Within-Precinct-Error Correlation BY Precinct (NEW)

[On the Road]

Mystery Pollster
Demystifying the Science and Art of Political Polling - By Mark Blumenthal
May 16, 2005

AAPOR: Exit Poll Presentation

---snip

Mitofsky spoke first and explicitly recognized the contribution of Elizabeth Liddle (that I described at length a few weeks ago). He described "within precinct error" (WPE) the basic measure that Mitofsky had used to measure the discrepancy between the exit polls and the count within the sampled precincts: "There is a problem with it," he said, explaining that Liddle, "a woman a lot smarter than we are," had shown that the measure breaks down when used to look at how error varied by the "partisanship" of the precinct. The tabulation of error across types of precincts - heavily Republican to heavily Democratic - has been at the heart of an ongoing debate over the reasons for the discrepancy between the exit poll results and the vote count.

Mitofsky then presented the results of Liddle's computational model (including two charts) and her proposed "within precinct Error_Index" (all explained in detail here). He then presented two "scatter plot" charts. The first showed the values of the original within precinct error (WPE) measure by the partisanship of the precinct. Mitofsky gave MP permission to share that plot with you, and I have reproduced it below.



The scatter plot provides a far better "picture" of the error data than the table presented in the original Edison-Mitofsky report (p. 36), because it shows the wide, mostly random dispersion of values. Mitofsky noted that the plot in WPE tends to show an overstatement mostly in the middle precincts as Liddle's model predicted. A regression line drawn through the data shows a modest upward slope.


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

[ccarter84]

1. ooooh pretty graph....n/t

[On the Road]

3. Yes.

Do you understand it and why it was presented.

[TruthIsAll]

2. Very pretty. Now let's see the data- all of it. And not massaged

like the final 613 in the National exit poll.

Ribofunk, you ain't got squat.

Oh, and send our regards to Sawyer and Febble.
They did a great job.

Trolls?
No, just moles.

[On the Road]

4. I Posted This

because I hadn't seen it yet on DU and because it's an important chart.

One of the criticisms of Liddle's hypothesis was that you could only really see the effect when the numbers were broken down by precinct. In a post on DU, Liddle said she didn't have the precinct data but was hoping Mitofsky would do that. That's what the chart represents.

Liddle's insight was that WPE would be largest in precincts which were competitive and lowest in strongly partisan precincts. The data bear that out. Her insight also predicted that there would be an upward slope so that WPE was slightly greater in heavily Republican precincts. The data bear that out.

Now, the "Liddle effect" is open to discussion. But once you see it, it's pretty much undeniable. The math is pretty clear, and is not that difficult.

The next step would be to take that effect into account in analyzing the exit polls to see if they still suggest localized fraud in heavily Republican districts. I hope they do this analysis. USCV would like to do this, but they don't have all the detail. Maybe it's being presented in the polling conference. I hope so. I'll post it if it is.

[tommcintyre]

27. "Maybe it's being presented in the polling conference." Yeah, right!

"The next step would be to take that effect into account in analyzing the exit polls to see if they still suggest localized fraud in heavily Republican districts. I hope they do this analysis. USCV would like to do this, but they don't have all the detail. Maybe it's being presented in the polling conference. I hope so. I'll post it if it is."

Do YOU really think Mitofsky will release ANYTHING that helps to build a case of significant Bush election fraud?

Or, perhaps you already KNEW what he intended to present at the conference, as ****** did?

Well, I'm not privy to such inside information, but I'll bet he DIDN'T present such information, did he?

C'MON! GIVE US A BREAK!

[On the Road]

29. OK, So When Data *Is* Presented You Don't Believe It

Personally, with an issue like this, I think the raw data should be made available. It should be mandated. If not to the general public, then at least to any group like USCV that meets certain minimum requirements.

However, it isn't. Pollsters are very protective of their data. if you look at the scatterplot, what Mitofsky released was much more damaging to himself professionally than any indication of fraud. That would have vindicated his polls. Look at that variance -- it's all over the place! There must be 20-30% errors in both directions all over the place. It makes you wonder what was going on on the ground to cause that kind of result.

Or, perhaps you already KNEW what he intended to present at the conference

Why would I know anything about what Mitofsky was going to present? I'm a poster on DU like you are. I work for the phone company. I did some analysis after the election and gave some money to recount efforts. Geez!

[LightningFlash]

30. There is no original "scatterplot"

We don't even see the original sources or calculation based question information for the scatterplot at all, just a jumbled bunch of assumptions that create the scatterplot graph.

Obviously based on the only data currently available, there are a number of tests already done that refute any basis of the scatterplot.

In conclusion, it very well has happened in a few places and response bias did take place. But response bias uniform and calculated like a calculator did not take place, and refutes its own argument that says it took place.

Unless a full analysis of the data can be performed and tested, we will not know how much and what level of fraud happened in the elections. We will know based on all the other reports that fraud happened and Kerry won based on the amount of votes switched. We will not at all know beyond that just how much declarative fraud happened in all the races, at all levels and that is the underlying issue.

Nobody can take assumptions with this. It is absolutely clear that the real numbers and full questions need to be released.

[On the Road]

32. If You Want to Have a Discussion

please say something that makes sense.

The scatterplot shows the within-precinct error by precinct plotted against % partisanship. USCV used the same kind of graph to argue for fraud. This is a more detailed version, using the detailed data Mitofsky didn't release. I would like to see him release it, but short of that the graph is helpful.

It shows that WPE are higher in competitive districts. It shows a slight increase in average error in more Republican districts. And it shows the huge range of error once you get down to the lowest level.

