79471 AUTHOR ACCEPTED MANUSCRIPT FINAL PUBLICATION INFORMATION The Effects of School-based Management in the Philippines : An Initial Assessment Using Administrative Data The definitive version of the text was subsequently published in Journal of Development Effectiveness, 4(2), 2012-06-19 Published by Taylor and Francis THE FINAL PUBLISHED VERSION OF THIS ARTICLE IS AVAILABLE ON THE PUBLISHER’S PLATFORM This Author Accepted Manuscript is copyrighted by the World Bank and published by Taylor and Francis. It is posted here by agreement between them. Changes resulting from the publishing process—such as editing, corrections, structural formatting, and other quality control mechanisms—may not be reflected in this version of the text. You may download, copy, and distribute this Author Accepted Manuscript for noncommercial purposes. Your license is limited by the following restrictions: (1) You may use this Author Accepted Manuscript for noncommercial purposes only under a CC BY-NC-ND 3.0 Unported license http://creativecommons.org/licenses/by-nc-nd/3.0/. (2) The integrity of the work and identification of the author, copyright owner, and publisher must be preserved in any copy. (3) You must attribute this Author Accepted Manuscript in the following format: This is an Author Accepted Manuscript of an Article by Khattri, Nidhi; Ling, Cristina; Jha, Shreyasi The Effects of School-based Management in the Philippines : An Initial Assessment Using Administrative Data © World Bank, published in the Journal of Development Effectiveness4(2) 2012-06-19 http://creativecommons.org/licenses/by-nc- nd/3.0/ © 2013 The World Bank The Effects of School-based Management in the Philippines: An Initial Assessment Using Administrative Data Nidhi Khattri, Senior Evaluation Officer, Independent Evaluation Group, The World Bank, Washington D.C. email: nkhattri@worldbank.org Cristina Ling, Evaluation Officer, World Bank Institute Evaluation Group, The World Bank, Washington D.C. email: cling@worldbank.org Shreyasi Jha,Monitoring and Evaluation Specialist, UN Women S. Asia Sub-regional Office, New Delhi, India. Email: shreyasi.jha@unwomen.org We would like to thank the Department of Education of the Philippines and Professor H. Nuqui, University of the Philippines, for providing the data and information on SBM. The authors also gratefully acknowledge the support of Dingyong Hou (Senior Operations Officer, EAPHD, and Task Team Leader for the Third Elementary Education Project), Lynette Dela Cruz Perez (Senior Operations Officer, EASHE), Christopher Thomas (Manager, EAPHD), and Eduardo Velez Bustillo (EASHE). This work was partially supported by the Education Program Development Fund. We are indebted our reviewers for providing helpful comments and suggestions on previous drafts. All remaining errors are ours alone. The opinions and conclusions expressed in this paper do not necessarily reflect those of the Executive Directors and officials of the World Bank or of its member governments. 1 Although some other school districts in the Philippines had ongoing SBM initiatives, they are not included in this analysis as insufficient information is available about the nature and timing of those initiatives. Also, while this report focuses on TEEP, it is important to note previous projects that implemented SBM such as the Secondary Education Development and Improvement Project and Basic Education assistance in Mindanao in 2001. This paper estimates the effect of school-based management on student performance in the Philippines using the administrative dataset of all public schools in 23 school districts over a 3-year period, 2003- 2005. The authors test whether schools that received early school-based management interventions (training in school-based management and direct funding for school-based reforms, based on school improvement plans) attained higher average test scores than those that did not receive such inputs. The analysis uses school-level overall composite test scores (comprising all subject areas tested) and test scores in three separate subject areas: English, math, and science. Their preferred estimator, difference-in- difference with propensity score matching, shows that the average treatment effect of participation in school-based management was higher by 1.5 percentage points for overall composite scores, 1.2 percentage points for math scores, 1.4 percentage points for English scores, and 1.8 percentage points for science scores. These results suggest that the introduction of school-based management had a statistically significant, albeit small, overall positive effect on average school-level test scores in 23 school districts in the Philippines. The paper provides a first glimpse of the potential for school-based management in a Southeast Asian context based on available administrative data. The authors suggest that the next order of research is to answer policy-related questions regarding the reforms: what aspects of the reform lead to desired results; are there differential effects across subpopulations; and what are the potential downsides to the reforms? The authors recommend that countries embarking on implementation of school-based management reforms specify their school-based management model and theories of change clearly and advance mechanisms for rigorous evaluations simultaneously. Such evaluations should not only provide more accurate estimates of the effectiveness of the reforms, but also help answer policy-related questions regarding design and implementation of those reforms in different socio-cultural contexts. Keywords: school-based management, decentralization, evaluation, propensity score matching, administrative data 1. Introduction Decentralization is a key feature of institutional reform throughout the world (King and Ozler, 1998). The main argument underpinning decentralization policies is that they empower people to be part of the local decision-making process –they improve government performance by alleviating information asymmetries and costs by bringing decision-making closer to the people concerned. However, decentralization can also worsen the provision of public goods in the presence of externalities, lack of technical capabilities by local governments, or capture of lower-level administration by local elites (Galiani and Schargrodsky, 2001). In the context of the education sector, decentralization typically includes one or more of the following features: decentralized revenue generation, curriculum design, school administration, and teacher hiring and management. Decision-making authority for these types of functions is devolved to regional/municipal governments or to schools themselves. The policy of allowing schools autonomy in decisions in these areas is referred to as school-based management (SBM), school based governance, or school self management (De Grauwe, 2004; Gertler, Patrinos, and Codina, 2006). Responsibility and decision-making over different types of school operations are transferred to individuals at the school level, who in turn must conform within a set of centrally or state-level determined policies. The popularity of SBM is evidenced by the large number of development agencies promoting it as a key component of the decentralization reforms and the growing number of countries that have adopted aspects 1 of this approach. SBM reforms began in the 1970s in Australia. Since then, a wide range of countries have experimented with or introduced SBM in all regions of the world, including Hong Kong (China), Indonesia, El Salvador, Nicaragua, Kenya, Kyrgyz Republic, Nepal, Paraguay and Mexico (World Bank 2007). Nevertheless, the impact