JFern's "Statistics"

[J. J.]

Yep, I had a good feeling that you'd miss the point.

If we have a sample of 1000, the 95% confidence interval is around +-3%. That means if Kerry has a lead of 6 or fewer points, it's not a statistically significant lead. If Kerry has a lead of 88 points (94-6), then it's quite statistically significant.

I know a lot more statistics than you, you are a fraud for blaming your lack of knowledge of statistics on my supposedly being wrong.

JFRAUD,  I assigned the confidence interval of 3%, the confidence level has nothing to do with it.  You can set both.  I can set the confidence interval at 0.1%, which requires a sample size 952775.

Even with tha virtually No margin of error (0.1%) and that sample size, the confidence level would still be 95%.  In other words, if I took 20 polls with a sample size of 952775, it is probable that at least one would not show the "actual" results within 0.1%

Likewise, I could be looking for numbers within +/-10% of the actual number and only poll 96 people, 20 different times.  The would still be the probability that one poll in those 20 wouldn't match the actual results. 

The response does affect the confidence interval, but not the confidence level.

[○∙◄☻¥tπ[╪AV┼cVê└]

Yep, I had a good feeling that you'd miss the point.

If we have a sample of 1000, the 95% confidence interval is around +-3%. That means if Kerry has a lead of 6 or fewer points, it's not a statistically significant lead. If Kerry has a lead of 88 points (94-6), then it's quite statistically significant.

I know a lot more statistics than you, you are a fraud for blaming your lack of knowledge of statistics on my supposedly being wrong.

JFRAUD,  I assigned the confidence interval of 3%, the confidence level has nothing to do with it.  You can set both.  I can set the confidence interval at 0.1%, which requires a sample size 952775.

Even with tha virtually No margin of error (0.1%) and that sample size, the confidence level would still be 95%.  In other words, if I took 20 polls with a sample size of 952775, it is probable that at least one would not show the "actual" results within 0.1%

Likewise, I could be looking for numbers within +/-10% of the actual number and only poll 96 people, 20 different times.  The would still be the probability that one poll in those 20 wouldn't match the actual results. 

The response does affect the confidence interval, but not the confidence level.

Of course if it's really a dead heat, 1 in 20 polls, will, by definition, show a statistically significant different at the 95% confidence level.  If you don't like that, then choose a widen confidence interval, like a 99.999999999% confidence interval, which still doesn't include Kerry leading Bush 94%-6%.

I can't believe you're still confused about this.

[○∙◄☻¥tπ[╪AV┼cVê└]

BTW, in case anyone caresa 1-4*10^-204 confidence interval would barely not include Kerry 94%, Bush 6%. That's 99.99...96%.

Shortening number to avoid horizontal scroll. -Alcon

[J. J.]

Yep, I had a good feeling that you'd miss the point.

If we have a sample of 1000, the 95% confidence interval is around +-3%. That means if Kerry has a lead of 6 or fewer points, it's not a statistically significant lead. If Kerry has a lead of 88 points (94-6), then it's quite statistically significant.

I know a lot more statistics than you, you are a fraud for blaming your lack of knowledge of statistics on my supposedly being wrong.

JFRAUD,  I assigned the confidence interval of 3%, the confidence level has nothing to do with it.  You can set both.  I can set the confidence interval at 0.1%, which requires a sample size 952775.

Even with tha virtually No margin of error (0.1%) and that sample size, the confidence level would still be 95%.  In other words, if I took 20 polls with a sample size of 952775, it is probable that at least one would not show the "actual" results within 0.1%

Likewise, I could be looking for numbers within +/-10% of the actual number and only poll 96 people, 20 different times.  The would still be the probability that one poll in those 20 wouldn't match the actual results. 

The response does affect the confidence interval, but not the confidence level.

Of course if it's really a dead heat, 1 in 20 polls, will, by definition, show a statistically significant different at the 95% confidence level.  If you don't like that, then choose a widen confidence interval, like a 99.999999999% confidence interval, which still doesn't include Kerry leading Bush 94%-6%.

I can't believe you're still confused about this.

You seem to be confusing the sample confidence interval with the confidence level.  What ever the result, there is still the chance that it's wrong; from a statistical standpoint, results that are below 95% are not statistically significant.

Let's say that there is another poll, conducted randomly with the same sample size at the same time.  Could that show Bush 94%, Kerry 6%?  Yes. 

One of two polls is obviously wrong, but it's wrong because of the nature of statistics.  The pollster randomly polled in a bad sample.  About one in twenty will be those bad samples; this probably accounts for some wide swings in the tracking polls.  When the sample passes through, the numbers drop back to where they were.  We really couldn't tell which of these polls.

What the poll result really says is that the poll, in 19 out of 20 cases, shows that the result is +/- 3 points of the reported result, if we poll 1067 people.  The problem is, we don't know it the 20th case or not.

That statement is true if the result 50/50 or 99/1.  What it does is change the confidence interval, known to most of us as the margin of error.  A 50/50 result of polling 1067 people would yield a MOE of +/- 3 points.  A 99/1 result of polling 1067 people would yield a MOE of +/- 0.6 points.  Both of those results would still be accurate 19 out of 20 times.

