The household and the society — Society in Silico

Consider a single parent in New York with one child who has a disability, earning $30,000 a year. How much do they owe in taxes? How much do they receive in benefits? What happens if they take on an extra shift?

These are not abstract questions. They decide whether the parent can make rent, whether a raise is worth chasing, whether a job across town pays off once it changes their benefits. The answers run through federal income tax, state income tax, payroll taxes, the Earned Income Tax Credit, the Child Tax Credit, Supplemental Security Income, SNAP, and a dozen other programs — each with its own definition of income, its own household unit, its own phase-out schedule, none of them designed to fit together.

No one can calculate this in their head. Most accountants would struggle. Yet the answer shapes every financial decision the family makes. This is the problem PolicyEngine's household calculator was built to solve ^[1] — and it is where the engine's work begins: with what it can show one family, before it can show anything true about a country.

What you owe, what you get

A typical low-income family might draw on the EITC, a partially refundable Child Tax Credit, SNAP, SSI for a disabled member, Medicaid, a housing voucher, a state child-care subsidy, and free school meals at the same time — eight programs, each written independently, each with its own idea of income and its own household unit, accreted over decades of legislation with no coordinating hand. The result is a system no participant fully understands. Caseworkers specialize in one program. Tax preparers work one side of the ledger. Benefits counselors know eligibility rules but not tax consequences. No one sees the whole picture. The fragmentation is not harmless: it is why eligible families leave benefits unclaimed, why a raise can quietly cost more than it pays, and why the same decision can look smart to a tax preparer and ruinous to a benefits counselor.

Seeing the whole picture is the calculator's single promise: enter your circumstances and it shows your taxes and benefits together, under current law. It walks through the household — how many adults, how many children, what ages — then income, wages, self-employment, investment, retirement, then the specifics of housing cost, child-care expense, disability. From those inputs it computes dozens of outputs: federal and state income tax, payroll taxes, the alphabet of credits, the benefit programs a household may qualify for, and a single net figure for what the family keeps after everything .^[2] PolicyEngine's US model spans federal income and payroll taxes, the EITC and CTC, SNAP, SSI, WIC, TANF, and dozens of state programs; the UK model spans income tax, National Insurance, Universal Credit, Child Benefit, and the full run of means-tested support. Each program is a separate module, its logic drawn from statute, regulation, and agency guidance.

The comprehensiveness is the whole point, because the programs interact. A family's SNAP benefit depends on its net income after taxes and deductions; its tax depends on credits that depend on the same children who might also generate SSI or a child-care subsidy. Change one input and the effect ripples across the others. No single-program calculator captures that. The microsimulation engine does, because it evaluates every program at once for the same household. And once the whole calculation is visible, a family can act on it — plan around a raise that won't lift take-home pay as much as expected, or claim a program they qualify for and never knew to ask about.

Rates above a billionaire's

Plot a household's net income against its earnings and the hidden structure of the system becomes visible. For many low-income families the line is not a smooth upward slope. It has flat stretches where more work barely raises take-home pay, and here and there a vertical drop — a cliff — where crossing a threshold means losing a benefit worth more than the raise that triggered it .^[3]

A cliff is just that: the point where earning more leaves a household worse off, because the benefits withdrawn and the taxes owed on the next dollar add up to more than the dollar itself. Analysts have worried about cliffs for decades. But until tools like this one made them visible, most of the people standing on one never saw it coming.

