I am including your post intact with my responses, to avoid any confusion.
....................................................
Can you please explain something to me?
http://www.truthout.org/docs_04/122704W.shtml There were three exit polls conducted in the Ukraine.
1. The Ukrainian Institute of Social Research and Social Monitoring Center poll sampled 13,500 voters at 360 polling stations.
2. The Razumkov Center of Political Studies and Kiev International Institute of Sociology interviewed 30,000 voters are 360 polling stations.
3. Frank Luntz and Douglas Schoen polled "about 10,000" voters, but the number polling stations is omitted from the article.
Please whip out your excel spreadsheet and calculate the confidence interval along with me...(assume 95% CL)
What did you get? I got:
n=13,500 / MoE=0.84%
n=30,000 / MoE=0.57%
n=10,000 / MoE=0.98%
Unfortunately, the story only reported the confidence intervals for two of these polls.
Why do you suppose that the confidence interval for both polls with sample sizes 13,500 and 10,000 was +/-2% when my handy dandy excell spreadsheet says it is MUCH lower? Are they lying to the public?
Inquiring minds want to know.
TIA:
First of all, 2% is not a confidence interval. Do you mean margin of error? If THAT is what they say it was, I suggest they speak to Edison/Mitofsky. They claim that the MOE for the NEP 13,047 randomly selected sample (which Kerry won 50.8-48.2) was 1.0%. Can you explain that? I can’t. I calculated the MOE as 0.875%.
.........................................................
Now, another question.
Mitofsky has told me in an e-mail that the design effect for the national election poll varies by the average number of interviews per precinct. Remember. There were not 13,660 separate interviews on election day, in fact, there were more like 12,200 (see MP's latest on the quadruple count of phone interviews).
Consider now that the NEP methods statement (
http://www.exit-poll.net/election-night/MethodsStatemen ... ) indicates that 250 precincts were sampled. This gives us 48.9 interviews per precinct. At 49 interviews per precinct, Mitofsky indicates that the square root of the design effect calculated for that poll was approximately 1.6. (For the moment, you don't have to buy the "clustering" factor...just follow).
Now take the sample size and calculate the confidence interval assuming no clustering. You get +/-0.89%. No problem. Now derive the standard error for this sample size s.e.=SQRT(0.25/n) and apply Mitofsky's design effect. Then calculate again for confidence interval. You get +/-1.42%. So - - - What do 0.89% and 1.42% have in common? They both round to +/-1%.
TIA:
So the 1% MOE was really 1.42%? But it was rounded off to the nearest 1%? I guess he didn’t want to scare anyone with those decimal places. After all, what’s the difference, a lousy 0.42%?
.........................................................
Basic math principle here: as n increases, the impact that the design effect square root (whatever it actually is) has on the standard error and confidence interval decreases.
Based on this principle, you can deduce that the US Exit polls were designed better than the Ukranian exit polls (i.e., the design effect for the national US exit poll had a smaller design effect than the two Ukranian exit polls). I posted on this comparison of the Ukrainian and US exits (
http://stones-cry-out.blogspot.com/2004/12/ukrainian-v -... ")
TIA:
As I said, I have not studied the Ukraine exit polls. I only know that BushCo kept alluding to them but said nothing about the discrepancies right here.
.........................................................
BTW TIA, I'm encouraged that you saw my point re: the impact of the rounding issue on precise probability calcs. Apparently, Simon/Baiman still do not.
TIA:
So why don’t the pollsters carry it out to at least 1 decimal place, like Zogby does? Could it be they need jiggle room? Or maybe they don’t want us to be too accurate in our analysis? Or both.
Simon/Baiman were actually too conservative. I get the feeling they don’t want to lay it on too heavy. For instance, in my calculations, the National Exit Poll deviation of 3.01% to Bush, based on the 1.0% MOE, gives odds of 547 million to 1. Of course, I am using the 3,01% deviation from 48.22% in the exit poll to 51.23% in the vote. The actual Bush percentage is probably slightly lower, as the late votes, which were strong for Kerry, have narrowed the gap.
In that regard, here’s a table which shows the odds of the Bush 3.01% deviation for various MOE’s.
MOE Probability Odds: 1 out of
3.00% 2.46184478% 41
2.50% 0.91413933% 109
2.00% 0.15899670% 629
1.50% 0.00419521% 23,837
1.00% 0.00000018% 547,044,797
0.87% 0.00000000% 166,787,631,560
.........................................................
JUST to stop everyone from freaking out. I am not arguing that the national exit polls were within the margin of error. They were obviously well outside the margin of error. MP doesn't even argue this point. Mitofsky doesn't even argue this point. My point is that the evidence I've presented in this post damns TIA's persistent denial of the design effect.
TIA:
I don’t deny any design effect. Just like I hope you won’t deny that the exit polls are according to Mitofsky: randomly-selected samples. And that the MOE, as published on the WP site, clearly read 1.0%. So maybe he already took the design effect into account. I have just calculated the MOE as 0.875%. Perhaps Mitofksy accounted for a 13% design effect and boosted the MOE up to 1.0%
.........................................................
TIA: If you provide a fair attempt at an answer to my questions, I will be more than happy to have a dialogue (so long as the kind moderators don't censor me for challenging one of their most adored contributors). I'll keep it polite; I promise. We are in search of truth -right?
TIA:
I appreciate your post and tone. I doubt that I’ll be satisfied with any attempts to explain away the exit poll discrepancies when the “official” results are released. There were too many documented instances of machine “glitches” which turned Kerry votes to Bush votes, voter disenfranchisement, touch screens with no paper trail (thanks to Republican control of HAVA), spoiled punched cards, Blackwell shenanigans, historical precedent (FL 2000, 2002 senate), ridiculous vote totals which defied voter registration numbers in FL, OH, NM and other states and, of course, the exit polls - state and national. We know that since 1988 the exit polls have consistently given Democrats higher margins than the recorded vote tally. According Greg Palast, prior to the election, there were at least a million spoiled ballots in mostly minority precincts (where punched cards are used). It happens in every election. I guess we’ll just have to live with it.
………………………………………………………………………………………………………………………………….