TO BAIMAN, FEBBLE, OTOH, TFC, O'DELL, etc: IS rBr PLAUSIBLE?
Edited on Sun Jun-05-05 03:15 PM by TruthIsAll
Don't you think its time we moved off the pablum and got down
to real analysis?
Don't you think it's time to stop criticizing USCV's findings
and get back to work?
Don't you think it's time to test out REAL assumptions and do
some basic algebraic analysis - using constrained
Don't you think it's time to try a different approach to
determine whether or not rBr is feasible?
Don't you think it's time to respond to my challenge to stress
test the Exit Poll Optimizer?
Don't you think it's time to cut the through the rBr fog and
see if it is worth your time to defend it?
Don't you think it's time?
Let's cut through the jargon.
Let's do some work.
Let's have your inputs.
I am at your service.
I will run any and all scenarios you can throw at the model to
see if there is ONE plausible scenario which justifies the rBr
Here is the model:
EXIT POLL RESPONSE OPTIMIZATION MODEL
Objective: Determine values of constrained variables required
to derive a target Kerry/Bush percentage split using aggregate
exit poll response data.
Constraints on Precinct Variables:
1-Response rate: constrained to weighted average within (MinW,
2-Kerry win percentages constrained to (Min, Max)
3-Alpha (K/B response ratio) constrained to weighted average
4-WPE enter your own (Min, Max) range. You can choose E-M
WPE's if you wish (as shown).
RESPONSE INPUT CONSTRAINTS
Enter a value:
53.0% Wtd Average Response rate
1.12 Wtd Average Alpha (K/B )
1250 Strong Bush Strong Kerry
Prcts 40 415 540 165 90
Kerry Win Pct:
MinW 0% 20% 40% 60% 80%
MaxW 20% 40% 60% 80% 100%
MinR 0% 0% 0% 0% 0%
MaxR 100% 100% 100% 100% 100%
Min 0.10 0.10 0.10 0.10 0.10
Max 100.00 100.00 100.00 100.00 100.00
E-M -10.0% -6.1% -8.5% -5.9% 0.3%
Min -10.0% -6.1% -8.5% -5.9% 0.3%
Max -10.0% -6.1% -8.5% -5.9% 0.3%
OPTIMIZER OUTPUT SUMMARY
Number 40 415 540 165 90 Wtd Average
Resp. 45.8% 69.7% 39.5% 51.2% 63.5% 53.0%
Dev -7.2% 16.7% -13.5% -1.8% 10.5%
K/B 1.48 1.21 1.07 1.04 1.00 1.12
Dev 33% 8% -5% -7% -11% 0%
WPE -10.0% -6.1% -8.5% -5.9% 0.3% -6.77%
E-M -10.0% -6.1% -8.5% -5.9% 0.3% -6.77%
Diff 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
Kerry 52.16% of exit poll responders
Bush needed 55.28% of refusers to match his vote
Categ. HighB Bush Even Kerry HighK . Total/Avg
Prcts 40 415 540 165 90 1250
Kerry 11% 26% 53% 78% 80% 48.19%
K/B 1.48 1.21 1.07 1.04 1.00 1.12
AvgDev 33% 8% -5% -7% -11% 0%
Total 18 289 213 85 57 663
Pct 46% 70% 40% 51% 63% 53.0%
Kerry 3 91 122 69 46 330
Pct 17% 32% 57% 81% 80% 52.2%
Bush 15 198 92 16 12 332
Pct 83% 68% 43% 19% 20% 47.8%
Total 22 126 327 80 33 587
Pct 54% 30% 60% 49% 37% 47.0%
Kerry 2 27 163 61 26 279
Pct 8% 22% 50% 75% 80% 44.7%
Bush 20 99 163 20 7 308
Pct 92% 78% 50% 25% 20% 55.3%
Kerry 5 118 285 129 72 610
Pct 11.9% 28.5% 52.8% 78.3% 79.9% 48.77%
Bush 35 297 255 36 18 640
Pct 88.1% 71.5% 47.2% 21.7% 20.1% 51.23%
Kv-Bv -76.2% -42.9% 5.7% 56.6% 59.9% -2.46%
Kp-Bp -66.2% -36.8% 14.2% 62.5% 59.6% 4.31%
WPE -10.0% -6.1% -8.5% -5.9% 0.3% -6.77%
E-M WPE -10.0% -6.1% -8.5% -5.9% 0.3% -6.77%
1) Kerry win percentages over the five precinct groupings. You can use the ranges given or input your own. The model will determine the appropriate Kerry (and Bush) win percentages for each precinct grouping.