The gentler version is everywhere. As the New York parent's earnings climb from $20,000 toward $30,000, several things happen at once: SNAP tapers, the EITC phases out, the child's SSI falls, and income and payroll taxes begin to bite. Economists watch marginal rates because they shape behavior — keep ninety cents of your next dollar and you might work more; keep twenty and you might not. For high earners those numbers are debated endlessly: the 37 percent top federal rate, capital gains, state tax. The counterintuitive reality is that a low-income worker often faces a higher marginal rate than any billionaire, because benefit phase-outs stack on top of taxes. SNAP withdraws about thirty cents per additional dollar of net income. The EITC phases out at roughly sixteen cents per dollar for a family with one child, twenty-one for two or more. SSI falls by about fifty cents per dollar of countable income. Stack them and the combined rate can approach eighty percent — more than double what a hedge-fund manager pays on his last dollar. This is not an edge case but the ordinary structure of a low-income budget, and it is exactly the kind of pattern that disappears into any average. At a true cliff the chart spikes past 100 percent, the unmistakable signature of earning more and keeping less. The calculator shades the earnings range where extra work barely moves net income — a dead zone made visible — and that visibility serves two ends at once: a family can plan around a cliff instead of walking off it, and a policymaker can see the work disincentives that well-meaning programs create only in combination, never one at a time.

The mechanics show up as specific traps, each one the calculator can make exact:

The SNAP categorical cliff. Receiving SSI confers categorical eligibility for SNAP — you qualify regardless of the usual income test. But once earnings push SSI to zero, that categorical eligibility vanishes and the household must suddenly satisfy SNAP's own income test. The cliff lands at precisely the earnings level where someone is trying to become self-sufficient.

The marriage penalty. Two single parents, each earning $25,000 with one child, each qualify separately for substantial EITC and CTC. Marry them and their combined $50,000 pushes both credits into higher phase-out ranges, and the household can end up with less in total transfers than the two separate households received. The penalty is written in no single line of the tax code; it emerges from several programs' different treatment of household structure, and the calculator surfaces it by running the two adults separately and then as one.

The aging-out cliff. A family receiving SSI for a disabled child hits a discontinuity when the child turns eighteen and SSI stops counting the parents' income and starts counting the child's own — often raising the SSI payment while stripping family-based benefits, as the same birthday moves the child out of pediatric Medicaid provisions, school meals, and child-care subsidies. Nudge the child's age in the model and the whole cascade appears.

The what-if machine

The calculator is not only a reader of current law. Open the policy editor and you can change the rules — raise the EITC, smooth a SNAP cliff, add a child allowance, make the Child Tax Credit fully refundable, set a universal basic income, flatten the income tax — then return to your household and watch what moves. Baseline shows in gray, reform in blue; the dead zones and the marginal-rate spikes redraw themselves, revealing whether the reform helps your specific family or quietly weakens its incentive to work. The reform space is as deep as the model is wide: someone exploring basic-income proposals can set a flat tax at any rate against a universal benefit at any amount and see at once whether the simplification leaves their household ahead or behind, and whether it smooths the cliffs or just trades them for a higher average rate. The charts update as you type — two net-income lines, two sets of dead zones, the marginal-rate curves stacked so it is immediately clear where a reform flattens the benefit system and where it carves a new cliff.

This is the "what if" ordinary people never had. What would a CTC reform mean for your family? Under 2026 law the Child Tax Credit is $2,200 per child ;^[4] any reform is a change measured against that. Take the clearest recent one. In 2021 Congress temporarily raised the credit to $3,600 for children under six and $3,000 for those six to seventeen, and made it fully refundable, so families who owed no federal income tax received the full amount. For a single parent earning $15,000 with two young children, that was no marginal adjustment: the engine computes roughly $1,800 under the old rules, held down by the phase-in on earnings above $2,500, against $7,200 under the expansion. When the expansion lapsed at the end of 2021, the same family's income fell by more than $5,000. The tool shows that not as a national average but as a dollar figure for one household.

It's your life, under different rules.

The same capability changes how policy gets designed. Rather than propose a reform and wait weeks for a score, a legislative aide can model dozens of variants in an afternoon: set the benefit at $3,000 per child or $4,000, phase it in from the first dollar or from $2,500, cap it at $75,000 or $150,000, and read a different household chart for each — design trade-offs invisible in the legislative text and immediate in simulation.

