County returns · 2000–2024
Seven presidential elections of county returns say richer places lean more Republican once you hold education constant. And the part of the Trump-era map that education cannot explain is, overwhelmingly, Hispanic and immigrant America.
Put a county's median household income next to its presidential margin and the past quarter-century looks like a class inversion. In 2000, the vote-weighted correlation between income and the two-party margin across 3,112 counties was a faint +0.10. By 2024 it had quadrupled to +0.40. Rich places vote blue now — every story about the suburban realignment says so, and the raw numbers appear to agree.
The raw numbers are an illusion. High-income counties are mostly high-diploma counties, and it is the diploma, not the paycheck, that moved. Put both variables in the same vote-weighted regression and income flips sign in every single cycle from 2000 through 2024: holding a county's share of college graduates constant, each standard deviation of extra income predicts a margin between 6.5 and 10.7 points more Republican, depending on the year. In 2024 the education coefficient was +27.6 points per standard deviation; the income coefficient was −7.8. The "Merchant Right" that Thomas Piketty and his coauthors documented in survey data across Western democracies has been sitting inside the American county returns the whole time, hiding under the Brahmin Left.
That is the first finding. The second is about when the diploma stopped explaining anything. Education accounts for the county map's 2012-to-2024 drift almost single-handedly — but it explains almost none of what happened between 2016 and 2024. A model using only county BA share captures 2.9 percent of the vote-weighted variation in the 2016→24 shift. Add two Census variables — Hispanic share and foreign-born share — and it captures 61.6 percent. The latest chapter of the realignment is ethnic, not educational. The map below shows exactly where the education model fails.
Where the diploma model breaks down
Each county's 2016→24 margin shift was predicted from its BA share alone (vote-weighted). The map colors what's left over: brown counties moved right of their education prediction, teal counties moved left of it. Hover any county for details.
The brown belt along the Rio Grande is the headline. Webb County, Texas — Laredo, 95.0 percent Hispanic — shifted 51.2 points to the right of where its diploma rate said it should land. South Florida, the Bronx and Queens, urban New Jersey, California's Imperial Valley: the counties education most badly over-predicted for Democrats are a roll call of Hispanic and immigrant America. The teal counter-current is smaller but just as legible — a ring of Black-suburbanizing exurbs around Atlanta, and the Mormon corridor's persistent distaste for Trump.
Why does the bivariate income correlation keep climbing if income's conditional effect points the other way? Composition. Educated counties got dramatically bluer, educated counties are richer, and so income gets dragged along for statistical credit it didn't earn. The regression below separates the two: in every cycle, the education coefficient towers above zero and keeps growing, reaching +27.6 points per standard deviation by 2024, while the income coefficient stays pinned below zero the entire time.
Education pulls counties left. Income, net of education, pulls them right.
Coefficients from a vote-weighted regression of county two-party margin on standardized BA share and standardized median household income, per presidential cycle. Whiskers are 95 percent confidence intervals (HC-robust).
By 2024 this two-variable model explains a remarkable share of the county map — a weighted R² of 0.464 — and the income term is −7.8 points per standard deviation with a standard error of 1.8. This is the county-level shadow of what Amory Gethin, Clara Martínez-Toledano and Thomas Piketty called the Brahmin Left versus Merchant Right pattern: educated elites moving left while wealth, net of education, stays right. In the American county returns the two coexist in the same regression, every cycle, without exception.
The education realignment itself has a precise timeline, and it is not the one the post-election commentary assumes. The vote-weighted margin gap between the most-educated and least-educated tenth of counties was 18.5 points in 2000. It reached 39.0 by 2012, exploded to 60.7 in 2016, peaked at 67.0 in 2020 — and then narrowed to 65.2 in 2024, the first contraction in the series. At the county level, the diploma divide has been frozen for two cycles.
The jaws opened from 2012 to 2016 — and stopped
Vote-weighted two-party margin in the top and bottom deciles of county BA share, with the gap labeled. Deciles re-cut each cycle.
Cumulatively the realignment is enormous: from 2008 to 2024 the top BA decile of counties moved +2.4 points toward Democrats while the bottom decile moved −28.4 — a 30.8-point divergence — and county BA share correlates with that sixteen-year shift at +0.67. For the 2012→24 window, education does literally all of the predictive work: in a head-to-head model of the shift, BA share earns +8.9 points per standard deviation (se 0.77) while income earns +0.10 (se 0.95) — a coefficient of nothing — with a weighted R² of 0.39.
