Written and maintained by CASRAI Editorial Board
Last updated
The CMS Overall Hospital Quality Star Rating is a single 1-to-5 star summary published on Care Compare, built by combining dozens of existing publicly reported quality measures into one number. It is not a payment programme — it doesn’t withhold or return a percentage of a hospital’s Medicare payments the way Hospital Value-Based Purchasing does, and it isn’t the same calculation as the Hospital Readmissions Reduction Program’s excess-readmission penalty. It is a public-reporting summary, and its mechanics — a latent variable model, fixed group weights, peer grouping by how many measure groups a hospital reports, and k-means clustering to assign the actual star — are what decide whether a hospital’s rating moves in a given release, often independent of whether care at that hospital changed at all.
This page works through that mechanism end to end: what goes into each of the five measure groups, how CMS combines correlated measures into a group score with a latent variable model rather than a simple average, how the group scores are weighted into one summary score, why hospitals are peer-grouped before they’re ever clustered into stars, and what k-means clustering actually does that a fixed percentile cutoff didn’t.
First, which programme is this? The Star Rating is reporting, not payment
Quality and risk teams routinely conflate three CMS mechanisms that all draw on overlapping Hospital Inpatient Quality Reporting (IQR) data but do different things with it. Keeping them separate matters because the fix for a bad number is different in each case.
| Overall Hospital Star Rating | Hospital VBP (Total Performance Score) | HRRP | |
|---|---|---|---|
| What it is | Public-reporting summary on Care Compare | Payment adjustment programme | Payment penalty programme |
| Effect on payment | None directly | Adjusts base DRG payments ± up to 2% | Cuts base DRG payments up to 3% |
| Core mechanism | Latent variable model + k-means clustering | Achievement/improvement points, linear exchange function | Excess readmission ratio per measure |
| Comparison group | Peer group by measure-group count (3, 4, or 5) | All eligible IPPS hospitals | Dual-eligible peer decile |
See Total Performance Score: How Hospital VBP Turns Four Domains Into a Payment Adjustment and Hospital Readmissions Reduction Program: Mechanics, Penalty Arithmetic, and the Critique for those two. A hospital can hold steady or improve on VBP and HRRP in the same cycle its Star Rating drops, because the Star Rating is answering a different statistical question — not “did this hospital earn back its withhold,” but “where does this hospital’s combined, standardized performance sit relative to its peer group this cycle.”
Date this page, because the weights and thresholds are not permanent
The Overall Hospital Quality Star Rating is codified at 42 CFR § 412.190, in the same subpart I of 42 CFR part 412 that houses Hospital VBP (412.160–412.168) and HRRP (412.150–412.154). The architecture described on this page — five measure groups, a latent variable model, peer grouping by measure-group count, k-means clustering — has been stable since CMS’s July 2021 methodology refresh and has not changed again since. The specific inputs to that architecture — which measures sit in which group, the exact group weights, and the minimum-count thresholds — are set through CMS’s public reporting rulemaking and technical documentation cycle, not through the regulation text itself, so they can and do move from one refresh to the next. Verify the current release’s technical notes on QualityNet before treating any specific weight or threshold below as this cycle’s number rather than the standing methodology.
The five measure groups, and where each one’s data actually comes from
Every measure that feeds the Star Rating is already publicly reported elsewhere on Care Compare — the Star Rating doesn’t collect anything new, it recombines what Hospital IQR, HCAHPS, and related CMS reporting programmes already require.
| Measure group | What it covers | Primary source of the underlying measures |
|---|---|---|
| Mortality | 30-day risk-standardized mortality for a defined set of conditions/procedures | Hospital IQR claims-based outcome measures |
| Safety of Care | Healthcare-associated infections and the AHRQ Patient Safety Indicator composite (PSI-90) | NHSN HAI measures + claims-based PSI-90, which is itself built partly from present-on-admission coding |
| Readmission | 30-day risk-standardized unplanned readmission for defined conditions/procedures | Hospital IQR claims-based outcome measures (a different measure set than HRRP’s own six, though built the same way) |
| Patient Experience | The HCAHPS composite and global measures | HCAHPS survey data — see HCAHPS Survey Administration |
| Timely and Effective Care | Process and effectiveness-of-care measures, largely eCQMs | Hospital IQR process/eCQM reporting — see eCQM Reporting for Hospitals |
For a patient-safety or infection-prevention team specifically, the practical implication is that NHSN surveillance data quality and PSI-90/POA coding accuracy feed the Safety of Care group score directly — a validation or coding gap upstream in infection surveillance doesn’t stay contained to the individual HAI measure’s own public display, it also moves the Safety of Care group score that the Star Rating consumes.
