Expanding school enrollment by subsidizing private schools: Lessons from Bogotá

Comparative Education Review

Volumn 50, Issue 2, 2006, Pages 241-277

  Uribe, Claudia ^a   Murnane, Richard J ^a   Willett, John B ^a   Somers, Marie Andrée ^a  

^a NONE

Author keywords

[No Author keywords available]

Indexed keywords

EID: 33746018300     PISSN: 00104086     EISSN: None     Source Type: Journal    
DOI: 10.1086/500695     Document Type: Review

References (91)

1. Edward B. Fiske and Helen F. Ladd, When Schools Compete: A Cautionary Tale (Washington, DC: Brookings Institution, 2000);
2. Chang-Tai Hsieh and Miguel Urquiola, "When Schools Compete, How Do They Compete? An Assessment of Chile's Nationwide School Voucher Program," NBER Working Paper no. 10008 (National Bureau of Economic Research, Cambridge, MA, 2003).
3. Estelle James, "Public Subsidies for Private and Public Education: The Dutch Case." In Private Education: Studies in Choice and Public Policy, ed. D. C. Levy, 113-37 (Oxford: Oxford University Press, 1986).
4. Indeed, Estelle James ("Why Do Different Countries Choose a Different Public-Private Mix of Educational Services?" Journal of Human Resources 28 [Summer 1993]: 571-92) finds that countries with the largest percentage of private school enrollments are those with excess demand for schooling.
5. There is some evidence that private schools are more cost-effective than private schools. For example, using data from Indonesia, Estelle James, Elizabeth M. King, and Ace Suryadi ("Finance, Management, and Costs of Public and Private Schools in Indonesia," Economics of Education Review 15 [October 1996]: 387-98) find that the cost per student of achieving a given level of academic performance is lower in private schools than in public schools.
6. Patrick McEwan and Martin Carnoy, "The Effectiveness and Efficiency of Private Schools in Chile's Voucher System," Educational Evaluation and Policy Analysis 22 (Autumn 2000): 213-39.
7. Secretaría de Educación del Distrito, Informe de Gestión: 1998-2000 (Bogota: Secretaría de Educación del Distrito, 2000).
8. The SED established a minimum of 35 students per class in grades 1-5 and of 40 in grades 6-11.
9. In comparison, only 21 percent of all pre-university-level students in the country were enrolled in private schools in the same year; from Secretaría de Educación del Distrito, Boletin Estadístieo: Versión Borrador (Bogotá: Secretaría de Educación del Distrito, 2001).
10. The yearly tuition fee for private schools in Bogota in the 2002 school year ranged from $30 to $2,400, with an average of $285; Secretaría de Educación del Distrito, Estadísticas Educativas, 1998-2003 (Bogota: Secretaría de Educación Distrital, 2003), This is considerably less than the $439 average expenditure per student in public schools. However, these figures are not completely comparable, because the public-sector expenditure figure includes the costs of school supplies, transportation, and nutritional supplement, while the private school tuitions do not. Also, many private schools impose additional fees for school activities and expect parents to make "voluntary" donations (bonos).
11. Bogotá has a six-level system of socioeconomic stratification, with six as the highest and one as the lowest. Roughly, estratos 1 and 2 correspond to families earning up to two times the minimum wage (the 2003 minimum wage is approximately US$100/month). Estralos-4 and 5 families earn approximately five to ten times the minimum wage.
12. Based on data from A. Sarmiento, B. L. Caro, J. I., González, E. Castaño, and J. Espinosa, Evaluación de la calidad de la educación primaria en Santa Fé de Bogota 1998: Factares asociados al logro; Informe final versión definitiva (Bogotá: Departamento Nacional de Planeación. Secretaría de Educación de Bogotá, Corporación Mixta para el Desarrollo de la Educación Básica, 1999).
13. New regulations that came into effect in 2002 introduce performance accountability mechanisms for promotions and reduce the weight of academic credentials and experience in determining promotions. These changes are likely to bring about changes in the city's teaching force over the coming years.
14. Sarmiento et al., Evaluación de la calidad
15. Alejandro Gaviria and Jorge Hugo Barrientos, "Determinantes de la calidad de la educación en Colombia," Planeación y Desarrollo 32 (Autumn 2001) : 339-86. A 1995 constitutional court ruling (related to article 198 of the 1991 constitution) required private schools to pay teachers according to the public teacher salary scale, though private schools still have greater flexibility than public schools in hiring teachers with modest credentials and in dismissing teachers at will.
16. For an older review of research in developing countries, see Bruce Fuller, "What School Factors Raise Achievement in the Third World?" Review of Educational Reseanh 57 (Autumn 1987): 255-92.
17. The only class size experiment is Project STAR in Tennessee; see A. B. Krueger, "Experimental Estimates of Education Production Functions," Quarterly Journal of Economics 114 (May 1999): 497-532. In terms of peer effects, the Moving to Opportunity project randomly awards housing vouchers to residents of low-income neighborhoods in various cities in the United States.
