Forum Navigation
Topics
Log In
Log Out
:
Forum Search
New Today
New This Week
Advanced Search
Tree View
Forum Account
Edit Profile
Register
Forgot Password
Forum Tools
Help/Instructions
Policies
CLICK STATE TO SEE:
"WATCH LIST"
Marked with: 
"OPEN & HONEST"
Marked with: 
...
|
| (US) 11/06 - EDA Report: 3 million v... |
|
| Author |
Message |
    
Daniel R Spak
Voting Rights Forum Participant
Username: Mapleboard
Post Number: 21
Registered: 12-2004
Best of Black Box? N/A
Votes: 0 (A keeper?) | | Posted on Saturday, November 18, 2006 - 6:29 am: | 
|
Edison/Mitofsky "edit" their exit polls, once again, to match final count, like 2004 presidential election...
http://electiondefensealliance.org/major_miscount_of_vote_in_2006_election |
Saul Iversen
Voting Rights Forum Participant
Username: Malachite
Post Number: 53
Registered: 07-2006
Best of Black Box? N/A
Votes: 0 (A keeper?) | | Posted on Saturday, November 18, 2006 - 9:34 pm: |
|
To avoid any sense of conspiracy or wrongdoing, it should be noted that the "National Election Pool" polls were never intended to be used to verify the accuracy of the actual vote. Other polls were used with this end in mind. See the following analysis of a statistician contractor hired by Ediston-Mitofsky. She demonstrates the false sense of utility that the public often attributes to these polls.
http://www.dailykos.com/story/2006/11/4/135126/905 |
Charles Christopher
Voting Rights Forum Participant
Username: Ilikeinfo
Post Number: 11
Registered: 11-2006
Best of Black Box? N/A
Votes: 0 (A keeper?) | | Posted on Sunday, November 19, 2006 - 12:50 am: |
|
http://www.dailykos.com/story/2006/11/4/135126/905
Nice but leaves out the most critical peice of the the answer, Google "confidence interval" and, while your eyes may glaze over, you'll find the most important part of the puzzle that *NOBODY* ever mentions. Yes, shes right that you need an even sample *BUT* that guarentees *NOTHING*.
If I recall the tables correctly (it's been a while) -
Stats say you can sample a very small part of a literally unlimited group of objects/votes/whatevers and get accurate statistics.
Stats then gives the following number of samples to take and you'll note that most polls *DO* adhere to these numbers (~600 or ~1200):
Sample 600 of that group, where each had the same chance of being picked, and you have a 95% that the result will be withing 5% of the acutal number.
Sample 1200 of that group, where each had the same chance of being picked, and you have a 95% that the result will be withing 3% of the acutal number.
The part that is missed is that, in both cases, there is a 5% chance your result is total nonsense .... That's why it's called a "confidence INTERVAL" and thats why the numbers are often so far off. In other words *NOBODY* I've ever read or heard *EVER* properly explains what the margin of error TRUELY is in these poll.
And that did'nt take pages and pages to explain. |
Catherine Ansbro
Frequent Voting Rights Forum Participant
Username: Catherine_a
Post Number: 3468
Registered: 12-2004
Best of Black Box? N/A
Votes: 0 (A keeper?) | | Posted on Sunday, November 19, 2006 - 3:28 am: |
|
In addition to your point, Charles, there is a major weakness in Elizabeth Liddle's analysis. She points to it herself:
Liddle (Febble):
quote:There is absolutely no correlation between the magnitude of the precinct-level discrepancies in 2004 and change in Bush's vote share relative to 2000 (what UK commentators call "swing"). If a single factor, e.g. fraud, was responsible both for both the discrepancy, and for inflating Bush's vote, then you would expect the two to be positively correlated. In fact, the correlation is slightly, but insignificantly, negative. If fraud was responsible for the discrepancy in 2004 then either it was absolutely uniform (which is not the case normally made) or it was carefully targeted in precincts (not states) where Bush was expected do badly. Either way, the data does not support the inference that the discrepancy was due to fraud; heroic assumptions need to be made to make it even consistent with fraud on a very large, nationwide scale.
Liddle was very limited in how she looked for fraud to show up; assuming that fraud would not be "absolutely uniform" or not be "carefully targeted in precincts where Bush was expected to do badly" are just some of the weaknesses. There was little or no understanding of how fraud can be applied using different methodologies in different precincts and in various combinations. The assumption that one can look for a simplistic uniform correlation even at the precinct level is naive at best and shows little understanding of the real world of election manipulation. The idea of comparing to 2004 election "results" as any kind of indicator of fraud or not is absurd.
