Francis-Tan, Tannuri-Pianto - 2018 - Black Movement Using Discontinuities in Admissions To Study The Effects of College Quality and Affi

Accepted Manuscript
Black Movement: Using discontinuities in admissions to study the effects of college

quality and affirmative action
Andrew Francis-Tan, Maria Tannuri-Pianto
PII: S0304-3878(18)30486-3
DOI: 10.1016/j.jdeveco.2018.06.017
Reference: DEVEC 2259
To appear in: Journal of Development Economics
Received Date: 4 September 2017

Revised Date: 20 June 2018
Accepted Date: 28 June 2018
Please cite this article as: Francis-Tan, A., Tannuri-Pianto, M., Black Movement: Using discontinuities
in admissions to study the effects of college quality and affirmative action, Journal of Development
Economics (2018), doi: 10.1016/j.jdeveco.2018.06.017.
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and all
legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
Black Movement:
Using discontinuities in admissions to study the effects of
college quality and affirmative action
Andrew Francis-Tan∗
PT
National University of Singapore
Maria Tannuri-Pianto
RI
University of Brasilia
The recent adoption of race-targeted policies makes Brazil an insightful place to study
SC
affirmative action. In this paper, we estimate the effects of racial quotas at the University of
Brasilia, which reserved 20% of admissions slots for persons who self-identified as black. To do
so, we link the admissions outcomes of high-performing applicants in 2004-2005 to their
U
education and labor market outcomes in 2012. We adopt methods that make use of sharp
discontinuities in the admissions process. In summary, the policy of racial quotas mostly
AN
improved outcomes for the targeted group. Quota applicants, specifically males, enjoyed an
increase in years of education, college completion, and labor earnings. More broadly, the results
for quota and non-quota applicants confirm the importance of college quality in a setting outside
of the U.S.
M
Keywords: affirmative action, racial quotas, mismatch, educational policy, college quality,
minorities, Brazil.
D
JEL codes: I23, I25, I26, J15, J24.

TE
C EP
AC
* Andrew Francis-Tan, PhD, Visiting Associate Professor at the Lee Kuan Yew School of Public Policy, National
University of Singapore, 469B Bukit Timah Road, #02-02, Singapore 259771; andrew.francis@nus.edu.sg; +65
6601 1179 (corresponding author). Maria Tannuri-Pianto, PhD, Associate Professor of Economics at the University
of Brasilia (tannuri@unb.br). We appreciate the many insightful comments that we have received from the Editor,
anonymous reviewers, and conference participants at the meetings of the Society of Household Economics (2017),
Population Association of America (2017), LACEA Labor Network (2017), Pathways to Social and Economic
Welfare (British Council Researcher Links, 2016), and Economics and Politics Research Group (University of
Brasilia, 2016) as well as seminar participants at the Lee Kuan Yew School of Public Policy (2016) and the
University of São Paulo (2017).
1
ACCEPTED MANUSCRIPT
Black Movement:
Using discontinuities in admissions to study the effects of
college quality and affirmative action
Andrew Francis-Tan∗
PT
National University of Singapore
Maria Tannuri-Pianto
RI
University of Brasilia
The recent adoption of race-targeted policies makes Brazil an insightful place to study
SC
affirmative action. In this paper, we estimate the effects of racial quotas at the University of
Brasilia, which reserved 20% of admissions slots for persons who self-identified as black. To do
so, we link the admissions outcomes of high-performing applicants in 2004-2005 to their
U
education and labor market outcomes in 2012. We adopt methods that make use of sharp
discontinuities in the admissions process. In summary, the policy of racial quotas mostly
AN
improved outcomes for the targeted group. Quota applicants, specifically males, enjoyed an
increase in years of education, college completion, and labor earnings. More broadly, the results
for quota and non-quota applicants confirm the importance of college quality in a setting outside
of the U.S.
M
Keywords: affirmative action, racial quotas, mismatch, educational policy, college quality,
minorities, Brazil.
D
JEL codes: I23, I25, I26, J15, J24.

TE
C EP
AC
* Andrew Francis-Tan, PhD, Visiting Associate Professor at the Lee Kuan Yew School of Public Policy, National
University of Singapore, 469B Bukit Timah Road, #02-02, Singapore 259771; andrew.francis@nus.edu.sg; +65
6601 1179 (corresponding author). Maria Tannuri-Pianto, PhD, Associate Professor of Economics at the University
of Brasilia (tannuri@unb.br). We appreciate the many insightful comments that we have received from the Editor,
anonymous reviewers, and conference participants at the meetings of the Society of Household Economics (2017),
Population Association of America (2017), LACEA Labor Network (2017), Pathways to Social and Economic
Welfare (British Council Researcher Links, 2016), and Economics and Politics Research Group (University of
Brasilia, 2016) as well as seminar participants at the Lee Kuan Yew School of Public Policy (2016) and the
University of São Paulo (2017).
1
ACCEPTED MANUSCRIPT
1. Introduction
Affirmative action policies – which aim to promote the status of historically
disadvantaged groups in educational and labor market settings – have roots in 1960s America
(Holzer and Neumark, 2000; Arcidiacono and Lovenheim, 2016). After decades of expansion,
PT
court rulings and state-specific laws have diminished the scope of these policies since the 1990s.
RI
While affirmative action may be on the decline in the United States, it is on the rise in other
countries including India, South Africa, Israel, China, Malaysia, and Brazil (Darity et al., 2011).
SC
Most of the scholarly literature concerns the U.S. Yet, contexts outside the U.S. are better suited
for identifying the causal effects of Educational Equity Policies.1
U
We study affirmative action in Brazil. About 10 times more slaves arrived in Brazil than
AN
in British Mainland North America (Eltis, 2001). For this reason, Brazil has had a large black
and mixed-race population. In 2014, about 45.5% of the population of 203 million was "branco"
M
or light-skinned, 45.0% "pardo" or brown-skinned, and 8.6% "preto" or dark-skinned (IBGE,

D
2015). Despite persistent disparities along the lines of skin color, most public policies in Brazil
TE
have been race-blind. However, in the 2000s, several universities began to adopt race-based
affirmative action in admissions.

EP
In this paper, we estimate the effects of racial quotas at the University of Brasilia (UnB),
which reserved 20% of admissions slots for persons who self-identified as black. To do so, we
C
link the admissions outcomes of high-performing applicants in 2004-2005 to their education and
AC
labor market outcomes in 2012, when they are 30 years old, on average. The university provided
complete admissions records for these applicant cohorts, and the government provided
comprehensive data on all formal workers in the country. Note that an applicant's entrance exam
1
Internationally, the term "Educational Equity" may be more fitting than "Affirmative Action." The term is less US-
centric and does not carry the connotations built up over years of national political tensions.
2
ACCEPTED MANUSCRIPT
score is the primary basis for admission, and the minimum score for admission differs by major.
We focus on applicants who obtained an entrance exam score "close" to their major-specific
cutoff. Nearly all the applicants in our estimation sample ranked above the 85th percentile in the
full applicant pool. We adopt regression methods to compare the education and labor market
PT
outcomes of high-performing applicants above and below admissions cutoffs. We also adopt
RI
regression methods to determine whether returns to admission vary according to the degree of
"mismatch" in academic credentials between quota and non-quota students upon matriculation.
SC
In models with both genders pooled, we find that quota applicants experience much
larger gains in educational outcomes than non-quota applicants. Relative to quota applicants not
U
attending, quota applicants attending UnB have significantly more years of education and are
AN
more likely to complete college. Additionally, attending UnB raises the likelihood that quota
applicants work as a director or manager. Attending UnB also raises the likelihood that non-
M
quota applicants obtain formal employment as well as a public sector job. When both genders are
D
pooled, UnB attendance significantly raises the labor earnings of non-quota applicants but not of
TE
quota applicants. Nevertheless, patterns are strikingly dissimilar by gender, as results reveal the
gains are mostly concentrated among male applicants. Male quota applicants attending UnB have
EP
significantly higher educational outcomes, while female quota applicants do not. Moreover,
attendance significantly raises the earnings of both male quota applicants and male non-quota
C
applicants but does not raise the earnings of female applicants. Further analyses demonstrate that
AC
some differences in gains between males and females exist in both selective and non-selective
majors, especially for quota applicants, and that, in some specifications, mismatch in academic
credentials is associated with a sizable reduction in the benefits of admission for female quota
students.
3
ACCEPTED MANUSCRIPT
Several features distinguish our study from the literature. Previous studies of college
quality and affirmative action primarily concern the U.S., where the admissions process is
complex and researchers do not observe all relevant variables. Regardless of their geographic
focus, few studies investigate the impact of affirmative action on labor market outcomes, and
PT
those that do typically rely on samples of individuals who self-report earnings. In our context,
RI
admissions is simple and our admissions data is complete. This enables the use of regression
discontinuity (RD) design as a method to estimate the effect of college attendance. Additionally,
SC
we have exceptional access to administrative data on labor market outcomes. In sum, this study
estimates the effects of affirmative action in college admissions on both education and labor
U
market outcomes, implementing an RD design with administrative data. Studying affirmative
AN
action in Brazil is valuable not only because of the methodological advantages but also because
affirmative action is new, many people are affected, and much is at stake.
M
The remainder of the article is organized as follows. Section 2 reviews related studies,
D
Section 3 describes the data and methods, Section 4 presents the results, and Section 5 discusses
TE
the conclusions.
EP
2. Related studies
2.1 Effect of affirmative action on educational outcomes

C
Most previous studies on affirmative action in college admissions focus on educational

AC
outcomes.2 These studies concern the academic performance of underrepresented minorities and
their fit with their institution of study. At the center of debate in the literature is the mismatch
hypothesis, which contends that "many of the beneficiaries of preferences are so misplaced
2
An interesting literature investigates the effect of eliminating affirmative action in the U.S. (e.g., Antonovics and
Backes, 2014; Backes, 2012; Cortes, 2010; Dickson, 2006; Hinrichs, 2012).
4
ACCEPTED MANUSCRIPT
academically that they would actually be better off in the absence of affirmative action"
(Arcidiacono and Lovenheim, 2016). Some studies of U.S. colleges do not find mismatch (Alon
and Tienda, 2005; Fischer and Massey, 2007; Massey and Mooney, 2007). For example, Alon
and Tienda (2005) use data on students who attended college in the 1980s and early 1990s. They
PT
claim that mismatch can be rejected if statistically similar underrepresented minorities are
RI
equally or more likely to graduate from selective institutions than from non-selective institutions.
In models that account for selection on unobservables, attending a selective institution appears to
SC
raise the likelihood that black and Hispanic students graduate within six years of entering
college. Thus, the authors find little evidence of mismatch. In an important recent publication,
U
Arcidiacono et al. (2016) study the science graduation rates of minorities enrolled in the
AN
University of California system during the late 1990s. To do so, they estimate a model of student
persistence in college majors and graduation. They account for possible selection bias associated
M
with choice of UC campus by controlling for the set of campuses to which a student applied and
D
the set of campuses which accepted him or her. They find that less prepared minorities at higher
TE
ranked campuses would have had higher science graduation rates if they counterfactually had
attended lower ranked campuses.

EP
Not only does the literature seek to estimate the effects of affirmative action in U.S.
colleges but also U.S. law schools. Sander's 2004 study on affirmative action is a seminal
C
contribution in this area. He draws on the Bar Passage Study, a national longitudinal survey of
AC
persons who entered law school in 1991. He characterizes the extent to which racial preferences
impact the sorting of black students into law schools and documents large disparities in academic
outcomes between black and white law students. Sander claims that mismatch is so pernicious
that eliminating affirmative action in law schools would actually raise the total number of black
5
ACCEPTED MANUSCRIPT
lawyers. Rothstein and Yoon (2008a, 2008b) use the same data as Sander but reach different
conclusions. Their methodology involves comparing blacks and whites with similar admissions
credentials. Differences in outcomes may be attributed to mismatch as blacks tend to attend
higher-ranked schools due to admissions preferences. Their findings reveal no evidence of
PT
mismatch for blacks with higher credentials. However, blacks with relatively low admissions
RI
credentials do underperform, which may be due to mismatch or another explanation. Rothstein
and Yoon conclude that the number of black lawyers would fall substantially if affirmative
SC
action were eliminated.
Fewer studies examine affirmative action outside of the United States (Alon and
U
Malamud, 2014; Francis and Tannuri-Pianto, 2012, 2013; Frisancho and Krishna, 2016; Bagde et
AN
al., 2016). For example, Bagde et al. (2016) consider India where affirmative action in
admissions favors individuals from underrepresented castes. They have access to data on
M
applicants to over private 200 engineering colleges in one Indian state. The data include college
D
entrance exam scores matched with high school exam records as well as graduation records and
TE
performance on a college exam administered at the end of the first year for those students who
chose to enroll at an engineering college. The findings indicate that for targeted students,
EP
affirmative action raises college attendance, boosts first-year academic achievement, and does
not adversely impact graduation rates.

