Methodology and Full Findings

How to use this appendix

This appendix is the technical record behind the seven-page brief. The brief carries the findings; this document carries the definitions, the data universe, the per-finding methods, the sensitivity tests, the scorer validation, the null model that removed a finding, and the complete state tables.

Definitions. Matched cohort: facilities matched on rurality, size band, acuity band, ownership group, and payer band, widened progressively until at least 20 peers exist. Material flip: a verdict that crosses the cohort median and moves at least 10 percentile points. Fee-for-service Medicare: Title XVIII days paid under traditional Medicare, excluding Medicare Advantage.

Data universe. 13,970 Medicare-certified SNF cost report filers; 91.2 percent anchored on FY2024 filings; medians and percentiles, never averages; real, suppressed, and not-reported values never flattened.

The finding sections are numbered to match the brief. Each carries its claim, its method note, and its exhibits. The validation sections that follow the methods chapter hold the permutation null model, the PBJ period alignment, the materiality tiers, the scorer confusion matrix, and the anchor mix by state. The state tables close the document, and the version page records corrections.

Four things are wrong at once

A public SNF filing can be completed exactly as instructed and still hand management the wrong conclusion.

Under those four conditions, the signal a nursing home receives from public data is incomplete until the payer definition, the peer group, the rating rules, and the reporting conventions are made visible.

This study covers 13,970 skilled nursing facilities, anchored on FY2024 federal cost reports. It is not a collection of benchmarks. It is an argument about why the benchmarks in circulation do not work.

The deepest chapters of the best-known national SNF benchmarking reports are Medicare chapters. Medicare cost per patient day. Medicare length of stay. Medicare therapy cost. The median American nursing home draws 7.3 percent of its resident-days from fee-for-service Medicare and 62.7 percent from reported Medicaid. The industry has built its entire measurement apparatus around the smallest thing it does.

That is the first problem. There are three more, and each one undermines the fix for the last.

The peers are wrong. When you judge a facility against its census region instead of against facilities that actually resemble it, 63.6 percent of American nursing homes get the opposite verdict on at least one core metric. Not a different number. The opposite verdict.

Peer selection changes the verdict on at least one metric for 63.6 percent of facilities. On staffing alone, 26.0 percent experience a material reversal that crosses the median and moves at least ten percentile points. The two figures answer different questions.

The scorecard is wrong. Nursing homes deliver between 16.0 and 19.3 billion dollars a year of nurse staffing that the Five-Star rating cannot see, and 2,545 facilities could buy a perfect staffing score, spend 3.31 billion dollars doing it, and not move their overall rating by a single star.

And the data underneath all three does not mean what it says. In Vermont, nineteen of thirty-two nursing homes report zero percent Medicaid to the federal government. They are full of Medicaid residents. The cost report has a column named Medicaid and that column does not contain Medicaid. Separately, the two federal staffing datasets disagree about the same building in 35.8 percent of nursing homes, and the disagreement runs at 99.0 percent inside one national chain and near zero inside others. What the federal record says about a nursing home depends on which state it sits in and who signs its checks.

Under those four conditions, the signal a nursing home receives from public data is incomplete until the payer definition, the peer group, the rating rules, and the reporting conventions are made visible.

That is not a rhetorical flourish. It is the finding.

How to read this study

An average is the one statistic guaranteed to describe no actual facility.

We publish medians and percentiles. We do not publish averages. Cost report data is fat-tailed, and an average is the one statistic guaranteed to describe no actual facility. Where a distribution matters, we print the distribution.

Every number in this study carries an n and a coverage percentage. Where a figure could not be computed from data that exists, it is absent. Nothing here is imputed, modeled from a national average, or filled forward from a prior year.

The comparison anchor is FY2024. That is a cohort anchor, not a claim that every facility filed a FY2024 cost report. Each facility carries its own latest settled report. 91.2 percent of the universe is anchored on FY2024, 4.9 percent on FY2023, and 3.9 percent earlier. The per-state mix is published in the method section, because in a handful of states FY2024 coverage is thinner than the national figure suggests and you are entitled to know that before you use the number.

