The_effects_of_family_background_and_sch.pdf

The Effects of Family Background and School Quality on Low and High Achievers: Determinants of Academic Achievement in the Philippines Marigee P. Bacolod* University of California, Irvine Elizabeth M. King The World Bank September 30, 2003

ABSTRACT Quantile regressions are applied to Philippine data to estimate the differential impact of inputs on students at various points on the conditional achievement distribution. Variation in the students who attend schools outside their district, students who do not attend the nearest school, and students who transferred schools are used to identify these differential impacts and control for selection. Results suggest a policy of reducing student to teacher ratios has a positive effect on raising students’ math achievement, but may benefit high achievers more than the average or low achievers. In contrast, the impact of class size on English achievement is greater for the average or median student. Keywords: education, school quality JEL Classifications: I21, I28

*

Corresponding author. Please address comments to: UC Irvine, Department of Economics, 3151 Social Science Plaza, Irvine, CA 92697-5100. Tel (949) 824-1990. Fax (949) 824-2182. Email: [email protected]. We would like to thank Jere Behrman, Janet Currie, Paul Glewwe, Joe Hotz, Ken Sokoloff and participants of the UCLA Applied Microeconomics Proseminar and the Population Association of America 2000 Meetings for comments on this paper. The views expressed herein are solely of the authors and not of The World Bank.

Education is increasingly viewed as an essential mechanism for raising incomes in developing countries, both at the national and individual levels. A rich literature using data from many countries provides estimates of the average or marginal effects of completed years of schooling on individual’s earnings (see for example, Psacharopoulos 1994). The literature has also gone beyond estimates of the relationship between earnings and number of years of schooling to estimates of the effect of learning itself, as measured by student performance, on earnings (Behrman and Birdsall 1983). A number of studies have also focused on estimating the effect of school inputs on later earnings (e.g. Card and Krueger 1992, Grogger 1996, Betts 1995). This broad literature points to academic achievement as a good predictor for later success in the labor market. Since the Coleman Report (Coleman et. al. 1966) claimed that family background and the characteristics of other students in the school are more important than school differences—in school inputs and in how schools operate—toward explaining differences in student performance, a large and growing literature has tried to isolate the impact of school quality. More recently, Glewwe (2002) provides a comprehensive review of this literature, which have included reviews of past studies (Fuller and Clark 1994 and Hanushek 1995), as well as what he considers the “best ‘conservative’ studies” (Harbison and Hanushek 1992 on Brazil, Glewwe and Jacoby 1994 on Ghana, Glewwe et al. 1995 on Jamaica, and Kingdon 1996 on India), those that have estimated education production functions. The verdict on these studies is that they do not sufficiently address important estimation problems, such as omitted variable bias, measurement errors, and sample selection problems that arise when, among other things, parents and students choose their school, that could lead to misleading conclusions about appropriate education policies or programs. Glewwe (2002) also reviews studies that used randomized trials (Jamison et al. 1981 on Nicaragua; Heyneman, Jamison and Montenegro 1984 and Tan, Lane and Lassibille 1999 on the Philippines; Kremer et al. 1997, Glewwe, Kremer and Moulin 2001, and Glewwe et al. 2002 on Kenya1; and Kagitcibasi, Sunar, and Bekman 2001 on Turkey) which are able to avoid many of the estimation problems of the education production function studies but nonetheless face other problems. The primary downside to these experimental studies is small samples, because these studies are costly to implement and nonrandom sample attrition also leads to sample selection bias. These studies examined the effects of a range of educational inputs on student performance—additional textbooks, pedagogical materials for teachers, educational radio, pre-school programs, and school-feeding programs—but, in general, have not yielded strong recommendations on these inputs. Two other sets of studies were reviewed: one that uses 1

A more recent related study is Glewwe, Ilias and Kremer (2003) which found that bonuses paid to teachers in program schools yielded better student test scores, although students did not retain these gains after the program ended, indicating that this incentive program primarily encouraged teachers to prepare students for the test.

2

natural experiments to estimate the impact of student-teacher ratios (Case and Deaton 1999 on South Africa) and class size (Angrist and Lavy 1999 on Israel) on student performance; and another that uses quasi-experimental methods to measure the impact of changes in how schools are governed (Jimenez and Sawada 1999 on El Salvador, and King and Ozler 2000 on Nicaragua). This growing and improving literature, however, has focused on the average effects of programs and policies. Yet, the provision of educational inputs that might not have a large or significant effect for students at the median of the achievement distribution might be shown to have a larger effect at the tails of the distribution, even after controlling for student characteristics and family background. This is the hypothesis we test in this paper. In designing education policies, it is important to recognize and understand why their effects may not be uniform across the conditional distribution of student achievement, and therefore necessary to understand for whom inputs matter. There are several reasons for this. For example, reducing class size may not matter in raising average test scores, but may be shown to raise test scores for the lower tail of the conditional distribution. If so, then a central question would be: If policymakers want to raise the average (by providing inputs that matter to the average student), should they use compensating (by providing inputs that matter more to low achievers) or reinforcing (by providing more inputs that matter to the high achievers) policies? In this paper we find some differential impacts of school and teacher quality and parental inputs on the conditional distribution of child cognitive outcomes.2 Utilizing a unique longitudinal dataset from Cebu, Philippines, and quantile regression techniques with selection corrections for school choice in an educational production framework, we find that certain inputs matter more to children at the upper part of the achievement distribution than to those in the lower end.3 Teacher experience has a greater positive impact at the top end of the conditional distribution, but student-to-teacher ratio has a larger impact at the median of the conditional distribution for English language and at the upper tail for mathematics. These results are robust to several model specifications and corrections for selection bias. We have come across only one other study that uses quantile regressions to estimate the relation between school quality and test performance. Eide and Showalter (1998) base their estimates on data from the High School and Beyond and find results suggesting that the marginal dollar allocated towards per pupil district expenditures raises Math test score gains at the bottom of the conditional distribution, yet it doesn’t affect the average test score gains. The paper is then organized as follows. Section II presents the data, highlighting what can be learned from the CLHNS. Section III lays out our identification strategy and examines potential threats to 2

We explore outcomes as both achievement in levels and in test score gains.

3

We address parents’ endogenous school choice by using propensity scores and forming various non-experimental control groups.

3

validity, after which Section IV presents the estimation results. Finally, Section V concludes the paper and provides suggestions for further research. II.

DATA We use data from the Cebu Longitudinal Health and Nutrition Survey (CLHNS), which was

carried out in the Metropolitan Cebu area on the island of Cebu, Philippines.4 Metro Cebu includes Cebu City, the second largest city in the Philippines, and several surrounding urban and rural communities. The CLHNS tracks a sample of 3,080 children born between May 1, 1983 and April 30, 1984, in randomly selected barangays (districts).5 The interviews began before birth, and health and nutrition data were collected every two months for the first two years of the child’s life. In 1991-92 and 1994-95, follow-up surveys of mothers and children were conducted. A third follow-up survey (1998-99) is currently underway, when most adolescents are in their second or third year in high school. Data collection should be complete by the end of 1999. Our study does not include the latest survey round. In all the surveys, detailed information on socioeconomic and demographic information for each household were collected, as well as community-level information on services, schooling, and labor market opportunities. The later follow-up survey also gathered detailed information on all schools in the Metro Cebu area, including academic inputs and GPS location. In addition, performance on the National Elementary Assessment Test (NEAT) was also collected from the school’s administrative files. This is a test designed to assess abilities and skills of Grade VI pupils in all public and private elementary schools.6 In 1994-95 and also in 1996-97, English reading comprehension and mathematics tests were developed for the surveys based on official school curricula at various grades. The tests were administered to the index children (that is, the children surveyed starting in the first round) and to their younger sibling of schooling age. These follow-up surveys also collected detailed schooling history of each index child and, if in school, his or her younger sibling. Excluding children with missing test scores leaves us with 2186 index children and 1140 siblings for our analysis.7 Means of some key variables are shown in Table 1.1 by age in 1996. Younger siblings aged 8 or 9 have higher enrollment rates in school year 1996-97, greater than 95% relative to the older 4

The survey was jointly conducted by the Office of Population Studies at the University of San Carlos, Philippines, the Nutrition Center of the Philippines, and the Carolina Population Center of the University of North Carolina (Chapel Hill). See Appendix 1.1 for sample attrition and selection.

