R7 replication: earnings sharing on real PSID couples vs Favreault–Steuerle (2007)

[MaxGhenis]

R7 replication (reported, not gated): earnings sharing on real PSID couples

Does our stack — real PSID marriage histories (mh85_23) + real earnings (the family panel) + the statutory chain (42 USC 415(a)/(b)) + the #80 spouse/survivor auxiliary-benefit plumbing — reproduce the incidence structure of Social Security earnings sharing: who wins and who loses, by sex × marital status?

Anchor: Favreault & Steuerle (2007), Social Security Spouse and Survivor Benefits for the Modern Family (Urban Institute 311436, DYNASIM3 runid 440v2), Table 3. Registration: issue #42 comment 4911171806. Reported, not gated — reads no gate, changes no gate, publishes regardless. Real couples only (no generated component). Package 1b (earnings sharing, no survivor benefit, +4.5% global increase) is primary; 1a is descriptive; 1c is documented as not cleanly implementable.

Three-way table — package 1b (DYNASIM / ours / floor), anchor-weighted %

Floor = 5-seed person-disjoint half-split noise on each share. Cells with n in the low tens carry large floors (noise-dominated) by design.

group	n	lose ≥20% (D/O/fl)	lose ≥5% (D/O/fl)	gain ≥5% (D/O/fl)	gain ≥20% (D/O/fl)
married / male	378	18 / 45 / 3	49 / 86 / 4	30 / 4 / 2	17 / 2 / 1
married / female	355	8 / 7 / 3	22 / 24 / 6	60 / 63 / 4	44 / 49 / 2
divorced / male	52	8 / 34 / 15	30 / 64 / 13	21 / 4 / 7	6 / 4 / 7
divorced / female	66	9 / 0 / 0	18 / 20 / 12	45 / 56 / 18	17 / 26 / 11
widowed / male	15	36 / 54 / 27	78 / 89 / 17	4 / 10 / 18	1 / 0 / 0
widowed / female	45	38 / 22 / 14	67 / 59 / 14	11 / 32 / 18	2 / 7 / 10
never-married / male	35	0 / 0 / 0	0 / 0 / 0	3 / 0 / 0	0 / 0 / 0
never-married / female	70	0 / 0 / 0	0 / 0 / 0	1 / 0 / 0	0 / 0 / 0

The flagship, best-populated cell — married women, the big gainers — lands on DYNASIM: gain ≥20% 49 vs 44, gain ≥5% 63 vs 60, lose ≥5% 24 vs 22, each well above its floor.

Registered directional expectations — all four held

• Single-earner-couple husbands among the losers — married men lose ≫ gain (lose ≥5% 86% vs gain ≥5% 4%).
• Divorced women among the gainers — gain ≥5% 56% vs lose ≥5% 20%.
• Survivor variant helps widowed women — widowed-women ≥20% losses fall from 22% (1b) to 21% (1a); the direction holds (muted by our women's strong own careers).
• Never-married near-zero by construction — 0% winners/losers at every ≥5% threshold (the uniform +4.5% is below 5%).

Where levels diverge — every gap is a named delta (direction, not level, is the test)

• Cohort: observed PSID retirees eligible 2005–2019 (born 1943–57) vs DYNASIM's projected 1960–80 cohorts in 2049. Our older cohorts have more single-earner couples → married men lose harder (45% vs 18% ≥20%) and fewer married men are the couple's lower earner (DYNASIM's married-men gains come from lower-earning husbands + within-couple claim-timing, neither prevalent here).
• Selection: PSID coverage-≥0.8 long-stayers. Our women have strong own careers, so rely less on survivor benefits and lose less when 1b removes them (widowed women 22% vs 38% ≥20%).
• Spouse observation: the study person is a long-stayer, but the spouse's earnings are whatever PSID observed (often sparser) — biasing the higher earner's loss upward.
• No dynamics: single expected claim-age reduction; no within-couple claim-timing, mortality projection, or behavioural response.

Magnitude within ±10pp on the largest cells (DYNASIM share ≥20%): 4/12 — and all four are the well-populated married/widowed women cells; the misses are the male and thin cells driven by the cohort/selection deltas above.

Package definitions & the modeling judgment on the +4.5% scalar

Package 1b's cited definition (Table 2, printed p.15) includes a +4.5% global benefit increase (a scalar on the PIA formula factors) that makes the package approximately cost-neutral. It is directly visible in the anchor's never-married Table-3 row (97–99% show a "gain <5%" — the uniform bump), so it is treated as part of package 1b. A scalar-off sensitivity is reported alongside; the scalar-off variant moves never-married to 100% "no change" (its fingerprint). The scalar is DYNASIM-calibrated, not re-derived for our sample — in our older cohort the aggregate benefit change under 1b is −4.23% with the scalar (−8.35% without), a named cost-neutrality delta.

Package 1c (self-financed survivor annuity) is transcribed for provenance but not scored — it needs cohort-specific unisex actuarial annuity pricing at 4.6% real that the paper specifies only as "roughly actuarially fair," which would inject an unvalidated mortality model.

Coverage, artifact, tests

• Study population: 1,549 Phase-A career-frame persons (coverage ≥0.8 of ages 22–61, eligibility 2005–19), 1,016 scored (both-spouse-computable real couples). Both-spouse coverage 65.7% (married 72%, divorced 37%, widowed 54%, never-married 100%); 173/258 career-frame widowhood episodes have a computable deceased spouse.
• Artifact: runs/replication_r7_sharing_v1.json (reported_not_gated: true; registration pointer; anchor provenance with page-cited package definitions and the full verbatim Table 3 + Table 5; every pooled stat recomputable from per_seed).
• Tests: tests/test_replication_r7_sharing.py — an always-runnable consistency tier (schema, provenance, floor/verdict recompute, never-married near-zero, pure-helper units) + a skipif-PSID live seed-0 pin.

One run. black -l 79 + ruff clean.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
social-security-model	Ready	Preview, Comment	Jul 8, 2026 4:42am