You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

CLIFF Index engine outputs, rule year 2025

These are the household-level files behind "How many households hit a benefits cliff?", the first results from the CLIFF Index's new microsimulation, published November 4, 2026. The Poverty Solution ran the households in the Census Bureau's American Community Survey through the Atlanta Fed's Policy Rules Database under 2025 rules. For each household, the files give its benefits, taxes and costs at today's earnings, after one earnings increase of $1,000 to $10,000, and at every $2,000 of earnings from $0 to past 450 percent of its poverty guideline. The files call that last set the ladder, and each $2,000 level a rung. The method is on the release's methodology page.

The geography is the Public Use Microdata Area (PUMA). The survey places each household in a PUMA, an area of at least 100,000 people, rather than a county. County in these files is derived from the PUMA, in two places.

  • int_puma_crosswalk__county_allocation splits each PUMA among its counties by their share of its 2020 population. Of the 2,462 PUMAs, 610 span more than one county. A county figure built from these files is an allocation: weight each household by allocation_factor.
  • county_fips and countyortownName in int_pums__prd_inputs name the county each household is priced in. That is the county with the largest share of its PUMA's population (in New England, that county's largest town).

Which run this is

The files come from one run of the pipeline, made October 1 to 4, 2026. The release was built from it.

Engine code commit aaedd3e of the CLIFF Index code repository
Rules Policy Rules Database, commit a402dc07 (August 18, 2026), applied for 2025
Survey American Community Survey five-year Public Use Microdata Sample (PUMS), 2020 to 2024
Release build commits d34479e, 8e18f0b and a1b2cfb of the same repository
This dataset tag run-aaedd3e

The release's one-increase and climb figures were built from these files by the code at the release-build commits. The release's other measures, such as which program decides a cliff and the employer-coverage cliffs, are defined on the methodology page. The code repository is private for now; write to info@thepovertysolution.com for access.

An earlier run, from August 2026 (commit dfdf249), was posted here first. Its files stay available at the tag run-dfdf249. The release was built from the October run only, and figures rebuilt from the August files differ from the release's. Of the 83 changes the release's run applies to the database (see Rules and assumptions), 24 were added after the August run. They include changes to the database's defaults, among them each member's own earnings in place of an equal split among a household's adults, and child care priced only once the parents earn. They also include child-care copay rules, two-parent TANF in Louisiana, New Hampshire and North Dakota, Texas's Medicaid limits for parents, SNAP's tests for elderly and disabled members, and the marketplace's contribution bands. The ladder also changed: it now holds other income at its reported amount, where the August ladder set it to zero.

Access is gated with automatic approval. Sign in to Hugging Face and request access on this page. The request shares your username and email with us; we use them only to tell you when the files change. Then:

hf auth login
hf download The-Poverty-Solution/cliff-index-engine-outputs --repo-type dataset --revision run-aaedd3e

The files

A household here is the Census Bureau's Supplemental Poverty Measure resource unit: the people who live together and share income and expenses. Its key is spm_unit_key.

The results cover every household whose total income was at or below 300 percent of the 2024 poverty guideline for its size, at any age: 2,782,236 households in the sample. A rate computed from these files is a rate among households at or below 300 percent.

Each household carries the survey weight of its reference person, pwgtp, and that person's 80 replicate weights, pwgtp1 to pwgtp80. They are in int_pums__prd_inputs.

File, under rule_year=2025/ Rows One row per Contents
int_universe__prd_actual_income 16,693,416 household and income_point Results at today's earnings (baseline) and after one increase: step (+$2,000), step_1000, step_3000, step_5000, step_10000
int_universe__prd_ladder_outputs 141,862,625 household and node_index Results at each rung of the ladder, where earnings are node_index times $2,000
int_universe__dollar_ladder 141,862,625 household and node_index The inputs at each rung, and the household's poverty guideline, fpl_line
int_pums__prd_inputs 6,670,898 household Inputs for every household in the sample, before the income limit: members, incomes, each member's share of earnings and of an increase, state (state_fips, stateAbbrev), PUMA (puma_key) and weights
int_universe__cohort_flags 6,670,898 household Two group flags: is_abawd_post_obbba (a member aged 18 to 64 who reports SNAP and no disability, in a household with no child under 14) and is_medicaid_parent_caretaker (a member aged 19 to 64 on Medicaid, in a household with a child under 19)
int_pums__prd_inputs_tax_units 3,229,676 tax filing unit Inputs for each filing unit, in households that file more than one return, before the income limit
int_universe__prd_actual_income_tax_units 7,268,490 filing unit and income_point Taxes, credits and health coverage computed for each filing unit
int_universe__prd_ladder_outputs_tax_units 85,563,581 filing unit and node_index The same, at each rung
int_puma_crosswalk__county_allocation 4,618 PUMA and county allocation_factor, the county's share of the PUMA's 2020 population
int_pums__spm_unit 15,209,501 person The household each person belongs to
int_pums__tax_units 15,209,501 person The filing unit each person belongs to
int_pums__persons 16,095,728 person The survey's person records
int_pums__households 4,142,895 housing unit The survey's housing records
int_cep__state_share 51 state Share of public-school students in schools with the Community Eligibility Provision