5

Our sample excludes multiple births. Inclusive would be 3289 children.

6

The NEAT consists of four subjects areas: Science, Mathematics, English and Heograpiya, Kasaysayan, Sibika (HeKASi) which are based on minimum learning competencies. See http://www.fapenet.org/public/netrc.htm, Philippine Dept of Education, Culture and Sports National Educational Testing and Research Center.

7

For more details on sample selection, see Appendix 1.1.

4

siblings and index children. Table 1.1 also shows that 29% of thirteen-year-old index children and 2227% of their siblings aged ten and older have delayed enrollment. In Cebu at the time these children started to enter school, the minimum age of school enrollment was six and a half. We thus define a child to be delayed if he or she first enrolls in first grade at seven and a half years of age or older.8 The high rates of delay among the thirteen-year-old index children relative to the twelve-yearolds can be explained by the timing of the Philippine school year, which typically begins in late June. These children were born later in the calendar year than the twelve-year-olds, so their parents may have perceived them to be less ready for school at the time they could have first enrolled.9 There are also a number of repeaters among the thirteen-year-old index children, with 27 percent having repeated a grade at least once. The high rate of repetition and delay for this cohort probably explains why their average grade level in school year 1996-97 was almost the same as the twelve-year-olds. Of the twelve-year-old index children, 44 percent are in high school while only 33 percent of the thirteen-year-olds are.10 There is usually at least one public primary school in each barangay. Table 1.1 shows that around 45 percent of siblings aged nine to twelve attend schools outside their barangay, while only relatively few of them attend private schools. The higher percentage of index children attending schools outside their attendance district may be partly due to a lack of high school in their residence barangay. Similarly, only about 50-55 percent of the sibling children attend the school nearest to their residence. We will attempt to use these sources of variation as potential sources of identification and will discuss this in more detail in the following section. A majority of the children reside in single or extended nuclear families. Around 40-50 percent of both fathers and mothers stopped schooling at elementary, 27 percent stopped at high school, and only 1314 percent reached college. III.

EMPIRICAL FRAMEWORK: The Education Production Function If learning is a principal goal of education, then measures of cognitive achievement are a direct

output of the educational process which involves a series of inputs applied in school, at home and the community at large. Following Hanushek (1986), scholastic outcome at time t, Υt, is some function Υt = f(Χt, Χt-1, Χt-2, ... , Χ0)

(1)

8

This is also how Glewwe, Jacoby and King(1999) define delayed enrollment.

9

Recall that the 12 year olds were born in the early part of 1984 while the 13 year olds were born after May 1983.

10

The typical pre-college education in the Philippines includes 6 grades in primary school and 4 years of high school.

5

where Χt, Χt-1, Χt-2, ... , Χ0 are the streams of family, community, school and teacher inputs obtained at each time t. This is because attainment at any date will depend on the entire history of these inputs prior to that date, although their value in explaining output may diminish over time. Since we also observe Υt -1, we can rewrite (1) as Υt = g(Χt, Υt -1)

(2)

Most studies measure scholastic outcome (Υ) by achievement scores, dropout rates, grade repetition rates, grade continuation rates, and transition rates to higher education cycles. In this paper, we focus on standardized student test scores in math and language (English). Notice that model (2), assuming Υt –1 enters linearly in g, is equivalent to the “gain” (or value added) model with the restriction that the coefficient on Υt –1 equals 1. That is, Υt = βΥt –1 + h(Χt) is equivalent to Υt - Υt –1 = h’(Χt)

(3)

with the restriction that β = 1. We estimate both restricted and unrestricted models. Hereafter we refer to (2) as the LEVELS model and (3) as the GAIN model. Under certain assumptions, the gain and level estimators bracket the causal effect of our parameters of interest (see Appendix 1.2). A central issue in estimating the education production relationship is that the levels of inputs applied to the learning process are in themselves choice variables. In particular, while a student and her parents may not directly affect inputs in any one school, by choosing one school over other schools, a student and her parents are effectively selecting the amount and combination of inputs in the schooling process. How relevant this issue is depends on the extent to which school choice exists. School choice exists if there is more than one school (public or private) in the community or if a student is able to use public schools outside the community of residence or if the student and her family can migrate closer to the desired school. In any of these scenarios endogeneity bias is a problem because the school inputs that determine cognitive achievement (or any other education outcome) are themselves systematically influenced by the same household characteristics that influence achievement. In the Cebu area that we study, 45 percent of enrolled students attend public schools that are located outside their barangay and 9.6 percent were enrolled in private schools.11 Hence, even ignoring the issue of who might have migrated in order to be closer to preferred schools, school choice exists.12

11

Overall 52.6 percent of enrolled students (46.5 percent of all children) attend schools outside of their barangay.

12

There is relatively little migration across barangays over the surveys. Only 18.6 percent of the children reside in a barangay in 1996 that is different from the one the index children were born in, and only 3.67 percent of all children change barangay of residence between 1994 & 1996, the periods under study.

6

More formally, if we estimate Οit = Χit’δ + u it

(4)

where Οit = Υit in the levels model and Χit includes Υi,t-1, or Οit = Υit - Υi,t-1 in the gain model, the endogeneity bias is given by13 plim δ-hat = δ + Qxxcov(u, Χ). Where parents have reinforcing tastes in that they will choose high inputs (X↑) for high achieving or highly-motivated children (u↑), the estimate for δ, δ-hat, is an overestimate since cov(u, Χ) > 0. Where parents have compensating or equalizing tastes in that they will instead choose high inputs to compensate for low ability children, we expect δ-hat to be an underestimate since cov(u, Χ) < 0. Hence, if parents desiring more learning for their children choose schools so as to optimize access to better inputs, then part of the effect of observed household characteristics on achievement is through their effect on school choice. A number of studies have handled this by explicitly modeling school choice as another equation, as in two-stage least squares. The disadvantage with using instrumental variables is that unobserved school characteristics that affect school choice may be correlated with unobserved student characteristics that determine student performance. Furthermore, if the correlation between the instruments and school inputs are weak, then even a small correlation between the instruments and unobserved student characteristics can produce a large inconsistency in two-stage least squares (Bound, Jaeger and Baker 1995).14 There are alternative methods of dealing with the issue of endogeneity of school characteristics in the education production function. The most common approach in the education production literature has been to ignore the problem. When student achievement data are available for more than two years, one approach has been to estimate a value-added function. Under the assumption that the unobserved variable in student achievement is time-invariant, the unobservables are thus differenced out. Hence, the error in the gain (or loss) in student achievement is not correlated with school choice (see for example Guryan 1999). One problem with this approach is that school choice is assumed to be set at the initial period, whereas students may, in fact, reconsider school choice depending on their performance in each period. When achievement data are available only for one year, one approach has been to model school choice explicitly, such as the choice of attending a public or private school, and then to estimate the production function with a school selection correction (e.g. Jimenez and Sawada 1999). The problem with 13

where Qxx=plim (X’X/n)-1 a positive definite matrix.

