Pooled ASEC 2022/2023 vintages lack the NOW_* coverage recodes: reported Medicaid at interview thins to 24.6M vs ~58M survey under 65 on the certified artifact

Summary

On the certified default artifact (hf://policyengine/populace-us/populace_us_2024.h5@populace-us-2024-buildp-sparse-rmloss100-cae8640-20260728T011454Z), the reported-coverage person input has_medicaid_health_coverage_at_interview sums to 24.6M weighted under-65 persons versus ~57.5M survey-reported (CPS ASEC 2023 NOW_MCAID, MARSUPWT). Root cause: the pooled support spine stacks three ASEC vintages keyed by income year (source_year 2022/2023/2024 = the ASEC 2023/2024/2025 survey files), and the income-year-2022 and -2023 raw H5 inputs carry only 2 of the 18 NOW_* at-interview coverage recodes (NOW_GRP and NOW_MRK). The CPS-carried derivation maps missing/NaN source values to False, so roughly two-thirds of the weighted population — every 2022- and 2023-vintage person, 173.2M weighted under 65 — is structurally unable to report Medicaid (or TRICARE, VA, CHAMPVA, IHS, other-means-tested, or non-marketplace direct-purchase) coverage.

has_esi and has_marketplace_health_coverage_at_interview escape only because NOW_GRP and NOW_MRK happen to be the two NOW_* columns the archived raw-lane converter carried for all vintages.

Downstream consequence (policyengine-scorecard PR #69, annotation kff-medicaid-reported-denominator-sanity): the "no reported coverage" construct identifies 80.9M under-65 persons versus 25.9M survey uninsured-at-interview (NOW_COV==2, ASEC 2023) and ~27M ACS nonelderly uninsured, so uninsured-population constructs on the certified artifact are not survey-faithful. Modeled enrollment is separately CMS-calibrated and does not backfill the reported flag.

No artifacts were published or promoted as part of this investigation.

Reproduction

Weighted person counts on the certified artifact (person-table columns × household weight; the columns are physically present in the h5 and the engine surface reads them as-is), against full-file CPS ASEC 2023 (MARSUPWT/100):

The marketplace flag tracks NOW_MRK one-to-one on all vintages, so its level above the ASEC 2023 survey reflects vintage composition and calibration-era weighting, not the missing-column mechanism (out of scope here).

Artifact totals: 340.1M persons, 278.9M under 65 (57,240 households, 166,321 person rows; pe-us pin 1.764.6 per the build manifest).

import os, h5py, numpy as np
p = "~/.cache/huggingface/hub/datasets--policyengine--populace-us/snapshots/26dcad66867687f15735dc4926523e3741920836/populace_us_2024.h5"
f = h5py.File(os.path.expanduser(p), "r")
per, hh = f["person/table"][:], f["household/table"][:]
hw = dict(zip(hh["household_id"], hh["household_weight"]))
w = np.array([hw[h] for h in per["person_household_id"]])
u65 = per["A_AGE"] < 65
flag = per["has_medicaid_health_coverage_at_interview"].astype(bool)
print(w[flag & u65].sum() / 1e6)          # 24.6
for yr in (2022, 2023, 2024):             # vintage split
    m = per["source_year"] == yr
    print(yr, w[m & flag & u65].sum() / 1e6, np.isnan(per["NOW_MCAID"][m]).mean())
# 2022  0.0  1.00 / 2023  0.0  1.00 / 2024  24.6  0.00

Root cause

1. The pooled spine stacks three ASEC vintages keyed by income year. source_year on the certified person table: 2022 = ASEC 2023 survey (54,464 rows, 106.8M weighted), 2023 = ASEC 2024 (54,654 rows, 105.7M), 2024 = ASEC 2025 (57,203 rows, 127.6M). Pooling unions per-year columns via pd.concat (asec_pool.py#L141), leaving NaN where a vintage lacks a column.

2. The 2022/2023 raw H5 inputs carry only 2 of the 18 NOW_* fields. On the certified frame, NOW_MCAID, NOW_CAID, NOW_COV, NOW_MRKS, NOW_MRKUN, NOW_PCHIP, NOW_CHAMPVA, NOW_VACARE, NOW_MIL, NOW_IHSFLG, NOW_NONM, NOW_OTHMT, NOW_PUB, NOW_PRIV, NOW_DIR are populated on exactly the 57,203 2024-vintage rows and NaN on all 109,118 rows from 2022/2023. Only NOW_GRP and NOW_MRK are populated for all vintages. This is a property of the SHA-locked inputs themselves: policyengine_us_data/storage/census_cps_2022.h5 (146,133 persons = pppub23, 100% PERIDNUM match; 137 person columns; NOW_* = {NOW_GRP, NOW_MRK}), census_cps_2023.h5 (144,265 persons = ASEC 2024; 138 columns; the same two), census_cps_2024.h5 (142,125 persons = ASEC 2025; 162 columns; all 18 NOW_* fields). The 2022/2023 inputs match the archived policyengine-us-data extractor's column list before 410f63a7 ("Add CPS current health coverage anchors", 2026-03-20, which added the broader block); the 2024 input was extracted with it. The input-coverage manifest already documents this vintage asymmetry for other fields ("The 2022 and 2023 inputs also omit PECOHAB, A_EXPRRP, and A_FAMREL; only 2024 carries those alternatives" — is_unmarried_partner_of_household_head exclusion; same pattern in the employer_sponsored_insurance_premiums exclusion, #32).

