Perhaps this year’s most intriguing gubernatorial contest is in Colorado, which had once looked safe for the Democratic nominee, Mayor John Hickenlooper of Denver. A series of recent polls, however, have shown sharp gains by Tom Tancredo, the former Republican Congressman, who is running on the ticket of the American Constitution Party.
Mr. Tancredo’s campaign had once seemed little more than a curiosity, his support starting out in the low 20s in polls, and having dropped to as low as 9 percent in a one survey in late August. Typically, once a third party candidate’s numbers decline into the single digits or the low teens, they have little hope of coming back.
Mr. Tancredo had the good fortune, however, of being on the ballot alongside an exceptionally marginal Republican nominee in Dan Maes. Mr. Maes had never earned more than 33 percent of the vote in any given survey. About a month ago, the standing of the two candidates crossed in the polls — perhaps as conservative Coloradans realized that Mr. Tancredo’s chances against Mr. Hickenlooper, while not strong, were no worse than Mr. Maes’s.
What all the pollsters agree upon is that Mr. Tancredo has continued to gain ground since then. They disagree, however, on exactly how much of a threat he represents to Mr. Hickenlooper.
Six polling firms have been active in Colorado within the last two weeks. Four of them — Public Policy Polling, Pulse Opinion Research (a subsidiary of Rasmussen Reports), Magellan Strategies, and SurveyUSA — use automated scripts to conduct their surveys rather than live interviewers. These surveys show Mr. Tancredo trailing Mr. Hickenlooper by an average of just 5 percentage points.
But two other polling firms, CNN and Ipsos, conducted traditional polls, and did not find Mr. Tancredo quite so close. They gave Mr. Hickenlooper leads of 14 and 11 percent, respectively.
I can think of two theories to explain the discrepancy, one of which would be favorable for Mr. Hickenlooper and the other for Mr. Tancredo.
Automated surveys are associated with lower response rates (indeed, probably very low response rates.) It can feel like a bit of a chore to (literally) push buttons to answer a “robopoll,” particularly in the absence of a live human being to help coax you along.
The lower the response rate in a survey, the greater the risk that that the people who do complete the poll are not representative of the electorate as a whole — the name given to this is response bias.
Although response bias can cause several problems, one of them may be that you will get only the most enthusiastic participants to complete your survey. While getting enthusiastic respondents is not a wholly negative thing — at least such people are quite likely to vote – such polls can, in effect, overcompensate for the “enthusiasm gap.” There is some evidence, in fact, that that factor is having a systemic effect on the polling, as automated surveys the year have been around 3-4 points more favorable to Republican candidates than traditional polls.
But those effects could especially skew polling for a controversial candidate like Mr. Tancredo, who — while garnering enthusiastic responses from those who like his positions — may have support that is deep but not broad. If, for instance, only 37 percent of the electorate would be willing to vote for Mr. Tancredo, those people might be more likely to tout their support to pollsters, and surveys prone to response bias could then exaggerate his standing.
The argument on the other side has to do with another potential flaw in polling: social desirability bias. When completing a poll, people can sometimes be reluctant to articulate support for a candidate or position that might seem socially undesirable or “politically incorrect.” This may be out of fear of embarrassment, conveying offense to the interviewer, and so forth.
Purveyors of automated surveys have sometimes suggested that their technique can reduce the impact of social desirability bias, since voters may be more willing to give honest responses when talking to a computer rather than another human being, which can create a greater feeling of anonymity. Mr. Tancredo is nothing if not “politically incorrect” in the view of many. Perhaps some voters are planning to vote for him but are unwilling to say as much to (human) pollsters?
In this instance, I tend to discount the latter hypothesis for the following reason. The most natural response of someone who was unwilling to disclose their support for a particular candidate because of social desirability bias would ordinarily say they were undecided — rather than going through the trouble feigning support for one of the other candidates. The so-called Bradley effect, for instance, was thought to manifest itself in voters telling pollsters they were undecided rather than recording their opposition to a black candidate.
If there were social desirability bias at work in the case of Mr. Tancredo’s polling, and the live-interviewer surveys were more prone to it, what we might expect to see is a higher number of undecideds in those polls.
That isn’t what we see, however. Instead, just 2 percent of voters in the CNN poll, and just 5 percent in the Ipsos poll, failed to choose one of the three gubernatorial candidates. (The automated polls also have a very low rate of undecided voters.)
So, this seems more likely to be a case where the automated pollsters are getting it wrong. But the race clearly has many singular qualities to it, and the 17 percent chances the model assigns to Mr. Tancredo seems like a reasonable hedge.
*-*
The balance of the gubernatorial polling was undramatic today, but Republicans got good news in two states.
In Oregon, polls from Rasmussen Reports and the Portland Tribune — the latter has a good track record in the state — show the Republican, Chris Dudley, with a 3-point lead over the Democrat, the former governor John Kitzhaber. This reverses a trend in which Mr. Kitzhaber had held a very small lead in a few consecutive surveys, and the model now has Mr. Dudley’s winning chances up to 45 percent — essentially making the race a tossup — from 31 percent last night.
In Florida, a poll by Susquehanna Polling for Sunshine State News gives the Republican, Rick Scott, a 2-point lead over Democrat Alex Sink. There is a modest trendline here: the same polling firm had shown a tied race in a poll released a few days ago, before the candidates’ debate on Monday. During a break in that debate, Ms. Sink caused a stir when one of her assistants passed her a text message in violation of debate rules. Although surely minor by the standards of political scandals, the race between Ms. Sink and Mr. Scott had been so competitive that even a half-point impact from “debategate” could conceivably determine the winner of the race — and one should be at least a little more willing to “buy” a trend when there is some cogent real-world explanation for it. The model — which just looks at the polling and has no knowledge of the incident –now has Mr. Scott as a 55 percent favorite.
This post has been revised to reflect the following correction:
Correction: October 28, 2010
A previous version of this post misstated the first name of John Hickenlooper.