Wednesday, July 19, 2006

Survey (USA) Says

Survey USA has released an interesting look at Bush's Job Approval numbers around the nation. The overall results are predictable, especially when you notice that the polls were taken mostly by television stations. In Rhode Island, for example, Bush has a weak-sounding 23% Job Approval, but the results came from WLNE-TV, who surveyed 600 "adults", with predictable overweighting on the young, Democrats, "Pro-Choice" adults and people who seldom attend church, if ever. On the other end of the spectrum, in Utah KSL-TV took the poll, resulting in 57% Approval in a respondent pool which was heavily Republican and who regularly attend church, but again who tend to be very young, who are not necessarily registered to vote, and which includes a sample of "pro-choice" respondents higher than the demographic norm for Utah.

What we see here is a hint to the method of poll-taking by television stations. The established nominal standard for national poll-taking is known as RDD, for Random Digit Dialed, where people are called randomly and asked identical questions. This is the process Survey USA used, per their Statement of Methodology. I would quibble just a bit, however, about their margin of error. Since each state was polled separately, each is a different poll and so the margin of error for the whole poll must reasonably be the highest MOE for any of the state polls taken, not derived from the total number of respondents nationally, because the poll was managed by fifty-two different groups (Survey USA plus 51 different TV stations), calling into question the constancy of the overall polling. I am not saying the numbers are not valid, but I am simply restricting claims to accuracy to the actual levels of procedure used. And reminding the reader again, to measure polls in their faithfulness to known demographics.


Anna said...

Yes, WKYC-TV in Cleveland is notoriously anti-Bush. Overall in '04, only 5 counties were strongly Democrat. The rest were strongly Republican (11) to independent.

Steve Falcon said...

DJ, you do some of the best poll analysis I've ever seen. I really enjoy the way you break down the numbers so that even novice (like myself) can understand.

I've pretty much stopped paying attention to all these polls for one main reason that you cite constantly: they are easily manipulated.

My college Statistics instructor used the old addage, "Figures lie and liars figure."

Steve Falcon said...

BTW, I (finally) fixed that link problem on the Bush and the Sh*t Speech article. Sorry it took so long but I've been gone all day and just got in.

Andrew Oh-Willeke said...

Sorry. The TV station only shows who paid for the poll. Survey USA does it the same way in every state- it robocalls and gives you push button answers, and selects the numbers to call in the same way in every state.

Andrew Oh-Willeke said...

The overall poll result should be based on a state population weighted average of the state's surveyed.

The overall poll MOE should be derived correctly by valuing each person surveyed in CA as a full person for sample size purposes, and discounting people surveyed in less populous states based on their population. So, the Colorado poll would be worth an increase of the sample size of about 66 people. Done this way, the total sample size exceeds the sample size in any given state, so the MOE for the entire poll should be smaller than the MOE for the highest state. Using that method (and with 600 persons surveyed in each state), the effective sample size for the nation as a whole is roughly 5,400 which produces a MOE of +/- 1.33 percentage points, considerably more accurate than the MOE in the largest state which is +/-4%.