Methodology

The method behind "How many households hit a benefits cliff?", the first results from The Poverty Solution's new CLIFF Index microsimulation.

What the index measures

The CLIFF Index estimates how many households hit a benefits cliff when their earnings rise. It applies federal and state program rules, from the Atlanta Fed's Policy Rules Database, to the households in the Census Bureau's American Community Survey. For each household it computes the benefits and taxes the household would have under 2025 rules, at today's pay and after an earnings increase. Because the index applies the rules to each household rather than counting what households report receiving, its estimates differ from program enrollment counts. The index finds how many households an increase would leave worse off, while surveys of people receiving benefits are the evidence on how many limit their income or pass up an earnings increase to keep their benefits.

Who is counted

Throughout, a household means the Census Bureau's Supplemental Poverty Measure resource unit: the people who live together and share income and expenses. It is broader than a tax filing unit or a Census family, adding cohabiting partners, foster children under 22 and unrelated children under 15. Because a benefit supports everyone who shares resources, each cliff is measured over this unit. The Department of Health and Human Services' Office of the Assistant Secretary for Planning and Evaluation (ASPE) uses the Census household instead in its marginal tax rate series (Brief 5, p.1). People in group quarters are not counted.

The universe is every household whose total income was at or below 300 percent of the 2024 poverty guideline for its size, in the survey's 2024 dollars. Reported public assistance counts toward that income for this test only. Within it, the households counted are those of working age: those whose reference person, the unit's first-listed member, is 18 to 64.

The sample holds 1,712,757 such households. Each carries the survey weight of its reference person, who is the householder in the householder's own unit. Every count and share is weighted, so the sample stands for about 42.5 million working-age households at or below 300 percent of the poverty guideline.

Data and rules

The households come from the American Community Survey five-year Public Use Microdata Sample (PUMS) for 2020 to 2024, with incomes carried from 2024 to 2025 dollars by one price index, the Bureau of Labor Statistics' R-CPI-U-RS. The survey places each household in a Public Use Microdata Area rather than a county, so each is priced in the county that holds the largest share of its area's population (in New England, that county's largest town).

The rules come from the Policy Rules Database at commit a402dc07 (August 18, 2026). The database is a collaboration between the National Center for Children in Poverty and the Federal Reserve Bank of Atlanta, originally developed by Elias Ilin and Ellyn Terry. The Atlanta Fed's own CLIFF tools, CLIFF Snapshot, CLIFF Dashboard and CLIFF Planner, run on the same database (Technical Manual, p.3). The CLIFF Index is The Poverty Solution's, built on the database with our changes to it.

The database has carried 2026 rules for most programs and taxes since its February and March 2026 releases; we apply 2025 rules, the most recent year that is complete and settled, and the latest year our corrections are verified for. We leave the database's files as published at that commit and apply 83 changes as it loads:

• 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;
• 3 do neither.

Some of the choices adapt the code to run survey households in large batches, and some pick which of two schedules in force during 2025 to carry for the whole year. Others change the database's defaults to use what the survey records, among them:

• each member's own earnings, where the database splits a family's income equally among its adults;
• child care priced only once the parents earn, where it assumes all adults work;
• each household's reported housing cost, where it uses the United for ALICE budget's rent.

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 also 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.

What 2025 means differs by program. SNAP's 2025 income limits and benefit amounts are those of federal fiscal year 2025, October 2024 to September 2025, and taxes and marketplace coverage follow calendar 2025. Where a state changed a child-care schedule or began free school meals for all students during 2025, the model applies one rule to the whole year: in some states the one in force from midyear, in others the earlier one. New Mexico, for one, removed the income limit on its child-care subsidies on November 1, 2025 (New Mexico Early Childhood Education and Care Department). The model keeps the earlier rule, so on the climb some New Mexico households hit a child-care cliff the state has since removed.

Some costs in the database come from earlier years. We run the database's United for ALICE "survival" budget setting, under which its code carries them to 2025 at fixed yearly rates:

• child-care prices, from 2024, at the database's 2 percent inflation assumption;
• premiums for a job's plan and out-of-pocket health costs, from 2024, at 1.3 percent;
• what school meals are worth, from 2023, at 3.5 percent;
• what WIC's food package is worth, from 2018, at 3.5 percent.

The 1.3 and 3.5 percent rates are ALICE's own growth rates with the database's 2 percent inflation assumption taken out. In the test of whether a job's plan rules out marketplace credits, the database carries one person's premium at 2 percent. Our correction for the family-coverage rule in force since 2023 tests a family on a family premium that the code doesn't carry forward, so it stays at its 2024 level.

