Essays in Economics of Education

A DISSERTATION SUBMITTED TO THE FACULTY OF THE GRADUATE SCHOOL OF THE UNIVERSITY OF MINNESOTA BY

Claudia Bueno Rocha Vidigal

IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY

Paul Glewwe, Adviser

February 2019

I would like to express my deepest gratitude to my advisor, professor Paul Glewwe, for his continuous support, guidance and encouragement. I had the great fortune to have worked with him and I am deeply indebted to his patient advice over the years. I am also grateful to my committee members, professors Deborah Levison, Joe Ritter, and Marc Bellemare, for their thoughtful comments and suggestions on my dissertation. I am also thankful for having had the opportunity to meet amazing colleagues and professors in my doctoral program. This long process would not have been possible without their support and friendship. I am profoundly grateful to my family, especially my parents, who have been a great source of support throughout my academic journey. Lastly, but most importantly, my husband and classmate, Vinicius Vidigal, deserves heartfelt thanks for his love, encouragement, and support.

i DEDICATION

To my mother, Katia Lucila Bueno, and my grandmother, Maria Angela Sotovia Bueno, whose love, encouragement and support made all this possible.

ii ABSTRACT

This dissertation consists of three essays in the economics of education. It investigates the effects of educational programs in primary and secondary schools in Brazil, as well as the effects of racial and low-income quotas in Brazilian universities. The first essay analyzes the impact of Brazil’s Multifunctional Resources Classroom Inclusion Program on the academic outcomes of disabled and non-disabled students in primary and secondary schools. School fixed effects estimations show that, in general, the Brazilian inclusion program benefits disabled students, especially those enrolled in grades 6-9 and 10-12, with no negative spillover effects onto non-disabled students. The second essay investigates the impact of the Mais Educação Extended School Day Program on academic outcomes in Brazil. The results suggest that this Brazilian longer school day program reduces the dropout rates of students in all grade levels, raises the enrollment of students in grades 6- 9, but reduces the enrollment of students in grades 10-12. Moreover, the estimates indicate that the impact on grade promotion is positive for students in grades 6-9, but negative for students in lower grades. Finally, the program seems to increase repetition rates for students in all grade levels. The third essay evaluates the impact of racial and low-income quotas on the academic performance of senior students in Brazilian colleges and universities. Using a panel data approach with school fixed effects, the results show that both the proportion of racial quota students and the proportion of low-income quota students have no statistically significant impact on the academic performance of either quota or non- quota students.

iii CONTENTS

List of Tables……………………………………………………………………… vi

List of Figures…………………………………………………………………….. ix

1 Introduction……………………………………………………………...... 1 2 Inclusive Education and Academic Outcomes: The Impact of Brazil’s Multifunctional Resources Classroom Inclusion Program…………………….. 3 2.1 Introduction……………………………………………………………….. 3 2.2 The Brazilian Multifunctional Resources Classroom Inclusion Program………………………………………………………………… 6 2.3 Data and Descriptive Statistics……………………………………………. 8 2.4 Empirical Framework……………………………………………………... 12 2.4.1 Equations for Estimation……………………………………………. 13 2.4.2 Identification Strategy………………………………………………. 16 2.5 Results…………………………………………………………………….. 17 2.5.1 Main Results………………………………………………………… 18 2.5.2 Impact by Time of Program Adoption and Cumulative Impact……… 2 0 2.5.3 Robustness Checks………………………………………………….. 22 2.6 Conclusion………………………………………………………………… 23 3 Impact of the Mais Educação Extended School Day Program on Academic Outcomes in Brazil……………………………………………………………... 42 3.1 Introduction………………………………………………………………… 42 3.2 Description of the Mais Educação Extended School Day Program………… 45 3.3 Data and Descriptive Statistics……………………………………………... 46 3.4 Empirical Strategy………………………………………………………….. 50 3.4.1 Estimation and Identification Strategies……………………………… 50 3.5 Results……………………………………………………………………… 55 3.6 Conclusion…………………………………………………………………. 60 4 Affirmative Action in Brazilian Universities: The Impact of Racial and Low- Income Quotas on Academic Performance……………………………………... 75

iv 4.1 Introduction………………………………………………………………… 75 4.2 Racial and Low-Income Quotas and the Admission Process in Brazilian Universities………………………………………………………………... 79 4.3 Data and Descriptive Statistics……………………………………………... 82 4.4 Empirical Strategy………………………………………………………….. 85 4.4.1 Estimation and Identification Strategies…………………………….... 86 4.5 Results……………………………………………………………………… 89 4.6 Conclusion…………………………………………………………………. 92

5 Conclusion……………………………………………………………………… 99

References………………………………………………………………………… 100

Appendix A……………………………………………………………………….. 107

Appendix B……………………………………………………………………….. 108

v LIST OF TABLES

2.1 Number of Schools in Brazil’s School Census from 1999 to 2015……………... 25

2.2 Descriptive Statistics for Eventually Treated and Never Treated Schools…...... 26

2.3 Estimates of the Program Impact on Log of Enrollment of Disabled and Non- Disabled Students: Results for Schools with Grades 1-5, 6-9 and 10-12, 1999- 2015……………………………………………………...... 28

2.4 Estimates of the Program Impact on Log of Enrollment of Disabled and Non- Disabled Students at the Municipio-Level: Results for Schools with Grades 1- 5, 6-9 and 10-12, 1999-2015…………………………………………………... 28

2.5 Estimates of the Program Impact on Dropout, Repetition, and Grade Promotion Rates: Results for Schools with Grades 1-5, 6-9 and 10-12, 1999-2015……….. 29

2.6 Estimates of the Program Impact on Log of Enrollment of Disabled and Non- Disabled Students by Time of Adoption and Cumulative Impact: Results for Schools with Grades 1-5, 6-9 and 10-12, 1999-2015…………………………... 30

2.7 Estimates of the Program Impact on Dropout Rate by Time of Adoption and Cumulative Impact: Results for Schools with Grades 1-5, 6-9 and 10-12, 1999- 2015…………………………………………………………………………… 31

2.8 Estimates of the Program Impact on Repetition Rate by Time of Adoption and Cumulative Impact: Results for Schools with Grades 1-5, 6-9 and 10-12, 1999- 2015…………………………………………………………………………… 32

2.9 Estimates of the Program Impact on Grade Promotion Rate by Time of Adoption and Cumulative Impact: Results for Schools with Grades 1-5, 6-9 and 10-12, 1999-2015………………………………………………………….. 33

2.10 Placebo Test: Estimates of the Program Impact for Schools with Grades 1-5, 6-9, and 10-12, 1999-2005 (Schools with Program in 2007 Assigned to 2005) 34

2.11 Robustness Check: Estimates of the Program Impact after Excluding the Smallest Schools with Grades 1-5, 6-9, and 10-12, 1999-2015…………….… 35

3.1 Number of Schools Participating in the Mais Educação Extended School Day Program, 2008-2014…………………………………………………………... 62

vi 3.2 Number of Schools in Brazil’s School Census from 1999 to 2014……………... 62

3.3 Descriptive Statistics for Eventually Treated and Never Treated Schools……... 63

3.4 Estimates of the Program Impact on Log of Enrollment: Results for Schools with Grades 1-5, 6-9 and 10-12, 1999-2014…………………………………... 64

3.5 Estimates of the Program Impact on Log of Enrollment at the Municipio-Level: Results for Schools with Grades 1-5, 6-9 and 10-12, 1999-2014………………. 64

3.6 Estimates of the Program Impact on Dropout, Repetition, and Grade Promotion Rates: Results for Schools with Grades 1-5, 6-9 and 10-12, 1999-2014……….. 65

3.7 Estimates of the Program Impact on Log of Enrollment by Time of Adoption and Cumulative Effects: Results for Schools with Grades 1-5, 6-9 and 10-12, 1999-2014…………………………………………………………………….. 66

3.8 Estimates of the Program Impact on Dropout Rate by Time of Adoption and Cumulative Effects: Results for Schools with Grades 1-5, 6-9 and 10-12, 1999- 2014…………………………………………………………………………… 67

3.9 Estimates of the Program Impact on Repetition Rate by Time of Adoption and Cumulative Effects: Results for Schools with Grades 1-5, 6-9 and 10-12, 1999- 2014……………………………………………………………………………. 68

3.10 Estimates of the Program Impact on Grade Promotion Rate by Time of Adoption and Cumulative Effects: Results for Schools with Grades 1-5, 6-9 and 10-12, 1999-2014………………………………………………………... 69

3.11 Placebo Test: Estimates of the Program Impact for Schools with Grades 1-5, 6-9 and 10-12, 1999-2007 (Schools with Program in 2008 Assigned to 2007) 70

4.1 Definition of Subsamples and Corresponding Fields…………………………... 93

4.2 Number of Colleges in Brazil’s Higher Education Census, 2010-2015………... 93

4.3 Number of Senior College Students under Racial Quota: Enrolled Students in Traditional (Non-Distance) Public and Private Learning, 2010-2015…………. 94

4.4 Number of Senior College Students under Low-Income Quota: Enrolled Students in Traditional (Non-Distance) Public and Private Learning, 2010- 2015…………………………………………………………………………... 94

vii 4.5 Number of Colleges with at Least One Quota Senior Student, 2010-2015……... 94

4.6 ENADE Test Scores for Colleges from Groups A, B, and C, Racial Quota……. 95

4.7 ENADE Test Scores for Colleges from Groups A, B, and C, Low-Income Quota…………………………………………………………………………. 95

4.8 Estimates of the Impact of Racial and Low-Income Quotas on Academic Performance on the ENADE Test: Results for Groups A, B, and C…………… 96

4.9 Estimates of the Impact of Racial and Low-Income Quotas on Academic Performance on the ENADE Test for Public Colleges: Results for Groups A, B, and C………………………………………………………………………. 97

A.1 Average Enrollment of Disabled and Non-Disabled Students in Treated and Never Treated Schools with Grades 1-5, 6-9 and 10-12, 2007-2015………….. 107

B.1 Descriptive Statistics for Groups A, B, and C: Racial Quota…………………... 108

B.2 Descriptive Statistics for Groups A, B, and C: Low-Income Quota…………… 110

B.3 Estimates of the Impact of Racial and Low-Income Quotas on Racial Composition: Results for Groups A, B, and C………………………………... 112

viii LIST OF FIGURES

2.1 Number of Schools (and Inclusive Classrooms) in the Multifunctional Resources Classroom Inclusion Program, 2005-2015………………………... 36

2.2 Number of Students with Special Educational Needs Enrolled in Regular Public Schools and in Special Public Schools, 2003-2014……………………. 36

2.3 Average Enrollment of All Students in Treated and Never Treated Schools with Grades 1-5, 6-9, and 10-12, 1999-2015………………………………… 37

2.4 Average Enrollment of Disabled Students in Treated and Never Treated Schools with Grades 1-5, 6-9, and 10-12, 2007-2015…………………………. 38

2.5 Dropout Rates for Treated and Never Treated Schools with Grades 1-5, 6-9, and 10-12, 1999-2015………………………………………………………… 39

2.6 Repetition Rates for Treated and Never Treated Schools with Grades 1-5, 6-9, and 10-12, 1999-2015…………………………………………………………. 40

2.7 Grade Promotion Rates for Treated and Never Treated Schools with Grades 1- 5, 6-9, and 10-12, 1999-2015………………………………………………….. 41

3.1 Total Enrollment for Treated and Never Treated Schools with Grades 1-5, 6- 9, and 10-12, 1999-2014………………………………………………………. 71

3.2 Dropout Rates for Treated and Never Treated Schools with Grades 1-5, 6-9, and 10-12, 1999-2014………………………………………………………….. 72

3.3 Repetition Rates for Treated and Never Treated Schools with Grades 1-5, 6-9, and 10-12, 1999-2014…………………………………………………………. 73

3.4 Grade Promotion Rates for Treated and Never Treated Schools with Grades 1- 5, 6-9, and 10-12, 1999-2014………………………………………………….. 74

4.1 Stylized Admission Process for Most Brazilian Universities…………………... 98

B.1 Stylized Admission Process for Some Brazilian Universities…………………. 113

Chapter 1

Introduction

This dissertation consists of three essays in the economics of education. It investigates the effects of educational programs in primary and secondary schools in Brazil, as well as the effects of racial and low-income quotas in Brazilian universities. Specifically, this dissertation provides estimates the impact of two nationwide primary and secondary school programs on academic outcomes – enrollment, grade promotion, repetition, and dropout rates – and the impact of quotas on the academic performance of students in higher education. The first essay aims to answer the following question: Does inclusive education affect the academic outcomes of students with special educational needs, and the academic outcomes of their peers? This question is addressed in the context of Brazil’s Multifunctional Resources Classroom Inclusion Program. Using school census panel data from 1999 to 2015, and administrative data from the program’s management at Brazil’s Ministry of Education, this essay uses a school fixed effects approach to estimate the impact of this program on academic outcomes. The school level estimates indicate that the Brazilian inclusion program raises the enrollment of disabled and non-disabled students in schools with grades 1-5 and 6-9. A more aggregated examination reveals an increase in overall enrollment due to the program, rather than migration of students across schools with and without the program. Moreover, the results show that the program: 1. Reduces the dropout rates of disabled students in grades 6-9 and 10-12; 2. Reduces the repetition rates of disabled students in grades 6-9; and 3. Raises the promotion rates of disabled students in grades 6-9 and 10-12. The program has also positive spillover effects onto the promotion rates of disabled and non-disabled students in grades 10-12. This study also investigates the program impact by time of adoption, as well as cumulative effects. It finds that the full impact of the program is not complete in the first year of program adoption, which points to the existence of long-term effects of the Brazilian inclusive program, which can be felt, for some outcomes, even after eight years of program.

The second essay investigates the impact of the Mais Educação Extended School Day Program on the academic outcomes of primary and secondary school students in Brazil. It uses data from Brazil’s school census for the period 1999-2014, and administrative data on school program participation for the period 2008-2014. The estimates of the program’s impact are identified by controlling for school fixed effects, state-year and initial enrollment level-year fixed effects, and separate time trends for schools that eventually participate in the program and for schools that never participate. The results suggest that the Brazilian longer school day program is effective at reducing dropout rates for students in all grade levels. The program also raises the enrollment of students in grades 6-9, but reduces the enrollment of students in grades 10-12 in schools where the program is available. Moreover, the estimates indicate that the impact on grade promotion is positive for students in grades 6-9. The impact on grade promotion is negative, however, for students in lower grades. Finally, the program seems to increase repetition rates for students in all grade levels. Estimates of the program impact by time of adoption and its cumulative effects indicate that, for most outcomes and groups of grades, the full impact of the program is not felt in the year of program adoption, but rather is accumulated over time. The third essay evaluates the impact of racial and low-income quotas on the academic performance of senior final year students from Brazilian colleges. It uses data from the Higher Education Census and information on academic performance from the National Examination of Student Performance (ENADE) for the period 2010-2015. While there are some concerns about the adoption of the racial quota in Brazilian universities, the results indicate that the proportion of racial quota students seems to have no statistically significant impact on the academic performance of quota and non-quota students. Likewise, the proportion of low-income quota students does not affect the scores of quota and non-quota students on the ENADE test. Finally, when the analysis is restricted to public colleges and universities, the estimates also suggest that both the proportion of racial quota students and the proportion of low-income quota students have no statistically significant impact on academic performance of quota and non-quota students, regardless of the groups of majors.

Chapter 2

Inclusive Education and Academic Outcomes: The Impact of Brazil’s Multifunctional Resources Classroom Inclusion Program

2.1 Introduction

Inclusive education has been gaining ground in educational policy agendas around the world in recent decades (Soukakou, 2012; Kanter et al., 2014). The inclusion of children with disabilities into regular classrooms has been a means of reducing social and academic exclusion as well as stigma and discrimination. The impact on academic outcomes of disabled and non-disabled students, however, is controversial, with no clear evidence in favor of inclusion programs (Göransson and Nilholm, 2014; Dyson, 2014; Lindsay, 2007; Farrell et al., 2007). Knowledge of the impact of inclusion programs on the academic outcomes of both students with disabilities and their non-disabled peers contributes to the overall debate regarding the adoption of inclusion policies in schools. This paper, therefore, aims to contribute to this discussion by evaluating the impacts of an inclusive education program in a developing country, Brazil. As part of a national effort to include children with disabilities in regular education, the Brazilian government launched in 2007 the Multifunctional Resources Classroom Inclusion Program (Programa Implantação de Salas de Recursos Multifuncionais). It is a nationwide program in which participating schools are provided with specialized pedagogical materials, furniture, and computers, to equip an inclusive classroom, which is used by students with disabilities and special educational needs to improve these students’ learning environment, socialization and the overall academic performance and personal development. Students from participating schools must be enrolled in regular classes for the regular school day and, in a different, after-school session, they can attend the “inclusive classroom.”

Several studies have argued that both students with disabilities and special educational needs and regular students can benefit from inclusive education programs, but the development of an appropriate inclusion policy is still a challenge for education systems (Ainscow and César, 2006; Ainscow, 2005; Farrell, 2000). One of the main concerns regarding the inclusion of students with special educational needs in regular classrooms is that there may be negative effects on the academic achievement of other students. The argument is that students with disabilities may require more attention from teachers at the expense of their classmates which, ultimately, can reduce the effectiveness of the class learning process for the students without special educational needs. The recent literature presents mixed evidence on the impacts of inclusive education on both students with and without special educational needs. Ruijs and Peetsma (2009) provide an extensive review of papers that evaluate the effects of inclusion on the cognitive and socio-emotional development of both students with special educational needs and other students. They find that, overall, the results of those papers suggest positive (or neutral) effects of inclusive education. Using a panel dataset of students in upper secondary education in Norway, Myklebust (2007) finds that students with special educational needs obtained better vocational competence under inclusive education. Hanushek et al. (2002) also find a positive impact of an inclusive education program in Texas on the academic achievement of special-education students, especially those classified as learning-disabled or emotionally disturbed, while not detracting from the performance of regular students. Friesen et al. (2010) use data from the Canadian province of British Columbia to investigate peer effects associated with disabled students in public schools. They find that attending school with a higher percentage of students with learning disabilities or behavioral disorders has a small and statistically insignificant impact on the reading and math test scores of non-disabled students. Using data from the Netherlands, Ruijs (2017), also finds no statistically significant effects of placing students with special educational needs in regular classrooms on the academic achievement of their peers.1

1 For more discussion of the spillover effects of having classmates with disabilities or special educational needs on the achievement of regular students, see Kalambouka et al. (2007), Cole et al. (2004), Demeris et al. (2007), Dyson et al. (2004), and Huber et al. (2001).

In contrast, other empirical research on inclusive education indicates that there may also be negative peer effects of classmates with disabilities or special educational needs onto the academic outcomes of students without disabilities. Gottfried (2014) uses a quasi- experimental method and longitudinal data for the United States and finds that students with a greater number of classmates with disabilities are negatively affected. This result is in line with the negative impacts observed by Fletcher et al. (2010) and Kristoffersen at al. (2015). Despite the number of empirical studies on inclusive education, most of the literature has focused on developed countries. Moreover, in the scant literature on developing countries, no empirical study has evaluated the impact of Brazil’s inclusion program. Therefore, using a school level panel data set, this study aims to evaluate the impact of the Brazilian Multifunctional Resources Classroom Inclusion Program on students’ educational outcomes – specifically, on total enrollment, grade promotion, repetition, and dropping out rates – at the primary and secondary levels. The impact of the program is assessed by using school census data and administrative data from 1999 to 2015. The data cover more than 250,000 schools in each year, with more than 50 million students and over 2 million teachers. In order to identify the impact of the program, this study relies on the assumption that, after controlling for school fixed effects, state-year fixed effects, initial enrollment level-year fixed effects, separate time trends for schools that eventually participate in the program and for schools that never participate, and observable school and student characteristics, the implementation of the program in a given school is unlikely to be correlated with unobserved variables that affect the academic outcomes evaluated. The remainder of this chapter is organized as follows. Section 2 presents a description of the Brazilian Multifunctional Resources Classroom Inclusion Program. Section 3 describes the data and provides descriptive statistics. The empirical framework, along with the estimation and identification strategies, is presented in Section 4. The results are presented and discussed in Section 5, and final conclusions are drawn, and suggestions for future research are made, in Section 6.

2.2 The Brazilian Multifunctional Resources Classroom Inclusion Program

Brazil’s Multifunctional Resources Classroom Inclusion Program, named Programa Implantação de Salas de Recursos Multifuncionais, was launched in 2007. It is a federal (national) program in which participating schools receive pedagogical materials, furniture, computers and specialized resources, to equip the “inclusive classroom” for use by students with mental and physical disabilities or pervasive developmental disorders, as well as students deemed to be super-gifted. 2 The program targets regular public schools that have students with special educational needs or disabilities. These students must be enrolled in regular classes and, in a period outside of normal class time, they participate in programming in the inclusive classroom, which was constructed exclusively for their use. Therefore, participating disabled students are provided with regular classroom instruction, placed together with non-disabled students (mainstreaming), along with specialized instruction, after school, in the inclusive classroom. The pedagogical instruction in the inclusive classroom is framed to meet disabled students’ individual needs, through individual or group activities. Specialized instructors or teachers are responsible for determining what activities will be developed to better serve each student’s needs. All regular schools with special educational needs students have them mainstreamed with other students, which means that the only difference between participating and non-participating schools is that the former implemented the inclusion classroom program. All regular public schools with at least one student with a disability or special educational needs are eligible to participate in the program. The eligibility criteria do not mean, though, that all schools with disabled or special educational needs students participate in the program – there are also schools with special-education students that did not implement the program. Furthermore, in schools where the program was implemented, it is possible that not all disabled or special-education students are treated due to some

2 The presence of super-gifted students is relatively rare in primary and secondary education in Brazil. Although the number of students diagnosed with super-giftedness has been increasing over the past years, they still represent a very low proportion of students. In 2014, for instance, there were 13,308 super-gifted students in the Brazilian school census, which corresponded to 0.03% of all students.

space or resource restrictions; this means that there may be a waitlist to get into the inclusive classroom. Thus, treated schools may have some disabled or special educational needs students not being treated by the program, but all such students are mainstreamed into regular classrooms. The decision on whether a specific school with special-education students will participate in the program is made by the state or local Department of Education. Then, based on the existence of at least one student with a disability or special educational needs enrolled in that school, which is recorded annually in the school census, the decision to adopt the program is made without any consultation with the school principal. Depending on the available funds, the national Ministry of Education, which is responsible for administering the program, can establish a quota for the number of inclusive classrooms for each municipality. Thus, the decision on whether a school will implement the program should be driven by the number of students with special educational needs in each school. Participating schools are provided with the materials and furniture only once, although in 2012, some schools received supplemental material. Figure 2.1 presents the number of schools (and inclusive classrooms) in the Multifunctional Resources Classroom Inclusion Program since 2005. Note that, in response to the number of students with disabilities or special educational needs and the available infrastructure, a few schools have more than one inclusive classroom. Although the program was officially launched in 2007, a few schools already had inclusive classrooms in 2005 and 2006, presumably under a different and smaller program. It is important to highlight that since 2013 no new schools have been added to the program due to lack of funds; nevertheless, the program is still operating for schools that had implemented the inclusive classroom program in any previous year. According to the Ministry of Education, the number of schools that had implemented the program by 2013 corresponded to 48.5% of all regular public schools with disabled and special educational needs students in that year. While the program officially started in 2007, the number of special-education students enrolled in regular public schools has been increasing since 2003. In 2014, the total enrollment of special-education students in regular public schools (with or without

the inclusion program) was 655,375, which is 378% higher than in 2003, as seen in Figure 2.2 (MEC, 2015).

2.3 Data and Descriptive Statistics

This study uses Brazil’s school census data and administrative data from the Brazilian Ministry of Education. Annually, the school census collects data on school, teacher and student characteristics, covering more than 250,000 public and private schools. All Brazilian schools are required to answer the census questions, which include information on academic outcomes such as total enrollment, and grade promotion, dropout and repetition rates.3 The school census is conducted in two phases. The first occurs at the beginning of the academic year and data on total enrollment and school, teacher, and student characteristics are collected. In the second phase, at the end of the academic year, data on promotion, repetition, and dropout rates are assembled to reflect the status of the students enrolled at the beginning of the school year. The census also provides information on school infrastructure (including accessibility for disabled students), student race (since 2005), area of students’ residence, and students’ disability conditions (since 2007). Data on Brazil’s Multifunctional Resources Classroom Inclusion Program, which are not available in the school census, were obtained directly from the program management at the Ministry of Education and comprise all schools that implemented the program in each year. Unfortunately, the data do not indicate how many disabled students participate in the inclusive classroom, so this chapter estimates the impact of the existence of the program, rather than the impact of participation in the program. In order to make the data comparable across schools and time, and to construct a panel of schools, school census data from 1999 to 2015 were used. The year 2006 was excluded from the panel since, due to methodological changes in the school census format,

3 The Brazilian school census does not provide information on students’ academic performance. Thus, the data used in this study cannot identify the impact of the inclusion program on academic performance. To do so is beyond of the scope of this paper, as the data on academic performance come from a different source and they are available for all schools starting only in 2007. Data on math and reading test scores of students in elementary and middle high schools can be obtained from the System of Assessment of Basic Education (Sistema de Avaliação da Educação Básica – SAEB), at the Ministry of Education.

there is no information on most educational outcomes evaluated in this study for that year.4 Columns 2 and 3 of Table 2.1 present the total number of schools and schools with at least one student enrolled in grades 1-5, 6-9, and/or 10-12 for each year. The total number of these schools declined from 1999 to 2007. As pointed out by Glewwe and Kassouf (2012), this reflects a policy of ending the activities of schools with unsatisfactory outcomes and merging small schools into larger ones. In contrast, from 2008 to 2015, the number of schools increased. One of the main reasons for that is the higher level of investments in education after the creation of the Fund for the Development of Basic Education (Fundo de Manutenção e Desenvolvimento da Educação Básica – FUNDEB) in 2007, which requires states and municipalities to invest 20% of their tax revenues into this fund. To analyze the impact of Brazil’s Multifunctional Resources Classroom Inclusion Program, treated schools are defined as those in which the program was implemented and all the remainder are defined as untreated schools. Columns 4, 5, and 6 in Table 2.1 show, respectively, the number of schools with grades 1-5, 6-9, and/or 10-12 with panel data from 1999 to 2015,5 the number of treated schools in each year for the balanced panel of 98,307 schools, and the proportion of the latter in relation to the final number of schools in the panel. To build the panel data set, only schools with regular education for grades 1-5, 6-9, and 10-12 were considered, which reduced the sample by 3-4%. The decision to keep only schools with regular education was driven by the fact that there are no detailed data on school, student and teacher characteristics for the other modalities of education, such as special education and youth and adult education, in the school census. In 2007, only 183 schools participated in the program, which represented less than one percent of schools

4 In 2006, some methodological changes were introduced in the school census format preventing the collection of many education outcomes. Before 2007 schools were used as the basic unit of analysis in the school census. After that year students became the basic unit of analysis and the school census started to collect individual student information, along with teacher, cohort, and school characteristics. Estimates that include 2006 using imputation for the missing data are very similar to those presented in this chapter and are available from the author upon request. 5 In Brazil, schools may offer classes for more than one grade level simultaneously. Considering the data used in this paper, there are 90,761 schools with students from grades 1-5. Among these schools, 52,632 offer both grades 1-5 and 6-9, and 15,844 of these also offer classes for students in grades 10-12. Additionally, from the total of 59,503 schools with grades 6-9, 21,954 also have students in grades 10-12. Finally, out of the 98,307 schools with panel data, 15,526 have grades 1-5, 6-9, and 10-12 simultaneously.

with panel data. This proportion is substantially higher in 2015, with 31.4% of schools with panel data participating in the program. Figures 2.3 to 2.7 present the educational outcomes evaluated in this study for eventually treated and never treated schools with grades 1-5, 6-9, and 10-12, from 1999 to 2015. Figure 2.3 shows the average enrollment of all students for both treated and untreated schools, over the period 1999-2015. Regardless the level of schooling, the average enrollment decreased over time. The decline was larger in schools with grades 6-9, with a reduction of 35.0% in treated schools and of 47.3% in untreated schools. The average enrollment of disabled students in eventually treated and untreated schools from 2007 to 2015 is presented in Figure 2.4. The first year in this figure is 2007 because data on students with disabilities were not collected in the school census until that year. Over the period 2007-2015, the average enrollment of disabled students increased in all groups of grades and in both treated and untreated schools. Schools with grades 10-12 had the largest percentage increase in enrollment of disabled students (453% and 380% in treated and never treated schools, respectively). For schools with grades 1-5 and 6-9 that implemented the program, the increase was from 3.1 to 7.3 (132%) in the former and from 1.3 to 6.0 (359%) in the latter. The increase for never treated schools with grades 1-5 and 6-9, in turn, was 129% (from 0.8 to 1.9) and 260% (from 0.7 to 2.6), respectively. Figure 2.4 also shows that treated schools with grades 1-5 had the highest average enrollment of this kind of student in all years of the study when compared to schools with higher grade levels. Table A.1 of the appendix presents the average total enrollment of disabled and non- disabled students considering all schools and also separately for eventually treated and never treated schools with grades 1-5, 6-9, and 10-12 from 1999 to 2015. The dropout rates for treated and never treated schools are presented in Figure 2.5. Treated schools with grades 1-5 had lower dropout rates than never treated schools before the program’s implementation. However, the difference between the rates of both types of schools has reduced over the years, particularly after 2007, suggesting that the program also increased dropout rates for schools with grades 1-5. Treated schools with grades 6-9 and 10-12, in contrast, had higher dropout rates than never treated schools before 2007. The difference, nevertheless, also decreased after that, suggesting a beneficid effect of the inclusion program.