of SBM on education quality, including student outcomes, remains a contentious issue, with some researchers arguing that SBM leads to enhanced educational outcomes (e.g., Gertler, Patrinos, and Codina, 2006), while others contending that SBM leads to the deterioration of educational quality especially among the weakest schools (Bardhan, 2002). The range of SBM approaches and the contexts in which they are implemented makes the debate about SBM quality an intricate one. The evaluation of SBM is complicated by the diversity of approaches to and elements of decentralization that collectively constitute “SBM” and by the institutional and socio-cultural contexts in which they are implemented. Nonetheless, some studies in recent years have found that SBM reforms are associated with improved education outcomes and processes (e.g., Skoufias & Shapiro, 2006; Sawada & Ragatz, 2005; Gunnarsson et al., 2004; Eskeland & Filmer, 2002). However, rigorous evidence base for the effectiveness of SBM in boosting student performance is thin. A recent review of the empirical literature on SBM since 1995 indicates that only 14 studies utilized rigorous methods to assess the impact of SBM, and only six reported positive impacts on students’ test scores (Barrera-Osorio et al., 2009) . Eleven studies are country-specific from Latin America, one from Kenya, and two exploit data from multiple countries. No empirical evidence is available from East Asia. This paper contributes to the small but growing empirical literature on the effectiveness of SBM by extending the research to Southeast Asia (see Yap and Adorio (2008) for a qualitative study). The paper provides an initial analysis of the potential of school-based management for improving educational outcomes in the Philippines, using aggregated school-level test scores and administrative data. In this 2 analysis, SBM is narrowly defined (based on the components of the Third Elementary Education Project) as the provision of training for school leaders to create school improvement plans (SIP) and annual implementation plans (AIP) and a small amount of funding for schools to use towards implementing the SIP. The data are quite limited and do not allow for a thorough analysis of the processes and approaches through which SBM reforms affect outcomes. Nonetheless, the paper attempts to use existing data to answer an initial question: did the introduction of SBM lead, on average, to enhanced educational outcomes? It also demonstrates how administrative data can be mined for exploratory assessments of potentially larger programs. The remainder of the paper is organized as follows: the next section provides a brief overview of the SBM program in the Philippines; section 3 outlines the methodological approaches used for the analysis; section 4 describes the data; section 5 discusses the empirical results; and section 6 provides a summary of our conclusions, discusses the limitations of the study, and highlights the implications for future, more rigorous SBM evaluations in the Philippines. 2. SBM Program in the Philippines SBM was implemented in between 2003 and 2005 in 23 districts participating in the Third Elementary Education Project (TEEP) supported by the World Bank.i The project provided funding for school infrastructure, training, curriculum development, and textbooks. SBM was introduced as an integrating framework for obtaining school-level project inputs and building school capacity for education planning and program implementation beginning in school year 2003-04. Schools participating in SBM were required to design a five-year School Improvement Plan (SIP) in partnership with parents and the community using data such as student achievement and students’ learning needs assessments, with the school principal or head teacher leading the process. Based on the SIP, schools developed an Annual 3 Implementation Plan (AIP) at the beginning of the school year and a report card to be shared with the community at the end of the school year. Project inputs for infrastructure, training, textbooks, and so forth, were partially based on the SIP. Principals and head teachers received training in leading the development and implementation of the SIP and the AIPs in collaboration with teachers and key members of the larger community. SBM schools also received funds for maintenance and operating expenses directly in cash rather than in kind, as had been the case previously. These cash funds could be used by the schools based on their AIP. The cash allocation was based on a formula that provided each school with a flat amount of funds plus a prorated figure based on the number of student and teachers as well as other criteria, such as percentage of indigenous student population in the school. Schools not participating in the SBM received no SBM-related training and no cash funds, and they were not required to develop SIPs and AIPs. The SBM program was designed to improve student outcomes through two main venues: by empowering the school community to identify education priorities and to allocate the school maintenance and operating budgets to those priorities (such as curriculum enrichment programs); and by enhancing transparency and accountability through the annual implementation plans and school report cards.ii However, the SBM program articulated no explicit assumptions regarding the timeframe within which improvements in student achievement were expected to take place. Systematic data on the level of uptake and implementation of the key features of the reforms are also not available. The SBM training, funds, and requirements, such as the development of the SIP and AIP, were rolled out in three batches and eventually covered almost all (84 percent) of the 8,613 schools in the 23 project districts. The first batch comprised 1,666 schools in 2003-2004, largely because they were perceived to be more capable, although no explicit assignment mechanism was designed.iii The next batch of 2,700 schools was targeted for SBM rollout in 2004-2005, and another batch of 1,529 was included in 2005- 2006. (All other TEEP inputs were implemented at the same time across all districts.) 4 3. Analytic Approach We assess the effects of SBM on student performance using average school-level test scores from all schools and school-level indicators between 2003 and 2005 from the 23 TEEP districts in the Philippines. We base our analysis on a pipeline comparison strategy. Since we have student achievement data over a three year period, but schools were inducted into the program in three batches over time, we identify the treatment group as the first batch of schools that received SBM funds in 2003-2004. The treatment group consists of the schools that did not receive SBM funds and training in 2003-2004. Therefore, for this study, we operationalized SBM batch one schools as the treatment group (n=1666) and all other schools (batches that received the SBM intervention later) as the control group (n=3501).iv In our analysis, the treatment group (batch one) therefore had exposure to SBM for a period of two years, 2003-04 and 2004- 05. Because selection into SBM was not voluntary, there is likely to be placement bias in our control and treatment groups. We therefore use two non-experimental evaluation techniques to estimate program impact: difference-in-difference estimators and difference-in-difference with propensity score matching estimators. Using nonparametric kernel matching techniques, we correct for potential sources of bias. Specifically, we tested whether the composite test scores and test scores individually in three subject areas (math, science and English) were higher after SBM was introduced for schools in the treatment group compared to those in the control group. 