Net household members' Medicaid MAGI before the zero floor, and give each spouse their own deductions

[MaxGhenis]

Summary

Medicaid MAGI household income for married couples on one return had two interacting bugs:

1. medicaid_magi_person floored each person's MAGI at zero, and medicaid_household_income added the floored amounts. One spouse's loss therefore never offset the other spouse's income, although the couple's MAGI is that of their joint return.
2. medicaid_adjusted_gross_income_person split every above-the-line deduction other than the self-employment ones 50/50 between head and spouse. That included an IRA deduction, student loan interest, and business and capital losses (loss_ald).

They have to be fixed together. With the per-person floor, giving a loss to the spouse who has it makes the floor bite harder. An October 2026 attempt to switch Medicaid AGI to own attribution alone raised medicaid_magi_person by $8.1B and cut Medicaid eligibility by about 137,000 people (hub runner, 2024), so it was backed out.

Example from the brief, a joint couple: spouse A earns $30,000 of wages and spouse B has a −$20,000 Schedule C loss, so joint AGI is $10,000.

	A	B	Household income
main (50/50 split, per-person floor)	20,000	−10,000 → 0	20,000
own attribution, per-person floor	30,000	−20,000 → 0	30,000
this PR	30,000	−20,000	10,000

For a deduction the earner owns (A's $6,000 IRA, B with no income), main gives $27,000 and this PR gives $24,000.

Law and guidance

• 42 CFR 435.603(d)(1): "household income is the sum of the MAGI-based income, as defined in paragraph (e) of this section, of every individual included in the individual's household." It is a plain sum, with no floor for each member. (eCFR)
• 435.603(e): MAGI-based income is "income calculated using the same financial methodologies used to determine modified adjusted gross income as defined in section 36B(d)(2)(B) of the Code."
• 26 USC 36B(d)(2)(A): household income is the taxpayer's MAGI plus the aggregate MAGIs of dependents required to file. (LII)
• Instructions for Form 8962 (2025), line 3: "Add the amounts on lines 2a and 2b. Combine them even if one or both of them are negative. If the total is less than zero, enter -0- on line 3." Line 2a is the couple's joint MAGI and line 2b the dependents' combined MAGI. (IRS, p. 8)
• 435.603(f)(4): each spouse of a married couple living together is in the other's household "regardless of whether they expect to file a joint tax return."
• CMS: the MAGI-Based Household Income Eligibility Training Manual (Dec. 2020) adds the spouses' incomes for joint filers ("Sam scenario", p. 83). I found no CMS guidance that addresses negative income offsetting another member's income. State MAGI manuals treat losses as offsetting other income:

  • Montana CMA 503-1: "Any loss reported on Schedule E, Line 32 can be used to offset any other income and decrease MAGI income."

  • New York GIS 14 MA/07: when "there is other taxable income in the household, the negative income is used to reduce the individual's other taxable income source(s)," and when "the only income source for the household is the self-employment that nets a loss, then the income would be set to zero."

    New York also bars a loss from offsetting nontaxable Social Security. That departs from the 36B worksheet (Form 8962 Worksheet 1-1 adds nontaxable benefits to an AGI that may be negative), which this model follows.

Where the floor goes. The floor goes on the household's total, the same as Form 8962 line 3: household income = max(0, J + D), where J is the couple's MAGI and D is the MAGI of dependents required to file. Applied to each path:

• Joint filers: the couple's combined MAGI is netted with counted dependents' MAGI, then floored.
• Non-filers (435.603(f)(3)): with co-resident parent IDs, the merged candidate sums the signed MAGI of the same linked members counted by medicaid_household_size, then floors the total. Medicaid MAGI household composition from co-resident parent ids (fixes #9404) #9406's membership excludes relatives outside that household, so linked parents', children's and siblings' negative amounts can offset its other counted income. Without parent IDs, the existing conservative fallback remains: an adult's own and their spouse's MAGI enter signed; family-wide relatives enter only when positive, each child's amount separately and married parents' amounts netted together. This includes a non-filer child's own amount among the family-wide children. Unmarried parents' losses and children's or siblings' negative amounts do not offset income in that fallback.
• Cohabiting spouses filing separately: each spouse's household includes the other (f)(4), so their two returns' MAGIs are summed, then floored. A dependent of either spouse is in that spouse's household (f)(2), so the dependent's household gets the other spouse's MAGI too. Household size already counted that spouse; before this PR, household income left them out.
• Dependents required to file: their own MAGI, possibly negative, enters the claimant's household before the floor, as on Form 8962 line 2b.

