NURSE STAFFING HOURS AND OVERALL FIVE-STAR RATINGS IN US NURSING HOMES — AN EMPIRICAL CROSS-SECTIONAL ANALYSIS

Abstract

Star ratings published by the federal Centers for Medicare and Medicaid Services condense the quality of every certified United States nursing home into a single score of one to five stars, and families rely on that score when choosing a facility. Staffing is the input that facilities and regulators can act on most directly, which makes the strength of the association between published staffing levels and the overall rating a question of practical consequence. The study quantifies how strongly reported nurse staffing hours are associated with the overall five-star rating across the certified nursing home population. The analysis uses the June 2026 federal provider file covering 14,695 facilities, of which 14,018 with complete information form the analysis sample. The overall rating is regressed on reported registered nurse hours and total nurse staffing hours per resident per day, controlling for facility size and ownership, with heteroskedasticity-robust standard errors and an ordered-logit specification as robustness, at the 5 percent significance level. Both staffing measures are strongly associated with the rating: half an additional registered nurse hour per resident per day corresponds, at constant total hours, to about one fifth of a star, and an hour of registered nurse time carries roughly 2.3 times the association of an hour of other nursing time. Non-profit and government facilities rate higher than for-profit facilities even at equal staffing and size. Staffing explains a meaningful but modest share of rating variation, and part of the gradient reflects the deliberate construction of the rating itself.

Keywords: nursing homes; nurse staffing; five-star ratings; quality of care; long-term care; ownership

1. Introduction

1.1. Problem and Aim

Choosing a nursing home is among the most consequential consumer decisions a family makes, and it is typically made under time pressure, with limited information, and on behalf of a person who is acutely unwell. To support that decision, the federal Centers for Medicare & Medicaid Services (CMS) publishes a five-star quality rating for every nursing home certified to participate in Medicare or Medicaid, displayed prominently on the public Care Compare website. The overall star rating condenses inspection findings, staffing data, and clinical quality measures into a single number between one and five, and it has become the dominant shorthand for nursing home quality in the United States, used not only by families but also by hospital discharge planners, insurers, lenders, and researchers (Centers for Medicare & Medicaid Services, 2026a). The stated goal of the system is to give consumers an easy way to understand nursing home quality and to make meaningful distinctions between high and low performing homes (Centers for Medicare & Medicaid Services, 2026a). A facility's stars consequently carry economic weight: they shape referral flows, occupancy, and reputation, and they define the public benchmark against which any staffing regulation will be judged.

Among the ingredients of the rating, staffing occupies a special position. Inspection outcomes and clinical indicators are consequences of many processes, but nursing hours are a direct managerial choice and a direct object of regulation: a facility can hire, and a legislator can mandate hiring. Nurse staffing levels have accordingly been at the center of renewed regulatory attention to minimum staffing standards in recent years. For both audiences, the practical question is the same: across the actual population of certified facilities, how strongly do published staffing hours align with the overall quality score, how large is the premium on skilled registered nurse time relative to nursing time in general, and how much of the variation in ratings does staffing leave unexplained?

The study answers these questions on the complete current census of rated United States nursing homes, using the June 2026 federal provider file, with facility size and ownership held constant, and with explicit attention to a complication that is easy to overlook: the overall star is partly constructed from staffing information, so the association measured here mixes a quality signal with the arithmetic of the rating itself.

1.2. Research Question and Hypotheses

The question is a practical one: are reported registered nurse hours and total nurse staffing hours per resident per day associated with the overall five-star rating of United States nursing homes, conditional on facility size and ownership type? H0 answers no: conditional on the number of certified beds and ownership type, neither staffing measure is associated with the overall rating, and the regression coefficients on both staffing variables are zero. Against this, H1 requires only that at least one staffing coefficient be positive. A secondary expectation, motivated by the skill-mix literature, is that an hour of registered nurse time is associated with a larger rating difference than an hour of other nursing time. Decisions on significance are taken at the 5 percent level.

