Skip to content

Add premium tax credit rules for 2014–2017 - #10846

Merged
MaxGhenis merged 20 commits into
mainfrom
add-pre2018-ptc
Nov 17, 2026
Merged

MaxGhenis merged 20 commits into
mainfrom
add-pre2018-ptc

Conversation

@MaxGhenis

@MaxGhenis MaxGhenis commented Nov 12, 2026 •

Copy link
Copy Markdown
Contributor

Calculate premium tax credits for 2014–2017 using annual applicable percentages, prior-year poverty guidelines, and estimates of historical benchmark premiums.

Congress supplied the original percentages in 26 U.S.C. 36B(b)(3)(A).
The IRS indexed them for 2015, Rev. Proc. 2014-37 §5.01, 2016, Rev. Proc. 2014-62 §2.01, and 2017, Rev. Proc. 2016-24 §2.01.
For example, the unit tests supply $30,000 MAGI, 250% FPL and a $12,000 benchmark.
Subtract $30,000 times 8.05%, 8.10%, 8.18% or 8.21%: credits equal $9,585, $9,570, $9,546 and $9,537.

Estimate historical benchmarks from 2018 rating-area premiums and the existing national premium index parameter.
Apply the CMS 2013 age curves, retaining the 2018 premium normalization.
Use the federal curve for historical New Jersey estimates.
In CMS's 2014 individual-market plan file, every New Jersey plan with adult rates follows the federal age ratios at each listed age from 21 to 60.
CMS's 2017 table labels the separate New Jersey column Small Group.

Use 2013 and 2014 HHS guidelines for coverage in 2014 and 2015.
Read the guideline HHS had most recently published when open enrollment began, under 36B(d)(3)(B), with the coverage year's tax-family size under 36B(d)(1).
For example, an Alaska couple with $49,150 wages in 2015 uses $14,580 + $5,080 = $19,660, so income equals 250% FPL (2015 instructions, Table 1-2).
Read the poverty-line state for the coverage year in every year, as the IRS instructs for Form 8962 line 4 (2016, 2017).
Main has read it that way since #10860 merged.
For example, an Alaska adult with $37,100 MAGI in 2017 who supplies only 2017 residence uses $14,840, so income equals 250% FPL (2017 instructions, Table 1-2).
Subtract $37,100 times 8.21% from the estimated $6,344.08 benchmark: the credit equals $3,298.17.
HHS defines no guideline for Puerto Rico or other outlying jurisdictions (2013, 2014).
aca_fpg returns a poverty line of zero for a household in Guam, Puerto Rico or the Virgin Islands in every coverage year, as it does on main for coverage from 2015.
It looks those households up under the contiguous guideline, because the shared guideline parameter holds no value for the three areas at the 2013 and 2014 dates.
That parameter keeps main's entries for the three areas, so every other formula that reads them gets main's result.
Preserve main's coverage-family pricing in the baseline and minimum-premium reform (#10041).
Read supplied benchmarks directly when every person belongs to the tax family.
Skip Medicaid category calculations when neither the federal nor the Arkansas work requirement applies.

Floor supplied negative MAGI at zero in 2014–2017, following 2014 Form 8962 line 3.
The IRS continues that instruction in 2018; correcting the existing 2018+ calculation needs a follow-up.
Retain the filing condition and coverage-family benchmark limits, 26 CFR 1.36B-3(d)–(f), including in the minimum-premium reform.
Keep main's allocation of a shared policy's premium in proportion to each taxpayer's benchmark under §1.36B-3(h)(1).

Historical benchmark changes also affect Medicaid cost indices, Basic Health Program costs and state premium programs.
The 2013 and 2014 guidelines also change other 2013 and 2014 results, and 2015 results that read a lagged guideline.
SNAP income standards read the guideline at each October 1, so they change from October 2013 through September 2015.
For one person in the 48 states, that guideline rises from $10,890, the 2011 amount, to $11,490 for fiscal year 2014 and $11,670 for fiscal year 2015.

Invariants and validation

  • Given a nonnegative benchmark, historical credits stay between zero and the benchmark, including for supplied negative MAGI.
  • Through 400% FPL, required contributions never decrease with income, and credits never increase while the benchmark and eligibility stay fixed.
  • Applicable percentages follow each annual table at its boundaries and within each tier.
  • When every person belongs to the tax family, the PTC benchmark equals slcsp for any enrollment.
  • When neither the federal nor the Arkansas work requirement applies, the Medicaid work bar returns false for every category.
  • Credits remain zero before 2014 and above 400% FPL in 2014–2017.
  • Preserve the 2018+ percentage tables, premium parameters and credit arithmetic inherited from main.
  • In every coverage year, the poverty line depends on the coverage year's state, as on main since Use the Alaska and Hawaii poverty guidelines in the premium tax credit's income fraction #10860 merged; a state supplied for the year before changes nothing.
  • In Guam, Puerto Rico and the Virgin Islands, aca_fpg returns zero in every coverage year.
  • The guideline parameter's entries for Guam, Puerto Rico and the Virgin Islands equal main's at every date.