Where the system is most complex

The household calculator originated in the UK, where PolicyEngine launched in 2021 before crossing to the US ,^[5] and the UK version proved a rule the US one later confirmed: the household view is most powerful where the system is most tangled. Universal Credit was meant to simplify welfare by folding six legacy benefits into one; in practice its taper — the rate at which the award is withdrawn as earnings rise — interacts with income tax, National Insurance, and council-tax support to push some working families' marginal rates above 70 percent. When Chancellor Rishi Sunak cut the UC taper from 63 to 55 percent in the Autumn 2021 Budget, the tool could show what that meant household by household: a single parent working 25 hours a week at minimum wage would keep about £1,000 more a year citation pending, while a two-earner couple would gain less — distributional detail that aggregate statistics hid. HM Treasury later published a formal evaluation benchmarking PolicyEngine UK against its own internal models :^[6] a research prototype the government took seriously enough to test against its own machinery.

From calculator to primitive

Every household calculation carries a caveat, shown plainly: PolicyEngine provides estimates, not benefit determinations .^[1] The model encodes the rules as written; an actual determination turns on discretion, documentation, asset tests, and local variation that no model fully captures. So the team validates continuously against official calculators and published tables ^[7] and treats every discrepancy as a bug to investigate — while admitting that perfect accuracy in a system this complex is unattainable, and that claiming it would be worse than naming the gap. Showing the logic openly is the point: a user can judge whether the calculation matches their own situation, which a black box asserting false precision would never allow.

That honesty is what let the calculation become infrastructure. In 2025 MyFriendBen, a benefits-screening service, launched in North Carolina, Illinois, Colorado, and New York on PolicyEngine's API :^[8] plain-language questions in a dozen languages, and in about six minutes a household learns which programs it likely qualifies for. In Colorado, users surfaced an average of $1,500 a month in benefits they appeared eligible for but had not claimed .^[9] PolicyEngine's own note reported that the estimates matched expected amounts more than 90 percent of the time — a self-reported figure, not an independent audit .^[8] The engine underneath is the same whether a policy wonk runs a scenario on the website or a navigator helps a family in Charlotte. A prototype pushes it further still, reframing the whole calculation around life events — a birth, a move, a marriage, a job loss, turning 65 — because that is how people actually meet policy, not by editing tax parameters .^[10]

The household view, it turned out, was a primitive — and in two senses at once. It is the basic computation that benefits screening, financial coaching, guaranteed-income design, and legislative analysis all separately need: given this household, what are its taxes and benefits? And it is the atomic unit of ground-truth verification — the level at which an encoded rule can be proven against a reference calculator or a real benefit determination. A national estimate is nothing but a weighted sum of household calculations; if the household calculation is wrong, everything built on top of it is wrong in ways no aggregate will reveal. Getting the single family right is what makes the country-level numbers mean anything.

Which is the bridge to the society view. A macroeconomic model might say a tax cut costs $100 billion and lifts GDP by 0.3 percent. Microsimulation can say that and also tell you the same cut hands $12,000 to a high-income household in Connecticut and $200 to a low-income household in Mississippi — and that the Mississippi family now faces a higher marginal rate, because the cut nudged it into a phase-out. The aggregate is assembled from millions of those specific calculations.

From one household to many

The society view answers what the household view cannot: not "what would this mean for me?" but "what would this mean for everyone?" It has a lineage. In 1974 Joseph Pechman and Benjamin Okner of Brookings published Who Bears the Tax Burden?, the first comprehensive attempt to allocate US tax burdens across the income distribution using household-level data .^[11] Merging records for 72,000 households, they found the system roughly proportional across most of the distribution — a result that irritated partisans on both sides — Their finding challenged conservatives who called the system too progressive and liberals who called it too regressive at the same time, and their method of running micro-data through incidence assumptions became the template for every distributional analysis since, at Treasury's Office of Tax Analysis, the Joint Committee on Taxation, and the CBO.