The repetition rates for treated and untreated schools are shown in Figure 2.6. As can be seen in the figure, the repetition rate for treated schools with grades 1-5 was lower than that for untreated schools before the program’s implementation. After 2007, however, the outcomes for both types of schools got closer, suggesting that the program increased the repetition rate for grades 1-5. Similarly, the program seems to have increased the repetition rate of treated schools with grades 10-12, but not for schools with grades 6-9. Figure 2.7 presents the grade promotion rates for eventually treated and never treated schools from 1999 to 2015. For both types of schools and for all groups of grades, the grade promotion rates are higher in 2015 than in 1999. For schools with grades 1-5, however, the program, which started in 2007, seems to have reduced this outcome. Yet this does not appear to be the case for schools with grades 6-9 and 10-12. Descriptive statistics averaged over the years 1999 to 2015 for all outcomes and explanatory variables are shown separately for treated (schools that eventually implemented the program) and never treated schools with grades 1-5, 6-9, and 10-12 in Table 2.2. For all grade levels, dropout, repetition, and grade promotion rates add up to 100% as enrolled students have only three possible outcomes at the end of the academic year: withdraw from school, fail to progress to the following grade, or advance to the next grade.6 Note that simple comparisons between treated and untreated schools may lead to misleading interpretations. For instance, in schools with grades 10-12, treated schools have lower grade promotion and higher repetition rates than never treated schools, which suggests that the inclusion program reduced promotion and increased repetition, which may not be true since time trends before the program implementation should be taken into account. In addition, for most variables, treated and untreated schools are statistically significantly different, which is controlled for during the estimation process. Table 2.2 also shows descriptive statistics for student characteristics that are available for all years, as well as those that are available only from 2005 to 2015 and only from 2007 to 2015, such as student race, disability conditions, and area of residence. For these variables, which are missing data for some years of the period of analysis (1999-2015), many approaches were

6 As a consequence, the estimated coefficients on the impact of program adoption on dropout, repetition, and grade promotion should add up to zero.

attempted to impute values for the missing data. Although not ideal, the imputation for missing data was necessary to allow the use of relevant student characteristics. Thus, the mean values for the period 2005-2015 (2007-2015) were calculated and assigned to observations from 1999 to 2004 (from 1999 to 2005). Schools that had no information on those variables for the entire period 2005-2015 (2007-2015) were excluded from the regression estimation since it was not possible to calculate the mean values for them. For disability conditions, only categories that were available for all years in the school census were considered.

2.4 Empirical Framework

This section presents the empirical strategy used to estimate the impact of the Brazilian Multifunctional Resources Classroom Inclusion Program on educational outcomes – specifically, on total enrollment, grade promotion, repetition, and dropout rates. In order to identify the effects of the program, this paper relies on the assumption that, after controlling for school fixed effects, state-year fixed effects, initial enrollment level-year fixed effects,7 separate time trends for schools that eventually participate in the program and for schools that never participate, and finally observable school and student characteristics, the adoption of the program in a given school is unlikely to be correlated with unobserved variables that determine the educational outcomes evaluated. This assumption could be violated, though, if the decision on implementing the inclusion program in a specific school is driven by unobserved factors. For instance, if the program is implemented in that school not only due to the enrollment of students with special educational needs or disabilities, but also due to political interests affecting local governments’ decisions, and these same political interests also directly affect the educational outcome of interest, than the estimated impact of the program will be biased.

7 In Brazil, there are 27 states, which implies that the interaction between states and 16 years generates 432 different fixed effects. Because schools had different sizes in 1999, all estimated regressions are also controlled for initial enrollment level-year fixed effects, allowing general trends in the educational outcomes to differ over time for different initial school sizes. All schools are arranged in ten different categories based on their initial enrollment levels, and then these categories are interacted with years, which creates a fixed effect for each interaction between categories and years.

2.4.1 Equations for Estimation

The estimation strategy presented in this section relies on a panel data approach with fixed effects to evaluate the impact of the Multifunctional Resources Classroom Inclusion Program on educational outcomes. The main program impacts one would like to estimate are: (i) the average treatment effect (ATE) – the impact of the inclusion program on all students, participants and non-participants, including those with no special educational needs; and (ii) the average treatment effect on the treated (ATT) – the treatment effect on students who participate in the inclusion program. However, the existing administrative data on the inclusion program do not include information on the proportion of students with special educational needs who actually participate in the program, which implies that the ATT cannot be estimated. Although the proportion of treated students is not available in the school census data, since 2007 the proportion of students with disabilities or special educational needs in each school is provided, which can be used to estimate the effect of program eligibility on academic outcomes. The availability of data on the proportion of eligible students allows the estimation of two kinds of treatment effects: an average spillover effect onto ineligible (and onto eligible)8 students, which I will abbreviate as ASE, and an intent to treat effect for eligible students, which I will abbreviate as ITT. This ITT is slightly different from the standard ITT since, in the case of the Multifunctional Resources Classroom Inclusion Program, schools may be involved in the decision about which eligible students get treated, not just the students and their parents.

Let Yist be an academic outcome (enrollment, grade promotion, repetition, or dropout) for a student (child) i in school s at time t. Suppose that Yist is a function of: student and household variables (Cist) other than variables indicating special needs; school and teacher characteristics (Sst); whether a school has the program at time t (Pst); and whether a student is eligible to participate in the program (!"#$ ). The linear model is given by:

8 In the context of inclusion programs, eligible students may also be affected by the presence of other disabled classmates. The externalities on both types of students may take the form of resource spillovers, changes in the curriculum or pedagogy to accommodate a more diverse classroom, increase in the frequency of disruptions, or reduction of teacher’s attention due to disabled students who require extra help.

( ( ( ( %"#$ = ' )"#$ + ,!"#$ + - .#$ + /0#$ + 1 ()"#$ × 0#$) + 5 (.#$ × 0#$) + 6!"#$0#$ ( ( ( + 7 ()"#$ × !"#$) + 8 (.#$ × !"#$) + 9 ()"#$ × 0#$ × !"#$)

+ :′ (.#$ × 0#$ × !"#$) + <"#$ (2.1)

where <"#$ is an error term with mean zero. Equation (2.1) allows the impact of the program to vary by student and school characteristics. Ideally, one would take the mean of Equation (2.1) at the school level. The problem, however, is that it is not possible to take the mean of )"#$ × 0#$ × !"#$ and

.#$ × 0#$ × !"#$, since there is no school level information on )"#$ and .#$ separately for special educational needs students. That is, for any given school it is not possible to observe

)"#$ × !"#$ and .#$ × !"#$. Therefore, the following equation is estimated, which does not include the last two terms of the above equation:

( ( ( %"#$ = '′)"#$ + ,!"#$ + - .#$ + /0#$ + 1 ()"#$ × 0#$) + 5 (.#$ × 0#$)

+ 6!"#$0#$ + <"#$ (2.2) where / measures the average spillover effect on eligible and ineligible students (ASE) and / + 6 measures the intent to treat effect for eligible students (ITT). Moreover, 1 provides estimates of how ASE varies over )"#$ and 5 measures how it varies over .#$. Equation (2.2) can be aggregated up to the school level as:

( ( ( %=#$ = '′)>#$ + ,!=#$ + - >.#$ + /0#$ + 1 ()>#$ × 0#$) + 5 (.>#$ × 0#$) (2.3)

+ 6!=#$0#$ + <#$̅

When the outcome variable is enrollment, the school level equation is slightly different, since the left-hand side variable is the total enrollment of disabled or non-disabled students, rather than the average enrollment, for school s at time t. Thus, the school level equation for enrollment is:

( ( ( %#$ = '′)>#$ + - .>#$ + @0#$ + 1 ()>#$ × 0#$) + 5 (>.#$ × 0#$) + <#$̅ (2.3’)

where %#$ is the total enrollment of disabled or non-disabled students for school s at time t, and @ is the program impact on eligible students (ITT), if the dependent variable is the enrollment of disabled students, or the spillover effect onto non-disabled students (ASE), if the dependent variable is the enrollment of ineligible students. Both Equation (2.3) and Equation (2.3’) are estimated using the merged school census and administrative data, and it is still the case that / is an estimate of the ASE and / + 6 is an estimate of the ITT in Equation (2.3), and @ is an estimate of the ASE or ITT, according to the dependent variable, in Equation (2.3’). In addition, 1 still measures how ASE varies with student characteristics, and 5 measures how it varies over school characteristics. OLS estimates of Equation (2.3) and Equation (2.3’) will produce unbiased estimates of all parameters in these equations only if the error term <#$̅ is uncorrelated with student characteristics ()>#$), school characteristics (.>#$), the fraction of students who are eligible for the program (!=#$), and the existence of the program at school s at time t (0#$). This assumption, however, is very unlikely to hold, since there may be unobserved school and student characteristics that affect the academic outcomes (%#$). For instance, )>#$ may include student innate ability and parental preferences for schooling, and .>#$ may include principal and teacher motivations. Since these are not observed, they become part of <#$̅ , and because they could be correlated with observed variables in )>#$ and >.#$, the error term

<#$̅ could be correlated with )>#$ and .>#$. Thus, to minimize bias in the estimated impacts of the Brazilian Inclusion Program on educational outcomes, Equation (2.3) and Equation (2.3’) add as controls school fixed effects, state-year fixed effects, initial enrollment level- year fixed effects, and separate time trends for schools that eventually participate in the program and for schools that never participate. After controlling for these fixed effects and time trends, Pst and all the other observed variables are less likely to be correlated with unobserved variables that determine the academic outcomes. Clustered standard errors at the school level were used for all specifications of the above equation. Equation (2.3) and Equation (2.3’) can also assume more flexible forms, including, for instance, not only linear time trends. In the case of the Brazilian Inclusion Program, the

coefficients on ITT and ASE are still identified if quadratic and higher power time trends are used. Moreover, since learning accumulates over time and changes in the number of students in one year may have implications for future academic outcomes, the total enrollment, dropping out, repetition, and grade promotion in any year can also be affected by whether the program operated in previous years. Therefore, it is important to consider that the full impact of the program may not be felt in its first year of implementation. This can be done by including lagged terms, denoted as 0#,$BC, 0#,$BD, etc., in Equation (2.3) and Equation (2.3’). To investigate the dynamics of the program implementation and its impact on academic outcomes in each year after program adoption, lags of the program adoption variable are added to the estimates of Equations (2.3) and (2.3’). Specifically, indicator variables are added for years 0-8 after program adoption. These indicator variables are dummies that equal one only in the relevant year, capturing the impacts of the program in each year after implementation. Finally, these yearly effects can be added up to obtain the cumulative effects. Because the proportion of treated schools increases considerably from 2007 to 2015, the estimates of the impacts by time of program adoption and the cumulative effects are weighted by the proportion of schools that implemented the program in each year.9

2.4.2 Identification Strategy

The identification strategy used in this paper is designed to minimize the three potential sources of statistical endogeneity: (i) omitted variable bias; (ii) reverse causality; and (iii) measurement error. Omitted variable bias refers to the problem of unobserved variables that may be correlated with the implementation of the program and with other observed variables as well. Using time invariant school fixed effects, state-year fixed effects, initial enrollment level-year fixed effects, and separate time trends for schools that eventually

9 The weights were obtained from the ratio of treated schools in each year and the total number of schools in the last year of the panel (2015). In 2012, for instance, 22,460 schools with grades 1-5 implemented the inclusion program, which represented 92.1% of the 24,400 treated schools with grades 1-5 in 2015. Thus, the coefficient on the program variable lagged three years for schools that had the program in 2012 was weighted by 0.921.

participate in the program and for schools that never participate, should minimize bias due to unobserved heterogeneity. Although desirable, it is not possible to control completely for heterogeneity between students; many student characteristics are not available or are difficult to measure, such as student innate ability and motivation. Regarding reverse causality, in which the educational outcomes could also cause program availability, there appears to be little reason to worry about it. As mentioned above, the decision on whether a specific school will participate in the program is made based on the existence of at least one student with special educational needs enrolled in that school. The State or local Department of Education are the only ones responsible for choosing and registering the schools that will implement the program; there is no participation by school principals or parents in this process. Finally, measurement error in the treatment variable is likely to be minimal since this paper uses administrative data for the program implementation variable. Only the administrative data, along with the original files from the Brazilian Ministry of Education, were used to determine the school treatment variable in order to avoid any possible misreported information in the school census. Still, other information in school census may have measurement error, but there is little to do to correct any possible attenuation bias.

2.5 Results

The estimated results of the impact of Brazil’s Multifunctional Resources Classroom Inclusion Program on (log) enrollment, dropout, repetition, and grade promotion rates, are presented in this section. Tables 2.3 to 2.9 report the regression results for estimates of Equations (2.3) and (2.3’), in which each educational outcome is regressed on school, teacher, and student characteristics.10 School fixed effects, state-year fixed effects, initial enrollment level-year fixed effects, as well as separate time trends for schools that eventually participate in the program and for schools that never participate, are included in all of the regressions. Moreover, all estimates use clustered standard errors at the school

10 While one might worry about possible collinearity of the program participation variable and the accessible restroom and accessibility variables, regression estimates excluding the accessibility variables do not change the results. These estimates are available from the author upon request.

level, and the dropout, repetition, and grade promotion variables are measured on a 0-100 scale.

2.5.1 Main Results

Table 2.3 reports the estimates of Equation (2.3’) in which the outcome variable is the (log) enrollment of disabled or non-disabled students in grades 1-5, 6-9, and 10-12. For each grade level, two regression specifications were estimated, the second one including all school and student variables. Because the estimation results are very similar for both specifications, the discussion below will focus on the most complete specification. The results do not show the estimated coefficients on the control variables. These coefficients were omitted in order to economize on space and, more importantly, to focus on the variable of interest: the school program participation variable. The estimates indicate that the program raises enrollment of grades 1-5 disabled students by 5.9%. The impact of school program participation is also positive and statistically significant for disabled students in grades 6-9, implying that the program raises their enrollment by 2.4%. The estimates are statistically insignificant, however, for disabled students in grades 10-12. Thus, the program seems not to affect the enrollment of older disabled students. Table 2.3 also presents the spillover effect of the inclusion program on the enrollment of non-disabled students. The results indicate that the program is effective at raising the enrollment of non-disabled students by 3.1% in schools with grades 1-5 and by 2.1% in schools with grades 6-9. The program has no effect, though, for students with no special educational need in grades 10-12. This positive spillover effect suggests that the presence of disabled students in regular schools should not necessarily raise concerns about special-education placements. Furthermore, a growing awareness of the importance of inclusive education may be leading parents to enroll their children in schools with the program. To investigate whether the positive effects of the inclusion program on the enrollment of disabled and non-disabled students are driven by migration of students across

schools with and without the program rather than by increases in the overall enrollment, the data were aggregated up to the municipio level (Table 2.4). The results provide evidence that the program raises the total enrollment of disabled students in grades 1-5 and 6-9 in municipios with participating schools. More specifically, in municipios with schools that implemented the program, the enrollment of disabled students in grades 1-5 and 6-9 increased by 11.0% and 10.5%, respectively. By contrast, the enrollment of disabled students in grades 10-12 in municipios with participating schools is reduced by 4.6 percentage points. Note that these impacts are for all disabled students, not just students participating in the program, as some eligible students may not be treated in schools with the program. Regarding the negative results for grades 10-12, students’ education at higher levels may often be sacrificed due to many reasons, such as work, marriage, and housework, and a longer school day for disabled students may discourage parents from enrolling their children in schools with the program. The impact of the inclusion program on the enrollment of non-disabled students at the municipio level is presented in the bottom panel of Table 2.4. The results show that the program is effective at raising the enrollment of non-disabled students in grades 1-5 by 3.1%. Thus, these estimates suggest that the program raises the enrollment of non-disabled students in those grade levels by attracting more students who were out of school into participating schools. The estimated effects of the program on the dropout, repetition, and promotion rates are shown in Table 2.5. For each outcome and grade level, two specifications are presented. In the first specification, the interaction term between school program participation and proportion of eligible students (!=#$0#$) is omitted from Equation (2.3). Hence, the coefficient on the school program participation variable measures the impact of school program adoption on the dependent variable for eligible and ineligible students combined. In the second specification, the impact of the program is estimated including the interaction term between the school program participation variable and the proportion of disabled students. Thus, the coefficient on the interaction term indicates the differential effect of the school program participation for disabled students in relation to non-disabled peers in schools that adopted the program; the overall impact for disabled students is obtained by

summing the coefficients on the interaction term and on the school program participation. Lastly, the coefficient on school program participation measures the average spillover effect onto all students, both eligible and ineligible. The regression estimates show that the program is effective at reducing the dropout rate of disabled students by 0.03 percentage points for grades 6-9 and by 0.1 percentage points for grades 10-12. The results also indicate that the program reduces the repetition rate of eligible students in grades 6-9. The impact of the program on grade promotion rate is positive and statistically significant for students in grades 6-9 and 10-12. The program raises the grade promotion rate of all grade 6-9 students (both eligible and ineligible) by 0.2 percentage points. When accounting for differential effects, the estimates indicate that disabled students in grades 6-9 are the only ones benefitting from the program. For students (eligible and ineligible) in grades 10-12, the program raises the grade promotion rate by 0.4 percentage points. The results also show that the inclusion program raises the grade promotion rate of disabled students in grades 10-12 by 0.5 percentage points (0.314 + 0.135 = 0.449). The spillover effect for students in those grade levels is also positive, with the program raising grade promotion by 0.3 percentage points.

2.5.2 Impact by Time of Program Adoption and Cumulative Impact

To explore the dynamics of program implementation and its impact on the academic outcomes of disabled and non-disabled students, as well as the program cumulative impacts, additional estimates including lags of program adoption are presented in Tables 2.6 to 2.9. These estimates include indicator variables for years 0-8 after program adoption. These variables for program adoption are dummies that equal one only in the relevant year, indicating the impact of the program for each year after program adoption. Tables 2.6 to 2.9 also show estimates of the cumulative effects, that are obtained by summing all individual effects for each year.11 Since the proportion of disabled students was not constant over the years, the estimates of the impact of the program on dropout, repetition,

11 Due to lack of space, Table 2.6 presents the cumulative impact of school program participation on the enrollment of disabled and non-disabled students for only eight years of the program.

and grade promotion rates, are focused on the whole set of students, with no distinction between the impact for eligible students and the spillover effect onto ineligible (and onto eligible) students. Table 2.6 presents the estimates of the program impact on (log) of enrollment of disabled and non-disabled students by time of adoption and the cumulative effect after eight years of program implementation. Two different specifications for each group of grades are presented, the second including all schools and student characteristics. In what follows, the discussion will focus on the most complete specification. The disaggregation of the program impact by time of adoption reveals that the inclusion program raises the enrollment of disabled students in grades 1-5 from the year of program adoption to the third year after that. The impact is negative, however, after the fifth year. After eight years of program, the cumulative impact for schools with grades 1-5 is positive, raising the enrollment of disabled students by 13.4%. For grades 6-9, the impact of the inclusion program also accumulates over time. The program raises the enrollment of disabled students from the year of adoption to the eighth year after program implementation, reaching a 31.1% increase after eight years of program. For non-disabled students in grades 1-5 and 6-9, the impact of the program is positive and statistically significant from the year of adoption to the eighth year after program implementation. The results indicate that the cumulative effect is higher for non-disabled students in grades 1-5 and 6-9 than for disabled students. The intuition behind this difference is that the inclusion program has a very large positive spillover effect onto ineligible students, increasing also overall enrollment. Estimates of the impact of the program on dropout rates over time are shown in Table 2.7. The results suggest that the program raises the dropout rates of students in grades 6-9 in the third, fourth, fifth, and sixth years after program adoption. After eight years of program, the estimates indicate that the dropout rate among these students has increased by 0.8 percentage points. Note that this is the impact for all students, with no distinction between the effect for eligible and ineligible students. For students in grades 10-12, the program seems to reduce the dropout rates from the first to the eighth year after program adoption. The cumulative impact suggests a 4.2 percentage point reduction in dropping out after eight years of the program.

Table 2.8 reports the estimates of the program impact on repetition rate in each year after program adoption, as well as the cumulative effects. In the year of adoption, the program is effective at reducing the repetition rate of students in grades 6-9 by 0.8 percentage points. For those students, the effect is higher and statistically significant in the subsequent three years, then it reaches 0.3 percentage points in the sixth year after program adoption. The results suggest that the negative impact for students in grades 6-9 accumulates over time, reaching a 1.8 percentage point reduction in the repetition rate after eight years of the program. In contrast, there is no effect of the program on repetition in grades 1-5 or 10-12. Finally, Table 2.9 displays estimation results for the grade promotion rate. The coefficients on program effects are statistically significant only for students in grades 6-9 and 10-12. For schools with grades 6-9, the program raises the promotion rate by 0.2 percentage points in the year of program adoption, after which the impact reaches 0.4 and 0.2 percentage points in the second and third years after program adoption, respectively. The cumulative effects show that the extent of the dynamics of the grade promotion response to adoption of the inclusion program is determined within five years for grades 6-9. At most, the program increases the grade promotion rate by about 1%. In grades 10- 12, the program raises promotion until the seventh year after implementation. Moreover, the cumulative impact for schools with grades 10-12 is quite large, raising grade promotion in schools that implemented the program by 5.1 percentage points after eight years.

2.5.3 Robustness Checks

In order to check whether the estimated impacts of the program observed in Tables 2.6 to 2.9 are reliable, two robustness checks were conducted. The first is a placebo test that checks for the existence of unobserved changes in schools close to the time of the implementation of the program that affect the educational outcomes but are not fully accounted for by the control variables (Table 2.10). The test was performed by using data from only 1999 to 2005 and creating a placebo variable that equals one in 2005 for the 943 schools, out of 98,307 schools in the sample, that had the program in 2007 and zero

otherwise. If unobserved changes occurred in 2005 and are not captured by the control variables, then regressing the educational outcomes on that placebo variable and the covariates would result in a significant impact of the program placebo variable. Table 2.10 shows that the coefficients on the placebo variable are not statistically significant, which suggests that the program effects observed in Tables 2.6 to 2.9 are due to the program itself. Second, since smaller schools have a higher probability of having the inclusive classroom but no participating students, this could lead to a downward bias in the previously estimated results. Thus, all regressions presented in Tables 2.6 to 2.9 were replicated after excluding the smallest 15% and 30% of schools with grades 1-5 and the smallest 10% and 15% of schools with grades 6-9 and 10-12 (Table 2.11). The choices on the percentiles of the distribution to be excluded for each group of grades was based on the distributions of treated schools. For each group of grades, the number of students in the smallest 5% and 10% of schools that were treated was identified and, afterwards, these numbers were used as cutoffs when deciding the percentiles to be removed from the data including treated and untreated schools. The findings presented in Table 2.11 indicate that this paper’s estimates are robust to these sample restrictions.

2.6 Conclusion

This paper investigates the effects of Brazil’s Multifunctional Resources Classroom Inclusion Program on the enrollment, dropping out, repetition, and grade promotion of disabled and non-disabled students in primary and secondary schools. The program provides schools with specialized pedagogical materials, furniture, and computers to improve the learning environment, socialization and overall academic and personal development of students with disabilities and special educational needs. Based on school level data and fixed effects estimations, this study finds that the inclusion program is effective at raising the enrollment of disabled and non-disabled students in grades 1-5 and 6-9. The municipio level estimates demonstrate that most of these results reflects an increase in the overall enrollment due to the program rather than migration of students across schools with and without the program. The study also finds

that the program reduces the dropout rates of disabled students in grades 6-9 and 10-12, reduces the repetition rates of disabled students in grades 6-9, and raises the promotion rates of disabled students in grades 6-9 and 10-12. Moreover, the program has positive spillover effects for students in grades 10-12, raising their promotion rate by 0.3 percentage points. Further investigation of the program impact by time of adoption and its cumulative effects shows that the full impact of the program is not completely felt in the year of program adoption, but rather accumulates over time. These results point to the existence of a long-term effect of the Brazilian inclusive program, which can last, for some outcomes, up to eight years. The findings of this study suggest that, in general, Brazil’s Multifunctional Resources Classroom Inclusion Program benefits students with special educational needs or disabilities, especially those enrolled in grades 6-9 and 10-12, with no negative spillover effects onto non-disabled students. Thus, the results provide further evidence that inclusive education may generate positive impacts for disabled students with no negative externalities on the academic outcomes of regular students. From a policy perspective, it should be noted that the program’s current design still requires improvements in order to target younger disabled students enrolled in grades 1-5, whose academic outcomes (dropout, repetition, and grade promotion rates) were not affected by the program. Finally, although this study evaluates the impact of the inclusion program on several academic outcomes of disabled and non-disabled students, future research can extend the analysis by assessing also the program’s impact on academic performance.

Table 2.1 – Number of Schools in Brazil’s School Census from 1999 to 2015 Schools with Total number 1st to 5th Schools with % of treated of treated Years Total number and/or 6th to panel data schools (after schools (after of schools 9th and/or 10th (from 1999 to balancing the balancing the to 12th grade current year) panel) panel) classes (1) (2) (3) (4) (5) (6) 1999 266,645 209,280 209,280 0 - 2000 261,988 206,235 177,104 0 - 2001 264,735 201,479 166,886 0 - 2002 256,986 195,465 157,350 0 - 2003 253,405 191,379 150,350 0 - 2004 248,257 188,493 144,845 0 - 2005 248,103 184,513 138,494 183 0.19 2007 237,387 176,614 126,506 943 0.96 2008 250,350 175,985 122,964 4,259 4.33 2009 255,445 173,855 118,345 15,773 16.04 2010 259,831 170,801 114,041 18,210 18.52 2011 263,833 168,570 110,670 27,699 28.18 2012 268,244 167,358 107,941 27,699 28.18 2013 272,049 164,571 104,356 30,835 31.37 2014 276,331 161,906 101,063 30,835 31.37 2015 272,996 160,605 98,307 30,835 31.37 Note: Column 6 is obtained by dividing column 5 by the final number of schools with panel data (98,307).

