Does Greater School Autonomy Make a Difference? Evidence from a Randomized Natural Experiment in South Korea Youjin Hahna* a

School of Economics, Yonsei University. Postal address: 50 Yonsei-ro, Seodaemun-gu, Seoul 03722, South Korea. Phone: +82 2 2123 2474. Email: [email protected]. * Corresponding author

Liang Choon Wangb b Department of Economics, Monash University. Postal address: Department of Economics, Monash University, Clayton, VIC 3800, Australia. Phone: +61 3 9905 5448. Email: [email protected].

Hee-Seung Yangc c

Department of Economics, Monash University. Postal address: Department of Economics, Monash University, Clayton, VIC 3800, Australia. Phone: +61 3 9905 2431. Email: [email protected].

Abstract We exploit the unique features of high schools in Seoul to study the effects of school autonomy on student outcomes. Under South Korea’s equalization policy, both private and public schools in Seoul admit students that are assigned randomly to them, receive equal government funding, charge identical fees, and use similar curricula. However, private schools have greater flexibility in personnel decisions, and their principals and teachers face stronger incentives to perform. We find that private high schools have better student outcomes than public high schools. Our results suggest that autonomy in personnel decisions explains the positive student outcomes in private schools.

Keywords: Private schools, public schools, randomization, school autonomy JEL Classification: I2, J2, H4

1.

Introduction The purported benefits of school autonomy underpin various recent school reform efforts

around the world, such as the establishment of charter public schools in the United States, free schools in England, independent public schools in Australia, and community-managed schools in many developing countries. Although some studies show that these recent reforms improve student outcomes, there is sparse evidence on the positive effects of school autonomy on student outcomes. In particular, most of these studies focus on recent reforms that often include a range of policy measures in addition to school autonomy, making it difficult to disentangle the effect of school autonomy. Thus, the focus of this paper is to understand whether giving schools greater autonomy would improve student outcomes in the longer term. In the current study, we exploit a randomized natural experiment in Seoul, South Korea, to understand the effects of giving schools greater autonomy in making personnel decisions on student outcomes. In 1974, the South Korean government implemented in Seoul what the country calls its ‘equalization policy’, and a lottery-based student enrolment system. High schools governed by this policy have several important features. First, the schools subject to the policy, whether privately owned or publicly owned (hereafter, private or public schools), receive equal government funding, charge the same fees, and follow the same national curriculum. Second, private schools maintain autonomy over their personnel decisions, while public schools do not. This autonomy allows private schools to recruit, retain, promote, retrench, remunerate, and organize a workforce more effectively and flexibly to achieve their educational objectives. Thus, motivated agents (i.e., school board members, principals, and teachers) can select or be selected into private schools that offer the best preference matches. In contrast, public schools do not directly control their workforce. Public school principals or teachers are recruited by the Seoul Metropolitan Office of Education as civil servants and rotate to a different high school every four years. Third, students are assigned randomly into equalization policy schools within their school districts.1 Although parents may choose to live in a neighborhood with high-quality schools, they do not have control over which schools their children will attend within the school district. Students are generally not allowed to transfer to another school within the same school district, and when students and their families move to another school district, they are reassigned

1

There are special-purpose and autonomous high schools that take priority in student selection and which operate outside the equalization policy (see Section 2 for details).

1

randomly to a school in the new district (Kang 2007). This setting contrasts with those in other countries, where randomization may be applied only to schools that face excess demand and to students who express school preferences through enrolment applications. Although the equalization policy removes various differences in factors that are commonly attributed to the positive effects of private and other forms of independent schooling, the autonomy unique to private schools allows them to vary resource allocations, incentive structures, and teacher compositions. Specifically, their principals and teachers face less job security and greater incentives to deliver good student outcomes. They are more likely to hold principals accountable and have a higher share of teachers with career and promotion concerns. Private schools also have a greater component of performance pay within teacher compensation and larger teacher salary dispersions, but they spend less per student, run larger classes, and have fewer experienced and highly educated teachers. Thus, it is an empirical question of whether private schools improve student outcomes. We find that private school students in Seoul are no more likely than public school students to drop out or graduate from high school, but they are 4.4 percentage points more likely to attend colleges and 66 percent less likely to be cited for disciplinary problems. The increase in college attendance rates is driven by the increase in four-year college attendance rates and the decrease in two-year junior college attendance rates. Private school students also outperform public school students in standardized examinations by roughly 0.07 to 0.13 standard deviations. Although we cannot pinpoint every dimension of school autonomy that explains the outcome differences, we show that private schools with a greater share of teachers with career and promotion concerns produce better student outcomes. We further rule out sorting of teachers across sectors, the longer history of private schools, single-sex schooling, the religious affiliation of private schools, and private tutoring as primary channels of the positive effects of private schools in Seoul. Our study is broadly related to two strains of research: studies that examine the effects of private and charter schooling and studies that examine the effects of school autonomy. While much research has been done on private and charter schooling, it does not precisely identify the effects of school autonomy on student outcomes. Although studies based on the random assignment of private school vouchers or oversubscribed charter school slots to low-income applicants show the positive effects of attending these schools on student outcomes, it is often

2

uncertain which precise aspects of these schools explain the outcome differences.2 When various studies compare outcomes between receivers and non-receivers of vouchers or enrolment slots, the estimated effects not only reflect differences in school autonomy, but also differences in student composition, peer quality, resources, and other dimensions of school quality between highly sought-after schools and default traditional public schools (e.g., Angrist et al. 2002, Angrist et al. 2006, Hoxby and Murarka 2009, and Abdulkadiroglu et al. 2011). There are few studies that specifically focus on the effects of school autonomy. Two earlier studies, Jimenez and Sawada (1999) and King and Ozler (2005), use non-experimental approaches to examine the effect of autonomy on student outcomes and find that students in the schools that are given greater autonomy have better test scores. Research that uses experimental and quasi-experimental settings to study the effects of school autonomy is scant. One notable exception is Clark (2009), which exploits a U.K. reform in 1988 that allowed public high schools to opt out of local authority control if they won a majority vote among parents. He shows positive effects of giving schools greater autonomy on achievement gains in a regression discontinuity framework. The effect of autonomy might be different in the less competitive, more regulated, and conventional school catchment area setting where the majority of public schools around the world operate. The equalization policy in Seoul is ideal for studying the effects of school autonomy in these other settings and complements Clark’s (2009) findings. More importantly, since the policy in Seoul had been in place for more than three decades and most of the estimates regarding school autonomy and charter schooling came from recent reforms, our findings shed light on the longer term impact of school autonomy. The policy experiment in Seoul indicates that even when high schools are guaranteed funding and enrolment, faced with little competition, and heavily regulated by the government, providing them with a high level of autonomy in their personnel decisions can improve a range of student outcomes.

2

Earlier observational studies—such as those of Coleman et al. (1982) and Alexander and Pallas (1983)—find that in the United States, private schooling is more effective in improving test scores than public schooling, even after controlling for the joint influences of private school choice and achievement. Later studies by Figlio and Stone (1999), Krueger and Zhu (2004), and Altonji et al. (2005a, 2005b), however, show the mixed effects of private and charter public schooling on achievement. In contrast, observational studies that focus on the effects of private or Catholic schooling on high school completion and college attendance—such as those of Evans and Schwab (1995), Neal (1997), Altonji et al. (2005a), and Vella (1999)—consistently show positive private school effects.

3

2.

Secondary Schools and Equalization Policy in Seoul The South Korean government first implemented the equalization policy among high

schools in Seoul and Pusan in 1974. In Seoul, the equalization policy removed the competitive high school entrance examination and introduced the random assignment of students across schools within school districts.3 In 2008, there were 208 high schools in Seoul subject to the equalization policy; these schools could be private or public, and coeducational or single-sex. Additionally, private schools can be religiously affiliated or secular.

2.1. School Types and Randomization in Seoul Although parents cannot select the preferred equalization policy high schools in which their children will be enrolled, parents with strong preferences for school quality have choices outside these schools. The government permits roughly 20 selective high schools in Seoul to operate outside the equalization policy and take priority in student selection. These selective schools are either special-purpose high schools that specialize in sciences, sports, arts, music, and foreign languages, or they are autonomous high schools. 4 They select students based on academic performance, may charge higher tuition, and enjoy a greater level of autonomy in designing and implementing their own school curriculum than private schools bound by the equalization policy (Paik 2013). Students can opt for these selective high schools before being subject to the lottery-based enrolment system, but they must attend a randomly assigned equalization policy high school if they fail to enter a selective high school. Thus, special-purpose and autonomous high schools function more like the typical private high schools in other countries, while the equalization policy private schools are essentially government-funded schools with some school autonomy. In Seoul, after special-purpose and autonomous high schools admit their students (roughly 5 percent of high school students), the remaining students are assigned randomly into the various general academic high schools within the 11 school districts. Because Seoul’s population density is high (i.e., 10 million people in a 605-km2 area), students need not travel far to attend one of 3

In South Korea, primary and middle school education are compulsory (i.e., up to grade 9). Although high school education is not mandatory, 99.7 percent of all middle school graduates entered high schools in 2008 (data sourced from the Statistical Yearbook of Education, Korean Educational Development Institute, available at http://cesi.kedi.re.kr/eng/publ/view?survSeq=2008&publSeq=2&menuSeq=0&itemCode=02&language=en#). 4 Special-purpose high schools are mostly private and were established after the 1970s, while autonomous schools were introduced in 2010.

4

the several equalization policy high schools within their school district. 5 Prior to 2010, new entrants into equalization policy high schools were assigned randomly into schools, unconditional on any potential school preference they had within the school districts; however, since 2010, school districts have partly taken into account the preferences of middle school students and their parents. 6 As we are interested in examining the causal effects of school autonomy, we focus on general academic high schools that operate under the equalization policy and the students who were admitted prior to 2010 when school choice was more restricted. The complete disregard of preferences in the randomization process was stringent to the extent that students must attend the high school randomly assigned to them, even if they were assigned to a private school with religious affiliation different from their own. This practice is not too surprising in the Korean context as the equalization policy was introduced when the country was still under Park Chung-Hee’s military dictatorship. The practice of religious schools offering religious courses to students who are not religious or have a different belief was more frequently questioned in recent years, eventually resulting in a lawsuit and a Supreme Court’s decision.7 As a result of the Supreme Court’s decision in 2010, religious private high schools that offer any religion course must offer options of other religious and non-religious courses to cater for students with a different belief. 8

2.2. Commonalities for public and private equalization policy schools With the introduction in the 1970s of the equalization policy, all historically private schools were added to South Korea’s existing system of centralized public school finance. Several commonalities were introduced to public and private schools, such as uniform and 5

As a point of comparison, the population density values of London and New York City are roughly 50 and 40 percent that of Seoul, respectively. 6 The Seoul Metropolitan Office of Education confirmed that for our sample period (2008–2010), the stated school preferences or any other factors (such as distance to school, siblings, and religion) were not considered in the randomization process in all other school districts. The exception is the small central school district called Jungbu comprising three administrative districts (Jongno-gu, Jung-gu, and Yongsan-gu), where stated school preferences are considered. Our results are not sensitive to dropping this school district (Table A1, Appendix I). 7 A student named Eui-seok Kang claimed that the policy disregarded individuals’ freedom of religion and was unconstitutional. In 2004, Kang was expelled from Daegwang High School, a Protestant-run private school in Seoul, for protesting against forced religious education. In 2005, he sued the school for its religious education and eventually won the case in the Supreme Court in 2010. For details of this court case and its implications see the discussion by Lee (2011). 8 For most religious private schools, students are usually given the choice to participate in religious services and classes. Non-Christian students could opt for self-studying while other students are taking their religious classes. This was the case even before the Supreme Court’s decision.

5

centralized policies over fees and tuition, curricula, and teachers’ salary scales. Although the tuition fees charged are low, families who cannot afford the fees are exempt from paying them. 9 Teachers must instruct students in accordance with the unified national curriculum, based on designated or certified textbooks (Kim et al. 2007). There is a minimum teacher qualification requirement for both public and private schools. An individual considering becoming a school teacher must first obtain a teacher certificate by graduating from an accredited teacher education institution and acquiring the required number of credit hours (Kim and Han 2002). The government fully funds the teacher salaries and operating expenditures of private schools, based on the standard budget required for equivalent public schools (Paik 2013). Both private and public school teachers are guaranteed equivalent salary schedules based on their experience and qualifications, and the structures of pension systems are also similar for private and public school teachers because they are both based on the Government Employees Pension system (Moon 2002).10 Principals in both public and private schools have control over their daily operations and how they allocate their overall budget and resources. For example, principals in both private and public schools can decide how to organize their classrooms and teachers.

2.3. Differences in Teacher Labor Markets and Personnel Decisions Public school principals in Seoul are appointed by the Seoul Metropolitan Office of Education. Principals begin their career as school teachers. To become a regular public school teacher, an individual holding a teacher certificate (and before reaching a certain age) must pass an open competition employment examination administered by the metropolitan and provincial offices of education (Kim and Han 2002). Before passing the employment examination, the person may be employed in a public school as a short-term contract teacher for no more than three years. Once passing the employment examination, the person’s tenure and status as a regular public school teacher is guaranteed. Promotion to a more senior position in a public school depends largely on the basis of the length of service and participation in performance development training (Kim and Han 2002). Public school principals have little control over their staffing decisions. Public school principals and teachers, are government employees recruited by the Seoul Metropolitan Office of Education and they must rotate to different schools every four 9

The annual tuition fee in 2009 for both public and private high schools was set at about 1,300 USD (1.45 million KRW). The admission fee was less than 15 USD (14,100 KRW) (source: http://www.law.go.kr). 10 Detailed information on Teachers Pension Law is available at http://www.law.go.kr.

