Some of the close states also did not have very many (or any) polls in the last week before the election. Since states could shift by more or less than national polling shifts, the lack of state polling meant that estimates for the states in question had to come from national polling shifts since the last (stale) state polls, resulting in greater potential polling error.
Their pre-election writeups also did specifically call out that they believed that the potential polling error was large enough to have a significant chance of changing the election result from what would be expected if the polls were accurate. Their last pre-election poll model had Clinton +3.6, while the actual election result was Clinton +1.7. This 1.9% polling error was somewhat greater than in 2012, where their polling model suggested Obama +2.5 but the actual election result was Obama +3.9, a polling error of 1.4%. They also had pre-election writeups specifically on the possibility of a popular versus electoral vote split, specifically indicating that Clinton would be more at risk of that than Trump would be.