All that is does is change the MOE; it doesn't reflect on the possiblity that the sample size is the 20th case.

You don't seem to understand the difference between MOE and the accuracy of the poll.

BTW:  Any one interested in reading about it can go to the web site http://www.surveysystem.com/sscalc.htm  They can run the numbers themselves.

I'm more than happy to let anyone interested to read it and make their own judgment.

[ATFFL]

If the actual population is 51% Bush, 49% Kerry, the odds are extremely small that one of 100 polls would give something like 94% Kerry, 6% Bush. In fact, the MOE is pretty small on a 94% Kerry, 6% Bush poll (it would be 0.75%)

Are you saying that the MOE is determined by the results?  Or did you mean to say the chance of a poll showing 94-6 when the actual population is 49-51 is .75%?

I'd appreciate this getting clarified.

[muon2]

There's another factor involved in the example that is al;so important. That factor is the systematic uncertainty. Up to now the discussion has been primarily on the statistical uncertainty, but systematic uncertainty is just as important in reporting a conclusion from measured data.

All measurements make assumptions about the sample being measured. That includes the suppositions about what the complete population is that is being sampled. All measurements have some inability to predict the complete population, so measuement technique has the possibility of introducing a systematic bias.

Like statistical uncertainty, systematic uncertainty affects the confidence level of a measurement predicting the actual value. Unlike statistical uncertainty, systematic uncertainty can not be predicted merely from looking at the number of trials in a measurement sample. To find the systematic uncertainty of a measurement takes careful understanding of the measurement, and thorough cross measurements to isolate the sourecs of systematic uncertainty and their relative contributions.

Even in a high-quality experiment, it is not unusual for the systematic uncertainty to be greater than the statistical uncertainty. Polls are notorious for introducing systematic errors, and then only using statistical factors to report a confidence level. Such polls have confidence intervals substantially greater than what is expected by statistics alone.

[○∙◄☻¥tπ[╪AV┼cVê└]

Yep, I had a good feeling that you'd miss the point.

If we have a sample of 1000, the 95% confidence interval is around +-3%. That means if Kerry has a lead of 6 or fewer points, it's not a statistically significant lead. If Kerry has a lead of 88 points (94-6), then it's quite statistically significant.

I know a lot more statistics than you, you are a fraud for blaming your lack of knowledge of statistics on my supposedly being wrong.

JFRAUD,  I assigned the confidence interval of 3%, the confidence level has nothing to do with it.  You can set both.  I can set the confidence interval at 0.1%, which requires a sample size 952775.

Even with tha virtually No margin of error (0.1%) and that sample size, the confidence level would still be 95%.  In other words, if I took 20 polls with a sample size of 952775, it is probable that at least one would not show the "actual" results within 0.1%

Likewise, I could be looking for numbers within +/-10% of the actual number and only poll 96 people, 20 different times.  The would still be the probability that one poll in those 20 wouldn't match the actual results. 

The response does affect the confidence interval, but not the confidence level.

Of course if it's really a dead heat, 1 in 20 polls, will, by definition, show a statistically significant different at the 95% confidence level.  If you don't like that, then choose a widen confidence interval, like a 99.999999999% confidence interval, which still doesn't include Kerry leading Bush 94%-6%.

I can't believe you're still confused about this.

You seem to be confusing the sample confidence interval with the confidence level.  What ever the result, there is still the chance that it's wrong; from a statistical standpoint, results that are below 95% are not statistically significant.

Let's say that there is another poll, conducted randomly with the same sample size at the same time.  Could that show Bush 94%, Kerry 6%?  Yes. 

One of two polls is obviously wrong, but it's wrong because of the nature of statistics.  The pollster randomly polled in a bad sample.  About one in twenty will be those bad samples; this probably accounts for some wide swings in the tracking polls.  When the sample passes through, the numbers drop back to where they were.  We really couldn't tell which of these polls.

What the poll result really says is that the poll, in 19 out of 20 cases, shows that the result is +/- 3 points of the reported result, if we poll 1067 people.  The problem is, we don't know it the 20th case or not.

That statement is true if the result 50/50 or 99/1.  What it does is change the confidence interval, known to most of us as the margin of error.  A 50/50 result of polling 1067 people would yield a MOE of +/- 3 points.  A 99/1 result of polling 1067 people would yield a MOE of +/- 0.6 points.  Both of those results would still be accurate 19 out of 20 times.

All that is does is change the MOE; it doesn't reflect on the possiblity that the sample size is the 20th case.

You don't seem to understand the difference between MOE and the accuracy of the poll.

BTW:  Any one interested in reading about it can go to the web site http://www.surveysystem.com/sscalc.htm  They can run the numbers themselves.

I'm more than happy to let anyone interested to read it and make their own judgment.

Read a statistics book so you'll stop telling me I'm wrong when I right. If it's outside the 95% confidence interval, it's statistically significant at the p=5% level. In this case, it really doesn't matter what sort of confidence interval you choose, it's even outside of the 1 - 10^-200 confidence interval, so i't statistically significant at the p=10^-200 level.

