The uneven recovery of nursing-home census

Read those against each other, not against the national figure. The tier indices exclude 964 facilities that reported a census in 2020 but were never given a metropolitan flag, almost all of which left before CMS began populating that field. The exclusion is one-sided and lifts every tier: the national index is 95.3 across all facilities and 99.3 across the flagged ones. The gap between the tiers is unaffected. Their distance from the national number is not.

Outside work reaches the same gradient by a different route. McGarry BE, Wilcock AD, Gandhi AD, Grabowski DC, Geng F, Barnett ML. Changes in US Skilled Nursing Facility Capacity Following the COVID-19 Pandemic. JAMA Internal Medicine 2026;186(3):285-292. measured operating capacity, the highest occupancy a facility accommodated in the prior 60 days, rather than census. Over 2019 to 2024 they found licensed beds down 2.5% and operating capacity down 5.0% net, with 14% of rural counties losing a quarter or more of capacity against 6% of urban ones. Their 5.0% and the 4.7% here are different quantities on different windows, so the two do not replicate each other. What they share is direction and the rural tilt, from an independent measure.

What this cannot tell you is why the buildings went up where they did. The decomposition is an accounting identity, not a mechanism: same-store plus openings minus closures equals the net change by construction, whatever the reasoning of whoever built. Whether metro development responded to residents leaving existing homes, or to land, labor, payer mix or capital availability, is outside this data.

Nor can we say these are new buildings. What is observable is certification, not construction. Of the 331 facilities that begin reporting a census over the window, 81% hold a first Medicare approval date inside it. The remainder were certified earlier and simply were not reporting a census at the start, so they are re-entries rather than additions. And of the ones approved inside the window, nothing here separates a newly built facility from an existing building newly participating in Medicare. The column is labelled newly certified for that reason.

Robustness: state composition

This is the finding we stress-tested hardest, because a gradient this clean is often two or three states driving an average. This one is not. Rural trails urban in 39 of the 48 states with at least 150 residents in each tier. Remove the three worst and the gap narrows from 3.8 points to 3.5. Remove the five worst, 3.4. Remove the ten worst and rural still trails by 2.7 points. No small group of states accounts for it.

Robustness: facility type

Rural nursing homes are not a random sample of nursing homes. They are twice as likely to sit inside a hospital (6.3% against 3.3%), more than twice as likely to be government-owned (11% against 5%), and much less likely to be part of a continuing care retirement community (6.9% against 11.9%) or to be Medicare-only (0.5% against 4.9%). Any of those could be the real driver. A hospital swing-bed unit and a 240-bed suburban facility are different businesses.

So we removed them, one at a time and then all together:

Facilities included	Urban	Rural	Gap	Count
All (as published)	100.0	96.2	+3.8	14,727
Excluding CCRCs	100.6	96.6	+4.0	13,177
Excluding hospital-based units	100.0	96.3	+3.8	14,119
Excluding government-owned	100.4	96.7	+3.7	13,773
Excluding Medicare-only	100.0	96.2	+3.8	14,182
Excluding under 30 beds	100.1	96.3	+3.9	14,374
Excluding all five at once	100.8	97.2	+3.6	11,550

Dropping all five simultaneously removes 3,177 facilities, more than a fifth of the panel, and the gap moves from 3.8 points to 3.6. The gradient survives all of them. Free-standing, privately owned, dually certified nursing homes of ordinary size show the same gap the full panel does.

Ownership explains part of it

Ownership is the one test that does not come back clean. Rural counties hold fewer for-profit homes than urban ones, 67% against 77%, and more than twice the share of government-owned facilities, 11% against 5%. That matters because the three ownership classes did not recover alike, and the differences between them are larger than anything geography produces:

Ownership	Urban	Rural	Gap	Facilities
For profit	102.2	98.3	+3.8	10,899
Non profit	92.4	90.9	+1.4	2,874
Government	94.0	92.3	+1.7	954

A for-profit home in an urban county sits at 102. A non-profit in an urban county sits at 92. About 10 index points separate them, against the 3.8 that separates urban from rural. On this evidence ownership is the stronger predictor of the two, which is not a finding this study set out to make and is the one most worth following.

