77302 THE WORLD BANK ECONOMIC REVIEW. VOL. 13. NO. 3: 4 15-41 Do Community-Managed Schools Work? An Evaluation of El Salvador's EDUCO Program Emmanuel Jimenez and Yasuyuki Sawada This article examines how decentralizing educational responsibility to communities and schools affects student outcomes. It uses the example of El Salvador's Community- Managed Schools Program ^Education con Participation de la Comunidad, EDUCO), which was designed to expand rural education rapidly following El Salvador's civil war. Achievement on standardized tests and attendance are compared for students in EDUCO schools and students in traditional schools. The analysis controls for student character- istics, school and classroom inputs, and endogeneity, using the proportion of EDUCO schools and traditional schools in a municipality as identifying instrumental variables. The article finds that enhanced community and parental involvement in EDUCO schools has improved students' language skills and diminished student absences, which may have long-term effects on achievement. Central governments in developing countries usually play a major role in allocat- ing educational resources. Even when authority is delegated to subnational lev- els, such as provinces or municipalities, individual school administrators and parents play only a limited pact. This kind of centralized structure may work best for regulating and administering large systems uniformly, but it may also be in- effective and expensive when school needs differ widely across communities and when there are diseconomies of scale. Moreover, a centralized system can stifle the initiative of those who are most critical in affecting school outcomes— teachers, principals, and parents. Despite the compelling case for school-based management, there is relatively little empirical evidence documenting its merits in developing countries.1 The main reason is that these administrative arrangements have only recently been 1. Two exceptions are James, King, and Suryadi (1996) for Indonesia, and Jimenez and Paqueo (1996) for the Philippines. Both studies conclude that community-based involvement improves efficiency. Emmanuel Jimenez is with the Development Research Group at the World Bank, and Yasuyuki Sawada is with the Department of Advanced Social and International Studies at the University of Tokyo. Their e- mail addresses are ejunenez2@worldbank.org and sawada@stanfordalumni.org. This project has been financially supported by the Development Research Group and the Research Support Budget (RPO 679-18 and 682-08) of the World Bank. The authors gratefully acknowledge the comments and support from the Evaluation Unit of the Ministry of Education of El Salvador, which collected the data. They also thank Marcel Fafchamps, Paul Glewwe, Elizabeth King, Takashi Kurosaki, Martin Ravallion, Laura Rawlings, Fernando Reimers, Diane Steele, the anonymous referees and editors, and participants in seminars at the World Bank, Stanford University, the University of the Philippines, International Child Development Centre (United Nation's Children's Fund, Florence), and the Institute of Developing Economies (Japan) for useful discussions and comments. © 1999 The International Bank for Reconstruction and Development/THE WORLD BANK 415 416 THE WORLD BANK ECONOMIC REVIEW. VOL. 13. NO. i implemented (World Bank 1994, 1995, 1996). One celebrated example is El Salvador's Community-Managed Schools Program (Education con Participation de la Comunidad, EDUCO). EDUCO is an innovative program for both preprimary and primary school designed to decentralize education by strengthening the di- rect involvement and participation of parents and community groups. A prototype of today's EDUCO schools emerged in the 1980s, when public schools could not be extended to rural areas because of El Salvador's civil war. Some com- munities took the initiative to organize their own schools, which an association of households administered and supported financially. Although these early attempts were constrained by the low income base in rural areas, they demonstrated com- munities' strong inherent demand for education and desire to participate in the governance of their schools. In 1991 El Salvador's Ministry of Education (MINED), supported by aid agencies such as the World Bank, decided to use the prototype that communities themselves had developed as the basis of the EDUCO program. Today, EDUCO schools are managed autonomously by community education associations (asociaciones comunales para la education, ACEs), whose elected members are parents of the students. Reimers (1997) describes community asso- ciations as being composed of literate members of the community who are given basic training in school management. These associations meet periodically with teachers and also provide them with teaching materials. In EDUCO schools the ACEs are in charge of administration and management; MINED contracts them to deliver a given curriculum to an agreed number of students. The ACEs are then responsible for hiring (and firing) teachers, closely monitoring teachers' perfor- mance, and equipping and maintaining the schools. The partnership between MINED and the ACEs is expected to improve school administration and manage- ment in that the ACES can better gauge local demand. In the future MINED intends to introduce community management to all traditional schools. The EDUCO program was conceived as a way to expand educational access quickly to remote rural areas. Initial evidence indicates that it has accomplished this goal (El Salvador, MINED 1995; Reimers 1997). The question that remains is whether this expansion has come at the expense of learning. Professional administrators in the center are less involved in the day-to-day running of schools, which are now in the hands of local communities. But many of the parents in these communities have an inadequate education themselves. Thus it remains to be seen whether moving away from traditional, centralized programs and toward greater community and parental involvement also improves students' learning. This article assesses the impact of EDUCO schools. We estimate school produc- tion functions using three measures of educational outcomes for third-grade stu- dents. 2 Two of the measures are standardized test scores in mathematics and 2. This study is part of a larger effort by the World Bank to distill the lessons of decentralized education (see World Bank 1996). Eventually, we want to determine whether all students in EDUCO schools achieve better educational outcomes at comparable costs relative to their counterparts in traditional public schools. This article has a more limited objective: it uses school production functions to compare three measures of educational outcomes among third-grade students only. Jimenez and Sawada 417 language. These may be good indicators of educational outcomes. However, they may also be relatively unresponsive in the short run to changes in school gover- nance. Thus we also use an indicator that can be considered more of an interven- ing variable in determining student achievement but is likely to exhibit a short- run response: the number of school days that a student has missed. As with all comparisons of educational achievement, the key is to quantify how much of the differential in academic achievement can be explained by differ- ences in household background, the schools' quantitative inputs, and, most im- portant, organizational factors attributable to intangible differences in the way that traditional and decentralized schools are run.3 We also address parents' en- dogenous school choice by explicitly considering how the government selected which municipalities would be the first to have EDUCO schools. I. THE CONCEPTUAL AND EMPIRICAL FRAMEWORK Educational outcomes are products of the complex interactions of agents who participate in the schooling process. Students' characteristics and motivation are key, but so are the actions of individual parents, parent groups (such as parent- teacher associations), teachers, and administrators from the school level up to the education ministry. In addition, agents not directly connected to the educa- tional system can affect educational outcomes if they influence the environment in which students learn. For example, decisions about road infrastructure in a locality could afford access to certain types of schools, or the provision of elec- tricity in a municipality could enable students to study at night. The Basic Model It would be impossible to model the structural relationships that capture the behavior of each relevant agent.4 Instead, we postulate a simple reduced-form model of educational outcomes (Y). Most studies measure educational output by using students' achievement scores, attendance rates, repetition rates, decision to continue in school, or dropout rates. These variables are thought to capture pros- pects of future earnings in the labor market. In this article we focus on two com- ponents of Y: scores on standardized achievement tests (S) and days absent from school {A). Studies of education production functions have had mixed success in explain- ing S.5 Aside from measurement and estimation issues, outcomes may be deter- mined by endogenous choices. For example, some of the explanatory policy vari- ables that determine S, such as type of school, may be systematically related to unobservable characteristics, which themselves may not be random across obser- vations. This could lead to bias. As explained below, we attempt to correct for 3. See Levin (1997) for a good review of these intangibles. 