The brief's property, max(0, J) + D, agrees with this whenever J ≥ 0 and no counted dependent's MAGI is negative. The two differ only when a negative joint MAGI meets a dependent's positive income, or a dependent's MAGI is negative. Form 8962 line 3 settles that case by combining before the floor.

The fix

• medicaid_adjusted_gross_income_person = medicaid_irs_gross_income − above_the_line_deductions_person.
  • Each spouse's own IRA deduction, student loan interest, alimony and business and capital losses now lower only their own amount, and the head's and spouse's amounts add up to the return's AGI.
  • A tax unit dependent's amount is their own income less their own person-level deductions. That includes alimony paid, which main already attributed to its payer. A dependent's own business and capital losses aren't modeled, as before: a dependent's gross income counts only positive amounts, and above_the_line_deductions_person gives dependents no loss deduction.
  • Taxable basic income (UBI Center reform) is divided equally between head and spouse with filer_share.
• medicaid_magi_person: no floor. Person-level MAGI additions (tax-exempt interest) are the person's own. Additions recorded only for the tax unit (foreign earned income exclusion, nontaxable Social Security) are divided equally between head and spouse with filer_share, as before for joint returns.
• medicaid_household_income: members' amounts are summed signed across the tax household, the claimant household, cohabiting separate spouses and the linked non-filer household. The merge retains Medicaid MAGI household composition from co-resident parent ids (fixes #9404) #9406's claimant and spouse membership helpers. Only the household total is floored at zero. The fallback for non-filers without parent IDs retains the positive family-wide relative amounts described above.
• HSA integration from main: health_savings_account_ald_person remains attributed to its owner by the existing allocator. When only a tax-unit HSA deduction is supplied, the existing equal filer fallback remains. The refresh changes no HSA formula: it removes stale equal-only metadata and documentation and corrects the independent spouse-deduction property reference for its tax-unit-only HSA fixture.
• medicaid_person_is_required_to_file: unearned income was gross income less earned_income. Gross income counts each source's positive amount, so a self-employment loss made other income look unearned: a dependent with a $5,000 loss and $100 of interest had $5,100 of "unearned income" and was required to file. With members no longer floored, that would have counted the dependent's negative income. Unearned income is now gross income less the positive wage and self-employment sources it contains (Pub. 501 (2025), Table 2). This also stops 401(k) deferrals, which box 1 wages leave out, from hiding unearned income.
• above_the_line_deductions_person: amounts in gov.irs.ald.filer_amounts_recorded_on_dependents (education savings bond interest, employer adoption assistance) are the filer's even when recorded on a dependent. They no longer reduce the dependent's own amount, which Medicaid now reads. Main's Medicaid formula already left them off dependents.

Dependency

This PR remains stacked on #9890 (each spouse's own above-the-line deductions in per-person AGI), which is open at eaabd725476ecb797eab2df217efaaf2552f79ac. That PR adds above_the_line_deductions_person, loss_ald_person, limited_capital_loss_person, limited_business_loss, alimony_expense_ald_person and filer_share. This branch retains its stacked history. Merge #9890 first. The filer_amounts_recorded_on_dependents fix to above_the_line_deductions_person described above is part of #9890.

Membership merge

#9406 (Medicaid MAGI households from co-resident parent IDs) merged in f33a4b3b1666b07f99da83ea771a0819a416c571. This branch merges main with a history-preserving merge and keeps its claimant, co-resident parent and spouse membership together with this PR's signed income sums and one household zero floor. Members' MAGI has no per-person floor.

Previously approved candidate head and release merge: 052420824c4b62bc4ff14c59607b354982de9515, with parents ad2d76c1b0de997a748be051d8c801ef618d5eb9 and current main e44d7c367327108858e06a22160b0ab5d7189c41. Earlier main merge 79ae598674c711d3d9961e698d4984a45e0ef04c retains main 7b956eb4014e3a189075c79f9b1f0a8f1d3b02ab; dependency merge 6b81b743c8b8238ecc8523c1cb936f0540af0a17 is preserved in that history.

Main merge (resolved conflict with #9857)

Merged main at 0a8c3c2c43e12bef720aba38dbb33b2acdada1fb into approved head 052420824c4b62bc4ff14c59607b354982de9515 with merge commit 6c5f5cddd8ad7cb43c21cd84ff281d382e9ecee7. The D.C. resolution preserves #9857's filer-only losses and proportional split, with #9860's business-first limited_business_loss cap and signed Medicaid household income netting. An additional Ohio test conflict retains both independent regression cases and their original expectations.