The urban-rural gap does survive inside every class, at 1.4 to 3.8 points, so it is not ownership in disguise. But it narrows: standardizing the rural mix to the urban one takes the gap from 3.8 to about 3.3. Roughly an eighth of the raw gap is rural counties holding more of the ownership types that recovered worst. The remainder is geography.

Robustness: weighting

Every index here is a ratio of sums, so large facilities carry more of it than small ones. Comparing facilities present in both periods, the gap is 3.3 points weighted by census, 4.1 as an unweighted mean of facility-level change, and 3.0 at the median facility. It is not an artifact of the big buildings; the unweighted version is larger.

Robustness: facility size

Rural homes are smaller, with a median of 81 certified beds against 108 in urban counties. If small facilities recovered worse, “rural” could be standing in for “small” throughout this study.

The obvious objection is that this is the hospital units again, since a distinct-part unit inside a hospital is usually small and rural facilities carry twice as many of them. It is not. Excluding them, the medians are 109 urban against 83 rural, which is one bed away from the figures above, and the rural share of each size band moves by at most a point. Hospital units are smaller in both tiers, 80 beds urban and 52 rural, but at 9% and 13% of the under-60 band they are too few to move a median. Rural nursing homes are smaller as free-standing facilities.

Small facilities did recover worse. So did the largest ones:

Certified beds	Facilities	Census index	Occupancy then	Occupancy now	Beds
Under 60	2,693	97.2	81.5%	83.2%	−4.7%
60 to 99	4,627	100.7	81.0%	81.8%	−0.3%
100 to 149	5,016	100.4	80.6%	79.8%	+1.5%
150 or more	2,387	97.3	82.1%	79.6%	+0.3%

The pattern is an arch, not a slope. Both ends of the size range sit at 97 while the middle is back above 100, so there is no simple “bigger is better” effect for rural to be hiding behind. The occupancy columns say why the two ends got there differently. Small homes cut 4.7% of their beds and their occupancy rose to 83.2%, the highest of any band. The largest homes added beds and their occupancy fell 2.5 points to 79.6%. The two ends of the size range have different problems.

And the urban-rural gap holds inside every band:

Certified beds	Urban	Rural	Gap	Rural share of band
Under 60	99.4	94.0	+5.4	40%
60 to 99	101.8	98.6	+3.2	35%
100 to 149	101.6	95.5	+6.1	22%
150 or more	97.5	94.1	+3.5	10%

A rural home trails an urban home of the same size in all four bands, by between 3.2 and 6.1 points, and the gap is widest in the 100-to-149 band where rural facilities are only 22% of the total. A 120-bed nursing home recovers differently depending on the county it sits in. This holds the business model roughly constant and varies only geography.

Holding everything constant at once

Everything above removes one thing at a time. That shows the gap survives; it cannot say how large the gap is once the other differences are accounted for together, because the differences overlap. A single specification can. The outcome is the change in a facility's own census between pre-pandemic 2020 and Q4 2026, for the 14,246 facilities reporting in both periods. The predictors are whether it is rural, its size, its type, its ownership, and which state it is in, carried as 53 fixed effects. Standard errors are clustered on state, since facilities in one state share whatever that state did.

Specification	Rural	Std. error
Rural alone	−3.80	0.63
Plus state fixed effects	−3.10	0.58
Plus state and facility controls	−4.31	0.63
The same, weighted by census	−4.15	0.64

Controlling for everything at once makes the rural gap bigger, not smaller, which is not what we expected. State fixed effects absorb about a fifth of the raw difference and ownership a further tenth. Size pushes the other way: rural facilities are smaller, and smaller facilities recovered better, so being small had been hiding part of the rural penalty. Holding state, size, type and ownership constant, a rural facility lost about 4.3% more of its census than a comparable urban facility in the same state.

How to read the table: the SE column is the uncertainty around each estimate. An estimate more than about twice its SE is one this data can separate from zero; one smaller than its SE is not.