1.3. Contribution

For policy and practice, the article delivers a current, population-wide quantification of the staffing-rating gradient, built on three design features. First, it uses the complete census of rated facilities in a single monthly vintage of the federal provider file, 14,018 complete cases from 14,695 listed facilities, so the estimates describe the industry rather than a sample of it. Second, it takes the construction of the dependent variable seriously: the staffing star rating is excluded as a regressor by design, because the overall star is assembled in part from that star, and the analysis instead uses the underlying reported hours, with the mechanical channel identified and discussed rather than ignored. Third, it exploits an exact reparametrization of the main regression to separate the association of registered nurse hours from that of other nursing hours, yielding a directly interpretable skill premium: the same fitted model, expressed in two ways, prices an hour of registered nurse time against an hour of licensed practical nurse and aide time. An ordered-logit specification confirms that the conclusions do not depend on treating the five-point rating as a cardinal scale.

2. Background

2.1. The Five-Star Rating System and Its Construction

CMS introduced star ratings for nursing homes on its public reporting website in December 2008. The current system assigns each facility a rating in three domains and an overall rating (Centers for Medicare & Medicaid Services, 2026a). The health inspection rating is based on deficiencies from the two most recent standard onsite inspections plus three years of complaint and infection control surveys, weighted by scope and severity. The staffing rating is based on six measures: case-mix-adjusted total nurse staffing hours per resident per day, case-mix-adjusted registered nurse hours, case-mix-adjusted weekend total staffing, and three measures of staff turnover covering nursing staff, registered nurses, and administrators. All six are derived from the Payroll-Based Journal system, through which facilities submit daily staffing hours each quarter, combined with a daily resident census derived from federally mandated resident assessments; since July 2024 the staffing levels have been case-mix adjusted using the nursing case-mix groups of the Patient-Driven Payment Model, so that facilities serving more care-intensive residents are not penalized for the same nominal hours (Centers for Medicare & Medicaid Services, 2026a). Points are assigned to each measure by position in the national distribution, with maxima of 100 points for each of the two weekday staffing levels, 50 points for weekend staffing and for each of the two nursing turnover measures, and 30 points for administrator turnover; the summed score of up to 380 points is then mapped to the one-to-five-star staffing rating. Facilities that fail to submit staffing data by the deadline, or that report four or more days in a quarter with residents present but no registered nurse hours, receive a one-star staffing rating outright. The quality measure rating summarizes 15 clinical indicators, nine for long-stay and six for short-stay residents.

Two construction features matter for any statistical analysis of the overall rating. The first is norm-referencing. The health inspection rating is assigned relative to other facilities in the same state, with the best 10 percent receiving five stars, the middle 70 percent split evenly across two, three, and four stars, and the bottom 20 percent receiving one star; the staffing and quality domains are scored against national point thresholds (Centers for Medicare & Medicaid Services, 2026a). As a consequence, the national distribution of ratings is close to uniform by design, and a star expresses relative position, not an absolute standard. The second feature is the composite arithmetic. The overall rating starts from the health inspection rating; one star is then added if the staffing rating is five stars and subtracted if it is one star; one star is added if the quality measure rating is five stars and subtracted if it is one star; and the result is bounded between one and five, with an additional cap for facilities whose inspection rating is one star (Centers for Medicare & Medicaid Services, 2026a). Staffing therefore enters the overall score mechanically at the extremes of the staffing distribution. Any regression of the overall rating on staffing variables estimates the sum of a substantive channel, in which staffing improves the inspection and clinical outcomes that the other domains measure, and a construction channel, in which high or low staffing moves the overall star directly. Section 3 explains how the design of this study responds to that fact; in the analysis file, the overall rating and the staffing star rating correlate at 0.48.

2.2. Nurse Staffing and Quality of Care

That nurse staffing relates to nursing home quality is among the oldest and most examined propositions in long-term care research, and the accumulated evidence supports a positive but complicated relationship. Harrington et al. (2000) related staffing hours to deficiency citations across all certified United States facilities and found that fewer registered nurse hours were associated with more quality-of-care deficiencies, while also noting that facility characteristics and state effects predicted deficiencies more strongly than staffing did and that the explained share of variance was small. Konetzka, Stearns & Park (2008) addressed the endogeneity that plagues cross-sectional comparisons, using longitudinal variation and instrumental methods, and found that increases in registered nurse staffing and a richer skill mix improved resident outcomes. Castle & Anderson (2011) followed 2,839 facilities with quarterly staffing data from 2003 to 2007 in a dynamic panel framework covering staffing levels, turnover, agency use, and professional mix, and concluded that a change to more favorable staffing is generally associated with a change to better quality. The regulator's own evidence base points the same way: the technical documentation of the rating system cites a clear association between nurse staffing ratios and quality of care found in the federal staffing study, together with a growing body of evidence linking higher staff turnover to poorer quality, which is why turnover measures were added to the staffing domain (Centers for Medicare & Medicaid Services, 2026a).

