The birth of microsimulation

In 1957, an economist named Guy Orcutt opened a paper with a verdict on his own profession: "Existing models of our socio-economic system have proved to be of rather limited predictive usefulness" .^[1] The journal was the Review of Economics and Statistics — respectable, not glamorous — and the paper, "A New Type of Socio-Economic System," was dense with equations and read by almost nobody. What it proposed was a simulated United States, built one household at a time, on computers that could not yet run it.

The idea arrived in three acts. In 1957 Orcutt proposed it. In 1961 he and three co-authors published a working demonstration — limited, but running. In 1975 the Urban Institute completed DYNASIM, the first full-scale version, eighteen years after the proposal. Whatever Orcutt saw in 1957, he judged it worth a career.

The engineer

Orcutt took a physics degree from the University of Michigan in 1939, finished it in three years, and began graduate work in economics the same fall; his Ph.D. came in 1944 .^[2] to verify He never really left the lab bench. Inspired by Jan Tinbergen's econometric models of national economies, he designed an analogue electrical-mechanical "regression analyzer" in his dissertation and built the prototype at MIT — a machine of circuits and dials that ground out statistical estimates decades before software did .^[2][3] Near the end of his life he wrote of his early fascination with science and his transition from engineering to economics, and the order of those words is right: the engineer never left the economist .^[4]

By his early thirties he had the credentials of a rising star in exactly the tradition he would later attack. In 1949, with Donald Cochrane, he published a method for handling serial correlation in regression — the Cochrane-Orcutt procedure, still taught today .^[5] He directed the Littauer Statistical Laboratory at Harvard. He consulted for the Federal Reserve Board and the International Monetary Fund .^[2] The historian Chung-Tang Cheng, who wrote the intellectual biography, describes the ambition Orcutt carried through those years as a "Tinbergen dream": one model that could capture an entire national economy .^[3]

The dream soured on contact with the instruments. The dominant approach in 1950s economics was macroeconomic modeling — systems of equations describing relationships among aggregates: total consumption as a function of total income, investment as a function of interest rates, employment as a function of output. The Cowles Commission, then the leading center for mathematical economics, was refining Keynesian simultaneous-equation models of exactly this kind, and Milton Friedman was arguing that their forecasts beat naive extrapolation not at all. Orcutt's complaint cut deeper than Friedman's. The models could say something about whether GDP would grow. They could say nothing about what would happen to actual people — which families a recession would break, which a tax change would reach — and that silence covered precisely the questions a government exists to answer. He had spent years refining instruments that answered the wrong question with steadily improving precision.

The aggregation problem

Orcutt located the flaw in the act of aggregation itself — not a limitation of the aggregate models but a mathematical error inside them.

Take a hundred households and a simple tax: 20 percent of income above $10,000. Give every household $15,000 and each owes $1,000 — $100,000 in all. Now split them: fifty households at zero, fifty at $30,000. Average income is still $15,000, so a model built on the average still collects $100,000. Compute household by household instead. The fifty at zero owe nothing; the fifty at $30,000 owe $4,000 apiece; the total is $200,000.

The aggregate is identical; the truth is twice as large.

Whenever the relationship between input and output bends — whenever it is nonlinear — the average of the outcomes is not the outcome of the average. And tax-and-benefit law is nonlinearity in bulk: brackets, thresholds, phase-outs, eligibility tests, cliffs where a dollar of earnings switches off thousands of dollars of support. A rule like that applied to an "average household" produces an answer for a household that does not exist, while the households that do exist land on every side of every threshold. This is the aggregation problem, and Orcutt stated it without mercy: "There is an inherent difficulty, if not practical impossibility, in aggregating anything but absurdly simple relationships about elemental decision-making units."

His alternative was to stop aggregating relationships and start aggregating people. Build the model from "interacting units which receive inputs and generate outputs" ^[1] — individuals, households, firms — each carrying characteristics drawn from real data, each following rules of behavior. "The most distinctive feature of this new type of model," he wrote, "is the key role played by actual decision-making units of the real world such as the individual, the household, and the firm" .^[1] The totals economists cared about would still exist, but they would come out of the model instead of going in: "Predictions about aggregates will still be needed but will be obtained by aggregating behavior of elemental units rather than by attempting to aggregate behavioral relationships."

That inversion is the whole breakthrough. Compute the law on each simulated family; sum the results; the aggregate is whatever the people add up to. Orcutt named the method "microanalytic simulation." Practitioners shortened it to microsimulation, and the name stuck.