Once you look at all the data this way (especially the adjusted version), there are still some outliers. Maybe there was fraud in some of those errors. But keep in mind there are almost as many outliers below the line.

[LightningFlash]

34. "USCV used the same kind of graph..."

Actually USCV has likely used the same data by now, if they were really shown the data by this time.

And the interesting WPE chart is based un un-named official sources. There is no access to the MP's actual data, and so his entire plot by nature is automatically suspect.

Likewise, in the full updated paper the new graph contends this was not a flat line. It was a peaked line, thereby shredding the argument that uniform responder bias was displayed all over the chart.

[Blue Shark]

56. Cooler heads...

...appreciate your work here. Keep on Keepin' on ribofunk.

[Melissa G]

5. As I understand this even Liddle says the is irrelevant to the Fraud issue

This is a Red Herring in the discussion. We are discussing a variable the neither confirms or denies FraWd. Even Febble says so...

[On the Road]

15. It's Very Relevant to One Issue

One of USCV's major arguments was that heavily partisan precincts were anomalies and were prima facie evidence for fraud, since the results required extraordinary differences in response to pollsters. What Liddle did was show that the effects could be explained without recourse to fraud or suspiciously high Democratic response rates. And that takes away what had looked like a statistical smoking gun.

On Edit: Elizabeth Liddle just PM'd me (first time -- it was a surprise). Here's what she had to say on that:

>My point, of course, is that of course it doesn't make the net
>bias go away. It simply shows that it doesn't correlate with
>vote-count margin in the manner inferred by USCV, and taken to
>indicate fraud.
>
>It could still be fraud - but not the pattern postulated by
>USCV. And if people look at the plots, it may dawn on them
>that if Republican fraud explains the high values, maybe
>Democratic fraud may explain the low values....

[Melissa G]

16. The issue is the House appears to BE ON FIRE

There is Smoke pouring out of all the windows.. It smells like fire..
It feels like fire.. No one has gone in the house to see if it is on Fire( as they are withholding the keys, data, for some inhumane reason) People( or our democracy) could be in the house that could be on FIRE. Lizzie is discussing something like the possible colors of the shirts on the possible people that walked away from the fire. It is almost but not quite irrelevant TO THE FACT THAT THE HOUSE IS LIKELY ON FIRE!
That other folks are not all that interested in her point might have something to do that while the picture she made is really pretty, The BLIPPING HOUSE IS likely on FIRE!
Those who want to deflect from the Fire are using Lizzie to distract us from the House BURNING DOWN. I hope Lizzie does the humane thing and helps put out the fire instead of letting the folks with the Kerosene get away.

[On the Road]

17. However Much Panic You Feel,

the goal is to win the battle of public opinion. And shouting "the house is on fire" without nailing the evidence is convincing large numbers of people (including sympathetic Democrats) that 2004 election fraud is a conspiracy theory on a par with Vince Foster. If you've heard a deafening silence recently, and your representative doesn't seem concerned, this is one of the reasons.

I do not want it suffer that fate. However boring or perverse it might appear, the way to have a chance of winning this thing is to ruthlessly pare the issues down to ones with the best evidence (or with the possibility of getting the evidence) and push those. Many of those issues may be small, but that's not the goal at this point. The goal is to frame the issue and put it in a place where it can critical mass and start to snowball. I don't know whether it's too late for that or not.

Richard Feynman said as a scientist, his first priority was to try to prove himself wrong as quickly as possible. Once you've taken that step, you're left with much stronger arguments and much better material. What Liddle is doing is the right way to pursue the statistical issue. And I'm starting to see that even when her effect is taken into account, the remaining data might still contain evidence of fraud. It's just hard to see with all the noise.

The country has survived far worse election fraud than is likely to have happened last year. It's important that it be pursued, and that a consensus emerge that the Republicans are the party of corruption. But reform movements win. The future is not dark.

[Melissa G]

19. I am not in a state of panic..ribofunk

Neither am I interested in wandering down irrelevant paths or following those who choose to try and lead the discussions in fruitless digressions. Life is too precious and the central issues are too important.
Glad you are finally noticing that this has little impact on the fraud discussion..
You mention Framing...pursuing Liddle's point has some negative impact on the framing discussion. I'm sure you also would like to stay focused on the probability of fraud. As you noted, our country may have survived... but if there is FraWd, democracy has Already Perished.
you write..
The country has survived far worse election fraud than is likely to have happened last year. It's important that it be pursued, and that a consensus emerge that the Republicans are the party of corruption.

I'm not convinced about what you assert re there ever having been worse fraud. I'm already convinced that those republicans who are currently in charge are fraWdulent. I do not however, paint all republicans with the same brush. The R's I feel saddest for are the ones who voted against b*** in this election and probably had the votes changed back against their will. Their votes were the most vunerable.

Editing to say I have confidence that much evidence will be appearing in print shortly. Truth will out.

[On the Road]

36. As Far as Framing Goes,

USCV was going public with a claim based on a flawed mathematical model -- that the higher WPE in partisan Bush precincts required an unbelievably high rate of participation from Kerry voters.

In this case, the peer review process corrected it. That's how it's supposed to work. And vilifying the person who made the observation puts partisanship above the truth.

The worst thing that could happen was for the USCV claim to hit the front pages and then be discredited (which it certainly would have been). Fortunately, few people other than political junkies were aware of it. Maybe USCV can revise it and find something else.

People on this forum seem to want to go public with any old argument and when it's criticized, just turn up the volume.