4. McMillan (1999) presents an interesting model of the interaction of parental and school preferences in determining educational outcomes. 5. See Hanushek (1995) for a review. 418 THE WORLD BANK ECONOMIC REVIEW. VOL. 13, NO. 3 this problem by modeling and estimating the choice of school type and then using that estimate in the production function to control for participation. It is often difficult to identify such models. But we are able to use the participation rule that the Salvadoran authorities used in choosing where to place EDUCO schools as the identifying restriction that directly affects choice but not outcomes. It may take time for a policy change such as decentralization to affect school performance, which tends to be a cumulative measure. We thus also consider an important intervening variable that eventually influences student outcomes: ab- sence from school (A). Students may be absent for a number of reasons, some of which, such as illness, have nothing to do with decentralization. But other rea- sons may be tied to school organization and management. Students (or their parents) may not be motivated to ensure regular attendance because the quality of schooling is poor or because parents do not feel involved in the education process. Also contributing to student absence is teacher absence, an important reason why students do not attend school in El Salvador. If teachers are absent, classes are usually canceled, since there is no tradition of using substitute teach- ers. Although teachers are sometimes absent for legitimate reasons, such as sick- ness, more often they are simply not fulfilling their duty. Teacher absence is an issue in many countries besides El Salvador: Lack of motivation and professional commitment produce poor attendance and unprofessional attitudes towards students. Teacher absenteeism and tardiness are prevalent in many developing countries . . . absenteeism is especially acute in rural areas. Students obviously cannot learn from a teacher who is not present, and absenteeism among teachers encourages similar behavior among students. In some countries . . . parents react to high rates of teacher absenteeism by refusing to enroll their children in school. (Lockheed and Vespoor 1991: 101.) Teacher absence could be minimized if teachers were appropriately monitored. We would expect that, in a decentralized school, parental involvement would mitigate such behavior. We assume that the components of Y = [S A] can be independently estimated. A will likely affect S, and we assume an implicit recursive process S = S(A), in which the residuals from the different equations are independent of each other, and the matrix of coefficients of endogenous variables is triangular. Each struc- tural equation can thus be estimated by ordinary least squares (OLS), equation by equation (Greene 1997). A simple model for the rth student in the nth school in the wth community is (1) X&TMf = fP^inm) C m , Djnm) where X is a vector of student and household characteristics, C is a vector of community variables for municipality m, and D is the type of school, either a decentralized EDUCO school or a traditional school. In this model the type of school is assumed to determine most of the school characteristics that affect stu- Jimenez and Sawada 419 dent outcomes. This model is the ultimate reduced form—it assumes that the effect that a school's observed characteristics, such as class size and teacher char- acteristics, have on achievement is fully determined by the school's management structure (that is, whether it is a decentralized EDUCO school or a traditional school) and the characteristics of the students and parents who participate in decisions concerning the school. We can often observe the effects of management structure through differences in school and classroom inputs, such as teacher-pupil ratios, teacher remunera- tion, or the educational background of teachers and administrators. But even if we were to enter as many observable school characteristics as we could in equa- tion 1, the type of school may still be significant because it captures unobserved managerial inputs (Levin 1997). Indeed, in reviewing 96 studies on the effects of five educational inputs on student performance in developing countries, Hanushek (1995) concludes that there are no clear and robust technical relationships be- tween key school inputs and student performance.6 Thus differences in resources might not be important determinants of school outputs, implying that schools in developing countries are paying for inputs that have little consistent effect on student performance. We distinguish the unobserved effect of community par- ticipation from the other unobserved effects of management by explicitly taking into account differences in the level of community involvement. Accordingly, we also derive an alternative model: = \^l 'iron l(^imn> *-'m» '-'imt' ^*nm> '•rmJ where Z is a vector of observed school and classroom characteristics, and P is the intensity of community participation. Since Z and P vary by school rather than by student, equation 2 expresses the achievement of the rth student in the wth school. To simplify notation, we drop the school and community subscripts in the rest of this article. Empirical Specification Linearizing and adding a stochastic term, which represents a well-behaved measurement error term, to equation 1, we derive the following regression formula: (3) Y( D takes a value of 1 if the rth student attends a decentralized EDUCO school and 0 if the student attends a traditional, centralized school. By assumption, E(«,) = 0 and Var(«,) = Oj. We add school and classroom characteristics and the intensity of community participation to derive the empirical version of equation 2. 7 6. Hanushek (1995) does, however, suggest that a minimal level of basic school resources, such as textbooks and facilities, is important to student achievement. 7. To simplify notation, we do not add the error terms associated with the school and municipal-level variables. We handle the school variables by using a program participation model and the municipal-level variables by using a municipality-level fixed-effects model. 420 THE WORLD BANK ECONOMIC REVIEW, VOL. 13. NO. 3 Observed household and student characteristics reflect the ability of parents to provide a supportive environment for their children. If capital markets were perfect, then life-cycle consumption and human capital investments could be determined independently. Parents would simply borrow to finance the home inputs needed to maximize their children's learning. But since credit markets are far from perfect in El Salvador, the economic circumstances of the household become important. In this article we use asset variables to control for the at- tributes that are hypothesized to be positively correlated with schooling out- comes (homeownership and the availability of electricity, sanitary services, and piped water). In addition, we control for parents' education, which may also directly affect living standards and preferences for children's education. We cannot measure students' innate ability directly. However, student charac- teristics that may be important include gender, since parents or teachers my treat boys and girls differently; age, since older students, while more mature and more likely to score higher, may be self-selected as underachievers and left behind by their cohort; and number of siblings, since the greater the number, the less time parents have to devote to any one child, that is, there are resource competition effects. We capture community characteristics, C, by municipality-level