If Kerry's support is really 50%, his polled support will be outside of that 95% confidence range range 1/20th of the time, and it'll be statistically significant at the 5% level. I'm not sure how many more times I have to say the same thing before you understand.

If the actual population is 51% Bush, 49% Kerry, the odds are extremely small that one of 100 polls would give something like 94% Kerry, 6% Bush. In fact, the MOE is pretty small on a 94% Kerry, 6% Bush poll (it would be 0.75%)

Are you saying that the MOE is determined by the results?  Or did you mean to say the chance of a poll showing 94-6 when the actual population is 49-51 is .75%?

I'd appreciate this getting clarified.

No, the odds of Kerry beating Bush by at least 94% - 6% in a random poll of 1000 people, when Kerry really trails Bush 49-51 can be found by a sum of the the binonmial distribution (or apporximated with the normal distribution, and an integral to find the normal distrubution cumulative density), and it's less than 2*10^-204.

[○∙◄☻¥tπ[╪AV┼cVê└]

There's another factor involved in the example that is al;so important. That factor is the systematic uncertainty. Up to now the discussion has been primarily on the statistical uncertainty, but systematic uncertainty is just as important in reporting a conclusion from measured data.

All measurements make assumptions about the sample being measured. That includes the suppositions about what the complete population is that is being sampled. All measurements have some inability to predict the complete population, so measuement technique has the possibility of introducing a systematic bias.

Like statistical uncertainty, systematic uncertainty affects the confidence level of a measurement predicting the actual value. Unlike statistical uncertainty, systematic uncertainty can not be predicted merely from looking at the number of trials in a measurement sample. To find the systematic uncertainty of a measurement takes careful understanding of the measurement, and thorough cross measurements to isolate the sourecs of systematic uncertainty and their relative contributions.

Even in a high-quality experiment, it is not unusual for the systematic uncertainty to be greater than the statistical uncertainty. Polls are notorious for introducing systematic errors, and then only using statistical factors to report a confidence level. Such polls have confidence intervals substantially greater than what is expected by statistics alone.

That's why I said we're assuming a random sample. If everyone voted on Diebold machines, you could take a random sample of 1000 of the actual vote, and not have to worry about any systematic errors. You'd have a completely random sample of 1000 cast votes (which might not be who they wanted to vote for, but that's irrelevant).

And to make things completely random you could have a quantum random number generator. You can buy those now and plug them into your computer.

Obviously, some polls have a much larger systematic error than the statistical error you'd get if they were truly random, a good example is the 1936 Landon landslide poll by Liteary Digest.

[J. J.]

Read a statistics book so you'll stop telling me I'm wrong when I right. If it's outside the 95% confidence interval, it's statistically significant at the p=5% level. In this case, it really doesn't matter what sort of confidence interval you choose, it's even outside of the 1 - 10^-200 confidence interval, so i't statistically significant at the p=10^-200 level.

If Kerry's support is really 50%, his polled support will be outside of that 95% confidence range range 1/20th of the time, and it'll be statistically significant at the 5% level. I'm not sure how many more times I have to say the same thing before you understand.

JFRAUD, I'll stop telling you that your wrong the decade that stop being wrong.  You are saying, in effect, that because the poll numbers show a certain thing, that means the poll is accurate. 

You are looking at one poll and assuming that the poll is one of the 19 that is correct, and your basing that on the numbers within that poll.  The results of the poll have no effect on if the sample was an "accurate" sample.  The results do effect MOE, but the only conclusion that can be reached is that Kerry is 95% likely to have support within the MOE.

Now that said, the MOE will be smaller as the support for one cadidate moves away from 50%, but that has nothing to do with if the 95% likelihood that the poll is correct.  For example, if 100 people were polled, we could say that we are 95% confident that Kerry has 94% +/- 4.65%.  Likewise, if 100,000 were polled we could say that we are 95% that Kerry has 94 +/- 0.47%.

If the numbers change the MOE changes.  Assume that Kerry has 40% of the vote, according to the poll.  If 100 people were polled we could say that we are 95% confident that Kerry has 40% +/- 9.6%.  If 100,000 were polled we could say that we are 95% confident that Kerry has 40% +/- 0.3%.

You'll note that in all cases the confidence level stays the same.  None of this affects the confidence level, only makes the MOE shrink or grow.

You are going to be able to determine the confidence level from the results of the single pole.  That is why your basic question illustrates your ignorence of the subject.  You cannot determine if this one of those randomly wrong polls from the internal numbers (though it will effect MOE).

[○∙◄☻¥tπ[╪AV┼cVê└]

Read a statistics book so you'll stop telling me I'm wrong when I right. If it's outside the 95% confidence interval, it's statistically significant at the p=5% level. In this case, it really doesn't matter what sort of confidence interval you choose, it's even outside of the 1 - 10^-200 confidence interval, so i't statistically significant at the p=10^-200 level.

If Kerry's support is really 50%, his polled support will be outside of that 95% confidence range range 1/20th of the time, and it'll be statistically significant at the 5% level. I'm not sure how many more times I have to say the same thing before you understand.

JFRAUD, I'll stop telling you that your wrong the decade that stop being wrong.  You are saying, in effect, that because the poll numbers show a certain thing, that means the poll is accurate. 