6

years. Such a rotation system is not unique to Korea; similar systems are also present in countries such as Australia, Canada, China, and Japan. 11 Public school principals can work as regular teachers after their term as a principal ends.12 Thus, public school principals’ and teachers’ jobs are fairly secure. Private school principals are appointed by the owner or board of directors of the school (school board). The school board determines the appointment and term extension of principals.13 Private school principals recommend to the school board whom they wish to hire as teachers and the length of the teachers’ contracts (i.e., short-term teachers or regular teachers); the school board then makes approvals. To become a private school teacher, a person holding a teacher certificate directly applies to private schools and is hired on a case-by-case basis. There is no examination requirement for a private school position. Individual private schools determine their own terms of employment. Private schools have flexibility in promoting teachers from shortterm contract teachers to regular teachers and from regular teachers to high-paying senior administrative positions (e.g., vice principals) on the basis of their performance. Private school principals’ terms can always be renewed, but they are not guaranteed a position after their employment term ends. Thus, private school principals and teachers generally face less job security than do those in public schools. Private school board members can be paid a salary as part of the operating expenditures. However, the scope for such salary payment is limited, and usually only one or two permanent board directors are paid a salary. Although good academic performance may not financially benefit the school owner and board members, delivering good student outcomes may help their reputation as community leaders and be in line with their educational philosophies. In addition, the Ministry of Education, Technology and Science (METS) monitors school operation and performance, and may intervene if private schools are poorly managed and the educational outcomes of their students suffer.

For example, Akiba and LeTendre (2009) claim that “Japan has a more stringent and standardized teacher rotation policy administered by the prefecture boards of education. Rotation policy – Tenkin – is a common practice among civil servants and industry, and is widely practiced throughout Japan.” OECD (2015, p.87) and Huang et al. (2016, p.49) report that in China, there have also been compulsory cross-school rotation and transferring of teachers and principals. In the Australian Capital Territory, public school teachers must also rotate every few years. Similarly, in British Columbia, Canada, principals must also rotate every few years (Coelli and Green 2012). 12 Public school principals can have two four-year terms. 13 Private School Law, Korea Ministry of Government Legislation (source: http://www.law.go.kr). 11

7

In sum, both public and private equalization policy high schools are regulated to operate in a similar manner, but they differ in terms of the autonomy to make personnel decisions and their labor markets. If private school principals can organize their schools more effectively—with a higher level of autonomy and stronger performance incentives—they may have greater success in delivering good student outcomes.

3.

Data We draw data from several sources. 14 First, we use publicly disclosed school-level

information pertaining to enrolment, dropouts, transfers, graduates’ destinations, number of teachers, incidents of disciplinary violations, expenditures, and other administrative records from the METS website. METS collects these data from schools each year. Second, we use individuallevel performance data from eleventh graders (i.e., second year of high school) on the National Assessment of Educational Achievement (NAEA), administered in 2010 by the Korea Institute for Curriculum and Evaluation (KICE).15 The NAEA is a relatively low-stakes test designed by the KICE to identify the factors that affect student achievement. The NAEA data also provide some useful student and school survey information. Third, we use twelfth graders’ individuallevel performance data on the College Scholastic Ability Test (CSAT) in 2009 and 2010 sourced from KICE as an additional measure of student achievement. 16 The CSAT is one of the prerequisites for college admissions. We focus on 198 general academic schools that have data available for all of the key outcome variables in 2008–2010. Finally, we surveyed high school principals in 2013 and obtained information from 111 school principals on their perceptions of differences between public and private schools.

[Table 1]

Table 1 provides the distribution of equalization policy high schools by school district and type. Nearly two-thirds of the high schools are privately owned. About 30 percent of the private schools are religiously affiliated, with the majority (90 percent) being Christian (Table A2, 14

Appendix II provides a detailed description of the data used and their sources. We use the NAEA 2010 test score data only because: (a) the NAEA 2009 measured test performance in grade rather than in score; and (b) the NAEA 2008 sampled 5% of 10th graders (1st year of high school) for the test only. 16 We do not use CSAT 2008 test score data because CSAT test performance is reported in grade rather than score in that year. 15

8

Appendix I). The schools are quite evenly distributed amongst coeducational, all-boys, and allgirls schools. Figure 1 shows a map of Seoul with the boundaries of school districts and the locations of private and public schools in our sample.

[Figure 1] Table 2 provides the summary statistics of students’ predetermined characteristics by private and public schools, as well as verifies random assignment. Panel A reports school-level student characteristics by school type and Panel B reports individual-level student characteristics by school type. The last two columns in Table 2 report statistics regarding mean differences between private and public schools after controlling for a set of school districts by year fixed effects. If randomization of students within districts is strictly enforced in Seoul, each fixedeffect mean difference will not be statistically different from zero. Our predetermined student characteristics—all of which cover a range of socioeconomic status and are predictive of student outcomes—include whether a student lives in a single-mother or dual-parent household, the share of students on public welfare support (a proxy for poverty), and the share of students receiving school lunch support (a proxy for low income).17 Although transfers are not common and students who transfer to another school district are subject again to random assignment, we also examine whether the net transfer rate is different between private and public schools.

[Table 2]

Panel A in Table 2 shows that the likelihood of a student being in a family on welfare assistance, or on lunch support is similar between private and public schools, and that net transfer rates do not differ across school types.18 Panel B shows that the likelihood of a student coming from a single-mother or dual-parent household is similar between private and public

17

We show in panel A of Table A3 (Appendix I) that these predetermined characteristics are correlated with student test scores and thus passing our tests of randomization implies that a student’s private school attendance status is orthogonal to the predetermined influences of outcomes. 18 This is not overly surprising, as individuals with strong preferences for school quality would have opted for selective high schools before the randomization. Moreover, as the differences between private and public high schools are subtle to students, there is no obvious reason for non-compliance. Indeed, some authors of this paper did not know their own school types until they came to work on this study.

9

school students. The coefficient estimates of the private school effect are, in all cases, close to zero; this is consistent with randomization.19

4.

Evidence of Greater Autonomy and Incentives to Perform in Private Schools Table 3 reports average school characteristics by school type. It shows that private and

public high schools differ significantly in their distribution of teacher types and resource allocation. Private and public schools spend similar amounts on teachers and staff on a perstudent basis, but private schools hire fewer teachers per student, keep larger fractions of (junior) regular teachers without an advanced certificate (a proxy for teaching experience) and teachers on short-term contracts, employ a lower fraction of teachers with an advanced certificate, and have relatively fewer teachers with a graduate degree. The differences are mainly in staffing decisions, rather than in the quality of infrastructure and the use of ability tracking. 20 More importantly, larger variations in the characteristics of and spending on teachers in private schools versus those in public schools are consistent with the extent of autonomy that private schools enjoy.

[Table 3]

Since private school principals are directly responsible for the recruitment and selection of teachers, they can more flexibly recruit, retain, and promote teachers who are most suitable to deliver the outcomes they desire. Table 3 shows that private school principals hire a larger fraction of teachers on short-term contracts and regular teachers without an advanced certificate, who generally face less job security and greater incentives to perform for career and promotion concerns. Less experienced teachers may have greater incentives to perform given that their base salary is relatively low. Similarly, short-term contract teachers may be more effective at delivering better student performance, as Duflo et al. (2015) and Muralidharan and Sundararaman (2013) have shown. In private schools, short-term teachers can be promoted to be regular teachers, depending on their performance; in public schools, however, short-term 19

We also show in Table A4 (Appendix I) the balanced predetermined characteristics of students across school types, using additional student-level data from the NAEA 2008 tenth-grade student survey (Panel A) and the Korean Education and Employment Panel’s (KEEP) middle school student sample (Panel B). 20 Some past studies indicate that student outcomes improve with better school infrastructure (Branham 2004; Glewwe et al. 2011) and the use of ability tracking (Duflo et al. 2011).

10

teachers cannot become regular teachers unless they pass the employment exam.21 Thus, shortterm teachers in private schools face stronger incentives to deliver better student outcomes, and regular teachers in private schools are more likely to be a selected group of teachers who have proven themselves through on-the-job performance. These regular and short-term teachers may have characteristics that are less readily observable to researchers, but which past research has shown to be more reflective of teacher quality. Compared to public schools, private schools pay their teachers and staff higher average total financial compensation. The difference is statistically significant and is about 3 percent of the average total financial compensation packages of public school teachers and staff (Table 3). As the average base salary of teachers and staff in private school is lower, this higher average total financial compensation is driven by higher variable/performance pay component. The higher variable/performance pay may work as an incentive that pushes them to deliver better student outcomes, as previous studies have shown (Duflo et al. 2015; Muralidharan and Sundararaman 2011). The dispersion in salary is also higher in private schools (Table 3). Although private school teachers are guaranteed the same pay-scale schedule as public school teachers who have the same credentials and years of teaching,22 their likelihood of within-school promotion to a senior, high-paying administrative position (e.g., vice principal) depends more on their performance, while within-school promotions (other than a stepwise increase that comes with teaching experience and credentials) are rare in public schools. Furthermore, private schools hire a higher proportion of short-term contract teachers and junior regular teachers (without an advanced certificate) who tend to be at the bottom end of the salary distribution and have strong incentives to perform for promotion and career concerns. The salary dispersion is also consistent with private schools having a larger within-school variance of teachers with an advanced certificate (proxy for experience). The heterogeneity in the productivity of team workers may increase average productivity through collaboration and mutual learning among different types of workers (Hamilton et al. 2003). Thus, private schools may have an environment more conducive to delivering good student outcomes.

21

Thus, short-term teachers in public schools focus on passing the examination. Enforcement Decree of the Private School Teachers and Staff Pension Act, Article 3. Source: http://www.law.go.kr/eng/engLsSc.do?menuId=1&query=private+school+act&x=0&y=0#liBgcolor0 22

11

[Table 4]

Table 4 reveals that private schools set quantifiable accountability measures for their principals and that principals’ perceptions regarding differences in autonomy, incentives, and accountability between private and public schools are consistent with the administrative data. 23 For example, private school principals place more emphasis on entrance into a prestigious university—something that is relatively easy to measure—while public school principals emphasize more the development of student creativity—something that is generally difficult to measure. They put roughly equal emphasis on good discipline and behavior. Private schools are more likely to be perceived as flexible with school policies. Private school principals’ and teachers’ jobs are also less likely to be perceived as secure, while their incentives to produce good academic performance are also perceived to be stronger. Private school teachers are also perceived to face stronger punishment for poor performance and also be given stronger encouragement to be innovative with their teaching. Thus, private school principals and teachers focus on delivering good academic outcomes that are quantifiable, and also face stronger incentives and pressures to do so. Although private schools have relatively fewer experienced and highly educated teachers and run larger classes—conditions that may have negative effects on student performance—their principals and teachers work in an environment with stronger incentives to deliver positive student outcomes. Thus, it is an empirical question as to whether private schooling leads to better student outcomes.

23

Table 4 reports results drawn from the anonymous principal survey we conducted in 2013, where we sent survey questionnaires to 162 schools rather than to all 198 school principals, because 34 schools changed their status from equalization policy high schools to autonomous schools after 2010 and we were informed by school district administrators to further exclude 2 private schools that were having principal or management changes at the time of survey. First three rows in Table 4 show that the survey and response rates do not differ by school type. Panel B in Table A1 (Appendix I) further demonstrates that our main findings hold for these 162 schools. Both private and public schools converted into autonomous schools are allowed to select new students (grade 10) and modify some of the curricula. Private schools that converted into autonomous schools can also charge higher fees, whereas public schools that converted into autonomous schools will get additional resources from the government and charge the same fees as all other equalization policy schools. Table A5 (Appendix I) shows the outcome differences between converted schools and non-converted schools by public (Panel A) and private (Panel B) status. Public schools that converted have much worse outcomes compared to the other public schools prior to conversion. Private schools that converted are not too different from the other private schools prior to conversion. It is important to note that despite the exclusion of low-performing public schools from the survey, Table 4 still indicates that principals and teachers in public schools generally face less incentives and pressures to deliver good student outcomes.

12

5.

Impacts of Private Schooling on Student Outcomes

5.1. Drop Out, Graduation, Disciplinary Problems, and College Attendance We estimate the causal effects of private schooling on student outcomes, using the following regression specification: 𝑦𝑗𝑘𝑡 = 𝛽 ∙ 𝑃𝑟𝑖𝑣𝑎𝑡𝑒𝑗𝑘 + 𝛿𝑘𝑡 + 𝜖𝑗𝑘𝑡 ,

(1)

where 𝑦𝑗𝑘𝑡 denotes an outcome of students in school j of school district k in year t. The schoollevel outcome variables include (1) the percentage of students dropping out of high school, (2) the percentage of high school seniors graduating, (3) the number of disciplinary problems reported per student, (4) the percentage of high school seniors attending any college, (5) the percentage of high school seniors attending two-year colleges, and (6) the percentage of high school seniors attending four-year colleges. 24 As four-year colleges are more academically oriented and generally more difficult for students to enter than two-year colleges, we examine separately the attendance rates of the two college types. 𝑃𝑟𝑖𝑣𝑎𝑡𝑒𝑗𝑘 is an indicator for whether or not school j is private. As students are assigned randomly into schools within a district, we include school district-year fixed effects 𝛿𝑘𝑡 to ensure that the selection into school districts is controlled for and that the coefficient of interest 𝛽 captures the causal effect of attending a private school on student outcomes. The term 𝜖𝑗𝑘𝑡 denotes all other unobserved influences of the outcomes. We weight all school-level regressions by the number of students in the denominators of dependent variables (the estimates are not sensitive to using weights). Figure 2 shows the distribution for each school-level outcome by school type. It appears that public schools have higher disciplinary problems per student and two-year college attendance rates, and lower four-year college attendance rates and overall college attendance rates than private schools. Private and public school students have similar dropout rates and graduation rates.