reference/rule_year=2025/ holds source tables the run read: the poverty guidelines, the survey's person and housing records, the database's county, Medicaid and SNAP tables, the Census tract tables behind the PUMA crosswalk, and state tables from CMS (Medicaid enrollment), USDA (SNAP), NCES (school enrollment), FRAC (Community Eligibility Provision students) and KFF (Medicaid expansion status).

Rules and assumptions

The rules. We leave the database's files as published at commit a402dc07 and apply 83 changes as it loads. Of these, 71 correct the database (they bring a rule into line with a law, a regulation or an agency's published schedule, or fix a defect in its code or tables), 17 carry a choice of ours, 8 of them within changes that also correct a rule, and 3 do neither. Each change is an entry in pipeline/engine/corrections_ledger.yml in the code repository.

Earnings. Each member has their own earnings, where the database splits a family's income equally among its adults. An increase is shared among the household's members aged 19 to 64 who aren't modeled as receiving SSI or SSDI, in proportion to what each earns, or equally if none earns yet. In a household with no such member, its other adults share it the same way. The shares are in int_pums__prd_inputs, as earn_share1 to earn_share12 and raise_share1 to raise_share12.

Child care. A child's care is priced only once every parent earns, where the database assumes all adults work.

Other income. Only earnings change from row to row. Other income stays at its reported amount, on the ladder as at today's earnings.

Enrollment. Every household eligible for a program is counted as enrolled, to measure the cliffs the rules set for everyone they make eligible. SSI and SSDI are the exceptions: they follow what households report.

Households that file more than one tax return. Where tax_unit_basis is true, the row's income and payroll taxes, tax credits, marketplace credits, Medicaid, job-based coverage and health care cost were computed for each filing unit and summed. The _tax_units files hold the per-unit results.

Rebuilding the release's figures

  1. Keep working-age households. The release counts households whose reference person is 18 to 64. The reference person is the member with the lowest sporder in int_pums__spm_unit, and age is agep in int_pums__persons. Join the two on serialno and sporder.

  2. Compute net resources for each row. This is what the household has to live on: income and benefits, minus taxes and what it pays out of pocket for child care and health care.

    net resources =
        income + income_gift + income_investment + income_child_support
      + value_snap + value_wic + value_schoolmeals + value_tanf + value_ssi + value_ssdi
      + value_eitc_fed + value_eitc_state + value_ctc_fed + value_ctc_state
      + value_cdctc_fed + value_cdctc_state
      - tax_income_fed - tax_income_state - tax_fica
      - greatest(exp_childcare - value_ccdf - value_headstart - value_earlyheadstart - value_prek, 0)
      - greatest(exp_healthcare - value_aca - value_medicaid - value_employerhealthcare, 0)
    

    income is the household's earnings on that row. Today's earnings are income on the baseline row. Every household has a value for net resources at every income point.

  3. One increase. In int_universe__prd_actual_income, compare each increase row with the household's baseline row. The household hits a cliff if net resources fall, by any amount.

  4. The climb. The first step runs from the baseline row to the step row, the $2,000 increase. It counts wherever it falls. After it, read the ladder steps in int_universe__prd_ladder_outputs whose lower rung is at or above today's earnings. Each compares a rung with the one $2,000 above it. For a household earning $31,000, the steps run from $31,000 to $33,000, then $32,000 to $34,000, then $34,000 to $36,000, and so on. A ladder step counts only when its upper rung is at or below 4.5 times fpl_line, in every state. fpl_line is the household's 2025 poverty guideline, in int_universe__dollar_ladder; join on spm_unit_key. A household counts once, at its first cliff. The ladder itself runs to the first rung at or above 4.5 times the guideline, 4.8 in the District of Columbia and 5.8 in Vermont, so its top rows sit above the line the release reads to.

