[MaxGhenis]

Summary

For tax years 2013–2023, the bundled taxsimtest (cd2026081819) puts the Additional Medicare Tax (AddMed) in fiitax. From 2024 it leaves it out. In every year it also counts AddMed in tfica.

I checked whether the emulator should copy that year split. It should not:

• On Inclusion of the additional medicare in fiitax in TAXSIM #416, TAXSIM's author said AddMed belongs with FICA.
• On Does taxsimtest count the Additional Medicare Tax twice (in fiitax and in fica/tfica)? #1225 (2026-09-24) he called its inclusion in fiitax "an error that was present in 2000-2023" and said he believes he has corrected it. The tax itself starts in 2013.
• The bundled binary predates that correction.

The Microsimulation runner already leaves AddMed out of fiitax (since #963). This PR:

• documents the convention with the actual evidence and pins it with tests. The old comments cited form lines that were never extracted. They also presented the emulator's layout as taxsimtest's ("AddMed … flows through v44"), but taxsimtest's line 44 is AddMed alone, and the emulator had no addmed column.
• adds the missing addmed output column.
• fixes the single-household path's frate. It still added AddMed in every year, which contradicted that path's own fiitax and the runner.

Evidence

Binary behavior (engine output, cd2026081819). Single filer, TX, $400K wages, idtl 2:

year	fiitax	v28	addmed	tfica
2021	111,951.75	110,151.75	1,800.00	16,453.60
2023	108,847.00	107,047.00	1,800.00	17,532.40
2024	105,264.75	105,264.75	1,800.00	18,053.20
2025	104,034.75	104,034.75	1,800.00	18,518.20

Probe sweeps across wages and self-employment income, several filing statuses, all 52 state codes, AMT and credit cases, and idtl 0/2/5 found that fiitax − (v28 + v27 + niit − credits − cares) equals addmed for every year 2013–2023 and is 0 from 2024. 2012 has no AddMed. The net investment income tax is in fiitax in every year. The older resources/taxsim35 build (cdate-2025Aug23) still puts AddMed in 2024 fiitax and leaves it out of tfica.

TAXSIM's author:

• Inclusion of the additional medicare in fiitax in TAXSIM #416, 2025-08-20: AddMed "is actually part of FICA not general revenue. I will fix that."
• Does taxsimtest count the Additional Medicare Tax twice (in fiitax and in fica/tfica)? #1225, 2026-09-24, closing it: "it did demonstrate an error that was present in 2000-2023. I believe I have corrected it corrected now. 1) The .038 tax on excess property income is in fiitax 2) The .009 tax on excess earnings income is in fica and tfica".

That is what the emulator does:

• PE-US income_tax includes NIIT through income_tax_before_refundable_credits.
• taxsim_tfica includes AddMed once, through employee_payroll_tax, whose components include additional_medicare_tax. That variable is computed on wages plus self-employment income. self_employment_tax holds only regular SECA.
• The emulator's tfica matches the binary's to the cent in the table above.

Documentation is stale on this point. https://taxsim.nber.org/taxsimtest/, "Output Results" (Last-Modified 2026-09-04):

fiitax = Federal income tax liability including capital gains rates, surtaxes, Maximum Tax, NIIT, AMT, Additional Medicare Tax and refundable and non-refundable credits including CTC, ACTC and EIC etc, but not including self-employment or FICA taxes.

Where the split likely comes from. This is a pointer only; the binary was built from newer, unpublished source. In the amalgamated taxsim.f in public feenberg/taxsim at 359c71f (2026-04-29):

• nlaw (lines 39383–39387) sends in(lawyr,1987,2023) to law87 and later years to law24.
• law87 ends with tax=tax+addmtx (line 39214). law24 has no such line.
• sstax adds addmtx to c(75) and c(190) (lines 40002–40003).

The standalone nlaw.for and sstax.for in the same commit are older and differ. The public code also computes 1,524.60 of AddMed on $400K of wages, where the binary gives 1,800.

