Skip to content

Model the Medicare Part D low-income subsidy as a CBO means-tested transfer - #10127

Open
MaxGhenis wants to merge 47 commits into
mainfrom
medicare-part-d-lis
Open

MaxGhenis wants to merge 47 commits into
mainfrom
medicare-part-d-lis

Conversation

@MaxGhenis

@MaxGhenis MaxGhenis commented Oct 11, 2026 •

Copy link
Copy Markdown
Contributor

This PR builds on #10118 (#10111 merged on 2026-11-13, #10107 on 2026-10-12). It merges #10118's head b867d6d and main as of 94054aa, and it targets main so that CI tests the combined tree: CI runs only on pull requests to main (.github/workflows/pr.yaml). Until #10118 merges, the diff here includes #10118's changes; this PR's own work is the commits after b867d6d. Once #10118 squash-merges, this branch merges main again before it lands.

Summary

#10118 makes the default CBO household income count Medicare the way CBO does: net of premiums, and without the Part D low-income subsidy (LIS, "Extra Help"). CBO counts the LIS as a means-tested transfer instead. pe-us had no LIS variable, so after #10118, CBO income after transfers leaves out the LIS: about $18 billion in 2026, which #10118 takes out of Medicare at $255.86 per enrollee.

This PR models the LIS and counts it in two places:

  • gov.household.cbo_means_tested_transfers, CBO's means-tested transfers;
  • the household health benefit lists, gov.household.household_health_benefits and healthcare_benefit_value, beside medicaid_cost and msp_cost.

The model pieces:

  • part_d_lis_eligible covers Medicare enrollees in the 50 states and DC whom the law deems eligible or who pass the income and resource tests. Medicaid enrollees, SSI recipients and people eligible for a Medicare Savings Program count as deemed.
  • part_d_lis_enrolled pays only people in a Part D plan (medicare_part_d_enrolled, from Charge the Part D income-related premium only to Part D enrollees #10115). In a plan, Medicaid enrollees and SSI recipients get the subsidy automatically; other eligible people need takes_up_part_d_lis_if_eligible.
  • part_d_lis gives each enrollee the Medicare Trustees' average subsidy per LIS enrollee, CBO's measure: $1,254 in 2026.
  • part_d_lis_countable_rental_property_equity: rental property equity less the exclusion for income-producing property (see Real estate).
  • Ten parameters under gov.hhs.medicare.part_d.low_income_subsidy, one calibration series (2013–2035), BLS's September 2026 CPI-U, a projection of the resource limits after 2026 in uprating_extensions.py, a programs.yaml entry, a take-up flag and a source-conflict candidate.

What changes:

  • household_net_income and household_benefits don't change. The health benefit lists reach them only when gov.simulation.include_health_benefits_in_net_income is on, and it's off by default.
  • healthcare_benefit_value, and so household_net_income_including_health_benefits, rise by the subsidy.
  • Neither list holds medicare_cost, whose spending per enrollee already includes all Part D spending, the LIS among it. So nothing counts twice.

In the 2026 microsimulation, 9.89 million people receive $12.40 billion. CBO income after transfers and taxes and healthcare_benefit_value each rise by that amount (see Impact).

Why an average, not a computed benefit

Most of the LIS is the cost-sharing subsidy. That depends on each enrollee's drug use, and pe-us has no Part D claims. The premium subsidy depends on the region's benchmark premium and the plan chosen. CBO values the LIS the same way for everyone:

  • Habib 2018, CBO Working Paper 2018-07, file page 46: "CBO removes LIS benefits from its total Medicare benefits and assigns them separately. ... A pool of eligible recipients is created by identifying Medicare recipients who also receive at least one means-tested transfer ... Recipients are chosen at random from that pool until the total LIS population is accounted for, and an average LIS benefit is assigned to each imputed recipient."
  • CBO, January 2026: "LIS benefits are assigned to some Medicare recipients who, CBO estimates, also receive other means-tested transfers." CBO allocates insurance at "the average resource cost, which is total spending divided by the number of recipients."

This PR keeps CBO's average but takes recipients from the law, not from CBO's pool. CBO draws recipients until their count matches the administrative total. The model pays every eligible person in a Part D plan who is on Medicaid or SSI or takes the subsidy up, so its count comes from the data (see Impact).

The Trustees publish the average directly: Table IV.B9's "Subsidy amount" per LIS enrollee. The series takes 2016–2035 from the 2026 report (file page 158), 2015 from the 2025 report and 2013–2014 from the 2023 report. Enrollment times the amount reproduces Table IV.B10's payments, for example 14.4 million × $1,254 = $18.1 billion in 2026.

Institutionalized enrollees. CBO removes institutionalized LIS recipients and their benefits (Habib 2018, file page 52). Its adjusted totals (file page 51) are:

Year CBO recipients CBO benefits CBO per recipient Trustees IV.B9 per enrollee
2015 10.7 million $22.6 billion $2,112 $2,112
2016 11.1 million $26.5 billion $2,387 $2,126

CBO's 2016 benefits come from its own baseline. Its recipient counts are 88–90% of the Trustees'. So the adjustment shows up mostly as fewer recipients, which the model's noninstitutional microdata already reflects.

Eligibility

Rule Source Code
Medicare enrollee, enrolled in a PDP or MA-PD 42 U.S.C. 1395w-114(a)(3)(A)(i); 42 CFR 423.773(a); 20 CFR 418.3101(b) defined_for = "medicare_enrolled"; part_d_lis_enrolled requires medicare_part_d_enrolled
Income below 150% of the poverty guideline for the family size (a)(3)(A)(ii); 423.773(a)(1); POMS HI 03020.055 part_d_lis_income_eligible
Family size: the person, a spouse living with them, and relatives who live with them and depend on them for at least half their support 42 CFR 423.772 part_d_lis_family_size: the marital unit plus the dependents on the couple's returns, on both returns when the spouses file separately
Income is SSI income, without in-kind support and maintenance, of the person and of a spouse living with them "regardless of whether the spouse is also an applicant"; all interest and dividends excluded; the $20, $65 and one-half exclusions apply once to the couple (a)(3)(C)(i); 423.772; 20 CFR 418.3310, 418.3325, 418.3350(d); POMS HI 03020.050 part_d_lis_countable_income: SSI's unearned sources less income/excluded_sources.yaml, then _apply_ssi_exclusions on the couple's totals
Rental income counts net, and expenses above the rent don't offset other income POMS SI 00830.505 each unearned source counts at no less than zero, for each person
Resources of the person and spouse ≤ the (E) standard: $10,000 / $20,000 in 2006, CPI-indexed; $16,590 / $33,120 in 2026 and $17,210 / $34,360 in 2027 (a)(3)(E); 423.773(b)(2)(iii), (d)(2); POMS HI 03030.025 limit/individual.yaml, limit/couple.yaml; extend_part_d_lis_resource_limits after 2026
Plus $1,500 per person for burial funds, unless the applicant says they won't use any for burial 20 CFR 418.3425(j); POMS HI 03030.020, .025 burial_exclusion.yaml
Countable resources: liquid resources and real estate other than the home; life insurance excluded; each amount at no less than zero (a)(3)(G); 423.772; 20 CFR 418.3405 sources.yaml (bank, stock and bond assets) and household_sources.yaml (household_other_real_estate_equity and part_d_lis_countable_rental_property_equity, for the household head's couple)
Income-producing property: up to $6,000 of equity excluded if its net annual income is at least 6% of the excluded equity; none if the return falls short 42 U.S.C. 1395w-114(a)(3)(D)–(E) → 1382b(a)(3); 20 CFR 416.1222(a); POMS HI 03030.020 A, SI 01130.503 part_d_lis_countable_rental_property_equity; rental_property/equity_exclusion.yaml, rental_property/minimum_return.yaml
Deemed: full-benefit dual eligibles, SSI recipients, QMB/SLMB/QI (a)(3)(B)(v); 423.773(c)(1) part_d_lis_deemed_eligible
50 states and DC only (a)(3)(F); 20 CFR 418.3010(b)(11) ineligible_jurisdictions.yaml

Before 2024, income from 135% to 150% of the guideline, or resources between the lower (D) and the higher (E) limit, gave a partial subsidy. The Inflation Reduction Act (§ 11404) made all of it a full subsidy from 2024. The model uses the outer bound in every year, because CBO and this PR give every recipient the same average.