14

This is essentially what we find when we use the characteristics of the nearest schools or the average in the barangay as instruments for characteristics of the school actually attended.

7

this method is finding an appropriate instrument that will identify the school choice equation, a problem that is often unsatisfactorily addressed. The ideal experiment to address this endogeneity would be to identify school effects over students who are randomly distributed across schools, and hence the impact of school choice is isolated from the impact of the school inputs themselves on achievement. A similar non-experimental method is to compare students who have made a certain choice in school with a matched group of students who have made other choices. This is the method we use in this paper. In its simplest form the identification strategy in this approach of comparing matched samples is to use a difference estimator.15 Applying this to the Philippine example, consider two children, A and B, with similar observable individual and household characteristics. As illustrated in Figure 1.1 - Panel A, child B lives in barangay Lahug but attends school in barangay Tisa, while child A resides and attends a school in barangay Lahug. In addition, suppose that the predicted probabilities (i.e., propensity score, which is described below) of child A and child B to attend a school outside their barangays are equal. We then use the outcome of child A to represent what would have happened to child B had child B attended a school in barangay Lahug. As an alternative, we also consider children who choose to attend a school farther than the school they live the closest to.16 In this case, child A and B reside on the same GPS coordinate, and while child A attends the nearest school, child B does not. The outcome of child A then represents what would have happened to child B had child B attended the closer school. We also consider children who transferred schools between periods t and t-1 as an alternative source of identification. Suppose now that Child A and Child B, again with similar observable individual and household characteristics, both attend School J in period t-1. Between period t and t-1, Child B transfers schools and attends School K in period t, while Child A remains in School J (see Figure 1.1-C). Suppose also that the predicted probabilities of Child A and Child B to transfer to a different school in period t are equal. We then use the outcome of Child A to represent the outcome of Child B had Child B remained in School J. In our analysis, using evaluation terminology, there are essentially four different sets of matching “treatment” and “control” groups, as further tabulated in Table 1.3: ƒ

In the first matching, we match children we refer to as MOVERS to children we refer to as NONMOVERS. The treatment group are all children who attend schools outside their barangay of

15

For an illustration, refer to diagrams A-D in Figure 1.1.

16

GPS readings of the household residence location as well as all schools in the Metro Cebu area were taken in the 1994/95 survey. Distances from the household residence to every school in Metro Cebu are then calculated as straight lines connecting household and school coordinates.

8

residence in school year 1996-97, while the control are those who attend schools in their barangay of residence in school year 1996-97. ƒ

In the second matching, we match children who attend the school nearest to where they live and who we refer to as NEAR, to children we refer to as NOT-NEAR. The treatment group are children who attend schools that is not the nearest in school year 1996-97, while the control are those who attend the nearest school.

ƒ

In the third matching, the treatment group are all children who transferred schools between school year 1994-95 and school year 1996-97, referred to as TRANSFERS, while the control group are all children who stayed in the same schools in both periods (NON-TRANSFERS), regardless of whether they attend schools within or outside their residence barangay in either periods.

ƒ

Finally, for a fourth matching, we use as treatment group all children who, between school years 1994-95 and 1996-97, transferred to schools located outside their barangay of residence (TRANSFERS*MOVERS). These are the children whose indicator variable for transferring, interacted with a dummy for attending a school in 1996-97 that is outside their barangay of residence, equals 1. The control group then includes all children who didn't transfer schools between school years 1994-95 and 1996-97, or who attend schools within their barangay of residence in school year 1996-97. The empirical model to be estimated is then: Οit = Χit’δ + αdit + u it

(5)

INDEXit = Zit’η + vit

(6)

dit = 1 iff INDEXit >0 and 0 otherwise

where:

Οit = Υit (outcome in levels) and Χit includes Υi,t-1 or

Οit = Υit - Υi,t-1 (outcome in gains) Χit = student’s individual, household, and school characteristics Zit = characteristics that determine school choice dit.17

We use propensity score analysis to estimate this model. First introduced by Rosenbaum and Rubin (1985), propensity score analysis is used to construct samples from non-experimental data so that these samples behave like experimental data.18

17

In identifying over “movers”, Zit include i’s test scores from period t-1, grade in t, other individual and household characteristics, and characteristics of schools in the barangay i resides in. In identifying over “not-near”, Zit also include past achievement, other characteristics of i and his household, and characteristics of the nearest school to i. For the “transfers”, Zit include i’s test scores from period t-1, grade in t-1, other household characteristics, and characteristics of the school i attended in t-1.

9

In particular, we estimate the propensity to make choice d using a probit model: Pr(dit = 1) = Φ(Zit’η) Using the predicted probabilities of making choice d, we calculate a propensity score for each observation and then drop from the sample those observations whose propensities are not within a given range. This range is defined by the minimum propensity score in the “treatment” group (dit=1) and the maximum propensity score in the “control” group (dit=0). The purpose of defining this limit is to identify students in the sample who have a similar propensity as the treatment group to choose a school outside their community of residence. Having obtained a sample of students with similar propensities in not attending the nearest school or moving or transferring or transferring to a school outside their residence barangay, we estimate (4) conditional on the predicted propensity score for each student.19 In including the propensity score as a regressor, our model is thus closer to those of selection-type models than traditional matching, thereby controlling for correlation between the error terms in equations (5) and (6). An advantage in using propensity scores is that we have overlapping support on the selectioncorrection term, which is important for reducing biases and achieving identification. More importantly, identification is achieved off of information on schools the child could have attended, not necessarily the school the child attended. IV.

RESULTS Our results include estimates of the effects of individual, household and school characteristics on

student performance. In the discussion we focus on the effects of school characteristics, as those are the crux of the study. In general, the results regarding household background are as expected. Parents’ education has a positive effect on achievement levels in both English and math, with father’s years of schooling having a larger effect than mother’s. Parents’ education, however, is not a generally significant factor in explaining gains in achievement. To the extent that the aspiration index is a factor independent of parents’ education, it exerts a positive effect on both achievement levels and gains.20 The coefficients of household income, measured by log(consumption expenditure per capita), are significant only for English achievement levels. 18

More recent work by Heckman, Ichimura and Todd (1997) looks at all matching estimators, including nonparametric estimators.

19

We condition on the predicted propensity score in two ways. First we include the propensity score as a regressor. We also divide the scores into five intervals or bins and estimate (4) separately for each bin. The results are qualitatively similar at the corresponding bins & quantiles. Results reported include the propensity score as regressor.

20

The aspiration index is a categorical variable asked of the parents, “What level of education do you want your child to complete?”: 0=some elementary, 1=elementary graduate, and so on.

