then it could mean fraud.

My dictionary doesn't have a definition for "all over the place" but I'm glad to know how you meant it. :)

If I could see a pattern in that mess I'd be delighted.

:)

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

If you want to see Ron's comments about the formula they are here.

He also commented here on his general response to AAPOR:

http://www.mysterypollster.com/main/2005/05/aapor_day_two.html#comments

You may not like the source, but it's definitely Ron, because you can check his email address.

Lizzie

Just what IS the mathematical formula you came up with that now tells the real WPE since Mitofsky's own report did not have it?

You stated the formula which his report had in writing, and that formula is wrong. Your specific formula has not been given.

Bonus question: Also do you deny that Rick Brady and Mark Bluementhal gave you the formula Febble? I have seen all the comments on Mystery Pollster blog and I know for a fact, that they gave it to you. I want to know why you assumed this information for the formula gives absolute results?

Bias = the natural log of: R(1+margin-WPE)/D(1-margin+WPE) where R is the proportion of votes for the Republican candidate, D is the proportion of votes for the Democratic candidate, margin is the vote-count margin (D-R) and WPE is the WPE. Before you take the log, we've all called it alpha, which is why USCV call it ln(alpha) for natural log of alpha.

More simply, if one had the raw data, it is the log of the Democratic response rate divided by the Republican response rate.

The Mitofsky formula is not wrong. It gives the WPE, which has been widely used as a measure of bias. It is not a good measure, for the reasons I have given. It can be converted into a measure of bias using my formula, or derived directly from the raw data.

I have no idea why you think either Rick or Mark "gave" me the formula. We were in constant email contact during the development of my paper, and I acknowledged them in my paper. But I'm the geek, and I did the algebra. But if you want to think they did the algebra, feel free. I can't stop you. It wouldn't matter, just be odd.

But as it happened, I did the math, wrote the code for the simulations, plotted the graphs and wrote the paper.

I'm a geek.

The truth is simple.
The REAL analysts at DU have called you out time and again.

You have boxed yourself into a position which is untenable. The recorded timeline of your posts shows vacillation and incongruities.
Your relationship with MP and the rest belies your objectivity.

At DU, we work for the truth.
You work for those who seek to hide the truth.

The time is rapidly approaching, if it is not here already, that your exercise will be reduced to absurdity. It will be known as "Febbles Last Stand", the point at which your House of Cards
falls apart completely. Are you prepared for it?

If you were a true scientist, you would have started off with a different "null": The analysis of state and national exit polls indicates beyond any reasonable doubt that the election was stolen.

The null hypothesis still stands.
So your whole premise is a "null".

"It's not the problem Elizabeth that you got it wrong, or made any mistakes no the reason people have actually had contention with you is your inability to disclose the secret relationships.

First it came with not telling anyone you worked with Mitofsky. Then it came with not telling anyone how you were deploying information for him"

Mitofsky applied my formula quite independently. My paper was publicly available. It caused a fair stir in the polling world. He confirmed he had used my formula, and the pattern was as I had said it might be. I was not deploying information for him.

"then it came with the close relationship observed and documented with Mystery Pollster. "

Mark Blumenthal is acknowledged, with others, including Rick Brady, on my paper. DemfromCT (from Daily Kos) also. Both wrote blog pieces describing the development of the work in detail. The work emerged from a four-way discussion both on and offline. There is no secret relationship.

"It suddenly comes with the changing of your story, which is that now what Mitofsky showed was flawed to begin with."

In my paper (page 4) (updated 27th April)

http://www.geocities.com/lizzielid/WPEpaper.pdf

I wrote:



You wrote:

“Now if you don't mind me asking, where does the actual WPE_Index formula come from? Where is the number algorithm that reflects this precisely? I think it would put the experts especially USCV at ease, to know what the BASIS is for those said formulas”.

USCV have used similar formulae in their Appendix B. They acknowledge my contribution. They saw my paper before it was posted. The derivations of the formulae are in my paper.

