The federal healthcare record, studied.

Source attribution

Where every number comes from

Every figure in these studies resolves to a full source record, not a footnote. The record pins the federal source the number came from, the dated copy of the file it was read from, the method and dataset it belongs to, and the licence that lets you reuse it. Around that record, each study also publishes the join that produced the figure and the limitation that bounds it — so a reader can trace any single number back to the public file it was read from, and reproduce it. Nothing renders on a study page unless a displayable source record stands behind it. We report the federal record and the gaps in it; we do not fill those gaps with estimates and we do not assert a provider fact a public source cannot support.

That standard is what makes the studies citable. Each ships with a frozen snapshot date, the SQL-level joins behind every figure, a machine-readable dataset record, and an append-only corrections log — history is never overwritten, only added to. When a federal file updates, the next snapshot carries the new figure and the old one stays on the record with its date. Read the full standard in the methodology, or trace any field through the chain-integrity log.

Reproducibility

How a study gets built

A study begins with a snapshot: a federal file is pulled on a known date and frozen, so the numbers cannot move underneath the analysis after publication. From that frozen copy, every figure is produced by an explicit join — the exact rule that turns raw rows into the reported result, written down rather than implied. Those joins are published with the study, so a reader with the same public file can run them and land on the same number. When a figure depends on more than one source, the join names each one and how they were matched, because a quiet mismatch between two federal files is the most common way an honest number goes wrong.

What we publish is deliberately narrower than what the data could be made to say. Where a federal file carries a suppression flag, a lagged reporting window, or a known coverage gap, the study names the limitation in plain language and declines to estimate past it. A headline that does not survive its own limitations section does not run. That is the difference between a figure that reads well and a figure a reader can stand behind.

The registry

What earns a place in the dataset registry

The 35 dataset records at /data are not a scrape of whatever was reachable. A source earns a place in the registry only when it clears a short, public test: it is published by a named federal body, it can be pulled and re-pulled on a stated cadence, it carries dates so a snapshot means something, and its fields can be displayed without inventing a claim the file does not make. Sources that fail that test — paywalled rosters, jurisdiction-by-jurisdiction lookups with no bulk path, files with no refresh — stay out, even when they would be convenient. The registry grows source family by source family, and a study cites one only once a dated snapshot stands behind it. The full set of registered sources, with each one’s publisher and cadence, is listed on the sources page.

Upcoming research

Specialist Access Maps — planned across Healthcare access, Care quality.