You are looking at one poll and assuming that the poll is one of the 19 that is correct, and your basing that on the numbers within that poll.  The results of the poll have no effect on if the sample was an "accurate" sample.  The results do effect MOE, but the only conclusion that can be reached is that Kerry is 95% likely to have support within the MOE.

Now that said, the MOE will be smaller as the support for one cadidate moves away from 50%, but that has nothing to do with if the 95% likelihood that the poll is correct.  For example, if 100 people were polled, we could say that we are 95% confident that Kerry has 94% +/- 4.65%.  Likewise, if 100,000 were polled we could say that we are 95% that Kerry has 94 +/- 0.47%.

If the numbers change the MOE changes.  Assume that Kerry has 40% of the vote, according to the poll.  If 100 people were polled we could say that we are 95% confident that Kerry has 40% +/- 9.6%.  If 100,000 were polled we could say that we are 95% caconfident that Kerry has 40% +/- 0.3%.

You'll note that in all cases the confidence level stays the same.  None of this affects the confidence level, only makes the MOE shrink or grow.

You are going to be able to determine the confidence level from the results of the single pole.  That is why your basic question illustrates your ignorence of the subject.  You cannot determine if this one of those randomly wrong polls from the internal numbers (though it will effect MOE).

Damn, you're still calling me Jfraud? That's ing pathetic.

At the 1-p confidence level, we have a false statistically significant difference at most p of the time. That's for a two sideded test, so if all we're testing is does Kerry have a statistically significant lead, it drops to p/2. Ok, now let p=5%. Yes, we get a false (statistically significant) positive at most 1 in 20 times.  Now, if you don't like that, fine, let p=10^-200. Kerry leading Bush 94-6 is still outside of that.

The MOE is sqrt(p*(1-p)/n), so the 100,000 MOE should be sqrt(1000) times smalelr than the 100 MOE, not 10. 



Let me repeat why I can be damn sure a 94%-6% poll is a statistically significant lead by Kerry. If it wasn't, the best Kerry could hope for was a 50-50 tie. The 1-10^-200 confidence interval for a 50-50 tied poll does not include Kerry beating Bush 94%-6%. Therefore, it's statistically significant even at the absurdly small level of p=10^-200.

[J. J.]

Damn, you're still calling me Jfraud? That's g pathetic.

At the 1-p confidence level, we have a false statistically significant difference at most p of the time. That's for a two sideded test, so if all we're testing is does Kerry have a statistically significant lead, it drops to p/2. Ok, now let p=5%. Yes, we get a false (statistically significant) positive at most 1 in 20 times.  Now, if you don't like that, fine, let p=10^-200. Kerry leading Bush 94-6 is still outside of that.

The MOE is sqrt(p*(1-p)/n), so the 100,000 MOE should be sqrt(1000) times smalelr than the 100 MOE, not 10. 



Let me repeat why I can be damn sure a 94%-6% poll is a statistically significant lead by Kerry. If it wasn't, the best Kerry could hope for was a 50-50 tie. The 1-10^-200 confidence interval for a 50-50 tied poll does not include Kerry beating Bush 94%-6%. Therefore, it's statistically significant even at the absurdly small level of p=10^-200.

Okay, first, I refer to you as JFRAUD because of "facts" you attempt to use in your posts, and I'm not the only one that has noticed that.

As to the question you asked:

https://uselectionatlas.org/FORUM/index.php?topic=20462.195

You must be logged in to read this quote.

You've asked here, "if there is a poll of 1000 random likely voters, if Kerry leads Bush with 94% of the vote to 6%, that means that Kerry has a statistically significant lead. "  The answer as you just admitted is, " Yes, we get a false (statistically significant) positive at most 1 in 20 times. "  We have no way, within that poll result, of determining it this a bad sample or not.  The only thing that can be legitimitely determined from this that we are, based on the poll results,  95% confident that Kerry is at 94% within the MOE.  If the result showed Kerry at 12%, all we could say is that we 95% confident that Kerry is at 12% within the MOE.  Neither of these results affect the confidence level; if you don't understand that, you don't have a solid grasp on how statistics work.  Why don't you ask one of your teachers.

Now if had two polls, possibly with different sample sizes, taken at the same time, we might be able  reach a conclusion, but we can not with only one poll.

You also move from straight statistical signifigance into probabilities and attempt to relate MOE to this.  Yes, MOE, the range where the 95% certainty score is, shrinks as you move away from 50%, but that has noeffect on the confidence level.  Whatever the result is, it will be a 95% confidence level.  A 50/50 split will have a different MOE, but will still be the same statistical significance.

I'm actually using the link to the site you posted; that is what that description says.

[○∙◄☻¥tπ[╪AV┼cVê└]

Damn, you're still calling me Jfraud? That's g pathetic.

At the 1-p confidence level, we have a false statistically significant difference at most p of the time. That's for a two sideded test, so if all we're testing is does Kerry have a statistically significant lead, it drops to p/2. Ok, now let p=5%. Yes, we get a false (statistically significant) positive at most 1 in 20 times.  Now, if you don't like that, fine, let p=10^-200. Kerry leading Bush 94-6 is still outside of that.

