Class-Partisan Polarization Intensified in the 2020 Cycle
OK, this may come as a surprise. Trump's vote share increased between 2016 and 2020 in 2,252 or 72.4 percent of the 3,110 US counties for which we have data. On average, he gained 1.41 percent in vote share by county. Weighted by population, of course, he lost 0.97 percent on average. But still, incredibly, there was a swing towards the GOP in nearly three-quarters of US counties.
The cross-section of electoral results by US counties contains information on the social basis of Trumpism as a mass society phenomena. There are two major readings of the central problem of mass society. Both share a common diagnosis. This diagnosis says that, in a modern mass society, individuals are dis-embedded from traditional institutions (kinship and religious groups etc), and atomized and isolated from each other. The atomized individuals of a modern society constitute a mass of freely-floating elementary particles. In the conservative reading, this atomization exposes the masses to seduction by demagogues who ride the rage of discontented masses into power. This "threat from below" is the lens I found in Eric Hoffer, which is the first frame I reached for to understand Trumpism.
What happens in a mass-movement is that differentiated individuals are marshalled by a demagogue into an undifferentiated, solid mass of identical particles, that is then deployed as a bludgeon to demolish a decaying political order.
Policy Tensor. "Riding the Great White Beast." May 2016.
The second, liberal reading of the problem of mass society holds that this mass of elementary particles is easily manipulated by the elites. As Richard F. Hamilton put it in Class and Politics in the United States, the theory says that 'the elites, through their use of the mass media, succeed in dominating even the consciousness of the masses' (1972, p. 47). This is the frame illuminated by Chomsky and Herman's Manufacturing Consent: The Political Economy of the Mass Media (1988).
Both of these readings are problematic. Both assume that the masses are gullible and easily controlled, either by a demagogue or by clever elites. Neither formulation works because, in fact, it is not the masses but the elites who are the most indoctrinated by the mass media.
It is the upper-middle class that is "most vulnerable" (or to put it another way, "most available") to the most explicitly political of the mass media. It seems likely, despite their high levels of educational attainment, that they are the most seriously "manipulated" of any group in society.
Richard F. Hamilton. Class and Politics in the United States, 1972, p. 47.
This upside-down differential exposure to indoctrination is largely due to the fact that the political content of the media consumed rises with status. Take the harebrained scheme to defund the police. Whether you thought it a good idea was essentially a function of your exposure to the prestige media, itself a function of status, which is in turn a function of the prestige of the university you attended. Anyone outside the prestige school-prestige media professional class echo chamber, and especially the two-thirds of the populace that did not go to university, was quite unimpressed. It takes a college education to unlearn common sense realism.
The role of the mass media is more restricted than we imagine. On issues where the media discourse can be checked against everyday reality, propaganda efforts by the media usually fail. But on issues remote from everyday life, such as central banking or foreign policy, the masses, as it were, defer to the elites largely because they have no choice. Sometimes the tension between the logics of technocracy and democracy can come to a boil. This is basically what happened over COVID-19 restrictions, police-minority relations, and the Trump White House. The elites were convinced that draconian measures were necessary to protect society from the pandemic; that American police departments were systemically racist; and that Trump was the most incompetent president in living memory, if indeed not a threat to American democracy. The masses were less sold on the necessity of social restrictions to deal with the pandemic; they were less convinced that blue collar police officers were racist; and given that job opportunities and wages were growing late in the cycle, they were less convinced that Trump had performed poorly.
So despite the all-out media offensive against Trump, despite the unity of interests that came together behind Biden, despite the unprecedented sums spent by the Democrats, despite the toll taken by the pandemic on his watch, Donald Trump came within a whisker of being reelected. What was revealed in 2020, was the deadlock of trench warfare in the class-partisan confrontation.
As I examine the numbers now, it has become increasingly clear that the class-partisan confrontation has not abated. It has not been tempered. No, sir. It has intensified. In order to document this fact, I look at the gradients of the change in Trump's vote share between 2020 and 2016 in the cross-section of US counties. Note that class-partisan polarization was already unprecedented in 2016. Since our response is the change in Trump's vote share, we are looking at the correlates of additional class polarization that has obtained over the past four years.