The YAML tests comprise 249 cases in 11 files.

File Cases
is_medicaid_ineligible_due_to_work_requirement.yaml 6
aca_ptc.yaml 74
aca_required_contribution_percentage.yaml 44
family_tier_applies.yaml 4
family_tier_multiplier.yaml 14
slcsp_age_0.yaml 32
slcsp_age_curve_multiplier.yaml 10
tax_unit_fpg.yaml 12
aca_fpg.yaml 17
aca_magi_fraction.yaml 17
ptc_minimum_premium.yaml 19

Four residence cases supply Alaska residence for 2016 or 2017 coverage with no state, or another state, for the year before.
Four cases check the zero poverty line: a 2015 Guam return, a 2014 Puerto Rico return, a 2015 Virgin Islands return and a 2017 Puerto Rico return.
The three committed property tests check table boundaries, contribution monotonicity, credit bounds and compatibility across 2018–2027.
The compatibility test uses frozen 2018–2027 percentage tables and supplied benchmark premiums.

Pull request run 43910644935 at f5cb72ef24 passed all 37 jobs on GitHub's merge with main 8e1837790b.

Methodology choices

  • Estimate 2014–2017 rating-area premiums from 2018 data; these estimates do not reproduce each area's historical prices.
  • Keep the model's 2018 normalization: divide historical federal age factors by 0.765 and AL/MS/OR factors by 0.635.
  • Date NY/VT family tiers to 2014, supported by NY DFS Exhibit 25 and Vermont's 2013 order.
  • Estimate 2014 NY extended-dependent premiums with the ordinary child tier and no added factor. Apply the Age 29 factor, first dated 2023, to 2015–2022 through the model's backdating; the checked sources do not establish a uniform historical surcharge.
  • Keep the wider model's 2015 parameter floor. Ordinary 2014 households need supplied MAGI and coverage eligibility; retain verified_start_year: 2024.
  • Read residence in the coverage year in every year, as the IRS instructs for Form 8962 line 4, and lag only the guideline parameters. Main has read residence that way since Use the Alaska and Hawaii poverty guidelines in the premium tax credit's income fraction #10860 merged and closed ACA applicable percentage uses the contiguous-states poverty guideline for Alaska and Hawaii #9482; this PR lets the lagged guideline reach 2013 where main stops at 2015.
  • Retain current interpolation precision and income truncation. The IRS directed 2014 filers to round income percentages, line 5; implementing that year's rounding needs a follow-up.
  • Preserve inactive minimum-premium behavior in 2014 without deriving historical bronze premiums. The 2014 before-filing calculation omits the minimum-premium cap; an explicitly activated 2014 minimum needs separate work.
  • Leave exact-400% eligibility to Treat income exactly at 400% FPL as premium tax credit eligible #10312.

Microsimulation impact

None from 2018 on, by a code read. This PR changes results only for 2013 through 2017: every changed parameter keeps its values from January 1, 2018, and test_post_2017_credit_matches_existing_rules holds later credits to the existing rules. The default dataset holds 2024 records, so we report no aggregate for 2013 to 2017.

Axiom parity

axiom: TheAxiomFoundation/rulespec-us#2291 queued (the 2014–2017 applicable percentage tables, 2013–2015 HHS guidelines and 2014–2016 Form 8962 lines 3–5; the corpus lacks the revenue procedures, HHS notices and instructions, so the signed encoder can't run yet)

🤖 Generated with Claude Code

MaxGhenis and others added 3 commits November 12, 2026 01:40
Date the annual applicable percentages and estimate historical benchmark
premiums from the existing 2018 rating-area data. Add historical CMS age
curves and poverty guidelines, and retain the filing condition.

