Kamala Harris narrowly ahead of Donald Trump in YouGov's first MRP estimates for the 2024 presidential election

This is the first release of our model estimating votes for the 2024 presidential election in every state, based on nearly 100,000 recent interviews with registered voters. We show that Kamala Harris leads Donald Trump 50% to 47% just before their first debate. However, the race will be decided by who gets the most electoral votes, not the popular vote, and as it currently stands, the race is neck and neck.

Harris leads in 22 states and Washington DC with 256 electoral votes and Trump leads in 25 states with 235 electoral votes.

The election requires 270 votes, and according to our model, the outcome will be determined by races we currently classify as undecided: Arizona, Georgia, Pennsylvania, and Nebraska's 2nd District. All of these contests are within the margin of error and could go either way.

The model is based on 95,643 interviews from YouGov's US panel conducted between August 23 and September 5, 2024 as part of the SAY24 project, a collaboration between Stanford, Arizona State and Yale universities. We have been interviewing most of these panelists regularly since December 2023 and are conducting regular repeat interviews – either monthly or quarterly, depending on the panelist. This is a much larger sample than is typically used for opinion polls and allows us to make estimates for each state. Panelists were asked about their likelihood of voting and their voting intention, as well as a number of other questions.

A unique feature of this study is that we can follow the same people over time and see how they change—or don't change—as the campaign progresses. Unlike most polls, which draw a new sample with each new poll, we can distinguish between voters who switch between candidates and do not vote, and variations due to changes in sample composition. So far, we've seen remarkable stability in voters' candidate preferences in 2024. Even after the first presidential debate on June 27—between President Joe Biden and Trump—or the attempted assassination of Trump on July 13, there was little change in voting intention.

Another important feature of this dataset is that we matched participants to TargetSmart's national voter file. This means that we have a sample of verified registered voters, excluding those who tell us they plan to register between now and the election. We also linked voters to the counted vote in the precinct where they are registered. This ensures that the sample is representative of various geographic areas, some of which have been underrepresented in previous years.

Even with a sample of this size, we still lack data in some key areas and hard-to-reach demographic groups. In Maine and Nebraska, for example, electoral votes are allocated by congressional district, and we have to estimate the vote for each individual district. In small states, we have correspondingly small samples. And in large states, some groups – such as younger voters and voters from rural areas – may be in short supply.

Our approach to estimating the electoral college votes is based on a Multilevel regression with poststratification (MRP) model. We have used this approach successfully in past elections in the US and elsewhere. It uses a statistical model to predict the votes of all people in the national voter file, regardless of whether they are part of the YouGov panel or not. The 95,643 interviews of our panelists are used to train a model that classifies people as likely Harris, Trump or third-party voters – or non-voters. This model is then applied to the entire voter file. We then aggregate these predictions – in what is called poststratification – to estimate the votes of all registered voters.

The model consists of three stages: (i) estimating the probability of voting; (ii) conditional on voting, what the probability is of voting for a major party or third party candidate; and finally (iii) predicting support for Harris and Trump among major party voters. We also used MRP models to estimate votes in the 2020, 2018, and 2016 elections.

In early October, we will update our vote model based on about 30,000 interviews we conduct this month. Then, in late October, we will re-interview the entire panel and release a final forecast before the election.

In addition, we have built models for every race for Senate and House seats in the U.S. We expect to release the first estimates next week, and we will update them regularly as the campaign progresses.

We note that these models are based on what people tell us about their plans. A small portion of registered voters (6%) say they are undecided, but the majority tell us they have made up their mind. However, people can change their minds, and when they do, those changes should be reflected in our model updates. These results reflect our best assessment of the current state of the race.