School Choice , Student Performance , and Teacher and School Characteristics : The Chilean Case

The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the view of the World Bank, its Executive Directors, or the countries they represent. Policy Research Working Papers are available online at http://econ.worldbank.org.


Introduction
A great deal of research on the effects of school choice in Chile has centered on the question of whether private voucher schools are more effective than public schools.By "effective," we mean higher levels of student achievement.Chile has received much attention because it has one of the oldest and largest voucher programs in the world.
Unlike the small-scale voucher programs in several U.S. cities, Chile began implementing a nationwide voucher program in 1980.
Studies on the effectiveness of private schools in Chile have provided mixed results.Although early researchers found some positive effects of private voucher schools (mainly because they failed to control for selection bias; see, for examples, Rodriguez 1988;Aedo and Larrañaga 1994;and Aedo 1997), the most recent research indicates no significant differences in student achievement among public and private voucher schools (Mizala and Romaguera 2000;McEwan andCarnoy 1999, 2000;Carnoy and McEwan 2001;McEwan 2001;Hsieh and Urquiola 2001).When researchers differentiate among private voucher schools with religious affiliations, however, they tend to find that Catholic voucher schools are more effective than public schools (McEwan andCarnoy 1999 and2000;Carnoy and McEwan 2001;McEwan 2001).
How schools change in response to the increased competition generated by voucher programs Chile and elsewhere has received less attention from researchers.In the United States, Hoxby (2000) explores the effects of school choice on the teaching professio n and finds that school choice results in increased demand for teachers with several characteristics generally associated with increased learning.In the Chilean context, Hsieh and Urquiola (2001) explore the effect of school choice on student sorting and find that choice results in a great deal of sorting by socioeconomic background, thus leading the authors to question the positive effects of private schools on student performance.
While most of the research on Chile has focused on the effects of public and private schools on performance, more research is needed on how schools in different sectors end up producing different results.In other words, how does school choice affect the kinds of teachers that schools employ, the types of students they serve, and the management strategies that schools adopt?More importantly, to what extent do differences among schools in these factors affect student outcomes?I use a unique data set of teachers in Chile that provides information on teacher demographic and labor market characteristics, as well as teachers' perceptions on school management.I match these teacher data with school-level data on student achievement from a national assessment data set (SIMCE).I find that after a decade of reform, public and private schools in Chile are more similar than they are different in terms of teacher characteristics and school management policies.In fact, there is greater variation among the schools within a sector than among sectors in teacher, student characteristics as well as in school management measures.Interestingly, I find that regardless of sector, schools that provide teachers with greater autonomy and, simultaneously, have decentralized decisionmaking tend to have higher student outcomes as measured by standardized test scores.

The Chilean Voucher Program
In 1980, the Chilean central government transferred school administration to municipal governments and transformed education financing.Before 1980, the central government, through the Ministry of Education, was responsible for school administration (including teacher hiring, promotion, and firing) and for assigning school budgets.Under the 1980 reform, municipal governments took over school administration and began receiving monthly payments from the Ministry of Education based on a fixed amount per student multiplied by the number of students enrolled in each school.This fixed amount was identical for municipal and private schools that did not charge tuition.Thus, the reform established a base voucher level, which varies according to school location and the level of schooling (McEwan and Carnoy 2000). 1   Because this reform is one where money follows the student, it involves real school choice.Under the voucher system, families can choose to send their children to free subsidized schools, either municipal or private, or they can choose fee-paying private schools if they can afford the tuition fees (Mizala and Romaguera 2000).One result of the reform has been a substantial expansion of the private subsidized school system.Figure 1 shows the distribution of primary education enrollment by sector in 1981 and 1999, the most recent year for which data are available.In 1981, around 15 percent of students were enrolled in private voucher schools and almost 80 percent in public schools.By 1999, around 35 percent of enrollments were in private voucher schools, and enrollment in municipal schools had dropped to 54 percent (Chilean Ministry of Education 2001). 1 Specifically, the base voucher is adjusted by grade level and selected municipalities receive compensation for high poverty or isolation (McEwan 2001).

Empirical Strategy
The empirical strategy is twofold.I first explore the sources of variation in student outcome measures, school and teacher characteristics.This provides the most unrestricted way to assess how the variables of interest vary among sectors, among schools within a sector, and within schools themselves.Second, I conduct weighted least squares (WLS) regression analyses of the relationship between student outcomes and the teacher and school characteristics described above.I use the number of students taking the tests in each school as weights in the analyses.This allows me to determine how much of the sectoral differences in student outcomes are explained by observable teacher and school management characteristics.However, it restricts the variables to have the same effect across schools.
At the outset, it is worth noting that the student outcome and socioeconomic background measures are school-level averages, while the teacher and school management variables are individual-level data (of only a few teachers in each school) from which I constructed school-level averages.As a result, the student outcome and socioeconomic background measures do not contain within-school variation but the teacher and school management variables do.Moreover, as explained in the Data Appendix, the teacher and school management averages are noisy measures of a school's true teacher-related and other policies.This noise means that the measured within-sector variation in teacher and management variables will exaggerate the true within-sector variation, especially relative to student outcomes and socioeconomic background.The noise will also mean that the estimated effects of teacher and management variables will be attenuated versions of the true effects.That is, teacher and school management policies are measured with error, which generates attenuation bias.If I find any effects of teacher-related and other policies, it is despite attenuation bias.