[Figure 2]

Because private school concentration rates differ across school districts (Table 1), we also assess if the private school effect estimated using specification (1) is similar to the average effect 24

Disciplinary problems include physical violence, bullying, harassment, verbal violence, threat, intimidation, harassment, and cyber-bullying, among other forms.

13

estimated using a specification without district-year fixed effects and the average treatment effect estimated by weighting a set of within-district specific private school effects by their respective student population size. First, if the estimates without district-year fixed effects are considerably different from the estimates with district-year fixed effects, then the sorting of households across districts on the basis of private school concentration rates is likely prevalent in Seoul. If motivated parents select into school districts with higher than average probability of getting into private schools, then district-year fixed effects estimates provide the weighted average of the effect of being in private schools in districts with motivated students and in districts with unmotivated ones. Although this form of sorting does not imply bias in our preferred districtyear fixed effects estimates, it may affect the interpretation of our results. In particular, the classroom composition and private school effects under this form of sorting and within-district randomization might be very different to the classroom composition and private school effects when students were randomized across districts. Second, because the share of private schools differs across school districts, school districts with close to 50 percent private schools get more weight than school districts with nearly all private or nearly all public schools under the districtyear fixed effects specification (1). If the private school effect is not homogenous across districts, then the district-year fixed effects specification will produce misleading estimates of the overall quality of private schools. To assess whether the district-year fixed effects specification is appropriate, we estimate private school effects within each district, and then produce an estimate of the average treatment effect by taking an average of these estimates weighted by student population size for comparison.

[Table 5]

Table 5 reports the results. Panel A reports the estimates without district-year fixed effects, panel B reports the preferred district-year fixed effects estimates, and panel C reports the estimates for the weighted average treatment effects. In general, the estimates in panel A are fairly similar to the district-year fixed effects estimates in panel B, suggesting negligible effects of sorting across districts.25 The district-year fixed effects estimates are also similar to estimates 25

Table A6 (Appendix I) also shows that there is no systematic relationship between the share of private schools and the mean social and economic characteristics of households in the district. Thus, there is no evidence that households sort into districts by social economic characteristics.

14

for the weighted average treatment effect in panel C. Thus, it is appropriate to focus our analysis using the district-year fixed effects specification. Columns 1 and 2 in panel B of Table 5 indicate that private school students and public school students are equally likely to drop out of and graduate from high school. In Seoul, the high school dropout rate is less than 2 percent and the graduation rate is close to 98 percent, and so there is not much room for improvement in these outcomes. Column 3 shows private schooling reduces the average number of disciplinary problems per student by approximately 0.001 (one per 1,000 students). Compared to the average number of disciplinary problems per student in public schools—namely 0.0015—private schooling reduces the number by almost 66 percent. This estimate is comparable to that derived by Cullen et al. (2006), who found that selfreported arrest rates are reduced by nearly 60 percent among the students who win lotteries to attend high-achieving Chicago high schools, compared to those who do not. Although due to data limitation we can only examine severe forms of disciplinary problems, such as physical violence, bullying, harassment, and intimidation, which are rare in Seoul’s high schools, the presence of these incidents may indicate that other less-extreme forms of behavioral problems are also pervasive. Having fewer disciplinary problems per student suggests that private school students tend to have lower levels of other behavioral issues and enjoy safer school environments. Table 5 also shows that private schooling increases the college attendance rates of high school seniors by moving them into four-year universities and away from two-year junior colleges and other options.26 Private schooling increases high school seniors’ college attendance rates by 4.4 percentage points, or 8 percent higher than public schools, where roughly 56 percent of public high school seniors enter colleges. The results indicate that private schooling significantly increases the likelihood of four-year college attendance and reduces the likelihood of two-year college attendance.27 Our estimated effect size is about one-third that of Evans and Schwab (1995) of the effect of Catholic private schooling: they found that Catholic school attendance in the United States increases the probability of entering a four-year college by 13 percentage points, or 46 percent of the college attendance rate of public school students. However, Neal (1997) shows that the estimated Catholic school effects become smaller when the

26

Table A7 (Appendix I) also reports the results by gender. College attendance and graduation rates are calculated by using the number of high school seniors in the denominators. If we use in the denominators the initial number of tenth-grade students, the estimated effects remain similar. 27

15

alternative public schools are more similar, while Altonji et al. (2005a) argue that past estimates are based on potentially problematic instrumental variables. Thus, our estimate is fairly sizable, considering that private and public schools in Seoul differ only in terms of school autonomy and that students are assigned randomly into schools. There may be concerns that schools might misreport some of the outcome measures presented above. Misreporting is problematic for our results if private and public schools do so differentially and that private schools inflate their performance. Differential misreporting is unlikely for several reasons. First, it is not possible for private schools to gain financially from better student outcomes as the government does not allow them to charge higher fees, admit more students, nor receive extra funding for better student outcomes. Second, the Korean Education and Employment Panel (KEEP) surveys, collected by independent researchers, also reveal that private school students report higher likelihood of college attendance than public school students do.28 As students face a different reporting environment in KEEP surveys, the likelihood for them to lie differentially by school type in the survey is expected to be minimal. Third, schools are legally obliged to truthfully report to METS and METS allows third parties an online platform to report any erroneous information reported by a school.

5.2. Standardized Test Scores We use eleventh-grade individuals’ test scores in the subjects of Korean, mathematics, and English in the NAEA 2010 and also twelfth-grade individuals’ test scores in the subjects of Korean and English in the 2009 and 2010 CSAT to assess the effects of private schooling on student achievement. It is important to note that NAEA is a relatively low-stakes test and students cannot strategically select subjects the way they would for the high-stakes College Scholastic Aptitude Test (CSAT). 29 Although all students take the same CSAT Korean and English subjects, other test components depend on their tracks and their choice of mathematics and electives. To enter into their preferred universities and college majors, students may strategically choose the seven CSAT component subjects through their track selection and KEEP (high school sample) data show that the college attendance rates in 2005 (based on students’ responses) are 0.52 and 0.44 for private schools and public schools, respectively. Similarly, private school students are also less likely (1.3 percentage points less) than public school students to be reported by their homeroom teachers for subjecting to disciplinary action. The mean of private school students is roughly 67 percent lower than that of public school students. 29 High stakes tests raise concerns about gaming, low stakes tests raise concerns that some students do not care enough to put forth effort. 28

16

selective preparation for certain subjects. Student’s in-school performance and choice of college major are also taken into consideration in the admission process (though CSAT scores are heavily weighted). Ultimately, it is college admissions that matter the most to students, rather than the mean CSAT scores. We also check whether private school students and public school students are equally likely to take each test, to ensure that the estimates do not suffer from any selection bias. Columns 1–3 of Table 6 show that private school students are 2 percentage points less likely to miss the NAEA tests, indicating higher test-day absenteeism among public school students. Similarly, columns 4 and 5 of Table 6 show that private school students are 3 percentage points less likely not to take the CSAT. The greater likelihood of private school students to take the CSAT is also consistent with their higher college attendance rates given that taking the CSAT is one of the pre-requisites for college admissions. If non-random selection into test-taking exists by school type, the estimated effects of private schooling on achievement based on the sample of test takers will suffer from non-random selection bias. For example, the estimated effects of private schools on test scores will be downwardly biased if public schools tend to make less academically inclined students miss the test.

[Table 6]

We use two methods to address this concern. First, we replace each missing value with the average within-district test score of the student’s school type. The assumption is that students who missed the tests are similar to the average students of a particular school type in the district. As weaker students are more likely to have higher absenteeism and miss the tests, this approach provides conservative estimates of private schooling effects. Second, we use Lee’s (2009) sharpbound estimators to bound the effects of private schooling on test scores. The sharp-bound estimators trim the private school sample on the basis of the selection rate (i.e., the probability of missing the test) of the public school sample relative to that of the private school sample, so that the two samples are comparable. Assuming that high (low) performers in public schools miss the test, the upper (lower) tail of the private school test score distribution will be trimmed to give the lower-bound (upper-bound) estimate of the private school effect.

17

[Figure 3]

Figure 3 shows the test score distribution for each test, by school type. Overall, the distributions of private school students’ test scores are to the right of those of public school students.

[Table 7]

Panel A of Table 7 reports the estimates without district-year fixed effects, panel B reports the district-year fixed effects estimates, and panel C reports the estimates for the weighted average treatment specification, all based on the ‘replacement of missing students’ method. Panel D of Table 7 report the lower-bound and upper-bound estimates of the effects of private schooling on Korean, mathematics, and English test scores. The estimates without district-year fixed effects, weighted average treatment effect estimates, and district-year fixed effects estimates are all similar. The preferred district-fixed effects estimates in panel B show that private school students outperform public school students in NAEA by 0.13 standard deviations for Korean, by 0.12 standard deviations for math, and by 0.12 standard deviations for English. The estimated private school effects on CSAT Korean and English scores are smaller, at 0.07 and 0.08 standard deviations respectively. The point estimates for the lower sharp-bounds are 0.07, 0.07, and 0.08 for NAEA Korean, mathematics, and English scores, respectively. These estimates are statistically greater than zero, indicating that even in the worst-case scenario— where the brightest public school students are selected out of test-taking—the estimated effect of private schooling on NAEA test performance is positive. However, the lower sharp-bound estimates for CSAT scores are not statistically different from zero. This means that if the bestperforming public school students are selected out of CSAT, we do not have evidence that private students outperform public school students in CSAT on average. We argue that this is unlikely the case and the lower-bound estimates are less relevant, especially given that KEEP data do not show that high-ability public school students have a greater likelihood of missing CSAT.30 For example, the relationship between missing CSAT Korean scores and the five-point self-rating of a student’s own Korean ability is significantly negative at the 5% level. The relationship between missing CSAT English scores and the five-point self-rating of English ability is not statistically different from zero at the 5% level. We also 30

18

The effect size based on the district fixed effects estimates is roughly one-half that found by Angrist et al. (2002) for the private school effect in Colombia. Unlike the situation in Colombia, however, private and public high schools in Seoul must admit similar students, use similar curricula, and charge the same fees, and so there are fewer factors that can influence outcomes that differ by school type. Nevertheless, the effect is substantial, considering that a one-standard-deviation change in classroom quality is associated with a 0.1 SD change in test scores in Araujo et al.’s (2016) study. 31 Private schools in Seoul achieve this effect without increasing average expenditures per student. As NAEA data also contain information about the household structure of test takers, we can investigate whether private schooling differentially benefits students from different types of households. Table 8 reports the heterogeneous effects of private schooling by student household type (i.e., dual-parent or not). The effects of private schools are larger for students in non-dualparent households, but the differences are not statistically different. Thus, there is weak evidence that private schooling benefits students from disadvantaged or low social economic backgrounds more.

[Table 8]

In summary, private high schools have fewer student disciplinary problems and are more likely to place their students in higher-education institutions, especially in four-year universities. Private high schools also improve students’ standardized test performance.

6.

The Roles of Autonomy in Personnel Decisions and Sorting of Teachers Across Sectors In this section, we examine the roles of autonomy in personnel decisions and sorting of

teachers across sectors in explaining the private school effects.

6.1. The Role of Autonomy in Personnel Decisions replace the missing values for public school students’ test scores by progressively assigning values that are higher than the average to examine when the private school effect will become statistically not different from zero at the 10% level. It turns out that public school students who missed NAEA would have to be roughly 1.4 SD in English, 1.6 SD in Korean, and 1.7 SD in mathematics above the mean for this to happen. For public school students who missed CSAT, they would have to be 0.17 SD above the mean in both Korean and English for this to happen. 31 Rockoff (2004) and Hanushek and Rivkin (2010) report the effects of changes in teacher quality on test scores that are also in a similar range.

19

To examine whether the private school effects are consistent with private schools’ greater autonomy in personnel decisions, we adopt an approach that closely resembles one that Angrist, Pathak, and Walters (2013) utilize to examine factors explaining the superior performance of urban charter schools in Massachusetts. We first estimate each private school’s effect on an outcome by regressing the outcome measure against a private school indicator and a set of control variables using a sample that includes the private school and all public schools in Seoul in the following specification: (2)

𝑦𝑗𝑡 = 𝜏 ∙ 𝑝𝑟𝑖𝑣𝑎𝑡𝑒𝑗 + 𝛿 ′ 𝑋𝑗𝑡 + 𝜖𝑗𝑡 .

The coefficient estimate for τ gives us a specific private school’s effect on an outcome y. We include in X a set of school district-year fixed effects and the percentage of students on lunch support to ensure that the public schools serving as the counterfactual have similar student characteristics. For a school level outcome, the regression is weighted using the denominator of the outcome measure. We then estimate a set of private school specific teacher characteristics in a similar fashion, replacing the dependent variable in equation (2) by teacher characteristics. This gives each private school’s teacher characteristics relative to the counterfactual. The teacher characteristics that we focus on include the share of short-term contract teachers, share of regular teachers without an advanced certificate (junior regular teachers), and student-teacher ratio. We focus on these teacher characteristics as we have shown earlier that while having fewer teachers per student, private schools tend to hire more short-term contract and less experienced teachers who face greater incentives to perform for career and promotion concerns.32 Once we obtain these measures about each private school’s characteristics relative to the counterfactual public schools, we regress the estimated private school effect on these teacher characteristics (and weight each observation by the variance of the private school effect) to evaluate whether they explain the private school effects. If the private school effects are driven by their personnel decisions, then we expect private school effects to be stronger when the share of teaching staff with greater career and promotion concerns is larger, while holding student-teacher ratio constant.