Floor supplied negative MAGI for the historical years and preserve the
calculations used from 2018 onward. Cover the tables, premiums, eligibility
ceiling, filing behavior and compatibility with YAML and property tests.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Set the state input for the coverage year in December cases. This lets
the New Jersey case exercise its historical curve.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Compare credits with the calculated annual benchmark after monthly input
conversion. Release simulations and dated parameter caches between checks
to bound memory use.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MaxGhenis and others added 2 commits November 14, 2026 22:37
Preserve coverage-family pricing and all historical cases while combining the family-tier helpers.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MaxGhenis and others added 11 commits November 15, 2026 03:20
Retain verification from 2024 and clarify that the IRS negative-MAGI instruction continues after 2017.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Read the 2013 and 2014 guideline parameters with coverage-year household inputs. Preserve coverage-family tests and the original 2017 input cases alongside the pre-coverage guards.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Remove the unsupported small-group curve from historical individual premiums. Identify the Default column in the Alabama, Mississippi and Oregon references.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Use the coverage family benchmark in 2014–2017 and its allocated premium for the minimum-premium cap from 2015. Cover claimable spouses in each historical formula.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Avoid enrollment and premium calculations when every potential enrollee belongs to the tax family. Retain allocation for families with excluded enrollees.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Read the existing applicability flags before deriving eligibility categories. Return false when both flags are inactive so historical ACA enrollment does not require later Medicaid categories.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The IRS instructs 2016 and 2017 filers to enter the Form 8962 line 4 amount for their state of residence in the coverage year, using the prior year's poverty lines. Read the state for the coverage year for every coverage year before 2018, so an Alaska or Hawaii household that supplies only its coverage-year state receives its own poverty line. Keep the existing guideline-year lookup from 2018.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The factor's first value dates from 2023, and the model backdates it to 2015. State in the family-tier documentation that calculations for 2015 through 2022 apply that value by choice.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Drop a 2013 case identical to a later one and a 2017 case whose input and assertion the next case repeats. Supply Marketplace enrollment to the property tests only for 2014, which precedes the model's 2015 parameter floor.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
@MaxGhenis
MaxGhenis marked this pull request as ready for review November 15, 2026 23:21
@MaxGhenis MaxGhenis closed this Nov 16, 2026
@MaxGhenis MaxGhenis reopened this Nov 16, 2026
@MaxGhenis MaxGhenis closed this Nov 16, 2026
@MaxGhenis MaxGhenis reopened this Nov 16, 2026
DTrim99 added a commit that referenced this pull request Nov 16, 2026
Main added 2017 coverage cases to aca_magi_fraction.yaml (the 2016 HHS
table, #10847/#10846 line); keep both sides' cases. Main's fpg helper now
reads the 2016 table but keeps its signature and the territory placeholders,
so aca_fpg and the zero-guideline guard are unaffected.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MaxGhenis and others added 2 commits November 16, 2026 09:24
Since #10860, main reads the coverage year's state for the poverty line in
every year. Resolve the conflict in aca_fpg.py: take that rule, drop this
branch's 2018 residence boundary, and keep the 2013 guideline floor, so
2014 and 2015 coverage use the 2013 and 2014 guidelines.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
For 2014 and 2015 coverage, aca_fpg reads the poverty guideline at 2013
and 2014 dates. The Guam, Puerto Rico and Virgin Islands placeholders
started in 2024, and the model backdates parameters to 2015 only, so the
lookup found no value for a household there in those two years. Main
reads the backdated zero for 2015.

Start the six placeholders in 2013 and keep their 2024 entries, where
uprating begins. No value changes from 2015 on. Add a 2015 Guam case and
a 2014 Puerto Rico case.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
MaxGhenis and others added 2 commits November 16, 2026 11:40
Remove the 2013 zero placeholders for Guam, Puerto Rico and the Virgin
Islands from gov.hhs.fpg, so the parameter's territory entries equal
main's again. Other formulas read that parameter. For example, with
those zeros tanf_non_cash_fpg returned a monthly guideline of zero for a
Guam or Virgin Islands unit in 2015, where main raises an error.

aca_fpg looks a household in those areas up under the contiguous
geography, which has a value at every guideline date, and returns zero
for it. A lookup on the area's own geography raises an error at the 2013
and 2014 dates that 2014 and 2015 coverage reads. For coverage from
2016, main's placeholder gives the same zero.

Add a 2015 Virgin Islands case and a 2017 Puerto Rico case, and restate
the comments on the 2015 Guam and 2014 Puerto Rico cases.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
@MaxGhenis
MaxGhenis merged commit f8fce90 into main Nov 17, 2026
37 checks passed
@MaxGhenis

Copy link
Copy Markdown
Contributor Author

US + core hub merge audit, us#10846 at f5cb72ef240bef51a5be246073644c36b6ba5d2a (squash): Add premium tax credit rules for 2014–2017.