The systematic review of Spilsbury et al. (2011) tempers simple readings of this literature. Across the studies reviewed, the authors identified 42 distinct measures of quality and 52 distinct ways of measuring staffing, and they concluded that a focus on the number of nurses alone fails to capture the influence of turnover, agency use, training, experience, and the organization of care. Two implications follow for the present design. The expected sign of the staffing coefficients is positive, but the expected explanatory power is moderate: staffing is one input among many. And the distinction between registered nurse hours and total nursing hours is substantive, because the literature locates much of the quality association in the skilled end of the staffing mix; the reparametrization introduced in Section 3.2 is built to measure that distinction directly.

A further lesson of this literature is econometric. Staffing is chosen by facilities in response to their circumstances, so cross-sectional staffing coefficients blend the effect of staffing on quality with the effects of everything that drives both, from payer mix to local labor markets; the longitudinal and instrumental designs of Castle & Anderson (2011) and Konetzka, Stearns & Park (2008) exist to break that blend. A single-vintage census such as the present one cannot do so, and it does not try. Its comparative advantage is coverage and currency: it measures the gradient across the entire rated industry at one recent date, as a description of where the industry stands rather than an estimate of what a staffing mandate would cause.

2.3. Ownership and Quality

Roughly three quarters of United States nursing homes operate for profit, and the association between ownership and quality has a long research history. Mor et al. (2004) described the industry as a two-tiered system in which the lower tier, the roughly 15 percent of non-hospital-based facilities serving predominantly Medicaid residents, has fewer nurses, lower occupancy, and more deficiencies, is concentrated in the poorest counties, and disproportionately serves African-American residents. Facilities in that tier are also more likely to be terminated from the Medicare and Medicaid programs, and Mor et al. (2004) warned that public quality reporting, by steering demand toward better-rated homes, could drive lower-tier facilities out of business and concentrate the consequences on residents of poor communities, a warning issued four years before the star system launched and directly relevant to how its ratings are read. Because Medicaid payment rates are the lowest in the payer mix, payer composition constrains the resources available for staffing, and ownership, payer mix, and staffing are entangled in the cross-section. Grabowski & Stevenson (2008) add an important caution about causal readings of ownership differences: examining nearly 3,000 ownership conversions between 1993 and 2004, they found that converting facilities differed from their peers before conversion, with facilities converting from non-profit to for-profit status being consistently low performers and those converting in the opposite direction consistently high performers, while the conversion itself changed performance little. Ownership differences in quality thus partly reflect selection of facilities into ownership forms rather than effects of ownership as such.

For the present study, ownership serves two purposes. As a control, it absorbs a coarse but powerful correlate of resources and mission when estimating the staffing gradient. As a descriptive finding in its own right, the conditional ownership differences, the rating gap that remains after staffing and size are held constant, indicate how much of the raw ownership gap the observable inputs explain.

3. Data and Method

3.1. Data

The data source is the Provider Information file of the CMS Provider Data Catalog, the public dataset underlying the Care Compare website, with one row per certified nursing home and 99 variables covering ratings, staffing, capacity, and ownership (Centers for Medicare & Medicaid Services, 2026b). The vintage analyzed carries the processing date 1 June 2026 and was retrieved on 1 July 2026, the reference date of this study. The file lists 14,695 facilities. Five variables enter the analysis under their exact published names: Overall Rating, the one-to-five-star composite described in Section 2.1; Reported RN Staffing Hours per Resident per Day, in which RN abbreviates registered nurse; Reported Total Nurse Staffing Hours per Resident per Day, which comprises registered nurse, licensed practical nurse, and nurse aide hours; Number of Certified Beds; and Ownership Type. The staffing variables are the reported payroll-based hours divided by the assessment-derived resident census, before the case-mix adjustment that CMS applies when scoring the staffing domain; the reported version is used here because it is expressed in natural units of care time per resident per day and does not embed the rating system's own adjustment model.

