Methods & limitations
Every chart should show where its certainty ends
The Data Lab publishes reproducible aggregates from public records. This page documents the source version, join logic, missing values, derived measures and known coverage limits behind the interface.
01 · Sources
County Voting Atlas
Election administration: U.S. Election Assistance Commission, 2024 Election Administration and Voting Survey, public release v2. The source is jurisdiction-level. We aggregate jurisdictions using the first five digits of the reported FIPS code.
Community profile: U.S. Census Bureau 2020–2024 American Community Survey five-year estimates, retrieved through the open Census Reporter API for county geography. The selected variables cover population, age, household income, education, poverty, nativity and household vehicle access.
The two sources are joined on five-digit county FIPS. Wisconsin is a documented exception: EAVS supplies municipal codes, so single-county jurisdictions are allocated from the county named in the source label. Its 58 “multiple counties” rows cannot be divided without inventing values and are excluded. One statewide Maine UOCAVA row is also excluded from county totals.
The current release contains 3105 matched profiles. Alaska and Connecticut are absent from the matched atlas because the election and ACS files use incompatible county-equivalent structures in this release.
02 · Semantics
Missing is not zero
EAVS uses negative sentinel values such as -77, -88 and -99, as well as blanks. The pipeline converts all of them to null before aggregation. A true reported zero remains zero.
Each county retains field-level coverage: the share of its source jurisdictions that reported a usable value. National totals are therefore labelled “reported jurisdictions only.” Rates are calculated only when the required numerator and denominator are both present and arithmetically possible.
03 · Comparison
County twins across state lines
Each county is compared only with counties in other states. Seven ACS measures are standardized across the matched dataset: log population, median age, log median household income, bachelor’s degree share, poverty share, foreign-born share and households without a vehicle.
The closest Euclidean neighbor becomes the twin. The displayed similarity score is a readable transformation of that distance. It is not a probability, causal estimate, political label or prediction of election results.
04 · Ballot paths
Ballot DNA classification
The source is the MIT Election Data and Science Lab 2020 cast-vote record archive, release 3.0. It covers 20 states and 362 county partitions, with substantial variation in local coverage.
A down-ballot contest qualifies only when the record contains at least one Democratic and one Republican candidate. A presidential ballot is classified “down-ballot Democratic” or “down-ballot Republican” only when every recorded choice in those eligible contests belongs to that party. All other paths are mixed.
Strict crossover counts only Biden/Democratic presidential ballots classified entirely down-ballot Republican and Trump/Republican ballots classified entirely down-ballot Democratic. Ballot-level identifiers never leave the ETL workspace; the public dataset contains aggregate counts only.
05 · Quality control
Validation before publication
- Known federal sentinel codes remain null while reported zero remains zero.
- National and state Ballot DNA flows reconcile exactly to their represented ballot totals.
- No ballot identifier is present in a public output.
- County twins always cross a state boundary and reference an existing profile.
- Impossible source relationships, such as returned ballots exceeding transmitted ballots, keep their raw counts but suppress the affected rate and add a visible flag.
- Processed files include source hashes and release metadata for reproducibility.
06 · Read carefully
Known limits
EAVS is an administrative survey, and reporting completeness varies by state, county and field. Aggregation can conceal differences between multiple jurisdictions inside one county. ACS estimates have sampling uncertainty that is not currently shown in the atlas.
Ballot DNA is a coverage archive, not a probability sample of the United States. It must not be used to infer how an identifiable person voted, and the aggregate paths do not explain why a voter made a choice.
Corrections can be submitted through the site’s contact page. Material corrections will be documented in the newsroom’s corrections process.