Thursday, April 17, 2008

Breaking down our Pennsylvania Poll

Thank you to everyone who e-mailed and called yesterday to let us know that our Pennsylvania poll was an outlier. We were kind of aware of that :)

It's an uncomfortable position because certainly if we're wrong we're going to get a heap of abuse for it. That said, we interviewed over 1,000 people each of the last three weeks and these are the results we've gotten. It's not like we're going to change them so they look more like everyone else's.

One reason I have faith in our numbers is that in six out of seven Democratic polls so far we've underestimated Obama's performance. The only place where we said he would do better than he ended up doing was Ohio where we said he would lose by 9 and he ended up losing by 10. So I feel a lot more comfortable as the outlier showing Obama doing better than everyone else than I would if it was the other way around.

Here are some of the common reasons for a poll being off:

-Bad demographic composition. I don't think that will be the case for us if we are wrong because we are largely in line with everyone else. We are showing a stronger black electorate (17%) than SurveyUSA (14%) but that only accounts for a little of the 17 point difference our companies are showing in Pennsylvania right now. We're largely in line with other companies on the percentage of the electorate under 45 (likely Obama supporters) and over 65 (likely Clinton supporters, although her margin with that group was surprisingly low in our last poll.)

-A bad sample. Simply put, sometimes you can get a bad poll because you got a bad sample. I don't think that's the reason for our disparity either though since we've shown it within the margin of error three weeks in a row now. Plus weighting should correct some of that even if you do get a bad sample.

-Bad turnout model. My guess is that if we end up being wrong in Pennsylvania our model of assessing likely voters (people who voted in the 2004 or 2006 primary or the 2006 general) will be the culprit. This worked for us in Texas, Ohio, and Wisconsin but every state is different and it could be that for whatever reason it's not a good way of doing it in Pennsylvania. One difference between PA and those three states is that it has a closed primary whereas those places have open primaries. I don't know why that would throw it off but it's a possibility.

We'll have one more Pennsylvania poll out Monday, but barring some major event in the next few days I imagine we'll continue to show one of the candidates with a lead under five points. Hopefully the results will bear out our polls but if not we'll figure out where we went wrong and try to fix it so we do better in the future.

1 comment:

Sean Braisted said...

Its kind of refreshing to see a polling firm acknowledge criticisms and describe their thinking/methodology process. Seems like others just throw it out there and then blame the voters if their polls turn out to be wrong.

Web Statistics