Jon Stewart, at packed “Daily Show” party, on exit polls: “We thought they were scientific. Turns out they just ask a few guys who are hanging around after they vote.” (Wonkette)

It’s 3:44 AM. I was about to go to bed, but while I was brushing my teeth I had a startling thought. Why are we surprised Bush won? The exit polls can’t work.

First, let’s talk about what an exit poll is. According to the PBS online glossary, an exit poll is

“a poll taken of voters as they leave the place of voting. The pollster asks a small percentage of the voters whether they voted for or against particular candidates and issues.”

So, after people voted, the pollsters would ask some of them how they voted. This, of course, explains exactly why the exit polls displayed anormalities. In a normal election with a heavy preference for one candidate over all others, the normal errors in exit polling are negligable to the margin between the candidates. But, in a closely contested election, like Bush v.s. Kerry 2004, some additional polling errors manifest.

First, there’s an issue of selection. Are pollsters more likely to interview Democrat than Republican voters? It’s hard to say–suffice that there are more reported Democrats than Republicans, although the inverse is actually true.

Second, is motivation. Are the people you select more likely to report their Kerry vote, or their Bush vote. Assume for a second that Kerry voters have different psychological tendencies than Bush voters. Then, when you ask a voter to report their votes, a Bush voter might slink into the background, while the Kerry voter proclaims his vote for America’s New Leader. It’s safe to say we might feel ashamed voting for Bush–and it’s obvious what this would do to an exit poll.

For the future, let me propose a more robust voter polling algorithm:

1. If possible, assign each incoming voter a unique identifier
2. Select voters at random to poll from the pool in 1
3. If the voter accepts the poll, record his vote into a pool of votees
4. If the voter does not accept the pole, ignore his vote, and either
   • remove one vote from the pool at random
   • ignore the next vote

You need to correct both for the pollsters’ preferences, and the voters’ preferences–this algorithm correctly does that. If you don’t, you’ll get the kind of media-skewed polling data that the blogosphere is so upset over.

More on the discussion here, there, and a really good one here.

blog comments powered by Disqus