The LIS income test takes SSI's exclusions "without regard to the application of section 1396a(r)(2)" ((a)(3)(C)(i)). So part_d_lis_countable_income keeps SSI's $20 general exclusion even in Illinois and Mississippi, whose Medicare Savings Program income takes larger 1902(r)(2) exclusions.

Rental losses. SSI counts net rental income, and POMS SI 00830.505 says not to "use [excess expenses] to offset other income". part_d_lis_countable_income therefore counts each unearned source at no less than zero, for each spouse. pe-us's own SSI and Medicare Savings Program income sums the sources with their signs, so a rental loss there still offsets Social Security; a separate task covers those.

Real estate. 423.772 counts "real estate that is not the applicant's primary residence", and 20 CFR 418.3405(b) counts it whether or not it can be sold within 20 workdays. pe-us holds real estate equity by household, and a household amount doesn't say whose it is. So part_d_lis_resource_eligible counts the household amounts for the household head's couple (the oldest member's, when no one is head), once per household:

  • household_other_real_estate_equity in full;
  • rental property equity through part_d_lis_countable_rental_property_equity. SSA applies the SSI exclusions to Extra Help (POMS HI 03030.020 A), including the one for nonbusiness income-producing property (SI 01130.503, 20 CFR 416.1222(a)): up to $6,000 of equity is excluded when the property's net annual income is at least 6% of the excluded equity, and none when it falls short. SI 01130.503's first example is a rented former home with $10,000 of equity, of which $4,000 counts. The model compares the household's net rental income (rental_income, summed over members) with 6% of the excludable equity.

What the model can't see:

  • a rental used in a trade or business, which SI 01130.501 excludes regardless of value. The model treats every rental as nonbusiness property, which SI 01130.503 tells SSA to check on Schedule E;
  • separate properties, each of which faces the 6% test on its own before their equity is totaled against the $6,000. The model treats the household's rentals as one property;
  • a shortfall in the return for reasons beyond the owner's control, which still qualifies (SI 01130.503 A.3).

Each amount, the household's and each person's, counts at no less than zero, so property worth less than its debt or an overdrawn account doesn't offset other resources.

The default dataset holds neither real estate input, and no negative bank, stock or bond amount (the person and household tables of populace_us_2024.h5@populace-us-2024-spm-20260909, snapshot 9a814a3b, checked 2026-11-14). So in the microsimulation these rules change nothing.

Part D plan. The statute's subsidy eligible individual "is enrolled in a prescription drug plan or MA–PD plan" ((a)(3)(A)(i)). CMS's rule reads "enrolled in, or seeking to enroll in a Part D plan" (423.773(a)), and SSA's rule tells people found eligible before they enroll that they "must enroll in one of these plans to actually receive a subsidy" (418.3101(b)). The model can't see who seeks to enroll, so part_d_lis_eligible holds the deemed, income and resource tests, and part_d_lis_enrolled adds the plan: only people with medicare_part_d_enrolled (from #10115) receive the subsidy.

Enrollment. CMS tells deemed people they don't need to apply (423.773(c)(2)). It enrolls subsidy-eligible people who don't choose a plan, and any of them can decline (42 CFR 423.34(d)(1), (e)). Qualifying covered retirees, whose employer retiree drug coverage qualifies for the retiree drug subsidy, are the exception: CMS enrolls only the full-benefit dual eligibles among them whose plan sponsor it approved for that subsidy, and only if they confirm that they want a plan (423.34(d)(3)). So:

  • part_d_lis_enrolled is automatic for medicaid_enrolled or ssi > 0 people in a Part D plan;
  • other eligible people in a plan need takes_up_part_d_lis_if_eligible, which defaults to True. pe-us models Medicare Savings Program eligibility, not enrollment, so MSP-deemed people go through the flag too;
  • microdata that set takes_up_medicare_part_d_if_eligible should set it True for Medicaid enrollees and SSI recipients other than those who decline and covered retirees who don't choose a plan. A person whose Part D input is False gets no subsidy. part_d_lis_enrolled's documentation states this contract.

Microcosm sets neither takes_up_medicare_part_d_if_eligible nor takes_up_part_d_lis_if_eligible (git grep on microcosm origin/main 3db65154b, 2026-11-07); its only mention of the first is a comment. Both default to True. In the microsimulation, therefore, every Medicare enrollee holds a Part D plan, and every eligible one receives the subsidy. microcosm#1206 proposes a Part D take-up input.

Values

Resource limits through 2026.

  • Individual: all 21 years, 2006–2026, match POMS HI 03030.025, which before 2024 lists two limits a year; the parameter takes the higher one. Each equals the statute's rule applied to the previous year's limit: × the September-to-September change in the not seasonally adjusted CPI-U (BLS CUUR0000SA0), never lowered, rounded to $10.
  • Couple, 2006–2014: SSA's amounts, which the same rule reproduces.
  • Couple, from 2015: the statute's rule, not SSA's amounts. For 2015, CMS's memo applies the factor 1.01658 (238.031 / 234.149) to 2014's $23,860, which gives $24,255.58 and rounds to $24,260. The memo prints $24,250, and POMS lists $24,250 for 2015 and 2016. SSA has raised its own published amount every year since. The parameter applies the statute's rule to the previous year's limit as the statute sets it. So it runs $10 above SSA's amounts in 2015 and 2016 and $20 above from 2017: $33,120 against SSA's $33,100 in 2026.
  • The ruling: source-conflict candidate us:part-d-lis-couple-resource-limit-2015 (source_conflicts/candidates.yaml), which follows the statute under the statute-first rule.
    • The memo's factor reproduces its other three 2015 amounts ($7,280, $10,930 and $12,140).
    • Every other limit SSA published from 2007 to 2026 equals the rule applied to the previous published limit. That covers individuals and couples, at the lower (full subsidy) and the higher level: 74 year-limit pairs, one mismatch (~/reviews/medicare-part-d-lis-2026-10/scripts/verify_published_chain.py, from POMS HI 03030.025's prior-year list and BLS's Septembers).
    • So $24,250 is the one amount the stated method doesn't give. The gap reaches only couples whose countable resources fall within $10 to $20 of the limit.
    • The rulespec-us register entry follows after merge (CONTRIBUTING.md, "Source conflicts", steps 2–4).

Resource limits after 2026. extend_part_d_lis_resource_limits (uprating_extensions.py) projects each later year:

  • the previous year's limit × the unrounded ratio of September of the year before to September two years before, never lowered, rounded to the nearest $10 with halves up;
  • get_september_cpi_u_values takes both Septembers from gov.bls.cpi.cpi_u_nsa_september when BLS has published both, and otherwise both from the model's CPI-U series, never one of each. Since main's change to that series, it is the not seasonally adjusted CPI-U (BLS through October 2026), then CBO's annual averages, which have no seasonal pattern;
  • the PR adds BLS's September 2026 value, 336.940 (API, retrieved 2026-11-08). For 2027 that gives $16,590 × 336.940 / 324.800 = $17,210.08, so $17,210, and $34,360 for couples. CMS's CY 2027 memo (November 5, 2026) prints $17,210 and $34,340; its couple amount builds on SSA's $33,100, so it stays $20 below the model's, as the source-conflict candidate describes;
  • CMS's CY 2026 memo reports the 2026 increase as 3.01%, but its published $16,590 needs the unrounded ratio: $16,100 × 1.0301 = $16,584.61 would round to $16,580;
  • the arithmetic is exact (Python fractions of the decimals as printed), with halves rounding up. Round 3 found that a finite-precision decimal ratio puts some exact $5 products a hair below $5: $16,590 × 332.3 / 331.8 = $16,615 exactly, which rounds to $16,620, came to $16,614.99… and so $16,610;
  • the limits hold from 2100, END_YEAR.