Cost report figures are each facility's latest settled filing. Care Compare and Payroll-Based Journal figures are the current snapshot. Those two clocks do not run at the same speed, and where that matters to a finding, we say so on the page rather than in a footnote.

No state median is published where fewer than 75 percent of eligible filers have reported. Where the gate blocks a figure, the cell is empty and the coverage is printed. An empty cell is a finding.

Nothing in this study is a causal claim. Where two things travel together, we say they travel together.

The industry benchmarks fee-for-service Medicare. It is 7.3 percent of the days.

The median facility draws 7.3 percent of its resident-days from fee-for-service Medicare and 62.7 percent from reported Medicaid.

The industry measures the thing with the good documentation and calls it a benchmark.

Now look at what a nursing home actually is. Across 13,970 facilities, the median Medicare fee-for-service day share is 7.3 percent. The tenth percentile is 2.4 percent. Even at the ninetieth percentile it is only 20.1 percent. Medicaid is the business, at a median of 62.7 percent of days, with private and other pay filling the remaining 27.3 percent.

That 62.7 percent is a floor, not a measurement, and Finding 5 explains why. The federal cost report books only Medicaid fee-for-service in its Medicaid column. In states that run long-term care through Medicaid managed care, and most of them now do, those days land in the private and other bucket instead. The true Medicaid dependence of this industry is higher than 62.7 percent, and nobody can say by how much. Which makes the point of this finding stronger, not weaker.

The reason the industry benchmarks Medicare is not that Medicare is large. It is that Medicare is legible. Medicare has a cost report worksheet, a per diem, a case-mix model, and a public dataset. Medicaid has fifty-one state programs, a supplemental payment layer nobody publishes, provider taxes, and directed payments. So the industry measures the thing with the good documentation and calls it a benchmark.

63.6 percent of nursing homes get the opposite verdict depending on who they are compared to

63.6 percent of American nursing homes are told the opposite thing about at least one core metric depending on whether they are benchmarked against their census region or against facilities that actually resemble them.

A biased benchmark can be corrected. A benchmark that is wrong in both directions at once cannot be corrected by anyone.

So we ran every facility in America twice. Once against its census region, the way the published reports do it. Once against a cohort matched on five dimensions: rural or urban, bed size band, acuity, ownership type, and Medicaid dependence, widened until at least fifteen genuine peers exist.

Then we counted the facilities that cross the median line. Above the median under one framing, below it under the other. Not a different percentile. A different answer to the only question a benchmark exists to answer: am I better or worse than my peers?

63.6 percent of 13,365 facilities flip on at least one core metric. Against state benchmarks rather than regional ones it is worse, at 70.6 percent. The median absolute percentile shift between the regional read and the true-peer read is 8.44 points.

A biased benchmark can be corrected. You learn the tilt and you lean against it. A benchmark that is wrong in both directions at once, at roughly equal rates, cannot be corrected by anyone, because there is no way to know from the inside which kind of wrong you are holding.

Take Lawrence Hall Health and Rehabilitation in Arkansas. Rural. 66.1 percent Medicaid. Four of its six core metrics reverse verdict between the regional framing and the true-peer framing. That facility has, at some point, sat in a board meeting and been told how it compares to other nursing homes. There is a better than even chance it was told the opposite of the truth on the metric that mattered.

And it is not evenly distributed across metrics. Cost lines are relatively stable. Staffing is not. Total nurse hours per resident-day flips for 27.1 percent of facilities, and non-nursing operating cost flips for 20.1 percent. Those are the two levers an operator actually pulls.

Cohort match on rural, size band, acuity band, ownership group, payer band, with progressive widening to a minimum of 15 peers. Regions use the eight-region grouping standard in published SNF benchmarking. A flip is a crossing of the cohort median. n varies by metric because coverage varies; every n is printed.

