Mark Blumenthal | August 1, 2007
The release of new surveys yesterday by the American Research Group (ARG) in Iowa, New Hampshire and South Carolina has generated quite a bit of buzz as well as questions about ARG of the sort I alluded to yesterday in my post on the screens pollsters use to select likely primary voters and caucus attendees. That post is the first in a series that will look at the methodological differences among the various pollsters active in the early primary states. For now, however, let me take up two issues specific to Iowa and ARG's surveys there.
The first question comes from reader AL:
I think your averaging is wrong on the Iowa race for the democrats. I think it should be Edwards at 25.7 and Hillary at 25.4. You have the averaging mixed up. You should correct it.
I am not sure how AL arrived at those numbers, but the current Pollster.com estimates in Iowa (Clinton 25.7%, Edwards 25.4%) are not "mixed up." The confusion may arise from the fact that our estimates are regression based estimators rather than true averages. Professor Franklin explains the difference in detail here.
The second issue concerns the Iowa surveys by ARG. As commenter jsamuel put it, "ARG seems to almost always under poll Edwards." While I lack Professor Franklin's flair for graphics and regression trend lines, some simple averages show that jsamuel is right. Sen. Clinton does consistently better ARG's surveys than those from other pollsters, while former Sen. Edwards does consistently worse:
We logged in six new Iowa surveys during June and July. The two from ARG show Clinton ahead of Edwards by an average six points (31% to 25%), while the four from other pollsters give Edwards an average lead of five points (27% to 22%).
Surveys from February, March, April and May show essentially the same pattern. Clinton leads by two points (30% to 28%) in four surveys conducted by ARG, while Edwards leads by an average six points (27% to 21%) in 13 surveys by other pollsters.
Unfortunately, this is the sort of scenario in which averaging (or simple regression based estimation) can be potentially misleading. One pollster (ARG) is getting consistently different results and contributing a large number of polls to our overall estimate. So if ARG is both different and wrong, their polls are throwing off our estimates.
We know ARG's results in Iowa are different. Why? And what should we make of that difference? For some clues, stay tuned to my series on primary screens.