Manufacturing Employment Can’t Explain Donald Trump’s Win.
There are a number of narratives that developed pretty quickly after this year’s election. One in particular — that Trump did well by appealing to workers in districts wrecked by job offshoring — has received a good deal of attention, both during the run-up to the election and immediately after. For instance, Jed Kolko at FiveThirtyEight writes that “the swing toward Trump [relative to Romney in 2012] was much stronger in counties with a higher share of routine jobs,” like those in manufacturing. In this short post I’m going to argue that the relationship isn’t quite so simple.
What I’m interested in is less the things that explain Republican vote share in presidential elections, and more the things that explain why Trump did so much better in some places than Romney did. After all, in 2012 Obama won pretty convincingly, which means that Trump needed to rewrite the electoral map. And rewrite it he did. Here are the county-level percentage changes in Republican vote share from 2012 to 2016 (where red means more Trump-friendly and blue means less).
If you’re going to explain this election, you clearly have to explain what happened in the midwest — that is, the four states that Trump flipped from 2012. To do that, we need to…
Much better. We can pick out a few patterns here. Northern Pennsylvania shifts pretty heavily to Trump, as does the south-eastern third or so of Ohio, the northern part of Michigan, and the western half of Wisconsin.
What I’m going to do now is plot the distributions of four demographic variables that have all been purported to explain the Trump phenomenon. (All are scaled relative to the relevant national average.) The twist is, none of them are labeled. See if you can identify the one that best explains the map above.
Conventional wisdom pegs the Trump swing as being generated by the top-right map, which represents the share of each district employed in manufacturing (clockwise from top-left: income, manufacturing, % white, and education. Remember these are scaled to the national average. The midwest is Very White). A quick visual inspection can’t rule it out. And if we run a bare-bones regression estimating the effect of manufacturing employment on a district’s Romney-to-Trump swing, there is in fact a positive, significant effect.
What if we complicate this simple regression, though? I estimated a fuller version, controlling for a district’s racial makeup, median income, and its share of voters with a bachelor’s degree — in other words, the rest of the variables mapped above. In this model, the effect of manufacturing employment is still significantly related to Trumpiness — but in the opposite direction! In other words, after controlling for these other variables, districts with high levels of manufacturing employment shift less of their votes to Trump.
So what happened? The culprit is the education variable. It turns out that a district’s manufacturing employment and its education level are closely related, as this scatterplot illustrates (points scaled by number of district votes).
What the regression analysis is telling us, then, is that districts with lots of manufacturing employment aren’t on average shifting their votes to Trump; they just appear to be shifting their votes to Trump because in many cases these districts also have below-average education levels, and these districts absolutely are shifting votes to Trump. In fact, after accounting for the effect of education on vote change, districts with manufacturing employment are shifting less of their votes towards Trump. (I ran a hundred or so permutations of this model, adding and subtracting different covariates. For instance, if education is left out but whiteness and income are included, manufacturing is still positive. I feel pretty confident that education is responsible for the flipped sign on manufacturing.)
Let’s be clear about what this is and isn’t saying, because I can hear the social scientists screaming about the ecological fallacy through my computer. These results aren’t saying that individuals are voting for Trump because they’re uneducated, and they definitely aren’t saying that individuals are voting for Trump because they’re stupid. In fact, these data don’t let us say anything at all about why individual voters vote the way they do, because they aren’t composed of observations of individuals. Instead, they show that the county-level relationship between manufacturing employment and Trump voting is an artificial one: it doesn’t hold up once we control for other factors that correlate with employment type.