Table 2.2 – Descriptive Statistics for Eventually Treated and Never Treated Schools Schools with Schools with Schools with

grades 1-5 grades 6-9 grades 10-12 Variables Eventually Never Eventually Never Eventually Never

Treated treated Treated treated Treated treated School variables Total enrollment (units) 235.4 123.7 300.5 256.1 421.8 364.3 (201.5) (157.5) (263.9) (259.5) (411.6) (369.6) Dropout rate (%) 4.7 5.7 7.8 6.7 14.2 8.6 (8.0) (9.7) (9.3) (10.0) (10.7) (10.3) Repetition rate (%) 11.4 12.3 12.7 9.8 10.1 9.2 (9.5) (12.9) (9.5) (9.9) (8.3) (8.5) Grade promotion rate (%) 84.0 82.0 79.5 83.6 75.7 82.3 (13.8) (17.9) (13.3) (14.6) (12.7) (13.7) Geographic region (%) North 11.2 15.2 10.4 10.6 11.3 4.9 (31.5) (35.9) (30.6) (30.8) (31.7) (21.7) Northeast 36.6 48.3 32.8 33.5 24.3 21.7 (48.2) (50.0) (46.9) (47.2) (42.9) (41.2) South 21.1 10.0 27.2 12.3 32.9 10.6 (40.8) (30.0) (44.5) (32.8) (47.0) (30.8) Central-West 8.9 3.5 11.3 5.2 14.2 4.9 (28.5) (18.4) (31.7) (22.2) (34.9) (21.5) Southeast 22.2 23.0 18.3 38.4 17.3 57.9 (41.6) (42.1) (38.7) (48.6) (37.8) (49.4) Rural (%) 33.2 61.1 25.9 29.6 7.9 4.8 (47.1) (48.7) (43.8) (45.7) (26.9) (21.4) Electricity (%) 97.4 85.4 99.2 97.0 99.7 99.5 (16.0) (35.3) (8.9) (17.0) (5.4) (6.8) Water (%) 98.3 95.8 98.9 98.3 99.5 99.5 (12.8) (20.0) (10.3) (12.8) (7.4) (7.2) Sewage (%) 97.7 89.4 98.7 96.6 99.2 99.3 (15.1) (30.7) (11.2) (18.2) (9.1) (8.5) Offers meal (%) 98.7 88.2 98.7 79.6 95.4 66.6 (11.4) (32.3) (11.4) (40.3) (21.0) (47.2) Library (%) 48.2 27.6 66.1 56.8 83.3 73.6 (50.0) (44.7) (47.3) (49.5) (37.3) (44.1) Accessible restroom (%) 21.0 8.5 26.4 17.5 32.6 25.7 (40.7) (27.8) (44.1) (38.0) (46.9) (43.7) Accessibility (%) 17.3 7.1 22.5 14.9 28.0 21.8 (37.8) (25.8) (41.8) (35.6) (44.9) (41.3) Computer lab (%) 39.9 24.2 53.6 53.3 70.5 78.4 (49.0) (42.8) (49.9) (49.9) (45.6) (41.1) Science lab (%) 9.1 8.9 20.5 26.8 44.8 50.5 (28.8) (28.5) (40.4) (44.3) (49.7) (50.0) Computer (units) 7.8 5.6 11.1 13.0 16.4 23.3 (17.8) (19.1) (20.7) (26.6) (21.6) (38.5) Internet (%) 41.9 29.9 51.3 56.2 65.7 80.4 (49.3) (45.8) (50.0) (49.6) (47.5) (39.7) Teacher with college (%) 50.8 38.6 77.9 76.9 89.5 92.8 (38.5) (40.3) (31.4) (33.6) (19.8) (16.4) Student variables Female (%) 46.9 46.6 49.5 49.2 54.2 53.0 (5.9) (8.8) (6.1) (8.2) (6.6) (7.5) Evening class (%) 1.5 1.7 10.1 8.9 46.4 35.4 (6.9) (7.9) (21.4) (22.4) (31.0) (34.7) Skin color (%) White 25.3 21.3 21.5 23.2 25.9 29.6 (25.0) (25.1) (24.2) (25.3) (26.7) (26.3) Black 3.5 3.4 2.9 2.9 2.7 2.7

(6.0) (7.3) (5.5) (5.8) (4.9) (4.6) Pardo 32.2 35.7 26.4 26.7 24.8 21.3 (25.7) (28.6) (24.2) (24.8) (22.8) (21.3) Yellow 0.5 0.5 0.5 0.5 0.8 0.5 (2.3) (2.7) (2.8) (2.5) (3.4) (2.1) Indigenous 0.6 1.5 0.5 1.1 0.6 0.6 (5.5) (10.4) (5.1) (8.7) (5.6) (5.8) Non-declared skin color 38.1 37.6 48.2 45.6 45.2 45.2 (32.3) (32.6) (35.2) (34.0) (35.5) (33.7) Lives in rural area (%) 4.4 7.1 4.0 3.9 2.7 1.3 (19.4) (25.2) (17.9) (18.5) (12.8) (8.8) Disability1 (%) 3.2 1.2 1.6 0.7 0.7 0.4 (7.5) (5.1) (3.8) (2.6) (2.2) (1.1) Vision problems or blind (%) 0.3 0.1 0.3 0.1 0.1 0.1 (1.4) (1.0) (1.5) (1.0) (1.0) (0.5) Hearing problems or deaf (%) 0.2 0.1 0.1 0.0 0.1 0.0 (1.5) (1.0) (1.1) (0.6) (0.7) (0.4) Deaf and blind (%) 0.0 0.0 0.0 0.0 0.0 0.0 (0.1) (0.1) (0.0) (0.0) (0.0) (0.0) Physical disability (%) 0.3 0.2 0.2 0.1 0.1 0.1 (1.2) (0.9) (0.8) (0.6) (0.5) (0.3) Mental disability (%) 2.2 0.8 1.0 0.4 0.3 0.2 (5.8) (4.0) (2.6) (1.8) (1.4) (0.8) Multi disability (%) 0.2 0.1 0.1 0.0 0.0 0.0 (1.4) (0.9) (0.6) (0.5) (0.3) (0.1) Gifted (%) 0.0 0.0 0.0 0.0 0.0 0.0 (0.3) (0.2) (0.6) (0.2) (0.3) (0.2) Observations 419,117 949,803 290,226 443,577 94,708 198,773 Number of schools 26,195 59,363 18,139 27,724 5,919 12,423 Notes: Standard deviations in parentheses. These are averages for 1999 to 2015. Some student variables are averages over fewer years. Skin color variables are available only for 2005-2015, and proportion of students living in rural area and proportion of students with disabilities are available only for 2007-2015. 1The disability variable comprises the following categories: vision problems or blind, hearing problems or deaf, deaf and blind, physical disability, mental disability, and multi disability.

Table 2.3 – Estimates of the Program Impact on Log of Enrollment of Disabled and Non- Disabled Students: Results for Schools with Grades 1-5, 6-9 and 10-12, 1999-2015

Dependent variable: log of enrollment of disabled students Grades 1-5 Grades 6-9 Grades 10-12 Variable (1) (2) (1) (2) (1) (2) School program participation 0.060*** 0.059*** 0.025*** 0.024*** -0.006 -0.007 (0.006) (0.006) (0.007) (0.007) (0.010) (0.010)

Observations 1,354,224 1,136,939 720,981 704,511 288,417 288,407 R-squared 0.832 0.832 0.799 0.798 0.736 0.736

Dependent variable: log of enrollment of non-disabled students Grades 1-5 Grades 6-9 Grades 10-12 Variable (1) (2) (1) (2) (1) (2) School program participation 0.033*** 0.031*** 0.025*** 0.021*** 0.008 0.009 (0.003) (0.003) (0.003) (0.003) (0.006) (0.006)

Observations 1,353,713 1,136,491 720,960 704,500 288,415 288,405 R-squared 0.908 0.904 0.910 0.908 0.900 0.902

School and student characteristics No Yes No Yes No Yes School fixed effects Yes Yes Yes Yes Yes Yes State-year fixed effects Yes Yes Yes Yes Yes Yes Enrollment 1999-year fixed effects Yes Yes Yes Yes Yes Yes Trend x ever program adoption1 Yes Yes Yes Yes Yes Yes Notes: Robust standard errors clustered at the school level. ***Significant at the 1% level. **Significant at the 5% level. *Significant at the 10% level. 1Time trends from the 1st to the 5th power are used.

Table 2.4 – Estimates of the Program Impact on Log of Enrollment of Disabled and Non- Disabled Students at the Municipio-Level: Results for Schools with Grades 1-5, 6-9 and 10-12, 1999-2015

Dependent variable: log of enrollment of disabled students Grades 1-5 Grades 6-9 Grades 10-12 Variable (1) (2) (1) (2) (1) (2) School program participation 0.119*** 0.110*** 0.112*** 0.105*** -0.045** -0.046** (0.039) (0.039) (0.032) (0.032) (0.021) (0.021)

Observations 73,123 72,714 80,386 80,046 79,315 78,315 R-squared 0.872 0.871 0.854 0.854 0.812 0.811

Dependent variable: log of enrollment of non-disabled students Grades 1-5 Grades 6-9 Grades 10-12 Variable (1) (2) (1) (2) (1) (2) School program participation 0.042*** 0.031** 0.016 0.007 0.012 -0.004 (0.016) (0.014) (0.016) (0.009) (0.011) (0.008)

Observations 73,123 72,714 80,385 80,045 79,315 78,315 R-squared 0.975 0.974 0.940 0.968 0.911 0.957

School and student characteristics No Yes No Yes No Yes School fixed effects Yes Yes Yes Yes Yes Yes State-year fixed effects Yes Yes Yes Yes Yes Yes Enrollment 1999-year fixed effects Yes Yes Yes Yes Yes Yes Trend x ever program adoption1 Yes Yes Yes Yes Yes Yes Notes: Robust standard errors clustered at the municipio level. ***Significant at the 1% level. **Significant at the 5% level. *Significant at the 10% level. 1Time trends from the 1st to the 5th power are used.

Table 2.5 – Estimates of the Program Impact on Dropout, Repetition, and Grade Promotion Rates: Results for Schools with Grades 1-5, 6-9 and 10-12, 1999-2015

Dependent variable: dropout rate Grades 1-5 Grades 6-9 Grades 10-12 Variables (1) (2) (1) (2) (1) (2) School program participation -0.027 -0.049 -0.066 -0.037 -0.201* -0.162 (0.034) (0.036) (0.051) (0.052) (0.120) (0.121) School program participation x disability 0.009 -0.028** -0.097** (0.006) (0.011) (0.046)

Observations 1,132,780 1,132,780 702,322 702,322 286,078 286,078 R-squared 0.551 0.551 0.627 0.627 0.652 0.652

Dependent variable: repetition rate Grades 1-5 Grades 6-9 Grades 10-12 Variables (1) (2) (1) (2) (1) (2) School program participation 0.031 0.047 -0.111 -0.021 -0.167 -0.152 (0.054) (0.059) (0.071) (0.073) (0.105) (0.106) School program participation x disability -0.007 -0.090*** -0.038 (0.010) (0.016) (0.038)

Observations 1,132,780 1,132,780 702,322 702,322 286,078 286,078 R-squared 0.571 0.571 0.542 0.542 0.527 0.527

Dependent variable: grade promotion rate Grades 1-5 Grades 6-9 Grades 10-12 Variables (1) (2) (1) (2) (1) (2) School program participation -0.003 0.002 0.177** 0.058 0.368** 0.314** (0.065) (0.070) (0.085) (0.087) (0.145) (0.147) School program participation x disability -0.002 0.118*** 0.135** (0.012) (0.019) (0.059)

Observations 1,132,780 1,132,780 702,322 702,322 286,078 286,078 R-squared 0.686 0.686 0.644 0.644 0.682 0.683

School and student characteristics Yes Yes Yes Yes Yes Yes School fixed effects Yes Yes Yes Yes Yes Yes State-year fixed effects Yes Yes Yes Yes Yes Yes Enrollment 1999-year fixed effects Yes Yes Yes Yes Yes Yes Trend x ever program adoption1 Yes Yes Yes Yes Yes Yes Notes: Robust standard errors clustered at the school level. ***Significant at the 1% level. **Significant at the 5% level. *Significant at the 10% level. The estimated coefficients are the direct effects on the dependent variables, since the latter were multiplied by 100. 1For grades 1-5, time trends from the 1st to the 5th power are used; for grades 6-9, time trends go from the 1st to the 4th power; and for grades 10-12, time trends from the 1st to the 3rd power are used.

Table 2.6 – Estimates of the Program Impact on Log of Enrollment of Disabled and Non- Disabled Students by Time of Adoption and Cumulative Impact

Dependent variable: log of enrollment of disabled students Grades 1-5 Grades 6-9 Grades 10-12 Variable: School program participation (1) (2) (1) (2) (1) (2) Year of program adoption 0.060*** 0.056*** 0.028*** 0.025*** 0.007 0.006 (0.005) (0.006) (0.006) (0.006) (0.009) (0.009) 1st year after program adoption 0.048*** 0.043*** 0.041*** 0.037*** -0.011 -0.011 (0.007) (0.007) (0.007) (0.007) (0.012) (0.012) 2nd year after program adoption 0.051*** 0.042*** 0.054*** 0.044*** -0.022 -0.023 (0.008) (0.008) (0.009) (0.009) (0.015) (0.015) 3th year after program adoption 0.035*** 0.021** 0.056*** 0.043*** -0.026 -0.028* (0.009) (0.009) (0.009) (0.009) (0.016) (0.016) 4th year after program adoption 0.037*** 0.015 0.078*** 0.059*** -0.006 -0.009 (0.009) (0.009) (0.010) (0.010) (0.018) (0.018) 5th year after program adoption -0.002 -0.021*** 0.046*** 0.031*** -0.003 -0.005 (0.007) (0.007) (0.007) (0.007) (0.012) (0.012) 6th year after program adoption 0.007 -0.015** 0.064*** 0.049*** 0.017 0.015 (0.007) (0.007) (0.007) (0.007) (0.011) (0.011) 7th year after program adoption 0.004 -0.004 0.024*** 0.019*** 0.001 0.000 (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) 8th year after program adoption -0.001 -0.003*** 0.005*** 0.004*** 0.000 0.000 (0.001) (0.001) (0.001) (0.001) (0.001) (0.001)

Cumulative impact (8 years of program) School program participation 0.238*** 0.134*** 0.395*** 0.311*** -0.042 -0.055 (0.045) (0.048) (0.049) (0.049) (0.080) (0.080)

Observations 1,354,224 1,136,939 720,981 704,511 288,417 288,407 R-squared 0.832 0.832 0.799 0.798 0.736 0.736

Dependent variable: log of enrollment of non-disabled students Grades 1-5 Grades 6-9 Grades 10-12 Variable: School program participation (1) (2) (1) (2) (1) (2) Year of program adoption 0.033*** 0.032*** 0.033*** 0.027*** 0.011* 0.012* (0.003) (0.003) (0.003) (0.003) (0.006) (0.006) 1st year after program adoption 0.043*** 0.042*** 0.049*** 0.041*** 0.011 0.012 (0.004) (0.004) (0.005) (0.005) (0.009) (0.009) 2nd year after program adoption 0.056*** 0.055*** 0.076*** 0.059*** 0.014 0.013 (0.006) (0.006) (0.007) (0.006) (0.012) (0.012) 3th year after program adoption 0.076*** 0.072*** 0.095*** 0.071*** 0.018 0.016 (0.008) (0.007) (0.008) (0.008) (0.015) (0.015) 4th year after program adoption 0.083*** 0.080*** 0.125*** 0.096*** 0.021 0.016 (0.009) (0.008) (0.010) (0.009) (0.019) (0.018) 5th year after program adoption 0.064*** 0.063*** 0.101*** 0.077*** 0.019 0.016 (0.007) (0.007) (0.008) (0.008) (0.014) (0.014) 6th year after program adoption 0.063*** 0.066*** 0.105*** 0.082*** 0.019 0.018 (0.007) (0.007) (0.008) (0.007) (0.013) (0.013) 7th year after program adoption 0.022*** 0.022*** 0.035*** 0.027*** 0.002 0.001 (0.003) (0.003) (0.003) (0.002) (0.003) (0.003) 8th year after program adoption 0.004*** 0.004*** 0.008*** 0.006*** 0.000 0.000 (0.002) (0.002) (0.001) (0.001) (0.000) (0.000)

Cumulative impact (8 years of program) School program participation 0.446*** 0.436*** 0.627*** 0.486*** 0.115 0.105 (0.044) (0.043) (0.048) (0.047) (0.088) (0.087)

Observations 1,353,713 1,136,491 720,960 704,500 288,415 288,405 R-squared 0.908 0.904 0.910 0.908 0.900 0.902

Table 2.7 – Estimates of the Program Impact on Dropout Rate by Time of Adoption and Cumulative Impact

Program Effects Grades 1-5 Grades 6-9 Grades 10-12 Program adoption 0.001 0.047 -0.341*** (0.034) (0.048) (0.133) 1st year after program adoption -0.017 0.035 -0.317* (0.046) (0.057) (0.170) 2nd year after program adoption -0.038 0.079 -0.579*** (0.061) (0.070) (0.218) 3th year after program adoption -0.008 0.142* -0.680*** (0.074) (0.074) (0.253) 4th year after program adoption -0.004 0.170** -0.840*** (0.083) (0.084) (0.306) 5th year after program adoption 0.021 0.106 -0.737*** (0.067) (0.066) (0.223) 6th year after program adoption 0.027 0.170*** -0.613*** (0.067) (0.063) (0.215) 7th year after program adoption -0.005 0.054** -0.122** (0.023) (0.022) (0.056) 8th year after program adoption -0.004 0.007 -0.015* (0.007) (0.007) (0.008)

Observations 1,132,780 702,322 251,159 R-squared 0.551 0.627 0.659

Cumulative Program Effects Grades 1-5 Grades 6-9 Grades 10-12 Program adoption 0.001 0.047 -0.341*** (0.034) (0.048) (0.133) 1 year of program -0.016 0.082 -0.658** (0.077) (0.098) (0.282) 2 years of program -0.054 0.161 -1.238*** (0.134) (0.160) (0.480) 3 years of program -0.062 0.304 -1.918*** (0.206) (0.228) (0.714) 4 years of program -0.066 0.474 -2.758*** (0.286) (0.306) (0.999) 5 years of program -0.045 0.580 -3.495*** (0.350) (0.366) (1.204) 6 years of program -0.018 0.750* -4.109*** (0.414) (0.423) (1.401) 7 years of program -0.023 0.804* -4.231*** (0.435) (0.441) (1.443) 8 years of program -0.027 0.811* -4.245*** (0.441) (0.445) (1.448)

School and student characteristics Yes Yes Yes School fixed effects Yes Yes Yes State-year fixed effects Yes Yes Yes Enrollment 1999-year fixed effects Yes Yes Yes Trend x ever program adoption 1 Yes Yes Yes Notes: Robust standard errors clustered at the school level. ***Significant at the 1% level. **Significant at the 5% level. *Significant at the 10% level. The estimated coefficients are the direct effects on the dependent variable, since the latter was multiplied by 100. 1For grades 1-5, time trends from the 1st to the 4th power are used; for grades 6-9, linear time trends are used; and for grades 10-12, time trends go from the 1st to the 3rd power.

Table 2.8 – Estimates of the Program Impact on Repetition Rate by Time of Adoption and Cumulative Impact

Program Effects Grades 1-5 Grades 6-9 Grades 10-12 Program adoption 0.002 -0.275*** -0.085 (0.055) (0.067) (0.116) 1st year after program adoption 0.061 -0.414*** -0.238 (0.071) (0.077) (0.151) 2nd year after program adoption 0.030 -0.290*** -0.226 (0.091) (0.091) (0.195) 3th year after program adoption 0.106 -0.314*** -0.180 (0.109) (0.095) (0.222) 4th year after program adoption 0.142 -0.161 -0.202 (0.118) (0.106) (0.273) 5th year after program adoption 0.084 -0.088 0.035 (0.094) (0.082) (0.198) 6th year after program adoption 0.002 -0.270*** 0.056 (0.095) (0.078) (0.194) 7th year after program adoption 0.015 -0.011 -0.048 (0.032) (0.030) (0.050) 8th year after program adoption 0.002 -0.003 0.004 (0.011) (0.010) (0.009)

Observations 1,132,780 702,322 251,159 R-squared 0.551 0.542 0.659

Cumulative Program Effects Grades 1-5 Grades 6-9 Grades 10-12 Program adoption 0.002 -0.275*** -0.085 (0.055) (0.067) (0.116) 1 year of program 0.063 -0.689*** -0.324 (0.117) (0.129) (0.244) 2 years of program 0.093 -0.979*** -0.549 (0.200) (0.205) (0.415) 3 years of program 0.198 -1.293*** -0.729 (0.302) (0.284) (0.615) 4 years of program 0.340 -1.454*** -0.931 (0.413) (0.375) (0.862) 5 years of program 0.424 -1.542*** -0.896 (0.500) (0.444) (1.039) 6 years of program 0.426 -1.812*** -0.840 (0.587) (0.508) (1.211) 7 years of program 0.440 -1.823*** -0.888 (0.615) (0.528) (1.248) 8 years of program 0.442 -1.827*** -0.884 (0.622) (0.532) (1.253)

Table 2.9 – Estimates of the Program Impact on Grade Promotion Rate by Time of Adoption and Cumulative Impact

Program Effects Grades 1-5 Grades 6-9 Grades 10-12 Program adoption -0.003 0.228*** 0.426*** (0.065) (0.079) (0.159) 1st year after program adoption -0.044 0.379*** 0.556*** (0.086) (0.093) (0.206) 2nd year after program adoption 0.008 0.211* 0.805*** (0.112) (0.111) (0.267) 3th year after program adoption -0.097 0.171 0.860*** (0.136) (0.117) (0.310) 4th year after program adoption -0.138 -0.009 1.042*** (0.149) (0.132) (0.378) 5th year after program adoption -0.105 -0.018 0.702** (0.120) (0.104) (0.275) 6th year after program adoption -0.029 0.099 0.558** (0.121) (0.099) (0.267) 7th year after program adoption -0.010 -0.043 0.170** (0.041) (0.037) (0.071) 8th year after program adoption 0.002 -0.003 0.011 (0.013) (0.012) (0.011)

Observations 1,132,780 702,322 251,159 R-squared 0.686 0.644 0.684

Cumulative Program Effects Grades 1-5 Grades 6-9 Grades 10-12 Program adoption -0.003 0.228*** 0.426*** (0.065) (0.079) (0.159) 1 year of program -0.047 0.607*** 0.982*** (0.142) (0.157) (0.341) 2 years of program -0.039 0.818*** 1.787*** (0.246) (0.253) (0.585) 3 years of program -0.136 0.989*** 2.647*** (0.375) (0.356) (0.871) 4 years of program -0.274 0.980** 3.689*** (0.517) (0.473) (1.223) 5 years of program -0.379 0.963* 4.391*** (0.630) (0.563) (1.475) 6 years of program -0.408 1.062 4.949*** (0.743) (0.648) (1.719) 7 years of program -0.418 1.019 5.119*** (0.779) (0.674) (1.773) 8 years of program -0.415 1.015 5.130*** (0.789) (0.679) (1.779)

Table 2.10 – Placebo Test: Estimates of the Program Impact for Schools with Grades 1-5, 6-9, and 10-12, 1999-2005 (Schools with Program in 2007 Assigned to 2005) Log of Log of enrollment of enrollment of Grade Variables Dropout Repetition disabled non-disabled promotion students students Schools with Grades 1-5 School program participation t+1 -0.000 0.002 0.151 0.147 -0.298 (0.009) (0.011) (0.210) (0.256) (0.325)

Observations 597,951 597,782 597,214 597,214 597,214 R-squared 0.985 0.951 0.544 0.578 0.680

Schools with Grades 6-9 School program participation t+1 0.006 0.019 -0.085 0.059 0.026 (0.012) (0.019) (0.273) (0.337) (0.420)

Observations 286,109 286,102 285,477 285,477 285,477 R-squared 0.978 0.936 0.688 0.587 0.705

Schools with Grades 10-12 School program participation t+1 -0.037 -0.005 0.787 -0.038 -0.749 (0.022) (0.042) (0.902) (0.684) (1.081)

Observations 109,442 109,442 107,929 107,929 107,929 R-squared 0.973 0.941 0.696 0.559 0.709

School and student characteristics Yes Yes Yes Yes Yes School fixed effects Yes Yes Yes Yes Yes State-year fixed effects Yes Yes Yes Yes Yes Enrollment 1999-year fixed effects Yes Yes Yes Yes Yes Trend x ever program adoption1 Yes Yes Yes Yes Yes Notes: Robust standard errors clustered at the school level. ***Significant at the 1% level. **Significant at the 5% level. *Significant at the 10% level. The estimated coefficients are the direct effects on the dependent variables, since the latter were multiplied by 100 (except log of enrollment). 1For grades 1-5, time trends from the 1st to the 5th power are used; for grades 6-9, time trends go from the 1st to the 4th power (except for the equations of enrollment of disabled and non-disabled students, which used time trends from the 1st to the 5th power); and for grades 10-12, time trends from the 1st to the 3rd power are used (except for the equations of enrollment of disabled and non-disabled students, which used time trends from the 1st to the 5th power).

Table 2.11 – Robustness Check: Estimates of the Program Impact after Excluding the Smallest Schools with Grades 1-5, 6-9, and 10- 12, 1999-2015 Log of enrollment of Log of enrollment of Dropout Repetition Promotion Variables disabled students non-disabled students (1) (2) (1) (2) (1) (2) (1) (2) (1) (2) Schools with Grades 1-5 School program participation 0.064*** 0.068*** 0.039*** 0.039*** -0.007 -0.003 0.088 0.107* -0.081 -0.105 (0.006) (0.007) (0.003) (0.003) (0.036) (0.036) (0.056) (0.056) (0.067) (0.068) School program participation x disability 0.004 0.003 -0.012 -0.005 0.008 0.002 (0.006) (0.007) (0.008) (0.008) (0.011) (0.011)

Observations 1,047,593 930,166 1,047,355 929,985 1,044,867 927,878 1,044,867 927,878 1,044,867 927,878 R-squared 0.828 0.822 0.905 0.901 0.579 0.594 0.602 0.623 0.708 0.720

Schools with Grades 6-9 School program participation 0.029*** 0.032*** 0.018*** 0.017*** -0.026 -0.012 -0.034 -0.024 0.060 0.036 (0.007) (0.007) (0.003) (0.003) (0.050) (0.051) (0.073) (0.075) (0.086) (0.088) School program participation x disability -0.039*** -0.047*** -0.090*** -0.089*** 0.130*** 0.136*** (0.013) (0.013) (0.016) (0.017) (0.020) (0.021)

Observations 646,327 602,772 646,322 602,767 644,698 601,344 644,698 601,344 644,698 601,344 R-squared 0.795 0.794 0.904 0.898 0.660 0.674 0.568 0.579 0.672 0.684

Schools with Grades 10-12 School program participation -0.005 -0.007 0.008 0.008 -0.183 -0.190 -0.131 -0.132 0.314** 0.322** (0.011) (0.011) (0.005) (0.005) (0.121) (0.123) (0.108) (0.111) (0.147) (0.149) School program participation x disability -0.127** -0.115** -0.035 -0.048 0.162** 0.163** (0.051) (0.055) (0.045) (0.049) (0.066) (0.072)

Observations 266,574 246,699 266,574 246,699 264,556 244,900 264,556 244,900 264,556 244,900 R-squared 0.733 0.733 0.900 0.892 0.657 0.655 0.537 0.540 0.686 0.682

School and student characteristics Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes School fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes State-year fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Enrollment 1999-year fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Trend x ever program adoption1 Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Notes: Robust standard errors clustered at the school level. ***Significant at the 1% level. **Significant at the 5% level. *Significant at the 10% level. The estimated coefficients are the direct effects on the dependent variables, since the latter were multiplied by 100 (except log of enrollment). Grades 1-5: (1) Estimates after excluding the smallest 15% of schools; (2) Estimates after excluding the smallest 30% of schools. Grades 6- 9 and 10-12: (1) Estimates after excluding the smallest 10% of schools; (2) Estimates after excluding the smallest 15% of schools. 1For grades 1-5, time trends from the 1st to the 5th power are used; for grades 6-9, time trends go from the 1st to the 4th power (except for the equations of enrollment of disabled and non-disabled students, which used time trends from the 1st to the 5th power); and for grades 10-12, time trends from the 1st to the 3rd power are used (except for the equations of enrollment of disabled and non-disabled students, which used time trends from the 1st to the 5th power).

