Right. It applies to exit polls, or to pre-election polls. The sample that you take has to be representative of the actual population. Picking any old 1,000 voters at random is not likely to match the actual electorate, so you'll have a biased sample if you do that. In years past, that meant pollsters would pick a certain number from certain precincts, based on previous years' voting patterns, so that they can get a good estimate.
My previous, specific, post was about the problem of using only election day exit polls for your projections. In years past, that was a small problem, because only 1 or 2 percent of the people voted absentee.
This year, there's a huge number who will vote early or absentee. 25 million have already voted, and the election is still more than two weeks away. It's easily possible that more than 1/3 of the votes will be cast before election day. There has already been 25 million votes cast, compared to 135 million total votes cast in 2016.
Democrats really are more likely to vote absentee/early/insert your state's method for voting before election day here. So, if you only poll election day voters, as in a traditional exit poll, your sample skews Republican. The pollsters have to take that into account somehow, but to do so they need estimates of how to find the people who voted the different methods. It's a hard problem.
And what keeps me up at night, figuratively for now, but by November 2 perhaps literally, is that I know that more mistakes are made with absentee votes than in person votes, which means more ballots will be rejected this year, and that sample will skew Democratic, even without any Republican cheating.
