Closer Look at Polls,
Surveys and Elections

Outcome of 2010 governor’s race is not a foregone conclusion

by Nat Ehrlich
August 16, 2010

In fewer than three months Michigan will elect its 48th governor. It may be only human nature, but wouldn’t we all like to know ahead of time whether “one tough nerd” Rick Snyder or “angry” Lansing Mayor Virg Bernero will win the seat held by Democrat Jennifer Granholm the past eight years?

To answer that question between now and Tuesday, November 2, thousands of people will be asked who they intend — at that moment — to vote for. They’ll be asked that question in any number of ways, when in fact, the question already has been asked — and answered numerous times.

Take a look:

These polls were conducted before the August primary, but based on the polls of summer, anyone might be hard-pressed to argue that the race will be close. For whatever reason, the polls indicate that Republican Snyder — with an average of about 60 percent more support than Democrat Bernero — is the favorite going into the final months of a busy campaign season.

The fact that Snyder is the favorite is now well-established news. People familiar with the details of how polls are conducted are also well aware that the four polling firms above use different modes of data collection: automated recorded interviews, live interviews, interviews drawn from lists of randomly selected registered voters or randomly selected households with telephones. No matter how the opinions were generated, however, the trend has been established and well reported via television, radio, chat room and the print media.

So, can the election’s outcome be a foregone conclusion? In a word: NO!

It would be a mistake to overlook the fact that awareness of one poll’s results has some effect on what people tell pollsters. Especially in an election where there is no incumbent, campaign history tells us that the public will be influenced by their perception of momentum.

The pre-primary campaigning found two strong candidates on the Democratic side and three on the Republican side. In both races, the early gubernatorial favorites — Democratic House Speaker Andy Dillon and Republican U.S. Rep. Peter Hoekstra — were overtaken.

Even the best of polling can be overtaken by unknown or unpredictable factors. Time is a notorious factor. There’s plenty of time for unknown circumstances before November’s gubernatorial election. The ancient Chinese proverb also gives wise guidance: “prediction difficult, especially as regards future.”

The proverbial addendum yields its own warning: “more future, more difficulty.”
 
If survey research tells us anything, it leaves us with this conclusion: the more time left before the actual election, the less reliable the prediction.

Right about now in an election cycle, people motivated to answer surveys about the governor’s race are those who are more committed politically. Views expressed in early surveys tend to be more polarized. As election day approaches, the pool of respondents comes closer to those voters who will actually show up at the polls, in part because they have heard and seen more of the candidates.

News once crossed the country on horseback. Now it flashes by in less than a Facebook second, with its own set of influences on polling and polls’ ability to predict the future. For example, as we began the 21st Century, blogs numbered in the low hundreds. In 2006 (only four years ago in real time), we were told that:

• Technorati tracked its 50 millionth blog
• The blogosphere was doubling about once every six and a half months (if this is true, we now have about 600 BILLION blogs)
• About 175,000 new blogs were created each day
• More than two blogs were created each second of each day, with about 1.6 million postings per day — the equivalent of 18.6 posts per second.

Or look at Facebook itself, which began in 2004 and currently has more 500 million users worldwide. As this piece is being written:

• 27,234 are fans of Rick Snyder
• Bernero has 1,421 fans.

Why does this matter? Ten years ago, in the Bush/Gore contest, the Friday-before-election revelation of Bush’s DUI conviction was only briefly touched on by the Sunday political talk shows. In today’s atmosphere, the news would have “gone viral” and very likely have changed the ultimate result.

Top of mind
Making sense of political polls can be confusing. We are all naturally suspicious when we read results — based on telephone interviews of 500 people — that predict how several million of us will vote. But as the next few weeks unfold, there are a few concepts to keep top of mind when poll reading.

Survey researchers account for error in their analysis. The more people questioned, the lower the likelihood of error. A sampling of the opinions of 500 people yields error of 4 percent on either side of the result. Adding another 500 people to the survey narrows the error to plus or minus 3 percent. With 2,500 people participating in the survey, the error goes down to plus or minus 2 percent.