For instance, you can change Kerry's win percentage constraint from the (40-60%) range to (50-52%). The model will force the result to fall within this range for the given grouping.
2) Exit Poll Response rates. The model will generate responses by precinct which satisfy the weighted rate (53%) constraint.
3) Alpha. The model will find Alphas to satisfy the weighted K/B alpha (1.12).
4) WPE. Same as above. The defaults are E-M WPE's.
For any of these precinct min-max ranges, should you wish to constrain the result to a single value (x), then enter x for both min and max.
Th model is VERY flexible. It's a work in progress which has already yielded preliminary confirmation of Ron Baiman's findings at USCV.
The goal here is to stress test the model by throwing various combinations of input variable constraints in order to see what it comes up with - namely output response curves, alphas, WPE's, as well as the percentage of Bush and Kerry refusers in each precinct required to achieve the target vote split.
If the model cannot find a FEASIBLE SOLUTION (input variable constraints cannot ALL be simultaneously satisfied), then a partial solution will be displayed at the point where the model gave up the search.
As I understand it: your target is the two vote percentages, and your constrained variables are the mean category WPEs from the E-M report and the mean completion rates (presumably responders/(responders+refusals+misses). You allow vote percentages to vary within the range for each category.
Tell me if I've got this wrong.
And two questions:
Is your output a series of paired completion rates for Kerry voters and Bush voters, one pair for each category?
And can you tell me where the category numbers (number of precincts in each category) fit in to the model?
What I don't understand yet is whether each precinct within a category can have a different value for WPE and/or completion rate. If it can, the Solver output should give you some estimate of the variance in each for each category. You might even be able to constrain the variance in some way.
But if the precincts can't have different WPEs and response rates within each category, then I don't see how it can model the actual data very well, as the actual data has lots of variance. We know this from the absolute WPEs, and we also know that the distribution in the high Bush category was skewed, because the median was lower than the mean. Ron discusses this point in the original USCV paper. The main topic of my own paper was the distribution of the WPE variance.
If your optimizer can output variance, I'll have a go. But if it can't, no deal, I'm afraid!
11. It's time for us all to go to school: Where are they?
This was supposed to be a research project and you've provided both a reasonable proposal and the horse power to conduct the project.
Where are they? Is it only on threads that pertain to intra-mural conflict that the researchers appear?
Is there any reason they wouldn't want to come her and show you, TIA, just how right they are?
And researchers, no excuses about TIA being "lovable bug gruff." This is about a passion for knowledge serving a passion for democracy. Besides researchers, your folks can be pretty rough in the clinches as we've all observed.
Do you just chat on your own threads (febble, OTOH, Tfc, O'Dell)?
Diversity is a prime value of the Democratic Party.
Leave the 'burbs and come on into the big city.
We could even have this thread juried and published on some internet journal site. Now there's a benefit to the main reason to do participate here--let's see what you've got to say here.
In an organizational chart format, that Bush would get 55%(plus) refusers all for him which were pre-dominantly women, AND get the changed re-weighted average after 1:41 AM when the Kerry gender weight changed?
What is the real approximate probability?
If you can calculate that, then slide it through the states simulator.....It's now mute. Move on to determine what amount of bias plus fraud costed Kerry the election.
37. You don't have to know how the optimizer works.
Edited on Mon Jun-06-05 06:47 PM by TruthIsAll
Just feed it input and check the output.
Then see if the results make mathematical and intuitive sense. Why the problem?
