Employer–firm relationships: US-first plan, evaluation gates, and two-workstream split

[vahid-ahmadi]

Employer–firm relationships in Populace (US-first)

📊 Visual plan (live): https://employer-firm-plan.vercel.app

Per Max's ask, this is the plan for attaching job spells with employer attributes to the Populace dynamic person panel — US first, following the ECPS recipe (impute from panel data → calibrate to administrative aggregates) and this repo's evaluation discipline (pre-registered gates, noise floors, one-shot candidate runs). The plan has been through two independent adversarial reviews (a data-claims audit with every source web-verified, and a modeling/precedents review against DYNASIM, MINT, PENSIM, Census SSB/SynLBD, AKM/BLM); rev 1 of the page incorporates both.

Architecture

flowchart LR
  subgraph A["Workstream A — Daphne (person side)"]
    SIPP["SIPP job-level panel<br/>job IDs, tenure, industry,<br/>establishment size (not firm)"]
    CPS["CPS ASEC + tenure supp.<br/>NOEMP = true firm-size label"]
    PSID["PSID / NLSY<br/>long-horizon tenure"]
    IMP["microimpute QRF<br/>spells onto CPS persons +<br/>transition hazards (phase 1)"]
    SIPP --> IMP
    CPS --> IMP
    PSID --> IMP
  end
  subgraph B["Workstream B — Vahid (firm side)"]
    MARG["SUSB · BDS · CBP · SynLBD<br/>firm-type margins & dynamics"]
    FLOW["QWI · J2J · JOLTS · BED<br/>flow & earnings aggregates"]
    REG["BLM firm-type register<br/>(types, not firm IDs)"]
    MARG --> REG
  end
  MATCH["match layer (interface)<br/>C1 spell schema is the contract;<br/>calibrated to QWI/J2J,<br/>disjoint from gate cells"]
  IMP --> MATCH
  REG --> MATCH
  FLOW --> MATCH
  HARNESS["dynamics harness<br/>E1–E12 battery, noise floors,<br/>pre-registered gates"]
  PE["PolicyEngine-US<br/>employer-keyed rules (ESI, mandates)"]
  MATCH --> HARNESS
  MATCH --> PE

Loading

Phasing

• Phase 0 (static): employer attributes (industry, firm-size class, tenure) imputed onto CPS persons; calibrated to SUSB. Enough for firm-size-threshold policy cross-sectionally.
• Phase 1 (dynamic): job-spell transition hazards (separation, hire, job-to-job, destination attributes) trained on SIPP, scored against SIPP holdouts, calibrated to QWI/J2J.
• Phase 2 (two-sided, gated): BLM-style firm-type register — firm types (industry × size × state), not firm identities, since rosters are unidentified from public margins. Goes ahead only if phase-1 gates pass and the E12 identification story holds; otherwise claims scope down to firm-size-keyed policy only.

Evaluation criteria (to be pre-registered, gates.yaml style)

E1–E12 battery (full table on the live page), highlights:

Gate	Moment	Reference
E1	employment share by firm-size × industry	SUSB
E2	separation/hire/J2J rates by age × sex	LEHD J2J
E4/E5	employer-retention pairs + multi-window attachment runs (catches chained-model persistence understatement)	SIPP holdout
E9	earnings-change distribution conditional on transition type (layering-coherence gate)	SIPP holdout
E10	regression gate: existing locked PSID earnings gates must still pass after the spell layer attaches	existing gates.yaml
E11	J2J flows by origin × destination firm size (the size ladder)	LEHD J2JOD
E12	within/between-firm earnings variance + coworker correlation — the moments that would be wrong under random worker-to-firm assignment	bias-corrected AKM (Song et al.; KSS), BDS

Protocol rules from review: calibration cells pre-registered disjoint from gate cells; a SIPP-vs-J2J seam-reconciliation run before thresholds lock; noise floors before any threshold number.

