Camila Comunello (Universidad Torcuato Di Tella) Presenting Gustavo Q. Saraiva (PUC de Chile)
Set of students: \(S\), Set of Colleges: \(C\)
Each college \(c\) has capacity \(q_c\)
Students’ preferences: \(\succ_s \text{ over } C \cup \{\emptyset\}\)
Colleges’ preferences: \(\succ_c \text{ over }S \cup \{\emptyset\}\)
Matching \(\mu\):
Definition: A matching is stable if:
Blocking pair: \((s,c)\) is a blocking pair if \(c \succ_s \mu(s)\) and \(|\mu(c)| < q_c\), or \(|\mu(c)| = q_c\) but \(s \succ_c s'\) for some \(s' \in \mu(c)\).
In other words:
This algorithm yields the best stable matching allocation to students.
A similar algorithm can be implemented in which colleges are the ones proposing (the COSM).
Using data from SISU (Sistema de Seleção Unificada), we find that approximately half of students admitted through the Brazilian centralized college admission system do not enroll during the first phase of the mechanism.
We show the extent to which this generates instabilities and gives students incentives to behave strategically.
We find that the leading cause of low enrollment rates is the existence of admissions off-platform.
Students accepted to top choices are more likely to enroll.
Students from different states are less likely to enroll.
Distortions can be reduced by providing students with more information on enrollment probabilities.
Kapor, Karnani, and Neilson (2024) and De Groote et al. (2025): Document inefficiencies from the Chilean and French college admission systems attributed to off-platform enrollment.
Fabre et al. (2024) and Meisner et al. (2025): Study mistakes made by students (mostly due to irrational behavior)
Barahona, Dobbin, and Otero (2023) and Machado and Szerman (2021): Impact of these affirmative actions on students’ performance/composition.
Aygün and Bó (2021) and Bó and Souza Senkevics (2023): Document inefficiencies in the Brazilian quota system.
Hakimov, Kübler, and Pan (2023) and Bó and Hakimov (2022): Analyze dynamic matching mechanisms.
Saraiva (2021) and Kojima and Pathak (2009): Small number of stable matches in large markets.
Students take the National admission exam (ENEM) in November
In January they learn their test scores
Then they submit their preferences: up to two choices on SISU (the platform)
During 5 days students can adjust their choices. At the end of each day a tentative allocation is provided (the allocation is determined by DA). Students receive this feedback and the only reports that matter in the end are the final ones made by students
Unassigned students select one of their 2 choices to be wait-listed.
Students assigned to their 2nd option cannot be wait-listed to their 1st option.
Only public universities participate in SISU, with many choosing not to participate.
Like in Chile, students apply to a program (e.g., Business at USP).
In Brazil, at least 50% of vacancies are also reserved for Affirmative Action (AA) students.
So, when eligible, students may also apply to an AA-vacancy, which we treat as a different program.
We now describe how, under the current regime, the Brazilian college admission system incentivizes students to behave strategically.
We then discuss some policies that could be implemented to reduce the distortions associated with low enrollment rates.
But we show that when:
The probability that students have incentives to wait until the last minute to report their preferences converges to zero.
But there is another type of manipulation (manipulation via cutoffs): a group of high-scoring students applies to a program only to inflate its cutoff. The inflated cutoff discourages other students from applying. On the last day, the group switches to the programs they really want, and a lower-scoring friend who knew the plan takes the vacated seat.
Therefore, students probably engage in straightforward behavior during the dynamic phase.
But students need to gather information on enrollment probabilities in order to effectively engage in adjusted straightforward behavior.
In the previous example, a student could improve her chances during the waiting list by making other students worse off and, thus, less likely to enroll.
It is, however, also possible that a student may benefit by making other students better off, as doing so could change the configuration of students who get matched to their acceptable choices.
Some high schools have been sued for misleading advertising.
Basically, some high schools that have entrances on 2 different roads can create 2 different CNPJ’s (Cadastro Nacional de Pessoas Jurídicas) using these different addresses.
They then pool their best students into one of these addresses, and rely on this technicality to market a higher average performance on ENEM.
We flag private high schools that engage in this strategy by looking at school names, and checking for similarity in their names.
If two private high schools with similar names were sufficiently close to one another (400m), both schools were flagged as suspicious.
In total, 37 private high schools were flagged.
Several policies should be implemented in parallel to address these issues (e.g., allow private universities to participate in the centralized mechanism).
In this presentation I will focus in one policy: providing students to better access of information on admission probabilities.
As seen earlier, the distortions from low enrollment rates could be mitigated if, during the dynamic phase, students had access to data that allowed them to engage in an adjusted straightforward behavior, where they ignore students to calculate adjusted cutoff grades.
We will see that providing this information to students is feasible.
The proof is similar to the one provided in Bó and Hakimov (2022).
Students have a threshold: if an option in the centralized mechanism falls below this threshold, the student does not enroll.
Students, however, attempt to get the best possible match in the centralized mechanism, even if it involves choosing options that fall below their threshold (this is because better outside options may not be known at the time).
The platform informs students about these students who do not enroll, who are then treated as “phantoms”.
A version of the following app could be provided to students to help them make more accurate predictions about program attainability:
In the limit, if students can perfectly predict how many spots will actually be taken, the final allocation should be stable.
Low enrollment rates plague the Brazilian college admission system, generating instabilities and giving students incentives to behave strategically
The main culprit are off-platform enrollments
One policy that could effectively mitigate this issue is the provision of better information by the matchmaker
AI applied in conjunction with students’ microdata (e.g., whether students are currently enrolled) could substantially reduce justified-envy.
Private Universities should be allowed and given incentives to participate in the centralized mechanism (e.g., provision of scholarships).
Because students accepted to lower options are less likely to enroll, changing the mechanism to a standard unrestricted one-shot DA could be catastrophic.
Gustavo Q. Saraiva