Behind the Final Prediction

In Gallup's final prediction of the popular vote in the presidential contest, 49% of voters are expected to support President George W. Bush, 49% Sen. John Kerry, and 2% other candidates, including 1% for independent Ralph Nader.

Voter Presidential Preference

These numbers are based on a two-point Kerry lead among all registered voters, a two-point Bush lead among a subset of "likely voters," and a judgment by Gallup researchers as to how the undecided likely voters might actually vote. This allocation of the undecided vote converted Bush's two-point lead among likely voters to a dead heat. 

Some explanation behind these numbers:

Likely Voters

Since the 1948 election debacle, when all of the "scientific pollers" of the day predicted Republican Thomas E. Dewey to beat incumbent Democratic President Harry S. Truman, Gallup has employed a likely voter model to help differentiate the general public from those who actually turn out to vote. Part of the 1948 error was due, George Gallup felt, to the fact that he had measured the presidential preferences of all Americans, many of whom did not vote. The non-voters, he discovered in post-election polls, were more pro-Dewey than the voters.

The current likely voter model includes seven questions, all of which are asked only of people who first indicate they are registered to vote. Four of the questions elicit information about the respondent's past voting behavior, while three measure the respondent's current intentions (and intensity) to vote. Each question is worth a point, so that each respondent can be given a score from 0 to 7 on the likely voter scale.

During the campaign season, many political observers criticized this likely voter model because it screens out people who may express an intense commitment to voting, but whose past (self-reported) voting behavior is less than stellar. Indeed, all older people who admit they have not voted in previous presidential elections are excluded from Gallup's likely voter model, no matter how fervent their stated intention to vote in this election. By contrast, people who could not have voted in the 2000 election because they were too young are included in the model based solely on their stated intentions to vote, not on past behavior. Thus, the model does include new voters, but almost all are younger than 25.

Critics argue that this year, the intensity of the campaign and the get-out-the-vote efforts could render this model obsolete. At Gallup, we are skeptical of these criticisms, because the model takes an estimated turnout into account. In our final prediction, we assumed that 60% of our sample -- which means 60% of the general public -- will vote. Thus, the top 60% of respondents on the likely voter scale are included in our final likely voter estimate -- 56% who scored a "7" and 4% who scored a "6." 

To explore the criticisms, however, we created a "high interest" voter model, which includes all registered respondents who score highly on intentions to vote regardless of their past voting behavior. If they say that on a scale of 1 to 10, their likelihood of voting is a 10, and if they also say they have given the campaign "quite a lot" of thought rather than "only a little," they are classified as a high interest voter. This category includes 72% of the sample, which would translate into a 72% turnout rate. If that level of turnout actually occurred, it would be a record since the advent of the secret ballot. 

These high interest voters support Kerry by 48% to 47%, marginally better (but decisively, if the numbers are precise) than the likely voter model showing Bush up by 49% to 47%. If the high interest voter numbers are allocated the way we allocated the likely voter numbers, that would predict a Kerry popular vote victory of 51% to 47%, with 2% going to Nader and other candidates.

These results raise the questions of why and how we allocate the undecided vote.

Allocation of the Undecided Voters

George Gallup started the tradition of allocating undecided voters with his very first election prediction in 1936, believing that he had to prove to a skeptical public that this new method of "scientific polling" was an accurate representation of public opinion. In some years, the undecided vote can be as high as 10 percentage points, but this year only 3% of voters indicated they had not decided.

Pollsters generally use three ways to estimate how people who say they are undecided will actually vote.

• A statistical model is used to predict a person's preference by taking into account that person's party affiliation, ideology, gender, age, and other characteristics that correlate with voter preferences among respondents who have said how they will vote.
  
  
• The undecided voters are assumed to vote for the candidates in the same proportion as those who have indicated a preference. Practically, this means ignoring the undecided and calculating the final numbers based solely on those who have reported their preferences.
  
  
• In an election with an incumbent, undecided voters are predicted to vote overwhelmingly for the challenger. The theory is that if, after four years and a campaign, the voters have not been persuaded to support the incumbent, then they will almost surely vote for someone else.

This year, we mostly followed the third method -- by allocating all the undecided vote to the challengers. But we also added a complicating factor. A quarter of likely voters, 25%, indicated they had already voted. They supported Bush by 49% to 45%, with 1% for other and 5% who refused to reveal their vote. Instead of assuming that this group of 5% would vote for the challengers, we assumed they would vote in the same pattern as the other "early voters." Among people who had not voted, if they refused to indicate a choice or indicated they were undecided, they were assumed to be voting for either Kerry or other candidates.

Thus, our net result: Among the 3% undecided, 1% support was allocated to Nader and others, and 2% support was allocated to Kerry -- giving a 49% to 49% dead heat.

Now we wait to see how well our likely voter model, and our allocation decision, will correspond with the "real world."