Writing in defense of the Humphrey Institute Poll – which indicated our tie governor’s race was headed for a 12 point blowout – Professor Larry Jacobs says:
Careful review of polls in the field conducting interviews during the same period indicates that the MPR/HHH estimate for Emmer (see Figure 2) was within the margin of sampling error of 3 of the 4 other polls but that it was also on the “lower bound” after accounting for random sampling error. (Its estimate was nearly identical to that of MinnPost’s poll with St. Cloud.)
Which showed the race a ten point blowout for Dayton.
Jacobs is, in effect, saying “yeah, our poll was a hash – but so was everyone else’s”.
This pattern is not “wrong” given the need to account for the margin of sampling error, but it is also not desirable. As part of our usual practice, the post-election review investigated whether there were systematic issues that could be addressed.
Research suggests three potential explanations for the MPR/HHH estimate for Emmer; none revealed problems after investigation.
Here are the three areas the Humphrey Institute investigated:
Weighting: First, it made sense to examine closely the weighting of the survey data in general and the weighting used to identify voters most likely to vote. Weighting is a standard practice to align the poll data with known demographic and other features such as gender and age that are documented by the U.S. Census. (Political party affiliation is a product of election campaigns and other political events and is not widely accepted by survey professionals as a reliable basis for weighting responses.)
Our own review of the data did not reveal errors that, for instance, might inflate the proportion of Democrats or depress that of Republicans who are identified as likely to vote. To make sure our review did not miss something, we solicited the independent advice of well-regarded statistician, Professor Andrew Gelman at Columbia University in New York City, who we had not previously worked with or personally met. Professor Gelman concluded that the weighting was “in line with standard practice” and confirmed our own evaluation.
“And an expert said everything’s hunky dory!”
Our second investigation was of what are known as “interviewer effects” based on research indicating that the race of the interviewer may impact respondents.11 (Forty-four percent of the interviewers for the MPR/HHH poll were minorities, mostly African American.) In particular, we searched for differences in expressed support for particular candidates based on whether the interviewer was Caucasian or minority. This investigation failed to detect statistically significant differences.
And the third was the much higher participation in the poll from respondents in the “612” area code – Minneapolis and its very near ‘burbs. Jacobs (with emphasis added by me):
When analyzing a poll to meet a media schedule, it is not always feasible to look in-depth at internals.
It’s apparently more important to make the 5PM news than to have useful, valid numbers.
With the time and ability that this review made possible, we discovered in retrospect that individuals called in the 612 area code were more prone to participate than statewide — 81% in the 612 area as compared to 67% statewide in the October poll.13 Given that Democratic candidates traditionally fare well among voters in the 612 area code, the higher cooperation rate among likely voters in the 612 area code may explain why the estimate of Emmer’s support by MPR/HHH was slightly lower than those by other polls conducted at around the same time. This is the kind of lesson that can be closely monitored in the future and addressed to improve polling quality.
Except we bloggers have been “closely monitoring” this for years. It’s been pointed out in the past; on this very blog, I have been writing about this phenomenon since 2004 at the very latest. Liberals looooove to answer polls. Conservatives seem not to.
That Jacobs claims to be just discovering this now, after all these years, is…surprising?
Frank Newport at Gallup critiques Jacobs’ report:
The authors give the cooperation rate for 612 residents compared to the cooperation rate statewide. The assumption appears to be that this led to a disproportionately high concentration of voters in the sample from the 612 area code. A more relevant comparison would be the cooperation rate for 612 residents compared to all those contacted statewide in all area codes other than 612. Still more relevant would be a discussion of the actual proportion of all completed interviews in the final weighted sample that were conducted in the 612 area code (and other area codes) compared to the Census estimate of the proportion of the population of Minnesota living in the 612 area code, or the proportion of votes cast in a typical statewide election from the 612 area code, or the proportion of the initial sample in the 612 area code. These are typical calculations. The authors note that residents in the 612 area code can be expected, on average, to skew disproportionately for the Democratic candidate in a statewide race. An overrepresentation in the sample of voters in the 612 area code could thus be expected to move the overall sample estimates in a Democratic direction.
That Jacobs finds an excuse for failing to weight for higher participation in a city that is right up there with Portland and Berkeley as a liberal hotbed would be astounding, if it weren’t the Humphrey Institute we’re talking about.
The authors do not discuss the ways in which region was controlled in the survey process, if any. The authors make it clear that they did not weight the sample by region. This is commonly done in state polls, particularly in states where voting outcomes can vary significantly by region, as apparently is the case in Minnesota.
Summary: The HHH poll is sloppy work.