50,000 45,000 40,000 35,000 30,000 25,000 20,000 15,000 10,000 5,000 0 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015

Number of schools Number of inclusive classrooms

Figure 2.1 – Number of Schools (and Inclusive Classrooms) in the Multifunctional Resources Classroom Inclusion Program, 2005-2015

800,000

700,000

600,000

500,000

400,000

300,000

200,000

100,000

0 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014

Total Enrollment in Regular Public Schools Enrollment in Special Public Schools

Figure 2.2 – Number of Students with Special Educational Needs Enrolled in Regular Public Schools and in Special Public Schools, 2003-2014

Schools with Grades 1-5

300.0

250.0

200.0

150.0

100.0

50.0

0.0 1999 2000 2001 2002 2003 2004 2005 2007 2008 2009 2010 2011 2012 2013 2014 2015

Treated Never treated

Schools with Grades 6-9

400.0

350.0

300.0

250.0

200.0

150.0

100.0

50.0

0.0 1999 2000 2001 2002 2003 2004 2005 2007 2008 2009 2010 2011 2012 2013 2014 2015

Treated Never treated

Schools with Grades 10-12

600.0

500.0

400.0

300.0

200.0

100.0

0.0 1999 2000 2001 2002 2003 2004 2005 2007 2008 2009 2010 2011 2012 2013 2014 2015

Treated Never treated

Figure 2.3 – Average Enrollment of All Students in Treated and Never Treated Schools with Grades 1-5, 6-9, and 10-12, 1999-2015

Schools with Grades 1-5

8.0

7.0

6.0

5.0

4.0

3.0

2.0

1.0

0.0 2007 2008 2009 2010 2011 2012 2013 2014 2015

Treated Never Treated

Schools with Grades 6-9

7.0

6.0

5.0

4.0

3.0

2.0

1.0

0.0 2007 2008 2009 2010 2011 2012 2013 2014 2015

Treated Never Treated

Schools with Grades 10-12

4.0

3.5

3.0

2.5

2.0

1.5

1.0

0.5

0.0 2007 2008 2009 2010 2011 2012 2013 2014 2015

Treated Never Treated

Figure 2.4 – Average Enrollment of Disabled Students in Treated and Never Treated Schools with Grades 1-5, 6-9, and 10-12, 2007-2015

Schools with Grades 1-5

14.0

12.0

10.0

8.0

6.0

4.0

2.0

0.0 1999 2000 2001 2002 2003 2004 2005 2007 2008 2009 2010 2011 2012 2013 2014 2015

Treated Never treated

Schools with Grades 6-9

16.0

14.0

12.0

10.0

8.0

6.0

4.0

2.0

0.0 1999 2000 2001 2002 2003 2004 2005 2007 2008 2009 2010 2011 2012 2013 2014 2015

Treated Never treated

Schools with Grades 10-12

25.0

20.0

15.0

10.0

5.0

0.0 1999 2000 2001 2002 2003 2004 2005 2007 2008 2009 2010 2011 2012 2013 2014 2015

Treated Never treated

Figure 2.5 – Dropout Rates for Treated and Never Treated Schools with Grades 1-5, 6-9, and 10-12, 1999-2015

Schools with Grades 1-5

18.0 16.0 14.0 12.0 10.0 8.0 6.0 4.0 2.0 0.0 1999 2000 2001 2002 2003 2004 2005 2007 2008 2009 2010 2011 2012 2013 2014 2015

Treated Never treated

Schools with Grades 6-9

16.0

14.0

12.0

10.0

8.0

6.0

4.0

2.0

0.0 1999 2000 2001 2002 2003 2004 2005 2007 2008 2009 2010 2011 2012 2013 2014 2015

Treated Never treated

Schools with Grades 10-12

14.0

12.0

10.0

8.0

6.0

4.0

2.0

0.0 1999 2000 2001 2002 2003 2004 2005 2007 2008 2009 2010 2011 2012 2013 2014 2015

Treated Never treated

Figure 2.6 – Repetition Rates for Treated and Never Treated Schools with Grades 1-5, 6- 9, and 10-12, 1999-2015

Schools with Grades 1-5

100.0 90.0 80.0 70.0 60.0 50.0 40.0 30.0 20.0 10.0 0.0 1999 2000 2001 2002 2003 2004 2005 2007 2008 2009 2010 2011 2012 2013 2014 2015

Treated Never treated

Schools with Grades 6-9

88.0 86.0 84.0 82.0 80.0 78.0 76.0 74.0 72.0 70.0 68.0 1999 2000 2001 2002 2003 2004 2005 2007 2008 2009 2010 2011 2012 2013 2014 2015

Treated Never treated

Schools with Grades 10-12

90.0

85.0

80.0

75.0

70.0

65.0

60.0 1999 2000 2001 2002 2003 2004 2005 2007 2008 2009 2010 2011 2012 2013 2014 2015

Treated Never treated

Figure 2.7 – Grade Promotion Rates for Treated and Never Treated Schools with Grades 1-5, 6-9, and 10-12, 1999-2015

Chapter 3

Impact of the Mais Educação Extended School Day Program on Academic Outcomes in Brazil

3.1 Introduction

The extension of the school day has been widely discussed as a mechanism to improve the quality of education (Holland et al., 2015; Bellei, 2009; Link and Mulligan, 1986; Llach et al., 2009). Increasing the time students spend in school is an important decision for policy makers and school administrators, especially in face of the argument that a longer school day could produce better academic, social and labor market outcomes. While there is a consolidated literature that provides evidence about the impact of school inputs on academic outcomes, researchers have a limited understanding of the effects of extending the school day, especially in developing countries (Hincapie, 2016; Patall et al., 2010; Cerdan-Infantes and Vermeersch, 2007). Therefore, this study attempts to fill this void by evaluating the Brazilian Mais Educação Extended School Day Program. Despite being an old practice in many Latin America countries, the extension of the school day is relatively new in Brazil. The Mais Educação (“More Education”) Extended School Day Program was implemented in public primary and secondary Brazilian schools in 2008, and participating institutions lengthened the school day from a half day (4-5 hours) to a full day (at least seven hours). The policy is part of a national effort to reduce educational disparities and improve learning outcomes, especially for disadvantaged students. Advocates of lengthening the school day argue that spending more time in school allows teachers to better deliver the curriculum and dedicate more time supporting students with low academic performance or who are struggling in the learning process. Moreover, students will likely spend more time in academic activities and less time with no supervision, reducing their exposure to potential risks, for instance, violence, crime, and

substance abuse (Jacob and Lefgren, 2003; Lochner and Moretti, 2004; Berthelon and Kruger, 2011). Finally, with longer school days, schools may offer extracurricular activities, such as music, sports, and digital culture, increasing the interest and demand for school, and consequently reducing dropout rates (Pires and Urzua, 2015). On the other hand, the extension of the school day may not improve educational outcomes. For instance, it may be that the instruction during the additional time is of poor quality. Teachers may be required to teach extra hours with little or no additional pay and consequently may reduce the amount of effort they put into providing better classroom instruction. Furthermore, if less prepared teachers are hired to teach during the additional hours of school time, spending more hours in school may not directly translate into more learning. Finally, if the extra hours in school are used mainly for non-academic activities, the impact on educational outcomes could be neutral or even negative. For instance, students may dedicate more out of school time to non-academic activities, to the detriment of time spent on homework. Therefore, if not well implemented, longer school days may have little effect on academic outcomes, and could even lead to reduced promotion rates and/or increased repetition and dropout rates. The recent literature provides mixed evidence regarding the impact of extended school day programs. Using a difference-in-difference approach, Bellei (2009) estimates the effect of a Chilean extended school day program on high school students academic performance. He finds that the program had a positive effect on academic achievement in both mathematics and language. The author highlights that the program had larger positive impacts on rural students and students enrolled in public schools. More evidence on the Chilean context is found by Pires and Urzua (2015) and Valenzuela (2005), both of whom find positive impacts of extending the school day on academic achievement. Hincapie (2016) analyzes the effect of longer school days on the achievement of students in 5th and 9th grades in Colombia. The paper uses school fixed effects models and finds positive impacts of the policy on math and language scores, especially for students in grade nine. The effects of full-day kindergarten on academic achievement in the United States were evaluated by Cooper et al. (2010) and Rathbun (2010). Both studies find that the implementation of full-day kindergarten (as opposed to half-day kindergarten) has a positive effect on academic performance.

In the opposite direction, a few studies on the impact of lengthening the school day have found no effect or even negative effects on academic outcomes. Arzola (2010) evaluated the program Jornada Escolar Completa in Chile and found no statistically significant effect of the program on students’ academic performance on math and language tests. Using propensity score matching and panel data, Aquino (2011) estimates the impact of an extended school day program in Brazil that was implemented only in public institutions in the State of Sao Paulo. The author finds no statistically significant effect of the program on math test scores and promotion rates, but there is a small and statistically significant impact on language test scores. Xerxenevsky (2012) also uses propensity score matching to estimate the effect of the Mais Educação Extended School Day Program on academic performance, restricting the analysis only to the State of Rio Grande do Sul. The author finds no significant effect for 8th grade students, but finds a statistically significant negative effect on the math test scores along with a positive impact on the language test scores of 4th grade students. The Brazilian program was also evaluated by Itau Social Foundation (Fundação Itau Social, 2015). The study provides an extensive review of the program and investigates the impact on dropout rates and academic performance using propensity score matching and a difference-in-difference approach. The study finds no statistically significant impact on academic performance in language test scores and dropout rates and a negative, statistically significant, effect on academic performance in math test scores. The empirical studies that have evaluated Brazil’s Mais Educação Extended School Day Program estimated the impact of the program for small selected samples and their analyses are limited to a short time interval, which does not allow them to capture cumulative effects of the program. Therefore, due to the lack of evidence about the impact of extending the length of the school day, especially in developing countries, and the limited research on the Brazilian program, this study aims to contribute to the literature by investigating the causal relationship between a longer school day and academic outcomes in Brazilian schools, using an identification strategy that mitigates potential endogeneity problems found in the literature. Moreover, most studies focus on the effect of extended school day programs on academic performance in language and math test scores, rather than other academic outcomes. Thus, this study contributes to the existing literature by

investigating the impact of Brazil’s Mais Educação Extended School Day Program on the enrollment, dropout, repetition, and grade promotion rates of students in grades 1-5, 6-9, and 10-12 in Brazilian schools, using school census data and administrative data from 1999 to 2014. The rest of this chapter is organized as follows. Section 2 provides the description of the Mais Educação Extended School Day Program. In Section 3, the data and descriptive statistics are presented. Section 4 provides the empirical framework, along with the estimation and identification strategies. The results are presented and discussed in Section 5, and Section 6 concludes and makes suggestions for future research.

3.2 Description of the Mais Educação Extended School Day Program

The Mais Educação Extended School Day Program was launched in 2007, and was implemented in 2008 in Brazilian public schools. It is a federally funded nationwide program in which participating schools extend the school day to at least seven hours, as opposed to 4-5 hours, through optional extra academic and non-academic activities. Schools receive financial resources, including payments for extra teaching time, to implement the program on a year-by-year basis, and every year they must indicate six activities they will include in the following areas: academic tutoring; environmental education; sports; human rights education; arts education; digital culture; health promotion; communication and media; natural sciences education; and economic education. All public schools with at least 100 enrolled students are eligible to participate in the program. Due to financial resource restrictions, however, priority is given to schools with a low Development Index of Basic Education (Índice de Desenvolvimento da Educação Básica – IDEB), which measures the quality of education based on students’ performance on the System of Assessment of Basic Education (Sistema de Avaliação da Educação Básica – SAEB). Schools indicate whether they want to participate in the program and the Ministry of Education makes the decision on which schools will implement the program. In schools with Mais Educação, it is possible that not all students are treated. Schools may decide to exclude some students, and students and their parents

may decide not to participate. When this decision is made by schools, they must indicate which students will participate in the activities of the extended school day, and those with low socio-economic background must be prioritized. Table 3.1 shows the total number of participating schools in each year from 2008 to 2014, the corresponding proportion relative to the total number of all schools, and the proportion relative to the total number of public schools. In 2008, less than 1% of all Brazilian schools participated in the program. This proportion was significantly higher in 2014, by which time 18.4% of all schools had implemented the extended school day program. If only public institutions are considered, the proportion of schools that implemented the program in 2014 increases to 23.7%.

3.3 Data and Descriptive Statistics

In order to estimate the impact of the Mais Educação Extended School Day Program on academic outcomes, this study uses Brazil’s school census and administrative data from 1999 to 2014. Each year, the Brazilian school census collects data on school, teacher, and student characteristics, along with education outcomes, such as total enrollment, promotion, repetition, and dropping out rates. The census is administered to over 250,000 public and private schools, from preschools to high schools, as well as vocational education and adult education schools. The school census is conducted in two steps. In the first, at the beginning of the academic year, data on total enrollment and on school, teacher, and student characteristics are collected. In the second step, at the end of the academic year, data on promotion, repetition, and dropout rates are assembled to reflect the status of the students who were enrolled at the beginning of the academic year. Data on Brazil’s Mais Educação Extended School Day Program, which are not available in the school census, were obtained directly from program management at the Ministry of Education and comprise all schools that implemented the program in each year since 2008. The available data do not indicate how many students participate in the program, which implies that this study can estimate only the impact of program adoption, rather than the impact of participation in the program.

To create a panel of schools, school census data from 1999 to 2014 were used. The year of 2006 was excluded due to missing data on most outcomes evaluated in this paper.12 Columns 1 and 2 of Table 3.2 show the total number of schools in Brazil from 1999 to 2014 and schools with at least one student enrolled in grades 1-5, 6-9, and/or 10-12.13 The total number of schools decreased from 1999 to 2007, reflecting the policy of merging small schools into larger ones and closing schools with unsatisfactory outcomes. From 2008 to 2015, in contrast, the number of schools increased. One of the reasons was the higher level of investments in education after 2007, when the Fund for the Development of Basic Education (Fundo de Manutenção e Desenvolvimento da Educação Básica – FUNDEB) was created. This Fund requires an investment of 20% of tax revenues from states and municipalities. Column 4 of Table 3.2 presents the total number of schools for which there are panel data. Therefore, in 2014 there are 239,975 schools with grades 1-5, 6-9 and/or 10- 12, of which 101,063 have data for all years from 1999 to 2014. The panel of schools was built using data only on schools with regular education for grades 1-5, 6-9, and/or 10-12, which reduced the sample by less than 4%. The reason for this restriction is that there is no detailed information in the school census on school, teacher, and student characteristics for the other modalities of education, such as youth and adult education and special education. Lastly, columns 5 and 6 show, respectively, the number of treated schools in each year after balancing the panel and the proportion of treated schools in relation to the final number of 101,063 schools in the panel. To evaluate the impact of the Mais Educação Extended School Day Program, only participating schools that implemented the program in any year and continued to participate

12 In 2006, some methodological changes were introduced in the school census format, preventing the collection of many education outcomes. Before 2007 schools were used as the basic unit of analysis in the school census. After that year, students became the basic unit of analysis and the school census started to collect individual student information, along with teacher, cohort, and school characteristics. Estimates that include 2006 using imputation for the missing data are very similar to those presented in this paper and are available from the author upon request. 13 In Brazil, schools may offer classes for more than one group of grades simultaneously. Considering the data used in this paper, there are 93,444 schools with students from grades 1-5. Among these schools, 52,797 offer both groups of grades 1-5 and 6-9, and 15,821 of these also offer classes for students in grades 10-12. Additionally, from the total of 59,712 schools with grades 6-9, 21,900 also have students in grades 10-12. Finally, of the 101,063 schools with panel data, 15,503 have simultaneously grades 1-5, 6-9, and 10-12.

over time were considered as treated schools, which reduces the final panel data to 90,721 schools. The exclusion of schools that left the program after adopting it avoids the issue of considering them in the control group even after belonging to the treatment group in one or more previous years. Table 3.3 shows the total number of schools and schools participating in the program for the period 2008 to 2014, after balancing the panel. In 2014, 54.5% of schools with grades 6-9 participated in the program. The proportion is lower for schools with grades 1-5 and 10-12: 41.2% and 30%, respectively. Descriptive statistics for all outcomes and school and student variables included in the analysis are also presented, separately for eventually treated and never treated schools with grades 1-5, 6-9, and 10-12, in Table 3.3. The first four lines contain summary data on education outcomes, including grade promotion, repetition, and dropout rates. Because enrolled students have only three possible outcomes at the end of the academic year – advance to the next grade, fail to progress to the following grade, or withdraw from school – grade promotion, repetition, and dropout rates add up to 100%. As a result, the estimated coefficients on the impact of program adoption on these three outcomes will add up to zero. It is important to highlight that comparisons between treated and untreated schools may lead to misleading interpretations if the existence of time trends before the program implementation is not taken into account. For instance, one may infer from Table 3.3 that the program increases dropout rates in all grade levels, yet this may not be true. Moreover, most school and student characteristics of the treated and untreated schools are statistically significantly different, so these characteristics are controlled for during the estimation process. Some student variables, such as race and area of residence, are available only from 2005 to 2014 or from 2007 to 2014. To address the missing data problem, the mean values for the period 2005-2014 (2007-2014) were calculated and assigned to observations from 1999 to 2004 (1999 to 2005). Although not ideal, this imputation process was necessary to allow the use of relevant student characteristics. Figures 3.1 to 3.4 show the educational outcomes evaluated in this study for treated and never treated schools with grades 1-5, 6-9, and 10-12. The total enrollment for eventually treated and never treated schools over the period 1999 to 2014 is presented in Figure 3.1. Regardless of the level of schooling, the total enrollment decreased over time. The decline was larger in schools with grades 6-9, with reductions of 38.4% and 41.4%

between 1999 and 2014 in treated and never treated schools, respectively. The total enrollment in treated schools with grades 1-5 and 6-9 is higher than in untreated schools (except in 1999 for schools with grades 6-9). In contrast, treated schools with grades 10- 12 have lower total enrollment than never treated schools. Overall, Figure 3.1 shows no differential changes in trends after the program started. Figure 3.2 shows the dropout rates for eventually treated and never treated schools. For both types of schools and all grade levels, the dropout rates decreased over time. Schools with grades 1-5 had the largest decline: 10.3 percentage points in treated schools and 8.5 percentage points in untreated schools. As can be seen in the figure, the dropout rates for treated schools are higher than for never treated schools in all grade levels and over the whole period of the analysis. After 2008, however, the outcomes for both types of schools with grades 10-12 got closer, suggesting that the program reduced dropout rate for those grade levels. The repetition rates for treated and never treated schools are presented in Figure 3.3. Over the period 1999 to 2014, the repetition rates increased in both treated and never treated schools with grades 6-9 and 10-12. The increase was larger in the schools with grades 10-12: 5.2 percentage points in treated schools and 3.2 percentage points in never treated schools. A reduction in repetition rates is observed, however, in schools with grades 1-5. The outcome decreased by 6.6 percentage points and by 7.4 percentage points in eventually treated and never treated schools, respectively. Figure 3.3 also reveals that treated schools with grades 1-5 had higher repetition rates than never treated schools, especially after the program’s implementation, suggesting that the program increased repetition. Finally, Figure 3.4 presents the grade promotion rates for eventually treated and never treated schools from 1999 to 2014. Treated schools in all grade levels have lower grade promotion rates than never treated ones. For both types of schools and all groups of grades, however, the grade promotion rates are higher in 2014 than in 1999. Schools with grades 1-5 had the largest increase in this outcome over the period 1999 to 2014, rising 16.9 percentage points in treated schools and 15.9 percentage points in never treated schools. These trends show no clear impact of the program for any grade level.

3.4 Empirical Strategy

This section explains how the impact of the Mais Educação Extended School Day Program on academic outcomes is estimated. In order to identify the effects of the program, this study relies on the assumption that, after controlling for: i. school fixed effects; ii. state- year fixed effects; iii. initial enrollment level-year fixed effects;14 iv. separate time trends for schools that eventually participate in the program and for schools that never participate; and v. observable school and student characteristics; the adoption of the program in a given school is unlikely to be correlated with unobserved variables that determine the academic outcomes evaluated.

3.4.1 Estimation and Identification Strategies

The estimation strategy presented in this section relies on a panel data approach with fixed effects to evaluate the impact of the Mais Educação Extended School Day Program on 15 educational outcomes. Let Yist be an educational outcome (total enrollment, grade promotion, repetition, or dropout) for a student (child) i in school s at time t. Suppose that

Yist is a function of student and household variables (Cist), school and teacher characteristics

(Sst), and whether a school is treated by the program at time t (Pst). The linear model at the individual level is given by:

!"#$ = & + (′*"#$ + +′,#$ + -.#$ + /′(*"#$ × .#$) + 4′(,#$ × .#$) + 5"#$ (3.1)

14 Because Brazil has 27 states, all regressions controll for 405 state-year fixed effects generated by the interaction between states and 15 years of data. To account for different levels of initial school size and to allow the program impact to be different over time due to changes in unobserved student and school characteristics that influence the impact, all regression estimates also controll for initial enrollment level-year fixed effects. All schools are classified into ten different categories according to their initial enrollment levels, and then these categories are interacted with years, which implies that each interaction between categories and years gets a fixed effect. 15 The estimation framework used in this study follows that presented in Glewwe and Kassouf (2012). The authors developed a detailed empirical model to estimate the impact of Brazil’s Bolsa Familia program on Brazilian children’s education outcomes.

where & is a constant and 5"#$ is an error term with mean zero. Equation (3.1) allows the impact of the program to vary by student and school characteristics. Most importantly, if one of the variables in *"#$ is program participation, then the coefficient / on the interaction term corresponding to that variable gives the average impact of program participation (ATT), and - measures spillover effects onto non- participants and onto participants in schools with the program. However, if no variable in

*"#$ is program participation, then - measures the intent to treat effect (ITT). The interaction terms *"#$ × .#$, in turn, still capture differential impacts according to student characteristics and ,#$ × .#$ capture heterogeneous effects according to school characteristics. The data provided by Brazil’s school census do not allow the construction of a panel data set at the student level. Therefore, the impact of the program is estimated by aggregating Equation (3.1) up to the school level:

!#$ = & + (′*#$ + +′,#$ + -.#$ + /′(*#$ × .#$) + 4′(,#$ × .#$) + 5#$ (3.2)

:;< :;< :;< where !#$ = (1⁄7#$) ∑"=> !"#$; *#$ = (1⁄7#$) ∑"=> *?@A; 5#$ = (1⁄7#$) ∑"=> 5"#$; and 7#$ is the number of students in school s at time t. In this aggregated school level estimation, redefining the Cst and Sst variables as deviations from their means implies that - measures the intent to treat effect (ITT) – as it measures the impact of the availability of the program, since there are no available data on students’ program participation – and / and 4 estimate how different this effect is for the observed student and school variables, respectively. Note that, when the outcome variable is enrollment, the school level equation is slightly different, since the left-hand side variable is total enrollment for school s at time t rather than the average enrollment. Therefore, when estimating the impact of the program on

:;< :;< enrollment, the only difference is that !#$ = ∑"=> !"#$ rather than !#$ = (1⁄7#$) ∑"=> !"#$. The school census provides some student and school characteristics, but many other student and school variables are not provided. Thus, there may be unobserved variables that influence both academic outcomes (!#$) and the program availability in a given school

(.#$), such as student innate ability, parental preferences for schooling, and principal and teacher motivations, and thus are related to the error term. To see the implications of having ∗ ∗ unobserved variables for estimation, let *#$ and ,#$ be unobserved student and school variables, respectively, while *#$ and ,#$ now indicate only observed student and school variables, respectively. Equation (3.2) can be modified to:

C ∗ ∗ C ∗ ∗ ∗ ∗ !#$ = & + ( *#$ + ( ′*#$ + + ,#$ + + ′,#$ + -.#$ + /′(*#$ × .#$) + / ′(*#$ × .#$) + ∗ ∗ 4′(,#$ × .#$) + 4 ′(,#$ × .#$) + 5#$ C C C C ∗ ∗ = & + ( *#$ + + ,#$ + -.#$ + / (*#$ × .#$) + 4 (,#$ × .#$) + [( ′*#$ + ∗ ∗ ∗ ∗ ∗ ∗ + ′,#$ + / ′(*#$ × .#$) + 4 ′(,#$ × .#$) + 5#$] (3.2’)

In Equation (3.2’) all unobserved variables, together with 5#$, are the overall error term. To obtain consistent estimates of Equation (3.2’) using OLS, one would have to assume that the terms in the brackets are uncorrelated with all observed student and school variables. This assumption, however, is very doubtful. Therefore, to control for unobserved characteristics that affect the academic outcomes and may be correlated with program implementation and observed variables, this paper uses school fixed effects, state-year fixed effects, separate time trends for schools that eventually participate in the program and for schools that never participate, and initial enrollment level-year fixed effects.16 Thus, the student and school unobserved characteristics are approximated by:

∗ *#$ = FG,# + IG,J,$ + KG,#$ ∗ ,#$ = F#,# + I#,J,$ + K#,#$ (3.3)

where FG,# and F#,# are school fixed effects, which pick up time invariant differences of schools and their students; IG,J,$ and I#,J,$ capture both year fixed effects that vary over states and over initial enrollment level and time trends for schools that eventually

16 School invariant and student invariant time fixed effects are not used since they are already captured by the state-year fixed effects.

participate in the program and for schools that never participate (where k denotes state, school enrollment level, or school program participation); and KG,#$ and K#,#$ are error terms with means of zero. Substituting Equation (3.3) into Equation (3.2’), the terms in brackets become:

F# + IJ,$ + K#$ + F#(L).#$ + MIJ,$(L) × .#$N + (K#$(L) × .#$) (3.4)

∗ ∗ ∗ ∗ where F# = O ′FG,# + P ′F#,#, a school fixed effect; IJ,$ = O ′IG,J,$ + P ′I#,J,$, so the

IJ,$ terms denote state-year and initial enrollment level-year fixed effects and separate time trends for schools that eventually participate in the program and for schools that never ∗ ∗ participate; Q#$ = O ′KG,#$ + P ′K#,#$, an average of the deviation terms; F#(L) = ∗ ∗ ∗ ∗ ∗ ∗ R ′FG,# + S ′F#,#; IJ,$(L) = R ′IG,J,$ + S ′I#,J,$; K#$(L) = R ′KG,#$ + S ′K#,#$. Note that the F#(L) term is accounted for in the estimation of the program’s impact only when the program is available (.#$ = 1). It allows the program’s effect to vary through interactions between program adoption and unobserved student and school characteristics. The IJ,$(L) term represents state-year, initial enrollment level-year fixed effects, and separate time trends for schools that eventually participate in the program and for schools that never participate, that are used in the estimation when the program is adopted. The term is important in the sense that it allows the program impact to vary over time, at different rates in each state and initial enrollment levels, caused by changes in unobserved determinants of that impact. Finally, the equation estimated in this paper is given by the substitution of Equation (3.4) into Equation (3.2’):

C C C C !#$ = ( *#$ + + ,#$ + -.#$ + / (*#$ × .#$) + 4 (,#$ × .#$) + [F# + IJ,$ + K#$ +

F#(L).#$ + MIJ,$(L) × .#$N + MK#$(L) × .#$N + 5#$] (3.2”)

Equation (3.2”) allows for estimation of the unbiased impact of the Mais Educação Extended School Day Program on educational outcomes and addresses endogeneity

problems by controlling for school fixed effects, state-year and initial enrollment level- year fixed effects, separate time trends for schools that eventually participate in the program and for schools that never participate, and student and school observed characteristics. Therefore, it is assumed that, after controlling for these fixed effects, time trends, and observed variables, Pst is unlikely to be correlated with unobserved variables that determine the academic outcomes. Clustered standard errors at the school level were used for all specifications of Equation (3.2”). Estimates of Equation (3.2”) can be made more flexible. For instance, the separate time trends for schools that eventually participate in the program and for schools that never participate need not be linear. In the case of the Mais Educação Extended School Day Program, the coefficients of Equation (3.2”) are still identified if quadratic and higher power trends are used. Furthermore, since learning accumulates over time and educational policies in one year may affect future academic outcomes, the total enrollment, grade promotion, repetition, and dropping out in any year can also be affected by whether a school adopted the program in previous years. Thus, the full impact of the program may not be felt in its first year of implementation, which means that including program lagged terms, denoted as .#,$T>, .#,$TU, etc., is important to capture the long-term or cumulative program effects. In order to explore the dynamics of the program implementation and its impact on academic outcomes over the years, lags of the program adoption variable are included in the estimates of Equation (3.2”). These lagged variables take the form of indicator variables that are dummies for years 0-6 after program adoption and are equal one only in the relevant year. Thus, they capture the impacts of the program in each year after implementation and these impacts can be added up to provide the cumulative effects. Finally, since the proportion of treated schools increases significantly over time, the estimates of the impacts by time of program adoption and the cumulative effects are weighted by the proportion of schools that implemented the program in each year.17

17 The weights were calculated based on the proportion of treated schools in each year in relation to the total number of treated schools in 2014 (last year of the panel). For instance, in 2013 there were 25,388 schools with grades 1-5 that implemented the program, which were equivalent to 79.5% of the 31,944 treated schools with grades 1-5 in 2014. Then, the coefficient on the program variable lagged one year for schools that had the program in 2013 was weighted by 0.795.

In summary, this study uses an identification strategy intended to minimize the three potential sources of statistical endogeneity: (i) omitted variable bias; (ii) reverse causality; and (iii) measurement error. The problem of omitted variable bias is addressed by applying a fixed effects panel data framework to the data. With regards to reverse causality or simultaneity, there is little reason to believe it is an issue. In the context of the Brazilian extended school day program, it is unlikely that education outcomes in a given year also cause program adoption in the same year. Although schools with higher proportion of students with poor academic performance are prioritized by the state or local Department of Education for program’s implementation, current decisions are made based on outcomes from earlier periods. In order to deal with the possibility of measurement error, especially in the program adoption variable, this paper uses administrative data along with the original records from the Brazilian Ministry of Education. Data from the school census may still be misreported, however there is little to do to correct any possible bias.

3.5 Results

Estimates of the impact of the Mais Educação Extended School Day Program on (log) enrollment, grade promotion, repetition, and dropout rates, are presented in Tables 3.4 to 3.6. For each academic outcome and group of grades, two distinct specifications are estimated, the second one being the most complete by incorporating all school, teacher, and student characteristics. Both specifications control for school fixed effects, state-year and initial enrollment level-year fixed effects, and separate time trends for schools that eventually participate in the program and for schools that never participate. All estimates use clustered standard errors at the school level, and the grade promotion, repetition, and dropout variables are measured on a 0-100 scale. In what follows, the discussion of the results will emphasize the most complete specification. Table 3.4 reports the estimated results for (log) enrollment for schools with grades 1-5, 6-9, and 10-12. The estimates indicate that the program raises enrollment of grade 6- 9 students by 1.7%. The impact is negative, however, for students in grades 1-5 and 10-12. The total enrollment was reduced by 0.2% in schools with grades 1-5 and by 1.5% in

schools with grades 10-12. The estimates are not statistically significant, however, for grades 1-5. The negative effect for students in grades 10-12 should not be surprising since, in face of the longer school day, some older students may decide to enroll in schools without the program in order to dedicate more time to working. It is important to highlight that the negative impact for enrollment of students in grades 10-12 does not necessarily mean that those students are dropping out of school. Since the estimates compare schools with and without the program, the results may reflect that students in those grade levels are migrating from schools with longer school days to schools with fewer hours of classes. Similarly, the positive effect for students in grades 6- 9 does not automatically imply that the enrollment in those grade levels is higher in the whole country due to the program. The only interpretation one may infer is that the program raised the enrollment of students in grades 6-9 in schools that implemented the longer school day, compared to schools without the program. To explore whether the estimates are driven by migration of students across schools with and without the program rather than by changes in overall enrollment, the data were aggregated up to the municipio level. The estimates of the program impact on log of enrollment using municipio-level data are found in Table 3.5. The results provide evidence that the program raises the total enrollment by 3.0% in municipios with participating schools with grades 6-9. Hence, the estimates suggest that the program not only raises enrollment of students in grades 6-9 in schools that implemented the longer school day, compared to schools without the program, it is also effective at attracting more students to those grade levels. For grades 10-12, the results indicate that the program has a positive impact at the municipio level, despite of having a negative impact at the school level. Thus, it appears that the program brings more students from grades 10-12 to school in municipios with a higher proportion of schools that implemented the program, but those students prefer to enroll in schools without the program. The estimated effects of the program on the dropout, repetition, and promotion rates are shown in Table 3.6. For all grade levels, the estimates indicate that the program reduces dropout rates. Students in grades 6-9 benefited the most from the program, with a 0.5 percentage point reduction in the dropout rate. One possible reason for these effects in all

grade levels is that the extended school day program may prevent students from dropping out by offering extracurricular activities that increases the interest and demand for school. The second panel of Table 3.6 provides the results for repetition rates. The regression estimates indicate that the program increases repetition rates regardless of the grade level. This effect is highest for students in grades 6-9, for whom the program increases the repetition rate by 0.3 percentage points. As previously mentioned, the program may worsen academic outcomes of students that already performed badly since they may have less time to dedicate to academic assignments if non-academic activities are developed in the extra hours of schooling. Moreover, students with low performance during the school year may have decided to drop out without the program, but are persuaded to stay in school due to the program, and ended up failing the grade. Finally, if less prepared teachers are hired to teach during the additional school time, students may be negatively affected by worse classroom instruction. The impact of the program on grade promotion rates is presented in the bottom panel of Table 3.6. The regression estimates suggest that the program is effective at raising the promotion for grade 6-9 students by 0.2 percentage points. In contrast, the program seems to reduce the grade promotion for students in grades 1-5 by 0.1 percentage points, reflecting a small and marginally statistically significant negative effect of the extended school day program for those students. Although unexpected, this effect could be explained by the fact that the priority for the program’s implementation is given to schools with a low Development Index of Basic Education, in which students have poorer academic performance. Thus, in the presence of the program, students that already had low educational achievement may dedicate more time to non-academic activities during the extra school time, which could harm their academic performance and prevent them from being promoted to the next grade. As presented in Section 2, schools that have adopted the extended school day program must indicate six activities (academic or non-academic) they will develop in the extra hours of schooling. There is no requirement in relation to the number of academic activities schools must choose, which means that all activities can be non-academic. Because the estimates presented in Tables 3.4 and 3.6 do not provide an understanding of the dynamics of program implementation and its impact on academic

outcomes over the years, new estimates including lags of program adoption are found in Tables 3.7 to 3.10. These new estimates use the same control variables as the previous ones, but also include indicator variables for years 0-6 after adoption. The indicator variables for program adoption are dummies that equal one only in the relevant year. Furthermore, Tables 3.7 to 3.10 contain estimates of the program impact on enrollment, dropout, repetition, and grade promotion rates, respectively, accounting for cumulative effects. For each academic outcome and group of grades, two specifications are estimated, each of which provides the impact by time of program adoption and the cumulative effect. The analyses below focus on the second specification, which is the most complete as it includes all school, teacher, and student variables. The estimates of the program impact on (log) of enrollment by time of adoption and its cumulative effects are found in Table 3.7. The coefficients on program adoption lagged terms for schools with grades 1-5 indicate that the program has no statistically significant effect in the year of adoption, but it reduces the enrollment by 0.7% in the first year after program adoption. The negative effect reduces over the subsequent years, being statistically significant until the sixth year after program adoption. The cumulative effect for schools with grades 1-5 is also negative, reducing the enrollment in schools that implemented the program by 2.4% after six years. In contrast, the program raises the enrollment for grades 6-9. In the year of adoption, the program raises the enrollment by 1.5%, after which it fluctuates between 1.4% and 0.6% over the subsequent four years; then it reaches 0.2% in the fifth year after program adoption. Thus, it seems that the extent of the dynamics of the enrollment response to adoption of the extended school day program in schools with grades 6-9 is determined within five years. After six years of program implementation, the enrollment for grades 6-9 is raised by 6.1%. For schools with grades 10-12, most of the program effect on enrollment is perceived in the year of adoption. The results suggest that the program reduces the enrollment in grades 10-12 by 1.8% right after its implementation. Moreover, the negative impact accumulates over time, reaching 5.0% after six years of program. Table 3.8 contains the estimates of the program effect on dropout rate in each year after adoption and its cumulative effects. In the year of adoption, the extended school day program raises the dropout rate in schools with grades 1-5 by 0.1 percentage points. The

estimates indicate, however, that the program reduces the dropout rate in the first year after program adoption and increases it in the second year. Moreover, in the sixth year after program implementation there is a small, but statistically significant, negative impact. Hence, although the program raises the dropout rate in some years, the cumulative impact after six years of program implementation is negative. That is, the program seems to be effective at reducing the dropout for grade 1-5 students by 0.3 percentage points after six years. For schools with grades 6-9 and 10-12, the program also reduces the dropout rates. For both groups of grades, most of the impact is perceived by the third year after program adoption, but it is still present until the sixth year. The cumulative effect is also higher for those grade levels after six years of program: 1.9 percentage points and 2.2 percentage points for grades 6-9 and 10-12, respectively. Estimates of the impact of the program on repetition rates by time of adoption and the cumulative effects are shown in Table 3.9. The dynamics of the program impact are similar for all groups of grades. The results indicate that the program raises the repetition rates in all grade levels and that the impacts decrease over time. After six years, the program seems to raise the repetition rates by 1.1 percentage points for grades 1-5 and 10-12, and by 1.0 percentage point for grades 6-9. Lastly, Table 3.10 reports the estimates of the impact on grade promotion in each year after program adoption and the cumulative effects. In the year of adoption, the program reduces the promotion rate in schools with grades 1-5 by 0.2 percentage points. The results indicate that there is no statistically significant effect in the first year after program adoption, but after that the program negatively affects grade promotion, until the sixth year after adoption. Thus, the cumulative impact for schools with grades 1-5 is negative, reducing the promotion rate in schools that implemented the program by 0.8 percentage points after six years. For schools with grades 6-9, the impact is positive and statistically significant from the year of adoption to the sixth year after program implementation. The highest impact occurs in the second year after program adoption, after which it decreases over the subsequent four years. After six years, the program raises the promotion rate by 0.9 percentage points. The estimation results also show that the program raises the grade promotion for schools with grades 10-12, but just after the third year. The

cumulative effect is positive and statistically significant only after six years of program, when it reaches 1.1 percentage points. In order to check the validity of the results in Tables 3.4 and 3.6, a placebo test is conducted. The test was performed using data only from 1999 to 2007 and consists of assigning the values of the treatment variable from the first year of the program (2008) to the year immediately before that, checking for any statistically significant effect. Ideally, one would not find any effect of the program placebo variable, since it would represent that unobserved changes prior to the treatment are not fully captured by the control variables. Table 3.11 shows the results for the placebo test with no program lagged terms for all grade levels and all outcomes. For grades 1-5, a statistically significant effect at the 10% level is observed for the placebo program variable affecting repetition rate, but there are no significant effects for the other three outcome variables. The placebo program variable is also statistically significant at the 5% level for grades 6-9 when repetition and dropout rates are evaluated, but not when enrollment and grade promotion are the dependent variables. Finally, three out of four coefficients in grades 10-12 are statistically significant, two at the 1% level and one at the 5% level of significance. Thus, although the placebo test does not provide definitive conclusion on the validity of the results, it suggests that some of the estimates for grades 10-12 and for grades 6-9 should be treated with caution.

3.6 Conclusion

This study presents estimates of the impact of the Brazilian Mais Educação Extended School Day Program on the academic outcomes – specifically on total enrollment, grade promotion, repetition, and dropout rates – of students in grades 1-5, 6-9, and 10-12. Using school census data and administrative data on program’s implementation from 1999 to 2014, the results of this chapter suggest that the Brazilian longer school day program is effective at reducing dropout rates for students in all grade levels. The program is also effective at raising the enrollment of students in grades 6-9, but reduces the enrollment of students in grades 10-12 in schools where the program is available. Moreover, the estimates indicate that the impact on grade promotion is positive for students

in grades 6-9. It is negative, however, for students in lower grades. Finally, the program seems to increase repetition rates for students in all grade levels. Estimates of the program impact by time of adoption and its cumulative effects indicate that, for most outcomes and groups of grades, the full impact of the program is not received in the year of program adoption, but it accumulates over the time. Therefore, the present results provide evidence that there are long-term effects of the extended school day program and, although they decrease over time, they continue to have an impact for up to six years. The overall results of this study suggest that the extended school day program in Brazil has been effective at reducing dropout rates, but at the same time, it has increased repetition rates in all grade levels. It is possible that schools are adopting non-academic activities rather than academic ones, becoming more attractive for students and reducing dropping out. However, those activities may reduce the time students dedicate to homework which, ultimately, could increase repetition rates and reduce grade promotion. Therefore, while the findings of this paper indicate that the Brazilian extend school day program provides positive effects for some outcomes, especially for students in grades 6- 9, they raise some concerns about the mechanisms through which the program has been implemented. This means that school administrators and policy makers should take into account possible problems in the actual program design. Finally, further research, perhaps using a randomized control trial, would be useful to assess the impact of the Brazilian program and to provide better understanding of its impacts.

Table 3.1 – Number of Schools Participating in the Mais Educação Extended School Day Program, 2008-2014 Treated schools (all Years % of all schools % of public schools grade levels) 2008 1,380 0.55 0.68 2009 4,638 1.82 2.27 2010 9,661 3.72 4.67 2011 13,032 4.94 6.24 2012 28,075 10.47 13.30 2013 40,007 14.71 18.79 2014 50,874 18.41 23.70

Table 3.2 – Number of Schools in Brazil’s School Census from 1999 to 2014 Schools with Total number 1st to 5th Schools with % of treated of treated Years Total number and/or 6th to panel data schools (after schools (after of schools 9th and/or 10th (from 1999 to balancing the balancing the to 12th grade current year) panel) panel) classes (1) (2) (3) (4) (5) (6) 1999 266,645 250,173 250,173 0 - 2000 261,988 246,601 177,104 0 - 2001 264,735 241,809 166,886 0 - 2002 256,986 237,292 157,350 0 - 2003 253,405 234,761 150,350 0 - 2004 248,257 233,266 144,845 0 - 2005 248,103 232,182 138,494 0 - 2007 237,387 233,704 126,506 0 - 2008 250,350 235,107 122,964 438 0.43 2009 255,445 236,584 118,345 1,949 1.93 2010 259,831 237,357 114,041 3,765 3.73 2011 263,833 238,092 110,670 6,060 6.00 2012 268,244 238,868 107,941 17,623 17.44 2013 272,049 239,453 104,356 29,003 28.70 2014 276,331 239,975 101,063 37,652 37.26 Note: Column 6 is obtained by dividing column 5 by the final number of schools with panel data (101,063).

Table 3.3 – Descriptive Statistics for Eventually Treated and Never Treated Schools Schools with Schools with Schools with

grades 1-5 grades 6-9 grades 10-12 Variables Eventually Never Eventually Never Eventually Never

treated treated treated treated treated treated School variables Total enrollment (units) 190.5 112.6 263.5 251.2 353.4 378.5 (180.5) (158.7) (257.6) (244.8) (312.5) (407.6) Dropout rate (%) 5.9 5.3 8.7 4.9 13.6 7.8 (8.8) (10.0) (10.0) (9.3) (11.0) (9.7) Repetition rate (%) 13.2 11.7 11.9 8.4 9.7 8.6 (10.7) (13.5) (9.7) (9.3) (8.5) (8.0) Grade promotion rate (%) 80.8 82.9 79.4 86.7 76.7 83.6 (15.3) (18.5) (13.8) (13.6) (12.9) (13.1) Geographic region (%) North 13.2 14.2 12.0 7.8 10.8 4.3 (33.9) (34.9) (32.5) (26.8) (31.0) (20.3) Northeast 52.0 41.0 43.9 19.8 25.2 18.9 (50.0) (49.2) (49.6) (39.9) (43.4) (39.1) South 12.0 14.6 16.0 23.4 18.7 19.5 (32.4) (35.3) (36.6) (42.3) (39.0) (39.6) Central-West 5.4 4.7 7.5 8.0 10.4 7.1 (22.6) (21.2) (26.3) (27.1) (30.5) (25.7) Southeast 17.4 25.5 20.7 41.0 34.9 50.2 (37.9) (43.6) (40.5) (49.2) (47.7) (50.0) Rural (%) 48.8 62.5 38.7 21.7 10.9 3.7 (50.0) (48.4) (48.7) (41.2) (31.2) (18.9) Electricity (%) 94.4 82.1 98.0 97.2 99.7 99.6 (23.0) (38.3) (13.9) (16.6) (5.3) (6.5) Water (%) 97.3 96.0 98.3 99.1 99.6 99.5 (16.2) (19.5) (13.0) (9.7) (6.2) (7.0) Sewage (%) 95.5 87.8 97.5 97.1 99.3 99.3 (20.7) (32.7) (15.5) (16.9) (8.5) (8.1) Offers meal (%) 98.6 85.1 98.4 69.9 97.4 59.9 (11.9) (35.6) (12.7) (45.9) (15.8) (49.0) Class size (units) 23.6 18.9 27.0 26.4 32.1 31.6 (9.6) (11.9) (8.8) (9.3) (8.6) (9.6) Library (%) 34.1 29.5 52.5 67.3 73.7 78.1 (47.4) (45.6) (49.9) (46.9) (44.0) (41.4) Computer lab (%) 28.6 24.4 44.4 59.8 68.9 78.6 (45.2) (43.0) (49.7) (49.0) (46.3) (41.0) Science lab (%) 5.7 10.8 14.7 35.9 38.4 55.4 (23.2) (31.1) (35.4) (48.0) (48.6) (49.7) Computer (units) 5.1 6.1 8.3 16.0 14.0 24.9 (14.7) (20.8) (17.5) (30.7) (16.1) (41.2) Internet (%) 31.2 30.3 43.2 62.4 65.7 79.6 (46.3) (46.0) (49.5) (48.4) (47.5) (40.3) Teacher with college degree (%) 41.2 39.2 71.4 81.8 89.5 92.7 (39.1) (41.0) (36.0) (30.1) (21.3) (16.4) Student variables Female (%) 46.6 46.6 49.5 49.0 53.8 53.0 (6.1) (9.6) (6.9) (8.7) (7.1) (7.7) Skin color (%) White 18.8 25.0 17.0 30.0 24.9 32.5 (21.3) (27.5) (20.9) (28.3) (24.9) (27.7) Black 3.8 3.1 3.1 2.2 3.0 2.2 (6.9) (7.2) (6.1) (4.8) (4.9) (3.9) Pardo 36.0 31.9 28.7 20.6 24.6 18.9 (26.5) (28.4) (24.8) (22.7) (21.9) (20.2) Yellow 0.5 0.5 0.6 0.5 0.8 0.6 (2.6) (2.7) (2.9) (2.5) (3.4) (2.3) Indigenous 0.7 1.5 0.6 1.2 0.8 0.5 (6.3) (10.7) (5.8) (9.2) (6.7) (4.8) Non-declared skin color 40.2 37.9 50.1 45.4 45.9 45.3 (32.5) (33.2) (34.6) (35.3) (34.9) (34.4) Lives in rural area (%) 6.7 8.1 5.8 3.6 3.1 1.6 (24.2) (26.8) (22.1) (17.4) (14.1) (9.4) Observations 503,800 702,647 310,156 276,572 69,439 161,041 Number of schools 33,587 46,843 20,677 18,438 4,629 10,736 Notes: Standard deviations in parentheses.

Table 3.4 – Estimates of the Program Impact on Log of Enrollment: Results for Schools with Grades 1-5, 6-9 and 10-12, 1999-2014 Grades 1-5 Grades 6-9 Grades 10-12 Variables (1) (2) (1) (2) (1) (2) School program participation -0.010*** -0.002 0.028*** 0.017*** -0.021** -0.015* (0.004) (0.003) (0.004) (0.003) (0.009) (0.008)

Observations 1,205,726 836,948 582,397 536,662 229,830 225,983 R-squared 0.911 0.915 0.913 0.924 0.904 0.917

School and student variables No Yes No Yes No Yes School fixed effects Yes Yes Yes Yes Yes Yes State-year fixed effects Yes Yes Yes Yes Yes Yes Enrollment 1999-year fixed effects Yes Yes Yes Yes Yes Yes Trend x ever program adoption 1 Yes Yes Yes Yes Yes Yes Notes: Robust standard errors clustered at the school level. ***Significant at the 1% level. **Significant at the 5% level. *Significant at the 10% level. 1Time trends from the 1st to the 6th power are used.

Table 3.5 – Estimates of the Program Impact on Log of Enrollment at the Municipio- Level: Results for Schools with Grades 1-5, 6-9 and 10-12, 1999-2014 Grades 1-5 Grades 6-9 Grades 10-12 Variables (1) (2) (1) (2) (1) (2) School program participation -0.001 -0.013 0.032** 0.030** 0.040*** 0.033*** (0.016) (0.015) (0.014) (0.012) (0.012) (0.011)

Observations 69,109 67,588 74,235 72,825 69,550 68,549 R-squared 0.972 0.975 0.956 0.964 0.951 0.959

School and student variables No Yes No Yes No Yes Municipio fixed effects Yes Yes Yes Yes Yes Yes State-year fixed effects Yes Yes Yes Yes Yes Yes Enrollment 1999-year fixed effects Yes Yes Yes Yes Yes Yes Trend x ever program adoption 1 Yes Yes Yes Yes Yes Yes Notes: Robust standard errors clustered at the municipio level. ***Significant at the 1% level. **Significant at the 5% level. *Significant at the 10% level. 1Time trends from the 1st to the 6th power are used.

Table 3.6 – Estimates of the Program Impact on Dropout, Repetition, and Grade Promotion Rates: Results for Schools with Grades 1-5, 6-9 and 10-12, 1999-2014

Dependent variable: dropout rate Grades 1-5 Grades 6-9 Grades 10-12 Variables (1) (2) (1) (2) (1) (2) School program participation -0.117*** -0.092** -0.542*** -0.461*** -0.427*** -0.302** (0.038) (0.036) (0.061) (0.057) (0.155) (0.154)

Observations 1,201,584 834,429 579,925 534,782 227,617 223,926 R-squared 0.476 0.558 0.589 0.611 0.651 0.658

Dependent variable: repetition rate Grades 1-5 Grades 6-9 Grades 10-12 Variables (1) (2) (1) (2) (1) (2) School program participation 0.257*** 0.216*** 0.333*** 0.299*** 0.239* 0.244* (0.055) (0.053) (0.072) (0.071) (0.144) (0.143)

Observations 1,201,584 834,429 579,925 534,782 227,617 223,926 R-squared 0.496 0.601 0.519 0.543 0.529 0.530

Dependent variable: grade promotion rate Grades 1-5 Grades 6-9 Grades 10-12 Variables (1) (2) (1) (2) (1) (2) School program participation -0.140** -0.124* 0.209** 0.162* 0.189 0.058 (0.069) (0.066) (0.092) (0.089) (0.191) (0.190)

Observations 1,201,584 834,429 579,925 534,782 227,617 223,926 R-squared 0.620 0.699 0.619 0.644 0.675 0.682

School and student characteristics No Yes No Yes No Yes School fixed effects Yes Yes Yes Yes Yes Yes State-year fixed effects Yes Yes Yes Yes Yes Yes Enrollment 1999-year fixed effects Yes Yes Yes Yes Yes Yes Trend x ever program adoption1 Yes Yes Yes Yes Yes Yes Notes: Robust standard errors clustered at the school level. ***Significant at the 1% level. **Significant at the 5% level. *Significant at the 10% level. The estimated coefficients are the direct effect of school program adoption on the dependent variable, since the latter was multiplied by 100. 1Time trends from the 1st to the 6th power are used.

Table 3.7 – Estimates of the Program Impact on Log of Enrollment by Time of Adoption and Cumulative Effects: Results for Schools with Grades 1-5, 6-9 and 10-12, 1999-2014 Grades 1-5 Grades 6-9 Grades 10-12 Variables (1) (2) (1) (2) (1) (2) Program Effects Program adoption -0.009*** -0.002 0.024*** 0.015*** -0.025*** -0.018*** (0.003) (0.003) (0.003) (0.003) (0.007) (0.007) 1st year after program adoption -0.011*** -0.007** 0.024*** 0.013*** -0.013** -0.010 (0.004) (0.003) (0.004) (0.003) (0.007) (0.006) 2nd year after program adoption -0.008** -0.003 0.028*** 0.014*** -0.010 -0.005 (0.003) (0.003) (0.004) (0.003) (0.007) (0.006) 3th year after program adoption -0.011*** -0.004*** 0.017*** 0.010*** -0.009** -0.005 (0.001) (0.001) (0.002) (0.002) (0.004) (0.004) 4th year after program adoption -0.010*** -0.004*** 0.010*** 0.006*** -0.009** -0.007* (0.001) (0.001) (0.001) (0.001) (0.004) (0.003) 5th year after program adoption -0.006*** -0.003*** 0.005*** 0.002*** -0.006*** -0.004* (0.001) (0.001) (0.001) (0.001) (0.002) (0.002) 6th year after program adoption -0.002*** -0.001*** 0.001 0.000 -0.003** -0.002 (0.000) (0.000) (0.001) (0.000) (0.002) (0.001)

Observations 1,205,726 836,948 582,397 536,662 229,830 225,983 R-squared 0.914 0.916 0.915 0.925 0.905 0.918

Cumulative Program Effects Program adoption -0.009*** -0.002 0.024*** 0.015*** -0.025*** -0.018*** (0.003) (0.003) (0.003) (0.003) (0.007) (0.007) 1 year of program -0.021*** -0.008 0.048*** 0.028*** -0.038*** -0.028** (0.007) (0.006) (0.007) (0.006) (0.014) (0.013) 2 years of program -0.029*** -0.012 0.076*** 0.043*** -0.048** -0.033* (0.010) (0.009) (0.010) (0.009) (0.020) (0.019) 3 years of program -0.040*** -0.016* 0.093*** 0.053*** -0.056** -0.037* (0.011) (0.010) (0.012) (0.010) (0.023) (0.022) 4 years of program -0.049*** -0.019* 0.103*** 0.059*** -0.066** -0.044* (0.012) (0.010) (0.013) (0.011) (0.027) (0.025) 5 years of program -0.056*** -0.023** 0.108*** 0.061*** -0.072** -0.048* (0.012) (0.011) (0.014) (0.012) (0.028) (0.026) 6 years of program -0.058*** -0.024** 0.109*** 0.061*** -0.075*** -0.050* (0.013) (0.011) (0.014) (0.012) (0.029) (0.027)

School and student characteristics No Yes No Yes No Yes School fixed effects Yes Yes Yes Yes Yes Yes State-year fixed effects Yes Yes Yes Yes Yes Yes Enrollment 1999-year fixed effects Yes Yes Yes Yes Yes Yes Trend x ever program adoption 1 Yes Yes Yes Yes Yes Yes Notes: Robust standard errors clustered at the school level. ***Significant at the 1% level. **Significant at the 5% level. *Significant at the 10% level. 1Time trends from the 1st to the 6th power are used.

Table 3.8 – Estimates of the Program Impact on Dropout Rate by Time of Adoption and Cumulative Effects: Results for Schools with Grades 1-5, 6-9 and 10-12, 1999-2014 Grades 1-5 Grades 6-9 Grades 10-12 Variables (1) (2) (1) (2) (1) (2) Program Effects Program adoption -0.102*** 0.058* -0.517*** -0.433*** -0.323** -0.221 (0.035) (0.034) (0.057) (0.055) (0.150) (0.149) 1st year after program adoption -0.130*** -0.129*** -0.518*** -0.479*** -0.527*** -0.428*** (0.040) (0.038) (0.061) (0.058) (0.123) (0.122) 2nd year after program adoption -0.057* 0.066** -0.461*** -0.431*** -0.470*** -0.394*** (0.033) (0.031) (0.055) (0.052) (0.116) (0.115) 3th year after program adoption 0.026* -0.009 -0.220*** -0.201*** -0.432*** -0.386*** (0.014) (0.013) (0.026) (0.025) (0.078) (0.077) 4th year after program adoption -0.013 0.000 -0.161*** -0.150*** -0.343*** -0.309*** (0.012) (0.011) (0.021) (0.020) (0.063) (0.063) 5th year after program adoption -0.007 0.001 -0.119*** -0.112*** -0.255*** -0.234*** (0.008) (0.007) (0.014) (0.014) (0.041) (0.041) 6th year after program adoption -0.007*** -0.005* -0.067*** -0.063*** -0.230*** -0.218*** (0.003) (0.003) (0.009) (0.008) (0.029) (0.029)

Observations 1,201,584 834,429 579,925 534,782 227,617 223,926 R-squared 0.476 0.558 0.590 0.611 0.651 0.658

Cumulative Program Effects Program adoption -0.102*** 0.058* -0.517*** -0.433*** -0.323** -0.221 (0.035) (0.034) (0.057) (0.055) (0.150) (0.149) 1 year of program -0.232*** -0.187*** -1.034*** -0.912*** -0.850*** -0.649*** (0.072) (0.069) (0.112) (0.107) (0.252) (0.250) 2 years of program -0.289*** -0.253*** -1.496*** -1.343*** -1.320*** -1.042*** (0.103) (0.098) (0.162) (0.154) (0.346) (0.343) 3 years of program -0.314*** -0.262** -1.716*** -1.544*** -1.752*** -1.428*** (0.115) (0.109) (0.184) (0.175) (0.404) (0.400) 4 years of program -0.328*** -0.262** -1.877*** -1.693*** -2.095*** -1.737*** (0.125) (0.119) (0.201) (0.191) (0.452) (0.447) 5 years of program -0.334*** -0.261** -1.997*** -1.806*** -2.350*** -1.971*** (0.131) (0.124) (0.212) (0.202) (0.477) (0.472) 6 years of program -0.341*** -0.266** -2.064*** -1.868*** -2.580*** -2.189*** (0.132) (0.126) (0.216) (0.206) (0.491) (0.486)

School and student characteristics No Yes No Yes No Yes School fixed effects Yes Yes Yes Yes Yes Yes State-year fixed effects Yes Yes Yes Yes Yes Yes Enrollment 1999-year fixed effects Yes Yes Yes Yes Yes Yes Trend x ever program adoption 1 Yes Yes Yes Yes Yes Yes Notes: Robust standard errors clustered at the school level. ***Significant at the 1% level. **Significant at the 5% level. *Significant at the 10% level. The estimated coefficients are the direct effect of school program adoption on the dependent variable, since the latter was multiplied by 100. 1Time trends from the 1st to the 6th power are used.

