We “data people” have all been burnt by polling – and it hurts. In Britain, America and elsewhere, there have been enough high profile “misses” by the industry to give anyone cold feet about relying on polls as the sole metric for how an election might play out. On balance, however, polling has good form for calling most races right – and is becoming increasingly accurate.
If this comes as a surprise, it might be because humans often process negative memories (poll misses) more thoroughly than they do positive (accurate polls). Perhaps conscious that history may repeat itself, betting markets are ranking a Trump victory as more likely than what our model and the models of other outlets forecast.
In the US, the antipathy towards polling is heightened because they have a president most polls did not see coming. In 2016, survey data understated Trump’s margin against Hillary Clinton by between 3pts and 6pts in most key battleground states. Some of this miss could be attributed to margin of error, but for polls to have underestimated Trump in a uniform way across different states indicates something else was going on; and, in particular, that the pollsters struggled to correctly quantify insurgent movements.
The question for 2020 is: will it happen again? Polling this time has been much more stable. But intense uncertainty remains about their value and accuracy. Especially when it comes to whether there is a shy Trump vote.
Is 2020 the new 2016?
The cocktail of factors which led to an almost uniform under-representation of the Trump vote in 2016 is unlikely to be repeated this time round however.
One of the reasons for that miss four years ago was the assumption by pollsters that undecided voters would break evenly for both the Democratic and Republican candidates in the final days. History had shown that to be the conventional wisdom at the time; but in 2016, it was proven wrong.
Another error was to assume that, of those voters declaring themselves “undecided”, only those who had history of eventually turning out would actually make it to the polling booth. History, again, suggested this was a reasonable assumption. But 2016 featured an inordinately high number of undecided voters and those plumping for third party candidates.
In the final few weeks, third party supporters did not break for the two main candidates as much as expected (and had they done so, it would have been disproportionately for Clinton). Undecided voters likewise – who were in 2016 disproportionately white and less likely to have a college education than in a typical US election – broke overwhelmingly for Trump.
Why can we have confidence this won’t happen again? There are two reasons. The first is that we have next to no undecided voters (the latest YouGov has the number unsure about their vote at 6 per cent). The second is that we know now that a combination of race and educational attainment is a better indicator of voting probability than income. This helps pollsters model who and how likely people are to turn out to vote.
The undecided voters who do exist this time round are in demographic groups less likely to favour Trump. Though his base is more diverse than the punditry often credits (15 per cent of his vote in 2016 came from Latino and African American voters), he still trails badly among them.
Could there still be a shy Trump vote?
On the night of 7 May 2015, UK pundits, activists and voters alike watched agape as the British Conservative party emerged from the election campaign with 20 extra seats and an overall majority. Most hadn’t believed this a realistic prospect. Polling in the weeks and months prior had prime minister David Cameron’s Tories neck-and-neck with the opposition Labour party and understated the party’s lead by as much as 6pts. This miss led to the launch of an inquiry which concluded there existed in polls a systematic over-representation of Labour supporters and under-representation of Tory equivalents.
Though the inquiry did not offer a definitive answer as to why, a number of suggestions were floated both in the report and, in the aftermath, by UK academics. One which has stuck is the idea of a “shy Tory effect”: voting Conservative is somehow seen as less socially desirable, so poll respondents are less likely to admit that is what they are going to do. Might it be possible that America is experiencing a similar phenomenon, the shy Trump voter?
The problem with this argument however, is not only that “social desirability” is itself a slippery concept, but that there is a large question mark over whether it is a factor in Trump voters’ minds. This isn’t a slight on Republicans, but an acknowledgment of a bitterly divided electorate. Few backers of Biden and Trump have close friends in the other camp. A Pew survey conducted in August found just 3 per cent of supporters of both Trump and Biden report having a lot of friends who support the opposing candidate.
The question for the Shy Trump Voter theory is: why would they be shy? Why would they feel any stigma attached to voting for Trump, when they move almost exclusively in circles who share their preference?
Not too long ago this theory was put to the test. In a Morning Consult survey, Trump supporters were just as likely to profess support when asked by a real person in a telephone interview as when filling in an online survey. The evidence for a Shy Trump Voter is as yet unproven, erring on unlikely.
But isn’t Trump still an error margin from victory?
A piece I wrote on the eve of the first presidential debate claimed Trump was an error margin from victory – just as Biden was an error margin from a landslide. At the time, these assertions were correct, but since then the polling has moved somewhat.
Because a poll grabs only a sample of a population for its survey, it cannot be 100 per cent accurate. The margin of error in a typical US survey would range between two and three percentage points, up and down. What this means is that if a poll with an error margin of 3pts puts candidate Joe Bloggs on 52 per cent, and opponent Emmett Emmerdale on 48 per cent, we would reasonably assume the poll’s “true” (ie. representative of the population) value for Bloggs would range between 49 per cent and 55 per cent. Bloggs, while ahead of Emmerdale, would nonetheless still be an error margin away from defeat.
In 2016, Trump was, in most battleground states, behind Clinton but a mere error margin away from snatching victory. This time round, things are not quite so simple. Not only across the US but also in the swing states that matter, we have polls which give Biden a much clearer lead over Trump, and this lead has only intensified in recent weeks. According to NS model data seen on October 12, just eight swing states are now within the margin of error for Trump, and were Trump to win all of them, he’d still be marginally behind Biden on the electoral college.
Trump was a margin of error from victory. Now he’s not.
Could he come back? Maybe. Polling has been wrong enough times to emphasise uncertainty. But this time there are strong factors in its favour: the size and stability of the lead; the improvements in polling; the lack of evidence for shy Trump voters; the lack of undecided voters, and the fact those who exist are less likely, demographically, to break for Trump.
Our election model puts Biden’s chances of winning at 88 per cent – and for good reason.