18. For results from Boston, see Lawrence F. Katz, Jeffrey R. Kling, and Jeffrey B. Liebman, "Moving to Opportunity in Boston: Early Results from a Randomized Mobility Experiment," Quarterly Journal of Economics 116 (May 2001): 607-54. The one random assignment experiment on teacher quality that we are aware of finds that Teach for America participants are somewhat more effective in teaching math (but not reading) to low-income children than the other teachers serving in these schools.
19. See Paul Decker, Daniel Mayer, and Steven Glazerman, The Effects of Teach for America on Students: Evidence from a National Evaluation (Princeton, NJ: Mathematica Policy Research, 2004).
20. Miguel Urquiola, "Identifying Class Size Effects in Developing Countries: Evidence from Rural Schools in Bolivia," Review of Economics and Statistics 88 (February 2006): 171-77.
21. Miguel Urquiola, "Identifying Class Size Effects in Developing Countries: Evidence from Rural Schools in Bolivia," Review of Economics and Statistics 88 (February 2006): 171-77 Ibid.
22. Joshua D. Angrist and Victor Lavy, "Using Maimonides' Rule to Estimate the Effect of Class Size on Scholastic Achievement," Quarterly Journal of Economics 114 (May 1999): 533-75.
23. Such policy rules have also been used in research from the United States, but contrary to research from developing countries, the effect of class size is not statistically significant. See Caroline M. Hoxby, "The Effects of Class Size on Student Achievement: New Evidence from Population Variation," Quarterly Journal of Economics 115 (November 2000): 1239-85.
24. By including dummy variables for every student and every teacher in the sample, studies using such data sets are able to eliminate all unobserved time-invariant student characteristics and teacher characteristics as sources of bias. Eric A. Hanushek, John F. Kain, and Stephen G. Rivkin ("Teachers, Schools, and Academic Achievement," NBER Working Paper no. 6691 [National Bureau of Economic Research, Cambridge, MA, 1998]) find that a 10-student difference in class size is associated with only a 0.04 to 0,06 SD difference in student achievement.
25. Using panel data from New Jersey, Jonah E. Rockoff ("The Impact of Individual Teachers on Student Achievement: Evidence from Panel Data," American Economic Review 94 [May 2004]: 247-52) finds that the effect of class size on mathematics achievement is not statistically significant.
26. See Hanushek et al., "Teachers, Schools, and Academic Achievement";
27. Rockoff, "Impact of Individual Teachers."
28. Bryan Graham, "Using Conditional Excess Variance Contrasts to Identify Social Interactions: Theory and an Application on the Relationship between Peer Groups and Academic Achievement," unpublished manuscript (Department of Economics, Harvard University, Cambridge, MA, 2004).
29. Patrick J. McEwan, "Peer Effects on Student Achievement: Evidence from Chile," Economics of Education Review 22 (April 2003): 131-41.
30. McEwan ("Peer Effects on Student Achievement") also compares the test scores of siblings and twins who are assigned to different classrooms, i.e., by using siblings/twins fixed effects, and finds that a one standard deviation difference in the mean mother's education of one's classroom peers is associated with a difference in mathematics achievement of 0.371 and 0.536 SD for the sample of siblings and twins, respectively. These estimates will be biased, however, if siblings differ in ability and/or if students are nonrandomly assigned to classrooms.
31. For reviews of the evidence, see Patrick J. McEwan, "The Potential Impact of Large-Scale Voucher Programs," Review of Educational Research 70 (Summer 2000): 103-49;
32. Richard J. Murnane, "The Role of Markets in American K-12 Education," in The Limits of Market Organization, ed. Richard R. Nelson, 161-84 (New York: Russell Sage, 2005).
33. For evidence pertaining to Latin America, see Marie-Andrée Somers, Patrick J. McEwan, and J. Douglas Willms, "How Effective Are Private Schools in Latin America?" Comparative Education Review 48 (February 2004): 48-69.
34. George Psacharopoulos, "Public vs. Private Schools in Developing Countries: Evidence from Colombia and Tanzania," International Journal of Educational Development 7 (January 1987) : 59-67;
35. Donald Cox and Emmanuel Jimenez, "The Relative Effectiveness of Private and Public Schools: Evidence from Two Developing Countries," Journal of Development Economics 34 (November 1991): 99-121.
36. See Somers et al., "How Effective Are Private Schools," for further discussion.
37. Joshua Angrist, Eric Bettinger, Erik Bloom, Elizabeth King, and Michael Kremer, "Vouchers for Private Schooling in Colombia: Evidence from a Randomized Natural Experiment," American Economic Review 92 (December 2002): 1535-58.
38. For experiments in the United States, see W. G. Howell, P. J. Wolf, D. E. Campbell, and P. E. Peterson, "School Vouchers and Academic Performance: Results from Three Randomized Field Trials," Journal of Policy Analysis and Management 21 (Spring 2002): 91-217;