"Finally it would be prudent to know if W-M uses those formulas or not, as it seems Mystery Pollster is the source of the formulas."

Mystery Pollster has the formulae. He contributed to the development of the argument. However, I derived the formulae myself.


“Elizabeth if you got "the formula from Mark, why the cover story or deception? Why say it "was your work? I've seen that similar formula on Mark's website....Do you actually work independantly?”

See above. There was no cover story because there was no deception.

But I should not have blogged after the start of my very recent association with Mitofsky (which came about because of my work, not the other way round). I agree.

I should also probably not have blogged at all, once I knew that my formula had been applied at precinct level, and what it showed. I apologise, once again.

"Until I realised that the WPE is a flawed measure, because for a given amount of true bias, it varies according to the proportion of votes counted for each candidate."

Could you please give us an example?

currently hosted at MP.



It shows the WPE you would get for a given amount of bias, for precincts with varying proportions of the Democratic vote.

The example given was for a large bias of 2:1 In other words, 2 Democrats for every Republican were polled.

As I said, the index is fraud neutral. A bias like this could arise either if Democrats were twice as eager to be polled, or were twice as likely to have their votes "lost" in some way.

You can see that the WPEs are larger in the middle, and also larger at the Bush end than the Kerry end. However, the bias is the same in each case.

My formula computes the actual bias from the WPE and the vote count percentages.

I understand that the WPE's will skew more negative where the precincts are more competitive, as a purely mathematical artifact... But explain to me why your graph presents an opposite picture of the following:

Let's assume 3 precincts... Precinct A is heavily in favor of Bush, Precinct B is competitive, and Precinct C is heavily in favor of Kerry.

Now, let's assume that it was also Mr. Mitofksy's intention to poll 20% of the voters in each precinct. So he sends his pollsters out and gets the following responses in the "Official Poll column":



Now, let's assume, as you did in your example, that Kerry supporters were inclined to respond twice as often as Bush voters. So, to get what the TRUE percentage of voters was in the election we should divide the Kerry responses in half, as I did in the "True Poll" column.

Now, if we assume that the Mitofsky pollsters actually managed to interview 1/5th of the raw numbers of people leaving the booths, but the Democrats were twice as likely to respond, the "Official Vote" would have reflected the split in my "True Poll" column. (i.e. the total vote count is 5 times the total poll count, but the actual split is reflective of the adjusted, true intention of the voters).

But, as you can see, the Official WPE's are OPPOSITE of what you contend: in an uncorrupted election, as I've portrayed here, there is a higher negative in the heavy Kerry precinct.

What am I missing here? Or, perhaps, what are YOU missing here?

Also, I should relay that this is not exactly a fraud neutrel argument.

Try applying your algorithm to the following table based on their segmented statistics.



Can the mean actually justify or "measure" purged and deleted votes?

Let's put that into more perspective, using additional information from the table.



Precinct/city demographics for the state of Florida showing both statistical mean and a weigh point.

If one was to justify and use the Z-score algorithm by TruthIsAll or Stephen Freeman, those votes could be readily accounted for. I am certain then it would be easy to measure the precise amount of votes tossed aside and/or switched. But using your algorithm, the mean stays as a flat line and in my estimation could not measure whether said totals were off the mark, or not.

This is what all the main contention is really about, it is imperative that the formula work to measure the exact amount of fraud or problem recorded. With something this large at stake, that's why the proposal of at least 6 PHDs working together to create the formula is neccesary.

which I used to calculate the maximum number of returning 2000 voters, from which one can determine the minimum number of New voters - and Kerry won them by near 60-40%

And Kerry won at least 70% of Nader 2000 voters.
And I am convinced by pure common sense that he won at least 95% of Gore voters, not the 92% they show in the how voted in 2000 demographic.

Its a very simple analysis based on factual data at the macro level.

that detected fraud, that would be fantastic.

I totally agree.

