Methodology

Overview

The CLIFF Index assigns a composite severity score (0–100) to every U.S. county and state, measuring how severely benefits cliffs affect working families in each geography. The score is decomposed into four dimensions — population exposure, cliff severity, economic vulnerability, and state policy environment — computed from 17 metrics drawn from 12 federal and state data sources.

The index was built to fill a gap: existing tools in the field (Atlanta Fed CLIFF Calculator, state-level benefit simulators, APHSA pilot dashboards) help individual families or track state policy, but none measure geographic severity at the community level. The CLIFF Index answers: which places are hit hardest, which programs drive the biggest losses, and how state policy interacts with local conditions.

What it is not: The CLIFF Index does not predict individual outcomes, does not model SSI/SSDI disability benefit formulas, and does not capture informal benefits, employer-provided coverage, or county-level policy variation below the state level. Scores represent community-level structural conditions, not the experience of any individual family. The index measures where cliffs are most severe — it does not measure whether families are actually experiencing cliff losses, which would require administrative earnings data that does not exist at scale.

The CLIFF Score

Each county receives a score from 0 (least severe) to 100 (most severe), computed in four steps.

Step 1: Percentile-rank each metric

Each of the 17 metrics is percentile-ranked across all 3,144 counties. A county at the 85th percentile on a given metric scores worse than 85% of U.S. counties on that measure. This follows the CDC Social Vulnerability Index (SVI) approach: distribution-free, outlier-resistant, and interpretable. The EIG Distressed Communities Index and County Health Rankings use the same method.

Step 2: Sum percentile ranks within each component

Percentile ranks for metrics within a component are summed, then the sum is re-ranked across all counties to produce a 0–100 component score. All metrics within a component carry equal weight.

Step 3: Aggregate three county components via power mean

The power mean with p=0.5 sits between the geometric mean and the arithmetic mean. It maintains non-compensability (a low score in one dimension cannot be fully offset by high scores elsewhere) without being as aggressive as the geometric mean, which crushes counties where any single component is low.

The 70/30 blend with the maximum component ensures that extreme severity in any single dimension gets partial recognition. Without it, a county with Exposure=8, Severity=88, Vulnerability=12 would score near zero despite extreme cliff mechanics. The 30% max-component term lifts imbalanced counties (e.g., Hennepin County MN: 8→35) without materially affecting balanced ones (e.g., Wayne County MS: 92→89). Component scores are floored to 1 before entering the power mean to prevent a zero value from collapsing the formula.

Step 4: Apply policy multiplier

Policy is applied as a multiplier (0.75–1.25), not as an equal component. Federal benefit rules create cliffs in every state regardless of state policy. Good state policy can moderate cliff severity by up to 25%, but it cannot eliminate structural cliffs. This means a state cannot achieve a zero cliff score through policy alone. The multiplier range was calibrated empirically to produce an 18-point median gap between Medicaid expansion and non-expansion states, consistent with observed outcome differences.

Tier system

Near-uniform distribution is expected and correct for a percentile-ranking methodology. No artificial clumping or gaps.

Component 1: Exposure

How many people are positioned in the income range where benefit phase-outs occur, and what is the composition of that population. Four metrics capture the size, relative dominance, family composition, and program enrollment of the cliff-affected population.

Design decisions

• E2 counterbalances E1. Cliff zone share (E1) correlates with poverty (r=+0.54). The cliff-to-poverty ratio (E2) anti-correlates with poverty (r=−0.53). Together they prevent the exposure component from functioning as a poverty index.
• E3 is cliff-zone-specific. The single-parent rate within the cliff zone can differ substantially from the county-wide rate. Wayne County MS: county-wide 16.5%, cliff-zone 8.8%. Henry County GA: county-wide 15.3%, cliff-zone 20.6%.
• E4 uses equal program weights. Each program creates a real cliff. Weighting by cliff severity would cross into the Severity component. Exposure measures breadth of enrollment, not magnitude of each cliff.
• Dropped metrics: Working rate (nearly tautological in the cliff zone), child share (already captured via prevalence-weighted METR in Severity), disability rate (SSI/SSDI cliff mechanisms are not modeled in the benefit engine — measuring prevalence of an unmodeled population is a half-measure; disability rate is retained as a display metric but not scored).

Component 2: Severity

How severe benefit reductions are when families cross program thresholds. Six metrics measure the steepness of benefit loss, the cliff depth (how far families fall), how many programs reduce benefits at the same income point, whether housing and childcare costs amplify the loss, and whether families losing health coverage have alternatives.