What else the specification says

Term	Coefficient	Std. error	t
Rural	−4.31	0.63	−6.8
Size, per log unit of beds	−6.77	1.11	−6.1
Non-profit	−9.00	0.83	−10.8
Government-owned	−7.59	2.00	−3.8
Continuing care retirement community	−3.42	1.05	−3.3
Hospital-based	−2.26	1.53	−1.5
Medicare-only	−4.33	3.63	−1.2

Ownership is the largest term in the model. A non-profit facility lost 9 points more than an otherwise identical for-profit one in the same state, and a government-owned facility 7.6 points more. Both are roughly twice the rural effect and both are precisely estimated. The raw ownership split reported earlier was not a composition artifact; it survives controls for state, size and type.

The state effects are wider still. Their standard deviation is 5.3 points and they run from 27 points below the reference state to 4 above. Which state a facility operates in moves its census by more than whether that facility is rural, which is the same conclusion the divergence map reaches by a different route.

What this specification does not settle

It describes, it does not identify. Nothing here is an experiment, and every variable in it could stand in for something unmeasured: ownership for payer mix, state for Medicaid policy, rural for distance to a hospital. The R-squared is 0.070, so almost all of the variation between facilities is unexplained.

The sample is also restricted to facilities reporting in both periods. Entry and exit are 58% of the national shortfall and are absent from the model entirely, so this measures the gap among survivors and says nothing about which facilities left.

One control was dropped deliberately. Including how full a facility was in 2020 moves the rural coefficient from 4.31 to 6.27, which looks like a stronger result and is not one: baseline occupancy is census divided by beds, and the outcome is the change in census, so baseline census sits on both sides of the equation. A facility whose 2020 census is high through reporting noise will mechanically show both higher occupancy and a smaller increase. The 6.27 is partly arithmetic. The figure reported here excludes it.

Median occupancy of each home type, pre-pandemic against the latest quarter. 30% of homes are at least 5 points emptier than in 2020 and 28% at least 5 points fuller. State medians run from Oklahoma's 64.6% to West Virginia's 94.5%. The national 80.4% sits below the median because large homes are emptier than small ones and the national figure counts beds, not homes.

Home type	Pre-pandemic	Latest
All homes	85.4%	85.7%
Dually certified	85.5%	85.8%
Medicare only	81.8%	83.4%
Medicaid only	88.7%	85.2%
Inside a hospital	83.5%	80.8%
Part of a CCRC	86.8%	86.1%
For profit	84.6%	86.1%
Non profit	88.2%	86.2%
Government	83.3%	78.0%
Urban county	86.6%	87.5%
Rural county	81.6%	80.3%

Across 48 states, 39 have urban above rural, with a median gap of 3.8 points. Divergence is the norm rather than the exception. It reaches 32 points in Alaska, where an urban tier at 132 sits against 309 rural residents statewide. The remaining 9 run the other way, with rural ahead of metro, by as much as 8.3 points in Washington.

36 of 48 states are still below their 2020 census. 20 of them shrank across both tiers; 16 shrank in rural while their metros held. Of the 12 that grew, 8 grew on their metros alone and 4 grew across both. Four groups, and it takes both dimensions to put a state in the right one.

Either axis on its own hides half the picture. Colorado and Washington both sit at 94 on total census, so a growth-only chart treats them as the same state, yet Colorado's metros run 12.3 points ahead of its rural while Washington's run 8.3 points behind. It runs the other way too: Wisconsin and New Mexico have almost the same gap, about 1.5 points, yet sit at 79 and 103 on level.

Contracting rows are listed worst first, growing rows strongest first. The two axes cannot share a panel: the gap needs a metropolitan flag, and that column is empty in the 2020 snapshot, so tier work reaches only facilities present in the latest snapshot. Growth is therefore measured on the true total with closures included, as everywhere else in this study. Measured on the flagged panel instead, 22 of 48 states would clear 100 rather than 12, because the homes a state lost are missing from its own baseline.

Agency nursing went from 3.7% of direct-care hours in 2020Q2 to 10.7% at the 2022Q4 peak and back to 5.4% by 2026Q1. That arc is sector-wide and on its own says nothing about census: every tier rose and every tier fell back.

What would bear on census is a difference between tiers, and it runs the wrong way. Rural facilities used less agency than urban through the whole rise and the peak, 9.1% against 11.0% at the top, and the lines did not cross until 2025Q2. If leaning on agency cost a facility census, rural should have leaned harder. It leaned less, while its census fell further.