Program rules come from the database's 2025 rows, except the state supplements to SSI, which it has only for 2022. Marketplace premiums are 2025 figures, and the 2025 premium-credit schedule includes the temporarily enhanced credits, which lapse after 2025, so these results describe 2025 law only.

What counts as a cliff

A household hits a benefits cliff when an earnings increase leaves it financially worse off than before. We count its income and benefits, minus its taxes and what it pays out of pocket for health care and child care: what it has to live on, its net resources. The measure is modeled on the Census Bureau's Supplemental Poverty Measure. The benefits counted are SNAP, WIC, school meals, TANF, SSI, SSDI, and the federal and state earned income, child and child-care tax credits. Health coverage and child-care subsidies count through the out-of-pocket costs they cover.

A household's net resources fall when the benefits it loses, the extra taxes it pays and any rise in its out-of-pocket costs add up to more than the increase. In the field's terms, its effective marginal tax rate, the share of an increase lost to all three, is above 100 percent. Any fall counts, however small, and a fall counts whatever causes it, such as one program's benefit ending at its income limit, or several benefits phasing out together with the tax on the increase.

The marginal tax rate chart. The rate is charted for the next $2,000 increase, with households grouped in 25-point bands of what they earn today, as a percent of the poverty guideline. Each band has two lines: the average across all its households, and the median, the rate of the typical household.

One increase from today's pay. The model raises a household's pay once, by $1,000, $2,000, $3,000, $5,000 or $10,000, and compares each size separately with today's pay. The increase is the household's, added to the earnings it reports for the year. It 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. At the $2,000 increase, 3.4 percent of households hit a cliff. These are the households right next to a cliff. On the narrower measure set out under Earlier national counts of cliffs, which leaves out health coverage and child care, that share would be 1.5 percent.

Which program decides a cliff. A program decides a cliff when its loss is the only one whose removal, by itself, would leave the household no longer worse off. The test removes each of twelve losses in turn: health coverage, child care, SNAP, WIC, school meals, TANF, SSI, SSDI, the earned income, child and child-care credits, and income and payroll taxes together. Each loss is added back with every other change left as computed; the rules aren't rerun, so interactions between programs stay as they were.

At the $2,000 increase, one program decides the cliff for about 2 in 3 of the households that hit one. For them, the program that decides is also the largest loss. PHI, a national organization on the direct care workforce, uses the largest loss as its measure in Tipping Point (p.6), so where it counts the same losses, its measure and ours name the same program. For the other third, a largest-loss rule still names a program, though undoing that loss alone either isn't the only fix or isn't enough.

The share with one deciding program falls as the increase grows, from 76 percent at $1,000 to 55 percent at $3,000 and 29 percent at $10,000; a larger increase can take in more of a household's losses at once. Which program decides most often shifts with it: with the employer-coverage cliffs counted (see Programs and take-up), health coverage at $1,000 to $3,000, SNAP at $5,000 and child care at $10,000; with them set aside, SNAP up to $5,000 and child care at $10,000.

Part of a stack. Where no single program decides, a program is part of the stack when adding its loss to some of the household's other losses leaves the household worse off, and those other losses alone wouldn't. A household counts under every program in its stack, so the stack counts overlap and can't be added. Every one of these households also pays more Social Security and Medicare tax on the increase. Income and payroll taxes together are part of the stack for 415,000 of the 463,000 households where no single program decides. Because taxes are in about 9 in 10 stacks, they're left out of the chart in Which programs drive cliffs.

The climb: $2,000 at a time

One increase shows what that increase does on its own. The climb shows what happens to a household as its pay keeps rising: pay rises $2,000 at a time from today's pay until it reaches 450 percent of the poverty guideline, and each level is compared with the one $2,000 below it. At each step the same test applies as for one increase: any fall in what the household has to live on counts as a cliff. A household counts once, at its first cliff.

The first step adds $2,000 to today's pay, exactly as the one-time $2,000 increase does. After it, each step runs between fixed earnings levels $2,000 apart, the same for every household: $0, $2,000, $4,000 and so on. The climb moves from today's pay onto that grid at the first level at or above it, so for a household with earnings the first two steps overlap. 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; in the post's first table, the row "Within $12,000" covers the steps up to $42,000. Only earnings change along the climb; other income stays at its reported amount.

On the climb, 49.3 percent of households hit a cliff before their earnings reach 450 percent of the guideline. The other 50.7 percent, about 21.5 million, reach that point without one. In every state, only cliffs below that line are counted. The line sits above every state's 2025 limit for children's Medicaid and CHIP; New York's, the highest, is 405 percent. In the District of Columbia and Vermont, though, child-care limits run higher (460 and 575 percent), so the child-care exit there falls above the line and isn't counted.

