Virtually all US polls in the 2016 Presidential Election predicted that Hillary Clinton would win—they were wrong. The same polls forecast a resounding Democratic win in the 2020 Presidential election, but instead, Biden barely won and the Democrats squeaked through with tiny majorities in both chambers of Congress. In the 2018 Midterm election, polls often predicted the wrong winner.
Why are polls frequently wrong? Will they be wrong again in 2022?
The public has a fundamental misconception about polling. Because polls give their results in specific figures, “this candidate has a 52% chance of winning over that opponent at 48%”, people think polling is an exact science. Nothing could be farther from the truth. What do polls try to do? To identify a sample of the population that is an exact reflection of those who will finally vote in an election, and to obtain from such a sample precise knowledge as to whom they will vote for.
An impossible equation
It is an impossible equation, particularly in a population that is shifting as broadly as is the US electorate, where in a Midterm Election only 40%-45% of eligible voters cast a ballot, so already identifying a sample that is a precise reflection of the final pool of voters is impossible, pollsters have to make estimates of the most likely sample. And even if they could define a theoretical sample very close to the future final result, they would have enormous difficulty to find eligible voter that would be the exact mirror image of the pre-defined sample.
Pollsters reach out to people by telephone, with at least obvious risks of bias, i) who pollsters are able to reach, that is who answers their phones, and ii) whether the answers given to the question are true. Pollsters constantly seek to measure where they have been wrong in the past and adjust to compensate for biases, but of course this remains a highly subjective, approximate business.
In fact, recent polling in the US has tended to improve in accuracy. For example, in one study, polls for the House Election in the last Midterm Election, in 2018, were accurate within 2.8 percentage points, a remarkable result and an improvement from the recent discrepancy estimated at 3.8% (Senate poll results were off by only 4.2%, as compared to 5.2% historically). So for races where the difference in candidates will be greater than 3%-4%, we can probably expect poll accuracy to be quite high.
But there is another, fundamental reason why polls are likely to continue to be wrong in predicting US election results: in contrast to most European countries, the US electorate is split into two virtually equal camps, Democrats and Republicans, and many elections are decided by differences that are tiny. Elections in the US are often decided by a tiny number of votes: the 2020 US Presidential Election was decided by votes in 4 critical states representing less than 0.0006% of total votes cast. In other words, polls would have to be 100 times more accurate than their current margin of error to reliably predict the results of US presidential elections, an impossible objective.
Given the structural approximation of the very process of polling, the nature of the US election system with most elections operating at the state level or congressional district level, rather than single national elections typical in Europe, and the virtual parity in the US electorate between Democrats and Republicans, US polls will continue to sometimes be wrong.