I've long told people since I was a senior in HS that looking at state polls should really only give a surface-level view of elections, especially something like the presidential election. With the possibility of another Trump presidency or Biden presidency just days away I thought I'd look over what they are saying.
The main problem I have with how people take these polls in is that they are focused on the 50 states and not the 3,141 counties and county equivalents that matter to the election. See the maps below for a comparison of 2016 from state to county breakdown, and even by precinct.
As you can see, it provides a far different outlook on the election. Even just looking at the estimated 152.666M registered voters, most polls that estimate state's opinions usually gather 600 to 1,200 likely/registered voters. See here for a collection of these polls; click "All *insert state* Polls" to see them once the page loads.
Breaking this down with a 99% Confidence Level (CL) (because I don't like using 95% CL) produces that almost all state voting polls will have a Margin of Error (MOE) of 3.5 to 5.5%. To achieve at least 3% MOE the range is 1,837 to 1,849 samples, 2% MOE being 4,097 to 4,160 samples, and 1% MOE lastly being 15,669 to 16,624 samples. I will attach an Excel Spreadsheet if anyone really wants to look below. So, doing some quick math, worst case scenario is that between 2-choice polls the polls might be off by 7 to 11%.
It boggles my mind that most people don't consider this.