The self-selection problem and school choice in Chile
To detect the relationships among sector, teacher, and school characteristics and student outcomes, it is important to reduce (ideally, eliminate) the effect of student selfselection.There are three potential sources of student self-selection that may affect my estimates: 1. Students (and their parents) may choose a specific school based on its resources.In this case, the observed effects of sector, teacher and school characteristics may simply reflect the effect of greater resources.This is a serious problem in countries such as the United States, where there are substantial differences in resources per student among public and private schools.In Chile, however, municipal and private voucher schools receive the same amount of resources per pupil.Consequently, the selection problem due to differences in resources is mainly a problem of identifying the effect of private paid schools and not so much of identifying the effects of private non-religious and Catholic voucher schools.Because, in terms of policymaking, our interest is in detecting the effects of differences in teachers and school policies among schools with similar resources, I am less interested in identifying the effect of private paid schools than in identifying the effect of voucher schools.Thus, not being able to address this type of selection with my data does not pose a major problem to my research.
2. Students (and their parents) choose specific schools based on arbitrary differences, such as geographic locatio n.I expect that controlling for student socioeconomic background will eliminate most of the potential bias due to this issue.
3. Students (and their parents) choose specific schools based on unobserved differences, such as their own motivation, which are very difficult or impossible to measure.This is a problem if, for instance, motivated parents systematically tend to choose schools in one sector (for example, Catholic voucher schools).In the United States, this issue of sample selection bias-resulting from the reality that students are not randomly assigned into schools in different sectors-has been a subject of much controversy among researchers.Ideally, if there were systematic selection into sectors by unobserved variables other than cost or geography (which I am controlling for by using measures of student socioeconomic background), I would need to identify such variables.In a recent study of the effectiveness of Catholic schools in the United States, Altonji, Elder and Taber (2000), however, contend that selection on the observables is likely to be stronger than selection on the unobservables.Consequently, results that indicate positive effects of Catholic schools should be interpreted as a lower bound estimate of the effect of Catholic schools.More importantly, if, for example, motivated parents choose a particular sector (e.g., Catholic voucher schools) because they believe they provide the best teachers, or have better school management policies, then this is not much of a problem for my research.In fact, if schools did not vary in their teachers and school management policies, then parents would not be selecting among different schools and sectors.
Besides student self-selection, there is also teacher self-selection into different types of schools.Indeed, this is precisely what I am trying to investigate.That is, in addressing my research questions, I am determining, via data analysis, the nature of the assignment of teachers to schools.There are at least three types of teacher self-selection: 1. Teachers may select schools based on different per-pupil resources.Again, in Chile, this is mainly a problem for private paid schools, as municipal and voucher schools receive the same per pupil resources.
2. Schools have different policies that may affect teachers' observable characteristics.I include measures of teachers' observable characteristics in my analyses in order to control for this type of selection.
3. Schools may do different things that affect teachers' unobservable characteristics.
My data cont ain valuable information regarding school management policies that may affect teachers' unobservable characteristics.
Because of the unobserved differences among students and family background, it is possible that my research does not establish the causal effects of the variables of interest on student outcomes.However, I attempt to control for student background to the full extent possible, and we can be confident that there is no systematic selection related to cost among the public, religious voucher, and non-religious voucher sectors.That is, observable or unobservable variables that would affect how a family would react to a school's cost cannot be generating different student outcomes between the public, religious voucher, and non-religious voucher sectors.As a result, assuming my controls for student background are effective, the remaining selection problem becomes rather small: unobserved background variables that affect school selection in some way that is unrelated to cost.Therefore, I am able to come close to identifying the causal effects of sector, teacher characteristics, and school management indicators.
Variation in student outcomes, student background, teacher characteristics and school management I investigate the variation in student outcomes, student background, teacher characteristics and school management in two ways.First, I plot the data to explore the distributional variation of the variables of interest among the four institutional sectors.Second, I decompose the variance in the variables of interest with and without controlling for student socioeconomic background.In this variance decomposition, I am interested in identifying the proportion of the total variance in a variable that comes from differences among sectors, differences among the schools within a sector, and differences among the teachers within a school.The Data Appendix includes a detailed description of this variance decomposition.