C
Francis and Tannuri-Pianto (2012, 2013) examine the racial quotas policy at the
AC
University of Brasilia. They focus on students who matriculated between 2003 and 2005, a
period including two admissions cycles before and three admissions cycles after the
implementation of quotas. The authors obtained administrative records as well as conducted a
detailed survey with student photos. They report evidence that racial quotas increased the
6
ACCEPTED MANUSCRIPT
proportion of black students, and that displacing applicants were, by many measures, from
families with lower socioeconomic status than displaced applicants. Difference-in-difference
regressions suggest that the policy did not reduce the pre-university effort of applicants or
students and did not exacerbate racial disparities in college academic performance.
PT
RI
2.2 Effect of affirmative action on labor market outcomes
The literature on returns to college quality is relevant (Andrews et al., 2016; Brewer et
SC
al., 1999; Dale and Krueger, 2002, 2014; Black and Smith, 2004, 2006; Hoekstra, 2009). In sum,
these studies demonstrate the positive economic returns to college quality. Illustrative of this
U
point, Hoekstra (2009) adopts a regression discontinuity design to estimate the effect of attending
AN
a U.S. flagship state university on the earnings of white men who applied in the late 1980s. He
finds that attending a selective state university increases earnings by about 20%. Dale and
M
Krueger (2014) link the college characteristics of persons who entered college in 1976 and 1989
D
with their earnings in 2007. The returns to attending a selective college are close to zero in
TE
models that adjust for selection. However, returns are large for blacks, Hispanics, and those
whose parents have relatively low education. Thus, it can be inferred that affirmative action may
EP
raise earnings by moving minorities from lower to higher quality colleges.
An emerging, but related, literature examines heterogeneity in the returns to college

C
degree. In the context of Chile, Hastings et al. (2014) estimate the long-run effects of college
AC
degree on earnings and investigate whether these returns vary by degree selectivity or student
background. Like students in Brazil, students in Chile apply simultaneously to a university and
field of study (which together constitute a "degree"). Linking college application records for
students applying during 1982-2011 with their tax records for 2005-2012, the paper adopts a
7
ACCEPTED MANUSCRIPT
regression discontinuity design. In doing so, it is the first paper to use many discontinuities to
identify the returns to college by field of study and by selectivity. The paper finds large positive
returns to selective degrees and to degrees in health, science/technology, and social science.
Moreover, for the most part, returns to selectivity do not vary by student socioeconomic status
PT
(SES). Therefore, earnings gaps across fields are not merely an artifact of selection.
RI
Other studies explore the labor market effects of affirmative action more directly
(Arcidiacono, 2005; Loury and Garman, 1993, 1995; Wydick, 2002). Loury and Garman (1993,
SC
1995) use data on the high school class of 1972 to investigate the effect of attending a selective
institution on the earnings of black and white males. Among other results, their analysis shows
U
that blacks enjoy larger returns to college selectivity, but that these gains are offset when their
AN
SAT scores are much lower than the median SAT of the college they attend. For this reason, the
authors speculate that "mismatched" blacks would have had higher earnings if they had attended
M
less selective colleges. In a seminal contribution, Arcidiacono (2005) estimates a forward-

D
looking dynamic discrete choice model in order to examine the relationship between racial
TE
preferences in admissions and labor market earnings. Arcidiacono reports that the counterfactual
elimination of affirmative action would only slightly reduce the earnings of blacks. The effect is
EP
modest because affirmative action mostly impacts where students attend college, not if they
attend, and because the returns to college quality are relatively low.
C
Not only does the literature examine racial preferences in U.S. college admissions but
AC
also in U.S. graduate school admissions. In his 2004 study, Sander analyzes the labor market
outcomes of black law students in addition to their academic outcomes. He regresses earnings on
law school prestige, law school GPA, race, and other covariates. Among other results, he finds
that the returns to prestige are smaller than the returns to GPA. He concludes that despite the
8
ACCEPTED MANUSCRIPT
increase in prestige that blacks receive, affirmative action must lower black earnings because
they have substantially lower grades in law school. Rothstein and Yoon (2008a, 2008b) take
issue with Sander's findings. Regressing labor market outcomes on race and law school
credentials, they argue that any racial differences in outcomes could possibly be attributed to
PT
mismatch. However, they do not find any evidence of black underperformance in the labor
RI
market. Grove and Hussey (2014) adopt Rothstein and Yoon's estimation strategy to examine
mismatch in U.S. MBA programs. Exploring the question with longitudinal data on individuals
SC
who registered for the GMAT in the 1990s, they do not find any mismatch effects for blacks or
Hispanics.
U
Few studies concern the effect of affirmative action on labor market outcomes outside of
AN
the U.S. (Bertrand et al., 2010; Frisancho and Krishna, 2016). The most related paper is by
Bertrand et al. (2010) who investigate the effects of an affirmative action policy that reserves
M
more than 50% of admissions slots for lower-caste groups at engineering colleges in one Indian
D
state. Engineering college aspirants in the state must pass a first-round exam and take a second-
TE
round exam. Beginning with those with the highest second-round scores in each caste,
individuals are invited to decide whether to attend an engineering college, which one to attend,
EP
and which major to study. The authors had exam scores and personal information for applicants
who took the second-round exam in 1996. To obtain information on labor market outcomes, the
C
authors conducted a survey of applicants roughly 8-10 years after they took the entrance exam.
AC
In the analysis, they regress earnings on engineering college attendance and controls including
gender, age, parent education, and so on. A binary indicator for having an exam score above the
admissions cutoff is used as an instrument for college attendance. The results demonstrate that
affirmative action successfully targets the economically disadvantaged, although the policy may
9
ACCEPTED MANUSCRIPT
reduce the number of females in engineering colleges. Lower-caste applicants obtain positive
returns to admission, but income gains for displacing applicants appear to be smaller than
income losses for displaced applicants.
PT
3. Data and methods
RI
3.1 Policy context
Brazil has been a nation historically characterized by race-blind public policies despite
SC
persistent disparities along the lines of skin color. Myriad factors led to the recent adoption of
race-based affirmative action: increasing social awareness of racial inequality, formal
U
acknowledgement of race issues by the Cardoso administration (1995-2003), and rise of the
AN
Black Movement, a social movement working to secure rights for black Brazilians (Bailey,
2009). In 2001, the Ministry of Agrarian Development became the first federal ministry to adopt
M
racial quotas in employment.3 Although a number of universities already had quotas based on
D
income or public school attendance, it was not until 2001 that two state universities in Rio de
TE
Janeiro became the first to adopt quotas based on race. Other universities followed, including the
University of Brasilia.
EP
The University of Brasilia (UnB), a tuition-free public institution, is one of the best
universities in Brazil. It is located in Brasilia, a metropolitan area of 3 million and the nation's
C
capital. Undergraduates are admitted biannually through the "vestibular" system. Applicants
AC
select one major and take an entrance exam specific to UnB (the vestibular). It is a two-day exam
with questions on a variety of subjects. Applicants who apply in the same admissions cycle take
the same exam, regardless of major. The overall score on the vestibular is the primary basis for
3
Note that there is no affirmative action in the labor market generally.
10
ACCEPTED MANUSCRIPT
admission, and the minimum score for admission differs by major.4 Applicants are either
admitted to their selected major or they are rejected. The acceptance rate varies widely by major.
For example, the acceptance rate for medicine is roughly 1%, while the rate for economics is
10%. Approximately 84% of those applicants who are admitted choose to matriculate at UnB.
PT
This yield rate is slightly higher than the yield rate at Ivy League colleges in the U.S.
RI
In July 2004, the University of Brasilia implemented racial quotas, becoming the first
federal university in the country to have a race-targeted admissions policy.5 UnB's policy of
SC
racial quotas was adopted by the administration without public vote or debate. 20% of each
major's vestibular admissions slots are reserved for applicants who identify as black (negro).
U
Most pretos and some pardos (typically pardos who are darker-skinned) self-identify as negro.
AN
To be considered for these reserved slots, applicants must opt into the quota system, identify as
black, and select one major when they register for the entrance exam. Admitted quota applicants
M
are interviewed by a university panel in order to verify black identity. Upon matriculation, quota
D
students have access to university-sponsored programs intended to support their academic and
TE
social development, including tutoring services, public seminars on the value of blacks in
society, and a campus meeting space to study and interact.

EP
The policy is implemented by allowing admissions cutoffs for quota and non-quota
applicants to differ within the same major. In principle, this might produce differences in
C
academic credentials between quota and non-quota students. Table 1 displays admissions
AC
statistics by major and quota status for students who were admitted between 2004 and 2005.
Majors are sorted by selectivity rank, which is based on the average non-quota cutoff score
4
As part of the exam, applicants write an analytical essay, which is only graded if their overall score qualifies them
for admission. About 1.5% of applicants who have scores above the admissions cutoff are not admitted due to poor
essays.
5
Quotas at UnB are featured in an excellent PBS documentary entitled "Brazil in Black and White."
11
ACCEPTED MANUSCRIPT
during this time period. More selective majors include medicine, law, and various engineering
majors. Less selective majors include humanities and teaching degree majors. The table
compares the UnB acceptance rate between quota and non-quota students. The acceptance rate
for quota students exceeds that for non-quota students for almost every selective major and for
PT
about half of less selective majors.
RI
The table also provides three measures of academic mismatch. The first of these is the
percentile of the median quota student's entrance exam score in the non-quota distribution of
SC
scores. For all selective majors (except law), this percentile is lower than 30. For about half of
selective majors, it is lower than 10, which indicates that quota students populate the bottom of
U
the exam distribution. For less selective majors, the percentile is much higher. The second
AN
measure of mismatch is the average difference between non-quota and quota cutoff scores. For
all selective majors and nearly all less selective ones, the difference in cutoffs is positive. The
M
third measure is the percentage of "displacing" quota students, those who would not have been
D
admitted by UnB in the hypothetical absence of quotas. To identify displacing students, we

TE
assumed that if there were no quotas, the same students would have applied, and they would
have received the same exam scores. Across all majors, the average percentage of displacing
EP
quota students is about 31%. In selective majors, roughly 40% of quota students would not have
been admitted in the absence of quotas.

C
To summarize, due to the admissions policy, the differences in academic credentials

AC
between quota and non-quota students are vast, especially in selective majors. This raises the
question of whether there are any effects of mismatch on education and labor market outcomes.
3.2 Administrative data on applicants
12
ACCEPTED MANUSCRIPT
The analysis focuses on persons who took the UnB vestibular exam between 2004 and
2005, a period which includes three admissions cycles following the implementation of racial
quotas (2nd semester 2004, 1st semester 2005, 2nd semester 2005). The university granted us
complete access to admissions records for this period. We had comprehensive information on
PT
major, system (quota/non-quota), semester of exam, vestibular scores, and outcome
RI
(accepted/rejected). Note that the admissions track – the group of applicants who are directly
competing for a fixed number of admissions slots – is determined jointly by major, system, and
SC
semester of exam. Recall that the cutoff score for admission varies by admission track, and
applicants in each track are either accepted or rejected. We also had responses to a short
U
voluntary questionnaire administered upon registration for the vestibular, which asked questions
AN
regarding parental and personal characteristics. About 45% of non-quota applicants and 80% of
quota applicants completed the questionnaire.

M
With these admissions records, we identified applicants who obtained exam scores
D
"close" to their track-specific cutoff: (a) accepted applicants and (b) high-performing rejected
TE
applicants. High-performing rejected applicants were those who scored at or above the 90th
percentile among all rejected applicants with a valid score in the same admissions track.
EP
Excluded from the sample were observations of applicants who applied to a major that required a
special exam in addition to the vestibular (music, architecture, performing arts, and visual arts);
C
applied to a track that either admitted or rejected all candidates; were rejected then but were
AC
accepted during another admissions cycle; and were accepted then but were also accepted during
a more recent admissions cycle.
All in all, the full sample includes 7,747 records of high-performing applicants in 318
admissions tracks. The median number of observations in an admissions track is 16. Figure 1
13
ACCEPTED MANUSCRIPT
displays the fraction of applicants included in the sample by entrance exam score percentile rank
in the full applicant pool (accepted and rejected). 95% of applicants in the estimation sample
ranked at or above the 85th percentile, and 81% ranked at or above the 90th percentile. Figure 2
displays the fraction of high-performing applicants admitted and enrolled around UnB
PT
admissions cutoffs. An applicant's normalized entrance exam score is the difference between his
RI
or her exam score and the corresponding track-specific cutoff score. As the figure shows, almost
all applicants who scored above their track-specific cutoff were admitted. About 84% of
SC
applicants who were admitted chose to enroll at UnB.
U
3.3 Administrative data on formal sector workers
AN
In this paper, we link the admissions outcomes of high-performing applicants in 2004-
2005 to their education and labor market outcomes in 2012. We use an administrative dataset
M
called "RAIS" which was collected by the Brazilian Ministry of Labor (Ministério do Trabalho e
D
Emprego, 2009-2012). Employers annually provide information on employees to the Ministry of

