Policy Work

For many decades Brazil’s federal universities—publicly funded and tuition-free—were attended predominantly by students from the country’s socio-economically privileged, largely white minority. In 2012, Congress approved the Federal Quotas Law (Lei 12.711/2012), requiring that at least half of all seats be reserved for graduates of public high schools, with racial and income-based sub-quotas inside that 50%.

While still Ph.D. students at Boston College, Orhan Aygün and I dissected the implementation rules and discovered a paradox: under the then-prevailing procedure, some Black or low-income applicants could be rejected for quota seats even when their ENEM scores exceeded those of white or high-income students admitted to “open” seats. The problem lay in forcing each applicant to commit ex-ante to a single sub-quota, which created unequal cut-offs across seat classes.

In College admissions with multidimensional reserves: the Brazilian affirmative action case (American Economic Journal: Microeconomics, 2021), we proposed a minimal fix: rank every candidate first in the open-seat list and, if not admitted there, cascade them to each reserved list for which they qualify—an ordering of slot-specific priorities that guarantees no protected applicant is ever displaced by a non-protected one with a lower score. We also documented that in nearly half of all programmes the original rule had indeed produced “unfair rejections.”

Academic reception was enthusiastic but political uptake was slow. A 2015 visit to the Ministry of Education (MEC) yielded no traction. By 2022, however, momentum shifted. At the initiative of Representative Tábata Amaral, discussions with economist Úrsula Mello and INEP researcher Adriano Senkevics led to a policy-oriented working paper that quantified the damage: in the first 2019 SISU round alone, > 10,000 quota-eligible students were wrongly denied admission, 8,000 of whom received no offer anywhere. The paper triggered national press coverage and social-media pressure on MEC.

Throughout 2023 I collaborated with MEC’s Secretaria de Educação Superior, supplying simulations for cabinet meetings and even a joint session in the Casa Civil. In August, Deputy Dandara—herself a former quota beneficiary—brandished our paper on the House floor moments before the vote on a new Quotas bill.

Starting in the national selection process for public universities of 2024, the method used for implementing the quotas now follows our proposal. With this, the possibility of unfair rejections is now completely eliminated, and its more than 1.2 million candidates can safely declare all of their affirmative action-related characteristics.

Governments hire differently from private firms. A company that wants two software engineers can interview a handful of people and pick whoever it likes. A government usually cannot. To curb favouritism, reduce corruption, and make hiring decisions defensible to the public—who tend to view permanent civil-service jobs almost as a public asset—most countries fill these posts through formal, multi-stage selection processes with criteria fixed in advance, often by law. The same basic recipe shows up in Brazil, the UK, France, India, and China: written and multiple-choice exams, qualification checks, sometimes physical or psychological tests, and a published ranking that determines who is hired.

In Brazil this process is called a concurso público. The structure is two-stage by necessity. Everyone takes a multiple-choice exam, which a computer can grade for hundreds of thousands of candidates at once. The next step is a written exam, which has to be graded by hand by a small army of examiners; doing that for every candidate is impossibly expensive. So the multiple-choice exam is used to screen: the top scorers—typically a fixed multiple of the number of vacancies, say three times—have their written exams graded, and the final ranking combines both scores. The cutoff is transparent, the rule is meritocratic, and the number of written exams to grade is known in advance. Until recently, every ministry and agency ran its own concurso for its own jobs, on its own calendar.

The Concurso Público Nacional Unificado (CPNU), introduced in 2024, replaced that patchwork with one nationwide exam. Its first edition drew over 2 million registered candidates competing for 6,640 vacancies across 173 jobs in 21 federal bodies—a logistical operation involving 350,000 supporting staff across the country. Jobs were grouped into eight thematic blocks (infrastructure and engineering, environment and agriculture, health, and so on); each block had its own exam, and a candidate chose one block to compete in. Within a block, a candidate could rank many jobs at once, and each job decided how to weight the parts of the exam: the same raw answers could place a candidate near the top for a legal post and far lower for a statistical one.

This is where the screening problem stops being routine. With one job and one ranking, calling “the top 3N” for grading is obviously fine. With dozens of jobs that share candidates and weight the exam differently, the question becomes: which candidates should we grade, knowing that the eventual matching of candidates to jobs will be done across all of them at once?

The reason it matters is the fairness rule the system is required to satisfy. Each candidate should be assigned to the job they most prefer among those for which their final score is high enough—and no candidate should be passed over for a job by someone with a lower final score for it. (This is the standard stability requirement from matching theory.) The catch is that the final scores depend on the written exams, which we have only graded for a subset of candidates. If we grade the wrong subset, then once the scores come in we can find ourselves in this kind of situation:

Suppose two jobs each have one vacancy, and we grade just two candidates, A and B. A prefers job 1; B also prefers job 1 but, by qualification rules, can only take job 1. After grading, A turns out to score higher than B at job 1. Stability now forces us to assign A to job 1 and leaves B unmatched—and job 2 unfilled, even though somewhere in the pool of un-graded candidates there was someone who would have qualified and wanted it.

The point of the example is that whether a job ends up empty depends on how the written-exam scores fall, which we cannot see at the moment we choose whom to grade. So the screening decision has to be robust to all the ways the final scores might come out. We need to bring enough of the right candidates into the graded pool that, no matter how the written exams turn out, the fair assignment still fills every job and respects every job’s qualification and quota rules.

Independent per-job calls collapse under this. When each job picks its own top scorers for grading, the same high scorers show up on list after list and crowd out the distinct candidates that other jobs, specialities, and reserved vacancies need. In our simulations on CPNU data, calling nine times the number of vacancies per job yielded only about 1.5 times the total vacancies in distinct graded candidates.

The procedure I designed treats screening and allocation as one coordinated problem. The first stage selects who will sit the written exams (and undergo title evaluation and other later assessments) while preserving enough qualified candidates for each job, quota category, and speciality to keep the eventual fair assignment feasible. The second stage takes the graded pool and assigns each candidate to at most one job: the most preferred job for which they have sufficient priority. The same logic governs the waiting list, so the government can handle later vacancies and movements between jobs in one coordinated pass instead of through slow sequential calls.

Affirmative action sits inside the procedure rather than on top of it. Reserved vacancies for Black, disabled, and Indigenous candidates enter the selection for grading, the final allocation, and the waiting list on the same terms as the rest of the design. Quotas have to coexist with transparency, merit-based priority, job-specific qualifications, and candidates’ stated preferences, and the procedure is built so they do.

The broader contribution is to apply market design to a large-scale public-administration problem in which information is costly, decisions are sequential, and small errors leave jobs unfilled or candidates treated inconsistently. Candidates receive one instruction: rank the jobs in the order you actually prefer. The government, in turn, gets a procedure that fills jobs, respects qualifications and quotas, and can explain to any candidate why they were or were not selected.