Key review findings already folded in

1. SIPP 2014+ measures establishment size, not firm size (EJB1_EMPSIZE is location size) → CPS ASEC NOEMP is the firm-size training label.
2. QWI publishes mean earnings, never medians → E7 restated on means.
3. QWI/J2J dimensions are pairwise-restricted (firm size XOR firm age; one demographic pair; sub-sector industry × size at state level only).
4. No one has published a credible synthetic matched US employer–employee file — Census released synthetic persons (SSB) and synthetic establishments (SynLBD) but never the link. Phase 2 would be first; the paper must say so.

Division of labor

Two independent workstreams meeting at three frozen interface contracts (details on the live page):

• @daphnehanse11 — Workstream A (person side): SIPP job-level reader (label-verified, family.py pattern), CPS NOEMP/tenure loaders, harmonized PSID/NLSY tenure series, QRF spell imputation, phase-1 transition hazards; owns SIPP noise floors and gates E3–E5, E8–E10 plus the seam-vs-J2J reconciliation run. Deliverable to B: the person panel emitting spells in the frozen C1 schema.
• @vahid-ahmadi — Workstream B (firm side): external-target pipeline (SUSB/BDS/CBP/QWI/J2J/JOLTS/BED extracts committed with provenance notes), canonical firm-size banding decision (C2), calibration with the target/gate partition, BLM firm-type register; owns gates E1, E2, E6, E7, E11, E12 and the aggregate-side floor studies.

Interface contracts: C1 spell schema and C2 canonical firm-size banding freeze in week 1 (changed only by joint PR); C3, the jointly-authored employer gate block, locks after a referee round in week ~4 — no one-shot candidate runs before C3 locks. Sync points are weeks 1, 4, and ~11 (phase-2 go/no-go); everything else is asynchronous — A can develop against fixture spells, B against a synthetic C1-conforming file.

cc @MaxGhenis — process update: plan built, adversarially reviewed twice (data claims + modeling/precedents), revised, split into the two workstreams above, and deployed for browsing at the Vercel link on top. Next concrete step is the week-1 items: readers + ADR (A), target pipeline + banding proposal (B).

🤖 Generated with Claude Code

[daphnehanse11]

Reviewed the plan and the live page — overall this is strong, and the repo checks out against it: the locked PSID earnings gates E10 anchors to are real (gate_1, gate_2, gate_w1 in gates.yaml), and the harness already has mobility_matrix/transition_rates/zero_spells views that E6/E8 can reuse. Five points before the week-1 freeze, each verified against primary sources. The first three affect the contracts.

1. C1 needs a class-of-worker field (and an explicit geography join). SUSB excludes government establishments, Public Administration (NAICS 92), crop/animal production, and employers without paid employees; a QWI in-scope job is at a non-federal employer (LEHD definitions). The plan page already flags "private ownership only" as a QWI caveat on the target side; the gap is the spell side: as specified (person_id, spell_id, start_period, end_period, industry, firm_size_band, earnings_share, primary_job), C1 can't exclude the ~15% of workers who are out of the calibration universe, and self-employed spells have no defined firm_size_band. QWI/J2J targets are also state-level, and C1 carries no geography — fine if the join key lives on the person table, but the contract should say so.

2. No current representative source observes firm size × tenure jointly — the phase-0 conditioning DAG needs to say where that dependence comes from. ASEC FIRMSIZE/NOEMP refers to the employer from the preceding calendar year (IPUMS); tenure is a January-biennial supplement referring to the current employer, with no firm-size question (IPUMS supplement notes); SIPP 2014+ has tenure but only establishment size (per the plan's own finding); NLSY observes both but only for two non-representative cohorts. One dated exception, which the plan page itself points at ("the redesign dropped the all-locations question"): pre-redesign SIPP — e.g. the 2008 panel — asked worker-reported firm size at all locations alongside job spells and tenure, so it observes the joint distribution in a representative panel, circa 2008–2013. Suggest the ADR name the bridge explicitly — 2008-panel SIPP joint structure (aged), SIPP establishment size as a proxy chain, or partial January→March rotation-group linkage — rather than leaving the QRF to imply it. The reference-period mismatch (last year's longest job vs current job) should be pre-registered alongside.