39. Cecilia E. Rouse, "Private School Vouchers and Student Achievement: An Evaluation of the Milwaukee Parental Choice Program," Quarterly Journal of Economics 113 (May 1998): 553-602;
40. Eric Bettinger, "The Effects of Educational Vouchers on Academic and Non-academic Outcomes: Using Experimental Economic Methods to Study a Randomized Natural Experiment," unpublished manuscript (Weather-head School of Management, Case Western Reserve University, 2003).
41. Somers et al., "How Effective Are Private Schools."
42. For administrative purposes, Bogota is stratified into six distinct socioeconomic strata with six being the highest (see n. 10). Schools in Bogota are highly segregated by SES.
43. The personal and family characteristics used in this procedure included student's age, gender, number of days absent from school, mother's education, mother head of household, family's ownership of an encyclopedia, computer, car, refrigerator, washing machine, vacuum cleaner, phone, and having access to electricity in the home. All of the latter have been used as indicators of the student's SES. We also fitted our final models on the smaller sample of students for whom we had complete data. The results were not substantively different from those reported in the paper. In order to evaluate the generalizability of our study's findings to the population from which our sample was drawn, we assessed whether the schools and students that did not respond in the second wave of data collection differed in any systematic manner from those that did. On average, nonmatched students are more likely to be male, to have obtained lower scores on the tests, to come from single mother households, and to have been absent from school more often. Although proportionally more public schools dropped out of the study than did private schools, we found no statistically significant differences in school average pretest scores or school SES between public schools that participated both years and those that dropped out of the study after the first year. However, among private schools, those schools that participated in both years had, on average, higher pretest scores and student SES than those schools that dropped out after the first year.
44. This test was developed in 1997 for an international comparative study of third and fourth graders organized by UNESCO's Laboratorio Latinoamericano de la Calidad de la Educación. Thirteen Latin American countries, including Colombia, participated in this study. Curriculum and testing experts from all participating countries, with the help of the Educational Testing Service, developed and pilot tested the instruments to make them reflect the curriculum of the first 4 years of primary education in the participating countries. This test has two parallel forms (form A and form B), each containing 32 questions. We used both forms of this test in the study, randomly distributing each form to approximately half the students in each classroom. We included in all models a dichotomous predictor to distinguish the form of the fifth grade test taken by each student. The reported Cronbach's alpha reliability for forms A and B of the test is .86 and .88, respectively, as reported in UNESCO, Primer estudio international comparative sobre lenguaje, matemática y factures asociados en tercero y cuarto grade: Informe preparado par el Laboratorio de Evaluación de la Calidad de la Educación (Santiago: UNESCO/OREALC, 1998), 67.
45. The fourth-grade tests used in this study were developed by Universidad Nacional for Bogota's Department of Education 1998 student assessment program, where they were used to test third graders. The pretests had three parallel forms, each containing 17 questions on mathematics and 17 on language. We administered all three forms of the pretest, randomly distributing the forms to approximately one-third of the students in each classroom. Items were scored separately by subject, on a scale ranging from 0 to 306. Students achieved an average of 200.2 points (SD = 50) with scores ranging from 12 to 306 points (six students in the sample obtained perfect scores). There are no published reliability or validity estimates for the pretests. We verified that the substantive results were not different when the models were fit with both the fourth-grade reading score and math score included separately as control variables.
46. We chose to use mother's education rather than father's education for two reasons. First, there are a number of studies that suggest that mother's education is more important than father's education in determining IQ; see Arleen Leibowilz, "Home Investments in Children," Journal of Political Economy 82 (March 1974) : S111-S131. Second, there was a high percentage of missing data on father's education in our sample. In this sample, 356 observations (9.9 percent of the student sample) were missing information on mother's education. We imputed these missing values by substituting the median value for the student's classmates (we imputed the median rather than the mean because we did not want the imputed values to be sensitive to outliers). Likewise, 178 observations (4.9 percent of the student sample) were missing data on whether their mother was head of the household (MHH). Given that we constructed this variable based on students' responses to whether their father lived with them or not, in the cases where they failed to answer we assumed that their father did not.