Changes

• variable_mappings.yaml: new addmed → additional_medicare_tax output, idtl 2 only. taxsimtest prints an addmed column only at idtl 2.
• core/marginal_rates.py: single-household frate no longer adds AddMed. At $400K wages it was 35.9 and is now 35.0. That matches the runner and the binary's own +$100 difference on the same basis (2024: 105,299.75 − 105,264.75).
• runners/policyengine_runner.py:
  • comments now state the verified basis;
  • mtr_vars is an ordered tuple instead of a set, so the frate/srate output column order no longer changes with hash randomization;
  • the fiitax computation itself is unchanged.
• README: an "Additional Medicare Tax and fiitax" section and an addmed row. The section scopes the convention to PolicyEngine-computed rows: the default CLI's StitchedRunner takes pre-2021 rows from the bundled binary, and those carry the binary's inclusion until it is refreshed.
• Dashboard docs: the fiitax description no longer says it includes AddMed, and addmed is marked implemented.
• output/*.csv: the e2e artifacts are regenerated. They add the addmed column and fix the frate/srate order; no values change.
• tests/test_addmed_excluded_from_fiitax.py, rewritten. The old docstring gave addmed = 2,355.75 for the 2024 $500K case; the binary prints 2,700.00. The tests:
  • cover TX single $400K in 2021, 2023, 2024 and 2025, and CA single $500K and MFJ $1M in 2023 and 2024. Expected fiitax is the binary's fiitax minus its addmed for 2013–2023, and the binary's fiitax as-is from 2024. Each case also asserts addmed and tfica against the binary, and frate against the binary's +$100 difference;
  • check that fiitax at idtl 0 equals idtl 2 and that idtl 0 has no addmed column;
  • check that NIIT stays inside fiitax;
  • check that the single-household path gives the same fiitax, addmed and frate. This test fails on main with frate 35.9;
  • add a binary guard that taxsimtest keeps AddMed out of 2024+ fiitax;
  • add a strict xfail (raises=AssertionError) for the binary's 2013–2023 behavior. It will XPASS, and fail CI on purpose, once resources/taxsimtest is refreshed with the Does taxsimtest count the Additional Medicare Tax twice (in fiitax and in fica/tfica)? #1225 correction. The marker and every bundled-binary note it lists get removed at that point. Open Update bundled macOS and Linux TAXSIM test binaries to the October 27, 2026 build #1206's September builds predate Does taxsimtest count the Additional Medicare Tax twice (in fiitax and in fica/tfica)? #1225, so they should keep xfailing.
• tests/test_performance.py: the fake-sim unit test now runs for 2021, 2023 and 2024.

Testing

• The full pytest tests/ suite passes locally: Python 3.14, PE-US 2.11.3, macOS. The only non-pass is the intended xfail.
• A reviewer agent also ran the AddMed module in a Python 3.10 environment matching CI's 3.10 pins (PE-US 1.783.0, pandas 2.3.3): 14 passed, 1 xfailed.
• CI showed that the bundled Windows binary (cd2026081318) reports addmed 1,524.60 on $400K of wages, where the macOS and Linux builds (cd2026081819) report 1,800.00. That is 0.009 × (0.9235 × 400,000 − 200,000): the self-employment factor applied to wages. The binary guard therefore checks only the convention (from 2024, fiitax == v28 with addmed > 0) and does not pin the amount. The emulator tests pin the macOS/Linux values and don't run the binary. The Windows build still xfails the 2013–2023 check.
• The first CI run's three Python 3.11 cells were cancelled by the matrix's fail-fast after the Windows 3.10 failure.

Not in this PR

• v43/v44 differ from taxsimtest's idtl 5 lines 43 and 44 (NIIT; AddMed alone). The emulator's v44 is employee Medicare tax plus AddMed. Queued as a separate task.
• The emulator appears to compute mstat 6 as joint. Queued as a separate task.
• Three other binary observations are drafted as questions for NBER, not posted:
  • a negative otherprop raises AGI in 2024+;
  • the $200K MFS AddMed threshold;
  • frate = 0 in cd2026081819.
    The uncapped employer OASDI in fica is already Does taxsimtest's fica skip the Social Security wage base for the employer share? #1224.

🤖 Generated with Claude Code

[vercel]

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

Project	Deployment	Actions	Updated
policyengine-taxsim	Ready	Preview	Sep 24, 2026 9:39pm UTC