How do Gallup's likely voter models work?

Turnout in typical presidential elections is related to a voter's current interest in the election, self-reported intention to vote, and previous voting history.

Based on these relationships, Gallup has in past elections successfully created a pool of likely voters (the "traditional" likely voter model) using data on an individual voter's values on these dimensions. The voting intentions of this pool of likely voters, measured in the final survey before an election, typically provide a more accurate representation of the final popular vote outcome than is accorded by the polled vote patterns of all registered voters. For example, in 2004, Gallup's final poll of registered voters before the election had John Kerry leading by 48% to 46%. Among likely voters, defined using Gallup's "traditional" likely voter model, George Bush was ahead by 49% to 47%, a margin quite close to the final popular vote outcome.

The voter turnout pattern in the 2008 presidential election may end up following the same tendencies that have occurred in previous presidential elections. If so, the "traditional" likely voter model will provide the best estimate of the final popular vote.

By many accounts, however, turnout this year has the potential to follow somewhat different patterns. New voter registration is reported to be up significantly in a number of states, and news accounts discuss evidence that the newly registered voters are disproportionately Democratic. Gallup has also monitored significantly higher enthusiasm about this year's election among Democrats compared to Republicans. Given that Barack Obama is the first major-party black candidate in U.S. history, it would not be unexpected to see unusually high turnout among black voters, and there continues to be discussion of unusually high registration and turnout among young voters.

These considerations have led Gallup to develop a second likely voter model (the "expanded" model) that takes into account current voting intentions, but does not incorporate past voting behavior. This model assumes there will be significant numbers of first time or infrequent voters in the final electoral pool, and that past voting history is less important as a predictor of voting.

Thus, at this point, Gallup is providing our data users with several ways to model the electorate, taking into account different assumptions about potential turnout. The base registered voter model reports the current voting intentions of all registered voters -- the data Gallup has been tracking all year. The "traditional" model assumes that both past voting history and current voting intentions are important determinants of likelihood to vote. The "expanded" likely voter model assumes that current voting intentions are the important determinant of likelihood to vote, and that past voting history will not be as much of a factor as it has been in previous elections.

Gallup is monitoring a number of elements that comprise the likely voter models in order to detect any unusual patterns historically that would suggest traditional assumptions about turnout may not apply this year. One key will be the effort to see if certain subgroups of the population, including young people and minorities, are disproportionately likely to report being registered to vote, disproportionately likely to report high levels of interest in the campaign, as well as self-reported intention to vote when compared with previous years. This monitoring will provide the best empirical estimate of the turnout patterns that will occur on Election Day. The analyses of these indicators will be reported periodically at gallup.com.

Details on how each likely voter model is computed are outlined below.

Computation of Gallup's Traditional Likely Voter Model for Presidential Elections

In each sample, a subsample of likely voters is identified by taking into account each respondent's score on a 0 to 7 scale that assesses a person's likelihood to vote. The top scorers on this scale -- equal to the projected turnout -- are identified as likely voters.

I. Computing the Likely Voter Score for Each Respondent

The seven questions comprising the likely voter scale consist of the following measures (see the full wordings at the end of this page). Respondents get a point on the likely voter scale if they give the response listed in parentheses.

1. How much thought given to election (quite a lot, some)

2. Voted in election precinct before (yes)

3. Know where people in neighborhood go to vote (yes)

4. Plan to vote in 2008 election (yes)

5. How often vote (always, nearly always)

6. Voted in last presidential election (yes)

7. Likelihood to vote on scale from 1 to 10 (scores of 7 to 10).

The following adjustments are made to these raw scores:

· Respondents who are not registered to vote are assigned a score of 0.

· Respondents who do not say they plan to vote (see item 4 above) are assigned a score of 0.

· Younger respondents' scores are adjusted to account for the fact they were not able to vote in some or all past elections. In other words, even though the model identifies voters on the basis of past voting history, younger voters are not penalized because they were not of voting age in past election years.

Ø If age 18-19, their scores are converted as follows: 1=2, 2=4, 3=5, 4+=7

Ø If age 20-21, their scores are converted as follows: 1=1, 2=3, 3=4, 4=6, 5+=7

· Respondents who indicate that they have already voted are assigned a score of 7.

II. Computing Likely Voters as the Proper Proportion of the Sample to Equal Estimated Turnout

Gallup begins by assuming turnout of 50% of all national adults. This estimate is adjusted up or down depending on Americans' reported interest in the election and voting intention. Since survey indicators monitored by Gallup show that interest in the 2008 election is as high as, or slightly higher than, it was in 2004, Gallup is assuming 60% turnout for the 2008 election. This figure may be adjusted until the final pre-election poll depending on the current level of interest in the election.

Thus, Gallup models this turnout level by counting the highest 60% of scorers as likely voters. In most instances, this means that all whose score is 7 on the likely voter scale will be included and some proportion of those whose score is 6.