  5. Weight. Sum pwgtp for counts and shares. Standard errors use the 80 replicate weights, by successive difference replication: the variance is 4/80 times the sum, over the replicates, of the squared difference between the replicate estimate and the full-sample estimate.

The post's first table prints the figures a rebuild should match: the share of households that hit a cliff at each of the five increases, and on the climb. The methodology page's Who is counted gives the number of working-age households, in the sample and weighted. The exact code for these steps is in pipeline/marts/cliff/ at the release-build commits. Which program decides a cliff, the programs that decide one only together, and the employer-coverage cliffs are computed separately from that code; the methodology page defines them.

Resource columns

The files' own net_resources column is a different measure from the release's net resources. It is the database's NetResources: income and benefits, minus taxes, minus all eight of its expense lines, including food, housing and transportation. It is null where an expense line is missing.

int_universe__prd_ladder_outputs also carries disposable_resources (income and benefits minus taxes, with no expense taken out), disposable_resources_ex_healthcare, prd_modeled_expense_basket_total (the eight expense lines summed) and alice_survival_total (the eight lines at the United for ALICE Survival Budget's amounts).

Dollar amounts are annual. Incomes are carried from 2024 to 2025 dollars by the Bureau of Labor Statistics' R-CPI-U-RS. Reported housing cost stays in 2024 dollars.

Limits

Some households fall outside the range the model handles. These are households with more than six adults or six children (Technical Manual, p.6), or with no member 19 or older. For these households net_resources is null. The release keeps them in its counts, which lowers the count at the $2,000 increase and raises it on the climb (methodology, Assumptions and limits). To leave them out, drop the households whose net_resources is null in int_universe__prd_actual_income.

Job-based coverage is priced once per filing unit. On rows where tax_unit_basis is true, a household that files several returns can carry several plan prices, which overstates its health care cost. How far this moves the cliff counts isn't measured. The schema file at pipeline/intermediate/cliff/schemas.py lists this and the other known differences on those rows.

Of the 3 changes that neither correct the database nor carry a choice, one is an error of ours that we haven't yet fixed. It pays TANF in Louisiana to married couples whom the state pays only when a parent is incapacitated or caring for a disabled family member, so it overstates TANF there. A second may do the same in New Hampshire, where the state's current rules don't settle whether such couples are paid. The third no longer changes anything, since the database has removed the line it fixed.

We know of at least 18 issues in the rules the model applies for 2025 that we haven't corrected. Each one's record says which way it moves the benefit, where that is known. None has been run through the counts. The list of changes and the list of open issues are available on request.

Counting everyone eligible as enrolled raises the count for some programs and not for others. At the $2,000 increase it pushes the count up for SNAP and cash assistance, and barely moves or lowers it for child-care subsidies. The other programs aren't checked, so overall it could go either way (methodology, Programs and take-up).

Reported earnings can be off by more than a $2,000 increase. Any one household's true distance from a cliff is uncertain. Counts from these files are estimates for households as a group.

The methodology's Assumptions and limits section lists the rest, among them that assets are set to zero, immigration status isn't used, and job-based coverage is read as an offer.

License and citation

The files are licensed CC-BY-4.0. Credit both the CLIFF Index and the Policy Rules Database.

The Poverty Solution (2026). CLIFF Index engine outputs, rule year 2025.
Hugging Face dataset The-Poverty-Solution/cliff-index-engine-outputs, revision run-aaedd3e.

Cite the database as the Federal Reserve Bank of Atlanta's Policy Rules Database, commit a402dc07 (August 18, 2026). The CLIFF Index runs it with our 83 changes. The database's own code is licensed GPL v3, and our Python pipeline code Apache-2.0.

Sources

  • U.S. Census Bureau. American Community Survey, 2020-2024 five-year Public Use Microdata Sample.
  • Federal Reserve Bank of Atlanta. Policy Rules Database, commit a402dc07 (August 18, 2026). https://github.com/Research-Division/policy-rules-database/tree/a402dc076a8064d3b72ae8d2793992f46c5c9cc3. A collaboration with the National Center for Children in Poverty; originally developed by Elias Ilin and Ellyn Terry.
  • U.S. Department of Health and Human Services. Poverty Guidelines.
  • United for ALICE. Survival Budget figures, as carried in the Policy Rules Database.

Questions and corrections: info@thepovertysolution.com.

Downloads last month
572