Thursday, August 2, 2007

Random thoughts on political polling

The Gallup poll method is reputedly tried and true, but I am wondering whether it is really the best approach. Seems needlessly complicated to me.

However, the book I looked at is a bit outdated, so perhaps Gallup's method has changed. According to the book, Gallup's presidential campaign poll is based on a questionnaire designed to screen out people who aren't likely to vote. But the extensive questionnaire could itself introduce bias -- i.e., what type of voter is likely to have time or patience for such a session?

Here's how I would design a political poll. Gather data from several previous national elections to see the percentages of people who identify themselves as Democrat, Republican and independent who vote versus those who don't.

Take a random sample of Americans (only a few thousand are needed for a population of 300 million) and ask them how they plan to vote, and perhaps one or two other questions. Then weight the sample outcomes according to the percentages who in recent history actually vote.

This method should be highly accurate (95% confidence level) because it preserves the randomness while taking account of those who express an opinion but who end up not voting.

One can use this method nationally, for the popular vote, and state by state, for the electoral college result. What about big cities, with lots of Democrats, versus suburbs and rural areas, with lots of Republicans? As long as the sample for a state is random, these differences should tend to cancel out.

New exit poll strategy
Really, I doubt the claims that the 2004 exit polls were way off base because youthful poll takers across the nation nearly all ignored the randomization requirement and buttonholed "their type" who were more likely to be Kerry supporters.

However, a way to counter such a problem would be to change the randomization procedure, perhaps interviewing fewer people. The method used in 2004 required the interviewers to approach every 10th voter exiting a polling station. Yet, a proper random sample usually doesn't need 10 percent of the specific population. Neither is it necessary to check every precinct, though this makes the average person feel that the method is better.

Just as in pre-election polling, the key is a sufficiently large (but not huge) random sample from around the state. That is, X number of precincts are chosen at random. Interviewers are sent only to those precincts. There, Y number of exiting voters are queried. Quantities X and Y are easily determined using statistical methods.

But how do we ensure that interviewers don't bias the results?

Randomly sample the interviewers and have a monitor check for interviewer selection bias. Let the monitor assign a weight to the bias. So then a specific percentage of interviewers can be expected to have an average bias (standard deviation) in favor of one or the other candidate (assuming a two-way race).

Additionally, a panel of independent statisticians could review the exit poll results, with the option of a minority dissent.

No comments: