Cross-check: Populace US 2024 net income & tax ~37–43% below Enhanced CPS (benefits match)

[MaxGhenis]

Summary

Cross-checking the current Populace US 2024 build against the legacy Enhanced CPS with the same policyengine_us engine (v1.746.0), same year (2025), same variables, the two layers disagree sharply on income and tax while benefits match:

Aggregate (2025, weighted national)	Populace US 2024 (managed default)	Enhanced CPS 2024 (legacy)	Gap
household_net_income	$14.59T	$23.09T	−37%
household_tax	$3.45T	$6.03T	−43%
household_benefits	$1.99T	$1.98T	+0.4% (match)
weighted population	343.2M	~same	match

Because tax is lower in Populace yet net income is also lower, the driver is almost certainly much lower market income in Populace (lower market income → both lower tax and lower net income, benefits ~unchanged). For reference, US aggregate household net income is plausibly ~$18–23T (BEA personal income ≈ $23T pre-tax), so $14.6T looks low.

Why it matters

pe.us.managed_microsimulation() defaults to populace_us_2024, so the production web app and any default-path analysis report the $14.6T figure. This surfaced during prep for the IMA World Congress 2026 PolicyEngine tutorial (the web app and the Colab notebook quote these aggregates).

Confidence (not a sampling/loading/year artifact)

• Full Populace, direct: Microsimulation(dataset=populace_us_2024.h5) over 224,026 households / 572,780 people → net $14.586T.
• Uniform 20k-household subsample → net $14.589T (reproduces full to ~0.02%).
• Managed default uses the same registered populace_us_2024 file; managed_microsimulation reduces to the identical Microsimulation(dataset=...) call.
• Gap present for both 2024 and 2025 (not an uprating artifact).
• eCPS via pe.us.ensure_datasets(["enhanced_cps_2024"], years=[2025]) (101,384 people) → net $23.09T.

Likely related

This looks connected to the income/weight regressions already tracked in #67 (income tax regressed ~15% vs prior build) and #64 (child weights inflated ~46%). Sharing the external (eCPS) benchmark in case it helps triangulate magnitude/root cause.

Suggested diagnostics

1. Compare aggregate market-income components (employment_income, self_employment_income, capital gains, pensions, …) between Populace and eCPS, and against SOI/BEA, to locate the ~$8T net gap.
2. Confirm whether national income/AGI calibration targets are active in the 2024 build (the benefits-match-but-income-low pattern suggests income targets may be missing or under-weighted).
3. Cross-check the web app production aggregates and any published PolicyEngine US totals.

Deferring to the data team on root cause — flagging the benchmark.

[MaxGhenis]

Checked against the latest release populace-us-2024-sparse-l0-refit-57k-71a0887-national-only-20260701 (2026-07-01): simulated the release artifact with policyengine-us 1.752.2 and read the release's 32,637-target calibration diagnostics.

The aggregates in this issue reproduce (2025: household_net_income $14.50T, household_tax $3.43T, household_benefits $2.00T, population 343.2M). But the calibration hits its administrative levels almost exactly:

target	value	rel. error
irs_soi.ty2022.historic_table_2.us.all.income_tax_liability_amount	$2.105T	+0.01%
irs_soi.ty2022.historic_table_2.us.all.adjusted_gross_income	$14.78T	+0.2%
irs_soi.ty2023.table_1_1.all.adjusted_gross_income	$15.29T	−3.1%
cbo.revenue_projection.ty2023.income_by_source.adjusted_gross_income	$15.35T	−3.5%
irs_soi.ty2022.historic_table_2.us.all.ctc_amount	$82.9B	+2.8%
EITC returns (filing season week 47)	23.8M	−2.3%

So Populace is not ~40% low against administrative anchors. The Enhanced CPS side of the original comparison ($23.09T net income / $6.03T tax) carried the May 19 calibration's over-weighted synthetic top tail (AGI $24.9T vs ~$15T SOI; see policyengine-us-data#1107) — the artifact that remained latest on HF through this comparison window. Most of the gap in this issue is the benchmark, not Populace. Benefits matched throughout ($2.00T vs $1.98T).

Two real residuals remain, better tracked in their own issues:

1. Target vintage/aging. SOI TY2022/TY2023 dollar levels are applied at 2024 without aging, so simulated 2025 runs ~6–10% under current-year projections (AGI $16.0T vs CBO TY2025 ~$17T+; income tax $2.15T vs FY2025 receipts ~$2.4T). Period-aware transformation needed for SOI EITC-by-AGI calibration targets #116 covers this for EITC-by-AGI; the same period-aware transformation applies to every SOI dollar target and belongs in Ledger target aging, the same mechanism as the Translate old-vintage CD targets to current district geography #205 geography-vintage crosswalk.
2. Concept bridge + nonfiler income. household_net_income vs BEA personal income differs by employer contributions, imputed income, and survey wage undercount (employment_income $10.3T vs BEA wages & salaries ~$12.1T). Add NIPA and program-admin nonfiler controls to the US target profile #30 (NIPA and program-admin nonfiler controls) is the targeted fix — adding records/mass where support is thin rather than weight, per the policyengine-us-data#1107 lesson.

Closing in favor of #116 (generalized to all SOI dollar targets) and #30.