4. Data and Statistical Specification 5 We use school administrative data from all 5,167 schools in the 23 TEEP districts to examine the effect of SBM program on student test scores. The dataset includes information on school personnel, students, facilities, and enrollment and dropouts for all schools in the district over a 3-year period, 2002-2003 to 2004-2005. The data were collected by the Philippines Department of Education as part of the management of the Third Elementary Education Project (TEEP) of which SBM was one component. Student achievement is measured using National Achievement Tests conducted in 2002-03 (year one – pre-intervention) in Grade 4, 2003-04 (year two) in Grade 5, and 2004-05 (year three, post-intervention) in Grade 6 in English, mathematics, science. However, only the school-level Mean Percentile Score (MPS) data are available in the dataset, limiting the analysis to the school level. We utilize the composite MPS computation (based on tests in all subjects), as well as the English, mathematics, and science MPS for 2002-03 (pre-intervention) and 2004-05 (post-intervention). Our method of analysis consists of two quasi-experimental evaluation techniques. We have data on test scores over three years for all schools. We also have time-invariant data on school characteristics. Once we determined the control and treatment groups as described in the preceding section, the preferred estimator to use in this context was the difference-in-differences (DID) estimator. However, in order for the DID estimator to provide credible estimates of the program impact, it should be free of any biases inherited by pre-existing (pre-intervention) differences between treatment and control groups. Because of problems incurred with using the DID estimator due to way in which SBM program was rolled out (i.e., no clear assignment mechanism), we also estimate and discuss DID in conjunction with propensity score matching (PSM) to estimate average treatment effects. (i) Difference-in-differences (DID) estimation The ideal experiment for identifying the effects of SBM on student test scores would have been to randomly assign schools to control and treatment groups. The SBM effect could then have been estimated 6 by comparing schools that received SBM with their peers in the comparison group. Unfortunately, such a design is not feasible in this case because of the nature of the SBM program rollout. Therefore, we assigned the set of schools that received early intervention in 2003 as the treatment group. The control group comprised all other schools that did not receive SBM support in 2003 – i.e., schools that received the intervention later. The SBM effect on student test scores can be evaluated as follows: SBM effect = (TestScores SBM .POST  TestscoresSBM , PRE )  (Testscores NONSBM, POST  TestScores NONSBM.PRE ) --------------------(eq. 1) Converting equation (1) above to an estimating model, we get: Ys ,t   0  1Yeart   2 SBM s   3 (Yeart * SBM s )   4 Z ts   p   ts -------------------(eq. 2) where Y is the outcome measure. We include school-level composite test scores as well as test scores in math, science, and English as outcome measures. Year is a dichotomous variable that equals 1 for post- intervention and equals 0 for pre-intervention. SBM is a dichotomous variable that is equal to 1 indicating that a school received SBM support during all of the treatment years (t =2002-2004). Z is a vector of school characteristics.  p is district-specific fixed effects intended to capture district-specific aggregate fixed effects correlated with schooling outcomes (demographic trends or changes in government, for example).  t , s is the school level error term that includes all the unobserved school characteristics that we assumed are uncorrelated with the explanatory variables for the time being.v The vector of school characteristics includes total enrollment, total dropout, student-teacher ratio, textbooks per student, teacher manual per teacher, school type, school head type, whether the school was an elementary leader school (a school considered to be a leader among others in the district) and whether the school had received SIIF funds.vi School type consists of four categories: complete combination and multi-grade schools; complete large and medium school; complete small schools; and primary schools. 7 School head type consisted of three categories: head principal, head teacher, and teacher in charge. Table I summarizes the descriptive statistics for all schools and by treatment status of the school.  0 is the intercept showing the average test scores of the SBM and non-SBM schools;  1 shows the change between 2002-03 and 2004-05;  2 is the difference in the treatment and control group; and  3 is the DID estimator capturing the differences between the control and treatment groups, before and after the intervention. More specifically, it measures the SBM effect as the change in average school-level test scores among two groups before and after the 2003 SBM intervention. These groups included: the “first batch,” or the schools that received SBM first (treatment group) to schools that received the intervention later and those that did not receive the intervention at all (comparison group). In order to obtain the DID estimates, we created a panel dataset from our original dataset of 5,167 schools. Since we had test scores for all three years, we simply replicated this data for every year to obtain a panel dataset with 15,501 observations. In this dataset, each school is observed in each of the three years (2002-03 to 2004-05). In order to use the DID estimates in this context, we need to be certain that the difference in post- intervention outcomes between SBM and non-SBM schools would have been identical in the absence of the intervention. However, this assumption is impossible to test because we do not actually observe the counterfactual. We can nonetheless test whether pre-intervention and “mid-term” educational outcomes under study were similar between the treatment group and the proposed control group. If the pre- intervention and “mid-term” (past baseline and probably without implemented intervention) outcome measures were not significantly different between treatment and control schools, there is no compelling reason to believe they would be significantly different in the post-intervention periods had the SBM program not been put in place. However, if we observe changes between pre-intervention and mid-term, 8 we can consider the possibility that the two groups would have diverged with respect to their outcomes even in the absence of the SBM intervention. We can test this assumption for the proposed treatment and control groups by running the following equation on the pre-SBM data (i.e. 2002-2003): Yst   0  1 (Year20020 / 20031 )   2 (Year20020 / 20031 * SBM 1 / 0 )   p   ts -------------------(eq. 3) Y is the outcome measure including: school level composite test scores and test scores individually in math, science, and English. Year is a dummy variable that takes value = 1, if year =2003-04 and =0 if year=2002-03 (both pre-intervention and implementation years). SBM is a dichotomous variable that is equal to 1 if the school s received SBM support (SBM=1). As in eq. (2),  p is intended to capture district- specific aggregate fixed effects and  t , s is the error term. We are interested in two issues: first, was there a pre-existing difference between SBM and non-SBM schools before the intervention; and two, did the difference increase over time (between 2002-03 and 2003-04). The coefficient for the interaction term can be interpreted as follows: if the coefficient  2 =0, then the pre-SBM outcomes for schools that would eventually receive SBM funds are not significantly different from those in the control group (i.e., schools that did not receive SBM funds in 2003). We run equation (3) separately for years 2002-03 and 2003-04. When year=2002-03 (pre-intervention), we do not have a coefficient for  1 . Table II reports the results from equation (3). For 2002, we report 2 coefficient which shows the difference between SBM and non-SBM schools at the baseline (2002-03). 