On 3 Dec 2002 15:49:50 -0800, [EMAIL PROTECTED] (Humberto Barreto) wrote: > Hi, > > I can see that there are many problems for Gallup: who's got phones, how to > define likely voter (BTW, the article does say they create three groups and > weight them 1, less than 1, and 0), and on and on. > > But maybe it's a case of looking at the other side of the proverbial coin: > even with all of the warts, the sampling procedure is doing remarkably well. > That's the end point of the article. It's the Literary Digest story all over > again. People don't believe you can do a pretty good of predicting an election > where millions vote by using a sample of a few thousand. I think scientific > polling shows you can.
Isn't THAT a crummy summary! Here are better details. Literary Digest had a sample of *many* 10s of thousands, if I remember what I've read. a) They were highly *selected* -- for being well-off, more than for being 'literary'. In mid-Depression (1936), Literary Digest subscriptions would be a bit of a luxury. That is the problem noted most often, since it is the one that was obvious, and sufficient to account for the error. b) Perhaps equally important: The poll asked, "Who do you think will win?" -- not, "Who will you vote for." Your OPINION was still the topic in 1948, when polls failed to predict that Truman could beat Dewey. - The folks of those days, 1930s -40s (researchers included) considered opinions and prejudices (and thus, by extension, voter choices) to be serious, inherent, permanent, and personal. Thus, it was too intrusive to ask "Who will you vote for?" -- rather like asking, "Can I test your IQ?" today, since so many people think IQ is a meaning and permanent number. - Since they figured opinions to be permanent, it made sense, in the 1948 Truman-Dewey election, that the pollsters should stop their public polling with six weeks remaining to the election . They learned that lesson, too. > > I want to teach my students that it is hard to sample and that the "simple" in > SRS does not mean easy, but I end up teaching them that it is impossible to > do. All I do is create a bunch of critics who, no matter the design, rip it. > How can I strike a balance? I want them to doubt, but not to refuse to > believe. > The questions about "non-representative" are two-fold. First, How biased is the sample, in terms of what you want to measure? If Republicans and Democrats are equally prone to be missed, you might do okay in the two-party race. But 3rd parties are notorious for having peculiar socio-economic and age selections, I think, so the wireless phones (for instance) are a problem whenever you have to worry about the usually-most-important confounding variables. Second, Can you compensate for exactly *who* it is, that is being missed? Jesse Ventura was elected governor when 68% of his state's voters turned out, instead of the 38% voting elsewhere on that election day. I think the pollsters missed those new voters *because* of their questions like, "Did you vote in any of the last 3 elections?" - which pick out the 'likely' voters. Their phone habits would probably identify them, too -- more, today, than 4 years ago. < okay, I get carried away by details of election polling.> -- Rich Ulrich, [EMAIL PROTECTED] http://www.pitt.edu/~wpilib/index.html . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
