How We Determined Our 2016 Likely Voter Model

For those who follow Morning Consult’s election polling, you may have noticed something different about our results today. We have shifted to likely voters. Through the election, Morning Consult will release likely voters results along with the broader registered voter population.

Most pollsters do not release details describing how they created their likely voter model, which can make it difficult to determine why polls differ. At Morning Consult, we want to be transparent about our methodology to encourage a discussion about voter turnout and hopefully improve election polling.

How We Determined Our Likely Voter Model

The main challenge of likely voter modeling is determining which Americans are likely to vote on Election Day and which ones are not.

Between 86 percent and 90 percent of registered voters, and between 55 percent and 58 percent of all adults, cast a ballot in the four presidential elections since 2000, according the Current Population Survey (see table below). Morning Consult’s standard national polls each contain 2,000 registered voters, so approximately 10-15 percent of our overall sample will not cast a ballot in November.

Morning Consult has asked about 10 questions about vote history, vote intention, interest in the current election, and knowledge about election topics on each national survey in recent months in anticipation of releasing a likely voter model after Labor Day. We used these questions, along with results from nearly 20,000 interviews throughout August 2016, to subset the broader sample of registered voters to the narrower sample of likely voters – about 85 to 90 percent of individuals.

Our likely voter model includes any registered voters who say they are at least a 7 on a 1-10 point scale, where 1 means an individual will definitely not vote in the November 2016 presidential election and 10 means an individual will definitely vote in November. We selected this model for a series of reasons:

• Proportion of registered voters in sample: In our previous 10 national polls, between 86% and 89% of registered voters fit this screen. This is directly in line with the percent of registered voters who cast a ballot in the previous four November presidential elections.

• Clarity: This single variable prediction model is much easier to interpret than a series of indices (e.g., Perry-Gallup) or probabilistic models that we considered. We found that results from a single variable or an index approach did not vary considerably in our August 2016 polling.

• Directness: The 1-10 scale explicitly asks respondents how likely they are to vote in November. Some indices consider questions less directly tied to future vote intention, such as whether you know where your election precinct is, whether you have voted in previous elections, and how much you are thinking about the upcoming election. These indices are valid alternatives, but we aimed for directness and clarity.

Our previous month’s polling suggests that the differences in reporting between registered voters and likely voters will be relatively small. This is largely because the vast majority of registered voters cast a ballot in November, so the likely voter modeling only reduces the overall sample by 10-15 percent. Unlike previous election cycles, the results suggest that likely voter samples are not uniformly more pro-Republican than the broader registered voter sample. The table below, for example, shows the vote margin in the 2-way horse race for all polls in August broken down by registered voters (Column 2), our likely voter model (Column 3), and a likely voter model based on the Perry-Gallup scale (Column 4).