Friday, November 10, 2006
It should be noted
...that this time, the polls locally got it right. (The StarTribune seems to have a problem with missing the Hatch result, but I don't think that's a bad miss at all for the pollsters.) It shouldn't really come as a surprise to even those of us who are skeptics; you could just say a stopped clock is right twice a day (or the blind squirrel finds an occasional acorn, or...), or you could guess that if the model was bad last time, pollsters have some incentive to fix the model. I'm going to guess the second.
I have a general question about polls however, and I invite longish comments on this topic.
To what extent are polls predictive? I note that most pollsters only seem to compare their last poll to the outcome. I could do this too. I could give you a forecast of GDP or CPI or some other announced number (as opposed to a continuously observed number like interest rates) one minute before the announcement. I would have relatively complete information at that time. But few forecasters in economics do this: Our forecast horizons are almost always a matter of months, if not a year or two. Pollsters seem to give themselves a pass for a poll result three weeks before the race, or even the Friday before Election Day, by saying things like 'late breaking voters moved towards Republicans'. I could say "well, late breaking production of hula hoops, tanks, and everything in between caused my GDP forecast to be off 3.5%," and you'd rightly laugh at me. When the Wall Street Journal holds a forecaster poll and announces a "winner" of the poll, six months have elapsed, and you don't get to resample economic data to improve your forecast.
I think part of my issue is the understanding of what constitutes a likely voter. Is that someone who would be likely to vote if the election were today (when you survey), or someone who you forecast will most likely vote on Election Day? Is a "late decider" in the sample of likely voters on October 15?
In general, is a poll on October 15 supposed to be used in any way to say something about sentiment on Election Day, or is it only a statement of sentiment on October 15?
I have a general question about polls however, and I invite longish comments on this topic.
To what extent are polls predictive? I note that most pollsters only seem to compare their last poll to the outcome. I could do this too. I could give you a forecast of GDP or CPI or some other announced number (as opposed to a continuously observed number like interest rates) one minute before the announcement. I would have relatively complete information at that time. But few forecasters in economics do this: Our forecast horizons are almost always a matter of months, if not a year or two. Pollsters seem to give themselves a pass for a poll result three weeks before the race, or even the Friday before Election Day, by saying things like 'late breaking voters moved towards Republicans'. I could say "well, late breaking production of hula hoops, tanks, and everything in between caused my GDP forecast to be off 3.5%," and you'd rightly laugh at me. When the Wall Street Journal holds a forecaster poll and announces a "winner" of the poll, six months have elapsed, and you don't get to resample economic data to improve your forecast.
I think part of my issue is the understanding of what constitutes a likely voter. Is that someone who would be likely to vote if the election were today (when you survey), or someone who you forecast will most likely vote on Election Day? Is a "late decider" in the sample of likely voters on October 15?
In general, is a poll on October 15 supposed to be used in any way to say something about sentiment on Election Day, or is it only a statement of sentiment on October 15?