47. Sarmiento et al., Evaluation de la Calidad.
48. In the 77 cases (2.8 percent of the student sample) in which information was missing on this variable, we substituted the sample mean for the missing value.
49. In the data analyses, we converted this ordinal variable into a system of dichotomous (dummy) predictors, with HWHELP3 (always receives help) being omitted from the fitted models as a reference category. In the 19 cases (0.7 percent of the student sample) in which information was missing for this variable, we substituted the value 2, which was the modal response.
50. For the 15 cases (0.5 percent of the student sample) in which information was missing on this variable, we assumed that there were no books in the home.
51. See Michael A. Boozer and Stephen E. Cacciola, "Inside the 'Black Box' of Project Star: Estimation of Peer Effects Using Experimental Data," Economic Growth Center Discussion Paper no. 832 (Yale University, New Haven, CT, 2001).
52. Averaged over students rather than classrooms, the mean is 36 students.
53. We chose a log function over a quadratic function of class size for several reasons. First, a quadratic functional form results in predicted values that first increase and then decrease very sharply with class size. Therefore, we chose a log functional form because it constrains achievement to be a monotonic function of class size. Second, we could not reject the null hypothesis that the predicted values of student achievement from models with the alternative functional forms for class size were not different, when evaluated at either the tenth or ninetieth percentile of the class size distribution. Third, the model's fit was slightly better with the log specification than with the quadratic specification. Note that we followed Frederick Mosteller and John W. Tukey's advice and "started" the transformation by adding a small amount (0.167) to the difference in order to avoid infinities when the student's class size equaled the maximum; see their Data Analysis and Regression (Boston: Addison-Wesley, 1977), chap. 4. To verify the appropriateness of this recommendation, we estimated a nonlinear model of student achievement against In (Max Class Size + k - Class Size) in order to determine the value of k that maximized the model fit. This procedure did indeed return a value very close to the recommended value.
54. Hanushek et al., "Teachers, Schools, and Academic Achievement";
55. Richard J. Murnane and Barbara R. Phillips, "Learning by Doing, Vintage, and Selection: Three Pieces of the Puzzle of Relating Teaching Experience and Teaching Performance," Economics of Education Review 1 (Autumn 1981): 453-65;
56. Rockoff, "Impact of Individual Teachers." As in the case of class size, we chose the log functional form over the quadratic functional form because the pattern of predicted values resulting from the latter seemed unlikely, given previous research (such as that by Hanushek et al.). The difference in predicted values between the log and quadratic functional forms is not statistically significant at either the tenth or ninetieth percentile of experience. Moreover, model fit is virtually identical for the two models.
57. If student achievement, net of background characteristics, did not vary among classrooms, our research could not provide any results. However, as explained below, student achievement is higher in some classrooms than in others, and this is even true among classrooms taught by the same teacher.
58. Stephen W. Raudenbush and Anthony S. Bryk, Hierarchical Linear Models: Applications and Data Analysis Methods (Thousand Oaks, CA: Sage, 2002).
59. To verify our respondents' statements that, with the exceptions noted above, students taught by the same teacher were not assigned to classes on the basis of ability or SES, we utilized a strategy proposed by McEwan ("Peer Effects on Student Achievement"). This involved creating "virtual class-rooms" by randomly reassigning students from separate classrooms into "new" classrooms. We then compared the distribution of class size and peer group composition in the virtual classrooms with those in the actual classrooms. We found almost no differences in the two distributions. Of course, this means that there is only modest variation in class size and peer group composition among classrooms taught by the same teacher.
60. This random effects specification is essentially a special case of the most general specification used by Raudenbush and Bryk, Hierarchical Linear Models:, see the Analytic Appendix for details.
61. 2 = 11.91, p = 0.4528, df = 12, based on model 3b). We therefore include only the main effect of PRIVATE in our reported analyses.
62. F(96,2986) = 4.30; p < 0.001.
63. UNESCO, Primer estudio internacional, 258.
64. Caroline M. Hoxby, "Peer Effects in the Classroom: Learning from Gender and Race Variation," NBER Working Paper no. 7867 (National Bureau of Economic Research, Cambridge, MA, 2000);