fixed effects. In El Salvador municipalities are the next administrative level below the depart- ment level. There is substantial variance in the distribution of resources across municipalities, which could affect students' access to ancillary services, such as the availability of electricity needed to study, which, in turn, could affect school- ing outcomes. Endogenous Program Participation A key estimation issue is endogenous program participation. Endogeneity may arise because parents choose which type of school their children attend (condi- tional on their choosing to send them to school, since we do not have informa- tion on children who are not in school).8 If attendance at an EDUCO school is systematically based on unobserved characteristics that could also influence stu- dent achievement, then the OLS estimates of the effect of EDUCO would be biased. That is, a in equation 3 may not accurately measure the value of attending an EDUCO school. The direction of the bias is ambiguous. If the important unobserved character- istics are students' motivation to learn and parents' commitment to education, and these variables are positively correlated with participation in EDUCO, then comparing outcomes, even after holding constant for observed characteristics, 8. Although EDUCO schools were targeted to areas with limited primary school coverage, parents would still have had a choice of whether or not to send their children to school. They could have had their children commute, albeit over long distances (child fosterage for schooling is not uncommon in developing countries; see Ainsworth 1992 and Glewwe and Jacoby 1994). Or, they could have changed residences (Salvadoran migration rates are high). Unfortunately, the school-based nature of the sample prevents us from including nonatrendance as an option. Jimenez and Sawada 421 would overestimate the effect of EDUCO. This bias, however, may be mitigated by the fact that EDUCO targets economically disadvantaged communities. To take the possibility of bias into account, we explicitly model program par- ticipation (that is, whether or not a student enrolls in an EDUCO school rather than in a traditional school). Using a familiar method for obtaining so-called treatment effects, we then estimate this model to obtain the parameters needed to correct equation 3. 9 WHAT DETERMINES PROGRAM PARTICIPATION? We assume that governments set priorities regarding which municipalities will receive an EDUCO school. Households then use that information to choose the type of school that maximizes their indirect lifetime utility, V. Parents make this choice by weighing the benefits and costs of an EDUCO school relative to other types of schools. The benefits of EDUCO depend on households' perceptions of the virtues of a decentralized program. Some of these preferences can be captured by measurable household characteristics, X, but others are unobserved. The cost of an EDUCO school relative to a traditional school depends on rela- tive direct costs, such as tuition payments, books, and other fees. The most im- portant components of cost are largely the same for both types of schools: all schools and books are free in first through sixth grade. But there are differences in the other direct costs, EDUCO students do not pay a registration fee, do not buy uniforms, and receive a basic package of school supplies, such as pencils, rulers, and markers.. Students in traditional rural schools must bear all of these costs.10 However, EDUCO parents must devote a substantial amount of time to the school by providing school meals and by building, maintaining, and administering the school.11 The principal cost differential between EDUCO and traditional schools comes from differences in access, given the relative paucity of schools in rural areas. We do not have information on households' schooling options (such as the distance from households to feasible EDUCO or traditional schools) because our data are school-based, not household-based. However, we assume that a household is more likely to choose an EDUCO school if the government considers the munici- pality a priority for the program and thus an EDUCO school is available in the community. The government gives priority to municipalities considered to be neediest according to a classification system developed by MINED and the Minis- try of Health. Municipalities' uneven access to social services has always been a serious issue in El Salvador. However, poverty is more widespread in smaller municipalities, 9. See Greene (1997: 981-82) for a clear discussion of this estimation strategy. 10. We are grateful to Diane Steele of the World Bank for this information, which she received from a phone interview with MINED staff. 11. We do not have data on the magnitudes of these costs. We assume in this article that these cost differentials are roughly offsetting for decentralized and traditional schools. We will verify this assumption with data from survevs that were fielded onlv in 1999. 422 THE WORLD BANK ECONOMIC REVIEW. VOL. 13. NO. I which usually lack the financial and institutional capacity to administer and manage social services. The EDUCO program was developed in 78 of the country's poorest municipalities. It started in 1991 with six ACES in three departments; by the end of 1992, the program had extended to all 14 departments. The key variables in the targeting system are the incidence of severe malnutri- tion (the percentage of undersize children in the municipality), the repetition rate, the percentage of overage students, and the net enrollment rate. Higher values for the first three variables, and a lower value for the last, make a municipality a higher priority. In the next section we discuss how this prioritization affects our choice of instruments. THE FORMAL MODEL OF PROGRAM PARTICIPATION. In the model a household chooses the type of school that yields the highest level of indirect utility, V J - There are two options: / = D if the household chooses a decentralized EDUCO school o r / = T if the household chooses a traditional rural school. V. depends on the relative benefits and costs of attending an EDUCO school as perceived by parents. Parents choose EDUCO if, for the rth student: (4) D,»=VDl-VTl>0 where D* is a latent variable that describes the likelihood that a child is in an EDUCO school. It is determined by: (5) D,* = W,co + e, D, = 1 if D? > 0; 0 otherwise. In equation 5, E(e,) = 0, Var(e,) = c2e, 0) = E(H.I W.co + e > 0) * 0. Jimenez and Sawada 423 where XDi and A.T/are selection terms estimated from Mills ratios.14 The difference in expected performance between EDUCO participants and nonparticipants, con- ditional on having chosen a type of school, can be obtained by subtracting equa- tion 7 from equation 6: (8) E( Y,| D, = 1) - E( Y,l D, = 0) = a + aut (\Di + XTi) where a is the coefficient of the EDUCO intercept and is usually referred to as the "true" program effect (see Maddala 1983). Thus if we define e, = u, - cu^.Di A + °«e^T/ (1 - A h a t e r m whose expectation is 0 for each of the cases D = (1,0), the following regression would yield unbiased estimators: (9) Yi=X^ + Cmy+Dia + aue[XDiDi-\Ti(\-Di)) + e, If we omitted the selection correction terms (in brackets) from this regression, the difference in equation 8 would be equal to what is usually estimated as the least squares coefficient on the treatment dummy variable. But this expression would overestimate or underestimate the treatment effect, depending on the direction of the participation bias. To estimate this model, we employ a two-step method.15 In the first step we estimate equation 5 as a probit model and then use the results to calculate the inverse Mills ratios XD and \T. In the second step we use the estimated inverse Mills ratios to form the participation terms in equation 9. We then estimate equa- tion 9 with municipal dummies to capture regional fixed effects. If the error terms in the probit and outcome equations are negatively corre- lated, that is, if am < 0 (this would occur if an unobserved variable, such as student motivation, negatively affeaed the likelihood of attending an EDUCO school but positively affeaed student achievement), then equation 6 implies that the prediaed score of a student drawn randomly from the population would be un- derestimated in the case of EDUCO schools if we use sample mean scores. This can be easily verified by the relationship, E(Y,I D, = 1) < X,P + Cmy + a if aKE < 0. A similar calculation can be done for equation 7 in the case of traditional schools. In a linear model, estimation and parameter identification are possible only if the veaors [X Cm] and W have no elements in common and are linearly indepen- dent. However, in the model above, even if [X Cm] and W are identical, equation 9 is estimable. This is because the first-stage estimation results are entered as a nonlinear function in the second stage (equation 9). The nonlinearity helps to identify the model. 