[Melissa G]

38. You still seem kinda confused about the irrelevancy of the

Febble factor to the discussion of fraud.. and I thought you told me you were catching on?
Framing again? Thanks for the perfect example (once more..) of misleading framing and wandering down irrelevant paths pretending that they mean something. Your post was a classic example. I will bookmark this for folks wanting to see what it looks like.

[On the Road]

39. I Have No Idea What You Just Said

This is a controvery about whether the exit polls are evidence of election fraud. The most specific and well-argued claim that they were was made by USCV, who analyzed the WPE within partisan Republican precincts. However, the model that they used was mathematically flawed.

The person who demonstrated the error was the object of baseless ad hominem attacks. The work was called irrelevant, which baffles me because it addresses the most robust argument from the exit polls. The only thing that's missing has been discussion of the actual issue.

Are you claiming that the data USCV presented still constitutes evidence of fraud? Maybe it does. I wish someone would make the argument, because all I'm hearing on DU lately is hot air. That does not help in any way.

[eomer]

40. The USCV paper itself contains the argument.

And the argument is not refuted by Febble's function. It is an argument that USCV developed after full knowledge of Febble's function and represents moving one step further beyond what Febble was looking at.

I don't see how anyone could make the argument any more clear than the paper does.

[Chi]

41. Got a question

Heya ribofunk
"However, the model that they used was mathematically flawed."

I believe what USCV said was, response rates were higher in heavy Bush precincts, than in heavy Kerry precincts,
so I'm curious as to what 'flaw' you are talking about.

[On the Road]

42. This is How I Understand the Situation:

USCV looked at the public exit polling data and broke it down by partisanship. They argued that in order to get the polling results in heavily Bush districts, an overwhelming number of Kerry voters would have to have been polled, in a few cases over 100%. Here's the chart USCV used:



USCV said: The required pattern of exit poll participation by Kerry and Bush voters to satisfy the E/M exit poll data defies empirical experience and common sense under any assumed scenario.

Personally, I agree -- that would have been prima facie evidence of fraud. As far as I know, USCV did not publish their formula. But apparently it was based on a straight-line method when it should have been the curved function proposed by Febble. Once you account for the curve, that "required response rate" drops off to a more reasonable level. USCV may still find some anomalies when this is taken into account. If they do, they'll be on much better ground.

I only know a little statistics and can't follow all the details of the USCV and Febble papers. I suspect most DUers can't either. Please correct me if I'm wrong. I would like to prove fraud, but if the core of an analysis is flawed I don't think it should be pushed in public.

And BTW thanks for the question. It really helps to keep this a civil discussion.

[eomer]

43. I think I understand it differently, but need you to clarify.

Do I understand you to say you think this quoted part needs to be corrected by Febble's function and then it would flatten out and essentially go away (as an indicator of fraud)?

If that's what you're saying then I don't believe that is right. I believe that the pattern of response rates shown in that graph are the rates that would be required to reproduce the mean WPE, media WPE and overall response rates and that this conclusion is not invalidated by, in fact not affected by, Febble's function.

Also, I think the points made earlier in the paper (pages 7 & 8 of the version currently on the USCV website) are just as important, if not more than, this one.

(BTW, I think you're not looking at the latest version of the paper, but the part you quoted seems to be essentially the same as the latest.)

[On the Road]

48. Yes, I Was Under the Impression

that the high WPEs in heavily Republican precincts were mostly an artifact of their formula and would return to a more normal range if the Febble function were applied. At the very least, that adjustment had to be made in order to convincing make their case. But I'm not absolutely sure.

As it is, looks like both side may may been partly right. The adjusted data move in the right direction (chart at the top of p.10), but there still seems to be an anomaly there. The top of the range has moved down from 90%, which is unbelievable, to about 70%, which is a little more plausible but still awfully high.

I just started looking at the new USCV paper based on your comment. I haven't absorbed all of it yet -- it looks like from their charts that that there are still serious questions, such as extremely partisan Republican areas (chart, top of p. 7). That may be where USCV should focus. Hopefully Mitofsky will cooperate in running the numbers against the full precinct-level data and provide specifics to take local action.

[Chi]

44. I believe it's simpler than you make it

Always glad to be civil, I have no need for hostility 8)

From USCV reply to Mitofsky's report.

=====================================================================
No data in the report supports the hypothesis that Kerry voters were more likely than Bush voters to
cooperate with pollsters, and the data suggests that the opposite may have been true:

(the graph is on page 4, I could not copy it <shrug> )

This chart was constructed from data within the report (p. 37) that is not analyzed or mentioned in the text.
This data bears directly on the plausibility of the report’s central hypothesis, and it goes in the wrong
direction. In other words, in precincts with higher numbers of Bush voters, response rates were slightly
higher than in precincts with higher number of Kerry voters."
=====================================================================

All they did was take the info straight from Mitofsky's report.
Which showed Bush supporters more likely to fill out exit polls in heavy Bush areas, than Kerry supporters in heavy Kerry areas.

So I think to portray it as a 'flaw' or 'error' is unfair.

[eomer]

45. Hey Chi,

sorry, I think I jumped in the middle of a conversation and am breaking the flow.

ribofunk, how about I copy my last post into a reply to the original post and that way we won't get these two things mixed up.

On edit: ribofunk, please pick up our conversation by replying to post #46. Thanks.

[Chi]

53. Not a problem

You are more versed in the Febble/USCV/Mitofsky papers, so have at it.
Looks like I thought ribo was on a different paper anyway.

One thing I would try and discuss, at this point is...
whether or not the peaks and valleys profile fits rbr, or not.
To me, rbr sinks fast with an erratic profile, like the one USCV got when it did the 20% chunks.

