ABSTRACT
 Title of dissertation: CONTRACTING OUT PUBLIC SCHOOLS
 AND ACADEMIC PERFORMANCE:
 EVIDENCE FROM COLOMBIA
 Juan D Bonilla-Angel
 Doctor of Philosophy, 2011
 Dissertation directed by: Professor Judith K. Hellerstein
 Department of Economics
 Contracting out public schools to private institutions is an instrument for
 reforming public education as it may facilitate academic innovation and im-
 prove student academic performance through higher school accountability and
 autonomy. The degree of autonomy that different providers have may vary sub-
 stantially depending on the contractual and institutional arrangements they are
 subject to. In principle, contractual differences should generate different sets
 of incentives for providers that may ultimately affect student academic perfor-
 mance. One can expect, for example, that programs with limited achievement
 accountability rules might invest sub-optimally in resources aiming at improv-
 ing the academic performance of their students.
 In this dissertation, I evaluate short- and longer-run achievement effects
 of the Colegios en Concesi?n (CEC) program, a large-scale initiative implemented
 in 2000 in Bogota, Colombia, which contracted out the administration of some
 traditional public schools (TPS) to reputed, not-for-profit private schools and
universities. This program allows participating schools to operate outside public
 schools? collective bargaining provisions in return for being accountable, among
 other things, for the academic performance of their students in the ICFES test, a
 high-stakes college entry national standardized test.
 The major empirical challenge in studies of alternative school models is
 selection bias. Students who attend CEC schools may differ in a number of
 ways from public school students. To overcome potential selection bias of CEC
 attendance, I exploit variation in distance from a student?s residence to the clos-
 est CEC institution as an instrument for CEC attendance. While distance may
 in theory be correlated with unobservable characteristics of students, I demon-
 strate using a variety of empirical strategies that this instrument is conditionally
 exogenous of unobserved determinants of academic achievement.
 I first evaluate the effects of attending a CEC school on ICFES test scores.
 Instrumental variables results indicate that CEC students exhibit important and
 significant gains in test scores on the ICFES test. That is, the two-stage least
 squares estimates obtained indicate that CEC students score 0.6 and 0.2 standard
 deviations higher in math and verbal tests, respectively, relative to TPS students.
 I provide evidence that the positive test score results of CEC attendance are not
 driven by unintended strategic responses by CEC schools such as excluding low-
 performing students from the pool of test-takers or via test specialization in the
 curriculum, or by significant differences in education inputs such as teachers?
 education, student-teacher ratios, or expenditures per student. I also provide
 suggestive auxiliary evidence that the estimated results are a consequence of an
institutional arrangement that makes CEC schools accountable for the academic
 performance of their students.
 I also evaluate whether attending a CEC school translates into longer-run
 gains in potentially more meaningful outcomes such as increasing the probabil-
 ity of investing in post-secondary schooling, attending a more selective tertiary
 institution, or being admitted in high-return academic programs. The results
 on college attendance indicate that CEC students exhibit a significantly higher
 probability of attending a higher education institution and to attend a vocational
 program relative to TPS students. Moreover, CEC students have a slightly higher
 probability of attending a selective public institution and are not more likely to
 drop out from college relative to TPS students.
 The overall results provide compelling evidence that the contractual ar-
 rangement that defines the operation of CEC schools are successful at improving
 the academic performance of their students relative to TPS.
CONTRACTING OUT PUBLIC SCHOOLS AND ACADEMIC
 PERFORMANCE: EVIDENCE FROM COLOMBIA
 by
 Juan D. Bonilla-Angel
 Dissertation submitted to the Faculty of the Graduate School of the
 University of Maryland, College Park in partial fulfillment
 of the requirements for the degree of
 Doctor of Philosophy
 2011
 Advisory Committee:
 Professor Judith K. Hellerstein, Chair
 Professor Melissa Kearney
 Professor Robert M. Schwab
 Professor Raymond Guiteras
 Professor Steven J. Klees
c Copyright by
 Juan D. Bonilla-Angel
 2011
Dedication
 Siempre me ha gustado la idea de elogiar a la dificultad con la vida misma. (As?
 tambi?n me enamor? de ti. Luchando en la corriente de un amor temeroso que se perd?a
 entre palabras ausentes). No fueron pocas las veces del discurso de motivaci?n ante las
 dificultades de estos a?os de estudio. Tan presentes. Tan copiosas. Estamos aqu? a pesar
 de haber desconfiado tanto. De esa terrible man?a de atribuirle lo propio a la suerte y
 lo impropio a s? mismos. Pero hoy no. Hoy quiero celebrar que hallamos tomado una
 de las mejores decisiones de nuestras vidas. Pese a los costos. Enormes. Pese a las
 l?grimas y a algunos momentos para el olvido. Pero juntos. Juntos, porque solo ni a
 palo. Aprendiendo a cocinar. A conocernos en otra lengua y a hacer amigos de mundos
 ajenos. A disfrutar de todo ese tiempo juntos que siempre se queda corto. Porque siempre
 me quedan faltando minutos para estar contigo. Pidi?ndole tiempo al juez al final del
 encuentro diario. A ti, mi Carito, te dedico esta disertaci?n.
 ii
Acknowledgments
 I am specially indebted to my advisor, Judy Hellerstein, for giving me sup-
 port and encouragement that I had only received before by members of my own
 family. For contributing to my education as an economist and for generously
 sharing her time and knowledge in that sweetest way of hers. For reading my
 mind in those days I was lost in translation, respecting the long silences and
 disappearances, and for looking after me in so many ways. I just hope that one
 day I can be half the advisor she was to me and indirectly pay for all I have
 received from her during these years.
 I am grateful to Raymond Guiteras, Melissa Kearney, and Bob Schwab for
 their invaluable contributions to this dissertation, for kindly questioning me on
 unresolved issues and for enthusiastically supporting me when I needed. I am
 also indebted to Abby Alpert, Mark Duggan, Teresa Fort, Carolina Gonz?lez,
 and Gabriel Lara for invaluable conversations and contributions to improve the
 contents of this dissertation and to Steven J. Klees for agreeing to participate in
 the dissertation committee. Finally, this dissertation would have never been pos-
 sible without the support and generosity of Maria Camila Uribe, Fabio S?nchez,
 and, most specially, Juli?n Mari?o. I have the greatest admiration for the work
 they do. For making things happen in a country so much in need of people of
 their qualities.
 Understandably, some may use these lines to publicly acknowledge those
 who, in one way or another, made one?s dissertation possible. For a public recog-
 nition is just morally due in these circumstances. But I may do something else
 as well. For I would like to use this space too to celebrate the little occurrences
 of the common life in the company of those with whom we spent most days
 and nights and sometimes made this dissertation happily impossible. To Abby
 Alpert, Paul Bailey, Teresa Fort, and Aaron Szott; Sergio D?az and ?lvaro La
 Parra ?los vecinos?; and Daniel Hernaiz, Gabriel Lara, and C?sar Sosa-Padilla
 ?los chicos del Chateau?. I am thankful for having shared with them "all those
 frivolous nothings which fill up the void of human life" and hope, with Carolina,
 to remain a part of their lives despite the distance.
 To my parents, I owe a powerful idea. A beautiful idea that was never
 explicitly taught, yet too present to be overlooked: that there is nothing more
 satisfactory in life than helping others ?within our own limitations. Perhaps I
 should not be surprised by how much this has defined my professional journey
 and I hope it will keep defining my research agenda. Finally, I want to thank my
 mother and my sister Cristina for believing so much in me and letting me know
 it. For their unconditional support that always keeps me hopeful that everything
 ?no matter what? is going to be alright.
 iii
CONTENTS
 1. The Effects of Contracting Out Public Schools: The Case of Colombia?s CEC
 Program . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
 1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
 1.2 The Colegios en Concesi?n (CEC) Program . . . . . . . . . . . . . . 5
 1.2.1 CEC School Evaluations . . . . . . . . . . . . . . . . . . . . . 9
 1.2.2 Related Literature . . . . . . . . . . . . . . . . . . . . . . . . 12
 1.3 Short-Run Effects of CEC attendance . . . . . . . . . . . . . . . . . . 15
 1.4 Longer-Run Effects of CEC attendance . . . . . . . . . . . . . . . . . 19
 1.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
 Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
 Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
 2. Short-run Effects of the CEC Program . . . . . . . . . . . . . . . . . . . . . . 27
 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
 2.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
 2.2.1 Public Education Sector in Bogot? . . . . . . . . . . . . . . . 29
 2.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
 2.4 Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
 2.4.1 Structural Equation . . . . . . . . . . . . . . . . . . . . . . . . 36
 2.4.2 Estimating the Causal Effect of CEC School Attendance . . 37
 2.4.3 Using Distance to the Closest CEC School as an Instrument 39
 2.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
 2.5.1 First Stage Estimates . . . . . . . . . . . . . . . . . . . . . . . 47
 2.5.2 2SLS Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . 48
 2.6 Robustness Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
 2.6.1 A Placebo Test . . . . . . . . . . . . . . . . . . . . . . . . . . 51
 2.6.2 Characterizing Treatment Effect Heterogeneity . . . . . . . . 54
 2.7 Mechanisms for Higher Academic Achievement . . . . . . . . . . . 56
 2.7.1 Strategic Responses . . . . . . . . . . . . . . . . . . . . . . . 57
 2.7.2 Alternative Mechanisms . . . . . . . . . . . . . . . . . . . . . 61
 2.7.3 A Contract for Academic Achievement . . . . . . . . . . . . 62
 2.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
 Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
Appendix A. Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
 Appendix B. Full Set of Estimation Results . . . . . . . . . . . . . . . . . . . . . 77
 Appendix C. Calculating Relative Distance to Closest CEC School . . . . . . . . . 79
 3. Longer-Run Effects of the CEC Program . . . . . . . . . . . . . . . . . . . . . 84
 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
 3.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
 3.2.1 Higher Education in Colombia . . . . . . . . . . . . . . . . . 90
 3.3 Data and Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . 94
 3.4 Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
 3.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
 3.5.1 First Stage Estimates . . . . . . . . . . . . . . . . . . . . . . . 101
 3.5.2 2SLS Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . 102
 3.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
 Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
 Appendix D. Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
 v
LIST OF TABLES
 1.1 Evaluation Methodology since 2007 . . . . . . . . . . . . . . . . . . . . 24
 2.1 Sample Exclusion Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . 67
 2.2 Descriptive Statistics by School Type . . . . . . . . . . . . . . . . . . . . 68
 2.3 Exogeneity of Relative Distance to Closest CEC . . . . . . . . . . . . . . 69
 2.4 First Stage Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
 2.5 OLS and 2SLS Results of CEC Attendance on ICFES Scores . . . . . . . . 70
 2.6 Placebo Test: Reduced Form Effects of Distance . . . . . . . . . . . . . . 71
 2.7 Treatment Effect Heterogeneity . . . . . . . . . . . . . . . . . . . . . . . 72
 2.8 School Inputs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
 2.9 First Stage Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
 2.10 Full Set of 2SLS Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
 2.11 Matched and Unmatched Samples by School Type . . . . . . . . . . . . 83
 3.1 Selected Descriptive Statistics by College Attendance . . . . . . . . . . . 109
 3.2 Descriptive Statistics for Higher Education Outcomes . . . . . . . . . . . 110
 3.3 First Stage Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
 3.4 OLS, Reduced Form, and 2SLS Estimates of CEC attendance . . . . . . . 112
 vi
LIST OF FIGURES
 1.1 Location of TPS, New TPS, and CEC Schools . . . . . . . . . . . . . . . 25
 1.2 CEC?s ICFES Performance Relative to TPS in same Locality . . . . . . . 25
 1.3 CEC Academic Performance Component 2007-2009 . . . . . . . . . . . 26
 1.4 CEC Non-Academic Indicators 2007-2009 . . . . . . . . . . . . . . . . . 26
 1.5 CEC Overall Evaluation 2007-2009 . . . . . . . . . . . . . . . . . . . . 26
 2.1 Location of CEC Schools . . . . . . . . . . . . . . . . . . . . . . . . . . 34
 2.2 Distribution of Relative Distance to Closest CEC . . . . . . . . . . . . . . 45
 2.3 Dropout and Transfer Rates by School Type . . . . . . . . . . . . . . . . 60
 2.4 UPZs with CEC Schools . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
 2.5 Non-linearity of RD on CEC Attendance Probability . . . . . . . . . . . 74
 2.6 Proportion of Students by Intervals of Actual Distance . . . . . . . . . . 75
 2.7 Location of Public and CEC Students? Residence . . . . . . . . . . . . . 75
 2.8 Commuting Patterns of 5 CEC Schools . . . . . . . . . . . . . . . . . . . 76
 2.9 Distribution of Relative Distance in 2000 and 2008 . . . . . . . . . . . . . 76
 3.1 ICFES Composite Score Distribution by School Types . . . . . . . . . . 113
 vii
1. THE EFFECTS OF CONTRACTING OUT PUBLIC SCHOOLS:
 THE CASE OF COLOMBIA?S CEC PROGRAM
 1.1. Introduction
 The question of whether private schools provide better education than
 public schools has been a long and much debated issue. This question is, for
 example, at the center of the debate over the effects on academic performance
 of vouchers, charter schools, and other programs that allow the participation of
 the private sector in the provision of public education.
 In order to improve the quality of public education, policy makers and
 researchers have examined the way in which public education is provided. Pro-
 ponents of public-private partnerships in education argue that the political insti-
 tutions that govern public schools inhibit the existence of academically effective
 organizations, as services are provided by organized labor unions that tend to
 put too much weight on the welfare of their members in detriment of education
 quality (Chubb and Moe, [1990]). Opponents, on the other hand, assert that
 due to contractual incompleteness, contractors engage in cost-saving strategies
 that affect those aspects of quality which are essentially non-contractible and
tend to avoid problem (i.e. unprofitable) students that the public authorities are
 responsible for serving. It has also been argued that private schools may have
 curriculums that do not fully reflect the social goals of education such as serving
 as an agent of socialization into a common set of moral values (Benveniste, et.
 al, [2003]).
 While the movement towards choice and decentralization in public educa-
 tion has included programs as varied as vouchers, magnet/pilot schools, and
 intra-district plans, contracted-out public schools appear to be one of the most
 replicated strategies among the reform initiatives.1 Unlike voucher programs
 ?where the control over the quality of education provided by participating
 schools is mostly left to private users? contracting out public schools is an
 attractive choice to policy makers as it allows the state to retain a fair measure
 of control over service delivery by monitoring performance, imposing penalties
 or by replacing providers in case of performance failure. To some extent, char-
 ter schools in the Unites States and in Alberta (Canada), Academies in England,
 Free schools in Sweden, and CEC schools in Bogota (Colombia) are salient exam-
 ples of this type of education provision. That is, initiatives aiming at broadening
 the decision-making autonomy of schools and promoting accountability, while
 remaining under ?admittedly varying levels of? public control.
 In the literature on public-private partnerships in education, both charter
 schools in the US and voucher programs in Latin America have received spe-
 1 Charter schools in the US are a prominent example of this type of strategies with more
 than 5500 institutions by 2010 up from 1,200 schools in 1999 ?a five-fold increase (Center for
 Education Reform, www.edreform.com last visited on May 15, 2011).
 2
cial attention. Charter schools in the United States are privately run schools,
 partially funded by the state and attended by choice. With greater flexibility,
 charter schools are able to run longer school days, have an academic curriculum
 different from public schools, and finance their activities through a mix of public
 and private funds. Also, they are able to operate outside teachers? collective bar-
 gaining provisions that constrain seniority and tenure conditions in traditional
 public schools (for a description of US charter school experiences see, for ex-
 ample, Hoxby [2009], Matthews [2009], and Wilson [2008]). Nevertheless, as I
 explain in more detail below, the evidence on charter schools in the US is rather
 mixed with positive effects being found for some schools and some grades only.
 The evidence on voucher programs in Latin America is not conclusive ei-
 ther. Using panel data for a large voucher program in Chile, Hsieh and Urquiola
 (2006) find no positive effects on mean test scores, repetition rates, or years of
 schooling. They find, however, evidence of increased sorting, as the best pub-
 lic school students moved to private schools. A second voucher program that
 has received attention in the literature is the lottery-allocated PACES program
 during the 1990s in Colombia (King, [1997]; Angrist et al, [2002]; Angrist et al,
 [2006]; Bettinger et al [2010]), one of the largest voucher school programs to have
 ever been implemented. The short- and long-run goals of this program were to
 rapidly raise secondary-school enrollment rates through private academic in-
 stitutions and to improve the education quality through school competition for
 voucher recipients (Kling et al [1998]). Angrist et al (2002) found that lotteried-in
 students were more likely to finish 8th grade ?mainly because they were less
 3
likely to repeat grades? and to score higher on achievement tests.2 Moreover,
 Angrist et al (2006) find that voucher winners were 15 to 20 percent more likely
 to graduate from high school than lotteried-out participants. Interestingly, these
 results seem to have been driven by students who switched to more expensive
 private schools but who would have attended private schools had they not re-
 ceived a school voucher. In fact, 85% of lottery losers started secondary school in
 a private institution. What these results indicate is that lottery participants did
 not come from low-income families, perhaps the group of first-order importance
 for this type of programs.
 The inconclusive evidence on the effectiveness of the private provision of
 public education has been used by some authors to question all forms of public-
 private partnerships in education such as charter school programs (Benveniste,
 et. al, [2003]; Levin, [2003]; Ravitch, [2010]). However, perhaps what the mixed
 evidence on these public-private partnerships in education suggests is that the
 debate on education provision should not be so much about privately versus
 publicly run schools but about designing the right incentives in order to induce
 public schools to be responsible for the academic performance of their students
 ?for example, by creating school accountability policies for both school admin-
 istrators and teachers.
 In this dissertation, I study the short- and longer-run causal effects of the
 Colegios en Concesi?n (henceforth, CEC) Program, an initiative implemented in
 Bogota (Colombia) in the year 2000 that aimed at improving the quality of ed-
 2 The effect on test scores is significant at a 10% confidence level.
 4
ucation imparted by public schools by contracting out some public schools to
 private operators while retaining a fair amount of control over the services pro-
 vided.
 1.2. The Colegios en Concesi?n (CEC) Program
 The ability of the Department of Education of Bogota (hereafter, SED) to
 implement policies that improve academic performance depends, to a great ex-
 tent, on its ability to direct and coordinate the actions of the complex public
 education sector in Bogot?, which consists of more than 26,000 teachers oper-
 ating within a national collective bargaining agreement. The existing institu-
 tional arrangement leaves the SED with little room to introduce major changes
 aimed at improving quality of education. First, wages are determined through
 a bargaining process between the teachers union and the national government
 ?not the local government? and are determined by education and tenure as
 opposed to performance. Second, the SED has little ability to coordinate the
 actions of traditional public schools (hereafter, TPS) through school principals,
 and principals themselves do not have social or legal instruments to exert au-
 thority and demand the fulfillment of assigned tasks to teachers. Indeed, al-
 though the SED has the power to appoint and transfer school principals and
 teachers, dismissals or other disciplinary sanctions are the responsibility of the
 Ranking Board, a committee where teachers? interests had traditionally been
 over-represented (Gandour et. al, [2000]).
 5
These institutional limitations prompted the creation of the CEC school
 program. The argument was that this program would allow the SED to be able
 to monitor and demand academic quality in some schools given the lack of high
 power incentives to encourage teacher cooperation in TPS. Thus, all the new
 public schools constructed by the SED between 1999 and 2003 were contracted
 out to private academic institutions. The SED established 15-year contracts with
 private schools and universities through a bidding process based on the superior
 academic results of their own institutions in the ICFES test, the proposed profile
 of potential teachers and yearly cost per-student. A total of 16 schools opened
 primary school grades (grades 1 to 5) in April 2000 and then added secondary
 education (grades 6 to 11) the year after. Six more schools began operating in
 March 2001, providing up to 10th grade that year and graduating their first class
 in 2002. The last 3 schools opened in 2003 for all academic grades.
 The CEC program is of interest for a number of reasons. First, to the best of
 my knowledge, this initiative is one of the largest programs of its kind to have
 been implemented by a local administration in a developing country, serving
 more than 40,000 students in all academic grades (from pre- to high-school).
 Second, the CEC program is a publicly designed program aimed at im-
 proving education quality. In fact, this program is the first education policy in
 Colombia designed to hold public schools accountable for the academic perfor-
 mance of their students. This is unlike similar public-private partnerships like
 U.S. charter schools, which have been implemented as a private response to pub-
 lic education. The Center for Education Reform, a U.S. charter advocate, states
 6
that "Virtually anyone can submit an application to open and operate a char-
 ter school: parents, educators, museums, civic groups, business leaders, service
 organizations and teachers have started schools in United States". As a result,
 charter schools in the U.S. are diverse and many are specifically created to be
 an alternative for those interested in subjects missing in traditional education:
 special arts or music programs, language, adult education, dropout prevention
 programs, and special education needs (Hoxby and Rockoff, [2004]). As a result,
 some charters may not see academic performance as their primary goal.
 An implication of the CEC program as a public policy is that all CEC
 schools are very similar along most relevant dimensions. First, the geographic
 location of each contracted out school was chosen by the Department of Educa-
 tion (SED) according to the demand for public education. As a result, all CEC
 schools were located in low-income areas of the city where excess demand for
 public education was highest. Second, the SED constructed all of these schools
 and made them similar in terms of their physical characteristics. Third, CEC
 schools must follow the same academic curriculum of TPS so that, other things
 equal, there should be no difference in academic achievement due to differential
 academic content. Fourth, CECs do not get to choose their own students but
 instead have to take those students allocated by the SED. Fifth, expenditures per
 student are practically the same as those of TPS. For the purpose of this study, a
 clear methodological advantage of having a relatively homogeneous program is
 that it reduces concerns about estimating their joint effect on academic perfor-
 mance. Also, the fact that CEC schools are similar to TPS in dimensions such as
 7
academic curriculum and per-student expenditures allows us to rule out some
 of the mechanisms behind academic performance.
 On the other hand, CEC schools are different from TPS in some important
 ways. First, CEC schools are able to hire teachers at regulated salaries on ten-
 month renewable contracts from a non-unionized pool of applicants. That is,
 unlike TPS, CEC schools have the flexibility to hire and fire teachers as needed
 on a yearly basis. In addition, the contracts establish that teachers should have at
 least a college degree in Education and have some years of relevant experience.
 Teacher wages are regulated by law so that CEC teachers earn at least what TPS
 teachers earn according to education level and experience. Interestingly, given
 the differences in job stability for public and CEC teachers, CEC school teachers
 often find it optimal to move to TPS, with some CEC schools having turnover
 rates of as much as 40% per year. Second, while many TPS provide either pri-
 mary or secondary grades, CEC schools provide all academic grades, a strategy
 that not only aims to reduce dropouts, but perhaps more importantly a strategy
 that provides incentives for long run student investments by the school. Third,
 the average CEC school has better school facilities than the average TPS as a re-
 sult of being more recently built by the SED. Nevertheless, as many as fifty new
 TPS constructed after 2004 and located in similar low-income neighborhoods
 (see Figure 1.1) have facilities comparable to CEC schools.3
 3 These new schools were not contracted out since the CEC program was phased out in 2004
 with the election of a new city mayor.
 8
1.2.1. CEC School Evaluations
 Perhaps the most important difference between CEC and TPS schools is
 that the former are subject to performance evaluations by organizations con-
 tracted by SED for this purpose. Importantly, only after 2007 a standardized
 and more comprehensive annual evaluation was implemented ?including ad-
 ditional indicators in terms of a national standardized test known as ICFES. The
 ICFES test4 is a high-stakes, multiple-subject, Colombian national standardized
 test taken by the vast majority of high school students in their last year of high
 school and serves as the main criterion to determine admission into higher edu-
 cation institutions.
 Before 2007, the evaluations conducted did not aim at becoming a regu-
 lar instrument for tracking CEC performance partly since these were conducted
 by different organizations with different methodological approaches. Starting
 in 2007, the current evaluation has used quantitative and qualitative indica-
 tors on multiple outcomes related to academic quality such as student academic
 achievement, teachers? education and experience, participation of students, par-
 ents and teachers in school bodies, quantity and quality of food provision, and
 transfer and dropout rates. As shown in Table 1.1, the current methodology is
 divided into components and each component is in turn divided into indicators.
 Note that both indicators and components are weighted in order to generate
 a composite grade for each CEC school. An important feature of the current
 4 The current name of this test is SABER 11
 9
evaluation method is that it allows for individual school comparisons over time.
 Regarding the academic achievement component (with 20% of the weight),
 the current evaluation looks at each CEC school average performance on the
 ICFES test using a weighted average of their overall and relative performance.