65. McEwan, "Peer Effects on Student Achievement";
66. Boozer and Cacciola, "Inside the 'Black Box.'"
67. See the Analytic Appendix for a more detailed explanation of our complementary use of both fixed effects and random effects models (the latter being most similar to the hierarchical linear models described in Raudenbush and Bryk, Hierarchical Linear Models). In essence, the parameter estimates from fixed effects models are less biased than those from analogous random effects models. However, because there are only 13 teachers in our sample who teach multiple classes, the standard errors of the estimates from fixed effects models will be larger than those from random effects models, resulting in a loss of power when making statistical inferences. Fortunately, in the case of model 2, the estimates from the fixed effects and random effects models are very similar, so we can conclude that the random effects model provides unbiased estimates after all. As such, we can use the (more precise) standard errors from the random effects model to make statistical inferences.
68. McEwan, "Peer Effects on Student Achievement."
69. Given our identification strategy, it is possible that the peer effect actually reflects differences in the effectiveness with which the same teacher teaches groups of students with different characteristics.
70. We also tried estimating peer effects using an alternate measure of peer composition: the percentage of students in classroom j with two parents. In the fixed effects specification (model 2b), a 1 percent difference in the number of children from a two-parent family results in a 35.3 point difference in a classroom's average test score (and is statistically significant at the 5 percent level). Thus using this alternate measure also leads us to conclude that peer effects are statistically significant in our data set.
71. As explained earlier, this pattern stems primarily from the public school teacher salary scale and its applicability to private schools. While the salary schedule provides strong incentives for public school teachers to gain educational credentials and to remain in teaching, the salary scale creates incentives for private schools to avoid hiring highly experienced teachers with significant educational credentials because they are very expensive.
72. 0k are too numerous to list in a table; however, they can be provided by the authors upon request.
73. Somers et al., "How Effective Are Private Schools."
74. Angrist et al., "Vouchers for Private Schooling."
75. The statistically significant variation in predicted average student achievement among students taught by different private school teachers indicates how sensitive the size of the private school effect is to the composition of the private school sample.
76. 0k from fitted model 2b.
77. 0k from fitted model 2c.
78. It is important to point out that we do not know for sure that the average differences in achievement among students in classrooms taught by different teachers represent teacher effects. Although we do control for class size and one measure of peer group composition, it is possible that there are also unmeasured classroom-specific differences that account for some of the classroom-specific achievement differences.
79. Other studies that show this same pattern include Hanushek et al., "Teachers, Schools, and Academic Achievement";
80. Murnane and Phillips, "Learning by Doing";
81. Rockoff, "Impact of Individual Teachers."
82. On New Zealand, see Fiske and Ladd, When Schools Compete; on Chile
83. see Chang-Tai Hsieh and Miguel Urquiola, "When Schools Compete, How Do They Compete?"
84. Caroline M. Hoxby, "Ideal Vouchers," unpublished manuscript (Department of Economics, Harvard University, Cambridge, MA, 2001);
85. Dennis Epple and Richard Romano, "Education Vouchers and Cream-Skimming," NBER Working Paper no. 9354 (National Bureau of Economic Research, Cambridge, MA, 2002).
86. See the discussion in Murnane, "The Role of Markets in American K-12."
87. Raudenbush and Bryk, Hierarchical Linear Models.
88. Fixed effects models are often used with panel data. However, they are also commonly used with cross-sectional data. Julian R. Betts ("Does School Quality Matter? Evidence from the National Longitudinal Survey of Youth," Review of Economics and Statistics 77 [May 1995]: 231-50) incorporates fixed effects for states within the United States.
89. Patrick J. McEwan ("Peer Effects on Student Achievement") uses school, twin, and sibling fixed effects.
90. Jere R. Behrman and Barbara L. Wolfe ("Does More Schooling Make Women Better Nourished and Healthier? Adult Sibling Random and Fixed Effects Estimates for Nicaragua," Journal of Human Resources 24 [Autumn 1989]: 644-63) use both sibling fixed effects and sibling random effects models.
91. We used a Hausman specification test to evaluate whether estimates from the fixed effects and random effects models differed based on model 2c and find that they do at all standard levels of significance. This indicates that estimates from the random effects model are indeed biased.