Patrick Flynn
The first public US presidential election MRP poll from Focaldata finds that Donald Trump is narrowly on course for a second term in the White House, defeating Democratic candidate Kamala Harris.
The former president leads in states totalling 291 electoral votes to Harris’ 247. Trump is set to flip Arizona, Georgia, Pennsylvania and Wisconsin, carrying 56 electoral votes between them.
Nationally, Harris leads the popular vote by two points, 47.5% to 45.9%, but loses the tipping-point state Pennsylvania by 1.4 points. This would mark the eighth popular vote win for the Democrats from the last nine presidential elections, and the third time since 2000 the party has won the popular vote but lost in the Electoral College.
Based on 2020 turnout, just 63,000 votes could change the outcome. If Harris were able to hold Pennsylvania and Wisconsin, she would take the White House and become the first female president of the United States.
The House of Representatives is on a knife-edge, but the Democrats are favoured to regain control, with 223 seats to the Republicans’ 212. In contrast to the presidential race, this comes in spite of the Republicans winning the popular vote.
The six closest states are:
Kamala Harris leads with voters under 45, but sees a significant drop in vote share with 18–34 year olds compared to Joe Biden in 2020. Harris wins 53% of the vote among this age bracket, down 13 points on her party’s 2020 result. Democratic losses are biggest with black and Hispanic voters. The party falls from 91% to 81% with the former group, and 64% to 56% with the latter.
Our results point towards an Electoral College bias (i.e. the difference between the state which gets the winning candidate over 270 votes and the national popular vote result) of three points in Trump’s favour. We estimate that Harris would need to win by at least three points nationally to win the election. This figure is in line with 2016 and 2020 (2.9 points and 3.8 points, respectively).
Concurrent with our MRP poll, we also conducted regular polling of voters in seven key swing states to find out their views on the election. Note: these findings may not align entirely with the MRP findings, due to different methodologies and respondents. They are designed to give us more insight into voter behaviour.
We find that Tim Walz is the most popular candidate on either ticket, with an average net favourability rating of +7. Kamala Harris is at +3, JD Vance is -2 and Donald Trump is -3. JD Vance is particularly popular in the South, outrunning Trump as the most popular person on either ticket in both Georgia and North Carolina, but is more of a drag in Western states like Arizona and Nevada.
Inflation / the cost of living and immigration are the top two issues in all swing states, but both are particularly acute in Arizona.
Finally, this is for us the most important graphic of this election. Donald Trump is seen as the best person to handle the most important issues facing the country, often by quite a distance.
"The House and Presidential votes both look set to be knife-edge, fall different ways, and in both instances fall to the party that loses the popular generic ballot vote. Perhaps what the electorate are saying is they are unsure who should win, and whoever does should be massively checked - whether that's the executive and legislature. That's a more subtle story than just saying America is divided 50/50. People are internally and externally conflicted."
"This MRP model really confirms what other state polling has shown. Namely, that over the last 25 years the Republicans have sacrificed 27 Electoral College votes (Colorado and Virginia) in all scenarios to gain 52 votes (Ohio, Florida and Iowa) in almost every possible simulation of an election. This systemic 25 Electoral College vote improvement is haunting the Democrats in an environment where they are set to win the popular vote but lose the presidency again."
"This is a knife edge election - decided again by a clutch of 5-7 states and tens of thousands of votes. Just a 1% swing to the Democrats flips Arizona and Pennsylvania in our model towards 277 votes for the Democrats."
"MRP is one of the best methods there is to estimate results at small geographical levels, and will play a major part in the future of the polling industry as costs and processing time both come down. In the UK, we have seen huge growth in MRP polling over the last few years – our own MRP poll in the recent election predicted just one fewer seat correct than the official on-the-day exit poll, so we are excited to be publicly releasing the first national MRP of the Harris-Trump race."
"We find ourselves in the bizarre situation in which the Democrats are favoured to win the House of Representatives but lose the House popular vote; and lose the presidency despite winning the presidential popular vote."
