Tools
Just 24 hours before Seattle's August 18 primary, the national polling group SurveyUSA announced that incumbent mayor Greg Nickels was in the lead with 26 percent of the vote, while his challengers, Joe Mallahan and Mike McGinn, were both at least four points behind him. A few days previous, SurveyUSA had also announced its findings in the race for King County executive: Democrat Dow Constantine was getting 20 percent of the vote while Republican Susan Hutchison was getting 37 percent.
Both of those polls were way off.
Stranger Personals
McGinn, who SurveyUSA had in last place going into the primary, ended up coming in first with 27.7 percent of the vote—almost seven points more than SurveyUSA had predicted. To its credit, SurveyUSA had called that race "a jump-ball." But in the executive's race, where no such disclaimer was made, SurveyUSA seriously underestimated Constantine's strength (he finished with 27 percent of the vote, seven points more than predicted) while overestimating Hutchison's (she finished with 33 percent, four points less than predicted). "We should do better than that," admitted Jay Leve, spokesman for SurveyUSA.
In bigger elections, and in bigger cities, this wouldn't be a problem. Other polling companies would be in the mix, and hopefully at least one would produce accurate results. But in this city, in this election, SurveyUSA has been pretty much the only game in town—in part as a result of the down economy. Public polls are expensive, and KING 5 Television, which pays for SurveyUSA's local polling, is essentially the only news outlet in Seattle with money to spare.
When the polling is as wrong as it has been so far, old complaints about the company's conservative tilt surface—usually based on studies conducted after the 2006 and 2008 federal elections that found SurveyUSA's polling was biased toward Republicans. Leve pointed out, however, that a 2004 study showed SurveyUSA's polling skewed too liberal that year, and added: "I have zero stake in steering it one way or another."
Another common complaint: SurveyUSA can't reach cell-phone users. "That's an absolutely legitimate concern," Leve admitted.
Many younger voters use only cell phones, and an inability to survey these voters—who also tend to be more liberal and excited about grassroots campaigns—may account in part for SurveyUSA's trouble predicting actual support for Constantine and McGinn.
Fortunately, as the general election nears, others are stepping into the fray. A poll put out by the University of Washington on October 27 found Constantine beating Hutchison by 13 points—a very different finding than that of SurveyUSA's most recent poll, which had Constantine behind by five points.
Not that new polling pleases everyone. The UW's poll also found McGinn trailing by 10 points, and McGinn's campaign immediately attacked one of the pollsters, saying, "His polls don't seem to reflect reality" and mocking misspellings in his online presentation of the full poll details.
The pollster, Matt Barreto, called the attacks "hilarious." 
Share via
1
So, for a realistic sample group, backcheck simple stats figures for oversampling of rich white men with expensive houses who are old - they do vote a lot, but they also answer the landline a lot and do surveys, but women screen and most young people only have cell phones and won't answer polls or calls from people they don't know.
Easy way to backcheck is the standard econ bracket data (income for family is (5 ranges)) with age range and gender.
To get a good sample, you'll still need about 512 people for a political poll (other data has controls, so they need twice that, but people are basically considering Voters).
Another backcheck is the LD distribution - if you find respondents cluster in rich LDs, you've oversampled due to non-answering of other people.
Please not telephone to voter record has a high mismatch rate for cell numbers, and a high error rate of wrong numbers - for all you know someone in 253 is in Bellingham - this is ok for statewide since it maps to pop, but bad for county or city measures which have restricted geographies.
If you're answering questions for a Seattle vote, 100 people is still too small, even with a reasonable match on age/sex/econ/LD distributions.
2
So, for a realistic sample group, backcheck simple stats figures for oversampling of rich white men with expensive houses who are old - they do vote a lot, but they also answer the landline a lot and do surveys, but women screen and most young people only have cell phones and won't answer polls or calls from people they don't know.
Easy way to backcheck is the standard econ bracket data (income for family is (5 ranges)) with age range and gender.
To get a good sample, you'll still need about 512 people for a political poll (other data has controls, so they need twice that, but people are basically considering Voters).
Another backcheck is the LD distribution - if you find respondents cluster in rich LDs, you've oversampled due to non-answering of other people.
Please note telephone to voter record has a high mismatch rate for cell numbers, and a high error rate of wrong numbers - for all you know someone in 253 is in Bellingham - this is ok for statewide since it maps to pop, but bad for county or city measures which have restricted geographies.
If you're answering questions for a Seattle vote, 100 people is still too small, even with a reasonable match on age/sex/econ/LD distributions.
6
But you then turn around and laud the Washington Poll for "fortunately stepping into the fray" and yet never disclose that their primary polling
but had McGinn coming in FIFTH place ahead of Kwame (Wyking) Garrett. They said McGinn would get 9% and instead he got 28%...a 19 point gap.
So if 7 points is "way off", what is 19 points? And why wasn't it relevant to your story?
RSS
Facebook
Twitter
Comments (9) RSS