Ownership Type distinguishes 13 legal subtypes, which are grouped into three categories: for-profit (10,855 facilities in the raw file, dominated by limited liability companies and corporations), non-profit (2,897, mostly corporations, with church-related and other organizations), and government (943, spanning county, city, state, hospital district, and federal facilities). The overall rating is missing for 125 facilities; under the rating rules these are mainly facilities too new to have two standard inspections and facilities in the Special Focus program, for which CMS suppresses ratings (Centers for Medicare & Medicaid Services, 2026a). Reported staffing hours are missing for 578 facilities, reflecting the payroll-data submission and validity exclusions of the staffing domain, and the bed count is missing for eight facilities. The analysis sample comprises the 14,018 facilities, 95.4 percent of the file, with complete information on all five variables: 10,375 for-profit, 2,764 non-profit, and 879 government facilities. In the published file the five rating categories from one to five stars contain 2,873, 3,025, 2,844, 2,783, and 3,045 facilities respectively, a near-uniform spread that follows from the norm-referenced construction described in Section 2.1.

3.2. Method

The study is a cross-sectional analysis of the certified nursing home population at a single reference date, reported in line with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) recommendations for cross-sectional studies (von Elm et al., 2007); the preceding subsection states the data source, the eligibility of facilities, the handling of missing data, and the resulting sample flow from 14,695 listed to 14,018 analyzed facilities. Because the data are a complete administrative census rather than a survey sample, no sampling weights or design corrections arise; the uncertainty expressed by the standard errors refers to the processes generating ratings and staffing rather than to sampling from a larger population. The main specification is the linear regression

Rating(i) = a + b1 RN(i) + b2 TOTAL(i) + b3 BEDS(i)/100 + b4 NONPROFIT(i) + b5 GOVERNMENT(i) + u(i) (1)

where Rating(i) is the overall star rating of facility i, RN(i) and TOTAL(i) are the reported registered nurse and total nurse staffing hours per resident per day, BEDS(i)/100 is the number of certified beds in hundreds, and the two ownership indicators take for-profit facilities as the reference category. Estimation is by ordinary least squares (OLS) with heteroskedasticity-robust standard errors. Because TOTAL(i) includes RN(i) as a component, the coefficient b1 measures the additional association of a registered nurse hour beyond that of a generic nursing hour, and the same model can be exactly reparametrized by replacing TOTAL(i) with the non-registered-nurse remainder, TOTAL(i) minus RN(i); in that form the coefficient on RN(i) equals b1 plus b2 and measures the full association of a registered nurse hour, while the coefficient on the remainder equals b2 and measures the association of an hour of licensed practical nurse or aide time. Both forms are reported, since together they express the skill premium directly. Collinearity is assessed with variance inflation factors.

Treating a five-point ordinal rating as a cardinal dependent variable is a simplification, so the analysis is repeated as an ordered logit,

LATENT(i) = c1 RN(i) + c2 TOTAL(i) + c3 BEDS(i)/100 + c4 NONPROFIT(i) + c5 GOVERNMENT(i) + e(i) (2)

in which the observed one-to-five-star rating arises from the latent quality score LATENT(i) crossing four estimated thresholds and the error term follows a logistic distribution; slope coefficients are reported alongside their implied odds ratios, which give the multiplicative change in the odds of being at or above any given rating threshold per unit of the regressor. Two design decisions respond to the construction problem of Section 2.1. The staffing star rating is not used as a regressor, because the overall star is assembled partly from it, which would build a mechanical association into the estimate by definition. And the staffing variables are the reported hours rather than the case-mix-adjusted hours that enter the rating algorithm, keeping the regressors in observable natural units. Neither decision removes the construction channel entirely, since reported and adjusted hours are strongly related; the interpretation in Section 5 keeps both channels in view. A battery of further checks, reported with the results, probes the estimates along the remaining lines the file permits: standard errors clustered by state and state fixed effects, the health inspection and quality measure ratings as alternative outcomes, exclusion of extreme staffing ratios, occupancy and case-mix-adjusted controls, the published turnover measures, and a threshold-by-threshold check of the proportional-odds assumption. The 5 percent significance level governs all tests.

4. Results

4.1. Descriptive Statistics

Table 1 presents the analysis sample overall and by ownership (see Table 1). The mean overall rating is 3.04 stars with a standard deviation of 1.42, and each of the five rating categories contains between 2,668 and 3,007 facilities in the analysis sample, the near-uniform spread that the rating design engineers. The counts run 2,668, 2,896, 2,738, 2,709, and 3,007 facilities from one to five stars, so the exclusions for missing data thin the lower rating categories somewhat more than the upper ones, a first indication that facilities with reporting problems are not a random draw. The average facility reports 0.68 registered nurse hours and 3.88 total nurse staffing hours per resident per day; averaged across facilities, registered nurses account for 17.0 percent of a facility's reported nursing time, which corresponds to 17.6 percent of aggregate reported hours; the median registered nurse figure of 0.58 lies below the mean, and the maxima of 8.19 and 16.22 hours indicate a long right tail produced by small facilities, where a low resident census inflates per-resident ratios. The average facility has 107.6 certified beds.