TE
Labor, which utilizes the information to study the labor market and determine certain labor
benefits. RAIS covers formal workers in the private and public sector. It includes self-employed
EP
persons who own a formal company. The Ministry of Labor estimates that RAIS covers about
97% of all formal workers in Brazil, approximately 50 million formal sector jobs. However,
C
RAIS does not cover informal workers, i.e., persons who are employed without a labor card or
AC
are self-employed but have not formally established a company.
RAIS is used to construct outcomes. "Years of education" is the number of years of
education completed, based on the highest level of education reported by a person's employer
during 2009-2012. Persons with no records in RAIS have missing values for this variable.
14
ACCEPTED MANUSCRIPT
"College completion" is a binary indicator that equals one if a person's employer during 2009-
2012 reported that he or she had completed college and equals zero if no employer reported he or
she completed college. "In RAIS" is a binary indicator that equals one if a person had any formal
job between 2009 and 2012 and equals zero if he or she had no records in RAIS during this
PT
period. "Formal employment" is a binary indicator that equals one if a person had any positive
RI
earnings in 2012 and equals zero if he or she had no earnings in 2012 or had no records in RAIS.
"Director/manager" is a binary indicator that equals one if a person's main occupation in
SC
2012 was "director or manager" and equals zero if his or her occupation was classified otherwise.
"Professional in sciences/arts" is a binary indicator that equals one if a person's main occupation
U
in 2012 was "professional in sciences or arts" and equals zero if his or her occupation was
AN
classified otherwise.6 According to the classification system in RAIS, other occupational
categories include military/police, mid-level technician, administrative service worker, and other
M
types of workers. "Tenure in main job" is equal to the number of months that a person had been
D
continuously employed by his or her main employer. "Public sector job" is a binary indicator that
TE
equals one if a person's main job in 2012 was in the public sector (e.g., state and federal
government, public universities, public foundations) and equals zero if his or her job was in the
EP
private sector. Public jobs, especially federal jobs, pay well and are plentiful in Brasilia. "Log
annual labor earnings" is the natural logarithm of all earnings reported in 2012. Earnings are
C
expressed in real 2014 reais. Persons with no earnings in 2012 or no records in RAIS have
AC
missing values for this variable. "Annual contractual hours" is equal to the total number of
contractual hours in 2012, based on information on weekly contractual (not actual) hours and
duration of employment for each job.
6
The most common occupations in the "professional in sciences/arts" category include engineers, accountants,
administrators, and medical professionals.
15
ACCEPTED MANUSCRIPT
3.4 Sample construction
UnB admissions records were linked with RAIS to create the estimation sample. Table 2
reports descriptive statistics. Matching the administrative datasets was challenging because UnB
PT
did not collect applicant employment identification numbers (called "CPF"). In short, we
RI
matched by name, benefiting from the fact that Brazilian names tend to be long (and unique), and
that the state where the vast majority of applicants resided, Distrito Federal (DF), is relatively
SC
small. It is assuring to examine the accuracy of our CPF match results. Fortunately, UnB did
collect the employment identification number of persons who matriculated. Our match result is
U
correct if either the CPF that we found is the same as the CPF that UnB provided or we did not
AN
find a CPF and the CPF that UnB provided does not exist in RAIS. Based on this calculation, the
overall match accuracy rate is 91%. In particular, match accuracy is 91% for non-quota students
M
and is 89% for quota students. "Exact name" match accuracy is almost 95%, while "partial
D
name" accuracy is 67%. In the next subsection, we describe an exercise that explores the
TE
robustness of our results with respect to match quality.
Overall, 73.6% of persons in the sample frame were found in RAIS, i.e., they had a
EP
formal sector job between 2009 and 2012. It is important to discuss what it signifies when a
person does not appear in RAIS. It means that he or she is (a) not working, (b) working in an
C
employment category not reported in RAIS, (c) or working in the formal sector but not found in
AC
RAIS due to our error.
Several pieces of evidence suggest that, most of the time, not appearing in RAIS indicates
not working. To begin, errors in finding persons in RAIS do not likely play a significant role.
Comparing our CPF match results with CPF information provided by the university,
16
ACCEPTED MANUSCRIPT
approximately nine of ten UnB students we did not find in RAIS were indeed not in RAIS.
Direct statistics on the labor force participation of former UnB students come from a small
survey that we conducted in 2012. We found that 79% of former UnB students reported they had
a remunerated job, formal or informal, while 21% of former students did not have a job. A
PT
nationally representative dataset like PNAD (Pesquisa Nacional por Amostra de Domicílios) can
RI
shed light on the size of the formal sector relative to the informal sector for persons who
completed at least secondary school and were born between 1978 and 1988. According to PNAD
SC
2012, about 83% of employed persons living in DF were working in a category reported in
RAIS.
U
An important issue is the extent to which there are differences in finding applicants in
AN
RAIS above and below admission cutoffs. For non-quota applicants, 76% of those above the
cutoff and 66% of those below the cutoff were found in RAIS. For quota applicants, 81% of
M
those above the cutoff and 77% of those below the cutoff were found in RAIS. These
D
differences, which are greater for non-quota applicants, may simply reflect the effect of UnB
TE
enrollment on labor force participation. However, two concerns related to sample selection are
pertinent.
EP
The first concern is whether match accuracy is correlated with being above or below the
cutoff. There is reason to believe that it is. Below-cutoff applicants are less likely to appear in the
C
sample than above-cutoff applicants, in part because they are less likely to reside in DF. It was
AC
much easier for us to identify persons in RAIS when they worked in DF, a less populous state.
The main reason why we could not obtain the CPF for all applicants employed in the formal
sector was the existence of multiple people with the same name, and the problem was greater
with matches found outside of DF. Nevertheless, this type of selection cannot be large. Overall
17
ACCEPTED MANUSCRIPT
match accuracy is high, and the vast majority of applicants, regardless of admissions score,
reside in DF. Also, below-threshold applicants may be positively selected. According to PNAD
2012, average formal (and informal) sector earnings were higher in DF than outside of DF for
persons who completed at least secondary school and were born between 1978 and 1988. Thus,
PT
if anything, our effects on earnings are underestimated.
RI
The second concern is whether working in a position unreported in RAIS is correlated
with being above or below the cutoff. There is little reason to believe that below-cutoff
SC
applicants are more likely to participate in the informal sector and hence less likely to appear in
the sample. According to PNAD 2012, levels of informal sector employment are very similar for
U
college and secondary school graduates in DF. Moreover, based on the UnB applicant survey,
AN
below-cutoff applicants without a CPF and above-cutoff applicants without a CPF have
comparable socioeconomic backgrounds as measured by mother's education. For non-quota

M
applicants not observed in RAIS, 60.4% above the cutoffs and 59.4% below the cutoffs had a
D
mother with a college degree. For quota applicants not observed in RAIS, 46.2% above the
TE
cutoffs and 44.0% below the cutoffs had a mother with a college degree. Even if this type of
selection exists, it would not be sizable, given the modest share of applicants in the informal
EP
sector, and below-threshold applicants would be positively selected. According to PNAD 2012,
average formal sector earnings were higher than average informal sector earnings. Thus, if
C
anything, our effects on earnings are underestimated.

AC
Taken together, the evidence implies that some amount of sample selection is possible,
for NQ applicants mostly, but it is small and likely biases our estimates toward zero.
3.5 Statistical procedure
18
ACCEPTED MANUSCRIPT
We implement a regression discontinuity (RD) design with the sample of high-
performing applicants. The treatment is attendance at UnB. It is useful to see this as a bundled
treatment, since it includes academic training, socialization, and university programs like
tutoring services for quota students. Generally, the advantage of an RD design is that it is like a
PT
randomized experiment in the neighborhood of the admissions cutoff. Thus, each of the 318
RI
admissions tracks can be thought of as a separate experiment. As discussed above, Figure 2
illustrates the discontinuity in admission at track-specific cutoffs. This discontinuity is almost
SC
perfectly sharp, as only a few applicants with scores above the cutoff were not admitted because
their analytical essay received a failing grade.
U
The key assumption of an RD design is that individuals are unable to precisely
AN
manipulate the assignment variable. In our context, applicants did not know admissions cutoffs
before they took the entrance exam. Cutoffs, which varied from exam to exam, were a function
M
of the number of available slots and the distribution of exam scores. Figure 3 plots the
D
distribution of normalized entrance exam scores.7 An applicant's normalized entrance exam score
TE
ܵ௜ is the difference between his or her overall vestibular score ܺ௜ and his or her track-specific
cutoff ‫ܥ‬௧ , i.e., ܵ௜ = ܺ௜ − ‫ܥ‬௧ . The figure shows there is no dramatic jump around the
EP
discontinuity.8 It is also possible to test the validity of the RD design by confirming that the
baseline covariates are balanced at the admissions threshold. Table 3 performs the tests both for
C
all applicants and for only applicants in RAIS. Figures 4-5 depict the tests graphically. The
AC
7
Applicants with normalized entrance exam scores of zero are excluded from the figure. The jump at exactly zero is
mechanical since every admissions track must have at least one normalized score of zero, by construction. Please
refer to Fort et al. (2016) for a discussion of this issue.
8
Modest discontinuities that remain between the bins to the left and right of cutoffs are undoubtedly the result of
decisions made by university admissions officials. Scores are especially close in the vicinity of cutoffs, and it
appears that officials set cutoffs so as to admit those applicants with scores extremely close to one another.
19
ACCEPTED MANUSCRIPT
results indicate there are no systematic differences in demographic characteristics between high-
performing rejected and accepted applicants.
Part of the empirical analysis employs an "intent-to-treat" model. Outcomes ܻ௜ are
regressed on an above cutoff indicator ‫ܣ‬௜ , i.e., ‫ܣ‬௜ = 1ሼܵ௜ ≥ 0ሽ, normalized entrance exam score
PT
ܵ௜ , and an interaction between the above cutoff indicator and normalized score. Regressions also
RI
include a set of binary controls for track ߠ௧ , which is jointly determined by major, quota/non-
quota system, and semester of exam. Thus, the following equation is estimated:
SC
ܻ௜ = ߚ଴ + ߚଵ ‫ܣ‬௜ + ߚଶ ܵ௜ + ߚଷ ܵ௜ ∙ ‫ܣ‬௜ + ߠ௧ + ߝ௜ .
The parameter of interest is ߚଵ. It can be interpreted as the intent-to-treat estimator or as the
U
causal effect of receiving an admission score above the cutoff. Depending on the comparison in
AN
focus, some regressions are stratified by gender and include regressors fully interacted with
M
indicators for quota/non-quota system, while other regressions are stratified by system and
include regressors fully interacted with indicators for gender.

D
Most of the analysis, however, employs a "fuzzy RD" design. The idea is to use the
TE
above cutoff indicator as an instrument to estimate the effect of attendance at UnB. It is assumed
that crossing the admissions threshold cannot cause some persons to take up and others to reject
EP
attendance at UnB, and that crossing the threshold can only affect the outcome through
attendance at UnB. Estimation is done by Two-Stage Least Squares (2SLS). The first-stage
C
equation is the following:

AC
‫ܧ‬௜ = ߙ଴ + ߙଵ ‫ܣ‬௜ + ߙଶ ܵ௜ + ߙଷ ܵ௜ ∙ ‫ܣ‬௜ + ߛ௧ + ߤ௜ ,
where ‫ܧ‬௜ denotes the indicator for attendance at UnB, and ߛ௧ denotes track-specific fixed effects.
The second-stage equation is then:
ܻ௜ = ߚ଴ + ߚଵ ‫ܧ‬෠௜ + ߚଶ ܵ௜ + ߚଷ ܵ௜ ∙ ‫ܣ‬௜ + ߠ௧ + ߝ௜ .
20
ACCEPTED MANUSCRIPT
The parameter of interest is ߚଵ . It can be interpreted as the causal effect of attending UnB. As
mentioned previously, some regressions are stratified by gender and include regressors fully
interacted with indicators for quota/non-quota system, while other regressions are stratified by
system and include regressors fully interacted with indicators for gender.
PT
Our paper exemplifies an increasingly common type of RD design in which multiple
RI
programs with different thresholds are pooled together through normalization of the running
variable. An emergent literature in econometrics is beginning to understand the particular
SC
features of these "stacked designs" (Bertanha, 2017; Cattaneo et al., 2018; Fort et al., 2016).
Cattaneo et al. (2018) establish that for stacked designs the RD estimand is a weighted average
U
across programs of the local average treatment effect at each program's threshold.
AN
In a standard RD design, it would be best to use a data-driven approach to obtain the
M
optimal asymptotic MSE-minimizing bandwidth based on the entire sample of observations
(Calonico et al., 2014). However, in our situation, this was not possible. Given the laborious
D
process of searching for applicants in RAIS, it was only feasible to focus on accepted applicants
TE
and the top 10% of rejected applicants. Anyways, one bandwidth for all tracks may not be
optimal. Bertanha (2017) argues that the issue of optimal bandwidth in the case of stacked
EP
designs is an open research question. Table A1 (see Appendix) demonstrates the robustness of
our main results with respect to bandwidth. In the exercise, the estimation sample is
C
progressively narrowed towards admissions thresholds by excluding the bottom X% of

AC
applicants below cutoffs and the top X% of applicants above cutoffs for each track.
Fort et al. (2016) examine a further challenge with stacked designs. Under certain
conditions (i.e., persons are located exactly at program thresholds, persons are heterogeneous
across programs, and programs have different probabilities to be observed in a neighborhood of
21
ACCEPTED MANUSCRIPT
zero-distance from the threshold), stacking makes outcomes as well as pre-treatment covariates
point-discontinuous functions of the running variable at zero distance from the threshold, even if
the treatment effect is zero at each separate threshold. In this situation, they suggest dropping
observations at zero distance from thresholds. Thus, we introduce Table A2, which shows the
PT
results when applicants exactly at admissions cutoffs are excluded.
RI
Other robustness exercises are conducted. Censoring of income may be a potential
concern. After all, some applicants were identified in RAIS, as they had a formal job at some
SC
point during 2009-2012, but did not work in 2012. To investigate the issue, we construct labor
market outcomes from the most recent year available between 2009 and 2012. Controls for the
U
year that the outcome is observed are added to models. Table A3 displays the results, which are
AN
extremely similar to our main findings.
Recall that we were not perfectly able to match admissions and labor market records. It
M
may be useful to restrict the estimation sample to high-quality matches. Table A4 displays the
D
main results with only applicants whose CPF was identified via "exact name" match. Recall also
TE
that the above models control linearly for the normalized entrance exam score and its associated
interactions. It may be informative to explore robustness with respect to the inclusion of

EP
quadratic controls. Table A5 displays the results with quadratic controls.

C
4. Results
AC
4.1 Main results
Table 4 displays fuzzy RD regressions with all high-performing applicants as well as
regressions stratified by gender. The estimated coefficients displayed in the table are the ones on
22
ACCEPTED MANUSCRIPT
the UnB attendance indicator for non-quota applicants and the attendance indicator for quota
applicants. Figures 6-8 depict graphically the results for key outcomes.
To begin, we consider regressions with both genders. The estimates reveal some effects
on education, which vary by quota status. Relative to non-quota applicants who do not attend,
PT
non-quota applicants who attend UnB have about 0.09 more years of education and are 4
RI
percentage points more likely to complete college, though neither of these coefficients is
statistically significant. In contrast, quota applicants experience significant gains in educational
SC
outcomes. Relative to quota applicants who do not attend, quota applicants who attend UnB have
about 0.49 more years of education and are about 11 percentage points more likely to complete
U
college. Wald tests show the difference in gains between non-quota and quota applicants is
AN
significant for years of education.
The estimates reveal some effects on labor market outcomes. For non-quota applicants,
M
attending UnB increases the likelihood that they have had formal employment and a public
D
sector job. It also increases their labor earnings, which rise by more than 12% (100*[exp(0.119)-
TE
1]). For quota applicants, UnB attendance does not significantly affect formal employment,
public sector employment, or labor earnings. Wald tests suggest that the difference in earnings
EP
gains between non-quota and quota applicants is not significant. However, attending UnB raises
the likelihood of working as a director or manager by 7 percentage points for quota applicants
C
and lowers the likelihood by 4 percentage points for non-quota applicants. This difference is
AC
statistically significant.
Regressions stratified by gender yield insightful results. They demonstrate that the
educational gains to UnB attendance are largely concentrated among male quota applicants.
Male non-quota applicants who attend UnB have about 0.03 more years of education and are 3
23
ACCEPTED MANUSCRIPT
percentage points more likely to complete college, while male quota applicants who attend UnB
have 0.77 more years of education and are 17 percentage points more likely to complete college.
These differences between non-quota and quota applicants are significant. Additionally, female
non-quota applicants who attend UnB are 6 percentage points more likely to complete college.
PT
Yet, attending UnB does not significantly improve educational outcomes for female quota
RI
applicants.
Furthermore, both male non-quota applicants and male quota applicants enjoy significant
SC
increases in monetary returns. Male non-quota applicants attending UnB have 20% higher labor
earnings, and male quota applicants attending UnB have 42% higher earnings, an effect which is
U
significant at the 10% level. The difference in gains between male non-quota and quota
AN
applicants is not significant. In contrast, females do not appear to receive any positive and
significant monetary returns. Attending UnB raises the likelihood of formal and public sector
M
employment for male non-quota applicants but not for female non-quota applicants. Females do
D
experience some changes in occupation, though. Attending UnB increases the likelihood of
TE
working as a director or manager for both male and female quota applicants. Plus, female non-
quota applicants attending UnB are more likely to work as a professional in the sciences or arts.
EP
4.2 Gender and major selectivity

C
It is difficult to know whether these results are explained by gender directly or by

AC
something correlated with gender. Most obviously, it might be that applicants in selective majors
are more likely male, and studying in a selective major tends to confer larger benefits.9 To
investigate the issue, Table 5 displays separate effects by gender and major selectivity. Majors
9
Females make up 37% of selective majors and 42% of non-selective majors among non-quota applicants in the
estimation sample, and they make up 45% of selective majors and 44% of non-selective majors among quota
applicants in the estimation sample.
24
ACCEPTED MANUSCRIPT
are divided into selective and non-selective categories according to the selectivity ranking in
Table 1. Dividing the estimation sample into two equal halves according to the number of
applicants in the sample, majors ranked 1-22 are considered selective, while those ranked 23-55
are considered non-selective.
PT
Estimates for educational outcomes exhibit interesting patterns. Regardless of quota
RI
status, the magnitude of effects is bigger for males in selective majors than for males in non-
selective majors, while the magnitude of effects is bigger for females in non-selective majors
SC
than for females in selective majors. For female non-quota applicants in non-selective majors,
the effects are significant at the 10% level. Attending UnB boosts their years of education by
U
0.20 and their likelihood of college completion by almost 7 percentage points. The effects are
AN
also significant for male quota applicants in both selective and non-selective majors.
Specifically, attending UnB increases years of education by 0.84 and increases the likelihood of
M
college completion by 23 percentage points for male quota applicants in selective majors, while
D
attending UnB increases years of education by 0.64 and increases the likelihood of college
TE
completion by 15 percentage points for male quota applicants in non-selective majors.
Estimates for labor outcomes also exhibit interesting patterns. Male non-quota applicants,
EP
both those in selective majors and those in non-selective majors, are more likely to have a public
sector job if they attend UnB. Among non-quota applicants, males in selective majors receive the
C
highest earnings gains with an increase of 26%. Attending UnB increases the likelihood of
AC
working as a professional in the sciences or arts for females in non-selective majors. In addition,
among quota applicants, males in selective and non-selective majors have large earnings gains as
both estimates exceed 28%, though the coefficients are not significant. Contrastingly, estimates
are negative for females in selective and non-selective majors. Attending UnB may reduce
25
ACCEPTED MANUSCRIPT
annual contractual hours among female quota applicants in non-selective majors, as the
coefficient is negative and significant at the 10% level. Moreover, attendance boosts the
likelihood that males and females in selective majors work as a director or manager, and it also
boosts the likelihood that males in selective majors work as a professional in the sciences or arts.
PT
In sum, while there are some differences in the magnitude of gains between applicants in
RI
selective and non-selective majors, certain differences between males and females are substantial
and, for quota applicants, occur in both selective majors as well as non-selective ones.
SC
4.3 Mismatch effects
U
As we have seen in Table 1, the quota system generated differences in academic
AN
credentials between quota and non-quota students. Such differences were evident in each of the
measures of mismatch, and they varied greatly across majors. Then, how do differences in
M
academic credentials impact the effect of UnB admission on education and labor market
D
outcomes? More pointedly, does mismatch reduce the benefits that quota students receive or
TE
even make them negative? To explore potential mismatch effects, interactions between
mismatch, the above cutoff indicator, and quota status are added to the intent-to-treat
EP
specification.10 Three measures of mismatch, which vary by major, are used separately in
regressions: the percentage of NQ students with admission scores higher than the median Q
C
student's score, the average difference between NQ and Q cutoff scores, and the percentage of Q
AC
students who would not have been admitted in the hypothetical absence of quotas.
Table 6 and Table A6 display the coefficients on the interactions between mismatch and
having a score above the admissions threshold. In short, there is no evidence of mismatch effects
10
Note it is insightful to estimate mismatch effects also for non-quota students. These results may provide evidence
on whether mismatch directly helps non-quota students and whether an omitted variable may be driving the
mismatch results for quota students.
26
ACCEPTED MANUSCRIPT
for male quota applicants. None of the coefficients are significant for them. However, there is
some evidence of mismatch effects for female quota applicants. Looking across the entire set of
results, female non-quota applicants have three coefficients that are negative and significant,
while female quota applicants have six coefficients that are negative and significant, though most
PT
of these are only significant at the 10% level. Mismatch appears to impact both education (years
RI
of education, college completion) and labor outcomes (formal employment, public sector job).
For example, as the percentage of displacing students rises by 10 points, the effect of UnB
SC
admission on years of education decreases by 0.18 and the effect on public sector employment
decreases by 9 percentage points. The magnitude of these particular mismatch effects is large
U
enough to make the magnitude of the overall effects of admission negative for female quota
AN
students enrolled in majors with moderate to severe mismatch.
M
5. Discussion
D
5.1 Summary of results

TE
In this paper, we estimate the effects of affirmative action at one of the best universities
in Brazil with data linking the admissions outcomes of high-performing applicants in 2004-2005
EP
to their education and labor market outcomes in 2012. The treatment is attendance at UnB, which
is effectively a bundled treatment, as it includes academic training, socialization, and university

C
programs like tutoring services for quota students. In fuzzy RD models with both genders
AC
pooled, quota applicants experience much larger gains in educational outcomes than non-quota
applicants. Relative to quota applicants not attending, quota applicants attending UnB have
significantly more years of education and are more likely to complete college. Additionally,
attending UnB raises the likelihood that quota applicants work as a director or manager.
27
ACCEPTED MANUSCRIPT
Attending UnB also raises the likelihood that non-quota applicants obtain formal employment as
well as a public sector job. When both genders are pooled, UnB attendance significantly raises
the labor earnings of non-quota applicants but not of quota applicants. Nevertheless, patterns are
strikingly dissimilar by gender, as results reveal the gains are mostly concentrated among male
PT
applicants. Male quota applicants attending UnB have significantly higher educational outcomes,
RI
while female quota applicants do not. Moreover, attendance significantly raises the earnings of
both male quota applicants and male non-quota applicants but does not raise the earnings of
SC
female applicants. Further analyses demonstrate that some differences in gains between males
and females exist in both selective and non-selective majors, especially for quota applicants, and
U
that, in some specifications, mismatch in academic credentials is associated with a sizable
AN
reduction in the benefits of admission for female quota students.
M
5.2 Implications
D
The findings shed light on the effects of attendance at a competitive undergraduate

TE
institution. The vast majority of applicants in our sample were ages 24-34 in 2012, 7-8 years
after taking the UnB entrance exam. Given the persistence of earnings trajectories over the
EP
lifetime, we observe them at a particularly crucial time early in their careers. Our labor market
results underscore the importance of college quality in a setting outside of the U.S. (e.g., Brewer
C
et al., 1999; Black and Smith, 2004, 2006; Hoekstra, 2009). To compare, Hoekstra (2009) reports
AC
that attending a selective U.S. university increases the earnings of white men by about 20%,
while we report that attending UnB increases the earnings of male non-quota applicants by 20%
and male quota applicants by 42%. Indeed, our evidence is consistent with previous findings that
the returns to college quality may be relatively large for certain U.S. minorities and persons with
28
ACCEPTED MANUSCRIPT
disadvantaged backgrounds (Andrews et al., 2016; Dale and Krueger, 2014). Examining the
analogous case of Chile, Hastings et al. (2014) estimate returns to college quality as well as field
of study. Like them, we also find large earnings gains to obtaining a degree in a selective field of
study, albeit only for male applicants. They report that returns to selectivity do not generally vary
PT
by SES. Our results are consistent with this insight. Male quota applicants (with relatively lower
RI
SES) and non-quota applicants (with relatively higher SES) receive similar earnings gains in
selective majors, while male quota applicants receive higher earnings gains than non-quota
SC
applicants in non-selective majors, though the difference is not significant.
The findings also shed light on the effects of affirmative action in college admissions. We
U
document large differences in academic credentials between quota and non-quota students upon
AN
matriculation. Despite such initial disparities, the evidence suggests that the admissions policy
improved outcomes for males in the targeted group. Similar to the gain in educational
M
achievement that quota applicants in India receive (Bagde et al., 2016), male quota applicants at
D
UnB receive an increase in educational attainment. Like quota applicants in India (Bertrand et al.
TE
2010), male quota applicants at UnB have positive labor market returns, but unlike their Indian
counterparts, they appear to receive higher returns than non-quota applicants. Nevertheless, we
EP
fail to find evidence of improved outcomes for females, perhaps with the exception of
occupational status. In fact, we uncover several negative mismatch effects for them. Thus, our
C
results for men are consistent with studies that do not find substantial mismatch effects (e.g.,
AC
Alon and Tienda, 2005; Bagde et al., 2016; Fischer and Massey, 2007; Grove and Hussey, 2014;
Rothstein and Yoon, 2008a, 2008b), while our results for women are consistent with studies that
do (e.g., Arcidiacono et al., 2016; Frisancho and Krishna, 2016; Loury and Garman, 1993, 1995;
Sander, 2004).
29
ACCEPTED MANUSCRIPT
However, it remains an open question why female applicants benefit little from attending
UnB. This is especially puzzling given the employment environment is generally favorable
towards women. Maternity leave is generous, and weekly hours are subjected to a legal
maximum. One potential explanation for the women's results is that the positive impact of UnB
PT
attendance on outcomes is offset by other forces. It may be that, through changes in marriage and
RI
childbearing, UnB attendance might lower labor force participation among women with the
highest potential earnings and/or raise participation among those with the lowest potential
SC
earnings. Though, according to PNAD 2012, labor force participation was high among college-
educated women who were aged 24 and 34 and residing in DF, regardless of marital status. Even
U
79% of married or cohabiting women with children were working. Another potential explanation
AN
is that UnB attendance really has no positive impact on women's outcomes. Our evidence on
mismatch effects comes to bear here. Female quota students were the group most impacted by
M
the differences in academic credentials that were created by the quota system. Yet, it is not clear
D
why there was a gender disparity in mismatch effects. In light of these findings and the
TE
employment environment for women in Brazil, the topic of gender and labor market outcomes is
worthy of future inquiry.

EP
5.3 Contributions and caveats

C
In sum, this study estimates the effects of affirmative action in college admissions on
AC
both education and labor market outcomes, implementing an RD design with administrative data.
Studying affirmative action in Brazil is valuable not only because of the methodological
advantages but also because affirmative action is new, many people are affected, and much is at
stake. It is also good to acknowledge the limitations of the study. One limitation is that we do not
30
ACCEPTED MANUSCRIPT
have information on informal sector participation, as RAIS only covers the formal sector. We do
not know what rejected applicants end up doing academically, that is, what college they attend or
what major they pursue. Another limitation is that admissions records did not contain
employment identification numbers, so we had to link admissions and labor market data using
PT
applicant names. Therefore, measurement error affects the results to some degree.
RI
Future research may be able to extend the analysis. It would be interesting to study the
long-run effects of the UnB quota policy on labor, demographic, and health outcomes. In light of
SC
recent legislation, it appears likely that group-targeted educational programs will continue to
expand in Brazil and elsewhere, and future research can study these Educational Equity Policies
U
as they emerge. Indeed, understanding more about the impact of educational policy on life
AN
outcomes is important and timely, as affirmative action remains an issue of current debate and
affects the lives of millions of people worldwide who belong to historically disadvantaged
M
groups.
D
TE
C EP
AC
31
ACCEPTED MANUSCRIPT
References
Alon, S., Tienda, M., 2005. Assessing the 'mismatch' hypothesis: Differences in college
graduation rates by institutional selectivity. Sociol. Educ. 78 (4), 294-315.
Alon, S., Malamud, O., 2014. The impact of Israel's class-based affirmative action policy on
admission and academic outcomes. Econ. Educ. Rev. 40 (June), 123-139.
Andrews, R.J., Li, J., Lovenheim, M.F., 2016. Quantile Treatment Effects of College Quality on
PT
Earnings. J. Hum. Resour. 51 (1), 200-238.
Antonovics, K., Backes, B., 2014. The Effect of Banning Affirmative Action on College
Admissions Policies and Student Quality. J. Hum. Resour. 49 (2), 295-322.
RI
Arcidiacono, P., Lovenheim, M., 2016. Affirmative Action and the Quality-Fit Trade-off. J.
Econ. Lit. 54 (1), 3-51.
Arcidiacono, P., Aucejo, E.M., Hotz, V.J., 2016. University differences in the graduation of
SC
minorities in STEM fields: Evidence from California. Am. Econ. Rev. 106 (3), 525-562.
Arcidiacono, P., 2005. Affirmative Action in Higher Education: How Do Admission and
Financial Aid Rules Affect Future Earnings? Econom. 73 (5), 1477-152.
Backes, B., 2012. Do Affirmative Action Bans Lower Minority College Enrollment and
U
Attainment? Evidence from Statewide Bans. J. Hum. Resour. 47 (2), 435-455.
Bagde, S., Epple, D., Taylor, L., 2016. Does Affirmative Action Work? Caste, Gender, College
AN
Quality, and Academic Success in India. Am. Econ. Rev. 106 (6), 1495-1521.
Bailey, S.R., 2009. Legacies of Race: Identities, Attitudes, and Politics in Brazil. Stanford:
Stanford University Press.
Bertanha, M., 2017. Regression discontinuity design with many thresholds. Working Paper.
M
Bertrand, M., Hanna, R., Mullainathan, S., 2010. Affirmative Action in Education: Evidence
From Engineering College Admissions in India. J. Public Econ. 94 (1-2), 16-29.
Black, D.A., Smith, J.A., 2004. How robust is the evidence on the effects of college quality?
D
Evidence from matching. J. Econom. 121 (1-2), 99-124.

