Blog/Analysis

According to most accounts, Biden’s campaign is starting to falter, with Trump maintaining a slight edge in the national polls. But anyone tempted to write off Biden should do so with caution. Here at Focaldata, we’ve conducted our own polling in six key swing states – Michigan, Pennsylvania, Arizona, Georgia and Nevada. We’ve used this polling to dig into US-specific sampling trends and identify the likely source, and direction, of polling error come November 2024.  

Overall, we find that:

1. Trump has only a very small lead across the swing states.
2. Biden has a hidden advantage as the incumbent.
3. There’s a possibility of a polling error against the Democrats.
4. Key battleground issues are yet to be won.

The state of the swing states

We ran separate state polls in each of America’s key six swing states. The first group was Michigan, Pennsylvania and Wisconsin — the “rust belt” that swung for Trump in 2016, but went blue again for Biden. We also polled Arizona, Georgia and Nevada, the first two of which were edged by the Democrats in 2020, for the first time since Bill Clinton, while Nevada has been more reliably Democrat but remains extremely close.

On the surface, the results from our polls show that Trump has a lead in all six states. He is just 2 points ahead in Michigan, and as many as 9 in Georgia — but a lot of small-to-medium leads add up to a strong advantage overall.

Figure 1: Presidential Vote Intention, by state, all respondents

Source: Focaldata polling, n = 5004 in Michigan, Wisconsin, Pennsylvania, Georgia, Arizona, Nevada, fieldwork 17 - 23 Jan 2024

However, when we narrow things down to people who are likely to vote, things narrow significantly. Trump is still well ahead in Georgia — the closest state in 2020 — but his lead in the other states has shrunk. Biden now even edges ahead in Michigan.

Figure 2: Presidential Vote Intention, by state, 'likely' respondents

Source: Focaldata polling, n = 5004 in Michigan, Wisconsin, Pennsylvania, Georgia, Arizona, Nevada, fieldwork 17 - 23 Jan 2024Source: Focaldata, n = 5004, Fieldwork: 17 - 23 Jan 2024

Most obviously, this shift reflects Trump’s popularity among politically disengaged groups who are less likely to turn out to vote in November. It leaves the former president with the challenge of motivating these people to show up: can he recapture the enthusiasm of 2016?

Adding further uncertainty is the fact that a large percentage of respondents — 11-15%, depending on the state, which is enough to swing the election — favour another candidate, or say they “don’t know” how they will vote. This far out from November, that’s normal — but how will it change in the coming months?

To find out, we asked these people a follow-up question: “Although you do not intend to vote for Donald Trump or Joe Biden, which of the two do you lean towards, even if only a little?”. The results were striking.

Figure 3: Presidential Vote Intention, by state, 'likely' respondents - 2 way choice

Source: Focaldata polling, n = 5004 in Michigan, Wisconsin, Pennsylvania, Georgia, Arizona, Nevada, fieldwork 17 - 23 Jan 2024

When we combine these responses with the Trump/Biden choices of other likely voters, the election looks much closer. Trump’s lead is now 2 points or less in Wisconsin, Arizona and Nevada, and just 3 points in Georgia and Pennsylvania. We think that this is closest to the real “state of play” in each of the swing states: Trump is only marginally ahead.

Trump has the better map

Of course, the presidency is not only about voter percentages, — it’s about the electoral college. HiIlary Clinton and Al Gore both won the popular vote, but not the keys to the White House. And this time, the math is against the Democrats. In other words, Biden could easily have a very good campaign and still lose.

Some of the six states are worth more than others. Pennsylvania, in particular, is worth more electoral college votes than Arizona and Nevada combined. Given how very close Georgia was last time (Biden won by just 0.23%), we might well expect him to lose it this year. That, however, makes Pennsylvania the key swing state. Even if Biden wins Wisconsin, Nevada, Arizona, and Michigan, he would still lose (by 270 to 268).

Figure 4: Hypothetical Electoral College Map

This particular map, where Trump flips only Georgia and Pennsylvania, is fascinating. Trump would win by 270-268, but it’s close enough to be decided by the changes in electoral college vote allocation since 2020, when Biden would have won with this configuration. What’s more, Maine’s attempt to bar Trump from the ballot would become relevant: if that were upheld by the courts, Trump would lose one vote to Biden, tying the electoral college.

It’s extremely difficult to put a hard number on things, but taking the only-Biden-or-Trump figures above as a baseline, and accounting for expected polling error, we would say that Trump probably can count on a 90% win rate with the polls where they are. However, things are close enough that if there were to be just a 1.3% swing to the Democrats, Biden would be favourite.

The power of the incumbent

If history is our guide, then a 1.3% swing is hardly inconceivable. Since Reagan was inaugurated in 1981, incumbent presidents have put on an average of 1.75% from January to election day. Moreover, once you exclude the mad 1992 election, in which Ross Perot took nearly 19% of the vote, and incumbent G.H Bush lost 7% during those months, the average incumbency pickup is 3.75%.

