Column: We are changing faster than polls keep up

• These are anxious days for presidential election pollsters.
• Polling was easier when voters were more likely to hold still.
• Low response rates are a major issue with most modern polls.

These are anxious days for presidential election pollsters.

It's becoming harder to get it right. The speed of modern communication makes campaigns more volatile. The great number of polls improves the chance that somebody will call the Obama-Romney outcome right on the money -- and somebody else will get it terribly wrong. But who?

Polling was easier when voters were more likely to hold still. You could base a sample on housing units or telephone numbers and give every voting-age person a known chance of being included.

But increasing mobility and the great variety of communication methods, especially cellphones, make it far more complicated. And yet the leading polls do about as well as they ever did. In 2008, the Gallup team missed Barack Obama's winning margin by only 2 percentage points. It called the 2000 and 2004 elections with even less error.

Today, there are signs of worry. Some midcourse tweaks were announced earlier this month by Gallup editor Frank Newport. The sample in the daily tracking poll now includes more cellphones. And a weight was added for the population density of the area where the respondent lives.

Looking for accuracy

As a jaded poll-watcher, I pay most attention in the current campaign to just two polling-related sources: Nate Silver's blog for The New York Times and the RAND Continuous Presidential Election Poll. Both have been showing a fragile advantage for Obama.

Silver, following the principles described in his book The Signal and the Noise, publishes daily reports of candidate standings. He expresses them as a probability of winning and estimates the electoral vote. Silver does no polls of his own but cranks others' results, including Gallup's, into his own statistical model. This model takes into account past reliability, sample sizes, historic trends, "house effects" and other factors.

House effects are the tendencies of certain pollsters to lean one way or the other that everybody knows about but no one can explain.

The RAND Corporation is a California-based non-profit set up after World War II to do policy-related research. It uses a large sample, about 3,500, and makes it a panel, interviewing the same people week after week.

This sample is drawn with more care than most and enjoys a high response rate. Interviewees answer online and are paid for their trouble. Those without Internet access when they were chosen were given the necessary equipment, an innovation tested in Holland in the 1990s.

More people refuse polls

Low response rates are a major issue with most modern polls, especially those that use automated interviews. Some voters don't like talking to machinery. Of course, those who hang up are replaced with more willing citizens, but there is no way to tell whether they are the same.

The RAND panel allows direct tracking of changes in candidate standings, but it comes at a cost. Participation in the panel arouses voter interest so panel members tend to pay more attention, possibly becoming less representative of the voting population.

One way to check on that is to keep pumping new people into the system so that experienced respondents can be compared with fresh ones. But the RAND people are not doing that for their election panel, according to Arie Kapteyn, who directs the project.

RAND's other interesting innovation is in its questions. Instead of being offered a direct choice, panel members are asked to estimate their probability of voting for either candidate plus the probability that they will vote at all.

George Gallup's organization has the longest history of presidential election polls, and it has been embarrassed by only one presidential call. In 1948, its final pre-election poll had New York Gov. Thomas Dewey beating President Harry Truman. It understated Truman's vote total by 5 percentage points.

There were two other presidential elections where Gallup was off by 5 or more points, but nobody minded because the right winner was predicted. The most recent case was 1992, when Gallup overstated Bill Clinton's lead over George H.W. Bush by 5.7 points. Clinton still won handily.

In George Gallup's first national effort, in 1936, the poll overestimated Franklin D. Roosevelt's winning margin by 6.8 percentage points. That poll nevertheless made Gallup famous because the most prestigious poll of the time, by the Literary Digest, had called the election for Kansas Gov. Alfred M. Landon.

The Literary Digest, a weekly magazine with circulation of 1million, never recovered. It shut down in 1938.

And Gallup prospered. In this election, such a career-changing fate could strike somebody again.

Philip Meyer, author of Paper Route: Finding My Way to Precision Journalism, is a member of USA TODAY's Board of Contributors.