For sixty years, economists have written about college admissions as a matching market. For ten years, one paper has defined what an agent-based model of that market looks like. The field is smaller than you would expect — and most of what has happened to American admissions since 2016 is missing from it.
The literature begins, as much of modern microeconomics does, with a 1962 paper that turned a marriage metaphor into mathematics. Gale and Shapley showed that a stable matching always exists between two sides with preferences, and that a deferred-acceptance algorithm produces the student-optimal match.
Forty years later, Abdulkadiroglu and Sonmez carried that machinery into school choice and showed why some real mechanisms invite manipulation. Epple, Romano and Sieg added an equilibrium framework for tuition and aid. Together these three papers fix a benchmark: a frictionless market with truthful preferences and need-based aid.
Real American admissions departs from that benchmark in every direction at once. Holistic review introduces noise. Binding ED distorts truthful preference revelation. Hooks — legacy, athlete, donor — decouple academic quality from institutional preference. The agent-based literature exists to model what theory cannot.
Then, in 2016, four researchers built the model that everyone else now cites.
Reardon, Kasman, Klasik and Baker published Agent-Based Simulation Models of the College Sorting Process in the Journal of Artificial Societies and Social Simulation. The architecture is austere: 8,000 students, 40 stylised colleges, three stages per simulated year — apply, admit, enroll — iterated to equilibrium over thirty years.
Each student carries two attributes: caliber (a noisy academic signal) and resources (an SES proxy). Each college carries one: quality, updated each year as the rolling average caliber of recently enrolled students. From those minimal pieces, a sorting hierarchy emerges that matches IPEDS patterns.
The methodological contribution is the portfolio algorithm. Students do not enumerate every possible application set; they select colleges in decreasing marginal expected utility until adding another no longer helps. The process is O(N·K) instead of combinatorial — cheap enough to run a hundred Monte Carlo replications per parameter cell.
The model's purpose was not to recreate admissions. It was to take SES apart.
Why do low-SES students sort into less selective colleges even after controlling for academic achievement? Reardon et al. proposed five candidate mechanisms and removed them one at a time, then in combination. The bar chart at right shows what happens to the 90-10 enrollment gap when each is shut off.
Resource-caliber correlation dominates. Removing it alone reduces the top-decile advantage from roughly twenty-fold to four-fold. The simple gap in academic achievement does most of the work.
But not all of it. The four behavioral mechanisms — application enhancement, information quality, application volume, and a near-zero contribution from utility preferences — together approximately equal the achievement gap's effect.
Equalising access to counselling, information and application resources could achieve roughly half of the stratification reduction that closing the achievement gap would. — Implication of Reardon, Kasman, Klasik & Baker (2016), Section 2.4
Two years later, the same team added the dimension American politics most demanded: race.
Reardon, Baker, Kasman, Klasik and Townsend (2018) extended the framework with a race attribute and ran a policy counterfactual that would become directly relevant five years later: can socioeconomic affirmative action substitute for race-based?
The answer in the model: not on its own. SES-based AA admits more low-income students, but because race and income are only partially correlated, it misses a large fraction of middle-class Black and Hispanic applicants who would benefit from race-based AA. Combining SES-based AA with race-targeted recruiting closes part of the gap. Removing affirmative action entirely produces a sharp drop in URM representation.
A spillover finding has aged especially well. When elite colleges shift to SES-based AA, they pull high-achieving low-SES students up the quality distribution, reducing diversity at the next tier of colleges that don't follow suit. The cascade goes downhill.
Outside the United States, the literature is still thinner.
Citations to Assayed et al. 2024 conflate three distinct papers from The British University in Dubai. The original (Assayed and Maheshwari, 2023a) is a NetLogo model of Jordanian medical college admissions with two student attributes — GPA and family income — and a single application round. The 2023b and 2025 papers are surveys, not new simulations.
