Limitations and scope
Medicare-Synth is a development and research-fixture system. It is not a replacement for restricted CMS data and it is not an inferential model of the Medicare population.
What a release can support
- prototyping schemas, joins, episodes, and analytic code;
- testing code against valid and intentionally invalid records;
- exercising deterministic export, checksum, and CI workflows;
- teaching contemporary Medicare-shaped file relationships; and
- evaluating structural, relational, temporal, and declared accounting rules.
What a release cannot establish
- national or subgroup-level population estimates;
- clinical effectiveness, policy effects, or causal conclusions;
- complete distributional, longitudinal, or provider-level fidelity;
- formal privacy guarantees solely because records are synthetic;
- that matching published aggregates or study summaries recovers the real Medicare joint distribution or legitimizes synthetic estimates as empirical Medicare findings; or
- availability of a field, file, or relationship in a restricted CMS product without checking that product’s current documentation.
Current source boundary
- The default baseline is the official 2021 CMS Synthetic Claims collection,
identified as
CMS-2021-SYN-CLAIMS. - The additive 2022 CMS Synthetic Medicare Claims PUF path is bounded to the beneficiary and carrier files. Other PUF claim families and unsupported years are outside that importer contract.
- Raw source files are acquired into ignored
data/directories. Tracked manifests and evidence snapshots preserve retrieval metadata and checksums.
Generated values and unknowns
Vertical synthesis, horizontal re-keying, scenario generation, and external calibration each make different claims. The provenance status and release profile must say which method produced a field. If evidence is missing or conflicting, the project should preserve the uncertainty instead of silently choosing a meaning.
Calibrated MVP boundary (Phases 0–5)
The pinned context mvp-2021-om-ffs-diabetes-ip can match selected public
moments under disclosed transport and evidence-grade assumptions. That does not
imply:
- universal condition/year support;
- realistic multiyear trajectories, repeated admissions, or provider effects;
- diagnosis-specific inpatient analyses;
- one-click website downloads of calibrated CSV;
- completion of Phase 6 independent research-utility evaluation; or
- permission to publish synthetic estimates as Medicare findings.
See interpreting results and the diabetes–hospitalization example.
License and redistribution
The code is licensed under Apache-2.0. CMS synthetic source material is acquired through manifest verification rather than redistributed in this repository. A release owner is responsible for checking the source terms and publishing the manifest, checksums, and disclosures that accompany any bundle.