Table 3.9 – Estimates of the Program Impact on Repetition Rate by Time of Adoption and Cumulative Effects: Results for Schools with Grades 1-5, 6-9 and 10-12, 1999-2014 Grades 1-5 Grades 6-9 Grades 10-12 Variables (1) (2) (1) (2) (1) (2) Program Effects Program adoption 0.252*** 0.222*** 0.295*** 0.259*** 0.134 0.142 (0.052) (0.051) (0.071) (0.070) (0.140) (0.140) 1st year after program adoption 0.243*** 0.209*** 0.321*** 0.323*** 0.317*** 0.319*** (0.057) (0.055) (0.073) (0.073) (0.117) (0.117) 2nd year after program adoption 0.272*** 0.265*** 0.230*** 0.189*** 0.246** 0.250** (0.046) (0.044) (0.064) (0.063) (0.106) (0.106) 3th year after program adoption 0.142*** 0.153*** 0.096*** 0.093*** 0.136* 0.142** (0.019) (0.018) (0.032) (0.031) (0.071) (0.071) 4th year after program adoption 0.146*** 0.148*** 0.069*** 0.067*** 0.087 0.090* (0.015) (0.015) (0.025) (0.025) (0.054) (0.054) 5th year after program adoption 0.089*** 0.088*** 0.046*** 0.047*** 0.109*** 0.109*** (0.010) (0.009) (0.017) (0.017) (0.036) (0.036) 6th year after program adoption 0.023*** 0.022*** 0.008 0.011 0.089*** 0.090*** (0.004) (0.004) (0.010) (0.010) (0.027) (0.027)

Observations 1,201,584 834,429 579,925 534,782 227,617 223,926 R-squared 0.496 0.601 0.519 0.543 0.529 0.530

Cumulative Program Effects Program adoption 0.252*** 0.222*** 0.295*** 0.259*** 0.134 0.142 (0.052) (0.051) (0.071) (0.070) (0.140) (0.140) 1 year of program 0.495*** 0.430*** 0.616*** 0.582*** 0.451* 0.461** (0.101) (0.098) (0.131) (0.130) (0.232) (0.232) 2 years of program 0.767*** 0.695*** 0.846*** 0.771*** 0.697** 0.711** (0.140) (0.135) (0.184) (0.182) (0.315) (0.315) 3 years of program 0.909*** 0.848*** 0.942*** 0.863*** 0.833** 0.853** (0.154) (0.149) (0.206) (0.205) (0.365) (0.365) 4 years of program 1.055*** 0.996*** 1.010*** 0.931*** 0.920** 0.942** (0.165) (0.159) (0.223) (0.221) (0.402) (0.402) 5 years of program 1.145*** 1.084*** 1.057*** 0.978*** 1.029** 1.052** (0.171) (0.165) (0.233) (0.231) (0.421) (0.421) 6 years of program 1.168*** 1.106*** 1.065*** 0.989*** 1.118*** 1.142*** (0.173) (0.167) (0.237) (0.236) (0.432) (0.432)

School and student characteristics No Yes No Yes No Yes School fixed effects Yes Yes Yes Yes Yes Yes State-year fixed effects Yes Yes Yes Yes Yes Yes Enrollment 1999-year fixed effects Yes Yes Yes Yes Yes Yes Trend x ever program adoption 1 Yes Yes Yes Yes Yes Yes Notes: Robust standard errors clustered at the school level. ***Significant at the 1% level. **Significant at the 5% level. *Significant at the 10% level. The estimated coefficients are the direct effect of school program adoption on the dependent variable, since the latter was multiplied by 100. 1Time trends from the 1st to the 6th power are used.

Table 3.10 – Estimates of the Program Impact on Grade Promotion Rate by Time of Adoption and Cumulative Effects: Results for Schools with Grades 1-5, 6-9 and 10-12, 1999-2014 Grades 1-5 Grades 6-9 Grades 10-12 Variables (1) (2) (1) (2) (1) (2) Program Effects Program adoption -0.150** -0.163*** 0.222** 0.174** 0.189 0.079 (0.064) (0.062) (0.088) (0.086) (0.182) (0.180) 1st year after program adoption -0.114 -0.079 0.197** 0.156* 0.210 0.109 (0.072) (0.069) (0.093) (0.091) (0.152) (0.151) 2nd year after program adoption -0.215*** -0.199*** 0.232*** 0.242*** 0.225 0.144 (0.059) (0.056) (0.083) (0.081) (0.143) (0.142) 3th year after program adoption -0.116*** -0.143*** 0.125*** 0.108*** 0.295*** 0.243*** (0.025) (0.024) (0.041) (0.040) (0.094) (0.094) 4th year after program adoption -0.133*** -0.148*** 0.092*** 0.082*** 0.256*** 0.220*** (0.020) (0.019) (0.032) (0.031) (0.075) (0.075) 5th year after program adoption -0.083*** -0.089*** 0.073*** 0.065*** 0.146*** 0.124*** (0.013) (0.012) (0.021) (0.021) (0.048) (0.048) 6th year after program adoption -0.016*** -0.017*** 0.059*** 0.052*** 0.141*** 0.128*** (0.005) (0.005) (0.013) (0.012) (0.032) (0.032)

Observations 1,201,584 834,429 579,925 534,782 227,617 223,926 R-squared 0.620 0.699 0.619 0.644 0.675 0.682

Cumulative Program Effects Program adoption -0.150** -0.163*** 0.222** 0.174** 0.189 0.079 (0.064) (0.062) (0.088) (0.086) (0.182) (0.180) 1 year of program -0.264** -0.243** 0.418** 0.330** 0.399 0.188 (0.128) (0.123) (0.168) (0.164) (0.310) (0.308) 2 years of program -0.478*** -0.442*** 0.650*** 0.572** 0.623 0.332 (0.180) (0.172) (0.239) (0.234) (0.427) (0.426) 3 years of program -0.594*** -0.585*** 0.774*** 0.680*** 0.919* 0.575 (0.200) (0.191) (0.271) (0.265) (0.498) (0.496) 4 years of program -0.727*** -0.734*** 0.867*** 0.762*** 1.175** 0.795 (0.216) (0.206) (0.294) (0.288) (0.555) (0.552) 5 years of program -0.810*** -0.823*** 0.940*** 0.828*** 1.321** 0.919 (0.226) (0.215) (0.309) (0.302) (0.585) (0.582) 6 years of program -0.827*** -0.839*** 0.999*** 0.880*** 1.462** 1.047* (0.228) (0.217) (0.315) (0.308) (0.601) (0.598)

School and student characteristics No Yes No Yes No Yes School fixed effects Yes Yes Yes Yes Yes Yes State-year fixed effects Yes Yes Yes Yes Yes Yes Enrollment 1999-year fixed effects Yes Yes Yes Yes Yes Yes Trend x ever program adoption 1 Yes Yes Yes Yes Yes Yes Notes: Robust standard errors clustered at the school level. ***Significant at the 1% level. **Significant at the 5% level. *Significant at the 10% level. The estimated coefficients are the direct effect of school program adoption on the dependent variable, since the latter was multiplied by 100. 1Time trends from the 1st to the 6th power are used.

Table 3.11 – Placebo Test: Estimates of the Program Impact for Schools with Grades 1-5, 6-9 and 10-12, 1999-2007 (Schools with Program in 2008 Assigned to 2007)

Variables Grades 1-5 Grades 6-9 Grades 10-12 Dependent variable: log enrollment Program adoption (placebo) -0.015 0.009 0.093** (0.021) (0.019) (0.041)

Observations 473,960 263,966 112,187 R-squared 0.942 0.943 0.931

Dependent variable: dropout rate Program adoption (placebo) -0.142 -1.017** 3.839*** (0.329) (0.428) (0.987)

Observations 472,498 262,895 110,500 R-squared 0.583 0.680 0.686

Dependent variable: repetition rate Program adoption (placebo) 0.673* 1.171** -0.102 (0.374) (0.472) (0.763)

Observations 472,498 262,895 110,500 R-squared 0.632 0.579 0.552

Dependent variable: grade promotion rate Program adoption (placebo) -0.531 -0.154 -3.737*** (0.451) (0.553) (1.029)

Observations 472,498 262,895 110,500 R-squared 0.718 0.698 0.698

School and student characteristics Yes Yes Yes School fixed effects Yes Yes Yes State-year fixed effects Yes Yes Yes Enrollment 1999-year fixed effects Yes Yes Yes Trend x ever program adoption 1 Yes Yes Yes Notes: Robust standard errors clustered at the school level. ***Significant at the 1% level. **Significant at the 5% level. *Significant at the 10% level. The estimated coefficients are the direct effect of school program adoption on the dependent variable, since the latter was multiplied by 100 (except log enrollment). 1Time trends from the 1st to the 6th power are used.

Schools with Grades 1-5

500 450 400 350 300 250 200 150 100 50 0 1999 2000 2001 2002 2003 2004 2005 2007 2008 2009 2010 2011 2012 2013 2014

Treated Never treated

Schools with Grades 6-9

500 450 400 350 300 250 200 150 100 50 0 1999 2000 2001 2002 2003 2004 2005 2007 2008 2009 2010 2011 2012 2013 2014

Treated Never treated

Schools with Grades 10-12

500 450 400 350 300 250 200 150 100 50 0 1999 2000 2001 2002 2003 2004 2005 2007 2008 2009 2010 2011 2012 2013 2014

Treated Never treated

Figure 3.1 – Total Enrollment for Treated and Never Treated Schools with Grades 1-5, 6- 9, and 10-12, 1999-2014

Schools with Grades 1-5

20 18 16 14 12 10 8 6 4 2 0 1999 2000 2001 2002 2003 2004 2005 2007 2008 2009 2010 2011 2012 2013 2014

Treated Never treated

Schools with Grades 6-9

20 18 16 14 12 10 8 6 4 2 0 1999 2000 2001 2002 2003 2004 2005 2007 2008 2009 2010 2011 2012 2013 2014

Treated Never treated

Schools with Grades 10-12

20 18 16 14 12 10 8 6 4 2 0 1999 2000 2001 2002 2003 2004 2005 2007 2008 2009 2010 2011 2012 2013 2014

Treated Never treated

Figure 3.2 – Dropout Rates for Treated and Never Treated Schools with Grades 1-5, 6-9, and 10-12, 1999-2014

Schools with Grades 1-5

20.0 18.0 16.0 14.0 12.0 10.0 8.0 6.0 4.0 2.0 0.0 1999 2000 2001 2002 2003 2004 2005 2007 2008 2009 2010 2011 2012 2013 2014

Treated Never treated

Schools with Grades 6-9

20 18 16 14 12 10 8 6 4 2 0 1999 2000 2001 2002 2003 2004 2005 2007 2008 2009 2010 2011 2012 2013 2014

Treated Never treated

Schools with Grades 10-12

20 18 16 14 12 10 8 6 4 2 0 1999 2000 2001 2002 2003 2004 2005 2007 2008 2009 2010 2011 2012 2013 2014

Treated Never treated

Figure 3.3 – Repetition Rates for Treated and Never Treated Schools with Grades 1-5, 6- 9, and 10-12, 1999-2014

Schools with Grades 1-5

100 90 80 70 60 50 40 30 20 10 0 1999 2000 2001 2002 2003 2004 2005 2007 2008 2009 2010 2011 2012 2013 2014

Treated Never treated

Schools with Grades 6-9

100 90 80 70 60 50 40 30 20 10 0 1999 2000 2001 2002 2003 2004 2005 2007 2008 2009 2010 2011 2012 2013 2014

Treated Never treated

Schools with Grades 10-12

100 90 80 70 60 50 40 30 20 10 0 1999 2000 2001 2002 2003 2004 2005 2007 2008 2009 2010 2011 2012 2013 2014

Treated Never treated

Figure 3.4 – Grade Promotion Rates for Treated and Never Treated Schools with Grades 1-5, 6-9, and 10-12, 1999-2014

Chapter 4

Affirmative Action in Brazilian Universities: The Impact of Racial and Low-Income Quotas on Academic Performance18

4.1 Introduction

Brazilian society is renowned for its long history of socioeconomic and racial disparities. In spite of recent improvements in some indicators, such as reductions in both income inequality and poverty and greater participation of non-white individuals in the labor force, many disparities persist and thus greatly affect the Brazilian educational system, notably in higher education. Access to higher education in Brazilian universities is a very competitive process in which only those who are well prepared or have higher socioeconomic status are likely to gain admission. According to the Brazilian Institute of Geography and Statistics (IBGE), 51.3% of young people between 18 and 24 years old are enrolled in post-secondary education in Brazil (IBGE, 2017). This level is much lower than in developed countries, where this percentage exceeds 70% (World Bank, 2013). While more than half of the Brazilian population is composed of non-white people, among those who are enrolled in universities only 34.5% are black and brown (5.2% black and 29.3% brown).19 In addition, no more than 8% of students enrolled in public universities are from

18 Some contents of this chapter have been published in the paper “Racial and low-income quotas in Brazilian universities: impact on academic performance” (https://doi.org/10.1108/JES-10-2016-0200). This article is © Emerald Publishing Limited and permission has been granted for this version to appear here [https://www.emeraldinsight.com/doi/abs/10.1108/JES-10-2016-0200]. Emerald does not grant permission for this article to be further copied/distributed or hosted elsewhere without the express permission from Emerald Publishing Limited. 19 In Brazil, various terms are used to identify different shades of skin color, but in the decennial censuses individuals are asked in a multiple-choice question to self-identify their skin tone or race based on only five possible answers. Respondents can self-identify as white, black, brown (mixed skin color), yellow (Asian descent), and indigenous. In 2010, 47.5% of the Brazilian population was white, 7.5% was black, 43.4% brown, 1.1% yellow, and 0.4% was indigenous (IBGE, 2017). According to Carvalho et al. (2004), the identification of skin color and race in Brazil is very a subjective process, based not only on physical appearance, but also on other factors such as income, education, and social status. For instance, a person with dark skin who is also poor is likely to identify himself or herself as black. In contrast, an individual of the same skin tone but with a higher status is more likely to identify himself or herself as brown or another color closer to the white end of the color spectrum.

the poorest 20% of the population, a proportion that is even lower in private universities (3.7%). In order to increase less advantaged students’ access to higher education, the Brazilian Government and some private universities have adopted affirmative action policies, such as racial and low-income quotas in the admissions process. These quotas target disadvantaged students based on criteria such as income, race, and being graduates of public secondary schools. The main goal of the quota policy is to offer better social and economic opportunities for disadvantaged students and, as a consequence, to reduce the socioeconomic and racial disparities in access to higher education and, eventually, in labor market outcomes. Given that quota policies have occupied a place of increasing prominence in the discussion on reducing socioeconomic and racial disparities, this chapter aims to evaluate the impact of racial and low-income quotas on the academic performance of students in public and private Brazilian universities. Although there is a consensus regarding the extent of racial and social inequalities in Brazil, there is no agreement about the benefits of the quota policy in Brazilian universities (Bailey and Peria, 2010). Some critics argue that students admitted under quotas do not have the necessary academic background to perform well in college. The argument is that the adoption of quotas would result in poor academic performance of the beneficiaries, with negative effects on overall student achievement. Empirical studies for the Brazilian case, however, show that, on average, students who are admitted under quotas tend to have satisfactory grades (Cardoso, 2008; Francis and Tannuri-Pianto, 2013; Childs and Stromquist, 2014). Most of the literature related to the impact of affirmative action on college enrollment and academic performance has focused on the United States, but a growing body of research has analyzed developing countries, such as Brazil and India (Desai and Kulkami, 2008; Frisancho and Krishna, 2016; Bertrand et al., 2010; Bailey and Peria, 2010). The recent literature for the United States has investigated the practice of affirmative action in higher education and its effect on the admission and enrollment of minority students (Tienda et al., 2008; Epple et al., 2008; Blume and Long, 2014; Howell, 2010; Grodsky and Kalogrides, 2008; Hinrichs, 2012; Dickson, 2006).

Hinrichs (2012) evaluates the impact of affirmative action bans on college enrollment, educational attainment, and the demographic composition of students in American universities. He finds that, for a typical student and a typical college, affirmative action bans have no effect. The ban, however, decreases underrepresented minority enrollment and increases white enrollment at selective colleges. Blume and Long (2014), in turn, estimate the impact of the change in affirmative action policies in college admission decisions; that is, they evaluate how affirmative action bans affect the relative likelihood of admission of minority and nonminority applicants. They find substantial declines in admissions of disadvantage students between 1992 and 2004 in highly selective colleges in the states where the policy was banned (i.e., Alabama, California, Florida, Georgia, Louisiana, Mississippi, Texas, and Washington). The authors also demonstrate how these declines affect not only students who live in the states where the policy was banned, but also students in adjacent states that lack highly selective colleges. In the Indian context, Bertrand et al. (2010) evaluate the impact of a low-income quota program in engineering colleges, and its implications for labor market outcomes. They find that the policy successfully targets poorer students, who gain economic benefits, such as higher incomes, from attending those colleges under quotas. Frisancho and Krishna (2016) similarly find that affirmative action effectively targets minority students who are poorer than the average displaced non-minority student. In addition, their results suggest that students admitted under quotas, especially those in more selective majors, have lower performance than non-quota students and are more likely to get worse jobs. Since the implementation of a national quota system in Brazilian universities in 1999, when the Special System of Quotas was established, several studies have been conducted on the topic (Francis and Tannuri-Pianto, 2012; Vidigal, 2018; Ferman and Assunção, 2005; Cunha, 2006; Pedrosa et al., 2006; McCowan, 2007). Francis and Tannuri-Pianto (2012), for instance, conducted a large study of the University of Brasilia, one of the first Brazilian universities to use quotas, in order to evaluate their impact. Using a difference-in-differences approach, they find, among other results, that the quota policy raised the proportion of black students from families with lower socioeconomic status at the University of Brasilia. Moreover, using data from a university-specific entrance exam, they find that racial quotas did not reduce the pre-university effort of either applicants or

students and that admission under the racial quotas does not diminish the performance of black and brown quota students relative to white students. Vidigal (2018) uses a propensity score matching approach applied to student-level data from Brazilian universities to evaluate the impact of racial and low-income quotas on the academic performance of quota students. She finds that there is no statistically significant difference in academic performance between students admitted under the racial quota and those who had the regular admission (non-quota students). In contrast, students admitted under the low- income quota perform worse than eligible non-quota students, as the scores of the former are, on average, 14% lower. Using data from Brazil’s Higher Education Census and from the National Examination of Student Performance (ENADE) to construct a panel data set for the years from 2010 to 2015, this study aims to identify the impact of these racial and low-income quotas on the academic performance of students enrolled in public and private colleges in Brazil. Specifically, the objective is to evaluate the effect of the proportion of quota students on the academic performance of both quota and non-quota students. This paper contributes to the previous literature by using a broader data set with college-level variables and by estimating separate effects of quotas on the achievement of treated and untreated students. Another contribution is to investigate the impact of the quotas on academic performance across different groups of majors in order to evaluate possible differential effects. The identification of the impact of the quota policy rests on the assumption that, after controlling for very detailed observed student and college characteristics, along with college and state-year fixed effects, the admission under quotas in a given college in a given year is unlikely to be correlated with unobserved variables that determine student’s performance and dropout rate. The remainder of this chapter is organized as follows. Section 2 describes the context of the racial and low-income quotas and the admission process in Brazilian universities. Section 3 presents the data and some descriptive statistics. The empirical strategy, along with the identification strategy, is provided in Section 4. Section 5 presents estimation results, and Section 6 summarizes the findings and provides suggestions for future research.

4.2 Racial and Low-Income Quotas and the Admission Process in Brazilian Universities

Economic disparities between white and African descendants are prominent in Brazil, and they are even higher when considering black and brown individuals from lower income levels. In particular, there is a notable discrepancy between the wages earned by white and black or brown people in Brazil. The difference between white and black employees’ wages is, on average, 46.4%. This disparity is even greater for black women. A black woman earns, on average, a wage 32.7% lower than the wage of a white woman and 50.1% lower than the wage of a white man. Moreover, only a small share of employers is African- Brazilian, representing no more than 30% of the total (RAIS, 2013). Various policies have been put in place since the early 1990s to reduce the aforementioned inequalities. The first Brazilian affirmative action initiatives related to education, which were implemented in 1992, were programs developed by non- governmental organizations (NGOs). Such organizations offered college preparatory courses for low-income and/or African-Brazilian students. The implementation of a national affirmative action quota system occurred seven years later, in 1999, when the Special System of Quotas was established in federal public universities for students from public primary and secondary schools.20 The most recent affirmative action initiative was implemented in 2012, and by 2020 it will guarantee that 50% of enrollment spaces in federal public universities21 are set aside for eligible students who received all their upper secondary education (grades 10-12) in public schools. According to the Brazilian Ministry of Education (MEC), admissions under the new quota system also take into account both family income level and race (MEC, 2013).

20 The policy targeted students from public primary and secondary schools due to the fact that in Brazil these schools tend to have lower quality compared to private schools. As a consequence, students who enroll in public schools have, in general, lower school performance than those who enroll in private institutions (Oliveira et al., 2013; Morais and Belluzzo, 2014). The Brazilian public school system (from elementary to high school) is funded by state governments, which generally provide poor infrastructure and services due to budgetary constraints and political choices. Therefore, students from families that can afford to pay for private schools receive a higher quality education compared to those from lower socioeconomic status families. 21 In 2013, 301 universities, out of 2,391, were public institutions. From these 301 universities, 106 were federal public institutions (INEP, 2017).

The admission process in Brazilian universities is highly competitive, especially in public universities, as they are generally considered to be better than private institutions. In addition, public universities do not charge tuition or fees, which makes them even more attractive. Contrary to many countries, where the admission process uses multiple criteria, selection in most Brazilian universities is based only on scores in the vestibular exam, which is a sequence of exams that applicants are required to take (Arcidiacono, 2005; Carvalho, 2014). Although most universities use the vestibular to select students, the process is decentralized and each institution may formulate its own admission exam. Before taking the exams, students must choose a single undergraduate major of interest, and then they compete only with those who chose the same major. After 2009, however, some universities incorporated in the admission process the National High School Exam (Exame Nacional do Ensino Médio – ENEM), which evaluates the performance of students from public and private high schools. Universities may use only this exam, or may combine it with other mechanisms in the selection process. Moreover, they can use a quota system for the purpose of selecting disadvantaged students based on criteria such as income, race, and/or whether the student had his or her entire upper secondary school education (grades 10 to 12) from public schools. As in the case of vestibular, each university can adopt its own criteria for quota admissions. The eligibility requirement is, in general, to be self-declared as black, brown or indigenous in the case of racial quotas, or to have a monthly family income per capita no higher than 1.5 times the Brazilian minimum salary, which currently corresponds to 365 US dollars,22 for the low- income quotas. In some cases, the racial and low-income quotas also require students to have had their entire secondary school education from public schools. In addition to the requirement that students specify only one major during the university application process, they can also indicate whether they want to be considered for admission under quotas, when this policy is available. In most universities, students who select this alternative compete with all the other applicants, as if they had not chosen the quota option. That is, they go through the regular admission process. If their scores are below the cut-off level for regular admission, they are then considered for admission on a

22 Exchange rate in August 17, 2018: 1 Brazilian Real = 0.26 US Dollar.

quota basis. This mechanism of admission implies that students admitted under quotas inevitably have lower admission scores than those admitted under the regular process. Therefore, estimates of the effect of quotas on academic performance that compare the scores of quota and non-quota students, will generate a selection effect because quota students typically have lower performance at the time of their entry to the university. The admission process for eligible and non-eligible students in those universities is represented in Figure 4.1. Conversely, in some universities, students who specify that they want to be considered for admission under quotas compete only with applicants who have chosen that same option. If their scores are below the cut-off level for quota admission, however, they can be considered for the regular admission, competing with all other students (Figure B.1 in the Appendix). This mechanism, though, is relatively uncommon when quotas are available. In contrast to other countries, in Brazil students are explicitly told whether they were admitted under a quota system. The official communication informing them of the final result of their application specifies whether they were admitted under quotas or under the general admission process. The implementation of a quota system in Brazilian universities, in general, has no direct costs since only a percentage of the available seats are reserved for disadvantaged students. That is, there is no increase or decrease in admission slots as a consequence of the adoption of quotas. Some social or indirect costs are generated, however, when students who would have been admitted in the absence of the quota system are displaced by quota students. In most cases, the displaced students are those who have low scores in the vestibular exam (although higher than the admission scores of quota students) and, as a consequence, are likely either to enroll in another, less selective, college or to retake the exam. Regarding the benefits of the quota system, the policy can be an important instrument for reducing poverty and income inequality by providing quota students with higher quality professional training and thus an advantage when they participate in the labor market. Furthermore, the policy has increased underrepresented minority enrollment in both public and private universities (UERJ, 2015).

4.3 Data and Descriptive Statistics

In order to estimate the impact of the racial and low-income quota policy on the academic performance of students in Brazilian colleges, this study uses data from Brazil’s Higher Education Census and from the National Examination of Student Performance (Exame Nacional de Desempenho de Estudantes – ENADE) for the period 2010-2015. The examination is part of the National System of Higher Education Assessment (Sistema Nacional de Avaliação da Educação Superior – SINAES), and it evaluates the performance of undergraduate senior students from public and private colleges. The exam is mandatory for a random selection of senior students from selected majors, covering the standard syllabi of the majors in which students are enrolled, and it takes place at the end of the academic year (typically in December). Every three years ENADE examines the same undergraduate majors from six different fields. Therefore, all majors offered in the country are divided into three groups and students from each of them are evaluated every three years. It is possible, however, that the same field is evaluated in two or three consecutive years, because each field has several different majors, and some of these majors could be evaluated in different years than other majors. The Higher Education Census has been conducted every year since 1995. It collects information on all public and private colleges, along with professor and student characteristics, which more recently include the proportion of students admitted under racial and low-income quotas. More specifically, since 2004 has provided information on students’ admission process, but this was limited to public institutions until 2008. Most importantly, data on admission under low-income quotas are available only starting in 2009, when the questionnaire was modified. These restrictions on the Higher Education Census imply that, although desirable, it is not possible to determine exactly when each institution adopted quotas for the first time. Even though the Higher Education Census offers information at the student level, the available data on academic performance from ENADE prevents the construction of a panel at that level. Indeed, not only does the ENADE exam evaluate different majors only every three years, on the years it is administered, it is taken by only a random selection of

senior students. Thus, this study uses data from 2010 to 201523 to construct a panel of Brazilian colleges and splits this panel into three subsamples to account for differences across groups. These subsamples and their respective fields are presented in Table 4.1. It is important to stress that, although the ENADE exam data set provides student level data on scores, it does not provide individual characteristics, including race. Thus, all college and senior student variables come from the Higher Education Census and are aggregated up to the college level to match the ENADE data. The main implication is that it is not possible to distinguish the ENADE test scores of white and non-white students or even quota and non-quota students. Columns 2 and 3 of Table 4.2 present the number of colleges in Brazil’s Higher Education Census, and the number of colleges for which there exist panel data for the period 2010-2015. In 2015 there were 2,359 colleges, of which 2,053 (87%) have data for all years from 2010 to 2015. Table 4.2 also shows the total number of colleges with ENADE test score panel data for each of the groups evaluated. In the first subsample, group A, there are 684 colleges with panel data. In groups B and C, there are 1,200 and 1,485 colleges with panel data, respectively. Table 4.3 presents the number of senior college students admitted under racial quotas in public and private colleges. Most racial quota students are enrolled in public colleges and they represent only 0.74% of all students in 2015. The proportion is higher, however, when restricted to the subgroup of students in colleges where the racial quota is available. It is important to highlight that, despite the low number of senior college students admitted under racial quotas, over the period 2010 to 2015 it increased by 512% in private colleges and by 152% in public colleges. Table 4.4 shows the number of senior college students who are admitted under the low-income quota in public and private colleges. Similarly to what is observed for the racial quota, most low-income quota students are enrolled in public colleges. The proportion of low-income quota students in relation to all senior college students, however, is much

23 While there is information on low-income quotas admission since 2009, the responses on quota admission in that year are somewhat suspicious. It is apparently highly mismeasured as they have very large numbers compared to the following years, besides the fact that responses appeared to be switched for public and private colleges.

lower. For instance, in 2015, only 0.2% of all senior college students were low-income quota students. This proportion is 10 times higher, though, when restricted only to students in colleges with low-income quotas, 2.1%. In order to estimate the impact of racial and low-income quotas separately, two types of colleges can be distinguished: colleges with at least one student admitted under the racial quota and colleges with at least one student admitted under the low-income quota. It is possible, however, that the same college has both racial and low-income quota students. This means that, if a college has at least one racial quota student and at least one low-income quota student, it is included for both types of colleges. Institutions that had quota students in the first year, but at a later point in time did not have any quota students, were excluded from the sample.24 The number of public and private colleges with at least one senior quota student for the period 2010-2015 is presented in Table 4.5. Table 4.5 shows that only a few institutions have quota students. For example, in 2015 only 5.1% of all colleges had at least one racial quota student. This proportion is even lower for low-income quotas. In that same year, only 3.1% of all colleges had at least one low-income quota student. Since there is no financial incentive to adopt these types of quotas, most colleges decide not to have their admission based on that. For public institutions, only public federal universities are required by 2020 to set aside 50% of their enrollments for students who have received all their upper secondary education in public schools and have low family income. Therefore, for most institutions the decision of whether to adopt quotas is mainly based on social pressure from the academic community. Tables 4.6 and 4.7 present the average ENADE test scores for colleges from groups A, B, and C for racial and low-income quotas, respectively. The exam is based on a 0-100 scale, although the average scores tend to be very low, regardless of the group of majors, as senior students do not have any incentive to achieve a high performance in the exam since their scores neither influence their grades nor appear in their transcripts. The ENADE

24 For Group A, four (ten) colleges that had at least one racial (low-income) quota student in 2010 and none in 2013, were dropped, which constitutes 2.0% of the original sample; for Group B, eight (eight) colleges that had at least one racial (low-income) quota student in 2011 and no quota student in 2014, were excluded, which is equivalent to 1.3% of the original sample; for Group C, three (six) colleges were dropped, since they had racial (low-income) quota students in 2012, but none in 2015. This exclusion is 0.6% of the original sample.

test scores are higher for colleges that have at least one racial quota student, as can be seen in Table 4.6. Table 4.7 also shows that colleges from all groups with at least one low- income quota students have higher ENADE test scores. It is important to mention, though, that these differences in scores between colleges with and without quota students are not very informative, since students with better academic background can also be considered for admission under quotas and some public colleges, which tend to have better quality than private colleges, are required to have quotas. Therefore, differences in performance due to the quotas can be explained only if changes over time are also considered. Finally, summary statistics for all student and college variables, separately for colleges with racial and low-income quotas, can be found in Tables B.1 and B.2 of the Appendix. These variables are type of college (public or private), proportion of professors with a graduate degree, proportion of baccalaureate students, proportion of students who have evening classes, proportion of racial and low-income quota students, proportion of students with additional activity, proportion of students in each field of science, age, gender, and race.