Much information is available about the real world of election manipulation because there are so many known examples, even if we have no way of putting names on the perpetrators.
Unfortunately statistical experts such as Liddle are not sufficiently well-informed about election fraud, and most of the election integrity folks are not statisticians with specific expertise in election polling. Both fields are fairly specialized and there is not yet enough overlap. |
Saul Iversen
Voting Rights Forum Participant
Username: Malachite
Post Number: 56
Registered: 07-2006
Best of Black Box? N/A
Votes: 0 (A keeper?) | | Posted on Sunday, November 19, 2006 - 9:24 am: |
|
I'll have to study what you have posted. It's quite interesting. Let me ask this in the mean time: do you agree that the NEP polls are not intended to audit the accuracy of the vote? And therefore any hint of skepticism of foul play on their part is misplaced? And would be better targeted with another poll? |
Catherine Ansbro
Frequent Voting Rights Forum Participant
Username: Catherine_a
Post Number: 3472
Registered: 12-2004
Best of Black Box? N/A
Votes: 0 (A keeper?) | | Posted on Sunday, November 19, 2006 - 12:08 pm: |
|
I don't know enough about the NEP polls to have an opinion. I saw the assertion that you mention, but I don't know what to make of it.
I would be a little surprised if someone claimed they were expected to have no relation to election results. I suppose one could claim that none of these polls have been set up to audit the vote in the strictest sense of the word "audit". That said, I don't think it follows to assume there should be no expected relationship, if both polls and elections are being carried out properly. |
Saul Iversen
Voting Rights Forum Participant
Username: Malachite
Post Number: 72
Registered: 07-2006
Best of Black Box? N/A
Votes: 0 (A keeper?) | | Posted on Wednesday, November 22, 2006 - 11:36 am: |
|
Here are Ms. Liddle's responses to some comments above:
------------------
OK, let's take Charles Chistopher first:
Nice but leaves out the most critical peice of the the answer, Google "confidence interval" and, while your eyes may glaze over, you'll find the most important part of the puzzle that *NOBODY* ever mentions. Yes, shes right that you need an even sample *BUT* that guarentees *NOTHING*.
I don't leave out "confidence interval" - I simply used the term "Margin of Error". And I am not sure what is meant by an "even sample".
If I recall the tables correctly (it's been a while) -
Stats say you can sample a very small part of a literally unlimited group of objects/votes/whatevers and get accurate statistics.
Yes.
Stats then gives the following number of samples to take and you'll note that most polls *DO* adhere to these numbers (~600 or ~1200):
Sample 600 of that group, where each had the same chance of being picked, and you have a 95% that the result will be withing 5% of the acutal number.
Sample 1200 of that group, where each had the same chance of being picked, and you have a 95% that the result will be withing 3% of the acutal number.
Well, not exactly. The formula is rather more complex than that, particularly with a clustered sample (sampled voters from sampled precincts), but yes, the MoE is calculable on the assumption of a random sample.
The part that is missed is that, in both cases, there is a 5% chance your result is total nonsense .... That's why it's called a "confidence INTERVAL" and thats why the numbers are often so far off. In other words *NOBODY* I've ever read or heard *EVER* properly explains what the margin of error TRUELY is in these poll.
No, this is not quite right. There is a 5% chance that the "true" proportion will lie outside the 95% confidence limits of your estimate. There is a 0.5% chance it will lie outside your 99.5% confidence limits. There is a 20% chance that it will lie outside your 80% confidence limits. There is no clear magical divide between sense and nonsense. The "true" proportion will have a binomial probability distribution around your estimate.
But what Charles Christopher ignores is the fact that these confidence intervals assume a random sample. They are the confidence limits of your answer given sampling error alone . They do not allow for "non-sampling error". No poll can be free from "non-sampling error". Some sources of non-sampling error will cancel out over many polls (and the exit poll is, in effect, many polls at precinct level) but some may not. In other words some sources of non-sampling error may produce a systematic bias when aggregated over many polls. This is a problem that all surveys face. Ironically the problem is more apparent in large surveys, as the sampling error is much smaller. In large surveys, a larger proportion of the error variance will be non-sampling error, which means that small sources of non-sampling error may often be "statistically significant" whereas in a smaller poll, with a larger MoE, such errors would be less likely to push the sample proportions outside it.
