Focaldata
Since the start of the year, the team at Focaldata has been working on our state polls and MRP for the US presidential election, alongside analysing other methods of forecasting the election like early voting and fundamentals.
We are releasing our final MRP results and state polls; the results of which can be found here. Read on for our full report on the election.
--
Since the start of the year, the team at Focaldata has been working on our state polls and MRP for the US presidential election, alongside analysing other methods of forecasting the election like early voting and fundamentals.
After months of slight variation in the methods, our signal has now become much clearer: Kamala Harris is the narrow favourite to be elected president, but the race remains very close.
Assigning each state to the most likely winner, our current projection estimates that Kamala Harris will win 276 Electoral College votes to Donald Trump’s 262, on a national popular-vote lead of 2–3 percentage points.
A Harris victory would be driven by wins across the Rust Belt – Michigan, Pennsylvania and Wisconsin – while Trump is set to pick up southern states like Arizona and Georgia.
Both our MRP and swing state polls have converged on this path over the course of the campaign. Almost all of our MRP runs since September have pointed towards a Trump victory, but our final update has picked up on movement towards the current vice president in key states.
Due to sensitivities around polling error and fundamentals, our call remains ‘lean Harris’ by the barest of margins. Some of Harris’ most crucial leads come in states which had a significant pro-Republican polling miss in 2016 and 2020. A margin shift of just 1.4 points in the tipping-point state of Pennsylvania in our swing state polls would see Trump take the win – in raw numbers, this means just 40,000 voters changing their minds, or 0.02% of the national vote (for context, the average polling error in swing states was 3.5 points in Trump’s favour in 2016, and 3.7 points in 2020).
Both candidates are a normal-sized polling error away from a large win in the Electoral College. A 2020-style polling error would see Trump back in the White House with 312 electoral votes, winning every swing state. On the flipside, there is also a credible path for Harris to outperform Joe Biden. The vice president is within touching distance in North Carolina and a polling error of just 1–2 points in her favour across the swing states would see Harris sweep to victory by a 100-vote margin in the Electoral College.
In reality, not really. The polls underestimated Trump in 2016 and 2020, but it’s rare for polls to be out in the same direction for three elections in a row, which is not to say that it definitely won’t happen, but that the direction of past polling error tells us very little about its direction going forward. In fact, we could make the argument that pollsters may be overcorrecting for underestimating Trump in the last two elections and may miss a Harris victory.
Our early voting analysis (below), however, can help to identify where a polling error may occur. The prime suspect in this case is Michigan, which ranks behind Wisconsin as the swing state with the largest total polling error over the last four elections. Given the demographic makeup of the state, we think Michigan may be susceptible to a polling error which favours Republicans.
STATE POLLING
Our final wave of swing state polling puts Kamala Harris narrowly ahead in the key Rust Belt states of Michigan, Pennsylvania and Wisconsin, along with Nevada. Trump, meanwhile, leads in Arizona, Georgia and North Carolina. If the leader in each poll went on to win the respective state (however narrow their lead!), Harris would win the election with 276 Electoral College votes to Trump’s 262.
Harris’ vote share is holding up in the Rust Belt, fueled by strong numbers among older white women – a group in which Trump underperforms a generic House Republican. The Republican nominee has made ground since 2020 with younger black and Hispanic men, however, which explains why Arizona and Georgia are currently showing narrow leads for the former president.
See the data tables here.
MRP
Our final MRP model of the campaign paints a similar picture, with leads for Harris in most of the swing states and a 3-point lead at the national level (50% for Harris, 47% for Trump). Over the course of the campaign, most runs of our MRP showed Trump ahead, but we have captured a late move towards Harris.
Our state polls and MRP have converged over the course of the campaign, and now agree in all swing states except Arizona:
While both methods show Harris ahead, it’s worth reiterating that both candidates are a normal-sized polling error away from clear wins. A 2020-style error would see Trump winning all seven swing states, whereas a 2012 error would secure a solid Harris victory.
THE DEMOCRAT COALITION IS MORE REAGAN-ESQUE
We have been tracking the changes in the US party coalitions over time. The plots below show how the counties which vote Republican and Democrat have shifted over time – both in terms of the size of each electoral coalition, but also the types of voters who are opting for each party. From 2000–2008 the overlap between the Republican and Democrat coalitions was much greater, with much less differentiation by density and education, and to a degree income. Since then, we have seen the Democrat coalition become much richer, much more diverse and urban between 2008 and 2016. In 2020, we saw an interesting inflection in which the Democrat coalition became higher income, less diverse, and even more educated than before – losing ground in ethnically diverse and working class areas, and picking up support in whiter, richer areas.
Compared to a generic Democrat, Harris is disproportionately popular with white women, whereas Trump gains with Hispanic men compared to his Republican House counterparts. This may see the Republican coalition becoming more diverse again in 2024, and will no doubt be the subject of a serious post-mortem in the event of a Trump victory.