Medicaid home equity, as a side effect. get_cms_cpi_u_percentage_increase now reads its Septembers through the same helper. With September 2026 published, CMS's rounded percentage for the 2027 Medicaid long-term care home equity limits becomes 3.7% (336.940 / 324.800) instead of 3.1% from the projected series. The 2027 limit rises from $1,165,000 to $1,172,000.

The helper also reads only September instants now, and a year needs both of its Septembers published to use them. Before the merge with main, pe-us backdated cpi_u_nsa_september's first value to 2015-01-01, which the old code counted as a published 2015 September. main now adds September 1988 (119.8), so the series runs 1988, then 2018 on, and no backdated entry exists; the check still keeps a gap year from reading the September before it.

Per recipient. IV.B9, 2013–2035:

Year LIS enrollees (IV.B9) Per recipient Per Medicare enrollee (#10118)
2013 11.5 million $2,023 $441.87
2014 11.8 million $2,052 $449.04
2015 12.1 million $2,112 $460.52
2020 13.1 million $2,507 $524.75
2024 14.5 million $2,833 $605.29
2025 14.2 million $1,226 $251.12
2026 14.4 million $1,254 $255.86
2030 15.8 million $1,833 $377.29
2035 16.7 million $2,076 $431.59
  • The amount falls from $2,833 in 2024 to $1,226 in 2025. The 2026 report says the Inflation Reduction Act redesigned the standard Part D benefit, "with the full effects of the benefit redesign implemented in 2025" (file page 119). Table IV.B9 gives no cause for the drop.
  • From 2036 the series grows with calibration.gov.hhs.medicare.per_capita_cost, like the per-enrollee series beside it.
  • Like Count Medicare net of premiums in CBO household income #10118's series, it starts in 2013. CBO means-tested transfers already fail before 2015 on the base, through Medicaid and SNAP parameters. So the CBO test file gets no 2013–2014 case, and programs.yaml gives 2015 as the first verified year.

Impact

2026, the default dataset (populace_us_2024.h5@populace-us-2024-spm-20260909: 57,240 households, 166,321 people), one full run per side through the shared heavy lock. Base: this branch's merge-base with #10118's branch, c9dfbc2 (#10118's head then). PR: fe4af16 (the later commits through 2c447a3 change only documentation and tests). Each run imported its own git archive snapshot. Both sides predate the later merges of #10118's head and of main (through 94054aa). Those bring #10837, which changes SSI's general exclusion under spousal deeming and so can move who receives SSI, and with it the subsidy automatically (0.17 million recipients come by SSI alone here). This PR's own code has changed since fe4af16 only by round 4's fixes, which move nothing on this dataset.

Base This PR Change
CBO income before transfers and taxes $22,718.2 billion $22,718.2 billion 0
cbo_medicare_benefit $929.6 billion $929.6 billion 0
medicare_cost $928.1 billion $928.1 billion 0
CBO income after transfers and taxes $19,497.8 billion $19,510.2 billion +$12.40 billion
healthcare_benefit_value $967.2 billion $979.6 billion +$12.40 billion
household_net_income $17,162.3 billion $17,162.3 billion 0
part_d_lis $12.40 billion
part_d_lis_enrolled 9.89 million
  • The changes in CBO income after transfers and taxes and in healthcare_benefit_value equal total part_d_lis to within $220 and $112, float rounding on $12.4 billion.
  • 7.44 million of 124.56 million households gain, and none lose.
  • Every recipient gets $1,254, and no one outside part_d_lis_enrolled gets anything.

Recipients by path:

Path Recipients
Automatic: Medicaid enrollees 5.69 million
Automatic: SSI recipients without Medicaid 0.17 million
Medicare Savings Program eligible, neither Medicaid nor SSI 2.84 million
Income and resource tests only 1.19 million
Total 9.89 million

By quintile, ranked as CBO ranks: households by size-adjusted income before transfers and taxes, in quintiles of equal numbers of people. Households with negative income stay out of the lowest quintile; they gain $51 million.

Quintile Change Per household Share of income after transfers and taxes
Lowest +$3.85 billion $168 0.27%
Second +$4.78 billion $195 0.26%
Middle +$1.89 billion $75 0.07%
Fourth +$1.09 billion $44 0.03%
Highest +$0.74 billion $28 0.01%

What the review fixes moved. Round 4's changes (5ef50e7: rental equity, zero floors, exact rounding) move nothing on this dataset, which has no real estate input and no negative asset amount (see Real estate), and the exact rounding gives the same limits in every year the model fills. Round 3's, against the previous head (e415392), record by record on the same dataset:

  • the rental-loss floor raises countable income for 4.63 million people (2.16 million of them Medicare enrollees), and 2,457 people lose the subsidy, $3.1 million;
  • the couple limit's statutory amount changes the resource test for 4 records, none of them Medicare enrollees;
  • separate filers' family size and real estate change nothing here: no record's family size moves, and the dataset has no real estate input.

Against the Trustees. The model reaches 69% of IV.B9's 14.4 million LIS enrollees and 68% of IV.B10's $18.1 billion. Two things account for much of the gap:

  • No Medicare enrollee in the dataset is under 65. All 24,147 Medicare-enrolled records are 65 or older: 58.5 million people. The 2026 Trustees Report counts 69.3 million enrollees in 2025, "62.2 million aged 65 and older, and 7.1 million disabled" (file page 13), and 70.7 million in 2026 (Table V.B3). pe-us reaches Medicare under 65 only through 24 months of SSDI (months_receiving_social_security_disability), and no Microcosm build sets that input (Medicare: add a Part B enrollment input and impute SSDI months so Part B rules reach the right people #10457, microcosm#876). So no disabled beneficiary under 65 can receive the subsidy here.
  • CBO drops institutionalized recipients. Its 2015–2016 counts came to 88–90% of the Trustees'. Applying that ratio to 2026 gives 12.7–12.9 million.

With #10118. #10118 takes $255.86 per Medicare enrollee out of cbo_medicare_benefit, which comes to $14.96 billion over the model's 58.5 million enrollees. This PR adds $12.40 billion back to the 9.89 million people it finds eligible.

Scripts and outputs: ~/reviews/medicare-part-d-lis-2026-10/impact/ (pair_inner.sh, compare.2026.fe4af16ec1.json, aggregates.2026.fe4af16ec1.json, r2_vs_r3.2026.json).

Invariants and tests

YAML: 82 cases across 10 files in tests/policy/baseline/gov/hhs/medicare/part_d/low_income_subsidy/. They include:

  • the exact boundaries: income at 150% of the guideline fails, and resources at the limit plus the burial exclusion pass while one dollar more fails, in 2026 and at the projected 2027 limits;
  • Alaska, Puerto Rico and DC;
  • the 2023 partial-subsidy range, and the 2013 and 2014 amounts;
  • a Medicaid enrollee who opted out of Part D, a Medicaid enrollee whose Part D input is False, and an eligible person outside Part D, none of whom gets the subsidy;
  • rental losses that don't offset Social Security or a spouse's income;
  • spouses who file separately, alone and with a child each;
  • non-home real estate, for the household head's couple only, for the older of two flagged heads and, without a head, for the oldest member's;
  • rental property equity: SI 01130.503's two examples ($10,000 of equity returning 6% counts $4,000; $3,000 returning $750 counts nothing), a return of exactly 6% and one a dollar short, a rental loss, negative equity, rent summed over the household, and the resource test at the limit and a dollar over with rental equity;
  • negative real estate and an overdraft, which count as zero;
  • the round 3 reviewers' two households: $100,000 of rental equity returning 6% counts $94,000 and fails the resource test, and −$5,000 of real estate no longer brings $20,000 under the limit;
  • households computed from their inputs through MSP deeming and through the LIS's own tests.