The overall star rating is insensitive to 16 billion dollars of nursing care

Nursing homes deliver 16.0 to 19.3 billion dollars a year of staffing that their overall star rating does not register.

It is not a savings opportunity. It is a measurement of the resolution of the instrument.

The Five-Star staffing rating runs on cut points. Cross a threshold, gain a point. Sit anywhere between two thresholds and the rating is indifferent to where you sit.

So we asked a question the rating system has never been asked. For every facility in America, what is the least staffing it could carry and still hold the overall star rating it has today? Call that the hold floor. Then price everything above it.

The answer is 16.0 to 19.3 billion dollars a year, across 11,874 facilities. That is care being delivered, at the bedside, on the payroll, that the federal scorecard is structurally incapable of registering. The range reflects whether you price those hours at the facility's employed wage or its all-in wage including purchased nursing.

The Five-Star staffing domain is a six-measure, coarse-grained rating built to sort facilities into five buckets. It was never designed to be an accounting of care delivered. But it is the instrument the entire sector is judged by, the one families read, the one referral partners screen on, and increasingly the one state Medicaid programs pay against. And it is blind to sixteen billion dollars.

Apply a prudence floor, and never model a facility below the numeric minimums CMS rescinded in February 2026, and the figure falls to 2.2 to 2.8 billion across 4,743 facilities. That is the conservative read and it is still a very large number.

The finding cuts the other way too, and this half deserves as much attention. 582 American nursing homes are staffed thinner than their own overall rating implies. The rating is telling families those buildings are better than the staffing data says they are. Blindness is not a one-directional error.

Five-Star staffing scored from the frozen CMS Technical Users' Guide. The scorer reproduces published Care Compare overall ratings at 100 percent and staffing stars at 93.8 percent. Hold floor is the minimum staffing preserving the current overall rating. Hours priced at reported wages. The prudence-floor variant never models below 0.55 RN and 3.48 total HPRD, the minimums rescinded effective 2026-02-02, which are not current law. No numeric federal HPRD minimum is in force. State minimums were not evaluated.

A staffing star costs $485,511 a year. For 2,545 facilities it cannot move the rating.

The next staffing star costs $485,511 a year at the median. For 2,545 facilities, a perfect staffing score would not move the overall rating at all.

You can purchase a perfect five-star staffing rating and watch your overall rating sit exactly where it was.

It costs $485,511 a year at the median. That is $17.48 per resident-day. Climbing all the way to a five-star staffing rating costs $1,101,243 a year at the median.

But the more useful half of this finding is the half that tells 2,545 facilities to stop.

The overall Five-Star rating is not an average of its parts. The health inspection rating sets the base, and staffing and quality can only move you off that base within limits. Which means that for a facility whose survey performance is the binding constraint, buying staffing does nothing. You can purchase a perfect five-star staffing rating and watch your overall rating sit exactly where it was.

2,545 facilities are in that position. 18.2 percent of the entire industry. If every one of them bought the staffing star they are chasing, they would spend 3.31 billion dollars collectively and move the number the public actually sees by zero.

Some of them are spending toward it right now. They are solving the wrong problem, expensively, and no benchmark in circulation is capable of telling them so.

Climb cost from the cohort staffing ladder, priced at reported wages. The unbuyable set is computed by applying the CMS overall-rating composition rules against each facility's own health inspection base. A facility is unbuyable if a 5-star staffing rating produces no change in overall rating.

The legacy Medicaid column does not contain Medicaid

Nineteen of Vermont's thirty-two nursing homes report exactly zero Medicaid days to the federal government. They are full of Medicaid residents.

The days did not disappear. They were reclassified by a form.

Vermont's median nursing home reports a Medicaid day share of 0.0 percent. Not suppressed. Not missing. Filed, as a real value, by facilities whose residents are overwhelmingly on Medicaid. Nineteen of the state's thirty-two nursing homes report exactly zero Title XIX days.

