When you see an election poll, the headline usually focuses on which candidate is ahead. But before pollsters can estimate the preferences of the electorate, they face another difficult question:
Who is actually going to vote?
That is where voter turnout prediction comes in.
Not everyone who is registered will cast a ballot. Some people who say they plan to vote ultimately will not, while others who initially appear unlikely to participate may turn out after all. Because the people who vote can differ from the broader registered-voter population, estimating turnout is an important part of election polling.
Pollsters use turnout models, past voting behavior, survey responses, voter files, and other information to estimate the composition of the likely electorate.
Voter turnout prediction is the process of estimating which eligible or registered voters are likely to participate in an upcoming election.
Rather than assuming every registered voter has an equal chance of voting, pollsters may consider several indicators, including:
The objective is not necessarily to predict whether each individual person will vote. Instead, pollsters are trying to estimate what the eventual electorate will look like.
This distinction matters because registered voters and actual voters are not the same population. The U.S. Census Bureau’s Voting and Registration data, for example, separately tracks registration and reported voting participation in national elections.
It might seem that the easiest way to predict turnout would be to ask respondents:
“Do you plan to vote?”
The problem is that stated intentions do not always match eventual behavior.
A respondent might fully intend to vote when answering a poll but ultimately not participate. Another person who seems uncertain may decide to cast a ballot as Election Day approaches.
Because of this uncertainty, many pollsters use several indicators rather than relying on a single question.
They might ask:
Combining these indicators can provide a more detailed estimate of a respondent’s likelihood of participating.
A likely voter model attempts to distinguish between people who are more likely and less likely to participate in an election.
There is no single model used by every polling organization.
Some models use a scoring system. Respondents receive points based on answers to questions about voting intention, previous participation, political interest, and other factors.
Someone who says they are certain to vote, regularly participates in elections, and closely follows the campaign might receive a relatively high likelihood score.
Other models use statistical techniques to estimate each respondent’s probability of voting.
The American Association for Public Opinion Research (AAPOR) notes that estimating which respondents are likely to vote is a major challenge in election polling because turnout itself can be unpredictable.
Previous participation is one of the pieces of information pollsters can use when estimating future turnout.
Voter files may contain records showing whether registered voters participated in previous elections.
Importantly, these records generally indicate whether someone voted—not whom they voted for.
A person who has consistently participated in previous elections may have a higher probability of voting again than someone who rarely participates.
But previous behavior is not destiny.
An infrequent voter may become highly motivated during a particular election. Likewise, someone who normally votes could decide not to participate.
This is one reason predicting voter turnout remains difficult even when pollsters have access to detailed historical information.
Pollsters can approach voter turnout modeling in different ways.
One method is a deterministic model.
Under this approach, respondents who meet certain criteria are classified as likely voters. Those who do not meet the threshold may be excluded from the likely-voter estimate.
Another approach is a probabilistic model.
Instead of simply classifying respondents as voters or nonvoters, the model estimates the probability that each person will participate.
For example:
Voter A: 90% estimated probability of voting
Voter B: 65% estimated probability of voting
Voter C: 35% estimated probability of voting
Those probabilities can then influence how much each respondent contributes to the final estimate.
Both approaches involve assumptions. Neither can determine future behavior with certainty.
A turnout model cannot solve every problem in the original survey.
Before estimating who will vote, pollsters first need a sample that appropriately represents the population they are trying to understand.
That makes political poll sampling a fundamental part of election polling.
Pollsters may sample from voter files, probability-based panels, telephone samples, or other sources depending on the methodology.
If important groups are underrepresented at the sampling stage, the pollster may need to make adjustments later. Turnout modeling is therefore only one component of a broader polling methodology.
One important distinction when reading an election poll is whether its results represent registered voters or likely voters.
Registered voters are people who are registered to participate in an election.
Likely voters are the portion of that population a polling methodology estimates is more likely to actually cast a ballot.
The difference can affect polling estimates.
Imagine Candidate A has stronger support among occasional voters, while Candidate B receives more support from people who participate in nearly every election.
A poll of all registered voters could therefore produce a different estimate from a poll using a likely-voter model.
That does not necessarily mean one of the polls is incorrect. They may simply be estimating preferences among different populations.
Pollsters also have to consider the overall level of participation expected in an election.
Suppose turnout is expected to be relatively low.
A likely-voter model may produce an electorate containing a larger proportion of highly engaged and habitual voters.
If turnout is much higher, occasional voters may make up a larger share of the electorate.
These differences matter when voter groups have different candidate preferences.
The challenge is that the exact turnout level is unknown until the election occurs.
As a result, voter participation estimates involve assumptions about both individual behavior and the overall composition of the electorate.
Turnout modeling is closely related to—but not identical to—survey weighting.
Survey weighting methods are used to adjust survey data so that the sample better represents the population being studied.