Table 1: Descriptive statistics, mean (standard deviation), of the analysis sample of United States nursing homes with complete data, by ownership category; RN denotes registered nurse; data are the June 2026 Provider Information file. Source: Own illustration based on Centers for Medicare & Medicaid Services (2026b).

The ownership pattern in the raw means already reproduces the ordering the literature leads one to expect. Non-profit facilities average 3.63 stars, government facilities 3.34, and for-profit facilities 2.85, a raw gap of 0.78 stars between non-profit and for-profit homes. The staffing gradient runs in parallel: non-profit facilities report 0.94 registered nurse hours against 0.60 in the for-profit sector, and 4.41 against 3.70 total hours, while government facilities lie between the two on every staffing measure. Non-profit facilities are also markedly smaller, with 91.4 beds on average against 111.3 in the for-profit sector, while government facilities show the widest dispersion in both size and staffing. The staffing mix shifts with ownership as well: at the group means, registered nurse time accounts for 16.2 percent of reported nursing hours in the for-profit sector and 21.3 percent in the non-profit sector, so the sectors differ not only in the volume of care hours but in their composition. Whether the ownership rating gap survives conditioning on these staffing and size differences is the question the regression settles.

4.2. Regression Results

Table 2 reports the estimates (see Table 2). In the main specification, both staffing measures carry positive and highly significant coefficients: 0.380 for registered nurse hours, with a robust standard error of 0.037 and t = 10.26, and 0.284 for total nurse hours, with a standard error of 0.019 and t = 14.72, both with p < 0.001 on 14,018 observations. In practical units, an additional 0.5 reported registered nurse hours per resident per day, roughly the distance from the for-profit mean to well above the non-profit mean, is associated with 0.19 additional stars over and above its contribution through total hours, and an additional 0.5 total nurse hours with 0.14 stars. In standard-deviation terms, one standard deviation of registered nurse hours corresponds to 0.18 stars and one standard deviation of total hours to 0.26 stars.

Table 2: Regressions of the overall five-star rating on reported staffing hours, facility size, and ownership for 14,018 United States nursing homes; coefficients with heteroskedasticity-robust standard errors in parentheses in columns (1) and (2), maximum-likelihood estimates with standard errors in column (3); threshold parameters of the ordered logit omitted; RN denotes registered nurse; all staffing and size coefficients and ownership contrasts have p < 0.001; t statistics referenced in the text are computed from unrounded coefficients and standard errors. Source: Own illustration based on the author's data (Section 3).

Facility size is negatively associated with the rating, at −0.252 stars per 100 certified beds with t = −11.02, and the conditional ownership differences remain substantial: at equal staffing and size, non-profit facilities rate 0.403 stars and government facilities 0.265 stars above for-profit facilities, both with p < 0.001. The coefficient of determination is 0.145.

The reparametrized column translates these estimates into the skill premium. An hour of licensed practical nurse or aide time is associated with 0.284 stars per resident-day hour, while an hour of registered nurse time is associated with 0.664 stars, the sum of the generic-hour association and the registered nurse increment, with a standard error of 0.031 and t = 21.12. A registered nurse hour thus carries roughly 2.3 times the rating association of a non-registered-nurse hour. Read cumulatively, adding half a registered nurse hour per resident per day, which raises total hours by the same amount, is associated with 0.33 additional stars, the product of the full registered-nurse coefficient of 0.664 and the increment of 0.5 hours. The two columns are the same fitted model, so the fit, the controls, and the residuals are identical; only the accounting of the staffing mix differs. H0 is rejected on every count: staffing levels are strongly associated with the overall rating, and the skill mix matters beyond the volume of hours. At the same time, the modest coefficient of determination is itself a finding. Staffing, size, and ownership together explain 14.5 percent of the variation in overall ratings; the remaining variation reflects the inspection and clinical-outcome domains, which respond to far more than staffing inputs, much as the deficiency literature anticipated (Harrington et al., 2000).

