Polls have been a mainstay of American politics since at least the 1930s, when George Gallup starting running surveys on elections. But with state-of-the-art technology, improved polling techniques, and ever greater scrutiny of political news, the endless march of polls has become almost deafening. How do you know which polls are most reliable? How do pollsters choose whom to call? And what the heck is a likely voter? Michael Dimock, associate director of research at the Pew Research Center for the People and the Press, answered some common questions about how to wade through the data.
Let’s start with a question on many people’s minds: what’s the difference between registered-voter polls and likely-voter polls? Why would anyone conduct a registered-voter poll, and why should we pay attention to them?
Most of the polls you see reported in the news are based on surveys of all adults, who, in the course of the survey, are asked a series of questions about whether they are registered and whether they plan to vote, as well as about their past voting habits and their interest in the election. Results from the polls could then be reported on all adults (which you rarely see, since fewer than half of eligible adults will actually vote), or limited to those who say they are registered, or limited to an even smaller subset who are registered and report levels of enthusiasm and engagement that suggest they are the most likely to turn out. Many pollsters report their election questions on all registered voters up through around September, and then begin to narrow their analysis to the subset of registered voters that appear to be most likely to vote. This is because any earlier than September, it is difficult to gauge who will and won’t cast a ballot with any degree of accuracy.
One often hears that likely-voter polls skew right. Is that always the case, or is it just a factor of an election cycle that seems to favor Republicans?
The reality is that voter turnout is skewed demographically toward older, higher-income, college-educated adults—and for the past few decades at least, people with these characteristics tend to be slightly more Republican in their orientation. In short, it’s not that Republicans are inherently more likely to vote based on their ideology, but that Republican identification is highest among the demographic groups that are most likely to vote. Therefore, this is not a “skew,” but a correction for differential turnout across demographic groups. If voters under age 30 would turn out at the same rate as voters age 65 and older, and voters with low incomes would turn out at the same rate as voters with high incomes, this might change.
How does Pew decide who qualifies as a likely voter? How did you arrive at that set of criteria, as opposed to other pollsters who use different criteria? Mightn’t people lie because they’re embarrassed about not voting?
We define likely voters as those registered voters who are the most committed to casting a ballot, who have given the most thought to the election, and who exhibit a record of voting in the past that suggests they will do so again in the future. These metrics have proven to be indicators of voting based on analyses we have done in the past where, in a local election where records could be searched, we validated who among the people we interviewed actually did cast a ballot and who didn’t. The need to estimate a person’s likelihood of voting based on numerous indicators is due to the fact that people overreport their intention to vote. In our most recent survey, 71 percent of registered voters told us that they would “definitely vote” this year. That would be a remarkably high turnout for a midterm election were it to actually come to pass.
There’s sometimes discussion of the mix of respondents in polls between Democrats, Republicans, and independents. How do you choose the right mix, and when does that change?
We never “choose” the right mix of Democrats, Republicans, and independents. We conduct a random sample of all Americans, and identify the likely registered voters from that random sample. The balance of party identification that results is one of the things we are measuring, not a parameter that we define. We believe it is impossible to fix what party identification should be because the balance of party identification shifts over time. Any pollster who assumes that the percent of Republican, Democratic and independent voters this year will be the same as in 2006 is ignoring the enormous changes the country has undergone in the intervening four years, and huge fluctuations in the intensity and engagement of partisan groups in different election cycles.
How do you determine a representative sample of voters, and does it vary between a presidential-election year and a midterm election?
Our survey samples are based on what is called “random digit dialing,” in which we randomly sample from active blocks of telephone numbers (a “block” is a set of 100 numbers, such as 202-555-12xx, where the xx ranges from 00 to 99). The last two digits of the block are randomly generated, to be sure that our surveys are not biased toward people with listed telephone numbers, or biased against people who have moved recently. Random digit dialing is more expensive than calling “listed samples,” but is not biased in any way.