10

OLS results Least squares estimates of our preferred specification are reported in Tables 1.2A-1.2B. In specifications without teacher characteristics, attendance at a private school rather than a public school means a higher average achievement level in English but a smaller gain. A negative but insignificant effect is found for math scores in the levels and negative significant effect in the gains functions. Adding teacher characteristics change these results. The effects of private school attendance on English achievement levels, math achievement levels and gains are significantly negative. Student-teacher ratios (approximately also equal to class sizes) are meant to measure the intensity of teacher attention that students receive in the classroom. The expectation is that the higher this ratio, the more diluted the attention and the poorer the learning experience is supposed to be. This ratio does not affect average achievement level in English but positively affects the average gain. In math, reducing class sizes is significant in raising both achievement levels and gains. Controlling for teacher characteristics heighten the effects of this variable. Younger teachers, conditional on experience, negatively influence average test scores and gains in those levels. Conditional on age, more experienced teachers raise average test score levels though at a decreasing rate, and have no effect on average gains in test scores. Movers, transfers, and stayers Table 1.2C shows OLS results including indicator variables for moving, not attending the nearest school, transferring, and transferring to a school outside the residence barangay. The least squares estimates of the effects of family background and school characteristic on achievement remain stable even when controlling for these indicators. Attending a school outside one’s barangay of residence (mover) is significantly positively related to English achievement. Meanwhile, not attending the nearest school is significantly negatively related to math achievement levels. Conditional on individual, household, and school characteristics, other indicators are insignificant on achievement. Overall, descriptive statistics in Panels 4A-4D in Table 1.4 show that movers, not-near, and transfers are strictly going to better schools, relative to the non-movers/near/non-transfers whom we refer to simply as stayers. Movers. There are a number of possible reasons why we observe movement of students across barangays. One reason may be that a student or her parents are drawn by the possibility of attending a much better school. Another possibility may be that a student’s parents’ workplace may be closer to the school of attendance. A third possibility may be that a student’s home is on the border and closer to the school in the neighboring barangay. Indeed, the nearest school for about 32 percent of movers is located outside

11

their barangay of residence, while this is the case for only 13.5 percent of stayers. Movers are also on average further away from the nearest school located in their barangays of residence. Among movers, the average distance to the nearest school in their barangay is 764.3 meters, while non-movers are on average only 569.4 meters from the school nearest to them. How far a parent has to travel to take his or her child to barangay schools may then be a factor leading the student to move. Movers also attend schools in barangays with higher average performance on the NEAT. The average NEAT score among schools in the barangay of residence for movers is only 90.2, and 106.9 among schools in the barangay of attendance. But then, the NEAT average for the school movers and stayers actually attend is about equal (78.4 for non-movers and 79.7 for movers). There are also more schools in the destination barangays than in the residence barangays for the movers, so the supply of school places could also be a driving force in students’ choices. On average, movers reside in a barangay with less number of schools (1.8) relative to non-movers (2.7) and attend barangays with 2.9 schools. These means seem to indicate that movers are moving to try to equalize to those characteristics found in the stayers’ barangays. In addition, movers attend schools in barangays with older schools. For movers there are on average 1.4 schools in their barangay of attendance that were established before 1960, while only 0.8 such schools in their barangay of residence. This suggests that school reputation, in that older schools may have a more established standing, may also be driving students to move. Overall, these unconditional means suggest movers are drawn to barangays with more established, slightly better schools.21 In our probit regressions (reported below), we control for the characteristics of the nearest school as well as all schools located in the residence barangay, to capture the “push” factors that may be driving students to move, as well as the “pull” factors by controlling for the characteristics of other schools in the choice set, that is the nearest school if the nearest is located outside the residence barangay. Note, too, that a higher proportion of movers are in grade levels above sixth grade (37 percent) compared to 15 percent among non-movers. As noted in Table 1.3, we estimate our models separately for all children and for children in primary grades only.

21

In order to resolve if flows across barangays are in fact driven only by particular schools in particular destinations, we looked at all school barangays where more than 3-4 percent of the children in our sample attend. There are on average about the same number of movers and stayers among the barangays that attract slightly more students than others. Given that only very few of the overall sample changed their barangay of residence between when the children were born and 1996, using mover status to represent school choice seems to be reasonable. The equal proportion of movers and stayers on average is partly driven by outliers amongst these barangays that attract a disproportionate share of the students. For example, about 9 percent of all enrolled children attend schools in Barangay Labangon, yet 86.2 percent of these children are stayers. On the other end, 7 percent of all enrolled children attend schools in Barangay Basak, yet only 3.9 percent of these children reside in that barangay. Most all (94 percent) of the flows into this barangay are students from two neighboring barangays.

12

Not-near. Those who do not attend the nearest school (referred to as not-near) are on average not further from the nearest school than those who decide to attend the nearest school(near). The choice between attending a private versus public school does not seem to unconditionally drive the choice to attend the closest school either. Among those who do not attend the nearest school, 42 percent reside nearest to a private school, yet a greater proportion of this group (21 percent) attend a private school, relative to those who do attend the school nearest to them. Only 2 percent of this latter group attends private schools. Overall, those who choose not to attend the nearest school seem to driven by the possibility of attending a better school—schools that are older, with higher enrollment and smaller class sizes, more experienced teachers—but not by school performance in itself as indexed by the NEAT. As in the movers matching group, more of the not-near group are in grade levels above sixth grade in 1996 (41 percent) compared to 2.3 percent among the near group. In the probit regressions below, we also estimate our models separately for all children and for children in primary grades only. Transfers. A student may transfer schools either because the family moves residences, or the school he or she is attending does not offer the next grade level, or parents can no longer afford to pay the tuition (if the school is private), or are looking for a better school for their child, or the school itself refuses the student continued admission. We find a substantial proportion of those who transfer had no choice but to transfer (72.1 percent) and subsequently exclude them from the means reported in Table 1.4C.22 Of these schoolchildren who transferred schools but had a choice, 17 percent changed barangay of residence between 1994 and 1996. Since the GPS measures were taken in the 1994 survey, this may explain why transfers are attending schools in 1996 that are on average further than the school they were attending in 1994. Among those who transfer, 23 percent also attend private schools in 1996 while only 16 percent among this group were enrolled in private schools in 1994. None of the other school characteristics seem to indicate, however, what is driving these children to transfer. Since 41 percent of those who transfer and had a choice are in high school in 1996, we also do the analysis separately for all children and for those in high school. Transfer*Movers. Similar to what determines a student to transfer schools, a student may transfer to a school outside the barangay of residence either because the family moves residences, or there are no schools in the barangay that offer the next grade level, or his/her parents want their child to attend a better school, or the parents find the cost of local private schools too high, or the school itself refuses the student continued admission. The means reported in Table 1.4D indicate that part of the story is that lack of private schools in their residence barangay lead parents to transfer their children to attend a school in barangays also with more private schools. Among those who transfer to schools outside the residence 22

A student is defined as having no choice but to transfer if his/her grade level in 1994 is not the highest offered at his/her school in 1994, nor is his/her grade level in 1995 not the highest offered at his/her school in 1995.