Mine certainly does not.

http://www.geocities.com/lizzielid/WPEpaper.pdf

You've got it though. The only thing is, you need to plot it against vote-margin, not poll margin, to get my plot. Try starting with the vote-count, and make that 90%, 50% and 10% Kerry, then poll twice as may Kerry voters as Bush voters. (Or assume there were twice as many Kerry voters with spoiled votes as Bush voters, if that makes you feel better!)

The math is not actually in dispute, which is why, I presume, it's got so much coverage. It's been pretty thoroughly checked. I think everyone, including USCV certainly, agrees that you don't get the same WPE for the same amount of bias, and that the pattern takes that shape. It's in their simulator on their website, and you can play with it.

My paper has a 3D plot if you are interested, showing how the WPE varies with different degrees of bias and different degrees of vote-count margin.

When you say "I got it." ???

I got something exactly OPPOSITE of what you're contending. I'll repeat it below.

I understand that the WPE's will skew more negative where the precincts are more competitive, as a purely mathematical artifact... But explain to me why your graph presents an opposite picture of the following:

Let's assume 3 precincts... Precinct A is heavily in favor of Bush, Precinct B is competitive, and Precinct C is heavily in favor of Kerry.

Now, let's assume that it was also Mr. Mitofksy's intention to poll 20% of the voters in each precinct. So he sends his pollsters out and gets the following responses in the "Official Poll column":



Now, let's assume, as you did in your example, that Kerry supporters were inclined to respond twice as often as Bush voters. So, to get what the TRUE percentage of voters was in the election we should divide the Kerry responses in half, as I did in the "True Poll" column.

Now, if we assume that the Mitofsky pollsters actually managed to interview 1/5th of the raw numbers of people leaving the booths, but the Democrats were twice as likely to respond, the "Official Vote" would have reflected the split in my "True Poll" column. (i.e. the total vote count is 5 times the total poll count, but the actual split is reflective of the adjusted, true intention of the voters).

But, as you can see, the Official WPE's are OPPOSITE of what you contend: in an uncorrupted election, as I've portrayed here, there is a higher negative in the heavy Kerry precinct.

What am I missing here? Or, perhaps, what are YOU missing here?

(well sort of - you got half values, but I think I know what you did).

The point is that to compare equivalent extreme precincts, you need to compare a precinct with 10% Kerry votes with a precinct with 10% Bush votes.

Not a precinct with Kerry 5% and Bush 82%.

Okay, I gotcha.

In fairness, I went through the exercise again, this time starting with the vote column, comparing 10%-50%-90% precincts and working backward to what the exit poll would have reflected if Kerry voters were twice as likely to respond, and below is what I got...

So, I hate to have to agree with you, but...

heavy Bush precinct. Lots of Bush voters, or just lots of Bush votes?

That's why we are plotting against "vote-count margin" - we don't KNOW if there were lots of Bush voters, or just lots of Bush votes. So to be neutral, we'll call it vote-count margin.

Try a precinct in which the Bush VOTES were 95% of the total, and compare it with a precinct in which the Kerry VOTES were 95% of the total.

The WPE will be greater in the 95% Bush precinct than in the 95% Kerry precinct, even if the bias is the same.

Okay, I'm working on that. :)

not to do it your way round, except that the data we had in the E_M report defined it that way round.

975 and 25 in your first column. You'll have to poll a fraction of a person, but it should give you a 95% Kerry vote precinct to match your 5% one.

OK, I went through the exercise you recommended, and now understand... but after I did this I had to remind myself that it doesn't prove any sort of Kerry bias... it just shows that the higher negative WPE's in pro-Bush precincts cannot be used as prima facie evidence of fraud.

It happens to us all ;)

I agree with TruthIsAll's chart too. I believe someone should do a Z-score analysis of several precincts.

this is of import, for me at least

" we don't KNOW if there were lots of Bush voters, or just lots of Bush votes. "

Ive been lurking and following along--
Pardon the interuption

You said you work for Mitofsky now.

Are you on his payroll while you're posting here?

That's a valid question since you see to do it all day long.

Just asking?

Are you British? Just curious.

;)