Targeted verification: 75 passed across 10 files, each run separately and sequentially in the foreground; make format passed. Runs used the supplied Python 3.13.9 environment (Core 3.32.15, pytest 9.1.1, Hypothesis 6.168.5), with a verified untracked raw-source cache after the first file. The committed tree is the exact tested tree. No partner expectations were edited. The existing hub impact section below was measured before this main merge; no new microsimulation was run. The PR remains draft for the hub's delta review.

Invariants

The existing generator draws joint filers with dependent children aged 5–17, non-filing couples, cohabiting spouses filing separately, and single filers without children, each household its own family. The merge adds a focused six-person linked non-filer property described below; YAML cases cover additional paths:

1. Joint filers: every member's household income = max(0, J + D). J is the couple's MAGI, figured independently from the tax unit's adjusted_gross_income plus the MAGI additions. D is the sum of the MAGI of children required to file. When J ≥ 0 and no counted child's MAGI is negative, this equals max(0, J) + D.
2. Accounting: the head's and spouse's medicaid_magi_person add up to J (differential against the tax-unit AGI path). A dependent's Medicaid AGI is their own gross income less their own deductions.
3. Couples outside a joint return: for non-filing couples and cohabiting spouses filing separately, household income = max(0, the two spouses' MAGI added). For single filers without dependents, max(0, their MAGI).
4. Floor: household income is never negative, equals max(0, the unfloored sum), and is zero whenever the unfloored sum is not positive.
5. Metamorphic:
   • Removing the spouse's own IRA deduction, early withdrawal penalty and alimony leaves the head's deductions unchanged. Student loan interest is excluded because its phase-out reads joint MAGI.
   • Swapping the spouses' incomes and deductions leaves every member's household income unchanged.
   • Lowering a spouse's self-employment income never raises household income.

These invariants are checked in policyengine_us/tests/core/test_medicaid_household_income_invariants.py, using 20 drawn batches of 10–40 households plus a seeded population of 200. The merge adds a property drawing signed MAGI for six co-resident people: two unmarried parents, their two children, a grandmother and an unrelated adult. An independent membership matrix checks both household size and max(0, signed member sum), including exclusion of relatives outside each person's household. Children's required-to-file flags are inputs, so this property isolates membership and netting. The fixture supplies person and parent IDs as dictionaries for the tested year (2025).

Tests

The list combines original PR coverage with the two HSA integration cases. Earlier run results below are historical; fresh commands and results for this candidate are recorded at the end of this section. YAML expectations are hand-computed:

• medicaid_household_income.yaml (new, 13 cases):

  • three CMS training-manual scenarios (Sam, Marty, Sadie/Stephanie), unchanged by this PR;
  • a joint loss;
  • one spouse's IRA deduction;
  • a negative joint MAGI with a dependent required to file (Form 8962 line 3);
  • a dependent with negative MAGI;
  • cohabiting spouses filing separately;
  • non-filer spouses;
  • a negative total floored to zero;
  • a grandparent outside a child's non-filer household (from review r1);
  • an adult dependent of a filer whose cohabiting spouse files separately (from review r1);
  • married non-filer parents whose loss offsets their wages in their child's household (from review r2).

  The claimant path with a cohabiting separate spouse is already covered by Case 11 of medicaid_uses_non_filer_rules.yaml.

• medicaid_magi_person.yaml (new, 2 cases).

• medicaid_adjusted_gross_income_person.yaml (10 cases in the candidate):

  • original PR additions: spouse-only deductions, a spouse's negative AGI, and savings bond interest recorded on a dependent;
  • new HSA integration case: a $4,000 supplied spouse deduction gives own Medicaid AGIs of $60,000 and $36,000;
  • new HSA fallback case: a $4,000 tax-unit deduction with no person amounts gives $58,000 and $38,000.

  Both HSA cases reconcile to $96,000 joint AGI. The HSA refresh preserves the preceding eight cases' outputs.

• medicaid_person_is_required_to_file.yaml (+2 cases): a self-employment loss; a 401(k) deferral.

Historical differential check before this merge refresh. The original PR ran the new and changed files against three checkouts:

Checkout	Failing cases
main, bab1a47	14
#9890 alone	14
original PR branch	0

The grandparent case passes on main and on #9890 alone, as it should: it guards against a leak that signed sums would otherwise open. The original four invariant test cases failed on #9890 alone. For example, the brief's loss case gives $20,000 there.