The same freeze shows up in the other great divide. The gulf between large-metro counties and rural ones widened from 28.9 points in 2000 to 57.9 in 2016, sat there through 2020, then narrowed to 54.7 in 2024 — and the narrowing came entirely from the city side, as large metros fell from D+19.0 to D+12.1 while rural counties kept sliding. Democrats' two-party share in the most rural counties has collapsed from 41.7 percent in 2008 to 28.7 percent in 2024; the consolation is that those counties cast a small fraction of the national vote, while the large-metro counties that just lurched right cast 56 percent of it.
So education organized the county map's drift for a decade. Between 2016 and 2024 it abruptly didn't. The vote-weighted correlation of BA share with the 2016→24 shift is just +0.17 — a slope of +1.3 points per ten points of BA share, a weighted R² of 0.029. The scatter shows why the line is so flat: a huge cloud of counties barely moved, and the counties that moved violently moved for reasons the x-axis doesn't know about.
After 2016, education stopped predicting the shift — ethnicity took over
Each circle is a county, sized by 2024 two-party votes, placed by BA share (x) and 2016→24 margin shift (y), and colored by Hispanic share of population. The fitted line is the vote-weighted education-only model. Hover for county detail.
Color the cloud by Hispanic share and the structure leaps out. Adding Hispanic share and foreign-born share to the education model lifts the weighted R² from 0.029 to 0.616 — from nothing to most of the map. With composition controlled, education's gradient actually reappears, at +0.40 points of shift per point of BA share; the foreign-born coefficient is −0.705 points per point. One honest wrinkle: Hispanic share and foreign-born share travel together across counties, and in the joint model the foreign-born term absorbs the Hispanic one (the Hispanic coefficient lands at −0.018, se 0.034). The education residual correlates at −0.61 with Hispanic share and −0.71 with foreign-born share — so attribute the break to the ethnic-immigrant cluster as a whole, not to one Census column.
The shift, sliced by Hispanic share
Vote-weighted mean 2016→24 margin shift by county Hispanic population share.
The magnitudes are stark. Counties that are majority-Hispanic shifted −19.8 points between 2016 and 2024, swinging from D+16.6 to a margin of −3.2 — a near-tie. Counties under 10 percent Hispanic, home to 64.6 million two-party votes across 2,222 counties, shifted +1.0 — toward the Democrats.
The named outliers read like a gazetteer of the border and the barrio. Webb County, Texas went from D+53.1 in 2016 to R+2.2 in 2024, a 55.3-point swing across 65,336 two-party votes, landing 51.2 points right of its education prediction. Starr County, upriver, went from D+61.4 to R+16.1 — a 77.4-point swing, the residual a staggering −72.3. Miami-Dade, with 1,085,945 two-party votes and a BA share of 34.4 percent, went from D+30.3 to R+11.5. Hidalgo and Cameron on the Rio Grande, Imperial in California, the Bronx, Queens, Hudson and Passaic in greater New York: every one of them sits far below the fitted line, and every one of them is heavily Hispanic, heavily foreign-born, or both.
Education explained the 2012–2016 chapter. It cannot explain the chapter that began in 2016 — because that chapter wasn't written in the language of diplomas.
The counter-current is just as specific. The counties that out-performed their education prediction most, among sizable places, are the Black-suburbanizing exurbs south of Atlanta: Henry County shifted +25.3 points (from +4.5 to +29.8, beating its prediction by 28.4), Douglas +20.6, Rockdale +21.2. Utah County and the Mormon corridor also sit in teal — places whose Republican margins under Trump never returned to their pre-2016 altitude.
One more way to see the same structure: run the education model state by state. In demographically uniform states, the diploma gradient on the 2012→24 shift is nearly deterministic — the within-state correlation is +0.96 in Minnesota, +0.93 in Kansas, +0.92 in Oklahoma. In diverse states it collapses: +0.35 in Florida, +0.47 in California, +0.62 in New York. Where the population is homogeneous, education is destiny for a county's trajectory. Where it isn't, ethnicity scrambles the gradient.
Everything here is about places, not people. "Richer counties lean more Republican once you control for education" is a statement about county aggregates; it does not mean a richer household votes more Republican than its degree predicts, and individual-level income gradients estimated from surveys can and do differ. The same caution applies to the ethnic results: a 95-percent-Hispanic county swinging right is overwhelming evidence about how that community's electorate moved, but a county where Hispanic residents are a minority of the population tells you less about who within it changed. County demographics are interpolated American Community Survey estimates at 2024 vintage, so some of the "education" composition is itself a consequence of politically sorted migration. And margins here are two-party margins computed from raw vote counts, which sidesteps a known share-column corruption in three states' 2024 records.
With those fences posted, the two findings stand. The county-level class inversion everyone can see in the raw data is a composition effect of the diploma realignment — beneath it, money has leaned the same direction for a quarter-century. And the realignment's newest chapter is not more of the diploma divide, which froze in 2020. It is something the education model literally cannot see: the political migration of Hispanic and immigrant America.