Well, there we are. With hundreds of thousands of votes left to process, leaving a number of states hanging in the balance, we still do not yet have a final result as the US approaches midday on Thursday 5 November. Joe Biden, nonetheless, is clearly the favourite, and is likely in the next few hours or days to pass the critical 270 electoral votes needed to win. The race appears to be a victory for the Democratic challenger (dubious-sounding legal challenges notwithstanding).
So where does this leave the pre-election polls and forecast models. Both predicted a Biden win. But clearly we have a race that has come down to the margins. In the states that will likely win it for the Democrats, the margin of victory is markedly tight; tighter than what the polling was saying just a few short days ago. Why?
***
Firstly, let’s be clear on the current situation. We are still waiting on a lot of ballots from a lot of states – a disproportionate number of which are more likely to be for Biden. In most cases the margins we see now are unlikely to significantly change, but it should still be noted that the figures below may, in ten days time, be a percentage point or two out of place.
Let’s start with Wisconsin, a state that has no more votes left to count. Joe Biden won it by 20,000 ballots – a lead over Trump of over half a percentage point. But the final polls, which the New Statesman’s election model was in part reliant on, had Biden with a lead of seven points. This translates to an error comparable to that seen in the state in the 2016 election.
Cross over to Michigan and the story is similar, if not quite as stark. Biden is currently winning the state with a 2.9pt lead; the polls suggested he win by, again, seven points. The error? 4.1pts.
Turn south to Ohio and Iowa. Polls said Trump would win both states by 2pts. He is so far winning them by 8pts.
And let’s not forget Florida! Polls had Biden on course to sneak the state with a lead of 3.4pts, but it was Trump who edged into the lead by 3.4pts. That leaves the pollsters with an error of 6.8pts.
Not ideal, but where did the pollsters do well? Clear nailed on successes are, it must be said, few and far between. In Arizona, Biden’s current lead of 2.4pts matches up well to the projected poll lead of 2.8pts. In Minnesota and New Hampshire, polls missed out by one percentage point. In Texas, however, polls missed out by 4pts.
While the error is not of a uniform magnitude, or on the scale of 2016, it is, nonetheless in a uniform direction. Of the states I’ve looked at, not one with substantial results has seen Joe Biden outperform his poll margin. In all states, Trump’s performance has been notably underestimated.
Why? It’s currently too early to give a definitive reason, but there are ideas worthy of consideration.
First, though, it needs to be understood that polling is not a precise predictor. It is a science, with a range of uncertainty. If a typical poll with a respectable sample says 50 per cent of people intend to vote for a candidate, the reality is that candidate is on course to get between 47 per cent and 53 per cent – a spread of six percentage points. There is always a margin of error, and yet readers still tend to think when a result touches the upper or lower level that the poll was somehow wrong.
To get a decent poll, the sample of people you gather needs to be representative of those who’ll eventually and actually turn out to the ballot booth. Almost everyone in a poll says they’ll vote, but that very rarely materialises. You, as a pollster, have to weigh responses down to a realistic level, but you need to know when and who to adjust.
Comparing an analysis of pre-election polls to on-the-day exit polling suggests pollsters struggled with regard to three vital groups: white women and, more generally, whites without a college degree; and America’s Hispanic population.
White women were a group throughout this campaign polls billed as very likely to swing substantially to Biden. Trump’s base – white America’s blue-collar populace – were also expected to swing big. What exit polls we can glean suggest that has not yet materialised, and the gains for Biden are much muted when compared to what the polls were pushing.
Hispanics – while not a homogenous group easily painted with broad-brush stroke statements – are nonetheless as a whole a historically hard group to poll. Disproportionate levels of political disengagement, and the added issue of a language barrier for some, means pollsters struggle to get representative samples. What polls we had in the run-up to the election saw muted increases in Trump’s support among Hispanics, but not anything like the gains we went on to see. In Florida, the swing among Hispanics to Trump was the largest in the country, and not something, I believe, we’ve seen in a presidential election before. Trump now ties with Biden on Hispanic support in the state. I’d wager it’s not a coincidence that Florida also, so far, has the biggest poll miss in the country this election.
[See also: Hispanics in Florida shift to Trump on US election night]
You can understand all this but still wonder how the polls could have missed as much as they did. One theory comes down to the pollsters themselves: herding. Like any other business, the US polling industry needs clients, and needs to appear “right”. If you were a pollster, it’d appear better to be among the pack, closer to the average of polls, rather than on the periphery.
Another theory rests not with pollsters, but voters. Voters inclined to voting Trump, or those that have been generally sampled poorly, are more likely to have low social trust, and less inclination to engage with others – including pollsters. “Social trust” is, as the Pew Research Center once wrote, a measure of a voter’s “faith in people” and a belief in another person’s integrity and reliability. Cynical voters are likely to be cynical in their dealings with pollsters, too.
Hispanics and those in non-professional social classes typically have high levels of low social trust, as do those who have no record in voting. In an election where turnout is reported to have been the highest in a century, with campaigns aimed at turning out the disengaged, it is understandable that pollsters may have struggled to sample or weight for those newly-engaged electors. In some instances, that may mean a newly-engaged section of a demographic whose historic voting patterns, simply put, do not exist. In that case, relying on the 2016 patterns is all you can do. This may in part explain why pollsters, to a degree, may have missampled Hispanics.
At any rate, we just don’t have enough evidence at this moment to definitively say why the polls were out. Newly-engaged Trump supporters may be a slightly harder demographic to sample than first thought, owing to being more likely to have high levels of low social trust, and therefore an unwillingness to engage with market researchers. The new voters who turned out for the current incumbent are not, also, as homogenous as first thought. Politically, they’re just as Trumpian as any of his other supporters; demographically they’re more diverse than ever. The old assumptions and weightings applied about where a candidate gets their support from – and the propensity for that support to turn out on the day – may warrant a rethink.
Trump is certain to have lost this election, but the legacy he will leave is a campaign of the insurgents, of breaking age old assumptions about who votes and what do they vote; that campaign, often understated in polls, may just hang about for a little while longer.
[See also: Five charts that tell us how America voted]
Content from our partners
Related
