Synthetic-firm layer scoping: phase-2 certification path and the E12 re-scope question

[daphnehanse11]

Motivation

Max's question (Slack, 2026-07-23): the current approach doesn't generate synthetic firm microdata — how do RAND/Urban do it for the health-insurance models? Researched against the primary methodology documents (HIPSM 2011/2013/2020 docs, RAND WR-650/OP313/IJM/PMC, CBO WP 2021-15 series, Census CES WP-14-46), with adversarial verification of 25 extracted claims (22 confirmed, 3 refuted). Full report posted on #192. This issue scopes what we adopt.

What the field actually does (verified)

No incumbent model has firm microdata or worker-firm linkage either. All three major models synthesize the firm side from survey inputs and certify it against aggregates:

• Urban HIPSM: firms built entirely from the person file — workers partitioned by ESI-offer status × region × industry × firm size; each worker seeds a firm ("core employee" / "nucleus") populated with statistically similar coworkers; firm weights calibrated to SUSB counts; offer/eligibility imputed by regression (2005 CPS supplement, the last with offer data) calibrated to published MEPS-IC tables; premiums from covered lives' expenditures + size/industry admin load, calibrated to MEPS-IC and Kaiser/HRET.
• RAND COMPARE: real Kaiser/HRET employer-survey records statistically matched to SIPP workers on region / firm size / industry / offer status; sequential reweighting (SIPP workforce = truth → Kaiser/HRET weights → firm counts to SUSB by size); premiums in region×size pools (P = E(me)×AV×(1+δ)); offer decision is firm-level utility maximization validated against %-of-payroll (~11%) and offer-elasticity (≈ −0.5) benchmarks.
• CBO HISIM2 (WP 2021-15; extracted but not fully verified — read before citing): worker-anchored synthetic firms where coworker-draw probabilities are estimated from linked administrative tax data — the strongest precedent for a defensibly non-random assignment.

None of them audits worker-to-firm assignment directly. Validation everywhere is status-quo replication + aggregate targets: SUSB/BDS firm counts by size class, MEPS-IC offer/take-up/premium tables by firm size, KFF/HRET premiums and self-funding shares, literature offer elasticities by size, %-of-payroll benchmarks. All public.

Proposed scope (not for the current lock)

Path A — HIPSM-style partition-and-populate, zero new microdata. Person-side partition axes from our existing readers (imputed offer status × region × industry section × IC2 firm-size band); seed firms from core workers; populate with similar coworkers; calibrate firm weights to SUSB counts by size/industry/region (targets already in the B-side stack). Division: partitioning + offer imputation = Workstream A; SUSB weighting + validation harness = Workstream B.

Path B (COMPARE-style match to KFF EHBS microdata) stays contingent on open question 2 below.

The E12 question — proposed as a §13 referee item

E12 (the gate that catches random worker-to-firm assignment) can't lock as a linkage gate, and the current end-state reading is "phase 2 explicitly uncertified." The research adds a third option: since no incumbent certifies assignment by linkage — the field-standard bar is aggregate replication — E12 could be re-scoped to an aggregate assignment audit: the synthetic assignment must reproduce QWI/J2J earnings-by-firm-size-by-industry joint distributions (targets already committed), which would exceed HIPSM/COMPARE practice. Options for the referee round:

1. Keep E12 as a linkage gate → unlockable → phase 2 stays uncertified (status quo, stricter than field practice — a legitimate choice, but it should be chosen knowingly).
2. Re-scope E12 to the aggregate assignment audit → phase 2 certifiable at (above) the field-standard bar.
3. Both: aggregate-E12 locks now as the certification bar; the linkage formulation is retained as a registered aspiration pending a data source that could ever power it.

Proposing this as an explicit §13 item so round 2 rules on it rather than hardening option 1 by default. No change to the current IC3 block is proposed — first-lock stays firm-size-keyed as drafted.

Open questions (close before design freeze)

1. Read CBO WP 2021-15 in full — does the admin-tax-estimated coworker draw imply an estimation our SIPP job-level data could support an analogue of?
2. Is KFF EHBS microdata obtainable on terms compatible with an open-source project? (Determines whether Path B exists.)
3. What does the current HIPSM vintage use in place of the 2005 CPS supplement for offer imputation — is there a newer public offer-data source?
4. Exact form of the aggregate-E12 statistic and its noise floor (same machinery as the existing floor batteries).

Sequencing

After the IC3 lock and the five _v1 floor promotions. Nothing here blocks or modifies round 2 except the proposed §13 item, which is a question, not a design.

🤖 Generated with Claude Code

[vahid-ahmadi]

Decision record — HIPSM-first baseline, RAND/KFF contingent extension

Adopting Path A (HIPSM-style partition-and-populate) as the public, reproducible baseline after the IC3 lock and required v1 evidence promotions. Path B is retained only as a contingent health-insurance extension.

