You must’ve have heard the idea that most Americans think of themselves as middle class. That’s not true. The General Social Survey has regularly asked Americans to choose from lower class, working class, middle class, or upper class since 1972. According to the latest numbers, those from the 2018 module, 44 percent call themselves middle class, another 44 percent call themselves working class. These are the two big self-identified classes in America and have been since 1972. Back then, for instance, 47 percent of respondents reckoned that they were working class while 44 percent choose middle class. What happened to the 3 percent? They now call themselves lower class — whose share has grown from 6 percent in 1972 to 9 percent in 2018. The share of the upper class has also grown from 2 percent to 4 percent. Apart from these small changes, the frequencies of self-identified class has been virtually constant over the past fifty years.
The best predictors of self-reported class are, no surprise there, what we think of as proxies for class — income, occupational prestige, and educational attainment. Just the bachelor’s degree is an excellent predictor. In order to compute odds ratios, we drop the minor classes at the top and the bottom of American class society from our sample. Some 59 percent of respondents without a college degree call themselves working class; only 26 percent with one do. These odds are extremely significant (OR = 4.15, P < 0.001). More generally, the relationship between highest degree attained and self-identified class can be read off the following heat map. Most graduates think of themselves as middle class; conversely, most high school graduates think of themselves as working class. More generally, as one can check by visual inspection, the frequency of self-identified middle class status has the correct order predicted by educational attainment: LHS < HS < JC < BA < GRAD. The frequency of working class is a bit more noisy but the order of GRAD < BA < JC < HS is the reverse of the hierarchical order of highest degree attained.
Self-identified class is a monotonic function of educational attainment for good reason. Education contains the strongest class signal. The professional class sorts itself on school prestige. A college degree is clearly required for any professional or managerial position. A high school diploma protects against a host of ills. Educational attainment is a strong conditioner of life outcomes — mortality, morbidity, suicides and drug overdose, smoking and other risk-taking, income, incarceration, wedlock, reproductive success, you name it — up and down the class hierarchy.
Race cannot be ignored at all. Whereas 38 percent of whites in the 2018 GSS module sport a college degree; only 29 percent of blacks do. The disparity is significant at the 5 percent level (OR = 0.65, P = 0.030). In order to examine the odds ratios by self-identified class and partisan orientation, we drop the lower and upper classes, races other than blacks and whites, and parties other than Democrats and Republicans. We first report the race-blind class-party odds ratio. Whereas 39 percent of college-educated Americans in the restricted sample are Republicans, 46 percent of those without a college degree are. The odds are significant but only marginally so (OR = 0.76, P = 0.049).
Some 55 percent of whites without a college degree are Republicans; 44 percent of graduates are. The odds are very significant (OR = 0.64, P = 0.004). Cf. [compare with] 7 percent of blacks with or without a college degree are Republicans (odds ratio computation unnecessary). So this is a whites-only business. We shall see that this is true going back fifty years. In order to document the diachronic pattern of class-party realignment, we report class-party odds ratios over 1972-2018. We document the class pattern with college degree and self-identified class. As we shall see, both tell the same story.
We begin, as before, with race-blind odds. Higher odds correspond to a Democratic advantage among the self-identified working class relative to the middle class. The dotted lines are 95 percent confidence intervals. We can see that there is a decline to a lower plateau around 1990, corresponding to social democracy’s capitulation upwards. The Clinton coalition was already fully neoliberal. Interestingly, there is a local maxima in 2008 — corresponding to Obama’s historic election. The global minimum is attained in 2018 — two years into the Trump presidency.
When we restrict attention to whites, we find that the white working class bias towards Democrats falls into insignificance in the early-1990s. Note the local maxima in 2008 (OR = 1.54, P < 0.001) — a pattern that carries over to the race-blind odds. The racial resentment theory could not get it any more wrong — the white working class was more biased towards Democrats than the white middle class when a black man was the presidential candidate than at any other time over the past thirty years. The decline, however, renews immediately. Apart from a marginally significant spike in 2014, the odds ratios fall off. The global minimum is attained in 2016 (OR = 1.07, P = 0.510).
Fluctuations that we can document in black class and partisan identification are largely not significant (P > 0.05). The exceptions are interesting: 2004 (OR = 3.25, P < 0.01), 2012 (OR = 5.41, P = 0.04), and — get this — 2018 (OR = 0.34, P = 0.04). The inversion has never been significant before. Note also that the local peak in 2008 in the race-blind, self-identified class-party odds ratio was not driven by working class blacks turning out democrats at higher rates — there was no class bias in black partisanship in 2008 (OR = 1.04, P = 0.94). Therefore, the local peak in 2008 in the race-blind class bias was driven entirely by whites. The white working class turned out for Obama.
Self-identified class is strongly associated with educational class. We begin again with the race-blind odds ratio. Again, higher odds show a working class bias for Democrats relative to the middle class. The pattern here is subtly different. Although the negative trend, recency of the global minima, and the local peak in 2008 are common features whether we use BA or self-identification, the college degree contains a more noisy signal of the class-partisan realignment. The steep decline of the early-1990s vanishes — we see only trend and cyclical components. The series becomes statistically significant around cyclical maxima. Any bias of Americans without a college-degree for Democrats falls into insignificance after 2008. The odds ratios invert in 2016.
Breaking it down by race, we see that working class bias for the Democrats falls into insignificance by 1996. In 2018, for the first time, it becomes significantly inverted (OR = 0.71, P < 0.01). We can see that, just like the odds ratios with self-identified class, the entire pattern of the BA/party odds ratios for whites is inherited by the race blind odds.