OK, here is a response to Catherine Ansbro:
In addition to your point, Charles, there is a major weakness in Elizabeth Liddle's analysis. She points to it herself:
Liddle (Febble): (quote)
There is absolutely no correlation between the magnitude of the precinct-level discrepancies in 2004 and change in Bush's vote share relative to 2000 (what UK commentators call "swing"). If a single factor, e.g. fraud, was responsible both for both the discrepancy, and for inflating Bush's vote, then you would expect the two to be positively correlated. In fact, the correlation is slightly, but insignificantly, negative. If fraud was responsible for the discrepancy in 2004 then either it was absolutely uniform (which is not the case normally made) or it was carefully targeted in precincts (not states) where Bush was expected do badly. Either way, the data does not support the inference that the discrepancy was due to fraud; heroic assumptions need to be made to make it even consistent with fraud on a very large, nationwide scale.
Liddle was very limited in how she looked for fraud to show up; assuming that fraud would not be "absolutely uniform" or not be "carefully targeted in precincts where Bush was expected to do badly" are just some of the weaknesses.
I did not make these assumptions. Indeed, I flagged these possibilities as possible mechanisms by which widespread fraud of a magnitude to account for the exit poll discrepancy might be consistent with the data. If you like, I provided a bar over which any widespread fraud hypothesis has to climb. However, for several reasons, my own estimate is that they are improbable. I am willing to consider evidence or hypotheses to the contrary.
To take the first: absolutely uniform - or indeed largely uniform - fraud might be consistent with the data. In this scenario, all the variance in redshift would be accounted for by sampling and non-sampling error, but the net redshift would be due to very widespread but largely uniform fraud. I conducted a number of Monte Carlo simulations, and found that only if fraud on a scale to account for the exit poll discrepancy was both extremely widespread (in around 80% of precincts in the poll) and had very little variance (whether calculated in terms of percentage of votes flipped, or size of margin changed) would there be a sporting chance of the correlation between "swing" and "redshift" being negative (and yes, I also factored in the extent of the generally low mean state WPEs for 2000). This means, in effect, that the fraudsters would have had to a) be lucky and b) to ensure that any increase in fraud over 2000 was perpetrated to the same magnitude on levers, punchcards, DREs and optical scanners - no more and no less. And even then, while this would account for the net redshift, statewide, it would leave many states with a "significant" bias, either in Kerry's direction or Bush's. Notably, it would still leave New York with a substantial pro-Kerry bias in the poll. Now of course I believe bias in any poll is perfectly possible. But if we accept that some polls in some states will have significant bias in favor of one or other candidate, then there is no logical reason to reject sampling bias as a cause of a net bias nationwide. And in fact, there is substantial evidence, including actual experimental evidence, for predominantly pro-Democratic bias in exit polls.
There was little or no understanding of how fraud can be applied using different methodologies in different precincts and in various combinations.
Well, I don't know on what basis Catherine Ansbro claims to understand my understanding. Indeed I did consider different methodologies in different precincts and in various combinations. It's why I consider uniform fraud pretty untenable as a hypothesis. Which brings me to consider the second: fraud precisely targetted only in those precincts where Bush was doing badly. This would require a dynamic algorithm, plus input regarding expected voteshare and turnout for each precinct. Again, this would have to be a strategy implemented across methodologies for it not to show up in the data. In which case, how might she suggest that such a dynamic algorithm could be implemented on levers machines? (And the answer isn't "tabulators", because these discrepancies were calculated, for most precincts from precinct counts, collected at the precinct on election night, not on county tabulations).
The assumption that one can look for a simplistic uniform correlation even at the precinct level is naive at best and shows little understanding of the real world of election manipulation.
Perhaps Catherine Ansbro would like to suggest a mechanism by which fraud on the scale of millions could have been perpetrated and NOT produced a "simplistic" positive correlation between swing and shift. The lack of correlation certainly does not rule out piecemeal fraud, and indeed there are many forms of election corruption that would not show up in the exit poll data anyway. So I am certainly not so 'naive' as to suppose that fraud cannot have occurred because it didn't show up as a correlation between shift and swing. What I do suggest, however, is that vote-switching (or differential vote deleting) fraud is unlikely to have occurred on the scale implied by those who infer it from the exit poll discrepancy.
The idea of comparing to 2004 election "results" as any kind of indicator of fraud or not is absurd.