Design decisions

• S1 uses structural METR at the county level, not participation-weighted. The peak is the peak. A family enrolled in CCDF + SNAP + Medicaid faces the full structural rate regardless of how many other families are enrolled. At the state level, S1 uses SP2K participation-weighted METR computed from the state’s own benefit landscape with population-weighted average housing costs. Enrollment breadth is captured in Exposure (E4).
• S6 balances the expansion gap. Expansion states show deeper benefit valleys (Medicaid creates a deeper cliff) but also have stronger coverage safety nets. Including uninsured rate ensures non-expansion states’ lack of coverage fallback is reflected in severity.
• S6 (Uninsured Rate) is a balancing metric. Without it, Medicaid expansion states would score higher on Severity because they have deeper benefit valleys (more Medicaid to lose) and more program stacking points. S6 corrects this by reflecting that non-expansion states lack coverage alternatives when families lose Medicaid. This is arguably a policy outcome rather than a cliff mechanism — it is retained because removing it would tilt the expansion gap in the wrong direction.
• Dropped metrics: Absolute FMR (created 37% metro/rural severity gap from dollar amounts), weighted average METR (captured between peak and valley), Medicaid cliff dollar value (enrollment already weighted into PW METR), EITC benefit intensity (gradual phase-out, not a hard cliff), healthcare shortage overlap (HRSA shortage designations are binary flags producing a coarse 0–3 integer — too blunt for county-level index construction).

Component 3: Economic Vulnerability

Whether the local economy positions families where benefit phase-outs are most likely to trigger, and whether advancement past them is realistic. Three metrics measure occupation mix, educational credentials, and whether local industries pay enough to clear benefit phase-outs.

Supplementary metric: Cliff-Wage Employment. A display-only metric measures the share of ALL county/state employment (not just cliff-zone workers) in industries paying $12–$18/hr (BLS QCEW). This captures the structural pipeline into the cliff zone — how many workers the local economy funnels toward cliff-sticky wages. This metric is displayed alongside V1–V3 but is not included in the scored Vulnerability composite.

Design decisions

• V2 (education) is tied to the advancement story, not general poverty. Education alone is a standard economic hardship metric. In this context, it measures whether cliff-zone workers have the credentials to access the industries that V3 identifies as paying above the break-even threshold. The industries that clear benefit phase-outs — healthcare, skilled trades, professional services — typically require at least an associate’s degree, vocational certificate, or apprenticeship.
• V3 uses average industry pay (QCEW), not entry-level. BLS QCEW reports average annual pay, which includes senior workers and management. Actual starting pay in higher-paying industries is lower than the average. This means the metric overstates advancement feasibility — the true share of industries accessible to cliff-zone workers is likely lower than reported. We use QCEW because no county-level entry-level wage dataset exists. BLS Occupational Employment Statistics data suggests entry-level pay (10th percentile) averages approximately 60–70% of the industry mean in cliff-relevant occupations. This metric likely overstates advancement feasibility by a corresponding margin — the true share of entry-level jobs paying above the threshold is lower than the industry-average measure indicates.
• V3 uses the SP2K break-even threshold because single parents with 2 children face the steepest cliffs. The threshold varies significantly by state: $10/hr in California (generous benefits push break-even low) to $27/hr in Georgia (weak benefits push break-even high). This means V3 partially captures state policy effects — states with poor policy have higher thresholds, making advancement harder regardless of local industry mix. This is intentional: the Policy component measures whether the statehas good policy; Vulnerability measures the consequence of that policy on local advancement pathways.
• Dropped metrics (previous version): Break-even multiple (advancement threshold ÷ minimum wage — nearly constant across states because the threshold clusters at ~$13.25, making it a minimum wage proxy rather than a vulnerability measure), wage ceiling (employment-weighted average industry pay — a blended average that obscures the distribution; advancement feasibility captures the same concept more honestly by showing what share of employment actually clears the threshold rather than averaging across industries). Part-time rate, car dependency, and low education rate were tested in earlier versions and did not differentiate within the cliff zone.

Component 4: Policy

State policies that moderate or amplify benefit cliff severity. Four sub-scores measure health coverage, income support programs, childcare subsidy access, and state tax burden. This score is identical for all counties within a state. Federal benefit rules create cliffs in every state; state policy acts as a multiplier, shifting cliff severity by up to 25% in either direction.