14. Assuming joint normality between uj and tn E(n,l W.co + E; > 0) = E(/i( I D* > 0) = C^p,, where XDl. = <|>(W(. co) / (W(co) / [1 - 4>(W,.Q>)]. 15. An alternative way to estimate program participation on unobservables is to use a maximum likelihood method without focusing on Mills ratios. However, this method is more burdensome computationally. 424 THE WORLD BANK ECONOMIC REVIEW. VOL. 13. NO. 3 II. DATA DESCRIPTION MINED collected the data in October 1996 with the assistance of the World Bank and the U.S. Agency for International Development (USAID). The survey covered 162 of the country's 262 municipalities. These municipalities share re- sponsibility with the central government for delivering social services. Since EDUCO was introduced only in 1991, it was not possible to compare the scores on achievement tests given in 1996—only five years later—of EDUCO stu- dents who were about to finish their primary education and students in tradi- tional schools. Instead, MINED decided to compare outcomes for third graders. MINED designed the sampling scheme so that the survey was nationally represen- tative. Moreover, the sample was selected so as to consider four types of schools: pure EDUCO, pure traditional, mixed, and private. We dropped students from private schools and traditional public urban schools from the sample, since they are not comparable to EDUCO students. Mixed schools have both EDUCO sections run by ACES and traditional sections. Some EDUCO programs rented space from traditional schools. The small number of students in these mixed schools attend either EDUCO or non-EDUCO classes located in traditional schools. Since the ad- ministration and management of mixed schools are different from the adminis- tration and management of pure EDUCO and traditional schools, and thus we should control for unknown management and school-level cross-effects, we could not include mixed schools in pure school samples. Nor could we isolate them as a separate category because of the small sample size. To ensure the robustness of our results, we based our estimations on pure schools only.16 This left us with 605 students in 30 EDUCO schools and 101 traditional schools. The survey comprises five questionnaires, one each for students, parents, school directors, teachers, and parent associations. The students' questionnaire requests information about students' relationship with their guardians, type of school, gender, and achievement test results. The parents' data include information on family background and living standards, such as parents' education level, the household's living standard, and asset ownership, as well as detailed socioeco- nomic information on students, including age, schooling, and health status. The questionnaire for the school director consists of questions about the director, student enrollment, the quality and quantity of teachers, school facilities, and finances. The data collected from teachers include their educational background, years of experience, and salaries, as well as information about the classroom, such as the availability of school materials and frequency with which members of the community association visit the classroom. Lastly, the community and par- ent association questionnaire contains qualitative information on how the asso- ciation is organized and how members participate in administration and man- agement of the school. The information on EDUCO schools was collected from 16. The results with both mixed and pure samples, which are not reported here, are consistent with the results for pure schools only. Jimenez and Sawada 425 ACES, and the information on traditional schools'was collected from a counter- part parent organization, the Sociedad de Padres de Familia. Dependent Variables MINED administered the achievement tests in October 1996 with the assistance of the Intercultural Center for Research in Education (El Salvador, MINED 1997). The tests were given nationally in the third, fourth, and sixth grades, but because EDUCO students had reached only the third grade when the data were collected, we use only the third-grade results in the analysis. Also, we focus only on scores for the mathematics and language tests, ignoring the social studies, science, health, and environment components. The mathematics test is composed of 30 questions covering ten key subjects— that is, three questions for each subject. Students have mastered a subject if they have answered two of three questions correctly. The language test includes 36 questions covering nine subjects—that is, four questions on each. Students have mastered a subject if they have answered three of four questions correctly. In our sample the average student was able to master 3.70 of 10 subjects in math, but only 1.75 of 9 subjects in language (table 1). These results are not out of line when compared with national averages (El Salvador, MINED 1997). Of greater interest are the average values for EDUCO and traditional schools. Students in EDUCO schools score marginally lower than students in traditional schools in both subjects, although the differences are not statistically significant (table 1). Our main concern is whether this similarity persists when we control for participation and student, school, and community characteristics. In addition to test scores, we examine another dependent variable from the parents' questionnaire. This is the response to the following question: "In the past four weeks, how many days of school did the child miss?" As mentioned earlier, we interpret student absence as an important intervening variable that eventually influences educational outcomes. Since we hold constant for student illness, we believe that this variable captures motivational factors. Comparing Table 1. Means and Standard Deviations of Variables by Type of School All. EDUCO Traditional Variable schools schools schools Output variables 3.70 3.59 3.73 Math achievement test score (2.54) (2.77) (2.47) 1.75 1.73 1.76 Language achievement test score (1.71) (1.85) (1.67) 0.95 0.95 0.95 Days absent from school in past four weeks (0.10) (0.11) (0.10) Child and household variables Gender (female = 1) 0.51 0.51 0.51 ChUd's age 10.58 11.01 10.44 (1.76) (1.97) (1.66) (Table continues on the following page) 426 THE WORLD BANK ECONOMIC REVIEW, VOL. 13. NO. 3 Table 1 (continued) All EDUCO Traditional Variable schools schools schools Child lives without parem(s)3 0.14 0.16 0.13 Child had respiratory illness or flu in the past two weeks* 0.60 0.63 0.59 Number of siblings (ages 4-15) 2.01 2.11 1.98 (1.54) (1.50) (1.56) Mother began basic education3 0.53 0.50 0.54 Mother's education missing* 0.08 0.06 0.09 Father began basic education3 0.39 0.38 0.40 Father's education missing9 0.04 0.03 0.04 Own house3 0.72 0.68 0.73 Electricity available3 0.58 0.28 0.67 Sanitary service available3 0.18 0.06 0.22 Water available3 0.06 0.01 0.08 School variables Teacher-pupil ratio (school level) 0.04 0.05 0.03 (0.056) (0.09) (0.041) Sanitation or latrine available at school3 0.93 0.89 0.94 Electricity available at school3 0.68 0.30 0.80 Piped water available at school3 0.32 0.12 0.38 Teacher and classroom variables Teacher finished university education3 0.46 0.75 0.37 Years of teacher experience 7.83 4.37 8.89 (6.44) (2.71) (6.87) Monthly base salary of teacher (thousands of colones) 3,035.21 2,919.23 3,070.71 (523.38) (269.40) (574.84) Teacher receives bonus3 0.64 0.74 0.61 All students have math textbook" 0.61 0.58 0.62 Math textbook information missing* 0.11 0.25 0.07 All students have language textbook* 0.59 0.59 0.59 Language textbook information missing3 0.12 0.28 0.07 Teacher instructs multigrade classroom* 0.24 0.39 0.20 Multigrade information missing* 0.01 0.04 0.00 Number of books in classroom library 74.32 114.63 61.98 (197.59) (272.84) (166.42) Classroom library information missing* 0.47 0.24 0.54 Community participation variable Number of parent association visits to classroom in the past month 2.52 5.65 1.56 (4.82) (6.59) (3.63) Regional school distribution Percentage of pure EDUCO schools in all primary schools within municipality 0.21 0.75 0.04 (0.34) (0.29) (0.11) Percentage of pure traditional schools in all primary schools within municipality 0.69 0.15 0.86 (0.37) (0.28) (0.20) Inverse Mills ratio 0.00 0.24 0.07 (0.33) (0.47) (0.22) Number of observations 605 142 463 Note: Standard deviations are in parentheses. a. Binary variable equals 1 if response is "yes," 0 otherwise. Jimenez and Sawada 42 7 sample means themselves indicates that, on average, students in both EDUCO and traditional schools missed 0.95 days in the four weeks before the survey. Explanatory Variables The means of the explanatory variables show the following. Both EDUCO and traditional schools have an equal number of girls and boys. A fairly large portion of students live without their parents, the proportion being slightly higher for EDUCO students. EDUCO students also have more siblings and are older, although the differences are not significant. Parents of traditional school students have more education than parents of EDUCO students. Fifty-four percent of mothers or female guardians of traditional students have had some basic education compared with 50 percent for EDUCO students. The gap also holds for fathers (40 and 38 percent). These differences in education are reflected in the asset variables. Fewer EDUCO parents are homeowners or have access to electricity, sanitary services, and running water, suggesting that EDUCO students come from poorer backgrounds than traditional school students. The socioeconomic characteristics of students are consistent with the charac- teristics of schools. While teacher-pupil ratios, access to textbooks, and the avail- ability of sanitary facilities are similar in both types of school, fewer EDUCO schools have access to electricity or piped water. However, more EDUCO teachers have finished university education, although they have less teaching experience. The EDUCO teaching corps consists of relatively young recent graduates who receive a bonus for teaching in the program. Another difference is that EDUCO parent asso- ciations visit classrooms more than once a week, which is three to four times more often than their traditional counterparts. The overall picture, then, is one of poor communities that have succeeded in mobilizing parents to become more involved in their children's education, de- spite their lower standard of living. What we want to know is how much of the differences in outcomes are due to EDUCO. Identification We account for possible program endogeneity by explicitly modeling the like- lihood of participation in EDUCO and using that information to correct the pro- duction function. The main challenge with such corrections is specifying the iden- tifying restriction that allows us to estimate the model. We include the percentages of EDUCO and traditional schools in all primary schools in each municipality to capture the relative cost of access to each type of school. Arguably, these percentages affect the likelihood that a student will at- tend an EDUCO school without directly affecting the education production func- tions at the student level. To isolate general community effects on achievement from the cost-of-access effect, we also include municipal fixed effects in the edu- cational output equation (equation 3). Although the percentages of EDUCO schools are linear combinations of these municipal dummies, we achieve identification because the probit participation equation has a nonlinear functional form. 428 THE WORLD BANK ECONOMIC REVIEW, VOL. 13. NO. 3 In order to test the robustness of our results, we also estimate specifications that do not rely exclusively on functional form for identification. For example, instead of using the proportion of EDUCO schools in each municipality, we in- clude the variables that the government uses to prioritize program placement: the extent of malnutrition, the proportion of overage students, repetition rates, and net enrollment rates. Because the government exogenously determines the prioritization formula, the variables included can be used to identify the partici- pation equation. We do not include them in the achievement equations, since, to the extent that local geographic conditions affect achievement, the municipal fixed effects capture their influence. Because our basic qualitative results do not change with these specifications, and to conserve space, we do not report the results here. They are, however, available from the authors. III. EMPIRICAL RESULTS: STUDENT ACHIEVEMENT The first step of the analysis is to estimate the determinants of participating in EDUCO to correct for possible endogeneity. The most significant variables are mother's education, household assets, and the geographical variables that cap- ture the cost of EDUCO schools relative to that of traditional schools. Mother's education, homeownership, and the availability of water are negatively corre- lated with EDUCO participation (table 2). Students from households that are bet- ter off have a higher likelihood of attending a traditional school. As expected, the availability of an EDUCO school within each municipality significantly increases the probability of enrolling in an EDUCO school. The next question is whether an EDUCO student (captured by an EDUCO dummy variable) achieves different test scores than a traditional school student. The re- gressions, which include the participation correction, use math and language achievement as dependent variables, and student and community characteristics (the latter captured by municipality fixed effects) as explanatory variables (table 3). The negative coefficient of the Mills ratio indicates that the error terms of the participation and achievement equations are negatively correlated. This means that EDUCO students have unobserved characteristics that are negatively corre- lated with achievement test scores. EDUCO's unconditional effect on language test scores is positive and signifi- cant, while its effect on math performance is positive and not significant (table 4). Thus the program has not lessened child learning (after correcting for partici- pation). In fact, it has improved performance in language. However, our measure of EDUCO's advantage in language may be imprecise. The estimate of the EDUCO coefficient is sensitive to the specification of the participation equation—it be- comes insignificant when we use in the first stage the municipal prioritization variables instead of the proportion of EDUCO schools in each municipality.17 17. However, the qualitative results described in the text hold. The results are available from the authors. Jimenez and Sawada 429 Table 2. Probit Analysis of School Choice Variable Coefficient Child and household variables Gender (female = 1) -0.14 (0.54) Child's age 0.11 (1.39) Child lives without parentfs)3 0.05 (0.12) Child had respiratory illness or flu in the past two weeks3 -0.47 (1.72)' Number of siblings (ages 4-15) 0.04 (0.43) Mother began basic education1 -0.54 (1.75)* Mother's education missing9 -1.29 (2.07)" Father began basic education3 -0.37 (1.18) Father's education missing3 -0.48 (0.63) Own house* -0.59 (2.08)" Electricity available3 -0.17 (0.61) Sanitary service available" 0.19 (0.46) Water available3 -3.00 (1.88)* Regional school distribution (proxies for cost variables) 5.96 Percentage of pure EDUCO schools in all primary schools within municipality (7.14)"» -1.66 Percentage of pure traditional schools in all primary schools within municipality (2.63)"* -1.19 Constant (1.11) Log likelihood -62.42 Pseudo R1 0.81 * Significant at the 10 percent level. ** Significant at the 5 percent level. *** Significant at the 1 percent level. Note: School choice is the dependent variable, which equals 1 for an EDUCO school and 0 for a traditional school, f-statistics are in parentheses. a. Binary variable equals 1 if response is "yes," 0 otherwise. Some of EDUCO's effects can be explained by observed differences in school inputs and community participation. In order to see the extent of these differ- ences, we estimate a model that includes school, classroom, and community par- ticipation effects (table 5). EDUCO's impact is lessened with the addition of these independent variables, indicating that some of the differences in test scores can be explained by differences in school inputs and the degree of community in- 430 THE WORLD BANK ECONOMIC REVIEW. VOL. 13. NO. 3 Table 3. Municipality Fixed-Effects Regressions of Student Achievement Mathematics Language Variable (V (2) (1) (2) EDUCO variables EDUCO school present11 0.45 2.17 (0.33) (2.32)** EDUCO school built in 1991-94" 0.74 2.16 (0.46) (1.91)* EDUCO school built in 1995° 1.68 2.91 (1.05) (2.62)* *» EDUCO school built in 1996a -0.37 1.73 (0.26) (1.72)* Year missing-1 -0.64 2.27 (0.35) (1.77)* Child and household variables Gender (female = 1) -0.69 -0.69 0.01 