I assume Catherine Ansbro means "2000". No, it is not absurd. The exit poll discrepancies in 2000 had a small net "redshift". The exit poll discrepancies in 2004 showed a large redshift. Bush increased his official margin from slightly less than zero in 2000 to several points in 2004. Some have alleged that both increases had the same cause - new levels of vote-switching fraud made possible by e-voting. Bush's precinct vote-share in 2000 was very highly correlated with his vote-share in 2004. However, in some precincts he increased his margin rather more than his average, and in some he increased it rather less (or reduced it). If a proportion of the variance in his performance was due to fraud of a type likely to be reflected in exit polls ( e.g. vote-switching), then variance in exit poll discrepancy will be shared with variance in the his "swing" from 2000. When variance in two variables is shared, it tends to produce a correlation. The failure to find such a correlation in such a large dataset (1250 precincts) strongly suggests that there was little or no variance shared between the two phenomena. This is not an "absurd" finding. It is simply information. If a fraud scenario is to be hypothesised that is consistent with the data it has to take this finding into account.
My point simply is that if vote-switching fraud was perpetrated on the scale indicated by the exit poll discrepancy, there is a high probability that it would have shown up as a correlation between "shift" and "swing", and a very low probability that that correlation would have been zero. Ergo, I infer that, horrendous as the insecurites in your voting systems are, it is improbable that votes were switched on anything like the kind of millions implied by the exit poll discrepancy. On the other hand, it is perfectly clear that in both 2004 and 2006, e-voting was unreliable, may well have been corrupt, and voter suppression tactics, including robocalls and push polls, were rife.
Feel free to post these reponses if you wish.
(Message edited by Malachite on November 22, 2006) |
Catherine Ansbro
Frequent Voting Rights Forum Participant
Username: Catherine_a
Post Number: 3493
Registered: 12-2004
Best of Black Box? N/A
Votes: 0 (A keeper?) | | Posted on Wednesday, November 22, 2006 - 4:12 pm: |
|
Elizabeth Liddle:
quote:If you like, I provided a bar over which any widespread fraud hypothesis has to climb.
To me, this sounds like Elizabeth makes an assumption of what "widespread fraud" would look like.
FWIW I don't personally consider exact amount of fraud over all precincts and all election systems (paper, various electronic, punch card, lever, etc.) to be a realistic scenario, either. On this we agree.
Eliabeth Liddle:
quote:Which brings me to consider the second: fraud precisely targetted only in those precincts where Bush was doing badly. This would require a dynamic algorithm, plus input regarding expected voteshare and turnout for each precinct. Again, this would have to be a strategy implemented across methodologies for it not to show up in the data. In which case, how might she suggest that such a dynamic algorithm could be implemented on levers machines? (And the answer isn't "tabulators", because these discrepancies were calculated, for most precincts from precinct counts, collected at the precinct on election night, not on county tabulations).
But why target "only those precincts in which Bush was doing badly"? The whole point is that many different kinds of fraud can be implemented in different locations, each to varying degrees, some simultaneous (and indeed by both parties). Some precincts may have had no fraud. Some may have had small amounts of fraud in ways not easily detectible (e.g. in precincts in which a candidate was particularly strong or weak--it could all depend on unique opportunities), or with a little bit a vote-shaving. Some precincts may have used combinations of machine-based DRE fraud with vote-shaving across a few precincts combined with various disenfranchisement tactics.
quote:So I am certainly not so 'naive' as to suppose that fraud cannot have occurred because it didn't show up as a correlation between shift and swing. What I do suggest, however, is that vote-switching (or differential vote deleting) fraud is unlikely to have occurred on the scale implied by those who infer it from the exit poll discrepancy.
See my comment below about why shift and swing in relation to past elections cannot be trusted. Those who suggest fraud occurred on a large scale include disenfranchisement as well--not just alleged voting machine fraud and not just electronic voting machine fraud.
There are many more kinds of voting machine fraud than vote switching or differential vote deleting. What about adding votes in punch cards or optical scan ballots? Lever machines and punch cards are also not immune to fraud of various kinds.
Yes, I did mean "2000".
Elizabeth and I disagree about the helpfulness of comparing to past election results. Our election systems have been broken so badly, for so long (virtually non-existant custody chain of ballots where any form of paper-based systems were used; one lever machine malfunction in 2005 was later discovered to have been malfunctioning for the previous 5 years), that comparisons to previous results are not meaningful. It might only indicate that cheating of one kind changed to another kind, or that a dishonest election official was replaced with someone more honest or less good at certain kinds of cheating, or that a new cheating strategy was tried here and there with varying degrees of success.
We have little evidentiary basis for believing past or current election results. We can't prove that fraud probably didn't exist on a large scale (and in many cases we do know that fraud on a large scale did indeed take place, such as in Chicago).