State-Level Scoring

States are scored against all 51 states (50 states + DC) using state-level metrics, not population-weighted averages of county scores. The three county-level components (Exposure, Severity, Vulnerability) are recomputed at the state level:

• Exposure metrics use population-weighted state averages of county-level rates (cliff zone share, single parent share, etc.)
• Severity metrics 1–3 (Peak METR, cliff depth, program stacking) are computed from the state’s own benefit landscape using population-weighted average housing costs, not averaged from county landscapes
• Severity metrics 4–6 and all Vulnerability metrics use population-weighted state averages of county values
• Policy is inherently state-level

State scores are percentile-ranked among 51 states. A state at rank #8 has a more severe cliff environment than 43 other states. Component scores are similarly ranked against 51 states.

Data Sources & Vintage

All data is public. No individual-level records are published. PUMS microdata is used for statistical aggregation only.

How Benefits Are Modeled

We model 10 programs using the Atlanta Fed’s Policy Rules Database (PRD v4.4.1) with state-specific parameters. Benefits are computed at each $0.50/hr wage step from $7.25 to $30.00 (2,080 hours/year), producing a benefit landscape of total resources (income + benefits) across the wage spectrum for each family type and county. The advancement threshold search extends to $35.00 using finer $0.25/hr steps.

Programs modeled

1. SNAP — Supplemental Nutrition Assistance Program
2. Medicaid — Adult and child coverage (expansion status varies by state)
3. CCDF — Child Care and Development Fund subsidies
4. EITC — Earned Income Tax Credit (federal + state where applicable)
5. CTC — Child Tax Credit (refundable portion)
6. TANF — Temporary Assistance for Needy Families
7. WIC — Women, Infants, and Children nutrition program
8. School Meals — National School Lunch & Breakfast Programs
9. ACA — Affordable Care Act marketplace subsidies
10. Section 8 — Housing Choice Voucher Program

Family types

Four family types are modeled: single parent with 2 children, single parent with 1 child, married couple with 2 children, and single adult (no children). Each type has different eligibility thresholds and benefit values. Prevalence-weighted metrics use PUMS-derived local family type shares to weight results.

County-specific inputs

Each county’s benefit landscape uses: state program rules (PRD), county-level fair market rents (HUD FMR, 5 bedroom sizes), county median household income, and childcare market rates (county-level from DOL NDCP where available, state-level fallback otherwise). This means the same family type facing the same program rules will see different cliff profiles in different counties due to local housing costs and income levels.

Benefit Valuation

Different benefit types are valued differently. This transparency is important because stacking valuations can inflate aggregate figures, particularly when Medicaid is included.

This follows the Atlanta Fed’s methodology. Per-program cliff values are shown individually. We do not present a single aggregate “total benefits lost” figure because the valuation methods differ.

Participation Weighting

The benefit landscape shows structural benefits — what happens if a family is enrolled in all programs they qualify for. But most families don’t receive all programs. Participation rates vary widely:

The participation-weighted METR adjusts the benefit landscape by scaling each program’s contribution by local enrollment rates. This produces realistic effective tax rates (typical range 1.7–4.0) rather than the theoretical maximums (1.7–5.6) that assume universal enrollment. This directly addresses a well-documented critique that structural cliff rates tend to overstate the problem.

Validation

External outcome correlations

All correlations significant at p < 0.001 and in the expected direction. These are correlations, not causal claims. The Opportunity Atlas measures outcomes for children born 1978–1983; current cliff severity may reflect persistent structural conditions rather than direct causation.

Independence from poverty

The CLIFF Score correlates with county poverty rate at r=+0.349 — moderately correlated but well below the 0.5 threshold that would indicate redundancy. The index captures cliff-specific conditions (program rules, benefit stacking, advancement pathways) that poverty rate alone does not explain.

Component independence (state level)

All inter-component correlations below 0.37 at the state level. The highest (E–V = 0.37) reflects that states with more people in the cliff zone also tend to have weaker local economies — a structural reality, not a scoring artifact. Severity remains nearly independent of all other components (highest |r| = 0.15).

Metro/rural balance

Metro counties: mean score 46.9. Rural counties: mean score 51.9. The 5-point gap reflects that rural economies genuinely have fewer industries that pay above family benefit thresholds. Earlier versions engineered near-perfect parity (1.5-point gap) through ratio-based metrics that washed out real structural differences. The current gap is a more honest representation of the advancement pathways available in rural vs metro economies.

Medicaid expansion gap

Expansion states: mean score 47.8. Non-expansion: 65.8. The 18-point gap confirms that the Policy multiplier adds meaningful signal. This is slightly narrower than the previous version (24 points in an earlier build) due to the removal of the healthcare shortage overlap metric, which disproportionately penalized non-expansion states.

Economic Impact Model

The economic impact section estimates the aggregate annual cost of cliff-avoidance behavior in each state. This is a modeled estimate, not an observed outcome. It combines survey-reported behavioral rates with county-level population data.

