Methodology

1. Fetch

A connector reads directly from the agency’s own API or bulk file. Nothing passes through a third-party aggregator, so there is no intermediary to introduce an error or a licence restriction.

2. Validate, before anything is written

A release must clear these before it may replace what is already published:

• Row count must be at least 80% of the previous release. A truncated download fails here.
• Required identifier and name fields present on at least 95% of rows.
• No duplicate canonical IDs.
• No going backwards: a release older than the one already published is skipped.

A failed run writes nothing and does not alter the live data. The previous version keeps serving. This is the single most important rule in the pipeline — a silent partial write is far more damaging than a stale page.

3. Normalise

Names are title-cased, addresses are kept as published, and each record is matched to its county and state using Census FIPS codes. Matching is by identifier first, then normalised name plus state. Anything below 90% confidence is left unmatched rather than guessed — currently 127 of 5,419 hospitals, mostly overseas territories and Virginia independent cities.

4. Compute

Only these are calculated by this site, never by the agency:

• National rank — standard competition ranking, honouring each measure’s direction. For a measure where lower is better, zero ranks first. The number of records sharing the same value is always shown alongside the rank.
• State median — the middle value across that state’s records.
• Percentage change — nominal, between the two periods named on the chart. Nothing is inflation-adjusted, because adjusting a published figure would mean this site had edited it.
• Data completeness score — 0 to 1, described below.

5. Publish, conditionally

Each record gets a completeness score from the share of measures present, whether it resolved to a geography, and how current the release is. Below 0.55 the page still renders and still passes links onward, but it carries noindex and is absent from the sitemap. Around one hospital in five currently sits below the line.

• No interpolation. A missing value renders as an em dash, never as zero.
• No smoothing or seasonal adjustment.
• No merging of a figure from one agency into a series from another.
• No generated prose. Summary sentences are conditional templates filled from the figures on the same page, so every clause is checkable against the table below it.
• No inferred ranks. If an agency does not publish a value, no rank is assigned.

• Per-resident spending divides by the 2020 decennial count, the population USAspending carries. It overstates per-head figures wherever a place has grown since.
• Federal obligations are recorded against a place of performance, which often reflects where a contractor is headquartered rather than where work is done.
• CMS quality measures are self-reported and vary in completeness by facility size. Low-volume hospitals report fewer measures — a reporting artefact, not a quality signal.
• County resolution is unavailable for 127 facilities, listed with their raw county name.