The move from one household to the nation sounds trivial — run the household calculation for everybody and add it up — and in principle it is exactly that. In practice it means solving some of the hardest problems in policy analysis, starting with a plain one: you cannot survey everybody. The monthly Current Population Survey samples about 60,000 households ;^[12] its Annual Social and Economic Supplement, the income workhorse most policy analysis draws on, reaches roughly 100,000 — either way a sliver of some 130 million American households. So each surveyed household stands in for many others like it, carrying a weight that says how many: a household in rural Wyoming might represent 5,000, one in Manhattan 500 — weights the Census computes to correct for the survey's sampling design and for who does not respond. To score a reform, the engine computes the change for each sampled household, multiplies by that weight, and sums across the sample — a national estimate that inherits all the uncertainty a weighted survey carries.

The data underneath

That uncertainty is the binding constraint, because the survey is flawed in ways that bias the answer rather than merely blur it. High incomes are top-coded and under-reported; benefits are undercounted — Bruce Meyer and colleagues found that 40 to 50 percent of SNAP recipients do not report their benefits, and the problem runs across programs .^[13] The sample is too thin for most state-level work, and asset questions are sparse, so wealth policies are hard to see at all. None of this is random noise: survey-based scores tend to understate both the cost of reforming means-tested programs and the revenue from taxing top earners, because the data undercounts exactly the benefits and incomes in play.

The answer is to stop treating the raw survey as ground truth and calibrate it to administrative reality — pulling the undercounted top incomes and underreported benefits back toward the totals the tax and benefit agencies actually record. PolicyEngine's earlier public milestone was the Enhanced CPS, released in August 2025: five datasets integrated, reported taxes and benefits replaced with computed amounts for internal consistency, income distributions corrected against IRS records, and survey weights re-solved to match 9,168 administrative totals — which cut deviations from those targets by about 97 percent .^[14] That line of work has since been folded into populace, a calibrated-microdata commons that treats data construction as shared infrastructure rather than a per-project chore: synthesize records where the survey is thin, impute missing variables with quantile-regression forests ,^[15] calibrate weights to administrative aggregates ,^[16] and publish certified releases the way software ships versioned builds. Local-area analysis then becomes a matter of filtering one national calibrated dataset rather than building fifty bespoke ones.

Cost, poverty, and who gains

The first question about any reform is what it costs. PolicyEngine scores it by running two simulations across the weighted sample — one under current law, one under the reform — and summing the difference in tax revenue and benefit spending .^[17] The arithmetic is simple; the computation is not, since each of the roughly 100,000 records triggers thousands of eligibility checks and bracket calculations, though the whole run still finishes in seconds. Scores like these are static — they hold behavior fixed; layering in labor-supply elasticities can pull an estimate down when people are assumed to work more, or push it up when they work less, which is one reason two honest analysts can publish different numbers for the same bill.

A single cost figure hides how contingent it is. The same reform carries different price tags depending on the year you score it in, as provisions phase in and thresholds inflate; on whether you let people change their behavior in response; and, most subtly, on the baseline you compare against. Baseline is the one that swings estimates by hundreds of billions. Before July 2025, scoring an extension of the 2017 tax cuts meant first deciding whether the baseline assumed the individual provisions would expire on schedule or continue — an assumption worth hundreds of billions before any policy changed. OBBBA settled that particular question by making the provisions permanent ,^[4] but it sharpened the general lesson: an estimate is only as meaningful as the world it is measured against.

Consider making the Child Tax Credit fully refundable — removing the refundable cap and the earnings phase-in so that families who owe no federal income tax receive the whole credit — scored against 2026 law, in which the credit is $2,200 per child .^[4] The net cost is roughly $23 billion for 2026, and child poverty falls by about 2.6 percentage points—on the order of 1.9 million children. The distributional picture is the opposite of what the headline might suggest: the families who gain are those whose credit currently exceeds the income tax they owe, and who therefore forfeit part of it today — low- and moderate-income working families with children. Higher-income households, already claiming the full credit, see nothing change.