2 For 2003-04 (mid-term), we report which shows the increase in test scores of SBM schools compared to non-SBM schools in 2003-04 compared to the baseline (2002-03). For both years Model A is a simple Model B includes district fixed effects. For overall test scores, Model B in Table II shows that OLS and 2 =2.0, implying that overall test scores for schools that would eventually receive SBM funds in 2003 9 were about 2 percentage points higher than non-SBM schools in 2002-03. In 2003-04 ( mid-term), the difference in overall test scores between SBM and non-SBM schools was about 3.0, implying that overall test scores for SBM schools 3 percentage points higher than non-SBM schools in 2003 compared to 2002. Table II shows there are significant differences in the pre-intervention period in test scores (math, science, English and overall) across schools that would eventually receive SBM and those that did not receive SBM in 2003. Based on these results, it appears that program placement bias is a serious concern in this context, and that difference-in-difference estimates may not give us unbiased estimates of the program effect. Therefore, in addition to difference-in-difference estimates, we also obtain propensity score matching estimates as a means of achieving unbiased identification of the SBM effect. (ii) Propensity score matching (PSM) estimation One of the potentially most serious problems in any evaluation study is the occurrence of placement bias. In this case, the problem arises because we would like to know students’ test scores with and without the SBM intervention. The results in Table II indicate that placement bias is a serious concern in this case because SBM and non-SBM schools are substantially different even in the absence of the treatment. The matching approach is one possible solution to the selection problem. This method originated from the statistical literature and shows a close link to the experimental context.vii The basic idea is to identify within a large group of non-participants those individuals who are similar to the participants in all relevant pre-treatment characteristics X. The assumption is that once this is done, differences in outcomes between the matched control group and the treatment participants can be attributed to the program. Since conditioning on all relevant covariates is limited in case of a high dimensional vector X (`curse of dimensionality'), Rosenbaum and Rubin (1983) suggest the use of so-called balancing scores b(X), i.e., functions of the relevant observed covariates X such that the conditional distribution of X given b(X) is independent of assignment into treatment. One possible balancing score is the propensity score, i.e., the 10 probability of participating in a program given observed characteristics X. Matching procedures based on this balancing score are known as propensity score matching (PSM) and we will use PSM matching in conjunction with DID as the preferred estimator in this paper. PSM proposes that all pre-treatment characteristics of a unit of observation is summarized into a single index (the propensity scores) and the units of observations are then matched on their propensity scores. According to Skoufias and Shapiro (2006), the propensity score is defined as the conditional probability of receiving a treatment (receiving SBM) given pre-treatment characteristics X, i.e. p(X) ≡ Pr (D = 1| X ) = E(D | X ) ------------------------------(eq. 4) where D is the indicator for receiving treatment (=1 if SBM, =0 if non-SBM). As shown by Rosenbaum and Rubin (1983), if the propensity score p(X) is known, then the PSM estimator for ATT (`average treatment effect on the treated’) can be estimated as follows: PSM TATT  E P ( X / D1) {E[Yi / D  1, p( X )]  E[Y0 / D  0, p( X )]} ------------------------------(eq.5) The potential outcomes are defined as Yi(Di) for each individual i, where i = 1….N and N denotes the total population. To put eq. 5 in words, the PSM estimator is simply the mean difference in outcomes over the common support, appropriately weighted by the propensity score distribution of participants. When estimating the propensity scores, two choices must be made. The first one concerns the model to be used for the estimation, and the second one requires identifying the variables to be included in the model. We use a logit model to estimate propensity scores. The variables in the model were selected to satisfy the conditional independence assumption (CIA). The matching strategy builds on the CIA, requiring that the outcome variable(s) must be independent of treatment, conditional on the propensity score. Hence, implementing matching requires choosing a set of variables X that credibly satisfy this condition. Only variables that influence simultaneously the participation decision and 11 the outcome variable should be included. We included five explanatory variables in the logit model: student-teacher ratio; size of school (total enrollment in 2002 base year); school type (incomplete vs. complete schools & combination/multigrade schools vs. mono schools); type of school head (principal vs. teachers and teacher-in-charge). The results from the logit model are presented in Table III. We impose the common support area condition to ensure that any combination of characteristics observed in the treatment schools can also be observed among the control group.viii We then apply the nonparametric kernel matching technique to match schools based on their propensity scores within school districts. We matched schools at the district level because the SBM intervention took place within school districts. Kernel matching uses weighted averages of all individual cases in the control group to construct the counterfactual outcome. Thus, one major advantage of these approaches is the lower variance which is achieved because more information is used. Since we do not condition on all covariates but on the propensity score, it has to be checked if the matching procedure is able to balance the distribution of the relevant variables in both the control and treatment group. Several procedures have been suggested to check the matching quality. The basic idea behind these approaches is to compare the situation before and after matching, and verify whether any differences remain after conditioning on the propensity score. If differences exist, matching on the score was not (completely) successful and remedial measures must be taken. We use a stratification test proposed by Dehejia and Wahba (2002) that divides observations into strata based on the estimated propensity score, such that no statistically significant difference remains between the estimated propensity score means of the treatment and control groups. We then use t-tests within each stratum to test if the distribution of X-variables is the same between both groups. Table IV presents the absolute t-scores for the X-variables across SBM and non-SBM school by quintiles. We notice that across variables there are 12 no significant differences between SBM and non-SBM schools on most variables except total enrollment.ix Finally, given the availability of data over three years we use the differences-in-difference estimator (developed by Heckman et al. in 1997, 1998a,b) to calculate the average treatment effect (ATE). This estimator compares the before-after test scores of SBM schools with the corresponding before and after changes among non-SBM schools, conditional on covariates X. This builds on the simple differences-in- difference statistic by controlling for covariates X and estimating the differences using nonparametric methods. The advantages of this estimator are similar to the advantages of the DID regression – it eliminates any unobserved factors that vary between observations, although not over time. This matching estimator allows selection on unobservables as long as the unobservable factors do not change between observations over time. Given the aforementioned advantages of the simple regression-based DID estimator, the DID with matching is our preferred estimator of program impact. 