13

barangay and who had a choice, 31 percent attend private schools in 1996 while only 18 percent among this group were enrolled in private schools in 1994. In addition, students who are transfer*movers on average reside in barangays with fewer private schools (0.7) and attend a school in a barangay with more private schools (1.03). None of the other school characteristics seem to indicate, however, what is driving these children to transfer. Since 53 percent of those who transfer and had a choice are in high school in 1996, we also do the analysis separately for all children and for those in high school. Propensity scores Turn now to the estimates of the probit models; these are reported in Tables 1.5A-1.5D. Figure 1.2 also plots the propensity scores for each group in the propensity score samples. The probit models are able to predict well moving, not attending the nearest school, and transferring for those in high school in 1996. There are also a good number of propensity score matches among the actual non-movers to match with actual movers, as well as actual near and actual not-near school children. This is not the same for transfers, however. In the propensity score sample, 92.9 percent of actual non-transfers have propensity scores less than 0.2. This means that in the p-score interval between 0 and 0.2, we are matching 1,970 students who didn’t transfer to only 94 who transferred. In addition, there are few remaining treated observations--only 9.7 percent of the propensity score sample transferred. The same problem arises in matching transfer*movers to non-transfer*movers. Even excluding those who transferred schools clearly because the schools they were attending prior to 1996 didn’t offer higher grades, there are obviously problems in matching all students who transferred and had a choice, to non-transfers, as well as matching all transfer*movers to non-transfer*movers.23 In light of this, we will focus only on matches using movers and not-near in discussing our subsequent results. Overall, the probit estimates confirm the story suggested by the conditional means in Tables 1.4A-1.4B and discussed in the previous section. Turning now to Table 1.5A, residing in an urban area strongly predicts attending a school outside the residence barangay. Those who changed barangay residences between 1994 & 1996 (3.5 percent of all enrolled children) are less likely to attend a non-local barangay school, an indication that parents want to minimize disruption (less than 60 percent of those who changed residences transferred schools). The supply of school places is also significantly driving students to move. Students are more likely to move from barangays with fewer local schools. Conditional on number of schools, students are also significantly more likely to stay in barangays with a greater number of older schools and where PTAs are more active. Having smaller class sizes and more

23

Among the students who transferred, only 18 percent in school year 1994-95 and 19 percent in 1995-96 attended schools that offered education beyond the primary grades (beyond grade six).

14

experienced teachers conditional on age, as well as better performance on the NEAT, also significantly determine attending local schools in the barangay. Residing in an urban area and having changed barangay residences also strongly predict attending a school other than the nearest one (Table 1.5B). Perversely, the closer a student’s residence is to the nearest school, and especially when that school is private, the more likely that student is to not attend that school, conditional on other characteristics. Note that this regression was estimated including controls for whether the nearest school should actually be in the feasible choice set of the student.24 Except for the nearest school’s performance on the NEAT in the probit estimated over the primary school children sample, none of the other nearest school characteristics seem to significantly indicate what is driving the choice to not attend the nearest school. The eventual results using the mover and not-near propensity samples are very similar, however (see discussion below). All in all, those who choose not to attend a local barangay school nor the nearest school seem to driven by individual circumstance and by the possibility of attending a better school than are offered by their local choices. Movers in particular are pushed by lack of schools, fewer private schools, and lower average school quality as indicated by teacher characteristics and performance on the NEAT. Individual circumstances for the most part seem to be determining whether students attend the school closest to them. Two-Stage Least Squares We earlier pointed out in the discussion on empirical strategy that one advantage in using propensity scores is that identification is achieved off of information on schools the child could have attended, not necessarily the school the child attended. Another way such information could be utilized is through two-stage least squares. In Tables 1.6A-1.6D Column 2, we report 2SLS estimates using the characteristics of the nearest school as instruments for the characteristics of the school actually attended by the student.25 Note that family background effects are in general stronger in the instrumental variables regressions relative to OLS, while hardly any of the school characteristics are significant. The weak second-stage results can be explained by the weak correlation between the characteristics of the nearest

24

Only the interaction for whether the nearest school offers grades 1 through 6 interacted with a dummy for whether the student is above sixth grade is significant. Dropping the other interactions do not change the results and some of these terms get dropped due to collinearity.

25

The implied assumption in the use of nearest school characteristics as instruments is that parents are not moving residences to be closer to better schools. As noted earlier in footnote 12, very few residential migration across barangays is going on. This suggests that most parents are sorting directly across schools, and not through residences.

15

school and the school attended over the entire sample of students.26 For a sub-sample of students that did attend the nearest school, this correlation is equal to one. Meanwhile, for the group that did not attend the nearest school, the correlation is closer to negative one (recall that the characteristics of the nearest school became “push” factors and were significant determinants in the decision to attend not-near schools). Hence, while the use of the nearest characteristics as instruments has intuitive appeal, in this particular context it is not empirically convincing. Quantile regressions Quantile regression estimates of the effect of family background and school characteristics for the th

th

10 , 50 , and 90th percentiles are reported in Appendix 1.3 using the various matching groups. We also report a subset of the same estimates in Table 1.6 together with OLS and IV estimates for a comparison. Results from matching across quantiles clearly provide more information than the least squares results. In the following we will discuss quantile regression estimates for the effect of family background and school characteristics on achievement using the mover and not-near matching groups separately. A. Results using movers to control for selection. Father’s and mother’s education have positive effects across the conditional English distribution in levels (Table 1.6A). The positive effects of mother’s education on English increase the higher a child is on the distribution, while the positive effects of father’s education on English levels are slightly stronger for the tail ends of the distribution relative to the median. Father’s education also has a slightly stronger effect on the tail ends of the conditional math achievement distribution in levels, while mother’s education is insignificant across the distribution. The effects of parent education are insignificant on math gains and father’s education is significant only for the lower quantile in English gains (Table 1.6B). Parental aspiration also has a positive effect on English and math achievement levels and on math gains, higher for the high achievers relative to the median and lower end. Per household member expenditure is positive and significant only for the upper tail in English levels. Higher student-to-teacher ratios have a significant negative effect on English achievement levels and gains only for the median (Tables 1.6A & 1.6B). Meanwhile, higher student-to teacher ratios are worse for the median to upper tails in math achievement levels and gains. We also find that streaming significantly benefits only the low-achievers in conditional English gains. Conditional on age, a teacher with more experience raises English outcome levels and gains and math levels at a diminishing rate from the median to upper tail. This positive effect of experience increases from the median to 0.9 quantile. 26

See Bound, Jaeger, Baker (1995).

16

Teacher age, conditional on experience, has a strong negative effect on English test score levels and gains, and on math levels for the median to lower tail and for the upper tail in math gains. B. Results using not-near to control for selection As in using non-movers as the control group, the positive effects of father’s education on English and math levels are slightly stronger for the tail ends of the distribution relative to the median (Table 1.6C). The positive effect of mother’s education is significant only for the English high achievers in levels. Parental aspiration also has a positive effect on math achievement levels and on math gains from the median to the upper tail, higher for the high achievers relative to the median. In English achievement in levels, parent aspiration is also strongly positive for the median. Per household member expenditure is positive and significant only for the median to upper tail in English levels. As in using non-movers as the control group, higher student-to-teacher ratios have a significant negative effect on English achievement levels and gains only for the median (Tables 1.6C & 1.6D). In math achievement levels and gains, higher student-to teacher ratios are worse for the median to upper tails. Conditional on age, a teacher with more experience raises English outcome levels and gains and math levels at a diminishing rate from the median to upper tail. This positive effect of experience increases from the median to 0.9 quantile. Teacher age, conditional on experience, has a uniform strong negative effect on English test score levels and gains, and a significant negative effect on math achievement levels for the median to lower tail and for the upper tail in math gains. In addition, we also find that streaming is positive and significant only for the low-achievers in English gains. Negative streaming. Tables 1.6A-1.6D also show that a school policy of streaming hurts students most at the middle to upper tail of the distribution in math achievement. Streaming is the practice of segregating students into classes on the basis of perceived ability and past achievement. The negative effect on 1996 math test scores and on test score gain becomes larger the higher a student is on the achievement distribution. In using not-near to control for selection (Table 1.6C or 6D), streaming is also negative and significant for the 0.9 quantiles on English achievement. The mean corrected math and English scores in schools that segregate based on ability are 26 and 22.2, respectively, while in schools that don’t segregate, the means are 29.4 and 24.5. Average class sizes are also bigger in schools that stream at 36.51 students per teacher compared to 25.34 among schools that don’t stream. Since streaming is a policy that may or may not be practiced by a school—and is thus endogenous—schools may be streaming because they are attracting low-achievers, rather than causing low achievement. The unclear direction of causality may be leading to the negative estimates.