Why $2,000 steps on the climb. The climb takes the $2,000 step that ASPE gave each household in its national count. PolicyEngine gave $2,000 to each adult instead, so a household with two adults got $4,000. With the employer-coverage cliffs counted (see Programs and take-up), smaller steps can only find as many cliffs or more: if resources fall across a $2,000 step, they fall across at least one smaller step inside it. And a loss that comes $500 into a $2,000 step can leave the household worse off at that point and better off by the step's end, a cliff only a smaller step finds. Larger steps find fewer. In a one-in-ten sample, $4,000 steps find a cliff before 450 percent for 33.4 percent or 33.1 percent of households, depending on where the steps start, against 49.1 percent with $2,000 steps. These counts include the employer-coverage cliffs; with them set aside, smaller steps could find fewer cliffs as well as more.

The SNAP-line comparison. Among the states that expanded Medicaid, SNAP's gross income line under the rules the model applies for 2025 is 200 percent of the guideline in 23 states and the federal 130 percent in 9 (see Where states draw the lines). Nearly every state with a 200 percent line expanded Medicaid, and comparing only expansion states keeps SNAP's line and Medicaid's from being mixed together. The model's 2025 rules put Alaska at the federal 130 percent, its line until July 2025. New Hampshire and New York are in neither group, since in the model each state's line depends on the household: New Hampshire's on whether it has a member under 19, a rule the database's code sets though USDA's 2025 chart lists 200 percent for all households, and New York's on whether it has child-care costs.

The two groups differ in who lives in them, so each group's rates are reweighted to one mix of households: by income in 25-point bands of the guideline, family size, and whether there is a child under 18. After reweighting, 2.9 percent of households in the states at 200 percent hit a cliff at the $2,000 increase, against 5.1 percent in the states at 130. Before reweighting, 2.8 percent hit one against 5.2 percent, so 93 percent of the gap remains after reweighting.

On the climb, at the same mix, 10.5 percent of households in the states at 200 percent hit a cliff where SNAP's benefit stops before their earnings reach 450 percent of the guideline, against 21.5 percent in the states at 130. In the states at 200 percent, 3.5 percent hit it at the gross income line itself, and 6.9 percent elsewhere, such as where the benefit runs out below the line or at the net income test. Counting a cliff from any program, 47.4 percent of households in the states at 200 percent hit one before 450 percent, against 51.5 percent in the states at 130.

The states in each group still differ in other rules, in living costs and in other traits of their households, so these are associations across states. Seven of the 23 states at 200 percent, and none at 130, give every student free school meals, so a child's free meals there don't end when SNAP does. Without those seven, at the same mix, 3.0 percent of households hit a cliff at the $2,000 increase, against 5.1 percent; in the other 16 states at 200 percent, 5.9 percent hit SNAP's cliff at the line. In three of the states at 200 percent, California, Colorado and North Dakota, the model also applies a net income test at 100 percent of the guideline to households with no elderly or disabled member; without those three, 2.7 percent hit a cliff before reweighting, against 5.2 percent.

Programs and take-up

The model covers, at 2025 rules: SNAP, WIC and school meals; Medicaid and CHIP for adults and children, and marketplace premium credits; TANF; child-care subsidies from the Child Care and Development Fund, Head Start, Early Head Start and state pre-K; SSI and SSDI; the federal and state earned income, child and child-care tax credits; and federal and state income tax and the worker's share of Social Security and Medicare tax. Housing vouchers and local rental programs are switched off, and energy assistance and summer meals aren't in the model.

We count every household eligible for a program 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, and SSI for members under 18 or 65 and over is carried at its reported amount.

Not everyone eligible enrolls. About 88 percent of eligible people received SNAP in fiscal year 2022 (Cunnyngham, Table B.1a note), about 62 percent of eligible citizen adults aged 19 to 64 were enrolled in Medicaid in 2016-2019 (Abdus), and about 56 percent of eligible individuals received WIC in 2023 (Kessler and others, Key Findings).

For SNAP and cash assistance, counting everyone eligible pushes the count up. With SNAP counted only where households report receiving it, 2.8 percent of households hit a cliff at the $2,000 increase, and with cash assistance counted only where it's reported, 3.2 percent do. Both are below 3.4 percent.

Counting everyone eligible for the Child Care and Development Fund's subsidy barely moves the count, or lowers it. Because the survey doesn't record child-care subsidies, the check takes that subsidy away altogether. Among households with earnings, that barely changes the count, from 4.1 percent to 4.0 percent at the $2,000 increase: about 150,000 cliffs where the subsidy ends go away, and about 128,000 appear where care would start without it. Among households without earnings, taking the subsidy away raises the count, since in the model a first job brings a child-care cost that the subsidy would otherwise pay.

