Is a Hospital's Excess-Readmission Flag a Real Signal, or Just Volume?

Data Sources and File Construction

Two public CMS hospital-level files were used: the FY2026 HRRP Supplemental Data File and the Unplanned Hospital Visits — Hospital file. HRRP reports, for six conditions and procedures (AMI, CABG, COPD, heart failure, hip/knee replacement, and pneumonia), each hospital's excess readmission ratio — predicted-to-expected 30-day readmissions given that hospital's case mix — along with predicted and expected rates, discharge counts, and readmission counts. The Unplanned Visits file reports standardized scores, denominators, and national comparison flags across fourteen measures, including excess-days-in-acute-care (EDAC) for AMI, heart failure, and pneumonia; a hybrid hospital-wide readmission/mortality measure; and outpatient procedure follow-up measures for colonoscopy, chemotherapy, and surgery.

Facility identifiers were reconciled to a common six-digit, zero-padded CMS Certification Number; the HRRP file's numeric ID was converted to this format before merging. Non-numeric placeholder text ("Not Available," "Too Few to Report") was forced to missing during numeric conversion, reflecting CMS's low-volume suppression rules.

Each source file was reshaped long-to-wide so one row represented one hospital, with condition- or measure-specific variables distinguished by suffix. The two wide files were merged 1:1 on CCN, then the merged file was reshaped back to long format on the six HRRP conditions — producing one row per hospital per condition. Hospital-level covariates from the Unplanned Visits file (EDAC scores, the hybrid mortality measure, outpatient procedure measures) are invariant across a hospital's six condition-rows and were retained unchanged through this reshape.

Outcome and predictors

The outcome was a binary indicator of excess readmissions — coded 1 if a hospital's excess readmission ratio exceeded 1.0 for a given condition, 0 otherwise. Records with a missing ratio (typically low-volume suppression) were excluded from the corresponding model. Discharge volume was log-transformed for right skew; condition entered as a categorical predictor (AMI as reference); the eight hospital-level quality covariates were standardized (z-scored) so odds ratios are interpretable per one SD change, given their non-comparable native scales (days per 100 discharges, percentages, and standardized ratios).

Modeling Strategy

Three sets of models were estimated. First, condition-specific logistic regressions predicted excess readmission status from log discharge volume alone, fit separately for each of the six conditions. Second, a pooled mixed-effects logistic regression (melogit) was fit across all six conditions jointly — log discharge volume and condition as fixed effects, a random intercept for hospital — to test whether hospital performance correlates across conditions net of volume and condition. Third, the pooled model was re-estimated with the eight standardized quality covariates added as fixed effects, to see how much of the hospital-level random-effect variance from the second model is explained by these measured quality domains. Predicted probabilities and empirical Bayes posterior means of the random effects (BLUPs) were extracted from the final model for hospital-specific case profiling.

Odds ratios below 1.0 indicate higher discharge volume associated with lower odds of excess readmissions — a volume effect strong enough to specifically flag hip/knee replacement and CABG as the most volume-sensitive programs.

N = 8,037 hospital-condition observations, 2,477 hospitals. Hospital-level variance = 0.962 (SE 0.096); LR test vs. standard logistic regression: χ²(1) = 246.02, p < .001 — hospital identity matters, net of volume and condition.

N = 3,879 hospital-condition observations, 929 hospitals. Hospital-level variance = 0.073 (SE 0.065); LR test vs. standard logistic: χ²(1) = 1.43, p = .116. "(z)" denotes a standardized (mean 0, SD 1) covariate.

From Model to Hospital: Palomar Health Downtown Campus

A statistical model is only useful if it can answer a question about one real hospital, not just a population average. Palomar Health Downtown Campus in California was selected at random from hospitals with complete model data and a mixed performance profile — some conditions above the national benchmark, others below — specifically because it was representative, not because its results were unusually strong or weak.

Instead of asking whether the hospital's readmission rate is above or below benchmark, the model asks a sharper question: given this hospital's patient volume and quality-measure profile, how much readmission risk should we expect — and do the actual results line up?

Two things stand out. First, heart failure was expected to be the lowest-risk of the four modelable conditions, yet it's the one where the hospital exceeds benchmark — a clear gap between expectation and reality, and one that lines up with a specific driver identified below. Second, CMS suppresses case counts for bypass surgery and hip/knee replacement at this hospital due to low volume, so the model can't evaluate those two — their results are real, but rest on a thinner evidence base than the other four.

Is this hospital unusual?

The model separates readmission performance into measurable factors (volume, related quality scores) and whatever hospital-specific effect is left over once those are accounted for — in effect, asking whether something unique to this hospital is driving its results. For Palomar Downtown, that residual hospital-specific effect (its empirical Bayes random-effect estimate) is essentially zero and not statistically distinguishable from an average hospital. There's no sign of a hidden, unmeasured issue — its performance is fully explained by volume and the related quality measures already in the model.

What drives the hospital's risk?

Patient volume is the strongest factor by a wide margin — it outweighs any single quality measure in the model. Overall hospital mortality performance is next, and works in the hospital's favor: hospitals with better mortality outcomes tend to have fewer excess readmissions, and that pattern holds here. The one factor working against the hospital is return visits after heart failure care — patients returning to the ED, observation, or inpatient care within 30 days of a heart failure discharge. That's the only measured factor raising this hospital's risk, and it's exactly what shows up in the heart-failure benchmark gap noted above. Return visits after heart attack care, by contrast, barely move this hospital's risk at all — even though they matter across hospitals generally, which is a useful reminder that an industry-level driver isn't automatically the driver for every individual hospital.

A plain-language version of this case profile, written for a non-technical hospital administrator audience, accompanies the full technical write-up in this portfolio's companion documents.