After the shocking result of the 2016 election, election data science was on its heels.
The polls almost all showed Hillary winning. How did Trump pull it off?
Because the Electoral College. The polls weren’t wrong; they predicted a 3% Hillary national popular vote win and were only off by 2; they just missed a few key states.
The primary problem is relying on national polls to predict an election that isn’t nationwide. The Electoral College system means that states elect the president, not voters nationwide. Hillary Clinton won the nationwide vote, but she lost the election, primarily because she ran up her margins in diverse major urban areas but lost almost the entire rest of the country.
Florida was the first shocker of election night. But polls showed Trump with a slim lead in Florida. Trump just slightly outperformed it.
Virginia looked closer than it should have been as returns started rolling in, but it ended up almost exactly where polls predicted.
North Carolina, which President Obama won in 2008 but narrowly lost to Romney in 2012 looked like it might flip back blue early on. Again, Trump lead the final poll average there too, so its final red hue shouldn’t have been surprising.
The real question then isn’t why polls were so wrong; most of them weren’t. Why were Rust Belt polls so wrong?