4.3. Robustness

The ordered logit in column (3) of Table 2 reproduces every qualitative conclusion without assuming cardinal ratings. The coefficient on registered nurse hours is 0.921 with z = 16.86; because total hours are held constant, this is the substitution effect of replacing an hour of other nursing time with an hour of registered nurse time, and its odds ratio of 2.51 says that such a substitution multiplies the odds of being at or above any given rating threshold by roughly two and a half. The full association of an additional registered nurse hour, which raises total hours as well, is the sum of the two staffing coefficients, 0.921 plus 0.441 = 1.362 with a standard error of 0.049, an odds ratio of 3.90, and a skill ratio relative to an hour of other nursing time of roughly 3.1 in the logit metric. The total-hours coefficient of 0.441 implies an odds ratio of 1.55 per hour, the beds coefficient of −0.338 per 100 beds an odds ratio of 0.71, and the ownership contrasts odds ratios of 1.62 for non-profit and 1.37 for government facilities, all with p < 0.001; McFadden's pseudo-R2 is 0.056. Collinearity between the two staffing measures is unproblematic: they correlate at 0.67, all variance inflation factors remain below two in both parametrizations, and the reparametrized form, whose regressors correlate at only 0.19, returns identical fitted values by construction. The estimated gradients are therefore not artifacts of the linearity assumption, the cardinal treatment of stars, or overlapping regressors. Two further observations reinforce the reading. The ownership and size coefficients are nearly identical across all three columns, indicating that the staffing gradient is not absorbing the ownership contrast, and the registered-nurse premium appears in both estimators, 0.664 against 0.284 in the linear model and, in the corresponding full-hour form, 1.362 against 0.441 in the ordered logit, so the skill-mix result is a feature of the data rather than of the estimation method.

Five further checks probe the estimates along the lines the data permit. First, because inspection ratings are assigned within states and staffing varies systematically across states, equation (1) was re-estimated with standard errors clustered at the state level, 53 clusters, and, separately, with state fixed effects: every conclusion is unchanged, with all coefficients significant at the 1 percent level under clustering and staffing coefficients of 0.419 for registered nurse and 0.248 for total hours once state effects are absorbed (R2 = 0.173). Second, replacing the composite outcome with its component domains shows that the gradient is not merely the arithmetic of the overall star: on the health inspection rating, which staffing does not enter mechanically, the registered nurse and total-hours coefficients are 0.190 (p = 0.008) and 0.175 (p < 0.001) with state-clustered errors, while on the quality measure rating the registered nurse coefficient is 0.355 (p = 0.013) and total hours are null. Third, excluding the 104 facilities above the 99.5th percentile of either staffing measure, the extreme per-resident ratios that small facilities produce, strengthens rather than weakens the gradient, with the registered nurse coefficient rising to 0.639 and total hours to 0.345. Fourth, adding occupancy as a control or replacing the reported hours with the case-mix-adjusted staffing measures leaves the ordering intact, and adding the two published turnover rates, available for 12,353 facilities, attenuates the registered nurse coefficient to 0.170 while both turnover measures enter negatively and highly significantly, consistent with turnover being one of the channels the staffing domain is built to capture. Fifth, a threshold-by-threshold check of the ordered logit, estimating a separate binary logit at each rating cutoff, shows the registered nurse coefficient declining from 1.375 at the lowest threshold to 0.823 at the highest, so the proportional-odds assumption holds only approximately while the association remains positive and large at every threshold.

5. Discussion

5.1. Reading the Staffing Gradient

The central estimates line up with five decades of staffing research: more nursing time, and especially more registered nurse time, accompanies better measured quality, with a skill premium of roughly 2.3 to one over other nursing hours. The finding echoes the longitudinal evidence that registered nurse staffing and skill mix drive outcomes (Konetzka, Stearns & Park, 2008) and the panel evidence that favorable staffing changes precede quality improvements (Castle & Anderson, 2011). Two channels, however, jointly produce the gradient measured here, and a single cross-section cannot fully apportion it between them. The substantive channel runs through care: staffing shapes the inspection findings and clinical outcomes that the other rating domains record. The construction channel runs through arithmetic: reported hours co-move with the case-mix-adjusted staffing measures that determine the staffing star, and the staffing star adds or subtracts a star from the overall rating at the extremes of the staffing distribution (Centers for Medicare & Medicaid Services, 2026a). The design choices of this study, excluding the staffing star as a regressor and using reported rather than adjusted hours, keep the mechanical channel from entering by definition, but they cannot remove it, because the channel exists in the rating itself. The component-domain checks of Section 4.3 nonetheless bound the two: the registered nurse gradient on the health inspection rating, which staffing does not enter mechanically, is roughly half the size of the composite gradient (0.190 against 0.380 in the incremental form), so a substantive channel demonstrably operates alongside the mechanical one, even though the two remain only partially separable. The correct reading of the coefficient is therefore conditional: among facilities of equal size and ownership, half a registered nurse hour per resident-day separates facilities by about one fifth of a star on the published scale, through care and construction together, because the rating was built so that staffing would matter.