 The overall measure ?receiving 40% of the weight within the academic perfor-
 mance component? ranks all private and public schools in the country accord-
 ing to their average performance in the ICFES test in 7 achievement levels. Since
 2004, the average CEC school has been placed in the fourth level with very little
 variation over time. Whereas the average CEC school in the period 2004-2007
 was placed in level 4.44, this figure was 4.59 for the 2008-2010 period. As ref-
 erence, the 2008 average CEC ICFES score difference between any two adjacent
 levels is 0.2 standard deviations. This may be an indication of how difficult it
 may be to move to a higher achievement level. This may be partly explained by
 the fact that this indicator does not control for SES differences between schools.
 Thus, in order to move to a higher achievement level, schools need to improve
 their performance relative to all schools including private schools.
 The contracts with CEC providers recognize that overall performance is an
 imperfect measure of school quality as academic achievement is highly corre-
 lated with students? background. Indeed, the contracts state that the SED can
 unilaterally break the contract whenever a school?s average results in the ICFES
 test are worse than the performance of the average TPS in their locality for two
 consecutive years.5 Consequently, the current evaluation takes this indicator into
 5 Bogota is administratively divided in 20 localities and there are CEC schools in 10 of them.
 10
account. In Figure 1.2, I show differences in means between each CEC school
 and the average TPS in their locality for both 2007 and 2008. For both years,
 the horizontal axis is sorted in ascending order of the 2008 CEC-TPS ICFES dif-
 ference to facilitate visualization. Note, first, that for 2008, seven CEC schools
 have scores statistically higher than the average TPS in their locality. Also, for
 the remaining schools, the difference in means is not statistically different from
 zero at a 95% confidence level. Moreover, although four CEC schools in 2007
 did perform below the average TPS in their locality, they managed to improve
 their performance in 2008. What these results suggest is that defining academic
 achievement relative to TPS in the same locality as stated in the provision con-
 tracts is a condition that not all CEC schools easily achieve.
 In Figures 1.3 to 1.5, I present the evolution of the current evaluation for
 the period 2007 to 2009. Figure 1.3 shows the results in percentage terms on
 the academic achievement component for all CEC schools for these years, where
 the horizontal axis represents each one of the CEC schools. As shown, the vast
 majority of schools have consistently maintained or improved their academic
 performance in these three years. However, that many schools are scoring just
 60% in this component suggests there may be some room for improvement in
 this area. Moreover, note that the positive trend in time is not only occurring
 in the academic component. The results are indeed similar in all components,
 including ?as shown in Figure 1.4? the one related to having low drop-out
 rates or the quality and quantity of the food provided to students. The overall
 grade for each CEC school in Figure 1.5 just confirms the pattern. I interpret
 11
these results as evidence that CEC schools are taking the current evaluation
 seriously, partly to comply with contractual demands, but also because they
 may be interested in showing merits in case they apply for contract renewal in
 the years to come.6
 1.2.2. Related Literature
 Charter Schools
 The CEC program has features in common with some charter school experiences
 in the US and therefore serves as a framework to compare the estimated effects
 of CEC attendance with those of charter school programs. Both CEC and US
 charter programs are supported by public funds, receive students interested in
 public education, cannot select their own students, and are not subject to teacher
 collective bargaining agreements. However, unlike CEC schools, charter schools
 are responsible for finding their own students, may target students interested in
 non-standard education programs (e.g., arts, music, language, special education,
 etc.), and are partially funded by private donors.7 CEC schools, on the other
 hand, are required to take the students allocated by the SED, have the same
 academic curriculum TPS have, and are fully funded by public resources.
 Most of the recent studies on charter schools use admission lotteries to
 circumvent the concerns about non-random selection. There is mixed evidence
 in the lottery-based literature of charter school attendance on academic achieve-
 6 The first contracts will expire in 2014.
 7 For example, the Harlem Children Zone Charter in New York City spends 25% more than
 the median per-pupil expenditure in New York State public schools (Dobbie and Fryer [2009]).
 12
ment, usually with positive effects being found for some grades only. The over-
 all evidence indicates that charter school students score from 0.3 to 0.4 standard
 deviations higher than lottery losers in math tests per year of treatment and
 around 0.1 to 0.2 standard deviations higher in reading tests per year (Angrist
 et. al [2010]; Abdulkadiroglu et. al, [2009]; Dobbie and Fryer, [2009]; Hoxby and
 Murarka, [2009]; Hoxby and Rockoff, [2004]). It is important to keep in mind,
 however, that charter schools that are forced to use lotteries to determine admis-
 sion may be the most effective charter schools ?those that parents find more
 appealing and continue to operate. As a result, the reported estimated effects
 may not speak of the overall effect of charter school attendance.
 In turn, studies that have used panel data to account for unobserved time
 invariant student characteristics in the US find that charters students experience
 poor test score growth in the first years after enrollment but later recover from
 this initial disruption, with no significant gain in test scores from charter school
 attendance after some time (Saas, [2004]; Bifulco and Ladd, [2004]; Hanushek
 et. al, [2007]; Booker et. al [2007]). There is some evidence, however, of positive
 charter effects on school attendance and lower disciplinary sanctions (Imberman,
 [2009b]).
 Test-Based Accountability
 A second group of educational programs that have similarities to the CEC school
 program are those that impose sanctions on low-performing public schools.8
 8 In the US, for example, the No Child Left Behind Act of 2001 (NCLB) established penalties
 for failing to meet performance standards in all states. In particular, the NCLB requires states
 13
The existing literature has traditionally focused on school responses to increased
 accountability, in particular, to what extent higher academic results are driven
 by unintended strategic responses by affected schools. An example of a posi-
 tive response to increased accountability is Chiang (2009). Using a regression
 discontinuity design and data from public schools in Florida, this study finds
 that threatened elementary schools raised math and reading test scores in 0.12
 and 0.1 standard deviations, respectively, by raising spending on technology,
 curricular improvements, and teacher training.
 Although schools face a powerful incentive to raise their academic perfor-
 mance ratings by reforming the inputs and processes of the educational produc-
 tion, most of the existing evidence suggests that schools opt to strategically re-
 spond to the accountability system in ways not intended by policy makers. That
 is, by raising test scores without contributing to students? knowledge and skills.
 Examples of these responses are to reclassify low-performing students into spe-
 cial education, or to impose longer disciplinary sanctions near testing dates so
 that these students do not affect school average test results (Deere and Strayer,
 [2001]; Figlio and Getzler, [2002]; Cullen and Reback, [2002]; Jacob, [2005]; Figlio,
 [2006]). Moreover, there is also evidence that higher test score results have been
 driven by increasing test-preparation at the expense of low-stakes subjects such
 as the natural sciences (Koretz and Barron, [1998]; Jacob, [2005]), or through
 teacher cheating (Jacob and Levitt, [2003]). This literature stresses the impor-
 to test students in grades 3 to 8 each year and to judge school performance on the basis of these
 test scores.
 14
tance of assessing potential strategic responses by CEC schools when interpret-
 ing the estimated effects of CEC attendance on test scores.
 CEC School Program
 There have been two previous evaluations of the CEC school program in Bo-
 got?. First, Sarmiento et al. (2005) surveyed 22 CEC schools as well as to 10 TPS
 with similar characteristics. That study concludes that CEC schools have a better
 academic environment, have more autonomy to manage both teachers and re-
 sources, and are more exposed to supervision by the SED. Barrera (2005), using
 a propensity score matching strategy, finds that CEC schools have lower dropout
 rates compared to similar TPS. He also finds evidence that students from CEC
 school students score 1 to 2 more points in the ICFES test relative to TPS stu-
 dents. Both of these studies are early evaluations of the effects of the program,
 which may be problematic given that students graduating from CEC schools in
 the first years of the program spent most of their elementary and middle school
 years at TPS.
 1.3. Short-Run Effects of CEC attendance
 In this dissertation, I start by examining the effects of attending a CEC
 school on test scores using data for all TPS and CEC students in Bogot? who
 took the ICFES test in the second semester of 2008.
 The major empirical challenge in studies of alternative school models is
 selection bias. Students who attend CEC schools may differ in a number of
 15
ways from public school students. Some of this bias is eliminated by control-
 ling for observable student and family characteristics, but the possibility of bias
 remains from unobserved variables that determine school choice such as a stu-
 dent?s ability, motivation or some other features of family background. For
 example, higher unobserved motivation among parents of CEC schools can er-
 roneously lead to the conclusion that CEC schools raise academic achievement
 relative to traditional public schools even if there is no difference in the qual-
 ity of these schools. Alternatively, estimates of the effects of CEC schools on
 achievement could be downward biased if CEC families have unobserved lower
 socioeconomic conditions relative to families of students attending traditional
 public schools.
 Therefore, establishing the causal effects of CEC school attendance on aca-
 demic achievement requires an exogenous source of variation in school choice.
 Basic economic models of education choice suggest that schooling decisions are
 based on the costs and benefits of alternative choices. In this dissertation, I
 exploit geographic differences in accessibility to CEC schools as a source of ex-
 ogenous variation. Using administrative data on student and school addresses,
 I conduct an instrumental variables approach that uses a function of residen-
 tial distance to the closest CEC school as an instrument for CEC attendance.
 As I explain in detail in Chapter 2, the proposed instrument solves the unob-
 served selection problem as long as the distance of a student?s residence to the
 closest CEC school has a strong effect on the probability of attending a CEC
 school, without affecting student outcomes directly conditional on other exoge-
 16
nous covariates. Intuitively, differential distances to CEC schools induce quasi-
 experimental differences in transportation and opportunity costs of attending a
 given CEC school, differences that are especially important for students with low
 socioeconomic background for whom the decrease in costs due to CEC proxim-
 ity may induce CEC school attendance.
 While residential distance to the closest CEC school may in theory be corre-
 lated with unobservable characteristics of students, I demonstrate using a variety
 of empirical strategies that the proposed instrument is likely to be conditionally
 uncorrelated to unobserved determinants of academic performance in this con-
 text. The most important factor driving this conditional exogeneity is that all
 CEC schools were constructed in neighborhoods with excess demand for pub-
 lic education in the poorest areas of the city.9 Thus, conditional on residing in
 these areas, I show that the distance from a student?s residence to the closest
 CEC school is not correlated with observed family or student determinants of
 academic performance and, therefore, it is likely to be uncorrelated with unob-
 served determinants as well. Moreover, to further check for the direct effect of
 distance on academic achievement, I perform a reduced-form placebo test using
 students? test scores for the year 2000, one year before any CEC school gradu-
 ated their first class and show that CEC schools were not systematically located
 in areas of the city with higher probabilities of success. In fact, if anything, the
 results indicate that my instrumental variables estimates are lower bounds of the
 9 The city?s Department of Education (SED) was responsible for locating and constructing all
 CEC schools in Bogot?. As a result, selected providers did not have any influence on the location
 of the schools.
 17
true causal effect on test scores as CEC schools may have been placed in neigh-
 borhoods with unobserved lower than average socio-economic characteristics.
 I show that, relative to public school students, CEC students scores are
 around 0.6 and 0.25 standard deviations higher in math and verbal test scores,
 respectively. Taken at face value, these effects reduce by a half the public-private
 school achievement gap in math and verbal test scores, a significant effect con-
 sidering the large differences in socio-economic characteristics between these
 two population groups. The estimated results are significantly higher than OLS
 estimates for CEC attendance, suggesting that the observed association between
 CEC attendance and test scores is not driven by omitted variables such as higher
 ability or highly motivated families among CEC students. Instead, the evidence
 points to the possibility that the unobserved variables that determine selection
 into CEC schools are negatively associated with academic performance.
 Unlike some programs that have implemented test-based accountability
 practices in public schools in the US (Koretz and Barron [1998]; Figlio and Get-
 zler [2002]; Cullen and Reback [2002]; Jacob [2003, 2004]), I show that the higher
 performance exhibited by CEC students are not driven by strategic responses
 by CEC schools such as forcing low-performing students to either transfer or
 dropout from CEC schools, or by preventing students in some other way from
 taking the test. I also provide evidence that the estimated effects are not driven
 by specializing in high-stakes subjects such as math and reading at the expense
 of ignoring other critical areas of learning like the natural sciences. First, unlike
 most standardized tests, ICFES test evaluates students on most of the curriculum
 18
subjects determined by the Colombian Ministry of Education,10 which reduces
 the effectiveness of test preparation. Second, neither the schools nor their stu-
 dents have incentives to exert more effort on only some specific subjects. From
 the point of view of the schools, the academic requirements defined in the pro-
 vision contracts are based on their students? ICFES composite test scores, calcu-
 lated for each student as a weighted average of all test sections. For the students,
 ICFES composite test score serves as the main admission criterion to higher edu-
 cation institutions in Colombia (ICFES, 1999). My instrumental variables results
 indicate that attending a CEC school increases composite test scores by 0.3 stan-
 dard deviations relative to public school students, suggesting that CEC schools
 are not specializing in certain academic subjects.
 1.4. Longer-Run Effects of CEC attendance
 Interestingly, most studies on the effects of contracting out public schools
 focus mostly on the short-run effects of attending privately run public schools
 on test scores, with most of the existing evidence coming from evaluating US
 charter schools (Angrist et. al [2010]; Abdulkadiroglu et. al, [2009]; Dobbie
 and Fryer, [2009]; Hoxby and Murarka, [2009]; Hoxby and Rockoff, [2004]; Saas,
 [2004]; Bifulco and Ladd, [2004]; Hanushek et. al, [2007]; Booker et. al [2007]).
 Missing from these studies is an assessment of longer-run outcomes of attending
 such schools such as earnings and college enrollment ?outcomes that are more
 10 ICFES test evaluates students on Math, Verbal, Biology, Chemistry, Physics, History, Geog-
 raphy, Philosophy, and English.
 19
clearly tied to economic success.
 In Chapter 3, I examine some longer-run effects of attending a CEC school.
 Given that the ICFES test is a compulsory requirement for students intending to
 enroll in college in Colombia and serves as the primary college admission crite-
 rion (ICFES 1999), it seems natural to evaluate whether CEC students have been
 able to translate higher ICFES test scores into high school graduation rates and
 higher college enrollment rates relative to Traditional Public School (TPS) stu-
 dents. The latter outcome is a relevant long term measure of school performance
 given that the returns from college in Colombia have been estimated to be twice
 as large as the returns from primary and secondary school (World Bank, 2003).
 In addition to this outcome, I also consider whether students who attended
 CEC schools are more likely to be admitted at selective universities as there is
 causal evidence for Colombia showing that students who attend selective uni-
 versities are more likely to have formal employment and have higher earnings
 than students attending non-selective institutions (Saavedra, [2008]). However,
 low-fee charging, selective, public universities in Colombia have managed to
 keep a low number of admitted students (World Bank, 2003). As a result, low-
 income students who traditionally graduate from TPS and CEC schools have a
 hard time competing with private school students for admission into selective,
 public institutions. Therefore, some public schools may be aiming at placing
 their students in two-year vocational programs, specifically designed to increase
 students? odds of getting formal employment. Thus, as a third outcome, I look
 at the probability of enrolling in 2-year technical and technological programs
 20
relative to TPS students.
 Lastly, I look at the college dropout behavior of CEC students. In Colom-
 bia, 50% of the students who enroll in higher education institutions drop out
 from their programs (MEN, 2010). While part of this may be due to finan-
 cial considerations ?including inadequate student financial aid for low-income
 students? it is often argued that the low-quality of secondary education is also
 an important reason for dropping out from college.
 To construct these higher education outcomes, I use administrative data
 from the Ministry of Education on college enrollment which has student-level
 information on institution and program enrollment as well as dropout behavior.
 Using the same identification strategy described above, I find that that CEC
 students in 2007 were 12 percentage points more likely to enroll in college. This
 represents more than a 50% increase over the average college enrollment rate for
 this sample.
 I also find that 2007 CEC students are one percentage point more likely to
 attend a selective and public institution compared to TPS students ?a relatively
 large effect compared to an unconditional mean of 3%. IV results also show that
 CEC students are almost 8.8 and 6.5 percentage points more likely to attend a
 vocational program than TPS students in the 2007 and the 2007-2008 samples,
 respectively. These results represent more than an 85% increase compared to the
 unconditional means of 2-year programs. This is a relevant outcome given that,
 according to the Ministry of Education, 74% of the graduates from technological
 programs in Colombia have formal employment ?just 3 percentage points fewer
 21
than graduates from 5-year colleges and universities in the 2001-2009 period
 (MEN, 2011). Also, graduates from technology programs have earnings equal to
 75% the earnings of university graduates.
 Lastly, regarding the probability of dropping out in the first two semesters
 after enrolling in higher education, I find that there is not a meaningful differ-
 ence in the college dropout probability.
 1.5. Conclusion
 In this dissertation, I evaluate a governance structure that aims at improv-
 ing education quality while allowing the state to retain a fair amount of control
 over the services provided. This governance structure can be broadly character-
 ized by elements of the contract design and the institutional framework in which
 the contracts are implemented. This program allows the SED to hold the CEC
 schools accountable for the academic achievement of their students. Moreover,
 the contractual arrangement aims to reduce quality-shading cost reductions by
 CEC providers in a series of ways, including selecting non-profit organizations
 with strong reputations.
 I provide strong evidence in terms of higher test scores by CEC program
 participants and some evidence on positive longer-run effects ?especially, in
 terms of higher college enrollment rate. The evidence, however, indicates that
 the higher college participation is, to a great extent, explained by a higher prob-
 ability of enrolling in two-year vocational programs. While there is some evi-
 22
dence that graduates from vocational programs may enjoy better labor market
 outcomes, for example, in terms of the probability of being formally employed,
 admission into vocational programs imposes lower requirements in terms of
 ICFES test scores as admission into these programs is not as competitive.
 What these results indicate is that the higher rates of college participation
 are not necessarily due to the better academic performance of CEC students and
 may be explained by alternative mechanisms. For example, the higher participa-
 tion in vocational programs by CEC students is consistent with the intention of
 some CEC schools to provide terminal vocational courses in order to smooth out
 their students? entrance to vocational programs at higher education institutions.
 Therefore, it will be interesting in the future to look at other labor market out-
 comes such as wages and employment which may be more relevant to evaluate
 the full impact of this program.
 Finally, it is also worth keeping in mind that the 2008 data on college en-
 rollment seems to be incomplete given the striking difference in enrollment rel-
 ative to 2007. Then, some of the results presented here on higher education
 ?particularly the ones using 2008 data? are to be considered preliminary as
 this requires a better understanding of this discrepancy.
 23
Tab. 1.1. Evaluation Methodology since 2007
 Component Selected Indicators Weight
 Participation of Students, Teachers, Principals, and 
Concesionnaires in school government, school 
manual, and pedagogical proposals
 Availability and Accessibility of education materials 
Dropout Rate
 Failing Rate
 Food Provision Quality and Quantity
 Absolute performance in ICFES test
 Relative Performance in ICFES test
 Programs for community and parental participation in 
school activities
 Institutional organization
 40%
 20%
 20%
 Other
 Academic Performance
 Non-Academic Indicators
 Pedagogical Model
 20%
 Source: IDEP (2010).
 24
Figures
 Fig. 1.1.Location of TPS, New TPS, and CEC Schools
 Notes: This figure shows the location of all public and CEC schools in the city as well as the students residences by type of school attended.
 CEC and TPS students? residences are depicted with dark and light gray dots, respectively. CEC and TPS locations are represented with
 crossed black dots and hollow circles, respectively.
 Fig. 1.2. CEC?s ICFES Performance Relative to TPS in same Locality
 Notes: This figure shows confidence intervals for the composite ICFES scores difference between each CEC and all TPS located in the same locality.
 For both years, CEC schools are sorted in ascending order of the 2008 CEC-TPS ICFES difference to facilitate visualization.
 25
Fig. 1.3. CEC Academic Performance Component 2007-2009
 0%
 20%
 40%
 60%
 80%
 100%
 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
 To
 ta
 l
 CEC School ID
 2007 2008 2009
 Fig. 1.4. CEC Non-Academic Indicators 2007-2009
 0%
 20%
 40%
 60%
 80%
 100%
 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
 To
 ta
 l
 CEC School ID
 2007 2008 2009
 Fig. 1.5. CEC Overall Evaluation 2007-2009
 0%
 20%
 40%
 60%
 80%
 100%
 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
 Tot
 al
 CEC School ID
 2007 2008 2009
 Source: "Resultados de la III fase de Evaluaci?n a Colegios en Concesi?n". IDEP (2010).
 26
2. SHORT-RUN EFFECTS OF THE CEC PROGRAM
 2.1. Introduction
 In this chapter, I assess the causal effect of attending CEC schools in Bo-
 gota, Colombia, on the ICFES test, a national standardized test taken by the vast
 majority of high school graduates in their last year of high school. For this, I
 use data for all public and CEC students in Bogota who took the ICFES test and
 graduated in the second semester of 2008. Two main reasons support choosing
 data from this school year. First, CEC students graduating in 2008 are likely
 to have attended the same CEC school through all their secondary school years
 (middle and high school).1 This point is important for at least two reasons.
 Cunha et al., (2006) show that the returns to education investments are higher
 the earlier these occur in the life cycle since they raise skill attainment at later
 stages and facilitate the productivity of later investments. In addition to this, the
 education production function depends on cumulative inputs (Hanushek [2006]).
 The second reason behind the choice of the 2008 cohort is that, unlike previous
 years, the ICFES test collected a large set of variables on student SES character-
 istics, which I use repeatedly in the empirical analysis.
 1 Secondary education in Colombia runs from 6th to 11th grades with middle school compris-
 ing 6th- to 9th and high school 10th and 11th grades.
As is customary in studies of school choice, we need to pay close atten-
 tion to the potential selection of program participants. Indeed, students who
 attend CEC schools may differ in a number of ways from traditional public
 school (henceforth, TPS) students. In this dissertation, I exploit geographic dif-
 ferences in accessibility to CEC schools as a potential source of exogenous vari-
 ation. In particular, I use an administrative dataset that allows me to conduct
 an instrumental variables approach, using distance to the closest CEC school to
 construct an instrument for CEC attendance. The proposed instrument solves
 the unobserved selection problem as long as the distance of a student residence
 to the closest CEC school has a strong effect on the probability of attending a
 CEC school, but does not affect student outcomes directly conditional on other
 exogenous covariates. Thus, special attention is paid to show that proposed in-
 strument is likely to be conditionally uncorrelated to unobserved determinants
 of academic performance.
 The two-stage least squares estimates obtained indicate that CEC students
 score 0.6 and 0.2 standard deviations higher in math and verbal tests, respec-
 tively, relative to public school students. As reference, these effects reduce by a
 half the public-private school achievement gap in math and verbal test scores, a
 significant effect considering the large differences in socio-economic characteris-
 tics between these two population groups. The estimated results are significantly
 higher than OLS estimates for CEC attendance, suggesting that the observed as-
 sociation between CEC attendance and test scores is not driven by omitted vari-
 ables such as higher ability or highly motivated families among CEC students.
 28
Instead, the evidence points to the possibility that the unobserved variables that
 determine selection into CEC schools are negatively associated with academic
 performance. I also show that that my instrumental variables results are not
 explained by treatment effect heterogeneity by students? income levels. Lastly, I
 provide evidence that the estimated results are not driven by unintended strate-
 gic responses by CEC schools, such as excluding low-performing students from
 the pool of test-takers or via test specialization in the curriculum.
 The rest of this chapter is organized as follows. Section 2.2 describes the
 institutional framework of Bogota?s public education system and briefly summa-
 rizes the literature on related school programs. Section 2.3 introduces the data
 sources and presents some descriptive statistics of the population of interest.
 Sections 2.4 and 2.5 describe the empirical strategy and the estimation results.
 Section 2.6 checks for the robustness of the results. Section 2.7 discusses the
 evidence on strategic responses to test-based accountability and the plausible
 mechanisms through which CEC schools are able to achieve higher academic
 results than TPS. Section 2.8 concludes.
 2.2. Background
 2.2.1. Public Education Sector in Bogot?
 There were 1.6 million children enrolled in primary or secondary school in
 Bogot? in 2008. Of these, 1.1 million children were part of the public education
 system: 80% of them attended more than 340 TPS, 5% attended 25 CEC schools
 29
and the remaining 15% attended 332 private schools paid by the City?s Depart-
 ment of Education (SED, hereafter) for taking public school students unable to
 obtain a seat in existing public schools.2
 Public schools in Bogot? largely educate the children of low income fam-
 ilies. As expected, families self-select into schools according to their economic
 means, which leaves the public schools with students from the bottom of the
 income distribution. For example, data on student socioeconomic characteristics
 collected at registration for ICFES test shows that 67% of students attending a
 public school come from families with a monthly household income less than
 two times the monthly minimum wage; this figure is just 30% for students at-
 tending private schools.