The MOE is sqrt(p*(1-p)/n), so the 100,000 MOE should be sqrt(1000) times smalelr than the 100 MOE, not 10. 



Let me repeat why I can be damn sure a 94%-6% poll is a statistically significant lead by Kerry. If it wasn't, the best Kerry could hope for was a 50-50 tie. The 1-10^-200 confidence interval for a 50-50 tied poll does not include Kerry beating Bush 94%-6%. Therefore, it's statistically significant even at the absurdly small level of p=10^-200.

Okay, first, I refer to you as JFRAUD because of "facts" you attempt to use in your posts, and I'm not the only one that has noticed that.

As to the question you asked:

https://uselectionatlas.org/FORUM/index.php?topic=20462.195

You must be logged in to read this quote.

You've asked here, "if there is a poll of 1000 random likely voters, if Kerry leads Bush with 94% of the vote to 6%, that means that Kerry has a statistically significant lead. "  The answer as you just admitted is, " Yes, we get a false (statistically significant) positive at most 1 in 20 times. "  We have no way, within that poll result, of determining it this a bad sample or not.  The only thing that can be legitimitely determined from this that we are, based on the poll results,  95% confident that Kerry is at 94% within the MOE.  If the result showed Kerry at 12%, all we could say is that we 95% confident that Kerry is at 12% within the MOE.  Neither of these results affect the confidence level; if you don't understand that, you don't have a solid grasp on how statistics work.  Why don't you ask one of your teachers.

Now if had two polls, possibly with different sample sizes, taken at the same time, we might be able  reach a conclusion, but we can not with only one poll.

You also move from straight statistical signifigance into probabilities and attempt to relate MOE to this.  Yes, MOE, the range where the 95% certainty score is, shrinks as you move away from 50%, but that has noeffect on the confidence level.  Whatever the result is, it will be a 95% confidence level.  A 50/50 split will have a different MOE, but will still be the same statistical significance.

I'm actually using the link to the site you posted; that is what that description says.

Funny how you always think I'm wrong when I'm right and you're wrong.

Anyways, you don't understand statistical significance. If I conclude that something is statistically significantly true at the 95% confidence level, that means that there is at most a 5% chance that it wasn't true. You can widen the confidence interval, which means you can conclude less things are statistically significant, but you can be more sure of the things that you do conclude are statistically significant.

The poll gives you information by itself. You use the standard error given by the poll to estimate in the error in the mean. The MOE is typically about 1.96 times the standard error for a two sided 95% confidence interval.

Before you reply to this post, learn what standard error is.

[J. J.]

JFraud, you are confusing validity of the poll, whether the poll is valid and the MOE.  You could poll one person and still make the statement that with 95% certainty that Kerry's result is 100% within the MOE.  The MOE in that case is 99%.  That was actuall SNL "Weekend Update" bit a few elections ago.  The result doesn't effect the confidence level.

You also are not looking at correlation between what causes the result and the result. 

Now, if you wish to ask if the number is outside the MOE, the answer is yes.

[muon2]

There's another factor involved in the example that is al;so important. That factor is the systematic uncertainty. Up to now the discussion has been primarily on the statistical uncertainty, but systematic uncertainty is just as important in reporting a conclusion from measured data.

All measurements make assumptions about the sample being measured. That includes the suppositions about what the complete population is that is being sampled. All measurements have some inability to predict the complete population, so measuement technique has the possibility of introducing a systematic bias.

Like statistical uncertainty, systematic uncertainty affects the confidence level of a measurement predicting the actual value. Unlike statistical uncertainty, systematic uncertainty can not be predicted merely from looking at the number of trials in a measurement sample. To find the systematic uncertainty of a measurement takes careful understanding of the measurement, and thorough cross measurements to isolate the sourecs of systematic uncertainty and their relative contributions.

Even in a high-quality experiment, it is not unusual for the systematic uncertainty to be greater than the statistical uncertainty. Polls are notorious for introducing systematic errors, and then only using statistical factors to report a confidence level. Such polls have confidence intervals substantially greater than what is expected by statistics alone.

That's why I said we're assuming a random sample. If everyone voted on Diebold machines, you could take a random sample of 1000 of the actual vote, and not have to worry about any systematic errors. You'd have a completely random sample of 1000 cast votes (which might not be who they wanted to vote for, but that's irrelevant).

And to make things completely random you could have a quantum random number generator. You can buy those now and plug them into your computer.

Obviously, some polls have a much larger systematic error than the statistical error you'd get if they were truly random, a good example is the 1936 Landon landslide poll by Liteary Digest.

My point is that it's hard to do in practice, even when all data has arrived through the same electronic means. Measurements of radioactively decaying isotopes are not free from systematic error, though the process is theoretically an entirely statistical one. A Diebold machine as an electronic database is fairly clear to query, and I would anticipate low systematic errors, but there will be some nonetheless. For instance the attempt to be random can be interferred with by the occurence of bad data - spoiled (but recountable) ballots, and machines that are offline when the sample data is collected to name two real effects.