Just a thought.

[eomer]

55. I agree, the shape argument is the main premise of the paper.

ribofunk and/or OnTheOtherHand (or Febble for that matter), have you looked at the argument USCV puts forward on pages 7 and 8 of their latest version? That is, I believe, the main argument they are making.

Their finding is that one way to get the simulator to generate a shape similar to that of the actual data is to assume vote switching from Kerry to Bush together with a response bias that is different than rBr.

The response bias in that successful simulation was one where Kerry voters are more likely to respond in Kerry strongholds and Bush voters are more likely to respond in Bush strongholds, which seems more intuitive than rBr.

They were not able to generate a similar shape using rBr alone.

[OnTheOtherHand]

47. as for the response rates...

The response rates reported in the January E/M report, and accurately reproduced by USCV, actually don't vary much across categories of partisanship -- I think the low is 52% and the high is 56%. What Mitofsky asserted at AAPOR last weekend (I think the first part of this claim also appears in the January report) is that there is no statistically significant relationship between partisanship and response rate, and in fact, there is no statistically significant difference between the observed rates and the slight decline that we would expect if Kerry voters were consistently more likely to respond.

You're right, the response rate argument has nothing to do with Febble's bias index.

But there are some other problems with USCV's response rate analysis. One is that they assume that the 'average' high-Bush precinct was 90% Bush, but probably the figure was more like 84%. May seem like a picky difference, but if you increase the supply of Kerry voters by over half, it has a big effect on the analysis.

I dunno. Some of those 40 high-Bush precincts are pretty bizarre no matter how you dice and slice them. As I think Eomer and I were agreeing elsewhere, what the scatterplot really emphasizes is that there are a lot of pretty bizarre precincts (with discrepancies favoring both Bush and Kerry) that aren't high-Bush.

[eomer]

50. It looks like the response rate range is wider than that

In the graph on page 10 of the latest, I see one that looks like 48 in the Bush participation line and one that looks like 69 or 70 in the Kerry participation line.

So the Bush range looks like from 48 to about 52 and the Kerry range looks like from 52 to about 69.

[OnTheOtherHand]

51. oops, we're talking past each other

I was referring to the five average response rates, for the five partisanship categories, as actually reported by E/M. In theory, those response rates are "actual," but the estimated response rates for Kerry and Bush voters are, well, estimated.

The rates in the graph on p. 10 -- first of all, elsewhere a few minutes ago I complained about a faulty assumption that generated an 84% top Kerry response rate. Here they make a better assumption and it's down to 69%. (I'm not drawing any strong conclusion from that, just trying to avoid possible confusion about my posts -- I think over there I was looking at the earlier version of the paper.)

I'm not sure what I think about these estimated "mean" response rates -- I'm not sure I understand the assumptions. Apparently Mitofsky at AAPOR presented a slide that showed the reported refusal rates ranging from near 0 to 80% and up. It's hard to believe that those are actual refusal rates. (That doesn't mean I think E/M made them up, but I have some serious doubts about the interviewers' getting stuff right.)

Basically, I think it's an odd result emerging from other odd results. If you there's a precinct where the recorded vote was 90% Bush, and the exit poll result was 80% Bush , you can calculate what "must" have been the Kerry response rate under certain assumptions, but I'm not sure what that will tell you, except what you already knew, namely that it's a weird result.

Yeah, I know, thanks for clearing that up (grin).

[Melissa G]

59. I said your post was a classic example of Bad Framing.

See post 58 if you want my elaboration.

[TruthIsAll]

22. More clap-trap.

"The country has survived far worse election fraud than is likely to have happened last year. It's important that it be pursued, and that a consensus emerge that the Republicans are the party of corruption. But reform movements win. The future is not dark".


FRAUD HAS NEVER BEEN ANYWHERE AS MASSIVE AS IT WAS IN THIS ELECTION.
DISENFRANCHISEMENT, TOUCHSCREENS, LONG LINES, OPTISCAMS, SPOILED PUNCH CARDS, POS SOS.

EIGHTY-SIX OF EIGHTY-EIGHT TOUCHSCREENS FLIPPED KERRY TO BUSH.
THAT'S A ONE IN 79 SEXTILLION SHOT:

1 IN 79* 10^21

CAN THE FEBBLE FUNCTION EXPLAIN THAT ANOMALY?

OVER 40,000 DOCUMETED ELECTION INCIDENTS, 99% FAVORING BUSH.
DO YOU WANT TO CALCULATE THOSE PROBABILITIES?

HOW DO YOU EXPLAIN IT?

[Melissa G]

23. Despite previous vapid msm media silence...

20 percent of Americans thought the election did not count their vote properly right after the election and by January that figure was around 30. I suspect is is significantly higher now.

you write
the goal is to win the battle of public opinion. And shouting "the house is on fire" without nailing the evidence is convincing large numbers of people (including sympathetic Democrats) that 2004 election fraud is a conspiracy theory on a par with Vince Foster. If you've heard a deafening silence recently, and your representative doesn't seem concerned, this is one of the reasons.

There are plenty of elected officials who are concerned with this issue now. I'm hearing significant rumblings in the Press... (Be sure to vote up the daily threads for Greatest)..Thanks to Tom Delay and re redistricting in Texas, my rep is no longer one of them. (I helped get Doggett elected anyway though so take that Tommy boy)... I am in contact with plenty of folks who are fighting for our rights. I hear no deafening silence. I hear the impending Thunder of Truth coming out as it always does..
What are you doing to to preserve democracy ribofunk?
Edit for clarity

[Blue Shark]

57. I Hear the Thunder too Melissa...

...and ain't it sweet?

