Thursday, May 22, 2008

Revisiting Old Polls

We have an outstanding new intern, John Willingham, who started this week. He's a graduate student in Political Science at NC State, which this Tar Heel will have to forgive him for.

I've been teaching John how to turn raw data into poll results, and it's actually been a great opportunity to take some of our older polls, weight them to the exit polls, and see what we find.

Here are the highlights:

-PPP infamously showed Bev Perdue with a 27 point lead over Richard Moore in early March, then had it down to 10 just two weeks later. As I've said before, I didn't feel good about that poll. We weighted the data from that survey again this week, using the final demographics from the exit poll of the primary two weeks ago, and found Perdue with a 22 point lead at that point. I think that's pretty believable given that Perdue was on the air at the time and Moore was not. So if you want a baseline of what rock bottom was for Moore before coming back over the month of March, it was probably a 22 point deficit.

-PPP came the closest to getting the margin of Barack Obama's victory (28 points) correct in South Carolina, but we were still way off given that we showed his lead at only 20 points. We went back and reweighted our final SC poll today with the exit poll demographics and got Obama 50, Clinton 23, Edwards 15. That's a lot closer to the final result of 55-27-18 than our final poll was (44-24-19).

What does that mean? It means our raw data was pretty darn good, but we didn't do enough weighting. If we had weighted our sample to more correctly match the age distribution of the SC electorate (we had it too old) then we would have been right on the mark.

Just goes to show you how much nuance goes into what makes for an on point poll and one that misses the mark. We've been able to learn from our mistakes over the primary calendar and make some improvements.

1 comment:

jr1886 said...

I am extremely interested in polling, do I need to graduate from college to get an internship with your polling organization?

Btw, Good job in Oregon

Web Statistics