17

V.

SUMMARY AND CONCLUSION We began this paper by asking a policy question: for whom do school inputs matter. Most studies

on the determinants of academic achievement have focused on estimating effects on the mean or average student. Using data from Cebu, Philippines, we apply quantile regressions to estimate the differential impact of inputs on students at different points in the conditional achievement distribution. We control for the endogeneity of school choice using propensity scores. We exploit variation in the students who attend public schools outside their barangay of residence, students who do not attend the nearest school, and students who transferred schools to identify these differential impacts. We find that father’s and mother’s education have positive effects across the conditional English achievement distribution in levels, with the effect of mother’s education increasing the higher a child is on this distribution. Mother’s education does significantly affect math achievement. The positive effects of father’s education are stronger for the tail ends of both the math and English levels distributions. The effects of parent education are mostly insignificant on gains. Per capita household expenditure (in logs) is significant only for the high achievers. Our results also suggest a policy of reducing student to teacher ratios will have a positive effect on raising students’ Math achievement in levels and gains, but may also benefit high achievers more than the average or low achievers. In contrast, the impact of class size reductions on English achievement and gain may impact the average or median student more, relative to the tails. Hiring more experienced teachers have a stronger impact on the high achievers, both in English levels and gain and in Math levels. Teacher age, conditional on experience, has an even negative impact across the distribution. We also find streaming benefit only the low achievers in English gains and to have strong negative effects on Math achievement in levels and gains, particularly for high achievers, which might be a result of the timing in setting school policy. Further investigation of the mechanisms by which school quality is formed, and how this may lead to the differential effects that we have estimated, may be warranted.

18

References Angrist, Joshua D. and Victor Lavy. 1999. “Using Maimonides’ Rule to Estimate the Effect of Class Size on Scholastic Achievement,” Quarterly Journal of Economics 114(2): 533-75. Behrman, Jere R. and Nancy Birdsall. 1983. “The Quality of Schooling: Quantity Alone is Misleading,” American Economic Review 73(5): 928-946. Betts, Julian. 1995. “Does School Quality Matter? Evidence from the NLSY,” The Review of Economics and Statistics, 77 (May): 231-250. Bound, John, David A. Jaeger, Regina M. Baker. 1995. “Problems with Instrumental Variables Estimation When the Correlation Between the Instruments and the Endogenous Explanatory Variable is Weak,” Journal of the American Statistical Association 90(430): 443-450. Buchinsky, Moshe. 1994. “Changes in the U.S. Wage Structure 1963-1987: Application of quantile regression,” Econometrica 62(2): 405-458. Card, David and Alan Krueger. 1992. “Does School Quality Matter? Returns to Education and the Characteristics of Public Schools in the United States,” Journal of Political Economy, 100(February): 1-40. Case, A., and A. Deaton. 1999. "School Inputs and Educational Outcomes in South Africa." Quarterly Journal of Economics 114(3):1047–84. Coleman, James, E. Campbell, et.al. 1966. Equality for Educational Opportunity U.S. Department of Health, Education, and Welfare, Office of Education. Washington, D.C.: U.S. Government Printing Office. Eide, Eric and Mark Showalter. 1998. “The effect of school quality on student performance: A quantile regression approach,” Economics Letters, 58(3): 345-350. Fuller, Bruce. 1985. “Raising School Quality in Developing Countries: What Investments Boost Learning?” World Bank Discussion Paper No. 2. Fuller, Bruce, and Prema Clark. 1994. “Raising School Effects while Ignoring Culture? Local Conditions and the Influence of Classroom Tools, Rules and Pedagogy,” Review of Educational Research. 64(1):119-157. Gertler, Paul, and Paul Glewwe. 1990. “The Willingness to Pay for Education in Developing Countries: Evidence from Peru.” Journal of Public Economics. 42(3):251-275. Glewwe, Paul. 2002. “Schools and Skills in Developing Countries: Education Policies and Socioeconomic Outcomes,” Journal of Economic Literature 40(2): 436-82. Glewwe, Paul, Margaret Grosh, Hanan Jacoby, and Marlaine Lockheed. 1995. “An Eclectic Approach to Estimating the Determinants of Achievement in Jamaican Primary Education.” World Bank Economic Review 9 (2): 231-258. Glewwe, Paul, Nauman Ilias, and Michael Kremer. 2003. “Teacher Incentives,” NBER Working Paper Series, No. 9671. [http://www.nber.org/papers/w9671] Glewwe, Paul and Hanan Jacoby, 1995, “An economic analysis of delayed primary school enrollment and childhood malnutrition in a low income country” Review of Economics and Statistics, vol. 77, pp. 156-169. Glewwe, Paul, Hanan Jacoby, and Elizabeth King. 1999. “Early Childhood Nutrition and Academic Achievement: A Longitudinal Analysis,” Journal of Public Economics, 81(3): 345-68. Glewwe, Paul, Michael Kremer and Sylvie Moulin. 2001a. “Textbooks and Test Scores: Evidence from a Randomized Evaluation in Kenya”. Development Research Group. The World Bank. Washington DC. Glewwe, Paul, Michael Kremer, Sylvie Moulin and Eric Zitzewitz. 2000b. “Retrospective vs. Prospective Analyses of School Inputs: The Case of Flip Charts in Kenya”. NBER Working Paper No. 8018. Cambridge, MA. Grogger, Jeff. 1996. “School Expenditures and Post-Schooling Earnings: Evidence from the High School and Beyond.” The Review of Economics and Statistics, 78(4): 628-37. Guryan, Jonathan. 1999. “Desegregation and Black Dropout Rates,” mimeo, MIT.