Elsewhere:

  • cbo_household_income.yaml gains a case where the LIS is the only means-tested transfer, and its list test now includes part_d_lis (and msp, which does not count);
  • healthcare_benefit_value.yaml gains a case with the LIS;
  • household_health_benefits.yaml gains two: the flag on counts the LIS, and the flag off leaves it out.

Python, tests/core/test_part_d_lis.py: 94 tests. At 4ecac84 the file took 24 s locally, 12–13 s of it the model load, and peaked at 1.46 GB against 1.13 GB for a bare model load; at 5ef50e7, with the registry tests in the same process, the peak was 1.45 GB. It runs in CI's core groups, which peaked at 6.0–9.7 GiB on main's 2026-11-07 run (42409791828). extend_part_d_lis_resource_limits runs on every model build and takes 0.05 s of it. Builds take 7–9 s locally: three fresh builds at the base and three at the head, with no difference beyond their spread (~/reviews/medicare-part-d-lis-2026-10/ci-cost/load_cost.txt).

  1. Transcription: the per-recipient series equals IV.B9 for 2013–2035.
  2. Conservation: IV.B9 enrollment × amount reproduces IV.B10 within the cells' rounding.
  3. Differential: the same payments over V.B3 total enrollment equal Count Medicare net of premiums in CBO household income #10118's per-enrollee series, which comes from different cells, within 0.5%. V.B3 for 2013–2014 comes from the 2021 report.
  4. Indexing, 2007–2027: both limits equal the statute's rule from the 2006 amounts in every year, with no exception, applied to BLS Septembers typed into the test. The typed Septembers match cpi_u_nsa_september wherever the parameter has them.
  5. Projection, 2028–2100: each limit equals the rule applied to the Septembers get_september_cpi_u_values returns, and the 2035 limit exceeds 2027's.
  6. Rounding: adjust_part_d_lis_resource_limit_for_inflation rounds to the nearest $10 with halves up and never lowers a limit, including three exact $5 products whose CPI ratios have no finite decimal; the published 2026 limits need the unrounded ratio.
  7. Synthetic trees: in the extension, YAML values stand, published Septembers come first, a fall in the index keeps the limit, a missing September sends both Septembers to the fallback series, and the value holds after end_year. In get_september_cpi_u_values, an instant other than September 1 (such as a backdated first value) doesn't count as a published September, and when a September is missing both come from the model's September values, not a neighboring month's.
  8. Bounds, 2013–2100: the amount stays positive and below Medicare spending per enrollee.
  9. Restated rules: a seeded sample of 300 households plus 12 boundary households (491 adults), each also run with more Social Security, more interest and more assets in one shared simulation. The test checks the model against rules it restates, not against the formulas:
    • countable income, with rental losses floored for each person, and family size;
    • the income and resource tests, with real estate and rental equity (after its exclusion) for the head's couple, each at no less than zero;
    • eligibility, the Part D plan condition and enrollment in the subsidy;
    • the benefit, which is zero or the year's amount.
  10. Interest and dividends never change countable income, eligibility or the benefit.
  11. Monotonicity: more Social Security never makes someone income eligible, and more assets never make someone resource eligible.
  12. Spouses get the same countable income, family size and test results, including spouses who file separately.
  13. Differential against MSP: with no interest, dividends or rental income, part_d_lis_countable_income equals 12 × msp_countable_income, a separate implementation of the same SSI exclusions (single people and couples who both have Medicare). The sample has no Illinois or Mississippi households, whose MSP income takes the larger exclusions.

The sample reaches every branch, and a test fails if one goes unreached. The test also checks that each variant raises its inputs and that each changes some result: more Social Security changes the income test for 14 households, more assets the resource test for 31, and counting the added interest would change the income test for 48. Counts at 5ef50e7 (scripts/reach_counts.5ef50e7a20.txt), adults for the per-person branches and households otherwise:

  • 179 households are couples, 113 have dependents and 33 live in Puerto Rico; 37 adults have no Medicare;
  • Medicaid alone deems 35 adults eligible, SSI alone 62 and an MSP alone 56;
  • 186 pass both tests, 154 fail only the resource test and 96 only the income test;
  • 171 receive the subsidy; 45 eligible people have no Part D plan, 47 have a plan but don't take it up, and 27 get it automatically despite a False take-up flag;
  • 4 households sit exactly on a resource limit and 4 one dollar over; 4 sit exactly on the income limit;
  • 10 separate-filer households have one dependent, and splitting it changes the income test for 3. The rental floor changes the income test for 9 households;
  • real estate changes the resource test for 27 households, and its zero floor for 4. Rental equity changes it for 19, and its exclusion for 3; 82 households hold rental equity whose return falls short, so all of it counts.

Mutations: each mutant edits a fresh git archive of a commit, which then runs the LIS YAML folder and the Python file, plus the suites the edit can reach (the household health benefit YAML, the Medicaid home equity YAML and the uprating tests). A mutant counts as caught only when a test fails, not when collection breaks. ~/reviews/medicare-part-d-lis-2026-10/mutants/ holds the results (results.txt, results_r3b.txt, results_r4.txt).

  • 20 at 33cc03d, all caught:
    • the Part D plan condition dropped, a strict resource test, income at the limit passing, a 145% income limit;
    • interest counted, the rental floor dropped, the spouse's income or unearned income ignored;
    • separate filers on their own return, a joint return's dependents doubled;
    • one burial exclusion per couple, real estate dropped or counted for everyone, the household head ignored;
    • no indexing, a lowered limit, a rounded ratio, SSA's 2015 couple limit, September 2026 dropped;
    • the subsidy left out of healthcare_benefit_value.
  • 23 at 4ecac84, aimed at the round 3 review questions; 21 caught:
    • both Septembers always from the model's series, no hold after the end year, banker's rounding, the LIS extension run before the CPI-U extension, the home equity Septembers swapped;
    • the youngest member holding the real estate when no one is head, a spouse's resources ignored, the couple limit for single people, no burial exclusion;
    • a rental loss offsetting the person's or the spouse's other unearned income, dividends counted, the spouse left out of family size, the guideline for one;
    • territorial residents qualifying, MSP or SSI no longer deeming, automatic enrollment dropped, take-up defaulting to False;
    • the subsidy left out of CBO's means-tested transfers or the household health benefit list.
  • 13 at 5ef50e7, aimed at the round 3 fixes; all caught. Each ran the one test file its edit targets, and each of those three files passes unedited (results_r4.txt):
    • the $6,000 exclusion without the 6% return test, rental equity counted in full, the return test on all the equity rather than the excluded part, a return of exactly 6% failing, one member's rent instead of the household's, negative rental equity passed through, a $5,000 exclusion, a 5% return;
    • rental equity left out of resources, the zero floor dropped for household or for personal amounts;
    • round 3's finite-precision decimal ratio, which fails only the three new tie cases, so the two methods give the same limits in every year the model fills; October in the fallback pair.
  • Two survived at 4ecac84; both fail at 2c447a3:
    • youngest_head: with two flagged heads, the younger one holds the real estate. No case had two heads; 2c447a3 adds one.
    • september_filter_dropped: an instant other than September 1 counts as a published September. On the real tree this changes nothing. The only such instant is the backdated 2015-01-01 entry, and the parameter's Septembers start in 2018, so no year gains a complete pair of Septembers through it. 2c447a3 adds a synthetic test of get_september_cpi_u_values.

