Why were the polls (especially national ones) so off?

[tagimaucia]

It isn't really talked about the way the rust belt polling miss in 2016 is, because all the quants pretty much predicted Obama would win and he did, which is a lot less interesting than an upset-- but the national polls were basically ~3 points off in 2012 on average.

Swing state polls tended to be a bit better, but they also tended to underestimate Obama by a point or two on average (I think, someone can fact check me on this if I'm wrong).

Has anyone done a deep dive into why the polls were so off, particularly national ones?  Can it mostly or entirely be explained by pollsters assuming that Obama couldn't duplicate 2008 turnout levels with black/minority voters, which he did?  Are there any other clear reasons?

[Redban]

Minus Rasmussen & Gallup, the polls in 2012 were accurate. I remember a headline that said, "The biggest winner in the 2012 election were the polls." Obama won by about 3.5%. Most polls had him ahead by 1-5%. We often forget that these polls have a margin-of-error; they are not meant to be razor precise.  If a poll with a MOE of 3.5% had Obama up by 1%, then the poll wasn't really wrong.

Now -- Rasmussen and Gallup were the big ones that got it wrong. Both of them had Romney ahead in their national polls, and Obama's 3.5% win was outside the margin of error for them.

 I don't know what happened with Gallup, but I know what happened with Rasmussen. Scott Rasmussen himself often went on Bill O'Reilly's show to explain. He assumed  that Obama wasn't going to duplicate the 2008 turnout, so his polls underweighted the Democrats in the methodology. A lot of conservatives latched on to Scott Rasmussen's thinking, leading to the claim that the mainstream polls were wrong and that Romney was going to win. This is what caused the "surprise loss" on election night.

After the election, some people mockingly asked, "Why did all those pollsters oversample the Democrats?"

[tagimaucia]

I'm not a statistician, but I don't think "margin of error" is really a concept to be applied to averages of polls.  It just represents a 95% confidence interval for each individual poll, doesn't it?

Obama won by 3.9% and even if you take away Rasmussen and Gallup, the average of the final polls included in the RCP average was 1.1%, which is a significantly bigger gap between the polls and the actual result than in 2016, 2008, or 2004.