Those who score a 7 on the likely voter scale receive a full weight factor of 1 in the likely voter sample. Those who score a 6 on the scale receive a weight factor greater than 0 but less than 1 (whatever proportion of 6's is needed to equal to 60% -- or whatever projected turnout is -- of the weighted sample size of national adults). Those who score a 5 or below on the scale receive a weight factor of 0 on this scale. [This is the typical pattern, though sometimes Gallup only needs those who score a 7 on the likely voter scale to meet the turnout level, and at other times may need to include some who score a 5 in addition to those who score a 6 or a 7].

It is important to point out that the likely voter sample calculations are based on weighted sample sizes. Gallup usually reports unweighted sample sizes to indicate how many respondents are in a particular sample. The two sample sizes can vary, sometimes significantly, in this likely voter model, depending on how many people scoring 6 are needed to reach the projected turnout level. The unweighted sample takes into account all respondents in a likely voter group even if those in that group are weighted such that their vote responses are given only proportionate value in the likely voter pool.

For a sample size of 3,000 national adults, the likely voter sample will have a weighted sample size of 1,800 national adults. The weighted sample size for likely voters will always be equal to 60% (or whatever turnout level is being estimated) of the weighted sample of national adults.

Computation of Gallup's Expanded Likely Voter Model for Presidential Elections

The expanded likely voter uses a subset of the same likely voter questions used in the traditional model. But there are two important differences between the models:

1. The expanded model does not assume a turnout level. Anyone who gives the proper answers to indicate a high probability of voting is considered a likely voter. Thus, in theory, 100% of a national adult sample can be likely voters in this model if they answered the questions in the proper manner.

2. The expanded model only uses questions that reflect current voting intention.

Only two questions are "scored" in this model. The criteria on these two questions are somewhat stricter to be considered a likely voter in the expanded model than are used in the traditional model; this is to ensure a very high level of commitment to voting in the election.

1. How much thought given to election (quite a lot)

2. Likelihood to vote on scale from 1 to 10 (scores of 9 to 10).

As in the traditional model, the following adjustments are made to these raw scores:

· Respondents who are not registered to vote are automatically excluded from the expanded likely voter sample.

· Respondents who do not say they plan to vote are automatically excluded from the expanded likely voter sample.

· Respondents who indicate they have already voted are automatically included in the expanded likely voter sample.

Since the expanded model does not assume a turnout level, no further computations are needed to create the likely voter sample, and vote choice is analyzed among this group of respondents. The range of the weighted national adult sample that has been considered likely voters using this approach has typically been around 66%, but has been as low as 65% and as high as 79% in various Gallup Poll samples.

Considerations on Sample Sizes for the Traditional and Expanded Models

Because the same questions are used in both models, nearly 100% of those who are considered a likely voter using the traditional model will also be considered a likely voter in the expanded model. [The only exceptions would occur if a respondent gives a weaker response on the thought item ("some" instead of "quite a lot") or the 10-point scale (7-8 instead of 9-10)].

The expanded model will thus include a larger proportion of likely voters than the traditional model (hence the name).

Despite this, it is possible the traditional model could have a higher unweighted sample size than the expanded model. This is because a large group of those scoring "6" may be included in the likely voter group, but with their responses weighted down to a small percent of the overall total. In other words, the number of respondents whose vote choice is included can be larger, but the impact of their vote choices can be quite minimal, depending on weighting. This is all dependent on how many of those scoring a 6 are needed to reach the 60% turnout level in the traditional model. But since these turnout levels are based on weighted sample sizes, the traditional model will always represent a lower turnout level.

Here are some examples from the 2008 Gallup Poll Daily tracking:

Notice that even though the unweighted N is larger in the traditional model in these examples (because of the need to use a fraction of the second-highest scoring group to meet the turnout estimate), the weighted N is always lower, and represents a lower turnout than the expanded model does.

Question Wordings

1. How much thought have you given to the upcoming election for president -- quite a lot, some, or only a little?

1 Quite a lot

2 Some (vol.)

3 Only a little

4 None (vol.)

5 Don't know

6 Refused

2. Do you happen to know where people who live in your neighborhood go to vote?

1 Yes, any response given

2 No

3 Don't know

4 Refused

3. Have you ever voted in your precinct or election district?

1 Yes, any response given

2 No

3 Don't know

4 Refused

4. How often would you say you vote -- always, nearly always, part of the time, or seldom?

1 Always,

2 Nearly always,

3 Part of the time, or

4 Seldom

5 Never (vol.)

6 Don't know

7 Refused

5. Do you, yourself, plan to vote in the presidential election this November, or not?

1 Yes

2 No

3 Don't know

4 Refused

6. In the election for president in November of 2004, did things come up that kept you from voting, or did you happen to vote?

1 Yes, voted

2 No, did not vote

3 Don't know

4 Refused

7. I'd like you to rate your chances of voting in November's election for president on a scale of 1 to 10. If '1' represents someone who definitely will not vote, and '10' represents someone who definitely will vote, where on this scale of 1 to 10 would you place yourself?

01 1 Definitely will NOT vote

02 2

03 3

04 4

05 5

06 6

07 7

08 8

09 9

10 10 Definitely WILL vote

11 Don't know

12 Refused