A science the opposite of Seldon's

Science fiction got to population-scale prediction first, and got it backwards in a way worth pausing on. Isaac Asimov's Foundation trilogy, published from 1951 to 1953, imagined "psychohistory," a mathematics that predicted the future of a galactic civilization — with a built-in limit. It worked only on masses. "The reaction of one man could be forecast by no known mathematics; the reaction of a billion is something else again" .^[6] Psychohistory was also a monopoly: legible to one man, steered by a secret foundation, invisible to the billions it described.

Orcutt's method runs the other way. It starts with the one man — the one household — and reaches the billion only by addition. Nothing about it requires secrecy or priesthood: if any household can be simulated, then anyone can, in principle, ask how a policy lands on families like their own and check the answer against a life they know. The fictional science was aristocratic by mathematical necessity. The real one is granular and, in principle, democratic. Whether it would be democratic in practice is the story of the next two chapters.

What it was in 1957 was impossible. Computers filled rooms, cost millions, and were programmed with punch cards in assembly language. The IBM 704, state of the art when the paper appeared, managed roughly 12,000 floating-point operations per second; a modern laptop does billions .^[7] Simulating millions of households, each with its own attributes and transitions, sat far beyond the machines — and Orcutt said so himself, with the driest understatement in the paper: "The problem of keeping track of all possible paths and their respective probabilities appears rather appalling." The ambition had company. In the spring of 1950, Jule Charney, Ragnar Fjørtoft, and John von Neumann had run the first numerical weather forecast on ENIAC — twenty-four hours of atmosphere in roughly twenty-four hours of computing, a race the machine barely tied .^[8] At Los Alamos, the MANIAC — the computer Nicholas Metropolis built, on von Neumann's architecture — was simulating neutron physics one random draw at a time with the Monte Carlo method Metropolis, Stanisław Ulam, and von Neumann had devised :^[9] follow the units and sum, applied to particles eight years before Orcutt proposed applying it to people. Herbert Simon and Allen Newell were writing the first artificial-intelligence programs at the Carnegie Institute of Technology. But company is not feasibility.

Orcutt knew it was hard. He published anyway.

First demonstrations

The first act took four years to reach the second. In 1961, with Martin Greenberger, John Korbel, and Alice Rivlin — Rivlin was his doctoral student — Orcutt published Microanalysis of Socioeconomic Systems: A Simulation Study, a book demonstrating a working, if limited, microsimulation of the American economy .^[10] Simulated people were born, died, entered and left the labor force, and earned income through time. It was small and it creaked, but it ran: the 1957 paper's central claim — that you could compute a society from its units — was no longer hypothetical.

Hold on to Rivlin's name. Fourteen years after the book, she became the first director of the Congressional Budget Office, the agency whose estimates now govern every serious fiscal argument in Washington; later she served as vice chair of the Federal Reserve. The line from an obscure 1957 paper to the machinery that scores every major American bill runs, in part, through one advisor and one student.

Government got its own first taste almost immediately, from a different direction. Between 1962 and 1965, a young economist named George Sadowsky brought computers to revenue estimation at the Treasury's Office of Tax Analysis .^[11] In about three months in 1963 — a Yale graduate student, consulting for the Treasury — he built a microanalytic model to analyze preliminary versions of what became the Revenue Act of 1964 .^[12] It was far simpler than Orcutt's dynamic vision: it aged nobody forward, simulated no births or marriages. But it did something no one had done before — it ran a proposed federal tax change against a sample of real taxpayers before enactment, so that Congress could see the consequences of a bill while it was still a bill. By the late 1960s, that kind of analysis was becoming standard in budget work. Sadowsky moved on through Brookings and then to a newly founded Washington think tank called the Urban Institute, where this story will find him again.

Eighteen years, start to finish

Orcutt spent the decade after 1958 trying to scale the demonstration into the real thing, and the decade defeated him. He moved to the University of Wisconsin–Madison and founded the Social Systems Research Institute in 1959, intending to build the full model there. The computing was too thin, institutional support wavered, and the ambition outran what a university lab could sustain. Cheng's biography calls the Wisconsin years a "failed trial" ^[3] — a label worth keeping, because the failure was of capacity, not of the idea, and the distinction decided what happened next.

