Let's clear up misinformation about Polling Margin of Error.

[TruthIsAll]

Let's clear up misinformation about Polling Margin of Error.

The greater the sample size the more accurate the estimate,
since the MoE decreases as the sample size increases. 

The MoE is equal to 1.96 * the standard deviation (Stdev)
The Stdev is a measure of dispersion around the sample mean. 


A typical national poll samples 1000-1200.

It makes NO difference whether one is sampling a city,  state
or the nation. The ONLY thing that matters is the number of
people sampled, not the SIZE of the underlying population
(city, state, or nation).

Exit Polls taken are more accurate than standard polls.
The exit polls sampled more than 2000 people in the critical
states.
An exit poll of 2000 people has an MoE of LESS than one
percent.

The statistics supplied below are applicable to STANDARD
pre-election polls. 

For any sample size, the sample mean will be within the MoE
of the actual population mean 95% of the time.

For a 1000 sample size, the sample mean will be within 3.10%
of the actual mean 95% of the time.

For a 2000 sample size, the sample mean will be within 2.19%
of the actual mean 95% of the time.

And so on...


Sample	Sample	Std
Size	MoE	Dev

1000 	3.10%	1.58%
1250 	2.77%	1.41%
1500 	2.53%	1.29%
2000 	2.19%	1.12%
3000 	1.79%	0.91%
		
4000 	1.55%	0.79%
5000 	1.39%	0.71%
6000 	1.27%	0.65%
7000 	1.17%	0.60%
8000 	1.10%	0.56%
		
9000 	1.03%	0.53%
10000 	0.98%	0.50%
11000 	0.93%	0.48%
12000 	0.89%	0.46%
13000 	0.86%	0.44%
		
14000 	0.83%	0.42%
15000 	0.80%	0.41%
16000 	0.77%	0.40%
17000 	0.75%	0.38%
18000 	0.73%	0.37%

[Hobbes199]

1. Thanks!

Can you compare that with the MOE/std-dev for post election polls? Most pre-election polls were in the 600 sample size range, no?

(edit)
I'm sorry, I'm completely neglecting the fact that you probably just spent a couple hours on that. Thanks for the info : )

[Straight Shooter]

2. Why aren't you on Air America Radio? We need your voice.

You make my brain hurt with your statistics

:hi:

[bleever]

4. As long as he's making the right brains hurt the right way

then it's all good.

Viral liberty, in action.

[BeFree]

3. Sample sites

Usually a precint is sampled. Precincts picked for exit-polling are picked because they have historically shown to closely follow previous national election numbers.

An example: Precinct 521 has historically voted the same as the rest of the country as a whole in the last 5 presidential elections.

In other words, precinct 521, percentage wise, matched the national numbers for at least the last 5 presidential elections.

So, voters at precinct 521 are exit-polled to determine how the rest of the nation will indeed vote.

Exit-polling has proven to be accurate within the margin of error for years and years, up to the year 2002. 2002 was the first year back box voting became widespread in the good ol' US of A.

[milkyway]

5. TIA, I know you're great with numbers, but intuitively it seems wrong that

the size of the underlying popultation doesn't matter. It seems that a sample of 1,000 out of a total population of 2,000 would be much more likely to be accurate than a sample of 1,000 out of a total population of 100 million. Could you explain this further? Perhaps I'm misreading something.

[TruthIsAll]

6. I know it seems strange, but its true.

If you take an unbiased sample 1000 people in a city of 1 million, you will achieve the same accuracy if you sampled 1000 people in a state of 10 million, or 1000 people in a country of 100 million.

Your poll will have an MoE of 3.1% in either case.

This is a statistical fact which one unfamiliar with statistical sampling theory will surely find hard to believe.

[Fortunato]

7. coin flips example

As long as the canvas is completely random, the 1000+ sample size holds pretty accurate. If it's easier, think about flipping a coin over and over.