Behavioral model

Working families

The economic impact model applies only to working families — those with at least one employed household member. Working rate is sourced per county from Census PUMS microdata (employed persons ÷ cliff-zone population). Where PUMS data is unavailable, it falls back to ACS household-level data (households with workers ÷ total households), then to 0.75 as a last resort. The national average working rate across cliff-zone families is approximately 75%.

Structural cross-check (deadweight loss)

As a cross-check, a separate estimate uses the public economics framework for deadweight loss from high marginal tax rates. It takes the average METR in the cliff zone, published labor supply elasticities (ε=0.15 intensive margin, ε=0.25 extensive margin, from Saez and Chetty), and total cliff-zone earnings. For most states, the structural estimate falls in the same range as the behavioral model, from completely different inputs.

Evidence Base

The cliff avoidance evidence base is thin but growing. A complete accounting of independent studies, null findings, and intervention outcomes follows.

Avoidance studies

Null & counter findings

Intervention outcomes

Known Limitations

Scoring limitations

• Equal weighting. All metrics within a component carry equal weight. Under moderate perturbation testing (Dirichlet α=10, &pm;10pp from equal), tier stability is ~44% and average rank shift is ~482 positions. This reflects the aggressive test design (uniform simplex sampling), not actual score instability. Always show component breakdowns alongside the composite score.
• Policy has zero intra-state variation. Every county in the same state receives the same Policy multiplier. Within-state differentiation comes entirely from Exposure, Severity, and Vulnerability.
• Advancement feasibility is partially policy-driven. The SP2K break-even threshold varies from $10/hr (CA) to $27/hr (GA) depending on state benefit generosity. States with weak policy have higher thresholds, making advancement harder regardless of local industry mix. This means Vulnerability partially captures policy effects that are also in the Policy component. The power mean aggregation prevents this from causing extreme distortion, but the overlap is a known design trade-off: Policy measures whether the state has good policy; Vulnerability measures theconsequence for local workers.
• Policy multiplier range is empirically calibrated. The 0.75–1.25 range produces an 18-point Medicaid expansion gap, consistent with observed outcome differences. The exact range is a design choice, not derived from a structural model.
• CCDF policy creates a non-compensatory cascade. A state with a restrictive CCDF income limit (e.g., 100% FPL) receives a high P3 (Childcare & Work) score, which increases the policy multiplier. That same restrictive limit also raises the advancement threshold in V3 — families must earn more to clear all phase-outs — increasing the base Vulnerability score. The policy multiplier then amplifies this already-elevated base. The estimated magnitude is 12–18 points of non-compensatory penalty. This is a deliberate design choice: the policy creates the problem (P3) and the local economy bears the consequence (V3). The power mean’s concavity partially mitigates extreme compounding.

Data limitations

• Data lag. Census ACS: 2020–2024 (center year ~2022). IRS SOI: TY 2021. Program parameters: 2026 rules. Population and program rules are not perfectly synchronized. Three specific temporal mismatches affect interpretation:
  • (a) IRS SOI & expired Child Tax Credit. IRS SOI data (tax year 2021) reflects the expanded Child Tax Credit ($3,600/child) that expired December 31, 2021. The current benefit engine uses the restored $2,200/child. EITC and CTC participation patterns in the IRS data reflect a policy environment that no longer exists, affecting the economic impact model’s inputs.
  • (b) North Carolina Medicaid expansion. North Carolina expanded Medicaid in December 2023. The Census ACS 2020–2024 5-year estimates (centered approximately on 2022) reflect pre-expansion enrollment patterns. The benefit engine and state configuration treat North Carolina as an expansion state with 138% FPL adult coverage. Population data and program rules are not synchronized for this state.
  • (c) DOL childcare prices. DOL childcare prices date from 2022 and do not include inflation adjustment. Post-pandemic childcare cost increases are not captured. The S5 (Childcare Cost Burden) metric likely understates current burden in most counties.
• PUMS sampling error. PUMS is a ~1% sample. Cross-tabulations for small counties (population <20,000) may have high margins of error. Counties with fewer than 10 unweighted PUMS records fall back to Census summary table estimates.
• Section 8 voucher rate exceeds 1.0 for 233 counties. HUD counts include project-based vouchers tied to buildings, not families in the cliff zone. Clamped to 1.0 but may slightly overstate housing program coverage.
• Childcare cost burden fallback. 43 counties (1.4%) lack both DOL county-level prices and state market rates, using a 0.15 fallback value.
• EITC participation rate. State-level only (no county-level source exists). All counties within a state receive the same rate.
• CCDF enrollment. State-level only (FFYF 2025). Ranges from 7% to 30% across states but does not vary within states.