Black, D.A., Smith, J.A., 2006. Estimating the returns to college quality with multiple proxies
TE
for quality. J. Labor Econ. 24 (3), 701-728.

Brewer, D.J., Eide, E.R., Ehrenberg, R.G., 1999. Does it pay to attend an elite private college?
Cross-cohort evidence on the effects of college type on earnings. J. Hum. Resour. 34 (1),
EP
104-123.
Calonico, S., Cattaneo, M.D., Titiunik, R., 2014. Robust nonparametric confidence intervals for
regression-discontinuity designs. Econom. 82 (6), 2295-2326.
Cattaneo, M.D., Keele, L., Titiunik, R., Vazquez-Bare, G., 2018. Interpreting Regression
C
Discontinuity Designs with Multiple Cutoffs. J. Polit. 78 (4), 1229-1248.

Cortes, K.E., 2010. Do bans on affirmative action hurt minority students? Evidence from the
AC
Texas Top 10% Plan. Econ. Educ. Rev. 29 (6), 1110-1124.

Dale, S., Krueger, A.B., 2014. Estimating the effects of college characteristics over the career
using administrative earnings data. J. Hum. Resour. 49 (2), 323-358.
Dale, S.B., Krueger, A.B., 2002. Estimating the payoff to attending a more selective college: An
application of selection on observables and unobservables. Q. J. Econ. 117 (4), 1491-1527.
Darity, W., Deshpande, A., Weisskopf, T., 2011. Who Is Eligible? Should Affirmative Action be
Group- or Class-Based? Am. J. Econ. Sociol. 70 (1), 238-268.
Dickson, L.M., 2006. Does ending affirmative action in college admissions lower the percent of
minority students applying to college? Econ. Educ. Rev. 25 (1), 109-119.
32
ACCEPTED MANUSCRIPT
Eltis, D., 2001. The Volume and Structure of the Transatlantic Slave Trade: A Reassessment.
William Mary Q. 58 (1), 17-46.
Fischer, M.J., Massey, D.S., 2007. The effects of affirmative action in higher education. Soc. Sci.
Res. 36 (2), 531-549.
Francis, A.M., Tannuri-Pianto, M., 2012. Using Brazil’s Racial Continuum to Examine the
Short-Term Effects of Affirmative Action in Higher Education. J. Hum. Resour. 47 (3), 754-
784.
PT
Francis, A.M., Tannuri-Pianto, M., 2013. Endogenous Race in Brazil: Affirmative Action and
the Construction of Racial Identity among Young Adults. Econ. Dev. Cult. Change 61 (4),
731-753.
RI
Fort, M., Ichino, A., Zanella, G., 2016. On the perils of stacking thresholds in RD designs.
Working Paper.
Frisancho, V., Krishna, K., 2016. Affirmative action in higher education in India: targeting, catch
SC
up, and mismatch. High. Educ. 71 (5), 611-649.
Grove, W.A., Hussey, A.J., 2014. Preferential Admission and MBA Outcomes: Mismatch
Effects by Race and Gender. B.E. J. Econ. Anal. Policy 14 (3), 613-668.
Hastings, J.S., Neilson, C.A., Zimmerman, S.D., 2014. Are Some Degrees Worth More than
U
Others? Evidence from college admission cutoffs in Chile. NBER Working Paper No. 19241.
Hinrichs, P., 2012. The Effects of Affirmative Action Bans on College Enrollment, Educational
AN
Attainment, and the Demographic Composition of Universities. Rev. Econ. Stat. 94 (3), 712-
722.
Hoekstra, M., 2009. The Effect of Attending the Flagship State University on Earnings: A
Discontinuity-Based Approach. Rev. Econ. Stat. 91 (4), 717-724.
M
Holzer, H., Neumark, D., 2000. Assessing Affirmative Action. J. Econ. Lit. 38 (3), 483-568.
IBGE (Instituto Brasileiro de Geografia e Estatística), 2015. Pesquisa Nacional por Amostra de
Domicílios: Síntese de indicadores 2014. IBGE: Rio de Janeiro. Accessed online at
D
http://biblioteca.ibge.gov.br/visualizacao/livros/liv94935.pdf.
Loury, L.D., Garman, D., 1993. Affirmative Action in Higher Education. Am. Econ. Rev. 83 (2),
TE
99-103.
Loury, L.D., Garman, D., 1995. College Selectivity and Earnings. J. Labor Econ. 13 (2), 289-
308.
EP
Massey, D.S., Mooney, M., 2007. The effects of America's three affirmative action programs on
academic performance. Soc. Probl. 54 (1), 99-117.
Ministério do Trabalho e Emprego (Ministry of Labor), 2009-2012. Relação Anual de
Informações Sociais (RAIS): microdados. Brasília, DF.
C
Rothstein, J., Yoon, A.H., 2008a. Affirmative action in law school admissions: What do racial
preferences do? Univ. Chic. Law Rev. 75 (2), 649-714.
AC
Rothstein, J., Yoon, A.H., 2008b. Mismatch in Law School. NBER Working Paper No. 14275.
Sander, R.H., 2004. A systemic analysis of affirmative action in American law schools. Stanf.
Law Rev. 57 (2), 367-483.
Wydick, B., 2002. Affirmative action in college admissions: Examining labor market effects of
four alternative policies. Contemp. Econ. Policy, 20 (1), 12-24.
33
ACCEPTED MANUSCRIPT
Figure 1. Fraction of applicants included in sample by entrance exam score percentile rank
1
0.9
Fraction included in sample
0.8
0.7
PT
0.6
0.5
0.4
RI
0.3
0.2
0.1
SC
0
0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95
Entrance exam score percentile rank
U
NOTE. Percentile rank is calculated by major, system, and semester of exam. The figure considers all applicants
AN
(accepted and rejected) who took the vestibular in 2nd semester 2004, 1st semester 2005, and 2nd semester 2005.
Figure 2. Fraction of applicants admitted and enrolled around UnB admissions cutoffs
M
1
0.9
0.8
D
Fraction of applicants
0.7
TE
0.6
0.5
0.4
0.3
EP
0.2
0.1
0
C
-80 -40 0 40 80 120

Distance from cutoff
AC
Admitted Enrolled
NOTE. The figure displays the fraction admitted and the fraction enrolled by entrance exam score bin. Applicants
are sorted into 10 bins of 8-point width below normalized cutoffs and 15 bins of 8-point width above normalized
cutoffs. The lowest and highest bins are wider to include the tails of the distribution. The vertical line indicates the
normalized admissions cutoff.
34
ACCEPTED MANUSCRIPT
Figure 3. Distribution of normalized entrance exam score
Quota applicants
0.08
0.07
PT
0.06
0.05
0.04
RI
0.03
0.02
SC
0.01
0
-120 -80 -40 0 40 80 120 160 200
U
AN
Non-quota applicants
0.08
0.07
M
0.06
0.05
D
0.04
0.03
TE
0.02
0.01
0
EP
-120 -80 -40 0 40 80 120 160 200

NOTE. Applicants who had a zero normalized entrance exam score are excluded from the figure. Bin width is 8
C
points.
AC
35
ACCEPTED MANUSCRIPT
Figure 4. Various demographic characteristics around UnB admissions cutoffs, non-quota

applicants
Non-quota Non-quota
0.32 0.36 0.40 0.44
1982.5
Birth year
Female
PT
1981.5
RI
-80 -40 0 40 80 120 -80 -40 0 40 80 120
Distance from cutoff Distance from cutoff
SC
Non-quota Non-quota
Completed QSC
U
Mother College
0.55
0.45
AN 0.45
0.35
-80 -40 0 40 80 120 -80 -40 0 40 80 120

M

D
Non-Quota Non-quota
TE
3000 3400 3800
0.45
Identified PPI
HH income
0.35
EP
-80 -40 0 40 80 120 -80 -40 0 40 80 120

C
NOTE. Lines plot fitted values from RD regressions. Points plot local averages of residuals from regressions of
AC
outcomes on track fixed effects. Applicants are sorted into 10 bins of 8-point width below normalized cutoffs and 15
bins of 8-point width above normalized cutoffs. The lowest and highest bins are wider to include the tails of the
distribution. The vertical line indicates the normalized admissions cutoff. A single measure of household income is
calculated using the midpoints of income ranges given in Table 2.
36
ACCEPTED MANUSCRIPT
Figure 5. Various demographic characteristics around UnB admissions cutoffs, quota applicants
Quota Quota
0.65
1983.0
Birth year
Female
0.50
PT
1981.5
0.35
-80 -40 0 40 80 120 -80 -40 0 40 80 120
RI
SC
Quota Quota
Completed QSC
Mother College
0.2 0.3 0.4 0.5

0.85
U
0.75
AN
-80 -40 0 40 80 120 -80 -40 0 40 80 120
M
Quota Quota
D
1.00
3000
Identified PPI
TE
HH income
0.98
2000
0.96
EP
-80 -40 0 40 80 120 -80 -40 0 40 80 120

C
AC
distribution. The vertical line indicates the normalized admissions cutoff. A single measure of household income is
calculated using the midpoints of income ranges given in Table 2.
37
ACCEPTED MANUSCRIPT
Figure 6. Years of education around UnB admissions cutoffs
Quota Non-quota
Years of Education
Years of Education
15.4
15.0
15.2
PT
14.0
15.0
-80 -40 0 40 80 120 -80 -40 0 40 80 120
RI
SC
Quota Male Non-quota Male
14.9 15.1 15.3 15.5

Years of Education
Years of Education
14.5 15.0 15.5
U
AN
-80 -40 0 40 80 120 -80 -40 0 40 80 120
M
Quota Female Non-quota Female

D
Years of Education
Years of Education
TE
15.0
15.4
14.0
15.0
EP
-80 -40 0 40 80 120 -80 -40 0 40 80 120

C
AC
distribution. The vertical line indicates the normalized admissions cutoff.
38
ACCEPTED MANUSCRIPT
Figure 7. College completion around UnB admissions cutoffs
Quota Non-quota
College Completion
College Completion
0.8
0.70 0.75 0.80

0.6
PT
0.4
-80 -40 0 40 80 120 -80 -40 0 40 80 120
RI
SC
0.80
0.9
College Completion
College Completion
U
0.70
0.7
AN
0.5
0.60
-80 -40 0 40 80 120 -80 -40 0 40 80 120

M

D
0.9
0.90
College Completion
College Completion
TE
0.7
0.80
0.5
0.70
0.3
EP
-80 -40 0 40 80 120 -80 -40 0 40 80 120

C
AC
39
ACCEPTED MANUSCRIPT
Figure 8. Log annual labor earnings around UnB admissions cutoffs
Quota Non-quota
Log Labor Earnings
Log Labor Earnings
10.6 10.8 11.0

11.0
10.6
PT
10.2
-80 -40 0 40 80 120 -80 -40 0 40 80 120
RI
SC
Log Labor Earnings
Log Labor Earnings

11.2
10.7 10.9 11.1
U
10.8
AN
10.4
-80 -40 0 40 80 120 -80 -40 0 40 80 120

M

D
10.4 10.6 10.8 11.0
Log Labor Earningsb
Log Labor Earnings
10.8
TE
9.8 10.2
EP
-80 -40 0 40 80 120 -80 -40 0 40 80 120

C
AC
40
Table 1 ACCEPTED MANUSCRIPT
UnB admissions statistics by major in 2004-2005
Selectivity UnB acceptance rate Measures of mismatch

rank Q NQ Percentile of Difference in Percentage
Major median Q score cutoffs displacing
Medicine 1 2.6 1.3 3.8 26.6 41.2

Law (evening) 2 3.0 2.4 51.2 31.4 28.6
PT
Law 3 1.5 1.2 71.4 9.6 0.0
Mechatronics Engineering 4 15.7 6.3 15.9 47.8 46.2
International Relations 5 11.8 3.2 17.4 9.7 19.0
RI
Information Systems Engineering 6 12.8 7.9 3.6 89.0 33.3
Electrical Engineering 7 19.0 12.6 28.3 13.9 18.8
Biology 8 7.6 5.0 1.5 124.6 73.3
SC
Computer Science 9 6.2 4.4 30.4 41.2 26.7
Dentistry 10 8.2 4.8 6.3 92.6 50.0
Communication Studies 11 6.8 4.1 3.7 55.0 51.5
U
Pharmaceutical Science 12 11.1 6.6 12.3 33.4 47.1
Mechanical Engineering 13 20.2 10.5 24.7 38.8 40.9
AN
Political Science 14 10.2 8.4 16.5 16.4 19.0
Economics 15 19.8 11.7 6.7 60.4 58.3
Civil Engineering 16 11.8 9.8 27.2 33.3 30.8
M
Nutrition Science 17 4.9 4.5 3.4 85.2 57.1
Industrial Design 18 11.5 8.3 2.3 55.2 55.6
Psychology 19 5.6 4.6 9.4 52.4 38.9
D
Physics 20 13.6 13.2 4.9 52.1 45.5
Biology (teaching degree) 21
TE
9.0 8.8 1.5 69.4 62.5
History 22 5.4 6.5 1.5 72.1 53.3
Veterinary Medicine 23 9.1 5.7 20.5 27.7 18.8
Business Administration (evening) 24 5.8 6.3 7.7 59.4 60.0
EP
Business Administration 25 10.4 7.1 27.0 32.6 23.1
Nursing 26 4.1 4.6 29.0 -2.9 20.0
Translation 27 12.6 9.0 1.8 43.8 45.5
C
Statistics 28 19.5 13.2 60.4 -0.5 20.0