3. C2 banding: the post-2019 ASEC label cannot resolve the 50-employee threshold. ASEC firm-size bands were 10–49/50–99 during 2011–2018 but reverted to 10–24/25–99 from 2019 on (IPUMS) — the ACA applicable-large-employer cut sits inside the 25–99 band. And that cut is 50 full-time employees including FTEs at 30 hrs/week (IRS), not headcount. QWI (20–49, 50–249) and detailed SUSB classes do have a 50 boundary. C2 on the plan page already settles the semantics half (administrative firm size, per review F5); what remains is (a) how the 50 cut is identified given the training label can't see it — calibration side, or pooling pre-2019 ASEC years across the band break — and (b) headcount vs FTE, presumably headcount bands plus an hours join on the person side.

4. QWI counts jobs, not persons. "If a worker has multiple employers in the same quarter, each person-employer relationship would be counted as a separate job" (LEHD). Phase 0 is primary-job-only, so calibrating person-spells to QWI cells carries a built-in wedge on the order of the multiple-jobholding rate (~5%, varies over time). The plan gestures at this in risk 4 — suggest making the job-count→person-count adjustment an explicit pre-registered item in C3 rather than a footnote.

5. Question on E10's scope. If spells attach strictly downstream of the locked earnings engine (spells condition on earnings, never the reverse), E10 is nearly a no-op. But E9 scores earnings-changes conditional on transitions — the moment transitions drive earnings, that's a new candidate on a locked-gate surface and the one-shot discipline applies. Which side of that line are phases 0–1 intended to sit on, and where does this slot relative to the M-sequence?

Happy to fold 1–4 into the C1/C2 drafts as part of the week-1 work.

[vahid-ahmadi]

Thanks @daphnehanse11 — all five verified points hold up, and yes please fold 1–4 into the C1/C2 drafts as week-1 work. Notes:

1–2. Agreed on class_of_worker in C1 and the explicit person-table geography join. On the conditioning DAG: the 2008-panel SIPP bridge is the right call — it's the only representative joint observation of firm size × tenure, and the ADR should name it (aged joint structure from 2008–2013, with 2014+ SIPP establishment size as the proxy chain), plus the ASEC reference-period mismatch as a pre-registered caveat.

3. Agreed the canonical band set must put an edge at 50 (QWI and detailed SUSB support it) and that post-2019 ASEC maps only coarsely — the mapping should surface the 25–99 straddle explicitly rather than silently assigning it. Headcount bands, FTE/hours via a person-side hours join. The 50-cut identification route (calibration-side vs pooling pre-2019 ASEC across the band break) gets pre-registered in C3 rather than decided implicitly.

4. Agreed — the QWI job-count→person-count adjustment moves from footnote to an explicit pre-registered C3 item.

5. The line sits exactly where you suspect, and it moves between phases:

• Phase 0 is strictly downstream: spells/attributes condition on (among other covariates) earnings; nothing feeds back. E10 is near-trivial there and is kept only as a cheap regression check that attaching columns didn't perturb the scored views.
• Phase 1 crosses the line by design: E9 requires transitions to drive earnings-changes, so the joint spell+earnings generator is a new candidate on the locked-gate surface, and the full one-shot discipline applies — registration on the Candidate 2 design: latent-permanent conditioned chained QRF (research memo) #42 pattern, run against both the existing locked PSID gates (that's E10's real job) and the new employer block after C3 locks. It does not amend any locked gate; the employer block is a separate pre-registration.
• Relative to the M-sequence: this sits outside it — the employer gate block is its own track (like the UK ADR-0002 posture), and nothing here touches gates.yaml until the C3 referee round.

Workstream B status: branch employer-firm-workstream-b is underway (target-pipeline extracts + provenance, firms/banding.py, loaders, ADR-0003 draft covering the above); PR to follow, with points 1–4 reflected in the ADR text for you to edit against.

🤖 Generated with Claude Code

[daphnehanse11]

Registering a downstream consumer for C1 before it freezes: SNAP ABAWD work-requirement compliance (populace side; context: PolicyEngine/policyengine-us#8961 and the Build K reform runs).

