The Search for Likely Voters

Finding a way to sample Americans who are most likely to vote is one of the most vexing problems facing public opinion researchers. We are going to hear more about this problem in the weeks ahead and as this year's presidential election draws closer.

Here's the situation in a nutshell. In pre-election polls, pollsters are most interested in the group of individuals who are going to vote on Election Day. The problem: There is no precise way to identify exactly who will or will not be voting on Election Day. 

We clearly don't want to publish the voting intentions of the entire adult population and argue that they are representative of what is going to happen on Election Day. We know that at best, about 35% of all adults will vote in midterm elections, and about 50% or slightly more will vote in presidential elections. The subgroup of those who actually turn out to vote can be, and often is, different from the total population. When it comes to elections, we want to be able to estimate the attitudes and perceptions of this voting group ("likely voters") in order to best understand what is happening in regard to the election.

Gallup researchers have spent more than 60 years trying to perfect the process of creating a sample of likely voters prior to the election that best represents the population of those who will vote on Election Day itself.

We have some starting points. We use specific, known requirements for voting to help narrow the population field. For example, voters must be 18 years old or older on or before Election Day, and must be U.S. citizens. Perhaps most importantly, voters in most states must be registered before they can vote. 

So, we begin with the goal of limiting our sample to U.S. citizens who are at least 18 and registered to vote. But even reaching this relatively simple starting goal is in some ways harder than it would seem.

Some pollsters obtain lists of registered voters from registrars. This seems like a reasonable way to get a good sampling frame, but there are problems. In a few states, voter registration continues right up until Election Day. Thus, any list of registered voters can quickly become outdated. Also, there are logistical problems with trying to bring together lists that contain all of the names of registered voters in an entire state or the entire nation -- this involves piecing together a patchwork quilt of information from thousands of different jurisdictions. The advent of electronic databases in recent years has made this less of a problem, but there is, as yet, no one available list of all the country's registered voters.

Another problem with registered voter lists: Most do not contain the phone numbers of all of the registered voters listed. One can match the names and addresses to obtain some phone numbers, but this does not get around the problem of unlisted numbers. So using registered voter lists is, for the most part, problematic.

A more typical procedure for survey researchers interested in obtaining a sample of Election Day voters at the state or national level is to use a screen-down process. That's what we do at Gallup. We begin with a basic sample of the general population and ask survey respondents whether they are registered to vote. Gallup Poll experience indicates that about 80% will say yes. So we cut out 20% of those we contact right off the bat.

Then, the challenge becomes one of further thinning the sample to those who have the highest probability of actually voting on Election Day. The fact remains that not all registered voters vote. In this country, only about 60% to 70% of registered voters will vote in any given election, a percentage that is somewhat higher in high-profile or hotly contested elections, and lower in off-year races or in elections without much interest or specific publicity hook.  

Some people's first reaction at this point is to suggest that pollsters simply ask those we contact: "Are you going to vote on Election Day?" and use those who say "yes" as the basis of the likely voter sample.

However, asking people a simple, straightforward question about their voting intentions doesn't work all that well in and of itself. The problem lies in the fact that many registered voters have the best intentions in the world about voting but don't actually make it to the election booth on Election Day. There is a social desirability and self-perception effect at work: The majority of registered voters who are interviewed before an election are naturally inclined to say that they intend to vote on Election Day. In fact, Gallup research indicates that on a routine basis, more than 90% of registered voters will tell an interviewer that they are very likely to vote on Election Day, when just about 60% to 70% actually vote.

We've discovered that if we ask a set of more indirect questions, we can better predict who is or is not likely to vote. We combine the answers to these indirect questions, assign every potential voter a score, and use this score as the basis for identifying likely voters. Those with the highest numbers have the highest probability of voting. These people answer almost every question in a way that indicates a strong likelihood of voting. They have voted in the past, know where to vote, and have a high degree of interest in the election. Those with low numbers have a low probability of voting. Despite their professed voting intentions, they don't know where they are supposed to vote, haven't voted in the past, and are less interested in the election. Our experience suggests that they probably won't show up at the polls. 

Gallup separates out a smaller subset of people in the sample at the high end of this "probability of voting" scale that we judge is most representative of the subset of people in the real world who would end up voting. We call these people "likely voters," and their answers are given full weight in the final sample. The answers of a second group are included, but given a smaller weight because their probability of voting, although high, is slightly lower than that of the top group. And the responses of a third group of people -- those who have the lowest probability of voting -- aren't considered at all in the final analysis. We estimate that the chances people in this third group will actually turn out to vote are so low that they should be totally excluded from our final sample of likely voters.

It's important to keep in mind that our final sample of likely voters in pre-election polls maintains the basic assumption of randomness that guides the entire sampling process. The likely voters in a sample are a randomly selected subset of likely voters in the real world. They thus become an extremely valuable group of people. Based on the mathematical properties of randomness and probability, it is almost certain that their responses will be representative of the responses of all voters across the country.

The final step in this process, of course, is to look closely at what the people in this carefully selected random sample say they are going to do on Election Day, most often a variant on this question: "If the election for president were being held today, would you vote for x or y?"

It's important to emphasize that we certainly don't want to use this likely voter sample for everything we do. In many instances, we are interested in the attitudes of the entire sample of Americans, not just those who go out and vote. But when it comes to understanding what is going to happen in specific elections, the likely voter sample becomes vitally important. The use of Gallup's likely voter techniques has been extremely successful over the years. In most instances, Gallup's likely voter estimates are quite close to the final "real-world" outcome.