Chemistry 29 12.0 12.9 40.5 0.4 25.0
AC
Social Science 30 10.4 11.6 37.3 16.1 20.0

Computer Science (teaching deg.) 31 9.5 8.5 15.4 8.7 25.0
Agronomy 32 10.0 7.5 19.8 35.2 21.7
Physics (teaching degree) 33 10.9 15.1 31.0 17.1 20.0
Philosophy 34 10.9 12.5 14.0 70.4 50.0
Portuguese 35 10.1 10.3 9.6 26.5 38.5
Accounting 36 10.6 11.3 50.6 5.2 9.5
41
French 37 12.8 12.5
ACCEPTED9.5
MANUSCRIPT 7.3 50.0
Mathematics (teaching degree) 38 10.6 11.7 1.4 42.8 50.0
Accounting (evening) 39 8.4 9.7 44.9 27.4 22.7
English 40 9.3 7.1 14.3 42.0 50.0
Forest Engineering 41 13.3 9.4 23.8 36.1 37.5
Geography 42 9.0 8.6 26.2 -1.9 0.0
Portuguese (teaching degree) 43 6.8 10.5 13.2 46.0 33.3
Chemistry (teaching degree) 44 9.7 12.4 33.8 6.8 26.7
Physical Education 45 4.1 3.9 46.8 7.3 17.4
Geology 46
PT
13.5 11.6 67.9 20.2 15.4
Mathematics 47 12.7 15.9 27.2 -11.3 30.8
Social Work 48 7.5 6.8 42.1 -2.7 0.0
Spanish (teaching degree) 49
RI
10.4 10.3 48.9 7.4 15.4
Archival Science 50 5.5 7.9 51.6 13.3 22.7
Library Science 51 8.7 10.1 54.8 -7.9 0.0
SC
Japanese (teaching degree) 52 17.0 18.2 18.8 53.4 33.3
Port. as Sec. Lang. (teaching deg.) 53 20.2 16.6 21.7 13.9 20.0
Education 54 9.4 10.5 50.6 6.6 4.7
U
Education (evening) 55 6.5 8.7 44.2 -5.2 0.0
AN
NOTE. Q indicates "quota" and NQ indicates "non-quota." Selectivity rank is based on the average NQ cutoff score during this time period. Three measures of mismatch are given: the
M
percentile of the median Q student's entrance exam score in the NQ distribution, the average difference between NQ and Q cutoff scores, and the percentage of "displacing" Q students, i.e.,
those students who would not have been admitted in the hypothetical absence of quotas.
D
TE
C EP
AC
42
Descriptive statistics for persons in estimation sample
NQ applicants Q applicants
Male Female Male Female
UnB admission outcomes

Standardized exam score (mean 0 & sd 1) 0.18 -0.07 -0.36 -0.42
PT
Admitted (%) 63.1 61.7 69.1 67.9
Enrolled (%) 50.7 53.3 63.1 62.7
RI
Education/labor outcomes in 2012
Years of education 15.2 15.5 14.8 15.3
College completion (%) 72.3 81.5 64.6 75.2
SC
In RAIS (any formal job during 2009-2012) (%) 72.3 72.4 81.3 77.7
Formal employment in 2012 (%) 65.4 64.3 74.0 69.4
Director/manager (%) 7.6 8.7 6.5 4.6
U
Professional in sciences/arts (%) 36.4 38.9 27.7 34.9
Tenure in main job (months) 48.8 33.8 46.9 34.3
AN
Public sector job (%) 52.4 53.1 53.8 51.8
Log annual labor earnings 11.0 10.6 10.8 10.5
Annual contractual hours 1876.1 1761.7 1907.8 1759.5
M
Demographic characteristics
Female (%) 0.0 100.0 0.0 100.0
D
Birth year 1981.8 1983.8 1982.2 1983.7
TE
Completed UnB admissions survey (%) 42.1 47.5 80.5 81.3
Mother college completion (%, UnB survey) 53.5 54.5 37.9 35.3
Family monthly income (%, UnB survey)
R$750 or less 7.9 9.0 14.3 20.4
EP
R$750-2500 28.3 29.9 40.1 42.3
R$2500-5000 25.0 27.1 21.4 21.5
R$5000 or more 26.8 24.8 12.8 8.5
C
Don't know 12.0 9.2 11.4 7.3

Identified as PPI (%, UnB survey) 42.2 39.7 99.3 99.4
AC
N 3,853 2,529 754 611
43
NOTE. Q indicates "quota" and NQ indicates "non-quota." The sample includesACCEPTED MANUSCRIPT
UnB applicants between 2nd semester 2004 and 2nd semester 2005 including (a) all persons who took the
entrance exam and were admitted and (b) all persons who were not admitted but had an exam score in the 90th percentile or higher. Earnings are expressed in real 2014 R$. Sample size is
smaller for variables derived from the UnB admissions survey.
PT
RI
U SC
AN
M
D
TE
C EP
AC
44
RD analysis of demographic characteristics around UnB admission cutoffs for applicants in 2004-2005
Full estimation sample Only applicants in RAIS

above cutoff above cutoff above cutoff above cutoff
Dependent variable x non-quota x quota N x non-quota x quota N
Female 0.015 0.012 7,747 0.017 0.018 5,705
PT
(0.019) (0.050) (0.025) (0.060)
Birth year 0.210 0.082 4,438 0.210 0.082 4,438
(0.375) (0.712) (0.375) (0.712)
RI
Completed UnB admissions survey 0.010 0.056 7,747 -0.003 0.080** 5,705
(0.022) (0.034) (0.025) (0.036)
Mother college completion -0.003 -0.004 3,915 0.017
SC
-0.062 2,908
(0.030) (0.048) (0.036) (0.056)
Family monthly income
R$750 or less -0.011 0.020 3,910 -0.021 0.056 2,902
U
(0.018) (0.040) (0.025) (0.044)
R$750-2500 0.025 -0.007 3,910 0.034 -0.001 2,902
AN
(0.030) (0.053) (0.040) (0.063)
R$2500-5000 0.009 -0.017 3,910 0.020 -0.036 2,902
(0.027) (0.046) (0.035) (0.057)
M
R$5000 or more -0.019 0.009 3,910 -0.028 -0.000 2,902
(0.027) (0.028) (0.034) (0.025)
Don't know -0.004 -0.004 3,910 -0.005 -0.019 2,902
D
(0.020) (0.032) (0.027) (0.034)
TE
Identified as PPI -0.027 0.001 3,540 -0.028 0.005 2,639
(0.035) (0.007) (0.043) (0.009)
EP
NOTE. The sample includes UnB applicants between 2nd semester 2004 and 2nd semester 2005 including (a) all persons who took the entrance exam and were admitted and (b) all persons
who were not admitted but had an exam score in the 90th percentile or higher. Additional controls include normalized entrance exam score x non-quota applicant, normalized entrance exam
C
score x quota applicant, indicator for above cutoff x normalized entrance exam score x non-quota applicant, indicator for above cutoff x normalized entrance exam score x quota applicant,
and a set of binary indicators for admissions track. Robust standard errors in parentheses are adjusted for clustering on tracks. ***, **, and * indicate statistical significance at the 1%, 5%,
AC
and 10% levels, respectively.
45
Fuzzy RD analysis of 2012 education and labor outcomes around UnB admission cutoffs for applicants in 2004-2005
Both genders Male applicants Female applicants

enrolled enrolled Wald enrolled enrolled Wald enrolled enrolled Wald
Dependent variable x non-quota x quota test N x non-quota x quota test N x non-quota x quota test N
Years of education 0.092 0.491*** 0.045 5,705 0.032 0.765*** 0.017 3,398 0.167 0.171 0.991 2,307
PT
(0.084) (0.180) (0.121) (0.281) (0.109) (0.291)
College completion 0.038 0.106** 0.222 5,705 0.025 0.168** 0.074 3,398 0.058* 0.014 0.618 2,307
(0.025) (0.049) (0.035) (0.072) (0.034) (0.082)
RI
In RAIS (2009-2012) 0.058** 0.014 0.306 7,747 0.062** 0.027 0.567 4,607 0.058 -0.002 0.436 3,140
(0.024) (0.036) (0.029) (0.055) (0.038) (0.067)
Formal employment 0.055** -0.003 0.226 7,747 0.065**
SC
0.028 0.580 4,607 0.045 -0.057 0.217 3,140
(0.025) (0.041) (0.031) (0.060) (0.038) (0.073)
Director/manager -0.036** 0.072*** 0.000 4,595 -0.024 0.086** 0.017 2,797 -0.046 0.083** 0.014 1,798
(0.016) (0.025) (0.020) (0.042) (0.033) (0.040)
U
Professional in sciences/arts 0.042 0.024 0.817 4,595 -0.020 0.058 0.467 2,797 0.140** -0.054 0.116 1,798
(0.040) (0.066) (0.048) (0.095) (0.061) (0.108)
AN
Tenure in main job 1.464 -7.053 0.363 4,601 5.722 -0.552 0.635 2,800 -4.980 -5.892 0.930 1,801
(3.704) (8.606) (5.052) (12.193) (5.004) (9.089)
Public sector job 0.078** -0.055 0.065 4,601 0.120*** -0.072 0.053 2,800 0.007 0.029 0.869 1,801
M
(0.030) (0.065) (0.039) (0.091) (0.052) (0.121)
Log annual labor earnings 0.119* 0.063 0.707 5,128 0.182** 0.353* 0.417 3,077 0.036 -0.034 0.760 2,051
(0.071) (0.133) (0.086) (0.193) (0.104) (0.202)
D
Annual contractual hours 60.486 -75.581 0.250 5,139 121.821* 89.721 0.817 3,085 -41.924 -182.514 0.420 2,054
TE
(48.429) (107.789) (71.785) (119.085) (70.466) (159.436)
EP
who were not admitted but had an exam score in the 90th percentile or higher. Instruments include indicator for above cutoff x non-quota applicant and indicator for above cutoff x quota
applicant. Additional controls include normalized entrance exam score x non-quota applicant, normalized entrance exam score x quota applicant, indicator for above cutoff x normalized
C
entrance exam score x non-quota applicant, indicator for above cutoff x normalized entrance exam score x quota applicant, and a set of binary indicators for admissions track. Robust standard
errors in parentheses are adjusted for clustering on tracks. ***, **, and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively. P-values associated with Wald tests of
AC
the difference between the non-quota and quota coefficients are listed.
46
Fuzzy RD analysis of 2012 education and labor outcomes around UnB admission cutoffs for applicants in 2004-2005, stratified by major selectivity
Non-quota applicants Quota applicants

Selective majors Non-selective majors Selective majors Non-selective majors
enrolled enrolled enrolled enrolled enrolled enrolled enrolled enrolled

