1. The raw view - and why it misleads

Start where every co-occurrence analysis is tempted to stop: just count how often topics appear together in angry reviews.

Why this is misleading. The raw ranking is wait times × bedside manner × front desk in every combination - the three most common complaint topics paired with each other and with everything else. Two mechanical forces produce this without any real affinity: popularity (common topics co-occur with everything) and review length (a review with five topics contains ten pairs; one with two topics, a single pair - and angrier reviews are longer). A meaningful "journey" claim needs a null model that grants the critic both of those facts and asks what still stands out.

2. A null that concedes popularity and length

The null model conditions on everything except togetherness:

1. Condition on the rating band - all comparisons live inside 1–2★ reviews, so "angry reviews share angry topics" cannot create associations.
2. Stratify by topics-per-review (2, 3, 4, 5+) - so review length is held fixed; within a stratum, every review has the same number of topics.
3. Permute each topic's column within each stratum (B = 499) - each topic keeps exactly its prevalence among reviews of that length; only which review it lands in is shuffled. Co-mention counts under this shuffle are what popularity + length alone produce.

A pair is accepted when its observed count exceeds the null at q < 0.05 (Benjamini–Hochberg over all 353 testable pairs), with lift > 1.15 and at least 30 observed co-mentions.

3. What survives: 31 real associations

Reading the survivors. The corrected ranking looks nothing like the raw one. The popular topics have vanished; what remains are specific pairings: billing × insurance at 4.6× chance - by a wide margin the strongest association in the corpus - followed by cleanliness × infection control (3.2×), facility condition × amenities, and a striking cluster of behavioral health with discrimination, security, and privacy. Wait times, the most common complaint of all, earns exactly one strong partner: staffing (1.36×, but on 2,075 observed co-mentions - the highest-volume true pair in the corpus). Patients connect the wait to the empty nurses' station.

4. Eight failure bundles

Treat the 31 significant pairs as a weighted network (edge weight = log₂ lift) and let greedy modularity clustering find its natural communities.

The eight bundles (greedy modularity on the significant-edge network):

Reading the bundles. The blocks along the diagonal are the journeys, and every one of them has a face:

• Financial friction is the tightest bundle in the corpus - billing and insurance are, to patients, a single experience that begins after discharge.
• Capacity & flow links what patients see (the wait) to what causes it (the staffing). This is the plan's hypothesized access breakdown, confirmed.
• Administrative access is the paperwork journey - scheduling, front desk, records, referrals - failure that happens before and between treatments.
• Diagnostic journey (assessment × imaging) is the diagnostic cascade hypothesis, confirmed at 1.6×.
• Clinical harm collects the operating room's failure modes: safety errors with surgery (1.6×), pain management, and maternity - where the stakes make reviews name safety explicitly (1.2×).
• Environment ties housekeeping to infection fears (3.2×): patients treat a dirty ward as a clinical risk signal, not a comfort issue.
• Dignity & vulnerability is the discovery. Behavioral-health complaints cluster with discrimination (2.0×), security/visitor control (2.0×), privacy (1.4×), and accessibility - a bundle about being processed by an institution while powerless. It is invisible in any per-topic ranking.
• Inpatient recovery joins nursing responsiveness to discharge and rehabilitation - the plan's responsiveness spiral, in weaker form (1.3–1.8×).

The faded off-diagonal cells matter too: bundles are not islands. Behavioral health links outward to medication and facility topics, and the administrative bundle touches records everywhere - journeys share stations.

Two of the bundles, in the reviewers' own words:

5. Journeys are worse than incidents

If these bundles are real journeys and not just vocabulary, a review that spans more of a bundle should be angrier than a review that touches it once. It is - across every bundle, on all 196,183 labeled reviews.

Reading the severity gradient. Every bundle gets angrier as it completes, but the slopes differ tellingly. Capacity & flow is the steepest cliff: a review mentioning waits or staffing averages 3.19★; one mentioning both averages 1.67★ - a −1.51★ compound penalty, the difference between a complaint and a catastrophe. Administrative access, the diagnostic journey, dignity & vulnerability, and clinical harm all lose another ~0.6–0.7★ when the journey completes. Financial friction starts at the bottom (1.76★ with one topic - billing alone is already rage) and still finds another −0.38★ when insurance joins. Environment and inpatient recovery are gentler slopes: annoyances that compound, not cascades.

This is the strongest evidence that the bundles describe experiences rather than writing style: co-mention is not just adjacent vocabulary - it marks measurably worse outcomes.

6. Robustness: the praise mirror and the ER check

Two ways the finding could be an artifact - and what the data says.

Could the bundles just be vocabulary adjacency? If two topics simply share sentences (billing and insurance live in the same paragraph whatever the sentiment), they should pair in praise too. Running the identical machinery inside 4–5★ reviews tests this.

Inside 4–5★ reviews the same test finds 16 significant pairs (vs 31 in 1–2★). 11 of the 31 failure pairs also pair in praise - these are partly structural vocabulary: billing × insurance (5.6× in praise - people also thank hospitals for handling both), cleanliness × infection control, facility × amenities, scheduling × records.

20 pairs are complaint-specific - they simply do not pair when people are happy. The complaint-only list is where the failure-journey reading is strongest, and it contains almost the entire dignity & vulnerability bundle (behavioral health × discrimination, × security, × privacy), the capacity pair (wait × staffing), patient safety × surgery, and scheduling × front desk. Institutional indifference, it turns out, has no praise-side mirror.