0.02 (3.19)*" (3.18)*" (0.08) (0.12) Child's age 0.19 0.19 0.04 0.04 (2.91)*" (2.88)*" (0.80) (0.80) Child lives without parent(s)* 0.38 0.35 0.43 0.42 (1.04) (0.97) (1.73) (1.68) Child had respiratory illness or flu in the past two weeks a 0.33 0.32 0.16 0.14 (1.42) (1.39) (0.98) (0.90) Number of siblings (age of 4-15) -0.05 -0.05 -0.02 -0.02 (0.65) (0.65) (0.40) (0.37) Mother began basic education* -0.09 -0.05 0.06 0.07 (0.35) (0.20) (0.32) (0.39) Mother's education missing3 -0.06 -0.06 0.33 0.30 (0.13) (0.14) (1-11) (1.01) Father began basic education9 -0.05 -0.04 0.19 0.20 (0.19) (0.16) (1.16) (1.21) Father's education missing9 0.54 0.43 -0.46 -0.49 (0.91) (0.72) (1.10) (1.19) Own house" -0.14 -0.17 0.13 0.13 (0.53) (0.65) (0.72) (0.70) Electricity available1 0.07 0.06 0.01 0.006 (0.24) (0.21) (0.03) (0.03) Sanitary service available1 0.55 0.49 0.25 0.22 (1.76)* (1.56) (1.13) (0.99) Water available3 -0.31 -0.25 -0.35 -0.32 (0.61) (0.50) (1.01) (0.91) Inverse Mills ratio -0.46 -0.27 -1.16 -1.05 (0.56) (0.33) (2.03)** (1.80)* Constant 1.82 1.85 0.51 0.50 (2.17)" (2.18)** (0.88) (0.84) Number of observations 605 605 605 605 Number of municipalities 90 90 90 90 R1 0.0242 0.0126 0.0002 0.0001 * Significant at the 10 percent level. ** Significant at the 5 percent level. *** Significant at the 1 percent level. Note: Dependent variable is score on mathematics or language test, r-statistics are in parentheses. a. Binary variable equals 1 if response is "yes," 0 otherwise. Jimenez and Sawada 431 Table 4. Summary of EDUCO Effects on Student Achievement Without school With school inputs but With school inputs inputs or community without community and community Subject participation variables participation variables participation variables Mathematics 0.45 0.40 -0.77 (0.33) (0.27) (0.47) Language 2.17 1.57 0.74 (2.32)" (1.51) (0.65) * "Significant at the 5 percent level. Note: f-statistics are in parentheses, volvement. Community involvement is captured by the coefficient on the num- ber of visits that members of the parent association made to classrooms. The coefficient is consistently positive and significant for the basic model with EDUCO dummy variables. This suggests that active community participation is crucial for improving students' achievement in EDUCO schools. An additional classroom visit per week could increase mathematics and language test scores 3.8 and 5.7 percent, respectively.18 Teacher monitoring by members of parent associations could also improve the quality of education, particularly in EDUCO schools. We try to distinguish between cohort years by including dummy variables for when the EDUCO program began: prior to 1995, in 1995, or in 1996. Our hypoth- esis is that the EDUCO effect may be stronger for schools that were built earlier, since they may have learned how to operate the system better. An alternative hypothesis is that newer schools would have better outcomes if there were a "Hawthorne" effect—that is, if the staff and students of newer schools were more motivated and ready to undertake reforms, the kind of enthusiasm that may wane over time. The coefficients for EDUCO are greater for entrants in 1995 and, in fact, are significant and positive for specifications without the community participation variable. This result is consistent with a Hawthorne effect. Still, most of the coefficients are not statistically significant. We can conclude from the OLS results, then, that the EDUCO program has not had a deleterious effect on student achievement, despite its rapid expansion. Looking at household background, we find that girls perform significantly worse than boys on the mathematics test. In contrast, there are no differences across gender in language. The coefficients on parents' education are not statisti- cally significant, possibly because this variable is likely to be highly correlated with some of the asset variables, and children from households with greater as- sets or access to infrastructure tend to have better outcomes. For example, per- formance in mathematics increases almost 15 percent of the mean if students come from households where sanitation is available. It is not surprising that homeownership is not significant—even poor rural families tend to own their own homes in El Salvador. 18. By estimating separate regressions, we find that there are significant effects when the EDUCO dummy is interacted with the participation variable. We do not present these results here. Table 5. Municipality Fixed-Effects Regressions of Student Achievement with School Inputs and Participation Mathematics Language Variable (1) (I) (3) (4) m (2) (3) (4) EDI/CO variables EDUCO school present3 0.40 -0.77 1.57 0.74 (0.27) (0.47) (1.51) (0.65) EDUCO school built in 1991-94" -1.93 -1.87 0.82 0.83 (0.97) (0.94) (0.59) (0.60) EDUCO school built in 1995" 3.21 5.50 3.26 3.85 (1.75)' (1.59) (2.55) (1.59) EDUCO school built in 1996" -0.49 -0.11 0.43 0.53 (0.28) (0.06) (0.35) (0.42) Year missing" —4.41 -4.82 0.45 0.34 (1.49) (1.60) (0.22) (0.16) Child and household variables £ Gender (female = 1) -0.57 -0.53 -0.51 -0.51 0.07 0.10 0.11 0.10 (2.61)"' f (2.40)" (2.31)** (2.33)" (0.47) (0.67) (0.69) (0.68) Child's age 0.17 0.17 0.18 0.18 0.02 0.02 0.02 0.02 (2.57)" (2.62)"* (2.66)* »» (2.65)"* (0.45) (0.50) (0.52) (0.52) Child lives without parcnt(s)J 0.42 0.42 0.40 0.39 0.46 0.46 0.44 0.43 . (1.17) (1.17) (1.11) (1.08) (1.85)* (1.85)' (1.73)» (1.72)' Child had respiratory illness or flu in past two weeks" 0.23 0.22 0.20 0.20 0.09 0.08 0.06 0.06 (1.01) (0.94) (0.88) (0.86) (0.56) (0.48) (0.39) (0.38) Number of siblings (ages 4-15)" -0.04 -0.03 -0.02 -0.02 -0.02 -0.01 -0.01 -0.01 (0.57) (0.38) (0.29) (0.33) (0.34) (0.14) (0.16) (0.17) Mother began basic education" -0.02 -0.02 -0.02 -0.02 0.07 0.08 0.08 0.08 (0.09) (0.09) (0.07) (0.06) (0.43) (0.43) (0.46) (0.47) Mother's education missing" -0.14 -0.12 -0.07 -0.07 0.34 0.35 0.33 0.33 (0.32) (0.28) (0.15) (0.16) (1.09) (1.13) (1.06) (1.05) Father began basic education" -0.07 -0.07 -0.06 -0.05 0.15 0.15 0.16 0.16 (0.31) (0.30) (0.25) (0.23) (0.90) (0.92) (0.96) (0.97) Father's education missing" 0.57 0.55 0.54 0.53 -0.37 -0.38 -0.42 -0.42 (0.94) (0.92) (0.89) (0.89) (0.89) (0.92) (1.00) (1.00) Own house" -0.19 -0.21 -0.23 -0.23 0.05 0.04 0.04 0.04 (0.74) (0.80) (0.89) (0.88) (0.26) (0.19) (0.21) (0.21) Electricity available" -0.02 -0.06 -0.15 -0.18 -0.01 -0.03 -0.06 -0.07 i (0.07) (0.18) (0.50) (0.57) (0.03) (0.15) (0.28) (0.31) Sanitary service available" 0.59 0.54 0.55 0.57 0.28 0.24 0.24 0.24 (1.84)* (1.68)* (1.73)* (1.78)* (1.25) (1.09) (1.05) (1.07) Water available" -0.26 -0.24 -0.24 -0.24 -0.39 -0.38 -0.36 -0.36 (0.50) (0.47) (0.48) (0.47) (1.10) (1.07) (1.01) (1.01) School variables Teacher-pupil ratio -27.39 -19.66 -28.18 -33.42 5.77 11.30 9.84 8.50 (1.16) (0.82) (1.17) (1.33) (0.35) (0.68) (0.59) (0.49) Sanitation or latrine available" 0.42 0.38 0.32 0.35 0.18 0.15 0.24 0.25 (0.52) (0.47) (0.39) (0.42) (0.31) (0.26) (0.42) (0.43) Electricity available" 0.16 0.34 0.95 1.09 0.27 0.39 0.57 0.60 (0.30) (0.60) (1.48) (1.64)' (0.69) (1.00) (1.27) (1.30) Piped water available" -0.19 -0.22 -0.13 -0.09 -0.21 -0.23 -0.19 -0.18 (0.37) (0.43) (0.24) : (018> (0.58) (0.64) (0.50) (0.47) Teacher and classroom variables Teacher finished university education" -0.57 -0.80 -0.66 -0.48 -0.15 -0.31 -0.06 -0.02 (1.33) (1.79)* (1.29) (0.87) (0.50) (1.00) (0.17) (0.04) Years of teacher experience 0.06 0.05 0.05 0.05 0.03 0.02 0.03 0.03 (1.35) (1.16) (1.15) (1.27) (1.07) (0.87) (1.08) (III) Monthly base salary of teacher (thousands of colones) -0.78 -0.72 -0.81 -0.87 -0.53 -0.49 -0.58 -0.59 (1.71)' (1.56) (1.70)» (1.80)* (1.67)' (1.52) (1.75)* (1.77) Teacher receives bonus" 0.53 0.51 0.53 0.55 0.49 0.48 0.46 0.47 (1.16) (1.12) (1.09) (1.13) (1.54) (1.51) (1.36) (1.37) (Table continues on the following page) Table 5. (continued) Mathematics Language Variable (V (2) 0) (4) (V (2) (V (4) All students have math textbook" 0.68 0.73 0.44 0.35 0.05 0.09 0.01 -0.01 (1.01) (1.08) (0.63) (0.49) (0.11) (0.18) (0.03) (0.02) Math textbook information missing" 2.29 1.90 -5.89 -8.42 0.40 0.13 -2.54 -3.19 (1.27) (1.05) (1.48) (1.64)- (0.32) (0.10) (0.92) (0.89) All students have language textbook" -0.75 -0.76 -0.95 -1.02 -0.16 -0.16 -0.25 -0.27 (1.14) (1.16) (1.42) (1.52) (0.34) (0.36) (0.55) (0.58) Language textbook information missing" -0.87 -0.54 7.82 10.65 0.28 0.52 3.SS 4.27 (0.47) (0.29) (1.90)* (1.94)» (0.22) (0.40) (1.24) (1.12) Teacher teaches in multigrade classroom" 0.75 0.93 0.68 0.55 0.49 0.62 0.53 0.50 (1.30) (1.58) (1.12) (0.88) (1.23) (1.51) (1.25) (1.14) Multigrade information missing" -0.01 -0.75 0.40 1.33 1.72 1.19 2.52 2.75 (0.00) (0.33) (0.16) (0.48) (1.10) (0.75) (1.45) (1.43) Number of books in classroom library (thousands of books) 0.91 0.90 0.70 0.63 1.54 1.53 1.47 1.45 (0.89) (0.87) . (0.68) (0.61) (2.14)** (2.13)" (2.05)" (2.01)" Classroom library information missing" 0.40 0.36 0.40 0.45 0.30 0.27 0.40 0.42 (0.73) (0.67) (0.69) (0.76) (0.79) (0.73) (0.99) (1.01) Community participation variable Number of parent association visits to classroom in past month 0.14 -0.11 0.10 -0.03 (1.72)* (0.78) (1.77)* (0.29) Inverse Mills ratio -0.58 -0.20 -0.05 -0.12 -1.09 -0.82 -0.73 -0.75 (0.68) (0.22) (0.06) (0.13) (1.84)* (1.34) (120) (1.22) Constant 3.97 3.45 4.13 4.47 1.12 0.74 0.83 0.92 (1.71)* (1.47) (1.75)* (1.87)* (0.69) (0.45) (0.51) (0.55) Number of observations 605 605 605 605 605 605 605 605 Number of municipalities 90 90 90 90 90 90 90 90 Overall R-' 0.0153 0.0129 0.0173 0.0175 0.0186 0.0010 0.0108 0.0077 * Significant at the 10 percent level. ** Significant at the 5 percent level. *•* Significant at the 1 percent level. Note: Dependent variable is mathematics or language test score, t-statistics are in parentheses. 