The joint specification says the same thing. Adding 2022 agency share to the model, refit on the 14,177 facilities with staffing data, moves the rural coefficient from -4.34 to -4.27, about 2% of the gap. The agency term is +0.030 per percentage point against an uncertainty of 0.026, and as a yes or no +1.52 against 0.61. Positive either way, which is the opposite of the hypothesis, and attached to whether a facility used an agency at all rather than how heavily. Capping the top 1% leaves it at +0.033 and dropping every facility above 50% agency at +0.045, so a handful of extreme users is not driving it.

What this cannot say: the staffing series runs 2020Q2 to 2026Q1, so agency is never observed at either end of the census window; agency share is a ratio whose denominator moves with census; and agency use and census could both be responding to local demand, which is not measured here.

The intuitive explanation is that some places are aging faster than others. It does not survive the test. Sorting the 2,513 counties with at least 50 residents into quartiles by growth in the Medicare-eligible population:

County quartile	65+ growth	Census change
Slowest-aging quarter	+3.5%	−7.3%
Second	+9.0%	−5.9%
Third	+13.7%	−4.9%
Fastest-aging quarter	+22.6%	−2.4%

The tilt is real and it runs the right way, but it is weak. A correlation of +0.08 across the window, which is well under 1% of the variation between counties. The second column is the point: even the fastest-aging counties, with 23% more seniors, still lost census. Aging does not account for the differences between counties in this window. One likely reason is that most of the growth in the 65-and-over count sits at the younger end of it, well below the age at which people enter nursing homes, but this study does not measure age within that population and cannot test that.

Census	The sum of each facility's reported average residents per day. A count of people, not of beds.
Occupancy	Census divided by certified beds. Moves when either one changes, so it is independent of census.
Census index	A census series rebased to 100 at its own pre-pandemic value. An index of 96 means that group has 4% fewer residents than it had, not 4% below any other group.
Gap	Urban index minus rural index, in index points, whatever level either reached.
Same-store	Facilities reporting in both periods of a comparison. Entry and exit are excluded, so the change reads as demand.
Newly certified	In the panel at the end but not the start, first approved inside the period. Certification, not construction: it may be a new building or an existing one newly in Medicare.
Exit	Present at the start, absent at the end. Includes closures, terminations and certification changes, which this data does not separate.
Urban and rural	In a metropolitan county or not, by the CMS flag. The three-tier gradient uses county population instead: 250k and above, 50 to 250k, below 50k.
Certified beds	Beds certified for Medicare or Medicaid, and the occupancy denominator. Neither licensed beds nor beds in service.
Hospital-based, CCRC, Medicare-only	CMS facility flags: a unit inside a hospital, part of a continuing care retirement community, certified for Medicare but not Medicaid.

Self-reported, unaudited	Compiled from facility submissions and not audited for this purpose. Across 420,543 facility-quarters, 0.5% carry no usable census and 1.2% report more residents than certified beds, which is impossible. Bed counts include 1 against more than 100 residents. CMS revises and republishes; each snapshot is read as published. Nothing is dropped or imputed, so these errors sit in the totals, bounded by the figures given.
Census is not occupancy	24 states cut beds while occupancy held or rose; California reached 88.5%. The urban-rural result holds on both measures. Individual states do not.
Tier levels sit on a different panel from the national figure	The metropolitan flag is populated only in recent files, so 964 facilities reporting a census in 2020 are absent from the urban and rural figures and none are at the endpoint. The gap survives it; the levels are not comparable to the national index. See the method note.
Two-quarter reporting lag	The 2020 decline appears at the Q4 2020 snapshot. Magnitudes are measured correctly; timing is not.
Entry and exit are panel events	A facility appearing or disappearing, not a building opening or closing. Net change is unaffected either way, as ownership transfers appear on both sides.
Facility mix is tested, not controlled	Type, size, ownership and state are now modeled jointly, and the gap holds at 4.31 points. That is a description, not an identification: every term could stand in for something unmeasured, and the model explains 7% of the variation between facilities.
Nursing homes only	Assisted living, home health and independent living are out of scope and may absorb displaced demand.
No causal claim	Medicaid rates, direct-care wages and state moratorium policy are not in this data.