Figure 5: Incumbent polling bounce, last 6 presidential terms  

Source: Gallup

Of course, this all depends on the scenario. Will Biden’s campaign follow Obama’s in 2012, where he broadly didn’t pick up any votes from February onwards? Or will it resemble Trump’s in 2020, when he actually lost votes towards election day? If the latter is the case, Trump 2024 is still the heavy favourite.  

Indeed, that 3.75% average is heavily tilted towards the massive incumbency gains made by Reagan in 1984 (7%), and Clinton in 1996 (9%), two of the greatest orators of the 20th century. Is that really replicable? When you step back, this election surely shares more in common with G.W. Bush’s in 2004, which resulted in a modest 2.5% pickup. And if that does happen, it could result in the closest electoral college map in history, with constitutional chaos bound to follow.

A very personal election

It’s going to be a defensive election for the Democrats. Biden, who represented a “fresh start” in 2020, is now feeling the costs of governing. The problem for Democrats is that the last election was very close — Biden won by less than a single percentage point in Arizona, Wisconsin and Georgia. And he can’t afford to lose them now.

That said, presidential vote choice now tends to be extremely static. Gone are the days when a large percentage of the electorate were genuinely prepared to vote for either of the two main parties. Over 90% of Republican and Democratic voters in 2016 voted for the same party in 2020. It suggests that this year’s election will be decided by turnout rather than switchers: just 3% of those we surveyed are intending to directly switch from Trump to Biden.

The question is: of those who voted Biden in 2020, how many are not planning to vote for him this time? We used a regression model — a way of teasing out using statistical analysis what drives voting behaviour without getting people’s explicit answers as people are not very good judges of their own motivations and impulses. The results were pretty striking.

Figure 6: Main predictors of Biden defection

The biggest factor behind defection from voting Democrat is, unsurprisingly, people’s approval of Biden as President. The next three factors are: belief that Republicans better handle the cost of living, high concern about Biden’s age, and low institutional trust. We know from polling over the weekend that concern about Biden’s age cuts across political camps: 73% of Democrats feel Biden is too old, a large share of the 86% who feel the same across the total population.

There are, however, also blockers that stop people from defecting: Biden’s decision to make climate change a top-three issue, and a high level of trust, as measured by Covid vaccination doses (more on this below).

There is a lot of collinearity here: those who rate Biden poorly are also very likely to think that Republicans will better handle the cost of living. But the overall message is that it’s a very personal election: people are not voting for Biden because they don’t think he’s done a good job on the cost of living. Reversing this is the key for the election.
And this is where things get tricky for Biden. For while his flagship policies involve investment in the economy, voters do not associate these policies with the cost of living. Half of respondents (51%) picked the cost of living among their top-three issues in our poll, almost 20% more than the next most important issues, immigration and healthcare. However, when respondents were prompted to pick a policy, they chose direct income — a higher minimum wage, or improving the Affordable Care Act — over industrial or green investment by a large margin.

Figure 7: Top voter policy priorities, by state

Source: Focaldata polling, n = 5004 in Michigan, Wisconsin, Pennsylvania, Georgia, Arizona, Nevada, fieldwork 17 - 23 Jan 2024

But before we get too bearish on Biden, it’s worth looking at the battleground maps. In any given election, looking at these quadrants is a staple of the war room. Each graph shows the salience of each issue — and which party is winning it. But as those below indicate, we are in a dog-fight of an election, and that’s probably going to remain the same right up to election day.  

Figure 8: Issue ownership by party, by state

Source: Focaldata polling, n = 5004 in Michigan, Wisconsin, Pennsylvania, Georgia, Arizona, Nevada, fieldwork 17 - 23 Jan 2024

Could we see a polling error?

For people like us who spend their day poring over political polling, the 2016 election night looms large in polling nightmares. The polling industry as a whole failed to pick up the extent of Trump’s support, and there was a large polling “miss” on a state level, as Trump unexpectedly won in a number of Democratic states — in particular, Michigan and Pennsylvania. Could that happen again?

For all the excitement, we don’t think such a big miss is likely. Since Trump and Biden have run against each other before, the overall pattern of their support is mostly a “known quantity”, unlike in 2016. There are, however, some signs that controlling for education — one of the key factors in the 2016 election surprise — may no longer be enough. This increases the possibility of a slight polling error — but this time against the Democrats.

Since at least 2000, the Democratic-Republican political divide has been moving away from income and “realigning” towards trust in the federal government and mainstream media. Poorer areas with high levels of suspicion towards the state have been shifting from Democratic to Republican, while wealthier and educated areas have become increasingly Democratic.

A key measure of this shift are Covid-19 vaccination rates. Many people in the US chose not to be vaccinated, or did not take at least one of the recommended doses. This choice was strongly linked to their trust in the government, media and scientists. The fact that vaccination data is available for relatively small geographical areas makes it an ideal tool to investigate how important “trust” is to politics.

