The MoE Formula: Mitofsky agrees its under 1.0% for over 8000 respondents

[TruthIsAll]

If the number (n) of exit poll respondents is over 8000, Mitofsky's own table states the MoE is 1%, regardless of voter characteristic percentage (p). The numbers are rounded to the nearest 1%.

The NEP table showing the MOE (for a 95% confidence interval):



Compare the table to the standard MoE Formula:
MOE = 1.96*sqrt(P*(1-P)/n)

Here are the corresponding formula-based percentages (using the middle value of the range):

%Voters
....... 100.... 150.... 350.... 725... 1650... 3800.. 6625.. 11500

05% 4.27% 3.49% 2.28% 1.59% 1.05% 0.69% 0.52% 0.40%
15% 7.00% 5.71% 3.74% 2.60% 1.72% 1.14% 0.86% 0.65%
25% 8.49% 6.93% 4.54% 3.15% 2.09% 1.38% 1.04% 0.79%
50% 9.80% 8.00% 5.24% 3.64% 2.41% 1.59% 1.20% 0.91%



These are the formula percentages for the corresponding number of exit poll respondents

%Voters
........ 3168 11027 13047 13660
.............. 7:38pm 12:22am 2:05pm

05% 0.76% 0.41% 0.37% 0.37%
15% 1.24% 0.67% 0.61% 0.60%
25% 1.51% 0.81% 0.74% 0.73%
50% 1.74% 0.93% 0.86% 0.84%

PartyID
39% 1.70% 0.91% 0.84% 0.82%

Gender
54% 1.74% 0.93% 0.86% 0.84%

New Voters
59% 1.71%

Note: there were 3168 respondents to the How Voted in 2000 category, of which 59% of 11027, 57% of 13047 and 54% of 13660 said they voted for Kerry.

[kiwi_expat]

1. This is how NEP prefaced the table:

“All samples are approximations. A measure of the approximation is called the sampling error. Sampling error is affected by the design of the sample, the characteristic being measured and the number of people who have the characteristic. If a characteristic is found in roughly the same proportions in all precincts the sampling error will be lower. If this characteristic is concentrated in a few precincts the sampling error will be larger. Gender would be a good example of a characteristic with a lower sampling error. Characteristics for minority racial groups will have larger sampling errors.

"The table below lists typical sampling errors for given size subgroups for a 95% confidence interval. The values in the table should be added and subtracted from the characteristic’s percentage in order to construct an interval. 95% of the intervals created in this way will contain the value that would be obtained if all the voters were interviewed using the same procedures.”
http://www.exit-poll.net/election-night/MethodsStatementNationalFinal.pdf


This is what NEP said about the sampling error in the FAQs for their 2004 exit polls:
“What is the Margin of Error for an exit poll?
Every number estimated from a sample may depart from the official vote count. The difference between a sample result and the number one would get if everyone who cast a vote was interviewed in exactly the same way is called the sampling error. That does not mean the sample result is wrong. Instead, it refers to the potential error due to sampling. The margin of error for a 95% confidence interval is about +/- 3% for a typical characteristic from the national exit poll and +/-4% for a typical state exit poll. Characteristics that are more concentrated in a few polling places, such as race, have larger sampling errors. Other non sampling factors may increase the total error.” http://exit-poll.net/faq.html#a15

[TruthIsAll]

2. Sounds like they are refuting their own table..

"The margin of error for a 95% confidence interval is about +/- 3% for a typical characteristic from the national exit poll and +/-4% for a typical state exit poll".

That is true for a typical National pre-election poll of 1000 respondents and pre-election state poll of 600 respondents. Just check Zogby, ARG, Gallup etc.

The National Exit Poll had 13047.
The state exit polls for battleground states ranged from 1500-2862 (FL). It was 1963 for OH.

Aren't exit polls supposed to be more accurate than pr-election telephone polls?