California reports 7.2 percent. Medi-Cal does not cover seven percent of California nursing home days.

Neither state is lying. Both are filling out the form correctly.

The CMS-2540 cost report, Worksheet S-3, splits resident days by payer. Its Medicaid column captures Title XIX fee-for-service. Medicaid managed care days do not go there. They fall into Other, the bucket everyone reads as private pay.

Vermont runs essentially all of its long-term care through Medicaid managed care. So Vermont's Medicaid column is zero and its Other column is 82.4 percent. California runs Medi-Cal managed care, so its Medicaid column reads 7.2 percent and its Other column reads 69.3 percent. The national median for Other is 27.3 percent.

The days did not disappear. They were reclassified by a form.

This study is not exempt. The 62.7 percent national median Medicaid day share in Finding 1 is a floor. We do not know the true figure and we are not going to estimate one, because estimating it would require assuming what fraction of each state's Other bucket is managed Medicaid, and that assumption would be doing the work rather than the data.

What we will say is this. If you benchmark a nursing home's payer mix against facilities in other states, you are partly measuring how far along those states are in moving Medicaid to managed care. That is not a fact about how anyone runs a building.

CMS agrees. Transmittal 1-Revised, issued November 27, 2024, revises Worksheet S-3 to collect Medicare and Medicaid HMO inpatient days and discharges, effective for cost reporting periods ending on or after September 30, 2025 (Form CMS-2540-24).

CMS-2540 Worksheet S-3 payer split. All Vermont and California values carry value_status real_value; none are suppressed or imputed. A full audit of the state table confirms zero nulls rendered as zeros across all eighteen columns. This is a structural property of the federal instrument, not a defect in this pipeline or in the filings.

The two federal staffing datasets disagree about the same building. And the disagreement runs by owner.

In 4,684 American nursing homes, the Payroll-Based Journal shows essentially no contract nursing hours while the cost report carries purchased-nursing dollars. The disagreement runs at 99.0 percent inside one national chain and near zero inside others.

One dataset says this building uses no agency. The other says it is paying for it.

The Payroll-Based Journal takes daily hours, split between employees and contractors, auditable against payroll. The Medicare cost report takes annual dollars, with purchased nursing services landing in a different column from salaries. The two are never reconciled against each other by anyone, because they live in different agencies' pipelines and no public tool has ever held both at facility level.

We held both. In 4,684 facilities, 35.76 percent of the 13,098 where both sources are present, the Payroll-Based Journal reports one percent or less of nurse hours as contract labor while the cost report carries enough purchased-nursing dollars to lift the all-in nursing wage at least five percent above the employed wage. The median premium is 11.3 percent. Nationally the wage delta sitting in that gap is 2.3 billion dollars a year.

One system reports almost no contract labor. The other records material purchased nursing services.

Our first pass at this number, computed on an older cost report substrate, returned 38.71 percent. On the FY2024 substrate it is 35.76 percent. About three points of the original figure was the calendar, not the discrepancy. The rest is real.

This does not smear evenly across the industry. It clusters, hard, by parent organization. PruittHealth: 96 of 97 facilities. Creative Solutions in Healthcare: 121 of 132. Life Care Centers of America: 119 of 192. The Ensign Group: 205 of 335. PACS Group: 88 of 278.

A national rate of 35.8 percent and a chain rate of 99.0 percent are not the same phenomenon. If this were an accounting convention baked into the cost report form, it would appear everywhere at roughly the same rate. It does not. It appears where a particular corporate parent's accounting function decides how a nursing hour gets recorded.

That is a question, not an accusation. It has a boring answer available: a management company that provides nursing staff to its own facilities can produce this pattern entirely legitimately, because the hours are employees of the parent and the dollars are purchased services to the facility. It also has less boring answers. We are not in a position to say which, and neither is anyone else, from public data.

But here is what it means for benchmarking, which is the point of this study. If you compare a facility's labor cost to its peers, and a meaningful share of those peers record nursing labor through a different convention than your facility does, then the comparison is not measuring performance. It is measuring corporate structure.

