Which Experiences Are Linked to Lower Hospital Ratings?
What the ratings of U.S. hospitals reveal about access, financial experiences and the concerns that need closer attention.
In Google reviews of U.S. hospitals and other healthcare facilities, concerns about billing, reception, scheduling and waiting remain closely linked to lower ratings after other reported experiences are considered. That gives hospital teams a reason to examine financial and access processes alongside clinical interactions. Deciding what deserves attention also requires knowing how often a concern occurs and what the full account describes.
The findings reveal different kinds of signal. Billing has the largest adjusted negative rating association, while waiting appears negatively in about three times as many accounts. Safety concerns usually appear in low-rated feedback alongside other complaints. Positive descriptions of bedside manner provide a substantial body of experiences to learn from and preserve.
These results use 196,183 labeled Google reviews from 2019 to early June 2026, covering 9,904 U.S. healthcare listings. They describe self-selected public feedback, with topics labeled by a language model. A listing may represent a hospital, clinic, emergency department or another service. Counts refer to accounts, rather than distinct patients or verified incidents.
Financial and access processes deserve direct attention
Negative billing mentions have an adjusted association of −1.05 stars, followed by front-desk conduct (−0.97), scheduling (−0.92) and waiting (−0.90). These are the four largest negative-mention estimates in the primary model.
The model compares feedback within the same Google listing and accounts for the other positive and negative topic mentions, text length, year and contributor group. It uses 195,616 reviews after removing listings represented only once. An adjusted association describes a difference in ratings conditional on those recorded features; it does not compare identical care experiences.
The useful finding is that financial and access concerns retain information about dissatisfaction even when the recorded clinical and interpersonal topics enter the comparison. A hospital team therefore has a reason to examine these parts of the experience directly. They can remain important questions even when clinical interactions receive praise elsewhere in the feedback.
The categories point to different situations. Billing covers charges, disputes, payment and collections. Front-desk conduct concerns administrative interactions. Scheduling describes appointment access, while waiting concerns the timeliness of the care process. Grouping all four under a broad experience score would make it harder to determine which process and supporting records need examination.
For a service lead, the next step is to identify which of those situations appears in validated local accounts. An unexpected charge, a difficulty obtaining an appointment and a dismissive administrative interaction require different investigation questions. The association identifies a topic worth examining; the account establishes what the person actually describes and what evidence could clarify it.
Explore the adjusted associations, including their uncertainty Dots are model estimates; lines are 95% confidence intervals clustered by listing. Start with the six largest associations in either direction, or show every topic. Hover for full topic names and mention counts.
Source: primary model of all 58 topic–sentiment indicators, with listing fixed effects and controls for review length, year, and reviewer group. Display requires at least 100 mentions in the labeled frame. Intervals describe model uncertainty; they do not include topic-label errors or selection bias.
Treat these results as a prominent group of signals. Weighting the sample toward the eligible source corpus retains the same top-four group but swaps the order of waiting and scheduling. Closely spaced estimates should not become a rigid league table. The intervals describe model uncertainty and do not include label errors or selection into writing feedback.
A strong rating signal and a frequent problem need different decisions
Waiting is mentioned negatively in 26,006 reviews, compared with 8,227 for billing. That is about 3.2 times as many accounts, although billing’s adjusted association is larger in magnitude. Front-desk conduct also supplies a substantial body of negative feedback.
| Concern mentioned negatively | Reviews | Share of labeled reviews | Adjusted association |
|---|---|---|---|
| Billing | 8,227 | 4.2% | −1.05 stars |
| Front-desk conduct | 16,965 | 8.6% | −0.97 stars |
| Scheduling | 7,279 | 3.7% | −0.92 stars |
| Waiting | 26,006 | 13.3% | −0.90 stars |
Counts and shares use all 196,183 labeled reviews; associations use the 195,616-review primary model. A review can contribute to several rows. These sampled counts describe the collection and cannot forecast a particular hospital’s workload.
The distinction matters when deciding where to spend investigative effort. A strong rating association identifies a concern that distinguishes dissatisfaction after recorded context is considered. A frequent topic provides a larger recurring set of experiences to understand. Local frequency, the situation described and the relevant process evidence are needed to turn either signal into a practical priority.
The raw rating gaps alone offer little separation: 23 of the 29 topics fall between −1.9 and −2.3 stars. A single low-rated account can contribute to several of those gaps. The chart shows what changes when the model considers the topics together.
Adjustment separates topics that look similar in the raw comparison Open dots: raw mean-rating gaps. Blue dots: adjusted associations for negative mentions. The lines connect two different comparisons; they do not divide a causal effect into components.
Source: saved notebook comparisons and primary model. The raw reference group has neither a positive nor a negative label for the topic; it can include neutral-only mentions. Selected examples illustrate the spread; switch to all topics for the complete comparison.
Billing and cleanliness have raw gaps of −2.25 and −2.04 stars; their adjusted associations are −1.05 and −0.43. Similar raw gaps can therefore accompany different relationships after other recorded features enter the comparison. The connecting lines show two statistical comparisons, rather than a division of responsibility for the low rating.
