A. Autonomous Private High School Policy

Since the 1970s, most high schools in Korea have been heavily regulated and subsidized by the government. The high school equalization policy of 1974 required that virtually all high schools in large cities, either private or public, follow a set of governmental rules which outlined nearly all aspects of their operation. For example, all high schools under the equalization policy must follow a uniform national curriculum, cannot select their own students, must charge tuition amounts set by the government, and must recruit teachers and principals certified by the government. Given that the equalization policy mandated private schools to charge an equal amount of tuition - to that of public schools, any deficits in the operating costs of private schools were fully refunded by the government. Consequently, private schools became nearly “public” in their operation, and this raised serious concerns about the lack of diversity and incentives in school education (e.g., Kim and Lee 2003).

In an attempt to diversify school education and spur competition among schools, the Korean government introduced the autonomous private high school policy in 2009. In this policy, autonomous private high schools are granted a greater degree of autonomy in their operation compared to traditional non-autonomous high schools. Essentially, these schools have substantial autonomy in many aspects of school management, including student/teacher/principal recruitment, tuition amounts, curriculum, textbooks, and academic terms (i.e., their choice of semester, trimester, or quarter systems). However, in order to enjoy this autonomy, the schools must bear the following responsibilities. First, autonomous schools cannot receive any financial subsidies from the government. Thus, they have a strong incentive to meet parental demand because schools failing to enroll enough students will not be able to finance their operations and will thus have to close. Second, they cannot charge more than three times the tuition of traditional non-autonomous schools. Third, they must reserve at least 20 percent of their places for students from low-income families. Fourth, they are allowed to select their own students, but not through entrance exams or interviews about academic knowledge. Instead, autonomous schools may select their students according to middle school grades, (non-academic) interviews, recommendation letters, and a lottery system. Particularly, the autonomous schools in Seoul in 2013, which constitute the main subject of this study, select their students in a lottery among applicants whose middle school grades are above the 50th percentile within their middle schools. Finally, the licenses of the autonomous schools must be re-evaluated by the local government (i.e., the education superintendent) every five years. If the schools do not pass this evaluation, they are converted to traditional non-autonomous schools.

Between 2009 and 2011, the Korean government designated 49 autonomous private high schools, mostly from existing traditional non-autonomous schools across the nation. More than half (25 out of 49) of the autonomous schools operate in Seoul, and the present study focuses on these. Since the introduction of autonomous schools, however, there has been heated debate as to whether the autonomous private high school policy should be maintained. Advocates of the policy argue that it can improve the overall quality of school education by spurring competition among schools. In contrast, opponents emphasize that autonomous schools can deteriorate the educational equity by providing better educational services only to those who can afford higher tuition levels of the autonomous schools. Particularly, as candidates pledging to abolish autonomous schools were elected in many cities and provinces in the 2014 local education superintendent elections,1 the conflict between advocates and opponents became even more serious. These conflicts garnered media headlines and were followed by a series of lawsuits involving parents, autonomous schools, local education superintendents, and the central government (i.e., the Ministry of Education).

B. Empirical Studies of Autonomous Private High School

In spite of the heated debate, surprisingly little is known about how autonomous schools affect the educational performance of students. Kim and Namkung (2014) evaluated the impact of autonomous private high schools on the academic achievement of their students. They compared the educational performance of students in autonomous schools with that of their peers in traditional non-autonomous schools after controlling for students’ family backgrounds. They conclude that there is a large gap in academic achievement (with standard deviations of approximately 0.6~0.7 and 0.7~1.0 for reading and math, respectively) between the two types of schools. However, given that autonomous schools select students based on their middle school grades, recommendations, interviews and related factors, there may be unobserved differences in students’ pre-determined academic quality levels that are not captured by their family backgrounds. In this respect, the estimated autonomous school premium reported by Kim and Namkung (2014) is likely to be overestimated.

Lee and Shin (2014) attempt to evaluate the spillover effect of autonomous schools. Specifically, they estimate how the designation of autonomous schools affects the academic achievement of the incumbent students in these schools (i.e., students entering the autonomous schools before the designation) and the incumbent students in the closest non-autonomous schools (i.e., students entering neighboring non-autonomous schools before the designation). They found that the designation of an autonomous school does not affect the academic achievement of incumbent students within the schools, whereas itnegatively affected the academic achievement of students in neighboring non-autonomous schools.