[Table 9] 32

We do not consider variables related to remuneration because our data are limited. For example, we only have total expenditures on teachers for 2009 and our data on average base pay are measured at the administrative district and school type level. Moreover, the teacher composition and student-teacher ratio variables that we consider determine the total expenditures on teacher compensation.

20

While the results reported in Table 9 should not be interpreted as causal, they are consistent with our hypothesis. For example, columns 1 and 2 show that as the share of junior regular teachers increases, the effect of private schools increases for college attendance rates and decreases for disciplinary problems per student. Similarly, columns 3 to 7 show that as the share of short-term contract teachers increases, the effect of private schools on students’ test scores increases. If we multiply the mean of these estimated teacher characteristics (0.061 for the share of junior regular teachers and 0.058 for the share of short-term contract teachers) by the coefficient estimates reported in Table 9, we can obtain private school effects that are not too far from those reported in Tables 5 and 7. For example, we obtain a private school effect of 5.7 percentage points for college attendance rates and 0.116 standard deviations for NAEA English. We further perform some simple back-of-the-envelope calculations to evaluate whether the estimated effects of private schooling are also consistent with findings in the literature regarding the effect of performance pay on test scores where comparable experimental estimates are available. For example, Muralidharan and Sundararaman (2011) find that after two years of a teacher incentive treatment in Indian primary schools, which rewarded teachers depending on their students’ performance gains, students in the treatment schools experienced test score improvement by 0.17 to 0.27 standard deviations. The average difference in bonus pay between the treatment group and control group is roughly three percentage points. In our setting, the average difference in variable or performance pay component between private and public schools is roughly 5 percentage points and the average difference in test score is between 0.07 to 0.13 standard deviations. Since high school students in Korea are near the top of the test-score distributions as evident in many international assessments, such as the Programme for International Student Assessment and the Trends in International Mathematics and Science Study, the room for test-score improvement in Seoul is likely much lower. Thus, our estimated effects on test scores appear reasonable.

6.2. Sorting of Teachers Across Sectors The greater autonomy in personnel decisions that private schools enjoy may lead to the sorting of teachers and principals across the private and public sectors. If better-quality teachers and principals sort into the private sector, then the positive effects of private schools on student 21

outcomes may reflect the effects of sorting rather than the causal effects of private schools and the value added from what private schools do with their autonomy. For sorting to drive the private school effects, there must be little overlap between sectors in the characteristics of teachers that predict student outcomes and also an over representation of teachers with better characteristics in the private sector. We examine whether these conditions hold in the following ways. First, we correlate teacher characteristics with student outcomes in public schools (the control group) to demonstrate whether teacher characteristics that we have data available are reasonable proxies of teacher quality. Second, we plot the distributions of these teacher characteristics by school type to assess the extent of overlap.

[Table 10]

Table 10 shows the relationships between various teacher characteristics and student outcomes in public schools. In general, teacher characteristics that are expected to predict better student outcomes display correlations in the expected directions. For example, public schools with a higher share of senior regular teachers and lower share of junior regular teachers tend to deliver better student test performance and fewer incidents of disciplinary problems. Figure 4 shows the distributions of these teacher characteristics by school type. In all cases, there is a large extent of overlap in distributions for each type of teacher characteristics. Moreover, it appears that public schools tend to have more experienced teachers and fewer short-term contract and junior regular teachers, characteristics that are associated with better student outcomes. Thus, it is unlikely that our findings are driven by sorting of teachers across school types.

[Figure 4]

7.

Alternative Channels of Private School Effects We have shown evidence that private schools have greater autonomy in their personnel

decisions and these differences are plausible explanations for the outcome differences. In this section, we examine several alternative (non-autonomy) explanations for why private schools may lead to the positive outcomes documented above.

22

7.1 History of Private Schools In Seoul, private schools typically have a longer history than public schools (Table 3). It is possible that private schools have better student outcomes on account of their greater expertise in operating schools, which has developed over time.33 Figure 5 shows the predictions from a local polynomial regression (Epanechnikov kernel) of each outcome variable against the years since establishment by private and public school. The predictions show that private schools outperform public schools for a wide range of school age. A few private schools that were established more than half a century ago actually have much worse average students’ test scores. Thus, the longer history of private schools does not explain their superior performance.

[Figure 5]

It is also plausible that private schools developed the culture of sending students to prestigious universities when they had to compete for students during the pre-equalization policy period (pre-1974). To address this history-cultural issue, we restrict the sample to 147 schools established after 1974 and estimate the main regression equation. The results (Panel A, Table 11) show that the private school effects remain strong.

[Table 11]

7.2 Religious Private Schooling About 30 percent of private schools in Seoul are religious schools (mostly of a Christian denomination). Past studies in other countries—such as the United States and Australia—have shown some benefits associated with Catholic private schooling for individuals (Evans and Schwab 1995; Neal 1997; Vella 1999; Altonji et al. 2005a, 2005b). The religious affiliation of private schools, rather than the ownership type per se, might explain outcome differences between private and public schools in Seoul. When we run regressions for each outcome variable against a religious affiliation dummy for the sample of private schools, the estimates show no differences between religiously affiliated private schools and secular private schools (Panel B, 33

For instance, private schools might develop practices that benefitted student outcomes in the more competitive pre-equalization policy environment, and these practices might persist even after the schools were taken into the state sector.

23

Table 11).34 Thus, the religious affiliation of schools plays no role in explaining private school effects in Seoul.

7.3 Single-Sex Schooling As private schools in Seoul are more likely to be single-sex schools as shown in Table 1, the private school effects may also capture single-sex school effects. We control for the gender type of a school and estimate the effects of private schooling. If the private school effects are present after controlling for the gender type of a school, then the effects are less likely to be driven by single-sex schooling. The analysis is pertinent, as Park et al. (2013) show that students in single-sex high schools outperform those in coeducational high schools in some academic outcomes. Panel C of Table 11 shows that after controlling for the gender type of a school, private schooling still reduces the likelihood of violent incidents and increases college attendance rates. The estimated effects of private schooling on test scores become smaller and somewhat noisier, especially for CSAT. 35 Thus, the effects of private schooling on college outcomes and disciplinary problems are unlikely to be driven by single-sex schooling, but the tendency for private schools to be single-sex may partially contribute to their superior test performance.

7.4 Substitution between Academic Performance and Creativity Development Our principal survey shows that public school principals place a greater emphasis on the importance of creativity development than private school principals do. If public school principals substitute between academic performance and creativity development, then our measures of student outcomes may not be fair to public schools. Although we do not have measures of students’ creativity, we can assess if our main findings are robust to controlling for differences in the emphasis school principal placed on creativity development between private

34

The main results are also robust to dropping religious schools or adding a religious-school indicator. Table A8 (Appendix I) reports a more stringent test for the effects of single-sex schooling, where we split the sample by the school’s gender type. The results also show that for both genders, private school students are more likely to enter colleges and less likely to be involved in violent incidents. The average test-score effects of private schools are noisier in coed schools, due to the smaller sample and the heterogeneous effects of private schools by household type. Private schools, whether coed or single-sex, are especially beneficial to students in non-dual-parent households (Table A9). 35

24

and public schools. 36 Panel D of Table 11 shows that our findings are fairly robust to the inclusion of this control variable, especially for the effects of private schooling on disciplinary problems per student and college attendance rates. There is some evidence of substitution between emphasis on creativity development and test performance, as the private school effects on test performance decrease after we control for the emphasis principal placed on creativity development. This perhaps reflects the greater autonomy that principals enjoy, and their consequent desire to satisfy parents’ preferences for academic achievement.

7.5 Private Tutoring South Korea has one of the most active private tutoring markets in the world (Bray 2009). Students may vary their use of private tutoring, depending on actual or perceived school quality (Hahn and Wang 2015; Wang 2015; Carr and Wang 2017). It is possible that private schooling may increase the likelihood of students taking private after-school tutoring. In this case, the effects of private schooling may be driven by the effects of private tutoring. We use NAEA data to estimate whether intensity of private tutoring is greater in private schools than public schools. Private tutoring reported in NAEA 2010 is a categorical variable that takes a value from 1 to 5 in increasing order, where 1 means no private tutoring, 2 means less than one hour, 3 means between one and two hours, 4 means between two and three hours, and 5 means more than three hours per week. The data show that the use of private tutoring is similar between private school and public school students. The mean for private tutoring is 2.93 for those in private schools and 2.96 for those in public schools (the p-value for the difference is 0.61).37 We also estimate the main model including private tutoring and its interaction term with private schooling as explanatory variables, as private tutoring might be more effective for students in private schools. The effects of private tutoring are smaller for private schools (Table A10, Appendix I). Thus, there is no evidence that private tutoring plays any role in explaining the private school effects.

8.

Conclusions

36

Note that because our principal survey is anonymous, we could only use the average value of emphasis placed on creativity development measured at the district-type level. 37 The results are not sensitive if we use dummy variables transformed from the original categorical variable.

25

We exploit the equalization policy and the random assignment of students into private and public high schools within Seoul’s school districts, to show that private schools—which have greater autonomy than public schools—deliver better student outcomes. Because of the equalization policy, we can rule out many factors that are commonly attributed to the effects of private schooling, such as peer quality, resources, curricula differences, and incentives for schools to compete for students and revenues. We show that private schooling leads to higher four-year college attendance rates, lower two-year college attendance rates, and fewer disciplinary problems per student. Private schooling has, however, no significant effect on dropout rates and high school seniors’ graduation rates. Furthermore, private school students outperform public school students in Korean, English, and mathematics standardized tests. The magnitudes of the estimated private school effects are consistent with the differences in teacher compositions between private and public schools. Additionally, we rule out the sorting of teachers across sectors, the longer history of private schools, single-sex schooling, religious affiliation of private schools, and private tutoring as the main channels of positive private school effects. Our results for college attendance and disciplinary problems are also robust to controlling for the greater emphasis that public school principals placed on the creativity development of students. There is some evidence that private school principals emphasize academic achievement at the expense of creativity development, perhaps reflecting their desire to satisfy parents’ preferences for academic achievement given their autonomy. Our findings suggest that autonomy in personnel decisions, together with strong incentives for principals and teachers to perform, may be effective in creating positive student outcomes. The results imply that it is possible to improve outcomes of students through greater school autonomy, even when the schools these students attend are in a school catchment area, guaranteed funding and enrolment, and heavily regulated by the government. Several caveats apply when drawing policy implications from this study’s findings. A key argument as to why private and other forms of independent schooling may improve outcomes stems from their potential to increase competition, and so our findings cannot imply that policymakers should randomize students across schools and eliminate the incentives for schools to compete for students. Moreover, our estimates are obtained in a partial equilibrium setting. The general equilibrium effects of giving all schools autonomy in their personnel and remuneration decisions can be quite different, as it will likely change the pool of available 26

teachers and principals. The findings in South Korea may have implications for countries with similar institutions and level of economic development, but may not be extended to other economies: Hanushek et al. (2013), for example, argue that the effect of school autonomy depends on the context under which the school operates. More evidence is needed to improve our understanding of how giving schools more autonomy may benefit students.

Acknowledgements We thank anonymous referees, the coeditor Jonah E. Rockoff, Tiffany Chou, Julie Cullen, Michael Ewens, Eric A. Hanushek, Hisam Kim, Jungmin Lee, John A. List, Pushkar Maitra, Karthik Muralidharan, Birendra Rai, and seminar and conference participants at Curtin University, Deakin University, Monash University, University of Adelaide, University of Melbourne, University of Western Australia, the Econometrics Society Australasian Meeting 2012, the Labour Econometrics Workshop 2012, Monash-Warwick Workshop 2013, the Asian Meeting of the Econometric Society 2013, and the annual meeting of the Society for Labor Economists 2014 for helpful comments. Byung Uk An, Lucy Eunju Kim, Hyuk Son, and Joonho Yeo provided excellent research assistance. This project would not have been possible without the funding from Monash University, survey assistance given by the Seoul Metropolitan Office of Education and Yeo Bok Yun, and data provided by the Korea Institute for Curriculum and Evaluation (KICE), the Korea Education and Research Information Service (KERIS), and the Ministry of Education, Science and Technology of the Republic of Korea. Institutional Review Board approval was obtained from the Monash University Human Research Ethics Committee (MUHREC).