When voting mechanisms change, we are still in the same boat--there are few if any meaningful baselines for comparison because of the extraordinary sloppiness of all aspects of our so-called democracy. The sloppiness we now know about goes back many decades and perhaps much more, and it is possible that we have only come to know about a few of the worst cases. We have little reason to believe that there were outstanding procedures in place anywhere. For this reason, I do not believe that looking for "shift" or "swing" in relation to previous election results is a reliable way to discover either indicators of likely fraud or of likely accuracy. This anaylysis would have to be based on the assumption that there was a past frame of reference that was both honest and correct. Yet this kind of baseless assumption is what is used by statisticians/polling experts to provide factors to adjust poll results to the reported returns.
I think the fundamental premise is faulty in relation to most US locations. US election systems have been too broken for too long (insufficient oversight over almost any aspect of the election process) for statistical work to provide a reliable basis for ruling fraud in or out in all but the most blatant (and therefore unlikely) examples. If rigorous statistical work were applied in a context in which prior results preceding a given date were known to have been both honest and accurate (because of consistent voting and auditing procedures and strict non-partisan oversight), that would at least provide a better starting point. But I do not know of any such starting point. We've gotten too used to elections based on nothing but trust, with virtually no public oversight. This makes statistical analysis problemmatic.
Exit poll analysis would only make sense in places where tightly scrutinized election protocols had been in place for a significant period of time in the past. Otherwise the factors used to adjust for WPE would be based on faulty data. Garbage In => Garbage Out. |
Saul Iversen
Voting Rights Forum Participant
Username: Malachite
Post Number: 73
Registered: 07-2006
Best of Black Box? N/A
Votes: 0 (A keeper?) | | Posted on Wednesday, November 22, 2006 - 8:03 pm: |
|
May I ask a trivial question, Catherine? How do you separate the quotations as you have done in your reply? I'd like to mimic that as appropriate. Thanks. |
Catherine Ansbro
Frequent Voting Rights Forum Participant
Username: Catherine_a
Post Number: 3497
Registered: 12-2004
Best of Black Box? N/A
Votes: 0 (A keeper?) | | Posted on Thursday, November 23, 2006 - 3:06 am: |
|
Saul, by coincidence I just posted about how to do that on a different thread here
(Message edited by Catherine_a on November 23, 2006) |
Saul Iversen
Voting Rights Forum Participant
Username: Malachite
Post Number: 76
Registered: 07-2006
Best of Black Box? N/A
Votes: 0 (A keeper?) | | Posted on Thursday, November 23, 2006 - 8:48 pm: |
|
More posting by proxy (from Ms. Liddle of course). Forgive me if I improperly attribute any of these quotes (see prefixes as appropriate):
-----
It sounds like I largely agree with Catherine, and that our only disagreement arises from the fact that I don't appear to have made my points very clear. I am certainly not arguing that 2004, or any previoius election, was clean. The evidence is clear that more people both nationwide and in Florida cast a vote for Gore than for Bush, and that if your election system had been fair, Gore would probably be your president still. I think there is copious evidence that the election system has systematically disenfranchised Democrats for years, possibly for ever, through voter suppression, "vote suppression" (more uncounted Democratic votes) and outright fraud (although I am less sure whether the evidence suggests a net Democratic disadvantage from outright fraud - I'd like to think so, though).
My point re the exit poll data is a specific point regarding a specific inference made from that data - that the unprecedentedly large discrepancy between poll and count in 2004 was caused by a new form of fraud, made possible by e-voting. I myself considered this a viable hypothesis at one stage, which was why I thought it should be tested. However, my findings strongly support the hypothesis that differential participation rates were responsible for the net "redshift" between poll and count in 2004, and, moreover, render improbable the hypothesis that an increased prevalence of vote-switching (or differential within-precinct deletion of Democratic votes) was responsible for a substantial proportion of the "redshift", and that therefore it is unlikely that such a form of fraud occurred on a scale of millions. However, it remains perfectly possible that the sum total of disadvantage to Democrats was very large - there are many forms of election corruption that would not show up in exit polls, including high residual vote rates in extreme Democratic precincts; fraud concentrated in key races; voter suppression; under-supply of voting equipment to Democratic precincts. What I am trying to do is to glean what can be gleaned from the exit poll data, specifically, as to what kind of fraud may have occurred on what kind of scale in 2004. In other words, to see what patterns of fraud might be consistent with the exit poll data. More inline:
quote:
quote:
Liddle: If you like, I provided a bar over which any widespread fraud hypothesis has to climb.