The reporting floor: a hospital needs at least three groups, and at least one outcome group
Not every hospital reports enough measures in every group to be scored on it — a small or specialty hospital may not have enough qualifying cases for a stable Mortality estimate, for instance. CMS’s minimum-reporting rule for a hospital to receive any Overall Star Rating at all is: at least three of the five measure groups populated, with a minimum measure count within each populated group, and at least one of those populated groups must be one of the three outcome groups (Mortality, Safety of Care, or Readmission) rather than only Patient Experience and Timely and Effective Care. A hospital under that floor gets no star rating published, not a rating built from an incomplete set.
Peer grouping by measure-group count — why a 3-group hospital is never compared to a 5-group one
Because the reporting floor above means real hospitals qualify for the rating with different numbers of measure groups, CMS doesn’t rank every rated hospital against the same national pool. It first sorts rated hospitals into peer groups based on how many of the five measure groups they qualify to be scored on — a peer group of hospitals reporting exactly three groups, one reporting exactly four, and one reporting all five. Every downstream step — standardizing scores and assigning the final star — happens within a hospital’s own peer group, not across all rated hospitals at once.
This is the answer to the single most common “why did our rating change” question that isn’t about the hospital’s own performance at all: a hospital’s summary score can hold flat while its star rating still moves, because the distribution of scores within its peer group shifted — other hospitals in the same measure-group-count peer group improved, declined, or moved into/out of that peer group entirely as their own reporting eligibility changed.
Inside a group: standardizing measures and the latent variable model
Within a measure group, the individual measures aren’t just averaged together as raw rates — they’re first standardized (converted to a common z-score-like scale so a mortality rate and an infection SIR are comparable in magnitude), and CMS then combines the standardized measures into one group score using a latent variable model rather than a flat average.
The distinction matters for a specific, well-documented reason: measures within a group are frequently correlated with each other (a hospital’s Mortality measures for related conditions tend to move together, for instance), and a simple average implicitly treats every measure as an independent, equally informative signal even when several of them are really telling you overlapping information. A latent variable model instead treats the group score as an unobserved (“latent”) underlying quality construct, and estimates that construct from the pattern of correlation across the observed, standardized measures — conceptually closer to a factor-analytic approach than to averaging. In practice this means a group’s score isn’t purely “however many points the measures in it add up to”; it reflects how those measures move together, which is also part of why two hospitals with visibly similar raw measure values can still land at different group scores.
This approach replaced an earlier, simpler averaging method specifically because CMS’s own technical documentation identified correlated and sometimes missing measures within a group as a source of distortion under flat averaging — the latent variable model was adopted as the more defensible way to combine a group’s measures into a single score. The exact statistical specification (which correlation structure, which estimation method) is documented in CMS’s own Overall Hospital Quality Star Rating technical methodology report on QualityNet rather than in the regulation text, and is worth pulling directly if a governance or quality committee needs to defend a specific group score rather than just describe the general approach.
From five group scores to one summary score: the weighting
Once each populated measure group has its own score, CMS combines them into a single hospital summary score using fixed group weights. Under the current architecture, when a hospital reports all five groups, Mortality, Safety of Care, Readmission, and Patient Experience are each weighted more heavily than Timely and Effective Care, which carries the smallest share of the five — the intent being that the three outcome-adjacent domains and the patient-experience domain carry most of the summary score, with the process/effectiveness domain contributing meaningfully but proportionally less. When a hospital doesn’t report all five groups, the weights are reallocated proportionally among whichever groups the hospital does qualify for, rather than simply dropped — a hospital’s summary score always adds the reported groups’ weights back up to a full 100%.
Do not treat any specific percentage split here as fixed indefinitely: like the domain weights on Hospital VBP, the exact group weighting is republished in CMS’s Star Rating technical notes on a recurring cycle and has been adjusted before. Confirm the live weighting for the current release before citing a specific number in a board or governance document.