Historical regression check before this merge refresh. Every YAML file under gov/hhs/medicaid/income/ passed on the original PR branch, as did is_older_child_for_medicaid, is_parent_for_medicaid, the CA CCHIP integration test, and the above-the-line deduction files. Those runs used single files and left full suites to CI.

Earlier merge verification (before the main merge with #9857): All 10 targeted files passed (120 cases), each run sequentially in the foreground. Runs used committed head ad2d76c1b0de997a748be051d8c801ef618d5eb9; its tested policyengine_us subtree is be1b6e89292e4880a3916f93836b7dc917e035ee. Final release merge 052420824c4b62bc4ff14c59607b354982de9515 preserves that exact subtree. The untracked source cache verified tracked source bytes and directory order; Hypothesis was the locked 6.168.3 version. Commands ran from .hubrepo with:

Each file ran on its own, sequentially in the foreground, against the repository's locked test dependencies.

Command	File	Passed	Options
pytest	policyengine_us/tests/core/test_medicaid_household_income_invariants.py	5	--durations=0
pytest	policyengine_us/tests/core/test_spouse_above_the_line_deductions_invariants.py	7	--maxfail=1
pytest	policyengine_us/tests/core/test_parent_links_properties.py	11	None
YAML	policyengine_us/tests/policy/baseline/gov/hhs/medicaid/income/parent_ids_composition.yaml	47	None
YAML	policyengine_us/tests/policy/baseline/gov/hhs/medicaid/income/medicaid_household_income.yaml	13	None
YAML	policyengine_us/tests/policy/baseline/gov/hhs/medicaid/income/medicaid_adjusted_gross_income_person.yaml	10	None
YAML	policyengine_us/tests/policy/baseline/gov/hhs/medicaid/income/medicaid_magi_person.yaml	2	None
YAML	policyengine_us/tests/policy/baseline/gov/hhs/medicaid/income/medicaid_person_is_required_to_file.yaml	8	None
YAML	policyengine_us/tests/policy/baseline/gov/states/dc/tax/income/dc_self_employment_loss_addition.yaml	10	None
YAML	policyengine_us/tests/policy/baseline/gov/states/dc/tax/income/dc_separate_capital_loss_adjustment.yaml	7	None

No partner contract expectations were edited. Full suites and microsimulation were not run for this merge refresh. Independent read-only source review: APPROVE for exact final local head 052420824c4b62bc4ff14c59607b354982de9515, with no unresolved P1/P2 findings. Approval covers the local source head only. The PR remains draft; current green CI and review of the current published head are required before readiness.

Methodology ruling (Max, d1120, 2026-10-09): approved as built. A non-filer Medicaid household sums its members' signed MAGI and floors only the household total at zero (42 CFR 435.603(d)(1); Form 8962 line 3). Unlinked relatives' income counts only when positive.

Hub microsimulation impact (current head)

A real microsimulation on the default dataset, main at the merge base 0a8c3c2c43 against this PR's head 6c5f5cddd8, for 2026. The runs went one at a time under the shared heavy-job lock. Sums are weighted.

Output	2026 change	2026 records changed (weighted)
medicaid	+2,560	5,273 (18,057,956)
is_medicaid_eligible	+194,970	73 (194,970)
is_chip_eligible	-19,449	11 (19,449)
aca_ptc	-342.79M	10 (23,336)
household_health_benefits	+0	0 (0)
household_net_income_including_health_benefits	-7.73M	3,652 (11,062,094)
state_income_tax	-909,087	340 (297,226)
income_tax	+27.46M	857 (2,200,173)
household_net_income	-20.74M	1,106 (2,310,990)
household_benefits	+5.81M	8 (5,668)

This run is at the merged head. It matches the pre-merge run at 052420824c to within $560 and one record in every row, so the #9857 merge doesn't change the PR's population effect. Medicaid spending barely moves because each state's total is a calibrated amount: more enrollees lower the amount per enrollee, not the total (see Impact below).

Comparison files are in fixes/9860-impact/ on the hub host.

Impact

Historical microsimulation, kept for the mechanism. The current-head estimate is the hub table above. The table below was measured on the default dataset with the hub runner (lockrun.sh, run.py, compare.py), comparing #9890's then-head 36058bac8d with this PR's then-head d4b561e7a1. It does not measure the current candidate against current main or current #9890. In those earlier runs, #9890 changed no Medicaid, CHIP or ACA variable in either year against its then-main baseline. One run per checkout per year. Dollar changes are weighted totals; "records" counts changed microdata records. The 2024 figures were identical, to the dollar, to the earlier pair of runs (986d04658a and 1681b9b487). No new microsimulation was run for this merge refresh.