Measurement window (100–199% FPL)

The Exposure metrics count population at 100–199% of the federal poverty level, where benefit phase-outs from multiple programs overlap most densely. However, some significant cliffs occur outside this window. In 30 states, CCDF (childcare subsidy) income limits exceed 200% FPL — the actual CCDF cliff hits at incomes the Exposure metrics cannot count. Connecticut and Vermont set CCDF limits at 400% FPL; New York at 349%; California at 342%. Conversely, in 8 non-expansion states, the primary Medicaid cliff occurs below 100% FPL (as low as 15% FPL in Texas).

The Severity metrics (METR, cliff depth) DO capture cliffs across the full $7.25–$30.00/hr wage range, including those above 200% FPL. The scope limitation applies specifically to the population count in the Exposure component.

Model limitations

• SSI/SSDI not modeled. The benefit engine does not model SSI earnings cliffs (SGA threshold) or SSDI trial work periods. The disability rate is shown as a display-only metric but is not scored, capturing the prevalence of this population without modeling their specific cliff mechanics.
• No sub-state policy variation. County or city-level benefit programs, local minimum wage ordinances, and employer-specific benefits are not captured.
• In-kind valuation. Medicaid and childcare values use per-capita cost, not market value to the household. Families may value these programs differently than what they cost the government.
• Economic impact is survey-derived. The avoidance rate comes from one nationally representative survey (CSD/WashU 2025, n=2,511). Self-reported behaviors, not validated against wage records. The direction is consistent across every study that has looked, but no study has matched survey responses to administrative earnings data.
• Eligibility ≠ enrollment. Population estimates count households by income. Not all are enrolled in every program. The scoring system addresses this through separate structural (Severity) and enrollment (Exposure) components.

Glossary

Benefits Cliff

A sudden and often unexpected decrease in public benefits that occurs when income crosses a program’s eligibility threshold. When benefit losses exceed the income gain, the family is financially worse off for earning more.

CLIFF Score

The composite severity score (0–100) assigned to each county and state. Computed from four components (Exposure, Severity, Vulnerability, Policy) using 17 metrics. Higher = more severe cliff environment.

Cliff Zone

The income range between 100% and 199% of the Federal Poverty Level — approximately $28,000–$55,000/year for a family of three in 2026. This is the range where most means-tested benefits phase out or terminate.

METR (Marginal Effective Tax Rate)

The share of each additional dollar earned that is lost to reduced benefits and higher taxes combined. At 50%, half of every new dollar disappears. At 100% or above, the family is no better off (or worse off) after a raise. The Upward Mobility Act of 2026 (S. 3583 / H.R. 6949) uses this same measure and proposes capping it at 50%.

Structural METR

The METR computed assuming a family is enrolled in all programs they qualify for. Represents the rate enrolled families actually face. Used in scoring (S1: Peak METR) and displayed in the benefit landscape chart.

Participation-Weighted METR

The METR adjusted by local program enrollment rates. Produces a population-average rate that accounts for the fact that most families do not receive all programs. Used in the economic impact model and available in the data pipeline.

FPL (Federal Poverty Level)

An income threshold set annually by HHS. For 2026, 100% FPL for a family of three is $27,320. Most benefit eligibility is expressed as a percentage of FPL (e.g., SNAP at 130%, Medicaid expansion at 138%).

Hard Cliff

A benefit that ends entirely at a single income threshold with no gradual phase-out. CCDF (childcare subsidies) in most states and Medicaid in non-expansion states are common examples.

Gradual Phase-Out

A benefit that decreases incrementally as income rises. SNAP and EITC are examples. Multiple gradual phase-outs occurring in the same income range can produce high combined METRs even without a single hard cliff.

Program Stacking

Multiple programs reducing benefits at the same income level. When SNAP, Medicaid, and CCDF all phase out near the same wage, the combined METR can exceed 200–300%.

Break-Even Multiple

The ratio of the wage where total income + benefits recover to pre-cliff levels divided by the starting wage (typically minimum wage). A multiple of 2.0× means a family needs to double their wage before the raise was “worth it.”

BBCE (Broad-Based Categorical Eligibility)

A state option that raises the SNAP gross income limit above the standard 130% FPL, typically to 200% FPL. States with BBCE have a wider income range over which SNAP phases out, reducing cliff severity for food assistance.

PRD (Policy Rules Database)

The Atlanta Fed’s database of benefit program parameters for all 50 states + DC. Version 4.4.1 (2026 rules) provides the eligibility thresholds, benefit formulas, and phase-out schedules used in the CLIFF Index benefit engine.