...The dry drunk and his band of boobs are at 43% and tanking. Iraq is a fuck-all, the press is getting courageous, a few nationally know Americans are screaming fraud (may God bless you Lampley and Koehler), and the absolute coruption that comes so surely from absolute power is beginning to supporate. The smell is there... the thunder is there... and the Truth is there.

...Now Harry Reid is gonna teach the new dogs some old tricks for a change. Bring it on fools.

[tommcintyre]

26. My my... What obvious spin and attempted damage control

It's so obvious when you start out with a desired conclusion - and work back-wards to prove it.

"If you've heard a deafening silence recently, and your representative doesn't seem concerned, this is one of the reasons."
Actually, lately, the trend is DEFINITELY in the other direction. But, no matter, if you only repeat it often enough, people will begin to believe it. WRONG!

Frankly, people are getting sick of such OBVIOUS BULLSH*T - after all, they've been putting up with it for a little over four years, and seem to have had about all of it they can take! So, this kind of attempted spin "just won't fly" around here.

"The country has survived far worse election fraud than is likely to have happened last year."
Aw c'mon, did you REALLY think you could slide THAT ONE past?

Frankly, when you make such statements, it calls into question EVERYTHING ELSE you have to say. Like I said, we've had over four of practice detecting this type of spin.

[Kip Humphrey]

20. And what Democratic fraud would that be???

When 99% of all exploits favored Bush at the expense of Kerry?

Oh yeah, I forgot: dead people on the voter roles (not that they voted, just that they were still registered - some fraud!)

[On the Road]

52. Here's How the Possibility of Democratic Fraud Fits In

I don't see any evidence of systematic Democratic fraud either. But it's very important in how the argument is made.

The difference between the exit polls and the official vote is being cited as evidence of Republican fraud. On a statewide level, there may be a red shift of 6%, which may seem suspiciously high. However, once you look at the precinct-level scatterplot, there's a problem with making that argument:



Once the data is broken down into small components, the results are all over the map. There are almost as many points below the line as above. Using 6% average as a yardstick, some of the outliers appear to be 20 to 30% off. If you've gone on record as stating that a 6% discrepancy is evidence of fraud, you're in a bind. You can no longer accuse only Republicans of fraud without appearing hypocritical.

If Democrats really didn't commit systematic fraud (which I have no reason to suspect), then the argument has to be changed. That's what I hope USCV is doing now. They appear to be taking this very seriously and have done a lot of follow up.

[eomer]

54. Noise versus fraud.

As Febble has pointed out, there is a lot of noise. The fraud would be mixed in with the noise. So I don't agree that a conclusion of Bush fraud requires a conclusion of Kerry fraud. You could have noise on both sides of the divide plus fraud on just one side.

USCV has a simulator that generates means and medians similar to that of the actual data. If you can duplicate both the mean and the median then you are to some extent approximating the scatter. It would be interesting to see the scatterplot generated by their simulator to see if it looks anywhere near as noisy as this one and, if so, what kinds of assumptions were necessary to get there.

[On the Road]

63. Yes, You Certainly Can Have Noise Plus Fraud

It's just that noise dilutes the argument. For example, if exit polls were accurate, a normal vote would be a reasonably straight line. A vote with good exit polls but electronic fraud would be a reasonably straight line with electronic districts in which fraud occured raised slightly above the line.

Instead, you have an extremely wide variance, and the results are all over the place. Electronic precincts by themselves do not appear to stand out. In the exit data, the exit polls float above the zero line, but do not demonstrate that the result was due to fraud.

[Melissa G]

58. USCV is unlikely to take this seriously because it is an Irrelevant

point. And BAD framing to focus on it.
ribofunk you write
And I'm starting to see that even when her effect is taken into account, the remaining data might still contain evidence of fraud.

Yet you descend right back into this nonsense of febble's function having any real impact on the Frawd discussion? USCV is asserting the probability of Fraud. They would not waste their time continuing to discuss a mostly irrelevant point. Is This Clear to you yet?
Why IF, as you write you are "starting to see that even when her effect is taken into account, the remaining data might still contain evidence of fraud."
You continue to discuss this as if it has meaning in the FraWd discussion? Are you 1) really this confused or 2) being deliberately misleading by continuing to frame the discussion around irrelevancies and asserting that they have meaning in the discussion?
Which is it ribofunk?

[On the Road]

60. If You Read the Revised USVC Article

you can see that they've adjusted for the Liddle effect. I applaud them for making the change. But they should acknowledge it:

Here's the original chart:


In the revised version, the uppermost value has declined from about 90% to about 70%. That's a significant difference. Unfortunately, the article in Adobe and the image can't be directly linked -- it's on the top of page ten. USCV can still make an argument for fraud from the date, but it's more localized.

This is part of the process of peer review and refining the model. It should not be taken as a justification for setting up opposing camps.

[OnTheOtherHand]

61. actually, I don't think that's the Liddle effect

These are the response rate estimates, right? (The chart above is sort of hard to read!) I think those are based on algebraic reasoning that isn't affected by the Liddle effect. The reason for the change, I think, is that the original chart shows the high-Bush precincts as being 90% Bush (is that right?), whereas the revised one shows them as being 83% Bush. The latter assumption is probably closer to the actual reported distribution -- as you can see from the scatterplots, most "high-Bush" precincts aren't that high-Bush.

I'm all for acknowledging Liddle's contributions, and I agree with you about avoiding opposing camps.