For blacks, the BA series is too noisy to ever be statistically significant. And the pattern of black working class bias towards Democrats is very different from that of whites.
We also document the two parties’ race-specific share of the two self-reported classes. We revert back to the full sample so as not to misrepresent the shares. Note that, compared to the odds ratios, the class-partisan realignment signal in these numbers is confounded by fluctuations in the fortunes of the two parties. The local maxima in the Democrat share of the white working class in 2008 stands out. As does the precipitous decline from the mid-1970s to the early-2000s. Again, these are not diagnostic measures of class-partisan realignment.
For completeness sake, we also document the partisan loyalties of blacks by self-identified class.
We have documented that self-identified class-party odds ratios have a very the strong signal of class-partisan realignment. Specifically, self-identified class contains a stronger signal than the college degree. And odds ratios are less confounded by fluctuations in aggregate partisan identification than party shares of the self-identified classes. This makes sense. Earlier, I had estimated class party realignment since 2000 from the gradient of college graduation rate in the cross-section of US counties. The realignment, I said, can be read off the rotation of the gradient vector in the following graph. The upper panel displays the county graduation rates for counties where the GOP and the Dems won; the lower panel displays the national means of the same.
The GSS data allow us to extend the analysis back to 1972 and work directly with class identities. The class-partisan sorting, we have shown, is confined to whites. And it reveals a local maxima around Obama’s historic election. It was not under Clinton that working class whites said they were Democrats at higher rates than middle class whites — it was under Obama. This calls into question on the racial resentment hypothesis, the dominant narrative of 2016. Moreover, in light of the pattern I have documented, we must revise our understanding of the Clinton coalition. Finally, recall that I am merely documenting facts using a shared reference frame. I am not making an argument or testing a hypothesis. This exercise is entirely model free. Instead, I am documenting the patterns that deserve explanation. If we are serious about understanding elite-mass relations in the United States and the shadow they cast on US politics, we have to begin by trying to understand the following graph.
Postscript. Quick cross-check of the 2008 local maxima in the white working class bias towards Democrats. Another set of variables asks whom the respondent voted in particular elections. The white working class is significantly overrepresented among Obama voters (OR = 1.20, P = 0.019). In the 2016 presidential election, the white working class is underrepresented among Clinton voters and overrepresented in the Trump coalition although the association is only marginally significant (OR = 0.77, P = 0.055). The racial resentment hypothesis of the Trump victory can therefore be ruled out.
Postpostscript. A short note on facts and hypothesis testing. This came out of endless discussion on Facebook about what the middle brow antiracist literature (DiAngelo, Kendi et al.) and the critical theory crowd calls “White European Enlightenment epistemology” — a derogatory term for the scientific method and evidence-based argument.
What is fact? Fact is an agreement on shared reference frames. Standards and measures are the classical instance. In physical anthropology, one reports morphometric quantities, say, the diameter of a molar measured by such and such technique, using such and such instrument. A factual claim is not a hypothesis about the world but a statement of brute fact such as blah is the speed of light in blah units. In fact, we can distinguish facts from hypotheses by checking whether the claim is about a causal structure in the world or a factual detail. The wager of a factual claim is that if any other competent person were to undertake the same measurement, they would find the same pattern. If not, then the factual claim would be debunked and the author understood to be incompetent.
Once you have established the facts, the patterns are revealed. If the pattern is interesting, they become explananda around which ‘ecologies of attention’ [Crawford] form. Hypotheses seek to explain patterns in the world already established using shared reference frames. What one immediately tries to do is to disprove hypotheses which can tested against the null.
The racial resentment hypothesis can be ruled out by merely establishing the facts of the matter. Specifically, we check whether the pattern is consistent with the predictions/causal descendants (X -> … -> y) of the racial resentment hypothesis. Odds ratio is one such reference frame — one shared right across the empirical disciplines. What is it? The basic design is this: it is the ratio of the ratios: treatment-success/treatment-failure and control-success/control-failure, where treatment/control and success/failure are any Boolean/zero-one/dummy variables.
I established the following fact that is inconsistent with the racial resentment hypothesis: The pattern of variation of relative working class support for Democrats shows a negative trend, a precipitous decline around 1990, a local maxima coinciding with the Obama candidacy in 2008, and a recent global minimum in 2018. If white working class people are more racist than the white middle class, then why did they surge for the Dems when the Democratic candidate for the presidency was a black man?
“By 2020,” Taibbi writes, “the official answer to What’s the Matter with Kansas? was Kansas is a White Supremacist Project and Can Go Fuck Itself.” Taibbi is being too kind to Frank. Frank doesn’t have a coherent answer to 2016. Why did Trump win? I have argued for the hypothesis that what put Trump in the white house is a breakdown of elite-mass relations that caused a class-partisan realignment. What is clear [a factual claim] is a rotation away from the self-identified working class whites and towards the middle class leading to the Clinton coalition; the rotation is partially reversed under Obama, but returns with a vengeance under Trump. The pattern is inconsistent with the racial resentment hypothesis. But it is a pattern that needs explanation. The assumption that the white working class is racist is not only a prejudice; it is what anchors the class work some of us do all the time, many of us do some of the time, and many of us — people who consider themselves working class and lower class — perceive in their lived experience as an insult. (I’ll have graphs to share on that last point soon.)