 Families applying for a public or CEC school for the first time must pro-
 vide a preference-ranked list of up to four schools.3 Before allocating children to
 schools, the SED sorts students according to the socioeconomic means of their
 families. That is, students from the most economically disadvantaged house-
 holds are allocated first to their most preferred options. If no seats are available
 for the preferred choices, students are placed in the closest available school to
 their residence with available seats. It is worth noting that this process is the
 same for students attending either public or CEC schools. That is, neither TPS
 or CEC schools can select their students using admission tests or similar criteria.
 2 This initiative is different from a voucher program since students are not free to choose the
 private school they attend, but must attend the one to which they are assigned by the SED.
 Students enrolled in participating private schools usually reside in areas with excess demand
 for public education.
 3 Students currently enrolled in a given school are guaranteed a place in their current aca-
 demic institution.
 30
As a result of the admission process, public schools exhibit important dif-
 ferences in the observed characteristics of their students and, very likely, in their
 unobserved characteristics as well. On the one hand, the allocation algorithm
 gives priority to families with the lowest socioeconomic status (SES). That is, for
 two students applying to a given school, the one with the lowest SES is more
 likely to attend it. Thus, over subscribed schools could systematically get stu-
 dents with lower than average SES. On the other hand, the SED allows students
 to apply to their most preferred institutions, which may lead to selection of
 most informed or motivated parents into better public schools. Consequently,
 it is unclear which selection mechanism dominates in this context and, in par-
 ticular, what exactly leads students to attend a CEC school. Because of this, in
 order to estimate a causal effect of CEC attendance on academic achievement, an
 identification strategy must account for the potential selection of CEC program
 participants into CEC schools.
 2.3. Data
 Administrative data for this chapter come from three main sources. The
 first source is ICFES, a public organization in charge of designing and admin-
 istering standardized tests for all schools and universities in Colombia. For the
 purposes of this study, we use the test known as Ex?men de Estado (commonly
 known as the ICFES test), taken by most high school students in their last year
 of school, which evaluates them on the multiple subjects mentioned above. A
 31
rich set of demographic data on students and their families is collected at the
 time of test registration, three months before the test is taken. Self-reported de-
 mographic information includes, for example, gender, age, parents? education,
 number of household members, asset ownership (e.g., car, computer, cell phone,
 TV, etc), household income, and residential address. In this chapter, I use the
 2008 ICFES results from students attending either TPS or CEC schools and set
 to graduate in the second semester of 2008. Using data from 2008, six years af-
 ter the last CEC school started operating, allows me to estimate the cumulative
 effect of CEC school attendance for students who presumably attended these
 schools all their middle and high school years at these schools.4
 The initial sample of public and CEC students graduating in the second
 semester of 2008 has 50,199 observations. After excluding students who attend
 night, technical, rural or military public institutions (because of concerns with
 the unobservable characteristics of these students) as well as observations with
 missing data for at least one of the variables (different from home address), I am
 left with 39,282 observations. The percentage of missing observations over total
 observations is around 6% for both CEC and TPS indicating that public students
 are not more likely to fail to report their personal information than CEC students
 (Table 2.1).
 The second administrative data source comes from UAECD (Unidad Ad-
 ministrativa de Catastro Distrital), a local government organization responsible for
 4 Dropout and transfer rates from these schools are very low as shown in Figure 2.3 (Sec-
 tion 2.7). Thus, even though individual years of exposure to treatment for each student is not
 available, data at the school level indicates this should not be a concern.
 32
maintaining a database of all legal buildings in Bogot? for urban planning and
 tax purposes. UAECD uses a geographic information system (GIS) that assigns
 coordinates (i.e., latitude and longitude) for all buildings in the city. Thus, using
 the residential address reported by students at the time of ICFES registration,
 I am able to calculate the distance from a student?s home to both actual and
 potential TPS and CEC schools in the city.5 As shown in Table 2.1, I am able to
 match 85% of students? addresses as reported at ICFES registration to addresses
 in the UAECD database.6 In general, the absence of meaningful demographic
 differences between the matched and unmatched samples suggests that we are
 not left with a non-random sample of students, something that could bias the
 estimated effects if, for example, students with lower SES are more likely to mis-
 report their home address. The final regression sample has 33,413 observations,
 including 31,783 public and 1,630 CEC students.
 Table 2.2 presents descriptive statistics for the year 2008 for public and
 CEC students. Panel A includes all the students used later for estimation (which
 I hereafter refer as the full sample). The results indicate that, relative to students
 from TPS, CEC students have slightly larger families, have less educated parents,
 and come from lower income households. Moreover, CEC students are more
 likely to live in low-strata neighborhoods. The stratum scheme is a six-level
 classification used by the local government to characterize the socio-economic
 5 Matching student addresses with the UAECD database requires addresses to be written in
 the same way UAECD addresses are written. Given that less than 15% of the addresses were
 exactly matched, a Stata program was created to standardize students addresses
 6 It is comforting that there are no economically significant differences in the demographic
 characteristics of the matched and unmatched samples as shown in Table 2.11 in Appendix C.
 The same patterns persist when looking at public and CEC students respectively.
 33
characteristics of a given dwelling. The higher the stratum, the better the socioe-
 conomic status of the dwelling. In general, residents of dwellings in the first two
 strata are considered the urban poor and pay, for example, subsidized prices for
 public utilities such as water and electricity. Whereas 94% of CEC students live
 in the lowest two strata of the city, 66% of public students do, indicating that
 the average CEC student lives in a neighborhood with very low socioeconomic
 conditions. Thus, even though the average TPS student comes from a low SES
 household, the average CEC school student has an even lower socio-economic
 background. These figures are not surprising given that CEC schools were sys-
 tematically located in strata 1 and 2 neighborhoods in peripheral areas of the
 city, as shown in Figure 2.1.
 Fig. 2.1.Location of CEC Schools
 Notes: This figure shows the location of all the residential units and the 25 CEC schools in Bogot?. Residential units are
 characterized by their stratum index. Stratum is a categorical index running from 1 to 6 and summarizes the socio-economic
 status of the dwelling and its neighborhood, 1 being the lowest. Dark gray areas correspond to strata 1 and 2 neighborhoods,
 medium gray to strata 3 and 4, and light gray to strata 5 and 6.
 34
To compare the socioeconomic characteristics of CEC and TPS students
 living close to CEC schools, I select all students living in UPZs with nearby
 CEC schools (henceforth, the UPZ sample ?see Figure 2.4). UPZs are subdivi-
 sions of the city which are used by the local administration for urban planning
 purposes. The 120 UPZs in the city are defined in such a way that the urban
 and economic characteristics of the housing units within them are very similar.
 Panel B of Table 2.2 presents descriptive statistics for public and CEC students
 in the UPZ sample. As shown, public and CEC students from the UPZ sample
 are very similar along most observable dimensions. This happens because the
 socioeconomic characteristics of TPS in Panel B are worse than those in Panel
 A. For example, the average public student in the UPZ sample has more family
 members and lower educated parents, lives in a lower stratum dwelling, and
 has a lower income relative to the average public student in the full sample.
 These figures provide evidence that CEC schools were located in economically
 deprived areas of the city.
 The last data source used in this chapter is the school surveys C600 for the
 year 2008 conducted by DANE (the Colombian national department of statistics).
 These surveys have detailed information for each school in the country on the
 number and education level of teachers by academic grade, number of students
 per grade, and the number of students who dropout or transfer at each grade.
 This information is used later in Section 7 when I consider potential explanations
 for CEC school higher academic performance.
 35
2.4. Empirical Strategy
 2.4.1. Structural Equation
 I begin by estimating the average differences in ICFES test scores as a func-
 tion of CEC school attendance, controlling for other student, family, and neigh-
 borhood characteristics. The structural equation for the outcome of interest Y
 for student i in UPZ u is given by:
 Yiu = rDiu + gu + XiG+ eiu (2.1)
 where gu is a UPZ fixed effect and Xi is a vector of potential observable
 determinants of academic performance, which includes all the socio-economic
 variables described in Table 2.2. The regressor of interest, Diu, is an indicator
 variable that equals 1 if a student attends a CEC school and is 0 otherwise.
 To estimate the causal effect of CEC school attendance on academic per-
 formance, the decision to attend a CEC school must be uncorrelated with un-
 observed determinants of academic performance captured in the error term, eiu,
 conditional on UPZ fixed effects and observable characteristics, Xi. If unobserv-
 able variables that determine the selection mechanism are omitted, estimates
 of CEC attendance (r) derived from OLS regressions in Equation 2.1 would be
 inconsistent.
 36
2.4.2. Estimating the Causal Effect of CEC School Attendance
 As stated above, the set of rules used by the SED to allocate students
 into schools may facilitate self-selection of highly motivated students into CEC
 schools since students are first assigned according to their stated school prefer-
 ences. Were this the case, OLS estimates of CEC attendance would be upward
 biased. On the other hand, the student allocation algorithm gives placement
 priority to students with the lowest socioeconomic characteristics. In fact, the
 evidence presented in Table 2.2 leans toward the possibility of negative selection
 into CEC schools. That is, compared to the average TPS student, CEC school
 students come from families with observably lower SES and live in low-income
 neighborhoods. It is entirely possible, then, that there are other unobservable
 characteristics which are both positively correlated with CEC school attendance
 and negatively correlated with academic performance. In this case, OLS esti-
 mates would be biased towards zero.
 A consistent estimate for the effect of CEC attendance on academic per-
 formance can be recovered if there is at least one variable Z that, in addition
 to being uncorrelated with the error term in Equation 2.1, is also a strong pre-
 dictor of the probability of attending a CEC school in the following first stage
 regression
 Diu = aZi + lu + Xib+ miu (2.2)
 37
where Diu is the indicator variable for CEC attendance of individual i in
 UPZ u, lu is a UPZ fixed effect, Xi is a vector of the student?s socioeconomic
 characteristics and miu is an error term. The causal effect of CEC attendance
 on academic achievement then can be estimated using Two-Stage Least Squares
 (hereafter, 2SLS).
 Note that, although a properly defined education production function should
 include school level variables that affect academic performance (e.g., teacher
 quality, student-teacher ratios, school facilities, ect.), the only school-level vari-
 able included in this model is the CEC attendance variable Diu. As a result, this
 variable should be interpreted as the aggregate effect of attending a CEC school
 on achievement without explicitly pinning down the mechanisms at work. Two
 reasons drive this methodological decision. First, including school level vari-
 ables will further complicate the identification strategy and the interpretation of
 the results as more instruments will be needed for every additional school level
 variable included. Second, even with additional valid instruments, including
 additional school level variables will prevent us from estimating the aggregate
 effect of participating in the CEC school program, which is one of the main goals
 of this research project. Nevertheless, I devote a great deal of attention to the
 differences in school level inputs in Section 2.7 and discuss what may be (and
 may not be) the mechanisms behind the estimated results.
 38
2.4.3. Using Distance to the Closest CEC School as an Instrument
 To estimate the causal effect of attending a CEC school on academic achieve-
 ment, I exploit variation in distance from a student?s residence to the closest CEC
 institution to construct an instrument for CEC attendance. For a function of dis-
 tance to the closest CEC school to be a good instrument for CEC attendance, two
 conditions are required. First, the distance of a student?s residence to the closest
 CEC school should have a strong effect on the probability of attending a CEC
 school. Second, for instrument validity, 2SLS should retain only the variation in
 the probability of CEC attendance that is generated by the quasi-experimental
 variation generated by the instrument. Intuitively, 2SLS allows us to compare
 the academic performance of students who differ in their propensity to attend
 a CEC school just because they reside various distances from these institutions.
 That is, distance to the closest CEC school should not be correlated with other
 unobserved determinants of academic performance. In this section, I show that
 the proposed distance instrument is both a valid and relevant instrument for
 CEC attendance.
 An Instrument for CEC Attendance
 The idea of using distance as an instrument has been exploited before. In the
 economics of education literature, Card (1993) estimated the returns to school-
 ing using variation in college proximity. Distance-based instruments have also
 been exploited in the health economics literature (McClellan, [1994]; Kessler and
 39
McClellan, [2000]; Freedman, [2010]), in studies of the effects of child subsidies
 on children wellbeing (Herbst and Tekin, [2010]), and the impact of community
 nurseries on low-income children nutrition in Colombia (Attanasio, [2010]).
 There are different ways to implement distance to the closest CEC as an
 instrument for CEC school attendance. For the empirical strategy, I define the
 instrument in terms of distance to the closest CEC school relative to the closest
 TPS. Two main reasons justify this decision. First, a relative distance instrument
 should have more power for those students who, in spite of living far from a
 CEC school, live even further away from a TPS. Second, it allows me to define a
 categorical instrument that takes the value of 1 if a student lives closer to a CEC
 school than to a TPS (i.e., for negative relative distances) and 0 otherwise. That
 is,
 Zi = 1fRDi < 0g, where RDi = Min
 c2CEC
 (dic) Min
 p2TPS
 (dip) (2.3)
 and dij is the distance from student?s i residence to school j 2 [c, p]
 Using a categorical instrument is partly driven by the non-linear relation-
 ship between relative distance to CEC school and CEC attendance probability.
 That is, relative distance in meters should be a strong predictor for CEC at-
 tendance only for students living within a certain distance from a CEC school,
 and be rather ineffective for students who live relatively far from it. Figure 2.5
 presents evidence on this point. It shows the fraction of total students attending
 a CEC school who live within a given interval of relative distance. This figure
 40
shows that whenever a student lives closer to a CEC than to a TPS (i.e., relative
 distance is negative), there is a higher probability of attending a CEC school.
 More importantly, for almost all the distance intervals for which a student lives
 closer to a TPS than to a CEC school (i.e. positive relative distance), the fraction
 of CEC students is not only small, but also non-decreasing in relative distance.
 To account for these non-linearities in the CEC attendance probability, I
 could partition the continuous instrument into several categorical variables for
 different intervals of relative distance. As it turns out, however, using a single
 categorical instrument that takes the value of 1 for students who live relatively
 closer to CEC schools (and 0 otherwise), exhibits the strongest effect on the
 probability of attending a CEC school as indicated by the F-test statistic in the
 first-stage specifications compared to a wide range of alternatives, including
 specifications with high order polynomials or multiple categorical instruments.
 Instrument Relevance
 Basic Economic models of education suggests that the school choice decision
 by a household is made by maximizing the difference between the expected
 returns and costs of attending a given school. The expected returns to attending
 a school are a function of the quality of the education provided. In turn, the
 costs of attending a public school are mainly determined by transportation and
 opportunity costs for both the students and their parents who usually need to
 walk their young children to school. Note that both of these costs are increasing
 in the distance from a student?s residence to a given school.
 41
The expected returns to attending a given school depend on the informa-
 tion set available to households. However, there is only partial public infor-
 mation on the quality of public schools in Bogot?. In fact, the only existing
 source of public information on school quality is average ICFES test scores at
 the school level, available online for each of the test subjects. Even if parents
 of public school students use this information to guide their school choice de-
 cision, Table 2.2 shows that unconditional test score means for CEC schools are
 less than 0.1 standard deviations higher relative to TPS. Moreover, given that
 few cohorts have graduated from CEC schools, there is also limited information
 from this source that might significantly affect the expectation of the returns
 to CEC school attendance. Thus, given the information set available to public
 school users, the expected returns to attending a CEC school may not be very
 different from attending a TPS.
 On the other hand, transportation and opportunity costs of sending a child
 to a school far from home can be prohibitive for most of these households.
 Monthly public transportation expenditures for a single person represent 15%
 of a monthly minimum wage. Also, due to traffic congestion, the average speed
 of public transportation in the city is around 14MPH,7 meaning that having to
 commute to a school far from home may significantly reduce the time that both
 children and their parents could otherwise use for work or housework. It can
 be argued that some households could find it optimal to assume one-time res-
 7 Source: Instituto de Estudios Urbanos, Universidad Nacional. www.redBogot?.com last
 visited on October 10, 2010.
 42
idence reallocation costs in order to reduce the costs of attending a preferred
 school. There is no evidence, however, that housing and school locations in Bo-
 got? are jointly determined by public student households. The 2007 Encuesta de
 Calidad de Vida, a locality-representative household survey in Bogot?, shows
 that, even though 21% of the households change their residence every 2 years,
 just 4% of them move due to education or health considerations.
 This simple framework suggests that parents? maximization problem in
 the current environment boils down to sending their children to schools where
 it is less costly to do so, that is, to schools close to their residence. Figure 2.6
 shows the proportion of students in the sample who live within a given distance
 interval from the school they actually attend. Approximately 15% of all students
 in the sample live within 300 meters (0.18 miles) from their school. Moreover,
 the cumulative distribution for both CEC and TPS students shows that 73% of
 them live fewer than 2000 meters (1.24 miles) from the school they attend. I
 interpret this as evidence that residential distance to schools determine the set
 of potential schools available to low income students in Bogot?.
 More importantly, this framework suggests that functions of distance to the
 closest CEC school could be used as an instrument for CEC school attendance as
 differential distances to CEC schools generates quasi-experimental differences in
 the costs of attending a given CEC school. The empirical evidence supports the
 idea that living close to a CEC school is a strong predictor of the probability of
 attending a CEC school. First, Figure 2.7 shows that CEC students? residences
 (dark gray dots) cluster around CEC school locations (crossed black dots). More-
 43
over, data collected by five of the 25 CEC schools on students? transportation
 choices indicates that around 90% of CEC students walk to school and for 95%
 of them it takes less than 25 minutes to get to school as shown in Figure 2.8.
 A third piece of evidence is presented in Figure 2.2, which shows the distribu-
 tion of students living at various relative distances from the closest CEC school.
 The figure shows that the empirical distribution for CEC students stochastically
 dominates the TPS distribution for negative values of relative distance (i.e., the
 closest CEC school is closer than the closest TPS). In contrast, those living closer
 to a TPS relative to a CEC school are more likely to attend the former. This
 figure motivates the use of the categorical instrument previously defined.
 More formally, first stage results presented in Section 5 show there is, in
 fact, a strong statistical relationship between distance to the closest CEC and
 CEC attendance probability.
 44
Z=1 Z=0
 0
 .000
 2
 .000
 4
 .000
 6
 .000
 8
  
?180
 0
 ?120
 0
 ?60
 0 0
 60
 0
 120
 0
 180
 0
 240
 0
 300
 0
 360
 0
 420
 0
 480
 0
 540
 0
 600
 0
 660
 0
 720
 0
 780
 0
 Relative Distance to Closest CEC (Meters)
 CEC Students
 Public Students
 Fig. 2.2.Distribution of Relative Distance to Closest CEC
 Notes: This figure shows the empirical distributions of relative distance in meters for both CEC and TPS students. Relative
 distance is defined as the distance to the closest CEC school minus the distance to the closest TPS. The figure graphically
 depicts the categorical instrument defined in Equation 3.1.
 Instrument Validity
 The second condition for distance to be a good instrument for CEC attendance is
 that the only relationship between the outcomes of interest and the instrument
 is through the first stage. This condition, known as the exclusion restriction, im-
 plies that the instrument should be uncorrelated with unobserved determinants
 of academic achievement. One concern is that families jointly determine their
 school choice and residence. Card?s (1995) college proximity instrument, for ex-
 ample, has been questioned on grounds of the joint decision between schooling
 and residence in the US since families with a higher taste for education may also
 choose to live near a college institution (Kling, 2001). Although the validity of
 an instrument is a non-testable condition, Altonji et al (2005) argue that if an in-
 45
strument is uncorrelated with a rich set of observable characteristics, it is likely
 to be uncorrelated with unobserved outcome determinants.
 To check for instrument exogeneity, Table 2.3 reports coefficients from OLS
 regressions of the socio-economic characteristics indicated in each row on the
 instrument, for specifications with and without UPZ fixed effects. The results in
 Column (2), which correspond to a specification without UPZ fixed effects and
 uses all the observations in the full sample, shows that living relatively closer to
 a CEC school is correlated with some of the students? observable characteristics
 such as low parent education levels, living in stratum 2, not living in stratum
 3, coming from a low monthly income household, and less likely to have a
 computer, home internet service or a car. These results are not surprising as
 they just confirm that CEC schools were located in low-income areas of the city.
 In fact, after conditioning on UPZ fixed effects in Column (3), the results indicate
 that relative distance is not systematically correlated with student socioeconomic
 characteristics. Moreover, using observations only from the UPZ sample, which
 is an alternative way of controlling for neighborhood characteristics near CEC
 schools, confirms the absence of correlation between the instrument and the
 rest of the exogenous covariates. I interpret this as evidence that the proposed
 relative distance instrument is exogenous conditional on UPZ fixed effects.
 46
2.5. Results
 This section estimates 2SLS specifications of CEC attendance on academic
 achievement. To account for the possibility of correlation of student outcomes
 within a given neighborhood, all regression estimates are clustered at the UPZ
 level using a paired bootstrap strategy that samples at the UPZ level.
 2.5.1. First Stage Estimates
 In this section I present the first stage estimates of the effect of living rela-
 tively closer to a CEC school than to a TPS on the probability of CEC attendance.
 I provide evidence that the instrument is strongly correlated with attending a
 CEC school and additional evidence that the proposed instrument satisfies the
 exclusion restriction.
 Table 2.4 shows the estimates of linear probability models for CEC school
 attendance on the categorical relative distance instrument and different sets of
 exogenous regressors.8 Columns (1) to (3) show the results for the full regression
 sample. Column (1), which only includes the instrument as a regressor, indicates
 that living closer to a CEC than to a TPS increases the probability of attending a
 CEC by 17 percentage points. Once UPZ fixed effects are added in Column (2),
 the estimated effect is reduced to 10 percentage points. Furthermore, adding the
 rich set of exogenous covariates in Column (3) do not change the point estimates
 of the instruments. I interpret the stability of these estimates as evidence in favor
 8 The full set of first stage estimates are presented in Table 2.9 of Appendix B.
 47
of the exclusion restriction conditional on UPZ fixed effects.
 At the bottom of the table, I also report the correct prediction rate for these
 linear probability models as well as F-statistics on the excluded instrument to
 assess the relevance of the instrument. The F-statistics for the specifications with
 UPZ fixed effects using the full sample are 27.64 and 27.83, well above the con-
 ventional critical values to assess finite sample bias due to weak instruments.
 Moreover, the correct prediction rates of these two specifications are above 71%.
 That is, both measures suggest that the proposed instrument is a strong predic-
 tor of CEC attendance. The results for Panel B, which includes only students
 living in UPZs with nearby CEC schools, largely replicate the results of the full
 sample, although the instrument is even stronger for this sample, as indicated
 by the large F-statistics in all three specifications in Columns (4) to (6).
 2.5.2. 2SLS Estimates
 Table 2.5 presents both OLS and 2SLS results of CEC school attendance
 on math, verbal, and the composite ICFES test scores measured in number of
 standard deviations for both the full and UPZ samples and for different sets of
 exogenous covariates and UPZ fixed effects.
 I report OLS and 2SLS estimates of CEC attendance on test scores in Table
 2.5. OLS estimates for all specifications (no covariates, UPZ fixed effects only,
 and covariates along with UPZ fixed effects) in both samples indicate that CEC
 students score 0.16, 0.18, and 0.13 standard deviations higher than TPS students
 in math, verbal, and the composite ICFES tests, respectively.
 48
Drawing from evidence I have provided that relative distance to a CEC
 school is exogenous conditional on UPZ fixed effects, I first focus on the results
 from Column (2), using observations from the full sample. 2SLS estimates show
 that attending a CEC school increases math, verbal, and the composite test scores
 by 0.52sd, 0.15sd, and 0.3sd, respectively, relative to TPS students, although the
 effect on the verbal test is imprecise. The 2SLS results in Column (3) shows
 that CEC attendance estimates are practically insensitive to the inclusion of the
 remaining exogenous covariates. As before, the stability of the estimates in Col-
 umn (3) provides evidence on instrument exogeneity conditional on UPZ fixed
 effects. Moreover, the 2SLS estimates using observations from the UPZ sample
 exhibit very similar magnitudes and significance levels.
 The lower gains in verbal test scores suggests that CEC schools have a
 stronger impact on some academic fields than others. Interestingly, this fact is
 consistent with the results found in both the charter school and test-based ac-
 countability literature in the US, where the gains in math are always larger than
 the gains in verbal test scores (see, for example, Abdulkadiroglu et. al., 2009;
 Chiang, 2009; Dobbie and Fryer, 2009; Jacob, 2005;). Note also that 2SLS esti-
 mates for the math and the composite tests are more than 4 and 3 times larger
 than the OLS estimates, respectively. This supports the claim that students at-
 tending CEC schools are different from public students in some unobservable
 dimensions that negatively affect test scores, which causes the OLS estimates to
 be downward biased. More formally, a Hausman exogeneity test of the equality
 of the OLS and 2SLS estimates is conducted through a paired-bootstrap proce-
 49
dure that samples at the UPZ level to account for correlation of student outcomes
 within UPZs.9 The results indicate that OLS and 2SLS estimates of CEC atten-
 dance for the math and composite tests are statistically different at the 5% and
 10% significance levels, respectively. While the 2SLS estimated effect for the ver-
 bal test is 58% higher than the OLS estimate (0.208 vs 0.131) in Column (3), I
 cannot reject that they are statistically equal.