5. Results Table V presents the simple means of school level test scores across SBM and non-SBM schools, before and after the intervention in 2003, without controlling for other explanatory variables. These figures do not control for any selection bias. Given that there is a strong possibility of selection bias, these estimates are helpful in showing trends and observing overall differences. However, we cannot make any claims of causality based on these comparisons. We examined test scores along four categories: math, English, science and overall (composite of all test scores). Comparing SBM schools over time, we find that SBM schools scored higher in 2004 (year 3) in all areas including: math by 16.57 percentage points, science by 11.74 percentage points, English scores by 18.57 13 percentage points and in overall scores by 16.16 percentage points. We find similar increases in test scores for non-SBM schools, although the increases are not as large as they are for SBM schools. Math scores increased by 15.46 percentage points, science scores increased by 10.37 percentage points, English scores increased by 16.66 percentage points and overall scores increased by 14.65 percentage points. Cross-sectionally, we find that SBM and non-SBM schools performed relatively similarly in 2002 (pre- intervention) (see Table V). In 2005, the SBM schools performed better than did non-SBM schools. For example, math scores between SBM and non-SBM schools increased to 1.58 percentage points, science scores were 2.03 percentage points higher, English scores were 2.20 percentage points higher and overall scores are 1.98 percentage points higher for SBM schools compared to non-SBM schools. A simple difference-in-difference comparison (without controlling for any factors) between SBM and non-SBM schools shows that test scores increased more rapidly among SBM schools. For instance, math scores increased by 1.12 percentage points for SBM schools, science scores increased by 1.36 percentage points, English scores by 1.91 percentage points and overall scores by 1.51 percentage points. Although these results point toward larger increase in test scores, we cannot make any causal inferences based on these estimates. Table VI presents the estimates of the average treatment effect derived from three multivariate analysis techniques. Since we are primarily interested in testing whether the introduction of SBM had an effect on test scores, we present only the coefficients associated with the variable measuring program impact. The detailed regression results are presented in the appendices. The first column presents the estimates of the average treatment effects obtained from running an ordinary least squares (OLS) model of equation (2) without any district level fixed effects. We find that there was a significant increase in overall test score (1.15 percentage points), as well as an increase in test 14 scores in specific subject areas: science (1.37 percentage points) and English (1.14 percentage points); the increase in math was not significant. Next, we ran equation (2) including district level fixed effects and find more or less similar results. We find that there was a statistically significant increase in overall combined test scores (1.06 percentage points) as well as an increase in science (1.28 percentage points) and English scores (1.05 percentage points). As with the OLS model, the increase in math scores for SBM schools compared to non-SBM schools was not significant. Finally, we present the results from DID estimators based on kernel matching. Generally, we find that the size of the impact was higher relative to DID without matching. Using kernel propensity score matching, we find that participating in the SBM program led to an increase of 1.45 percentage points in overall test score, 1.82 percentage points in science scores, and 1.32 percentage points in English scores. At 1.188 percentage points, the relative increase in math scores is also significant at the 10 percent level. Summing up the findings, we can conclude that SBM in the Philippines had a statistically significant impact on overall student test scores and test scores in English and science. The effect sizes were small to moderate: .10 for overall scores, .13 for science, .09 for English, and .07 for math.x These are not inconsistent with Borman’s (2002) analysis, which shows effect sizes of .17 to .14 after one and two years of implementation, respectively, and increasing with years of continued implementation. If we use the coefficients based on OLS and Fixed-Effects models, the effect sizes would be estimated to be lower. 6. Conclusions 15 This study of the SBM program in the Philippines shows that school-averaged student performance on national tests improved between 2002-03 and 2004-05 and that the level of improvement was higher for schools involved in SBM for two years compared with schools that had not yet received the intervention or received the intervention later. School-averaged student performance improved in math, science, and English and on the composite score. Improvement for schools that received SBM early was significantly higher in science and English and on composite test scores. The Philippines subsequently embarked on a wider effort to introduce and implement SBM. While this paper provides an early indication of the usefulness of SBM in a few districts in the Philippines and its promise for other parts of the country, it also highlights the importance of specifying the program intervention and its underlying theory of change clearly and of integrating a rigorous evaluation design into the rollout of a program. We discuss the limitations of the study and provide some recommendations for future evaluations of SBM. First although the study found significant differences between SBM and non-SBM schools on school- level outcomes, the possibility of unmeasured differences influencing outcomes exists. The findings in impact evaluations are critically dependent on the choice of the comparison group. In the context of this study, although a comparison group was identified, whether the group represented a good counterfactual is more difficult to argue. SBM funds were disbursed in batches and since we had data for all schools, we chose the first batch of SBM schools as the treatment school and created a counterfactual using available administrative data. Although our model controls for a number of important school-level factors such as student-teacher ratio, school size, school type, and school head type, it is quite possible that other unobservable variables affected the