Polls are clearly much harder than isotopes or Diebold machines for sampling since the sample of voters is not known by the polling sample. The Literary Digest is an extreme example, but there are far more subtle effects.

I'm not saying that a poll cannot control and then report meaningful systematic errors. I am saying that I don't see polls do that, and I suspect many have substantial contributions from systematics. Until the systematic error is also known it is less meaningful to get so precise on the statistical meaning of a result.

Many posters as readers of polls rely too heavily on the purely statistical meaning. It's not unlike the problem of reporting a result with five significant figures of accuracy when the input measurements have no better than two figures of accuracy. It invites conclusions beyond the meaningfulness of the data.

[○∙◄☻¥tπ[╪AV┼cVê└]

JFraud, you are confusing validity of the poll, whether the poll is valid and the MOE.  You could poll one person and still make the statement that with 95% certainty that Kerry's result is 100% within the MOE.  The MOE in that case is 99%.  That was actuall SNL "Weekend Update" bit a few elections ago.  The result doesn't effect the confidence level.

You also are not looking at correlation between what causes the result and the result. 

Now, if you wish to ask if the number is outside the MOE, the answer is yes.

If I have a random sample of 1 person, then the normal approximation completely fails. I can't really say much. I couldn't even say that I'm 95% sure that Kerry has better than 20% support. I don't see any point in working out example what happens with a sample of 1 or 2, since you definitely can't claim a statistically significant lead based on samples that small.

Again, you try to confuse the situatiion. I made it crystal clear we're talking about samples of 1000, not 1.

[J. J.]

JFraud, you are confusing validity of the poll, whether the poll is valid and the MOE.  You could poll one person and still make the statement that with 95% certainty that Kerry's result is 100% within the MOE.  The MOE in that case is 99%.  That was actuall SNL "Weekend Update" bit a few elections ago.  The result doesn't effect the confidence level.

You also are not looking at correlation between what causes the result and the result. 

Now, if you wish to ask if the number is outside the MOE, the answer is yes.

If I have a random sample of 1 person, then the normal approximation completely fails. I can't really say much. I couldn't even say that I'm 95% sure that Kerry has better than 20% support. I don't see any point in working out example what happens with a sample of 1 or 2, since you definitely can't claim a statistically significant lead based on samples that small.

Again, you try to confuse the situatiion. I made it crystal clear we're talking about samples of 1000, not 1.

I an not trying to confuse the situation.  You didn't raise the MOE point in your initial post.  The statement that, "You could poll one person and still make the statement that with 95% certainty that Kerry's result is 100% within the MOE," is a valid statement. 

Now, is there a relationship between MOE and poll results?  Yes.  Is there a relationship between confidence level and poll results?  No.  Do the results have anything to do with if the poll is statistically significant? No.

To take this to the other end, if we polled 10,000,000 voters , we could get an MOE of between 0.03 to 0.01.  The conclusion would still be that we are 95% confident that Kerry's vote is X +/- MOE, but we're still at the 95% confidence level.  The results of the poll do not effect that.

[○∙◄☻¥tπ[╪AV┼cVê└]

JFraud, you are confusing validity of the poll, whether the poll is valid and the MOE.  You could poll one person and still make the statement that with 95% certainty that Kerry's result is 100% within the MOE.  The MOE in that case is 99%.  That was actuall SNL "Weekend Update" bit a few elections ago.  The result doesn't effect the confidence level.

You also are not looking at correlation between what causes the result and the result. 

Now, if you wish to ask if the number is outside the MOE, the answer is yes.

If I have a random sample of 1 person, then the normal approximation completely fails. I can't really say much. I couldn't even say that I'm 95% sure that Kerry has better than 20% support. I don't see any point in working out example what happens with a sample of 1 or 2, since you definitely can't claim a statistically significant lead based on samples that small.

Again, you try to confuse the situatiion. I made it crystal clear we're talking about samples of 1000, not 1.

I an not trying to confuse the situation.  You didn't raise the MOE point in your initial post.  The statement that, "You could poll one person and still make the statement that with 95% certainty that Kerry's result is 100% within the MOE," is a valid statement. 

Now, is there a relationship between MOE and poll results?  Yes.  Is there a relationship between confidence level and poll results?  No.  Do the results have anything to do with if the poll is statistically significant? No.

To take this to the other end, if we polled 10,000,000 voters , we could get an MOE of between 0.03 to 0.01.  The conclusion would still be that we are 95% confident that Kerry's vote is X +/- MOE, but we're still at the 95% confidence level.  The results of the poll do not effect that.

The MOE is huge with one person. I think it would include everything with Kerry having more than 5% of the vote (not sure of that). The sample size gives you an upper bound on the MOE. It can be smaller than that, for example if there's only 1 Kerry voter out of 10 million, the MOE is a lot smaller than you said (the standard deviation or error would be 1 person in that case).

What you're missing is that all I need is one poll, and I have a 95% confidence interval, which is huge with a sample of 1, is about +- 3% if we're close to a tie with a sample of 1000, and is tiny for a sample of 10 million. If that interval doesn't include a tie, I can conclude we have a statistically signficant lead with 95% confidence. I can increase the confidence, but that increases the interval.

What part of this don't you understand?