4.4 Empirical Strategy

This section describes the empirical strategy used to estimate the effects of racial and low- income quotas on the academic performance of students in public and private colleges in Brazil. The identification of the impact of the quota policy assumes that, after controlling for college and student observed characteristics, college fixed effects, and state-year fixed effects, the proportion of quota students in a given college is unlikely to be correlated with unobserved variables that determine the academic performance, measured by the ENADE test scores. The estimation and identification strategies, along with data limitations, are presented below.

4.4.1 Estimation and Identification Strategies

The strategy used to estimate the impact of quotas on the academic performance relies on a panel data approach with fixed effects. The impact of the proportion of quota students can be estimated by the following student-level equation:

C C ̅ !"G$ = V S"G$ + ( *G$ + XYG$̅ + [\G$ + RY"G$ + -\"G$ + 5"G$ (4.1)

where !"G$ is the ENADE test score of a senior student i in college c at time t; ,"G$ is a 25 vector of senior student characteristics, such as gender, race, field of study, etc.; *G$ is a vector of college variables that include the proportion of professors with graduate degree and type of college (if public or private); YG$̅ is the proportion of senior students admitted ̅ under the racial quota; \G$ represents the proportion of senior students admitted under the low-income quota; Y"G$ and \"G$ are dummy variables indicating whether a student was admitted under a racial or low-income quota, respectively; and 5"G$ is an error term with mean zero. If data at the student level were available, Equation (4.1) could be estimated such that X would measure the spillover effect of the proportion of senior students admitted under the racial quota; [ would estimate the spillover effect of the proportion of low- income senior quota students in colleges with low-income quotas; and R and - would measure the impact of being admitted under the racial or low-income quotas, respectively. Three things about the specification of Equation (4.1) deserve comment. First, it is important to highlight the inclusion of the race variable as one of the student characteristics. On one hand, including race dummy variables would underestimate the effect of quotas since part of the effect may reflect changes in the racial composition of the students due to the quota policy. On the other hand, the inclusion of race variables may control for selection

25 The inclusion of the field of study variable could rise a concern about possible collinearity between that covariate and the quota variables, since the quotas could change the proportions of students across fields of study. However, that does not seem likely to happen as, in a given college, the same proportion of reserved seats is applied to all majors and, consequently, to all fields. Thus, all fields are equally affected by seat reservation established by the quota policy.

effect of quotas and so both the coefficient on the dummy indicating quota admission and the coefficient on the proportion of quota students are the actual impacts on the academic performance. That selection effect refers to the fact that the quotas (racial and low-income) bring some students into a college while displacing other students. To better illustrate this effect, consider the example of a college that has adopted racial quotas and admitted lower performing black students who displaced an equal number of higher performing white students (but these would be the lowest performing white students). Thus, the average quality of the black students would go down at that quota college while the average quality of the white students would go up. Although there is a change in the academic quality of those racial groups, it is not caused by the quota affecting any student’s individual academic performance, but purely because of the selection process implied by the quotas. Hence, in order to investigate differential effects due to race composition, the impact of quotas will be estimated with and without race variables.26 A second issue is also related to selection effect of quotas. Because the quota admission process generally implies that lower performing students in the vestibular exam displace higher performing students, one would expect that the impact of quotas on academic achievement would be negative, implying negative signs for X, [, R, and -. The ENADE exam, however, is taken by students in the last year of their undergraduate program, that is, four or five years after the admission process. Thus, in the case of this study, one may not necessarily expect the academic performance gap observed at the time of admission to last until students graduate, as quota students may strive for better educational achievement throughout their years in college. Finally, a possible strategy to deal with the aforementioned quota selection issue, would be the estimation of Equation (4.1) separately by race or by type of admission process. However, as explained in Section 3, that is not feasible since the ENADE data do not allow matching of individual student scores with student characteristics.

26 To measure the influence of the race coefficients, the impact of the racial and low-income quotas on the proportion of racial groups enrolled is also estimated (Table B.3). In general, the impact of the quotas on racial composition is statistically insignificant. The only exceptions are the impacts for black students in groups A and C (low-income quota) and group B (racial quota), and for yellow students in group C (both racial and low-income quotas). The estimated effects are relatively small, though, and no significant bias would be expected, since the proportions of black and yellow students are very small across groups of majors (less than 5.5% for black students and less than 2.6% for yellow students).

Unfortunately, the Higher Education Census and the ENADE available data prevent the construction of a panel data set for students. Therefore, the impact of racial and low-income quotas is estimated by aggregating Equation (4.1) up to the college level:

C C ̅ !]G$ = V ,^G$ + ( *^G$ + (X + R)YG$̅ + ([ + -)\G$ + 5G$̅ (4.2)

where the coefficient on YG$̅ measures the effect of the proportion of racial quota students, which includes both the spillover effect of quota admissions (X) and the direct effect of being admitted under the racial quota (R). Unfortunately, neither effect can be identified separately since only (X + R) can be estimated, so one cannot assess the signs of X and R (although the sign of X + R implies that either X or R, or both, have the same sign). ̅ Similarly, the coefficient on \G$ estimates the impact of the proportion of low-income quota students, which also includes the spillover effect of the low-income quota admission ([) and the direct effect of being admitted under that quota (-). Likewise, these effects cannot be identified separately. Thus, the estimated impacts of the proportion of low-income quota students on academic performance from Equation (4.2) include both the direct and the spillover effects.

If data were available for all variables in ,^G$ and *^G$, OLS would provide unbiased estimates of all parameters in Equation (4.2). This assumption is very likely to be violated, though, since many student and college variables are unobserved, such as student innate ability, social stigma, unobserved college incentives and motivation. Since these variables are not observed, they become part of 5G$̅ . Moreover, if they are correlated with observed

,^G$ and *^G$, then the error term will also be correlated with ^,G$ and *^G$, which would result in omitted variable bias for estimates of the impact of racial and low-income quotas. Therefore, to minimize this bias in the estimated impacts, Equation (4.2) adds as controls college and state-year fixed effects:

C C ̅ !]G$ = V ,^G$ + ( *^G$ + (X + R)YG$̅ + ([ + -)\G$ + _G + `a$ + 5G$̅ (4.2’)

where _G is a time invariant college fixed effect and `a$ is a college invariant state-year fixed effect (where j denotes state and t indicates time period). Thus, the identifying assumption is that after controlling for all observed student and college characteristics, along with college and state-year fixed effects, the proportion of senior quota students is unlikely to be correlated with the remaining error term which now includes unobserved elements in the S and C vectors. Finally, note that Equation (4.2’) is estimated using clustered standard errors at the college level in all specifications. The strategy used in this study to identify the impact of racial and low-income quotas on academic performance is designed to minimize the potential sources of endogeneity bias. This paper addresses the issue of omitted variable bias by using a panel data approach with college and state-year fixed effects. With regards to reverse causality, which would arise if the ENADE test score could also affect the proportion of quota students in a specific college, there is little reason to believe that the scores of senior students in the ENADE test would have any effect on the proportion of quota students. In particular, the quotas were determined at the time of admission, while the ENADE scores occur four or five years later, so there is no mechanism by which the ENADE score could affect the quotas that occurred several years previously. Lastly, measurement error in the proportion of quota students variables is unlikely to be an issue since colleges have no incentives to misreport the proportion of senior students that were admitted under quotas in the Higher Education Census. Measurement errors in the other student and college variables may exist but, unfortunately, little can be done to correct any potential bias. Even if those variables have measurement error, it should not be a concern since they are not the variables of interest.

4.5 Results

This section presents the estimates of the impacts of racial and low-income quotas on the academic performance of quota and non-quota students. For each group of majors, three different specifications of Equation (4.2’) are estimated. The first one does not include college and student variables. The second contains all college and student characteristics,

except race. And the third specification includes all college and student variables, including race. All specifications control for college and state-year fixed effects, and present robust standard errors clustered at the college level. Table 4.8 shows the regression estimates of the impact of the proportion of quota students on the ENADE test for groups A, B, and C for all public and private colleges. In interpreting the coefficients on the proportion of quota students variables, note that the explanatory variables are being reported in decimal (from 0 to 1) instead of percentage (from 0 to 100) units, and that the ENADE score variable is in a 0-100 scale. Consequently, the estimated coefficients on the proportions of racial and low-income quota students can be interpreted as changes in the ENADE score (in units) associated with one percentage point increase in the percentage of quota students, then the coefficients in Table 4.8 must be multiplied by 0.01. In general, the coefficients are statistically insignificant for all groups of majors. Moreover, they are different and vary in sign across groups, which indicates that the effect of a higher proportion of quota students on test scores depends on the majors that are evaluated. In the first specification, which does not include college and student characteristics, the estimates indicate that the presence of racial and low-income quota students does not affect the ENADE scores in all groups of majors. When college and student characteristics, except race, are included in the analysis (specification 2), most of the coefficients keep the signs of the impact, but change in magnitude. The only exception is for the effect of the proportion of low-income quota students in group B. The impact for this group becomes positive as college and student variables are incorporated, but it remains statistically insignificant. To investigate whether the coefficients are driven by changes in the racial composition of the students due to the quota policy, race variables that measure the proportion of students from each race category are included in specification 3. The race variables have little effect on the magnitude of the impacts and, in five out of six cases, keep the signs for both racial and low-income quotas, the exception being the sign change for the proportion of low-income quota students in group A. Most importantly, all of the coefficients on the proportion of quota students remain statistically insignificant. Considering the same specification, the results show that the only group for which the estimated impacts are positive for both the proportion of racial quota students

and low-income quota students is Group B. Nevertheless, these coefficients are statistically insignificant. Not only are the estimated impacts of the proportion of quota students statistically insignificant, they would not be large effects even if they were significant. For instance, for Group A, a one percentage point increase in the proportion of racial quota students would reduce ENADE exam scores in 0.02 points. Since the coefficients on the proportion of quota students include both the spillover effect of college quota admissions and the direct effect of being admitted under quotas, it is not possible to separately identify whether a student is positively or negatively affected by being a quota student. Similarly, it is not possible to individually identify whether there is any spillover effect of having a higher proportion of quota students on the academic performance of both quota and non-quota students. Even so, the overall results in Table 4.8 indicate that, regardless of the group of majors, there are no statistically significant effects of having higher proportions of racial or low-income quota students. These results for racial quota are consistant with the findings of Vidigal (2018), who found no evidence that the racial quota policy in Brazilian universities has any impact on test scores. In order to investigate heterogeneity in impacts by the type of college, estimates of Equation (4.2’) were performed only for public colleges.27 Table 4.9 provides the estimates of the impacts of racial and low-income quotas on the academic performance considering only public institutions. As in Table 4.8, three specifications are presented, the third one including all control variables. The coefficients on the proportion of racial quota students and on the proportion of low-income quota students are quite different across the specifications, which is expected since college and student variables control for observed characteristics that may affect academic performance. For majors from groups A and B, the results in the third specification indicate that higher proportions of racial quota students positively affect the academic outcomes. However, the coefficients are statistically insignificant. Even if they were significant, it would not be possible to determine whether there are spillover or direct effects. The estimates also indicate that, in all those groups of

27 Although desirable, it is not possible to estimate the impact of the proportion of quota students separately for private colleges since just a very few private institutions have at least one quota senior student, especially for racial quota. For instance, only one private college had at least one racial quota senior student in 2010, as shown in Table 4.5. Thus, due to the lack of variability on the proportion of private institutions with quotas, estimation of the impact of the quotas on the ENADE scores for that type of college cannot be performed.

majors, the academic performance of senior students is negatively affected by the proportion of low-income quota students, but as in the case of racial quotas, all the coefficients are statistically insignificant.

4.6 Conclusion

This study evaluates the impact of racial and low-income quotas on the academic performance of senior students from Brazilian colleges. Data from the Higher Education Census and information on academic performance from the National Examination of Student Performance (ENADE) for the period 2010-2015 were used to estimate the impact of the quota policy on the academic performance of three different groups of majors. While there are some concerns about the adoption of the racial quota in Brazilian universities, the results indicate that the proportion of racial quota students seems to have had no statistically significant impact on the academic performance of quota and non-quota students. Similarly, the proportion of low-income quota students does not affect the scores of these students on the ENADE test. Finally, when the analysis is restricted to public colleges, the estimates also suggest that both the proportion of racial quota students and the proportion of low-income quota students have no statistically significant impact on the academic performance of quota and non-quota students, regardless of the groups of majors. The overall findings of this study provide important evidence of the impacts of affirmative action in public and private Brazilian colleges. The main strength of this paper lies in the identification of differential effects of racial and low-income quotas for students across different groups of majors. Despite that contribution, there is a need for further systematic investigation of the impact of quotas over time, especially to identify the direct effect of the quota policy on quota students. Therefore, more research is needed in order to understand the mechanisms through which affirmative action, such as racial and low- income quotas, affect academic performance and generates externalities.

Table 4.1 – Definition of Subsamples and Corresponding Fields

Group Years Fields

- Agricultural Sciences; - Engineering and Technology; A 2010 and 2013 - Medical and Health Sciences*; - Social Sciences. - Engineering and Technology*; - Humanities; B 2011 and 2014 - Medical and Health Sciences; - Natural Sciences*; - Social Sciences.

- Engineering and Technology; C 2012 and 2015 - Humanities; - Social Sciences*.

* Fields with most majors within the group.

Table 4.2 – Number of Colleges in Brazil’s Higher Education Census, 2010-2015 Total number Colleges with Colleges with ENADE test score panel data Years of colleges panel data Group A Group B Group C 2010 2,373 2,373 771 - - 2011 2,361 2,278 - 1,356 - 2012 2,412 2,235 - - 1,627 2013 2,386 2,162 684 - - 2014 2,363 2,095 - 1,200 - 2015 2,359 2,053 - - 1,485 Notes: Group A (among the 684 colleges, 119 are public and 565 are private); Group B (from the 1,200 colleges, 173 are public and 1,027 are private); Group C (among the 1,485 colleges, 144 are public and 1,341 are private).

Table 4.3 – Number of Senior College Students under Racial Quota: Enrolled Students in Traditional (Non-Distance) Public and Private Learning, 2010-2015 Type of college Total of Total students in Total quota % colleges with % Years students Public Private students racial quotas (a) (b) (a/b) (c) (a/c) 2010 18,993 26 19,019 5,292,130 0.36 310,145 6.13 2011 24,035 117 24,152 5,609,199 0.43 345,410 6.99 2012 25,224 133 25,357 5,783,488 0.44 375,910 6.75 2013 42,234 267 42,501 6,029,722 0.70 637,380 6.67 2014 47,397 169 47,566 6,367,524 0.75 669,562 7.10 2015 47,824 159 47,983 6,502,344 0.74 1,006,476 4.77

Table 4.4 – Number of Senior College Students under Low-Income Quota: Enrolled Students in Traditional (Non-Distance) Public and Private Learning, 2010-2015 Type of college Total students in Total of Total colleges with quota % % Years students low-income Public Private students quotas (a) (b) (a/b) (c) (a/c) 2010 10,484 2,042 12,526 5,292,130 0.24 124,122 10.09 2011 1,702 1,328 3,030 5,609,199 0.05 86,094 3.52 2012 5,843 1,574 7,417 5,783,488 0.13 83,341 8.90 2013 10,508 1,201 11,709 6,029,722 0.19 230,955 5.07 2014 8,854 1,597 10,451 6,367,524 0.16 334,503 3.12 2015 11,738 1,559 13,297 6,502,344 0.20 633,346 2.10

Table 4.5 – Number of Colleges with at Least One Quota Senior Student, 2010-2015 Racial quota Low-income quota Years Type of college % of all Type of college % of all Total Total Public Private colleges Public Private colleges 2010 26 1 27 1.3 4 18 22 1.1 2011 26 4 30 1.5 3 16 19 0.9 2012 34 3 37 1.8 6 11 17 0.8 2013 61 5 66 3.2 14 14 28 1.4 2014 64 5 69 3.4 22 14 36 1.8 2015 101 4 105 5.1 49 15 64 3.1

Table 4.6 – ENADE Test Scores for Colleges from Groups A, B, and C, Racial Quota College has any racial quota student Years All colleges Yes No Difference Panel 1 – Colleges from Group A 2010 44.4 51.9 44.2 7.7*** (6.4) (3.5) (6.3) 2013 45.5 52.1 45.1 6.9*** (6.0) (4.3) (5.9)

Panel 2 – Colleges from Group B 2011 41.6 42.1 41.6 0.5 (6.6) (8.1) (6.6) 2014 43.6 46.4 43.5 2.9*** (5.5) (5.4) (5.4)

Panel 3 – Colleges from Group C 2012 38.1 44.8 37.9 6.8*** (5.3) (4.8) (5.3) 2015 44.0 51.4 43.7 7.7*** (5.8) (5.6) (5.5) Notes: Standard deviations in parentheses. ***Significant at 1% level.

Table 4.7 – ENADE Test Scores for Colleges from Groups A, B, and C, Low-Income Quota

College has any low-income quota student Years All colleges Yes No Difference Panel 1 – Colleges from Group A 2010 44.4 51.7 44.3 7.3 (6.4) (1.5) (6.4) 2013 45.5 52.5 45.4 7.1*** (6.0) (5.4) (5.9)

Panel 2 – Colleges from Group B 2011 41.6 48.7 41.6 7.1** (6.6) (4.1) (6.6) 2013 43.6 45.9 43.5 2.4** (5.5) (4.1) (5.5)

Panel 3 – Colleges from Group C 2012 38.1 41.0 38.1 2.9 (5.3) (7.8) (5.3) 2015 44.0 51.0 43.8 7.1*** (5.7) (6.8) (5.6) Notes: Standard deviations in parentheses. ***Significant at 1% level. **Significant at 5% level.

Table 4.8 – Estimates of the Impact of Racial and Low-Income Quotas on Academic Performance on the ENADE Test: Results for Groups A, B, and C

Group A Group B Group C Variables (1) (2) (3) (1) (2) (3) (1) (2) (3) Racial quota Proportion of quota students -5.088 -2.754 -2.110 12.136 13.187 13.960 -1.859 -3.034 -2.908 (6.796) (6.527) (6.411) (9.124) (8.933) (9.110) (4.794) (4.374) (4.268) Low-income quota Proportion of quota students 2.133 1.859 -0.657 -1.010 0.618 0.437 -1.542 -0.756 -0.237 (6.027) (5.165) (7.654) (2.626) (2.575) (2.633) (7.074) (6.250) (6.093)

Observations 1,322 1,322 1,322 2,326 2,326 2,326 2,886 2,886 2,886 R-squared 0.832 0.841 0.842 0.733 0.759 0.759 0.877 0.884 0.884

Control variables College and student variables No Yes Yes No Yes Yes No Yes Yes Race variables No No Yes No No Yes No No Yes College fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes State-year fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Notes: Robust standard errors clustered at the college level. ***Significant at the 1% level. **Significant at the 5% level. *Significant at the 10% level. Group A (2010 and 2013; omitted field: Medical and Health Sciences); Group B (2011 and 2014; omitted field: Social Sciences); Group C (2012 and 2015; omitted field: Social Sciences). Each estimated coefficient must be interpreted as the effect of a one percentage point increase in the proportion of quota students on the ENADE score (in 0-100 units), then such coefficients must be multiplied by 0.01.

Table 4.9 – Estimates of the Impact of Racial and Low-Income Quotas on Academic Performance on the ENADE Test for Public Colleges: Results for Groups A, B, and C

Group A Group B Group C Variables (1) (2) (3) (1) (2) (3) (1) (2) (3) Racial quota Proportion of quota students -1.438 2.967 4.818 -0.969 4.050 3.884 -2.525 0.601 -0.210 (8.377) (9.136) (10.345) (12.042) (9.442) (9.344) (4.197) (3.834) (3.466) Low-income quota Proportion of quota students 5.605 2.243 -16.834 2.725 -2.063 -5.660 -10.463 -10.540 -8.438 (4.611) (7.189) (19.667) (15.018) (15.590) (17.371) (6.557) (8.235) (10.775)

Observations 214 214 214 318 318 318 262 262 262 R-squared 0.859 0.874 0.880 0.872 0.906 0.908 0.912 0.924 0.925

Non-eligible student Eligible student

Want to be considered for admission under quotas

No Yes

Vestibular score

Admitted Rejected

Considered under quotas

Admitted Rejected under quotas

Non-eligible student and eligible student who does not want to be considered under quotas

Eligible student who wants to be considered under quotas

Figure 4.1 – Stylized Admission Process for Most Brazilian Universities

Chapter 5

Conclusion

This dissertation is comprised of three essays in the economics of education. In particular, the chapters in this dissertation investigate the effects of educational programs in primary and secondary schools in Brazil on academic outcomes, and the impact of recent quota policies in Brazilian universities on their students’ academic performance. Chapter 2 evaluated the impact of Brazil’s Multifunctional Resources Classroom Inclusion Program on the academic outcomes of disabled and nondisabled students. The results showed that, in general, the Brazilian inclusion program benefits students with special educational needs or disabilities, especially those enrolled in grades 6-9 and 10-12, with no negative spillover effects onto non-disabled students. The results provide further evidence that inclusive education may generate positive impacts for disabled students with no negative externalities on the academic outcomes of non-disabled students. Chapter 3 investigated the impact of the Mais Educação Extended School Day Program on the academic outcomes of students in primary and secondary schools in Brazil. The overall results of this study suggest that the Brazilian extended school day program has been effective at reducing dropout rates, but at the same time, it has increased repetition rates in all grade levels. Therefore, while the findings of this paper indicate that the Brazilian extend school day program provides positive effects for some outcomes, especially for students in grades 6-9, they raise some concerns about the mechanisms through which the program has been implemented, since unexpected effects are observed for most outcomes evaluated. Finally, Chapter 4 examined the adoption of racial and low-income quotas in Brazilian universities. It evaluated the impact of these quotas on the academic performance of senior college students. The study found that both the proportion of racial quota students and the proportion on low-income quota students have no statistically significant effect on the academic performance of quota and non-quota students.

References

Ainscow, M. (2005). “Developing Inclusive Education Systems: What are the Levers for Change?” Journal of Educational Change, 6(2), 109-124.

Ainscow, M., César, M. (2006). “Inclusive Education Ten Tears after Salamanca: Setting the Agenda.” European Journal of Psychology of Education, 21(3), 231-238.

Aquino, J. M. (2011). “A Ampliação da Jornada Escolar Melhora o Desempenho Acadêmico dos Estudantes? Uma Avaliação do Programa Escola em Tempo Integral da Rede Pública do Estado de São Paulo.” Ph.D. Thesis, University of Sao Paulo.

Arcidiacono, P. (2005). “Affirmative Action in Higher Education: How Do Admission and Financial Aid Rules Affect Future Earnings?” Econometrica, 73 (5), 1477-1524.

Arzola González, M. P. (2010). “Impacto de la Jornada Escolar Completa en el Desempeño de los Alumnos, medido con la Evolución en sus Pruebas SIMCE.” Master Thesis, Pontificia Universidad Catolica de Chile.

Bailey, S. R., Peria, M. (2010). “Racial Quotas and the Culture War in Brazilian Academia.” Sociology Compass. 4 (8), 592-604.

Bellei, C. (2009). “Does Lengthening the School Day Increase Students’ Academic Achievement? Results from a Natural Experiment in Chile.” Economics of Education Review, 28, 629-640.

Berthelon, M., Kruger, D. I. (2011). “Risky Behavior among Youth: Incapacitation Effects of School on Adolescent Motherhood and Crime in Chile.” Journal of Public Economics, 95 (1-2), 41-53.

Bertrand, M., Hanna, R., Mullainathan, S. (2010). “Affirmative Action in Education: Evidence from Engineering College Admissions in India.” Jounal of Public Economics, 94 (1-2), 16-29.

Blume, G. H., Long, M. C. (2014). “Changes in Levels of Affirmative Action in College Admissions in Response to Statewide Bans and Judicial Rulings.” Educational Evaluation and Policy Analysis, 20 (10), 215-229.

Cardoso, C. B. (2008). “Efeitos da Política de Cotas na Universidade de Brasilia: Uma Análise do Rendimento e da Evasão.” (Master's thesis), University of Brasilia, Brasilia, Brazil.

100

Carvalho, J. A. M., C., Wood, C. H., Andrade, F. C. D. (2004). “Estimating the Stability of Census-Based Racial/Ethnic Classifications: The Case of Brazil.” Population Studies, 58 (3), 331-343.

Carvalho, J. R. (2014). “College Choice Allocation Mechanisms: Structural Estimates and Counterfactuals.” Working paper [506], Toulouse School of Economics.

Cerdan-Infantes, P., Vermeersch, C. (2007). “More Time Is Better: An Evaluation of the Full-Time School Program in Uruguay.” World Bank Policy Research Working Paper, 4167.

Childs, P., Stromquist, N. P. (2014). “Academic and Diversity Consequences of Affirmative Action in Brazil.” Compare: A Journal of Comparative and International Education, 27, 1-22.

Cole, C. M., Waldron, N., Majd, M. (2004). “Academic Progress of Students across Inclusive and Traditional Settings.” Mental Retardation, 42, 136-144.

Cooper, H., et al. (2010). “Effects of Full-Day Kindergarten on Academic Achievement and Social Development.” Review of Educational Research, 80 (1), 34-70.

Cunha, E. (2006). “Sistema Universal e Sistema de Cotas para Negros na Universidade de Brasília: Um Estudo de Desempenho.” (Master’s Thesis), Department of Education, University of Brasilia.

Demeris, H., Childs, R. A., Jordan, A. (2007). “The Influence of Students with Special Needs Included in Grade-3 Classrooms on the Large-Scale Achievement Scores of Students without Special Needs.” Canadian Journal of Education, 30, 609-627.

Desai, S., Kulkami, V. (2008). “Changing Educational Inequalities in India in the Context of Affirmative Action.” Demography, 45(2), 245-70.

Dickson, L. (2006). “Does Ending Affirmative Action in College Admissions Lower the Percent of Minority Students Applying to College?” Economics of Education Review, 25, 109-119.

Dyson, A., Farrell, P., Polat, F., Hutcheson, G., Gallannaugh, F. (2004). Inclusion and Pupil Achievement. London: DfES.

Dyson, A. (2014). “A Response to Göransson and Nilholm.” European Journal of Special Needs Education, 29(3), 281-282.

Epple, D., Romano, R. and Sieg, H. (2008). “Diversity and Affirmative Action in Higher Education.” Journal of Public Economic Theory, 10(4), 475-501.

101

Farrell, P. (2000). “The Impact of Research on Developments in Inclusive Education.” International Journal of Inclusive Education, 4(2), 153-162.

