|
My previous post is not an argument, but an observation as to why you obtained the pattern that you did; and that it has nothing directly to do with Central Limit Theorem.
All the "statistical terminology" I used is descriptive.
Here is the argument, and thanks for the intro:
As to the big fat MOE argument, the assumption that allows the combination of smaller samples into a larger aggregate is that the within population variance for each sample population is greater than the between population variance of the sample and the aggregate. Those two sub populations (humps near the tails) I pointed to, might under an ANOVA, not be from the same statistical population.
The second assumption is that the variables affecting the variance within one sub population are independent from those affecting the variance in another population. If they are not independent, or are insignificant one may use the aggregate population's SD, but if not, then the population would be biased or skewed into one direction, and your MOE can only be calculated (if one is foolish enough to persist) by (MOE/pop 1+MOE/pop 2,+.....MOE n/pop n), which within this context comes to 4%.
Now, lets build this back wards--we know that NEP reports bias within its sample design--they did not adequately control for variability within the exit pollers, such that there was an apparent under sampling of Bush support in a sufficient number of precincts. We don't know which precincts, but the 767:341:354 WPE ratio cited suggests that this is a significant factor.
If bias is in the design, and unbiased design is what dictates the National Exit Poll MOE of 1%, then aggregation of precincts to create the national poll can only be represented by their aggregate MOE.
Now an argument can be made that for the national poll, only precincts within the 341 population were used. This would be reasonable, if the stated purpose of the exit poll was a national poll, but it is not. If anything such sorting would be to use the extremes for the national poll, and used the 341 with balanced use of the other two extremes to characterize state exit polls, leaving the residual for the national poll.
The other problem is we do not know how these precincts are spatially arranged. Their spatial arrangement may permit unbulloching the bias, but would elevate MOE.
The other problem (back of envelope calculation):
If we assume with Minivis' Ohio Exit Poll Thread that it is three precincts for each congressional district (435), then 155 precincts are left for a national exit poll. This would require a response rate of 88% if only a 100 voters of each precinct was approached to approach the total of 13,000 responses. Or, a response rate of 44% if 200 voters were approached. Does anyone begin to see a problem?
I have suggested that the national exit poll was a culling of single responses into a preweighted pattern, rather than sampling at precincts, and then aggregated into a larger sample population. It is how one with any statistical background would expect a statistician to proceed, and it explains the later additions as a response when the weightings had to be altered. This probable explanation possesses greater weight when the above is considered.
Now, can we get on to the more plausible investigations of how this election was stolen? Or do you have another take on how you can spin our wheels?
Mike
Mike
|