1. Who reviews the hospitals

The corpus is the complete scraped review record of 11,694 U.S. healthcare listings - hospitals, emergency rooms, medical centers, urgent care - with 3.14 million reviews from 2.94 million contributor profiles. The population of interest is the 2.17 million profiles behind the 2.31 million text reviews.

Reading the distribution. 94.7% of text-review profiles wrote about exactly one hospital. The 2–10-hospital group is 5.2% of profiles but - because each profile contributes several reviews - supplies 10.7% of all text reviews. Every ninth text review a patient reads was written by someone who also reviewed another hospital. The 11+ group is 46 profiles in three million: visible only under a square-root scale, and behaviorally so unusual (section 7) that we keep it out of the main comparison.

2. The gap

Multi-hospital reviewers rate lower, and the effect is monotone in how many hospitals a profile has reviewed - up to the point where the extreme group breaks the pattern entirely.

Reading the gap. Single-hospital reviewers average 3.74 stars; profiles with 2 hospitals average 3.53; with 4–5 hospitals, about 3.26 - every additional hospital in a profile's record comes with a lower average, and the 1–2★ share climbs from 30% to 41%. Then the 11+ group reverses everything: 4.00 stars, only 12% negative. Whatever that group is doing, it is not the same activity.

Pooled, the 2–10 group rates 0.24 stars below single-hospital reviewers (3.50 vs 3.74). Weighting each profile equally instead of each review moves the multi mean by only +0.014 - the gap is not an artifact of a few prolific writers. The obvious reading - "people who see more hospitals judge more harshly" - turns out to be mostly wrong, and the decomposition shows why.

3. Where they review, not how they judge

A reviewer group can rate lower for two very different reasons: because it scores the same hospital lower (a judging effect), or because it reviews different hospitals (a composition effect). The two have opposite implications - a judging effect biases comparisons between hospitals; a composition effect mostly does not.

Splitting the pooled gap at the hospital level separates the two.

Reading the decomposition. Of the −0.24-star pooled gap, −0.17 comes from composition - multi-hospital reviewers disproportionately review facilities whose single-hospital reviewers already rate them low - and only −0.06 from judging the same facility more harshly. The scatter shows the composition channel directly: a facility's multi-reviewer share correlates at r = −0.31 with its mean rating. High-multiplicity facilities are the big urban emergency rooms and referral centers; the places people pass through on complicated care journeys are the places everyone rates poorly.

Within a given facility the two groups largely agree (median within-facility gap −0.04 stars, and the gap is positive at 45% of facilities). This reframes the bias question: the multi-reviewer group doesn't so much distort hospital ratings as over-sample the painful end of the system - which is exactly what you'd expect from caregivers and referral patients, and why section 4 finds so little movement when they're removed.

4. Ratings without them

The direct test of the evaluation-bias concern: recompute every rating with multi-hospital reviewers excluded, and see what changes - nationally, per state, and per facility.

Stability in numbers: the median facility moves by 0.030 stars; 10.8% of facilities move ≥0.1 stars and only 0.8% move ≥0.25. Facility rankings barely notice: Spearman rank correlation between the public and single-reviewer-only rating is 0.998, and only 4.8% of top-decile facilities change under the alternative rating.

Reading the counterfactual. Nationally, the entire multi-reviewer phenomenon is worth +0.025 stars. At the facility level the story is more interesting than a simple null: the shift distribution is centered on zero - not on the +0.02 the pooled gap would suggest - because at nearly half of facilities the multi-hospital reviewers are actually the more generous group. The negativity concentrates between facilities, not within them.

There is a thin tail worth respecting: about one facility in nine moves ≥0.1 stars, and these are disproportionately smaller facilities where a handful of repeat reviewers carries real weight. For any per-hospital league table built from this corpus, a minimum-review threshold matters far more than reviewer-composition adjustment.

5. The geography of multiplicity

Reviewer composition is not uniform across the country - it is a regional variable, and any cross-state comparison of hospital ratings silently compares different reviewer mixes.

Reading the geography. The multi-reviewer share ranges from 5.0% (Hawaii) to 15.2% (Washington), with the Midwest (11.9%) and West (11.7%) systematically above the South (9.9%) and Northeast (8.9%). The pattern is consistent with market structure: multi-hospital reviewing requires multiple hospitals within one life - sprawling metros, referral networks, consolidated systems.

Two things stop this from becoming a bias story. First, even in the highest-multiplicity states the counterfactual shift tops out at +0.07 stars (Nebraska, Washington). Second, the within-state gap itself varies enormously - from −0.56 in Hawaii to effectively zero in Kansas and positive in Arkansas - so "multi-hospital reviewers are harsher" is not even directionally true everywhere. What is regionally consistent is their presence, which matters for any future reviewer-mix-standardized comparison (section 8).