Health coverage, Head Start, pre-K and the other programs aren't checked this way, and one program's effect can't be added to another's, so counting everyone eligible could raise or lower the count overall. Altig and others, a national study on other bases, tested the same choice. For the bottom fifth of households by lifetime resources, with heads aged 20 to 69, assuming that everyone eligible enrolls puts the median current-year marginal tax rate at 40.3 percent; matching each program's national take-up rate puts it at 29.7 percent (pp.17-18).

Medicaid, CHIP and marketplace eligibility are tested on modified adjusted gross income, the income measure those programs use. The SNAP shelter deduction uses each household's reported housing cost, kept in the survey's 2024 dollars. TANF is computed under each state's rules, so the public assistance a household reports isn't added on top. Cliffs are measured over a full year after every program has adjusted to the new income, so rules that hold benefits for a time after an increase, such as the 12 months of continuous eligibility that children's Medicaid and CHIP have had since 2024, aren't counted.

Child care is priced from the United for ALICE Survival Budget, the basis of the county-by-county Minimum Household Budget that the database draws on (Technical Manual, p.5); the model uses the county's daily rate for family child care, full time for children under 5 and half time for children 5 to 12. A child's care is priced only once every parent earns, or, where no parent lives with the child, once the householder and any spouse or partner earn. Because an increase goes to the adults who already earn, and is shared equally only where no adult of working age earns, a parent who doesn't work today starts work in the model only in a household where no working-age adult does. Where one of two parents works, the model never prices their children's care.

Child-care subsidies follow the database's rules for families already enrolled: a family keeps its subsidy until its income passes the top of its state's copay schedule. The Child Care and Development Fund is a capped block grant, so counting everyone eligible departs from actual receipt furthest here. Of children eligible under federal rules, 15 percent received a subsidy in an average month of fiscal year 2021, and of those eligible under their state's rules, 22 percent did (Chien, ASPE Data Point, pp.1, 3).

The survey records health coverage through a job but not what the plan costs the worker. The model reads that coverage as an offer and prices the plan at the worker's share of its premium from the database's ALICE figures, which equal the state's 2024 average for single, plus-one or family coverage in the federal survey of employers, MEPS-IC. As the Atlanta Fed's tools do, it assigns "the most cost-effective bundle for a household's health care needs" (Technical Manual, pp.5-6, 225): for some of these households, Medicaid, CHIP or a marketplace plan with a premium credit, which an earnings increase can end. Where that coverage goes to the member who reports the job's plan, the survey says the member already has the plan and would lose no coverage; where it goes to someone else, the loss can be real.

Employer-coverage cliffs. These are the cliffs that removing the loss of health coverage alone would end, in three kinds of household:

1. every child under 19 reports coverage through a job, and none reports Medicaid or CHIP;
2. an adult aged 19 to 64 reports coverage through a job, and the model gives an adult Medicaid;
3. the increase moves the household from a marketplace plan to the job's plan, with no one losing Medicaid.

A cliff counts here whether or not another loss alone would also end it, so these cliffs outnumber the ones health coverage decides. At the $2,000 increase, health coverage decides about 438,000 cliffs, and about 371,000 of them are employer-coverage cliffs; adding the employer-coverage cliffs that another loss alone would also end, there are about 523,000 in all.

With the employer-coverage cliffs set aside, 2.2 percent of households hit a cliff at the $2,000 increase instead of 3.4 percent, and 40.3 percent, about 17.1 million, hit one on the climb instead of 49.3 percent. Households with a child under 18 become 8.1 times as likely as those without one to hit a cliff at the $2,000 increase, against 3.4 times with the employer-coverage cliffs counted. With them set aside, SNAP decides the most cliffs among the households still worse off at the $2,000 increase, about 220,000, and health coverage falls to fourth, behind school meals and child care.

Past the climb's first step, the data don't record which coverage ends, so for the first two kinds of household a step's cliff is set aside whenever removing that step's loss of health coverage would by itself end it. The third kind is set aside at the first step only.

Uncertainty

Each rate carries a standard error from the reference person's 80 replicate weights, by successive difference replication, the method the Census Bureau uses for its published American Community Survey estimates. A rate is flagged when its coefficient of variation, the standard error divided by the rate, is above 0.30, or when it is 0 or 100 percent and its reliability can't be assessed. For each of the 51 states, neither the figure for the next $2,000 increase nor the climb figure is flagged as unreliable. The program breakdowns carry no standard errors. Nationally, the margin of error is small: at 95 percent confidence, 49.2 to 49.4 percent of households hit a cliff on the climb, and 3.40 to 3.47 percent with the next $2,000 increase. That margin covers sampling error only. The assumptions listed below can move the counts much further, as setting aside the employer-coverage cliffs does (see Programs and take-up).

