Over the most recent election cycle I had my students find polls in local newspapers or television via the internet. We found about polls on about 20 races (my students were not as successful in finding polls as I had hoped) which were either for US Congress, US Senate, or Governor. They recorded the percentages, margin of error and sample size for each poll. After the results were in we checked to see how close these polls had come. The problem in such comparision is that there are no 'undecided' votes, so, the undecided portion of each poll must be allocated to each candidate. I decided, out of ignorance, to assign it proportional to each candidate's listed percentage. That is no different than treating undecideds as non-voters on election day. I have later found that generally undecided voters tend to not vote for the incumbent, though in knowing that I still don't know how I would apportion the undecideds.
Anyway, we then compared the actual outcomes to these crudely adjusted polls. We found the outcome was within the margin of error in about 16 of the about 20 races. (I'm not at my office writing this, so I don't have the data in front of me. If y'all are interested, I can get the details and post them.) I came to the conclusion that the polls did a pretty good job, considering the crude nature of my adjustment (I had at that point learned that something more sophisticated was needed). These polls were from many different polling organizations, not just Gallup, or other national pollsters. Paul . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
