It was once a political truism that the GOP had the more regular and reliable voters, educated suburbanites who voted even in municipal elections. So Republicans tended to do better in lower-turnout elections — better in midterms than in presidential years, for instance.
Accordingly, when pollsters narrowed their voter universe by applying a likely voter screen — broadly speaking, any method that seeks to predict who is actually likely to vote from among those reached by a survey — it tended to help Republican candidates. There are many ways to select for “likely voters.” Some polling firms simply ask respondents whether they plan to vote, while others use voter registration files to identify whether respondents have voted in past elections or create probabilistic models to estimate how likely each respondent is to vote. It can be as much of an art as a science.
Whatever methods pollsters used, though, they once predictably helped the GOP. Traditionally, pollsters began switching from registered voter polls to likely voter polls after Labor Day, which often boosted Republican candidates. For instance, you can see this in the data for Bush-Kerry polling in 2004.1
In the Trump Era, this tendency is far more ambiguous. Democrats now dominate among the educated, suburban voter base — and excel in low-turnout environments like special elections. Instead, Donald Trump relies on irregular voters who wouldn’t be caught dead voting for city council — but who can enthusiastically turn out in presidential years, or at least will turn out for Trump.
We carefully track the differences between likely voter (LV) and registered voter (RV) polls because this is one factor our model accounts for. Because LV polls are theoretically more accurate than RV polls, we treat them a little differently. When a firm releases two versions of a poll, we use the LV version instead of the RV version. And we weight polls of likely voters higher than polls of registered voters in our polling average and forecast.
Furthermore, when a poll only publishes RV numbers, we adjust it to essentially transform it to an LV basis. The adjustment is based on polls that release both likely-voter and registered-voter versions of the same survey, e.g. in this Washington Post poll of North Carolina, Trump leads Kamala Harris by 4 points among registered voters, but his lead shrinks slightly to 3 points in the LV version.2
Is this typical? Does Harris generally do better when these likely voter screens are applied? The answer is pretty weird — in reviewing the data for this story, we found something we weren’t expecting.