19

Jepsen, Christopher. 1999. “The Effectiveness of Catholic Primary Schooling,” mimeo, Northwestern University. Jimenez, Emmanuel and Yasuyuki Sawada. 1998. “Do Community-Managed Schools Work? An Evaluation of El Salvador’s EDUCO Program,” Working Paper No. 8 in Series on Impact Evaluation of Education Reforms, The World Bank. Hanushek, Eric. 1995. “Interpreting Recent Research on Schooling in Developing Countries,” World Bank Research Observer, 10(2): 227-46. Hanushek, Eric. 1986. "The Economics of Schooling: Production and Efficiency in Public Schools." Journal of Economic Literature, 24(3): 1141-1177. Harbison, Ralph W. and Eric A. Hanushek. 1992. Educational performance of the poor: Lessons from rural northeast Brazil. New York: Oxford University Press for the World Bank. Heckman, James, Hidehiko Ichimura and Petra Todd. 1997. “Matching as an Econometric Evaluation Estimator: Evidence from Evaluating a Job Training Programme,” Review of Economic Studies, 64(4): 605-654. Heyneman, Stephen P., Dean T. Jamison and Xenia Montenegro. 1984. "Textbooks in the Philippines: Evaluation of the Pedagogical Impact of Nationwide Investment." Educational Evaluation and Policy Analysis 6(2):139-150. Imbens, Guido, Jeffrey Liebman, and Nada Eissa. 1997. “The Econometrics of Difference in Differences,” mimeo. Jamison, Dean, Barbara Searle, Klaus Galda and Stephen Heyneman. 1981. "Improving Elementary Mathematics Education in Nicaragua: An Experimental Study of the Impact of Textbooks and Radio on Achievement". Journal of Educational Psychology 73(4): 556-67. Kagitcibasi, Cigdem, Diane Sunar, and Sevda Bekman. 2001. “Long-Term Effects of Early Intervention: Turkish Low-Income Mothers and Children,” Journal of Applied Developmental Psychology 22(4): 333-61. King, Elizabeth M. and Berk Ozler. 1999. “What’s Decentralization Got To Do With Learning? The Case of Nicaragua’s School Autonomy Reform.” Working Paper No. 9 in Series on Impact Evaluation of Education Reforms, The World Bank. Kingdon, Geeta G. (1996). Student achievement and teacher pay: A case-study of India. London: London School of Economics and Political Science, The Development Economics Research Programme No.74. Kremer, Michael. 1995. “Research on Schooling: What we know and what we don’t, A Comment on Hanushek,” World Bank Research Observer, 10(2): 247-54. Kremer, Michael, Sylvie Moulin, Robert Namunyu and David Myatt. 1997. “The Quantity-Quality Tradeoff in Education: Evidence from a Prospective Evaluation in Kenya”. Unpublished. Office of Population Studies. 1989. "The Cebu Longitudinal Health and Nutrition Study: Survey Procedures and Instruments". University of San Carlos. Cebu City. Philippines. Psacharopoulos, George. 1994. “Earnings and Education in Latin America,” Education Economics, 2(2): 187-207. Rosenbaum and Rubin. 1985. “Constructing A Control Group Using Multivariate Matched Sample Methods that Incorporate the Propensity Score,” The American Statistician, 39: 33-38. Tan, Jee-Peng, Julia Lane and Gerard Lassibille. 1999. “Outcomes in Philippine Elementary Schools: An Evaluation of Four Experiments.” World Bank Economic Review. 13(3):493-508.

20

Table 1. Descriptive Statistics ALL Age as of 1996

Index Children 13 12

Younger Siblings 12 11 10

9

8

Number of Children

3326

1647

539

128

492

384

126

10

CHILD CHARACTERISTICS Percent Male Percent delayed Number of years delayed (mean for those who delayed)

52% 23% 1.16

52% 29% 1.2

52% 7% 1.31

57% 27% 1.18

53% 27% 1.12

51% 22% 1

48% 0% --

60% 0% --

Percent ever repeat any grade Percent enrolled (SY 1996/97)

19% 90%

27% 86%

11% 95%

23% 93%

14% 93%

9% 92%

2% 0% 97% 100%

Grade in SY 1996/97 (mean for those enrolled)

5.06

6.47

6.68

5.18

4.33

3.52

2.84

2.6

Mean 1996 - 94 Math score

8.27

10.28

12.05

4.86

4.17

3.1

1.46

1.75

Mean 1996 - 94 English score

6.28

7.74

10.02

3.32

3.01

2.53

-0.92

-2.57

Attending a private school* in SY 1996/97 Transferred schools between SY 94-95 & 96-97 Attends a school* outside barangay of residence Does not attend the nearest school* (by distance)

9.6%

11.8%

13.5%

3.4%

4.6%

5.6%

4.8% 10.0%

26.6%

38.8%

46.4%

5.9%

6.3%

7.6%

9.8% 20.0%

52.6%

53.9%

61.3%

44.5%

43.7% 43.7% 46.7% 20.0%

61.4%

66.4%

76.3%

48.7%

49.3% 45.4% 45.1% 50.0%

Type of Household Structure in 1994 Singe-person Household 0.1% One nuclear family 68.5% Extended nuclear family 18.9% Household of unrelated persons 12.5% No. of siblings as of 1994 4.17

0.1% 67.1% 19.7% 13.1% 4.13

0.2% 65.1% 21.0% 13.7% 3.67

-73.4% 19.5% 7.0% 4.12

0.2% 0.3% --72.4% 71.9% 69.8% 80.0% 16.7% 15.6% 18.3% 10.0% 10.8% 12.2% 11.9% 10.0% 4.48 4.62 4.28 5.1

0.9% 45.7% 37.9% 15.5%

1.3% 3.0% 0.8% 10.0% 55.7% 56.8% 57.0% 50.0% 30.6% 27.7% 30.6% 20.0% 12.3% 12.5% 11.6% 20.0%

1.8% 47.3% 36.4% 14.6%

1.8% 0.6% 3.5% 0.0% 51.0% 56.5% 43.0% 66.7% 32.2% 28.1% 32.5% 11.1% 15.0% 14.8% 21.0% 22.2%

Mother's educational attainment No Schooling 8.9% 2.3% 1.0% Elementary 50.1% 56.2% 48.0% High School 27.5% 27.8% 33.3% College or greater 13.0% 13.7% 17.7% Father's educational attainment No Schooling 14.9% 2.4% 0.6% Elementary 44.4% 52.3% 47.4% High School 26.6% 30.2% 31.3% College or greater 14.0% 15.1% 20.7% Note: *Percentages taken over students enrolled in SY 1996-97.

21

Table 2A. OLS Results in ENGLISH ACHIEVEMENT

(1) Past achievement 1994 Math score 1994 English score

0.235 (0.00)** 0.449 (0.00)**

Student Characteristics in 1996 -3.754 male age grade 4 grade 5 grade 6 above grade 6

(0.00)** 0.291 (0.18) 2.680 (0.00)** 6.014 (0.00)** 10.066 (0.00)** 11.595 (0.00)**

Household Characteristics aspiration index father’s educ in yrs mother’s educ in yrs log(per HH member expenditure)

1996 English score (2) (3) 0.215 (0.00)** 0.400 (0.00)**

0.213 (0.00)** 0.395 (0.00)**

0.211 (0.00)** 0.390 (0.00)**

-4.174 (0.00)** 0.691 (0.00)** 1.706 (0.00)** 4.764 (0.00)** 8.388 (0.00)** 9.550 (0.00)**

-4.187 (0.00)** 0.623 (0.01)** 1.954 (0.00)** 5.127 (0.00)** 8.767 (0.00)** 9.391 (0.00)**

-4.168 (0.00)** 0.574 (0.02)* 2.015 (0.00)** 5.253 (0.00)** 8.961 (0.00)** 9.483 (0.00)**

1.133 (0.00)** 0.280 (0.00)** 0.177 (0.00)** 0.506 (0.00)**

1.148 (0.00)** 0.276 (0.00)** 0.160 (0.00)** 0.442 (0.00)**

1.191 (0.00)** 0.287 (0.00)** 0.182 (0.00)** 0.396 (0.00)**

1.832 (0.03)* -0.910 (0.12) -0.009 (0.37)

-0.547 (0.57) -0.514 (0.35) -0.025 (0.10)

School Attended Characteristics private indicator if school streaming student per teacher Teacher Characteristics (School Avg) Age

Experience2 0.66

0.67

(5)

-2.971 (0.00)** 0.681 (0.00)** 1.640 (0.01)* 4.361 (0.00)** 7.171 (0.00)** 6.421 (0.00)**

1996-94 English score (6) (7)

0.67

0.68

-2.998 (0.00)** 0.822 (0.00)** 1.281 (0.14) 3.903 (0.00)** 6.488 (0.00)** 6.118 (0.00)**

-2.969 (0.00)** 0.796 (0.00)** 1.308 (0.13) 3.964 (0.00)** 6.588 (0.00)** 6.115 (0.00)**

0.579 (0.07) 0.100 (0.15) -0.056 (0.36) -0.054 (0.73)

0.576 (0.07) 0.112 (0.10) -0.035 (0.56) -0.016 (0.91)

0.592 (0.06) 0.117 (0.09) -0.025 (0.68) -0.046 (0.76)

-1.346 (0.11) -0.471 (0.32) -0.019 (0.03)*

-2.845 (0.00)** -0.216 (0.64) -0.026 (0.03)* -0.253 (0.03)* 0.286 (0.09) -0.003 (0.50)

0.11

0.11

Notes: 1. 2. 3. 4. 5. 6.

