Morning Jay: Obama's Falling Numbers, How To Read The Polls, The Jon Corzine Effect, and More!
6:30 AM, Sep 27, 2010 • By JAY COST
The most important way to diagnose the quality of a poll is to look at party identification spreads. There is no better predictor of vote choice than party identification. That’s been true for as long as political polling has been done, and it is now a better predictor than ever. The two sides now break for their own candidates on about a 90/10 basis these days, and comparing a poll to recent exit polling can give us a good sense of whether or not it is oversampling a particular side. While we can’t know for sure exactly what the partisan spread will be this year, we can be pretty confident that the electorate will be less Democratic than it was in 2008, and the exit polls from that cycle can all be found here. Instead, it will be closer to the 2004 electorate.
So, that offers an easy way to evaluate the accuracy of a poll. For instance, via Allahpundit, consider the Kentucky Senate poll by SurveyUSA, which gives the Democrats a fifteen-point edge in party identification, compared to the nine-point edge they actually enjoyed in 2008 or the four-point Democratic edge in 2004. Use those 2004 partisan identification numbers against the SurveyUSA results, and you find Rand Paul with a nine-point lead, rather than the advertised two-point lead. Now, SurveyUSA is a very fine pollster, which I have praised on this page. Importantly, the laws of statistics say very clearly you do enough polls, sooner or later you're going to produce results that are outliers, as this poll appears to be. Ultimately, it's up to each of us to be smart consumers of polling.
Looking at party ID can give a good sense of how accurate the poll is. Partisan identification on Election Day is fairly stable and predictable. In the last few cycles, good Republican years tend to result in roughly equal party strength, e.g. 2004. Good Democratic years tend to be D+5 or more nationwide. All accounts suggest that this will be a good Republican year, which means polls whose statewide party identification spreads are closer to 2008 than 2004 are probably over-sampling Democrats.
My advice: bookmark those links to the 2004 and 2008 exit polls, and use the data you find there to evaluate the polling that comes out from this point forward.
3. The Corzine Effect. Lately, I’ve been thinking about Jon Corzine. The former governor of New Jersey pops into my head every time I look at polls of the Wisconsin and Nevada Senate races. Both Russ Feingold and Harry Reid seem to be stuck in the same basic position that Corzine was in.
To appreciate what I mean, consider the following chart, which tracks the average monthly polling position of Corzine, Feingold, and Reid through the summer. Corzine’s numbers are obviously for 2009 while Feingold and Reid’s numbers are from this year.
I’m really struck by the fact that the numbers of all three incumbents were basically unchanged through the whole summers in question, with all three of them consistently under 50 percent.