outcomes. For example, the model does not capture student-level data (e.g., socio-economic background), which literature shows to have a significant relationship to student achievement. School 16 personnel quality may similarly play a role. Furthermore, although we rely on the assumption that propensity-score matching would have accounted for some unmeasured differences, important differences cannot be ruled out. While quasi-experimental evaluation methods are an improvement over simple multivariate techniques, they are not a substitute for more rigorous experimental designs Second, the time frame captured in this study is very short, only two years of SBM preparation and implementation. In order to fully assess the impact of SBM, the assumptions regarding the pathways through which the reforms play out over time and eventually affect student achievement will need to be articulated and examined explicitly. The possibility of a shorter-term “hawthorne” effect cannot be eliminated, and, potentially, an evaluation with a longer time-frame would permit the identification of longer-term effects. Behrman and King (2008) highlight the risks of a poorly timed evaluation, ranging from finding partial or no impacts, when they in fact would take a longer time to materialize, to the risk of scaling up a poor program, by waiting too long to evaluate. Articulating clear assumptions regarding the reforms would assist with implementing evaluations in a timely fashion. Third, the study does not examine the processes through which SBM affected school performance, a critical aspect of understanding how programs affect outcome and for policy design. Barrera-Osorio et al.’s (2009) review of the SBM literature conceptualizes the range of SBM approaches along two dimensions: who has the power to make decisions and the degree of decision-making devolved to the school level. The review notes that “With so many possible combinations of these two dimensions, almost every SBM reform is unique.” (pg. 4). While some programs transfer authority to principals or teachers only, others encourage or mandate parental and community participation, often through school committees. Furthermore, the contexts in which the authority is devolved are likely to play a role in the eventual effects. For example, engaging parents and the wider community in school matters may not be easy in certain contexts – studies indicate that in communities with social and political tensions, the school board may be used as an instrument by the elite to orchestrate greater inequities (Cladwell, 2005). 17 Evidence from New Zealand and Australia demonstrates the under-representation of minority groups in the composition of school boards. Several previous papers have examined some components of, and the processes through which, SBM affects student achievement (e.g., Jimenez and Sawada (1998), Eskeland and Filmer (2002) Gunnarsson et al.(2004), Sawada and Ragatz (2005)). Available reports on TEEP indicate that several aspects of SBM may have been conducive to improved student performance: improved school management through intensive principal and head-teacher training, identification of school-level needs and allocation of resources to those specific needs, greater community attention to schools’ and students’concerns. These elements need to be examined within specific cultural contexts. Fourth, the study does not examine the distribution of effects across different types of schools, both in terms of specified outcomes – student achievement – as well as unintended effects, such as toll on the principal’s and teachers’ time in community engagement. The Implementation Completion Report for the TEEP project indicates that, indeed, community engagement demanded that principals and teachers spend considerable time on community relations in addition to their administrative and pedagogical responsibilities, a commitment that several were beginning to find burdensome thus raising questions about its longer-term viability. Thus, answers to questions regarding the conditions under which the different SBM models work and which implementation processes are effective are critically important from a policy and program design and implementation perspective. Additional research is necessary in order to address these issues and draw generalizable conclusions about SBM policy in the Philippines and more broadly about similar SBM programs in different contexts. Future evaluations should be designed where the theory of change is laid out and tested systematically 18 through an experimental approach. It is understandable however that in real world conditions an experimental design can be challenging to implement. While may be difficult to specify control groups that may never receive program benefits (or may not emulate it to some degree), it may be possible to implement a pipeline approach and roll out the program in batches, using randomization or a clear assignment mechanism over time, to enable “control groups” to be built in. In addition, an integration of qualitative data collection, with a focus specifically on processes of implementation and short-term effects, would be necessary to provide insights into how various components lead to various outcomes and to depict examples of the change processes in detail. Despite these limitations, this study intends to contribute to the growing body of studies that examine school based management as a tool for improving student outcomes. It also advances the use of existing data to examine, to the degree possible, program effectiveness. 19 REFERENCES Bardhan, P. 2002. Decentralization of Governance and Development. Journal of Economic Perspectives, 16(4): 185–205. Barrera-Osorio, F., T. Fasih, & Patrinos, H.A. 2009. Decentralized decision-making in school. The theory and evidence on school-based management. Washington, D.C. The World Bank. Behrman, J. R. and King, E. M. 2008, Program Impact and Variation in Duration of Exposure, in Samia Amin, Jishnu Das, and Markus Goldstein, eds, Are You Being Served: New Tools for Measuring Service Delivery, Washington, DC: World Bank, 147-172. Berlinski, S. and Galiani, S. 2007. The effect of a large expansion of pre-primary school facilities on preschool attendance and maternal employment. Labour Economics, 14(3), 665-680. Borman, G.D., Hewes, G. M., Overman L.T., Brown, S. 2003. Comprehensive school reform and achievement: A meta analysis. Review of Educational Research, 73(2), 125-230. Caldwell, B. J. 2005. School-based Management. Education Policy Series. 3. Paris, The International Institute for Educational Planning, and Brussels, The International Academy of Education, Paris and Brussels. Cohen, J. 1988. Statistical Power Analysis for the Behavioral Sciences. 2nd ed. Hillsdale, NJ: Erlbaum De Grauwe, A. 2004. School-based management (SBM): does it improve quality? Paper commissioned for the EFA Global Monitoring Report 2005, The Quality Imperative. Dehejia, R. H. and Wahba, S. 2002. Propensity Score-Matching Methods for Nonexperimental Causal Studies. The Review of Economics and Statistics, 84(1), 151-161. Eskeland, G. and Filmer, D. 2002. Does Decentralization Improve Learning? Autonomy and Parental Participation in Argentine Schools. The World Bank. Washington D.C. Galiani, S. and Schargrodsky, E. 2001. Evaluating the Impact of School Decentralization on Education Quality. Economica. Vol. 2(2): 275-314. Gertler, P., Patrinos, H. and Rubio-Codina, M. 2006. Empowering Parents to Improve Education: Evidence from Rural Mexico. Impact Evaluation Series No. 4, World Bank Policy Research Working Paper No. 3935. Gunnarsson, Louise V., Orazem, P., Sanchez, M., and Verdisco, A. 2004. Does School Decentralization Raise Student Outcomes? Theory and Evidence on the Roles of School Autonomy and Community Participation. Iowa Stata University Department of Economics Working Paper Series 04005. Heckman, J., Hidehiko I., and Todd, P. 1997. Matching as an Econometric Evaluation Estimator: Evidence from Evaluating a Job Training Programme. Review of Economic Studies 64(4): 605- 54. Heckman, J., Hidehiko I., Jeffrey S., and Todd, P. (1998a.) Characterizing Selection Bias Using Experimental Data. Econometrica 66(5):1017-1098. 