Poverty is the next question, and it depends on how you measure. PolicyEngine uses the Supplemental Poverty Measure, which unlike the older official measure counts taxes and in-kind benefits like SNAP, adjusts for geographic differences in housing cost, and nets out work and medical expenses; for two adults and two children renting, the 2024 SPM threshold is about $39,400 .^[18] In a 2023 analysis PolicyEngine put roughly 9.6 percent of Americans below their threshold, with children poorer than working-age adults and seniors, and women poorer than men ^[19] to verify. Apply a reform and the engine recomputes each household's resources against its threshold; run the change by demographic group and a program's uneven reach appears — it can cut child poverty sharply while barely moving the rate for seniors, or help women more than men. WIC is a concrete case: it lowers overall poverty by about 0.8 percent and deep poverty by 2.2 percent, but child poverty by 2.6 percent, and women's poverty by 0.9 percent against men's 0.7 — a program that looks gender-neutral on its face landing unevenly once the microdata is split .^[19]

Beyond poverty lies the question of who wins, who loses, and what happens to the distance between top and bottom. Sorting households into income deciles and showing the share of each that gains or loses reveals whether a reform is progressive or regressive, and the same cut can be taken by wealth decile using imputed wealth, by age, by sex, by geography, and by race where the data supports it .^[17] Inequality gets its own measures. The Gini coefficient is the standard, but it is most sensitive to the middle of the distribution, so a reform that transforms life at the very bottom can barely move it; the Atkinson index, tuned by an inequality-aversion parameter, weights the bottom more heavily and responds more to anti-poverty reforms. The measures can disagree in instructive ways — a reform can cut poverty while raising inequality by helping the near-poor more than the poorest, or shrink inequality while barely touching poverty by redistributing within the middle. A cost figure means little on its own: fifty billion dollars concentrated on families in poverty is a different policy from fifty billion spread evenly across the income spectrum, and only the distribution says which one a reform is.

PolicyEngine does not tell you which of those to prefer. It tells you what the outcomes are.

The neutrality is deliberate, and imperfect. The tool shows budget, poverty, inequality, and distribution across several cuts and lets users weigh what they value; it adds no editorial verdict, documents its behavioral assumptions and its data limits, and keeps its methodology open to challenge. But the choice of what to display is itself normative — showing a poverty rate implies poverty matters, and framing by income decile frames the debate a particular way. The honest claim is not perfect neutrality, which is impossible, but a deliberate separation of analytical infrastructure from advocacy.

Where the numbers are solid, and where they're soft

Society-level estimates carry more uncertainty than household ones — the sample is weighted, variables are imputed, interactions compound — so it matters to say plainly where the outputs are trustworthy and where they are not.

Household calculations are the most reliable thing the engine produces, because they depend on the encoded rules rather than on the data distribution: for a family with stated characteristics, the answer is only as good as the statute encoding, and no better or worse for the survey behind it. Directional and comparative results are nearly as solid — whether a reform helps low-income families more than high-income ones, whether Reform A costs more than Reform B — because the shapes of distributions and the ratios between reforms hold steadier than their absolute levels, and the same data limits press on both sides of a comparison.

Absolute budget scores deserve the most caution. PolicyEngine's aggregate revenue estimates have historically run well below official totals — by roughly a third to verify — because survey microdata undercounts top incomes and underreports benefits, and a static model omits the behavioral and macroeconomic feedback that moves real collections. For an order-of-magnitude figure, or a comparison across reforms, that is good enough; for a precise revenue estimate, the Joint Committee on Taxation and the CBO remain the standard, because they work from confidential tax returns that capture the whole income distribution. The calibration itself is validated by holding out some administrative targets and checking the reweighted sample against them — the same procedure that cut deviations by about 97 percent — but even a 97 percent improvement from a large gap still leaves a meaningful gap. No microsimulation matches reality exactly. The question is whether the model's limits are knowable, and open models, uncomfortably, expose theirs where closed ones hide behind institutional authority.

Analysis as infrastructure

The society view enables a kind of analysis that used to live only inside government: a response to a policy proposal while the proposal is still live. Traditional analysis takes weeks or months — a team reads the legislation, codes it, runs the model, writes it up, clears review — and by the time it publishes, the legislative moment may have passed. The model's speed comes from pre-investment: building and maintaining the engine takes years, but once it exists, scoring a new reform takes hours, because the fixed cost is already paid and the marginal analysis is cheap. That changes who can take part in the debate, and when.