[Febble]

62. Actually, they did -

They took account of the "Liddle effect" in an update to their original paper (Appendix B) AND acknowledged me.

The only problem, as I subsequently demonstrated (in my paper), is that the WPE needs to be applied at precinct level, not at the level of the means. Originally, both Baiman and myself applied the correction to the means, and in fact it appeared to strengthen the case for greater bias in Bush strongholds. It was only when I realised it had to be applied at precinct level that I started to have doubts.

Mitofsky's plots show what happens when the correction is applied at precinct level.

[OnTheOtherHand]

64. yeah, they do that too

I think every version of the working paper has acknowledged Liddle and uses alpha. And some of the tables ponder "what if alpha were fixed?" But I don't think the chart you refer to is doing that; I think it is much like Table 1 on p. 19 except using different partisanship values.

Of course, the chart and Table 1 tell related stories. The chart claims that the Kerry response rate in high-Bush precincts is "too high"; the table also implies that alpha in the high-Bush precincts is "too high." (Table 1 looks strange to me -- why do the mean WPE of -10.0 and the median WPE of -5.8 yield almost identical response rates?)

[mod mom]

31. Where are the massive reports of Dem fraud? Repug spin like WMDs were

a reason to invade Iraq.

[Bill Bored]

6. Question for Febble or someone else who knows:

Prior to this, the only precincts in which the exit polls matched the vote count were the 90 Kerry strongholds.

But why does the Febble function reduce the WPE in the 40 Bush strongholds, and yet increase the WPE in the 90 Kerry strongholds?

It would seem that the partisanship effect should work both ways, i.e., Febble's function should either reduce or increase WPE at both ends of the curve. So why doesn't it?

[autorank]

7. Well, I don't think she ever explained that very logical question...

did she. If so, oh so sorry, but I doubt I'll be appologizing any time soon.
:hi:

[Bill Bored]

9. I haven't seen it, but it's probably in the formula itself.

I was just looking for an explanation in plain English.

It's counter intuitive for the error bounds to widen at one end of the spectrun (pro-Kerry), while narrowing at the other end (pro-Bush). Doesn't mean it's wrong, but it should be questioned. Otherwise no one other than the true statistics geeks will understand it.

As I understand Febble's function, it's supposed to take into account the natural numerical tendency for the errors to be higher in less-partisan precincts and lower in more partisan ones. Now, in Mitofsky's data, there are over twice as many strong Kerry precincts as strong Bush precincts (90 vs. 40 respectively). Nevertheless, it would seem that Febble's function should work equally on both ends of the scale, and in the same direction (i.e., an increase in error rates at both ends and a decrease in the middle).

I'd expect to see an increase in error percentage at BOTH ends of the partisanship curve, if adjusted with Febble's function, whatever that is exactly.

Instead, from the scatter charts, we see what looks like an increase in the error in the pro-Kerry precincts and a decrease in the error in pro-Bush precincts. I don't get that. If she doesn't answer this, maybe I'll PM her, assuming she hasn't been driven off the board.

[autorank]

18. Bill take a look at this.

"Having ruled out the rural/urban split, Liddle contends, voting machine type strongly correlates with the voting patterns.

Does this analysis prove that election fraud actually occurred? No, and for two reasons:

* Liddle does not disclose which counties she selected, so it is impossible to evaluate how successfully she controlled for geographic factors.
* The pattern could be explained by a third, unknown variable."
http://www.eriposte.com/election04/2004_results_1.htm

Maybe it's not in the formula. This is from eriiposte.com which I'm sure you know a about. I was interested in this comment about a lack of information provided to ascertain the validity of the study. Maybe it's been provided, maybe it's part of the problem.

You can wait until you know what freezes over to get numbers from some people. Theory is cheap as is sophomoric philosophizing. I'd rely on our native sources here and Freeman. Curiosity didn't kill the cat but it sure diverted it's attention needlessly.

I think your question is good but I think that some questions that come up not are based on needless "noise" in the air lately.

The numbers are all there and they point to fraud. If you were a prosecuting attorney, you'd run over your mother (oh, wait that's Colson), you'd give your new Shelby Mustang for this type of evidence.

:hi:

[Bill Bored]

24. Uh, this refers to Liddle's work on the FL Op Scan "Dixiecrat" effect.

Febble actually confirmed Kathy Dopp's work by ruling out county size (population) as a reason for the results. I think I saw the counties she slected on a map on Kathy's site.

I don't think this story held water when other investigators found similar patterns in 2000 though. (Not that that necessarily proves anything either, but the equptment may have been different in Y2K.)

[On the Road]

8. You are Correct. If Exit Polls Are Accurate,

the effect will be symmetrical at both extremes. If a candidate gets 100% of the vote, response bias will have zero effect. That's why the general diamond shape.

The effect is not linear because: "As the true vote for the leading candidate approaches the ceiling of 100%, there would be more room to underestimate the leader's margin than to overestimate it." The effect appears to peak at about 80% support, at least for the values that Liddle tested. That results in a slightly S-shaped curve.

So far, it's still symmetrical. But overlay a 5-point bias towards Kerry in the exit polls, and it results in a higher WPE in blue precicnts and a lower WPE in red precincts, where the two effects offset each other.

The methodology Liddle was using was to start with a null hypothesis that exit polls did not indicate fraud, and that test that against the data to see if it could be disproven. The results show a pattern basically consistent with an across-the-board bias towards Kerry combined with the statistical distribution of WPE by partisanship. While there's room for further work, at this point the null hypothesis is maintained.

