Wednesday, November 01, 2006

A two week snapshot 

(UPDATED AND BUMPED to account for more numbers and new poll -- see bottom)

The SCSU Survey is out, and it has very large leads for both Mike Hatch and Amy Klobuchar. Since two of the three directors are current and past department chairpersons, I know them pretty well, and in fact knew them before we took up those miserable tasks. So I wrote to them with questions. Both are busy and have only responded to one. Rather than rehash that email, I'm posting what I sent them after they were kind enough to send me the results of the survey (the "you" in this is one of the directors) -- the italicized remarks are not what I sent them:

  1. The survey is taken over two five-day weeks (off Friday and Saturday nights). Most of the polls we see get their samples in shorter periods, most within one five-day window. Is this normal for you, and do you think it makes any difference? I notice your report says you were trying on the last few nights to go back and convert particular groups to increase representation in the survey from "respondents who often are younger, more affluent". How is this known? Does this repeated calling of a particular subgroup cause any concerns? (more on this in question 5)
  2. Related to that, part of the period you sample has the after-effects of Foleygate, yet you didn't ask that question at all. Why? I notice that Jacobs made note of its effect on party ID. Do you think Foley is driving your party ID results?
  3. Do you see any effect of asking your right track/wrong track questions, the Bush JA question, the most important issue questions, etc., before you ask the electoral preference question? Do the earlier questions affect the later ones? And how long does a completed interview take on average? Note: scroll down this for the survey instrument. It is lo-o-o-ong. Its purpose is more than just an electoral preference poll ... which makes it not completely comparable to other polls.
  4. Even with your repeated calling, the response rate (if I understand how to calculate these things) is 17.2%. Do you consider this a normal rate? (Jacobs' September poll was quite similar.) They answered this by noting that the random digit dial method they use gets a lot of non-operating numbers; a phone book gets better response rates but misses unlisted numbers. They argue that their cooperation rate of 77% is well within norms. As I noted yesterday, though, Mark Blumenthal pointed towards 22% as a norm.
  5. Looking at your demographics: Your sample has 28% of households with income over $100k. The American Community Survey of the U.S. Census has only 18% such households. So richer households are going to participate more in this election? Why or why not? Was this what lead to you repeating calls to get compliance from younger, affluent households" as we discussed in my first question? Likewise, according to the same survey in 2005, 71% of Minnesotans 18 and over are employed; your sample has 65%, and a 6% unemployment rate (the state rate is below 4%.) I actually misspoke there. Since the unemployment rate is unemployed/(unemployed + employed), their unemployment rate in the sample is 8.5%. That's much more than the state rate, though I could believe that the unemployed vote disproportionately to their number.
In general, polls are to be a snapshot in time; that's a phrase I've heard Steve Frank use repeatedly in our emails, and it's my understanding of the polling business. What I find most troubling is the two-week window, which appears to be a result of pushing people into cooperating with the poll. In particular, they pushed a group of richer voters into the poll and ended up with a sample that looks heavy on rich people (or rich with heavy people, who knows!) The resulting sample gives them a heavy leaning towards Democratic voters, a party ID pattern that if true is going to make all these predictions of GOP doom a reality on November 8th no matter how much we dice and slice surveys. As noted earlier, it's hard to find other instances of such large swings in party ID.

The line I often use is, "buy the premise, buy the bit" (it's a Johnny Carson quote discussing joke-telling.) Every reasonable forecast, be it political or economic or psychological or whatever, has a premise, a storyline out of which all the results flow. This is part of what I teach forecasting students, to look for that storyline, make it lay down with the data and be sure it fits. And, at the end of the day, they have to ask themselves "Does this make sense to you? Are you willing to tell that story line in a press conference, a board meeting, or a speech?"

In looking at survey data I find much instruction from the marginal analyses as to the premise of the poll. If you think there are currently 12% more Democrat voters in Minnesota than Republican -- regardless of the number you identify as independent -- you don't need a poll any more. Your premise is one word: tsunami. You believe that there has been a large wave that moved Minnesota from battleground state to a very blue state.

The premise for the people who say this is crap is that those waves don't usually happen all at once. They erect stories about gerrymandering and get out the vote and then use them to turn the survey data into something that doesn't predict a sweeping DFL victory. That's not wishful thinking or whistling past the graveyard. That's people who look at history and say this would be unprecedented for so many people to vote against a party not because they are fed up with their incumbents but because they switched parties. And yet, just because something's unprecedented doesn't mean it can't happen.

Each story has the air of plausibility, and to their supporters develop a ring of truth (perhaps due to a version Eric Black's "confirmation bias" theory.) On November 7th, we find out which one has just the veneer, and which one actually comes to pass.

UPDATE: I think Steve is mad at me, because he's decided he's "moving on" and tells me to post what I want. But before that he pointed to the Humphrey poll that shows similar margins for Klobuchar over Kennedy and Hatch over Pawlenty. Different survey instruments, he says, with similar results for party ID and the horseraces. And then he gives me what I think is his premise story:

It appears to me (in an instant thought while writing this email) this is a classic referendum election where a great many citizens are transferring their views toward the President and the Iraq war to state elections. This is compounded by property tax increases, fees, roads that need work and so on. Plus some candidates are good and some not so good.

I pay a lot of attention to our feeling thermometer. Many voters vote for the candidate they �like�. Minnesotans like Klobuchar and many just don�t like Kennedy.

When we interviewed Ken Mehlman in September on the radio show he said something to the effect of "voters will have to focus on the choice" rather than a referendum. The referendum story is the same as the wave story, the same as the tsunami story. And that's the point I'm making: If you believe the Democrats have managed to nationalize the election, the tsunami story makes sense. That is the 1974 story (Watergate), the 1958 story (labor strikes). Which is why from the very beginning the Republicans have blunted the referendum. And they do that in part by message but more by redistricting and turnout. Steve recognizes that's possible, but doesn't give it much credence. The other Steve wouldn't be interviewed in the paper saying "I think the respondents are tired of Republicans." That's their premise, and if you buy it, Kennedy's down 25. If you don't, he's not.