PRINCETON, NJ -- Only a subset of the total population of citizens in a typical election will actually vote. This subset often has different characteristics than the total population. Thus, the results of a typical election are different than they would be if every citizen actually turned out and voted -- because the people who actually vote are different from the people who don't. If pre-election polls don't take this into account, they run the risk of estimating an election result that will differ from the actual vote on Election Day.
These facts of life have led to more than 60 years of work by Gallup pollsters to perfect the process of scientifically determining ahead of time who is most likely to turn out and vote on Election Day. The basic process attempts to winnow down the total adult population in the weeks and months prior to an election to those most likely to vote -- or in other words, likely voters.
It is important to note that in some elections, all of the trouble to develop a sample of likely voters may not matter. If the subset of the population that actually votes is no different in most ways from the total population, then interviews with a random sample of the total population will yield a result that is accurately predictive of the election results.
More typically, however, the pool of those who vote is not representative of the total population. Experience suggests that actual voters can -- and often do -- differ in significant ways from all eligible voters. These differences are usually the result of two factors.
First, there are demographic characteristics that have traditionally been associated with the propensity to vote. Everything else being equal, certain types of people are more likely to vote than others. The most important of these are age and socioeconomic status. Older and better-educated individuals are, in general, more likely to vote than younger and less well educated individuals. Generally speaking, because Republican voters tend to be better educated than Democratic voters, the actual voter pool on Election Day can be skewed more toward Republicans than the total population composition would suggest. This in turn helps explain why, in some elections, the actual vote on Election Day can be more Republican than pre-election polls predict.
The second set of factors which relate to voter turnout are idiosyncratic and can reflect one of a hundred characteristics that come into play in any given election. In some elections, specific factors may activate or excite certain groups of voters and cause them to vote at higher proportions than their representation in the general pool of registered voters might predict. An election that has racial overtones, for example, might activate minorities more than usual. An election that has hotly debated labor-related issues might activate union members.
The challenge for the pollster is to find the best procedures or mechanisms to use to estimate who will be in and who will be out of the population of voters, and then to find a way of sampling opinion just from those who have the highest likelihood of voting.
The first and easiest step for pollsters is screening for specific, known voting requirements. The fact that voters in most states must ultimately be registered before they can vote provides a starting point for researchers, enabling them to reduce the total population of all Americans down to the smaller group of registered voters. Although it is possible to obtain lists of registered voters in most states, Gallup most typically begins with a sample of the general population and simply asks them, once they are reached on the telephone, whether or not they are registered to vote. Gallup poll experience indicates that about 80% will say yes. Thus, as a starting point, Gallup eliminates 20% of those we contact from our pool of likely voters.
Beyond that point, some pollsters have historically used background or demographic characteristics to determine likely voters, eliminating from a sample, for example, younger voters or those with lower levels of education. But, the impact of such background characteristics varies from election to election. Gallup generally uses the views, opinions and self-perceptions of the voters themselves as a mechanism for isolating likely voters.
The simplest example of such a "likely voter" question is the most straightforward: "Are you going to vote on Election Day?" A pollster could simply accept the respondents at their own word, and include in the final likely voter sample those who say yes, and eliminate those who say no. The major problem with this procedure is the natural inclination of the majority of registered voters who are interviewed before an election to say that they are going to vote -- without giving it much thought. In fact, Gallup research indicates that, on a routine basis, about 90% of registered voters will tell an interviewer that they are very likely to vote on Election Day.
But, variants on the idea of asking individuals questions in order to determine their probability of falling into the sample of likely Election Day voters can be quite effective. Over the years, Gallup has developed a series of questions that provide a good prediction of the probability that an individual will end up voting. These questions include asking whether or not the individual knows the location of his or her voting place, whether or not the individual voted in the past election, how closely the person is following the election, and so forth.
Based on more than 65 years of experience predicting "likely voters," Gallup's analysts have developed computer programs which give each person interviewed a score based on how they answer these questions. Those with the highest score have the highest probability of voting. These people have voted in the past, know where to vote, and have a high degree of interest in the election. Those with low scores have a low probability of voting. Despite what they say they are going to do, they don't know where to vote, they hadn't voted in the past, and they have a lower level of interest in the election. Thus, the computer program judges them to have a lower probability of actually voting.
At this point, the process Gallup uses reduces the total sample (which represents the population) down to a size that best estimates the percentage of the voting age population that will vote on Election Day. If turnout is estimated to be roughly 50%, a sample of 1000 is reduced to about 500, throwing out those respondents who have the lowest "likely voter" scores. In short, the procedure involves mimicking the real-world process by developing a final "likely voter" sample that approximates the ratio of likely voters to the total adult population in the real world.
Some polling organizations do not attempt to reduce the sample to "likely voters" until close to Election Day itself. This year, Gallup has decided to use likely voters for the basic presidential ballot question all year, under the assumption that this makes our understanding of where the election stands as accurate as possible, as early as possible.
It is interesting to note, however, that so far this year the voting intentions of likely voters are not significantly different from the voting intentions of those who are less likely to vote. This may change. In 1996, for example, the voting intentions of likely voters tended to tilt more toward Republican Bob Dole than did the voting intentions of the total adult sample, so that the likely voter sample was more Republican than the total population sample. Gallup's final estimate of the vote before the November 1996 election -- based on a sample of likely voters -- was very accurate in predicting the vote percentages ultimately obtained by Clinton, Dole and Perot.