In order to compare apples-to-apples, we compute Cohen's d, a measure of effect size, for various features. Specifically, we discretize all features into quartiles (four equal bins) and then we ask: What is the difference in our response, change in Trump's vote share, if we condition on the bottom quartile of counties by a feature compared to the top quartile of counties by the same feature.
There's more graphs to come. But all the results I want to establish can be read off this summary graph of effect sizes. The strongest conditioner of the swing to Trump in 2020 was the share of county population without a college degree (d = +0.813). Interestingly, the share of whites without a college degree is a significantly weaker predictor of the Trump swing (d = +0.672). The third strongest feature is the share of the black population (d = -0.668) and the fourth is an ordinal variable, the rural-urban continuum, that measures effective distance from the Metropolis (d = 0.620). Finally, the fifth strongest conditioner is median household income (d = -0.578).
So the big story is that class-partisan polarization has intensified. The rotation of the working class away from the Democrats to the Republicans, that became glaringly obvious in 2016, has continued unabated. As the Journal noted recently:
Since the Reagan era, the two parties have essentially traded their core supporters. Democrats, the party of union halls and working-class America, now include much of the nation’s professional class. Republicans, once the party of Americans with four-year college degrees, are increasingly the party of the white working-class—and, the 2020 election showed, of a larger minority of Hispanic Americans. Counties that flipped from one party to the other in the 2020 election showed a clear divide between higher-growth, white-collar parts of America and a slower-growth set of communities more reliant on blue-collar jobs.
Wall Street Journal. "How the 2020 Election Deepened America’s White-Collar/Blue-Collar Split." Nov 24, 2020.
Documenting effect sizes is not enough, of course. It is possible that some of these estimates are driven by a small number of outliers. The best way to examine and check the signal with your eyes is to look at scatter plots one feature at a time. We show the major ones separately and the rest together in a panel.
We begin with polarization by educational class. Frankly, I was expecting this gradient to turn negative, mostly because Ruy Teixeira convinced me that pre-election polling reliably showed that Biden had made great gains among non-college whites. That turns out to have been badly mistaken. Working class counties swung to Trump even more in 2020, implying that the working class a whole was not affected by the relentless media propaganda. If anything, it seems to have pissed them off even more.
Interestingly, the income class gradient is nearly as robust. The social basis of Trumpism can be identified not only by education but also income. That should silence any challenges on the question of class proxies.
Where it gets really interesting is that the Trump coalition expanded beyond the white working class. Although the effect size of Hispanic share of the population is small (d = 0.030), that it is positive at all goes against all professional class antiracist expectations. Many, many counties with a predominantly Hispanic population swung hard towards Trump. This is a case where it has been claimed that the swing is an artifact of a few outliers in Florida and Texas. Not so. The mean swing to Trump in the 149 counties with a Hispanic share of at least 40 percent was +5.2 percent, compared to aforementioned mean swing of +1.4 percent across all US counties. It is a red herring to claim, as Shor and others have done, that the reason is that Democrats didn't do enough canvassing in southern Texas and Florida. The truth is that, as a multi-ethnic linguistic community, Hispanics do not like being called people of color. It is the thick antiracist discourse of professional class Democrats that pissed them off.
Similarly, it should come as a surprise that the gradient of the nonwhite population is positive, even if it is basically zero. Everyone said this was a confrontation over race. It wasn't — it was a confrontation over class.
Here are the rest of the scatter plots.
What is striking is that the pre-election polls were not just wrong about the blowout. They were also wrong about the details. The working class did not rotate back "home." It continued to rotate away to the GOP. It can no longer be denied that the GOP is the party of the working class and the Democratic Party is the party of the educated and well-heeled. This gives the Democrats an unprecedented advantage in firepower that can be expected to last. But it undercuts, completely, any pretense that the Democrats represent the working class of America.
Why has this obtained? The answer is simple and frustrating that for that very reason. The working class has further abandoned the Democrats because of relentless class work. One thinks here of the BLM protestor who yelled at the blue collar police officers: "Half of you don't even have a college education." More recently, as Matt Taibbi noted, the class contempt was manifest in David Atkins tweet storm, which began with the problem of "de-programming 75 million people."
Sometimes it feels like the problem of elite-mass relations in the United States has no solution. For the only solution, the tempering of class work, seems to be well beyond the capacity of the professional class — it is simply too busy with self-congratulation. But if we can't solve this problem then what hope is there of taking on the physical challenges facing American civilization?
