Well, now the Star Tribune’s Sunday poll isn’t a total outlier — a new MPR/Humphrey Institute gubernatorial survey shows Mark Dayton’s lead over Tom Emmer rising to 11 points. August’s poll showed a dead heat. And this is without the potentially DFL-leaning cell-phone-only voters.
The topline: Dayton 38, Emmer 27, Horner 16. Twenty percent are undecided.
Horner’s number doesn’t quite hit the 18 percent that the Strib and Rasmussen recently tallied, but it’s up from 13 percent a month ago. Emmer’s total is the lowest among recent polls because Horner pulls more than twice as many Republicans (22 percent) as DFLers (10 percent).
One of the wacky things about the 750-voter, landline-only, five-day survey is how much partisan ID shifted in just a month. August’s poll was 46 percent GOP, 41 percent Democrat — the only major survey with a Republican plurality. This one is 48 percent Dem, 38 percent GOP.
Humphrey Institute Prof. Larry Jacobs ascribes this to renewed DFL enthusiasm. However, the dramatic shift will inflame the doubters, particularly GOP partisans, whose rightest wing generally regards the HHH poll as the junior member of a Strib-led pro-DFL Gruesome Twosome.
This demands a little more discussion of how the poll determines likely voters, a topic I broached the other day.
U prof Joanne Miller described a four-question index, based on registration, self-reported voting likelihood, voting in the ’06 guv’s race and 2010 interest. Voters are rated on a 0-8 scale.
Miller subsequently explained that unlike some surveys — which only include voters over, say, an 8 on 10-point scale — the HHH poll actually includes everyone, but weights responses differently:
We’ve gone back to those polls and have calculated each respondent’s score on the index. Then we’ve gone to the voter rolls to see whether each person who was surveyed actually voted. Then we calculate the percentage of people who scored 0 on the index who actually voted, the percentage of people who scored 1 on the index who actually voted, and so on. We use those percentages to create likely voter weights for the current poll.
So, this methodology doesn’t “delete” anyone from the analysis. For example, our analysis shows that even a tiny percentage of the people who score 0 on the index do, in fact, vote. So we give them a tiny, non-zero weight.
My first question: how do you know who your respondents are? Usually, voters are anonymous beyond their phone numbers.
Miller says at the end of the survey, respondents are asked if they mind being contacted by a reporter. If they say yes, the Humphrey Institute has a name it can check against voter rolls. A reminder: those rolls only show if a person voted, not how they vote.
Miller says the approach more realistically models voter turnout; other surveys, including the Strib’s, show voter participation higher than historic off-year norms.
It’s an interesting method, but one that has not earned the HHH poll a high ranking on FiveThirtyEight.com’s pollster survey. A survey of recent results shows the final pre-election HHH poll, like the Strib’s, consistently over-estimates the DFL vote, at least compared to Election Day.
However, in two of the three most recent statewide elections (’06 and ’08 Senate), HHH was closer than SurveyUSA or Rasmussen, which FiveThirtyEight ranks high. Then again, HHH’s final ’06 gubernatorial poll was off by more (7 percentage points) than any competitor. Major caveat on this last one: the HHH poll closed earlier than competitors, so could have legitimately missed voters breaking to Pawlenty late.
The current survey has a margin of sampling error of 5.1 percentage points, plus or minus. Dayton’s lead is outside that. The Hump Institute is more honest about error margin than some, including design effects. The MPR story notes the “conventional” error margin would be 3.6 percent.