Relationship between student outcomes, sectors, student and teacher characteristics, and school management
To explore the extent to which differences in student outcomes among schools can be explained by differences in the students they serve, in the characteristics of the teachers they employ, and by differences in their management strategies, I conduct weighted least squares (WLS) regressions of average student test scores on sector, teacher, and school characteristics, controlling for average student socioeconomic background.The weights used are the number of students taking each test by school.I use these weights to account for the fact that the aggregated values of the outcome (average student test scores by school) and the predictors were based on different sample sizes and, thus, the residual variances would likely differ.
My ana lyses are similar to those used in investigating standard educational production functions, but they incorporate rich information on teacher and school characteristics not often available to researchers.Let T j be the average student test scores in school j, X j be a vector of average teacher characteristics in school j, S j be a vector of average school management characteristics in school j, V j be a vector of average student socioeconomic background in school j, and D j k be a set of dummy variables indicating the sector k (public [municipal or corporation], private paid, private voucher [or shared financing] and Catholic voucher [or shared financing]) to which school j belongs.Then, we can express the relationship among student test scores, teacher and school characteristics as: In this model, the parameters to be estimated are α, β, δ, and φ.The term ε j represents the unobserved variance, or error, in student outcomes by school.In fitting this model, I am particularly interested in estimating the parameter vector α, the effect of sector on student outcomes, and in how it changes upon inclusion of the rest of the parameters.For example, if before including the vectors of teacher characteristics and school management indicators (X j and S j ) the estimated coefficients in α were large and, upon inclusion of X j and S j , they were reduced, then I would conclude that much of the sector variation in student outcomes can be explained by differences in teacher characteristics and school management strategies among sectors.

Data, Sample and Measures
Two types of data are used: (1) school-level average data on student outcomes and student socio-economic background from a national-level educational assessment program administered by the Chilean Ministry of Education and (2) teacher-level data on teacher characteristics and school governance structures from a teacher survey conducted by local researchers in the metropolitan area of Santiago.
The data come from two sources.Student outcome data consist of 1999 average fourth-grade student test scores in mathematics, language, and reading by school from the Chilean Ministry of Education's Sistema de Medición de la Calidad Educativa (SIMCE).These data are publicly available at the Ministry of Education's website.Data on student socioeconomic background at the school level also come from this source, though they were originally collected by a separate government agency responsible for developing education and health programs targeted to disadvantaged children.These data are aggregated to the school level.
Information on teacher characteristics and teacher reports' of school management policies come from a teacher survey conducted in the 1998-99 school year by Alejandra Mizala, Pablo González, and Pilar Romaguera from the Center for Applied Economics of the Department of Industrial Engineering of the Universidad de Chile, under the supervision and financing of The Inter-American Development Bank.The teacher survey was conducted in the metropolitan area of Santiago, and therefore my study centers on a sample of schools in this area.For a detailed description of this survey and preliminary analyses, see Mizala, González, and Romaguera (1999).These data are at the teacher-level.My final sample consists of 901 teachers and 171 schools in the metropolitan area of Santiago.
Five types of measures are used.The outcome measure is student test scores aggregated to the school level.The principal question predictor is the sector to which a school belongs (municipal, private paid, private voucher, and Catholic voucher).In addition, I also include as question predictors several measures of teacher characteristics and school management from the teacher surveys.Because several teachers in each school were asked the same questions regarding their own characteristics, each teacher's individual response is not a fully representative measure of the average teacher characteristics in a school.Thus, for each school, I aggregate (by taking the average of) teachers' reported characteristics to the school level in order to use these data in the regression analyses.As a result, the indicators of teacher characteristics likely represent the mean teacher's characteristics with some error.
Similarly, because several teachers in each school were asked the same questions regarding how their schools are managed, each teacher's individual response is not an accurate measure of school management strategies.For each of the schools, I also calculate the school-level average of teachers' responses.As a result, my indicators of school management also likely represent the mean teacher's views with some error.
As a control predictor, I include student socioeconomic background.This information is originally aggregated to the school level.The measures are described in more detail in the Data Appendix.Appendix Tables A1 and A2 present descriptive statistics on the variables used in my analyses.

Findings
The focus of most previous research on school choice has been on differences between sectors in student outcomes and explanatory variables.I find that differences within sectors in student outcomes and student background, teacher characteristics and school management are often greater than the between-sector differences.My findings indicate that some teacher and school characteristics do affect student performance, but that a great deal of unexplained variance among sectors remains important in predicting student outcomes.Teacher education, decentralization of decisionmaking authority, whether the school schedule is strictly enforced and the extent to which teachers have autonomy in designing teaching plans and implementing projects all are predicted to affect student outcomes.However, there is and interaction between teacher autonomy and decentralization of decision-making authority in the effect of these variables on student outcomes.Schools where teacher autonomy is greater tend to have higher student outcomes only when decision-making authority is also decentralized.This finding suggests that decentralization of decision-making authority allows for greater supervision and support of teachers, which, in turn, allows teachers to make better use of autonomy in their classrooms.

