Methods

How every figure in this case study is made, in one place. Published November 1, 2026. Changes are dated and listed in the corrections log.

Sources

Results, 2026: each state's own official results files (unofficial until the state certifies), read automatically and checked before they are shown. States we do not read link to their official results.
Past results: MIT Election Data and Science Lab (CC0): U.S. House and Senate 1976 to 2024, county presidential returns 2000 to 2024. Totals were checked against official national and state figures.
What places are like: U.S. Census Bureau, American Community Survey 2020 to 2024 5-year estimates, with margins of error (90% confidence).
More recent figures: Bureau of Labor Statistics county unemployment (September 2026, preliminary, not seasonally adjusted) and Census Bureau population estimates (July 2025).
Who voted, by group: Census Bureau Current Population Survey, November 2024 supplement. It shows who voted, never for whom.
Campaign money: Federal Election Commission candidate summary file for 2025 to 2026: each candidate's campaign committee totals as reported. Party and outside spending are not included.
Boundaries: Census Bureau 2024 cartographic files and 2026 congressional districts. Missouri uses its 2024 lines.

Margin: the Democratic share minus the Republican share of all votes cast in the race, in points. Third-party and write-in votes count in the total. Change is this year's margin minus the earlier one.
Districts with new lines in 2026 (AL, CA, FL, LA, NC, OH, TN, TX, UT, 173 districts): Census figures are built by us from block groups, split by each block's 2020 population into the new districts; medians are estimated from income brackets. They are marked as our estimates. These districts are never compared with their 2024 House result, which belongs to a different shape.
Reliability: each Census estimate is rated by its coefficient of variation: under 15% reliable, 15 to 30% use with caution, over 30% unreliable. Unreliable estimates are shown but marked, and left out of correlations, averages and ranks.
Recent figures for districts and states: BLS and population figures for districts are built from the counties they contain, weighted by each county's share of residents in the district; states are summed from counties.
Turnout, 2024: votes cast for president divided by the Census estimate of citizens of voting age. It is an approximation, not a count of eligible voters on Election Day.

Correlation (r): Pearson's r across places, from -1 to 1. Strength words, always shown with the number: under 0.1 almost none, 0.1 to 0.3 weak, 0.3 to 0.5 moderate, 0.5 and over strong.
Each place once, and weighted: the main figures count each county once. The baseline's addendum also weights each county by its population, so each resident counts equally; the two can differ a lot.
Uncertainty: 95% ranges for r use Fisher's z, treating places as a sample. Places are not independent of one another, so the range is a guide to how settled a pattern is, not a formal test.
Thirds and fifths: places are split into equal-sized groups on a measure; where many places share the lowest figure (often 0), only places above it count as the highest group.
Trend line: a straight least-squares line, used only to describe the pattern and to find places furthest from it.
Similar places: the closest places on eleven standardized Census figures; votes are not used to find them.
Neighbours: places that share a border on the Census boundary files.

It does not call races, predict results, rate races or endorse anyone.
It does not estimate how groups of voters voted: votes are counted by place, and comparisons describe places, not people.
It does not use paid or unlicensed data. Polls appear only where the publisher allows reuse.
Every sentence of the findings is written by a fixed formula and checked against our list of words we do not use. The questions were fixed and fingerprinted before the vote: see What we will report.