A reader asked Nate Silver if the polls will underestimate Trump.

He answered that we have no way of knowing ahead of time. Poll bias is not consistent across elections.

And this is important he added that the 2016 and 2020 biases aren't necessarily the same phenomenon. So it would be wrong to just assume 2016 and 2020 will repeat simply because they both underestimated Trump.

2016 was plausibly an issue of polls that over sampled college educated voters (2016 was the election when education became a large partisan shift with Trump gaining ground on non-college educated whites voters and losing ground with college educated whites).

2020 was plausibly an issue of the pandemic affecting response rates with Biden supporters more likely to respond to polls while staying at home.

As he noted, pollsters change their methodologies across elections. As just one example, weighting on education became much more important after 2016. Likewise online and text based polling has become much more common across pollsters.

As you can see, the 2022 Midterm polling was pretty good on bias. And pundits if anything got it horribly wrong assuming Republicans were being underestimated.

These are small sample sizes as elections are so infrequent. So don't be making strong statements off of 2016 and 2020.

If we do have polling error it is likely to be systemic and correlated such that many of the states are directionally similar. Of course there are notable exceptions each election. And wide variance in that systemic error (Wisconsin was a very large polling error in 2016 even though the national polling error was relatively normal).