 To assess the magnitude of the estimates, I compare the effect of attending
 a CEC school with some of the estimates of the socio-economic determinants of
 test scores, presented in Table 2.10 in Appendix B. For example, the estimates
 for the math test in Column (1) show that having parents with at most primary
 education reduce math test results by 0.12 standard deviations relative to col-
 lege educated parents and living in a stratum 1 neighborhood decreases math
 test scores by 0.1 standard deviations relative to living in stratum 4. Similarly,
 coming from a household with a monthly income less than one monthly mini-
 mum wage reduces math test scores by 0.2 standard deviations relative to those
 earning 4 or more times the minimum wage. These effects are significantly lower
 than the estimated 0.6 standard deviations effect of CEC attendance on the math
 test.
 I also compare these results with respect to the average performance of pri-
 vate school students in Bogot? taking the test in the second semester of 2008.10
 9 Given that the standard errors allow for correlation at the UPZ level, the variance covariance
 matrices under the Hausman null hypothesis are not asymptotically efficient as required. Thus,
 a 5000-repetition, paired bootstrap strategy that samples at the UPZ level with replacement is
 conducted in order to construct a consistent difference between the 2SLS and OLS variance-
 covariance matrices.
 10 The ICFES test is conducted two times a year as there are 2 different school calendars, one
 50
As mentioned earlier, the choice between private and public schools in Bogot?
 is largely determined by income. On average, private school students score one
 standard deviation higher than TPS in math and 0.5 standard deviations higher
 in verbal and the composite test. Altogether, the estimated effects for CEC atten-
 dance suggest that attending a CEC school reduce by a half the public-private
 school achievement gap, a significant effect considering the large differences in
 socio-economic characteristics between these two population groups.
 2.6. Robustness Checks
 In this section I explore the robustness of the results in two ways. First,
 I conduct a placebo test to further check for instrument exogeneity. Second, I
 also check whether the estimated effects of CEC attendance are driven by some
 specific population groups.
 2.6.1. A Placebo Test
 The fundamental claim that justifies using an IV strategy to identify the
 causal effect of interest is that the only reason why test scores change as the
 instruments change is through the variation that the instruments have on CEC
 attendance. That is, there should be no relation between the instrument and test
 scores in samples where there is no relationship between CEC school attendance
 and the instruments.
 for the period February-December and the other for August-June. While all public and CEC
 schools operate in the first calendar, private schools operate in both.
 51
In order to further assess the validity of the instrument, I construct a
 placebo test using ICFES data from the year 2000, one year before CEC schools
 graduated their first class. The idea is to construct the same categorical relative
 distance instrument used earlier for estimation on a sample for which the in-
 strument should not be relevant. To do this, I first georeference the residential
 address of public students who took the ICFES test in the second semester of
 2000. Then, I construct a variation of residential relative distance to the closest
 CEC school as introduced earlier in Equation 3.1. Recall that relative distance
 is defined as the distance to the closest CEC school minus the distance to the
 closest TPS. Note, however, that it is not possible to calculate the actual distance
 to the closest CEC school for high school students taking the ICFES test in 2000
 because CEC schools were not available for these students in that year. Instead,
 I calculate the distance for all TPS students in the sample to the closest poten-
 tial CEC school to construct a measure of relative distance. Then, I define the
 instrument as an indicator variable that takes the value of 1 if potential relative
 distance is negative and 0 otherwise.
 The placebo test consists of estimating reduced form specifications, that is,
 estimating the effect of the dummy instrument on the math, verbal, and overall
 test scores. More specifically, the reduced form specification is given by:
 52
Yiu = dZi + fu + XiP+ xiu, where (2.4)
 Zi = 1fPRDi < 0g
 PRDi = Min
 c2CEC
 (pdic) Min
 p2TPS
 (dip)
 That is, PRDi is the potential relative distance for student i, pdic is the potential
 distance of student i to CEC school c, and dip is the actual distance from student?s
 i residence to TPS p. Note that a positive reduced form effect of the instrument
 on test scores would question the exogeneity of the instrument for it would
 imply, for example, that CEC schools were strategically located in areas of the
 city with a higher probability of academic success.
 Reduced form estimates of the instrument on ICFES test scores, d, are pre-
 sented in Table 2.6 for both 2000 and 2008 academic years. As before, I focus on
 the results from Column (3) and (6), which are the specifications with the full set
 of UPZ fixed effects and exogenous covariates. The reduced form estimates of
 the effect of living relatively closer to a CEC school on the math and verbal tests
 in the year 2000 are slightly negative and not statistically different from zero.
 The estimated effect on the composite test score in both samples is also nega-
 tive and not different from zero in the UPZ sample. In contrast, all the reduced
 form effects using 2008 data show a significant positive effect of living relatively
 closer to a CEC school. These results provide strong evidence that the estimated
 effects of CEC attendance are not due to schools being strategically located in
 53
areas with a higher probability of academic success. If anything, the negative re-
 duced form estimates for the year 2000 indicate that CEC schools may have been
 constructed in neighborhoods with lower than average SES even within UPZs.
 In that case, the 2SLS estimates for CEC attendance would be lower bounds of
 the true causal estimate. However, most of the 2000 reduced form estimates
 either in the specification in Column (3) in the full sample or in all the specifi-
 cations using the UPZ sample show the year 2000 reduced form estimates are
 insignificant. Thus, I interpret the results from this test as additional evidence
 on the statistical validity of the instrument. 11
 2.6.2. Characterizing Treatment Effect Heterogeneity
 Up to this point, the estimated effects of CEC attendance have been as-
 sumed to be the same for all students in the data. However, with treatment
 effect heterogeneity, IV estimates can be interpreted as a weighted average of
 causal effects for particular subgroups of the population (i.e., the compliers) for
 whom the treatment status is changed by the instrument. The estimated effect
 is known as the local average treatment effect, or LATE (Imbens and Angrist,
 [1994]). Card (1993), for example, argues that the effect of college proximity
 on college attendance should be more important for children of less wealthy
 households. That is, what pushes some low-income students into college is the
 11 One concern with this placebo test, however, is whether students graduating in 2000 actually
 lived close to where potential CEC schools were later located, that is, whether student are simi-
 larly distributed in the city in both years. To evaluate this point, Figure 2.9 shows the empirical
 distributions of relative distance in meters for both 2000 and 2008 academic years. The large
 similarities of these distributions suggest that results for year 2000 are not driven because of a
 disproportionately low amount of public students in future CEC school areas.
 54
reduction in the costs due to college proximity. As a result, Card?s IV estimates
 of the return of college attendance can be interpreted as the return to college
 only for students from less wealthy households who are credit constrained in
 financing college.
 In order to assess possible treatment heterogeneity by income levels, I first
 construct a socio-economic index on the sample of students who live closer to
 a TPS than to a CEC school using all the observable characteristics used earlier
 in estimation. This index is constructed as the predicted probability of being a
 very-low income household.12
 Panel A of Table 2.7 shows descriptive statistics of selected student socio-
 economic characteristics by quartiles of the constructed SES index. By construc-
 tion, students from the lowest quartile have larger families, lower parental ed-
 ucation and income, and a higher probability of living in low strata neighbor-
 hoods. Following Kling (2001), the overall weight received for each group q
 when using 2SLS is given by
 wqjx = WqlqjxDDqjx/ 
q
 WqlqjxDDqjx (2.5)
 That is, conditional on the exogenous covariates X, the weight given by
 2SLS to a specific quartile q is formed by three components. The first is the
 proportion of observations that belong to quartile q, Wq. Second, 2SLS weights
 12 In practice, this is equivalent to being a SISBEN 1 household. SISBEN is a poverty-targeting
 index used by the Colombian government to determine eligibility for social programs. The index
 takes on 6 values, where only levels 1 and 2 are eligible for most government subsidies and is
 constructed from a rich set of variables such as housing materials, access to public utilities,
 ownership of durable assets, demographic composition, educational attainment, and labor force
 participation.
 55
the observations by the average conditional variance of the instrument estimated
 over the empirical distribution of X for each q, that is, lqjx = E[P[ZjX, q](1 
P[ZjX, q])jq]. The last term reflects the average impact of the instrument on CEC
 attendance for quartile q, DDqjx = E[E[DjZ = 1, X, q] E[DjZ = 0, X, q]jq].
 Estimates of wqjx in Equation 2.5 and its components are presented in Panel
 B of Table 2.7. The results in this table show that the two lowest quartiles receive
 marginally higher weights than the last two quartiles. That is, they receive just
 58% of the total weight. I also estimate the causal effect of CEC attendance on the
 math, verbal, and composite ICFES scores by income quartiles. Both the weights
 that each quartile receive as well as the estimated causal effects by income quar-
 tiles speak of a rather homogeneous treatment effect of the CEC program. This
 does not contradict the idea that the way the instrument operates is by lowering
 the costs of attending a CEC school. In fact, given that all students attending
 TPS belong to households with low socio-economic characteristics, it is not sur-
 prising that they respond similarly to cost reductions in school attendance.
 2.7. Mechanisms for Higher Academic Achievement
 Understanding the mechanisms that improve a student?s academic out-
 comes is a relevant policy question. This section discusses some possible mech-
 anisms through which CEC schools achieve better academic performance. First,
 I evaluate whether CEC schools responded strategically to the imposed account-
 ability system in ways similar to the ones referenced in the test-based account-
 56
ability literature. Second, I explore other dimensions in which CEC schools
 are different from TPS (besides higher flexibility and accountability) to assess
 whether these differences are the driving force behind the higher academic
 achievement of CEC students.
 2.7.1. Strategic Responses
 One potential concern with evaluating schools according to the perfor-
 mance of their students in standardized tests is the possibility that schools would
 respond by devoting more instruction time to high-stakes subjects at the expense
 of other critical areas such as the natural sciences. That is, given that most stan-
 dardized tests only evaluate students? math and reading skills, there are strong
 incentives to increase test-specific skills in these areas (Johnson, [1984]; Koretz
 and Barron, [1998]; Jacob, [2005]).
 This concern is largely addressed in the present context by virtue of the
 design of the ICFES exam, which evaluates students on most subjects included
 in the standard academic curriculum defined by the Colombian Ministry of Edu-
 cation. Moreover, the composite ICFES test score is calculated as a weighted av-
 erage of ICFES test sections that also penalizes students for specializing in some
 subjects by using a linear function of the student?s standard deviation.13 As
 shown in Section 5, CEC students score around 0.3 standard deviations higher
 than TPS students in the composite test. These results suggest that CEC schools
 13 the overall test score for student i is given by Scorei =  cfWc  Scorec,ig  
si
 2 , where
 c=[Biology, Chemistry, Physics, Math, Verbal, Social Sciences, Philosophy, English], weights
 given by Wc = (1, 1, 1, 3, 3, 2, 1, 1) and si equal to student?s i standard deviation from all c test
 components.
 57
are not responding to increased test-based accountability by specializing in some
 academic subjects. Part of this is explained by students? incentives to perform
 well in all test sections since ICFES composite test score serves as the main (and
 sometimes only) criterion for admission to two-year and four-year colleges and
 universities in Colombia. But also, academic accountability for CEC schools is
 contractually defined in terms of this composite score.
 This is an important finding for policy reasons because it suggests that
 evaluating students on multiple subjects should make test-based accountability
 programs less likely to be explained by test specialization as long as incentives
 are defined in terms of composite scores.
 A second potential response to test-based accountability is to prevent low-
 performing students from taking the test by making them transfer or drop out
 from a given school. Unfortunately, information on dropout or transfer students
 is not available at the student level. To account for these possibilities, I use data
 from the C-600 national school survey in 2008, which collects data on teacher
 and student characteristics at the school level. In particular, this survey collects
 information on the number of students per academic grade as well as the num-
 ber of students per grade either transferring or dropping-out. For a given school
 and grade, the transfer (dropout) rate is calculated as the number of students
 who transfer (drop out) over the total number of students in that grade. Then,
 for each academic grade, I calculate the average transfer (dropout) rate among all
 schools which offer that academic grade14 weighted by the total number of stu-
 14 Unlike CEC schools, most TPS do not offer all academic levels from primary to high school
 58
dents in that school-grade combination. If CEC schools exhibit higher average
 transfer or dropout rates, this would suggest that higher academic performance
 in CEC schools is achieved by dismissing low-performing students.
 Results on average rates are presented in Figure 2.3 for three different
 school samples: all TPS, TPS in the UPZ sample, and CEC schools. Results
 for TPS in the UPZ sample are included to assess whether TPS in this sample
 UPZs are more similar to CEC schools than the average TPS in the city. The
 figure shows that the average dropout rate for primary (1st to 5th), middle (6th
 to 9th), and high school (10th to 11th) grades in TPS are 3%, 5%, and 4%, respec-
 tively. In contrast, CEC schools have average dropout rates slightly above 1% for
 all academic grades. Interestingly, there are no meaningful differences between
 the rates for TPS in the full and UPZ samples. Moreover, the results on average
 transfer rates also indicate that whereas TPS transfer 2.5% of their students in
 each grade, CEC schools transfer just 1% of their students. Rather than being
 a concern, these results indicate that CEC schools are actually more successful
 than TPS at keeping students enrolled in school.
 A third potential unintended response is to prevent low-performing stu-
 dents from taking the ICFES test or, equivalently, to have students self-select
 into taking the test. A low test taking rate by CEC students would suggest that
 the higher academic achievement of CEC schools is due to differences in the
 composition of test takers. To evaluate this possibility, I construct a test taking
 rate for each school in the sample. For a given school, the test taking rate is
 but are primary, middle, or high school only.
 59
Pu blic (Full Sample)                     Pu blic (UPZ Sample)                     CE C Schools
 0%
 1%
 2%
 3%
 4%
 5%
 6%
 1 2 3 4 5 6 7 8 9 10 11
 Academ ic Grad e
 D ropout Rates by Academic Grade
 0%
 1%
 2%
 3%
 4%
 1 2 3 4 5 6 7 8 9 10 11
 Academ ic Grad e
 Transfer Rates by Academic Grade
 Fig. 2.3.Dropout and Transfer Rates by School Type
 Notes: Author?s Calculations from 2008 School Survey C-600. This figure shows the average dropout (transfer) rate by grade
 and school type for both the full regression sample and the UPZ sample. Averages are weighted by the total number of
 students at each school. Dropout (Transfer) rate is defined as the number of students who dropped out (transfer) from a
 given school and grade over the total number of students in that grade-school combination.
 defined as the total number of students taking the test in the second semester of
 2008 over the total number of students in grade 11 as reported in the 2008 C-600
 survey. Test taking-rates for TPS is approximately 93% to 96% of CEC schools.
 Moreover, the high test taking rates exhibited by both school types provides
 evidence on how serious schools and students are about the ICFES test.
 The existing evidence does not indicate that CEC schools are strategically
 responding to increased accountability by preventing low-performing students
 from affecting school results ?or at least no more than the average TPS. It is
 possible that students who drop out, transfer, or do not take the test are more
 likely to come from the bottom part of the potential score distribution. How-
 ever, given that CEC schools outperform TPS in all these dimensions, then 2SLS
 estimates of CEC attendance on test scores would be lower bounds of the true
 60
causal effect of interest.
 2.7.2. Alternative Mechanisms
 Determining the mechanisms that drive the higher academic achievement
 of CEC schools is relevant for policy reasons. In this section I discuss some
 potential mechanisms that may explain the higher academic results of CEC stu-
 dents. First, I look at input differences between public and CEC schools that have
 been associated with education quality in the education literature. In particu-
 lar, I look at teachers? education, student-teacher ratios and annual expenditure
 per student. Table 2.8 summarizes the results. Regarding teachers? education,
 TPS have more highly educated teachers relative to CEC schools at all academic
 levels. For example, whereas 47% of public middle school teachers have a grad-
 uate degree, only 11% of primary CEC school teachers do. This result is not
 surprising since teachers? salaries in TPS are partly determined by educational
 attainment. Thus, TPS teachers find it optimal to invest in their own education,
 regardless of graduate program quality. In contrast, at least 75% of CEC school
 teachers have at most a college degree in Education.
 Second, the data indicates that that both TPS and CEC schools have almost
 identical student-teacher ratios in primary and middle school grades of 40 and
 30 students per teacher, respectively. For high school grades, CEC schools have
 a ratio of 21 students per teacher compared to 30 students in TPS. Third, the
 annual cost of a CEC student to the city is lower than the cost of a TPS student.
 The higher operational costs of TPS are primarily driven by the higher educa-
 61
tional attainment of their teachers relative to CEC schools (Contraloria Distrital,
 2003).
 In principle, none of these three factors can explain the higher academic
 achievement by CEC schools. If anything, having more educated teachers and
 higher spending per student should increase academic performance of TPS.
 Also, given that the difference in student teacher ratios between CEC and TPS
 only exists in the last two years of high school, it is unlikely that the estimated
 results are driven by this factor. In any case, the literature of the effects of school
 resources on achievement such as teacher?s education, student-teacher ratios
 and student expenditures are far from conclusive, even in studies that exploit
 exogenous sources of variation [Hanushek, 1986, 1996, 2006].15
 2.7.3. A Contract for Academic Achievement
 The CEC program implemented in Bogot? provides strong empirical sup-
 port for the case of private provision of public education through service-provision
 contracts. In particular, the existing institutional design exhibits gains in educa-
 tion quality, while limiting the opportunities for inefficient cost reductions.
 The factors that explain the gains in quality are primarily due to private
 operators being contractually subject to annual performance evaluations, which
 includes rules for the academic performance of their students in standardized
 tests. Moreover, in addition to holding CEC schools academically accountable,
 15 For example, in a review of 376 studies on resources and student performance using US data,
 Hanushek (2006) finds that in 86% of the studies teachers? education does not improve student
 performance; 72% do not find evidence of a positive effect of lower student-teacher ratios; and,
 66% do not find an effect of higher expenditure per student.
 62
schools themselves are able to hold teachers accountable for the academic per-
 formance of their students. It is unlikely that academic performance would
 improve were the schools not allowed to demand results from teachers.
 Nevertheless, the arguments against private provision of public services
 are related to the part of quality that is not contractible in nature. The CEC pro-
 gram addresses potentially harmful cost reductions in a variety of ways. First,
 students are allowed to choose the schools they want to attend and private op-
 erators are paid according to the number of students enrolled up to school ca-
 pacity. This creates a potential competition effect that softens quality-reducing
 decisions. Moreover, schools are not able to select their students in any way and
 were not able to select the school locations in order to prevent CEC schools to
 systematically select profitable students at the expense of unprofitable ones (e.g.,
 problematic, low-performing students).
 Second, to further attenuate incentives for cost reductions, contracts were
 assigned to not-for-profit, well regarded institutions. The critical characteristic of
 a non-profit firm is that it is barred from distributing any profits it earns to those
 who exercise control over it. As a result, profits are usually used to improve the
 quality of the service (Glaeser and Shleifer, [1998]). This is in fact the case for the
 CEC program. Although public and CEC schools in Bogot? have similar costs
 per student, TPS used 30 percent more of their annual budget to pay for teachers
 wages in 2002 (Cotraloria Distrital, 2003). Additional resources were invested by
 CEC schools in food provision and school supplies for students. Moreover, by
 assigning contracts to respected schools and universities, providers should face
 63
large costs from choosing lower non-contractible quality in the form of lower
 reputation of their own private institutions. Also, given that service provision
 contracts in Colombia are subject to renewal for at most 50% of the length of the
 original contract, lower reputation due to reduced quality inflicts large costs on
 CEC providers interested in contract renewal.
 It is also possible that not-for-profit, well regarded organizations have a
 genuine preference from providing goods of higher quality derived from altru-
 istic behavior (Rose-Ackerman, [1996]; Lakdawalla and Philipson, [1998]). In
 fact, inspection of CEC school budgets show that they are investing in potential
 quality-improving activities that are not part of their annual evaluation process
 such as psychological attention to children, remedial education programs, or
 sending school social workers to visit the household of students who have re-
 cently stopped attending school.
 Third, to avoid ex-post opportunistic behavior by CEC providers, owner-
 ship of school facilities and supplies ?the most specific assets of this contrac-
 tual arrangement? was retained by the state. By doing so, the SED aimed at
 retaining residual contract rights in uncontracted circumstances (Joskow, 1988).
 One final factor that prevents significant reductions in non-contractible
 quality lies in the nature of education provision along with the wage structure in
 the education sector in Colombia. Providing education is a labor intensive ser-
 vice with labor costs accounting for an important fraction of total expenditures.
 This is even more pronounced for CEC schools given that the state is the formal
 owner of the school facilities and is therefore responsible for major infrastructure
 64
investments. More importantly, wages in Colombia for both private and public
 teaches are regulated by law. As a result, although CEC schools could find cost
 reduction opportunities by hiring lower quality workers at lower wages, the le-
 gal constraints on teacher wages provide incentives for these schools to hire the
 best teachers willing to work for them.
 2.8. Conclusion
 This chapter presents estimates of the causal effects on academic achieve-
 ment of attending a contracted out public school in Bogot?, Colombia. Instru-
 mental variables estimates constructed by using a distance-based instrument
 controlling for neighborhood fixed effects and a rich set of exogenous covariates
 consistently show a large positive association between program participation
 and academic achievement. These effects are largest and significant for the math
 test as well as for the composite test, which evaluates students in most middle-
 and high-school academic subjects included in the Colombian standard curricu-
 lum. The results for the verbal test, while positive, are imprecise. I also provide
 a variety of empirical strategies to argue that, conditional on taking the test, the
 proposed instrument is indeed exogenous.
 Although standardized tests have been used as proxies for school quality,
 higher academic results in standardized tests do not necessarily translate into
 longer-run gains in potentially more meaningful outcomes. Evans and Schwab
 (1995), for example, argue that standardized tests are limited measures of a
 65
student?s ability, creativity or deeper problem-solving skills. In Chapter 3, I
 examine the longer-run effects of CEC school attendance using administrative
 data on college level enrollment (2- and 4-year programs), major choice, and
 college academic performance.
 This chapter also explores some plausible factors that may have driven the
 higher academic results of contracted out public schools. Using school-level
 data, I argue that the higher academic achievement of these schools is unlikely
 to be explained by unintended strategic responses to increased test account-
 ability. If anything, the lower socio-economic characteristics of student partici-
 pants and the lower dropout and transfer rates as well as the higher test taking
 rates of these schools indicate the estimated effects are lower bounds of the true
 causal effect. Moreover, using school level data on education inputs show that
 contracted-out schools exhibit higher levels of academic achievement, despite
 having lower educated teachers and similar student-teacher ratios and expendi-
 tures per student.
 66
Tab. 2.1.Sample Exclusion Criteria
 Public C EC Total
 (1) (2) (3)
 Initial Sample: Students taking 2008 ICFES  Test and Graduating in 2008 48,125 2,074 50,199
 Exclusion criteria (total cases)
 Nig ht Schools 4,033 0 4,033
 Technical Schools (including  Teacher  S c hools) 3,417 0 3,417
 Military Sc hools 239 0 239
 Rural Schools 151 0 151
 Observations with m issing  data (differ ent from address) 2,931 146 3,077
 Missing observations / I nitial Sample (%) 0.061 0.070 0.061
 Subtotal (Initial Sample m inus ex cluded observations) 37,354 1,928 39,282
 Minus observations with unmatched address 5,571 298 5,869
 Unmatched observations / Total non-excluded observations (%) 0.15 0.15 0.15
 Matched observations / Total non-excluded observations (%) 0.85 0.85 0.85
 Final Reg ression Sample 31,783 1,630 33,413
 Notes: This table presents the sample construction for regression analyzes. The initial sample of public and CEC
 students graduating in the second semester of 2008 has 50,199 observations. The sample reduces to 31,783 observa-
 tions after excluding students who attend night, technical, rural or military public institutions, have missing data for
 at least one of the variables, or were nor succesfully assigned a valid residential address from the UAECD database.