27

References Abdulkadiroglu, Atila, Joshua D. Angrist, Susan M. Dynarski, Thomas J. Kane, and Parag A. Pathak. 2011. “Accountability and Flexibility in Public Schools: Evidence from Boston’s Charters and Pilots.” Quarterly Journal of Economics 126: 699-748. Akiba, Motoko and Gerald LeTendre. 2009. Improving Teacher Quality: The U.S. Teaching Force in Global Context. Teachers College Press. Alexander, Karl L., and Aaron M. Pallas. 1983. “Private Schools and Public Policy: New Evidence on Cognitive Achievement in Public and Private Schools.” Sociology of Education 56: 170-81. Altonji, Joseph G., Todd E. Elder, and Christopher R. Taber. 2005a. “An Evaluation of Instrumental Variable Strategies for Estimating the Effects of Catholic Schools.” Journal of Human Resources 40: 791-821. Altonji, Joseph G., Todd E. Elder, and Christopher R. Taber. 2005b. “Selection on Observed and Unobserved Variables: Assessing the Effectiveness of Catholic Schools.” Journal of Political Economy 113: 151-184. 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 92(5): 1535-1558. 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 96(3): 847-862. Angrist, Joshua D., Parag A. Pathak, and Christopher R. Walters. 2013. “Explaining charter school effectiveness.” American Economic Journal: Applied Economics 5(4): 1-27. Araujo, M. Caridad, Pedro Carneiro, Yyannú Cruz-Aguayo, and Norbert Schady. 2016. “Teacher quality and learning outcomes in kindergarten.” Quarterly Journal of Economics 131(3), 1415-1453. Branham, David. 2004. “The Wise Man Builds His House Upon the Rock: The Effects of Inadequate School Building Infrastructure on Student Attendance.” Social Science Quarterly 85(5): 1112-1128. Bray, Mark. 2009. Confronting the Shadow Education System: What Government Policies for What Private Tutoring? International Institute for Educational Planning, Paris. Carr, Daniel, and Liang Choon Wang. 2017. “The Effect of After-School Classes on Private Tuition, Mental Health and Academic Outcomes: Evidence from Korea.” Sociology, forthcoming. Clark, Damon. 2009. “The Performance and Competitive Effects of School Autonomy.” Journal of Political Economy 117(4): 745-783. Coelli, Michael, and David A. Green. 2012. “Leadership Effects: School Principals and Student Outcomes.” Economics of Education Review 31(1): 92-109. Coleman, James, Thomas Hoffer, and Sally Kilgore. 1982. High School Achievement: Public, Catholic and Private Schools Compared. Basic Books, NY. Cullen, Julie Berry, Brian A. Jacob and Steven Levitt. 2006. “The Effect of School Choice on Participants: Evidence from Randomized Lotteries.” Econometrica 74(5): 1191-1230. Duflo, Esther, Pascaline Dupas, and Michael Kremer. 2011. “Peer Effects and the Impacts of Tracking: Evidence from a Randomized Evaluation in Kenya.” American Economic Review 101(5):1739-1774. 28

Duflo, Esther, Pascaline Dupas, and Michael Kremer. 2015. “School Governance, Teacher Incentives, and Pupil–Teacher Ratios: Experimental Evidence from Kenyan Primary Schools.” Journal of Public Economics 123: 92-110. Evans, William, and Robert Schwab. 1995. “Finishing High School and Starting College: Do Catholic Schools Make a Difference?” Quarterly Journal of Economics 110: 947-974. Figlio, David N., and Joe A. Stone. 1999. “School Choice and Student Performance: Are Private Schools Really Better?” Research in Labor Economics 18: 115-140. Glewwe, Paul W., Eric A. Hanushek, Sarah D. Humpage and Renato Ravina. 2011. “School Resources and Educational Outcomes in Developing Countries: A Review of the Literature from 1990 to 2010,” NBER Working Paper 17554, October. Hahn, Youjin, and Liang Choon Wang. 2015. “The Effectiveness of Single-Sex Schools through Out-of-School Activities: Evidence from South Korea.” http://dx.doi.org/10.2139/ssrn.2607313. Hamilton, Barton H., Jack A. Nickerson, and Hideo Owan. 2003. “Team Incentives and Worker Heterogeneity: An Empirical Analysis of the Impacts of Teams on Productivity and Participation.” Journal of Political Economy 111(3): 465-497. Hanushek, Eric A., Susanne Link, and Ludger Woessmann. 2013. “Does School Autonomy Make Sense Everywhere? Panel Estimates from PISA.” Journal of Development Economics 104: 212-232. Hanushek, Eric A., and Steven G. Rivkin, 2010. “Generalizations about using value-added measures of teacher quality.” American Economic Review 100(2): 267-271. Huang, Eva, John Benson, and Ying Zhu. 2016. Teacher Management in China: The transformation of educational systems. Routledge Hoxby, Caroline M., and Sonali Murarka. 2009. “Charter Schools in New York City: Who Enrolls and How They Affect Their Students’ Achievement.” NBER Working Paper 14852. Jimenez, Emmanuel, and Yasuyuki Sawada. 1999. “Do community-managed schools work? An evaluation of El Salvador's EDUCO program.” World Bank Economic Review 13(3): 415441. Kang, Changhui. 2007. “Classroom Peer Effects and Academic Achievement: Quasirandomization Evidence from South Korea.” Journal of Urban Economics 61: 458–495. Kim, Ee-Gyeong, and You-Kyung Han. 2002. Attracting, Developing, and Retaining Effective Teachers: Background Report for Korea. Korean Educational Development Institute, Seoul. Kim, E. G., Kim, K. S., Kim, D. K., and Kim, E. Y. 2007. Improving School Leadership: Country Background Report for Korea. Korean Educational Development Institute, Seoul. King, Elizabeth M., and Berk Ozler. 2005. “What’s Decentralization Got to Do With Learning? School Autonomy and Student Performance.” Kyoto University: Interfaces for Advanced Economic Analysis. DP 54 (2005): 51-60. Krueger, Alan B., and Pei Zhu, 2004. “Another look at the New York City school voucher experiment.” American Behavioral Scientist 47(5): 658-698. Lee, Chull. 2011. “Critical Study on the Opposition to Religious Education of Mission School in Consideration of the Decision of the Supreme Court on the Case of Eui-seok Kang.” Hyunsang-gwa-Insik 114(9): 59-74. Lee, David S. 2009. “Training, Wages, and Sample Selection: Estimating Sharp Bounds on Treatment Effects.” Review of Economic Studies 6(3): 1071-1102.

29

Moon, Hyungpyo. 2002. The Korean Pension System: Current State and Tasks Ahead. Korea Development Institute, Korea. Muralidharan, Karthik and Venkatesh Sundararaman. 2011. “Teacher Performance Pay: Experimental Evidence from India.” Journal of Political Economy 119(1): 39-77. Muralidharan, Karthik and Venkatesh Sundararaman. 2013. “Contract Teachers: Experimental Evidence from India.” NBER Working Paper No. 19440. Neal, Derek. 1997. “The Effect of Catholic Secondary Schooling on Educational Attainment.” Journal of Labor Economics 15: 98-123. OECD. 2015. OECD Economic Surveys: China. OECD Publishing, Paris. Paik, Sung Joon. 2013. 2012 Modularization of Korea’s Development Experience: Role of Private Schools in Korea’s Educational Development. KDI School of Public Policy and Management, Korea. Park, Hyunjoon, Jere R. Behrman, and Jaesung Choi. 2013. “Causal Effects of Single-Sex Schools on College Entrance Exams and College Attendance: Random Assignment in Seoul High Schools.” Demography 50(2): 447-469. Rockoff, Jonah E. 2004. “The Impact of Individual Teachers on Student Achievement: Evidence from Panel Data.” American Economic Review, Papers and Proceedings 94: 247-252. Vella, Francis. 1999. “Do Catholic Schools Make a Difference? Evidence from Australia.” Journal of Human Resources 34: 208-224. Wang, Liang Choon. 2015. “All Work and No Play? The Effects of Ability Sorting on Students’ Non-school Inputs, Time Use, and Grade Anxiety.” Economics of Education Review 44: 29-41.

30

Table 1 Distribution of School Types by School District in 2008 School district

Number of Private schools Public schools schools Coed. All-boys All-girls Coed. All-boys All-girls 1 25 3 6 6 7 2 1 2 25 4 9 6 5 0 1 3 19 4 3 4 5 1 2 4 11 1 4 3 1 1 1 5 17 3 8 5 0 0 1 6 18 0 3 2 10 1 2 7 16 2 4 3 5 1 1 8 9 1 1 2 4 0 1 9 13 0 4 4 4 1 0 10 24 3 5 6 9 1 0 11 21 0 8 9 1 3 0 Total 198 21 55 50 51 11 10 Notes: Distribution of religious school types by school district is reported in Table A2 (Appendix I).

Table 2 Student Characteristics by School Type and Test of Randomization Private Mean S.D.

Obs.

Public Mean S.D.

F.E. Difference Mean (p-value)

A. School-level student characteristics Share of students on welfare assistance 198 0.041 0.044 0.040 0.026 -0.002 (0.620) Share of students on lunch support 594 0.107 0.135 0.098 0.112 0.014 (0.215) Net transfer rate 594 0.001 0.012 0.002 0.011 -0.0002 (0.843) B. Individual-level student characteristics Single-mother household – Male 48566 0.075 0.264 0.074 0.262 0.001 (0.743) Dual-parent household – Male 48566 0.877 0.328 0.872 0.334 0.005 (0.339) Single-mother household – Female 41259 0.082 0.274 0.082 0.274 -0.0001 (0.974) Dual-parent household – Female 41259 0.876 0.330 0.874 0.332 0.002 (0.786) Notes: In panel A, school-level share of students on welfare assistance came from NAEA 2010 principal survey, whereas school-level share of students on lunch support and net transfer rate for 2008-2010 came from the Ministry of Education, Technology and Science’s (METS) website. Mean difference is tested by regressing each variable on a dummy of private school and a set of school district or school district-year fixed effects and the regression is weighted by the number of students enrolled in the school in each year; standard errors are clustered at the school level. In panel B, individual students’ household information came from the NAEA 2010 grade-11 student survey. Mean difference is tested by regressing each variable on a dummy of private school and a set of school district fixed effects; standard errors are clustered at the school level. Table A3 (Appendix I) shows that these predetermined characteristics are correlated with outcomes and Table A4 (Appendix I) reports additional tests for randomization using other data sources. Appendix II provides detailed data sources and variable definitions.

31

Table 3 School Characteristics by School Type Private Public F.E. Difference Obs. Mean S.D. Mean S.D. Mean (p-value) Teacher and staff salary per stud. (‘000KRW)# 197 3188 389.5 3250 246.0 -42.683 (0.410) Total enrolment 594 1408 339.7 1376 255.0 42.246 (0.301) Student teacher ratio 594 18.33 1.45 17.77 1.28 0.469 (<0.001) Teachers with a M.A. degree^ 198 0.330 0.130 0.472 0.120 -0.143 (<0.001) Senior regular teachers (with an adv. cert.) (proxy for exp.) 594 0.506 0.111 0.614 0.080 -0.100 (<0.001) Junior regular teachers (without an adv. cert.) 594 0.170 0.057 0.110 0.043 0.056 (<0.001) Short-term contract teachers 594 0.107 0.059 0.046 0.031 0.061 (<0.001) School passed infrastructure inspection 594 0.931 0.253 0.912 0.284 0.020 (0.457) # Average teacher and staff compensation (‘000KRW) 197 50881 4687 49277 3772 1730.08 (0.007) # Average base salary (‘000KRW) 198 25998 939 27637 1367 -1485.06 (<0.001) Average variable/performance pay (‘000KRW)# 197 24883 4446 21640 3014 3206.12 (<0.001) Within-school base salary dispersion (SD) 198 8.14 0.402 7.53 0.383 0.582 (<0.001) Within-school teacher type (adv. cert) dispersion (SD) 594 0.425 0.065 0.351 0.064 0.069 (<0.001) Years since establishment 594 30.83 14.27 19.50 11.63 9.424 (<0.001) Tracking – Korean^ 198 0.175 0.381 0.097 0.298 0.074 (0.185) Tracking – Mathematics^ 198 0.968 0.176 0.986 0.118 -0.021 (0.235) Tracking – English^ 198 0.960 0.196 0.972 0.165 0.022 (0.391) Tracking – Science^ 198 0.032 0.176 0.014 0.118 0.008 (0.680) Tracking – Social studies^ 198 0.016 0.125 0.028 0.165 -0.034 (0.239) Notes: School characteristics are drawn from www.schoolinfo.go.kr for the years 2008, 2009, and 2010, except that variables about teacher compensation (#) are only available for 197 schools in 2009 and variables about teachers with a M.A. degree and whether the school uses tracking (^) are drawn from the NAEA 2010 principal survey. Average variable/performance pay is computed by taking the difference between average compensation and average base salary (measured at the administrative district and school type level). Mean difference is tested by regressing each variable on a dummy of private school and a set of school district or school district-year fixed effects; standard errors are clustered at the school level. Appendix II provides detailed data sources and variable definitions.

32

Table 4 Perceptions of Principals by School Type Private

Public

F.E. Difference

Surveyed among the full sample

Obs. 198

Mean 0.833

S.D. 0.374

Mean 0.792

S.D. 0.409

Mean 0.030

(p-value) (0.648)

Responded to survey among the full sample

198

0.540

0.500

0.597

0.494

-0.052

(0.516)

Responded to survey among survey sample

162

0.648

0.480

0.754

0.434

-0.107

(0.216)

Good academic performance is important*

111

0.324

0.199

0.395

0.198

-0.173

(0.101)

Entering into prestigious university is important*

111

0.706

0.225

0.395

0.210

0.292

(0.007)

Good disciplines and behaviors are important*

111

0.853

0.132

0.884

0.199

0.010

(0.875)

Creativity development is important*

111

0.088

0.180

0.302

0.156

-0.148

(0.103)

Excel in extracurricular activities is important*

111

0.000

0.000

0.023

0.046

-0.008

(0.389)

School policies are more flexible^

111

0.677

0.205

0.076

0.089

0.625

(<0.001)

Principal job is more secure^

111

0.062

0.074

0.640

0.162

-0.604

(<0.001)

Teacher job is more secure^

111

0.070

0.112

0.445

0.173

-0.382

(<0.001)

Stronger principal incentives to deliver good outcomes^

111

0.704

0.158

0.051

0.077

0.699

(<0.001)

Stronger teacher incentives to deliver good outcomes^

111

0.736

0.157

0.059

0.063

0.697

(<0.001)

Stronger punishment on teachers for poor performance^ Stronger encouragement for innovation in teaching^

111

0.432

0.157

0.147

0.117

0.299

(<0.001)

111

0.588

0.188

0.129

0.121

0.446

(<0.001)

Teachers are rewarded for good performance

111

0.397

0.493

0.256

0.441

0.184

(0.079)

Teachers can be punished for poor performance

111

0.382

0.490

0.046

0.213

0.311

(<0.001)

Notes: The data came from the high school principal survey that the authors conducted in 2013. Each principal in the survey is asked to pick two most important outcomes out of five measures of student achievement indicated in the table. A variable indicated with an asterisk (*) takes the value of one when a principal picked a particular outcome as one of the two most important measures. Each principal is also asked to compare whether private schools are more than, less than, or equal to public schools in having a set of characteristics. A variable indicated with a hat (^) measures the percentage of a principal’s school type being perceived as having more of a particular characteristics in the school district. Mean difference is tested by regressing each variable on a dummy of private school and a set of school district fixed effects; standard errors are clustered at the school level. Appendix II provides detailed data sources and variable definitions.

