Use tax liability aggregates to validate/prune disaggregated PUF records

[MaxGhenis]

Summary

Use tax liability aggregates from the original PUF aggregate records as additional calibration targets when generating synthetic disaggregated records. This would prune implausible records via L0 regularization.

Motivation

The 4 aggregate PUF records contain not just income variables but also computed tax liabilities (E05800, E06200, E06500, E09600, E10300, etc.). These are per-return averages like everything else. Currently the disaggregation only calibrates to income totals — we could additionally require that the synthetic records, when scored through PolicyEngine, reproduce the known tax liability totals.

This is valuable because:

• A record with $500M AGI but implausible deduction structure would produce wrong tax liability
• Tax liabilities encode the joint plausibility of the entire return, not just individual variables
• L0 regularization naturally zeros out records that can't contribute to matching both income AND tax targets

Approach

1. Overgenerate synthetic records (e.g., 200-400 per bucket instead of 20-40)
2. Score each through PolicyEngine to compute tax liabilities
3. Set up reweighting problem: find non-negative weights that minimize distance to:
   • Income component totals (E00200, P23250, E00650, etc.)
   • Tax liability totals (E05800, E09600, etc.)
   • Count targets from GDB
4. L0 regularization zeros out records that don't help match both sets of targets
5. Output the surviving records with their optimized weights

Tax year considerations

The raw PUF is 2015 but gets uprated to 2021. The tax liability values also get uprated. We'd score the uprated synthetic records under 2021 tax law (which PolicyEngine supports) and compare to the uprated tax liability totals.

Integration point

This would replace the current Dirichlet + donor scaling approach in disaggregate_puf.py. The overgeneration step is fast; the PE scoring is the bottleneck (~1,000 records scored is feasible).

Dependencies

• Improve top-tail income representation in enhanced CPS #606 (aggregate record disaggregation — in progress)
• PolicyEngine-US must support the target tax year's law
• microcalibrate or custom L0 reweighting (already used in enhanced CPS)

References

• L0 regularization: already used in enhanced_cps.py for sparse reweighting
• Tax variables in aggregate records: E05800, E06200, E06500, E09600, E10300, E10700, E10900

[MaxGhenis]

Correction: The tax-scoring step should be done at 2015 (pre-uprating), not 2021.

The tax liabilities in the raw PUF aggregate records were computed under 2015 tax law. If we uprate to 2021 first, the dollar amounts scale but the law doesn't — the uprated liabilities would be inconsistent with 2021 law (e.g., TCJA bracket changes, SALT cap, etc.).

So the flow should be:

1. Disaggregate at 2015 dollar levels
2. Score synthetic records under 2015 tax law via PolicyEngine
3. Reweight/prune to match 2015 income AND tax totals
4. THEN uprate the surviving records to 2021

This requires PolicyEngine-US to support 2015 parameters, which it should since parameters are defined with effective dates. Need to verify.