PolicyEngine US domain knowledge
This skill is the US-specific layer for analysts using the policyengine package. It assumes
you already know the shared mechanics — pe.us.calculate_household, population Simulation,
reform dicts, datasets — from the policyengine skill; read that first for anything not
US-specific. For building new US variables or parameters, use policyengine-model-development.
For Medicaid / ACA / CHIP / Medicare, use policyengine-healthcare.
Verified against policyengine 4.21.0 / policyengine-us 1.764.6 (2026-07). Re-verify law-year values before reporting — the model updates continuously.
The six entities, and which program lives on which
US taxes and benefits are administered by different units, so the model has six entities. A program's variable is defined on exactly one of them, and reading it from the wrong entity is the most common US mistake.
| Entity | Plural key | What it groups |
|---|---|---|
| person | people | individuals |
| marital_unit | marital_units | a married couple (or a single person) |
| family | families | a nuclear family (Census-style) |
| tax_unit | tax_units | a tax-filing unit (a 1040) |
| spm_unit | spm_units | a Supplemental Poverty Measure resource-sharing unit |
| household | households | everyone at a physical address |
Which entity a program attaches to (verified — entity and definition period both matter):
| Program / measure | Variable | Entity | Period |
|---|---|---|---|
| Federal income tax | income_tax | tax_unit | year |
| EITC | eitc | tax_unit | year |
| Child Tax Credit | ctc | tax_unit | year |
| State income tax (aggregate) | state_income_tax | tax_unit | year |
| SNAP | snap | spm_unit | month |
| TANF | tanf | spm_unit | year |
| SSI | ssi | person | month |
| Net income | household_net_income | household | year |
| Total benefits | household_benefits | household | year |
| Total taxes | household_tax | household | year |
| In poverty (SPM) | in_poverty | spm_unit | year |
| Person in poverty | person_in_poverty | person | year |
| Is a child | is_child | person | year |
Rules of thumb: income tax and its credits (EITC, CTC, CDCC, education, and the aggregate
state_income_tax) are tax-unit variables; means-tested transfers keyed to a resource-sharing
unit (SNAP, TANF, school meals, housing) are spm_unit variables; SSI is a person-level
benefit (each individual's own SSI), rolled up into household_benefits. Poverty is measured on
the spm_unit. On a calculate_household result you read these as dot-attributes on the singular
entity — result.tax_unit.eitc, result.spm_unit.snap, result.person[0].ssi,
result.household.household_net_income.
Tax-unit roles and filing status
tax_unit={"filing_status": ...} accepts SINGLE, JOINT, SEPARATE, HEAD_OF_HOUSEHOLD,
SURVIVING_SPOUSE.
Three person-level role flags exist: is_tax_unit_head, is_tax_unit_spouse,
is_tax_unit_dependent. You usually do not set them for a simple family — minor children in
the people list are inferred as dependents automatically, the first adult is the head, and
JOINT designates the second adult as the spouse. Setting them wrong is worse than omitting them.
When you do need them: adult dependents. An adult in the people list is not auto-inferred
as a dependent. A 20-year-old student, a supported parent, or any qualifying relative must be
flagged is_tax_unit_dependent: True, or the model treats them as an independent unit member and
their credit vanishes. Verified — a HOH filer with a 20-year-old earning $4,000:
import policyengine as pe
def odc(dependent):
kid = {"age": 20, "employment_income": 4_000}
if dependent:
kid["is_tax_unit_dependent"] = True
return pe.us.calculate_household(
people=[{"age": 45, "employment_income": 40_000}, kid],
tax_unit={"filing_status": "HEAD_OF_HOUSEHOLD"},
household={"state_code": "CA"},
year=2026,
).tax_unit.ctc
assert odc(False) == 0 # adult not treated as a dependent -> no credit
assert odc(True) == 500 # $500 credit for other dependents (ODC)
Household calculation
state_code (not state_code_str) goes in the household dict. A low-income married couple with
two children, showing where each program lands:
import policyengine as pe
r = pe.us.calculate_household(
people=[
{"age": 35, "employment_income": 25_000},
{"age": 33, "employment_income": 0},
{"age": 8},
{"age": 5},
],
tax_unit={"filing_status": "JOINT"},
household={"state_code": "NY"},
year=2026,
)
assert round(r.tax_unit.ctc) == 4_400 # 2 children x $2,200 (OBBBA, 2026)
assert round(r.tax_unit.eitc) == 7_316 # federal EITC (tax_unit)
assert round(r.spm_unit.snap) == 7_364 # annual = sum of 12 monthly allotments
assert round(r.household.household_net_income) == 62_681
Here income_tax is about −$10,691 (refundable EITC + CTC exceed liability), so
household_net_income exceeds gross earnings. state_income_tax can likewise be negative when
refundable state credits exceed state liability.
