Model Card

What has been tested, what the tests found, and where the model is known to be weak. Every table below is read from the JSON its generating script wrote, so this page cannot drift from the analyses it describes — the failure mode that left the rest of these docs claiming the model had no multi-listing and no acceptance modelling long after both shipped.

Read this first

Most of what follows is an internal-consistency diagnostic: it checks that the model is honest about its own uncertainty and extracts the signal its inputs contain. That is not the same as evidence that it predicts real patient outcomes. The genuinely external checks are labelled.

The interactive version of this page lives at transplant.today/model-card.html and reads the same files.

How good could any model be here?

Before asking how well the model ranks centers, it is worth asking how well anything could. This study simulates centers whose truth is known, runs the full pipeline over them, and measures how much of the true ordering comes back.

The recoverable ceiling sits at ρ ≈ … at realistic cohort sizes, and falls to … for centers with fewer than 60 observed candidates. Every correlation reported elsewhere on this page should be read against that bound rather than against 1.0. It is also why small-cohort estimates are shrunk toward a national baseline and flagged, rather than reported at face value.

Are the intervals honest?

A 95% interval should contain the truth 95% of the time. Measured against held-out later SRTR releases:

Loading coverage-audit.json…

Raw intervals were under-covered, so the shipped intervals are widened by the inflation factors shown. Coverage decaying with horizon is a real property of the data rather than a fixable bug: an estimate describes the release it came from, and centers drift away from it.

Is the Bayesian machinery calibrated?

Simulation-based calibration draws parameters from the prior, simulates data, refits, and checks that the rank statistics come out uniform. A failure would mean the priors, the likelihood and the sampler disagree with each other.

All four monitored parameters pass (KS p between 0.41 and 0.57). This is a stronger statement than "the sampler converged": it says the posterior intervals mean what they claim. It says nothing about whether the model is right about reality.

Does the ranking match the registry?

Per-center Spearman correlation between predicted access and observed SRTR transplant rates runs 0.70–0.89 depending on organ, against a recoverable ceiling of 0.92. Note this is a cross-field consistency check: wait factors come from SRTR Table B10 and observed rates from Table B7 — the same registry, so it verifies the competing-risks model, not the registry.

Pediatric

Loading pediatric-calibration.json…

SRTR publishes no pediatric wait percentiles at all, so pediatric median waits are derived from published pediatric transplant rates. That derivation was validated on adults, where both quantities are published, and recovers order far better than magnitude. Heart fails the gate outright. Treat displayed pediatric medians as directional.

Would pooling across releases help?

Loading panel-fit.json…

No. Pooling loses to simply using the most recent release, for every organ tested, because centers drift and older releases describe programs that no longer exist. The single-release design is therefore an evidence-backed choice rather than a simplification awaiting improvement. A negative result kept visible is as much a part of the record as a positive one.

Known limitations

• Estimates describe the SRTR release they were built from, not real-time allocation.
• Center discretion enters only as a center-level average, so equity analyses understate between-group disparity at a given center (L-075).
• Pediatric median waits are derived, not observed (L-076), and pediatric cohorts are small enough that some figures are the national prior wearing a center's name (L-077).
• No patient-level clinical trajectory is modeled.
• This is a research and education tool. It is not a clinical decision aid and has not been reviewed by transplant faculty for face validity (#107).

The complete register is maintained in docs/limitations.md (77 entries) and docs/clinical-assumptions-register.md (221 assumptions, 129 still needing external justification).