2020 County Vote-Share Demographic Regression

What this research found

How much of the 2020 presidential vote is written into county demographics? Official county returns were merged with American Community Survey five-year estimates for 3,111 counties, and the Republican share of the two-party vote was regressed on just five covariates: college attainment, non-Hispanic white share, median household income, population density, and median age. The model explains 64.6% of the cross-county variance, and when it is retrained without ever seeing ten of the fifty states it still explains 45% of the variance in those states.

• Five demographic covariates account for 64.6% of the cross-county variance in Republican two-party share (R² = 0.6455, adjusted 0.6449, F = 1,130.8) with an in-sample root-mean-squared error of 0.097 across 3,111 counties.
• Education and racial composition dominate, and almost cancel each other out. One standard deviation more bachelor's-degree holders (about 9.7 percentage points) is associated with an 8.5-point lower Republican share, while one standard deviation more non-Hispanic white residents (about 20 points) is associated with an 8.4-point higher share.
• Population density adds an independent negative pull of −0.052 per standard deviation even with education, race, income, and age held fixed. Median age (−0.025) and log income (+0.027) are the weakest contributors.
• Clustering standard errors by state inflates them by factors of 2.3 to 6.1 relative to classical ones — most for the geographically clustered predictors, non-Hispanic white share (6.13 times) and log density (5.00 times). All five predictors remain significant at p < 0.01 under the conservative errors.
• Trained on 2,433 counties in 40 states and tested on 678 counties in the 10 withheld states, the model reaches a held-out R² of 0.4494 and an RMSE of 0.1022 — predictions typically within about ten percentage points. Training-set R² of 0.676 sits close to the full-sample value, so the gap reflects extrapolation across unseen state political cultures rather than overfitting.
• The largest errors expose the cost of one racial covariate. Miami-Dade County came in at 0.46 Republican share against a predicted 0.21, while Native American-majority counties went the other way: Sioux County, North Dakota at 0.24 actual versus 0.59 predicted, and Menominee County, Wisconsin at 0.18 versus 0.52.

How it was done

County returns for 2020, ACS 2016–2020 five-year demographic estimates, and the 2020 Census Gazetteer land-area file were joined on five-digit FIPS codes, with identifiers zero-padded to preserve leading zeros, retired and renamed codes remapped, and non-participating territories removed. The join retained 98.7% of participating election jurisdictions; the systematic non-matches are Alaska's forty state-legislative districts, which report presidential results without a county-level demographic analogue. Income and population density were log-transformed to tame right skew spanning nearly seven-fold and five orders of magnitude respectively, and all five predictors were z-scored so coefficients read as effect sizes. Ordinary least squares was fitted with state-level cluster-robust standard errors, and generalization was tested by shuffling the fifty states under a fixed seed into a 40-state training set and a 10-state test set, with standardization statistics taken only from the training counties.

Data sources

• MIT Election Data and Science Lab — County Presidential Election Returns 2000–2020 (3,152 county-equivalent jurisdictions in the 2020 contest)
• US Census Bureau American Community Survey 2016–2020 five-year estimates — nine raw variables at the county summary level
• US Census Bureau 2020 Gazetteer file — county land area for the density calculation