A useful local question is whether a recurring concern points to a defined process problem that can be investigated. For waiting, that could mean identifying the stage people describe and comparing it with available waiting-stage records. For billing, it could mean distinguishing charge accuracy, disputed coverage and an unclear explanation. Those are proposed applications to test against local evidence. Multiplying review counts by coefficients would not estimate the number of affected patients, recoverable stars or the return from a change.
A smaller coefficient can sit inside a very negative account
Negative safety mentions have an adjusted association of −0.64 stars. Yet 6,787 of the 7,377 accounts containing a negative safety mention, or 92%, have a one- or two-star rating. And 6,282, or 85%, contain at least one other topic mentioned negatively. Safety concerns are often embedded in accounts describing several sources of dissatisfaction.
Equitable-treatment concerns show a related pattern: 94% of the 4,266 accounts mentioning the topic negatively have a one- or two-star rating, while its adjusted association is −0.74 stars. These labels identify reported concerns and perceptions; they do not establish verified safety events or discrimination.
This context explains why a coefficient cannot stand in for the significance of an account. The overall rating summarizes the experience, while the model estimates the recorded topics jointly. Several concerns can share the same low rating, and a reviewer cannot choose a value below one star. A smaller adjusted coefficient can coexist with a strongly negative description.
The data do not isolate how much of a particular coefficient’s size is due to the rating floor, shared topics or other differences between accounts. The practical lesson is to preserve the concern and its surrounding context when feedback is summarized. Ranking only by adjusted stars could leave a safety, fairness or privacy issue less visible even though its substance merits specific assessment.
For those accounts, the next decision comes from human reading, label validation and appropriate supporting evidence. The question is what was reported, what needs verification and which existing service or quality process should assess it. The rating association remains useful context, while the significance of the particular concern requires its own judgment.
Positive feedback shows what is worth understanding and preserving
Bedside manner is mentioned positively in 66,094 accounts, or 33.7% of the labeled frame, with an adjusted association of +1.02 stars. Switch the coefficient chart above to “Positive mentions” to explore this complementary picture.
The large body of favorable descriptions gives hospital teams material for learning what people value in clinical interactions. The category includes empathy, respect, listening and interpersonal conduct. The model does not distinguish which specific behavior accounts for the association, so a useful next step is to read validated positive examples in a comparable service and identify the conduct they describe.
The largest positive point estimate belongs to behavioral health (+1.07 stars), but it comes from 979 positive-mention accounts, compared with 66,094 for bedside manner. Their confidence intervals overlap. Both may supply valuable examples, with different scopes; the slightly larger coefficient does not establish a universally stronger learning opportunity.
Positive and negative signals also need separate interpretation. Billing’s negative association is −1.05 stars and its positive association +0.50; scheduling’s are −0.92 and +0.46. The model estimates praise and complaints as separate features. Removing a complaint does not automatically produce the experience described in a positive account, and silence about a topic does not establish satisfaction with it.
For improvement work, this creates a second question alongside investigating dissatisfaction: what do people value that a proposed change should preserve? A team could select a small set of validated positive accounts from the relevant service, identify recurring behaviors and choose a specific practice to examine or test. The association helps select material for that learning; it does not supply a tested protocol or predict the rating gain from adopting it.
Use the signals to choose the next question
The findings support three questions for a hospital deciding where to look more closely:
- Which experiences distinguish dissatisfaction? Use the adjusted associations to identify topics that remain informative after the recorded context is considered.
- Which concerns recur in our own service? Use local feedback counts, periods and service context to locate a recurring set of experiences.
- Which accounts need specific assessment because of what they describe? Preserve the full concern and use relevant verification, irrespective of its position in a rating ranking.
For example, a hospital noticing recurring billing concerns could first establish what is disputed: the amount, the coverage decision or the explanation. That determines which records and participants are relevant. Only after clarifying the problem can a team choose a response and a measure of whether it helped. The national association supplies a reason to look; it does not select the local solution.
If a process changes, evaluate it using appropriate independent process and experience evidence, alongside what subsequent feedback describes. A change in review volume or star ratings alone can also reflect who contributes and which services are represented. Positive accounts can help specify what the revised process should preserve.
The topic-pair article adds another useful question: which linked concerns need to be read together? Their shared context matters for follow-up. The number of complaints in an account is not a linear measure of rating damage; the supporting interaction check below explains why.
Do connected complaints bring an additional rating penalty?
Five selected pairs, including billing with insurance and waiting with staffing, were examined in two listing halves. Each model includes the usual topic indicators and controls, plus total topic count and a term for each pair of negative mentions. That interaction is the additional association beyond the two separate terms.
None of the five selected pairs shows an extra negative interaction Blue: estimates in the half designated as holdout. Open gray dots: the other half. Lines show 95% confidence intervals clustered by listing. Values above zero indicate less negative associations than an additive model would predict.