436 THE WORLD BANK ECONOMIC REVIEW. VOL 13, NO. i Children with more siblings perform worse on both math and language tests, although the coefficients are not statistically significant. This result may indicate that parents devote less time to their children's individual needs. Older children do better in math than younger ones, even though they are in the same grade. However, age does not matter in determining language scores. The EDUCO effect can be mediated through school and classroom-level indica- tors and through the intensive involvement of parent associations. To capture these effects, we include school and classroom-level characteristics and a com- munity participation variable in the regressions.19 The EDUCO effect is less than that in regressions without school-level variables, indicating that a significant portion of the difference between EDUCO and traditional schools can be captured by differences in observable school characteristics and differences in community involvement (see table 4). The EDUCO coefficient, however, is still statistically insignificant. The basic results for the effects of socioeconomic characteristics do not change. Most of the school-level variables are not significantly different from zero. The two exceptions are teachers' base salary, which has a negative coefficient, and the availability of a classroom library, which is positively related to language achievement scores. EDUCO teachers receive a piece-wage rate, which the ACEs determine, while teachers in traditional schools have a fixed-wage scheme (World Bank 1995). The results for teachers' base salary might capture the inefficiency of fixed wage schemes in traditional schools. The positive effect on language scores of having a classroom library is also consistent with past evaluation of EDUCO (World Bank 1995: 19-20). It may be that classroom libraries help teach- ers to complete their lesson plans and stimulate students' interest and reading habits, both of which improve language scores. Most notably, the community participation variable has a positive and statis- tically significant coefficient for the basic specification with the EDUCO dummy variable (see column 2 of table 5). This finding indicates that the intensity of community involvement is significantly related to students' academic achieve- ments. Community participation might have a positive peer effect or work to monitor teachers. IV. EMPIRICAL RESULTS: STUDENT ABSENCE Parents' negative perceptions of education are an important issue, and this is true not only in El Salvador. In many rural areas throughout the developing world uneducated parents underestimate the value of education and thus do not send their children to school. The reasons may be cultural, social, or economic. The main problem, however, seems to be that parents are given poor incentives or poor information. Teacher absenteeism is also a chronic problem in the public schools 19. We enter them linearly and interact them with the EDUCO dummy, since EDUCO may change school characteristics. We do not report the regressions with the interaction terms here. They are available on request. Jimenez and Sawada 437 of many developing countries. Although excuses are sometimes legitimate, such as sickness, more often teachers are simply derelict. When teachers are absent, classes are usually canceled, since there is no tradition of using substitute teachers. Our hypothesis is that in a decentralized setting parents are better able and motivated to send children to school and to monitor teacher behavior. In fact, parents are more likely to send their children to school if they attend an EDUCO school, and teacher absenteeism is less prevalent in EDUCO schools, further reduc- ing student absence (World Bank 1997). The dependent variable in our regressions is the number of school days that the child missed in the past month. To control for student absence because of health problems, we add an additional indicator variable that equals 1 if a child suffered from a respiratory illness or flu in the past two weeks and 0 otherwise. The princi- pal result is that the coefficients on the EDUCO dummy variable are consistently negative and statistically significant, especially if we control for participation bias (table 6). A student in an EDUCO school is less likely to be absent even after we hold constant household, school, and participation characteristics.20 An important finding is that mother's education has a negative and statisti- cally significant coefficient in all specifications. Having an educated mother re- duces student absence approximately 70 percent. This implies that the mother's positive perception of education drawn from her own experiences contributes to better child attendance. We also differentiate the EDUCO dummy variables by year. The dummy vari- able for 1991-94 is insignificant—a third-grade student attending an EDUCO school during those years is just as likely to be absent as a student from a traditional school (table 6). However, the EDUCO coefficients are negative and significant for schools that were built in 1996 in the regressions with the participation correc- tion. The EDUCO effect is stronger for newer schools, which is consistent with the Hawthorne effect described earlier. Thus the negative effect of a decay in early enthusiasm of staff and students seems to dominate the positive effect of experi- ence on performance. The coefficient of the participation correction term is positive, although not significant. This means that unobserved characteristics of EDUCO students make them more likely to miss school. V. CONCLUSIONS El Salvador's EDUCO program has been remarkably successful in expanding educational opportunities for the poor living in rural areas. Decentralization has 20. Teachers are absent more days in traditional schools (1.4 days a month) than in EDUCO schools (1.16 days a month), implying that parent association monitoring works. This finding contrasts with that of Reimers (1997), who conducted interviews in 140 schools in 1993. That study was done on a earlier vintage than the schools in this study and did not correct for other variables that may affect outcomes. Sawada (1998) also extensively investigates the transmission mechanism from community participation to better educational outcomes in EDUCO schools by estimating teacher compensation and effort functions, as well as input demand functions. 