When you take the results of the 2012 and 2016 elections, and model changes in Democratic vote share, you see the biggest individual-level predictor for vote switching was education; college-educated people swung toward Democrats and non-college-educated people swung toward Republicans. But, if you ask a battery of "racial resentment" questions — stuff like, "Do you think that there are a lot of white people who are having trouble finding a job because nonwhite people are getting them instead?" or, "Do you think that white people don't have enough influence in how this country is run?" — and then control for the propensity to answer those questions in a racially resentful way, education ceases to be the relevant variable: Non-college-educated white people with low levels of racial resentment trended towards us in 2016, and college-educated white people with high levels of racial resentments turned against us.
You can say, "Oh, you know, the way that political scientists measure racial resentment is a class marker because college-educated people know that they're not supposed to say politically incorrect things." But when you look at Trump's support in the Republican primary, it correlated pretty highly with, uh … racially charged … Google search words. So you had this politician who campaigned on an anti-immigrant and anti–political correctness platform. And then he won the votes of a large group of swing voters, and vote switching was highly correlated with various individual level measures of racial resentment — and, on a geographic level, was correlated with racist search terms. At some point, you have to be like, oh, actually, these people were motivated by racism. It's just an important fact of the world.