If we weren’t looking at voting-intention polls, our prior would be a Trump win. The recent performance of incumbent governments across the Western world since the post-COVID cost of living crisis and Trump’s material lead on the top electoral issues both point towards a second non-consecutive term for the Republican nominee. On Harris’ side, however, is her superior favourability ratings in the Midwest.
WISDOM OF THE CROWDS
We asked respondents in the swing states the following two questions:
Crowdsourcing a forecast in this way can be a soft signal of election results. In every swing state we polled, a majority of respondents thought people in their local area would vote for Trump. When questioned about the national winner, a narrow majority of voters in six of the seven swing states thought Harris would win the election. Interestingly, in our MRP and ‘final call’ exercise on an Australian constitutional referendum the wisdom of crowds and voting intention broadly aligned – and accurately, we may add.
We do not see crowd-sourced forecasting as a hugely useful metric until we have more evidence, so it only plays a very small part in our final call on the election. If Trump goes on to win, however, the signs will have been there, but we simply may not have trusted non-traditional methods enough.
TOP ISSUES
Trump’s lead on the top issues has not shifted over the course of the campaign. However, we also find that some Trump-heavy issues like immigration have fallen in salience in recent weeks, with the former president unable to keep the issue as high up the agenda as it was a couple of months ago. In contrast, areas in which Harris is strong – particularly healthcare – have climbed up people’s priority lists.
The above chart shows the lead for each candidate among those who select an issue as one of their top priorities, so there is a degree of partisanship which can affect the outcome. For example, Republican voters are more likely to select immigration as a top priority, so it stands to reason that Trump’s lead on the issue will be high. However, even when we account for partisanship by asking all voters for their assessment of the candidates on the big four issues (inflation/cost of living; immigration; jobs and the economy; and abortion), Trump still led on three.
FAVOURABILITY
Harris boasts a higher approval rating than Trump in five of the seven swing states, with just Arizona and Georgia giving leads for the Republican nominee. Harris’ approval ratings are pretty consistent across the swing states, but in parts of the Rust Belt – specifically Michigan and Wisconsin – Trump’s favourability ratings are prohibitively low.
Early voting analysis is traditionally frowned upon, and trying to read the tea leaves through direct comparison with 2020 early voting stats (an election when the country was in the middle of a global pandemic and one candidate was openly stating that early voting was going to be rigged against him) is a fool’s game.
However, as a research company we have the advantage over other early vote analysis methods in that we have actually asked voters whether they would vote early or on election day, so no comparison with 2020 is necessary. We have compared our estimates of the early-voting electorate with the demographic breakdowns of votes so far to project the overall early vote shares for each candidate, allowing us to compare the estimated totals with hypothetical early vote shares if the race were tied overall.
For example, in a tied race in Arizona, we would expect Harris to lead the early vote by 2.8 points based on our polling. Based on the demographics of those who have actually voted early, however, we think that the early vote is basically tied (Harris leads by 0.6 points). Harris underperforming a tied race by 2.2 points means that we project Trump will take the state by that margin based solely on early voting analysis.
Like other methods, our early voting analysis points to an extremely close race, with a razor-sharp margin of just 0.3 percentage points favouring Harris in the tipping-point state of Georgia. Early voting stats do not give any indication of a landslide for either candidate.
The one state that stands out in our analysis is Michigan, where Harris is significantly underperforming expectations. We estimate that she leads the early vote by 12 points, when a tied race would show a Harris lead of 19. These stats have echoes of the Democratic primary in the state earlier this year, where an ‘uncommitted’ movement opposing Joe Biden picked up 13.2% of the vote from the Democratic electorate, driven partly by a rejection of the administration’s position on the Israel-Palestine conflict. Michigan has the fifth-highest Muslim population in the country at 2.4%, and as we saw in the recent UK general election, pollsters have serious issues with sampling ethnic-minority respondents – particularly those of Muslim background. These factors all create a perfect storm for a potential polling error in Michigan, meaning our final call in the state is only ‘likely Harris’ rather than ‘safe Harris’.
This year’s US presidential election has felt like a remarkably low-key affair given the stakes involved and the tightness of the race. Pollsters are afraid of another big miss on the election and underestimating Trump for the third time in a row. A degree of herding appears to be occurring in public polls. For non- poll junkies, herding is a phenomenon where pollsters aim to reduce the risk of being a notable outlier who calls the race wrong by weighting their results to more or less match other public polls. This reduces reputational risk for an individual pollster, but it makes the overall polling average more prone to a miss, and misinforms the public about the range of possible outcomes.