 67
Tab. 2.2.Descriptive Statistics by School Type
 PUBLIC CEC PUBLIC CEC
 Male=1 0.46 0.46  0.46 0.46  
Age in  Y ears 17.08 16.99 ** 17.11 16.97 ***
 (1.30) (1.75) (1.40) (1.79)
 Numb er Hous ehold Memb ers 5.16 5.32 *** 5.28 5.40 **
 (2.05) (2.06) (2.04) (2.10)
 Numb er of Rooms 4.24 4.32 * 4.23 4.36 **
 (1.87) (1.77) (1.86) (1.83)
 No Younger Siblings =1 0.34 0.27 *** 0.32 0.26 ***
 Dad is  Literate 0.98 0.98  0.98 0.98  
Mom is  Literate 0.99 0.99  0.98 0.99  
Father Education
 Primary 0.39 0.43 *** 0.48 0.44 ***
 Secondary 0.47 0.47 0.44 0.48 ***
 A s sociate 0.08 0.07  0.05 0.06  
College or More 0.06 0.04 *** 0.03 0.02  
Mother Education
 Primary 0.33 0.38 *** 0.41 0.40  
Secondary 0.51 0.49 *** 0.48 0.50  
Associate 0.11 0.09 *** 0.08 0.08  
College or More 0.05 0.05  0.03 0.03  
Household Income
 1 MW or less 0.14 0.13  0.17 0.13 ***
 1 < MW < 2 0.53 0.59 *** 0.57 0.60 ***
 2 < MW < 3 0.23 0.23  0.20 0.23 **
 MW > 4 0.09 0.05 *** 0.05 0.03
 Household Stratum
 One 0.15 0.24 *** 0.23 0.26 ***
 Two 0.51 0.70 *** 0.66 0.71 ***
 Three 0.32 0.05 *** 0.10 0.03 ***
 Four+ 0.01 0.01  0.00 0.00
 Floor Quality of Dwelling
 Low 0.09 0.05 *** 0.08 0.04 ***
 Med iu m 0.34 0.43 *** 0.45 0.43 **
 High  0.57 0.53 *** 0.47 0.53 ***
 Assets
 Computer at Home 0.49 0.47 *** 0.42 0.45 ***
 D VD  at Home 0.75 0.75  0.75 0.75  
Car Owners h ip 0.15 0.16  0.12 0.13 *
 Motorbike Own ers h ip 0.07 0.10 *** 0.07 0.08  
No Cellphone in Hous ehold 0.06 0.04 *** 0.06 0.04 ***
 One Cellphone in Hous ehold 0.22 0.21  0.23 0.19 ***
 Internet at Home 0.23 0.19 *** 0.17 0.17  
TV Service 0.61 0.56 *** 0.57 0.56  
Outcomes
 Math Test Score (SD ) 0.00 0.10 *** -0.05 0.10 ***
 (1.00) (0.94) (0.97) (0.94)
 Verbal Test Score (SD ) 0.00 0.08 *** -0.07 0.09 ***
 (1.01) (0.89) (0.97) (0.88)
 Overall Test Score (SD ) -0.33 -0.27 *** -0.39 -0.26 ***
 (0.75) (0.64) (0.71) (0.63)
 Distance
 R elative D istance (meters ) 1800 468 *** 546 10 ***
 (1774) (1385) (564) (522)
 Distance to Actual School (meters ) 2856 3244 *** 1803 1883  
(3920) (5712) (2662) (4901)
 Number of Schools 410 205
 N 31783 10664
 A. Full Sample B. UPZ  Sample 
25 25
 1630 1304
 Notes: Panel A includes all observations with valid non-missing data used later in estimation. Panel B includes all
 students whose residence is within the same UPZ of a CEC school. The stratum variables is an index running from 1
 to 6 which summarizes the socio-economic status of the dwelling and its neighborhood, 1 being the lowest. Household
 income is measured in number of monthly minimum wages (MW). Relative Distance is defined as distance to the closest
 CEC school minus distance to the closest TPS. Significance levels for the equality of means tests: *** 1%, **5%.
 68
Tab. 2.3.Exogeneity of Relative Distance to Closest CEC
 RD >0 
Means
 No 
Controls
 UPZ  Fi x ed 
Ef f ects
 RD >0 
Means
 No 
Controls
 UPZ  Fi x ed 
Ef f ects
 (1) (2) (3) (4) (5) (6)
 Student and Family Characteristics
 Male 0.46 -0.003 0.006 0.47 -0.009 0.004
 Age 17.07 0.038 0.023 17.09 0.014 0.014
 Number Hous ehold Memb ers 5.17 0.012 -0.032 5.31 -0.114 -0.075
 Father Ed ucation: Pri mary = 1 0.38 0.107*** 0.016 0.47 0.034 0.006
 Father Ed ucation: Secondary = 1 0.48 -0.048** -0.009 0.45 -0.019 0.000
 Mother Ed ucation: Pri mary = 1 0.32 0.077*** -0.002 0.40 0.020 -0.007
 Mother Ed ucation: Secondary = 1 0.52 -0.026 0.014 0.48 -0.004 0.019
 Dad is  Literate 0.98 -0.003 0.004 0.98 0.000 0.004
 Mom is  Literate 0.99 -0.005 0.002 0.98 -0.002 0.003
 No Younger Si bling s 0.34 -0.038*** -0.023* 0.32 -0.022 -0.026*
 Household Income
 Stratum 1 = 1 0.15 0.050 -0.010 0.24 -0.001 -0.018
 Stratum 2 = 1 0.50 0.214** 0.025 0.65 0.061 0.040*
 Stratum 3 = 1 0.34 -0.253*** -0.014 0.11 -0.060 -0.021
 House Floors - Low Quality 0.09 -0.031** -0.000 0.08 -0.018 0.006
 House Floors - Med Quality 0.34 0.111*** 0.005 0.44 0.021 -0.011
 Number of Rooms i n Hous ehold 4.24 0.016 -0.008 4.24 0.011 -0.017
 House In come: 1 MW or les s 0.14 0.014 0.004 0.16 -0.000 0.006
 House In come: 1 < MW < 2 0.53 0.034** -0.006 0.58 -0.005 -0.015
 House In come: 2 < MW < 3 0.23 -0.008 0.003 0.20 0.011 0.008
 Household Assets
 Computer at Home 0.50 -0.065** -0.007 0.43 -0.014 -0.005
 DVD at Home 0.75 -0.014 -0.004 0.75 -0.008 0.001
 Car Owners hi p 0.16 -0.031** -0.003 0.12 -0.000 -0.002
 Motorbike Owners hi p 0.07 0.002 0.001 0.08 -0.003 -0.002
 No Cellphone i n Hous ehold 0.06 -0.001 0.005 0.06 -0.007 0.008
 One Cellphone i n Hous ehold 0.22 0.005 0.007 0.22 0.001 0.002
 Internet at Home 0.24 -0.053** 0.011 0.17 0.003 0.020
 Cable TV at Home 0.61 -0.058* -0.018 0.58 -0.043 -0.030
 N 29971 33413 33413 9373 11968 11968
 Full Sample UPZ  Sample
 Balance Regres s ions Balance Regres s ions
 Notes: Columns (1) and (4) report means of the variable indicated in each row for students living closer to a TPS than to a CEC
 school for the full and UPZ samples, respectively. Columns (2), (3), (5), and (6) report coefficients from regressions of the variable
 indicated in each row on an indicator variable equal to one if the student leaves closer to a CEC school than to a TPS (i.e., the
 relative distance instrument). Columns (2) and (5) only include the instrument as a regressor. Columns (3) and (6) include also
 UPZ fixed effects. Significance Levels *** 1%, **5%, *10%. All standard errors account for clustering at the UPZ level.
 69
Tab. 2.4.First Stage Results
 (1) (2) (3) (4) (5) (6)
 {Relative Di stance < 0m} = 1 0.17*** 0.10*** 0.10*** 0.20*** 0.14*** 0.14***
 (0.03) (0.02) (0.02) (0.03) (0.02) (0.02)   
Exogenous Covariates No No Yes No No Yes
 UPZ Fi x ed  Ef f ects No Yes Yes No Yes Yes
 Correct Pred i ction Rate 0.890 0.753 0.717 0.789 0.746 0.759
 F-Statistic 42.92 27.64 27.83 52.43 46.81 47.77
 N 33413 33413 33413 11968 11968 119968
 A. Full Sample B. UPZ Sample 
Dep en d ent: 1{CEC Student =1}
 Notes: Panel A includes observations from the full sample. Panel B includes students only from the UPZ sample; that is,
 those whose residence is located in UPZ nearby CEC schools. Relative distance is defined as distance to the closest CEC
 school minus distance to the closest TPS. The excluded instrument for CEC attendance is an indicator variable equal to 1 if
 relative distance is less than or equal to zero. The exogenous covariates are described in Table 2.2. Significance Levels: ***
 1%, **5%, *10%. 5000-repetition, paired bootstrap standard errors sampled at the cluster level in parentheses.
 Tab. 2.5.OLS and 2SLS Results of CEC Attendance on ICFES Scores
 (1) (2) (3) (4) (5) (6)
 OLS 0.101*** 0.136*** 0.120*** 0.150*** 0.159*** 0.134***
 (0.032) (0.030) (0.030) (0.035) (0.035) (0.035)
 2SLS 0.044 0.525*** 0.559*** 0.309*** 0.600*** 0.609***
 (0.119) (0.191) (0.201)   (0.096) (0.134) (0.133)   
p-value Hausman 0.63 0.05 0.03   0.11 0.00 0.00
 OLS 0.088** 0.145*** 0.131*** 0.162*** 0.179*** 0.159***
 (0.038) (0.033) (0.036) (0.039) (0.038) (0.038)
 2SLS -0.089 0.156 0.208   0.195 0.239 0.260
 (0.145) (0.202) (0.213)   (0.146) (0.168) (0.169)
 p-value Hausman 0.22 0.96 0.71   0.82 0.72 0.54
 OLS 0.066** 0.113*** 0.098*** 0.126*** 0.137*** 0.114***
 (0.029) (0.026) (0.028) (0.031) (0.029) (0.029)
 2SLS -0.060 0.307** 0.347** 0.206* 0.333*** 0.343***
 (0.122) (0.151) (0.162)   (0.107) (0.110) (0.109)   
p-value Hausman 0.30 0.17 0.10   0.42 0.05 0.03   
Covariates No No Yes No No Yes
 UPZ FE No Yes Yes No Yes Yes
 N 33413 11968
 Panel A.   Full Sample Panel B.   UPZ Sample
  Math Test Score in Standard Deviations
  Verbal Test Score in Standard Deviations
  Overall Test Score in Standard Deviations
 Notes: See notes to Table 2.4 for details on samples and exogenous covariates. Relative distance
 is defined as distance to the closest CEC school minus distance to the closest TPS. The excluded
 instrument for CEC attendance is an indicator variable equal to 1 if relative distance is less than
 or equal to zero. 5000-repetition, paired bootstrap standard errors sampled at the cluster level in
 parentheses. Significance Levels: *** 1%, **5%, *10%.
 70
Tab. 2.6.Placebo Test: Reduced Form Effects of Distance
 (1) (2) (3) (4) (5) (6)
 ICFES 2000 -0.064*** -0.032* -0.026 -0.018 -0.017 -0.040   
(0.019) (0.016) (0.023) (0.032) (0.031) (0.046)   
ICFES 2008 0.007 0.053*** 0.057*** 0.062*** 0.085*** 0.086***
 (0.020) (0.020) (0.020) (0.017) (0.018) (0.018)   
ICFES 2000 -0.097*** -0.051*** -0.025* -0.020 -0.034 -0.029   
(0.016) (0.015) (0.014) (0.032) (0.035) (0.034)   
ICFES 2008 -0.015 0.016 0.021 0.039 0.034 0.037   
(0.026) (0.020) (0.021) (0.028) (0.023) (0.024)   
ICFES 2000 -0.136*** -0.084*** -0.050** -0.021 -0.047 -0.049   
(0.025) (0.024) (0.022) (0.050) (0.054) (0.048)   
ICFES 2008 -0.010 0.031** 0.035** 0.041** 0.047*** 0.049***
 (0.021) (0.015) (0.016) (0.020) (0.014) (0.014)   
Covariates No No Yes No No Yes
 UP Z FE No Yes Ye s No Yes Ye s
 N  ICFES 2000
 N ICFES 2008
  Overall Test Score in Standard D eviations
 20533 3776
 33413 11968
 Panel A.   Full Sample Panel B.   UPZ Sample
  Math Test Score in Standard D eviations
  Verbal Test Score in Standard D eviations
 Notes: This table shows reduced form effects of distance to math, verbal, and composite test scores using
 2000 and 2008 ICFES data. The year 2000 is one year before the first CEC schools graduated their first
 high school class. The instrument for both years is an indicator variable equal to 1 if the student lives
 relatively closer to a CEC school than to a TPS. For the year 2000, the instrument is constructed as the
 distance to the closest not-yet-available CEC school minus the distance to the closest TPS.
 71
Tab. 2.7.Treatment Effect Heterogeneity
 Lowest 2nd 3rd Highest
 5.62 5.25 5.03 4.74
 0.75 0.50 0.25 0.04
 0.65 0.42 0.19 0.04
 0.33 0.15 0.08 0.03
 0.30 0.16 0.08 0.02
 0.08 0.17 0.26 0.44
 0.86 0.37 0.10 0.02
 0.19 0.37 0.55 0.87
 0.05 0.09 0.16 0.33
 0.03 0.10 0.21 0.61
 3.91 4.12 4.34 4.64
 8706 8428 8248 8031
 Quartile W q l q|x DD q|x wq|x rq
 math
 rq
 v erbal
 rq
 Composite
 1 0.260 0.077 0.075 0.247 0.605 0.216 0.398
 (0.020) (0.299) (0.270) (0.218)
 2 0.252 0.065 0.124 0.334 0.563 0.170 0.375
 (0.026) (0.185) (0.198) (0.160)
 3 0.247 0.055 0.108 0.243 0.396 0.012 0.094
 (0.025) (0.223) (0.261) (0.172)
 4 0.240 0.042 0.106 0.176 0.693 0.549 0.557
 (0.022) (0.267) (0.272) (0.235)
 Internet at Home = 1
 Rooms in dwelling
 N
 Panel B. D ecomposition of IV Weighting by SES Index
 No. Household M em bers
 Father Education: Primary =  1
 Mother Education: Primary =  1
 Stratum 1 = 1
 House Income:  1 MW or less
 House Income:  2 < M W < 3
 Quality Floor Materials: Medium =1
 Computer at Home = 1
 Car at Home
 Panel A. Socio-Economic Characteristics by SES Index
 SES Index  Q uartiles
 Notes: This table decomposes IV CEC school attendance estimates by quartiles of a
 constructed socio-economic index. Index reflects the predicted probability of being
 a SISBEN 1 household using demographic characteristics used in estimation for the
 sample living closer to a TPS than to a CEC school. The SISBEN is a government
 poverty-targeting index that takes on 6 values, where only levels 1 and 2 are eligible
 for most government subsidies. Wq = P(Quartile = q) is the proportion of sample
 observations in quartile q = f1, 2, 3, 4g. lqjx = E[P(ZjX, Q)(1 P(ZjX, Q))jQ] is
 the expected conditional variance of Z for each quartile, where Z is the categorical
 relative distance instrument. DDqjx = E[E(DjZ = 1, X, Q) E(DjZ = 0, X, Q)jQ] is
 the average effect of the instrument on CEC attendance for a given quartile. wqjx =
 (WqlqjxDDqjx)/( q WqlqjxDDqjx) is the final weight received by quartile q.
 72
Tab. 2.8.School Inputs
 P ublic CEC P ublic CEC P ublic CEC
 T eachers' Education
 L ess than College 0.07 0.10 0.03 0.11 0.02 0.07
 Major in Education 0.48 0.72 0.45 0.76 0.41 0.74
 Other Major 0.04 0.01 0.05 0.01 0.06 0.03
 Graduate in Education 0.37 0.15 0.43 0.11 0.45 0.15
 Other Graduate 0.02 0.00 0.04 0.00 0.04 0.00
 Student-T eacher Ratio 40.37 38.09 31.3 31.72 30.26 20.87
 (24.5) (6.37) (11.74) (8.6) (27.23) (5.95)
 Cost Per Student Ratio
 P ublic /  CE C 1.11
 Primary School Middle School High School
 2001 2002 2003
 1.01 0.99
 Notes: Author?s Calculations from School Survey C-600 for the year 2008 for teachers education and
 student-teacher ratios. Source for cost per student is Contralor?a Distrital, 2003.
 73
Appendix A. Figures
 Fig. 2.4.UPZs with CEC Schools
 Notes: This figure shows the 120 UPZs in Bogot?. UPZs are administrative subdivisions of the city used for urban planning
 purposes. The figure also shows UPZs with nearby CEC schools used to define the UPZ sample used in regression analyzes.
 0.1.2.3
  
-1255
 -900
 -600
 -300
 0
 300
 600
 900
 1200
 1500
 1800
 2100
 2400
 2700
 3000
 3300
 3600
 3900
 4200
 4500
 4800
 5100
 5400
 5700
 6000
 6300
 6600
 6900
 7200
 7500
 >7800
 30
 0-
 M
 et
 er
  R
 el
 at
 iv
 e 
D
 is
 ta
 nc
 e 
In
 te
 rv
 al
 s
 P
 ro
 po
 rti
 on
  o
 f C
 E
 C
  S
 tu
 de
 nt
 s 
w
 ith
 in
  R
 D
  In
 te
 rv
 al
 P
 ro
 po
 rti
 on
  o
 f T
 ot
 al
  S
 tu
 de
 nt
 s
 G
 ra
 ph
  1
 . P
 ro
 po
 rti
 on
  o
 f S
 tu
 de
 nt
 s 
by
  In
 te
 rv
 al
 s 
of
  R
 el
 at
 iv
 e 
D
 is
 ta
 nc
 e 
(R
 D
 )
 Fig. 2.5.Non-linearity of RD on CEC Attendance Probability
 Notes: This figure shows the fraction of total students attending a CEC school who live within a 300-meter interval of relative
 distance. Relative distance is defined as the distance to the closest CEC school minus the distance to the closest TPS.
 74
.2.4.6.81
  
0.1.2
  
0
 600
 1200
 1800
 2400
 3000
 3600
 4200
 4800
 5400
 6000
 6600
 7200
 7800
 8400
 9000
 9600
 10200
 30
 0-
 M
 et
 er
  A
 ct
 ua
 l D
 is
 ta
 nc
 e 
In
 te
 rv
 al
 s
 C
 E
 C
  S
 tu
 de
 nt
 s
 P
 ub
 lic
  S
 tu
 de
 nt
 s
 C
 E
 C
  +
  P
 ub
 lic
  C
 D
 F
  (
 R
 ig
 ht
  S
 ca
 le
 )
 G
 ra
 ph
  1
 . P
 ro
 po
 rt
 io
 n 
of
  S
 tu
 de
 nt
 s 
by
  In
 te
 rv
 al
 s 
of
  A
 ct
 ua
 l D
 is
 ta
 nc
 e
 Fig. 2.6.Proportion of Students by Intervals of Actual Distance
 Notes: This figure shows empirical distributions of distance in meters to actual school attended for both CEC and TPS
 students. The gray dashed line depicts the joint empirical cumulative distribution for CEC and TPS students, measured in
 the right scale. Approximately 15% of all students in the sample live within 300 meters (0.18 miles) from the school they
 attend. 73% of all students in the sample live 2000 meters (1.24 miles) or less (right scale) from the school they actually
 attend.
 Fig. 2.7.Location of Public and CEC Students? Residence
 Notes: This figure shows the location of all public and CEC schools in the city as well as the students residences by type of
 school attended. CEC and TPS students? residences are depicted with dark and light gray dots, respectively. CEC and TPS
 locations are represented with crossed black dots and hollow circles, respectively.
 75
75%
 80%
 85%
 90%
 95%
 100%
 1 2 3 4 5
 School ID
 Transportation to School
 Walk Bus Bike
 0%
 20%
 40%
 60%
 80%
 100%
 1 2 3 4 5
 School ID
 Commutin g Time to School
 0- 10 Min 11- 20 Min 21- 30 Min
 Fig. 2.8.Commuting Patterns of 5 CEC Schools
 Notes: This figure describes commuting habits for students from 5 of the 25 CEC schools. Source: Alianza Educativa.
 0.02.04.06.08.1
  
-1500
 -1200
 -900
 -600
 -300
 0
 300
 600
 900
 1200
 1500
 1800
 2100
 2400
 2700
 3000
 3300
 3600
 3900
 4200
 4500
 4800
 5100
 5400
 5700
 6000
 6300
 6600
 6900
 7200
 7500
 7800
 30
 0-
 m
 et
 er
  R
 el
 at
 iv
 e 
D
 is
 ta
 nc
 e 
In
 te
 rv
 al
 s
 20
 00
 20
 08
 G
 ra
 ph
  1
 . P
 ro
 po
 rt
 io
 n 
of
  S
 tu
 de
 nt
 s 
by
  In
 te
 rv
 al
 s 
of
  R
 el
 at
 iv
 e 
D
 is
 ta
 nc
 e
 Fig. 2.9.Distribution of Relative Distance in 2000 and 2008
 Notes: This figure describes the empirical distributions of relative distance for the years 2000 and 2008. The year 2000 is one
 year before the first CEC schools graduated their first high school class. For the year 2000, relative distance is constructed as
 the distance to the closest not-yet-available CEC school minus the distance to the closest TPS.
 76
Appendix B. Full Set of Estimation Results
 Tab. 2.9.First Stage Results
 (1) (2) (3) (4) (5) (6)
 {Relative D istance < 0m} = 1 0.17*** 0.10*** 0.10*** 0.20*** 0.14*** 0.14***
 (0.03) (0.02) (0.02) (0.03) (0.02) (0.02)   
Male=1 0.00 0.00   
Ag e -0.00* -0.01   
No. Household Mem bers 0.00 0.00   
Father Education: Primary  = 1 -0.02** -0.03*  
Father Education: Se condary = 1 -0.00 -0.00   
Father Education: As sociate = 1 0.00 0.02   
Mother Education: Primary  = 1 -0.01 -0.01   
Mother Education: Se condary = 1 -0.01 -0.01   
Mother Education: As sociate = 1 -0.01 -0.01   
Stratum 1 = 1 0.01 0.09*  
Stratum 2 = 1 0.01 0.08*  
Stratum 3 = 1 -0.02 0.04   
House I ncome: 1 MW or les s 0.01 0.03   
House I ncome: 1 < M W < 2 0.01** 0.05** 
House I ncome: 2 < M W < 3 0.01** 0.04***
 Dad is  Literate -0.00 -0.01   
Mom is  Literate 0.02* 0.03*  
House Floors - Low Q uality -0.02*** -0.04***
 House Floors - M ed Q uality -0.01 -0.02** 
Computer at Home 0.00 0.00   
DVD  at Home -0.00 -0.01   
Car Owners hip 0.01 0.01   
Motorbik e Ow ners hip 0.01 0.01   
No Cellphone in Household -0.01* -0.02   
One C ellphone in Household -0.00 -0.02   
I nternet at Home 0.00 0.00   
Cable TV at Home -0.01* -0.01   
Number of Rooms in Household -0.00 0.00   
UPZ Fix ed Effects No Yes Yes No Yes Yes
 Adj R-Squared 0.06 0.11 0.11 0.07 0.1 0.11
 Correct Prediction Rate 0.890 0.753 0.713 0.789 0.746 0.762
 F-Statistic 42.92 27.64 27.92 52.43 46.81 47.27
 N 33413 33413 33413 11968 11968 119968
 Dependent: 1{CEC  Student =1}
 A.  Full Sample B.  UPZ Sample 
Notes: See notes to Table 2.4 for details on samples and exogenous covariates. Significance Levels: *** 1%,
 **5%, *10%.
 77
Tab. 2.10.Full Set of 2SLS Results
 Math Verbal Overall Math Verbal Overall
 (1) (2) (3) (4) (5) (6)
 CEC  School = 1 0.553*** 0.198   0.338** 0.610*** 0.256   0.337***
 Male=1 0.296*** -0.007   0.135*** 0.293*** -0.028** 0.127***
 Ag e -0.073*** -0.074*** -0.075*** -0.064*** -0.065*** -0.064***
 No. Household Mem bers -0.013*** -0.016*** -0.015*** -0.012** -0.018*** -0.016***
 Father Education: Primary  = 1 -0.124*** -0.205*** -0.175*** -0.103** -0.139*** -0.144***
 Father Education: Se condary = 1 -0.109*** -0.164*** -0.145*** -0.068   -0.081*  -0.096** 
Father Education: As sociate = 1 0.025   -0.024   -0.003   0.085   0.047   0.037   
Mother Education: Primary  = 1 -0.124*** -0.125*** -0.123*** -0.088   -0.127** -0.112** 
Mother Education: Se condary = 1 -0.075*** -0.094*** -0.084*** -0.054   -0.110** -0.083** 
Mother Education: As sociate = 1 0.023   0.019   0.005   0.013   0.024   0.005   
Stratum 1 = 1 -0.107*  -0.104*  -0.122** 0.100   0.110   0.119   
Stratum 2 = 1 -0.083   -0.042   -0.065   0.133   0.176   0.187   
Stratum 3 = 1 -0.014   0.028   0.001   0.174   0.269   0.251   
House I ncome: 1 MW or les s -0.199*** -0.194*** -0.184*** -0.212*** -0.193*** -0.180***
 House I ncome: 1 < M W < 2 -0.099*** -0.121*** -0.103*** -0.112*** -0.135*** -0.096***
 House I ncome: 2 < M W < 3 -0.061*** -0.051** -0.047*** -0.083** -0.081** -0.047*  
Dad is  Literate 0.054   0.038   0.041   -0.001   -0.020   -0.027   
Mom is  Literate 0.083*  0.084*  0.106*** 0.013   0.051   0.069*  
House Floors - Low Q uality -0.007   -0.047** -0.012   -0.019   -0.097*** -0.039   
House Floors - M ed Q uality -0.002   -0.013   -0.004   0.037*  0.008   0.021   
Computer at Home 0.073*** 0.077*** 0.078*** 0.097*** 0.080*** 0.088***
 DVD  at Home -0.039*** -0.028*  -0.031*** -0.037** -0.034   -0.029   
Car Owners hip -0.020   -0.032** -0.029** 0.000   -0.054** -0.040***
 Motorbik e Ow ners hip -0.098*** -0.068*** -0.096*** -0.057** -0.046   -0.068***
 No Cellphone in Household -0.019   -0.040   -0.032   -0.034   -0.050   -0.045   
One C ellphone in Household -0.047*** -0.023*  -0.031*** -0.045*  -0.030   -0.037** 
I nternet at Home 0.005   0.011   0.010   -0.017   0.007   0.003   
Cable TV at Home -0.044*** -0.060*** -0.056*** -0.044** -0.053*** -0.056***
 Number of Rooms in Household 0.012*** 0.009*** 0.011*** 0.009*  0.012*** 0.010***
 UPZ Fix ed Effects Yes Yes Yes Yes Yes Yes
 N
 D ependent:                                            
Test Score in SD
 A.  Full Sample B.  UPZ Sample
 33413 11968
 Notes: See notes to Table 2.5 for details on samples and exogenous covariates. Significance Levels: *** 1%, **5%, *10%.