With three people of the same age, the household amount goes to exactly one of them, and to a flagged head when there is one. That held for all 24 combinations of member order and head flags tried (scripts/tie_check.py).

Local runs at the head, 17646bf (main through 94054aa merged), one process at a time in a shared slot under run_capped.py (local-runs/summary.r5.17646bfd02.txt, core 3.36.0):

  • all ten LIS YAML files: 82 cases pass, peak 3.80 GB;
  • test_part_d_lis.py with the registry tests: 980 passed, 10 skipped, peak 1.41 GB;
  • uprating (home equity indexing and the uprating extensions): 188 pass;
  • code health: 547 passed, 1 skipped.

At 5ef50e7's tree (local-runs/summary.r4.b34287ff8f.txt), code health passed: 547 passed, 1 skipped.

Earlier, at 4ecac84 (local-runs/summary.4ecac84cc2.txt):

  • YAML: Medicare (279 cases in 5 batches), household income (69), the health benefit files (98), the EDAA reform (9) and Medicaid long-term care (62) all pass.
  • Python: the LIS file with the registry, aggregate-list, SPM-contract and Medicare-series tests (1,201 passed, 10 skipped), uprating (268), health (77) and code health (529 passed, 1 skipped) pass.
  • The code-health run's two failures, test_reform_folders_run_each_yaml_and_python_file_separately[True-contrib/crfb] and [True-reform], came from the local venv missing coverage. With uv.lock's 7.13.2 installed, all four of that test's cases pass.
  • At 2c447a3 the LIS YAML folder (68) and the Python file (90) passed.

Review

  • Round 2 (e415392), two independent reviews on Subfleet: one on the law, values and design, one on the formulas, tests and integration. Both asked for changes: 14 findings on the law side and 7 on the code side. ff54d4a, 33cc03d and fe4af16 address them:
    • the resource limits now index after 2026 (law 1, code 4);
    • non-home real estate counts (law 2);
    • spouses who file separately get one family size (law 3, code 2);
    • part_d_lis_enrolled documents the auto-enrollment contract (law 4);
    • the subsidy counts as a household health benefit (law 5);
    • part_d_lis says how its recipients differ from CBO's, and this body reports the impact (law 6);
    • the couple limit follows the statute, with a source-conflict candidate (law 7);
    • rental losses no longer offset other income (code 1);
    • the excluded sources come out of SSI's list instead of SSI's total (code 3);
    • more boundary cases and mutants for the Python file (code 5);
    • citations, descriptions, the registry entry and the changelog (law 8–13, code 6–7).
  • Round 3 (2c447a3), the same two lenses on Subfleet. Both found every round 2 finding resolved. New findings (full text in review/r3_law_full.md and review/r3_code_full.md):
    • P1, law 2: rental property equity was left out entirely. 5ef50e7 counts it less the 416.1222 exclusion (see Real estate).
    • P1, code 3: negative household equity offset other resources. 5ef50e7 floors each amount at zero, each person's too.
    • P1, law 1 and code 1: health benefits counted twice with the flag on. This predates the PR and stays with Fix household health benefit double counting #10403 (see Not modeled).
    • P2, code 2: a finite-precision decimal ratio rounded some exact $5 products down. 5ef50e7 uses exact fractions.
    • P2, code: the fallback's choice of September had no test (the october_fallback mutant survived). 5ef50e7 adds one.
    • P2, code: the sample's variants weren't checked to change anything. 5ef50e7 adds those reach checks.
    • P3: the registry's CBO wording, the take-up documentation's covered-retiree exception and the CY 2026 memo link (CMS's URL now serves CY 2027). All three are fixed.
  • Round 4: before the push, an in-session static check of 5ef50e7 (runs/r4-precheck/) found no P0–P2 issues. Its P3s, the statute citation for the exclusion and two comments, are fixed in 3022441. The US + core hub's independent review of b867d6d..17646bf (law, code, and the two merges) approved with no P0–P2 findings. 7764ee6 answers its P3s:
    • two at-limit cases show that a negative bank balance and negative real estate count as zero, not as their size. Both abs() floor mutants are caught, each by its own new case (mutants/results_r5.txt);
    • the rental variable and household_sources.yaml say why the exclusion follows the statute's section 1382b and POMS HI 03030.020 A rather than 20 CFR 418.3425 or 42 CFR 423.772 read alone, and cite 418.3425(c), 416.1222(a) and SI 01130.501. The review found no source conflict: 418.3425's list doesn't claim to be exhaustive;
    • the rental variable's documentation says trade-or-business rentals have no input, so every rental is held to the $6,000 exclusion;
    • at 7764ee6, part_d_lis_resource_eligible.yaml (17, the file with the new cases), the reference and source-conflict code-health tests (257) and the rental and integration YAML (24) pass locally (mutants/results_r5.txt, local-runs/summary.r5b.7764ee6248.txt).
    • The hub's delta review of 7764ee6 (Opus 5.5 on Subfleet) approved with no P1 or P2: both new cases are right by hand at the $18,090 limit, each catches its own abs() mutant, every new reference resolves, and only the three intended files changed. 5f60570 answers its two P3s: the codified 20 CFR 416.1222(a) still caps trade-or-business property at $6,000, and the documentation and comment now say that the last clause of 42 U.S.C. 1382b(a)(3) and POMS SI 01130.501 A.2 override that cap from May 1, 1990; the comment's point about the regulations now names the Extra Help regulations, since 416.1222(a) itself shows the rental exclusion for SSI.

Not modeled and follow-ups

  • Take-up. Microcosm seeds neither take-up flag (see Enrollment). microcosm#1206 proposes a Part D input, and a separate task covers the LIS flag, which should apply only to people with neither Medicaid nor SSI.
  • Medicare under 65. Medicare: add a Part B enrollment input and impute SSDI months so Part B rules reach the right people #10457 and microcosm#876 track the missing SSDI months.
  • Retirement accounts count as LIS resources (20 CFR 418.3405), but pe-us has no input for their balances.
  • Rental property: trade-or-business rentals, separate properties' 6% tests and temporary shortfalls (see Real estate).
  • The Social Security cost-of-living increase: SSA leaves it out of income through the month after HHS publishes the new poverty guidelines (20 CFR 418.3340(f)); the model counts the year's benefits in full.
  • Carry-over of deemed status: people deemed eligible from July to December stay deemed through the next year (423.773(c)(2)(ii)); the model tests each year on its own.
  • Institutionalized enrollees. See above; the average stays.
  • The premium and cost-sharing subsidies, including the pre-2024 partial subsidy's sliding scale, have no per-person computation.
  • SSI and the Medicare Savings Programs net rental losses and leave out non-home real estate in their own income and resource tests. Separate tasks cover them.
  • Health benefits counted twice with the flag on. With gov.simulation.include_health_benefits_in_net_income on, household_net_income already holds household_health_benefits, and household_net_income_including_health_benefits adds healthcare_benefit_value on top, so every health benefit counts twice there and in marginal_tax_rate_including_health_benefits. That includes Medicaid, MSP and CHIP, and now part_d_lis. Net income including health benefits counts health coverage twice when health benefits are in net income #9581 (open) tracks it, and Fix household health benefit double counting #10403 (open) fixes it for the whole list, the subsidy included, because this PR adds part_d_lis to both lists and keeps them identical, as Register the remaining five state premium programs in household_health_benefits #9514 made them. Fix household health benefit double counting #10403 conflicts with main's new premium_tax_credit_in_force subtraction and has had no CI run. The round 3 law and code reviews both raised this (P1); it predates this PR and touches code this PR doesn't change, so it stays with Fix household health benefit double counting #10403.
  • The source-conflict register: after merge, us:part-d-lis-couple-resource-limit-2015 goes into rulespec-us's known-source-conflicts.yaml, and a follow-up PR here marks the deciding line.

