Peter Levine

« keeping on top of those young voters | Main | the narrowing of the curriculum »

February 7, 2008

exit poll predictions

Early on Super Tuesday evening, I guiltily visited the Drudge Report to see the leaked exit poll results, which showed Senator Obama with a big lead in California. Of course, that lead evaporated. Why were the exit polls wrong? In short, because they are not designed to predict the outcomes of elections.

To give a more detailed answer: Commercial and academic pollsters never really try to draw a random sample of Americans from which they expect to be able to derive results directly. Instead, they always construct a sample using a source other than the poll itself to guide them. Most often, they try to fill quotas of interviewees from various categories (e.g., race, region, and gender). Even after they have done their best to fill their quotas using quasi-random methods, they almost always "weight" their data to reflect the demographic breakdown of the population, according to the US Census. For instance, if only seven percent of the sample is age 18-25, when the Census finds that 14 percent of adults are in that age range, then each interviewee of age 18-25 counts for two. (Doubling the weight of some respondents is not uncommon.)

In the case of exit polls, the results are not "weighted" to match Census demographics, because exit polls are intended to describe voters, not residents. Instead, results are weighted to reflect the actual election results, as reported by officials. Thus, if 51 percent of voters cast ballots for Senator Clinton, but 55 percent of people who were interviewed in the exit poll said that they preferred Senator Obama, then each Obama supporter in the poll counts for more than one in the results.

I'm sure that the early results leaked to Drudge were unweighted, because the ballots had not yet been counted. Once they were weighted, they showed a Clinton victory in California, because that's what election officials reported. This is also why exit polls showed Bush behind in Ohio in 2004, yet he won the election.

This is a legitimate method, because any poll is designed to address only some questions, not all questions. An exit poll cannot predict who will win, but it can estimate (with various sources of error) what kinds of people voted for each candidate. That's what it's for.

Some of us are overly credulous about surveys. Others, having realized that surveys are imperfect, discount them all. I recommend carefully considering what questions any given poll can answer, and using it only for those purposes. (Remind me to stay off Drudge on November fourth.)

February 7, 2008 6:38 PM | category: none

Comments

February 7, 2008 7:57 PM | Comments (1) | posted by Jimmy L