Variation in student background, student outcomes, teacher characteristics and school management
Not unexpectedly, municipal schools serve students of lower socioeconomic background, on average, than do private voucher and Catholic voucher schools.However, there is substantial variation in student background among the schools within a sector. Figure 2 presents the distribution of student socioeconomic background (as measured by the vulnerability index described in the Data Appendix) by sector.Although there are differences in the average student socioeconomic background by sector, there are even larger differences among the schools within a sector (see Table 1).For instance, while about 85 percent of the variance in student socioeconomic background comes from differences among schools within a sector, only about 11 percent is explained by between-sector differences.As has been found elsewhere (Rodriguez 1988;Aedo and Larrañaga 1994;Aedo 1997; Mizala and Romaguera 2000;McEwan andCarnoy 1999, 2000;Carnoy and McEwan 2001;McEwan 2001;Hsieh and Urquiola 2001), without accounting for student background there exist substantial differences in average student math scores by sector (see Figure 3).Private paid schools have much higher average student test scores than do schools in other sectors.Private and Catholic voucher schools have higher average student math test scores than do municipal schools.The proportion of total variance in student test scores that can be accounted for by between-sector differences ranges from 35 to 41 percent (see Table 2a).Importantly, there is also great variation in average student test scores among the schools within a sector.Differences among schools within a sector account for about 60 to 65 percent of the total variance in school-level average test scores (see Table 2a).
Controlling for student socioeconomic background reduces much of this betweensector variation.As Table 2b shows, when controls for student socioeconomic background are included, the proportion of variance that is explained by between-sector differences falls to between 17 and 21 percent.Figures 4-7 present the distribution by sector of the four teacher measures used in the analyses-the percentage of teachers with university education by school, teachers' average years of experience, teachers' self-reported high school grades, and teachers' average monthly salaries.In general, the figures show that the schools within each sector tend to choose teachers with similar characteristics.The figures also indicate that there are some important differences among sectors in most average teacher characteristics.For example, Figure 5 shows that teachers in private paid schools and Catholic voucher schools tend to report higher average high school grades than do teachers in municipal and private voucher schools.Similarly, Figure 6 shows that private and Catholic voucher schools tend to have a higher proportion of their teachers who have less than two years of teaching experience than do municipal and private paid schools.This is probably a result of voucher schools' being a relatively new sector in Chile.There also appear to be important differences by sector in average teacher salaries, with private paid and municipal school teachers earning higher average salaries than do their colleagues in private and Catholic voucher schools, as shown in Figure 7.  3a and 3b report the variance decomposition in teacher characteristics.Without controlling for student socioeconomic background, the between-teachers-within school variance in teachers' years of experience, average high school grades, and mean monthly salary accounts for 60, 75, and 70 percent, respectively, of the total variance.However, the between-schools-within-sector variance is not insignificant, accounting for 23, 28, and 30 percent of the total variance in teachers' average high school grades, years of experience and mean monthly salary, respectively.2,3Thus, just as there are important differences within schools, there are also large differences in teacher quality among the schools within a sector.For instance, the majority of the variance (about 80 percent) in teacher education comes from differences among schools within a sector (see Figure 4).In contrast, the proportion of variability in teacher characteristics that comes from differences between sectors is relatively small.Only about 12 percent of the estimated variance in teachers' years of experience comes from between-sectors.As seen in Figure 6, teachers in the municipal sector tend to have more years of experience than their colleagues in private fee-paying and voucher schools.The proportion of variation in high school grades and average monthly salary that comes from differences among sectors is even smaller, only around 2 percent.university education.
As explained in the Introduction, schools can adopt different management strategies that may affect teachers' unobservable characteristics and student outcomes.Schools can, for example, adopt very centralized or decentralized structures of decisionmaking authority, they can be more strict or lax regarding teacher absenteeism and timely attendance, they can enable teachers to have more or less autonomy over their jobs, and they can contribute to foster varying degrees of teacher job and career satisfaction.
Figures 8 and 9 present the distribution by sector in teachers' average responses by school of the degree of decentralization of decisionmaking authority and the extent to which their school schedule is strictly enforced.The figures suggest that there is variation by sector and among the schools within a sector in these two school management strategies.4a and 4b present a decomposition of variance of these school characteristics, with and without controlling for student socioeconomic background.Table 4a shows that teachers within a school have varying views regarding how their school is managed.This is indicated by the high proportion of total variance in school characteristics that is explained by differences among teachers within a school.In fact, without controlling for student socioeconomic background, the majority of the variance in the school management variables comes from between-teachers-within-school differences (between 66 and 70 percent of total variance).
In Table 4b, I present the same analysis controlling for student socioeconomic background.This reduces even more the proportion of variance explained by differences among schools within a sector and among sectors.In fact, after controlling for student socioeconomic background, the proportion of total variance in school management measures that is explained by differences among the teachers within a school ranges from 80 to 87 percent.As mentioned above, much of this variability among the teachers within a school is noise, and aggregating teachers' responses to the school level generates better measures of school management structures, even if they do contain error.

