Why pollsters once again struggled to measure the vote for Donald Trump

What goes through the minds of American voters when Donald Trump's name is on the ballot? This is the question that pollsters have been asking for eight years and the billionaire's first victory, which opinion polls were unable to anticipate in 2016. The day after Donald Trump's second election to the White House, the The question arises again, as the Republican's results exceed the level of voting intentions that the polls attributed to him until the day before the election.

At the national level, the average of polls calculated by the different media (the New York TimesFiveThirtyEight, The Silver Bulletin) gave Kamala Harris a lead of about one percentage point. But the provisional results, Wednesday, November 6, show that it was actually Donald Trump who won the popular vote, with a comfortable margin of around 5 million votes, or a gap of 4 to 4.5 percentage points. This is well above the average national gap since 1988 (2.3 points) and similar to the gap observed in 2020, when the media overestimated the popular vote for Joe Biden by 4 points.

At the level of pivotal states, the average of the 2024 polls underestimated the Trump vote in some of them and to varying degrees. Pennsylvania, where Kamala Harris long posted a slim lead of around 1 point, finally saw Donald Trump triumph by a margin of 3 points. Nevada displays a similar gap, since the Republican candidate maintains, at the time of writing, a lead of 4 points over his Democratic rival, while the polls compiled by the media gave a tie between the two candidates.

To a lesser extent, the two other swing states in the Midwest, Wisconsin and Michigan, show a substantial gap of 2 to 3 points between the majority of polls, which gave Harris a lead, and the provisional results, which show Trump winning. with several hundred thousand votes in advance. But these scenarios are observed frequently: since 2000, polls in undecided states generally “miss” the election result by 3.1 points on average. It was therefore relatively expected by specialists that 2024 would be no exception to this rule and that the polls would underestimate one or the other of the candidates by the same order of magnitude. In this regard, the differences observed in the pivotal Southern states of North Carolina, Georgia and Arizona (1-2 points) can therefore be considered particularly small.

However, the differences observed all point to a slight to moderate underestimation of the vote for Donald J. Trump. Which is not a first.

Trump voters respond less to polls

This is the third time in a row that pollsters have failed to correctly measure voting intentions in favor of Mr. Trump. In 2016 and 2020, opinion surveys suffered from the same defect in all the decisive states (with the exception of Nevada in 2016).

In 2020, to better represent the voters of the Republican candidate, the pollsters took into account the degree of education of the respondents, a variable which was not “predictive” of Republican or Democratic votes before 2016. They thus hoped to better represent the voters non-graduates, who had voted overwhelmingly for Trump four years earlier. In vain, since the differences were even greater than in 2016, without much impact on public opinion since Joe Biden, who was the favorite, still won the election.

The fact remains that pollsters still do not know with certainty why their surveys struggled to correctly anticipate the Trump vote in 2020 and 2016. Especially since during the mid-term elections in 2018 and 2022, the polls were relatively accurate.

The main hypothesis of specialists is that Donald Trump voters would have a lower propensity to respond to surveys when asked, what is called non-response bias. The institutes struggle to survey the least politically engaged voters, those who vote infrequently and irregularly. However, the data show that this less politicized electorate overwhelmingly supports the Republican candidate, especially during the presidential elections, since they hardly vote in the mid-term elections, which typically mobilize the most politicized Americans.

A smaller gap than in 2016 and 2020

And in fact, survey data shows that non-response bias does exist. In 2020, the New York Times had measured a 20% gap between the response rate of registered Democratic voters and that of Republican voters. Four years later, polls from the New York daily measured an equivalent differential (16%) among the white electorate, which made the chief political analyst of the Timesthe day before the election, that “polls could underestimate Mr. Trump once again”.

The changes in method made by pollsters this year to improve the precision of opinion surveys have probably made it possible to limit the extent of this underestimation, since it remains more limited in 2024 (2.75 points on average) than in 2020 (4.1 points) and in 2016 (3.6 points).