How many voters are enough to produce statistically sound results?
There is no hard cutoff in terms of what sample size is necessary to be representative, but bigger is always better. Using a strict definition, the margin of error around a survey of 800 people is plus or minus 4 percentage points. That means you could have an election margin of 54 percent Republican to 46 percent Democratic but still not be confident that it is a statistically significant lead, because there is a reasonable chance the margin is actually 50–50. But if your survey has 1,600 people, that margin of error is plus or minus 3 percentage points, making it easier to draw firm conclusions from the survey results.
How useful is the generic ballot in predicting election results? In an age where there are many pollsters who poll in nearly every competitive race and results can be easily distributed over the Web, does it still serve a purpose, especially late in the campaign?
The national generic ballot—i.e., “If the elections for Congress were being held today, would you vote for the Republican candidate or the Democratic candidate in your district?”—is only a broad indicator of national trends, and obviously each of the 435 districts has its own internal dynamics that our surveys can never pick up. However, the correlation between the overall balance of votes cast for Republican and Democratic candidates nationally is strongly correlated with the number of seats each party will win. Thus national generic ballots can give a reasonably good estimate of the size of the “wave,” even if they can’t predict precisely which districts will go which way.
How should readers weigh the accuracy of robo-polls or Internet polls against in-person phone interviews?
This is a great challenge—because robo-polls are automated, they are prohibited by law from contacting cell phones, so right off the bat they are not representing the views of 25 percent of adults. However, the pollsters using these methods have weighting techniques to try to adjust for this gap in coverage, and have done a decent job of estimating the vote in previous election cycles. Readers of these surveys simply have to have faith that the expert adjustments can account for what is, by definition, not a random sample of all eligible voters. Internet polling, too, has its challenges. Since not all Americans have computer and Internet access, pollsters who want to be able to reflect the views of all Americans typically have to develop “panels” of participants, and give the technology to some subset of people who don’t have it themselves. Then these people become part of an ongoing panel, and respond to multiple surveys over the course of a year. There are questions as to whether this type of empaneling changes the way people answer questions, and the costs of setting up this kind of Internet panel can be quite steep. But there are serious polling experts who have advanced this approach quite far, and have produced quite accurate polling results.
Under what circumstances, if any, should readers heed the results of internal polling from campaigns?
The most important thing to keep in mind about internal polling from campaigns is that the campaigns may well be running 10 polls and only releasing the one or two in which their candidate does best. In short, I would not accuse any professional campaign pollsters of fudging their numbers or producing false polls—they are professionals and it’s not in the interest of their candidates’ campaigns to get bad information. The question readers should always ask is, which internal polls will the campaigns choose to make public?
There’s some muttering among liberals that many polls are flawed because they don’t survey cell-phone users, or because they undersample these voters, who tend to be younger and farther left. How does Pew calibrate the ratio between cell-phone and land-line respondents, and have you adjusted it since the 2008 election?
Our view is that this is an important issue. Current estimates are that 25 percent of adults in the U.S. today live in households that are “cell only” and have no landline telephone. But because cell-phone numbers are logged into separated, but similar, blocks of numbers, we are able to randomly dial cell phones as well. One out of every three interviews we complete these days is conducted over a cell phone, to be sure we are accurately representing the cell-only population. We issued a report a few weeks ago that showed that had we conducted our preelection surveys without any cell-phone interviews this year, in three of our four preelection polls the estimates would have shown slightly, but significantly, more support for Republican candidates and less support for Democratic candidates. This is consistent with the demographics of “cell-only” Americans—younger Americans and minorities are significantly more likely to have “cut the cord,” and these are the two best voting blocs for Democrats. We are always calibrating the balance of cell-phone and landline phone interviewing to try to keep pace with Americans’ changing technology use, and we balance each of our surveys to the best available current estimates of how many Americans are cell-only, landline-only, or have access to both kinds of phones.