axiom: TheAxiomFoundation/rulespec-us#1681 queued (42 U.S.C. 1395w-114(a)(3) and 42 CFR 423.772–773; the corpus has both, but rulespec-us's pin, us-rulespec-2026-10-29-ecfr-section-text, includes neither; 20 CFR 418 ingestion is TheAxiomFoundation/axiom-corpus#818)

Sources

Source URL Retrieved
42 U.S.C. 1395w-114 https://www.law.cornell.edu/uscode/text/42/1395w-114 2026-10-11
42 CFR 423.772, 423.773 (eCFR, 2026-10-01) https://www.ecfr.gov/current/title-42/chapter-IV/subchapter-B/part-423/subpart-P 2026-10-11
42 CFR 423.34 (eCFR, 2026-11-05) https://www.ecfr.gov/current/title-42/chapter-IV/subchapter-B/part-423/subpart-B/section-423.34 2026-11-08
20 CFR 418 subpart D (eCFR, 2026-10-01) https://www.ecfr.gov/current/title-20/chapter-III/part-418/subpart-D 2026-10-11
SSA interim final rule, 75 FR 81843 (effective 2011-01-01) https://www.federalregister.gov/documents/2010/12/29/2010-32848 2026-10-11
SSA POMS HI 03020.010, .030, .050, .055; HI 03030.020, .025 https://secure.ssa.gov/poms.nsf/lnx/0603030025 2026-10-11; HI 03030.025 again on 2026-11-08
SSA POMS SI 00830.505 (rental income) https://secure.ssa.gov/poms.nsf/lnx/0500830505 2026-11-08
SSA POMS SI 01130.500, .501, .503 (property essential to self-support) https://secure.ssa.gov/poms.nsf/lnx/0501130503 2026-11-08; .503 on 2026-11-14
20 CFR 416.1222, 416.1224 (eCFR) https://www.ecfr.gov/current/title-20/chapter-III/part-416/subpart-L/section-416.1222 2026-11-14
CMS, 2015 LIS income and resource standards (Attachment 4, file pages 5–6) https://www.cms.gov/files/document/2015-lis-income-and-resourcespdf 2026-11-08
CMS, CY 2026 LIS resource and cost-sharing limits, October 31, 2025 (file page 2; sha256 d0e19604…0996). CMS's URL now serves the CY 2027 memo, so the parameters cite the Wayback capture of 2025-12-04, the same bytes https://web.archive.org/web/20251204185002/https://www.cms.gov/files/document/cy2026-lis-resource-limits-memo.pdf 2026-11-08; Wayback on 2026-11-14
CMS, CY 2027 LIS resource and cost-sharing limits, November 5, 2026 (sha256 90f92195…3ef2) https://www.cms.gov/files/document/cy2026-lis-resource-limits-memo.pdf 2026-11-14
2026 Medicare Trustees Report (sha256 ffa56b91…8fa8) https://www.cms.gov/oact/tr/2026 2026-10-11
2025 Medicare Trustees Report (sha256 1b3de7a4…4d95) https://www.cms.gov/oact/tr/2025 2026-10-11
2023 Medicare Trustees Report (sha256 89daaf29…fa67) https://www.cms.gov/oact/tr/2023 2026-11-07
2021 Medicare Trustees Report (sha256 ef3d1fa7…fc8e) https://www.cms.gov/oact/tr/2021 2026-11-07
Habib, CBO WP 2018-07 https://www.cbo.gov/system/files/2018-07/54234-workingpaper.pdf 2026-10-11 (browser)
CBO, January 2026 health insurance sensitivity report https://www.cbo.gov/publication/61910 2026-10-11 (browser, via #10118's notes)
BLS CPI-U, CUUR0000SA0 https://api.bls.gov/publicAPI/v2/timeseries/data/ 2026-10-11; September 2026 on 2026-11-08

🤖 Generated with Claude Code

MaxGhenis and others added 9 commits October 11, 2026 12:59
Replace KFF's 2021 traditional-Medicare figure and an unsourced $14,500
2025 estimate (held flat for every later year) with CMS National Health
Expenditure Table 17's Medicare spending per enrollee: historical
2013-2024, projected 2025-2034, extended to 2100 at the 2031-2032 rate
so the end of sequestration in 2033 does not compound.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Transcription checks against NHE Table 17 (and its spending over
enrollment), a cross-source level check against the Trustees Report's
per beneficiary cost, Hypothesis properties for order, the 2035-2100
closed form and calendar-year steps, and identities for
medicare_gross_cost, medicare_cost and qmb_cost_sharing. Existing
expectations move from $14,500/$11,080 to the CMS values.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
CMS Medicare spending per enrollee (calibration.gov.hhs.medicare.per_capita_cost,
#10107) includes Part D, which enrollee premiums
partly finance, but medicare_cost subtracted only Part A and Part B premiums.
It now also subtracts the Part D enrollment share times (Part D premiums per
Part D enrollee + the person's Part D IRMAA), both new calibration series from
the 2026 Medicare Trustees Report (Tables III.D3, V.E4 and V.B3).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Transcription from the Trustees' cells, conservation back to Table III.D3
premium income, bounds, the 2036-2100 extension's closed form, and a
Hypothesis differential test of medicare_cost against a reference
implementation (0 <= net <= gross, non-increasing in Part D IRMAA).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
CBO's standard household income counts Medicare at the government's
average cost per enrollee, net of premiums paid by enrollees, and counts
the Part D low-income subsidy separately as a means-tested transfer
(CBO, January 2026, publication 61910; Distribution of Household Income
in 2021, Appendix A; Habib 2018, Appendix B). The default list used
medicare_gross_cost, which is gross of premiums and includes the
subsidy.

Add cbo_medicare_benefit: Medicare spending per enrollee less
per-enrollee Part A, B and D premium income and the Part D low-income
subsidy per Medicare enrollee, from the 2026 Medicare Trustees Report.
Every enrollee gets the same value, as CBO assigns the average. Fix the
medicare_gross_cost docstring, which said gross matched CBO.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
YAML: the 2025, 2015 and 2035 values, and five people with different
premiums (Part A premium, income-related premiums, under 65, not
enrolled) who all get the same average. Python, parameters only:
transcription and conservation against the Trustees cells, the Part D
series against #10111's share x premium plus income-related premiums,
Part B premium income including income-related premiums, bounds through
2200, the post-2035 extension keeping each series' share of spending,
and January reads.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ans-tested transfer

Adds statutory eligibility for the Part D low-income subsidy (Extra Help):
countable income below 150% of the poverty guideline for the family size,
countable resources within the CPI-indexed limit plus the burial funds
exclusion, and deemed eligibility for Medicaid enrollees, SSI recipients and
Medicare Savings Program enrollees. The subsidy is valued at the Medicare
Trustees' average payment per subsidy enrollee (Table IV.B9), as CBO assigns
it, and joins gov.household.cbo_means_tested_transfers.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
YAML tests for every new variable, households computed from their inputs
(Medicare Savings Program deeming, the subsidy's own tests, the pre-2024
partial range, take-up), and the CBO means-tested list. Python invariants:
IV.B9 transcription, IV.B10 conservation, agreement with the per-enrollee
series cbo_medicare_benefit subtracts, the CPI rule behind the resource
limits, Hypothesis bounds, restated eligibility rules over a seeded sample
with income, interest and asset variants, and a differential test against
msp_countable_income.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…idy average