  • Reviews:
    • Four independent review rounds ran outside GitHub, with a fix round after the first, the second and the fourth, and a scoped closing check after the last fix. A merge of main and a territory fix followed the third round.
    • The first round, at 645fc88a45, read 36B(b)(3)(A), the three revenue procedures, the 2013 and 2014 HHS notices and the 2014 to 2017 Form 8962 instructions. It also read the CMS age-curve tables, Vermont's 2013 order and New York's Exhibit 25. All 48 applicable percentages, both guideline years, the age curves and the family tiers matched.
    • On the head, 160 YAML cases and three property tests passed. With main's newer aca_magi_fraction in place, two Alaska cases failed: its guard sent 2014 and 2015 coverage to the coverage year's guideline. It also found no source for a New Jersey individual-market age curve, and a verified_start_year of 2014 that ordinary 2014 households couldn't compute.
    • The first fix round, ending at ff06996e06, let the guideline lookup reach 2013 and dropped the New Jersey curve for the federal one. It kept verified_start_year at 2024, said that the negative-MAGI instruction continues after 2017, and corrected the Alabama, Mississippi and Oregon citations.
    • The second round, at ff06996e06, confirmed each first-round finding fixed. It recomputed every changed expectation and checked both merges of main case by case. It found every New Jersey plan with adult rates in CMS's 2014 plan file on the federal age ratios, 625 rows in all.
    • It found that 2016 and 2017 coverage read the state from the year before. By hand, an Alaska adult who gave only 2017 residence would get the 48-state line and a $2,749.09 credit instead of $3,298.17. It also asked for description corrections, wording on New York's Age 29 factor, merging two pairs of repeated cases and a parity issue.
    • The second fix round, ending at 1b4e15dfab, made the lookup read the coverage year's state before 2018 and added four cases with no state, or another state, for the year before 2016 or 2017 coverage. It called New York's 2015 to 2022 Age 29 factor a methodology choice, dropped the repeated cases and limited the property test's enrollment input to 2014.
    • The third round, at 1b4e15dfab, found no defect in the code. It recomputed the four new cases, audited both merges path by path, confirmed the CI run and counted 236 YAML cases. Its three findings concerned an unfiled follow-up draft, a non-government link in the description and the open parity line.
    • The draft overstated its effect, and Use the Alaska and Hawaii poverty guidelines in the premium tax credit's income fraction #10860 has since merged and closed ACA applicable percentage uses the contiguous-states poverty guideline for Alaska and Hawaii #9482, so it stays unfiled. The description now links the national index parameter on GitHub, and a new parity issue answers the third finding.
    • The third round couldn't reach the government hosts. A check at 1b4e15dfab then read the 2014 code edition of 36B, the three revenue procedures, 26 CFR 1.36B-1 and 1.36B-3, and the 2013 to 2015 guidelines. It also read the 2014 to 2016 Form 8962 instructions. Every table value and guideline in the PR matches.
    • The reviews listed corrections to the description, and I made each before merge. The description also now says that the new guidelines move SNAP from October 2013 through September 2015, and that the 1.36B-3(h)(1) allocation comes from main.
    • Use the Alaska and Hawaii poverty guidelines in the premium tax credit's income fraction #10860 then merged, and main now reads the coverage year's state in every year. A merge of main at fd0a748698, commit bcd8072d61, resolved the one conflict, in aca_fpg.py. The lookup takes main's state rule, drops this PR's 2018 boundary and keeps its 2013 guideline floor. Every case either side added stays, and no expectation moved.
    • A check of that merge compared every path with both parents and recomputed each case the merged lookup could move. It found no guideline value for Guam, Puerto Rico or the Virgin Islands at the 2013 and 2014 dates the lookup reads for 2014 and 2015 coverage. By a code read, the lookup raised an error for a household there, where main returns the zero placeholder for 2015.
    • A fix at 18c09a91d0 dated those zero placeholders from 2013 and added a 2015 Guam case and a 2014 Puerto Rico case. A second check recomputed both cases by hand, read the 2013 and 2014 HHS notices and confirmed that no parameter value moved from 2015 on. At that head, the baseline job that runs aca_fpg.yaml passed all 15 cases.
    • The fourth round, at 18c09a91d0, found the merge faithful and one defect in the territory fix. Other formulas read the same placeholders. With the 2013 zeros, tanf_non_cash_fpg returned a monthly guideline of zero for a Guam or Virgin Islands unit in 2015, where main raises an error. SNAP's broad-based categorical eligibility then got a gross income limit of zero there.
