Plot.plot({
...chartBase(narrow ? 350 : 280),
marginLeft: narrow ? 8 : 195, marginRight: 62,
x: {domain: [-0.26, 0.015], label: "Rating difference (stars)", grid: true, ticks: [-0.2, -0.1, 0]},
y: {domain: ratingGap.map(d => d.label), label: null, axis: narrow ? null : "left", paddingInner: 0.55},
marks: [
Plot.barX(ratingGap, {x: "value", y: "label", fill: d => d.label === "Overall rating gap" ? C.neutral : C.accent}),
Plot.text(ratingGap, {x: 0, y: "label", text: d => d.value.toFixed(3), textAnchor: "start", dx: 7, fill: C.ink}),
...(narrow ? [labelAbove(ratingGap, "label", "label", 350, 0.55)] : []),
Plot.ruleX([0], {stroke: C.ink})
]
})Whose Voices Shape Hospital Ratings
What contributor patterns reveal about experiences in U.S. hospitals and how to interpret the ratings.
A lower average rating gives a hospital a question to answer: which experiences need attention? In this dataset of U.S. hospitals and other healthcare facilities, most of the gap between contributor groups reflects which places receive their feedback. Differences in the topics they discuss also persist within similar rating bands. Understanding the setting and the content helps a team choose a relevant comparison and a specific problem to investigate.
Those are two different uses of feedback. For reporting, a hospital needs to understand the population behind a score. For improvement work, it needs to understand what happened in the experiences people describe. A stable average can support the broad picture while individual accounts still identify something worth checking.
The rating findings use 2,306,550 Google text reviews of U.S. healthcare listings, spanning 2007 to early June 2026. A single-listing contributor has eligible text about one healthcare listing in the collected dataset; a repeat contributor has text about 2–10. The tiny 11+ group is tracked separately. These are observed profiles, not verified patient identities or actual care journeys. The topic findings use a separate 2019+ labeled sample introduced below.
Understand the setting behind the rating gap
Repeat contributors give 3.50 stars on average, compared with 3.74 among single-listing contributors. About 73% of that gap sits in listing mix: approximately 0.173 of the 0.237-star difference. The within-listing component is 0.061 stars, with a small remainder.
Single-listing contributors also rate the repeat group’s mix of places lower than their own overall average. The pooled difference therefore contains information about where feedback comes from, as well as how people assess their experience.
Most of the group rating gap reflects which listings receive feedback Repeat profiles (2–10 listings) minus single-listing profiles. The listing-mix, within-listing, and remaining components add to the overall gap; negative values indicate a lower repeat-profile mean.
Source: saved decomposition over 8,707 listings with both groups, weighted by repeat-profile review volume. The residual completes the reported overall gap.
Setting makes the heterogeneity concrete. In the source record, the pooled group gap varies across Google listing categories:
| Listing category | Single-listing mean | Repeat mean |
|---|---|---|
| University hospital | 2.92 | 2.86 |
| Children’s hospital | 3.52 | 3.48 |
| Emergency room | 3.70 | 3.41 |
| Mental health clinic | 2.82 | 2.51 |
Source: frozen eligible text reviews; selected categories with at least 20,000 contributions. These are pooled, unadjusted category means. Google categories can misdescribe services.
There is little group separation in the first two settings and a larger difference in the latter two. A single label such as “repeat contributors are harsher” would miss that variation. The table does not establish its cause, or rank clinical quality.
For a hospital comparing its feedback with peers, the useful first step is to define comparable services and periods. A rating gap concentrated in one setting calls for closer reading there. Its explanation might involve the services represented, the experiences described, or when contributions were made. The decomposition helps locate the difference; the hospital still needs contextual evidence to understand it.
Similar ratings can describe different experiences
The separate topic sample contains 176,853 labeled reviews at 7,776 listings represented by both groups: 152,703 single-listing and 24,150 repeat contributions. Repeat contributors mention accessibility, equitable treatment and safety more often. Amenities also appear more often, while waiting appears less often.
Contributor groups describe a different mix of experiences Percentage of each group’s labeled reviews mentioning a topic, at listings with both groups. Open dots: single-listing profiles. Filled dots: repeat profiles (2–10). Topics are selected examples, not a complete ranking.
Source: saved topic-reach comparison. Labels are model-generated; restricting the comparison to shared listings does not equalize the groups’ mix of reviews within those listings.