[○∙◄☻¥tπ[╪AV┼cVê└]

Let's make this clear. Our null hypothesis is that Kerry and Bush are statistically tied, so we assume Kerry and Bush are both at 50%. Our confidence interval is 95% ( we could make it larger, if wanted).

For small samples, we have to use the binomial distribution. For samples of size 1-5, we can't conclude a statistically signifant difference, no matter with. If our sample is of size 6, we have a 1/64 chance that all 6 people support Kerry, and another 1/64 change that they all support Bush. The resuling 1/32 is less than our reject p=5%, so it would be statistically significant in that case.

For large samples, we get a standard error or deviation of sqrt(p(1-p)/n) = 0.5/sqrt(n). The radius 95% confidence interval is 1.96 standard deviations, or 0.98/sqrt(n). So basically, we have a statistically significant different if we're outside of
[50% - 1/sqrt(n), 50% + 1/sqrt(n)].  For n=1000, we get about
[46.9%, 53.1%].

[J. J.]

Let's make this clear. Our null hypothesis is that Kerry and Bush are statistically tied, so we assume Kerry and Bush are both at 50%. Our confidence interval is 95% ( we could make it larger, if wanted).

For small samples, we have to use the binomial distribution. For samples of size 1-5, we can't conclude a statistically signifant difference, no matter with. If our sample is of size 6, we have a 1/64 chance that all 6 people support Kerry, and another 1/64 change that they all support Bush. The resuling 1/32 is less than our reject p=5%, so it would be statistically significant in that case.

For large samples, we get a standard error or deviation of sqrt(p(1-p)/n) = 0.5/sqrt(n). The radius 95% confidence interval is 1.96 standard deviations, or 0.98/sqrt(n). So basically, we have a statistically significant different if we're outside of
[50% - 1/sqrt(n), 50% + 1/sqrt(n)].  For n=1000, we get about
[46.9%, 53.1%].

The null hypothesis is not a 50/50 split; that is where you are making a mistake. 

You asked a question about this would be statistically signifigant.  The numbers that poll genenerates will always be with 95% confidence, each candidates numbers are within the MOE of the result.  If you are asking if the 95% is outside of the MOE (or "confidence interval"), yes it is.  That does not effect the "confidence level."  Unless you arbitrarily decide to move it to 99% confidence level (or the third Standad Deviation), it will always be a 95% confidence level.

[○∙◄☻¥tπ[╪AV┼cVê└]

Let's make this clear. Our null hypothesis is that Kerry and Bush are statistically tied, so we assume Kerry and Bush are both at 50%. Our confidence interval is 95% ( we could make it larger, if wanted).

For small samples, we have to use the binomial distribution. For samples of size 1-5, we can't conclude a statistically signifant difference, no matter with. If our sample is of size 6, we have a 1/64 chance that all 6 people support Kerry, and another 1/64 change that they all support Bush. The resuling 1/32 is less than our reject p=5%, so it would be statistically significant in that case.

For large samples, we get a standard error or deviation of sqrt(p(1-p)/n) = 0.5/sqrt(n). The radius 95% confidence interval is 1.96 standard deviations, or 0.98/sqrt(n). So basically, we have a statistically significant different if we're outside of
[50% - 1/sqrt(n), 50% + 1/sqrt(n)].  For n=1000, we get about
[46.9%, 53.1%].

The null hypothesis is not a 50/50 split; that is where you are making a mistake. 

You asked a question about this would be statistically signifigant.  The numbers that poll genenerates will always be with 95% confidence, each candidates numbers are within the MOE of the result.  If you are asking if the 95% is outside of the MOE (or "confidence interval"), yes it is.  That does not effect the "confidence level."  Unless you arbitrarily decide to move it to 99% confidence level (or the third Standad Deviation), it will always be a 95% confidence level.

Again, you're wrong. If you're testing to see if there's a statiscally significant difference, the null hypothesis is that there is a 50-50 tie. Re-read what I said carefully, you seem to lack basic understanding of this.

[J. J.]

Let's make this clear. Our null hypothesis is that Kerry and Bush are statistically tied, so we assume Kerry and Bush are both at 50%. Our confidence interval is 95% ( we could make it larger, if wanted).

For small samples, we have to use the binomial distribution. For samples of size 1-5, we can't conclude a statistically signifant difference, no matter with. If our sample is of size 6, we have a 1/64 chance that all 6 people support Kerry, and another 1/64 change that they all support Bush. The resuling 1/32 is less than our reject p=5%, so it would be statistically significant in that case.

For large samples, we get a standard error or deviation of sqrt(p(1-p)/n) = 0.5/sqrt(n). The radius 95% confidence interval is 1.96 standard deviations, or 0.98/sqrt(n). So basically, we have a statistically significant different if we're outside of
[50% - 1/sqrt(n), 50% + 1/sqrt(n)].  For n=1000, we get about
[46.9%, 53.1%].

The null hypothesis is not a 50/50 split; that is where you are making a mistake. 