    • The third fix round, ending at f5cb72ef24, removed the 2013 zeros, so the guideline parameter's territory entries equal main's again. aca_fpg now looks a Guam, Puerto Rico or Virgin Islands household up under the contiguous guideline and returns zero for it. The round added a 2015 Virgin Islands case and a 2017 Puerto Rico case, and merged main at 8e1837790b with no conflict. A check of the round recomputed the four territory cases and the guideline for every area at each coverage year from 2014 to 2019, and found no required fix. At this head, the baseline job that runs aca_fpg.yaml passed all 17 cases.
    • A scoped closing check at f5cb72ef24 found that defect closed and no regression. It found the guideline parameter identical to its content at 1b4e15dfab, recomputed 72 poverty-line values for coverage 2014 to 2019 and the four territory cases, and rebuilt the merge of 8e1837790b with git merge-tree. Six CI jobs had not finished when it read the run, and all 37 jobs have since passed.
  • Methodology choices for maintainers:
    • The PR estimates 2014 to 2017 benchmarks from 2018 rating-area premiums and the national premium index, and these estimates don't reproduce each area's historical price. For example, the PR estimates a $6,344.08 benchmark for an Alaska 40-year-old in rating area 1 in 2017.
    • The model keeps the unrounded interpolated percentage, where 26 CFR 1.36B-3(g)(1) and the IRS's Table 2 round it to 0.01%. At 134% in 2014, the model uses 3.0588% and Table 2 uses 3.06%.
    • The model truncates the 2014 income percentage, where that year's line 5 instructions round it. By hand, a single filer in the 48 states with $28,700 of 2014 income gets 249% and an 8.015% rate, against the IRS's 250% and 8.05%. That filer's credit runs about $10 higher.
    • The lookup reads the coverage year's state in every year, as Form 8962 line 4 directs, and main has done so since Use the Alaska and Hawaii poverty guidelines in the premium tax credit's income fraction #10860 merged. This PR lets the lagged guideline reach 2013, where main stops at 2015.
    • HHS defines no guideline for Puerto Rico or other outlying jurisdictions. aca_fpg returns a poverty line of zero for a household in Guam, Puerto Rico or the Virgin Islands in every coverage year, as it does on main for coverage from 2015. The formula looks those households up under the contiguous guideline and then returns zero. The guideline parameter keeps main's entries for the three areas, so any other formula that reads them at a 2013 or 2014 date raises an error, as on main. For example, by a code read, tanf_non_cash_fpg raises one for a Guam unit in January 2015, on main and at this head.
    • The PR floors negative MAGI at zero for 2014 to 2017 only. The IRS keeps that line 3 instruction after 2017, and main's 2018 formula doesn't apply it.
    • Historical New Jersey estimates use the federal age curve. New York and Vermont family tiers start in 2014, and New York's Age 29 factor reaches 2015 to 2022 through backdating.
    • verified_start_year stays 2024. Below 2015, the wider model lacks some parameters, so an ordinary 2014 household needs supplied MAGI and eligibility.
  • Gates checked live at merge:
  • Impact: none from 2018 on, by a code read. The PR changes results only for 2013 to 2017. Every changed parameter keeps its values from January 1, 2018, and the PR's property test holds credits after 2017 to the existing rules. The default dataset holds 2024 records, so we report no aggregate for 2013 to 2017. No microsimulation pair ran.
  • Next:
  • Axiom: Encode the premium tax credit for 2014–2017: annual applicable percentage tables, the 2013–2015 poverty guidelines and Form 8962 lines 3–5 (26 U.S.C. 36B; Rev. Procs. 2014-37, 2014-62 and 2016-24; 26 CFR 1.36B-3(g)) (PE-US #10846) TheAxiomFoundation/rulespec-us#2291 queued, a new pe-parity issue for the 2014 to 2017 applicable percentage tables, 26 CFR 1.36B-3(g), the 2013 to 2015 HHS guidelines and Form 8962 lines 3 to 5 for 2014 to 2016. It carries companion rows from this PR's YAML and leaves the residence rule and the 2017 guideline table to Encode the premium tax credit's poverty guideline by state of residence, with Alaska, Hawaii and movers (26 CFR 1.36B-1(h)), Form 8962 Worksheet 2 truncation, and no credit in the territories (PE-US PR 10860) TheAxiomFoundation/rulespec-us#2210. The signed encoder can't run yet: no corpus manifest names the three revenue procedures, an HHS notice before 2024 or Form 8962 instructions before 2025.

🤖 Generated with Claude Code

@MaxGhenis

Copy link
Copy Markdown
Contributor Author

The follow-ups from the merge audit now have issues:

🤖 Generated with Claude Code

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.

ACA applicable percentage uses the contiguous-states poverty guideline for Alaska and Hawaii

1 participant