In 1968 the Urban Institute hired Orcutt to lead the project he had been proposing for eleven years ,^[2] and by 1969 it had the resources to begin in earnest. DYNASIM — the Dynamic Simulation of Income Model — aimed to simulate every major event in American lives: births, deaths, marriages, divorces, schooling, employment, disability, retirement, taxes, benefits. The team built it on a DEC System-10 mainframe inside a custom framework called MASH, written by George Sadowsky, the same economist who had computerized the Treasury's estimates a decade earlier .^[2] to verify The model carried 10,000 simulated people — enough to draw statistical inferences .^[13]

The first version was complete in 1975: eighteen years after the proposal, fourteen after the first demonstration .^[13] Its architecture holds up as a description of the field today — modules organized by domain, one for demographic events, one for the labor market, one for taxes, transfers, and wealth, the whole run as an integrated system in which a simulated recession changes simulated marriages. And it refused to die. The Urban Institute carried DYNASIM through a narrower retirement-income version in the early 1980s, a major overhaul around 2000 that projected the American population seventy-five years forward on the Social Security and Medicare trustees' assumptions, and a fourth generation still active today — among the longest-lived computational models in the social sciences .^[13]

The family it founded

DYNASIM was dynamic in the field's vocabulary: it aged its population forward through time, so you could ask what the retirement system looks like in 2050. A static model asks a nearer question — take today's population, change the law today, and compute who pays what tomorrow morning. Static models are where legislation gets scored, and the agencies followed the need: the IRS, the Congressional Budget Office, and the Treasury all built static microsimulation models of the tax system. (Budget fights would later attach a different meaning to "dynamic" — whether a tax cut changes the size of the whole economy — and the next chapter keeps the two senses apart.)

Around the original, a family grew. Steven Caldwell built CORSIM at Cornell, a direct descendant, which in turn seeded POLISIM for Social Security analysis and DYNACAN for Canadian pensions .^[14] Statistics Canada built SPSD/M, a static tax-and-transfer workhorse with a feature this book will circle back to: it ran on a deliberately synthetic database, records assembled from surveys and administrative data so that no confidential file ever left the agency .^[14] Sweden's SVERIGE modeled the country's entire population .^[14] And in 1974, Joseph Pechman and Benjamin Okner of Brookings merged data on 72,000 households to ask, for the first time with real coverage, who actually bore the burden of American taxes .^[15]

The method also escaped economics entirely. Health researchers project how a population ages, develops disease, and responds to a vaccination program. Climate models pair physical projections with economic ones to trace how warming lands on different households. Transportation planners simulate individual travelers choosing routes and modes; epidemiologists model transmission person to person across contact networks. Wherever aggregates hide the action, someone eventually rebuilds the problem from units.

What the units buy, concretely:

The last three rows are the reason the method exists. They are also the rows every closed institution in the next chapter will monopolize.

What the units turned out to be for

The constraint that made Orcutt's idea absurd in 1957 dissolved almost completely within his lifetime. DYNASIM simulated 10,000 people on a mainframe that cost a fortune to run; a present-day laptop simulates millions; computing power that cost millions of dollars in 1975 now fits in a pocket. Orcutt died on March 5, 2006, having lived from the regression analyzer to the eve of the smartphone .^[16] The work stayed in the family: his daughter, Alice Orcutt Nakamura, became an economist, and his granddaughter, Emi Nakamura, won the 2019 John Bates Clark Medal — awarded to the most promising American economist under forty — for work that answers macroeconomic questions with finely disaggregated data. Three generations spent on the same conviction about averages.

Cheng's summary of what Orcutt built is the right one: microsimulation was "an engine designed for not only scrutinizing the system but reengineering the society" .^[3] But the deepest consequence of the design took fifty years to surface, and it is the one this book is built on. Orcutt wanted units because averages lied. Units turn out to buy something he never advertised: a model assembled one household at a time can be checked one household at a time. This family's income tax can be compared against the statute that sets it, that family's benefit against a determination letter, and only then the sum against the totals a government publishes. The audit can reach all the way down. Aggregate models cannot offer that even in principle — there is no unit to check.

For the first fifty years, hardly anyone could run the check. The models worked, and the institutions that owned them answered to no outside examiner, because outside the institutions there was nothing to examine with. Orcutt's question — can you compute a society from its units? — was answered by 1975. The question that drives the rest of this book is the one his answer created: when the units, the rules, and the checks exist, who gets to hold them?