Wed Oct 24, 2012, 06:22 PM
cthulu2016 (7,946 posts)
A cool example of Poll Weighting in Action
Last edited Wed Oct 24, 2012, 07:47 PM USA/ET - Edit history (5)
The new Time magazine poll shows Obama up by 5% in Ohio.
It would be hard to poll a raw sample in a close state among whom Obama is ahead because older people are likelier to 1) be at home, 2) pick up the phone despite unknown caller ID, and 3) take the time to do the poll, rather than hanging up.
But the raw sample is weighted by age. The extra retired people answering the phone are discounted to match their population.
Unlike many poll reports that only give unweighted sample totals, this one appears to give both the raw sample sizes and the weighted sample sizes. The language is confusing and if I am wong I'm wong, but that's my reading of it.
Looking at Page 31.
I am giving the Total Sample first and the adjusted sample second. You get pretty much what one would expect.
65+ __ 296-178
Tons of older people did the poll, but they get cut almost in half.
Middle-aged people were in the sample only slightly over-represented. Small adjustment.
Young people do not answer the phone and take polls. Though there should have been 207, only 101 responded. So the weight of the young responses gets more than doubled.
White sample a little large, and adjusted downward slightly. Non-white sample dead on.
Okay, who is likelier to be at home when the pollster calls. The employed or the not employed?
Not Employed 401-300
Big over-sample of not employed and big under sample of employed, with big adjustments. But here's the cool part... this poll does not weight by employment. Nobody could do that. There is no reliable demographic information to weight by employment.
The big discrepancy and big correction arises from factors they do weight, most notably age. When you double 18-39 and cut 65+ almost in half you end up increasing employed and decreasing not-employed (including retired).
I like that... where weighting by the things you know seems to move other things into line.
Here's a good one... Have a school-aged child. Are people with a school-age child likelier or less likely to answer the phone and take the poll? Our intuition that they have their hands full appears to be correct:
A small sample that was adjusted upward, but that does not mean that they weight for having a school-aged child. (They could, but I strongly suspect not.) When you weight other demographic trends some things will fall into place.
Note: Not all categories add up to (what I am assuming to be) their weighted size. 783 total raw - 742 total weighted. Male/Female adds up exactly. Regions within Ohio does. Others do not. Don't know why. Probably non-responses to certain questions.
1 replies, 216 views
Always highlight: 10 newest replies | Replies posted after I mark a forum
Replies to this discussion thread
A cool example of Poll Weighting in Action (Original post)