The Geography of Despair

Deindustrialization is not the culprit

I recently discovered Justin R. Pierce’s work at the Fed. Together with his coauthors, he used proprietary microdata from the CDC to show that the China shock in 2001 had a systematic effect on deaths of despair. Specifically, they have shown that counties more exposed to import competition from China had greater increases in suicides and alcohol poisoning rates. Moreover, the effect was seemingly confined to non-Hispanic white men.

There is a risk that their results will be abused to further the misguised confrontation with China. So, it’s important to understand what their results actually mean; what the magnitude of the effect is; and whether we should expect more of the same.

As they note, they’re interested in identifying the effect of labor market disruption on people’s health and well-being. The China shock merely provides a tidy natural experiment; a sort of event study that allows one to identify the causal diagram—like that famous boatload of Cubans in Florida. As they acknowledge, this is not an estimate of the net impact of trade liberalization with China, simply because, like Autor et al., they’re merely adding up one side of the ledger. We too gained many jobs from our trade with China, perhaps millions. No one has yet attempted to carefully evaluate both sides of that equation. We won’t attempt to do that here. What we shall do is examine the broad geographic variation of deaths of despair in America, and try to understand its relationship with deindustrialization.

Their feature is a bit complicated. In any case, we don’t have access to their microdata. Instead, we obtain all-cause mortality and suicide rates by county from CDC Wonder. We obtain some standard covariates from Social Explorer. Although Social Explorer does have a dataset on business patterns, I have some doubts about the integrity of their data. (Results based on that dataset were a wash.) Instead, we obtain information on the number of industrial establishments operating in each county from the EPA’s Toxics Release Inventory. This allows us to estimate the number of plants closed for a large number of counties across all fifty states. We compute change in log number of establishments between 2000 and 2021, and between 1987 and 2000. The goal is understand the broad pattern and magnitude of the relationship between deindustrialization and despair. We restrict attention to non-Hispanic whites, both to prevent confounding by composition effects, and because, until quite recently, the phenomenon of deaths of despair was confined to working class whites.

The crude death rate, which we calculate as the population-weighted average of all-cause mortality rates across US counties, increased from 99 per million in 1999 to 105 per million in 2016. Moreover, as Case and Deaton have shown, college-educated whites continued to see gains in life expectancy—the dramatic increase in all-cause mortality was largely confined to working class whites.

Here’s the geography of this shift. High refers to the top 20% of counties by this feature; Low to the bottom 80%. Appalachia and the South stand out.

Of course, crude death rates are confounded by the changing age composition of the population. Using age-adjusted numbers, we obtain a cleaner signal.

In order to test the strength of the relationship with covariates, we proceed in a non-parametric fashion that is robust to non-linear patterns. Specifically, we first compute quintiles of the variables. Then we construct contingency tables for the top and bottom quintiles for every feature. We then compute odds ratios and carry out Fisher’s exact tests. What this tells us is whether counties that have, say, low incomes are disproportionately represented among counties that have high death rates. The odds ratio is easy to interpret and pretty standard, so everything that follows should be very clear.

Anyone familiar with Case and Deaton’s path-breaking work expects to see a strong class gradient. But I was not prepared to see these kinds of numbers. We find that counties in the bottom income quintile are six times as likely to have seen a large increase in age-adjusted mortality as the average county; while counties in the top income quintile are only a sixth as likely to have done so. This means that moving from the top to the bottom income quintile increases the odds of having seen a large increase in death rates by a factor of 40!

(As is standard practice, * = p < 0.10, ** = p < 0.05, and *** = p < 0.01. Here p refers to the p-value of Fisher’s exact test, which tests whether the odds ratio is significantly different from 1.)

The factor for college graduation rate is an astounding 84; meaning that counties with a low proportion of college graduates are eighty-four times as likely to have seen a large increase in the age-adjusted, all-cause mortality rate. The effects are very large and in the expected direction for median rent, population, and even population density.

For our main feature that tries to capture deindustrialization, we do not find very significant effects. The only statistically significant odds ratio is that for change in the number of industrial establishments in 1987-2000 (OR=1.63, p=0.02). In particular, the same for the period 2000-2021 is not even marginally significant (OR=1.21, p=0.43). So, Pierce’s China shock was not large enough to show up in the aggregate mortality numbers. The fact that our 1987-2000 feature is significant suggests instead that the neoliberal remaking of the economy, when industrial firms came to be disciplined powerfully by private equity firms and Wall St in general, was large and systematic enough to push the aggregate numbers around, although only a little bit.

It gets grimmer. We verify that the overall suicide rate for non-Hispanic whites has been going up. It has increased from 12.8 per 100,000 in 1999 to 17.7 by 2016, an increase of 38%.

Here’s the geography of suicides. It’s an all-American business. Most suicides are committed by whites using their own firearms.

In order to understand the socioeconomic correlates of suicide, as before, we look at the odds ratios. We find that density is the stronger correlate of suicides. Less densely-populated counties, likely rural, are three times as likely to have seen a large increase in suicides. College-education, a proxy for the upper classes, is also a strong correlate. Working class counties are more than twice as likely to have seen a large increase in the suicide rate.

We find an unexpected upside-down relationship with our proxy for deindustrialization. However, this is a statistical mirage; as an examination of the joint distribution shows. There is, in fact, little predictive information contained in the number of industrial establishments closing.

Here’s the change in suicide rates by state. Some surprises here. Utah, Colorado, New Hampshire and New Mexico have seen a greater increase in suicides than West Virginia!

This has been a grim undertaking. But we have made some progress. The preceding analysis confirms, first of all, that despair in America is strongly structured by the geography of class. What we’ve also shown is that, even if the causal effect of deindustrialization can be identified with sophisticated weapons, as Pierce has shown, the signal is simply not strong enough to shape broad patterns of despair in America.

This may be due to the fact that manufacturing employed less than a quarter of the civilian labor force even at its peak at midcentury. It had steadily fallen to 18% by 1980, 14% by 1990, and 12% by 2000, on the eve of the China shock; we’re now at 8%.

Were deindustrialization responsible for the despair, then the deaths of despair should’ve been rising through the 1970s, 1980s and 1990s, when the United States went through most of the deindustrializing. Why do we see little or no upward movement in working class mortality rates during those decades? Deaths of despair instead take off when deindustrialization begins to taper off. So, deindustrialization is unlikely to be the central culprit it is thought to be. What may be responsible is the broader hourglass economy, which comes together precisely when the structural break in deaths of despair occurs in the early-1990s, a decade before the China shock and many decades into the secular decline of manufacturing employment.

I fear that too many serious people are investing too much into re-industrialization as a strategy for solving the problem of working class despair—and ultimately, as a strategy of political stabilization. The decline in manufacturing per se almost certainly did not cause the despair, and it is quite unlikely to be the master solution so many are positing it to be. Meanwhile, our politicians seem hell-bent on wrecking relations with China, based on what appear to be poorly thought-through political and geopolitical calculations with a dash of even more intellectually-challenged mercantilism thrown in for good measure.