Monday, October 30, 2006
I have tried for weeks to get some pollsters in Minnesota on our radio show, but to no avail so far. One has been busy assisting WCCO with its Bachmann smear campaign (I have to wonder which editorial genius came up with the let's-refight-the-Reformation strategy), and another is no doubt preparing the last poll for the StarTribune. My friends at the SCSU Survey haven't released their poll yet either, though it seems they've been in the field a long, long time judging by the buzz of phone calls I've heard there the last two-plus weeks.
I'm not a pollster; I am not trained in sampling theory. I do use statistics extensively in my work, though, and I'm a forecaster. So when I read the best story on polling written this year, by no less an eminence than Michael Barone, I have to pay attention.
In 2004, the electorate that went to the polls or voted absentee was, according to the adjusted NEP exit poll, 37 percent Democratic and 37 percent Republican. In party identification, it was the most Republican electorate since George Gallup conducted his first random sample poll in October 1935.
But most recent national polls show Democrats with an advantage in party identification in the vicinity of 5 percent to 12 percent. Party identification usually changes slowly. Historically, voters have switched from candidates of one party to candidates of the other more readily than they have changed their party identification.
I have noted before that local polls have been heavily leaning turnout towards Democrats and women. Now, to be sure, the pollsters will respond that they didn't do that, these are the results of screens they run to determine who is a likely voter. I accept that explanation only to a point. Surely they know what Barone says in this last paragraph; do they really believe this shift is real? And if so, is it because they want to believe it, just as much as Charlie Brown both knows Lucy's history and yet wants to believe that this time the voters won't pull the ball back at the last second?In any forecasting exercise, one step after you "run the model" is to do a consistency check. I make students tell me a story for how all the things the model says will happen can happen at the same time. As I tell them, your client will forgive a bad forecast if you have a story for why it went wrong. But you have to then fix the model. If you keep giving the same bad forecast and keep telling the same story and yet your client doesn't fire you it's not your problem any more. Sometimes we just have stupid clients.
You have to wonder if that's the problem with the Democrats. Certainly they also know what Barone says here:
Serious pollsters concede that there are some problems with polling. Americans have fewer landline phones than they used to, and the random digit dialing most pollsters use does not include cell-phone numbers. Larger and larger percentages of those called are declining to be interviewed.
Interviewers can inject bias in the results. The late Warren Mitofsky, who conducted the 2004 NEP exit poll, went back and found that the greatest difference between actual results in exit poll precincts and the reports phoned in to NEP came where the interviewers were female graduate students -- and almost all the discrepancies favored the Democrats.
But if you know that, don't you fix the problem? No, if the polls keep allowing you to engage in a fantasy you can't resist -- that this time the Republicans have really gone too far, that this time the voters have seen through the lies, that this time you shall be delivered to the promised land of majority and speakership. And that's no less a fantasy, no less wish-fulfillment, than the guy who looks at the Sports Illustrated swimsuit issue each spring thinking this is the summer he will get a girl who looks like that. They don't study polls; they look at them as the dandy looks in the mirror in admiration and sure that this is the time all will recognize his beauty.