33

Table 5 The Effects of Private Schooling on School-level Students’ Outcomes

A. Specification without district-year FEs Private R-squared District-year F.Es B. Specification with district-year FEs Private R-squared District-year F.Es C. Weighted average treatment specification Private

(1) Percent dropout

(2) Percent graduation

(3) Disc. problems per student

(4) Percent College

(5) Percent Four-year

(6) Percent Two-year

-0.001 (0.001) 0.0004 No

0.001 (0.007) 0.0001 No

-0.001*** (0.0002) 0.082 No

0.032*** (0.010) 0.035 No

0.061*** (0.007) 0.183 No

-0.029** (0.012) 0.027 No

-0.001 (0.001) 0.041 Yes

0.002 (0.005) 0.066 Yes

-0.001*** (0.000) 0.146 Yes

0.044*** (0.008) 0.397 Yes

0.063*** (0.007) 0.392 Yes

-0.019** (0.009) 0.454 Yes

-0.001 0.001 -0.001*** 0.042*** 0.065*** -0.023*** (0.001) (0.005) (0.000) (0.008) (0.006) (0.009) Number of schools 198 198 198 198 198 198 11 11 11 11 11 Number of districts 11 Sample year 2008-2010 2008-2010 2008-2010 2008-2010 2008-2010 2008-2010 Grade level 10-12 12 10-12 12 12 12 Mean of dependent variable (public) 0.017 0.976 0.0015 0.562 0.347 0.215 Notes: All regressions in panels A and B are weighted by the total number of students in the denominators. Each regression specification in Panel C is the student population weighted average private school coefficient derived on the basis of a set of within-district regressions of the outcome variable against a private school indicator (also weighted by the total number of students in the denominator). Appendix II provides detailed data sources and variable definitions. Robust standard errors clustered by school are reported in parentheses. *** p<0.01, ** p<0.05, * p<0.1.

34

Table 6 Differences in Missing NAEA and CSAT Scores by School Type (1) NAEAKorean Missing Private

(2) NAEAMath Missing

(3) NAEAEnglish Missing

(4) CSATKorean Missing

(5) CSATEnglish Missing

-0.017*** -0.017*** -0.017*** -0.030*** -0.031*** (0.003) (0.003) (0.003) (0.007) (0.007) R-squared 0.003 0.003 0.003 0.034 0.033 District-year F.Es Yes Yes Yes Yes Yes Observations 89825 89825 89825 189501 189501 Sample year 2010 2010 2010 2009-2010 2009-2010 Grade level 11 11 11 12 12 Mean of dep. var. 0.037 0.037 0.036 0.129 0.134 Notes: Each of the dependent variables measures whether the student is absent on the day of the particular test. Robust standard errors clustered by school are reported in parentheses. Appendix II provides detailed data sources and variable definitions. *** p<0.01, ** p<0.05, * p<0.1.

35

Table 7 The Effects of Private Schooling on Students’ NAEA and CSAT Scores

A. Specification without district-year FEs Private R-squared B. Specification with district-year FEs Private R-squared C. Weighted average treatment specification Private

(1) NAEAKorean

(2) NAEAMath

(3) NAEAEnglish

(4) CSATKorean

(5) CSATEnglish

0.146*** (0.041) 0.005

0.138*** (0.042) 0.005

0.144*** (0.055) 0.005

0.097*** (0.033) 0.003

0.110** (0.046) 0.003

0.125*** (0.039) 0.024

0.116*** (0.030) 0.042

0.116*** (0.039) 0.066

0.070*** (0.025) 0.029

0.079*** (0.030) 0.064

0.128*** (0.037)

0.126*** (0.030)

0.117*** (0.038)

0.069*** (0.024)

0.076*** (0.028)

D. Sharp bounds specification Lower-bound

0.072* 0.069** 0.083** 0.002 0.002 (0.044) (0.030) (0.043) (0.032) (0.035) Upper-bound 0.151*** 0.140*** 0.126*** 0.107** 0.111** (0.048) (0.035) (0.047) (0.044) (0.045) Trim. proportion 0.016 0.016 0.015 0.030 0.031 Observations 89825 89825 89825 189501 189501 Mean of dependent variable -0.095 -0.090 -0.094 -0.062 -0.071 Notes: The NAEA and CSAT scores are normalized to have mean zero and variance one. Panels A, B, and C include missing values replaced with the average test score by school type within each district. Sharp bound effects are estimated using Lee’s (2009) sharp-bound estimators. Mean of dependent variable is for public schools (which includes replacement of missing values). Robust standard errors clustered by school are reported in parentheses. Bootstrapped standard errors (5000 repetitions) clustered by school are reported in parentheses for the sharp-bound estimates in panel D. Appendix II provides detailed data sources and variable definitions. *** p<0.01, ** p<0.05, * p<0.1.

36

Table 8 Heterogeneous Effects of Private Schooling on Test Scores by Household Type

A. Dual-Parent Household Private Observations R-squared B. Non-Dual-Parent Household Private

(1) NAEA – Korean

(2) NAEA – Math

(3) NAEA – English

0.115*** (0.039) 78621 0.025

0.109*** (0.031) 78621 0.044

0.109*** (0.039) 78621 0.070

0.191*** 0.136*** 0.146*** (0.045) (0.029) (0.040) Observations 11204 11204 11204 R-squared 0.019 0.027 0.042 Difference in private school effect by household type: B-A 0.076 0.027 0.037 (0.060) (0.042) (0.056) Notes: All specifications include a set of district fixed effects. Missing values of NAEA scores are replaced with the average test score by school type within each district. Robust standard errors clustered by school are reported in parentheses. *** p<0.01, ** p<0.05, * p<0.1.

37

Table 9 Teacher Characteristics that Explain the Private School Effects Estimated Private School Effect On (1) (2) (3) (4) (5) (6) (7) Disc. prob. Percent NAEANAEANAEACSATCSATper student College Korean Math English Korean English % Short-term contract teachers 0.000 0.110 1.035 0.784** 1.992* 2.385*** 1.729*** (0.005) (0.272) (1.498) (0.325) (1.059) (0.429) (0.389) % Junior regular teachers (without adv. cert.) -0.011** 0.935** 2.602 -0.446 0.747 0.547 1.085 (0.005) (0.411) (1.678) (0.409) (1.097) (0.941) (0.731) Student-teacher ratio 0.000 -0.016 -0.031 0.042*** -0.006 -0.068 -0.074* (0.0002) (0.014) (0.034) (0.009) (0.024) (0.054) (0.044) Observations 126 126 126 126 126 126 126 R-squared 0.110 0.193 0.177 0.329 0.224 0.749 0.670 Notes: The sample includes all private schools in Seoul. The dependent variable is the estimated effect of a private school on the outcome of interest. Each private school’s effect on an outcome is estimated in a district-year fixed effects specification that controls for the percentage of students on lunch support where all public schools in Seoul are served as the control group. Similarly, each private school’s teacher characteristics are also measured in relative term to all public schools. The share of short-term contract teachers, share of junior regular teachers, share of senior regular teachers, and share of other non-teaching staff members sum up to one. The share of junior regular teachers is highly negatively correlated with the share of senior regular teachers. Each regression is weighted by the inverse of the variance of the estimated private school effect. Robust standard errors clustered by school are reported in parentheses. *** p<0.01, ** p<0.05, * p<0.1.

38

Table 10 The Relationship between Student Outcomes and Teacher Characteristics in Public Schools (1) (2) (3) (4) (5) (6) (7) Disc. prob. Percent NAEANAEANAEACSATCSATper student College Korean Math English Korean English % Short-term contract teachers 0.001 0.528*** -1.818** -2.144* -2.274 -0.962* -1.780** (0.003) (0.198) (0.855) (1.136) (1.432) (0.505) (0.783) % Junior regular teachers (without adv. cert.) 0.011*** 0.379*** -1.857*** -2.755*** -3.548*** -1.376*** -2.274*** (0.003) (0.132) (0.562) (0.637) (0.798) (0.397) (0.590) % Senior regular teachers (with an adv. cert.) -0.005*** -0.170** 0.694* 0.946* 1.221* 0.655** 1.056** (0.002) (0.081) (0.397) (0.563) (0.664) (0.268) (0.408) Observations 216 216 31910 31910 31910 68345 68345 Notes: Each coefficient is obtained by regressing an outcome against a single school-level teacher characteristics of interest for the sample of public schools only. The share of junior regular teacher is negatively correlated with four-year college attendance rates and positively correlated with two-year college attendance rates; the net relationship with overall college attendance rates is positive (column 2). Regressions for school-level outcomes are weighted by the population size in the dominator. Robust standard errors clustered by school are reported in parentheses. *** p<0.01, ** p<0.05, * p<0.1.

39

Table 11 Robustness to Various Alternative Explanations (1) Disc. prob. per student

(2) Percent college

(3) NAEAKorean

(4) NAEAMath

(5) NAEAEnglish

(6) CSATKorean

(7) CSATEnglish

-0.001***

0.035***

0.142***

0.136***

0.156***

0.081***

0.100***

(0.0002)

(0.010)

(0.040)

(0.034)

(0.044)

(0.028)

(0.035)

R-squared

0.155

0.362

0.029

0.046

0.071

0.031

0.067

District-year F.Es

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Observations

441

441

67976

67976

67976

143421

143421

Mean of dep. var.

0.001

0.580

0.017

0.016

0.031

0.014

0.023

-0.0002

-0.013

0.021

0.005

0.009

0.033

0.032

(0.0002)

(0.014)

(0.060)

(0.042)

(0.052)

(0.042)

(0.043)

R-squared

0.101

0.412

0.019

0.036

0.060

0.031

0.067

District-year F.Es

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Observations

378

378

57915

57915

57915

121156

121156

Mean of dep. var.

0.001

0.598

0.063

0.078

0.086

0.035

0.039

-0.001***

0.050***

0.097**

0.085**

0.086*

0.039

0.041

(0.0002)

(0.010)

(0.049)

(0.036)

(0.048)

(0.031)

(0.035)

0.0003

0.012

-0.055

-0.063*

-0.061

-0.061*

-0.076**

(0.0002)

(0.010)

(0.047)

(0.037)

(0.049)

(0.031)

(0.036)

R-squared

0.154

0.400

0.025

0.043

0.066

0.030

0.065

District-year F.Es

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Observations

594

594

89825

89825

89825

189501

189501

Mean of dep. var.

0.001

0.586

-0.001

-0.001

-0.001

-0.0001

-0.001

-0.001***

0.051***

0.106**

0.096***

0.088*

0.042

0.042

(0.0002)

(0.009)

(0.049)

(0.034)

(0.046)

(0.030)

(0.033)

0.0002

0.049**

-0.120

-0.130*

-0.181

-0.179**

-0.243***

(0.0005)

(0.023)

(0.154)

(0.077)

(0.123)

(0.081)

(0.078)

R-squared

0.147

0.404

0.024

0.043

0.067

0.030

0.066

District-year F.Es

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Observations

594

594

89825

89825

89825

189501

189501

Mean of dep. var.

0.001

0.586

-0.001

-0.001

-0.001

-0.0001

-0.001

A.

Schools since 1974

Private

B.

Effects of religion

Religious

C.

Control for coed

Private Coeducational

D.

Control for creativity development

Private Emphasis on creativity

Notes: Panel A includes only 147 schools established after the equalization policy of 1974. Panel B includes only 126 private schools. Panels C and D include all 198 schools. The emphasis on creativity variable used in panel D came from the (anonymous) principal survey and is measured at the district and school type level. Robust standard errors clustered by school are reported in parentheses. *** p<0.01, ** p<0.05, * p<0.1.

40

Figure 1 Geographical Distribution of School Districts and Types in Seoul

Note: The square boxes indicate the names of school districts and the administrative districts that each school district contains.

41

Figure 2

.005 .01 .015 Disc. prob. per student

40 10 0

.1 .2 Dropout rate

.3

0

.2

.4 .6 .8 Graduation rate

Private

Private

Public

Public

Public

1

.3

.4 .5 .6 .7 .8 College attendance rates

0

0

0

1

2

2

2

3

4

4

4

6

6

Private

5

0

0

0

0

20

20

40

30

60

1000 200 400 600 800

80

Distributions of School-Level Outcomes by School Type

.1 .2 .3 .4 .5 .6 Four-year college attendance rates

0 .1 .2 .3 .4 Two-year college attendance rates

Private

Private

Private

Public

Public

Public

Note: Kernel densities of various school-level outcomes. College attendance rates is the sum of four-year college attendance rates and two-year college attendance rates.

