|
Just give some perspective and background information. NEP readily admits they toyed with the weights -- they planned to do so from the start, as it was part of the design of the survey to provide a final total that was adjusted to the final vote. Your sharp eyed analysis would suggest that on the Democratic side, they toyed with a lot of subgroup weights in very slight amounts, whereas on the Republican side, they "bumped" the weights on a very few subgroups by 1% increments. The 1% thing is fishy, no doubt, and the fact that the subgroup weighting is done differently on each side is also fishy, but it isn't like NEP couldn't figure out a way to explain it if they were put on the spot about it.
What makes the numbers especially fishy is that if you sum up the shift in each category, it should be the same, with a small deviation for those who left questions blank (the deviations could have an excuse to get very large if a question appeared on one survey but not on another, but a large deviation would suggest a perhaps intentionally extreme weighting of respondents that left questions blank.)
If we stick to questions where we know that almost all the respondents answered (individual N sizes for each question are available to make this assessment more accurate, and are provided below.) Just doing a few here:
Bush:
Region: 261 + 261 + 522 + 522 = 1566 Age: 391 + 522 + 391 + 261 = 1565 Gender: 391 + 391 = 782 Race/Gender: 522 + 391 + 261 + 261 = 1435
Kerry:
Region: 82 + 33 - 3 - 85 = 27 Age: 70 - 109 - 97 + 21 = -115 Gender: 0 - 122 - 79 = -201 Race/Gender: 199 + 150 - 295 - 122 = -68
...that shouldn't happen on either set of numbers, unless N varies or there are non-answers. There should be the same number of "people units" added or subtracted from each category. How can you gain 391 males and 391 females, and at the same time, when broken down by race, gain 652 females and 781 males? It isn't possible. A systematic error that excluded people or doublecounted them should have done so in all questions. They would have to cop to "accidentally" doublecounting or undercounting only on certain sets of questions, which is an incredibly sloppy mistake for so-called professionals.
But we still have the problem with N varying. How much did N for each above question really vary? Well, for the final weighting, Mitofski's own numbers are:
Gender: 13,600 = 60 less respondents Race/Gender: 13,419 = 241 less respondents Age: 13,580 = 80 less respondents Region: 13,660 = this answer exists for everyone (except, apparently, the extra 59 entries in the data file which totals 13,719 lines.)
...that would mean that the "region" can be taken as gospel -- the total shift in each category should match the total shift in the region, and if it doesn't, that means that in order to get the numbers to weight the way they did, they had to extremely heavily weight people that did not respond to one question or another on the survey.
Now what if they say that they held back surveys with non-responses and didn't process them in the weighting until the end? That could be a potential excuse. A more thorough working of the numbers might be able to preempt that excuse, however.
In theory, if we find a few questions with very high N's, but where the total shift is different from the "Region" shift, we will also find that the weight assigned to the people who left the question blank is very high or very low.
Don't have time to do that now, though.
|