Ansbro: To me, this sounds like Elizabeth makes an assumption of what "widespread fraud" would look like.
No, I am not assuming anything. What I am doing is starting with what the exit poll data look like, and trying to infer what kind of widespread fraud is consistent with it. Widespread vote-switching fraud on a large scale does not appear to me to be plausible given these data, unless it was a) uniform, which seems unlikely for reasons I gave or b) targetted only on precincts in which Bush was doing worse than his average, which also seems unlikely. Widespread voter suppression, on the other hand, or vote-switching concentrated in a few key areas IS, I think, consistent with the data.
quote:Ansbro: FWIW I don't personally consider exact amount of fraud over all precincts and all election systems (paper, various electronic, punch card, lever, etc.) to be a realistic scenario, either. On this we agree.
Yes. Thanks. I suspect we agree on more than this. I will try to be clear.
quote:
quote:
Liddle: Which brings me to consider the second: fraud precisely targetted only in those precincts where Bush was doing badly. This would require a dynamic algorithm, plus input regarding expected voteshare and turnout for each precinct. Again, this would have to be a strategy implemented across methodologies for it not to show up in the data. In which case, how might she suggest that such a dynamic algorithm could be implemented on levers machines? (And the answer isn't "tabulators", because these discrepancies were calculated, for most precincts from precinct counts, collected at the precinct on election night, not on county tabulations).
Ansbro: But why target "only those precincts in which Bush was doing badly"?
Only because such a strategy would tend to counteract any tendency for fraud to be correlated with advantage to Bush. I agree it is unlikely, which is why I give it a low prior. It is simply a hypothesis that might be consistent with the data, were it to be plausible, which I think it isn't.
quote:Ansbro: The whole point is that many different kinds of fraud can be implemented in different locations, each to varying degrees, some simultaneous (and indeed by both parties). Some precincts may have had no fraud. Some may have had small amounts of fraud in ways not easily detectible (e.g. in precincts in which a candidate was particularly strong or weak--it could all depend on unique opportunities), or with a little bit a vote-shaving. Some precincts may have used combinations of machine-based DRE fraud with vote-shaving across a few precincts combined with various disenfranchisement tactics.
This is all true, and I don't dispute it in the least. If fraud occurred, and I strongly suspect it did, I would expect it to take this form. But if the vote-switching component was on the scale of millions inferred by some from the exit poll data, then I would expect to see a correlation between redshift and swing in the exit poll data, and we don't. Therefore the probability of vote-switching being on the scale of millions is low.
quote:
quote:
Liddle: So I am certainly not so 'naive' as to suppose that fraud cannot have occurred because it didn't show up as a correlation between shift and swing. What I do suggest, however, is that vote-switching (or differential vote deleting) fraud is unlikely to have occurred on the scale implied by those who infer it from the exit poll discrepancy.
Ansbro: See my comment below about why shift and swing in relation to past elections cannot be trusted. Those who suggest fraud occurred on a large scale include disenfranchisement as well--not just alleged voting machine fraud and not just electronic voting machine fraud.
Yes, I understand this. It is specifically the hypothesis that the exit poll discrepancy in 2004 evinces massive, widespread vote-switching fraud that I think is contra-indicated by the data. I have been banging the voter-suppression drum since November 2nd 2004! I think it was a huge effect. And I also agree that e-voting is unreliable and vulnerable to fraud. I was following closely your own efforts to track this down in the months following that election. My only disagreement (not necessarily with you) is with the argument that the 2004 exit poll discrepancy gives us an index of the magnitude of electronic (or other vote-switching) fraud. I think it doesn't. In fact, I think it suggests that the scale of such fraud is likely to have been small and localized.
quote: Ansbro: There are many more kinds of voting machine fraud than vote switching or differential vote deleting. What about adding votes in punch cards or optical scan ballots? Lever machines and punch cards are also not immune to fraud of various kinds.
Indeed. And the exit poll discrepancy, in urban precincts, was actually higher where levers and punchcards were used. However, that wouldn't show up as a correlation between swing and shift, because unless it increased between 2000 and 2004, it wouldn't show up as an increase in swing.
quote: Ansbro: Yes, I did mean "2000".
Elizabeth and I disagree about the helpfulness of comparing to past election results. Our election systems have been broken so badly, for so long (virtually non-existant custody chain of ballots where any form of paper-based systems were used; one lever machine malfunction in 2005 was later discovered to have been malfunctioning for the previous 5 years), that comparisons to previous results are not meaningful.