20 Heckman, J., Hidehiko I., Jeffrey S., and Todd, P. 1996. Sources of Selection Bias in Evaluation Social Programs: An Interpretation of Conventional Measures and Evidence on the Effectiveness of Matching as a Program Evaluation Method. Proceedings of the National Academy of Sciences. Vol. 93 No. 23 13416-13420. Heckman, J., Hidehiko I., and Todd, P. 1998b. Matching as an Econometric Evaluation Estimator. Review of Economic Studies 65(2): 261-294. Jimenez, E. and Sawada, Y. 1998. Do community-managed schools work? An evaluation of El Salvador’s EDUCO program. The World Bank Development Economics Research Group Working paper No. 8, series on Impact Evaluation of Education Reforms. King, E. and Ozler, B. 1998. What decentralization got to do with learning? The case of Nicaragua’s school autonomy reform. The World Bank Development Economics Research Group Working paper No. 9, series on Impact Evaluation of Education Reforms Rosenbaum, P. R., and Rubin, D. B., 1983. The Central Role of the Propensity Score in Observational Studies for Causal Effects. Biometrika 70, 41–55. Sawada, Y. and Ragatz, A. 2005. Decentralization of Education, Teacher Behavior and Outcomes: The Case of El Salvador’s EDUCO Program. In: Emiliana Vegas, Ed., Incentives to Improve Teaching: Lessons from Latin America. Washington, DC. The World Bank. Skoufias, E. and Shapiro, J. 2006. Evaluating the Impact of Mexico’s Quality School Program: The Pitfalls of Using Nonexperimental data. Impact Evaluation Series No. 4, World Bank Policy Research Working Paper No. 4036. The World Bank. Washington D.C. World Bank. 2007. What is School Based Management? An Introduction. Education Human Development Network. The World Bank, Washington D.C. Yap, Ingrid R. and Adorio, M.P. 2008. School-Based Management: Promoting Special Education Programs in Local Schools. Education Quarterly 66(1), 50-70. 21 Table I: Descriptive Statistics, by All Schools, SBM, and Non-SBM All Schools SBM Schools Non-SBM School Variable N Mean SD N Mean SD N Mean SD Dependent Variables, Year= 2002 Math 5167 46.14 17.32 1666 46.45 17.25 3501 46 17.35 English 5167 43.09 15.11 1666 43.29 14.79 3501 43 15.25 Science 5167 44.69 13.53 1666 45.14 13.17 3501 44.48 13.7 Total 5167 44.65 14.06 1666 44.97 14.01 3501 44.5 14.09 Dependent Variables, Year= 2003 Math 5167 53.98 17.15 1666 54.95 16.99 3501 53.51 17.2 English 5167 50.01 15.11 1666 51.02 14.9 3501 49.54 15.19 Science 5167 49.86 13.35 1666 50.81 13.25 3501 49.41 13.38 Total 5167 51.28 14.26 1666 52.26 14.18 3501 50.82 14.28 Dependent Variables, Year= 2004 Math 5167 61.97 17.1 1666 63.06 16.54 3501 61.46 17.34 English 5167 55.51 14.02 1666 56.9 13.56 3501 54.85 14.18 Science 5167 60.38 14.88 1666 61.88 14.49 3501 59.66 15.01 Total 5167 59.79 13.83 1666 61.15 13.38 3501 59.14 14 Other School Characteristics Total enrollment 5167 266.06 219.64 1666 348.9 6.96 3501 226.86 2.86 Total Dropouts 5167 8.97 13.26 1666 11.78 16.19 3501 7.63 11.38 Student teacher ratio 5167 32.45 14.63 1666 31.64 11.35 3501 32.83 15.94 Textbooks per student 5167 3.58 2.58 1666 3.29 1.81 3501 3.71 2.86 Manuals per teacher 5167 4.05 2.98 1666 3.81 2.49 3501 4.17 3.18 Head teacher 5167 0.25 0.43 1666 0.29 0.45 3501 0.22 0.42 Head principal 5167 0.21 0.41 1666 0.33 0.47 3501 0.15 0.35 Complete combination and multigrade schools 5167 0.28 0.45 1666 0.16 0.37 3501 0.34 0.47 Complete Large and Medium Schools 5167 0.24 0.43 1666 0.39 0.49 3501 0.17 0.38 Complete Small Schools 5167 0.44 0.2 1666 0.44 0.13 3501 0.45 0.23 Primary Schools 5167 0.03 0.17 1666 0.01 0.09 3501 0.04 0.19 Elementary Leader Schools 5167 0.07 0.25 1666 0.15 0.36 3501 0.03 0.16 School received SIIF funds 5167 0.34 0.47 1666 0.48 0.5 3501 0.27 0.44 22 Table II: Differences in Scores between SBM and Non-SBM Schools during Pre- Intervention (2002-2003) and Implementation (2003-2004) Periods Math Score Science Score English Score Total Score 1A 1B 2A 2B 3A 3B 4A 4B Diff in 2002 0.463 2.112 0.289 1.845 0.669 2.148 0.469 2.022 [0.90] [4.35]*** [0.64] [4.35]*** [1.67]* [5.63]*** [1.11] [5.13]*** Diff in 2003 1.432 3.081 1.469 3.025 1.387 2.865 1.429 2.982 [2.79]*** [6.35]*** [3.27]*** [7.13]*** [3.47]*** [7.51]*** [3.39]*** [7.57]*** Division FE Incd Incd Incd Incd N 10334 10334 10334 10334 10334 10334 10334 10334 Note: Absolute value of t statistics in parentheses. Note: A denotes OLS; B denotes Fixed Effects *Significant at 10%; **Significant at 5%; ***Significant at 1% 23 Table III: Estimation of Propensity Scores Number of observations = 5167 LR Chi2(5) = 439.44 Prob>chi2 = 0 Pseudo R2 = 0.0676 Log likelihood = -3028.71 Dependent variable: SBM Variables Coefficient Std. Error z Student-teacher ratio -0.011 0.002 -4.68 School size 0.001 0.000 9.60 Complete schools 0.471 0.253 1.86 Monograde schools 0.544 0.083 6.55 School head principals 0.370 0.086 4.29 Intercept -1.848 0.271 -6.82 The common support option was selected. The region of common support is .119, .943 The final number of blocks is 10. This number of blocks ensures that the mean propensity score is not different for treated and controls in each block. Table IV: P-Score Quintiles, by SBM and Non-SBM Schools Quintile1 (n=1030) Quintile2 (n=1030) Quintile3(n=1029) Quintile4 (n=1030) Quintile5 (n=1029) Non-SBM SBM t-score Non-SBM SBM t-score Non-SBM SBM t-score Non-SBM SBM t-score Non-SBM SBM t-score Propensity Score 0.174 0.174 0.24 0.23 0.24 3.27*** 0.31 0.31 1.42 0.35 0.35 1.42 0.52 0.54 3.24*** Head_Principal 0.002 0.001 0.65 0.021 0.048 2.26** 0.01 0.003 1.27 0.14 0.11 1.42 0.88 0.84 1.57 Complete Schools 0.92 0.93 0.56 0.006 0.013 0.66 1 1 1 1 Monograde Schools 0.08 0.06 1.19 0.49 0.55 1.49 0.99 1 0.78 Student teacher ratio 36.78 37.24 0.36 32.21 33.04 0.59 28.95 30.46 2.34 31.5 31.4 0.16 31.6 30.2 2.28** Total Enrollment 2002 114.5 121.3 1.64* 151.38 165.2 2.32** 193.4 202.6 2.34** 280.4 290.7 1.92** 514.7 574.5 3.14*** *Significant at 10%; **Significant at 5%; ***Significant at 1% 24 Table V: Unconditional Test Score Means, By SBM/Non-SBM and Year 2002 2004 Diff t-stat Math Scores 46.14 61.96 15.82 46.71*** SBM 46.45 63.03 16.57 28.28*** non-SBM 45.99 61.45 15.46 37.29*** Diff 0.46 1.58 1.12 t-stat 0.89 3.11*** Science Scores 44.68 60.36 15.68 56.01*** SBM 45.14 56.87 11.74 34.82*** non-SBM 44.47 54.84 10.37 44.21*** Diff 0.67 2.03 1.36 t-stat 1.66* 4.99*** English Scores 43.09 55.50 12.41 43.29*** SBM 43.28 61.86 18.57 27.63*** non-SBM 42.99 59.65 16.66 33.66*** Diff 0.29 2.20 1.91 t-stat 0.65 4.88*** Overall Scores 44.64 59.78 15.13 55.14*** SBM 44.96 61.12 16.16 34.02*** non-SBM 44.49 59.14 14.65 43.63*** Diff 0.47 1.98 1.51 t-stat 1.12 4.82*** 25 Table VI: Treatment Effects Based on OLS, DID (with fixed effects), and DID with Kernel-PSM Effect Fixed Effect Kernel- Effect OLS size Effects size PSM size Scores 1 Math Scores 0.851 0.05 0.743 0.04 1.88* 0.07 2 English Scores 1.141** 0.08 1.046** 0.07 1.361** 0.09 3 Science Scores 1.368*** 0.10 1.283*** 0.09 1.818*** 0.13 4 Overall Scores 1.151** 0.08 1.058** 0.08 1.45** 0.10 *Significant at 10% **Significant at 5% ***Significant at 1% 26 Appendix A: OLS Regression Estimates of SBM Effects on School-Averaged Test Scores Dependent Variable Math English Science Overall Post-intervention (2004 – year 3) 12.357 9.188 13.14 12.076 [33.25]*** [28.85]*** [44.30]*** [40.02]*** SBM Batch1 School -0.381 -0.568 -0.75 -0.541 [0.98] [1.70]* [2.41]** [1.71]* SBM School x Post-intervention 0.851 1.141 1.368 1.151 [1.32] [2.07]** [2.66]*** [2.20]** Does the school have any dropouts? 