438 THE WORLD BANK ECONOMIC REVIEW. VOL. 13. NO. 3 Table 6. Municipality Fixed-Effects Regressions of Student Absence Variable (1) (2) (3) (4) EDUCO variables EDUCO school present" -3.01 -3.93 (1.79)* (1.91)* EDUCO built in 1 9 9 1 - 9 4 ' -3.11 -3.55 (1.53) (1.41) EDUCO built in 1995" -2.64 -0.93 (1.32) (0.21) EDUCO built in 1996 3 -3.20 -423 (1.76)" (1.84)* Year missing 3 -3.89 -5.89 (1.69)* (1.55) Child and household variables Gender (female = 1) -0.08 -0.08 -0.03 -0.03 (0.29) (0.30) (0.10) (0.13) Child's age 0.17 0.17 0.16 0.17 (2.10)** (2.07)** (1.98)** (2.04)** Child lives without parent(s)* -0.12 -0.12 -0.16 -0.21 (0.27) (0.28) (0.36) (0.46) Child had respiratory illness or flu in past two weeks3 -0.29 -0.29 -0.26 -0.26 (1.02) (1.02) (0.89) (0.90) Number of siblings (ages 4-15) -0.11 -0.11 -0.08 -0.08 (1.23) (1.23) (0.86) (0.86) Mother began basic education1 -0.70 -0.68 -0.67 -0.66 (2.24) • • (2.17)** (2.13)** (2.09)* * Mother's education missing3 -0.23 -0.22 0.13 0.16 (0.44) (0.40) (0.23) (0.28) Father began basic education3 0.41 0.41 0.36 0.35 (1.38) (1.38) (1.21) (1.18) Father's education missing3 -0.58 -0.61 -0.68 -0.75 (0.79) (0.81) (0.91) (0.99) Own house3 -0.17 -0.20 -0.24 -0.26 (0.53) (0.60) (0.74) (0.79) Electricity available3 -0.39 -0.39 -0.75 -0.79 (1.07) (1.08) (1.93)* (2.02)* * Sanitary service available3 -0.20 -0.22 -0.29 -0.29 (0.52) (0.56) (0.71) (0.72) Water available3 -0.58 -0.57 -0.68 -0.68 (0.93) (0.90) (1.06) (1.06) School variables Teacher-pupil ratio 27.18 23.95 (0.90) (0.76) Sanitation or latrine available3 -1.87 -1.79 (1.84)** (1.73)* Electricity available3 1.38 1.73 (1.95)** (2.06)** Piped water available3 0.09 0.25 (0.13) (0.37) Jimenez and Sawada 439 Table 6. (continued) Variable (1) (2) (3) (4) Teacher and classroom variables Teacher finished university education3 0.63 1.03 (1-13) (1.48) Years of teacher experience 0.001 0.004 (0.03) (0.08) Monthly base salary of teacher (thousands of colones) nes) 1.08 0.90 (1.88)* (1-47) Teacher receives bonus" -0.17 -0.29 (0.29) (0.48) All students have math textbook11 1.31 1.09 (1.54) (1.23) Math textbook information missing3 1.50 -3.55 (0.66) (0.55) All students have language textbook3 -0.73 -0.82 (0.88) (0.96) Language textbook information missing3 -0.23 5.21 (0.10) (0.76) Teacher teaches in multigrade classroom3 0.88 0.60 (1-19) (0.76) Multigrade information missing3 -2.93 -0.98 (1.02) (0.28) Number of books in classroom library (thousands of books) -0.24 -0.30 (0.19) (0.23) Classroom library information missing3 -0.50 -0.26 (0.74) (0.35) Community participation variable 0.07 -0.05 Number of parent association visits to classroom (0.74) (0.26) 0.96 1.02 1.48 1.59 Inverse Mills ratio (0.93) (0.98) (1.34) (1.43) 1.62 1.68 -2.30 -1.89 Constant (1.55) (1.59) (0.78) (0.63) Number of observations 605 605 605 605 Number of municipalities 90 90 90 90 R1 0.0022 0.0024 0.0067 0.0074 * Significant at the 10 percent level. ** Significant at the 5 percent level. *** Significant at the 1 percent level. Note: Dependent variable is the number of school days student missed in past month, (-statistics are in parentheses. a. Binary variable equals 1 if response is "yes," 0 otherwise. also been instrumental in getting families and communities more involved in their children's schooling. But has the program delivered more? This article has evalu- ated EDUCO by determining if it has raised achievement scores and lowered stu- dent absence. The average scores of EDUCO students on standardized mathematics and lan- guage tests are lower than those of students attending traditional schools. This is 440 THE WORLD BANK ECONOMIC REVIEW. VOL. IJ. NO. 3 not surprising since EDUCO students come from disadvantaged backgrounds. What is interesting is that, after we control for background and correct for participa- tion bias in the samples, these differences disappear. In fact, the average perfor- mance of EDUCO students on language tests is slightly better than that of tradi- tional school students. The similarity in outcomes holds regardless of how long a school has participated in the EDUCO program, although newer schools do show an advantage that is not significantly different from zero. There is considerable variance in performance even after holding constant for type of school. The most important socioeconomic variables that have a positive effect on student achievement are being male, coming from a family with access to sanitary services, being older, and having fewer siblings. At the school level, the availability of a classroom library has a positive effect on achievement. Most important, coefficients of the parent participation variable are positive and sta- tistically significant in the two specifications that include an EDUCO dummy vari- able. Test results are significantly and positively related to the number of visits by ACES or their equivalent. The number of days students miss is negatively related to intangible EDUCO effects. In a decentralized setting parents are more motivated to send children to school and better able to monitor teachers. Parents' education, especially mother's education, has a positive impact on student attendance. Moreover, teachers tend to miss fewer days when monitored by parent associations. This, in turn, implies that students miss fewer days. We conclude that the rapid expansion of rural education through EDUCO's decentralized mechanism has not lowered achievement levels in El Salvador, even in the most needy parts of the country. In fact, it has improved language scores. The important transmission mechanism is likely to be the community monitor- ing system. Parents' deeper involvement in their children's education may result in improved attendance and pressure providers to deliver observable inputs. Al- though teachers, parents, and parent associations are not given direct incentives to raise standardized test scores in mathematics and language, the EDUCO pro- gram has had an overall positive impact. Thus parents' participation in school- based management seems to be an appropriate way of improving education in poor communities. REFERENCES The word "processed" describes informally reproduced works that may not be com- monly available through library systems. Ainsworth, Martha. 1992. "Economic Aspects of Child Fostering in Cote d'lvoire." Liv- ing Standard Measurement Study Working Paper 92. World Bank, Washington, D.C. Processed. Cox, Donald, and Emmanuel Jimenez. 1991. "The Relative Effectiveness of Private and Public Schools: Evidence from Two Developing Countries." Journal of Development Economics 34(1 ):99-121. Jimenez and Sawada 441 El Salvador, MINED (Ministerio de Education). 1995. EDUCO Learns and Teaches. San Salvador: Algier's Impresores. . 1997. "Informe de evaluation del rendimiento en 3o, 4o, 6o grado de education basica en lenguaje, matematica, estudios sociales, y ciencia, salud, y medio ambiente basado en la aplicacion national de pruebas de octubre de 1996." Direction National de Evaluation e Investigation, San Salvador, February. Processed. Glewwe, Paul, and Hanan Jacoby. 1994. "Student Achievement and Schooling Choice in Low-Income Countries: Evidence from Ghana." journal of Human Resources 29(3):841-64. Greene, William H. 1997. Econometric Analysis. 3d ed. Upper Saddle River, New Jersey: Prentice Hall. Hanushek, Eric A. 1995. "Interpreting Recent Research on Schooling in Developing Coun- tries." The World Bank Research Observer 10(2):227-46. James, Estelle, Elizabeth King, and Ace Suryadi. 1996. "Finance, Management, and Costs of Public and Private Schools in Indonesia." Economics of Education Review 15(4):387-98. Jimenez, Emmanuel, and Vicente Paqueo. 1996. "Do Local Contributions Affect the Ef- ficiency of Public Primary Schools?" Economics of Education Review 15(4):377-86. Levin, Henry M. 1997. "Raising School Productivity: An X-efficiency Approach." Eco- nomics of Education Review 16(3):303-ll. Lockheed, Marlaine, and Adriaan Verspoor. 1991. Improving Education in Developing Countries. New York: Oxford University Press for the World Bank. Maddala, G. S. 1983. Limited-Dependent and Qualitative Variables in Econometrics. Cambridge, U.K.: Cambridge University Press. McMillan, Robert. 1999. "Parental Pressure and Competition: An Empirical Analysis of Public School Quality." Department of Economics, Stanford University, Stanford, Calif. Processed. Reimers, Fernando. 1997. "The Role of the Community in Expanding Educational Op- portunities: The EDUCO Schools in El Salvador." In James Lynch, Celia Modgil, and Sohan Modgil, eds., Education and Development: Tradition and Innovation, Equity and Excellence, vol. 2. London: Cassell. Sawada, Yasuyuki. 1998. "Community Participation, Teacher Effort, and Educational Outcomes: A Case Study of El Salvador's EDUCO Program." Department of Econom- ics, Stanford University, Stanford, Calif. Processed. Willis, Robert, and Sherwin Rosen. 1979. "Education and Self-Selection." Journal of Political Economy 87(5):57-36. World Bank. 1994. "El Salvador: Community Education Strategy: Decentralized School Management." Country Report 13502-ES. Washington, D.C. Processed. . 1995. "Staff Appraisal Report: El Salvador Basic Education Modernization Project." IBRD Report 14129-ES. Washington, D.C. Processed. . 1996. "Impact Evaluation of Education Projects Involving Decentralization and Privatization." Working Paper on Impact Evaluation of Education Reforms 0. El Sal- vador Evaluation Team, Washington, D.C. Processed. . 1997. "El Salvador's EDUCO Program: A First Report on Parents' Participation in School-based Management." Working Paper on Impact Evaluation of Education Re- forms 4. El Salvador Evaluation Team, Washington, D.C. Processed.