What the federal record says about a nursing home depends on who owns it.

Seam A: PBJ contract share of nurse hours at or below 1.0 percent AND all-in nursing wage at least 5 percent above employed wage. Employed wage is Worksheet A column 1 salaries over PBJ hours. All-in adds column 2 purchased services. Column 2 carries agency and other purchased nursing, and is not exclusively agency. The reverse seam, PBJ showing 5 percent or more contract hours with no cost report premium, occurs in 77 facilities, 0.59 percent. Only one frozen PBJ vintage exists, so quarter-level alignment to each facility's cost report period is not possible; the fiscal-year anchor split is the available timing control.

The signal is incomplete

Four independent failures, each of which would be serious alone, and they compound.

The signal is incomplete until the payer definition, the peer group, the rating rules, and the reporting conventions are made visible.

The payer is wrong. The industry measures Medicare, which is 7.3 percent of the days.

The peers are wrong. 63.6 percent of facilities get the opposite verdict depending on the comparison set.

The scorecard is wrong. It cannot see 16 billion dollars of delivered care, and it charges 2,545 facilities 3.31 billion dollars for a rating that will not move.

And the data underneath all three does not mean what it says. A column named Medicaid does not contain Medicaid. Two federal reporting systems disagree about the same building. A cost structure decided at a corporate office is filed as a fact about a facility.

These do not cancel out. They compound. The result is not that public data provide no signal. It is that the signal is incomplete until the payer definition, the peer group, the rating rules, and the reporting conventions are made visible. That is what this appendix exists to make visible.

None of these findings requires a single filing to be wrong. A form completed exactly as instructed can still support the wrong management conclusion.

The 2,545 facilities buying an unbuyable star are the clearest illustration. They are not lazy or badly run. They are responding rationally to the only information anyone has given them. The information is wrong.

There is no policy recommendation at the end of this study. We are not a trade association and we are not asking Congress for anything. This is a measurement problem, and the fix for a measurement problem is better measurement.

What we would say is this. If you operate a nursing home, the number you have been using to tell your board how you are doing is more likely to be wrong than you think, and you can find out which kind of wrong in an afternoon.

Start with your own building.

Every state, every core metric, FY2024

One row per state. Medians, not averages. No median is printed where fewer than 75 percent of eligible filers have reported, and where the gate blocks a figure the cell is empty and the coverage is stated.

FY2024 anchor coverage varies by state and it is printed in the method section. Colorado, Oregon, Vermont and Alaska sit below 86 percent FY2024 coverage. Read those rows accordingly.

Method, vintage, and what we will not do

Universe: 14,064 skilled nursing facility filers. 94 dropped as analysis-ineligible under the standing exclusion rule. Analytic universe 13,970.

Comparison anchor: FY2024. Each facility carries its own latest settled cost report. 91.2 percent of the universe is anchored on FY2024, 4.9 percent on FY2023, 3.9 percent earlier. The full per-state anchor mix is published below and is not summarized away.

Sources and vintages. Cost reports: HCRIS, rolling recut of 2026-07-07. Quality and staffing ratings: CMS Care Compare, 2026-06-23. Daily staffing: Payroll-Based Journal, 2026-06-27. Rates: PDPM FY2020 through FY2024, SNF wage index by county, Medicaid rates from MACPAC 2019.

That last one is a known weakness and we are going to say so rather than bury it. The best public state Medicaid rate data available is from 2019. The effective Medicaid payment a nursing home actually receives is base rate plus supplemental payments plus managed care directed payments, less the effect of provider tax financing. No public source assembles it. MACPAC has told CMS as much. Until that changes, any claim about Medicaid rate adequacy, ours included, rests on a five-year-old snapshot. We use it where it helps and we do not build a headline on it.

Peer cohorts match on five dimensions: rural or urban, bed size band, acuity band, ownership group, and payer band. The cohort widens progressively until at least fifteen genuine peers exist. Where it cannot reach fifteen, the facility is flagged thin and its percentiles are not reported as if they were stable.

