The election error in 2016 is harder to explain as 538 which was a notably outlier that was quite explicit about their modeling and the reasons for it. There are lots of reasons to think that state poll errors would be correlated - not just because of similarities in sample composition across state lines but also for the simple reason that even recent polls are retrospective and don't capture last minute swings.
It could be as simple as the fact that it's not easy to do what Nate Silver did, and it can become a rabbit hole of assumptions. I know from professional experience that it's disconcertingly common for decision-makers to eschew "fancy models" if they seem too hard, and just hope that whatever source of complication they're not tackling will prove to not be material.