Farrell, P., Dyson, A., Polat, F., Hutcheson, G., Gallannaugh, F. (2007). “Inclusion and Achievement in Mainstream Schools.” European Journal of Special Needs Education, 22(2), 131-145.

Ferman, B., Assunçâo, J. (2005). “Affirmative Action in University Admissions and High School Students' Proficiency.” Working paper, Department of Economics, Pontifical Catholic University of Rio de Janeiro (PUCRio).

Fletcher, J. (2010). “Spillover Effects of Inclusion of Classmates with Emotional Problems on Test Scores in Early Elementary School.” Journal of Policy Analysis and Management, 29(1), 69- 83.

Francis, A. M., Tannuri-Pianto, M. (2012). “Using Brazil’s Racial Continuum to Examine the Short-Term Effects of Affirmative Action in Higher Education.” Journal of Human Resources, 47 (3), 754-84.

Francis, A. M., Tannuri-Pianto, M. (2013). “Endogenous Race in Brazil: Affirmative Action and the Construction of Racial Identity Among Young Adults.” Economic Development and Cultural Change, 61 (4), 731-753.

Friesen, J., Hickey, R., Krauth, B. (2010). “Disabled Peers and Academic Achievement.” Education Finance and Policy, 5(3), 317-348.

Frisancho, V., Krishna, K. (2016). “Affirmative action in higher education in India: targeting, catch up, and mismatch.” Higher Education, 71 (5), 611-649.

Glewwe, P., Kassouf, A. (2012). “The Impact of the Bolsa Escola/Familia Conditional Cash Transfer Program on Enrollment, Dropout Rates and Grade Promotion in Brazil.” Journal of Development Economics, 97, 505-517.

Göransson, K., Nilholm, C. (2014). “Conceptual Diversities and Empirical Shortcomings: A Critical Analysis of Research on Inclusive Education.” European Journal of Special Needs Education, 29(3), 265-280.

Gottfried, M. A. (2014). “Classmates with Disabilities and Students' Noncognitive Outcomes.” Educational Evaluation and Policy Analysis, 36(1), 20-43.

Grodsky, E., Kalogrides, D. (2008). “The Declining Significance of Race in College Admissions Decisions.” American Journal of Education, 115, 1-33.

Hanushek, E. A., Kain, J. F., Rivkin, S. G. (2002). “Inferring Program Effects for Special Populations: Does Special Education Raise Achievement for Students with Disabilities?” The Review of Economics and Statistics, 84(4), 584-599.

102

Hincapie, D. (2016). “Do Longer School Days Improve Student Achievement? Evidence from Colombia.” IDB Working Paper, 679.

Hinrichs, P. (2012). “The Effects of Affirmative Action Bans on College Enrollment, Educational Attainment, and the Demographic Composition of Universities.” Review of Economics and Statistics, 94, 712-722.

Holland, P., Alfaro, P., Evans, D.K. (2015). “Extending the School Day in Latin America and the Caribbean.” World Bank Policy Research Working Paper, 7309.

Howell, J. S. (2010). “Assessing the Impact of Eliminating Affirmative Action in Higher Education.” Journal of Labor Economics, 28, 113-166.

Huber, K. D., Rosenfeld, J. G., Fiorello, C. A. (2001). “The Differential Impact of Inclusion and Inclusive Practices on High, Average and Low-Achieving General Education Students.” Psychology in the Schools, 38, 497–504.

IBGE (2017). Brazilian Institute of Geography and Statistics. Brazilian Census. Retrieved from: http://www.ibge.gov.br/home/estatistica/populacao/censo2010/default.shtm (accessed July 10, 2017).

INEP (2017). Inep Data, “Educação Superior.” Available at: http://inep.gov.br/web/guest/inep-data (accessed July 17, 2017).

Itau Social Foundation. (2015). “Relatório de Avaliação Econômica e Estudos Quanlitativos. O Programa Mais Educação.” Fundação Itau Social, São Paulo.

Jacob, B., Lefgren, L. (2003). “Are Idle Hands the Devil’s Workshop? Incapacitation, Concentration, and Juvenile Crime.” American Economic Review, 93(5), 1560- 1577.

Kalambouka, A., Farrell, P., Dyson, A., Kaplan, I. (2007). “The impact of Placing Pupils with Special Educational Needs in Mainstream Schools on the Achievement of their Peers.” Educational Research, 49, 365-382.

Kanter et al., (2014). “The Right to Inclusive Education Under International Law: Following Italy's Lead.” Journal of International Special Needs Education, 17(1), 21-32.

Kristoffersen, J. H. G., Krægpøth, M. V., Nielsen, H. S., Simonsen, M. (2015). “Disruptive School Peers and Student Outcomes.” Economics of Education Review, 45, 1-13.

Lindsay, G. (2007). “Educational Psychology and the Effectiveness of Inclusive Education/Mainstreaming.” British Journal of Educational Psychology, 77(1), 1- 24.

103

Link, C. R., Mulligan, J. G. (1986). “The Merits of a Longer School Day.” Economics of Education Review, 5(4), 373-381.

Llach, J., Adrogué, C., Gigaglia, M. (2009). “Do Longer School Days Have Enduring Educational, Occupational, or Income Effects? A Natural Experiment in Bueno Aires, Argentina.” Economía, 10(1), 1-43.

Lochner, L., Moretti, E. (2004). “The Effect of Education on Crime: Evidence from Prison Inmates, Arrests, and Self-Reports.” American Economic Review, 94(1), 155-189.

MEC (2013). Brazilian Ministry of Education, “Lei de Cotas nas Universidades Federais.” Available at: http://portal.mec.gov.br/cotas/legislacao.html (accessed January 28, 2016).

MEC (2015). Brazilian Ministry of Education, “Principais Indicadores da Educação de Pessoas com Deficiência.” Available at: http://portal.mec.gov.br/index.php?option=com_docman&view=download&alias =17655-secadi-principais-indicadores-da-educacao- especial&category_slug=junho-2015-pdf&Itemid=30192 (accessed August 20, 2018).

Mendes, K. D. (2011). “O Impacto do Programa Mais Educação no Desempenho dos Alunos da Rede Pública Brasileira.” São Paulo: University of São Paulo.

McCowan, T. (2007). “Expansion Without Equity: An Analysis of Current Policy on Access to Higher Education in Brazil.” Higher Education, 53(5), 579-598.

Morais, A. G. E., Belluzzo, W. (2014). “O Diferencial de Desempenho Escolar entre Escolas Públicas e Privadas no Brasil.” Nova Economia, 24 (2), 409-430.

Myklebust, J. O. (2007). “Diverging Paths in Upper Secondary Education: Competence Attainment among Students with Special Educational Needs.” International Journal of Inclusive Education, 11, 215-231.

Oliveira, P. R., Belluzzo, W. and Pazello, E. T. (2013). “The Public–Private Test Score Gap in Brazil.” Economics of Education Review, 35, 120-133.

Patall, E. A., Cooper, H., Allen, A. B. (2010). “Extending the School Day or School Year: A Systematic Review of Research (1985–2009).” Review of Educational Research, 80(3), 401-436.

Pedrosa, R. H. L. et al. (2006). “Educational and Socioeconomic Background of Undergraduates and Academic Performance: Consequences for Affirmative Action Programs at a Brazilian Research University.” Available at:

104

https://www1.oecd.org/site/imhe2006bis/37245034.pdf (accessed March 20, 2016).

Pires, T., Urzua, S. (2015). “Longer School Days, Better Outcomes?” Unpublished working paper. Available at: https://sites.google.com/site/tiagomcccpires (accessed August 7, 2017).

PNAD (2014). Pesquisa Nacional por Amostra de Domicílios (National Household Survey). IBGE, Rio de Janeiro.

RAIS (2013). Relação Anual de Informações Sociais, “Remuneração Média do Trabalho.” Available at: http://www.rais.gov.br/ (accessed February 10, 2016).

Rathburn, A. (2010). “Making the Most of Extra Time: Relationship Between Full-Day Kindergarten Instructional Environments and Reading Achievement.” American Institute for Research, Research Brief.

Ruijs, N. M., Peetsma, T. T. (2009). “Effects of Inclusion on Students with and without Special Educational Needs Reviewed.” Educational Research Review, 4(2), 67-79.

Ruijs, N. (2017). “The Impact of Special Needs Students on Classmate Performance.” Economics of Education Review, 58, 15-31.

Soukakou, E. P. (2012). “Measuring Quality in Inclusive Preschool Classrooms: Development and Validation of the Inclusive Classroom Profile (ICP).” Early Childhood Research Quarterly, 27, 478-488.

Tienda, M., Alón, S., Niu, S. X. (2008). “Affirmative Action and the Texas Top 10% Admissions Law: Balancing Equity and Access to Higher Education.” Working paper, Princeton University Office of Population Research.

UERJ (2015). Universidade do Estado do Rio de Janeiro, “Avaliação Qualitativa dos Dados sobre Desempenho Acadêmico: Relatório ano 2011.” Available at: http://www.caiac.uerj.br/rel.pdf (accessed January 11, 2016).

Valenzuela, J. P. (2005). “Partial Evaluation of a Big Reform in the Chilean Education System: From a Half Day to a Full Day Schooling.” Ph.D. Thesis, University of Michigan.

Vidigal, C. B. R. (2018). “Racial and Low-Income Quotas in Brazilian Universities: Impact on Academic Performance.” Journal of Economic Studies, 45 (1), 156-176.

Xerxenevsky, L. L. (2012). “Programa Mais Educação: Avaliação do Impacto da Educação Integral no Desempenho de Alunos no Rio Grande do Sul.” Master Thesis, Pontifica Universidade Catolica do Rio Grande do Sul (PUCRS).

105

World Bank (2013). World Data Bank, “School Enrollment: Tertiary.” Available at: http://databank.worldbank.org/ddp/home.do (accessed December 18, 2015).

106

Appendix A

Table A.1 – Average Enrollment of Disabled and Non-Disabled Students in Treated and Never Treated Schools with Grades 1-5, 6-9 and 10-12, 2007-2015 Schools with grades 1-5 All schools Treated schools Never treated schools Years Non-disabled Non-disabled Non-disabled Disabled students Disabled students Disabled students students students students 2007 158.8 1.5 237.9 3.1 123.6 0.8 (180.8) (5.3) (202.7) (7.6) (157.9) (3.7) 2008 154.5 1.6 231.5 3.5 120.3 0.8 (175.4) (5.3) (196.9) (7.6) (153.1) (3.6) 2009 149.6 2.0 224.4 4.1 116.5 1.0 (169.1) (5.5) (190.6) (7.8) (147.0) (3.7) 2010 143.5 2.5 214.8 5.4 112.0 1.2 (162.1) (6.1) (182.4) (8.4) (141.2) (4.1) 2011 139.2 2.9 208.9 6.2 108.6 1.4 (157.5) (6.2) (177.0) (8.6) (137.4) (4.0) 2012 135.2 3.1 203.1 6.7 105.6 1.5 (153.9) (6.4) (172.7) (8.8) (134.7) (4.0) 2013 132.9 3.2 200.0 6.9 103.9 1.6 (152.1) (6.4) (170.5) (9.0) (133.3) (4.0) 2014 132.2 3.4 197.9 7.1 104.0 1.8 (152.9) (6.5) (170.0) (9.1) (135.4) (4.1) 2015 130.1 3.5 194.6 7.3 102.5 1.9 (151.4) (6.5) (167.9) (8.9) (134.6) (4.2) Schools with grades 6-9 All schools Treated schools Never treated schools Years Non-disabled Non-disabled Non-disabled Disabled students Disabled students Disabled students students students students 2007 263.0 0.9 294.0 1.3 243.0 0.7 (248.4) (3.5) (251.7) (4.4) (244.2) (2.8) 2008 257.6 1.1 287.5 1.6 238.4 0.8 (243.0) (3.5) (245.7) (4.3) (239.3) (2.8) 2009 251.8 1.4 282.3 2.0 232.2 1.1 (237.1) (4.1) (238.6) (4.8) (234.0) (3.4) 2010 245.7 1.9 274.7 2.8 227.1 1.3 (230.0) (4.6) (230.0) (5.5) (228.2) (3.9) 2011 238.6 2.3 264.3 3.4 222.2 1.5 (223.1) (4.7) (222.5) (5.7) (221.9) (3.8) 2012 230.4 2.8 255.5 4.3 214.6 1.8 (215.6) (5.2) (215.6) (6.3) (214.1) (4.1) 2013 222.2 3.1 246.2 4.7 207.0 2.1 (208.0) (5.5) (208.9) (6.5) (206.1) (4.4) 2014 213.2 3.5 237.7 5.3 197.7 2.3 (196.9) (5.7) (200.8) (6.8) (192.8) (4.5) 2015 206.7 3.9 233.5 6.0 189.8 2.6 (188.8) (6.0) (194.7) (7.2) (183.0) (4.6) Schools with grades 10-12 All schools Treated schools Never treated schools Years Non-disabled Non-disabled Non-disabled Disabled students Disabled students Disabled students students students students 2007 369.1 0.5 420.4 0.7 344.7 0.4 (359.7) (2.4) (396.8) (3.6) (338.0) (1.6) 2008 354.6 0.6 400.6 0.9 332.9 0.5 (342.1) (2.1) (372.8) (3.1) (324.3) (1.5) 2009 347.5 0.8 395.8 1.2 324.9 0.6 (333.7) (2.3) (365.1) (3.2) (315.4) (1.6) 2010 341.6 1.0 385.0 1.5 321.2 0.7 (323.7) (3.1) (354.8) (4.6) (306.1) (1.9) 2011 337.5 1.2 377.5 1.8 318.9 0.9 (318.1) (2.6) (346.4) (3.4) (302.2) (2.0) 2012 333.6 1.6 374.0 2.4 314.8 1.2 (311.9) (3.1) (339.2) (4.1) (296.5) (2.4) 2013 327.9 1.8 369.0 2.7 308.8 1.4 (306.0) (3.3) (333.6) (4.2) (290.4) (2.7) 2014 324.1 2.2 365.2 3.2 305.1 1.7 (302.3) (3.6) (329.7) (4.5) (286.7) (3.0) 2015 311.1 2.5 350.5 3.7 292.8 1.9 (291.1) (4.0) (320.0) (4.9) (274.8) (3.3) Note: Standard deviations in parentheses.

107

Appendix B

Table B.1 – Descriptive Statistics for Groups A, B, and C: Racial Quota

Group A Group B Group C Variables All Any racial quota student All Any racial quota student All Any racial quota student colleges Yes No Diff colleges Yes No Diff colleges Yes No Diff ENADE score 44.9 52.0 44.6 7.4*** 42.6 45.1 42.5 2.6*** 41.1 49.6 40.8 8.9*** (6.2) (4.0) (6.1) (6.1) (6.5) (6.1) (6.3) (6.2) (6.1) College variables Public college 0.173 1.000 0.140 0.860*** 0.149 0.971 0.124 0.847*** 0.099 0.957 0.071 0.886*** (0.379) (0.000) (0.348) (0.356) (0.169) (0.329) (0.298) (0.204) (0.256) Professor with graduate degree 0.612 0.832 0.603 0.229*** 0.616 0.834 0.610 0.224*** 0.622 0.833 0.615 0.218*** (0.199) (0.140) (0.196) (0.200) (0.125) (0.198) (0.203) (0.113) (0.202) Student variables Baccalaureate students 0.716 0.632 0.719 -0.087** 0.589 0.546 0.590 -0.044 0.718 0.521 0.725 -0.204*** (0.265) (0.224) (0.266) (0.325) (0.297) (0.326) (0.290) (0.319) (0.287) Evening class 0.651 0.321 0.665 -0.344*** 0.785 0.369 0.798 -0.428*** 0.812 0.397 0.826 -0.429*** (0.302) (0.164) (0.299) (0.261) (0.204) (0.252) (0.242) (0.211) (0.231) Racial quota students 0.002 0.055 0.000 0.055*** 0.002 0.061 0.000 0.061*** 0.002 0.071 0.000 0.071*** (0.018) (0.075) (0.000) (0.017) (0.078) (0.000) (0.025) (0.120) (0.000) Low-income quota students 0.001 0.006 0.001 0.005** 0.001 0.009 0.001 0.007** 0.002 0.008 0.001 0.007* (0.017) (0.044) (0.015) (0.025) (0.041) (0.024) (0.034) (0.038) (0.034) Additional activity 0.182 0.178 0.182 -0.004 0.171 0.193 0.171 0.022 0.187 0.174 0.187 -0.013 (0.300) (0.169) (0.304) (0.310) (0.205) (0.313) (0.327) (0.208) (0.330) Natural Sciences (field 1) 0.000 0.000 0.000 - 0.222 0.322 0.219 0.104*** 0.000 0.000 0.000 - (0.000) (0.000) (0.000) (0.282) (0.184) (0.283) (0.000) (0.000) (0.000) Engineering and Technology (field 2) 0.120 0.065 0.122 -0.057 0.268 0.312 0.266 0.045 0.150 0.279 0.146 0.134*** (0.288) (0.237) (0.290) (0.353) (0.295) (0.355) (0.283) (0.422) (0.277) Medical and Health Sciences (field 3) 0.664 0.527 0.669 -0.142*** 0.091 0.040 0.092 -0.053** 0.000 0.000 0.000 - (0.383) (0.301) (0.385) (0.209) (0.044) (0.212) (0.000) (0.000) (0.000) Agricultural Sciences (field 4) 0.106 0.340 0.096 0.244*** 0.000 0.000 0.000 - 0.000 0.000 0.000 - (0.238) (0.307) (0.230) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) Social Sciences (field 5) 0.111 0.068 0.112 -0.044 0.293 0.149 0.297 -0.148*** 0.838 0.695 0.843 -0.147*** (0.254) (0.084) (0.259) (0.355) (0.111) (0.358) (0.288) (0.412) (0.282) Humanities (field 6) 0.000 0.000 0.000 - 0.114 0.178 0.112 0.066** 0.012 0.025 0.012 0.013* (0.000) (0.000) (0.000) (0.216) (0.129) (0.218) (0.069) (0.046) (0.070) Age 28.1 26.4 28.1 -1.7*** 28.8 27.0 28.9 -1.9*** 28.8 27.0 28.8 -1.9*** (2.5) (1.4) (2.5) (2.9) (2.4) (2.9) (2.8) (2.0) (2.8) Female 0.619 0.594 0.620 -0.026 0.590 0.563 0.590 -0.027 0.593 0.540 0.595 -0.055*** (0.118) (0.061) (0.120) (0.172) (0.103) (0.174) (0.121) (0.118) (0.121) White 0.282 0.227 0.284 -0.057 0.329 0.329 0.329 0.000 0.354 0.379 0.353 0.026 (0.310) (0.279) (0.311) (0.313) (0.314) (0.313) (0.316) (0.295) (0.317) Black 0.023 0.030 0.023 0.007 0.031 0.047 0.030 0.017*** 0.031 0.052 0.031 0.021***

108

(0.051) (0.035) (0.051) (0.049) (0.041) (0.049) (0.050) (0.046) (0.050) Brown 0.129 0.125 0.129 -0.004 0.153 0.134 0.153 -0.019 0.163 0.150 0.163 -0.013 (0.197) (0.203) (0.197) (0.207) (0.166) (0.208) (0.209) (0.155) (0.210) Yellow 0.008 0.009 0.008 0.001 0.010 0.008 0.010 -0.002 0.011 0.011 0.011 0.000 (0.020) (0.016) (0.020) (0.032) (0.012) (0.033) (0.031) (0.013) (0.032) Indigenous 0.001 0.003 0.001 0.002 0.002 0.003 0.002 0.001 0.002 0.003 0.002 0.001 (0.006) (0.008) (0.006) (0.016) (0.005) (0.016) (0.010) (0.008) (0.010) Non-declared skin color 0.556 0.606 0.554 0.052 0.476 0.479 0.476 0.004 0.440 0.404 0.441 -0.037 (0.397) (0.373) (0.397) (0.392) (0.362) (0.393) (0.387) (0.375) (0.387) Number of observations 1,340 51 1,289 2,370 69 2,301 2,952 93 2,859 Notes: Standard deviations in parentheses. ***Significant at 1% level. **Significant at 5% level. *Significant at 10% level.

109

Table B.2 – Descriptive Statistics for Groups A, B, and C: Low-Income Quota

Group A Group B Group C Variables All Any low-income quota student All Any low-income quota student All Any low-income quota student colleges Yes No Diff colleges Yes No Diff colleges Yes No Diff ENADE score 44.9 52.4 44.8 7.6*** 42.6 46.2 42.5 3.7*** 41.1 49.8 40.9 8.9*** (6.2) (5.1) (6.2) (6.1) (4.1) (6.1) (6.3) (7.6) (6.2) College variables Public college 0.173 0.875 0.165 0.710*** 0.149 0.676 0.141 0.536*** 0.099 0.800 0.086 0.714*** (0.379) (0.342) (0.371) (0.356) (0.475) (0.348) (0.298) (0.404) (0.281) Professor with graduate degree 0.612 0.819 0.610 0.209*** 0.616 0.794 0.614 0.180*** 0.622 0.817 0.618 0.198*** (0.199) (0.150) (0.198) (0.200) (0.161) (0.199) (0.203) (0.165) (0.202) Student variables Baccalaureate students 0.716 0.561 0.717 -0.156** 0.589 0.490 0.591 -0.101* 0.718 0.627 0.720 -0.093** (0.265) (0.273) (0.264) (0.325) (0.296) (0.325) (0.290) (0.249) (0.291) Evening class 0.651 0.394 0.655 -0.260*** 0.785 0.536 0.789 -0.253*** 0.812 0.471 0.818 -0.348*** (0.302) (0.157) (0.302) (0.261) (0.343) (0.258) (0.242) (0.274) (0.237) Racial quota students 0.002 0.017 0.002 0.015*** 0.002 0.025 0.001 0.023*** 0.002 0.028 0.002 0.027*** (0.018) (0.044) (0.017) (0.017) (0.071) (0.014) (0.025) (0.061) (0.023) Low-income quota students 0.001 0.084 0.000 0.084*** 0.001 0.096 0.000 0.096*** 0.002 0.099 0.000 0.099*** (0.017) (0.133) (0.000) (0.025) (0.187) (0.000) (0.034) (0.243) (0.000) Additional activity 0.182 0.113 0.182 -0.069 0.171 0.221 0.171 0.050 0.187 0.252 0.186 0.067 (0.300) (0.107) (0.302) (0.310) (0.302) (0.310) (0.327) (0.282) (0.327) Natural Sciences (field 1) 0.000 0.000 0.000 - 0.222 0.247 0.221 0.025 0.000 0.000 0.000 - (0.000) (0.000) (0.000) (0.282) (0.194) (0.283) (0.000) (0.000) (0.000) Engineering and Technology (field 2) 0.120 0.240 0.119 0.121* 0.268 0.252 0.268 -0.016 0.150 0.131 0.150 -0.019 (0.288) (0.400) (0.287) (0.353) (0.230) (0.355) (0.283) (0.279) (0.283) Medical and Health Sciences (field 3) 0.664 0.501 0.666 -0.165* 0.091 0.112 0.091 0.022 0.000 0.000 0.000 - (0.383) (0.325) (0.384) (0.209) (0.246) (0.208) (0.000) (0.000) (0.000) Agricultural Sciences (field 4) 0.106 0.205 0.104 0.100* 0.000 0.000 0.000 - 0.000 0.000 0.000 - (0.238) (0.260) (0.237) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) Social Sciences (field 5) 0.111 0.055 0.111 -0.057 0.293 0.222 0.294 -0.072 0.838 0.838 0.838 0.000 (0.254) (0.091) (0.256) (0.355) (0.231) (0.356) (0.288) (0.286) (0.288) Humanities (field 6) 0.000 0.000 0.000 - 0.114 0.168 0.113 0.054 0.012 0.031 0.012 0.019* (0.000) (0.000) (0.000) (0.216) (0.180) (0.217) (0.069) (0.071) (0.069) Age 28.1 27.4 28.1 -0.7 28.8 28.0 28.8 -0.8* 28.8 27.7 28.8 -1.1*** (2.5) (2.3) (2.5) (2.9) (2.5) (2.9) (2.8) (2.6) (2.8) Female 0.619 0.585 0.620 -0.035 0.590 0.603 0.589 0.014 0.593 0.585 0.593 -0.008 (0.118) (0.111) (0.118) (0.172) (0.087) (0.173) (0.121) (0.079) (0.122) White 0.282 0.263 0.282 -0.019 0.329 0.359 0.328 0.030 0.354 0.393 0.353 -0.039 (0.310) (0.235) (0.311) (0.313) (0.305) (0.313) (0.316) (0.251) (0.317) Black 0.023 0.036 0.023 0.012 0.031 0.036 0.030 0.005 0.031 0.039 0.031 0.007 (0.051) (0.040) (0.051) (0.049) (0.035) (0.049) (0.050) (0.035) (0.050) Brown 0.129 0.201 0.128 0.073 0.153 0.176 0.152 0.023 0.163 0.197 0.162 0.035 (0.197) (0.281) (0.196) (0.207) (0.211) (0.207) (0.209) (0.197) (0.209) Yellow 0.008 0.009 0.008 0.001 0.010 0.023 0.010 0.013** 0.011 0.026 0.011 0.015***

110

(0.020) (0.016) (0.020) (0.032) (0.104) (0.030) (0.031) (0.096) (0.029) Indigenous 0.001 0.002 0.001 0.001 0.002 0.001 0.002 -0.001 0.002 0.002 0.002 0.001 (0.006) (0.003) (0.006) (0.016) (0.003) (0.016) (0.010) (0.003) (0.010) Non-declared skin color 0.556 0.489 0.557 -0.068 0.476 0.405 0.477 -0.072 0.440 0.344 0.441 -0.098* (0.397) (0.365) (0.397) (0.392) (0.368) (0.393) (0.387) (0.328) (0.387) Number of observations 1,340 16 1,324 2,370 34 2,336 2,952 50 2,902 Notes: Standard deviations in parentheses. ***Significant at 1% level. **Significant at 5% level. *Significant at 10% level.

111

Table B.3 – Estimates of the Impact of Racial and Low-Income Quotas on Racial Composition: Results for Groups A, B, and C Group A

Black Brown Yellow Indigenous Non-declared White Racial quota 0.018 0.528 0.086 -0.001 -0.419 -0.211 (0.072) (0.406) (0.115) (0.007) (0.847) (0.540) Low-income quota -0.137*** 1.247 -0.527 0.001 -0.358 -0.225 (0.024) (0.892) (0.363) (0.003) (1.287) (0.151)

Observations 1,322 1,322 1,322 1,322 1,322 1,322 R-squared 0.894 0.869 0.718 0.616 0.873 0.893

Group B Black Brown Yellow Indigenous Non-declared White Racial quota 0.157** 0.032 -0.044 0.002 -0.100 -0.047 (0.075) (0.348) (0.038) (0.008) (0.510) (0.332) Low-income quota -0.025 0.186 -0.007 0.001 -0.204 0.048 (0.028) (0.209) (0.007) (0.003) (0.134) (0.096)

Observations 2,326 2,326 2,326 2,326 2,326 2,326 R-squared 0.759 0.849 0.656 0.581 0.831 0.851

Group C Black Brown Yellow Indigenous Non-declared White Racial quota -0.049 0.042 -0.037*** 0.011 -0.139 0.172 (0.056) (0.096) (0.011) (0.012) (0.187) (0.154) Low-income quota 0.141*** 0.658 0.041** -0.005 -0.998 0.163 (0.054) (0.659) (0.019) (0.005) (0.900) (0.283)

Observations 2,886 2,886 2,886 2,886 2,886 2,886 R-squared 0.748 0.831 0.681 0.537 0.803 0.844

College and student variables Yes Yes Yes Yes Yes Yes College fixed effects Yes Yes Yes Yes Yes Yes State-year fixed effects Yes Yes Yes Yes Yes Yes Notes: Robust standard errors clustered at the college level. ***Significant at the 1% level. **Significant at the 5% level. *Significant at the 10% level. Group A (2010 and 2013; omitted field: Medical and Health Sciences); Group B (2011 and 2014; omitted field: Social Sciences); Group C (2012 and 2015; omitted field: Social Sciences).

112

Non-eligible student Eligible student

Want to be considered for admission under quotas

No Yes

Vestibular score Vestibular score

Admitted Rejected

Rejected Admitted under quotas

Non-eligible student and eligible student who does not want to be considered under quotas

Eligible student who wants to be considered under quotas

Eligible student who wants to be considered under quotas and who was rejected under the quota process (student’s score can be considered under the regular process)

Figure B.1 – Stylized Admission Process for Some Brazilian Universities

113