Figure 9: Republican vote share, and swings from 2000 to 2016, plotted against vaccination rates

Source: CDC, MIT Election Lab

As the above graph reveals, trust is extremely important in understanding the shift in Republican support since 2000. The correlation between Trump’s vote in 2016 and the percentage who completed the initial course of two doses is a strong (at -0.62), as is the correlation to Republican gain between 2000 and 2016 (-0.46). Note this is the Republican vote share from 2016 — vaccination rates are used to track underlying social attitudes, rather than opinions relating to Trump’s handling of the pandemic. What vaccination rates give us is a granular variable that isn’t strictly demographic, but is in fact quite predictive of your cultural and values outlook. It helps us understand when a poll may be demographically balanced - i.e. have the right number of white people of 18-24, but not representative of other politically sensitive variables i.e. too many unvaccinated people or not enough two-dose Americans in a poll. When this happens you get a polling error. In the case of our own polling we found that panel data typically overestimated unvaccinated, and underestimated two-dose Americans - thereby potentially underestimating the Democrat vote share.  

Added to this factor is the possible existence of “non-response bias”, a result of the skew caused by certain kinds of people being more likely to answer surveys than others. In the US, the hardest to reach respondents tend to be the youngest (18-24) as well as those aged 65 and over. In surveys these groups are often represented by a particular subset who are more likely to answer surveys: namely, the 18-24s, and the most educated retirees.

This probably skews the results for those groups, making the youngest groups appear more Republican, and the oldest too Democrat. We think this is behind many polls showing Trump leading Biden among voters aged under 30 — something Sabato’s Crystal Ball has noticed. Although Trump has probably gained among this group since 2020, it would be very surprising if he had gone as far as to overturn Biden’s lead. We think that the age cross breaks of US polling are deeply vulnerable too polling error and data quality. We are sceptical that the age curve of the US has flattened to the degree the polls suggest.  

A similar effect, in reverse, may also be leading pollsters to overestimate Democratic support among the over-65s. To make matters thornier, this skew can interact with the vaccination angle. In our raw sample, unvaccinated 18-24s are overrepresented: just half have received two doses, a figure that should be closer to two-thirds. With fully vaccinated young Democrats being less represented in polls, and more authoritarian, anti-trust unvaccinated young adults - it is clear that the young are very likely to be less pro-Trump than polls suggest, but older voters marginally more supportive of Trump than public polls.

Figure 10: Vaccination weights by age group, unweighted sample

Source: Focaldata polling, n = 5004 in Michigan, Wisconsin, Pennsylvania, Georgia, Arizona, Nevada, fieldwork 17 - 23 Jan 2024

Another aspect of our polling we’ve been looking at is “white identity” — or specifically, the belief that there is discrimination against white Americans. In this survey, we measured this by asking: “In the USA today, how likely is it that many whites are unable to find a job because employers are hiring minorities instead?” The most common response was “Not at all likely” — but responding positively to any extent is a strong predictor of voting for Trump, both in 2016 and 2020.

Figure 11: Presidential vote intention by white identity, 'likely' voters

Source: Focaldata polling, n = 5004 in Michigan, Wisconsin, Pennsylvania, Georgia, Arizona, Nevada, fieldwork 17 - 23 Jan 2024

What’s really interesting here, however, is that polls may be overestimating the percentage of people — particularly white people — who are hostile to affirmative action, and who believe their job prospects are hindered by minorities. If that’s the case, then we might expect a very different kind of polling error this cycle. When we look at the incidence in our survey compared to those in the general population (as measured by the ANES 2020 survey), we are over-sampling those who answer “extremely” and “very likely” (and particularly so in Pennsylvania).

Figure 12: White identity, unweighted sample vs ANES 2020

‍Source: Focaldata polling, n = 5004 in Michigan, Wisconsin, Pennsylvania, Georgia, Arizona, Nevada, fieldwork 17 - 23 Jan 2024

Sample

Focaldata online polling of 5,004 respondents in Michigan, Wisconsin, Pennsylvania, Georgia, Arizona, Nevada. Fieldwork was carried out between 17/01/2024 and 23/01/2024. Data is weighted by age, gender, education, race, Hispanic identity and state-level vaccination rates

Enjoyed this week's newsletter but haven't subscribed yet? Why not subscribe and get the next one straight to your inbox. You can also follow Focaldata on Twitter and LinkedIn.

Feel free to check out our previous editions of Bi_Focal:

• Our two-part edition on the rise of Reform, and the shape of public attitudes on social and economic issues, available here (Part I) and here (Part II)
• A new theory of "elastic seats", a defining feature of pivotal areas with large variation in voter support over time - link here
• Our MRP of the Voice referendum in Australia last October (we were one of the closest pollsters to the final vote tally) - link here
• Our (now year old) stress test of predictions about a Labour landslide -- still valid a year on, we think! -- link here

Finally, if you or anyone you know regularly run polling, we'd be happy to run a free couple of questions for you -- we'd love to demonstrate the speed of our platform and showcase the technical capacity of our research team.

For the above, or if you have any thoughts or feedback, please hit reply or email me at james@focaldata.com.