Postscript. It is always possible that the method you're using fails for reasons you don't know. So it is always good to use multiple identification strategies to get at the real pattern underlying the data.
I was alerted to the possibility that the results could be an artifact of not weighting by population (or Democrats in denial could clutch at this straw). It is also possible that our method of computing Cohen's d is, for some unknown reason, not kosher. So I computed robustly standardized slope coefficients (betas) by first recentering and rescaling both the response and the features with the scale parameter estimated from the interquartile range, and then estimating OLS slopes. The following graph reports the results. The horizontal bars are standard errors.
Table 1 reports the results in detail. The population-weighted and unweighted betas are mostly close. There are some differences. And they reinforce the results reported earlier. The nonwhite share of the population is much more strongly associated with the Trump swing once we weight by population. The noncollege share is slightly less correlated with the swing, but the no-high school share is more correlated. Weighting by population thus increases the weight of the underclass in Trump's coalition relative to the working class. Interestingly, weighting by population increases the correlation between the Trump swing and the Hispanic share. So both the main results are reinforced: class-partisan polarization increased in the 2020 cycle and the Trump coalition expanded into racial minorities. I prefer the unweighted estimates because I think that variation contains a stronger class signal than when we weight by population. In any case, realism demands that we weight by electoral college votes rather than population. If we were to do that, the slopes would fall somewhere between the unweighted and population-weighted estimates. It's not going to change the headline results.
Post-postscript. It has been claimed that then additional class-partisan polarization documented above is confounded by geography. Specifically, that the working class swing towards Trump in 2020 was instead a rural story. This alternate hypothesis is easily tested. All we need to do is stratify by the Rural-Urban Continuum, a variable that combines population size with distance from the Metropolis. Stratifying by this variable reveals that the class gradients are not that different for metros, burbs, small towns, and rural areas.
More precisely, we estimate hierarchical regressions. We admit fixed effects for percent of the county population that is female, percent of county population that is black, and change in the vote share of the third party — these are our controls. We also admit random effects stratified by the Rural-Urban Continuum, and model them as linear functions of class proxies.
Using median household income as our proxy for class, we find some variation in gradients in different kinds of regions. But the slopes are negative everywhere.
Perhaps a better way to visualize the gradients is to graph them in a bar plot. Large suburbs (urban population above 20,000, adjacent to a metro) display the largest gradient (b = -0.64). Rural counties (urban population smaller than 2,500) next to metros display the smallest gradient (b = -0.20).
Educational attainment contains a stronger class signal than household income. Counties with low shares of the college-educated are more certainly predominantly working class than counties with low household incomes. Conversely, counties with high shares of the college-educated are more certainly predominantly professional class than counties with high household incomes.
Using percent of the county population without a college degree as our class proxy, we find larger gradients. And all of them sport the same sign. There is no confounding by settlement type.
There is an interesting pattern that emerges from this analysis. The class gradient is most pronounced in small towns (urban population of 2,500 to 19,999, not adjacent to a metro area) and small suburbs (urban population of 2,500 to 19,999, adjacent to a metro area), and less pronounced in small metros and particularly big metros (with populations greater than a million). This pattern is probably due to the differential weight of the working class by settlement type — the demographic dominance of the working class is more pronounced in smaller places than metropolitan centers for obvious reasons.
Note that we have recentered and rescaled all variables to have zero mean and unit variance. We can therefore interpret the gradients as betas or elasticities. The mean gradient for percent without a college degree across settlement types is 0.63. That's a lot. It means that a one standard deviation higher noncollege share, which comes to 9 percentage points, predicts a 0.63 standard deviations, or +1.8 percent, larger swing towards Trump in 2020 over the 2016 baseline.
Or take a different metric. The rank correlation between noncollege share and the Trump swing is nearly one-half (r = 0.48, P < 0.0001). Equivalently, the correlation between noncollege share and the level of Trump's vote share increased from r = 0.40 in 2016 to r = 0.45 in 2020 (P < 0.0001 for both). So no matter what metric you pick, you get the same result. The evidence is thus overwhelming that class-partisan polarization intensified across the board — the working class moved even further towards Trump.