[Bill Bored]

10. Well, compare this plot to the one you posted:

OK, here are both scatter plots together:

Pre-Febble:


Post-Febble:


I still can't grasp why the error range would increase on the left and decrease on the right.

If Kerry had 80% of the vote or more (left), the effect of the function should be the same as if Bush had 80% or more (right). In the first curve, which you posted (pre-Febble) there was a lower WPE rate on the left and a higher WPE rate on the right. The function should result in both ends either expanding or contracting. I'd think they'd expand since the baseline (pre-Febble) was for the errors to be naturally smaller in the more partisan precincts.

Sorry if I'm dense, but I still don't understand why the result is for the WPE to increase on the left and decrease on the right.

I have no problem with adjusting the WPE to compensate for natural laws of numbers, but I think we need a clearer understanding of why the adjustment has the effect it does.

Kerry's WPE was already very low. It should have increased but so should Bush's, which was already very high.

[On the Road]

13. It's a Combination of the Two Effects

in blue precincts, the alleged Kerry bias adds to the smaller pro-Kerry WPE proposed by Liddle. In the red states, the Liddle effect is in the opposite direction and the two would cancel each other out.

That explanation was based on your earlier description of WPE being higher in heavily blue precincts than in heavily red ones. Looking at the chart, the effect doesn't look that striking. There are so many values, it's difficult to eyeball what the averages are. It supports the prediction of higher WPEs in the middle, but it's not clear by just looking at it whether it supports the other half of the equation.

[Bill Bored]

11. Slow down!

You wrote:

"So far, it's still symmetrical."

Agreed.

"But overlay a 5-point bias towards Kerry in the exit polls, and it results in a higher WPE in blue precincts and a lower WPE in red precincts, where the two effects offset each other."

Presumably a 5-point bias towards Bush in the VOTE COUNT would have the same effect then, right???

"The methodology Liddle was using was to start with a null hypothesis that exit polls did not indicate fraud, and that test that against the data to see if it could be disproven. The results show a pattern basically consistent with an across-the-board bias towards Kerry combined with the statistical distribution of WPE by partisanship."

So what? Why is this even news? Does it matter if the fraud was widespread or not? The issue is to detect it. No one should be making any assumptions about exactly what it would look like. Detection might not even be possible with exit polls. Then it wouldn't be possible to use exit polls to rule out fraud either, unless they matched the vote count and this could be shown not to be coincidental (i.e., statistically significant). Since the polls did NOT match the count, we are nowhere near proving that there was no fraud.

"While there's room for further work, at this point the null hypothesis is maintained."

I disagree. There is a bias which could still either be due to vote count corruption or polling inaccuracy or both. There is evidence for both sides. The huge Absolute Mean WPEs in 1,160 out of 1,250 precincts, rarely mentioned by either side for example. And the large Red-shifted Median WPE in all but the 90 Kerry strongholds. And the systemic bias which Febble has actually proven with her function.

Only if one assumes that the fraud was limited to Bush strongholds would there be any reason to reject the fraud hypothesis. Who said this was how it had to be done? USCV? Mitofsky? Febble? Not me.

Here's what I think:

This Exit Poll analysis is a major brain drain!

No hacking, no Clint Curtises, no Feeneys were necessary to steal votes. The e-voting products facilitate count corruption in any desired direction by anyone who knows how to use them. If there were more Republicans than Democrats that knew this, it's likely these features were used to steal votes because the temptation would simply have been too great not to.

A social Psychologist such as Febble should be able to opine on how a group of ideologically motivated people who had control of such systems might use them to achieve their ends, especially if they believe these justify the means.

This, among other things, should be the focus of the investigation and so far it has not been. This could be due to laziness, ignorance or just the obsession with these damn exit polls -- the brain drain.

[OnTheOtherHand]

12. here's the thing...

USCV did say that the higher error rates in "Bush strongholds" were evidence against uniform bias (or, presumably, uniform fraud). Here's the first sentence of their current working paper:

"New evidence from mathematical simulations conclusively shows that any constant mean exit poll response bias hypothesis such as the 'reluctant Bush responder' (rBr) hypothesis is not consistent with the pattern shown by the Edison/Mitofsky exit polling data."

Well, the second scatterplot says that's just plain wrong.

If we can agree that that's just plain wrong, then we can talk about the rest of it. (Or if "just plain" is a bit too direct, we can throw in some caveats and curlicues.) You're right that a lot of energy has been wasted defending positions that never made much sense, and now look kind of silly.

I won't try to explain the whole thing about the scatterplot, but basically: Febble's paper showed that the line would slope upward from left to right even if there were no response bias at all. (So, it would be negative at left and positive at right.) Adding bias moves the whole line above the x-axis.

[On the Road]

14. Yes, a 5-Point Pro-Bush Error in the Vote Count Would Have the Same Effect

but the way you build a case from statistics is that you propose no effect (meaning no fraud in this case) and then try to disprove that statement. The fact that both fraud and no-fraud scenarios would result in the same effect means that the null hypothesis cannot be disproven, at least by this method.

It is possible, of course, to set fraud as the null hypothesis and try to disprove that, but that's just not the way it's done. Nobody will listen to that. It borders on assuming your conclusion.

Another obstacle to the fraud scenario is that the WPE seems broad-based across machine types. The original scenario was that the vote was hacked on electronic machines. Once the polls were found to show no difference in evoting areas, the hypothesis was changed to assume that all vote counting types were hacked. The evoting scenario was spelled out in great detail. The broader-based hypothesis is not -- it's not clear that this could be done consistently across all machine types and geographies without detection. And it suffers from being a fallback hypothesis.