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The branch was stacked on #10107, which has since merged to main in revised
form, so its shared lines (the per capita cost series, its extension and its
tests) take main's version.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#10115 made income_adjusted_part_d_premium_surcharge zero for anyone not
enrolled in Part D (medicare_part_d_enrolled). Multiplying it by the Part D
enrollment share then counted non-enrollment twice once take-up is modeled,
and charged an enrolled person only 82 percent of the adjustment they pay.
The share now scales only the average Part D premium, which matches
spending per enrollee averaged over every Medicare enrollee; the person's
own adjustment amount is subtracted as computed, as the Part B one is.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Part D series listed the report in several reference entries that
differed only by #page=; CLAUDE.md asks for one entry per source with the
pages in its title. The 2030 base beneficiary premium reset is on file page
154, not 152, and file page 219 says low-income subsidy enrollees pay
substantially reduced premiums, not none. The December 31 credits move to
the following year, not back a year.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…re-net-of-premiums

Brings in current main via #10111. Main renamed unemployment_compensation
to total_unemployment_compensation in the CBO social insurance list, so the
list now reads social_security, cbo_medicare_benefit,
total_unemployment_compensation, workers_compensation.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LwQq6HHj8EHZr9hwABoKJk
CLAUDE.md asks for one reference entry per document with the page list in
the title. The four Trustees calibration series listed the 2026 report
two to four times with hrefs that differed only by #page=; each now has one
entry naming every table and PDF page its comments cite (and the LIS series
keeps a separate entry for the 2025 report). The CBO social-insurance list
now cites Appendix D (PDF page 39) of the 2021 report alongside note 3, and
the 2026 report's PDF page, and cbo_medicare_benefit notes the extra PDF
pages of its 2021 and 2018 CBO references.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LwQq6HHj8EHZr9hwABoKJk
claude and others added 8 commits November 5, 2026 03:33
cbo_medicare_benefit subtracts these series from
calibration.gov.hhs.medicare.per_capita_cost, which starts in 2013, but
each series started in 2015. So cbo_household_income raised
ParameterNotFoundError in 2013 and 2014 for any household with a Medicare
enrollee, where main returns a value.

The 2022 to 2026 Trustees Reports drop 2013 and 2014 from Tables III.B4,
III.C7 and V.B3, so Part A, Part B and enrollment come from the 2021
report. Part D premium income comes from the 2026 report's Table III.D3,
whose footnote 6 lists no December 31 credit in either year, and the
low-income subsidy from the 2023 report's Table IV.B10, the newest that
shows both years. Each value is the aggregate over Table V.B3's total
enrollment, rounded to the cent. The 2021 report reproduces every stored
2015 and 2016 value to the cent.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01H3DAwxuFHxBaZvWgnqNt1E
Nothing tested a year before 2015: the Hypothesis bounds test started at
2015 and every YAML case was 2015 or later, so the 2013 and 2014 crash in
cbo_household_income went unseen.

- cbo_medicare_benefit.yaml: Cases 6 and 7 assert the 2013 and 2014 net
  values. Case 3 is renamed, since 2015 is no longer the series' first
  year.
- cbo_household_income.yaml: the review's 2014 witness, an enrollee with
  $10,000 of earnings, asserts the net benefit and total income.
- test_cbo_medicare_benefit_series.py: a transcription test recomputes
  each 2013 and 2014 value from the Trustees cells, and the bounds test
  now starts at 2013 and always tries 2013.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01H3DAwxuFHxBaZvWgnqNt1E
The documentation said "net of premiums paid by or on behalf of
enrollees", stating as fact a reading CBO's own wording ("premiums paid
by enrollees") leaves open. Say instead that the series net premium
income as the Medicare Trustees count it, which includes premiums state
Medicaid programs pay, and that PolicyEngine counts those in Medicaid:
medicaid_cost_if_enrolled spreads state totals whose MACPAC (2021-2022)
and KFF sources include Medicaid's payments of Medicare premiums.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01H3DAwxuFHxBaZvWgnqNt1E
…m titles

The date test now draws from 2013 and pins December 28, 2013, so a
mid-year entry under a 2013 value fails it. Three reference titles drop
the one page they cite; each href keeps it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Say in `medicare_cost`'s documentation that every enrollee nets the same average premium.
- Name retiree drug subsidy beneficiaries in both labels, and add page 153 to the premium reference.
- Drop the test's reference equality, keeping the bounds and the IRMAA slope.
- Add a 2036 case that pins where the premium's extension starts.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The subsidy pays part of a Part D plan's premium and cost sharing, so
42 U.S.C. 1395w-114(a)(3)(A)(i) defines a subsidy eligible individual
as one enrolled in a prescription drug plan or MA-PD plan, and 20 CFR
418.3101(b) says an applicant "must enroll in one of these plans to
actually receive a subsidy". medicare_part_d_enrolled reached main in
#10115 after this branch began, so part_d_lis_enrolled now requires it,
including for Medicaid enrollees and SSI recipients: Medicare enrolls
them in a plan automatically (42 CFR 423.34), but they can opt out of
Part D, for example for TRICARE for Life or VA drug coverage.

takes_up_medicare_part_d_if_eligible defaults to true and Microcosm does
not yet set it, so the microsimulation is unchanged; a household that
sets it to false now gets no subsidy. The seeded household sample in
test_part_d_lis.py now draws Part D enrollment and its restated rule
requires it; the YAML tests add an opt-out case for each path.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#10118 extended calibration.gov.hhs.medicare.part_d_low_income_subsidy_
per_enrollee and the other Medicare per-enrollee series back to 2013,
the first year of per_capita_cost. The per-recipient average started in
2015, so part_d_lis raised ParameterNotFoundError for a 2013 or 2014
enrollee. The 2023 Trustees Report's Table IV.B9 (file page 158), the
newest that shows both years, gives $2,023 per subsidy enrollee in 2013
(11.5 million enrollees) and $2,052 in 2014 (11.8 million); its Table
IV.B10 gives $23.2 billion and $24.3 billion, which the products match
within rounding, and the 2021 report's Table V.B3 enrollment (52,504 and
54,115 thousand) turns them into #10118's per-enrollee values within
0.4%. The invariant tests now cover both years.

Before 2015, cbo_household_means_tested_transfers still raises
ParameterNotFoundError through Medicaid and SNAP parameters, as it does
without this change.

The 2026 report's reference now lists both cited pages (119 and 158),
and the Habib reference both of its pages (51 and 52), per the
repository's reference rules.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@MaxGhenis

Copy link
Copy Markdown
Contributor Author

part_d_lis gives each recipient the Trustees' average subsidy per LIS enrollee, which includes nursing home residents. CBO removes institutionalized LIS recipients and their benefits from its totals, using the Master Beneficiary Summary File (Working Paper 2018-07, file page 52). #10126 scales the other Medicare values to enrollees outside nursing homes with a ratio from MedPAC's fee-for-service Part A and B spending, which doesn't cover the LIS. Matching CBO here would take an LIS average per recipient outside institutions.

