US fiscal-target build (0c38bc4): income tax regressed ~15% and SS dependents overshoot ~60% vs prior build

[MaxGhenis]

Summary

The new US fiscal-target build (populace-us-2024-f32c2e5-0c38bc47db89-20260616T124451Z, certified by policyengine.py 4.17.4) expands the target surface to 5,229 targets — adding Social Security components, SSI, TANF-by-state, and Medicaid. The Social Security calibration is a real, valuable improvement, but the expansion regressed income tax and overshot SS dependents. Flagging from CRFB taxation-of-benefits sentinels so the next iteration can land both.

The win (worth keeping)

Targeting Social Security fixed the elderly/population shortfall the prior income-tax-only build (0cdbb27) had:

	prior 0cdbb27	new 0c38bc4	~actual
population (2024)	360M	333M	~335M
65+ share	15.4%	17.3%	~17.5%
Social Security total	$1,363B	$1,542B	~$1,500B

For CRFB specifically, this raised the downstream Stage-B effective sample size from 126 → 1,759 (14×) when reweighting to the Trustees age distribution, and brought the taxation-of-benefits joint to within +4.7% of target (vs +18.8% on the prior base) before any adjustment. A big reliability win for any SS/benefit analysis.

The regressions (to fix before this is the long-term default)

1. Income tax regressed ~15%. Net income tax (PolicyEngine-US 1.729.0, via managed_microsimulation) fell from $2,260B (0cdbb27) to $1,934B (0c38bc4). The prior build hit the CBO individual-income-tax target (~$2.43T) at ~99% on a positive basis; overall fraction_within_10pct dropped from 85.8% → 27.7%, and income tax is among the targets that fell off. Adding the benefit-target surface appears to have pulled income tax away from its CBO target.

2. SS dependents overshoot ~60%. social_security_dependents calibrated to $81.5B vs a $51.1B target (+59%). (SSI also lands low, ~$57B vs $63B target, −10%.)

Ask

Re-balance the target surface so income tax stays within tolerance alongside Social Security — e.g., per-target tolerance/weighting, more optimization epochs, or a regression guard so income-tax (and other previously-converged) targets don't degrade versus the prior build. And check the social_security_dependents target/mapping for the +59% overshoot.

Related calibration-balance symptom: #64 (child weights inflated in 0cdbb27).

Data: CRFB taxation-of-benefits sentinels, 2024 + 2026, PolicyEngine-US 1.729.0, via the policyengine.py managed path.

[MaxGhenis]

Update from PR #79: the income-tax part of this issue was a target-semantics problem, not evidence that the microdata should chase CBO receipts. CBO individual income tax receipts are macro cash receipts; the hard micro calibration target for income_tax should be IRS SOI return-level liability.

PR #79 changes Populace so:

• CBO individual income tax receipts no longer compile into hard calibration targets.
• IRS SOI country/all income tax liability is the required federal_income_tax_total target.
• CBO receipts remain suitable as macro/reference diagnostics, not as a return-level calibration anchor.

This does not close the Social Security dependents overshoot in this issue; that remains a separate mapping/support bug to fix.

[daphnehanse11]

Validation note from an end-to-end local run on current main (843f1e0 locally):

1. I generated a 2024 Arch facts bundle for the active fiscal aliases.
2. The normal active bundle failed target coverage because the active CMS Medicaid alias emits 2025-12 facts, so I appended the valid cms-medicaid-chip-monthly-enrollment-december-2024 facts. With that addition, Populace target compilation passed:

facts: 6029
targets: 4988
gate_passed: True
cms_medicaid targets: 104
cms_aca targets: 102

3. I then ran:

uv run python tools/build_us_fiscal_refresh_release.py \
  --ledger-facts /private/tmp/populace-facts-2024-with-dec-medicaid.jsonl \
  --out /private/tmp/populace-us-health-check-20260617 \
  --release-id populace-us-health-check-20260617T1018Z \
  --skip-reform-validation \
  --skip-demographics

The run got through target compilation and ACA source refresh, then failed the critical target gate after calibration:

income_tax_liability_amount: relative_error=-0.77913
income_tax_liability_returns: relative_error=-0.839452
social_security_benefits.payment_amount: relative_error=-0.860073
ctc_amount: relative_error=-0.758743

I inspected the downloaded base H5 from policyengine/populace-us/populace_us_2024.h5. Its household_weight sums to only 25,433,517, which is far below national household scale. Since the release builder calls calibrate(..., mass="conserve"), the optimizer cannot scale the dataset up to national fiscal/count targets. That explains why the post-calibration estimates are around 20% of the relevant targets.

ACA-specific source diagnostics show the same denominator problem: the active ACA facts provide a 2024 APTC-recipient target of 19,743,689, but the source-refreshed frame only has 4,489,655 weighted eligible/APTC-support people, so the state take-up rates clip to 1.0 and still cannot reach the target under conserved mass.

So this no longer looks like an ACA-specific calculation bug. The next blocker is either:

• the base H5 published at policyengine/populace-us is underweighted / not the intended national-scale base, or
• the fiscal refresh builder should not conserve mass when starting from that base.

The new critical target gate is doing the right thing by failing this build.