3. The CPS-carried derivation silently maps missing to False. _fill_health_coverage_inputs maps has_medicaid_health_coverage_at_interview ← NOW_MCAID (L272) through _yes_code, where _source returns zeros for an absent column and fillna(0.0) for NaN. The certified flag equals NOW_MCAID == 1 exactly (weighted crosstab: 24.6M both-true, 0.0M in either disagreement cell) — no later stage rewrites it, and the export carries it (this is not the #361 export-drop class; the column is present).

4. The 2024-vintage slice is survey-plausible; selection and reweighting push the other way. Within the 2024 vintage the flag carries 24.6M of 105.7M under-65 = 23.3%, versus 21.1% in full ASEC 2023 (57.5M of 272.7M). Joining the 2022-vintage rows back to pppub23.csv by PERIDNUM (100% match, zero age mismatches), the under-65 Medicaid-at-interview rate moves from 21.1% in the full survey to 23.8% among the selected source persons, 24.5% with support clones, and 25.8% under final artifact weights — selection and calibration mildly enrich reporters, the opposite direction from the thinning. Counterfactual scale: restoring the raw codes on the artifact's existing rows and weights gives ≈ 70M under 65 (22.7M on the 2022 vintage, verified directly against pppub23; the 2023 slice extrapolated at the same enrichment; 24.6M already on 2024); at survey rates ≈ 0.211 × 278.9M ≈ 59M. A rebuild reselects and reweights, so neither is a guaranteed post-fix value — the point is the scale, ~2.5–3× today's 24.6M.

5. The Medicaid take-up stage consumes the already-thinned anchor. medicaid_take_up.py anchors on the flag (eligible reporters always take up) and hard-fails only if the column is absent (L173-L177) — there is no guard on per-vintage coverage. The CMS state-count fill compensates in enrollment terms, so enrollment-side calibration passes while the reported-coverage denominator stays thinned; the thinned anchor also shifts enrollment composition from reported toward filled and overstates saturation pressure (#170's diagnostic direction).

Why certification passed

• release_input_coverage requires the column to be present with non-default signal ("Column mass parity is not column coverage" — the gate closes the absence class, #368); 24.6M true satisfies it.
• degenerate_input_signal (175 columns) passes any column that is neither all-default nor constant.
• health_input_signal checks 2 columns (selected_marketplace_plan_benchmark_ratio, takes_up_aca_if_eligible) — no reported-coverage flags.
• ecps_parity is a populated-layer presence check (158 layers, 0 gaps).

Nothing measures per-vintage signal or survey-mass fidelity for the reported-coverage family, so a flag populated for one of three vintages certifies.

Relationship to existing issues

• #591 (pool default-value holes): same family — engine inputs the pool under-populates. Its table notes the sibling receives_medicaid is never emitted while has_medicaid_health_coverage_at_interview "is emitted"; this issue sharpens that: it is emitted for the 2024 vintage only. Partial-vintage holes are invisible to presence-style checks.
• #361 / #278 / #356 (export-side column drops): distinct mechanism — here the column survives export; the values were born thinned at source pooling.
• #296 (ingest ASEC 2024/2025 into the support pool): the natural vehicle for the data-side fix — any re-extraction should carry the full NOW_* coverage block for every vintage.
• #170 / #331 / #605 (Medicaid take-up and enrollment diagnostics): enrollment-side consumers of the thinned anchor.
• #230 (coverage-universe targets for uninsured and private coverage): targeting the uninsured universe would have surfaced this earlier and remains the calibration-side complement.

Proposed fix

1. Data side (the actual fix): re-extract the income-year 2022 and 2023 inputs (ASEC 2023/2024) with the full Census person column set — not a NOW_*-only patch, so the next missing field (#32's NOW_OWNGRP/NOW_HIPAID/NOW_GRPFTYP, #38's PERRP/A_EXPRRP/PECOHAB) does not repeat this cycle, while any new derivations stay in separate PRs — and rebuild the pooled source base. Do it as one schema migration with #296 (whose body predates the ASEC 2025 ingest: Build P already consumes income years 2022–2024) after reconciling survey-year vs income-year naming. The raw CSVs carry the data — verified for ASEC 2023 (pppub23.csv has NOW_MCAID, column 547). This runs in the gated data lane outside PR CI; no artifact publication is part of this issue.
2. Regression guard (PR #744): a per-vintage reported-coverage signal gate — for each pooled source_year, every reported-coverage input must carry non-default signal (the crisp invariant violated today: 2022/2023 vintages have exactly zero true mass for seven of the nine flags). Ships red against current bases by design, like the #368 asset gates, and turns green with the re-extraction.
3. Scorecard follow-up: once a fixed artifact certifies, the kff-medicaid-reported-denominator-sanity annotation and the reported-uninsured rows in policyengine-scorecard PR #69 can be revisited.

Numbers computed 2026-08-18 on the locally cached certified snapshot (26dcad668676…, sha256 48b9d479fb4f…) and asecpub23csv with policyengine-us 1.764.6 (the build pin).

Start with packages/populace-build/src/populace/build/us_runtime/asec_pool.py and cps_carried.py, then inspect the release_input_coverage gates and the reported-coverage inputs. Verify the pooled source-year slices and run the per-vintage signal check described in PR #744. Done means the 2022, 2023, and 2024 vintages carry non-default reported-coverage signals and the regression gate passes.