Add KFF eligible-uninsured moments and Medicaid 100% take-up reform rows

[MaxGhenis]

Summary

• Add KFF's 2024 state-indicator shares and 2022 flagship-brief counts for nonelderly uninsured people eligible for Medicaid or CHIP to the additive moments pipeline.
• Compute 2024 PolicyEngine counterparts on the certified Populace Build P bundle, retaining reported-uninsured as the primary construct and modeled-uninsured as an annotated alternative.
• Add a policyengine_us_inputs ReformRef route and stage the reform that forces annual Medicaid take-up to 100%, including enrollment, spending, and marginal-enrollee bridge rows for the US and every state.
• Preserve the descriptive Scorecard register: KFF joins are different_model, held_out, and either concept_mismatch for 2024 or not_computed with a dated 2024 reference for the 2022 claims.

Results

Measure	Result
Baseline Medicaid eligible, all ages	77.325M
Baseline Medicaid enrolled, all ages	72.336M
Baseline Medicaid spending	$880.019B
KFF 2024 US eligible share among remaining uninsured	24.9%
PE reported-uninsured eligible share, under 65	43.15% (34.921M / 80.932M)
PE modeled-uninsured alternative, under 65	5.71% (2.785M / 48.796M)
Enrollment change at 100% take-up	+4.989M
Medicaid spending change at 100% take-up	+$71.527B
Marginal reported-uninsured enrollees	2.792M (56.0%)
Marginal enrollees with other coverage	2.196M (44.0%)

The national eligible and enrolled estimates differ from the lane anchors by 0.03% and 0.05%, respectively. Enrollment change equals eligible minus enrolled, and the two bridge components sum to the enrollment change within floating-point precision.

The three largest absolute state differences in the primary eligible-share comparison are California (+31.99 percentage points PE minus KFF), Connecticut (+31.74), and Maryland (+28.25). These rows carry named construct differences rather than accuracy labels.

Construct annotations

• KFF applies MAGI Medicaid pathways and includes CHIP for children. PE includes SSI-linked, medically needy, optional senior/disabled, and other Medicaid pathways, while this construction does not add standalone CHIP.
• The primary PE uninsured definition is none of has_esi, Marketplace-at-interview, non-Marketplace direct-purchase-at-interview, Medicaid-at-interview, TRICARE, VA, CHAMPVA, other-means-tested, or IHS coverage, plus not modeled medicare_enrolled. The under-65 restriction is applied during moment aggregation.
• The modeled alternative additionally removes medicaid_enrolled, including anchor-and-fill enrollees presumed to have survey underreporting. KFF states no analogous underreporting correction.
• KFF uses ACS records and its own immigration and ESI-offer imputations. PE uses CPS-ASEC-based Populace; the certified artifact lacks years_since_us_entry, so the engine's five-year-bar leg uses that variable's default.
• KFF applies 2025 eligibility levels to the 2024 state indicator and 2023 levels to the 2022 brief. PE applies 2024 law to the single-period 2024 artifact.
• The reform bridge is all ages and all PE Medicaid pathways, while the KFF population is nonelderly MAGI Medicaid plus CHIP.
• Reform spending holds the baseline Medicaid state-allocation denominator fixed, matching policyengine-us baseline-branch semantics while the baseline and reform simulations run in separate processes.

Staging and ingest

• Regenerated app/public/data/moments.json: 603 total rows, including 52 KFF concept_mismatch rows and five KFF not_computed rows with 2024 references.
• Committed US/state diagnostic CSVs and 214 PE counterpart rows under data/pe/.
• Staged 208 reform rows under sources/campaign-20260818/us/, covering four components across the US plus 50 states and DC.
• Ingested those rows into data/scorecard.db as 208 exhibits under canonical ReformRef key 849973669b6526d6; offline re-ingest is idempotent.
• Person-level baseline and reform extracts remain outside the repository under /private/tmp.

Ready-to-run ingest command:

PYTHONPATH=. .venv-pe/bin/python -m scorecard_db.ingest_campaign data/scorecard.db sources/campaign-20260818/us

Validation

• Required deterministic 5,000-household sample first; the bounded-memory eligibility path matched the ordinary native engine output exactly across all 14,254 person rows and every extracted column.
• Full baseline and reform ran in separate OS processes through pe.us.managed_microsimulation() on policyengine-us 1.764.6 and certified bundle populace-us-2024-buildp-sparse-rmloss100-cae8640-20260728T011454Z (us-5.0.2).
• uv run --offline --no-project --active pytest tests/ -q: 179 successful tests.
• Repository-wide Ruff lint and formatting completed successfully (51 Python files).
• The app lint/build was not run because this worktree has no installed app/node_modules; network access was prohibited, so no dependency install was attempted.