The model is interesting for what it shows about emergent cutoffs: when low-income high-GPA students are prioritised, college reputation rankings reshuffle from cutoff GPA and student preference rather than from prestige. But the feature gap with the U.S. context is large — no ED, no hooks, no real college data, no validation against published admit rates.
Daemen and Leoni's 2025 Netherlands ABM adds peer effects, geography, and a wages-versus-loans counterfactual, but none of the U.S.-specific machinery. The international literature is structurally different and structurally smaller.
Most of the recent action has been in calibration, not architecture.
Three datasets define what a credible American ABM has to match. Chetty, Deming and Friedman (2023) linked 2.4 million tax records to admissions outcomes at 139 colleges. Arcidiacono, Kinsler and Ransom (2022) worked from the Harvard files disclosed during SFFA litigation. Avery and Levin (2010) documented the ED signaling premium long before binding rounds were a third of an Ivy class.
The Harvard numbers are the starkest. Forty-three percent of white admits are athletes, legacies, dean's-interest, or children of faculty. The athlete admit rate is 86 percent; the non-ALDC rate is below 5.5 percent. Roughly three quarters of white ALDC admits would have been rejected without the preference.
Chetty's headline figures speak to where the prestige premium comes from. Same SAT, top-1-percent income versus middle-class: the wealthier student is twice as likely to attend Ivy+. Forty-six percent of that advantage runs through legacy admissions; twenty-four percent through athletic recruitment; thirty percent through non-academic ratings of essays, recommendations and interviews. The three preference channels are uncorrelated with post-college outcomes.
All of this empirical detail was missing from the foundational ABM. So we built the gap.
The chart at right scores three models — Reardon 2016, Assayed 2023a, and college-sim — on twenty-one features. Green dots indicate full implementation; faded dots indicate partial or implicit treatment; absent dots indicate the feature is missing.
Reardon's strengths are conceptual: a clean five-mechanism decomposition, a robust portfolio algorithm, and Latin Hypercube sensitivity testing. Its silences are institutional: no ED rounds, no hooks, no financial aid, no waitlist, no real colleges, no validation beyond IPEDS aggregates.
Assayed adds NetLogo as a platform but loses scope — a single round, two student attributes, no validation, no archetypes. The college-sim project fills the institutional gaps with real CDS data, six rounds, nine calibrated hooks, archetype-based application behavior, and Chetty-derived income-bracket yield.
Filling those gaps lets us run the questions the literature couldn't.
The thinness of the field is also an opportunity. The 2016 review identifies five questions a richer ABM is positioned to answer — questions Reardon's stylised model couldn't, and that observational data alone can't.
How does the post-SFFA equilibrium evolve as colleges adapt? Does binding ED break the application-inflation spiral Sirolly described, by locking commitment in early? What happens to yield and class composition when MIT, Hopkins or Amherst eliminates legacy preferences unilaterally? Can targeted outreach reach Hoxby and Avery's missing low-income applicants? And how far does a single elite waitlist call cascade through the four tiers below it?
Each of those is a counterfactual that requires hooks, multi-round processing, real college parameters, and archetype heterogeneity simultaneously. The model the literature needs is the model the literature does not yet contain.
The American higher-education complex is a multi-trillion-dollar institution that touches roughly two million eighteen-year-olds every spring. The academic agent-based literature on its admissions process, ten years after Reardon, comprises perhaps a dozen papers and one canonical model.
That asymmetry is partly methodological: ABMs are demanding to specify, calibrate and validate. It is partly institutional: matching-theory economics rewards proofs more than simulations. And it is partly historical — the data needed to calibrate hooks did not exist in public form until SFFA forced disclosure in 2018-2019, and the Chetty linkage of tax records to college outcomes only landed in 2023.
The gap between what the literature contains and what the system now does is the working space for the next decade of admissions modeling. Six rounds, nine hooks, three hundred real colleges, eight archetypes, one post-SFFA legal landscape — that is roughly the scope of model the system requires. It is roughly the scope of college-sim.