3.443 2.901 2.222 2.883 [9.37]*** [9.21]*** [7.58]*** [9.66]*** No of dropouts, given that the school had >0 -0.336 -0.273 -0.264 -0.294 dropouts [11.94]*** [11.01]*** [11.44]*** [12.53]*** Size of School (total enrollment) 0 0 0 0 [.01] [0.55] [0.04] [0.23] Student Teacher Ratio 0.027 -0.003 -0.004 0.007 [2.73]*** [0.28] [0.49] [0.83] Textbook per student 0.742 0.589 0.641 0.656 [8.99]*** [8.30]*** [9.69]*** [9.76]*** Teacher Manual per teacher -0.522 -0.367 -0.417 -0.44 [7.28]*** [5.95]*** [7.26]*** [7.53]*** Head Teacher 2.872 2.713 2.885 2.77 [7.43]*** [8.18]*** [9.35]*** [8.82]*** Principal 1.299 2.057 2.186 1.896 [2.59]*** [4.72]*** [5.39]*** [4.59]*** 3.737 2.207 2.187 2.677 Complete Combination and Multigrade School [4.38]*** [3.02]*** [3.21]*** [3.87]*** 7.312 5.652 6.384 6.303 Complete Mono Large and Medium School [7.77]*** [6.40]*** [7.77]*** [7.54]*** Complete Mono Small School 5.304 3.365 3.908 4.097 [6.18]*** [4.53]*** [5.64]*** [5.82]*** Elementary Leader School 1.679 1.289 2.152 1.597 [2.70]*** [2.41]** [4.33]*** [3.16]*** Received SIIF Grant 0.756 0.903 1.187 0.955 [2.34]** [3.26]*** [4.59]*** [3.63]*** Constant 41.259 40.164 41.032 40.906 [44.63]*** [50.62]*** [55.53]*** [54.42]*** Observations 14823 14823 14823 14823 R-squared 0.13 0.11 0.21 0.18 27 Appendix B: Difference-in-Difference with District Level Fixed Estimates of SBM Effects on School-Averaged Test Scores (1) (2) (3) (4) Dependent variable Math English Science Overall Post-intervention (2004 – year 3) 12.484 9.328 13.23 12.186 [36.30]*** [31.69]*** [48.16]*** [44.30]*** SBM Batch1 School 1.071 0.63 0.49 0.706 [2.85]*** [1.96]** [1.63] [2.35]** SBM School x Post-intervention 0.743 1.046 1.283 1.058 [1.25] [2.05]** [2.70]*** [2.22]** Size of School (total enrollment) 0 0 0 0 [0.99] [0.54] [0.18] [0.11] Does the school have any dropouts? 3.78 3.322 2.442 3.171 [10.88]*** [11.17]*** [8.80]*** [11.41]*** No of dropouts, given that the school had >0 -0.208 -0.164 -0.162 -0.183 dropouts [7.64]*** [7.04]*** [7.45]*** [8.41]*** Student Teacher Ratio 0.017 -0.002 -0.004 0.003 [1.70]* [0.29] [0.46] [0.38] Textbook per student 0.229 0.159 0.229 0.198 [2.72]*** [2.20]** [3.40]*** [2.94]*** Teacher Manual per teacher -0.222 -0.139 -0.206 -0.184 [3.18]*** [2.33]** [3.70]*** [3.30]*** Head Teacher 2.213 2.171 2.261 2.177 [6.13]*** [7.02]*** [7.83]*** [7.53]*** Principal 1.832 2.281 2.304 2.193 [3.81]*** [5.54]*** [6.00]*** [5.70]*** 3.341 2.098 2.184 2.486 Complete Combination and Multigrade School [4.22]*** [3.09]*** [3.45]*** [3.92]*** Complete Mono Large and Medium School 6.937 5.254 6.001 5.902 [7.23]*** [6.39]*** [7.83]*** [7.69]*** Complete Mono Small School 5.89 3.975 4.541 4.685 [7.30]*** [5.76]*** [7.05]*** [7.26]*** Elementary Leader School 2.582 1.983 2.79 2.378 [4.46]*** [4.00]*** [6.03]*** [5.13]*** Received SIIF Grant 0.145 0.287 0.534 0.311 [0.47] [1.09] [2.17]** [1.26] Constant 41.426 39.704 40.672 40.764 [47.15]*** [52.79]*** [57.95]*** [58.01]*** Observations 14823 14823 14823 14823 Division fixed effects Incld Incld Incld Incld R-squared 0.14 0.12 0.23 0.2 28 Appendix C: DID with (Kernel) PSM of SBM Effects on School-Averaged Test (1) (2) (3) (4) Dependent variable Math English Science Overall Post-intervention (2004- year 3) 11.152 8.675 12.182 11.251 [26.46]*** [24.12]*** [36.27]*** [33.01]*** SBM Batch1 School -0.668 -0.77 -0.83 -0.737 [1.54] [2.07]** [2.39]** [2.10]** SBM School x Post-intervention 1.188 1.361 1.818 1.45 [1.66]* [2.22]** [3.18]*** [2.50]** Size of School (total enrollment) 0 0 0 0 [0.14] [0.55] [0.16] [0.20] Does the school have any dropouts? 3.338 2.897 2.345 2.917 [8.03]*** [8.17]*** [7.08]*** [8.68]*** No of dropouts, given that the school had >0 dropouts -0.313 -0.254 -0.239 -0.271 [10.32]*** [9.80]*** [9.87]*** [11.03]*** Student Teacher Ratio 0.012 -0.017 -0.02 -0.007 [0.93] [1.62] [1.96]** [0.69] Textbook per student 0.947 0.712 0.77 0.808 [9.14]*** [8.05]*** [9.32]*** [9.64]*** Teacher Manual per teacher -0.557 -0.349 -0.392 -0.446 [6.65]*** [4.88]*** [5.87]*** [6.58]*** Head Teacher 2.626 2.233 2.64 2.473 [5.93]*** [5.91]*** [7.48]*** [6.91]*** Principal 0.762 1.628 1.975 1.515 [1.32] [3.31]*** [4.30]*** [3.25]*** Complete Combination and Multigrade School 3.744 2.437 2.3 2.871 [3.90]*** [2.97]*** [3.01]*** [3.70]*** Complete Mono Large and Medium School 7.07 5.712 6.055 6.189 [6.11]*** [5.78]*** [6.56]*** [6.61]*** Complete Mono Small School 5.104 3.517 3.716 4.062 [5.22]*** [4.21]*** [4.77]*** [5.14]*** Elementary Leader School 1.47 0.794 1.814 1.245 [2.12]** [1.34] [3.28]*** [2.22]** Received SIIF Grant 1.678 1.777 1.998 1.84 [4.69]*** [5.83]*** [7.02]*** [6.36]*** Constant 41.001 39.425 40.452 40.294 [38.29]*** [43.15]*** [47.42]*** [46.54]*** Observations 11349 11349 11349 11349 R-squared 0.12 0.11 0.2 0.17 Absolute value of t statistics in brackets * significant at 10%; ** significant at 5%; *** significant at 1% 29 i Although some other school districts in the Philippines had ongoing SBM initiatives, they are not included in this analysis as insufficient information is available about the nature and timing of those initiatives. Also, while this report focuses on TEEP, it is important to note previous projects that implemented SBM such as the Secondary Education Development and Improvement Project and Basic Education assistance in Mindanao in 2001. ii The TEEP objectives also included improved completion rates and reduced drop-out rates in schools. iii No written materials were available on the identification strategy. Project officials interviewed as part of this paper mentioned that schools were chosen on the basis of the perceived strength of their capabilities. iv The number of schools in the analysis is not the same as the number of schools included in the SBM program due to missing information. The difference is 728 schools. v The intervention might alter the number of children enrolling in a particular school. If as a consequence the distribution of students’ skills changes in treatment schools (with respect to control schools), then the program impact estimates are likely to be biased. We will explore the existence of this bias in section five when discuss the caveats. The characteristics of the average student in the school are also included in the error term because of lack of data on individual students’ characteristics. vi School Improvement and Innovation Facility that granted funds to schools based on proposals for school improvement programs. vii Propensity Score Matching (PSM) was proposed by Rosenbaum and Rubin (1983). viii Given that common support area condition is met, PSM approach eliminates a substantial amount of selection bias that would alter conclusions in interpretation (Heckman, Hidehiko, Smith and Todd, 1996). ix Some of the cells do not show t-scores because there were no observations for that particular variable in that cell. We tried several combinations of variables, but due to a limited dataset and number of available variables, kernel matching using the variables noted above provided the closest matches. x See Cohen (1988) for a discussion on the magnitude of effect size. Effect size is calculated by dividing the coefficient for interaction term, SBMxPostintervention, by the standard deviation of the 2002-03 baseline score. 30