Source	CMS Provider Data (Care Compare), monthly snapshots
Period	Pre-pandemic 2020 to Q4 2026, 28 quarters
Panel	14,624 certified nursing homes; last snapshot in each calendar quarter
Census	Sum of average residents per day, capped at 110% of certified beds
Index	Each series set to 100 at its own pre-pandemic value
Cap	Binds on 0.43% of rows. The national index is 95.33 uncapped, 95.33 at 110% and 95.34 at 100%, so it is a guard rather than a lever. The tightest cap reading highest is not an error: it trims proportionally more from the 2020 baseline than from the endpoint. The extreme ratios are bad bed counts, which matters more for occupancy than for census.
Baseline	A single pre-pandemic snapshot. A trailing four-quarter blend moves the national gradient by under 0.2 points; a few small state-tiers move up to 4% on 2019 bed-supply changes.
Urbanicity	Fixed once per facility from its latest snapshot, since the CMS flag is populated only in recent files. County population tiers (2023 ACS, 96% join) give a 3.7-point gap between metro and rural; the CMS flag gives 3.8.
Same-store	Matched by CMS certification number. Same-store plus entry minus exit equals net change by construction.
Two panels, not one	The national figures run over every facility reporting a census: 15,360 at the baseline. The urban and rural figures can only run over facilities CMS gave a metropolitan flag, and that field is populated only in recent files, so 964 facilities reporting a census in 2020 have no flag and are absent from the tier panel. At the endpoint 0 are missing. The exclusion is therefore one-sided, and it flatters the tier panel: the same national index is 95.3 over all facilities and 99.3 over the flagged ones. The urban-rural GAP is unaffected, because both tiers are restricted identically and an unflagged facility has no tier to be placed in. The LEVELS are not comparable to the national figure, and should not be read as one tier beating the national rate.

Robustness of the urban-rural gap

Test	Result
Remove worst rural states	3.8 to 2.7 points at ten removed; never reverses
Exclude facility types	3.6 to 4.0 points
Within bed-size bands	3.2 to 6.1 points
Within ownership classes	1.4 to 3.8 points; standardizing the mix takes 3.8 to 3.3
Unweighted instead of census-weighted	4.1 mean, 3.0 median, against 3.3 weighted
All of the above at once, with state fixed effects	−4.31 points (se 0.63); the gap grows under controls rather than shrinking
Change baseline	Under 0.2 points
Switch metric to occupancy	Gap widens; rural occupancy falls further in 37 of 48 states

This establishes a pattern, rules out four explanations for it and finds that a fifth, ownership, accounts for roughly an eighth. It does not identify a cause, and the variables most likely to carry one are not in CMS Provider Data. In the order we would work through them:

Home and community-based services	1915(c) waiver expansion and ARPA funding move long-stay residents into the community, lowering census without any change in demand for care. Testable against CMS 372 reports and state expenditure data.
Level-of-care criteria	States set the functional threshold for nursing-facility eligibility. Changes are documented and dated per state, so this is a natural difference-in-differences.
Managed Medicaid long-term care	These plans carry an explicit mandate to keep members in the community. Penetration varies by state and is already in our data, making this the cheapest to test.
Home health and hospice supply	Agency density per capita and its change since 2020, to see whether displaced demand went somewhere measurable. Cost reports are loaded.
Medicare Advantage penetration	These plans shorten skilled stays and divert rehabilitation to home health. Penetration runs about 57% in Wisconsin and Minnesota against 51% nationally.
Medicaid rate adequacy	Per-diem against facility cost, from Medicare cost reports. The direct test of whether bed closures track financial viability.
Ownership and payer mix	Ownership is the largest term in the specification above, at 9 points for non-profits. Whether that is ownership or the payer mix that travels with it is the obvious next question, and Medicaid share is the variable that would answer it.
Direct-care workforce	Wages and turnover against local labor markets. Payroll data is quarterly, so this can be tested per facility rather than per state.
Assisted living supply	The substitution everyone assumes and nobody measures, since there is no federal registry. Would require assembling state licensure files.

Two things would strengthen this. Replicating every state-level result on occupancy would resolve the handful of states where the two measures disagree. And modeling entry and exit, rather than conditioning on survival, would close the gap between what the specification measures, which is survivors, and what the national shortfall is, which is 58% capacity.