Does the China Shock Explain the Trump Swing?
'It was in the counties where the highest number of jobs were lost because of the China shock,' Adam Tooze writes in the current edition of the London Review of Books, 'that Trump scored best in the 2016 election.' What Tooze seems to have thrown his immense intellectual weight behind is a theory of US politics that traces the white working class revolt that put Trump in the White House to the China shock. 'Thanks to the painstaking work of labour economists we can trace, county by county,' he writes, 'the impact of Chinese imports and the loss of factory jobs across the US.'
Actually, the economists concerned, Autor, Hanson and Dorn, did not work at the county level. Instead, they develop their measure of exposure to Chinese competition at the level of multi-county commuting zones. The logic of their argument is sound. Extraordinary productivity growth in China — the result of Stalinism with Chinese characteristics — had a dramatic impact on many American industries. Regional economies reliant on these industries suffered job losses in manufacturing and wage declines in non-manufacturing; as they have shown.
The Toozian hypothesis that the China shock put Trump in the White House is testable. Merely testing it against noise is not persuasive however. We must test it against an alternate hypothesis. We have previously argued that what put Trump in the White House is the class-partisan realignment — a logic endogenous to US politics. The realignment can be read off the following graph.
Here we test these hypotheses against each other. We obtain the Autor-Hanson-Dorn measure of exposure to import competition from China from Dorn's website. We obtain overdose deaths from the USDA and electoral data from MIT election lab by county. We then aggregate the data by commuting zones. Our main explanandum, the response variable, is the swing to Trump — the difference in GOP vote share between 2012 and 2016. We also look at overdose deaths as the response. Throughout, we estimate robust linear regression by IRLS using the Andrew Wave norm. We standardize both the response and the features using sklearn's RobustScaler function. The slope coefficients can therefore be thought of as elasticities or betas.
We begin by replicating the main Autor-Hanson-Dorn result that exposure to China explains the spatial variation in the decline in manufacturing employment across the 722 US commuter zones. We find a large and significant elasticity of -0.50 meaning that a one standard deviation change in exposure to Chinese imports predicts half a standard deviation decline in manufacturing employment. This is in line with the results presented in their AER paper.
We then jump straight into the simple linear models. The simple linear models allow us to identify the total effect of the features on the Trump swing. We find modest but statistically significant total fixed effects for the China shock (b=0.11) and change in manufacturing employment (b=-0.13). We also find that the total fixed effects of the class-partisan realignment (b=-0.38) and overdose death rates (b=0.40) are four times as large. Straight off the bat, then, we can see that our features give us a much stronger handle over the Trump swing than the features of the Toozian hypothesis.
Another way to examine the same one-on-one relationships is to examine the means of the features by Trump swing quintiles. The steeper and more monotonic the gradient, the more confidence we can have in the feature.
We can see that the gradient is much more robust for college graduation rate and overdose death rates compared to decline in manufacturing employment or the China shock.
The strength of the one-on-one relationships only take us so far, however. In order to identify the underlying causal diagram, we must control for class-partisan realignment. Table 3 reports the estimates. We can see that the fixed effect of the China shock, already pretty small, is attenuated even more once we control for college graduation rate. Same for decline in manufacturing employment. While both remain significant at the 5 percent level, the elasticities fall to around 0.06. The Toozian hypothesis is thus revealed as a second order correction to our model.
Our second feature, overdose deaths, is potentially causally downstream from the China shock. That is, overdose deaths may be a mediator between the China shock/manufacturing decline and the Trump swing. Is it? Table 4 shows our regression estimates with overdose death rates as the response and the China shock and change in manufacturing employment as the features. The pattern is consistent with the interpretation that the effect of the China shock on overdose deaths is significant (b=0.23) and mediated by decline in manufacturing employment (since it falls into insignificance once we control for the latter).
Since overdose deaths are a strong predictor of the Trump swing and the China shock propagated to overdose deaths, we have a clear causal channel for the Toozian hypothesis. Table 5 documents that overdose is an effective mediator between the China shock/decline in manufacturing and the Trump swing (since it "kills" their coefficients).
We have thus shown that Toozian causal vector between the China shock and the Trump swing is a second order correction to the class-partisan realignment. Moreover, we have seen that decline in manufacturing is a mediator between China and Trump, and overdose deaths (and likely deaths of despair more generally) is a mediator between decline in manufacturing employment and the swing to Trump. These facts are also clear from the kitchen sink regression.
What put Trump in the White House was thus not the exogenous shock of the rise of China as a global manufacturing superpower but a logic endogenous to American politics.
Postscript. Turns out, we were not class-reductionist enough. If we restrict attention to blue collar manufacturing employment — not the change but the level — we get another way to identify the working-class. The change in manufacturing employment, whether due to China, Mexico or automation, still doesn't give us a good handle on the Trump swing. But the size of the manufacturing working-class turns out to be the most important predictor of the Trump swing — even bigger than college graduation rate and overdose death rates. Note that all these variables contain a very strong class signal: recall that deaths of despair, overdose deaths in particular, are entirely a blue collar phenomena in the sense that it is confined to whites without college degrees.
The American industrial working-class resides in commuter zones exposed to China.
Our selected model suggests that 35 percent of the explained variation is explained by college graduation rate, 23 percent by overdose deaths, 42 percent by the size of the manufacturing working-class. Of course, 100 percent of the explained variation is explained by the class-party sorting documented in the very first graph of this essay; reproduced below.