Add state income tax revenue as calibration target

[MaxGhenis]

Problem

State/district datasets are not calibrated to actual state income tax revenue, leading to potentially large deviations from real-world collections.

Evidence: Ohio deviation

When modeling Vivek Ramaswamy's proposal to eliminate Ohio capital gains tax, we found:

Metric	PolicyEngine	Actual
Ohio income tax revenue	$24.4B	~$10B
Deviation	2.4x	-

Sources for actual Ohio income tax revenue:

• Policy Matters Ohio: $9.52B for FY2024
• Buckeye Reporter: $1.9B in Q1 2024

Current calibration targets

Per fit_calibration_weights.py, state/CD datasets currently calibrate to:

• SNAP
• Health insurance premiums
• Household counts (stratum_group_id 4)

But not state income tax revenue.

Impact

• Absolute revenue estimates for state tax reforms are unreliable
• Distributional shares (e.g., "83% of benefit goes to top 1%") are likely still valid since they're relative
• Users analyzing state tax policy may be misled by raw revenue numbers

Proposed solution

Add state income tax revenue as a calibration target using Census Bureau's Annual Survey of State Government Tax Collections.

Data source

The Census STC provides individual income tax collections by state annually since 1939. FRED also provides this data in a convenient format (e.g., OHINCTAX for Ohio).

Implementation

1. Create pull_state_income_tax_targets.py to fetch Census STC or FRED data
2. Add state_income_tax_state.csv calibration target file
3. Include state_income_tax in the target_filter variables list in fit_calibration_weights.py
4. Re-run calibration

Related

This was discovered while modeling Ramaswamy's Ohio capital gains tax elimination proposal for PolicyEngine analysis.

[MaxGhenis]

Pre-calibration comparison: Simulated vs Census STC targets

Full comparison spreadsheet

Three-way comparison of Census STC FY2023 targets vs national enhanced_cps_2024 vs current local (state-aggregated district) files:

Metric	National ECPS	Local ECPS
Total ($B)	$461 / $495	$492 / $495
Overall ratio	0.931	0.994
Median state ratio	0.925	0.932
Mean	error

Notable local file outliers:

• NJ: 1.53x ($27.5B vs $17.9B target)
• WA: $2.9B vs $0 target (no income tax state)
• CO: 0.51x ($7.0B vs $13.7B)
• AR: 0.49x ($2.1B vs $4.4B)

Ohio is at 1.02x in local files — the 2.4x from the issue description was likely an older dataset version.

PR #497 adds state_income_tax as a calibration target. Rebuilding to measure improvement.

[MaxGhenis]

Post-calibration results

After adding state_income_tax as a calibration target (PR #497) and running 5000 epochs of L0-regularized calibration:

Updated comparison spreadsheet

Metric	Pre-cal local	Post-cal
Total ratio	0.994	1.004
Median state ratio	0.932	1.005
Mean |error|	15.3%	0.7%

Every income-tax state is now within ±2% of target. The worst outliers were fixed:

State	Pre-cal local	Post-cal	Target
NJ	$27.5B (1.53x)	$18.3B (1.02x)	$17.9B
CO	$7.0B (0.51x)	$13.8B (1.01x)	$13.7B
AR	$2.1B (0.49x)	$4.4B (1.00x)	$4.4B
KY	$4.3B (0.69x)	$6.2B (1.01x)	$6.2B
ND	$0.3B (0.56x)	$0.5B (1.00x)	$0.5B

Note: The 25k-per-CD person limit in stacked_dataset_builder.py prevents building full state h5 files with these new weights (CD 101 exceeds the limit). This needs to be addressed before publishing. Filed #499 for the precomputation architecture improvement.