Variables outside the default output columns (poverty flags, is_child, state credits, most
intermediate variables) must be requested with extra_variables=[...], or attribute access raises
AttributeError listing what is available:
import policyengine as pe
r = pe.us.calculate_household(
people=[{"age": 35, "employment_income": 25_000}, {"age": 8}],
tax_unit={"filing_status": "HEAD_OF_HOUSEHOLD"},
household={"state_code": "NY"},
year=2026,
extra_variables=["in_poverty", "person_in_poverty", "is_child"],
)
assert r.person[1].is_child == 1
r.spm_unit.in_poverty # SPM-unit poverty flag
r.person[0].person_in_poverty # person-level projection
Current law is post-OBBBA — check the year
The One Big Beautiful Bill Act reshaped 2026 federal law; pre-2026 summaries are wrong. Verified 2026 headline values (each is a live parameter read, not a memorized number):
<!-- verify -->from policyengine_us import CountryTaxBenefitSystem
p = CountryTaxBenefitSystem().parameters
# Child Tax Credit: $2,200 per child in 2026 (not $2,000, not the $3,600 ARPA figure)
assert p.gov.irs.credits.ctc.amount.base[0].amount("2026-01-01") == 2_200
# Standard deduction, 2026
std = p.gov.irs.deductions.standard.amount
assert std.children["SINGLE"]("2026-01-01") == 16_100
assert std.children["JOINT"]("2026-01-01") == 32_200
assert std.children["HEAD_OF_HOUSEHOLD"]("2026-01-01") == 24_150
# SALT cap, by filing status: $40,400 in 2026, scheduled to revert to $10,000 in 2030
salt = p.gov.irs.deductions.itemized.salt_and_real_estate.cap
assert salt.children["JOINT"]("2026-01-01") == 40_400
assert salt.children["SEPARATE"]("2026-01-01") == 20_200
assert salt.children["JOINT"]("2030-01-01") == 10_000
The SALT cap is scheduled: $40,000 (2025) → $40,400 (2026), rising ~1%/yr through 2029, then a hard revert to $10,000 in 2030. Always read the value at your simulation year before interpreting a SALT reform.
The parameter tree
Federal and state law live under gov.*. The top-level agencies (verified children of gov)
include irs, ssa, hhs, usda, hud, dol, ed, doe, states, local, aca,
simulation. The ones you reach for most:
| Path | Contains |
|---|---|
| gov.irs.* | federal income tax: credits.ctc, credits.eitc, deductions.standard, deductions.itemized.salt_and_real_estate, payroll, capital_gains |
| gov.usda.snap.* | SNAP: allotments, deductions, income limits, expected_contribution |
| gov.ssa.* | Social Security (social_security), SSI (ssi), sga, wage indices |
| gov.hhs.* | TANF (federal frame), Medicaid, CHIP |
| gov.hud.* | housing assistance |
| gov.states.{xx}.* | one node per state, {xx} = lowercase two-letter code (ca, ny, dc, …); state income tax lives at gov.states.{xx}.tax, state benefits under the agency (e.g. gov.states.ca.cdss) |
Bracket/scale parameters index the scale node directly — there is no .brackets segment
(gov.irs.credits.ctc.amount.base[0].amount). Discover paths by browsing .children.keys() on a
node or grepping the YAML tree in policyengine-us/policyengine_us/parameters/gov/; a guessed name
raises a ValueError listing the real children. See the policyengine skill for the three
reform-dict formats — they all share these paths.