MaxGhenis and others added 3 commits November 8, 2026 01:05
- Index the resource limits after 2026 by the unrounded September CPI-U
  ratio (42 U.S.C. 1395w-114(a)(3)(E)(i)), and add September 2026's CPI-U.
- Follow the statute's chain for the couple limit from 2015, entered as a
  source-conflict candidate against CMS's 2015 memo.
- Count non-home real estate for the household head's couple, floor each
  income source at zero, count both returns' dependents for spouses who
  file separately, and document the auto-enrollment data contract.
- Fix the per-recipient and jurisdiction citations.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Add part_d_lis next to msp_cost in household_health_benefits and
  healthcare_benefit_value, which count only when
  include_health_benefits_in_net_income is on.
- YAML cases for the indexed 2027 limits, the statutory couple limit,
  non-home real estate and its attribution, rental losses, separate
  filers' family size and the Part D input of Medicaid enrollees.
- Python tests: the CPI rule through 2027 with no published exception,
  projected years and a synthetic tree for the extension, exhaustive
  loops for the bounds, and a sample with separate filers, rental
  losses, real estate and households on each limit, plus a reach check.
- Move the registry entry to the HHS block and fix its notes.
- Read only September instants as published CPI-U values.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MaxGhenis and others added 3 commits November 8, 2026 01:59
16,590 x 336.940 / 324.800 is 17,210.08 and 33,120 x 336.940 / 324.800
is 34,357.92. The rounded limits and the tests' inputs don't change.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
CMS enrolls a qualifying covered retiree only if they are a full-benefit
dual eligible whose plan sponsor it approved for the retiree drug subsidy
and they confirm that they want a plan; it enrolls no other covered
retiree. The docstring had limited the exception to retirees whose
employer gets the subsidy. Documentation and a test comment only.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A household with two flagged heads gives its real estate to the older
one, and a non-September instant in the September CPI-U node is not a
published September. Both rules had no test; the round 3 mutants that
reversed them survived.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MaxGhenis and others added 14 commits November 11, 2026 14:50
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The module docstring promised a reference-implementation comparison that the
file no longer contains. It now names the bounds check the test runs and
points to the YAML cases that pin the value.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Main's #10632 added two 2026 cases to medicare_cost.yaml. Keep them as
unit tests 14 and 15, after this branch's cases.

The temporary protected status case now subtracts the 2026 average
Part D premium: $18,983 - $2,434.80 Part B - 0.8195 x $289.79 = $16,310.72.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The Hypothesis test adds a third enrollee with different Part A and
Part B premiums and checks that all three people have the same gross
cost. Invariant 6 now names the three premiums the test varies.

Invariant 2 now says its $0.2 billion slack covers the December 31
credits that Table III.D3 books a year early.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Main's version bump consumed changelog.d/medicare-cost-part-d-premiums.fixed.md
into CHANGELOG.md. Merging #10111's branch head restored it, which would repeat
the entry in the next release.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
CBO's 2021 household income report defines its terms in Appendix D,
which starts on PDF page 41. Page 39 holds Table C-3. Cite page 41 in
the reference title and comment of cbo_social_insurance_benefits.yaml
and in cbo_medicare_benefit's reference comment.

Describe Medicare in cbo_household_social_insurance_benefits as
cbo_medicare_benefit does: average spending per enrollee less premium
income and the Part D low-income subsidy.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Conflicts, resolved keeping both sides:
- uprating_extensions.py: the Part D LIS helpers and main's spousal
  impoverishment functions both stay, and the LIS extension runs after
  the home equity limits and before the spousal impoverishment
  standards. get_september_cpi_u_values now describes main's not
  seasonally adjusted CPI-U series.
- cpi_u_nsa_september.yaml: main's description, since the series now
  also feeds the spousal impoverishment index.
- source_conflicts/candidates.yaml: main's #10076 note and new NIIT
  candidates, then the LIS couple-limit candidate.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Answers the round 3 reviews of #10127:

- Rental property equity counts as an LIS resource, less up to $6,000 when
  the household's net rent is at least 6% of the excluded equity (20 CFR
  418.3405(b), 416.1222(a); POMS HI 03030.020, SI 01130.503). Every rental
  is treated as nonbusiness property.
- Each resource amount, the household's and each person's, counts at no
  less than zero, so negative equity or an overdraft no longer offsets
  other resources.
- The resource-limit adjustment uses exact fractions, so exact $5 products
  round up; the finite decimal ratio put some just below $5.
- Tests: rental equity and zero-floor YAML, the reviewers' households, the
  exact-tie rounding cases, the model-September fallback, and reach checks
  that each sample variant changes a result.
- Citations: the CY 2026 CMS memo at its Wayback capture (CMS's URL now
  serves CY 2027), the registry's CBO wording and the take-up docs.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The rental property exclusion now cites 42 U.S.C. 1395w-114(a)(3)(E)(i),
which makes the subsidy's resources SSI's, and 1382b(a)(3), SSI's
exclusion for property essential to self-support. The variable's
documentation names the parameters instead of their 2006 values, and the
September filter's comment says what the gap in published years means.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@MaxGhenis
MaxGhenis changed the base branch from cbo-medicare-net-of-premiums to main November 14, 2026 22:12
@MaxGhenis
MaxGhenis marked this pull request as ready for review November 14, 2026 22:12
MaxGhenis and others added 2 commits November 14, 2026 17:36
Round 4 review P3s on #10127:
- Two at-limit cases show a negative bank balance or negative real estate
  counts as zero rather than as its size, so an abs() floor now fails.
- The rental variable and household_sources.yaml say why the exclusion
  follows the statute's section 1382b and POMS HI 03030.020 A rather than
  20 CFR 418.3425 or 42 CFR 423.772 read alone, and cite 418.3425(c),
  416.1222(a) and SI 01130.501.
- The rental variable's documentation says trade-or-business rentals have
  no input, so every rental is held to the $6,000 exclusion.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Round 4 delta review P3s on #10127: the codified 20 CFR 416.1222(a) still
caps trade-or-business property at $6,000, so the documentation and comment
now say that the last clause of 42 U.S.C. 1382b(a)(3) and POMS SI 01130.501
A.2 override that cap from May 1, 1990. The comment's claim that the
regulations don't show the rental exclusion now names the Extra Help
regulations, since the SSI regulation 416.1222(a) does show it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@MaxGhenis

Copy link
Copy Markdown
Contributor Author

Handing this PR to the US + core hub. Its remaining steps are CI, merge order and a decision for Max.

Head: 5f60570dcc81. CI run 43654498751 was running at hand-off: 3 checks passed, 24 pending. The PR's test merge, 956b77b, sits on main f67128e; the checkout log will confirm that once the run completes.

Review: every round has finished, with no open P0–P2.

  • Round 4, at 17646bf: the hub's law, code and merge lenses all approved, with no P0–P2 (~/reviews/us-hub/landing-2026-11-10/runs/opus-1114-reviews/pr10127-r4-{law,code,merge}.md). 7764ee6 answers its P3s.
  • Delta review of 7764ee6: approved, with no P1 or P2 (pr10127-r5.sol.out.md in the same folder). 5f60570 answers its two comment P3s, and the hub read that commit: documentation and comment only.

Passed locally:

  • At 17646bf: all ten LIS YAML files (82 cases), test_part_d_lis.py with the registry tests (980 passed, 10 skipped), uprating (188) and code health (547 passed, 1 skipped).
  • At 7764ee6:
    • part_d_lis_resource_eligible.yaml (17);
    • both abs() floor mutants, each caught by its own new case;
    • the reference and source-conflict code-health tests (257);
    • the rental and integration YAML (24).
  • Evidence is in ~/reviews/medicare-part-d-lis-2026-10/ (mutants/results_r4.txt, results_r5.txt, local-runs/).

Depends on: #10118, which squash-merges first. It waits on Max's d2930.

Landing steps (~/reviews/medicare-part-d-lis-2026-10/LANDING.md): after #10118 squash-merges, merge main into this branch with a plain merge. If a release has already folded #10118's fragment into CHANGELOG.md, delete changelog.d/cbo-medicare-net-of-premiums.fixed.md in that merge, as in 103cd4a. Keep changelog.d/medicare-part-d-lis.added.md. Then confirm the diff against main holds only LIS work, and merge on a green run with every check pass or skipping and the PR mergeable.

Follow-ups owed:

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants