Carmena’s OvertonSubscribe
Menu

Methodology

Every number on this site comes from a named source, every modelled figure is labelled as a model, and every caveat that changes the reading stays with the chart. This page explains how. Last updated Oct 23, 2026.

Sources

We use public records first. Where a state publishes voter-level or aggregate election data, we use it directly:

  • North Carolina: State Board of Elections absentee and early-vote file, statewide registration file and voter history file.
  • Ohio: Secretary of State absentee and early voting export, and statewide voter file.
  • Texas: Secretary of State Early Voting Roster and 2026 primary rosters; 2022 and 2024 primary rosters posted by Tarrant, Dallas, Bexar and Hidalgo counties.
  • Iowa, Kansas, Maine, Minnesota, Alaska, South Carolina, Michigan: state early-vote reports and files.
  • Past results: county and precinct returns from MIT Election Data and Science Lab, Harvard Election Data Archive and OpenElections.

Where no state record exists, we quote trackers and say so: the NBC News early-vote tracker (TargetSmart data), UF Election Lab, VoteHub, L2 and On Point. For polls we use the Silver Bulletin generic ballot average for 2026 and 538’s averages for 2018 and 2022.

Labels

Party of early voters means three different things, and we tag each:

  • reg. Party registration from a state file.
  • primary. Party of the last primary ballot, from state records. Used in Ohio and Texas, which have no party registration. Tested in North Carolina, it gets the D−R margin within about a point but understates unaffiliated voters by about 10 points.
  • model. A vendor’s estimate, usually TargetSmart via NBC. In 21 states, holding 48% of 2026 early votes on Oct 23, the NBC tracker’s party is modelled.

States NBC does not mark as modelled are described as “not marked as modelled”, not as recorded or official.

Party is not vote choice. Registration, primary history and modelled scores describe who voters are, not how they vote.

Comparisons

“11 days out” is Oct 23, 2026. We compare with the same point in earlier cycles: Oct 25, 2024 and Oct 28, 2022. Margins are Democratic share minus Republican share, in points, written D+ or R+.

Geographic estimates, which weight each county’s early votes by its 2024 result, understated the Democratic margin by 7 to 41 points in 8 states where we could test them. We use them only as a lower bound.

Caveats

A caveat that changes the reading of a chart appears on every copy of it, including thumbnails. Examples: party is modelled in 21 states; Ohio and Texas party figures come from state files older than the rest of the table.

Privacy of voter records

Several state files contain names and addresses. We reduce them to counts as they are read and keep no voter-level records. Everything we publish or license is aggregated.

Forecasts

Our 2026 House and Senate forecast is not yet published. When it is, we will publish for each chamber:

  • the median seat count and a range;
  • the chance of control;
  • the date and time of the forecast, and the data it used.

Each forecast will be published before results are known and kept online unchanged. If we update it, earlier versions stay visible with their dates.

After the result we will score it in public on the track record page: the seat error against the median, whether the result fell inside the stated range, and, for chances, a standard probability score (the Brier score). We will report misses the same way as hits.

Polls

We have not published a poll. Every poll we publish will carry a disclosure block, following the American Association for Public Opinion Research (AAPOR) disclosure standards:

  • who sponsored the poll and who conducted it;
  • fieldwork dates;
  • the population and the sampling frame, with known coverage gaps;
  • sample size, total and per mode, and for likely voters;
  • how respondents were recruited, eligibility rules, quotas and oversamples;
  • interview modes and languages;
  • whether the sample is probability or non-probability;
  • exact question wording and order, with the full questionnaire;
  • weighting variables, targets and their sources;
  • precision: margin of error with design effect for probability samples, or a documented credibility interval for non-probability samples;
  • data processing and quality checks, including how many interviews were removed and why;
  • a statement that question wording and practical difficulties can add error beyond sampling error;
  • how to contact us with questions about the method.