Relationship between student outcomes, sectors, student and teacher characteristics, and school management
The analysis of the relationship between student test scores, student and teacher characteristics and school management yields very similar results for all three subject tests -mathematics, reading, and language.Thus, to simplify the presentation, in Table 5 I report only the results from WLS regressions of 1999 school average test scores in mathematics.5  b.An F-test of the null hypothesis that the coefficients on (e), (f), and (g) are simultaneously zero in the population could not be rejected [F(2, 159) = 0.66, Prob>F = 0.518].c.In this model, I also tested whether the effects of average salary on student outcomes vary by sector.An F-test of the null hypothesis that the two-way interactions between the sector dummy variables and mean teacher salary are jointly zero was not rejected.In addition, I conducted an F-test of the null hypothesis that the two-way interactions between each sector dummy variable and mean teacher salary are equal; I was unable to reject the null hypothesis.
d.An F-test of the null hypothesis that the coefficients on two-way interactions between (j) and the sector dummies are simultaneously zero in the population could not be rejected [F(3, 158) = 0.77, Prob>F = 0.511].e.An F-test of the null hypothesis that the coefficients on the main effects and the two-way interaction of (j) and (o) are simultaneously zero in the population is rejected at the 5 percent level [F(3, 158) = 2.85; Prob>F = 0.039).
Student outcomes and sector.In columns ( 1) and (2) of Table 5, I explore differences in student outcomes by sector, first without controlling for student socioeconomic background and then controlling for it by including the vulnerability index described in the previous section.Without controlling for student socioeconomic background, private voucher/shared financing, private paid, and Catholic voucher/shared financing schools have higher values of student outcomes than do the municipal schools.The magnitude of this effect is largest for private paid schools, with an advantage of about 2 standard deviations, followed by Catholic voucher/shared financing schools (about 0.8 standard deviations higher) and then by private voucher/shared financing schools (about one-half of a standard deviation higher).
Controlling for socioeconomic background reduces the estimated differences in achievement among public and all other schools substantially.In fact, after controlling for student socioeconomic background, the advantage of the private voucher/shared financing schools disappears.For all other sectors, net of student socioeconomic background, the estimated effects are much smaller than the effects without controlling for student background.For example, the private paid school advantage is now about half of the effect without controlling for student background-slightly more than one standard deviation.Although there continues to be an advantage to Catholic voucher/shared financing schools after controlling for student socioeconomic background, this effect is almost one-fourth smaller than without controlling for student background.6 Student outcomes and teacher characteristics.I also assess the extent to which differences in student achievement can be explained by differences in often-researched measurable teacher characteristics such as teacher's educational attainment, years of experience, high school grades, and mean salary.As Hanushek (1986) and others have found, the results presented in columns (3)-( 6) suggest that, with the only exception of teacher education, these measurable teacher characteristics appear to contribute little to student achievement.
My findings indicate that percent of teachers in a school with university education is positively related to student outcomes.The estimated coefficient on the percent of teachers with university education is positive and statistically significant at the 5 percent level, or better.The estimate suggests that for a one percent difference in the percent of teachers with university education in a school, the average test scores of 4 th grade students are higher by more than 1 standard deviation.
Student outcomes and school management strategies.Do other characteristics related to how teachers work in schools, and to how much the school environment contributes to teacher effectiveness not often available to researchers affect student outcomes?As explained earlier, I explore the marginal effects on average student achievement of several school management indicators, including: the degree of decentralization of decisionmaking authority, strictness of enforcement of the school schedule, teacher absenteeism, a series of measures of teacher autonomy, and indicators of teacher career and job satisfaction.
In schools where the main decisionmaker is closer to the teacher, students tend to have higher test scores (see column (7)).For example, schools with a one-point difference in decentralization of decisionmaking authority (e.g., where the main decisionmaker is a school intermediary instead of the school principal), are estimated to have average math test scores that are almost 0.3 standard deviations higher.
The results presented in column ( 8) indicate that, in schools where teachers perceive the schedule to be strictly enforced, student outcomes tend to be higher.In particular, in schools where teachers report that the schedule is strictly enforced, average math test scores are estimated to be 0.3 standard deviations higher than in schools where teachers report that the schedule is flexible.
In columns ( 9), ( 10), and ( 11), I explore the effects on student outcomes of teacher absenteeism, teacher autonomy over school-level decisions and teacher autonomy over classroom-level decisions.The coefficient estimates are not statistically significant.
The results presented in columns ( 12) and ( 13) indicate that the effect of teacher autonomy in defining teaching plans and implementing projects varies depending on the degree of decentralization of decisionmaking authority.Specifically, the greater the degree of decentralization of decisionmaking authority, the greater the effect of teacher autonomy in defining teaching plans and implementing projects on student test scores.This relationship is depicted in Figure 10 for municipal schools.In the figure, the three fitted lines represent different levels of teacher autonomy in defining teaching plans and implementing projects: low autonomy, average autonomy, and high autonomy.At low levels of teacher autonomy, increased decentralization of decisionmaking authority has no effect (or even a slightly negative effect) on estimated math test scores.In contrast, the lines for average and high levels of teacher autonomy indicate that relatively high levels of autonomy with low degrees of decentralization of decisionmaking authority are associated with low estimated math scores.This result suggests that when decisionmaking authority is too centralized, this may lead to an inability to effectively supervise teachers.The positive slopes of the lines of average and high autonomy suggest that teacher autonomy over teaching plans and project implementation can improve student test scores when there is effective supervision.In other words, in schools where the main decisionmaker is closer to the teacher and where teachers can exert autonomy over planning and project implementation, student outcomes tend to be higher.a.To construct these prototypical fitted lines, I used the estimated coefficients presented in column (12) of Table 5, and substituted average values for all variables for public schools except for decisionmaking authority and teacher autonomy.The range of values of decentralization of decisionmaking authority is the sample -specific range for municipal schools.
Finally in columns ( 14) and ( 15), I explore whether schools with higher levels of teacher career and job satisfaction have higher estimated student outcomes, as Perie and Baker (1997) have suggested using U.S. data.My findings do not support this hypothesis.