Review findings (main session)

A post-lane probe of the certified artifact's person inputs found the
reported-uninsured denominator is not survey-faithful: 80.9M people under 65
carry none of the nine reported-coverage inputs, versus roughly 27M
ACS-reported nonelderly uninsured in 2024. Employer (146.9M under 65) and
Marketplace (21.4M) inputs sit near survey levels, but the reported
Medicaid-at-interview input sums to 24.6M under 65 against roughly 60M
survey-reported Medicaid — the sparse build thins it, and CMS-calibrated
modeled enrollment does not backfill the reported flag. Consequences, now
annotated on the joined rows:

• The primary 43.15% reported-uninsured share should be read with that
  denominator context; the modeled-uninsured variant (5.71%) and the
  eligible-and-not-enrolled counts (2.79M uninsured marginal enrollees) are the
  informative PE rows for this join.
• The large state deltas (CA +31.99pp, CT +31.74pp, MD +28.25pp) combine this
  denominator effect with a genuine construct difference: PE's eligibility
  formula includes state coverage regardless of immigration status (e.g.
  CA's is_ca_medicaid_immigration_status_eligible leg), while KFF's ACA-lens
  categories class those people as ineligible due to immigration status.
• Artifact-side follow-up to file in microcosm: reported-coverage input
  thinning in the sparse build (adjacent to the untargeted-input zeroing class,
  microcosm#361).

Follow-ups

• Re-verify the rendered KFF table's first-column label against KFF's downloadable CSV before narrowing Medicaid/Other Public to a more specific metric label.
• Complete the Medicaid/CHIP concept split tracked in microcosm#321, then add a CHIP-inclusive PE child counterpart.
• Add a 2025-rules PE variant when the certified artifact supports that policy period.

🤖 Generated with Claude Code

[vahid-ahmadi]

Review — KFF eligible-uninsured moments + 100% take-up rows

Ran the branch locally (179 passed / 4 skipped, ruff clean, CI green) and verified the external side against the live source:

Verified:

• KFF values re-fetched from the live indicator page: US 24.9 / AL 9.5 / CA 32.5 / AK 40.8 / MD 23.3 all match raw/state_indicator_2024acs_2025levels.md exactly, vintage confirmed (2024 ACS, 2025 eligibility levels). Bonus: the page's own column header reads "Medicaid/Other Public Eligible" — matching the staged STATE_SOURCE_COLUMN, so the kff-column-label-verbatim caution (annotations.json:47, source.json diagnosis_upstream) can be retired.
• Take-up vocabulary is the engine's real input: takes_up_medicaid_if_eligible (compute_kff_medicaid_extract.py:30) is in compute_counterparts.py's proven fulltakeup set (line 60), not a guessed name.
• Arithmetic identities hold: 77.325 − 72.336 = 4.989M enrollment change; bridge 2.792 + 2.196 = 4.988 ≈ Δ within fp.
• reform_ref exhibit route is fail-loud: mutual exclusion with external_claim_match, framework pin, reform_key cross-check all raise (ingest_campaign.py:176-199); idempotency stays exact-run_id, no prefix sweep.
• The 80.9M reported-uninsured denominator anomaly is measured and honestly diagnosed (annotations.json kff-medicaid-reported-denominator-sanity: thinned Medicaid-at-interview input, 24.6M vs ~60M survey).

Should consider:

1. Primary/alternative designation contradicts the lane's own sanity annotation. The denominator annotation concludes the reported-uninsured share "is not a survey-faithful uninsured measure on this artifact" and names the modeled variant + eligible-and-not-enrolled counts as "the informative PolicyEngine rows for this join" — yet reported-uninsured (43.15%) is the primary construct and modeled (5.71%) the annotated alternative. A consumer surfacing primary rows shows exactly the comparison the lane says is uninformative. Either swap primacy, or carry the annotation's verdict onto the primary rows' display status so the 43.15%-vs-24.9% pairing can't render unqualified.
2. Minor: meta["note"] (ingest_campaign.py:205,240) raises a bare KeyError on a note-less exhibit row rather than the labeled ValueError every sibling contract violation gets.

Solid lane overall — the construct-difference annotations are the most complete of any source in the repo.

[vahid-ahmadi]

Flagging a CI issue on this branch that isn't about its contents.

This PR is running an older workflow than main's, so its green tick means less than it looks. GitHub runs a pull_request workflow from the PR's own branch, and this one forked before #74 — so it still has the pre-#74 ci.yml, which has neither the determinism check (two builds must agree on the content hash) nor the no-drift gate (git diff --exit-code). It also still tracks data/scorecard.db, the committed database #74 removed.

I found this auditing all 16 open PRs after hitting it on two of my own (#49 and #64, both since fixed). Four branches were affected; this is one of the two remaining.

The fix is mechanical and nothing about your own changes needs to move:

git checkout us/kff-medicaid
git merge origin/main
rm -f data/scorecard.db          # gitignored since #74
PYTHONPATH=. python -m scorecard_db.build_db data/scorecard.db
pytest tests/ -q
git status --short               # must be clean

I haven't pushed to your branch. Happy to if you'd rather I did — say the word.

There's also a structural fix in flight at #95: a gate-freshness check that runs from the base via pull_request_target, so a stale head can't skip the gates in future. It reads main's ci.yml to derive what to require, rather than hardcoding it, so gates added later are enforced automatically.