Assumptions and limits

We haven't measured which way these move the counts:

• No program's work requirements are modeled, so every household is counted as meeting any that its programs set.
• The child-care exit sits at the top of each state's copay schedule.
• In states that set child-care subsidy limits as a share of state median income, the database's 2025 limits rest on a newer estimate of state median income than the states' own 2025 figures where we've checked them; the database's July 2026 update to that estimate raised most of those limits by about 4 to 7 percent, so the child-care exits there come later on the climb.
• Losing Medicaid costs a household only what it then pays out of pocket for the coverage the model gives it instead; Medicaid's spending per enrollee drops out.
• Medicare isn't modeled, so members who would have it are priced on the model's other coverage.
• Earnings and benefits are measured over a year, though some programs test a shorter period.
• Other income and household members stay the same as earnings rise.
• Households are counted by the earnings they report, which can be off by more than a $2,000 increase. Any one household's true distance from a cliff is uncertain, so the model can't say which particular households are near one; the counts are estimates for households as a group.

Earlier national counts of cliffs

Earlier national counts stand on other bases, and the basis moves the count. A narrower measure counts only income, taxes and tax credits, and the benefits paid in cash or food, leaving out health coverage, child-care help and what households pay out of pocket for health care and child care. On it, 1.5 percent of working-age households at or below 300 percent of the guideline would be worse off after a $2,000 increase, against 3.4 percent on the full measure.

Like this count, ASPE's figures show households with children more than three times as likely to hit a cliff after a $2,000 increase.

Data and citation

For the data behind these counts, or to work with us on a proposal, write to info@thepovertysolution.com. 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.

Sources

• Chien, Nina. Estimates of Child Care Subsidy Eligibility & Receipt for Fiscal Year 2021. Data Point. Office of the Assistant Secretary for Planning and Evaluation, U.S. Department of Health and Human Services, September 2024.
• Cunnyngham, Karen. "Empirical Bayes Shrinkage Estimates of State Supplemental Nutrition Assistance Program Participation Rates: Fiscal Year 2020 and Fiscal Year 2022." Prepared by Mathematica for the U.S. Department of Agriculture, Food and Nutrition Service, February 2025.
• Abdus, Salam. "Trends in Medicaid Take-Up Among Eligible Adults After the Affordable Care Act Medicaid Expansions: 2014-2019." Medical Care Research and Review 82, no. 1 (2025): 100-106. https://doi.org/10.1177/10775587241273429
• Kessler, Courtenay, Andrew Bryant, Kate Munkacsy, Sophia Maxson, Daniel Ressler, Rashi Saluja and K. Farson Gray. National- and State-Level Estimates of WIC Eligibility and Program Reach in 2023. Final Report, Volume I. Westat for the U.S. Department of Agriculture, Food and Nutrition Service, December 2025.
• Altig, David, Alan J. Auerbach, Laurence J. Kotlikoff, Elias Ilin and Victor Ye. "The Marginal Net Taxation of Americans' Labor Supply." NBER Working Paper 27164. May 2020, revised July 2023. http://www.nber.org/papers/w27164
• PHI, with LeadingAge Virginia and PositiveAge. Tipping Point: Exploring the Risk of Benefits Cliffs Among Direct Care Workers in Virginia. July 2025.
• Office of the Assistant Secretary for Planning and Evaluation (ASPE), U.S. Department of Health and Human Services. Marginal Tax Rate Series, March 2019: Chien, Nina, and Suzanne Macartney, "What Happens When People Increase Their Earnings? Effective Marginal Tax Rates for Low-Income Households," Brief #2; Giannarelli, Linda, Kye Lippold, Elaine Maag and others, "Estimating Marginal Tax Rates Using a Microsimulation Model: Technical Appendix," Brief #5. https://aspe.hhs.gov/marginal-tax-rate-series
• Ghenis, Max, and Nikhil Woodruff. "How Would Reforms Affect Cliffs?" PolicyEngine, February 2, 2023.
• 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. Policy Rules Database Technical Manual, updated July 7, 2024.
• New Mexico Early Childhood Education and Care Department. Preamble and Explanatory Statements for New Mexico Administrative Code Title 8, Chapter 9, Parts 3, 4, and 5. October 2025.
• U.S. Census Bureau. American Community Survey, 2020-2024 five-year Public Use Microdata Sample.