This study contributes to the literature by examining how autonomous schools affect the academic achievement of students. Although understanding whether and to what extent autonomous schools can improve the academic achievement of students would be the first step towards an evaluation of the desirability of the policy, there is surprisingly little evidence on this issue. Kim and Namkung (2014) reported a large impact, but their estimates are likely to be upwardly biased. Lee and Shin (2014) analyzed the spillover effect of autonomous schools but not the direct effect (i.e., how they affect students who enrolled in them after the designation), which constitutes the main objective of this study.

A. Identification Issue

Ideally, the causal effect of attendance at an autonomous school on student outcomes could easily be identified if admissions to autonomous schools were randomly determined. In fact, in 2013, the autonomous schools in Seoul admitted students by lottery. This indicates that, among the participants in the applicant lottery, admissions to the autonomous schools were randomly assigned. However, whether a student applied for entry into an autonomous school was clearly non-randomly determined. In 2013, only top 50 percent of students in terms of their middle school grades were able to apply for entry into an autonomous school. In addition, autonomous schools charged two to three times the tuition of non-autonomous regular high schools. These facts suggest that the identification of a causal effect of attending an autonomous school depends on how much one can control for middle school grades and the parental income of students as well as their preferences for an autonomous school.

B. Empirical Model

To address these concerns, I attempt to identify the autonomous school premium by controlling for ninth-grade (i.e., the third year of middle school) test scores of students along with other background characteristics that are likely to be correlated with their application decisions. Specifically, I estimate the following “value-added” model:

In equation (1), Y_i,m,s,10th indicates the tenth-grade (i.e., first year in high school) test scores of student i in high school s who graduated from middle school m. Autonomy_s is a dummy variable that takes a value of 1 if high school s is an autonomous school and 0 otherwise (i.e., a traditional non-autonomous school) as of the year 2013.2 Y_i,m,s,9th represents the ninth-grade (i.e., third year in middle school) test scores of student i. The lagged test scores are intended to capture the minimum required condition to apply for entry into an autonomous school (i.e., the top 50% in terms of middle school grades) and the difference in the pre-determined academic quality of students between the autonomous and the non-autonomous schools. X_i,9th refers to the baseline characteristics of student i measured in the ninth grade when the student decided upon the high schools to which he would apply. Specifically, X_i,9th includes variables on student characteristics (gender, disability, birth order) and family background (parental age, parental education, parental employment status, parental income, number of siblings, single parent). ρ_m represents middle school fixed effects, capturing any unobservable heterogeneity that students from the same middle schools have in common. In Seoul, nearly all students graduating from their elementary schools are assigned to their neighborhood middle schools. This suggests that ρ_m will also contain a substantial amount of information on the students’ residential locations. W_s refers to the characteristics of high school s, such as the gender composition (co-educational, male-only, female-only) and the establishment type (private or public). Given that autonomous schools are all private and mostly single-sex schools, controlling for these characteristics is particularly important for distinguishing the effect of school autonomy. Finally, ε_i,m,s,10th is an error term.

C. Estimation Results

I begin by estimating equation (1) with OLS, clustering standard errors at the middle school level. Table 3 shows the estimation results when the Korean, math, and English test scores are used as outcome variables. Column (1) of Table 3 shows the simple regression results without any covariate, which basically compares the average test scores of autonomous school students with those of non-autonomous school students. On average, students in autonomous schools achieve higher test scores than their peers in traditional non-autonomous schools by about 0.76, 1.00, and 0.97 standard deviations in Korean, math, and English, respectively. In column (2), I include student characteristics (gender, disability, first-born child), family characteristics (number of siblings, single parent, parental age, parental education, parental employment status, and parental income), high school characteristics (gender composition, establishment type), and dummies for the middle schools from which the students graduated (i.e., middle school fixed effects) as control variables. When these characteristics are controlled, the estimated test score gap between the two groups is slightly reduced to about 0.69, 0.88, and 0.84 standard deviations in Korean, math, and English, respectively. These estimates are comparable to those reported in Kim and Namkung (2014) (0.6~0.7 and 0.7~1.0 standard deviations in Korean and math, respectively), who mainly estimated the impact of autonomous schools on the academic achievement of students by regressing test scores on school types after controlling for student, family, and school characteristics. The estimation results in column (2), in conjunction with the results in column (1), also indicate that the student, family, and school characteristics can only account for approximately 10% of the observed test score gap between autonomous school students and non-autonomous school students.