Dependent variable x male x female x male x female N x male x female x male x female N
PT
Years of education 0.129 -0.003 0.013 0.203* 4,617 0.836** -0.144 0.640** 0.393 1,088
(0.131) (0.129) (0.145) (0.121) (0.346) (0.354) (0.287) (0.279)
RI
College completion 0.051 0.003 0.021 0.065* 4,617 0.226** -0.076 0.150** 0.065 1,088
(0.037) (0.038) (0.042) (0.037) (0.091) (0.100) (0.076) (0.080)
SC
In RAIS (2009-2012) 0.075** 0.037 0.061* 0.046 6,382 0.085 -0.026 0.011 -0.006 1,365
(0.038) (0.049) (0.034) (0.039) (0.065) (0.083) (0.055) (0.061)
Formal employment 0.079* 0.031 0.056 0.038 6,382 0.085 -0.091 0.007 -0.015 1,365
(0.042) (0.050) (0.034) (0.039) (0.077) (0.087) (0.058) (0.071)
U
Director/manager -0.048* -0.052 -0.015 -0.048 3,720 0.165*** 0.099** 0.061* 0.038 875
AN
(0.028) (0.038) (0.022) (0.030) (0.051) (0.050) (0.034) (0.041)
Professional in sciences/arts -0.015 0.074 0.008 0.139** 3,720 0.223** -0.008 -0.020 -0.018 875
(0.062) (0.069) (0.049) (0.063) (0.112) (0.124) (0.093) (0.101)
Tenure in main job 6.789 -2.809 4.930 -7.093 3,724 -16.384 -11.998 0.870 -9.624 877
M
(5.741) (5.942) (5.741) (5.445) (11.392) (10.372) (12.589) (10.340)
Public sector job 0.111** -0.009 0.105** 0.054 3,724 -0.293*** -0.086 0.006 0.021 877
D
(0.049) (0.059) (0.044) (0.052) (0.110) (0.119) (0.090) (0.125)
Log annual labor earnings 0.229** 0.060 0.110 0.066 4,146 0.255 -0.153 0.247 -0.154 982
TE
(0.099) (0.122) (0.088) (0.108) (0.238) (0.231) (0.185) (0.212)
Annual contractual hours 140.347 -3.177 77.618 -9.175 4,157 95.249 -10.306 27.711 -315.421* 982
(86.411) (83.735) (71.408) (75.550) (162.702) (189.979) (118.055) (167.193)
EP
C
who were not admitted but had an exam score in the 90th percentile or higher. Instruments include indicator for above cutoff x selective major x male applicant, indicator for above cutoff x
selective major x female applicant, indicator for above cutoff x non-selective major x male applicant, and indicator for above cutoff x non-selective major x female applicant. Additional
AC
controls include normalized entrance exam score x male applicant, normalized entrance exam score x female applicant, indicator for above cutoff x normalized entrance exam score x male
applicant, indicator for above cutoff x normalized entrance exam score x female applicant, indicator for female applicant, and a set of binary indicators for admissions track. Robust standard
errors in parentheses are adjusted for clustering on tracks. ***, **, and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively.
47
RD analysis of 2012 education and labor outcomes, interactions with major-specific mismatch
Mismatch = percentage of Q students who would not have been admitted in the hypothetical absence of quotas
mismatch x above cutoff mismatch x above cutoff mismatch x above cutoff
Dependent variable x NQ xQ N x NQ xQ N x NQ xQ N
PT
Years of education 0.000 -0.011 5,705 0.004 -0.006 3,398 -0.005 -0.018* 2,307
(0.003) (0.007) (0.004) (0.012) (0.004) (0.011)
RI
College completion -0.000 -0.003 5,705 0.001 -0.002 3,398 -0.002* -0.005 2,307
(0.001) (0.002) (0.001) (0.003) (0.001) (0.003)
In RAIS (2009-2012) -0.000 0.000 7,747 -0.001 0.000 4,607 -0.000 -0.001 3,140
SC
(0.001) (0.001) (0.001) (0.002) (0.001) (0.002)
Formal employment -0.001 -0.002 7,747 -0.001 -0.002 4,607 -0.001 -0.003 3,140
(0.001) (0.001) (0.001) (0.002) (0.001) (0.002)
U
Director/manager -0.001 0.001 4,595 -0.001 0.001 2,797 -0.000 0.001 1,798
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001)
AN
Professional in sciences/arts -0.001 0.001 4,595 -0.001 0.003 2,797 0.000 -0.002 1,798
(0.001) (0.002) (0.001) (0.003) (0.002) (0.004)
Tenure in main job 0.029 -0.176 4,601 0.026 -0.171 2,800 0.055 -0.158 1,801
M
(0.125) (0.248) (0.192) (0.450) (0.154) (0.254)
Public sector job -0.001 -0.004* 4,601 0.001 -0.003 2,800 -0.003* -0.009** 1,801
(0.001) (0.002) (0.002) (0.004) (0.002) (0.004)
D
Log annual labor earnings -0.001 0.003 5,128 0.001 0.008 3,077 -0.004 -0.000 2,051
(0.002) (0.004) (0.004) (0.006) (0.003) (0.006)
TE
Annual contractual hours -1.223 1.929 5,139 0.813 4.809 3,085 -4.129 0.081 2,054
(1.621) (3.041) (2.542) (3.893) (2.567) (5.250)
EP
NOTE. The analysis in this table is analogous to that in Table 4. Here, interactions between mismatch, above cutoff indicator, and quota status are added to models, and their corresponding
C
coefficients are reported in the table. The measure of mismatch, which varies by major, is the percentage of Q students who would not have been admitted in the hypothetical absence of
quotas. Robust standard errors in parentheses are adjusted for clustering on tracks. ***, **, and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively.
AC
48
Table A1 ACCEPTED MANUSCRIPT
Fuzzy RD analysis of 2012 education and labor outcomes, narrowing the estimation sample towards cutoffs
Estimation sample excludes the bottom 10% of applicants below cutoffs and the top 10% of applicants above
enrolled enrolled enrolled enrolled enrolled enrolled
Dependent variable x non-quota x quota N x non-quota x quota N x non-quota x quota N
PT
Years of education 0.095 0.426** 5,094 0.057 0.660** 3,003 0.115 0.051 2,091
(0.095) (0.186) (0.138) (0.290) (0.117) (0.292)
RI
College completion 0.034 0.088* 5,094 0.023 0.139* 3,003 0.040 -0.030 2,091
(0.029) (0.052) (0.040) (0.076) (0.036) (0.083)
In RAIS (2009-2012) 0.057** 0.022 6,945 0.065** 0.051 4,093 0.057 0.007 2,852
SC
(0.025) (0.039) (0.031) (0.061) (0.039) (0.072)
Formal employment 0.055** 0.010 6,945 0.065** 0.044 4,093 0.048 -0.046 2,852
(0.026) (0.045) (0.033) (0.066) (0.039) (0.078)
U
Director/manager -0.038** 0.060** 4,084 -0.021 0.056 2,461 -0.051 0.093** 1,623
(0.016) (0.027) (0.023) (0.047) (0.032) (0.047)
AN
Professional in sciences/arts 0.050 -0.004 4,084 -0.016 0.035 2,461 0.135** -0.137 1,623
(0.042) (0.069) (0.053) (0.095) (0.062) (0.114)
Tenure in main job 1.186 -3.456 4,089 5.577 2.957 2,464 -3.885 -0.713 1,625
M
(4.043) (9.370) (5.863) (12.031) (5.157) (10.684)
Public sector job 0.066** -0.076 4,089 0.091** -0.119 2,464 0.030 0.043 1,625
(0.034) (0.072) (0.043) (0.096) (0.057) (0.139)
D
Log annual labor earnings 0.106 0.055 4,569 0.181** 0.332* 2,716 0.025 -0.045 1,853
(0.077) (0.139) (0.092) (0.195) (0.118) (0.223)
TE
Annual contractual hours 83.346 -114.610 4,580 168.671** 16.865 2,724 -33.221 -205.100 1,856
(53.711) (110.980) (79.862) (123.747) (75.752) (151.246)
EP
C

AC
Years of education 0.078 0.408** 4,294 0.032 0.550* 2,533 0.114 0.062 1,761
(0.107) (0.195) (0.159) (0.305) (0.136) (0.320)
College completion 0.023 0.085 4,294 0.008 0.121 2,533 0.040 -0.028 1,761
(0.032) (0.055) (0.044) (0.082) (0.044) (0.090)
In RAIS (2009-2012) 0.053* 0.009 5,870 0.058* 0.027 3,445 0.060 -0.003 2,425
(0.027) (0.044) (0.035) (0.066) (0.043) (0.081)
Formal employment 0.047 -0.011 5,870 0.044 -0.001 3,445 0.065 -0.064 2,425
49
(0.029) (0.049)
ACCEPTED
(0.037)
MANUSCRIPT
(0.070) (0.045) (0.088)
Director/manager -0.049*** 0.075*** 3,432 -0.038 0.084** 2,073 -0.056* 0.102** 1,359
(0.018) (0.027) (0.026) (0.043) (0.032) (0.050)
Professional in sciences/arts 0.058 0.022 3,432 -0.016 0.067 2,073 0.185*** -0.070 1,359
(0.045) (0.075) (0.057) (0.104) (0.067) (0.124)
Tenure in main job -0.373 -3.672 3,436 3.725 7.642 2,075 -7.292 -7.400 1,361
(4.054) (10.314) (5.905) (14.718) (5.295) (9.809)
Public sector job 0.061* -0.139* 3,436 0.082* -0.160 2,075 0.035 -0.044 1,361
(0.033) (0.079) (0.044) (0.103) (0.061) (0.151)
PT
Log annual labor earnings 0.087 0.018 3,855 0.166* 0.332 2,295 -0.023 -0.118 1,560
(0.079) (0.145) (0.097) (0.203) (0.125) (0.234)
Annual contractual hours 80.293 -125.356 3,862 184.678** 27.628 2,299 -61.004 -213.590 1,563
RI
(55.750) (115.399) (84.312) (128.048) (80.956) (152.287)
SC
U
AN
Years of education 0.079 0.404* 2,958 -0.046 0.320 1,755 0.215 0.188 1,203
(0.118) (0.230) (0.170) (0.400) (0.160) (0.364)
College completion 0.017 0.095 2,958 -0.022 0.089 1,755 0.071 0.018 1,203
M
(0.037) (0.065) (0.052) (0.108) (0.051) (0.103)
In RAIS (2009-2012) 0.073** 0.003 4,022 0.081** 0.053 2,367 0.083 -0.039 1,655
(0.033) (0.049) (0.040) (0.082) (0.054) (0.095)
D
Formal employment 0.077** -0.006 4,022 0.079* 0.045 2,367 0.090 -0.098 1,655
(0.035) (0.056) (0.042) (0.088) (0.057) (0.099)
TE
Director/manager -0.041** 0.064** 2,355 -0.033 0.045 1,422 -0.021 0.100* 933
(0.020) (0.033) (0.029) (0.053) (0.039) (0.059)
Professional in sciences/arts 0.076* 0.008 2,355 -0.006 -0.009 1,422 0.154** -0.005 933
EP
(0.045) (0.084) (0.056) (0.132) (0.078) (0.139)
Tenure in main job 0.608 -6.521 2,357 6.420 2.275 1,423 -6.033 -14.830 934
(4.676) (9.117) (7.192) (15.083) (5.025) (11.876)
C
Public sector job 0.032 -0.211** 2,357 0.045 -0.220 1,423 0.052 -0.158 934
(0.041) (0.094) (0.052) (0.143) (0.074) (0.160)
AC
Log annual labor earnings 0.106 -0.090 2,654 0.112 0.277 1,575 0.121 -0.319 1,079
(0.084) (0.146) (0.108) (0.243) (0.135) (0.230)
Annual contractual hours 61.925 -186.262 2,660 123.806 22.936 1,579 -45.697 -114.897 1,081
(59.725) (122.512) (85.313) (147.298) (98.093) (144.975)
50
ACCEPTED
NOTE. The analysis in this table is analogous to that in Table 4. Here, the estimation MANUSCRIPT
sample is progressively narrowed towards admissions thresholds by excluding the bottom X% of
applicants below cutoffs and the top X% of applicants above cutoffs. Robust standard errors in parentheses are adjusted for clustering on tracks. ***, **, and * indicate statistical significance
at the 1%, 5%, and 10% levels, respectively.
PT
RI
U SC
AN
M
D
TE
C EP
AC
51
Fuzzy RD analysis of 2012 education and labor outcomes, excluding applicants exactly at the thresholds

Years of education 0.083 0.652*** 5,453 0.034 0.953*** 3,251 0.143 0.396 2,202
PT
(0.086) (0.183) (0.122) (0.303) (0.113) (0.284)
College completion 0.036 0.155*** 5,453 0.024 0.212*** 3,251 0.052 0.077 2,202
(0.026) (0.050) (0.035) (0.077) (0.035) (0.081)
RI
In RAIS (2009-2012) 0.055** 0.027 7,428 0.056* 0.038 4,425 0.057 0.024 3,003
(0.024) (0.039) (0.029) (0.058) (0.040) (0.072)
Formal employment 0.053** 0.004 7,428 0.063** 0.024
SC
4,425 0.040 -0.037 3,003
(0.025) (0.043) (0.032) (0.065) (0.039) (0.076)
Director/manager -0.034** 0.062** 4,395 -0.017 0.095* 2,680 -0.054 0.055 1,715
(0.017) (0.027) (0.021) (0.050) (0.035) (0.038)
U
Professional in sciences/arts 0.037 0.039 4,395 -0.025 0.056 2,680 0.134** -0.003 1,715
(0.041) (0.069) (0.049) (0.093) (0.063) (0.119)
AN
(3.745) (8.970) (4.994) (13.769) (5.116) (9.574)
Public sector job 0.079*** -0.040 4,401 0.129*** -0.057 2,683 -0.009 0.033 1,718
M
(0.030) (0.068) (0.039) (0.095) (0.053) (0.128)
Log annual labor earnings 0.132* 0.031 4,901 0.207** 0.324 2,944 0.045 0.043 1,957
(0.072) (0.144) (0.086) (0.208) (0.108) (0.224)
D
Annual contractual hours 64.612 -72.609 4,912 129.349* 68.233 2,952 -34.766 -137.480 1,960
TE
(49.187) (104.942) (72.372) (125.619) (71.837) (176.590)
EP
NOTE. The analysis in this table is analogous to that in Table 4, with one exception. Here, applicants who had a normalized entrance exam score of zero, exactly at the thresholds, are
excluded from the estimation sample. Robust standard errors in parentheses are adjusted for clustering on tracks. ***, **, and * indicate statistical significance at the 1%, 5%, and 10% levels,
respectively.
C
AC
52
Fuzzy RD analysis of labor outcomes in 2012 or most recent year around UnB admission cutoffs for applicants in 2004-2005

Director/manager -0.028* 0.059*** 5,280 -0.018 0.070** 3,180 -0.039 0.059 2,100
PT
(0.014) (0.021) (0.018) (0.035) (0.029) (0.036)
Professional in sciences/arts 0.046 0.030 5,280 -0.001 0.091 3,180 0.128** 0.013 2,100
(0.036) (0.061) (0.045) (0.084) (0.056) (0.100)
RI
(3.165) (7.812) (4.133) (11.116) (4.323) (8.085)
Public sector job 0.062** -0.034 5,287 0.095*** -0.039
SC
3,185 0.003 -0.003 2,102
(0.028) (0.059) (0.036) (0.085) (0.047) (0.107)
Log annual labor earnings 0.164** 0.064 5,698 0.221** 0.378** 3,394 0.084 -0.062 2,304
(0.071) (0.126) (0.087) (0.180) (0.102) (0.177)
U
Annual contractual hours 83.454* -68.400 5,698 143.858** 106.362 3,394 -29.894 -209.534 2,304
(46.442) (96.845) (67.324) (110.836) (65.678) (142.475)
AN
M
NOTE. The analysis in this table is analogous to that in Table 4. Outcomes are constructed from 2012 or the most recent year available between 2009 and 2012. Additional controls include a
set of indicators for the year in which the outcome is observed. Robust standard errors in parentheses are adjusted for clustering on tracks. ***, **, and * indicate statistical significance at the
D
1%, 5%, and 10% levels, respectively.
TE
C EP
AC
53
Fuzzy RD analysis of 2012 education and labor outcomes, excluding applicants with "partial name" CPF matches