I agree entirely about your broken election systems. But I disagree that past results are not helpful when testing a specific hypothesis a pattern of fraud postulated to have increased between 2000 and 2004, and to have been manifest in an increased exit poll discrepancy. I think such an analysis is not only helpful, but informative. I think helps define the likely scale of a specific problem. It does not solve the problem, nor does it tell us the scale of other problems. But defining and scaling a problem is an important step in trying to fix it.
quote: Ansbro: It might only indicate that cheating of one kind changed to another kind, or that a dishonest election official was replaced with someone more honest or less good at certain kinds of cheating, or that a new cheating strategy was tried here and there with varying degrees of success.
Indeed. And those changes are precisely what I was testing with that hypothesis.
quote:Ansbro: We have little evidentiary basis for believing past or current election results. We can't prove that fraud probably didn't exist on a large scale (and in many cases we do know that fraud on a large scale did indeed take place, such as in Chicago).
Of course not. But what we can say is that whatever fraud occurred in 2000 (the reference year) it was not of a type to be manifest as a large redshift in the exit poll data. 2000 is therefore an appropriate baseline for investigating fraud of a type that would be so manifest, given the large redshift in 2004.
quote:Ansbro: When voting mechanisms change, we are still in the same boat--there are few if any meaningful baselines for comparison because of the extraordinary sloppiness of all aspects of our so-called democracy. The sloppiness we now know about goes back many decades and perhaps much more, and it is possible that we have only come to know about a few of the worst cases. We have little reason to believe that there were outstanding procedures in place anywhere.
I agree.
quote:Ansbro: For this reason, I do not believe that looking for "shift" or "swing" in relation to previous election results is a reliable way to discover either indicators of likely fraud or of likely accuracy.
Well, see my reasoning above.
quote:Ansbro: This anaylysis would have to be based on the assumption that there was a past frame of reference that was both honest and correct. Yet this kind of baseless assumption is what is used by statisticians/polling experts to provide factors to adjust poll results to the reported returns.
No, it does not assume that the past frame of reference was either honest or correct. It assumes that in 2000 the scale of exit-poll visible fraud (e.g. vote-switching, differential vote-deletion within-precinct) was lower than that postulated in 2004, which is an inference from the observation that the mean redshift in 2000 was significantly lower than in 2004. That doesn't mean that 2000 was honest or fair. It patently wasn't - it returned the wrong candidate to office.
quote:Ansbro: I think the fundamental premise is faulty in relation to most US locations. US election systems have been too broken for too long (insufficient oversight over almost any aspect of the election process) for statistical work to provide a reliable basis for ruling fraud in or out in all but the most blatant (and therefore unlikely) examples. If rigorous statistical work were applied in a context in which prior results preceding a given date were known to have been both honest and accurate (because of consistent voting and auditing procedures and strict non-partisan oversight), that would at least provide a better starting point. But I do not know of any such starting point. We've gotten too used to elections based on nothing but trust, with virtually no public oversight. This makes statistical analysis problemmatic.
It does indeed, and is probably why so much bad statistical inference has gone on. It was also why I took so much care and so much time over the data analysis I did for Warren Mitofsky. But again, see my reasoning above vis a vis the rationale for the analysis.
quote:Ansbro: Exit poll analysis would only make sense in places where tightly scrutinized election protocols had been in place for a significant period of time in the past. Otherwise the factors used to adjust for WPE would be based on faulty data. Garbage In => Garbage Out.
Well, it's why any analysis needs to have a clear rationale. I too have seen a lot of GIGO analyses of the exit poll data. I do not believe mine comes into this category.
Cheers,
Lizzie |
Catherine Ansbro
Frequent Voting Rights Forum Participant
Username: Catherine_a
Post Number: 3499
Registered: 12-2004
Best of Black Box? N/A
Votes: 0 (A keeper?) | | Posted on Friday, November 24, 2006 - 12:03 am: |
|
Yes, there is much on which Elizabeth and I agree. For example, I don't think either of us thinks the exit polls of 2004 have enough information to reveal what degrees of fraud & disenfranchisement occurred, what multiple tactics were used (by either party), or on what scale. It's debatable to what extent electronic voting machines played a role in impacting election results in 2004. Debatable meaning not knowable. (BBV certainly uncovered hard evidence of fraud using electronic voting machines in targeted areas it investigated. Were these locations the tip of the iceberg? Or merely a few pockets that were far worse than elsewhere? Could a few pockets of fraud on unauditable machines produce enough change to affect an election result? Sure--but we'll never know for absolute certain that this happened. There is no way to know. It seems that election machines and election processes have been designed to prevent us from knowing.)