Wage standardization re-prices the labor-related share of cost to the national wage index. We use it to rank, because geography is not performance. We do not use it to dollarize, because a facility cannot spend a wage-standardized dollar.

The Five-Star scorer used in this study reproduces published Care Compare overall ratings at 100 percent and staffing stars at 93.8 percent, scored from the frozen CMS Technical Users' Guide.

Real values, suppressed values, and not-reported values are three different facts and are never flattened into one. Where a value is suppressed, we print the reason.

Nothing in this study is imputed. Where a figure could not be computed from data that exists, it is absent, and its absence is stated.

This study makes no causal claims. Where two things travel together, we say they travel together.

One independent check worth stating. Our median Medicare fee-for-service day share, computed from the cost reports, is 7.3 percent. MedPAC, working from claims and by a different method, reports 8 percent for the median freestanding facility. Two unrelated pipelines landing that close is the best evidence we can offer that the spine is sound.

The finding that did not survive

An earlier draft carried a seventh finding: related cost center pairs flagged together at 9.9 to 10.8 percent, read as a coding pattern. A random-pair permutation test returned a null rate of 10.81 percent against an observed 10.78 percent, P equal to 0.519, and the artifact held at every alternative threshold.

The finding was removed from the study. The test that removed it is published here.

PBJ period alignment on the seam

The seam rate in Finding 6 compares a calendar-year payroll system to fiscal-year cost reports. On the period-aligned subset, FY2025-anchored facilities against the CY2025 Payroll-Based Journal, Seam A runs 43.2 percent at the published thresholds, against 35.76 percent unaligned.

Full multi-year alignment requires the quarterly PBJ files from Q2 2020 through CY2024, which are not yet in the frozen raw layer. That ingestion is named, not skipped.

Materiality tiers on the peer flip

A median crossing can be a hair's width. The tiers below require the flip to be material: at least ten percentile points of movement, or a crossing between quartiles.

The brief's Finding 2 headline is the material flip on total nurse hours per resident-day: 26.0 percent.

Scorer validation and the restricted star findings

Every star finding is computed by reproducing the CMS staffing-star scorer. The reproduction matches the published staffing star exactly for 13,075 of 13,943 scored facilities, 93.8 percent.

Restricted to that exact-reproduction subset, the unbuyable-star population is 2,545 facilities and 3.31 billion dollars, against 2,717 and 3.41 billion unrestricted. The study prints the restricted figures.

Two counts published in an earlier draft, the facilities within five and within ten points of the next staffing star, failed an internal consistency check (a within-ten count cannot be smaller than a within-five count) and are withdrawn pending recomputation.

Anchor mix by state

Every facility is anchored on its most recent usable cost report. The mix below shows how current each state's anchor set is.

Where your building sits

It can tell you that 2,545 facilities are chasing an unbuyable star. It cannot tell you whether you are one of them.

This study can tell you that 63.6 percent of American nursing homes are getting the opposite verdict from their benchmark. It cannot tell you whether you are one of them.

It can tell you that the staffing seam appears in 35.8 percent of facilities and runs at 99 percent inside one chain. It cannot tell you whether it appears in your building.

It can tell you that 2,545 facilities are chasing an unbuyable star. It cannot tell you whether you are chasing one.

Those questions have answers, and they are facility-specific. The Astrelis SNF Benchmarking Report answers them for one building: your true peer cohort, your position in it, the dollar value of every gap, the buckets that are clean, and the ones that are a coding question rather than a performance question. Thirty pages, and the arithmetic reconciles to the dollar.

$2,500.

Version and corrections

Version 1.0, published 2026-07-13. Anchored on the FY2024 cost report cycle; public data through June 2026.

Corrections contact: mattborchardt@astrelis.co.

Material corrections to any figure in the brief or this appendix will be logged on this page, with the prior value, the corrected value, and the date.