By Ezra Klein

[...]

Back in 2016, Harry Enten, then at FiveThirtyEight, calculated the final polling error in every presidential election between 1968 and 2012. On average, the polls missed by two percentage points. In 2016, an American Association for Public Opinion Research postmortem found that the average error of the national polls was 2.2 points, but the polls of individual states were off by 5.1 points. In 2020, the national polls were off by 4.5 points and the state-level polls missed, again, by 5.1 points.

You could imagine a world in which these errors are random and cancel one another out. Perhaps Donald Trump’s support is undercounted by three points in Michigan but overcounted by three points in Wisconsin. But errors often systematically favor one candidate or the other. In both 2016 and 2020, for instance, state-level polls tended to undercount Trump supporters. The polls overestimated Hillary Clinton’s margin by three points in 2016 and Joe Biden’s margin by 4.3 points in 2020.