When Did the White Working Class Defect to the GOP?
Democrats have a story they tell themselves about how they lost the white working class to the GOP. The story goes that, upon signing the Civil Right Bill into law, Johnson turned to his aide and declared "now we have lost the South for a generation." Although widely believed, the story is a complete myth. It was not until much later that the Dems lost the South. But the South also stands in for the white working class. The moral of the story is not lost on anyone: Us, educated and civilized Americans, became good antiracists after the Boasian revolutions of the Sixties. Those guys over there, in Flyover Country, they're poorly educated. They didn't get the memo — they're still bigots. Why do you think they keep voting for the wrong party?
However contemptuous, the legend is a coherent hypothesis. It says that the Democratic New Deal coalition broke because Dems embraced a politics of multiracial inclusion and this drove racially-resentful working class whites into the arms of the GOP. I have been wondering how to test this hypothesis for a while. Earlier, I looked at the diachronic pattern in party identification by class using the ANES dataset.
I also looked at whether partisan polarization is predicted by ideological sorting or racial resentment. I concluded that racial resentment was a weak predictor of GOP partisanship:
We find that ideological polarization has increased dramatically across the board. However, it is much more pronounced among college-educated whites relative to the working class. The patterns again point to the early-1990s as a crucial inflection point. Partisan polarization comes to a boil over the past decade. Meanwhile, the evidence for racial resentment as a driver of partisan polarization is weak; and in as much as there is evidence of it at all, it is much more recent and not confined to the GOP. Instead, we find much stronger evidence that white antiracist Democrats have become much more partisan over the past decade. Moreover, the evidence for the hypothesis that racial resentment has made the white working class strong GOP partisans is practically nonexistent.
Now I want to subject the hypothesis to a different test. We shall examine the diachronic pattern of interstate variation in the GOP vote share by election year. We interrogate the presidential election returns for 1976-2016 from MIT election lab. The basic idea is party vote shares per se are confounded by the ebb and flow of electoral fortunes. That is, the information contained in the GOP's vote share is hard to extract directly. For instance, if we examine the GOP's vote share in the subset of states or counties most dominated by working class whites over time, that series will exhibit fluctuations unrelated to the rotation of the white working class to the GOP; it'll be confounded by such things as the rise and fall of leaders like Bush and Reagan. This is why we looked at the geographic bias (North less South) of the Dem vote share in that piece.
But the information we are interested in can be extracted from the time-variation in the cross-sectional pattern. Here's how. Our response is the GOP's vote share in presidential election by state and election year (obtained from MIT election lab). Our feature, that contains class information, is the percentage of adults without a college degree. We have data by state but not election year from Social Explorer. Instead, we have data by census year. We construct our feature from this raw data by linear interpolation. Once we have matched data for both the feature (percent of adults without a college degree) and the response (GOP vote share) by election year and state, we can proceed.
The information we are interested in is contained in the gradient of the response against the feature. However, the data is plagued by outliers, and given that our regressions are cross-sectional, the problem of heteroskedasticity can be expected to be acute. We therefore dispense with OLS and resort to robust estimation using the MCD estimator. (OLS yields similar but less reliable results).
The story of when the white working class migrated to the GOP is revealed by the rotation of the fitted line. See scatter plots below — read from the top left. The reference lines in the graphs are not least squares lines, but fitted regression lines using the MCD estimator. We can see that SLOPE was negative in 1976, flattens in the 1980s, becomes upward sloping in the 1990s, and acquires the modern pattern in the 2000s when it becomes sharply upward-sloping. (The outlier is DC.)
The following bar graph displays the MCD robust slope estimates, SLOPE. The pattern is interesting. The migration of the white working class is arrested in the 1996 and 2004 cycles. But between them lies the great jump of the year 2000. With the 2008 cycle, the rotation is resumed. In 2016, SLOPE was higher that it has ever been.
What is clear is that the migration of the white working class to the GOP is a secular realignment that exhibits a quantum jump in 1980 and 2000. More concretely, in 1976, Jimmy Carter performed better in states where the white working class was predominant than where it not. An advantage that he lost decisively in the 1980 election. Clinton managed to stem the migration of working class whites in 1996, but Gore lost them in decisive numbers in the 2000 cycle. There is a modest jump in SLOPE in 2008, suggesting that Obama's candidacy did not push the white working class into the arms of the GOP. Finally, Trump's election is no structural break but a natural progression of a long-standing movement.
We have county level data from 2000, again from the MIT election lab. That shows a much more dramatic jump in SLOPE in both 2008 and 2016. Since we are using the exact same methodology for both, the contrast between these two graphs is due entirely to within-state variation. The level of analysis makes a huge difference in interpretation. Trump 2016 and Obama 2008 now really stand out as dramatic moments of acceleration in class-partisan sorting. SLOPE was so massive in 2016 that it explained fully one-third of the cross-sectional variation in the GOP's vote share by county, up from one-tenth during the Obama elections.
In order to dig into this further, we sort the counties into quintiles by college graduation rate and obtain a model-free estimate of the gradient as the effect size. Specifically, we use Cohen's d, defined as the difference in the mean GOP vote share in the lowest and highest quintiles (Q1-Q5) divided by the standard deviation of the GOP vote share across counties. This measures the change in GOP vote share if we move from the top fifth of counties to the lowest fifth of counties by college graduation rate. This methodology reveals yet another pattern, but at least the secular trend is consistent.
Rank-based methods are perhaps more robust. Table 1 presents our estimates for Spearman's correlation coefficient. They are completely consistent with our slope estimates from the cross-section of counties above. Table 1. Spearman's correlation coefficients for GOP vote share and percent of adults without a college degree. r P 2000 0.023 0.191 2004 0.086 0.000 2008 0.222 0.000 2012 0.222 0.000 2016 0.373 0.000 Source: MIT Election Lab, Social Explorer, US Census, author's computations. N=3,154 US counties. Estimates in bold are significant at the 1 percent level. What is perplexing about these patterns is the incongruence between the interstate gradient and the inter-county gradients in 2000, and to a lesser extent 2004. The interstate SLOPE estimates suggest an acceleration of the rotation of the white working class to the GOP, while intercounty estimates — SLOPE, interquintile effect size, and rank correlation coefficients — suggest otherwise. Table 1. Spearman's correlation coefficients for GOP vote share and percent of adults without a college degree. Interstate Intercounty r P r P 1976 -0.347 0.013 1980 0.015 0.914 1984 0.128 0.370 1988 0.282 0.045 1992 0.443 0.001 1996 0.393 0.005 2000 0.547 0.000 0.023 0.198 2004 0.592 0.000 0.093 0.000 2008 0.671 0.000 0.228 0.000 2012 0.673 0.000 0.239 0.000 2016 0.779 0.000 0.404 0.000 Source: MIT Election Lab, Social Explorer, US Census, author's computations. N=3,154 US counties and N=51 states. The most parsimonious explanation of the incongruence in the 2000 cycle is that the class signal is getting swamped at the county level. Indeed, when we compare the ranking of states and counties, a clear pattern emerges in the former but the latter is completely random. Put another way, the signal-to-noise ratio is very low in the latter.
This is still a work in progress. But what is clear is that the class party realignment is a much more recent story than doesn't get going until the 1990s, if not the 2000s. I mentioned to someone on Facebook that perhaps Bernie was bidding for the wrong party's nomination. He did earnestly want to forge a mass working class movement. But the majority of working class folk now reside within the GOP. What explains this class-party realignment? Why did working class whites abandon the Dems? This is the principal explanadum of contemporary American politics.