Deviation from the ‘true’ value is inherent to polling and is very much a good thing. Due to their sample sizes, individual polls are not perfect predictors of the population they are trying to measure (this is why you might hear about margins of error), and they should not be presented that way. If an election were truly tied 50-50, a poll of 1000 people should show a rounded lead of six points or more for one of the candidates around 7-8% of the time, and a rounded lead of four points or more around a quarter of the time.
If we compare those estimates with actual poll results from the seven key swing states, we see that pollsters have released more than 50% more tied polls than we would have expected (13% expected vs 21% actual), and while over a quarter of polls (25.9%) ought to show rounded leads of four points or more, barely 1-in-10 (10.6%) polls in October did. While weighting schemes will tighten the distribution of our expected values, some firms are putting out polls where the variance between states and previous polls looks suspect.
The tight distribution of polls is giving the false impression of a high degree of certainty in a close race, which should not be the case. In reality, both candidates are a normal-sized polling error away from a clear win. A hard-to-read race does not necessarily mean it will be close. As mentioned earlier, there is a clear pathway for Harris to have a highly comfortable victory; we have her almost over the line in both Georgia and North Carolina state polls – 0.6 points or less, which would give her a win bigger than Biden’s 2020. Indeed our MRP simulations show that the number of large Democratic wins dwarfs the number of simulations with large Trump wins.
How does Focaldata’s methodology differ from anyone else’s?
We have been concerned about the sources of polling error among US pollsters over the last ten years, and have taken time to address it. We believe that viewing the race using both traditional polling and MRP modelling is the best way to get a clear view of the overall electoral picture.
Our turnout model is a big distinguishing feature in our polling. Our analysis shows that merely relying on someone’s self-described likelihood to vote does not provide you with the full picture, and that a ‘definite’ voter has different definitions across a variety of demographics.
Our turnout model uses validated voter panels to assess the effects of demographics and self-described turnout likelihood on voter participation, in order to get a better view of who will actually turn out in an election.
We also have an Automatic Data Quality (ADQ) engine, which screens out bots, duplicate respondents and those who race through the survey without reading the questions properly.
How does MRP differ from normal polling?
While state polls typically rely on samples of around 1,000 in each state, our MRP model has a ‘mega sample’ of over 20,000 voters nationally. One of the key advantages of MRP is that it makes use of data across the country to predict levels of support in each individual state. If a poll does not have many respondents from rural Pennsylvania for example, MRP can learn from the behaviour of similar respondents in similar states to make an informed estimate of how rural Pennsylvanians might vote, while retaining specific state-level effects. The modelling inherent to MRP makes it more immune to wild swings in polling merely caused by who does and does not respond to a survey (known in the industry as non-response bias).
On the other hand, this introduces an element of synthetic-ness to the MRP data, so having traditional polling conducted alongside it, in which all respondents come from the state in question, can provide a good sense check of the results.
Turnout
We find that our registered voter samples are much more Democratic than our likely voter ones after our turnout model has been applied. It’s possible that our likely voter model is penalising the Democrats too much. Our turnout model is trained on both the 2020 and 2016 elections, and we take an average of the predicted turnout for each respondent based on the two models. If turnout looks more like 2020, Harris will fare better, whereas if it looks more like 2016, Trump will.
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 our US MRP, however, would not make a difference to the overall outcome and no states would change hands, so we have opted not to take the added risk of implementing one. Some results at the extremes (e.g. District of Columbia) may display smaller leads for the Democrats than we actually see.
Late swing
While confidence intervals are included in the data tables, please note that these are confidence intervals within each poll or MRP model itself. Our 95% confidence intervals give us an idea of the likely range of results the same poll/model would produce if we ran it again with a different set of respondents. A candidate having a lead outside the margin of error should therefore not be read as a certain victory. These intervals do not account for other sources of error, like the polls moving in the days before the election following the completion of our fieldwork.
Non-response bias
Although our polls are weighted by past vote (with an adjustment for false voting recall so we don’t end up with too many 2020 non-voters), it’s possible that the kinds of respondents for each party who answer our surveys differ significantly from those who do not. This is what’s known in the industry as ‘non-response bias’. For example, if people who have switched their vote from the last election are less likely to respond to surveys than party loyalists, our poll results may underestimate the degree of change since the last presidential election.
Trust levels / vaccination status
One theory as to the Trump polling misses in 2016 and 2020 is the high degree of correlation between the Trump base and distrust in institutions, which was tied very closely to COVID-19 vaccination status in 2020. One of the biggest predictors of Trump support in that election at the county level was the proportion of the population who had received the vaccine.
It stands to reason that people with distrust in institutions would be less likely to respond to surveys – especially when surveys so notably underestimated Trump at the last two elections, perhaps sowing greater distrust in the industry among the MAGA faithful. Although weighting by vaccination status did not alter the outcome of our polls or MRP to any significant degree, it’s possible that hard-to-reach groups will surprise the pollsters once again.