You asked a question about this would be statistically signifigant.  The numbers that poll genenerates will always be with 95% confidence, each candidates numbers are within the MOE of the result.  If you are asking if the 95% is outside of the MOE (or "confidence interval"), yes it is.  That does not effect the "confidence level."  Unless you arbitrarily decide to move it to 99% confidence level (or the third Standad Deviation), it will always be a 95% confidence level.

Again, you're wrong. If you're testing to see if there's a statiscally significant difference, the null hypothesis is that there is a 50-50 tie. Re-read what I said carefully, you seem to lack basic understanding of this.

The null hypothesis, where you reject the validity of the poll, is that the result is less than the second standard deviation from the median.  The internal numbers of the poll have no effect on that.

[○∙◄☻¥tπ[╪AV┼cVê└]

Let's make this clear. Our null hypothesis is that Kerry and Bush are statistically tied, so we assume Kerry and Bush are both at 50%. Our confidence interval is 95% ( we could make it larger, if wanted).

For small samples, we have to use the binomial distribution. For samples of size 1-5, we can't conclude a statistically signifant difference, no matter with. If our sample is of size 6, we have a 1/64 chance that all 6 people support Kerry, and another 1/64 change that they all support Bush. The resuling 1/32 is less than our reject p=5%, so it would be statistically significant in that case.

For large samples, we get a standard error or deviation of sqrt(p(1-p)/n) = 0.5/sqrt(n). The radius 95% confidence interval is 1.96 standard deviations, or 0.98/sqrt(n). So basically, we have a statistically significant different if we're outside of
[50% - 1/sqrt(n), 50% + 1/sqrt(n)].  For n=1000, we get about
[46.9%, 53.1%].

The null hypothesis is not a 50/50 split; that is where you are making a mistake. 

You asked a question about this would be statistically signifigant.  The numbers that poll genenerates will always be with 95% confidence, each candidates numbers are within the MOE of the result.  If you are asking if the 95% is outside of the MOE (or "confidence interval"), yes it is.  That does not effect the "confidence level."  Unless you arbitrarily decide to move it to 99% confidence level (or the third Standad Deviation), it will always be a 95% confidence level.

Again, you're wrong. If you're testing to see if there's a statiscally significant difference, the null hypothesis is that there is a 50-50 tie. Re-read what I said carefully, you seem to lack basic understanding of this.

The null hypothesis, where you reject the validity of the poll, is that the result is less than the second standard deviation from the median.  The internal numbers of the poll have no effect on that.

If the sample is large, and the level of statistical signifiance we are using is 95%, then it's about 1.96 standard deviations. When we are outside of that range, we conclude that there's a statistically significant difference. You can choose 3 standard deviations or 99.7%, and so on.

[J. J.]

If the sample is large, and the level of statistical signifiance we are using is 95%, then it's about 1.96 standard deviations. When we are outside of that range, we conclude that there's a statistically significant difference. You can choose 3 standard deviations or 99.7%, and so on.

Actually it exceptionally close to two standard deviations; a full description can be seen here:


http://www.robertniles.com/stats/stdev.shtml

You must be logged in to read this quote.

This, however, is NOT the MOE.  The sample is always assumed to be accurate at, but it is not 5% of the time.  Even a 50/50 result will still not represent the population (even withing the MOE) 5% of the time.  The result of the poll, who has what percentage, will not change that.  A poll that shows 50% for each candidate will be just as statistically valid as a poll that shows 96% for one candidate; the MOE will be different.

[J. J.]

I'll ad that that if you had the entire population, i.e. all voters, you could make different assumptions.

[○∙◄☻¥tπ[╪AV┼cVê└]

If the sample is large, and the level of statistical signifiance we are using is 95%, then it's about 1.96 standard deviations. When we are outside of that range, we conclude that there's a statistically significant difference. You can choose 3 standard deviations or 99.7%, and so on.

Actually it exceptionally close to two standard deviations; a full description can be seen here:


http://www.robertniles.com/stats/stdev.shtml
[/qupte]
And it's even closer to 1.96. Do you have a ing point? You seem to randomly try to find things that you think makes me lookwrong. That doesn't mean that I'm not right and you're not an idiot.

You must be logged in to read this quote.

This, however, is NOT the MOE.  The sample is always assumed to be accurate at, but it is not 5% of the time.  Even a 50/50 result will still not represent the population (even withing the MOE) 5% of the time.  The result of the poll, who has what percentage, will not change that.  A poll that shows 50% for each candidate will be just as statistically valid as a poll that shows 96% for one candidate; the MOE will be different.

Again you're wrong, 3 standard deviations is more than 99%.You'tre hopeless confused about the MOE. If I have 1000 people polled, and 94% or 96% support one candidate, that's well outside the margin of error at the 95% level, the 99% level or the 99.9999999999999% level.

I'll ad that that if you had the entire population, i.e. all voters, you could make different assumptions.

When you say something is statistically significant, you do that based upon on poll, typically a sample of 1000 or so. You completely fail to understand statistics if you don't realize that a reasonably sized sample gives statistically significant information. Go ask your statistics 101 teacher about statistical significance of opinion polls.

I can't beleive you haven't admitted that you're wrong yet. This is ing pathetic, and a good example of why I hate Republicans. No matter how obvious I made it, they insist I'm wrong.  you all.[/quote]