5.2. Ownership and Scale

The conditional ownership gaps are, in some respects, the more striking result. Staffing and size differences together account for about half of the raw non-profit advantage, which falls from 0.78 stars in the means of Table 1 to 0.403 stars once staffing and size are held equal, a reduction of 48 percent, but a substantial premium remains, and the government premium of 0.265 stars behaves similarly. The government category is itself heterogeneous, pooling county, city, state, hospital-district, and federal facilities whose funding and missions differ, so its premium averages over dissimilar institutions. The pattern is consistent with the two-tier account of Mor et al. (2004), in which payer mix concentrates resource constraints in a lower tier dominated by for-profit, Medicaid-dependent facilities, so that ownership proxies for unmeasured resources, case mix, and mission beyond the staffing that this file records. The selection evidence of Grabowski & Stevenson (2008) warns against reading the premium causally: facilities sort into ownership forms, and converting facilities carried their performance with them. The negative size gradient, a quarter star per 100 beds, likewise admits several readings, from care organization in large facilities to their concentration in market segments with more constrained payers; the present design records the association without adjudicating its source. One further reading of the size gradient deserves note: the health inspection domain records deficiencies found on site, and larger facilities offer inspectors more residents, more staff, and more processes in which deficiencies can be found, so part of the association may be an exposure effect of scale in inspection-based measurement rather than a quality difference per bed.

5.3. Implications

The results carry three implications. For families and their advisers, the published staffing columns of the provider file are informative in the most practical sense: they align with the summary score, and the registered nurse column aligns most strongly, so a consumer comparing two facilities can read staffing tables as a meaningful signal rather than administrative decoration. For policy, the estimates describe the terrain on which minimum staffing standards operate: the for-profit sector, holding three quarters of facilities, reports on average 0.60 registered nurse hours per resident per day against 0.94 in the non-profit sector, so binding staffing floors would engage precisely the segment where both staffing and ratings are lowest, while the modest explained variance cautions that staffing mandates alone would leave most of the rating distribution in place. For research, the construction channel documented here argues for outcome-based dependent variables, deficiencies, hospitalizations, or clinical indicators, rather than the composite star, whenever the causal effect of staffing is the estimand; the composite is designed to respond to staffing, which is a virtue for consumers and a confound for causal inference (Spilsbury et al., 2011, draw the parallel methodological lesson that how quality is measured shapes what staffing studies can find). The June 2026 numbers give that agenda a concrete baseline: a rated industry averaging 0.68 registered nurse hours per resident per day, an association of about one fifth of a star per half registered-nurse hour at constant total hours, and three quarters of facilities in the ownership segment where both figures are lowest.

6. Limitations

Seven limitations qualify what these results can support. First, the design is cross-sectional: staffing, ownership, and ratings are observed at one reference date, and no causal claims are made; reverse pathways, in which highly rated facilities attract staff and residents, and omitted variables such as payer mix, resident acuity, and regional labor markets, plausibly contribute to every estimated gradient. In particular, facilities anticipating inspections may staff up in advance of surveys, and highly rated facilities command payer mixes that fund richer staffing, so the arrow between staffing and ratings runs in both directions even before any omitted variable is considered. Second, the construction channel discussed in Sections 2.1 and 5.1 means that part of the staffing coefficient is built into the dependent variable by design; the analysis quantifies the combined association but cannot decompose it. Third, the staffing regressors are reported rather than case-mix-adjusted hours, a choice made for interpretability, but consequently they do not account for differences in resident acuity across facilities, and ratio measures based on small resident censuses can take extreme values, as the maxima in Table 1 show. Fourth, the rating scale is ordinal and partly norm-referenced within states, so identical stars in different states need not denote identical absolute quality, and the linear model treats rating steps as equal; the ordered logit addresses the second concern but not the first. Fifth, the analysis sample excludes 677 facilities without ratings or staffing data, disproportionately new facilities, Special Focus facilities, and facilities failing payroll-data validity checks, groups that are unlikely to be missing at random, so the estimates describe the rated population rather than every certified home. The direction of that selection can be quantified: the excluded facilities are smaller on average, 90.6 against 107.6 certified beds, and the 552 of them that do carry an overall rating average 2.30 stars against 3.04 in the analysis sample. Sixth, the file is a single monthly vintage: ratings, staffing data, and even the rating methodology are revised over time (Centers for Medicare & Medicaid Services, 2026a), and the estimates carry the date of the data. Seventh, all variables are facility-level aggregates: the analysis says nothing about which residents within a facility receive the reported hours, and associations at the facility level need not hold at the resident level.

