Setting aside for the moment that the national-level polls were largely correct in 2016 (the final RealClearPolitics average had Clinton up by 3.2 points, her actual popular vote victory was 2.1 points), it is certainly true that some state-level polls portrayed a wildly incorrect picture of how the “blue wall” states would break. Trump outperformed his polls in Pennsylvania by three points, Michigan by about four points, Ohio by about five, and most egregiously in Wisconsin by seven. While many states had polling that was right on the money (Virginia and Florida come to mind), those Midwest states blew up the predictive models.
The American Association for Public Opinion Research investigated the causes of the polling error, and one key problem was polls in 2016 overrepresenting college-educated voters. Given Trump’s strength with non-college-educated voters, polls that didn’t include enough of them overstated the Clinton lead.
As a result, polls in 2020 should adjust their samples for education level. Many today do, such as the New York Times Upshot/Siena College polls that showed Biden with double-digit leads in Michigan, Wisconsin, and Pennsylvania.
But not all polls have made that adjustment, or they are still grappling with how to do so.