TABLE 3

ESTIMATION RESULTS FOR THE TENTH-GRADE (FIRST YEAR IN HIGH SCHOOL) TEST SCORES

Note: Robust standard errors clustered at the middle school level are in parentheses (*** p<0.01, ** p<0.05, * p<0.1). Student, family, and school characteristics include gender, disability, first-born child, number of siblings, single parent, parental age, parental education, parental employment status, parental income, high school gender composition (male-only, female-only, co-educational), high school establishment type (private, public), and dummies for middle schools from which the students graduated (i.e., middle school fixed effects).

Exploiting the longitudinal structure of my data, in column (3) I include students’ pre-determined academic quality levels as measured by their ninth-grade (i.e., third year in middle school) test scores as controls added to the list of controls used in column (2). As discussed in the chapter II, only the top 50% of students in terms of their middle school grades could apply for entry into autonomous schools. Hence, students in autonomous schools are likely to perform better than their peers in non-autonomous schools even before they enter high schools. Adding lagged test scores as an additional control could control for this differences in the pre-determined academic quality levels between the two groups. When the baseline test scores are further controlled, the estimated achievement gap between the two groups is reduced substantially to about 0.39, 0.51, and 0.53 standard deviations in Korean, math, and English, respectively. These results suggest that Kim and Namkung (2014) likely overestimated the achievement effect of the autonomous schools by ignoring the differences in the pre-determined academic quality between the autonomous school students and the traditional non-autonomous school students.

In columns (1) to (3), I estimated equation (1) with the OLS method using different sets of covariates. Econometrically, however, estimating equation (1) with OLS will result in an inconsistent estimate when the error term (ε_i,m,h,10th) is serially correlated with its lagged term (ε_i,m,h,9th) because equation (1) includes a lagged dependent variable (Y_i,m,h,9th) as a regressor. To address this issue, I instrument the potentially endogenous ninth-grade test scores (Y_i,m,h,9th) with seventh-grade test scores (Y_i,m,h,7th). This allows the error term (ε_i,m,h,10th) to follow a “mild” serial correlation (i.e., AR(1) or AR(2) process) but not a “severe” one (i.e., AR(p) process with p≥3). Column (4) of Table 1, which is my most preferred specification, shows the two-stage least-square (2SLS) estimation results using seventh-grade test scores (Y_i,m,h,7th) as an instrument variable for ninth-grade test scores (Y_i,m,h,7th) . The impacts of attending an autonomous school on Korean, math, and English test scores are estimated to be 0.18, 0.32, and 0.33 standard deviations, respectively. Comparing these 2SLS estimates with the OLS estimates reported in column (3) reveals that the serial correlation issue discussed above is indeed serious.

Table 2 shows that students in autonomous schools tend to spend more on private tutoring than their peers in traditional non-autonomous schools. To the extent that private tutoring may improve the academic achievement of students, as discussed in a number of recent studies (e.g., Kang, 2012; Ryu and Kang, 2013), the estimated autonomous school premium reported in column (4) of Table 3 could be spuriously driven by the differences in private tutoring investment. To check for this possibility, I add the amount of private tutoring expenditures for each subject as an additional control variable in column (5). Even after controlling for these non-school educational inputs, the estimated autonomous school premium remains similar, indicating that the estimates reported in column (4) are largely attributable to the type of high schools the students attend and not to differences in private tutoring expenditures. Finally, in Table 4, I repeat the above-mentioned analysis using eleventh-grade (i.e., second year in high school) test scores as an outcome variable. Specifically, I estimate equation (1) using Y_i,m,h,11th as the left-hand-side variable instead of Y_i,m,h,10th . The results are roughly similar to those reported in Table 3. In terms of my preferred specification (column 4), attending an autonomous school improve tenth-grade test scores by 0.18, 0.32, and 0.33 standard deviations and eleventh-grade test scores by 0.24, 0.28, and 0.34 standard deviations in Korean, math, and English, respectively.