These figures are estimates, not guarantees. From research history, we expect those limits to hold up 95 percent of the time, and be exceeded 5 percent of the time.

The calculation of error also assumes that every respondent is selected at random, that each one has an equal chance of being contacted to take the survey: Republicans and Democrats and Independents; men and women; people of all ages and races. But in practice, we don’t see that happening. Research tells us that women respond more than men. Republicans respond to survey questions more than Democrats do. Older people are more likely to take part in surveys than younger people. These deviations from a true random sample increase the size of the potential error.

Response rate can be very important but is rarely reported. In general, the higher the response rate, the better.

Here’s a sample of what a well-known and reputable survey firm reported: “[We] surveyed 606 American voters from August 6th to 9th. The survey’s margin of error is +/-4.0 percent. Other factors, such as refusal to be interviewed and weighting, may introduce additional error that is more difficult to quantify.” In other words, the margin for error will be somewhat higher than +/-4.0 percent, but we have no idea by how much.

Survey research is labor intensive and relies on people, first, to answer their phone and, second, to be willing to give honest answers to the hard-working interviewers who know that rejection may be the answer they hear most. Careful survey research, conducted over a period of several weeks, can hope to complete interviews with 50 percent or more of the people who answer a phone call from a list of randomly generated telephone numbers.

A political poll conducted over one to three days will be lucky to exceed a 5-percent response rate. It’s only common sense to expect those who respond to surveys to:

1. Change their minds between the time they’re interviewed and an election
2. Decide not to vote at all or
3. Some may even (gasp!) lie.

As an example, in 2006 Michigan voters rejected affirmative action by a 58:42 ratio, even though both candidates for governor came out in favor of keeping the status quo. Six months later, a survey found the same ratio, but reversed: 42 percent of the voters said they had voted to reject affirmative action and 58 said they voted to keep it.

This kind of after-the-fact reporting of personal experience is not uncommon. After all, “what’s done in the privacy of the voting booth…”

The method of data collection is another variable that affects the validity of poll results. Most polls will have been conducted by telephone, with human beings dialing from a list of randomly generated phone numbers. Some interviews are conducted by computer, with no live human interaction between the respondent and a live interviewer. Good interviewers are trained to engage those they call and to personalize the interaction. Robo-type automated calls can be off-putting, though there’s little research to show whether results are affected by the personal touch.

Other interviews will be conducted by Internet, completed by a click of the keyboard or movement of a mouse. Some will be the result of e-mail invitations sent to a select list of likely voters at one end of the continuum. Others will be drop-in surveys encapsulated in articles on a political website.

In our experience, the most accurate method is the live human dialing of randomly generated phone numbers and speaking to a person who identifies him/herself as a registered voter. The least accurate are the drop-in web surveys. With drop-in web surveys, there’s no definite way to tell who is actually answering the questions.

Looking ahead
I would be remiss if I failed to mention the widespread feeling that Michigan voters are poised to let a Republican have a try at running the state. Back to 1969, Michigan has alternated Republican/Democrat/Republican/Democrat in electing governors.

Right now, all signs point to a Republican victory in November — but the election is not right now. As this column is written, there are weeks of future to stymie the best of predictions. Such as:

• Events that overtake the candidate and carry voter preference to one candidate or another.
• Neither candidate is well known at this point. Each carries his party’s label, but not the ringing endorsement of the party.
• Bernero is the underdog. His status could work in his favor if he can be seen as building momentum.

Speaking style and appearance have been shown to be powerful factors, especially in races between relative unknowns. The contest is not between parties or ideologies, but between people. In most elections, the candidate who appeals on a personal level to more people is the more likely to win.

The proverb and its follow-up are givens in polling. But the fact remains, no matter what we read between now and November, there is only one poll that counts. It’s called the election.

Nat Ehrlich, Ph.D., is a survey research specialist at the Institute for Public Policy and Social Research at Michigan State University. He has authored more than a dozen papers in survey research methodology and statistics, and has conducted a wide variety of survey research. He is a specialist in statistics and previously taught psychology at the University of Michigan and City College, City University of New York.