Years of education 0.122 0.466** 5,109 0.085 0.833*** 3,071 0.156 0.097 2,038
PT
(0.089) (0.186) (0.128) (0.283) (0.109) (0.300)
College completion 0.045* 0.104** 5,109 0.033 0.191*** 3,071 0.059* 0.012 2,038
(0.026) (0.051) (0.036) (0.073) (0.034) (0.082)
RI
In RAIS (2009-2012) 0.061** 0.017 7,151 0.063** 0.025 4,280 0.065 0.000 2,871
(0.026) (0.037) (0.032) (0.057) (0.040) (0.074)
Formal employment 0.056** -0.002 7,151 0.065* 0.020
SC
4,280 0.045 -0.055 2,871
(0.027) (0.042) (0.034) (0.062) (0.041) (0.080)
Director/manager -0.037** 0.083*** 4,110 -0.031 0.103** 2,519 -0.035 0.068 1,591
(0.017) (0.026) (0.023) (0.047) (0.033) (0.042)
U
Professional in sciences/arts 0.049 0.011 4,110 -0.014 0.043 2,519 0.140** -0.077 1,591
(0.041) (0.067) (0.049) (0.097) (0.067) (0.117)
AN
Tenure in main job -0.476 -10.825 4,116 4.953 -4.343 2,522 -8.438* -19.508** 1,594
(3.574) (8.390) (5.247) (12.368) (4.689) (8.676)
Public sector job 0.072** -0.097 4,116 0.121*** -0.102 2,522 -0.006 -0.071 1,594
M
(0.031) (0.068) (0.040) (0.095) (0.054) (0.135)
Log annual labor earnings 0.155** 0.002 4,593 0.220** 0.285 2,782 0.079 -0.147 1,811
(0.075) (0.121) (0.088) (0.185) (0.110) (0.198)
D
Annual contractual hours 67.053 -111.740 4,601 122.626* 15.967 2,787 -23.891 -226.719 1,814
TE
(49.163) (109.213) (70.849) (115.802) (73.897) (164.450)
EP
NOTE. The analysis in this table is analogous to that in Table 4, with one exception. Here, applicants with "partial name" CPF matches are excluded from the estimation sample. Robust
standard errors in parentheses are adjusted for clustering on tracks. ***, **, and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively.
C
AC
54
Fuzzy RD analysis of 2012 education and labor outcomes, including quadratic controls in models

Years of education 0.054 0.412 5,705 -0.005 0.663* 3,398 0.063 -0.001 2,307
PT
(0.126) (0.251) (0.175) (0.387) (0.162) (0.432)
College completion 0.010 0.085 5,705 -0.007 0.146 3,398 0.018 -0.044 2,307
(0.039) (0.070) (0.051) (0.102) (0.051) (0.118)
RI
In RAIS (2009-2012) 0.067** 0.082* 7,747 0.081* 0.146* 4,607 0.060 -0.008 3,140
(0.033) (0.048) (0.041) (0.078) (0.051) (0.086)
Formal employment 0.073** 0.064 7,747 0.091** 0.133
SC
4,607 0.061 -0.071 3,140
(0.035) (0.057) (0.045) (0.089) (0.053) (0.095)
Director/manager -0.035 0.093*** 4,595 -0.043 0.134** 2,797 0.013 0.074 1,798
(0.022) (0.032) (0.029) (0.066) (0.039) (0.051)
U
Professional in sciences/arts 0.094** -0.098 4,595 0.024 -0.104 2,797 0.211** -0.134 1,798
(0.046) (0.086) (0.056) (0.125) (0.084) (0.143)
AN
Tenure in main job -1.716 -18.394* 4,601 3.298 -8.552 2,800 -7.083 -14.401 1,801
(4.769) (9.603) (6.365) (15.501) (6.156) (12.186)
Public sector job 0.063 -0.160 4,601 0.071 -0.170 2,800 0.071 -0.080 1,801
M
(0.040) (0.104) (0.050) (0.146) (0.074) (0.170)
Log annual labor earnings 0.002 -0.143 5,128 0.016 0.590** 3,077 0.039 -0.537** 2,051
(0.085) (0.175) (0.111) (0.254) (0.138) (0.257)
D
Annual contractual hours 40.616 -207.629 5,139 62.752 173.826 3,085 34.002 -419.738** 2,054
TE
(64.010) (142.051) (83.048) (165.527) (108.232) (199.893)
EP
NOTE. The analysis in this table is analogous to that in Table 4. However, the following controls are added to regressions: normalized entrance exam score squared x non-quota applicant,
normalized entrance exam score squared x quota applicant, indicator for above cutoff x normalized entrance exam score squared x non-quota applicant, indicator for above cutoff x
normalized entrance exam score squared x quota applicant. Robust standard errors in parentheses are adjusted for clustering on tracks. ***, **, and * indicate statistical significance at the
C
1%, 5%, and 10% levels, respectively.

AC
55
RD analysis of 2012 education and labor outcomes, interactions with major-specific mismatch
Mismatch = percentage of NQ students with admission scores higher than the median Q student's score
PT
Years of education 0.001 -0.007 5,705 0.004 -0.002 3,398 -0.003 -0.017 2,307
(0.003) (0.007) (0.004) (0.010) (0.004) (0.011)
RI
College completion 0.000 -0.001 5,705 0.001 -0.001 3,398 -0.002 -0.005 2,307
(0.001) (0.002) (0.001) (0.003) (0.001) (0.003)
In RAIS (2009-2012) -0.001 -0.000 7,747 -0.001 -0.001 4,607 -0.001 0.002 3,140
SC
(0.001) (0.001) (0.001) (0.002) (0.001) (0.002)
Formal employment -0.001* -0.001 7,747 -0.001 -0.001 4,607 -0.002 0.000 3,140
(0.001) (0.001) (0.001) (0.002) (0.001) (0.002)
U
Director/manager 0.000 0.001 4,595 0.000 -0.000 2,797 0.000 0.001 1,798
(0.000) (0.001) (0.001) (0.001) (0.001) (0.001)
AN
Professional in sciences/arts -0.001 -0.001 4,595 -0.002* -0.001 2,797 0.000 -0.002 1,798
(0.001) (0.002) (0.001) (0.003) (0.002) (0.003)
Tenure in main job 0.072 0.050 4,601 0.105 0.168 2,800 0.113 -0.083 1,801
M
(0.157) (0.241) (0.212) (0.380) (0.181) (0.285)
Public sector job -0.001 -0.003 4,601 0.001 -0.002 2,800 -0.003* -0.008** 1,801
(0.001) (0.002) (0.002) (0.003) (0.002) (0.004)
D
Log annual labor earnings 0.001 0.000 5,128 0.003 0.004 3,077 -0.003 -0.007 2,051
(0.002) (0.004) (0.004) (0.006) (0.003) (0.006)
TE
Annual contractual hours -0.641 3.621 5,139 0.818 4.818 3,085 -2.699 1.279 2,054
(1.473) (3.168) (2.042) (3.564) (2.512) (5.561)
Mismatch = difference between NQ and Q cutoff scores
EP
C
AC
Years of education 0.001 -0.006 5,705 0.003 -0.001 3,398 -0.001 -0.013* 2,307
(0.002) (0.004) (0.003) (0.008) (0.002) (0.007)
College completion 0.000 -0.002 5,705 0.001 -0.000 3,398 -0.001 -0.003* 2,307
(0.001) (0.001) (0.001) (0.002) (0.001) (0.002)
In RAIS (2009-2012) -0.000 0.000 7,747 -0.000 0.001 4,607 0.000 -0.002 3,140
(0.000) (0.001) (0.001) (0.001) (0.001) (0.001)
Formal employment -0.000 -0.001 7,747 -0.000 0.000 4,607 -0.000 -0.002* 3,140
56
(0.001) (0.001)
ACCEPTED
(0.001)
MANUSCRIPT
(0.002) (0.001) (0.001)
Director/manager -0.000 0.001 4,595 -0.000 0.001 2,797 -0.000 0.000 1,798
(0.000) (0.000) (0.000) (0.001) (0.001) (0.000)
Professional in sciences/arts -0.001 0.001 4,595 -0.001 -0.000 2,797 -0.000 0.001 1,798
(0.001) (0.002) (0.001) (0.003) (0.001) (0.002)
Tenure in main job 0.027 -0.177 4,601 -0.007 -0.076 2,800 0.102 -0.163 1,801
(0.072) (0.149) (0.106) (0.300) (0.092) (0.163)
Public sector job 0.000 -0.002* 4,601 0.001 -0.002 2,800 -0.001 -0.004 1,801
(0.001) (0.001) (0.001) (0.002) (0.001) (0.003)
PT
Log annual labor earnings -0.001 -0.001 5,128 -0.001 -0.001 3,077 -0.001 0.000 2,051
(0.001) (0.003) (0.002) (0.004) (0.002) (0.004)
Annual contractual hours -0.921 -0.638 5,139 -0.105 0.271 3,085 -1.965 -0.659 2,054
RI
(1.059) (2.019) (1.422) (2.935) (1.767) (3.286)
SC
NOTE. The analysis in this table is analogous to that in Table 6. Here, interactions between mismatch, above cutoff indicator, and quota status are added to models, and their corresponding
coefficients are reported in the table. The measures of mismatch are the percentage of NQ students with admission scores higher than the median Q student's score and the average difference
U
between NQ and Q cutoff scores. Robust standard errors in parentheses are adjusted for clustering on tracks. ***, **, and * indicate statistical significance at the 1%, 5%, and 10% levels,
respectively.
AN
M
D
TE
C EP
AC
57
ACCEPTED MANUSCRIPT
HIGHLIGHTS
Title: “Black Movement: Using discontinuities in admissions to study the effects of college
quality and affirmative action”
Authors: Andrew Francis-Tan and Maria Tannuri-Pianto
• Brazil is an insightful place to study affirmative action in college admissions

• This paper examines the effects of racial quotas at a selective university
PT
• A regression discontinuity design is implemented with multiple thresholds
• We find education and labor benefits, particularly for male quota applicants
• Findings confirm the importance of college quality in a setting outside of the U.S.
RI
U SC
AN
M
D
TE
C EP
AC

Francis-Tan, Tannuri-Pianto - 2018 - Black Movement Using Discontinuities in Admissions To Study The Effects of College Quality and Affi

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Francis-Tan, Tannuri-Pianto - 2018 - Black Movement Using Discontinuities in Admissions To Study The Effects of College Quality and Affi

Uploaded by

Copyright:

Available Formats

Accepted Manuscript

Black Movement: Using discontinuities in admissions to study the effects of college

Andrew Francis-Tan, Maria Tannuri-Pianto

To appear in: Journal of Development Economics

Received Date: 4 September 2017

JEL codes: I23, I25, I26, J15, J24.

JEL codes: I23, I25, I26, J15, J24.

Affirmative action policies – which aim to promote the status of historically

for identifying the causal effects of Educational Equity Policies.1

or light-skinned, 45.0% "pardo" or brown-skinned, and 8.6% "preto" or dark-skinned (IBGE,

affirmative action in admissions.

2.1 Effect of affirmative action on educational outcomes

Most previous studies on affirmative action in college admissions focus on educational

attended lower ranked campuses.

credentials. Differences in outcomes may be attributed to mismatch as blacks tend to attend

higher-ranked schools due to admissions preferences. Their findings reveal no evidence of

not adversely impact graduation rates.

implementation of quotas. The authors obtained administrative records as well as conducted a

families with lower socioeconomic status than displaced applicants. Difference-in-difference

raise earnings by moving minorities from lower to higher quality colleges.

An emerging, but related, literature examines heterogeneity in the returns to college

less selective colleges. In a seminal contribution, Arcidiacono (2005) estimates a forward-

income losses for displaced applicants.

race-based affirmative action: increasing social awareness of racial inequality, formal

society, and a campus meeting space to study and interact.

admitted by UnB in the hypothetical absence of quotas. To identify displacing students, we

been admitted in the absence of quotas.

To summarize, due to the admissions policy, the differences in academic credentials

3.2 Administrative data on applicants

quota applicants completed the questionnaire.

a more recent admissions cycle.

Emprego, 2009-2012). Employers annually provide information on employees to the Ministry of

are self-employed but have not formally established a company.

RAIS is used to construct outcomes. "Years of education" is the number of years of

"Director/manager" is a binary indicator that equals one if a person's main occupation in

duration of employment for each job.

3.4 Sample construction

robustness of our results with respect to match quality.

RAIS due to our error.

comparable socioeconomic backgrounds as measured by mother's education. For non-quota

anything, our effects on earnings are underestimated.

3.5 Statistical procedure

We implement a regression discontinuity (RD) design with the sample of high-

illustrates the discontinuity in admission at track-specific cutoffs. This discontinuity is almost

their analytical essay received a failing grade.

performing rejected and accepted applicants.

Part of the empirical analysis employs an "intent-to-treat" model. Outcomes ܻ௜ are

include regressors fully interacted with indicators for gender.

equation is the following:

‫ܧ‬௜ = ߙ଴ + ߙଵ ‫ܣ‬௜ + ߙଶ ܵ௜ + ߙଷ ܵ௜ ∙ ‫ܣ‬௜ + ߛ௧ + ߤ௜ ,

The second-stage equation is then:

variable. An emergent literature in econometrics is beginning to understand the particular

optimal asymptotic MSE-minimizing bandwidth based on the entire sample of observations

progressively narrowed towards admissions thresholds by excluding the bottom X% of

across programs, and programs have different probabilities to be observed in a neighborhood of

interactions. It may be informative to explore robustness with respect to the inclusion of

quadratic controls. Table A5 displays the results with quadratic controls.

4.1 Main results

Table 4 displays fuzzy RD regressions with all high-performing applicants as well as

statistically significant. In contrast, quota applicants experience significant gains in educational

4.2 Gender and major selectivity

It is difficult to know whether these results are explained by gender directly or by