42

Figure 3

-4 -2 0 2 4 Standardized NAEA Korean Score

.3 .2 .1 0

0

0

.1

.1

.2

.2

.3

.3

.4

.4

.4

Distributions of NAEA and CSAT Scores by School Type

-4 -2 0 2 4 Standardized NAEA Math Score

-3 -2 -1 0 1 2 Standardized NAEA English Score

Private

Private

Public

Public

Public

0

0

.2

.2

.4

.4

.6

.6

Private

-4 -2 0 2 Standardized CSAT Korean Score

-3 -2 -1 0 1 2 Standardized CSAT English Score

Private

Private

Public

Public

Note: Kernel densities of NAEA and CSAT scores. Test scores are standardized to mean of zero and standard deviation of one.

43

Figure 4

0

0

2

5

4

6

10

8

Distributions of Teacher Characteristics by School Type

0

.1 .2 Share of short-term contract teachers

0

Public

.1 .2 .3 .4 Share of junior teachers (without adv. cert.) Private

Public

0

0

1

1

2

2

3

3

4

5

4

Private

.3

.2

.4 .6 .8 Share of senior teachers (with adv. cert.) Private

0

Public

.2 .4 .6 Share of teachers with a MA degree Private

.8

Public

Note: Kernel densities of teacher characteristics. Data for the percentage of short-term contract teachers, percentage of junior regular teachers, and percentage of senior regular teachers are available for 2008-2010, while data for the percentage of teachers with a M.A degree are available for 2010 only.

44

Figure 5 Relationship between Student Outcome and School Age by School Type

0 -.4 -.2

0 -.2 -.1

.5 .45

0

CSAT English scores .2

.1 .2 .3

.55

.6

.002 .004 .006

20 40 60 80 100 Years around

0

20 40 60 80 100 Years around

0

20 40 60 80 100 Years around

0

20 40 60 80 100 Years around

Private

Private

Private

Private

Public

Public

Public

Public

NAEA math scores

NAEA English scores .2 0 -.4 -.2

-.4 -.2

-.4 -.2

0

0

.2

.2

.4

.4

NAEA Korean scores

.4

0

CSAT Korean scores .4

% College attendance .65

Disc. prob. per student

0 20 40 60 80 100 Years around

0 20 40 60 80 100 Years around

0

20 40 60 80 100 Years around

Private

Private

Private

Public

Public

Public

45

Appendix I: For Online Publication Table A1 Robustness of Main Results to Exclusion of Certain Schools

A. Dropping Jungbu district Private Observations Number of schools R-squared B. Dropping schools that changed status Private

(1) Disc. prob. per student

(2) Percent college

(3) NAEAKorean

(4) NAEAMath

(5) NAEAEnglish

(6) CSATKorean

(7) CSATEnglish

-0.001*** (0.000) 531 177 0.144

0.038*** (0.008) 531 177 0.401

0.125*** (0.039) 81882 177 0.026

0.133*** (0.030) 81882 177 0.046

0.130*** (0.039) 81882 177 0.071

0.065*** (0.025) 173517 177 0.031

0.081*** (0.030) 173517 177 0.070

-0.001*** 0.052*** 0.113*** 0.080** 0.087* 0.046* 0.047 (0.000) (0.010) (0.041) (0.035) (0.044) (0.027) (0.034) Observations 486 486 73710 73710 73710 156299 156299 Number of schools 162 162 162 162 162 162 162 R-squared 0.154 0.400 0.018 0.034 0.052 0.023 0.054 Sample years 2008-2010 2008-2010 2010 2010 2010 2009-2010 2009-2010 Notes: All specifications include district-year fixed effects. In panel A, the sample excludes 21 schools in the school district named Jungbu that comprises administrative districts Jongno-Gu, Jung-Gu, and Yongsan-Gu, where conditional randomization based on students stated school preferences is used. In panel B, the sample includes 162 of the 198 schools used for the main analysis. For the 36 schools that are excluded from this sample, 34 of them converted into autonomous schools after 2010 and 2 (private) of them experienced principal or management changes at the time of survey. Robust standard errors clustered by school are reported in parentheses. *** p<0.01, ** p<0.05, * p<0.1.

46

Table A2 Distribution of Religious and Secular Schools by District School District 1 2 3 4 5 6 7 8 9 10 11 Total

Number of Schools 25 25 19 11 17 18 16 9 13 24 21 198

Private Secular 14 17 10 7 10 4 6 3 3 9 8 91

47

Private Religious 1 2 1 1 6 1 3 1 5 5 9 35

Public Secular 10 6 8 3 1 13 7 5 5 10 4 72

Table A3 The Relationships between Pre-determined Student Characteristics and Outcomes (1) Single-mom: Male

(2) Dual-parent: Male

Dependent variable: Ave. NAEA std. score

(3) Single-mom: Female

(4) Dual-parent: Female

(5) Lunch support

(6) Welfare assistance

-0.289*** 0.364*** -0.287*** 0.352*** -1.005** -2.192*** (0.017) (0.016) (0.017) (0.016) (0.484) (0.808) R-squared 0.001 0.001 0.001 0.002 0.057 0.058 District-year F.Es Yes Yes Yes Yes Yes Yes Sample year 2010 2010 2010 2010 2010 2010 No. of observations 48566 48566 41259 41259 89825 89825 No. of schools 138 138 132 132 198 198 No. of coed schools 72 72 72 72 72 72 No. of all-boys schools 66 66 0 0 66 66 No. of all-girls schools 0 0 60 60 60 60 Notes: Each specification uses the average NAEA standardized test score as the dependent variable and includes each student characteristics, a private school indicator and district-year fixed effects on the right hand side. Average NAEA standardized test score is the average of Korean, math and English NAEA standardized test scores. Singlemother is a dummy that takes the value of 1 for students living with the mother only. Dual-parent is a dummy that takes the value of 1 for students living with both parents. Lunch support and welfare assistance are the share of student population in the school who are on lunch support and on government welfare assistance respectively. See Appendix II for detailed information on data sources and variable definitions. Robust standard errors clustered by school are reported in parentheses. *** p<0.01, ** p<0.05, * p<0.1.

48

Table A4 Additional Verification of Random Assignment

A. NAEA 2008 student-level data Private District-year F.Es Number of students sampled R-squared Sample year Grade level

B. KEEP student-level data Private

(1) Father College -0.012 (0.026) Yes 4385 0.062 2008 10

(2) Father Dropout -0.003 (0.007) Yes 4385 0.011 2008 10

Percentile rank in school 0.353 (2.337)

Often absent from school -0.0004 (0.019)

(3) Mother College -0.013 (0.026) Yes 4385 0.070 2008 10 Received disciplinary action 0.008 (0.007)

(4) Mother Dropout 0.008 (0.006) Yes 4385 0.015 2008 10 Ave. mthly household income 1.344 (16.634)

Metropolitan area F.Es Yes Yes Yes Yes 537 600 600 583 Number of students sampled 0.139 0.013 0.009 0.146 R-squared 2005 2005 2005 2005 Sample year 10 10 10 10 Grade level Notes: Although we examine the outcomes of tenth-grade students surveyed in NAEA 2008 (who later became high school seniors in 2010), NAEA 2008 only randomly sampled 5 percent of the students, and so we cannot directly correlate students’ predetermined characteristics with the outcome measures that we analyze. The problem with the KEEP dataset is that it lacks school district information and has a small Seoul-based sample. Panel A reports estimates using the NAEA 2008 grade-10 (first year of high school) student survey data (5% random sample). College is an indicator of whether the student’s parent has at least some college education; Dropout is an indicator of whether the student’s parent has less than a high school diploma. Panel B reports estimates (sampling weights adjusted) based on the Korean Education and Employment Panel (KEEP) data. The student’s middle school teacher provided information on percentile rank of academic performance, whether the student received any disciplinary action, and whether the student was often absent from school. The student’s parent or guardian provided information on average monthly household income. We restrict the sample to students living in the seven major equalization policy metropolitan areas and attending general academic high school in 2005. Because KEEP does not provide school district information, we regress each of the dependent variables against the private school dummy and a set of metropolitan area fixed effects. The results are similar when we restrict the sample to Seoul only (256 students). We focus on the major metropolitan areas with equalization policy as it increases the sample size and these areas are more similar than smaller provincial cities. These seven metropolitan areas are Seoul, Pusan, Daegu, Incheon, Gwangju, Daejeon, and Ulsan. School district information is not available in KEEP. Robust standard errors clustered by school are reported in parentheses. *** p<0.01, ** p<0.05, * p<0.1.

49

Table A5 Differences in Student Outcomes between Converted and Non-Converted Schools Converted Mean S.D.

Obs.

Non-Converted Mean S.D.

F.E. Difference Mean (p-value)

A. Public Schools Disciplinary problems per student 216 0.002 0.002 0.001 0.002 0.0009 (0.062) Four-year college attendance rates 216 0.312 0.043 0.352 0.064 -0.020 (0.070) Two-year college attendance rates 216 0.280 0.052 0.199 0.089 0.050 (0.007) College attendance rates 216 0.592 0.057 0.551 0.082 -0.030 (0.030) NAEA – Korean 31910 -0.286 0.982 -0.049 1.019 -0.148 (0.024) NAEA – Math 31910 -0.326 0.870 -0.034 0.998 -0.174 (<0.001) NAEA – English 31910 -0.396 0.901 -0.021 1.017 -0.227 (0.001) CSAT – Korean 68345 -0.262 0.945 -0.014 0.940 -0.176 (<0.001) CSAT – English 68345 -0.339 0.881 -0.005 0.943 -0.221 (<0.001) B. Private Schools Disciplinary problems per student 378 0.0007 0.001 0.0006 0.001 -0.0002 (0.378) Four-year college attendance rates 378 0.408 0.069 0.405 0.061 0.003 (0.779) Two-year college attendance rates 378 0.153 0.083 0.192 0.076 0.038 (0.002) College attendance rates 378 0.561 0.088 0.597 0.082 0.041 (0.001) NAEA – Korean 57915 -0.023 1.027 0.066 0.952 -0.084 (0.244) NAEA – Math 57915 0.100 1.049 0.037 0.973 0.063 (0.216) NAEA – English 57915 0.058 1.034 0.048 0.960 0.004 (0.952) CSAT – Korean 121156 0.041 0.950 0.034 0.923 0.013 (0.747) CSAT – English 121156 0.080 0.958 0.031 0.924 0.053 (0.222) Notes: Converted schools include 21 private schools and 15 public schools that were excluded from our principal survey. 34 of these schools converted into autonomous schools after 2010 and 2 private schools experienced principal or management changes at the time of survey.

50

Table A6 The Relationship between the Share of Private Schools and Mean Household Characteristics across Districts (1) Age Years of education Labor force participation rates

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

0.083 (0.075) -0.015 (0.104) -5.008 (4.581)

Ownership of multiple houses

-0.980 (5.070)

Residential size per HH member

-0.002 (0.002)

Annual HH’s labor income

-0.000 (0.000)

Yearly household expenditure

-0.001 (0.002)

Household financial asset

-0.000 (0.000)

Household real estate asset

0.000 (0.000) Constant -2.520 0.788 3.454 0.703 0.761*** 0.715** 0.853** 0.634*** 0.561** (2.827) (1.137) (2.586) (0.394) (0.192) (0.295) (0.323) (0.100) (0.173) Observations 11 11 11 11 11 11 11 11 11 R-squared 0.121 0.002 0.117 0.004 0.056 0.010 0.053 0.001 0.018 Notes: The dependent variable is the share of private schools in each district. The explanatory variables in columns 1-3 came from Census 2010 individual data; variables in columns 4-5 came from Census 2010 household data; variables in columns 6-9 came from Korean Labor and Income Panel Survey 2010 household data. Robust standard errors reported in parentheses. *** p<0.01, ** p<0.05, * p<0.1.

51

Table A7 The Effects of Private Schooling on College Attendance Outcomes by Gender (1) (2) (3) (4) (5) (6) Percent Percent Percent Percent Percent Percent College: Four-year: Two-year: College: Four-year: Two-year: Male Male Male Female Female Female Private 0.051*** 0.072*** -0.021** 0.022*** 0.049*** -0.027** (0.010) (0.008) (0.009) (0.008) (0.010) (0.012) R-squared 0.473 0.431 0.495 0.436 0.365 0.481 District-year F.Es Yes Yes Yes Yes Yes Yes Number of schools 138 138 138 132 132 132 Number of coed schools 72 72 72 72 72 72 Number of all-boys schools 66 66 66 0 0 0 Number of all-girls schools 0 0 0 60 60 60 Sample year 2008-2010 2008-2010 2008-2010 2008-2010 2008-2010 2008-2010 Mean of dependent variable (public) 0.528 0.335 0.192 0.616 0.367 0.249 Notes: All regressions are weighted by the total number of students in the denominators. Robust standard errors clustered by school are reported in parentheses. *** p<0.01, ** p<0.05, * p<0.1.

52

Table A8 The Effects of Private Schooling on Student Outcomes by Gender Type of School Disc. prob. per student Single-Sex Private Number of schools Sample years R-squared

-0.001* (0.000) 126 2008-2010 0.134

Percent College 0.060*** (0.019) 66 2008-2010 0.520

NAEAKorean 0.122* (0.071) 31630 2010 0.035

Male NAEAMath 0.121* (0.062) 31630 2010 0.060

Coeducational Private

NAEAEnglish 0.073 (0.081) 31630 2010 0.083

Percent College 0.051*** (0.014) 60 2008-2010 0.506

NAEAKorean

Female NAEAMath

0.164*** (0.060) 26878 2010 0.021

0.098* (0.054) 26878 2010 0.035

NAEAEnglish 0.110 (0.083) 26878 2010 0.061

-0.001 0.051 0.069 0.047 0.130 0.032 0.013 0.053 0.046 (0.000)* (0.016)*** (0.061) (0.063) (0.073)* (0.014)** (0.051) (0.066) (0.079) Number of schools 72 72 16936 16936 16936 72 14381 14381 14381 Sample years 2008-2010 2008-2010 2010 2010 2010 2008-2010 2010 2010 2010 R-squared 0.165 0.504 0.023 0.041 0.061 0.500 0.021 0.032 0.064 Notes: All specifications include a set of district-year fixed effects. We omit results using the percentage of seniors graduated or the percentage of high school dropout as the dependent variable because private and public schools do not differ in these outcomes. We do not split disciplinary problems per student by gender because the data were not reported by gender at www.schoolinfo.go.kr. Robust standard errors clustered by school are reported in parentheses. *** p<0.01, ** p<0.05, * p<0.1.