Discussion
There is a great deal of variation in student outcomes among sectors in Chile.For example, even after controlling for student socioeconomic background, Catholic voucher schools outperform municipal schools (and non-religious voucher schools) by about a third of a standard deviation, which most researchers would agree is not insubstantial (Mosteller 1995).In this paper, I have investigated the extent to which this variation in student outcomes among schools of different sectors can be explained by differences in the quality of the schools' teacher forces and by variation in the way that schools are managed.7 The first part of my analysis indicates that there is not a great deal of consistency among the schools within a sector in the teacher characteristics they hire and in the school management policies they put in place.In most variables, there is greater variation among the schools within a sector than among sectors.
In the second part of my analysis, I explored the extent to which differences in teacher characteristics and in school management policies affect student outcomes in Chile.In particular, my goal in this section was to assess the extent to which the inclusion of indicators of teacher and school characteristics contributes to explain the observed sector effects.In the extreme, if observable teacher and school characteristics were to fully explain the sector effects, then the estimated coefficients on the sector indicator (dummy) variables would be reduced to zero.
My findings indicate that some teacher and school characteristics do affect student performance, but that a great deal of unexplained variance among sectors remains important in predicting student outcomes.Teacher education, decentralization of decisio nmaking authority, whether the school schedule is strictly enforced and the extent to which teachers have autonomy in designing teaching plans and implementing projects all appear to affect student outcomes.Importantly, I found that teacher autonomy has positive effects on student outcomes only when there is also decentralization of decisionmaking authority.From a school management perspective, this implies that to improve educational outcomes, it is not enough to give teachers room for decisionmaking in the classroom but also to support them and guide them through effective supervision.
Moreover, my findings suggest that the way schools are managed-as measured by the variables mentioned above-is more strongly related to student outcomes than are observable teacher characteristics such as education, experience, and teachers' high school grades.This is an important contribution to the literature on education production.Future research should further our understanding of the characteristics of school manage ment that are related to student learning.
Finally, my results also suggest that, besides decentralization of decisionmaking authority, enforcing the school schedule, and providing teachers with autonomy, there are other, unobserved ways in which Catholic voucher schools operate differently from municipal schools that result in their having higher average student test scores.Further research is needed to identify these variables.

Outcome measure: average student test scores by school
Student test score data consist of school-level averages on the SIMCE 1999 tests of mathematics, language, and reading, which were administered to 285,094 fourth-grade students in 5,467 schools throughout the country.These scores are scaled using Item-Response Theory (IRT).In 1999, the Chilean government adopted a point scale with an arbitrary mean set at 250 points and a standard deviation of 50 points (Mineduc 2000). 8ll analyses are conducted using each test score as an outcome variable and the results are presented separately.For ease of interpretation, and so that my results may be readily compared to those of previous researchers, I follow McEwan and Carnoy (2000) in standardizing all test score variables to a mean of zero and a standard deviation of one prior to conducting my analyses.