If much manipulation was provided by disenfranchisement and manipulation of non-electronic machines (tactics common in the past so there's no reason to assume they suddenly stopped), and then a crucial extra amount of fraud with electronic machines was also involved, one could separate out each kind of fraud and rightly claim that it was the determining factor in affecting the results. (It was the DREs! It was the optical scanners! It was the tabulators! It was disenfranchisement! It was the punch cards! It was the ballot design!) You could make a case for each of these providing "the" crucial factor in the 2004 election results, not just in FL.
Individual tactics that could stay within margins of error would be unlikely to throw up red flags when probed statistically. Combined tactics that would produce results that would be seen as "reasonable" or "expected" may have been particularly important. But whether these expectations were based on accurate prior elections is another question.
Elizabeth:
quote:No, it does not assume that the past frame of reference was either honest or correct. It assumes that in 2000 the scale of exit-poll visible fraud (e.g. vote-switching, differential vote-deletion within-precinct) was lower than that postulated in 2004, which is an inference from the observation that the mean redshift in 2000 was significantly lower than in 2004. That doesn't mean that 2000 was honest or fair. It patently wasn't - it returned the wrong candidate to office.
This does not give me any reason to have confidence that there is historical evidence that can provide an accurate base line. The 2000 exit polls themselves would have been adjusted/corrected (WPE) as a result of other previous elections. And our reason for trusting the accuracy of those previous elections is . . .???
All the WPE adjustment factors are based on a layer of previous unverified and unverifiable election results, which were in turn based on yet more layers of previous elections' unverified results. Polling WPE correction factors have been based on multiple layers of myth from preceding elections, creating a virtual reality that could be pure fiction for all we know.
That doesn't mean they are all wrong. And it doesn't mean they are all right.
There is no way of knowing which past elections were good, which weren't, and which ones can therefore provide an accurate basis for appropriate WPE adjustments. Without a provable, known, accurate, fraud-free historical basis for adjusting for WPE, the usefulness of exit polls as an indicator of degrees of accuracy or inaccuracy in vote counting is greatly reduced. Exit poll statisticians do too little questioning of the accuracy of the historical basis on which their own results inevitably rely.
With few exceptions we have no basis for confidence in the results of elections. Exit polls whose WPE adjustments rely on previous election "results" are inherently flawed.
Hence the importance of reclaiming the public's right to observe, and therefore know, that elections are fairly run, and that final election results are complete and accurate reflections of the citizens who chose to participate. |
Saul Iversen
Voting Rights Forum Participant
Username: Malachite
Post Number: 78
Registered: 07-2006
Best of Black Box? N/A
Votes: 0 (A keeper?) | | Posted on Friday, November 24, 2006 - 7:48 am: |
|
From Ms. Liddle:
-----
Catherine, thanks for your response. However, a little clarification would seem to be in order:
The WPE is not "adjusted/corrected". It is simply the difference, in percentage points, between the margin between the two leading candidates in the exit poll responses and the margin in the official count (collected, in most cases, at the precinct, on election night). For 2000 I had the mean state WPEs. These were very much smaller for 2000 than for 2004.
That does not tell us that 2000 was "good", but it does tell us the kind of fraud that is likely to show up as a discrepancy at precinct level between poll and count (precinct level vote-shifting, or deletion of Democratic votes) was unlikely to have had been on the scale inferred by some from the 2004 exit poll discrepancy.
When considering the exit polls, it is important to distinguish between the projections of the the state results (which, unless the margin in the poll responses is extremely wide will incorporate vote returns, and the closer the race, the greater the number of returns that are incorporated before a confidence level of 99.5% is reached, and the state is "called"), the cross-tabulations (made from a subsample of questionnaire responses, and subjected to "post-stratification re-weighting" in line with the vote returns), and the raw responses themselves, which are compared with the count at precinct level after the election so that the pollsters can try to estimate where and why their responses diverged from the count. One measure of this divergence is the "WPE". I, together with my colleagues Mark Lindeman and Rick Brady,
http://inside.bard.edu/~lindeman/ASApaper_060409.pdf
devised what we believe is a better measure, and it was that measure I used to reanalyse Mitofsky's data, precisely to try to find out what might have contributed to the divergence. However, mean state WPEs remain a reasonably valid way of estimating the magnitude of either bias in the sample or corruption in the count (it can't intrinsically tell you which) in any given election.
As for your final point, of course, I completely agree. |
|
|