You have provided some very sophisticated analytical writing here, yet you claim ignorance of the model even after I have spelled it out for you very clearly. You can call me arrogant if you wish. But you are far off the mark.
I issued the call to you for involvement in the process, yet you (and not only you) seem to be reluctant to pursue the stress test.
Why is that?
Why are you so reluctant when it comes to giving me a few inputs, yet you gladly spend hours in esoteric discussions which lead nowhere.
Here is a chance to learn more about the process of modeling exit poll response.
I think it's quite remarkable that Ron Baiman agrees that the model confirms his analysis. And make no mistake about it, we are working with LIMITED data here.
THE OPTIMIZATION CONFIRMS THE USCV SIMULATION. TWO COMPLETELY DIFFERENT METHODOLOGIES DERIVE ESSENTIALLY THE SAME CONCLUSION.
AS I MENTIONED TO OTOH AND FEBBLE, DON'T FOCUS ON THE FACT THAT ALPHA VARIANCE IS NOT USED. FOCUS ON THE CONSTRAINTS USED IN THE MODEL. WHY THIS RELUCTANCE TO USE IT IF IT DOES NOT RUN THOUSANDS OF SIMULATIONS?
WELL, I HAVE NEWS FOR YOU. IT IS DOING SOMETHING QUITE REMARKABLE IN COMING UP WITH THE OUTPUT. IN MY MIND, IT'S A LOT MORE SOPHISTICATED AN APPROACH THAN TAKING THAT TRIP TO MONTE CARLO.
EVEN WITHOUT SIMULATING ALPHA USING INPUT VARIANCE, WE HAVE DERIVED RESULTS WHICH MAKE SENSE.
Your model appears very similar to what USCV has been doing. I would guess that it would produce the same result, i. e. that the constant mean bias conjecture, and particularly the K=.56 and B=.5 hypothetical, cannot possibly explain the E-M data.
I have done some statistical analysis (in previous post) that shows the same thing.
In any case the argument is now irrelevant as E-M has stated that:
"There is no constant mean bias conjecture on our part. This is wholly USCV's invention."
I have an email from them to this effect, it contains other personal references so I cannot make it public - unless I get permission from the sender.
It appears that you have considerable expertise at quantitative analysis. Would you like to join our list? Please contact me if you are interested. We could then subject your model to further investigation.
20. TIA, you should most definitely join the group.
It's clear that now we've narrowed it down to a probability analysis.
There must therefore be a certain determined "probability", as per the cencus results, for Bush to get 55% of refusers in the United States population (out of literally 125 million people- Half the total population) and to get those refusers in only the eastern states.
If it can be proven what probability that is, Reluctant Lie Refuser is dead and can not be ressurected. The only problem is, that no one has made a real comprehensive case against it yet.
Then, that would leave the only alternative explanation as fraud. However, there was fraud and sampling bias. There was not massive amounts of either one, just a million purged votes in this state, a million (60,000) in Ohio and elsewhere which fill up the means.
But I'm positive now there was not more Kerry bias than Bush bias, instead there was LESS Kerry voters than Bush voters......Which means there must be a huge archive of lost Kerry votes in that pool.
Warren Mitofsky says it was just reluctant Bush responders that screwed up his exit polls. If so, then why did he have to adjust the final national weighting to show as many Repubs as Dems, and in the Western Region, MORE Repubs than Dems?
If it was just a case of reluctant responders, the turnout of the 2 main parties should have been close to historical norms after he made his little adjustments, but to get the numbers to add up, he had to show that the Dems stayed home or the Repubs turned out in droves, and didn't vote for Kerry.
Warren says there's a reluctant Bush responder hypothesis, but his own numbers employ a reluctant Kerry voter hypothesis, don't they?
California's party registration is 43% democratic to 34.7% republican. The report has party identification at 38% to 33%, which is a good match to republican registration, but not democratic. What is also problematic is that a binary comparison is being made, when all parties need to enter the picture. What may explain the 5% difference may be stronger allegiance to a third party, such as the greens, although one is a registered democrat; or oversampling of republicans (as we had seen with Gallup during the election).
