Democrats have a 14-point advantage among likely voters in the race for control of Congress, according to a new poll. Or is that edge just 4 points? The answer depends entirely on the assumptions the pollster made about a question that is impossible to answer with certainty until all votes are tallied on Election Day: Who will actually show up to vote? With Democrats favored to win the House and state-level polls showing the Senate within the party’s reach, turnout remains the biggest unknown that will decide exactly what kind of election we have.

Republicans are surely hoping the polls this year are underestimating their support, as surveys did in 2016, 2020 and 2024, all presidential election cycles with Donald Trump on the ballot. It’s true that some GOP voters, particularly those drawn into politics by Trump, are less likely to take polls. That raises the specter that a critical chunk of the electorate is missing from the results even as most pollsters have adjusted their methods to try to account for these missing voters.

No one can know for sure who will turn out until after votes are counted, and it’s important to remember that polls are not predictions, only snapshots in time. Still, there’s a level of variation playing out in national surveys with implications for how we perceive the race for Congress, and that’s because of how pollsters go about identifying a population that doesn’t yet exist — the “likely voter.”Though we can’t always know for certain how survey research firms do this, we used the results of the new NBC News Decision Desk Poll released Sunday to illustrate just how much the choices of pollsters matter.

Depending on how we choose to identify likely voters, Democrats have anywhere between a 14- and 4-point advantage over Republicans when it comes to control of Congress in this year’s midterms. The reason for that difference? Two different forces showing up in the data, working against each other.

While Democrats have enthusiasm on their side, Republicans have hope that demographics provide a bulwark against any blue wave. Why identifying likely voters is so hard Predicting which party will prevail in the midterms isn’t as simple as picking up the phone and asking people how they will vote. About half of eligible voters are expected to turn out on Nov. 3, which means that a significant number of the people we reach in our polls will not vote. Each firm has its own method of determining who is likely to vote, but all likely voter polls rely on two ways to identify that population: letting the voters tell us what they will do, or predicting whether someone will vote based on who they are.

Sometimes these two things line up to produce similar outcomes, but this year, they push in opposite directions. Because firms have to decide how to mix these two factors, since neither works perfectly, the results of a likely voter poll are affected by the choices different survey firms make. A close look at the polls we include in the NBC News congressional ballot polling average illustrates the result of this dynamic, with the likely voter polls significantly more variable there than the registered voter or adult polls.

To produce the different results for our Decision Desk poll experiment, we have four scenarios for determining who is likely to vote. This first scenario simply uses registered voters as a stand-in for likely voters. Americans who aren’t registered can’t vote, so the usual first step of any likely voter identification is to refine the population to the people who are registered. Along with removing people who told us they aren’t registered, we weighted the registered voters to the correct proportions of age, education, race, sex, census regions and 2024 presidential vote. The second scenario is one where we take people at their word, only keeping the people who say they will definitely vote or claim to have already voted.

In our data these people heavily favor Democrats, which means a model that relies heavily on what people say they will do produces a wide Democratic margin. The problem with this approach is that some percentage of the people who aren’t sure they will vote will actually turn out, and those people largely prefer Republicans. The third scenario takes the opposite approach: We ignore intent completely and instead rely on who voters are to predict whether they will actually cast a ballot.

For this analysis, we looked at the 2022 Cooperative Election Study, an academic survey that, critically, goes back after the fact and records whether the respondents actually voted. Using those data, we built a model of what characteristics were related to turnout that year, and then used those results to predict who will vote in the 2026 data. This method results in a top-line result much closer to a tie.

A big reason is demographics. Older voters and white voters — demographics that lean more Republican — were significantly more likely to vote compared to younger and nonwhite voters in 2022. But another big reason is that this model assumes that Republicans will turn out to defend their current House majority in a way that is similar to how Democrats turned out in 2022, when Republicans wrestled control of the chamber away from the Democratic Party but only narrowly. For these types of models, it matters a lot which years are used to calculate who is more or less likely to vote.

If we instead pooled 2022 and 2018 (a much more Democratic-friendly year than 2022), our estimate would favor Democrats by 7 points. Finally, we can combine the two methods together. The last estimate again builds a model on 2022 data, but it takes into account each respondent’s characteristics and their intent to vote.

And we further adjust how we deal with intent: Instead of a hard line separating some people as definite voters and others as definite nonvoters, we assign a probability based on how likely it is people vote after giving a particular answer. For example, in 2022, around 15% of people who said they wouldn’t vote actually did, so in this model the people who say they won’t vote get at least a little weight. Unsurprisingly, mixing both these signals produces a result that sits somewhere in between the other two.

The need to balance all of these factors means that survey firms have enormous leeway to produce different outcomes from the same underlying data based on reasonable choices. The 2026 variation in action One way to really see the variations this year is by looking at pollsters who release results for both registered voters and likely voters from the same poll. Consider two prominent polling firms that both released surveys in mid-September. Both polls released results on 2026 vote choice for both registered voters and likely voters.

The two firms were close to agreement on the opinions of registered voters: The Economist/YouGov reported a Democratic 5-point advantage among registered voters, while New York Times/Siena reported 7-point edge. Where they differ is how they think those numbers shift when focusing on those likely to vote. The New York Times/Siena poll thinks that those who will turn out to vote are only slightly more pro-Democratic than all registered voters, producing a likely voter group with a 9-point advantage for Democrats.

This suggests that this pollster thinks that turnout between those who favor Democrats and those that favor Republicans will be pretty even this year, producing a result among likely voters that is broadly representative of all registered voters.

In contrast, the poll conducted by The Economist/YouGov suggests that likely voters will be significantly more pro-Democratic than all registered voters, with that group having a plus-12 edge for Democrats. This poll thinks that the registered voters who support the Democrats are more likely to turn out to vote compared with those who support Republicans. From what we know about how each determined their likely electorate, the New York Times/Siena poll describes a likely voter method that combines what respondents did in 2022 and 2018 with each voter’s self-described intent.

The Economist/YouGov poll reports taking into account intent. Both are reasonable methods, and the real result likely lands somewhere in the middle. No pollster knows exactly where that will be.