"This election may not be decided by grand ideas on issues like democracy, but from a handful of working parents in Pennsylvania who don’t have time to watch CNN or MSNBC and merely want to know who is going to bring the inflation rate of gas and groceries down. Many of those people look back and think ‘I was better off four years ago under Trump’. On that, Harris has work to do."
Focaldata surveyed a nationally representative sample of 19,428 respondents from online panels. Fieldwork was carried out between the 16 August - 2 September 2024. The sampling frame was constructed using: state, age, gender, education level, race, Hispanic identity, and 2020 presidential vote. The MRP modeling process effectively weights all of these, as well as marital status and the interaction of 2020 vote with age and education.
What is multilevel regression with poststratification (MRP)?
Multi-level regression with poststratification (MRP) is a statistical technique for estimating public opinion in small geographic areas or sub-groups using national opinion surveys. It originated in the US, and was used by academics to estimate state-level opinion cheaply, given the expense of doing polls throughout the country.
MRP has two main elements. The first is to use a survey to build a multi-level regression model that predicts opinion (or any quantity of interest) from certain variables, normally demographics. The second is to weight (post-stratify) your results by the relevant population frequency, to get population level (or constituency level) estimates.
At the end of this process the aim is to get more accurate, more granular (thus more actionable) estimates of public opinion than traditional polling. There are, however, significant technical challenges to implementing it effectively. These include large data requirements, dedicated cloud computing resources, and an understanding of Bayesian statistics.
Who built the model?
This new US model was built by Dr Ben Lobo and Dr Adam Higgins. Our modelers received technical support from Dr Pete Logg, Dr Matt Chennells, James Alster and the wider technical team, and domain support from our Chief Executive Justin Ibbett and Chief Research Officer James Kanagasooriam.
Who is Focaldata?
You can find out more about us here, or on our website.
Data tables
You can find our MRP data tables for all states and congressional districts here.
How did you create the model?
A brand new poststratification frame for the US was built from scratch. This was done by collecting US 2020 Decennial Census data, along with data from the 2020 & 2022 editions of the American Community Survey (ACS) at the census tract level. From this, proportional fitting was used to build (geo)demographic frames [congressional district, state, age, sex, race/hispanic status, education, marital status] for the congressional district boundaries on which the upcoming election will be fought.
A database of demographic change between election cycles was constructed by comparing editions of the ACS. From this, a model was built to map past election results to different boundaries, forming a library of ‘notional’ election results.
Using the Cooperative Election Study (CES/CCES) as training data, along with our library of notional results, new Voting-Behaviour and Turnout Models were developed, and then used to model vote and turnout history onto these demographic frames to finish the poststratification frame - the “P” part of MRP.
Attenuation bias
MRP has had some problems historically around regularisation and attenuation. For the UK general election, we deployed an ‘unwinding’ adjustment to account for this, which made a tangible difference to overall seat counts. Any unwinding adjustment for this MRP, however, would not make a difference to the overall outcome and no states would change hands, so we have opted not to perform it for this release. Some results at the extremes (e.g. District of Columbia) may display smaller leads than expected.
We are currently assessing whether to employ an unwinding adjustment for our ‘final call’ in late October prior to the election.
The polls could move
While confidence intervals are included in the data tables, please note that these are confidence intervals within the model itself. Our 90% confidence intervals give us an idea of the likely range of results the same model would produce if we ran it again with a different set of respondents. As such, these intervals do not account for other sources of error, like the polls moving in the run-up to the election. A candidate having a lead outside the margin of error should therefore not be read as a certain victory in November.
Over-estimating 'other' parties
Generally speaking, MRP is not optimised to handle the results of smaller parties. Part of the 'regularisation' inherent in multilevel modelling will increase the predicted share of the vote for small groups, like 'other' parties in our model.
We also see some respondents selecting 'other' as a quasi-don't-know option. Again, this is one of the things we are looking at for future iterations of the model.
Furthermore, some of our data was collected prior to Robert F. Kennedy, Jr. dropping out of the race, which will naturally inflate the vote share for third parties.
If you'd like a comment from us, a chat about this poll, or if you have any thoughts or feedback, please email patrick@focaldata.com.