Predictors
Sector.Mizala and Romaguera (2000) distinguish among municipal, private paid, and private subsidized (or voucher) schools.In addition to these categories, McEwan and Carnoy (2000) distinguish public corporations from municipal schools, and Catholic and Protestant voucher schools from non-religious voucher schools.Given the small number of Protestant and public corporation schools in my sample, I do not maintain these as separate categories.I distinguish among the following four sectors: municipal (including public corporation), private (non-religious) voucher (including shared financing schools, a relatively new option which allows private schools to charge a small fee in exchange for a reduced voucher), Catholic voucher (also including shared financing schools), and private paid.I include indicator (dummy) variables for each of the sectors, with the municipal sector as the omitted category.
Teacher characteristics.Measures of teacher characteristics used in my research include: educational attainment, years of teaching experience, teachers' self-reported high school grade averages, and teachers' reports of average monthly salary.Research on the effect of most observable teacher characteristics on student achievement has yielded inconsistent results (Hanushek 1986(Hanushek , 1997)).Nevertheless, it is worth exploring whether any of these characteristics, which we intuitively associate with student learning, are in fact related to student outcomes in the Chilean case.
My measure of educational attainment is an indicator (dummy) variable of whether a teacher has university education and beyond (1) or less than university education (0).Using this measure, I construct a school-level variable that indicates the percent of teachers with university education and beyond in each school.This school-level variable ranges from 0.30 to 1.
Teachers' self-reported average high school grades may provide information about teachers' cognitive skills, a variable that has been related to student test score gains in studies using data from the United States and from developing countries (Harbison and Hanushek 1992).This measure consists of a set of indicator (dummy) variables representing: (1) teacher's report of average high school grades is less than 60, and 0 otherwise; (2) teacher's report of average high school grades is 60-69, and 0 otherwise; (3) teacher's report of average high school grades is 70-84, and 0 otherwise; and (4) teacher's report of average high school grades is 85 and higher, and 0 otherwise.Using these teacher-level data, I constructed school-level aggregates representing the percentage of teachers within a school in each category.
Because the information on teacher's average high school grades comes from selfreports, it contains at least two types of measurement error.First, teachers attended schools that may have graded differently.Second, teachers may differ in their capacity to accurately recall their average high school grades.Thus, the results using this measure must be interpreted with caution.
The fourth variable, teachers' reports of average monthly salary, can be considered an alternative measure of educational attainment and experience given that in Chile, as in most countries, teacher salaries are tightly linked to education and experience.The original, teacherlevel salary measure is a continuous variable.I used this teacher-level measure to construct school-level averages of teachers' mean monthly salaries.School management measures.I explore the effects on student outcomes of four broad types of school management measures derived from teachers' reports in the survey: (1) degree of decentralization of decisionmaking authority within a school; (2) the extent to which the school enforces its daily schedule and the teacher attendance; (3) the extent to which teachers have autonomy over teaching methods, classroom-and school-level decision making; and (4) teachers' degree of job and career satisfaction.
(1) Degree of decentralization of decisionmaking authority.Recent research has focused on the relationship between decentralization of decisionmaking authority and student outcomes.King and Özler (1998) analyzed the case of Nicaragua's school autonomy reform and found that schools that exert greater autonomy regarding teacher staffing and the monitoring and evaluation of teachers appear to have higher student performance.Pães de Barros and Mendonça (1998) analyzed the effect of three institutional innovations-financial autonomy, selection of the principal, and the establishment of local school councils-on several indicators of school quality in Brazil.The authors detected an impact on school quality of each of the three innovations, but the size of the effects was reduced when teacher and household characteristics were controlled.Jimenez and Sawada (1999) analyzed the effect of community-managed schools on student and teacher absenteeism, as well as on student math and language achievement in El Salvador.They found that school decentralization was associated with lower teacher and student absenteeism.In addition, their study reports a small effect of decentralization of school management on students' scores in language tests.Navarro and de la Cruz (1998) analyzed the effect of concentration of decisionmaking on the internal efficiency of eighteen national, state, and Catholic schools in Mérida, a Venezuelan state.They found that when decisionmaking authority is concentrated at the school level (as opposed to the state or national level), schools appear to select better inputs, provide better discipline to teachers, and have more motivated teachers.
In the teacher survey conducted in Chile, teachers were asked to identify the most important decisionmaker for their school.Using teachers' responses to this question, I constructed a categorical variable that represents the degree of decentralization of decisionmaking authority within the school, with the following levels: (1) state/national government authority; (2) municipal government authority; (3) school principal; (4) school intermediary (e.g., area coordinator, subject area leader); and (5) teachers.I used this teacherlevel variable to construct a school-level variable representing the average of all the teachers' reports of decentralization of decisionmaking authority within a school.
(2) Enforcement of school schedule and teacher attendance.In most Latin American countries, it is often argued that schools are not able to enforce teacher attendance, much less their timely attendance.It is intuitive that schools where teachers are consistently present and arrive on time are more likely to have higher student achievement than are schools where teachers are frequently absent and/or late.One question in the survey asked teachers to report the extent to which they agree that teachers in their school are frequently absent.The categories are: (1) strongly disagree; (2) disagree; (3) neither agree nor disagree; (4) agree; and (5) strongly agree.I used this teacher-level variable to construct a school-leve l measure of teacher absenteeism by averaging teachers' responses to this question by school.