One possible explanation is that repeat contributors give more low ratings, and those accounts discuss different subjects. Comparing within rating bands tests part of that explanation:
| Topic | Repeat-to-single reach ratio, 1–2 stars | Repeat-to-single reach ratio, 4–5 stars |
|---|---|---|
| Accessibility/accommodations | 1.25× | 1.72× |
| Equitable treatment | 1.28× | 1.69× |
| Patient safety/errors | 1.17× | 1.29× |
Source: frozen valid labels at shared listings. Each ratio compares the share of each group’s reviews mentioning a topic within the stated rating band. Topic mentions can have any sentiment.
The direction persists in both bands. For equitable treatment, the 4–5-star shares are 0.36% versus 0.62%. The ratio is sizeable, but the absolute difference is about a quarter of a percentage point. What that difference means for a hospital depends on the situations those accounts describe.
A separate check gives both groups the same listing weights. The equitable-treatment gap narrows but remains, at 2.27% versus 3.03%. Listing mix and the overall rating distribution are therefore incomplete explanations in these separate comparisons. Neither check isolates a causal reason for the difference.
For a patient-experience team, the practical lesson is to read subject matter alongside stars. Reviewing only low-rated feedback narrows the material available for learning. Favorable assessments can also describe an accommodation, an interaction or a safety-related experience worth examining. The subject tells a team where to look; the wording and evidence determine whether it describes praise, a problem or a serious concern.
Exact topic shares and the complete 29-topic comparison
A review can mention several topics, so the shares do not add to 100%. Ratios compare repeat contributors with single-listing contributors; they are descriptive estimates without displayed confidence intervals.
The earlier exploratory report retains the full topic chart and within-topic sentiment comparison. The methods below distinguish the pooled rates in this table from the separate rating-band and weighting checks.
Turn the topic signal into a question about the experience
The aggregate data locate subjects for closer reading. To learn what a hospital could change, a team needs to distinguish the concrete situations grouped under each label. The full topic definitions provide a starting point:
| Signal | What to clarify in the accounts | Follow-up |
|---|---|---|
| Accessibility and accommodations | Physical access, mobility support, interpretation, or sensory communication accommodations? | Validate the described barrier and involve the relevant facilities, access or communication team. |
| Equitable treatment | What interaction was perceived as biased or unfair, and in which service context? | Examine the account and relevant contextual evidence through existing review processes. |
| Patient safety and errors | What error, omission or harm concern was reported? | Assess serious concerns through the existing quality/safety route, independently of the star rating. |
| Care coordination and referrals | Where did a referral, transfer, handoff or follow-up lose information or responsibility? | Review the described sequence with the relevant service teams. |
These are investigation questions informed by the definitions, rather than causes established by the group averages. For example, a checked account describing an interpretation problem would lead to a different follow-up from one describing a physical access barrier, even though both can receive the same accessibility label.
What might produce the group differences? Different services, exposure to several settings, different periods or expectations, and the amount people write are plausible explanations. The topic sample’s repeat contributors write somewhat longer texts; a separate length-mix check leaves the selected differences in the same direction. More than one explanation may contribute.
A profile’s activity over a long period cannot establish that its author is a caregiver, a frequent patient or someone following a referral pathway. The useful learning comes from validating what was described and connecting it to the hospital’s own evidence. Group patterns help select questions; they do not supply the answers on their own.
Recognize stable priorities and sensitive local scores
The contributor differences leave the broad picture largely intact. Restricting the full text-review population to single-listing contributors changes the mean from 3.711 to 3.737 stars, or +0.025 before rounding. In the separate topic sample, the same top five topics remain in the same order: bedside manner, waiting, front desk, communication and nursing responsiveness. No topic moves more than one rank.
Those results support a stable broad agenda in this sample. Removing repeat contributors would reduce the available examples and modestly change topic prevalence, but does not substantially reorder the issues. Their share of the record also helps explain the small change in the pooled rating.
Use the findings for comparison and for improvement
For interpreting a score, start with a comparable context. Define the service, period and included feedback, then inspect local sensitivity and retained counts. If a conclusion changes with contributor composition, examine the accounts and relevant contextual evidence before making a performance claim. A filtered score is a diagnostic comparison, not automatically a better representation of experience.
For deciding what to investigate, start with the content. Keep the broad priorities in view, read relevant examples across rating bands and groups, and validate the particular problem before assigning follow-up. An accommodation barrier can deserve a response even when addressing it would barely change the star mean. The goal is a clearer understanding of the experience and an appropriate next step; this study has not evaluated the effect of an intervention.
For the next reporting discussion, bring one comparison whose population is explicit and one checked topic example with a defined follow-up question. The topic-pair analysis then extends the reading to issues people describe together.