 78
Appendix C. Calculating Relative Distance to Closest CEC School
 This appendix describes the methodology used to construct the relative
 distance instrument used in Chapters 2 and 3 of this dissertation. As indicated
 above, the proposed instrument requires finding the closest CEC and the closest
 TPS for each student in the sample, regardless of the actual school they attended.
 For this, two data sources are used. First, I use administrative data on students?
 residential addresses collected by ICFES at the time of test registration. With
 these addresses available, the idea is to obtain geographic information ?that
 is, a latitude and a longitude? for each one of these addresses. To do so, I
 use administrative data from UACD, a local government organization in Bogota
 responsible for keeping a comprehensive register of the metes-and-bounds of
 all real property of the city. This dataset includes details of ownership, tenure,
 precise location (GPS coordinates), area, value, and UPZ location.
 For these two datasets to be successfully merged by address, both ICFES
 and UACD addresses may be written in a standardized form. While the ad-
 dresses in the cadastral dataset are properly written (in fact, UACD is respon-
 sible for assigning legal address to properties), students ?and city residents in
 general? usually use several ways of writing them. As a result, approximately
 15% of students? addresses can be merged using the data reported by students.
 Fortunately, addresses in Bogota have a rather specific alpha-numeric structure
 that can be used to infer the proper address reported to ICFES by students.
 The general address structure in Bogota can be best described by a numer-
 79
ical cartesian plane where streets increase in number as they move away from
 the center. In principle, streets that run from east to west are known as calles and
 streets that run from north to south are known as carreras. Nevertheless, as the
 city is far from being a simple raster, some streets are also referred as Avenida,
 Diagonal, or Transversal. A common address in Bogota starts with the name of
 the street type (i.e., calle, carrera, avenida, diagonal, transveral) where the property
 is located. After the street type, most addresses have next a number with at most
 3 digits which may be followed by a single letter in the set [A-D,F-M,O-R,T-Z]
 as well as by the word BIS, which may also have a letter on its own too. Then,
 addresses located in the south part of the city have the word SUR and addresses
 un the east part of the city the word ESTE. Addresses located in the north or
 west sections of the city do not have anything. After this part of the address is
 written, another at-most 3-digit number must follow with the possibility of hav-
 ing another single letter in the set [A-D,F-M,O-R,T-Z] as well as by the word BIS.
 Then a final at-most 3-digit number follows. At this point there may also be ad-
 dresses with either the word SUR or the word ESTE. After this point, properties
 that are apartments or houses within a residential complex may have additional
 information such as the apartment number.
 To make things clear, an example of an address in Bogota can be: KR 5A
 SUR 24B 58 . An address like that, for example, is located on carrera 5A in the
 south-west quadrant of the city, 58 meters away from the intersection of the 5A
 carrera and the 24B calle. Note that the letters that follow the streets indicate
 that there are many streets with the same number and the letters are used to
 80
distinguish between them. For example, a property on the calle 5B is located in
 between calle 5A and calle 5C (or calle 6 in case there is no 5C).
 Taking advantage of this common structure, I wrote a stata program that
 makes extensive use of regular expressions in order to rewrite all student ad-
 dresses in the UACD format to maximize the number of merged addresses. The
 main reasons why addresses do not exactly match between these two datasets is
 because students include either more information than needed (e.g., variations
 of the word n?mero to separate numerical parts within the address, or the name
 if the neighborhood), or because spaces between numbers and letters do not
 precisely coincide (KR 5 A as opposed to KR 5A). The written program allows
 me to assign geographic coordinates (e.g., latitude and longitude) to 81% and
 85% of the students? residences for the years 2007 and 2008, respectively. As
 shown in Table 2.11, there are no significant differences in the SES characteris-
 tics of the matched and unmatched samples for the year 2008, suggesting, for
 example, that students with lower SES are not more likely to misreport their
 home address. Unfortunately, this analysis can only be conducted for a subset
 of the SES variables in 2007 since most of the SES variables were not collected
 in this year. Nevertheless, the results for 2007 also show that there are no signif-
 icant differences in students? age and gender between matched and unmatched
 samples.
 Once I am able to geo-reference student residences, I proceed to calculate
 the linear (i.e., euclidean) distance from each student?s home to all TPS and CEC
 schools in Bogota. To do this, I use the Hawths tool in ARC-GIS to calculate
 81
the distance between all the points of two datasets, the ICFES and the UACD
 address files. Then, for each student I find what is the closest CEC school as
 well as the closest TPS. Finally, I use these two measures to calculate the relative
 distance to the closest CEC school as the difference between the distance to the
 closest CEC school minus the distance to the closest TPS.
 82
Tab
 .2.11
 .Matche
 d
 an
 d
 Unmatche
 d
 Sample
 s
 by
 Schoo
 l
 Typ
 e
 Unmat
 ched
 Mat
 ched
 Unmat
 ched
 Mat
 ched
 Unmat
 ched
 Mat
 ched
 Unmat
 ched
 Mat
 ched
 Mal
 e
 0.4
 6
 0.4
 7
 0.4
 7
 0.4
 6
 0.4
 7
 0
 .46
 0.4
 3
 0.4
 6
 Age in Yea
 rs
 16
 .30
 16
 .30
 1
 7.1
 4
 17
 .07
 17
 .15
 17
 .08
 16
 .86
 1
 6.9
 9
 (1.4
 6)
 (1.5
 4
 )
 (1.4
 9)
 (1.3
 2
 )
 (1.5
 1)
 (1.3
 0)
 (1.0
 5)
 (1.7
 5)
 No
 . Ho
 useho
 ld Members
 4.9
 4
 5.1
 7
 4.9
 3
 5
 .16
 5.1
 6
 5.3
 2
 (1.8
 6)
 (2.0
 5
 )
 (1.8
 6)
 (2.0
 5)
 (1.8
 5)
 (2.0
 6)
 Fat
 her
 's 
Edu
 ca
 ti
 o
 n
 Primary
 0.4
 0
 0.3
 9
 0.3
 9
 0
 .39
 0.4
 0
 0.4
 3
 Seco
 ndary
 0.4
 8
 0.4
 7
 0.4
 7
 0
 .47
 0.5
 1
 0.4
 7
 Asso
 ciat
 e
 0.0
 8
 0.0
 8
 0.0
 8
 0
 .08
 0.0
 6
 0.0
 7
 So
 me Co
 llege o
 r Higher
 0.0
 5
 0.0
 5
 0.0
 5
 0
 .06
 0.0
 3
 0.0
 4
 Mo
 the
 r's
  E
 du
 ca
 ti
 o
 n
 Primary
 0.3
 3
 0.3
 3
 0.3
 3
 0
 .33
 0.3
 9
 0.3
 8
 Seco
 ndary
 0.5
 0
 0.5
 1
 0.5
 0
 0
 .51
 0.4
 9
 0.4
 9
 Asso
 ciat
 e
 0.1
 1
 0.1
 0
 0.1
 1
 0
 .11
 0.0
 8
 0.0
 9
 So
 me Co
 llege o
 r Higher
 0.0
 6
 0.0
 5
 0.0
 6
 0
 .05
 0.0
 4
 0.0
 5
 Str
 at
 u
 m
 One
 0.1
 7
 0.1
 5
 0.1
 7
 0
 .15
 0.2
 3
 0.2
 4
 Two
 0.5
 5
 0.5
 2
 0.5
 4
 0
 .51
 0.7
 1
 0.7
 0
 Three
 0.2
 7
 0.3
 1
 0.2
 8
 0
 .32
 0.0
 5
 0.0
 5
 Fo
 ur+
 0.0
 1
 0.0
 1
 0.0
 1
 0
 .01
 0.0
 0
 0.0
 1
 Ho
 use
 ho
 ld Inc
 o
 me
 1 
MW o
 r less
 0.1
 4
 0.1
 4
 0.1
 5
 0
 .14
 0.1
 3
 0.1
 3
 1 
< MW < 2
 0.5
 4
 0.5
 3
 0.5
 3
 0
 .53
 0.6
 0
 0.5
 9
 2 
< MW < 3
 0.2
 3
 0.2
 3
 0.2
 3
 0
 .23
 0.2
 2
 0.2
 3
 MW > 4
 0.0
 9
 0.0
 9
 0.0
 9
 0
 .09
 0.0
 4
 0.0
 5
 N 
82
 28
 35
 23
 6
 58
 6
 9
 33
 41
 3
 55
 71
 31
 78
 3
 29
 8
 16
 30
 2
 0
 0
 8
 All 
Samp
 le2
 0
 0
 7
 All 
Samp
 le
 TPS
 CEC
 Notes
 :
 Thi
 s
 tabl
 e
 compa
 re
 s
 socio-economi
 c
 characteristic
 s
 bet
 w
 ee
 n
 student
 s
 whos
 e
 residence
 s
 w
 er
 e
 successfull
 y
 geo-
 refe
 rence
 d
 (matched
 )
 an
 d
 thos
 e
 wh
 o
 w
 er
 e
 no
 t
 (unmatched)
 .
 Th
 e
 stratu
 m
 variable
 s
 is
 an
 inde
 x
 runnin
 g
 fr
 om
 1
 to
 6
 whic
 h
 summarize
 s
 th
 e
 socio-economi
 c
 statu
 s
 of
 th
 e
 d
 w
 ellin
 g
 an
 d
 it
 s
 neighbo
 rhood
 ,1
 bein
 g
 th
 e
 lo
 w
 est
 .
 Househol
 d
 incom
 e
 is
 measu
 re
 d
 in
 numbe
 r
 of
 monthl
 y
 minimu
 m
 w
 age
 s
 (MW)
 .
 83
3. LONGER-RUN EFFECTS OF THE CEC PROGRAM
 3.1. Introduction
 The question of whether private schools provide better education than pub-
 lic schools has been a long and much debated issue. This question, for exam-
 ple, is at the center of the debate over the effects on academic performance of
 programs such as vouchers, charter schools, or education management orga-
 nizations. Unlike voucher programs ?where the control over the quality of
 education provided by participating schools is mostly left to private users?
 other programs have been designed so that the state retains a fair measure of
 control over service delivery by monitoring performance, imposing penalties or
 by replacing providers in case of performance failure. To some extent, charter
 schools in the Unites States and in Alberta (Canada), Academies in England,
 Free schools in Sweden, and CEC schools in Bogota (Colombia) are good exam-
 ples of this second provision type. In general, these initiatives aim at broadening
 the decision-making autonomy of schools and promoting accountability, while
 remaining under ?admittedly varying levels of? public control.
 Previous research on the effects of this second provision type focuses mostly
 on the short-run effects of attending privately run public schools on test scores,
with most of the existing evidence coming from evaluating US charter schools
 with rather mixed results (Angrist et. al [2010]; Abdulkadiroglu et. al, [2009];
 Dobbie and Fryer, [2009]; Hoxby and Murarka, [2009]; Hoxby and Rockoff,
 [2004]; Saas, [2006]; Bifulco and Ladd, [2004]; Hanushek et. al, [2007]; Booker
 et. al [2007]). Nevertheless, missing from these studies is an assessment of
 longer-run outcomes of attending such schools, such as earnings and college
 enrollment, outcomes that are more clearly tied to economic success.
 While generating higher earnings capacity is one of the most important
 goals of education, using earnings as an outcome to evaluate school quality is
 empirically difficult for a variety of reasons, including the fact that actual earn-
 ings do not show up until many years after individuals have finished their sec-
 ondary education. As a result, standardized tests have been used as proxies for
 school quality, even though the existing evidence of the relationship between test
 scores and earnings is not conclusive. While some have argued in favor of a pos-
 itive relationship between test scores and individual earnings and productivity
 (e.g., Hanushek, [2006]; Lazear, [2003]), others have asserted that standardized
 tests can only partially measure a student?s ability and cannot measure students?
 deeper problem-solving skills (Card and Krueger, [1992]; Evans and Schwab,
 [1995]). Consequently, these authors have stressed the importance of looking at
 alternative measures of educational attainment such as completing high school
 or going on to college to evaluate school performance. Looking at college en-
 trance is an important outcome because of the large pecuniary gains derived
 from higher education and its potential effects on reducing income inequality.
 85
This chapter examines the longer-run effects of attending contracted-out
 public schools using data from Bogota?s CEC school program, previously dis-
 cussed in chapters 1 and 2. To the best of my knowledge, this study is the first
 attempt to look at the causal relationship between contracted out public schools
 and higher education outcomes.
 In chapter 2, I looked at the effects on ICFES test scores, a high-stakes
 standardized test taken by the vast majority of students in their last year of high
 school. In 2008, for example, 93% of TPS and 96% of CEC students, respectively,
 took the test. Taking the test is a compulsory requirement for students intending
 to enroll in college in Colombia and serves as the primary college admission
 criterion (ICFES, 1999). Therefore, it seems natural to ask whether CEC students
 have been able to translate higher ICFES test scores, for example, into higher
 college enrollment rates relative to Traditional Public School (TPS) students. The
 latter outcome is a relevant long term measure of school performance given that
 the returns to college in Colombia have been estimated to be twice as large as
 the returns to secondary schooling (World Bank, 2003).
 In addition to this outcome, I also consider whether students who attended
 CEC schools are more likely to be admitted at selective universities, as there is
 causal evidence for Colombia showing that students who attend selective uni-
 versities are more likely to have formal employment and have higher earnings
 than students attending non-selective institutions (Saavedra, [2008]). However,
 low-fee charging, selective, public universities in Colombia have managed to
 keep the number of admitted students low (World Bank, 2003). As a result, low-
 86
income students who traditionally graduate from TPS and CEC schools may
 find it difficult to compete with private school students for admission into se-
 lective, public institutions. Consequently, some public schools may be aiming to
 place their students in two-year vocational programs, specifically designed to in-
 crease students? odds of getting formal employment. In fact, some CEC schools
 have developed terminal vocational courses in order to smooth out their stu-
 dents? entrance to vocational programs at higher education institutions. Thus,
 as a third outcome, I look at the probability of enrolling in 2-year technical and
 technological programs relative to TPS students.
 Finally, I look at the college dropout behavior of CEC students. In Colom-
 bia, about 50% of the students that enroll in higher education institutions drop
 out from their programs (MEN, 2010). While part of this may be due to financial
 considerations, including inadequate student financial aid for low-income stu-
 dents, it is often argued that the low-quality of secondary education is also an
 important reason for dropping out of college.
 To construct these higher education outcomes, I use administrative data
 from the Ministry of Education on college enrollment which has student-level
 information on institution and program enrollment as well as dropout informa-
 tion. Given that students who attend CEC schools may differ in unobservable
 ways from public school students, I obtain causal estimates of CEC attendance
 by exploiting variation in distance from a student?s residence to the closest CEC
 institution to construct an instrument for CEC attendance. In chapter 2, I care-
 fully argued that this instrument is conditionally uncorrelated with unobserved
 87
determinants of academic achievement.
 OLS and 2SLS estimates indicate that, relative to TPS students, CEC stu-
 dents are substantially more likely to enroll in college in general and to enroll in
 two-year vocational programs in particular. Moreover, while CEC students score
 significantly higher in the ICFES test than TPS students, these higher scores do
 not seem to be sufficiently larger to effectively compete for admission at se-
 lective, public institutions in Bogota. I also provide some evidence that CEC
 students are not more likely to drop out from college relative to TPS students.
 The remainder of the chapter proceeds as follows: Section 3.2 provides
 background information on the CEC program and on the higher education sec-
 tor in Colombia. Section 3.3 describes the sample and the construction of the
 higher education outcomes. Section 3.4 presents the empirical strategy. Section
 3.5 provides the OLS and IV results and Section 3.6 concludes.
 3.2. Background
 The demand for secondary education increased dramatically in the last
 two decades in Colombia. In 2002, there were 1.6 millon 15- and 16-year-olds
 in the country, of whom 57% were attending a high school institution. By 2010,
 this gross enrollment rate reached 79%, even though the population in that age
 bracket also grew by 8 percent during that period (MEN, 2010). As a result,
 the number of high school graduates increased by 67% between 2002 and 2009.
 Interestingly, most of the increase in high school enrollment was primarily ab-
 88
sorbed by the public education system, with 81% of the high school students
 attending public schools, up from 70% a decade earlier.
 What these figures suggest is that both local and national education author-
 ities in Colombia have implemented policies to meet the increasing demand for
 public education. However, local governments have also been exploring policies
 aimed at improving the quality of public education. Colombia has consistently
 performed below average in international standardized tests relative to countries
 with similar levels of income. Indeed, country average results on the 2006 PISA
 test for reading, mathematics and science are slightly below the Latin American
 average and much lower relative to other emerging OECD countries. Alternative
 international evaluations such as TIMSS 1995, TIMSS 2006, and Pisa 2009 show
 similar results.
 Some of the implemented policies to either increase enrollment or improve
 quality involved the participation of the private sector in the provision of public
 education. The PACES program, for example, during the 1990s in Colombia is
 known for being one of the largest voucher school programs to have ever been
 implemented (Angrist, 2002). That program was an effort to increase secondary
 enrollment rates among low income families. In addition to PACES, the local
 government of Bogota implemented in 1999 the Colegios en Concesi?n (CEC) Pro-
 gram, a large-scale initiative, which contracted out the administration of all the
 new public schools constructed in the city in the period 2000-2003.
 As discussed in detail in chapters 1 and 2, the CEC program allowed con-
 tracted out schools to operate outside teacher collective bargaining provisions
 89
in return for increased school accountability, including test-based accountability.
 More specifically, the program made CEC schools subject to an annual evalua-
 tion based on a series of factors related to education quality including, for exam-
 ple, teachers? education and experience and student dropout and transfer rates.
 Also, CEC schools failing to meet academic standards based on the school?s
 average score on the overall ICFES test may be subject to contract rescission.
 Instrumental variable results in Chapter 2 show that CEC students? ICFES
 scores in 2008 are approximately 0.6 and 0.25 standard deviations higher in
 math and verbal test scores, respectively, relative to TPS students. Moreover, the
 results on ICFES composite test score1 indicate that attending a CEC school in-
 creases composite test scores by 0.3 standard deviations relative to TPS students.
 This latter finding is of interest for the present chapter given that the ICFES com-
 posite test score is the single most important criterion used by higher education
 institutions to determine admission (ICFES, 1999). Then, the idea in this chapter
 is to determine to what extent the higher ICFES test scores exhibited by CEC
 students have translated into better education outcomes for them relative to TPS
 students.
 3.2.1. Higher Education in Colombia
 In Colombia, the proportion of 18- to 24-year-olds enrolled in college in
 2010 reached 37%, which represents a 85% increase relative to the rate in 1999
 1 the ICFES test evaluates students on most of the curriculum subjects determined by the
 Colombian Ministry of Education, namely, Math, Reading, Biology, Chemistry, Physics, History,
 Geography, Philosophy, and English.
 90
(MEN, 2011) ?although, it is still below the Latin American rate which has an
 average rate of 44%. Despite these gains in college enrollment in the last decade,
 there remains a significant difference between the gross college enrollment rate
 of 37% and the high school gross enrollment rate of 79%. However, as is true
 almost everywhere, college enrollment in Colombia rises dramatically with in-
 come. According to Saavedra (2008), 40% of the 18- to 24-year-olds from the
 wealthiest income quintile are enrolled in college, compared to just 6% from the
 lowest quintile.
 There are 289 colleges and universities in Colombia, 72% of which are
 privately owned. About 44% of college students attend private institutions indi-
 cating that private institutions serve fewer students on average relative to public
 ones. There are four types of tertiary education institutions: universities (27%
 of the total number of institutions), university institutions (40%), technological
 institutions (18%), and technical training institutions (15%). While the first two
 categories serve as teaching and some as research institutions, technical and
 technological institutions provide short, non-academic programs that respond
 flexibly to labor market demands ?similar to two-year colleges in the United
 States. Whereas in 2000, 86% of higher education students were enrolled in ei-
 ther universities or university institutions (World Bank, 2003), by 2010 65% of
 them are. This suggests that most of the large increase in college enrollment in
 Colombia in the last decade has been absorbed by technical and technological
 institutions (MEN, 2011).
 There are 59 selective higher education institutions in Colombia, that is,
 91
institutions with more applicants than slots. As expected, the most selective
 institutions are those with higher expenditures per student and higher fractions
 of full time and PhD faculty and admit students with higher ICFES test scores
 (Saavedra, 2008).
 As noted above, students intending to enroll in college in Colombia must
 be high school graduates and must have taken the ICFES test. Moreover, a
 high ICFES score almost certainly leads to a place in high-return programs at
 selective universities. In fact, of the 59 selective universities in Colombia, 25 use
 the ICFES test exclusively to determine admission; the remaining schools use
 it in combination with other requisites (Saavedra, 2008). The test is conducted
 twice a year, usually in April and September, given that some high schools in
 Colombia end their academic year in June while others do so in December. As
 a result, higher education institutions also have two different entry periods each
 year.
 In Colombia, the private returns to each year of tertiary education reached
 22% in 2001. Moreover, workers with tertiary education earn 275% more than
 average worker and more than 6.5 times the wage of a worker with no education
 (World Bank, 2003). In comparison, college graduates in the US earned 55%
 more on average than high school graduates in 2009 (Becker et al, 2010).
 The joint existence of high returns to higher education and low enrollment
 rates suggests there may be some systematic barriers that prevent students from
 investing in higher education. Although a comprehensive characterization of
 this phenomenon is outside the scope of the present study, some studies have
 92
argued that the low and unequal access to higher education in Colombia is partly
 originated in the basic education system (Bloom and Hansen, 2003). That is, the
 low quality primary and secondary education may prevent students from being
 admitted at selective, high returns college and majors, leaving low-quality, low-
 return institutions as their only choice. Indeed, whereas inexpensive, selective,
 public institutions have been in strong demand in the last two decades, private,
 non-selective institutions have faced oversupply (World Bank, 2003).
 In addition, there is also evidence suggesting that the probability of drop-
 ping out of college in Colombia is highly correlated with poor academic back-
 ground. First, only 14% of college dropouts passed all attempted credits in the
 period 1998-2005. Second, the largest college dropout rates occur in the first two
 semesters of college. Indeed, out of 100 students enrolled in higher education,
 25 dropped out at the end of the first semester ?by far the largest dropout rate
 for a given semester. This may be a sign that students enrolling in college have
 not been sufficiently prepared. Third, the highest dropout rates occur among
 those with the lowest composite ICFES scores as well as those enrolled in ma-
 jors with higher requirements for high-school background such as engineering
 (MEN, 2006).