State coverage
PolicyEngine-US models income tax for all 50 states and DC (gov.states has one node per state
plus dc); state_income_tax is computed everywhere and returns 0 in no-income-tax states (TX,
FL, WA, …). Many states also model refundable credits — state EITCs, CTCs, and rent/property
credits — as {state}_{program} variables. Request them with extra_variables:
import policyengine as pe
r = pe.us.calculate_household(
people=[{"age": 30, "employment_income": 20_000}, {"age": 4}],
tax_unit={"filing_status": "HEAD_OF_HOUSEHOLD"},
household={"state_code": "CA"},
year=2026,
extra_variables=["ca_eitc"],
)
assert round(r.tax_unit.eitc) == 4_427 # federal EITC
assert round(r.tax_unit.ca_eitc, 2) == 410.16 # CalEITC (state)
assert r.tax_unit.state_income_tax < 0 # refundable CA credits exceed liability
State-EITC variable names follow the pattern {xx}_eitc (ny_eitc, md_eitc, …); browse a
state's variables with [v for v in CountryTaxBenefitSystem().variables if v.startswith("ca_")].
SNAP: monthly, FPL-October, and the allotment hierarchy
SNAP is the US benefit whose timing most often surprises analysts. Everything below is verified against the model source.
SNAP is monthly. snap and its inputs have definition_period = MONTH. When
calculate_household reports an annual figure (like the $7,364 above), it is the sum of twelve
monthly allotments — not one annual calculation. That is why SNAP produces partial-year cliffs
that annual-only reasoning misses.
The FPL updates in October, on the federal fiscal year. snap_fpg (the poverty guideline used
for the income tests) reads the guideline dated October 1: for a month in Oct–Dec it uses that
calendar year's October figure; for Jan–Sep it uses the prior year's October figure. So a
household near the limit can fail the gross-income test for nine months and pass for the last
three, landing its annual SNAP at roughly a quarter of the full-year amount rather than at zero.
This is a real feature of the law, not a rounding artifact.
The allotment hierarchy (each name verified to exist as a monthly spm_unit variable):
snap = snap_normal_allotment + snap_emergency_allotment + dc_snap_temporary_local_benefit
|__ snap_normal_allotment = max( snap_min_allotment,
snap_max_allotment - snap_expected_contribution )
|__ snap_max_allotment (by household size and region)
|__ snap_expected_contribution = floor(snap_net_income) x 0.30
| |__ snap_net_income = snap_gross_income - deductions
| (snap_standard_deduction, snap_earned_income_deduction,
| snap_dependent_care_deduction, shelter, medical, child support)
| snap_gross_income = snap_earned_income + snap_unearned_income
|__ snap_min_allotment (small 1-2 person households)
is_snap_eligible = meets_snap_gross_income_test (<= 130% FPL, or categorical)
& meets_snap_net_income_test (<= 100% FPL)
& meets_snap_asset_test
& (meets_snap_categorical_eligibility can override the gross test)
The 30% expected-contribution rate is itself a parameter — gov.usda.snap.expected_contribution =
0.30 in 2026 — so it is reformable.
To debug one month, drop to the country package and calculate at a monthly period:
from policyengine_us import Simulation; Simulation(situation=...).calculate("snap_normal_allotment", "2026-11"), with .trace = True for the dependency tree. That direct-import surface is for tracing,
not for analysis results you report (see the policyengine skill).
Datasets and population analysis
US population runs use the certified populace_us_2024 bundle (and populace_us_2024_acs_local
for state/district grain). The mechanics — ensure_datasets, Simulation,
economic_impact_analysis, calculate_budgetary_impact, MicroSeries — are shared across countries
and documented in the policyengine skill; how Populace is built and calibrated is in the
policyengine-data skill. State and congressional-district breakdowns filter one national
dataset by its state_fips / congressional_district_geoid columns — there are no per-state or
per-district data files.