The first release-scale runs of HR1's ABAWD provisions surfaced two structural gaps in the static model, and both land exactly on Workstream A's phase-1 outputs:

1. Monthly hours. The ABAWD test is 80 hours per month (7 CFR 273.24(a)(1)); the static model proxies it with annual-average weekly hours, so intermittent workers — the population the time limit actually binds — pass all twelve months. Replacing that proxy requires month-resolved employment/hours from imputed job spells: precisely the phase-1 spell + transition-hazard deliverable.

2. The 3-in-36 clock. The static model cannot track countable months, so it treats noncompliance as immediate permanent ineligibility — its own docstring flags this as the biggest simplification, worst in FY2026 when HR1 restarts clocks for newly covered populations. Month-counting over a 36-month window (plus regaining via 80-hours-in-30-days and the one-time 3-consecutive-month allotment, 273.24(d)-(e)) is inherently a dynamic-panel computation. SNAP time limits are a natural flagship consumer for this panel.

Schema ask for C1: carry enough granularity per spell-month to derive 80-hours/month compliance — ideally monthly hours (or hours bands + part/full-time), not just spell start/end + annual earnings. class_of_worker (already agreed above) also matters here for the workfare/work-program interaction. If monthly hours are out of scope for the freeze, an explicit "hours granularity deferred" note in the ADR would keep the contract honest about which consumers it can't yet serve.

Secondary note for Workstream B: industry-conditioned separation hazards would sharpen this use case too — ABAWD churn concentrates in seasonal industries — but nothing there needs to change at C1/C2.

🤖 Generated with Claude Code

[vahid-ahmadi]

@daphnehanse11 ABAWD consumer registered — ADR 0003 (PR #195) now carries an explicit C1 deferral note: no hours column at freeze, so C1 cannot yet serve monthly-hours consumers (no annual-proxy workarounds), and hours_band/monthly hours is the first scheduled C1 amendment once phase-1 establishes what hours granularity SIPP supports. class_of_worker is already in the frozen schema for the workfare interaction. Noted on the B side: industry-conditioned separation hazards for seasonal-industry churn — the committed QWI/J2J extracts are already sector-resolved, so nothing changes at C1/C2.

🤖 Generated with Claude Code

[MaxGhenis]

Cross-PR integration review of the week-1 batch (#194, #199, #201, #195) — one hazard worth fixing this week, before the C1/C2 freeze:

The NOEMP↔FIRMSIZE code-space seam. #194 emits raw Census NOEMP codes (0–6) with native string labels; #195's cps_firmsize_to_canonical expects IPUMS FIRMSIZE numbering. Same concept, different code spaces — NOEMP 6 = 1000+ while FIRMSIZE 6 = 50–99 — so wiring A's output into B's mapper today silently mis-bands firm sizes. No adapter or seam test exists yet. Suggested contract addition: a named adapter (noemp_to_firmsize or a CodeSpace enum on the reader output) plus one seam test that round-trips every NOEMP code through to CanonicalBand and asserts the intended bands — freeze it with C2.

Related freeze-readiness notes:

• C1/C2 are not yet frozen (ADR 0003 still "Proposed"; Add label-verified SIPP job-level monthly reader (EJB records) with C1-preview spell collapse #201 is explicitly C1-preview) — fine for week 1, but the seam above should land inside the freeze, not after it.
• class_of_worker vocabularies differ across all three readers (LJCW / PEIO1COW / EJB_CLWRK) and none emits the ADR's canonical 5-way yet — the C1 freeze needs the mapping tables or a declared deferral.
• What already agrees everywhere and matters most: the establishment-size-vs-firm-size training-label discipline (Employer–firm relationships: US-first plan, evaluation gates, and two-workstream split #192 finding 1) is honored identically across all three readers and the C2 banding. The key scientific contract is holding.
• Mechanical, affects merge order: every branch's tests/tier_counts.json is stale against today's master (unit 420 / artifact 1011 after M6 factory: (0,24) projection-coverage pad + pe-us param-dir binding (3e patch: F1/F2/N1 + bridge test) #197). Whichever PR merges next recounts on the merged tree via the full-collection tier-policy test; each subsequent merge recounts cumulatively. Also: the three fork-based PRs get no pytest/lint CI (fork-gated — only Vercel runs), so local runs or maintainer-approved workflow runs are the check until that changes.

🤖 Generated with Claude Code