Variable	2024 change	2024 records	2026 change	2026 records
Medicaid eligible (people)	+83,335	75 (all up)	+224,472	81 (all up)
Medicaid enrolled (people)	+66,228	66 (all up)	+180,844	70 (all up)
CHIP eligible and enrolled (people)	−31,590	17	−40,096	14
CHIP spending	−$108M	17	−$127M	14
Medicaid spending	−$2,318 (none, in effect)	5,390	+$4,539 (none, in effect)	7,077
ACA PTC	−$17.8M	12	−$343M	12
ACA PTC eligible (people)	−1,551	22	−59,622	22
medicaid_magi_person (sum)	−$38.2B	31,557	−$34.1B	31,586
medicaid_household_income (sum over people)	−$27.7B	18,490 (5,297 up, 13,193 down)	−$34.4B	18,613 (5,970 up, 12,643 down)
medicaid_person_is_required_to_file (people)	−56,588	48	−58,988	44
Federal income_tax, state income taxes, AGI	none	0	none	0
household_net_income	+$4.75M	5	+$10.79M	9

Medicaid spending stays the same because each state's total is a calibrated amount (calibration.gov.hhs.medicaid.totals.spending). medicaid_cost_if_enrolled spreads it over that state's enrollees, weighted by cost index (medicaid_slcsp_state_denominator). More enrollees lower the amount per enrollee, not the total. medicaid_magi_person falls because members' negative MAGI is no longer floored. Two changes in this PR can also raise household income: dependents now count the claimant's separate spouse, and the filing-requirement fix can newly count a dependent's positive income. The increases were not decomposed between them.

In the historical comparison, the brief's attempt to switch Medicaid AGI to each spouse's own deductions while keeping the per-person floor cut eligibility by about 137,000. The historical combined fix instead raised eligibility. These figures do not establish the current merge's impact.

axiom: TheAxiomFoundation/rulespec-us#1507 queued (42 CFR 435.603(d)(1): us/policies/medicaid/magi_household_income_pipeline.yaml already sums members' MAGI with no per-member floor; the queued companion tests pin a spouse's negative MAGI offsetting the household)

🤖 Generated with Claude Code

[MaxGhenis]

Context for the hub from the original session, which is handing this PR over and will not push to it.

• The Medicaid MAGI household composition from co-resident parent ids (fixes #9404) #9406 resolution here matches mine. Before the hub adopted the PR, I resolved the same merge locally and never pushed it. My medicaid_household_income.py formula is identical, line for line, to the one at 6c5f5cd; only one comment differs. It is Medicaid MAGI household composition from co-resident parent ids (fixes #9404) #9406's parent-link membership (medicaid_non_filer_member_sum, medicaid_filer_spouse_sum, medicaid_tax_dependent_spouse_sum, the head's separate-spouse propagation) fed this PR's signed member income, with one floor on the household's total.
• Test evidence for that resolution, run 10/07 on Give each spouse their own above-the-line deductions in per-person AGI #9890 eaabd72 merged with main 07227ea (includes Medicaid MAGI household composition from co-resident parent ids (fixes #9404) #9406 and Cap the educator expense deduction per eligible educator (26 U.S.C. 62(a)(2)(D)) #9833), under heavy_run.sh:
  • seam YAML (educator, loss_ald_person, own deductions, DC, parent_ids.yaml): 54 passed;
  • gov/hhs/medicaid: 5 batches passed; gov/hhs/chip and states/ca/dhcs/medi_cal: passed;
  • property tests: test_parent_links.py, test_parent_links_properties.py, test_medicaid_household_income_invariants.py, the dependent, educator and determinism files: passed.
  • The only failure was Give each spouse their own above-the-line deductions in per-person AGI #9890's spouse property test, all 7 tests. Main's Subtract Mississippi's HSA deduction once, from the account holder #9865 added health_savings_account_ald_person, which Give each spouse their own above-the-line deductions in per-person AGI #9890 then still listed as equally divided. The hub's Give each spouse their own above-the-line deductions in per-person AGI #9890 head handles that now.
• The impact tables in this body (2024/2026) predate Medicaid MAGI household composition from co-resident parent ids (fixes #9404) #9406. They compare Give each spouse their own above-the-line deductions in per-person AGI #9890 36058ba with this branch at d4b561e. Medicaid MAGI household composition from co-resident parent ids (fixes #9404) #9406 changes Medicaid household membership, so they need a rerun at the current heads.

🤖 Generated with Claude Code