 Consequently, improvements in the quality of education received by low-
 income households at public schools should not only increase their odds of at-
 tending more selective institutions, but also the likelihood of staying enrolled
 while in college.
 93
3.3. Data and Descriptive Statistics
 The data used for this chapter are derived from three sources. First, I
 use administrative data from ICFES, which contains student level information
 on TPS and CEC students who took the test in the second semesters of 2007
 and 2008. For both of these years, I have individual scores for each one of the
 nine sections of the test, the student?s high school, age, gender, and residential
 address. For the year 2008, I also have information on students? parental edu-
 cation, residence stratum,2 household income, number of household members,
 and asset ownership (e.g., car, computer, cell phone, TV, etc.).
 The second source of data uses Bogota?s cadastral database, a comprehen-
 sive register of all the real properties in the city. This dataset ?which includes
 details of ownership, tenure, location (GPS coordinates), area, value, and UPZ
 location of all properties? allows me to assign geographic coordinates (i.e., lat-
 itude and longitude) to 81% and 85% of the students? residences for the years
 2007 and 2008, respectively. By geo-referencing students residences, I am able
 to calculate the distance from a student?s home to both actual and potential TPS
 and CEC schools in Bogota. As shown in Table 2.12 in Chapter 2, there are
 no economically significant differences in the SES characteristics of the matched
 and unmatched samples for the year 2008, suggesting, for example, that students
 2 The stratum is a six-level classification used by the local government to characterize the
 socio-economic characteristics of a given dwelling. The higher the stratum is, the better the
 socioeconomic status of the dwelling. In general, residents of dwellings in the first 2 strata are
 considered the urban poor and pay, for example, subsidized prices for public utilities such as
 water and electricity.
 94
with lower SES are not more likely to misreport their home address. Unfortu-
 nately, this analysis can only be conducted for a subset of the SES variables in
 2007 since most of the SES variables were not collected in this year. Neverthe-
 less, the results for 2007 also show that there are no significant differences in
 students? age and gender between matched and unmatched samples.
 The last data source uses information from SPADIES 2010, an administra-
 tive dataset from the Ministry of Education of Colombia that contains individual
 level information on college enrollment such as institution attended, major, date
 of first entry into higher education, and date of last semester the student is ob-
 served enrolled.
 Table 3.1 presents descriptive statistics for CEC and TPS students by year
 of high school graduation and college enrollment status. Some of these statistics
 are worth discussing. First, students enrolled in college have higher ICFES test
 scores than students who did not enroll. For example, for both 2007 and 2008
 and both school types (i.e., TPS and CEC), students attending a college insti-
 tution score 0.3 standard deviations higher in Math relative to students not in
 college. Second, as shown in columns 7 to 10, relative to non-college students,
 college students are younger, come from smaller families, have more educated
 parents, have a higher family income, and live in better neighborhoods as indi-
 cated by the stratum of the dwelling. In general, these figures provide evidence
 that college enrollment in Bogota is positively correlated with family income.
 Third, 34% of TPS students and 35% of CEC students are enrolled in higher
 education in 2007. It is rather surprising, however, that only 18% of TPS and CEC
 95
students were pursuing a college degree in 2008. It is not clear what the reason is
 for this lower college enrollment rate in 2008. Although the 2007 cohort has had
 three years to have enrolled in college by 2011 ?relative to two years for the 2008
 cohort? data on students? college entry dates shows this extra year explains
 little of the difference in college enrollment rates. That is, the 2007-cohort college
 enrollment rate is as high as 31% whenever the students who enrolled in college
 in their third year after high school graduation are considered as not enrolled. A
 more plausible explanation is that some higher education institutions have not
 reported their most recent information on enrollment to SPADIES. In any case,
 the striking difference in enrollment for these two cohorts cast some doubts on
 the use of the 2008 data on higher education. Consequently, in Section 5, I will
 be presenting separate results for 2007 cohort as well as joint results for the 2007
 and 2008 cohorts.
 Finally, I compare test scores and SES characteristics for TPS and CEC stu-
 dents who enrolled in college and whose residences are located in UPZs with
 nearby CEC schools (henceforth, the UPZ sample). UPZs are subdivisions of the
 city used by the local government for urban planning purposes and are defined
 in such a way that the urban and economic characteristics of the real proper-
 ties within them are similar. The descriptive statistics in column 12 show that
 CEC and TPS students have very similar SES characteristics, although CEC stu-
 dents are slightly more likely to live in low-strata neighborhoods. Nevertheless,
 CEC students exhibit higher ICFES scores than TPS students of about 0.15 to 0.2
 standard deviations.
 96
In Table 3.2, I also present descriptive statistics on college outcomes by
 high school graduation cohort and school type. Using observations from the
 full sample, one can see that higher education outcomes between TPS and CEC
 students in the full sample are fairly similar. First, TPS and CEC students have
 very similar college enrollment rates. Second, around 30% of both CEC and
 TPS students are enrolled at a selective institution.3 Third, they also have very
 similar major distributions with 2-year vocational programs (i.e., technical and
 technological) being the program most commonly chosen by students. Fourth,
 their ICFES composite test scores make them equally eligible for government
 sponsored ICETEX loans for higher education studies.4
 Fifth, based on students? ICFES scores and cut-offs used to determine ad-
 mission into engineering programs at the Universidad Distrital, I find that at most
 6% of TPS and CEC students are eligible for admission into selective, 5-year
 engineering programs at this institution. The Universidad Distrital is a low-fee,
 selective, public university in Bogota, although it is not the most selective, public
 university in the city.5 This shows that low-income students intending to enroll
 in selective, public institutions in Bogota face strong competition from private
 school students. Finally, CEC and TPS students have similar college dropout
 rates with a quarter or more dropping out at the end of their first semester of
 3 An institution is considered selective if the ratio of slots over applicants is less than one. I
 construct this ratio based on information from ICFES (2002)
 4 The ICFES score is just one of the criteria used to grant these loans. Students are also
 required to have guarantors with collateral, which in practice may have resulted in low-income
 students receiving a relatively low proportion of the granted loans.
 5 I use the Universidad Distrital as an example here since ?unlike other selective, public insti-
 tutions in Bogota? its admission process is only determined by ICFES test scores.
 97
enrollment and 40% or higher dropping out during their first college year.
 While the higher education statistics for the UPZ sample mostly reflect the
 patterns of the full sample, two differences are worth noting. First, CEC students
 are 9 percentage points more likely to enroll at a college institution in 2007.
 Second, CEC students are slightly more likely to be enrolled at a selective, public
 university than TPS students.
 3.4. Empirical Strategy
 I use the same identification strategy in this chapter as in Chapter 2, that
 is, an instrumental variables approach that uses distance to closest CEC school
 to construct an instrument for CEC attendance. As previously discussed, the
 algorithm used in Bogota to allocate students into public schools gives priority
 to families with the lowest socioeconomic status (SES) and, as a result, over
 subscribed schools could systematically get students with lower than average
 SES. On the other hand, students are allowed to apply to any school in the city,
 something that can lead to selection of most informed or motivated parents into
 better public schools. Whatever the case, an identification strategy must account
 for the potential selection of CEC program participants into CEC schools in order
 to estimate the causal effect of CEC attendance on higher education outcomes.
 A consistent estimate for the effect of CEC attendance on academic per-
 formance can be recovered if there is at least one variable Z that, in addition to
 being uncorrelated with unobserved determinants of college outcomes, is also a
 98
strong predictor of the probability of attending a CEC school. I exploit variation
 in distance from a student?s residence to the closest CEC institution to construct
 an instrument for CEC attendance. More specifically, I first calculate the distance
 to the closest CEC school relative to the closest TPS school, mostly to account
 for the fact that students who live far from a CEC school may still live even fur-
 ther away from a TPS and consequently are more likely to attend a CEC school.
 Thus, I define the instrument as a categorical variable taking the value of 1 if a
 student lives closer to a CEC school than to a traditional public school (i.e., for
 negative relative distances) and 0 otherwise. That is,
 Zi = 1fRDi < 0g, where RDi = Min
 c2CEC
 (dic) Min
 p2TPS
 (dip) (3.1)
 where dij is the distance from student?s i residence to school j 2 [c, p]
 The rationale behind the proposed instrument is that differential distances
 to CEC schools may induce quasi-experimental differences in transportation and
 opportunity costs of attending a given CEC school to students and their par-
 ents. These differences may be especially important for low-income students for
 whom the decrease in costs due to CEC proximity is what may induce choosing
 a CEC school. For the proposed instrument to be valid, it can only affect student
 outcomes through its effect on CEC attendance. Since CEC schools were located
 in low-income neighborhoods in Bogota, students who live closer to CEC schools
 are also more likely to have a lower SES relative to TPS students and, as a result,
 distance to CEC schools would also capture their lower socio-economic back-
 99
ground, violating the exclusion restriction. Indeed, in Table 3.1, columns 8 and
 10, the SES characteristics of CEC students who enrolled in college are slightly
 lower than those of TPS students. That is, CEC students have less educated par-
 ents, come from lower income families, and live in low-strata neighborhoods.
 Note also that CEC students have a lower SES than TPS students among those
 that did not enroll in college ?columns 7 versus 9.
 However, the SES characteristics of the 2008 CEC and TPS students who
 are enrolled in college and belong to the UPZ sample (columns 11 and 12) are
 remarkably similar. Moreover, correlations between each of the SES variables
 and the categorical instrument for the 2008 cohort (Table 2.4, chapter 2) suggests
 that the instrument is not correlated with SES variables in those specifications
 that control for either student UPZ fixed effects or include observations only
 from the UPZ sample. I interpret these results as evidence that the proposed
 relative distance instrument is exogenous conditional on UPZ fixed effects or
 conditional on residing in UPZs with nearby CEC schools (i.e., the UPZ sample).
 3.5. Results
 This section estimates 2SLS specifications of CEC attendance on higher
 education outcomes. To account for the possibility of correlation of student
 outcomes within a given neighborhood, all regression estimates are clustered at
 the UPZ level for the 2007 cohort and at the UPZ-year level for specifications
 with data from both 2007 and 2008 cohorts .
 100
3.5.1. First Stage Estimates
 Table 3.3 shows the estimates of linear probability models for CEC school
 attendance on the categorical relative distance instrument and different sets of
 exogenous regressors. For the 2007 cohort and using all observations in the
 sample, the probability of attending a CEC school is 16 percentage points higher
 whenever the closest school available to the student is a CEC school. In turn,
 using observations from the UPZ sample increases this probability by 26 per-
 centage points.
 Notice that whereas the inclusion of UPZ fixed effects slightly reduces the
 reported point estimates, controlling for SES student characteristics such as age,
 gender, and student?s residential stratum has little effect on the point estimates.
 That is, that the coefficient estimates and standard errors show little or no change
 across columns indicates there is little correlation between the instrument and
 the observable characteristics. This result supports the argument that the pro-
 posed instrument may be uncorrelated with unobserved characteristics as well.
 First stage estimates obtained from the joint 2007-2008 sample largely repli-
 cates the results of the the 2007 cohort. Notice that, whenever this sample is
 used, I also include a year fixed effect to account for differences in CEC at-
 tendance in time. Finally, note that the F-statistic on the excluded instrument
 in all specifications are well above the conventional levels used to assess finite
 sample bias from weak instruments and provides evidence on the strong associ-
 ation between the relative-distance based instrument and the probability of CEC
 101
attendance.
 3.5.2. 2SLS Estimates
 I now proceed to present ordinary least squares (OLS), reduced form (RF),
 and two-stage least square (2SLS) estimates of CEC attendance on higher edu-
 cation outcomes. In particular, I estimate the effects on ICFES test scores as well
 as the effects on the probability of college enrollment, the probability of being
 enrolled at a selective public institution, the probability of enrolling in a 2-year
 vocational program, and the probability of dropping out during the first year
 in college. Notice that in Table 3.4, I restrict the analysis to observations from
 the UPZ sample to better control for unobserved neighborhood characteristics.
 In addition to this, all specifications include UPZ fixed effects, age, gender, and
 student residential stratum as controls. For the joint 2007-2008 cohort specifica-
 tions, I also include a 2008 year fixed effect.
 In chapter 2, I provide causal evidence that CEC students graduating from
 high school in 2008 exhibited large and significant gains in test scores relative
 to TPS students. In the top panel of Table 3.4, I estimate the effects of CEC
 attendance for those graduating in 2007 as well as for the joint 2007-2008 cohorts.
 The point estimates for 2007 indicate that CEC students score 0.13 standard
 deviations higher on the ICFES test than TPS students. Similar point estimates
 are found for the math and verbal sections of the test. Note that these results
 are smaller than the ones found in Chapter 2 for the 2008 cohort and similar
 in magnitude to OLS estimates. In fact, the results using both 2007 and 2008
 102
cohorts show stronger results of ICFES scores in favor of CEC students mostly
 because of the large effects of the 2008 cohort. The lower estimates of the 2007
 cohort could be a result of a series of factors including CEC schools? learning
 curve. In addition, as discussed in detail in Chapter 2, it was only after 2007
 that a standardized and more comprehensive annual evaluation process of CEC
 schools? performance was implemented, including more demanding conditions
 in terms of ICFES test results. These evaluations have been conducted yearly
 since then to all CEC schools and allow for individual school comparisons over
 time.6
 Regarding higher education outcomes, I first report the estimates of CEC
 attendance on college enrollment. In 2007, CEC students are 12 percentage
 points more likely to enroll in college. This represents more than a 50% in-
 crease over the average college enrollment of 22% for this sample. For the joint
 2007-2008 cohorts, the increase in the probability of college enrollment of CEC
 students relative to TPS students is less pronounced, 4.6 percentage points, and
 less precisely estimated ?although similar in magnitude to the more precise
 OLS estimate. In any case, as stated before, it remains to be seen to what extent
 this lower estimated effect has been affected by the suspiciously low proportion
 of 2008 students found enrolled in higher education as discussed in Section 3.
 There is a rather low proportion of students in the sample who are enrolled
 at the selective public universities in Bogota, namely, the Universidad Nacional
 6 In future work, it will be of interest to determine how more recent cohorts of CEC students
 have performed in the ICFES test to further assess the evidence among these two plausible
 explanations.
 103
and the Universidad Distrital. This is not surprising given that these two uni-
 versities are the most highly selective, low-fee institutions in the city and, as a
 result, public high school students face strong competition from private school
 students. In any case, OLS and 2SLS estimates for the 2007 cohort indicate that
 CEC students are about 1 percentage point more likely to attend one of these
 institutions than TPS students ?a relatively large effect compare to an uncondi-
 tional mean of 3%.
 The low proportion of public high school students enrolled at selective,
 public institutions provides suggestive evidence that there is high demand for
 public, selective institutions in the city and, unless public school students have
 outstanding high ICFES scores relative to all college applicants, the probabil-
 ity of being admitted to programs at these institutions is quite low. Figure 3.1
 presents some evidence on this point. In particular, it shows kernel distribu-
 tions of all ICFES test takers in the second semester of 2008 in Bogota by school
 type. First, notice that ?despite large heterogeneity in test scores among pri-
 vate school students? the test score distribution for private school students is
 to the right of CEC and TPS distributions. As a result, private school students
 are more likely to be admitted at selective institutions thanks to having higher
 ICFES test scores. More importantly, note how similar the unconditional test
 score distributions for CEC and TPS are. However, in Chapter 2 I provided ev-
 idence that ?after accounting for negative selection into the CEC program?
 the average CEC student significantly outperforms the average TPS student on
 the ICFES test, especially for the 2008 cohort. What these two results indicate
 104
is that, although CEC schools attendance improve average performance on the
 ICFES test, these higher results are not sufficiently large so as to compensate for
 the negative selection of CEC participants so that they can compete with private
 school students for admission at selective institutions.
 Nevertheless, what the results on college enrollment indicate is that CEC
 students are significantly more likely to enroll in a college or university than TPS
 students, just not at selective institutions. I therefore consider, as a third outcome
 for higher education, the probability of attending a two-year vocational program
 by CEC students. The results indicate that CEC students are 8.8 and 6.5 percent-
 age points more likely to attend a vocational program than TPS students in the
 2007 and the 2007-2008 samples, respectively. These results represent more than
 an 85% increase when compared to the unconditional means of 2-year programs.
 The estimated results for this outcome are of interest because, according to the
 Ministry of Education, 74% of the graduates from technological programs in
 Colombia have formal employment ?just 3 percentage points less than gradu-
 ates from 5-year colleges and universities in the 2001-2009 period (MEN, 2011).
 Also, graduates from technology programs have earnings equal to 75% of the
 earnings of university graduates. Thus, these results suggest that attending a
 CEC school has important longer-run effects related to higher education, even
 though these effects are not so much obtained via higher ICFES test scores. Fi-
 nally, the evidence on the probability of dropping out in the first two semesters
 after enrolling in higher education suggests there is no statistically significant
 difference in this probability between CEC and TPS students.
 105
3.6. Conclusion
 In this chapter I estimate the causal effects of attending contracted out pub-
 lic schools on higher education outcomes. The issue of contracting out public
 schools has gained attention in the education literature as a potential alternative
 for the provision of public education with schools experiencing higher levels
 of autonomy while being contractually accountable for their students? perfor-
 mance. Given that the literature has mostly focused on the effects on standard-
 ized tests, this study is, to the best of my knowledge, the first attempt to look
 at the effects of these programs on longer-run outcomes and, in particular, on
 tertiary education outcomes.
 Given that in Bogota CEC students were not randomly allocated into schools
 and CEC schools were not randomly allocated in the city, I use an instrumental
 variables strategy that exploits exogenous variation in distance from a student?s
 residence to the nearest CEC school. In Chapter 2 and here, I provide evidence
 that the location of CEC schools was unrelated to the academic performance of
 potential students and that the proposed instrument is likely to be uncorrelated
 with unobserved determinants of academic performance.
 OLS and 2SLS estimates indicate that CEC students are substantially more
 likely to enroll in college in general and to enroll in two-year vocational pro-
 grams in particular. Moreover, while CEC students score significantly higher in
 the ICFES test ?the main criterion used to determine admission into higher ed-
 ucation in Colombia? than TPS students, these higher scores do not seem to be
 106
sufficiently higher to effectively compete for slots at selective, public institutions
 in Bogota. I also provide some evidence that CEC students are not more likely
 to drop out from college relative to TPS students.
 In this study, I provide some evidence on positive longer-run effects of
 attending a CEC school, especially the higher college enrollment rate of CEC
 students. The evidence also indicates that this higher participation is, to a great
 extent, explained by a higher probability of enrolling in 2-year vocational pro-
 grams. Estimates by the Ministry of Education suggests that graduates from
 vocational programs may enjoy better labor market outcomes in terms of the
 probability of being formally employed and earnings. However, admission into
 vocational programs imposes lower requirements in terms of ICFES test scores
 as admission is not as competitive. Therefore, what these results indicate is that,
 even though CEC students outperform TPS in the ICFES test, their higher college
 enrollment rates are not explained by their higher ICFES scores. This is consis-
 tent with the intention of some CEC schools of providing terminal vocational
 courses in order to smooth their students? entrance to vocational programs at
 higher education institutions.
 These results also suggest that perhaps there are other labor market out-
 comes such as wages and employment that will be relevant to evaluate the full
 impact of this program.
 Finally, it bears keeping in mind that the 2008 data on college enrollment
 seems to be incomplete given the striking difference in enrollment relative to
 2007. Thus, some of the results presented here ?particularly the ones using 2008
 107
data? are to be considered preliminary as this requires a better understanding
 of this discrepancy.
 108
Tab
 .3.1
 .S
 electe
 d
 Descripti
 ve
 Statistic
 s
 by
 Colleg
 e
 Attendanc
 e
 200
 7
 200
 8
 In
 Colleg
 e
 In
 Colleg
 e
 TP
 S
 CE
 C
 (UP
 Z
 Sample
 )
 TP
 S
 CE
 C
 (UP
 Z
 Sample
 )
 No
 t
 In
 Colleg
 e
 No
 t
 In
 Colleg
 e
 TP
 S
 CE
 C
 No
 t
 In
 Colleg
 e
 No
 t
 In
 Colleg
 e
 TP
 S
 CE
 C
 (1
 )
 (2
 )
 (3
 )
 (4
 )
 (5
 )
 (6
 )
 (7
 )
 (8
 )
 (9
 )
 (10
 )
 (11
 )
 (12
 )
 ICFE
 S
 Mat
 h
 Sco
 re
 in
 sd
 -0.1
 1
 0.2
 4
 -0.08
 0
 0.2
 2
 0.1
 3
 0.2
 0
 -0.08
 5
 0.3
 4
 0.02
 8
 0.4
 1
 0.3
 1
 0.4
 5
 (0.95
 )
 (1.06
 )
 (0.88
 )
 (1.12
 )
 (1.02
 )
 (1.13
 )
 (1.00
 )
 (0.95
 )
 (0.89
 )
 (1.07
 )
 (0.92
 )
 (1.05
 )
 ICFE
 S
 V
 erba
 l
 Sco
 re
 in
 sd
 -0.1
 4
 0.3
 2
 -0.1
 1
 0.2
 8
 0.1
 7
 0.2
 6
 -0.08
 7
 0.3
 5
 0.01
 1
 0.4
 1
 0.2
 9
 0.4
 8
 (0.97
 )
 (0.98
 )
 (0.96
 )
 (1.01
 )
 (0.98
 )
 (1.00
 )
 (1.02
 )
 (0.88
 )
 (0.87
 )
 (0.90
 )
 (0.88
 )
 (0.79
 )
 ICFE
 S
 Composit
 e
 Sco
 re
 in
 sd
 -0.5
 1
 -0.09
 5
 -0.4
 7
 -0.1
 4
 -0.2
 3
 -0.1
 6
 -0.4
 1
 0.01
 3
 -0.3
 3
 0.04
 7
 -0.03
 5
 0.09
 2
 (0.59
 )
 (0.69
 )
 (0.56
 )
 (0.71
 )
 (0.65
 )
 (0.70
 )
 (0.76
 )
 (0.62
 )
 (0.60
 )
 (0.69
 )
 (0.61
 )
 (0.61
 )
 I(Relati
 ve
 Distanc
 e
 to
 CE
 C
 <
 0)
 0.09
 4
 0.07
 5
 0.4
 7
 0.4
 4
 0.1
 3
 0.4
 7
 0.09
 0
 0.07
 0
 0.4
 3
 0.4
 1
 0.1
 6
 0.5
 3
 Mal
 e
 =
 1
 0.4
 6
 0.4
 9
 0.4
 8
 0.4
 9
 0.4
 6
 0.4
 8
 0.4
 6
 0.4
 8
 0.4
 6
 0.4
 9
 0.4
 7
 0.4
 8
 Ag
 e
 16.
 4
 16.
 0
 16.
 4
 15.
 8
 15.
 9
 15.
 8
 17.
 2
 16.
 6
 17.
 1
 16.
 6
 16.
 5
 16.
 5
 (1.61
 )
 (1.38
 )
 (1.39
 )
 (1.22
 )
 (1.81
 )
 (1.22
 )
 (1.33
 )
 (1.02
 )
 (1.60
 )
 (2.26
 )
 (1.09
 )
 (2.51
 )
 No
 .
 Househol
 d
 Member
 s
 .
 .
 .
 .
 .
 .
 5.2
 1
 4.9
 4
 5.3
 5
 5.1
 8
 5.0
 7
 5.2
 9
 No
 .
 Room
 s
 in
 Hous
 e
 .
 .
 .
 .
 .
 .
 4.1
 9
 4.4
 6
 4.2
 8
 4.4
 8
 4.4
 7
 4.5
 2
 Fathe
 r
 Educatio
 n
 Prima
 ry
 =
 1
 .
 .
 .
 .
 .
 .
 0.4
 2
 0.2
 6
 0.4
 5
 0.3
 1
 0.3
 5
 0.3
 1
 Fathe
 r
 Educatio
 n
 Seconda
 ry
 =
 1
 .
 .
 .
 .
 .
 .
 0.4
 7
 0.4
 8
 0.4
 5
 0.5
 2
 0.4
 9
 0.5
 5
 Mothe
 r
 Educatio
 n
 Prima
 ry
 =
 1
 .
 .
 .
 .
 .
 .
 0.3
 6
 0.2
 1
 0.4
 1
 0.2
 4
 0.2
 7
 0.2
 7
 Mothe
 r
 Educatio
 n
 Seconda
 ry
 =
 1
 .
 .
 .
 .
 .
 .