- Flip it once, and you won't even see one of the sides.
- Flip it five times and you could get five heads (or tails), which does not accurately depict the "true state of the coin" (which is 50% heads and 50% tails.)
- Flip it 100 times, and things will start to even out because the extremes will cancel each other.
- Flip it 1000 times and you should be pretty close to 500 for heads, 500 for tails.
- Flip it 10000 times and you will STILL be pretty close to half and half.

The larger the size, the closer things get to reflecting the whole accurately -- but note the point of diminishing returns. At a certain point, the "half and half" result is close to being a certainty, with very little change. The number of flips needed for this is comparable to our "minimum sample size" of 1000, while the MOE depicts how far off the certainty of this result will be with this sample size.

Anyway, if the survey neglects to reflect a certain segment of the population, however, the poll will be off. That's why so much care was taken to find precincts that seem to reflect the overall group being polled -- as an extreme example, polling a black urban precinct "randomly" when the rest of the county is white rural would result in a non-representative sample.

(I guess this is all pretty obvious already, but I felt like talking.)

[milkyway]

8. Thanks, the coin flip example really clarifys it. Now it makes sense.

[On the Road]

9. That's Correct, BUT...

Random error is NOT the only sources of discrepancy. Representative samples and response variables are much, much greater sources of error, especially in pre-election polls.

To be accurate, the pool of people contacted need to be a representative sample of the population voting on election day, and have to accurately communicate who they're voting for. Neither one of these is at all certain.

Discrepancies between pre-election polls and votes happen. Gore did better than pre-election polls in 2000, for example. It does not necessarily indicate fraud. Discrepancies between EXIT polls and votes do not necessarily indicate fraud either, even though they should match more closely. There are many response variables which lead to skewed results. Look at the early exit polls showing Kerry winning PA by 60-40. Those were not accuate.

When discussing polling discrepancies are, it is much better to portray them as discrepancies that raise concerns and call for further investigations, rather than mathematic proof of fraud. Over the last couple of weeks, the media has rightly discarded claims of mathematical proof and used that as an excuse for avoiding legitimate grounds for investigation.

[Nederland]

12. You are correct

Although I doubt that your comments will matter to those around here who are convinced that there is "mathematical proof of fraud".

The validity of the exit samples is key, and I've never seen anyone that has data on what the makeup of the 1pm and 4pm exit samples were. Its interesting to note, however, that the NEP told their customers that that those results (the 1pm and 4pm results) were inaccurate due to sampling errors. One has to ask why people here insist on believing that the 1pm and 4pm results were correct when the very company who produced them says that they were not.

The real evidence of fraud, IMHO, exists with what BevHarris is encountering down in Florida, not these types of posts.

[On the Road]

14. I Actually Think There MIGHT Be an Indication of Fraud

if eyes were trained on the right target. Two ways to commit fraud are ballot-box stuffing and destroying (or not counting) ballots. The exit polls sampled every nth person coming out of the polls and are the only independent measure of how many people voted that I'm aware of.

In some cases, the exit pollers seem to have been placed far from the exits and missed many of the people. That may have led to some of the distortions. But it still narrows the field by providing and independent source of data.

[dewaldd]

10. dude, it totally matters where the sample was taken from

1000 voters from a blue city do not reflect 1000 voters from a red rural areas

It's common sense as well as good statistics. Your statistics only apply to random samples--but that is an ideal unattainable in the real world.

[TruthIsAll]

11. Dude, you miss the point.

Of course, polling results from Blue City will not equal those taken in a Red rural area. That's obvious.

My point is that a sample of 1000 from either population will provide the same level of accuracy (3.10% MoE) as far as the Sample mean is concerned relative to the population mean.

In other words, assume a polling sample is taken Red rural area of 100,000 and another polling sample is taken from another Red rural area of 50,000.

Sampling 1000 from each will produce the same MoE of 3.10%

[Nederland]

13. Question

Will you admit that the MOE is only accurate if the exit poll sample correctly reflects the actual voting population?