From summary score to stars: why CMS uses k-means clustering instead of a fixed cutoff
The last step converts each hospital’s summary score into an actual 1-to-5 star rating. CMS does this separately within each peer group (see above) using k-means clustering: a statistical clustering algorithm that sorts the peer group’s hospitals into five clusters based on how their summary scores naturally group together, rather than by fixed percentile cutoffs (e.g., “top 10% always get 5 stars”).
The practical consequence: the score boundary between, say, 3 stars and 4 stars is not a fixed number that stays put release over release — it’s wherever the natural gap between clusters in that specific peer group’s score distribution happens to fall this cycle. A hospital’s own summary score can be essentially unchanged from the prior release and its star rating can still move, because the clustering boundaries shifted around it as other hospitals in its peer group entered, left, or changed their own scores. This is also why two hospitals with very similar summary scores can occasionally land on different sides of a cluster boundary and end up with different star ratings — clustering finds natural breaks in the data, it doesn’t guarantee even, predictable spacing between them.
So: why did our rating actually move? Three real drivers, not one
When a rating changes between releases, it is the product of up to three independent things, and distinguishing them is the actual diagnostic work for a quality director or risk manager asked to explain the number:
- The hospital’s own measure performance changed — a real shift in the underlying HAI rate, mortality rate, readmission rate, HCAHPS composite, or eCQM performance that feeds a group score.
- The hospital’s peer group changed — the hospital started or stopped qualifying for a measure group (crossing the reporting floor in either direction), moving it into a different measure-group-count peer group with a different comparison pool entirely.
- The clustering pattern around the hospital changed — the hospital’s own summary score held flat, but enough other hospitals in its peer group moved that the k-means boundaries shifted, changing which cluster (star) the hospital’s unchanged score now falls into.
A rating narrative built only on driver 1 — “our infection rates went up so our stars went down” — is sometimes the whole story and sometimes only part of it. Before presenting a rating change to a board or medical staff, it’s worth checking the hospital’s own group scores against the prior release first (driver 1), then checking whether the hospital’s populated-group count changed (driver 2), before assuming the movement reflects driver 1 alone.
Where this fits for a patient-safety programme specifically
The Star Rating itself doesn’t distinguish infection-prevention-specific inputs from the rest of the Safety of Care group — NHSN HAI measures and PSI-90 sit inside one combined group score. That means an infection preventionist’s surveillance and validation work has a downstream effect on a public-facing number well beyond the individual HAI measure’s own listing, and a QAPI or performance-improvement plan built around a Safety-of-Care-driven rating change should trace back to the specific HAI or PSI component actually responsible, not just the group score in aggregate. See QAPI Plan, QAPI Report, and PIP Write-Up for documenting that kind of targeted improvement plan, and Present on Admission (POA) Indicator for how coding accuracy specifically affects the PSI-90 component of that group.
Frequently asked questions
Is the Overall Hospital Star Rating the same thing as the Hospital VBP Total Performance Score?
No. They draw on overlapping IQR data but are structurally different: the Star Rating is a public-reporting summary using a latent variable model, peer grouping, and k-means clustering, with no direct payment effect. Hospital VBP is a payment-adjustment programme using achievement/improvement points and a linear exchange function. A hospital’s Star Rating and its VBP Total Performance Score can move in opposite directions in the same release.
Can a hospital’s star rating change even if none of its measures changed?
Yes. Because clustering happens within a peer group and the cluster boundaries are set from that peer group’s score distribution each cycle, a hospital with an unchanged summary score can still move stars if enough other hospitals in its peer group moved, or if the hospital’s own populated-measure-group count changed its peer group assignment.
What’s the minimum a hospital needs to report to get a star rating at all?
At least three of the five measure groups populated with the required minimum measure count, and at least one of those populated groups must be an outcome group — Mortality, Safety of Care, or Readmission. A hospital below that floor is not assigned a star rating.
How often does CMS refresh the Overall Hospital Star Rating?
CMS refreshes the rating on a recurring public-reporting schedule tied to Care Compare updates, using rolling periods of underlying measure data rather than a single fixed reporting year — check the current release’s technical notes on QualityNet for the specific reporting period and refresh date in effect, since that cadence is set administratively rather than in the regulation.
Why does CMS use a latent variable model instead of just averaging the measures in a group?
Because measures within a group are often correlated, a flat average can let a small number of related measures dominate a group score while treating each measure as if it carried independent information. The latent variable model instead estimates the group score as an underlying construct inferred from the pattern of correlation across the group’s standardized measures, which CMS adopted specifically to address that distortion.