7. Conclusion

Across 14,018 rated United States nursing homes in June 2026, reported nurse staffing hours are strongly associated with the overall five-star rating: 0.5 additional registered nurse hours per resident per day correspond to about one fifth of a star at constant total nursing time, facility size, and ownership, an hour of registered nurse time carries roughly 2.3 times the association of an hour of other nursing time, and the results hold in an ordered-logit specification that respects the ordinal rating scale. Non-profit and government facilities rate 0.403 and 0.265 stars above for-profit facilities even at equal staffing and size, while staffing, size, and ownership together explain 14.5 percent of rating variation, a share that is meaningful for a single input family yet leaves the bulk of variation with the inspection and clinical domains. The association is real, current, and by design partly mechanical, since the rating system was constructed so that staffing moves the star. For consumers the staffing tables are a usable quality signal; for causal questions about what staffing does to residents, outcome measures rather than the composite star remain the appropriate ground, and panel designs on successive vintages of these files are the natural next step.

References

Castle, N.G. & Anderson, R.A. (2011): 'Caregiver staffing in nursing homes and their influence on quality of care: using dynamic panel estimation methods', Medical Care, 49(6), pp. 545–552. https://doi.org/10.1097/MLR.0b013e31820fbca9

Centers for Medicare & Medicaid Services (2026a): Design for Care Compare Nursing Home Five-Star Quality Rating System: Technical Users' Guide, April 2026. Available at: https://www.cms.gov/medicare/provider-enrollment-and-certification/certificationandcomplianc/downloads/usersguide.pdf (Accessed: 2 July 2026).

Centers for Medicare & Medicaid Services (2026b): Provider Information [data set], Provider Data Catalog. Available at: https://data.cms.gov/provider-data/dataset/4pq5-n9py (Accessed: 1 July 2026).

Grabowski, D.C. & Stevenson, D.G. (2008): 'Ownership conversions and nursing home performance', Health Services Research, 43(4), pp. 1184–1203. https://doi.org/10.1111/j.1475-6773.2008.00841.x

Harrington, C., Zimmerman, D., Karon, S.L., Robinson, J. & Beutel, P. (2000): 'Nursing home staffing and its relationship to deficiencies', The Journals of Gerontology: Series B, 55(5), pp. S278–S287. https://doi.org/10.1093/geronb/55.5.S278

Konetzka, R.T., Stearns, S.C. & Park, J. (2008): 'The staffing-outcomes relationship in nursing homes', Health Services Research, 43(3), pp. 1025–1042. https://doi.org/10.1111/j.1475-6773.2007.00803.x

Mor, V., Zinn, J., Angelelli, J., Teno, J.M. & Miller, S.C. (2004): 'Driven to tiers: socioeconomic and racial disparities in the quality of nursing home care', The Milbank Quarterly, 82(2), pp. 227–256. https://doi.org/10.1111/j.0887-378X.2004.00309.x

Spilsbury, K., Hewitt, C., Stirk, L. & Bowman, C. (2011): 'The relationship between nurse staffing and quality of care in nursing homes: a systematic review', International Journal of Nursing Studies, 48(6), pp. 732–750. https://doi.org/10.1016/j.ijnurstu.2011.02.014

von Elm, E., Altman, D.G., Egger, M., Pocock, S.J., Gøtzsche, P.C. & Vandenbroucke, J.P. (2007): 'The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies', PLoS Medicine, 4(10), e296. https://doi.org/10.1371/journal.pmed.0040296