Top-pair stability across facility types (stratified lift, point estimates):

Reading the robustness. The strongest pairs replicate on both sides of the ER divide with similar magnitudes - the bundles are a property of hospital complaints generally, not an emergency-department artifact. And the praise mirror sharpens rather than weakens the story: it separates the pairs that are partly how people write (billing–insurance, cleanliness–infection) from the pairs that only exist when things go wrong - which is precisely the list an operations or equity team should read.

One caveat stands regardless: a co-mention is one person describing one visit. The data cannot say the wait caused the staffing complaint, or which came first - only that they arrive together, at scale, above chance.

7. Do the journeys breathe in calendar time?

Everything above lives inside single reviews. A second, independent test of the bundles: if two topics belong to one underlying experience, months in which one of them worsens (or draws more attention) should also be months in which the other does. And if anything leads - one topic rising before another - the review stream would work as an early-warning instrument.

Method: monthly series per topic, 2019–2026 (89 months). Two signals per topic - tone (negative share of directional statements, months with ≥30 statements) and attention (share of the month's labeled reviews mentioning the topic). Both are first-differenced before correlating, so shared long-run trends (the corpus-wide tone improvement) contribute nothing; what remains is genuine month-to-month co-movement.

Within-bundle vs cross-bundle month-to-month co-movement (mean pairwise correlation of first differences):

Reading the co-movement. The bundles pass the calendar test where the data is thick enough to run it. Attention co-moves four times more strongly within bundles than across them (+0.14 vs +0.04): when billing complaints take up more of a month's reviews, insurance complaints rise with them (+0.72), and the dignity-bundle topics rise and fall as a block. Tone is measurable monthly only for the high-volume topics, and there the same two pairs that anchor the co-occurrence map co-move again: scheduling × front desk (+0.44) and staffing × wait times (+0.31) - bad months at the front door are bad months on both of its topics at once.

One honest caveat: part of the attention co-movement is a composition factor - the mix of facilities writing reviews shifts over time (most dramatically in spring 2020, when ER-heavy topics collapsed and outpatient topics surged together). The within-vs-cross gap survives this shared factor, but individual attention correlations should not be read as pure operational coupling.

The same co-movement, seen raw - the actual monthly series behind the three strongest pairs:

Reading the raw series. The financial-friction panel is the cleanest: billing and insurance attention trace each other month by month for seven years - the same peaks, the same troughs, including a joint spike in spring 2020 (billing briefly reaching ~12% of all reviews amid pandemic cancellation and coverage chaos). That is what a +0.72 change-correlation looks like in the flesh.

The tone panels show co-movement as shared direction on different levels. The most striking joint feature is the early-2021 trough: staffing negativity - structurally pinned at 85–95% - plunges to ~67% while wait-time negativity dips in step, the corpus's brief window of pandemic gratitude toward visibly overwhelmed staff. Both then climb back together through the 2021–22 capacity crisis, and drift downward together from 2023. Scheduling and front desk ride the same waves throughout. The formal statement of "together" is the differenced correlation printed in each subtitle; the levels are structural to each topic.

The pandemic ran through the bundles in sequence

The window contains one natural experiment, and it produces the closest thing this corpus has to a leading indicator.

Reading the sequence. Three phases, in order:

1. The shock (spring–winter 2020). Infection-control attention jumps from under 1% of reviews to 5–6% - and, uniquely, its tone improves while it spikes (negativity falls from ~80% to ~54% at the peak: reviewers praising masks, screening, and cleaning). Visitor restrictions and scheduling chaos spike in the same quarter; wait-time mentions collapse from 28% to 17% of reviews - the emptied ERs of early COVID.
2. The hangover (from late 2021). Staffing attention climbs to its corpus maximum (≈4.1% of reviews) and wait times overshoot their pre-pandemic level (≈31%), both staying elevated through 2022.
3. The lag between them is measurable. Infection-control attention correlates with staffing attention at +0.5 to +0.6 when shifted forward 8–12 months (and near zero contemporaneously) - the capacity crisis followed the infection shock by about a year.

The honest framing: this is one macro event, not a validated early-warning system - a single pandemic cannot establish that infection-control chatter generally predicts staffing strain. What it does establish is that the review stream picks up system-level shocks in the right order and roughly in real time, which is exactly the property the official survey pipeline (with its rolling multi-quarter windows) cannot offer. With the full corpus labeled, weekly series and facility-level event studies become feasible.

8. What it means

For hospital operations. Complaints arrive in bundles, so fixes that target a single topic under-deliver: the wait-time complaint is a staffing complaint; the billing complaint is an insurance-navigation complaint. The severity gradient gives a triage rule - the compound forms of capacity & flow and financial friction mark the reviews that end relationships (1.4–1.7★), and preventing the second failure in a journey is worth more than preventing the first.

For the equity conversation. The dignity & vulnerability bundle is the analysis's most consequential finding: behavioral-health complaints travel with discrimination, privacy, and security complaints - and only in anger, never in praise. Psychiatric patients experience institutional failures as a package. Any hospital equity review that reads discrimination complaints in isolation will miss the setting where they cluster.

For the measurement program. The bundles are natural composite indicators for the upcoming HCAHPS comparison: several (financial friction, administrative access, dignity & vulnerability) live largely outside HCAHPS's domains, making them concrete candidates for the blind-spot analysis. And because bundles are defined by co-occurrence, not frequency, they are exactly the kind of structure a star rating cannot carry.

When more of the corpus is labeled, two extensions become possible: per-facility bundle profiles (which hospitals produce which journeys) and a temporal check on whether journeys strengthen under system stress (e.g., winter capacity).