[kiwi_expat]

3. A characteristic like "Bush-2000 voter" might have a 3% MoE....

...because so few respondents were sampled for the "Who did you vote for in 2000?" question. It was only on one of the four national questionnaires.

The "Kerry voter" characteristic would have a lower MoE, because the question "Who did you vote for in this election" was on all the national questionnaires.

However, there are other factors involved in determining a characteristic's MoE, than just sample size. For example, Mytofsky mentions that a low response rate or "if this characteristic is concentrated in a few precincts" will result in a larger sampling error.

[TruthIsAll]

4. Bush 2000 voter? The upper limit on THAT should be 39.82%..

In THAT characteristic, we already KNOW what the maximum (assuming 100% Bush 2000 voter turnout). We don't need a sample to tell us that. It's a fact.

The Maximum Bush 2000 voter/2004 turnout ratio X is:
X= (1-.035)*50.456/122.26
or
X= 48.69/122.26 =39.82%,
since approximately 3.5% of Bush 2000 voters have DIED.

So how could it have been 43%? That is IMPOSSIBLE. Even if it were a SAMPLED result, which it wasn't, that would put it way beyond the MoE. No, the 43% was a characteristic "adjustment" weighting that was made in order to MATCH the vote.

Exit polls are MORE accurate than standard surveys. And the MoE pollsters claim for standard pre-election polls EXACTLY matches the formula. You can look them up: Zogby, ARG, Gallup... ALL of them.

So why should the Standard MoE formula NOT be used for EXIT POLLS? If anything, it is CONSERVATIVE, since EXIT polls have been proven to be MORE accurate than STANDARD POLLS. Why so?

Because the EXIT POLL SAMPLE-SIZE IS LARGER AND THE RESPONDENTS KNOW WHO THEY JUST VOTED FOR.

Kiwi, believe whatever you want to believe.

[kiwi_expat]

5. I should have said "Here is my interpretation of NEP's preface (post#1)

to the table you show in your original post...". I don't know if NEP's calculation of MoEs is correct or not.

I am not a statistician.

[mgr]

8. Did the icky real world mess up your Gaussian curve?

I think I am the first to ask you if the Central Limit Theorem is what drives this, why are the MOEs higher for states that have larger samples than the national sample, I think you need to conclude that it is a necessary, but not a sufficient explanation.

From above: "Sampling error is affected by the design of the sample, the characteristic being measured and the number of people who have the characteristic. If a characteristic is found in roughly the same proportions in all precincts the sampling error will be lower. If this characteristic is concentrated in a few precincts the sampling error will be larger."

All sampling is constrained by time and space and thus are not perfectly random as mathematics would have it. A state has a smaller spatial extent than a country, and thus has fewer opportunities for a unique characteristic to be sampled. Think gender, and think ethnicity-- gender is can be considered evenly dispersed across the country, whereas African Americans, Jews, Pakistanis, Asian Americans, Hispanic-Americans (redundant) are not, they are clustered. You recall clustering effect don't you?

If you are confused by this, think how likely you are to sample an African American from Utah, and what type of error term would you construct around a sampling event that included one? With the US, you have larger bucket to select from.

What does "Sampling error is affected by the design of the sample,...." mean to you. What is the design of a sample meant to do-- any first year psychology or sociology major will tell you--to reduce bias; and what did NEP report? If there is bias, you cannot assume independence of the various samples from each other, but dependence, which does not allow recalculation of MOE with a larger combined sample size.

Mike

[TruthIsAll]

9. 8000-1%; 8000-1%; 8000-1%; 8000-1% < LYIN' EYES?

Bush world is not real.

Bush world is surreal.

You live in Bush world.

[LatePeriduct]

6. How do you get a bunch of smoothed over red state exit polls?

By taking the exit polls from the previous elections and expanding upon them!

kick

[LatePeriduct]

7. kick

[LatePeriduct]

10. kickin

[TruthIsAll]

11. The ONE Percent Solution

.