53

Table A9 The Heterogeneous Effects of Private Schools on NAEA Test Scores by Gender Type of School

Single-Sex A. Dual-Parent Household Private Observations R-squared B. Non-Dual-Parent Household Private Observations R-squared Coeducational A. Dual-Parent Household Private Observations R-squared B. Non-Dual-Parent Household Private

Korean

Male Math

Korean

Female Math

English

English

0.106 (0.072) 27762 0.036

0.106* (0.062) 27762 0.063

0.058 (0.078) 27762 0.088

0.118** (0.050) 23503 0.027

0.055 (0.038) 23503 0.046

0.041 (0.052) 23503 0.090

0.152** (0.071) 3868 0.026

0.117** (0.050) 3868 0.045

0.065 (0.087) 3868 0.058

0.259*** (0.069) 3375 0.022

0.101** (0.043) 3375 0.023

0.102 (0.076) 3375 0.049

0.105** (0.042) 14757 0.034

0.052 (0.051) 14757 0.061

0.148*** (0.052) 14757 0.088

-0.018 (0.039) 12599 0.034

0.021 (0.056) 12599 0.055

-0.007 (0.058) 12599 0.103

0.191*** 0.175*** 0.240*** 0.181** 0.106 0.085 (0.071) (0.065) (0.077) (0.081) (0.067) (0.090) Observations 2179 2179 2179 1782 1782 1782 R-squared 0.047 0.056 0.073 0.051 0.054 0.085 Notes: All specifications include a set of district fixed effects, as well as controls for the share of students on welfare assistance and the share of students on lunch support to tighten the standard errors. Missing values of NAEA scores are replaced with the average test score by school type within each district. Robust standard errors clustered by school are reported in parentheses. *** p<0.01, ** p<0.05, * p<0.1.

54

Table A10 Private Tutoring as a Channel of Private School Effects (1) Disc. prob. per capita

(2) Four-year college

(3) Two-year college

Private

(4) NAEAKorean

(5) NAEAMath

(6) NAEAEnglish

-0.004*** 0.240*** -0.150*** 0.211*** 0.180*** 0.185*** (0.001) (0.052) (0.045) (0.052) (0.032) (0.043) Private tutoring -0.001*** 0.087*** -0.176*** 0.149*** 0.206*** 0.199*** (0.000) (0.015) (0.015) (0.007) (0.007) (0.007) Priv.×Priv. tutoring 0.001** -0.059*** 0.042*** -0.027*** -0.020** -0.021** (0.000) (0.018) (0.015) (0.009) (0.009) (0.010) R-squared 0.170 0.445 0.673 0.066 0.131 0.148 District-year F.Es Yes Yes Yes Yes Yes Yes Sample year 2008-2010 2008-2010 2008-2010 2010 2010 2010 Mean of dep. var. 0.001 0.409 0.180 0.063 0.078 0.086 Notes: Estimates in columns 1-3 are based on school-level regressions, while estimates in columns 4-6 are based on individual-level regressions. Specifications in columns 1 to 3 include 594 observations each (198 schools over 3 years) and specifications in columns 4-6 include 89825 observations each. Because data on private tutoring are drawn from the NAEA 2010 student survey, we can only link them to test score data (columns 4-6). For columns 13, we calculate the average value of private tutoring of grade-11 students for each school and use it as a proxy for private tutoring of all students attending that school. Private tutoring reported in NAEA 2010 is a categorical variable that takes a value from 1 to 5 in increasing order, where 1 means no private tutoring and 5 means more than 3 hours per week. The mean for private tutoring is 2.93 for those in private schools and 2.96 for those in public schools. The p-value for the difference is 0.61. The results are not sensitive if we use dummy variables transformed from the original categorical variable. Robust standard errors clustered by school are reported in parentheses. *** p<0.01, ** p<0.05, * p<0.1.

55

Appendix II: Data Appendix for Online Publication Variables Share of students on welfare assistance Share of minority students Share of students on lunch support Net transfer rate Single-mother household Dual-parent household Private tutoring

Teacher and staff salary per student (‘000KRW) Student-teacher ratio Teachers with a M.A. degree

Years Data Source A. School-level student characteristics 2010 The NAEA school principal survey, the KICE website 2010 The NAEA school principal survey, the KICE website 2008, 09, 10 The METS website 2008, 09, 10 The METS website B. Individual-level student characteristics 2010 The NAEA student survey, the KICE website, 11th grade 2010 The NAEA student survey, the KICE website, 11th grade 2010 The NAEA student survey, the KICE website, 11th grade C. School characteristics 2009 The METS website 2008, 09, 10 2010

Data Type

Note

Propriety

1

Propriety

2

Public

3

Public

4

Propriety

5

Propriety

6

Propriety

7

Public

8

Public Propriety

9 10

Public

11

Senior regular teachers (regular teachers with an advanced certificate; proxy for experience) Junior regular teachers (regular teachers without an advanced certificate) Teachers on short-term contracts School passed infrastructure inspection Average teacher and staff compensation (‘000KRW) Average base pay (‘000) Average variable/performance pay (‘0000 KRW) Within-school base salary dispersion (SD.)

2008, 09, 10

The METS website The NAEA school principal survey, the KICE website The METS website

2008, 09, 10

The METS website

Public

12

2008, 09, 10

The METS website

Public

13

2008, 09, 10

The METS website

Public

14

2009

The METS website

Public

15

2009 2009

The METS website The METS website

Public Public

16 17

2009

Public

18

Within-school teacher type (advanced certificate)

2008, 09, 10

(1) The data for the number of teachers by pay grade and type of school in each administrative district are from http://statistics.sen.go.kr; (2) The data of base salary by pay scale are from http://www.mospa.go.kr The METS website

Public

19

56

dispersion (SD.) Years since establishment Tracking – Korean; Math; English; Science; Social studies Religious school Dropout rate Seniors’ graduation rate College attendance rate Four-year college attendance rate Two-year college attendance rate Disciplinary problems per student NAEA – Korean; Math; English standardized test score NAEA – Missing test score (Korean ; Math; English) CSAT test score

2008, 09, 10 2010

The METS website The NAEA school principal survey, the KICE website

2008, 09, 10 Individual schools' websites D. School-level student outcomes 2008, 09, 10 The METS website 2008, 09, 10 The METS website 2008, 09, 10 The METS website 2008, 09, 10 The METS website 2008, 09, 10

The METS website

2008, 09, 10

The METS website

E. Individual-level student outcomes 2010 The NAEA test score data, the KICE website, 11th grade 2010

The KICE website

2009, 10 The KICE website F. Perception of principals 2013 Authors’ survey of principals

Public Propriety

20 21

Public

22

Public Public Public

23 24 25 26

Public Public Public

28

Propriety

29

Propriety

30

Propriety

31

Good academic performance Authors is important administered Entering into prestigious 2013 Authors’ survey of principals Authors university is important administered Good disciplines and 2013 Authors’ survey of principals Authors behaviors are important administered Creativity development is 2013 Authors’ survey of principals Authors important administered Excel in extracurricular 2013 Authors’ survey of principals Authors activities is important administered G. Principals’ Perception on Differences between Private and Public schools More flexible and 2013 Authors’ survey of principals Authors autonomous school policies administered Principal job security 2013 Authors’ survey of principals Authors administered Teacher job security 2013 Authors’ survey of principals Authors administered Principals’ incentives to 2013 Authors’ survey of principals Authors deliver good student administered outcomes Teachers’ incentives to 2013 Authors’ survey of principals Authors deliver good student administered outcomes Punishment for teachers for 2013 Authors’ survey of principals Authors

57

27

32 33 34 35 36

37 38 39 40

41

42

poor performance Whether teachers are 2013 encouraged to implement innovative classroom practices and solutions H. Additional data used in appendix Father – College 2008 Father – Dropout

2008

Mother – College

2008

Mother – Dropout

2008

Percentile rank in school

2005

Often absent from school

2005

Received disciplinary action

2005

Average monthly household income

2005

Authors’ survey of principals

The NAEA student survey data, the KICE website, 10th grade The NAEA student survey data, the KICE website, 10th grade The NAEA student survey data, the KICE website, 10th grade The NAEA student survey data, the KICE website, 10th grade The Korean Education and Employment Panel data The Korean Education and Employment Panel data The Korean Education and Employment Panel data The Korean Education and Employment Panel data

administered Authors administered

43

Propriety

44

Propriety

45

Propriety

46

Propriety

47

Public

48

Public

49

Public

50

Public

51

Notes 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17

Number of students in families on welfare assistance divided by total enrolment Number of ethnic minority students divided by total enrolment Number of students on lunch support divided by total enrolment Number of students transferred in minus number of students transferred out divided by total enrolment Indicated that mother is present and father is not in the household Indicated that mother and father are both present in the household Originally a categorical variable that takes a value from 1 to 5 in increasing order, where 1 means never, 2 means less than 1 hour, 3 means 1-2 hours, 4 means 2-3 hours, and 5 means more than 3 hours per week. The results are not sensitive if we use a dummy variable taking the value of 1 when the categorical value is greater than either 2, 3, 4, or 5. Total salaries and other remuneration divided by the total enrolment Total enrolment divided by total number of teachers Number of teachers with a M.A. degree divided by the total number of teachers The product of the number of regular teachers and the share of teachers with an advanced certificate divided by the total number of teachers The product of the number of regular teachers and the share of teachers without an advanced certificate divided by the total number of teachers The number of teachers on a short-term contract divided by the total number of teachers Whether the school passed its annual infrastructure inspection Total salaries and other remuneration divided by the total number of staff Average base pay is constructed by administrative-district level and school type using administrativedistrict-level-by-school-type information on the number of teachers in each pay grade and the corresponding base salary for each pay grade. Average salary minus average base pay.

58

18

18

19 20 21 22 23 24 25 26 27 28

29

30 3136 37 3839 4041 42 43 44 45 46 47 48 49

Within-school base salary dispersion is constructed using administrative-district-level-by-school-type information on the number of teachers in each pay grade and the corresponding base salary for each pay grade. Because the base salary dispersion data are aggregated at the administrative district level by school type, rather than at the school level, some administrative districts include a few schools that not included in our sample. We calculate within-school dispersion of senior teachers (those with an advanced certificate) using the following steps. We first create teacher-level (empty) variable using the information on number of teachers so that total number of observations matches the total number of teachers. We fill this variable by assigning 1 for the first x% of observations in each school, where x is the share of senior teachers at each school. Remaining values are assigned 0. This dummy variable now indicates whether a teacher has an advanced certificate in a school. We take standard deviation of the dummy variable for each school to obtain within-school dispersion of senior-teachers. The year the school was established An indicator of whether the school implements tracking An indicator of whether the school is religiously affiliated Number of dropouts divided by total enrolment Number of graduates divided by total number of high school seniors Number of students attending four-year and two-year colleges divided by total number of graduates Number of students attending four-year colleges divided by total number of graduates Number of students attending two-year colleges divided by total number of graduates Number of violent incidents divided by total enrolment Original scores are in the range of 100 and 300. We standardized them by the sample mean and standard deviation. The data include all students who participated in NAEA. An indicator of whether a student missed the NAEA test. To create this dummy variable, we checked the total enrolment of eleventh graders in the METS dataset against the total NAEA participants. When the enrolment number exceeds the number of (any) test takers, we expand the data by the difference between the two figures, and assign the value 1 for missing test score. All test takers have the value 0 for the dummy variable “missing test score.” Private school students are more likely to take CSAT than public school students. We replaced the missing scores of individuals who did not take CSAT with the average scores of students of the same gender in the same type of school in the same school district. Each principal is asked to pick two most important outcomes out of five measures of student achievement in this section. This variable takes the value of 1 when a principal picked this outcome as one of the two most important outcomes. Each principal is asked to compare whether (a) public schools are more; (b) private schools are more; or (c) they are equally flexible Each principal is asked to compare whether (a) public schools are more; (b) private schools are more; or (c) they are equally secure Each principal is asked to compare whether (a) public schools have more; (b) private schools have more; or (c) they have equal incentives Each principal is asked to compare whether (a) public schools are more; (b) private schools are more; or (c) they are equally likely to punish Each principal is asked to compare whether (a) public schools are more; (b) private schools are more; or (c) they are equally likely to encourage An indicator of whether the father of the student has any college education An indicator of whether the father of the student did not complete high school An indicator of whether the mother of the student has any college education An indicator of whether the mother of the student did not complete high school The percentile rank of a student’s academic performance in final middle school year. An indicator of whether the student was ever absent from school during the final middle school year.

59

An indicator of whether the student ever violated any disciplinary rules during the final middle school year. 51 The average household income of the student reported in the final middle school year. Source: The Korea Institute for Curriculum and Evaluation (KICE) website: www.kice.re.kr. The Ministry of Education, Technology and Science (METS) website: www.schoolinfo.go.kr. 50

60