Teachers were also asked to define their school schedule as either strict or flexible.The original variable has values of 0 if the teacher reports that the school schedule is flexible, and 1 if s/he reports it is strict.I use this teacher-level variable to construct a school-level indicator of teachers' average reports to this question by school.
(3) Teacher autonomy.Hoxby (2000) found that school choice in the United States would change the teaching profession by, among others, demanding teachers with a greater degree of effort and independence.I explore whether the school choice reform in Chile has yielded similar results.Using a 5-point Likert-type scale, teachers in the survey were asked the extent to which they agree or disagree with the following statements: • I have been assigned the responsibility of coordinating programs or activities • I am able to make decisions in my job • I am able to plan my own activities.
I used principal components analysis to composite teachers' responses to these questions.I detected three dimensions of teachers' degree of autonomy in their jobs.As a result, I created three composites to summarize this information.The first composite measures the extent to which teachers have autonomy over defining their teaching plans and implementing projects in the school.The second composite represents the extent to which teachers have autonomy regarding decisions that affect the entire school.The third composite measures the extent to which teachers can exert autonomy regarding what goes on inside their own classrooms.9(4) Teacher career and job satisfaction.Labor economists have argued that more satisfied workers are less likely to leave their jobs and more likely to invest in improving their skills for the job (Hamermesh 1999).In education, researchers have also argued that satisfied and committed teachers are more likely to stay in the profession and to upgrade their skills (Perie andBaker 1997 andFirestone 1990).My data contain rich indicators of teacher career and job satisfaction.For example, teachers were asked: • If you could elect another career today, would you choose teaching again?
• Would you be happy if any of your children became a teacher?
• Have you ever considered leaving the teaching profession?
I used principal components analysis to construct a composite representation of teacher career satisfaction using teachers' responses to these questions.Similarly, I also used principal components analysis to develop a composite of teacher job satisfaction, using the extent to which teachers agree with the following statements: • My work environment is very professional • I have earned the respect of my colleagues • I have the capacity to ensure that my work is done properly • I am treated as a professional in my job • I have an impact on my students • I have the opportunity to grow on a daily basis as a result of wo rking with my students • I am able to do a good job • I can participate in important activities for the kids I used these teacher-level composites to construct average values by school of career and job satisfaction among the teachers in each school.10Student background.To identify the effects of the variables of interest on student outcomes, I must control for differences in students' socioeconomic background that may also affect their outcomes.As Coleman and others (1966) first showed, family background cha racteristics of students have important effects on student outcomes.My principal measure of average student socioeconomic background by school consists of an "index of school vulnerability" developed by the Junta Nacional de Auxilio Escolar y Becas (JUNAEB), a national government organization responsible for developing education and health programs targeted to disadvantaged children.This vulnerability index was constructed from student surveys conducted in 1998.It describes the extent to which poverty measures, such as low weight, height and excessive medical needs, are found among students within a school (Díaz Véliz 1998).The measures include students' weight, height and medical needs as well as mothers' education (Mizala and Romaguera 2000).The index ranges from 0 to 100.I replace values of the continuous index by a set of indicator (dummy) variables that include: (1) minimum vulnerability (0-32.4);(2) low vulnerability (32.5-44.7);(3) medium vulnerability (44.8-56.2);(4) high vulnerability (56.3-66.7);and (5) very high vulnerability (66.8 and higher). 11  To ensure that my principal measure of student socioeconomic background soaks up most of the variability due to differences among schools in student background, I first regressed the student outcome data (1999 test scores) on the principal measure of student socioeconomic background (the 1998 vulnerability index) and obtained the residuals.Then, I regressed these 11 I use a second measure of student socioeconomic background to explore the extent to which the first measure is effective in accounting for the variability in average student socioeconomic background.This second measure comes from 1994 SIMCE exam data.It consists of 4 categories, defined as follows (Mizala and Romaguera 2000): Socioeconomic level A: schools in which most parents have completed secondary education, or have some higher education (complete or incomplete), and whose educational expenses are greater than 25,052 Chilean pesos (about US$50).
Socioeconomic level B: schools in which most parents have incomplete or complete higher education or less and whose educational expenses are between 13,210 and 25,051 Chilean pesos (between around US$26 and US$50).
Socioeconomic level C: schools where parents have incomplete secondary education or less and whose educational expenses are between 5,284 and 13, 209 Chilean pesos (between around $11 and US$26).Socioeconomic level D: schools where most parents have incomplete primary education or less, and whose educational expenses are less than 5,283 Chilean pesos (about US$11).residuals on the alternative measure of student background from the 1994 SIMCE data.If the 1994 student background measure were to contribute useful information, the estimated coefficients on this regression would be statistically significant.They were not, leading me to conclude that my principal measure of student socioeconomic background is effective.Teachers were asked to report the extent to which they agree with the following statement: "In my opinion, the teachers in this school are frequently absent."The categories and their values are: (1) strongly disagree, (2) disagree, (3) don't agree or disagree, (4) agree, (5) strongly agree.

Figure
Figure 1.Distribution of enrollment by sector, 1981 and 1999

Figure 2 .
Figure 2. Distribution of student socioeconomic background, by sector

Figure 4 .
Figure 4. Percent of teachers with university education, by sector

Figure 5 .1Figure 7 .
Figure 5. Average high school grades for teachers, by sector

Figure 8 .
Figure 8. School average reports of decentralization of decisionmaking authority, by sector

Figure 10 .
Figure 10.Average math test scores, decentralization of decisionmaking authority, and teacher autonomy in defining teaching plans and implementing projects, municipal schools a