TABLE 4

ESTIMATION RESULTS FOR ELEVENTH-GRADE (SECOND YEAR IN HIGH SCHOOL) TEST SCORES

Note: Robust standard errors clustered at the middle school level are in parentheses (*** p<0.01, ** p<0.05, * p<0.1). Student, family, and school characteristics include gender, disability, first-born child, number of siblings, single parent, parental age, parental education, parental employment status, parental income, high school gender composition (male-only, female-only, co-educational), high school establishment type (private, public), and dummies for middle schools from which the students graduated (i.e., middle school fixed effects).

D. Falsification Test

Tables 3 and 4 show that students attending the autonomous schools outperform their peers attending non-autonomous schools after controlling for student, family and school characteristics (including middle school fixed effects), baseline academic performance, and private tutoring expenditures. However, whether the estimated achievement gap between the two groups of students reflects the causal effect of attending an autonomous school remains questionable. For example, it is still possible that the estimated achievement gap reflects unobservable differences in pre-determined academic quality levels between the two groups of students.

In order to determine whether equation (1) correctly identifies the causal effect of attending an autonomous school, I perform the following falsification test. Specifically, I estimate the impact of attending an autonomous school on the pre-determined academic performance of students. Specifically, I replace the outcome variable of equation (1) with eighth-grade test scores (Y_i,m,h,8th) . Given that the eighth-grade test scores were determined before the students entered high school, the autonomous school attendance of students cannot causally affect their eighth-grade test scores.

Table 5 shows the falsification test results. Columns (1) to (3) report positive and statistically significant impacts of the autonomous high schools, indicating that the corresponding regression equations are likely to be misspecified. On the other hand, in columns (4) and (5), my preferred specifications, I do not find any statistically significant effect for Korean and math. These results suggest that the estimated achievement gaps in the Korean and math test scores reported in columns (4) and (5) of Tables 3 and 4 are not driven by model misspecifications but instead reflect the causal effects of attending an autonomous school. For English test scores, however, the estimates from the falsification test are marginally significant at the 10% level, suggesting that the estimation results for English test scores should be interpreted with caution.

TABLE 5

FALSIFICATION TEST RESULTS FOR PRE-DETERMINED EIGHTH-GRADE (SECOND YEAR IN MIDDLE SCHOOL) TEST SCORES

Note: Robust standard errors clustered at the middle school level are in parentheses (*** p<0.01, ** p<0.05, * p<0.1). Student, family, and school characteristics include gender, disability, first-born child, number of siblings, single parent, parental age, parental education, parental employment status, and parental income, high school gender composition (male-only, female-only, co-educational), high school establishment type (private, public), and dummies for the middle schools from which the students graduated (i.e., middle school fixed effects).

E. Subgroup Analysis

As discussed in chapter II, autonomous schools can charge up to three times the tuition of traditional non-autonomous schools in exchange for receiving no governmental subsidies. This feature raises the serious public concern that the autonomous schools can only serve students from high-income families. In this respect, it would be worthwhile to determine how the observed autonomous school premium varies across students' family backgrounds.

To address this issue, I divide the estimation sample into the two subgroups of a high-income sample and a low-income sample. The high-income sample consists of students whose parental income in 2012, when the students were enrolled in the ninth grade, or their third year of middle school, is greater than or equal to or the median (KRW 4,500,000). Accordingly, the low-income sample consists of students whose parental income in the ninth grade is below the median. For each subgroup, I estimate the value-added model of equation (1) using the 2SLS method after controlling for the amount of private tutoring expenditures. This specification is comparable to the regression model used for column (5) of Tables 3, 4, and 5.

Table 6 summarizes the estimation results. Overall, I do not find any clear evidence that the autonomous school premium varies according to students' family backgrounds. These results suggest that the benefits that accrue from school autonomy and incentives can be enjoyed by all students regardless of their family background.

TABLE 6

ESTIMATED EFFECTS OF ATTENDING AN AUTONOMOUS SCHOOL BY INCOME LEVEL

Note: Robust standard errors clustered at the middle school level are in parentheses (*** p<0.01, ** p<0.05, * p<0.1). Student, family, and school characteristics of gender, disability, first-born child, number of siblings, single parent, parental age, parental education, parental employment status, parental income, high school gender composition (male-only, female-only, co-educational), high school establishment type (private, public), and dummies for middle schools from which the students graduated (i.e., middle school fixed effects), with the amount of private tutoring expenditures controlled. The low-income group refers to students whose parental monthly income is below KRW 4,500,000. The high-income group refers to those whose parental monthly income is greater than or equal to KRW 4,500,000.