6. A stable share, a widening gap, and the referral-center signature

Two robustness views: does the phenomenon change over time, and does it concentrate in particular care settings?

Reading time and setting. The multi-reviewer share has been stable between 10% and 12% for a decade - this is not a growing distortion. Both groups share the strong ratings inflation visible across the whole corpus (single-reviewer means rose from ~3.0 in 2016 to ~4.2 in 2026), but the gap between them has widened from near zero in 2016–17 to ~0.25 stars in recent years.

The category pattern carries the care-journey signature: university hospitals (13.5%) and children's hospitals (11.9%) attract the most multi-hospital reviewers - referral destinations, where arriving means you've been somewhere else first - yet show the smallest rating gaps (−0.06 and −0.04). The biggest gaps are at mental-health clinics (−0.31) and emergency rooms (−0.29), settings where a cross-facility perspective plausibly reflects genuinely worse experiences.

7. The 46 extreme profiles

Profiles that reviewed 11 or more hospitals break every pattern above, and earn their separate treatment.

8. What repeat reviewers push onto the record

Ratings say how unhappy people were; topics say what they were unhappy about. The topic-labeling run over the 10%, year- and facility-stratified coverage sample (196,567 labeled reviews, joined back to reviewer groups with zero key loss) lets us ask whether multi-hospital reviewers direct public attention differently.

Because multi-hospital reviewers frequent different facilities (section 3), all comparisons below are restricted to the 7,776 facilities with labeled reviews from both groups - 152,703 single-reviewer and 24,150 multi-reviewer labeled reviews. Reach is the share of a group's labeled reviews mentioning a topic.

Robustness - reach ratios within rating strata (the attention gap is not a byproduct of multi reviewers writing more negative reviews):

Reading the topic gap. Two findings, and they rhyme with the rating decomposition of section 3.

First, attention. Multi-hospital reviewers over-index most on the systemic, accountability topics: accessibility (1.51×), discrimination and equitable treatment (1.48×), privacy (1.33×), patient safety and medical errors (1.30×), infection control (1.28×), medical records (1.24×) - and, true to the care-journey hypothesis, care coordination and referrals (1.19×). Single-hospital reviewers own the everyday encounter: wait times (0.92×), the front desk (0.94×), maternity care (0.82×). The stratified check confirms this is not merely because multi reviewers write angrier reviews - inside 4–5★ reviews the over-indexing on accountability topics is stronger (discrimination 1.67×, infection control 1.66×): even when satisfied, repeat reviewers evaluate hospitals as institutions, not encounters.

Second, tone. Within a topic, the groups sound remarkably alike: every point hugs the diagonal, with the multi group running just above it (overall 50.4% vs 45.3% negative phrases; typical within-topic gap +2 to +6pp, largest for maternity, scheduling, bedside manner - and essentially zero on the topics that are near-universally negative for everyone, like billing and safety events).

Together: repeat reviewers change what the public record talks about far more than how it sounds. Their marginal contribution is coverage of the system-level failure modes - exactly the domains HCAHPS measures thinly or not at all, which makes this group analytically valuable rather than a nuisance population.

(Estimates from the 10% coverage sample; ~25k labeled multi-reviewer reviews. Topic-level reach ratios have 95% CIs of roughly ±0.05–0.10 at the thresholds used. The remaining corpus can be labeled later to tighten rare-topic estimates.)

9. What it means

For the evaluation-bias question this analysis set out to test: the public hospital rating is not materially bent by repeat reviewers. Their −0.24-star pooled gap is three-quarters composition - they frequent the hospitals everyone rates poorly - and the counterfactual record without them is nearly indistinguishable (rank correlation 0.998). Reviewer composition can be treated as a documented robustness result rather than a required correction, with one exception: thin facilities, where minimum-review thresholds do the real work.

For the care-journey hypothesis: the composition finding is itself evidence. Multi-hospital reviewers over-sample emergency rooms, referral centers, and university hospitals - the trajectory of a complicated illness. Their reviews are not noise to be filtered; they are the corpus's best window into cross-facility experience, which no per-hospital survey instrument observes.

For the upcoming HCAHPS comparison: state-level reviewer-mix differences (5% to 15%) are large enough to bother cross-state comparisons of raw Google averages. Any Google-vs-HCAHPS state analysis should either standardize to a common reviewer mix or include the multi-share as a covariate - this notebook provides the machinery for both. And section 8 sharpens the blind-spot argument: the reviewer group that over-samples complicated care journeys is also the one directing attention at safety, equity, and coordination - the domains official surveys capture most thinly.