Elections, 1976 to 2024

Data sources

• MIT Election Data and Science Lab, U.S. President 1976-2020 (extended to 2024)State returns by candidate and party line, 13 electionsWeights: none (full counts)
• Algara and Amlani county presidential returns 1868-2020 (Harvard Dataverse, doi 10.7910/DVN/DGUMFI), and MIT Election Data and Science Lab 2024County votes for the Democratic and Republican candidates and all candidates, 13 elections plus 1972 for the 1976 swingWeights: none (full counts); vote-weighted views weight by two-party votes
• County demographics panel 2000-2024Census Bureau decennial census and American Community Survey, USDA rural-urban codes, interpolated to election yearsWeights: none (full counts and interpolated estimates)
• American National Election Studies Cumulative Data File, 1948-2024Self-reported vote for President among voters, presidential years 1976 to 2024Weights: VCF0009z
• American National Election Studies 2024 Time SeriesUsed to rebuild 2024 church attendance onlyWeights: none applied here; estimates use VCF0009z from the cumulative file

The charts read manifest.json and one file per year for the states, the counties and the regressions, in the same folder. All are aggregates: no respondent-level rows.

Methods

Margins are two-party margins: 100 times the Democratic minus the Republican votes, over the votes for the two parties, so a positive margin favors the Democratic candidate; the margin over all votes is shown where it is offered. A swing is the change in that margin from the previous election. Returns are full counts, so they carry no sampling error. The county files leave out Alaska and merge Kansas City, Missouri into Jackson County in 2024.

County traits come from the Census Bureau and the USDA rural-urban codes, interpolated to election years. Each regression is a least-squares fit of the county two-party margin on five traits entered together. A coefficient is the difference in margin associated with a trait one standard deviation higher, with the standard deviation taken over the counties in the fit and the other four traits held fixed. Standard errors are clustered by state and intervals are 95% normal intervals. The vote-weighted view weights counties by two-party votes. Within-state fits use states with at least 15 counties with complete data, standardize the traits within the state and use HC3 robust standard errors. The correlation in each scatter plot has a 95% interval from a bootstrap over states (2,000 replicates, seed 12345).

Survey estimates are the share of the two-party vote among ANES respondents who reported voting for the Democratic or Republican candidate, weighted with VCF0009z. Intervals are Wilson intervals on the Kish effective sample size, at the 95% level. A group with fewer than 30 respondents in a year is not shown, and one with fewer than 100 is marked thin.

What these data cannot show

County and state results describe places, not voters: a county that is more Democratic where college share is higher does not show how college graduates in that county voted. The trait coefficients describe how county margins differ and are not estimates of what would happen if a county’s traits changed. The survey estimates use reported vote, cover respondents who said they voted, and have intervals that leave out the survey’s clustering and stratification, which can make them narrower or wider than a design-based interval.

How to cite

@misc{singh2026exploreelections,
  author = {Singh, Vishal},
  title = {Elections, 1976 to 2024},
  year = {2026},
  url = {https://vishalsingh.org/politics/explore/elections},
  note = {In The Public Mind. Published 30 September 2026.}
}