Poll Discrepancies

National polls of the presidential race have been unusually wide-ranging over recent weeks. Here is a summary of some of their results:

1. Economist, 9/15: Bush +1

2. Harris, 9/13: Kerry +1

3. ICR, 9/12: Bush +4

4. Pew, 9/14: Kerry +0

5. Rasmussen, 9/16: Bush +4

6. Zogby, 9/9: Bush +2

7. ABC, 9/8: Bush +6

8. CBS, 9/16: Bush +8

9. Fox, 9/8: Bush +2

10. Gallup, 9/14: Bush +8

11. IDB, 9/12: Kerry +2

12. News, 9/10: Bush +6

All of these polls were taken over a roughly one-week period, yet they show an range in results that is far outside of the typical poll’s margin of error (MoE). Why?

Chris Bowers at MyDD provides a thorough explanation for why polls #1-6 listed above are fundamentally different from polls #7-12, and why the first six are more likely to be accurate. The explanation is this: the first six polls weight their results by party ID, whereas the last six polls do not.

For example, if 50% of the poll respondents say that they’re Republican, then it’s likely that the aggregate figures will show a big lead for Bush. But there might be some reason to believe that you haven’t surveyed a representative sample, if you have other information that leads you to believe that the percent of Republicans in the population is really only 30%. Polls #1-6 above adjust their results to reflect this prior knowledge of how many Repubs, Dems, and Independents there actually are in the US. Polls #7-12 do not.

For some reason, several polls have recently had unusually high numbers of respondents self-identify as Republicans — around 40% of respondents to the Gallup poll were self-identified as Republican, even though nationally Republicans are generally only about 33-35% of the population. Polls that adjust for this discrepancy show the race nearly tied, while polls that don’t how a large lead for Bush. This suggests that it may make sense to weight the raw results by party ID. Earlier this week, Ruy Teixeira at the Emerging Democratic Majority wrote an excellent summary of that argument.

I am left with a lingering, but possibly crucial question, however. What accounts for the recent spike in people self-identifying as Republicans? I’d like to believe that the race is tied right now (which is what the weighted polls suggest), but if the answer is that some people change their party ID answer depending on which candidate they’re leaning toward, then it may not make sense to weight the raw data by party ID at all. It seems to me that answering this crucial question is fundamental to interpreting these widely varying poll results.