It also changes what the debate is about. When analysis is a one-time product — a report from a think tank — the argument is whether to trust that particular report. When analysis is infrastructure that anyone can run, through a web app, a Python package, or an API that other products build on, the argument shifts to whether the underlying model is sound. That is a more productive fight, because a model's quality is testable and improvable in ways a report's credibility is not. The same engine meets different users through different surfaces — a researcher iterating over thousands of scenarios in the Python package, an advocate embedding a shareable chart in an op-ed, a journalist who wants to show her work, a legislator comparing reform options side by side — and each new analysis adds an example, exposes an edge case, and compounds the value of the shared model.

Two reforms, seen clearly

The 2021 Child Tax Credit expansion is the clearest recent demonstration of what the society view catches — and of what goes dark without it. The American Rescue Plan raised the credit and, more consequentially, made it fully refundable, removing the earnings phase-in that had excluded the poorest families, and paid half of it out in monthly installments from July through December. Before enactment, Columbia University's Center on Poverty and Social Policy projected that the broader relief package could cut child poverty by more than half .^[20] After enactment, the data agreed: SPM child poverty fell from 9.7 percent in 2020 to 5.2 percent in 2021, a record low .^[21] The credit as a whole kept 2.9 million children out of poverty that year; the expansion alone accounted for 2.1 million of them ^[22] — two numbers the debate routinely conflates and the microdata keeps distinct. The one-year cost ran to roughly $105 billion citation pending. The predictions landed close not by luck but because the policy worked through mechanical channels — direct cash on simple eligibility rules — where behavioral uncertainty is low and microsimulation is at its best. When the expansion expired in January 2022, Columbia tracked child poverty nearly doubling within months; the rebound was legible only because the simulation had drawn the baseline. The episode carried a subtlety worth keeping: because the expansion also raised the maximum credit for every eligible family, the distributional tables showed the largest percentage gains at the bottom but real dollar gains further up the distribution — whether that read as a broad coalition or as loose targeting depended on values the model could not adjudicate.

The UK offers the contrast that shows the method's edge. In September 2022, Chancellor Kwasi Kwarteng's "mini-budget" proposed abolishing the 45p additional rate of income tax, cutting the basic rate, and reversing a National Insurance rise, all in the name of growth. PolicyEngine UK published household-level distributional analysis within hours — among the only independent estimates available in the critical first days — showing that the package overwhelmingly favored high earners, the top decile gaining about £2,500 a year while the bottom gained almost nothing .^[2] Media outlets cited it; within weeks the government reversed the 45p abolition, and within a month the prime minister had resigned. Set beside it a quieter change with the opposite lesson: that same UC taper cut, from 63 to 55 percent. Microsimulation drew the distinction that mattered — the taper cut reached only working recipients, while the pandemic's £20-a-week uplift had reached everyone on Universal Credit, so per pound the uplift cut poverty roughly 40 percent more effectively citation pending, even as the taper cut improved work incentives, dropping the share of workers facing marginal rates above 70 percent from 26 to 9 percent citation pending. Two policies, similar budgets, opposite distributional signatures — a trade-off invisible without the society view.

By 2026 this work had begun to change shape. Rather than produce a score when someone asks, the same infrastructure started watching legislation as it moved, flagging which bills the model could already handle and routing the rest toward encoding ^[23] — the on-ramp to encoding the policy corpus at scale, which is the subject of Part III.

One engine, three ingredients

Both views run on one engine. The household calculation and the national estimate are the same computation at two scales, and it works only because three things beneath it hold. The rules have to be encoded correctly. The data has to represent the people. And the behavior — the seam where the model stops describing the law and starts predicting how people respond to it — has to be handled honestly. Those are the three ingredients of any microsimulation, and the next chapter takes them one at a time.