Since voter registration is tied to motor vehicle registration, it is possible the party registration may be out of date, but since registration is annual, anyone feeling strongly on the matter would have switched over the last two years since 2001(the survey is 2003).
The other problem is the one characteristic of all surveys, and with the discussion of NEP's exit poll's merits needs to always be considered, and that is non-response rates. I was unable to download the data set for the survey, but if the response rates are in the 40% or greater then the error terms are going to be high; and the MOE wider.
Question 18 may be the basis for the report, but the numbers are trending away from republicans towards democrats in the post 9/11 period in 2003.
55. If there was any (net) rBr, it could be caused by aRv, or aIBv.
Edited on Wed Jun-08-05 08:47 PM by kiwi_expat
The second exit poll - showing Kerry and Bush about even - would been Mitofsky's attempt to match the recorded vote as it came in, wouldn't it?
If the difference between the first exit poll and second exit poll was at least partly the result of sampling bias (rBr), that portion of the red shift could have been caused by additional Republican voters or by additional Independent voters who voted for Bush. It would not necessarily mean that Democrats stayed home.
I agree with Ron B. that this line of analysis looks similar to what is in the USCV working paper.
What do you consider the important take-home message(s) from the model? Is it that alpha in high-Bush precincts must be very high?
O'Dell's simulation, like (apparently) your optimization, closely matches the observed E/M figures for the high-Bush precincts, and indeed both yield a high mean alpha for those precincts. However, O'Dell also shows that if one trims three high-negative-WPE outliers in the high-Bush precincts (possible pro-Bush fraud, error, or some combination), and three high-positive-WPE outliers in the high-Kerry precincts (possible pro-Kerry fraud, error, or some combination), then the natural log of alpha is about the same on average in the high-Bush and high-Kerry precincts. (See around p. 31 of his paper.)
One way of stating that result is that any pro-Bush fraud in the high-Bush precincts may have been balanced by pro-Kerry fraud in the high-Kerry precincts -- although that is taking the simulation too literally. But at any rate, a high mean alpha in high-Bush precincts by itself doesn't prove much. This is probably why Febble was urging you to model variance in alpha, as O'Dell's simulation does.
23. I will respond to your comments later this evening, when I have the time.
Edited on Mon Jun-06-05 09:04 AM by TruthIsAll
I asked you for inputs so YOU could see the take-home results for yourself.
I will say this: your interpretation may or may not be valid, or only partially so.
Keep in mind this is an optimizer not a simulator. We can restrict alpha in each precinct to a given range. That is an implicit variance, no? The model will determine the appropriate precinct partisanship alpha, subject to the weighted average alpha (1.12)
But I will experiment by adding a variance constraint to the model.
We can take our time on this and try to reach agreement on what is right -- measure twice, cut once. Even if we don't agree on all of it, at least we can agree on some.
I can see the results, but I'm not sure what you think they prove that is of interest. (I mean that literally, not snarkily.) If your main point is that the mean alpha is much higher in the high-Bush precincts, then I think you are right. I don't like the approach of categorizing entire precincts as e.g. "refusers," but I don't think it materially affects your conclusion about alpha. So, if that is your main point, I don't think I need to see the results for any other inputs.
"We can restrict alpha in each precinct to a given range. That is an implicit variance, no?"
I'm not sure. Is it possible for different high-Bush precincts to have different alphas? and if so, how would that variance be displayed in the results?
that strictly speaking, alpha could be not very high in the high-Bush precincts, but only if we postulated ridiculously skewed refusal rates? I'm not sure why that skew in refusal rates doesn't also show up in response rates. (I'm not too bothered, since I'm willing to stipulate a high mean alpha in those precincts anyway.)
(I'm not sure why you would bother to let alpha rise to 1.5 in the high-Kerry precincts, since obviously it doesn't have to.)
