Where every number comes from, and what it cannot tell you.

CMS withholds values where case counts are too small to report reliably. The rule is not one universal threshold, and we do not describe it as one.

• For the 30-day readmission measure, CMS assigns facilities with fewer than 25 eligible cases to a “too few to reliably report” category rather than publishing a rate.
• For the chart-abstracted measures — HBIPS-2 and HBIPS-3 among them — small-count suppression follows the CMS and Joint Commission measure specifications for that measure.

In either case the cell is unknown. It is never treated as zero, never imputed, and never dropped without notice. Every suppressed cell in a report is marked and footnoted, and any aggregate computed over a column containing suppressed cells states how many were excluded.

Restraint (HBIPS-2) and seclusion (HBIPS-3) are reported per 1,000 psychiatric inpatient patient-hours. That is an exposure denominator, not a per-patient one: a facility with longer stays accumulates more patient-hours, and the rate accounts for it. Readmission (READM-30-IPF) is measured per eligible discharge.

Rates from these two families are not interchangeable and are never combined. Every table shows the reporting volume behind a rate, so the sample size is visible next to the number.

Two different numbers are both correctly called “the national rate,” and they are not equal.

• The CMS-published national rate is case-weighted. Large facilities move it more than small ones.
• The facility-level median is the middle value across reporting facilities. Each facility counts once, regardless of size.

Both appear in the reports, each labeled with which one it is. They are never averaged together, and a comparison against one is never presented as a comparison against the other.

Percentile ranks are computed at the facility level. Each reporting facility is counted once. Facilities whose value is suppressed for that measure and period are excluded from the population rather than assigned a placeholder value, and the size of the population used is stated alongside the rank.

Flags mark findings worth examining. They are described qualitatively here because the report labels each flag with its own basis, and because a flag is a pointer to a source, not a substitute for it. The bases in use are:

• A restraint rate at or above the CMS-defined questionable restraint threshold.
• A restraint or seclusion rate at or above twice the same-year national facility median — the middle value across reporting facilities, not the case-weighted CMS-published rate. The two are different numbers; every comparison names which one it uses.
• Critically low floors on follow-up after hospitalization and on medication continuation.
• Assignment by CMS to the “Worse than the national rate” readmission category.

A deteriorating trend is deliberately not on that list; see the next section.

The last of these is a CMS-PUBLISHED FINDING and is directly citable. The others are PRODUCT-COMPUTED and are labeled as such wherever they appear. Context supplied by an operator — a press statement, a corporate integrity agreement summary — carries VERIFY BEFORE USE.

A run of consecutive worsening years looks like evidence. Usually it is not, and it is easy to check. Take a facility’s own five reported values, shuffle them into a random order, and count how often a run of a given length turns up.

• A three-year run appears in 5 shuffles out of 12. It is the single most likely outcome of pure chance.
• A four-year run appears in about 1 shuffle in 13.
• A five-year run appears in about 1 in 120.

Across CY2023–CY2026 CMS data, three-year worsening runs occur in 18–30% of all reporting facilities on every measure we track, against 15–27% expected by chance alone. On medication continuation they occur less often than chance.

So we do not report three-year runs. A run spanning four or more reported years is printed as context, not as a finding, with the chance figure stated beside it, and it is excluded from the litigation-relevant flag count. The five-year trend table still shows every value.

This removes findings that a competent expert would take apart on cross-examination. It is why a report from us carries fewer flags than one assembled from the same public data, and why the flags that remain are worth citing.

Facilities are attributed to an operator by matching that operator’s own published facility directory against CMS records, then corroborating each match by city and street address. Names alone are not sufficient: unaffiliated facilities share names, and operators rename facilities after acquisition.

Matches that remain ambiguous after address corroboration are sent to human review rather than guessed. A facility claimed by two operators’ directories is never resolved automatically: a joint-venture hospital has one street address and two corporate parents, so the address cannot decide between them.

Every attribution is reproducible. Each records the operator directory row that produced it, the CMS record it matched, and whether city or street address corroborated the match. An attribution whose evidence cannot be reproduced from the source files is removed before publication.

Facilities that cannot be confidently attributed are counted and disclosed in the coverage statement. A chain report states how many facilities are covered, how many are not, and why.