Sources, comparison checks and original visualizations
Scope and groups. The collected record contains 3,138,214 reviews of 11,694 U.S. healthcare listings, including hospitals, emergency departments, clinics, medical centers and urgent care. The rating population uses 2,306,550 reviews with at least 15 characters of text and 2,173,013 contributor profiles. Groups count distinct Google listings per profile across the full observed period, 2007 to early June 2026. They do not measure repeat status at each past date, actual visits or physical institutions. Separate listings at one site can increase the count. Public feedback is self-selected. These calculated text-review means differ in scope from Google’s displayed rating, which can also include ratings without text.
The 11+ group. Its 46 profiles supply 964 text reviews and average 4.00 stars. The pooled all-profile mean includes this group; the single-only comparison removes both repeat groups. The listing-level histogram, direction counts and volume breakdown compare single-listing profiles with the combined single-plus-2–10 population and exclude 11+ from that baseline.
Decomposition and setting. The 8,707 listings with both main groups contribute to the within-listing and listing-mix components, weighted by repeat-profile review volume. Within-listing −0.061153, listing-mix −0.172939 and remainder −0.003177 sum to the −0.237269 pooled gap; listing mix accounts for 72.9%. These are arithmetic components, not causal mechanisms. The new setting table selects four Google categories with at least 20,000 eligible text reviews. Its group means are pooled within category, rather than adjusted within the same listings or matched services. Category and rating differences do not establish clinical quality or actual referral journeys.
Topic sample and validity. The separately drawn 2019+ sample contains 217,343 reviews; 196,183 distinct listing-plus-review keys pass the labeling analysis’s validity gates. The contributor-topic comparison narrows to 176,853 keys at 7,776 shared listings and excludes the 11+ group. Topic reach divides reviews mentioning a topic by labeled reviews in that contributor group. A review can mention multiple topics. Validity gates check permitted names, IDs, phrases, review resolution and known sentiment. They do not establish labeling accuracy. Labels describe reported subjects, not verified events. The investigation questions above draw on full taxonomy definitions, not an inspection establishing which particular barrier explains the group gap.
Separate descriptive checks. The original topic chart pools rates across shared listings. The rating-band table compares 1–2 and 4–5 stars, with respective denominators of 55,673/9,838 and 92,968/13,449 single/repeat reviews. Mentions in those bands may have any sentiment. A separate standardization gives both groups identical listing weights, based on combined labeled-review volumes at each of the 7,776 listings. Equitable-treatment reach becomes 2.27%/3.03%, safety 3.95%/4.94%, and accessibility 0.93%/1.15% single/repeat. Equal-listing weights preserve the same directions. Another separate check standardizes the groups’ mix over four character-length bands: 15–99, 100–299, 300–999 and 1,000+. The selected directions persist. Median text lengths are 229/265 characters; mean distinct topics are 1.94/2.01. These checks assess particular compositional explanations separately. They do not jointly adjust listing, rating, date, service, text length and contributor characteristics, identify an intrinsic group trait, or establish statistical significance or causation.
Stable topic picture. Ranking all 29 topics by pooled reach before and after retaining single-listing profiles gives the same top five in the same order; the largest movement is one rank and Spearman correlation is 0.9985. For example, equitable-treatment reach changes from 2.368% to 2.223%, accessibility from 0.970% to 0.908%, and safety from 4.076% to 3.914%. Repeat contributors supply 13.7% of this sample, including 19.2% of accessibility, 18.9% of equitable-treatment and 17.1% of safety topic accounts. Those coverage shares do not establish that whole issues disappear, that information is unique to a group, or that every local service has the same priorities.
Local sensitivity and precision. The 6,224 eligible listings have at least 30 single-profile text contributions. The local shift equals the repeat share of single-plus-repeat contributions multiplied by the single-minus-repeat mean difference. This identity holds to numerical precision. Listing rank correlation is 0.998; the typical absolute shift is 0.030 stars. The preserved histogram uses shifts rounded to four decimals: 677 cross the 0.1-star boundary there, versus 674 using exact source means. Both are approximately 11%. The volume table and 45.6% share with lower single-only means use unrounded calculations. Neither the eligibility rule nor the marker validates a performance claim or determines whether feedback is important.
Preserved figures and evidence. All four original charts and the full topic table remain on this page. The earlier report retains detailed contributor distributions, geography, time and setting comparisons, 11+ profiles, and complete topic and sentiment plots. It is an exploratory archive; the interpretation here supersedes stronger claims there. Its time charts classify contributors using their final observed history, including later contributions. Download the original chart aggregates or expanded narrative evidence. The latter recomputes the source-based comparisons and retains their query definitions. All original group counts/means, decomposition components and 29 topic rates reconcile with the retained snapshots. No model is rerun, and no source text or contributor identifiers are exported.