For example, imagine a survey contains too many respondents from one age group and too few from another. Weighting may adjust how much influence respondents in each group have on the final results.
Turnout modeling answers a different question:
Which members of that population are likely to vote?
In election polling, these processes can work together.
A pollster may weigh a sample to address demographic or other imbalances and then use turnout information to estimate the preferences of the likely electorate.
Voter files can provide pollsters with another source of information for voter turnout prediction.
These databases can contain voter registration records and information about participation in previous elections.
When pollsters match survey respondents to voter files, they may be able to compare self-reported voting behavior with recorded participation history.
Voter files can also help researchers identify habitual voters, infrequent voters, and newly registered voters.
However, voter files have limitations.
Records can become outdated. Matching respondents to records may not always be successful. People move, registration information changes, and new voters enter the electorate.
Voter-file data should therefore be viewed as an input into turnout modeling rather than a perfect representation of future behavior.
The quality of a turnout model also depends on the quality of the information being collected.
If questions are confusing or push respondents toward a particular response, the data entering the model may be less reliable.
This is why researchers should pay attention to leading questions in surveys and other forms of question-wording bias.
Turnout questions should make it easy for respondents to answer accurately rather than encouraging them toward what they believe is the socially desirable answer.
For example, asking someone how certain they are that they will vote can provide more information than simply asking for a yes-or-no prediction.
People’s behavior can change between the day they answer a poll and Election Day.
Someone who seems disengaged during a survey may later become motivated to vote. Someone who says they are certain to participate may ultimately not cast a ballot.
Election participation can also be influenced by factors such as:
Another challenge is that many turnout models use historical information.
Past turnout can help identify patterns, but the next electorate will never be exactly the same as the previous one.
New voters become eligible. Some previous voters stop participating. Different elections can also motivate different groups.
A useful turnout model therefore has to learn from previous elections without assuming that the next electorate will simply repeat the past.
Research into likely-voter modeling has examined whether combining survey responses with verified voting history can improve estimates.
For example, Pew Research Center’s research on likely-voter models examined several approaches to identifying likely voters and found that incorporating information about previous voting behavior could improve some models.
The broader lesson is not that one specific model will always work best.
Rather, turnout estimation involves choosing among different indicators, assumptions, and statistical techniques—and those choices can affect the resulting poll.
Even a carefully designed turnout model cannot make an election poll perfectly precise.
Polling still involves sampling uncertainty and other potential sources of error.
This is where understanding the poll margin of error becomes important.
The margin of error primarily describes uncertainty associated with sampling under the assumptions used to calculate it. It does not capture every possible source of polling error.
For example, a poll could have a relatively small reported margin of error while still making an incorrect assumption about which groups will turn out.
Turnout-model error, nonresponse, question wording, coverage issues, and other methodological factors can introduce uncertainty beyond the reported sampling margin of error.
Yes.
Turnout models are estimates of future human behavior, so their assumptions can turn out to be incorrect.
A model could underestimate participation among one group while overestimating turnout among another.
If those groups have different political preferences, that error can influence the final polling estimate.
Different polling organizations may also use different likely-voter models.
One pollster might emphasize previous voting history. Another might place greater emphasis on stated voting intention and political engagement. Another could combine voter-file information with statistical modeling.
This helps explain why two reputable polls conducted during the same period can sometimes produce different estimates.
When looking at an election poll, don’t focus exclusively on the percentages beside the candidates’ names.
Look at the methodology as well.
First, determine which population the poll represents:
If the poll reports likely-voter results, look for information about how those voters were identified.
Did the pollster use previous voting history? Did respondents report how certain they were to vote? Was voter-file information used? Were turnout probabilities assigned?
Understanding these methodological choices provides important context for interpreting the results.
Voter turnout prediction is one of the most challenging parts of election polling because pollsters are trying to estimate behavior that has not happened yet.
Rather than simply asking people whether they intend to vote, turnout models in polling can incorporate previous participation, voting intention, political engagement, voter-file information, and statistical probabilities.
Those models also interact with other important elements of polling methodology, including political poll sampling, survey weighting, survey question design, and the interpretation of a poll’s margin of error.
No turnout model can know exactly who will vote before Election Day. But understanding voter turnout modeling, voter participation estimates, and the assumptions behind likely-voter polls makes it much easier to understand what election polls are actually measuring—and why their estimates can sometimes differ.
Sampling is the backbone of polling, determining how well a subset of people (the sample)…
Sampling is the backbone of polling, determining how well a subset of people (the sample)…
Sampling is the backbone of polling, determining how well a subset of people (the sample)…
Sampling is the backbone of polling, determining how well a subset of people (the sample)…
Sampling is the backbone of polling, determining how well a subset of people (the sample)…
Sampling is the backbone of polling, determining how well a subset of people (the sample)…