 0.5
 1
 0.5
 3
 0.4
 7
 0.5
 6
 0.5
 6
 0.5
 6
 Stratu
 m
 1
 =
 1
 0.1
 3
 0.06
 0
 0.2
 0
 0.1
 6
 0.1
 5
 0.1
 7
 0.1
 6
 0.08
 0
 0.2
 5
 0.1
 8
 0.1
 7
 0.2
 2
 Stratu
 m
 2
 =
 1
 0.4
 6
 0.3
 9
 0.6
 7
 0.6
 8
 0.6
 5
 0.6
 9
 0.5
 3
 0.4
 7
 0.6
 9
 0.7
 6
 0.6
 9
 0.7
 4
 Stratu
 m
 3
 =
 1
 0.3
 6
 0.4
 9
 0.1
 1
 0.1
 5
 0.1
 9
 0.1
 3
 0.3
 0
 0.4
 3
 0.05
 2
 0.05
 4
 0.1
 4
 0.03
 5
 Hous
 e
 Income
 :
 1
 M
 W
 or
 les
 s
 .
 .
 .
 .
 .
 .
 0.1
 6
 0.07
 8
 0.1
 5
 0.08
 8
 0.1
 1
 0.09
 2
 Hous
 e
 Income
 :
 1
 <
 M
 W
 <
 2
 .
 .
 .
 .
 .
 .
 0.5
 5
 0.4
 6
 0.6
 0
 0.5
 3
 0.5
 2
 0.5
 3
 Hous
 e
 Income
 :
 2
 <
 M
 W
 <
 3
 .
 .
 .
 .
 .
 .
 0.2
 2
 0.3
 0
 0.2
 1
 0.3
 0
 0.2
 8
 0.3
 2
 Compute
 r
 at
 Hom
 e
 =
 1
 .
 .
 .
 .
 .
 .
 0.4
 5
 0.6
 8
 0.4
 3
 0.6
 5
 0.6
 1
 0.6
 3
 Ca
 r
 Ownershi
 p
 =
 1
 .
 .
 .
 .
 .
 .
 0.1
 4
 0.2
 3
 0.1
 5
 0.2
 0
 0.1
 6
 0.2
 0
 Inte
 rne
 t
 at
 Hom
 e
 =
 1
 .
 .
 .
 .
 .
 .
 0.2
 0
 0.3
 8
 0.1
 7
 0.2
 8
 0.2
 8
 0.2
 7
 Obse
 rv
 ation
 s
 2198
 8
 1175
 9
 97
 1
 51
 8
 219
 6
 46
 5
 2578
 0
 600
 6
 133
 4
 29
 6
 161
 4
 22
 9
 P
 roportio
 n
 0.6
 6
 0.3
 4
 0.6
 5
 0.3
 5
 0.8
 1
 0.1
 9
 0.8
 2
 0.1
 8
 109
Tab. 3.2.Descriptive Statistics for Higher Education Outcomes
 2007 2008
 Full Sample UPZ Sample Full Sample UPZ Sample
 TPS CEC TPS CEC TPS CEC TPS CEC
 Enrollment and Majors
 Enrolled at Selective 0.28 0.28 0.27 0.28 0.31 0.35 0.31 0.36
 Enrolled at Selective and Public 0.18 0.22 0.19 0.21 0.22 0.26 0.25 0.27
 Two-year Technology Program 0.15 0.18 0.18 0.19 0.13 0.13 0.15 0.14
 Two-year Vocational Program 0.19 0.24 0.25 0.24 0.20 0.22 0.24 0.23
 Health 0.07 0.07 0.06 0.07 0.07 0.07 0.06 0.07
 Social Sciences 0.08 0.06 0.07 0.07 0.09 0.09 0.08 0.09
 Economics / Business 0.18 0.10 0.15 0.10 0.15 0.15 0.14 0.16
 Engineering / Architecture 0.17 0.13 0.13 0.13 0.18 0.16 0.15 0.16
 Education 0.12 0.16 0.13 0.16 0.14 0.15 0.15 0.14
 Student Loan and Selective Program Eligibility
 ICETEX Eligible for 5-year Program 0.80 0.77 0.74 0.76 0.83 0.83 0.81 0.83
 ICETEX Eligible for 2-year Program 0.96 0.95 0.94 0.95 0.97 0.97 0.97 0.97
 UD Eligible for Cadastral Engineering 0.15 0.16 0.10 0.16 0.15 0.17 0.14 0.18
 UD Eligible for Mechanical Engineering 0.05 0.05 0.03 0.05 0.05 0.06 0.04 0.06
 UD Eligible for Industrial Engineering 0.06 0.05 0.03 0.05 0.05 0.06 0.04 0.06
 UD Eligible for Industrial Technology 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00
 Dropout Behavior
 Drop out in 1st Semester 0.24 0.25 0.26 0.26 0.29 0.27 0.30 0.26
 Drop out in 1st or 2nd Semester 0.36 0.40 0.37 0.40 0.45 0.41 0.46 0.40
 Observations 11759 518 2196 465 6006 296 1614 229
 Proportion enrolled in College 0.35 0.35 0.25 0.34 0.19 0.18 0.15 0.18
 110
Tab
 .3.3
 .Firs
 t
 Stag
 e
 Result
 s
 200
 7
 2007-200
 8
 Ful
 l
 Sampl
 e
 UP
 Z
 Sampl
 e
 Ful
 l
 Sampl
 e
 UP
 Z
 Sampl
 e
 (1
 )
 (2
 )
 (3
 )
 (4
 )
 (5
 )
 (6
 )
 (7
 )
 (8
 )
 (9
 )
 (10
 )
 (11
 )
 (12
 )
 I(Relati
 ve
 Distanc
 e
 to
 CE
 C
 <
 0)
 0.16**
 *
 0.13**
 *
 0.13**
 *
 0.26**
 *
 0.24**
 *
 0.23**
 *
 0.17**
 *
 0.12**
 *
 0.12**
 *
 0.23**
 *
 0.19**
 *
 0.19**
 *
 (0.03
 )
 (0.03
 )
 (0.03
 )
 (0.05
 )
 (0.05
 )
 (0.04
 )
 (0.02
 )
 (0.02
 )
 (0.02
 )
 (0.03
 )
 (0.03
 )
 (0.02
 )
 I(Mal
 e
 =
 1)
 0.0
 0
 0.0
 1
 0.0
 0
 0.0
 0
 (0.00
 )
 (0.01
 )
 (0.00
 )
 (0.00
 )
 I(Stratu
 m
 1
 =
 1)
 0.0
 1
 -0.0
 1
 0.0
 1
 0.0
 3
 (0.02
 )
 (0.09
 )
 (0.01
 )
 (0.05
 )
 I(Stratu
 m
 2
 =
 1)
 0.0
 0
 -0.0
 2
 0.0
 1
 0.0
 3
 (0.02
 )
 (0.10
 )
 (0.01
 )
 (0.05
 )
 I(Stratu
 m
 3
 =1
 )
 -0.0
 1
 -0.06*
 *
 -0.0
 1
 -0.0
 3
 (0.01
 )
 (0.02
 )
 (0.01
 )
 (0.03
 )
 I(
 Y
 ear=2008
 )
 0.0
 1
 -0.
 00
 -0.0
 0
 -0.03
 *
 -0.06**
 *
 -0.06**
 *
 (0.01
 )
 (0.00
 )
 (0.00
 )
 (0.02
 )
 (0.02
 )
 (0.02
 )
 UP
 Z
 Fixe
 d
 E
 ffect
 s
 N
 o
 Y
 es
 Y
 es
 N
 o
 Y
 es
 Y
 es
 N
 o
 Y
 es
 Y
 es
 N
 o
 Y
 es
 Y
 es
 Exogenou
 s
 C
 ov
 ariate
 s
 N
 o
 N
 o
 Y
 es
 N
 o
 N
 o
 Y
 es
 N
 o
 N
 o
 Y
 es
 N
 o
 N
 o
 Y
 es
 Cor
 rec
 t
 P
 redictio
 n
 Rat
 e
 0.8
 9
 0.7
 1
 0.7
 1
 0.8
 1
 0.7
 4
 0.7
 5
 0.8
 9
 0.7
 2
 0.7
 0
 0.7
 9
 0.7
 1
 0.7
 2
 F-Statisti
 c
 of
 Exclude
 d
 Instrument
 s
 39.7
 2
 24.5
 9
 24.7
 3
 25.3
 3
 22.8
 6
 31.4
 6
 80.5
 5
 49.7
 6
 49.1
 3
 69.9
 9
 49.4
 7
 57.3
 7
 N
 3471
 0
 3471
 0
 3471
 0
 981
 7
 981
 7
 981
 7
 6734
 7
 6734
 7
 6734
 7
 2100
 8
 2100
 8
 2100
 8
 111
Tab. 3.4.OLS, Reduced Form, and 2SLS Estimates of CEC attendance
 2007 2007-2008
 Dependent Variable Mean OLS RF 2SLS Mean OLS RF 2SLS
 ICFES Test Scores
 Math Score in sd -0.071 0.114*** 0.021 0.092 -0.051 0.125*** 0.054*** 0.257***
 (0.040) (0.021) (0.090) (0.026) (0.015) (0.084)
 Verbal Score in sd -0.106 0.137*** 0.037 0.159 -0.078 0.145*** 0.036* 0.158
 (0.035) (0.027) (0.105) (0.026) (0.019) (0.102)
 Composite Score in sd -0.472 0.125*** 0.031* 0.133** -0.422 0.120*** 0.041*** 0.172***
 (0.026) (0.016) (0.059) (0.019) (0.012) (0.063)
 Higher Education Outcomes
 Enrolled in College 0.228 0.092*** 0.028* 0.120** 0.187 0.057*** 0.006 0.046
 (0.017) (0.015) (0.057) (0.012) (0.010) (0.048)
 Enrolled at Selective and Public 0.034 0.013*** 0.003 0.011 0.031 0.011*** 0.001 0.004
 (0.004) (0.004) (0.016) (0.004) (0.003) (0.017)
 Two-year Vocational Program 0.097 0.044*** 0.020* 0.088** 0.076 0.022*** 0.011* 0.065**
 (0.012) (0.011) (0.041) (0.007) (0.006) (0.030)
 Drop out in 1st or 2nd Semester 0.088 0.047*** 0.010 0.044 0.078 0.024*** -0.000 0.000
 (0.013) (0.012) (0.046) (0.009) (0.006) (0.033)
 UPZ Fixed Effects Yes Yes Yes Yes Yes Yes
 Exogenous Covariates Yes Yes Yes Yes Yes Yes
 N 9815 9815 9815 21006 21006 21006
 112
Fig. 3.1. ICFES Composite Score Distribution by School Types
 0
 .
 2
 .
 4
 .
 6
 .
 8
  
?
 3
 ?
 2
 ?
 1 0 1 2 3
 Composite ICFES Score in SD
 Private
 TPS
 CEC
 Notes: This figure presents kernel distributions of all ICFES test takers in the second semester of 2008 in Bogota.
 113
REFERENCES
 1. Abdulkadiroglu, Atila, Joshua Angrist, Susan Dynarski, Thomas Kane and
 Parag Pathak, (2009). "Accountability and Flexibility in Public Schools:
 Evidence from Boston?s Charters and Pilots". NBER Working paper 15549
 2. Altonji, Joseph G., Todd E. Elder, and Christopher R. Taber, (2005). "Selec-
 tion on observed and unobserved variables: Assessing the Effectiveness of
 Catholic Schools." Journal of Political Economy, 113, 151-184.
 3. Angrist, Joshua, Eric Bettinger, Erik Bloom, Elizabeth King, and Michael
 Kremer, (2002). "Vouchers for Private Schooling in Colombia: Evidence
 from a Randomized Natural Experiment". American Economic Review,
 Vol. 92, No. 5, 1535-1558.
 4. Angrist, Joshua, Eric Bettinger, and Michael Kremer, (2006). "Long-Term
 Educational Consequences of Secondary School Vouchers: Evidence from
 Administrative Records in Colombia". American Economic Review, Vol.
 96, No. 3, 847-862.
 5. Angrist, Joshua D., Susan M. Dynarski, Thomas J. Kane, Parag A. Pathak,
 and Christopher R. Walters (2010). "Who Benefits from KIPP?" NBER
 Working Paper 15740.
 6. Attanasio, Orazio P., Vincenzo Di Maro, and Marcos Vera-Hernandez, (2010).
 "Community nurseries and the nutritional status of poor children. Evi-
 dence from Colombia". Mimeo. University College London.
 7. Barera-Osorio, Felipe, (2006). "The Impact of Private Provision of Public
 Education: Empirical Evidence from Bogot??s Concession Schools" World
 Bank Policy Research Paper 4121 (February)
 8. Becker, Gary, William Hubbard, and Kevin Murphy (2010). "Explaining
 the Worldwide Boom in the Higher Education of Women" The Journal of
 Human Capital, Fall 2010.
 9. Benveniste, Luis, Richard Rothstein, and Martin Carnoy, (2003). "All Else
 Equal: Are Public and Private Schools Different?". RoutledgeFalmer ed.
10. Eric Bettinger, Michael Kremer, and Juan Saavedra, (2010). "Are Educa-
 tional Vouchers only Redistributive?". The Economic Journal, Vol. 120, No.
 546, F204-F228.
 11. Bifulco, Robert, and Helen Ladd, (2006). "The Impacts of Charter Schools
 on Student Achievement: Evidence from North Carolina," Education Fi-
 nance and Policy, 1 (1), 123-138.
 12. Blank, Rebecca, (2000)." When Can Public Makers Rely on Private Markets?
 The Effective Provision of Social Services". The Economic Journal, Vol. 110,
 March, C34-C49.
 13. Blom, Andreas and Thomas Hansen, (2003). "Economic Perspectives of
 Tertiary Education" in Tertiary Education in Colombia, The World Bank.
 14. Booker, Kevin, Scott Gilpatric, Timothy Gronberg, and Dennis Jansen (2007).
 "The Impact of Charter School Attendance on Student Performance", Jour-
 nal of Public Economics, 91, (5/6), 849-876.
 15. Card, David, (1995). "Using Geographic Variation in College Proximity to
 Estimate the Return to Schooling" in Aspects of Labour Economics: Essays
 in Honour of John Vanderkamp, edited by Louis Christofides, E. Kenneth
 Grant and Robert Swindinsky. University of Toronto Press.
 16. Card, David and Alan Krueger (1992). "Does School Quality Matter? Re-
 turns to Education and the Characteristics of Public Schools in the United
 States", Journal of Political Economy 100, 1-40.
 17. Chiang, Hanley, (2009). "How accountability pressure on failing schools
 affects student achievement". Journal of Public Economics, 93, 1045-1057.
 18. Contraloria Distrital, (2004). "Auditor?a Integral a la Secretar?a de Edu-
 caci?n."
 19. Cullen, Julie and Randall Reback, (2006). "Tinkering toward accolades:
 school gaming under a performance accountability system" in Gronberg.
 T. and D. Jansen (eds.) Improving School Accountability: Check-Ups or
 Choice, Advances in Applied Microeconomics 14. Amsterdam: Elsevier
 Science.
 20. Chubb, John E. and Terry M. Moe, (1990). "Politics Markets and Americat?s
 Schools", The Brookings institution Press, Washington.
 21. Deere, Donald and Wayne Strayer, (2001). Putting schools to the test:
 school accountability, incentives and behavior. Working paper, Department
 of Economics, Texas A&M University.
 22. Dobbie, W. and R, Fryer, (2009). "Are High Quality Schools Enough to
 Close the Achievement Gap? Evidence from a Social Experiment in Harlem".
 NBER Working paper 15473.
 115
23. Dombberger, Simon, and Paul Jensen, (1997). "Contracting Out by the
 Public Sector". Oxford Review of Economic Policy, Vol. 13, No. 4, 67-78.
 24. Duggan, Mark, (2004). "Does Contracting Out Increase the Efficiency of
 Government Programs? Evidence from Medicaid HMO?s". Journal of Pub-
 lic Economics, 88, 2549-2572.
 25. Duggan, Mark, (2000). "Hospital Ownership and Public Medical Spend-
 ing". Quarterly Journal of Economics, November, 1343-1373.
 26. Evans, William and Robert Schwab, (1995). "Finishing High School and
 Starting College: Do Catholic Schools Make a Difference?" The Quarterly
 Journal of Economics, Vol. 110, No. 4.
 27. Figlio, David and Lawrence Getzle, (2006). "Accountability, ability and
 disability: gaming the system?" in Gronberg, T. (ed.) Advances in Microe-
 conomics. Elsevier.
 28. Figlio, David N., (2006). "Testing, Crime And Punishment," Journal of
 Public Economics, 2006, Vol 90, May, 837-851.
 29. Freedman, Seth, (2010). "The Effect of Deregionalization on Health Out-
 comes: Evidence from Neonatal Intensive Care". Mimeo. Economics De-
 partment. University of Maryland.
 30. Gandour, M., Carolina Gonz?lez, and Maria Camila Uribe, (2000). "De-
 scentralizaci?n, Estructura de Gobernaci?n y Pol?ticas de Calidad para la
 Educaci?n de Bogot?". Misi?n de Reforma Institucional de Bogot?.
 31. Glaeser, Edward and Andrei Shleifer, (2001). "Not-for-profit Entrepreneurs".
 Journal of Public Economics, 81, 99-115.
 32. Grossman, Sanford, and Oliver Hart, (1986). "The Costs and Benefits of
 Ownership: A Theory of Vertical and Lateral Integration". Journal of Po-
 litical Economy, Vol. 94, 691-719.
 33. Hanushek, Eric (1986). "The Economics of Schooling: Production and Effi-
 ciency in Public Schools", Journal of Economic Literature, 24(3), 1141-1177.
 34. Hanushek, Eric (1996). "Measuring Investment in Education", Journal of
 Economic Perspectives, 10(4), 9-30.
 35. Hanushek, Eric (2006). "School Resources" in Handbook of the Economics
 of Education, Chapter 14, Vol. 2.
 36. Hanushek, Eric, John Kain, Steven Rivkin, and Gregory Branch. "Charter
 School Quality and Parental Decision Making With School Choice", Journal
 of Public Economics, 2007, 91 (5/6), 823-848.
 37. Hart, Oliver, and John Moore, (1990). "Property Rights and the Nature of
 the Firm". Journal of Political Economy, Vol. 98, 1119-1158.
 116
38. Hart, Olivert, Andrei Shleifer, and Robert Vishny, (1997). "The Proper
 Scope of Government: Theory and an Application to Prissons". Quarterly
 Journal of Economics, Vol. 112, no. 4, 1126-1161.
 39. Herbst, Chris and Erdal Tekin, (2010). "The Impact of Childcare Subsidies
 on Child Well-being: Evidence from Geographic Variation in the Distance
 to Social Services Agencies". NBER Workin Paper 16250.
 40. Hoxby, Caroline, (2003). "School Choice and School Productivity. Could
 school choice be a tide that lifts all boats?" in The Economics of School
 Choice, University of Chicago Press, 287-342.
 41. Hoxby, Caroline and Sonali Murarka, "Charter Schools in New York City:
 Who Enrolls and How They Affect Their Students? Achievement", NBER
 WP 14852, 2009.
 42. Hoxby, Caroline and Jonah E. Rockoff. 2004. "The impact of charter schools
 on student achievement. Mimeo. (November)
 43. Hsieh, Chang-Tai and Miguel Urquiola, (2002). "When Schools Compete,
 How Do They Compete? An Assessment of Chile?s Nationwide School
 Voucher Program". NBER Workin Paper 10008.
 44. ICFES (1999). "Antecedentes y Marco Legal del Examen de Estado". Bo-
 got?, Colombia.
 45. ICFES (2002). "Estad?sticas de la Educaci?n Superior". Subdirecci?n de
 Monitoreo y Vigilancia. Bogot?, Colombia.
 46. Imbens, Guido and Joshua Angrist (1994). "Identification and Estimation
 of Local Average Treatment Effects". Econometrica, Vol. 62, No. 2.
 47. Imberman, Scott, (2009). "The Effect of Charter Schools on Achievement
 and Behavior of Public School Students". Department of Economics, Uni-
 versity of Houston, Mimeo.
 48. Jacob, Brian, (2005). "Accountability, Incentives and Behavior: The Impact
 of High-Stakes Testing in the Chicago Public Schools". Journal of Public
 Economics, Vol 89, 761-796.
 49. Jacob, Brian, and Steven Levitt, (2003). "Rotten apples: an investigation of
 the prevalence and predictors of teacher cheating". Quarterly Journal of
 Economics Vol 118, 3, Aug, 843- 878.
 50. Johnson, Sylvia (1984). "Preparing Black Students for the SAT?Does It
 Make a Difference? An Evaluation Report of the NAACP Test Prepara-
 tion Project". Mimeo Howard University.
 51. Joskow, Paul, (1988). "Asset Specificity and the Structure of Vertical Rela-
 tionships: Empirical Evidence". Journal of Law, Economics, and Organiza-
 tion, Vol. 4, No. 1, Speing, 95-117
 117
52. Kessler, Daniel P. and Mark B. McClellan (2000). "Is hospital competition
 socially wasteful?" Quarterly Journal of Economics, 115, 577-615.
 53. Koretz, Daniel and Sheila Barron, (1998). The Validity of Gains in Scores
 on the Kentucky Instructional Results Information System (KIRIS). RAND,
 Santa Monica, CA.
 54. Kling, Jeffrey (2001). "Interpreting Instrumental Variables Estimates of the
 Returns to Schooling". Journal of Business and Economic Statistics, Vol.
 19, No. 3.
 55. La Porta, Rafael, and Florencio Lopez-de-Silanes, (1999). "The Benefits of
 Privatization: Evidence from Mexico". Quarterly Journal of Economics,
 November, 1193-1242.
 56. Ladd, Helen, (2002). "School Vouchers: A Critical View". Journal of Eco-
 nomic Perspectives, Vol 16, No 4, Fall, 3-24.
 57. Lakdawalla, Darius, and Thomas Philipson, (2006). "The nonprofit sector
 and industry performance". Journal of Public Economics, Vol. 90, Issues
 8-9, September, 1681-1698.
 58. Lavin, Henry, (2003). Public Schools Inc. in Frontline, PBS.
 59. McClellan, Mark, Barbara McNeil, and Joseph Newhouse, (1994). "Does
 more intensive treatment of acute myocardial infarction in the elderly re-
 duce mortality? Analysis using instrumental variables", JAMA, September
 21, Vol. 272, No. 11.
 60. Megginson, William, and Jeffry Netter, (2001). "From State to Market: A
 Survey of Empirical Studies on Privatization". Journal of Economic Litera-
 ture, Vol 39, June, 321-389.
 61. Milne, Robin, and Magnus McGee, (1992). "Compulsory Competitive Ten-
 dering in the NHS: A new Look at Some Old Estimates". Fiscal Studies,
 Vol. 13, Issue 3, August, 96-111.
 62. Ministerio de Educaci?n Nacional (2006). "Deserci?n Estudiantil: Priori-
 dad en la Agenda" in Educaci?n Superior. Bolet?n Informativo No. 7.
 63. Ministerio de Educaci?n Nacional (2010). "Revolucin? Educativa 2002-2010
 Acciones y Lecciones".
 64. Ministerio de Educaci?n Nacional (2011). "Estad?sticas de Educaci?n Su-
 perior" www.mineducacion.gov.co/sistemasdeinformacion/1735/ articles-
 212350_resumen.xls last visited June 8, 2011.
 65. Ravitch, Diane, (2010). "Why I Changed My Mind About School Reform".
 Wall Street Journal. Retrieved 2010-06-24.
 118
66. Saavedra, Juan. "The Learning and Early Labor Market Effects of College
 Quality: A Regression Discontinuity Analysis", Harvard Mimeo, (2008).
 67. Sass, Tim. "Charter Schools and Student Achievement in Florida", Educa-
 tion Finance and Policy, 2006, 1 (1), 123-138.
 68. Sarmiento A, C.E. Alonso, G. Duncan, and C.A. Garz?n, (2005). "Evalu-
 aci?n de la Gesti?n de los Colegios en Concesi?n en Bogot? 2000-2003".
 Documento de Trabajo. Departamento Nacional de Planeaci?n.
 69. Sass, Tim. "Charter Schools and Student Achievement in Florida", Educa-
 tion Finance and Policy, 2006, 1 (1), 123-138.
 70. Shleifer, Andrei, (1998). "State versus Private Ownership". Journal of Eco-
 nomic Perspectives, Vol 12, No. 4, Fall, 133-150.
 71. Snyder, Christopher, Robert Trost, and R. Derek Trunkey, (2001). "Reducing
 Goverment Spending with Privatization Competitions: A Study of teh De-
 partment of Defense Experience". The Review of Economics and Statistics,
 February, 83(1), 108-117.
 72. Rose-Ackerman, Susan, (1996). "Altruism, Nonprofits, and Economic The-
 ory". Journal of Economic Literature. Vol. 34, June, 701-728.
 73. World Bank (2003). "Tertiary Education in Colombia Paving the Way for
 Reform".
 119