The Readmission Penalty Paradox

FindingsPerformance doesn't track finances. Burden does.

The expected finding did not hold. Readmission performance is uncorrelated with financial health — r = , 95% CI , spanning zero. It survives five sets of controls and 21 robustness specifications. Reported as found.

The burden is what diverges. The penalty costs roughly the same dollars everywhere but consumes more of a struggling hospital's annual income than a healthy one's. Because performance is independent of finances, that gap is structural: a uniform rate meeting radically non-uniform balance sheets.

Analysis 01Financially weak hospitals do not readmit more

Three views of the same null, in increasing order of how hard they are to dismiss: the raw relationship, the relationship after controls, and whether the design could have detected an effect at all.

The fitted line is flat, and the confidence band is narrow enough to prove it

Each point is one hospital. The line is an OLS fit; the shaded band is its 95% interval. A 10-percentage-point improvement in margin moves the readmission ratio by — indistinguishable from zero on a measure centred at 1.0.

Adding controls does not move the estimate

The coefficient on total margin at each stage, with 95% confidence intervals. If the raw null were an artefact of confounding, the estimate would move as size, capacity, payer mix, service scope and geography are absorbed. It does not budge from zero. R² rises from 0.000 to 0.142 across the ladder — almost all of it at the state-fixed-effects stage, which is a finding in itself.

This design could have detected an effect six times smaller than nothing

At 80% power and n = 2,373, the smallest correlation this study could reliably detect is 0.058. The observed effect and its confidence interval sit entirely inside the undetectable zone — which is what separates evidence of absence from absence of evidence.

Analysis 02The same penalty costs 2% of income here, 40% there

Hospitals in five equal bands by total margin, modelled penalty 0.5% of net patient revenue. Switch to dollar cost and note how flat it is: the gap is entirely about capacity to absorb, not the size of the bill.

Bars show the median; the thin line spans the interquartile range. The weakest band has no bar because every hospital in it is already loss-making — there is no income for the penalty to be a share of.

How large the gap is depends on choices a reasonable analyst could make differently

The burden multiple under each specification tested. It holds between 14× and 18× across sample restrictions, and the penalty rate is irrelevant by construction. But banding hospitals on operating rather than total margin gives 3.7× — a different question about a different kind of weakness. The headline is therefore a range, not a point estimate.

Analysis 03One condition in six shows a real financial link

Testing all six conditions separately means a 26% chance of a false positive at conventional thresholds, so p-values are corrected for the family. Five conditions are null. Hip and knee replacement — the only elective procedure in the set — survives, in the opposite direction to the equity critique: financially stronger hospitals readmit more.

Only elective joint replacement separates from zero after FDR correction

Correlation between total margin and each condition's excess readmission ratio, with 95% intervals. Benjamini-Hochberg at q = 0.05; the corrected p-value is shown against each condition. Intervals crossing the zero line are nulls.

Flagged, not claimed. Volume and patient selection in elective surgery is the obvious mechanism — stronger hospitals do more joint replacement, possibly on more marginal candidates. This data cannot test that. Without multiplicity control the effect would have been buried among five nulls; with it, the effect is real but the explanation is a hypothesis.

Analysis 04Geography explains seven times more variance than hospital finances

The largest effect in the dataset, and the one with no explanation. States with at least 15 hospitals, ten highest and ten lowest.

New Jersey and Utah are five times apart, and their intervals do not overlap

Share of hospitals performing worse than expected, with 95% Wilson intervals — several states rest on fewer than 25 hospitals, so the intervals matter. The gap between the extremes survives that uncertainty comfortably.

This needs an explanation the data cannot give. State fixed effects lift model R² from 0.019 to 0.142 — geography outweighs every hospital characteristic measured here combined. It does not track margins: Florida hospitals are financially healthy (12.1% median margin) and still perform poorly. Medicare Advantage penetration, post-acute supply and regional admission practice are all plausible; none is testable here. No map is shown, because a choropleth would imply a spatial mechanism this analysis has not established.

Analysis 05A revenue-neutral cap would move $1.3B off the weakest hospitals

Three designs, each collecting the identical $5.70B modelled pool from the same hospitals. Only the distribution changes. This is the question a policy team would ask.

Capping the penalty at 25% of net income shifts 23% of the pool, at no cost to Medicare

Share of the total penalty pool borne by each financial band under each design. Bars are ordered weakest to strongest within each group; lightness distinguishes the three designs so the comparison survives greyscale.

MethodWhat this analysis can and cannot support

Join. CMS FY2026 HRRP (18,330 hospital-condition rows) to CMS HCRIS audited cost reports FY2023 on CMS Certification Number. 2,373 of 2,392 eligible matched (99.2%).
Validation gate. HCRIS columns are raw worksheet coordinates; misreading a line number produces plausible but wrong figures. Every derived field is checked against published national benchmarks and the pipeline halts on failure.
The null survives controls. Five-stage ladder — size, occupancy, payer mix, service scope, state fixed effects — leaves the coefficient at −0.00001 (p = 0.88), HC1 robust errors throughout.
Robustness. 21 specifications across five families. The null holds throughout. The burden multiple ranges 14–18× by sample and 3.7× under operating-margin banding, so it is reported as a range.
A discarded variable. HCRIS Medicaid patient days were the intended social-risk proxy. They gave a median share of 4.7% against a published ~19%. Dropped rather than rescaled — adjusting it would have manufactured agreement and any equity finding would then be an artefact of that adjustment.
Modelled, not observed. A uniform 0.5% penalty is applied. Actual payment adjustment factors are in the CMS HRRP Supplemental Data File. The claim is "an identical penalty lands unequally", not "these hospitals received these penalties".
No system affiliation. A thin standalone margin inside a large system understates real capacity. The most likely source of overstatement in the burden figures, and not observable here.
No causal claim, and no patient-level inference. A flat correlation rules out a simple story, not offsetting mechanisms that cancel. All analysis is hospital-level.