I take it that precincts can be assigned fractionally? So, for instance, when the table shows 6 out of 40 precincts as "Kerry votes," and that equalling 16.0%, actually the 16.0% is more precise, and it's about 6.4 precincts.
I will have to go back and look at that in slow motion. In my defense, it's probably harder to read the optimizer output without having looked at the innards. I may PM you in a bit with a question of clarification (or post it, whichever seems likely to be easier to follow).
Also my grades are a week overdue, so bear with me.
(it didn't take long once I cleared my desk) -- also registered your clarification downthread.
So, I take it the basic reason that the alpha in the middle precincts goes up is that it has to, in order for the overall alpha to remain equal to 1.12, since the high-Bush alpha is constrained to be lower than it 'wants' to be (1.48 in the original model).
That difference between 1.10 and 1.07 in the middle doesn't seem like a big deal, although it helps me understand (I think!) how your optimizer works. I don't entirely understand why the response rates are so volatile (why is there a 20-point difference in the middle precincts, from 39.5% in the original model to 59.6% in the constrained model?), but I certainly think we can rule out 21.9% for the high-Bush precincts.
(it would depend partly on whether we think of it as a population parameter -- in which case it represents millions of votes -- or a sample statistic -- in which case the difference may not be statistically significant and therefore we don't know what to think of it....)
But I don't see how the calculated alpha is consistent with the other figures. Looking at the middle precincts in the 1.2/1.2/1.5/1.5/1.5 model, I think I see 193 (or so) Kerry responders and 108 Kerry refusers, for a Kerry response rate of about 64.1%. And 129 Bush responders and 110 Bush refusers, for a Bush response rate of about 54.0%. So alpha would be 64.1%/54.0% = about 1.19. Which is close to what I would expect given the WPE of -8.5%, but a lot higher than 1.10. ???
I'm certainly confused about something -- maybe your formatting.
Kerry won 193 (60%) of the 322 who responded in the non-partisan middle group. Bush won 51% of the 218 refusers.
Bush won the 83-85% of those who refused in his partisan groups. He won 76-79% of the responders there as well.
alpha = 1.1, not .6/.4
Can we focus on the middle or "even" precincts for a moment, since those were the numbers I was citing?
Given your output, what do you think are the response rates K and B in the even precincts? I don't see where you are getting the 1.1. (I can see why .6/.4 is wrong, that isn't a problem -- no way is alpha 1.5 in the middle precincts.)
(K is the Kerry response rate, and B is the Bush response rate)
But why should it be true that Kerry Responders =(?) Kerry vote% * (K/B) * Total Responders??
I think you were thinking something like
Kerry Responders = Kerry vote% * (K/R) * Total Responders
where R would be the overall response rate. So, for instance, suppose we have 100 Bush voters and 100 Kerry voters, but K = 56% and B = 50%, so we end up with 56 Kerry responders and 50 Bush responders (106 total responders). Alpha = 0.56/0.50 = 1.12. By your equation, it should be true that
56 =(?) 0.50 * 1.12 * 106
but actually the right side of that equals 59.36.
If you do
56 == 0.50 * (0.56 / 0.53) * 106
K/B is close to K/R when B is close to R regardless of K because there are very few Kerry voters, as in the high-Bush precincts -- or when K and B are close to each other, as they apparently were in the high-Kerry precincts. So when I eyeballed those alphas, they seemed familiar. It's the one in the middle that really seems too low to generate a WPE of -8.5%. Ron Baiman figured it at 1.186.
They helped stage and allow 9/11 to happen, and have committed mass crimes and treason. Their lackies including Thomas Noe stole the election after 45% of the religious whackjobs marched for Bush. :mad: :mad:
Powered by DCForum+ Version 1.1 Copyright 1997-2002 DCScripts.com
Software has been extensively modified by the DU administrators
Important Notices: By participating on this discussion
board, visitors agree to abide by the rules outlined on our Rules
page. Messages posted on the Democratic Underground Discussion Forums are the
opinions of the individuals who post them, and do not necessarily represent
the opinions of Democratic Underground, LLC.