Claims and implementation boundaries for the design freeze:

1. The baseline will create real-survey person microrecords plus synthetic firm and roster microrecords. It will not claim observed worker–firm links.
2. First release axes must be supported by committed targets. The current SUSB stack is national sector × size, not region/state × sector × size; therefore region is excluded from the first calibrated design unless a suppression-aware regional target pipeline is added and reconciled to national totals.
3. Offer status requires a new public person/job input or an explicit imputation model with MEPS-IC published-table targets; it is not present in the current person-side readers.
4. For E12, adopt the “both” posture with a strict scope: an aggregate assignment audit may certify reproduction of observable size/industry employment, mean-earnings, and flow margins. It cannot certify true linkage, coworker sorting, within/between-firm variance, firm effects, or spillover claims. The original linkage/sorting E12 remains a registered aspiration and a no-go boundary for those stronger claims.
5. RAND-style KFF EHBS matching proceeds only after KFF approves the named researchers and statistical-match purpose under its DUA and confirms permitted derived use and redistribution. Raw KFF records will not be committed or used in public CI absent written redistribution permission; synthetic fixtures must cover that reader.
6. MEPS-IC employer microdata remain FSRDC-restricted; only published tables enter the open baseline. CBO’s confidential administrative-link coworker probabilities are a precedent, not reproducible from public SIPP.
7. The design will preserve separate person survey weights and analytic firm weights, deterministic seeded assignment, explicit universe reconciliation, calibration/gate separation, and negative tests preventing synthetic links from being described as observed.

Next artifact after IC3: a separate synthetic-firm ADR defining person, firm, and roster schemas; partition fallback rules; population and weighting algorithms; target/gate partitions; reproducibility; licensing; and the claims boundary above.

This decision does not modify or unblock the current IC3 ceremony and authorizes no candidate fitting before IC3 locks.

[vahid-ahmadi]

Open-question resolution — official-source review

Closing questions 1–3 for the public design, based only on primary official documentation:

1. CBO HISIM2 is a precedent, not a public-data recipe. CBO WP 2021-15 constructs 241 firm types from confidential 2017 employer/employee tax filings using each real firm’s joint employee age–earnings distribution, then estimates conditional coworker characteristics. Public SIPP job IDs do not identify a shared employer across respondents, so SIPP cannot estimate that firm-clustering or coworker-sorting kernel. Our open baseline remains HIPSM-style partition-and-populate with aggregate validation and no observed-sorting claim. Sources: CBO publication page, working paper.

2. KFF EHBS remains a restricted optional extension. KFF offers a deidentified file only through a purpose-limited DUA for named researchers, with access, security, destruction and redistribution restrictions. The posted terms do not clearly authorize release of fitted donor summaries, matched records or synthetic derivatives. Path B therefore requires KFF approval of the exact statistical-match purpose and written clarification of permitted derived/model redistribution. EHBS raw data stay out of Git, public CI and agent contexts; public code uses synthetic fixtures. Sources: 2025 EHBS, data-use form.

3. No newer documented public HIPSM offer donor was found. Urban’s latest complete official methodology located (2020) still trains offer/eligibility regressions on the 2005 CPS Contingent Worker Supplement—the last cited supplement with ESI-offer information—and calibrates to then-current published MEPS-IC tables. The current HIPSM page names no newer donor. The reproducible baseline will pin the 2005 donor plus latest compatible published MEPS-IC calibration; any claim that current internal HIPSM uses something newer remains unsupported until Urban documents it. Sources: HIPSM page, 2020 methodology.

Question 4—the exact aggregate-E12 statistic and floor—remains a post-IC3 synthetic-firm ADR/referee design item. These findings authorize no worker–firm fitting before IC3 locks and do not relax the true-linkage/coworker/firm-effect claims boundary.

[vahid-ahmadi]

Supersedes the Path A decision record above

The decision record posted here on 2026-08-01 adopted Path A (HIPSM-style partition-and-populate), under which "the baseline will create real-survey person microrecords plus synthetic firm and roster microrecords".

That premise no longer holds for the firm records. Public firm microdata exists and covers the calibration grid:

• Form 5500 — 828,330 real plan sponsors
• OSHA ITA — 383,283 real establishments

Form 5500 alone populates all 97 SUSB sector × canonical-band cells with at least ten records each, so the firm population is now observed records reweighted to SUSB, not generated ones. Implementation in #386 and #390; full finding on #192.

Confirmed with @MaxGhenis that real firm data is the preferred direction.

What changes

• Firm records: synthetic → observed. No firm row is generated.
• Roster (worker-to-firm assignment): unchanged — still constructed, because no public source links a worker to their employer.
• Claims boundary: unchanged. Points 1 and 4 of the original record still hold in full: no observed worker–firm links, no coworker sorting, no identified firm effects, no AKM decomposition.

What survives from the original record

• The E12 re-scope question — it concerns the assignment, not the firm rows, so it is untouched and still needs a referee ruling.
• Open question 4 (the aggregate-E12 statistic and its noise floor) — still a post-IC3 ADR item.
• Points 5 and 6 on KFF EHBS and MEPS-IC remain accurate; those paths were contingent extensions and stay contingent.

Leaving this issue open for the E12 question. Retitling may be worthwhile, since "synthetic-firm layer scoping" no longer describes what is being scoped.