Diabetes–hospitalization example
Question
Before analyzing restricted Medicare files, how could we test code that compares hospitalization occurrence among people with and without diabetes—without claiming the synthetic answer is the Medicare answer?
Cohort (artificial)
- Setting: 2021 Original Medicare fee-for-service–like population
- Age: 65+
- Enrollment: full-year Parts A/B under the project’s simplified enrollment rule
- Context ID:
mvp-2021-om-ffs-diabetes-ip
These are synthetic records for methods development. They are not reconstructed beneficiaries.
Core tables and fields
Typical analyst-facing pieces in this MVP:
| Piece | Role in the example |
|---|---|
| Beneficiary demographics | Who is in the artificial cohort |
| Enrollment | Simplified full-year A/B membership |
| Chronic-condition indicators | Diabetes (sp_diabetes) and heart failure (sp_chf) flags |
| Inpatient claims | Presence of at least one inpatient claim (all-cause) |
Important definition: the outcome is any inpatient admission, not a diabetes-principal diagnosis and not a diabetes-specific admission reason. Heart failure is available as a stratifier/indicator, not a fully modeled comorbidity system.
Public anchors (selected, approximate)
The Phase 5 worlds share three frozen calibration targets:
| Target | Approximate value | Source framing |
|---|---|---|
| Diabetes prevalence | 27.7% | 2018 CMS aggregate (transported into 2021 context) |
| Heart failure prevalence | 12% | 2022 CMS aggregate (transported) |
| Inpatient risk among people with diabetes | ~13% | Weaker industry prior (not a clean peer-reviewed Medicare estimand) |
Transport across years and populations is an explicit approximation. Matching these moments is not a comprehensive 2021 Medicare benchmark.
Illustration from stored Phase 5 worlds
All example worlds below use 6,000 simulated beneficiaries and the same selected calibration targets. Free assumptions still differ.
| World (stored example) | Method highlight | Free background IP assumption | Inpatient claim records (stored) |
|---|---|---|---|
| Reference generator | Adjust generator parameters toward targets | P(IP | no diabetes) = 0.04 | 404 |
| Alternate background risk | Same calibration targets | P(IP | no diabetes) = 0.08 | 572 |
These counts come from the checked-in Phase 5 family summary
(reports/phase5/world_family_summary.json). They are illustrative stored
outcomes, not universal expected counts for every run configuration.
What you can practice
- Grouping and merging beneficiary and claims-style tables
- Defining exposure (diabetes indicator) and outcome (any inpatient claim)
- Checking analytic units (person vs claim)
- Running simplistic associations as code tests
- Sensitivity checks when free assumptions change but calibration targets match
What you cannot infer
- A real Medicare diabetes effect on hospitalization
- Validated predictive or risk-adjustment performance
- Diagnosis-specific admission patterns
- Causal effects, policy conclusions, or population-representative rates
- That the Phase 6 independent-fidelity question has been resolved (the current disposition is no-go for lack of independent evidence)
Reproduction (developer path)
There is no no-code calibrated download on the website. The Phase 5 summary is a historical record. Current code writes a separate v2 output with different evaluation semantics. Ask a developer to replay the original producer in a separate checkout; the developer guide explains the boundary. A pinned Phase 6 report records a limited workflow check and an independent-fidelity no-go.
For a literature-paired walkthrough (Andes MMWR prevalence in-domain vs Drye
admission rates out-of-domain), see the checked-in example. Run it only in a
separate checkout if you need to preserve its stored COMPARISON.md; the script
rewrites that file:
uv run python examples/andes-mmwr-drye-diabetes/scripts/run_example.py
Details: Calibrated worlds (developers) ·
phase docs under docs/phase5/ ·
examples/andes-mmwr-drye-diabetes/.