If the Liddle effect is factored into the numbers and the pro-Bush precincts still look like anomalies, then the argument for localized fraud can resume, and in fact be stronger because it's met closer scrutiny. It's not clear to me whether it can do that.

Yes, it's true that fraud might not be detectable through exit polls -- they're a very blunt instrument. I think localized fraud is very likely. But 130,000 votes is a big gap. If the entire margin were due to fraud, it would have to be a large enough effect to show up in the exit polls.

[Bill Bored]

21. Fair enough.

You wrote,

"Another obstacle to the fraud scenario is that the WPE seems broad-based across machine types. The original scenario was that the vote was hacked on electronic machines. Once the polls were found to show no difference in evoting areas, the hypothesis was changed to assume that all vote counting types were hacked. The evoting scenario was spelled out in great detail. The broader-based hypothesis is not -- it's not clear that this could be done consistently across all machine types and geographies without detection. And it suffers from being a fallback hypothesis."

This is why I tend to think there is probably a combination of exit poll error and vote corruption. I admit this is the ultimate fallback hypothesis, but I was never an exit poll true-believer in the first place.

We now know that there were huge WPEs in many precincts based on both Median and Absolute Mean WPEs. We also know, or we should know, that it's pretty easy to hack the tabulators, and also that the same servers (tabulator PCs) are used to program many DREs and Op Scanners before the election and that this process is frequently outsourced to partisan vendors and potentially persuadable techs, not to mention BOE insiders who are also partisan by nature, even if ostensibly bi-partisan.

So, no reason why you couldn't have rBr, rKr, Exit Poll Error and Vote Count Error, in varying degrees in different counties or states.

I hope that once more of the mere potential for fraud is revealed to the MSM and the general public, something can finally be done to minimize it.

[LightningFlash]

25. I think this says it all.

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


One could probably consider this argument moot. There is no integrity to the discussion that a bush response bias can outpace its own data, to the point that the original data Edison-Mitofsky holds no longer is valid in itself. It's what I call double reasoning. Plus on top of that, there's no way that such a sampling on purpose or mistake could be orchestrated at this level. At least not in multiple elections and all using the same affect.

http://www.votefraud.org/News/2000/7/071800.html

I think the theory is effectively dead. Was there response bias, most absolutely. Just none so phenomenonal it somehow lines up exactly with Mitofsky's exit poll statistics. But my real question is, why hasn't Mitofsky given all of the precinct data over to the PHDs and statisticians? Why has he not released it to USCountVotes and the entire team? What is his motive or operandi, that while anyone who comes out of nowhere like ESI and even Liddle can view some of the real data....yet fully hard working and credentialed master PHD scientists like Kathy Dopp, Stephen Freeman, and Ron Baiman can not?

What exactly does he hide? It's needed to be proven Bush won, isn't it? How will anyone know if all the data isn't checked and released? How come Bluementhal and others got to see the data in 2002, 2000, and other elections before when the "responder bias" was documented, yet nobody within the scientific community even got a glimpse?

Wonders will never cease in this day and age.

[On the Road]

33. When Has ANY Pollster Released the Raw Data?

That's universal. The data is their chief business asset. It's protected like a major trade secret.

I want the data released, too. In a case like this, I think it should be mandatory. But the fact that Mitofsky isn't releasing it is not the slightest bit suspicious to me.

As far as USCV goes, I would love for them to prove fraud. On the other hand, they have a mixed record so far. As far as I know, none of them has ever been involved in actual polling. Their most specific claim -- that the exit polls in partisan Republican precincts required an enormous participation by Kerry voters -- has been shown to be based on an inaccurate mathematical model.

Once USCV recalibrates, they may find other anomalies that actually point to fraud. But to me the most noticeable things in the data are (1) how much variance there is, and (2) how widespread the blue shift was.

[LightningFlash]

35. "How widespread the blue-shift is"

Again, this is hardly surprising given the source. MP was former worker for Voter News Service(VNS) and thus his polling seems to be the only credentials he has supplied.

In affect, this blue-shift has been seen in all these past elections including the 2002 gubernational race. Mitofsky has documented the same hypothesis every time, for reluctant responder and largely the issue has been ignored.

So it is hardly surprising the same thing would become postulated again. What is really telling in the data, is that this time the recorded mean does not revert back to a straight line. If you apply the revised and fully awcknowledged simulations in USCV's paper, you see the same line spike upward, deleting the possibility for the mean error being uniform.

[Bill Bored]

37. Definition of Blue Shift?

I assume you mean the shift of the Exit Polls toward Kerry, right?
Most of us call this the Red Shift, meaning the vote was shifted toward Bush. I'm not going to debate this, just need to know if your term is synonymous to avoid confusion.

[On the Road]

49. Yes, I Did Mean What's Usually Called the Red Shift

[eomer]

28. Self delete (nt)

[eomer]

46. ribofunk, let's pick up our conversation here.

I think I understand it differently but need you to clarify.

Do I understand you to say you think this quoted part needs to be corrected by Febble's function and then it would flatten out and essentially go away (as an indicator of fraud)?

If that's what you're saying then I don't believe that is right. I believe that the pattern of response rates shown in that graph are the rates that would be required to reproduce the mean WPE, media WPE and overall response rates and that this conclusion is not invalidated by, in fact not affected by, Febble's function.

Also, I think the points made earlier in the paper (pages 7 & 8 of the version currently on the USCV website) are just as important, if not more than, this one.

(BTW, I think you're not looking at the latest version of the paper, but the part you quoted seems to be essentially the same as the latest.)