About this project
A nonpartisan record of what candidates for federal and state office say about health care on their own campaign websites.
Why
On a campaign website, a candidate says what they intend to do, on the record and in their own words. Reading every site for one policy area shows how health care is being presented to voters, which polls, voting records and news coverage do not. The dataset is meant for researchers, journalists and voters; it records positions but does not evaluate them.
What was done
Who. The two leading candidates in every 2026 U.S. House, Senate and Governor race, identified by prediction-market win probability (Polymarket) at the time of data collection: 1,002 slots, 982 identified candidates, 972 with a readable campaign website (492 Democrats, 473 Republicans, 7 independents).
What was read. Only the candidate's own campaign website: the home page and every issues, priorities, platform or plan page, including pages not labelled "health" (veterans, seniors, families, rural). Official government pages, news, Ballotpedia, social media and voting records were not used.
What was recorded. For each of 10 main issues (reproductive health, mental health, drug pricing, substance use, Medicare, VA health, the ACA, food and nutrition, Medicaid, universal coverage), whether the site mentions it (yes / no / unclear), the position taken (support or expand; protect or oppose cuts; reform; oppose or repeal; mentioned without a direction), a short description of the proposal, a verbatim quote of at most 15 words, and the page URL. Any other health topic the site names was recorded the same way with a standard label (for example rural hospitals, price transparency, paid leave).
How. Sites were coded by large language models following the written codebook, with two cross-model audits in which a second model re-read the sites and flagged disagreements. A research assistant is now manually checking sites and adding candidates; the data on this site are updated as that review proceeds. The paper describing the method, its validation and the findings is in preparation.
Caveats
- Sites were read in October 2026 and change over time; the version date tells you when.
- Twelve House candidates and one Governor candidate had no campaign website at data collection. Nineteen House slots had no identified nominee yet (for example, Louisiana's November 3 primaries).
- Only what a site says is coded. Positions are never inferred from party, votes or statements elsewhere.
- Coding of ambiguous language is a judgment call; "Unclear" rows are reported so that users can decide for themselves.
- Until manual review is complete, individual entries may contain errors. If you find one, open the candidate in Explore and use "Report a problem" at the top of their details, or email [email protected].
Updates and versions
Each data release is dated and kept permanently. The Data page lists every version with notes on what changed. The same structure will be used for future elections, so results can be compared across cycles.
Who
This project is by David Scheinker, Maya Shah, Anton Backer and Kevin Schulman. The views and any errors are the authors' own; the project does not endorse or oppose any candidate. Contact: [email protected].