Source: saved interaction models, including total topic count. Pair selection came from the earlier co-occurrence analysis of the same broader dataset; the split should not be treated as fully independent validation of pair discovery.
All five holdout estimates are positive, ranging from +0.19 to +0.74 stars, with individual 95% intervals above zero. The combined association is less negative than simple addition predicts. This describes a bounded rating model; it does not mean two concerns improve the experience or are less serious. It cannot establish the benefit of preventing a second problem.
Pair selection used the earlier co-occurrence analysis of the broader dataset, so the split is not fully independent validation of discovery. Both halves are fitted separately. The displayed intervals are individual intervals, without a simultaneous correction across the five pairs.
How stable are the rating signals?
The checks support the broad pattern across alternative weighting and rating specifications. Weighting retains billing, front-desk conduct, scheduling and waiting as the top-four negative-mention set, while waiting and scheduling exchange order. These results favor interpreting a group of prominent signals over assigning importance to small ranking differences.
| Check | Saved result | What it tells us |
|---|---|---|
| Low versus high ratings | Coefficient correlation −0.996 with the stars model | Topics associated with lower stars also tend to be associated with a higher probability of a 1–2-star rating, among 1–2- and 4–5-star reviews. The signs run in opposite directions. |
| Sampling weights | Coefficient correlation 0.9887; largest change 0.10 stars | Weighting toward the eligible corpus leaves the negative-topic estimates broadly similar. It does not remove self-selection into reviewing. |
| Ordered rating categories | Rank correlation 0.95 | In a 40,000-review subsample, a topic-only ordinal model gives a similar negative-topic ordering to a topic-only linear model. This is a simpler check, without the main model’s listing effects or other controls. |
The following original plot compares all 29 negative-topic estimates. The vertical axis is an association in percentage points from a low-versus-high rating model, rather than an overall probability of a low rating. Like the linear stars model, a linear probability model can produce predictions outside the outcome’s bounds.
Source: saved low-versus-high robustness model. It drops three-star reviews and uses the same topic indicators, listing effects and controls as the primary model. Weighting does not remove self-selection into reviewing, and the simpler ordinal check does not validate all features of the primary specification. None replaces human validation of the labels.
Methods, contributor comparison and aggregate data
Sample and units. The 217,343-review coverage sample spans 2019 to early June 2026. Of these, 196,183 reviews at 9,904 Google listings have at least one validity-gated label. The primary model retains 195,616 reviews at 9,337 listings after singleton removal. A review is keyed by listing plus review ID. Counts and shares in the text and coefficient tooltips refer to the full labeled frame; interaction counts refer to the corresponding split before singleton removal. Listings are not necessarily distinct physical hospitals.
Model and comparison. Ordinary least squares models the 1–5-star rating using 58 binary topic–sentiment indicators, log(1 + review word count), year and contributor-group indicators, and listing fixed effects. A topic can have both a positive and a negative flag; neutral statements set neither. For a raw negative-mention gap, the reference group has neither directional flag for that topic and can include neutral-only mentions. The raw and adjusted comparisons have different conditioning and slightly different populations, so their difference is not a causal decomposition. Listing effects account for stable differences between listings, rather than changing service or reviewer mix within them. Intervals are 95% and clustered by listing. R² is 0.782, an in-sample fit measure rather than predictive performance on new feedback.
New context checks. Negative-mention reach divides distinct negative-topic accounts by all 196,183 labeled reviews. Low-rating concentration divides one- and two-star negative-topic accounts by all accounts mentioning that topic negatively. Additional-negative-topic share records whether another topic also has a negative flag in the same account. These are descriptive context measures, not patient prevalence or estimates of verified incidents.
Contributor comparison. In this labeled frame, the correct multi-listing-minus-single-listing contrast is −0.0068 stars with listing effects alone (95% interval −0.0315 to +0.0178), and +0.0277 in the full primary model (+0.0147 to +0.0408). The adjusted difference is small and positive, with its interval excluding zero under the model. The contrast subtracts the two group coefficients using their full clustered covariance; their common omitted reference is the 11+ group. It does not establish identical experiences or explain away the different population comparison in Whose Voices Shape Hospital Ratings.
Interpretation. Topic labels describe the text accompanying the observed rating. Their associations are therefore diagnostic of the same reported experience. Automated validity checks establish source and format matching, not human agreement. Public feedback is self-selected; model intervals do not cover label errors, missing experiences or selection bias. The labels do not verify clinical events. Coefficients cannot serve as severity weights or expected intervention returns.
Data and reproducibility. Download the original aggregate chart estimates or the additional counts, context measures and verified contrasts. The original charts retain the executed notebook’s saved values, rounded to four decimals. scripts/generate_hospital_rating_insights.py reconstructs the frozen review-level inputs, recomputes the new descriptive measures and refits the primary, weighted, low/high and pair models for reconciliation. Their estimates match the saved chart precision. The ordinal model was inspected in the executed notebook and not refitted in this check. Exports contain no raw review text, contributor identifiers or listing identifiers.