Robust p-values are reported in parentheses. Columns (1)-(4) using the LEVELS specification; (5)-(8) using the GAIN model. * denote significant at 5% level; ** significant at 1% level Models estimated with robust standard errors assuming interdependence within schools, independence across. All regressions include controls for missing values and a constant, and are based on a sample of 3326 children. Base category: Female, in grades 1-3 or not enrolled.

22

(8)

-2.993 (0.00)** 0.760 (0.00)** 1.447 (0.04)* 4.130 (0.00)** 6.792 (0.00)** 5.968 (0.00)**

-0.412 (0.00)** 0.583 (0.00)** -0.009 (0.09)

Experience

Adj. R-squared

(4)

0.11

0.12

Table 2B. OLS Results in MATH ACHIEVEMENT (1) Past achievement 1994 Math score 1994 English score

0.392 (0.00)** 0.275 (0.00)**

Student Characteristics in 1996 -2.407 male age grade 4 grade 5 grade 6 above grade 6

(0.00)** 0.561 (0.03)* 3.420 (0.00)** 6.963 (0.00)** 12.686 (0.00)** 12.717 (0.00)**

Household Characteristics aspiration index father’s educ in yrs mother’s educ in yrs log(per HH member expenditure)

1996 Math score (2) (3) 0.382 (0.00)** 0.253 (0.00)**

0.381 (0.00)** 0.251 (0.00)**

0.379 (0.00)** 0.245 (0.00)**

-2.582 (0.00)** 0.779 (0.00)** 2.914 (0.00)** 6.307 (0.00)** 11.803 (0.00)** 11.687 (0.00)**

-2.606 (0.00)** 0.881 (0.02)* 2.645 (0.00)** 5.966 (0.00)** 11.290 (0.00)** 11.815 (0.00)**

-2.603 (0.00)** 0.832 (0.03)* 2.712 (0.00)** 6.116 (0.00)** 11.481 (0.00)** 11.935 (0.00)**

0.787 (0.04)* 0.167 (0.01)* 0.027 (0.59) 0.225 (0.18)

0.837 (0.03)* 0.189 (0.00)** 0.046 (0.36) 0.202 (0.21)

0.890 (0.02)* 0.202 (0.00)** 0.069 (0.19) 0.169 (0.29)

-0.560 (0.51) -2.310 (0.00)** -0.034 (0.03)*

-2.566 (0.02)* -1.925 (0.02)* -0.052 (0.00)**

School Attended Characteristics private indicator if school streaming student per teacher Teacher Characteristics (School Avg) Age

Experience2 0.62

0.63

(5)

-2.057 (0.00)** 1.156 (0.00)** 2.142 (0.03)* 4.062 (0.00)** 7.931 (0.00)** 5.631 (0.00)**

1996-94 Math score (6) (7)

0.63

0.63

-2.057 (0.00)** 1.439 (0.00)** 1.180 (0.11) 2.879 (0.00)** 6.454 (0.00)** 6.401 (0.00)**

-2.050 (0.00)** 1.406 (0.00)** 1.222 (0.10) 2.970 (0.00)** 6.556 (0.00)** 6.376 (0.00)**

0.589 (0.12) 0.037 (0.59) -0.080 (0.18) -0.302 (0.08)

0.667 (0.08) 0.065 (0.35) -0.043 (0.48) -0.284 (0.09)

0.707 (0.05) 0.074 (0.27) -0.024 (0.70) -0.318 (0.06)

-2.372 (0.02)* -1.847 (0.01)** -0.025 (0.01)*

-4.388 (0.00)** -1.450 (0.04)* -0.037 (0.00)** -0.295 (0.02)* 0.361 (0.09) -0.006 (0.28)

0.11

0.11

Notes: 1. 2. 3. 4. 5. 6.

Robust p-values are reported in parentheses. Columns (1)-(4) using the LEVELS specification; (5)-(8) using the GAIN model. * denote significant at 5% level; ** significant at 1% level Models estimated with robust standard errors assuming interdependence within schools, independence across. All regressions include controls for missing values and a constant, and are based on a sample of 3326 children. Base category: Female, in grades 1-3 or not enrolled.

23

(8)

-2.022 (0.00)** 1.130 (0.00)** 2.202 (0.03)* 4.194 (0.00)** 8.078 (0.00)** 5.901 (0.00)**

-0.304 (0.02)* 0.587 (0.01)** -0.013 (0.02)*

Experience

Adj. R-squared

(4)

0.12

0.12

Table 2C. OLS results including controls for mover/transfer/not-near English Achievement Family Background aspiration index father’s educ in yrs mother’s educ in yrs

(1)

(2)

(3)

(4)

1.191 (0.00)** 0.287 (0.00)** 0.182 (0.00)** 0.396 (0.00)**

1.177 (0.00)** 0.290 (0.00)** 0.183 (0.00)** 0.401 (0.00)**

0.592 (0.06) 0.117 (0.09) -0.025 (0.68) -0.046 (0.76)

0.581 (0.05) 0.119 (0.09) -0.022 (0.71) -0.048 (0.75)

(0.57) -0.514 (0.35) -0.025 (0.10)

-0.681 (0.48) -0.589 (0.30) -0.026 (0.06)

-2.845 (0.00)** -0.216 (0.64) -0.026 (0.03)*

-2.946 (0.00)** -0.243 (0.60) -0.029 (0.01)*

-0.253 (0.03)* 0.286 (0.09) -0.003 (0.50)

-0.268 (0.02)* 0.318 (0.06) -0.004 (0.39)

log(per HH member expenditure) School Attended Characteristics -0.547 private indicator if school streaming student per teacher

Teacher Characteristics (School Avg) -0.304 -0.422 Age (0.02)* 0.587 (0.01)** -0.013 (0.02)*

Experience Experience2

(0.00)** 0.612 (0.00)** -0.010 (0.06)

27

Indicator variables Mover

0.842 (0.02)* -0.311 (0.45) 1.410 (0.36) -0.928 (0.35) -0.130 (0.88)

Not attending nearest school No choice but to transfer Transferred schools Transfer*Mover Adj. R-squared

0.63

0.65

0.711 (0.07) -0.564 (0.15) 1.077 (0.42) -1.014 (0.28) -0.171 (0.84)

0.12

0.11

Notes: 1. 2. 3. 4. 5. 6.

27

Robust p-values are reported in parentheses. Columns (1)-(2) using the LEVELS specification; (3)-(4) using the GAIN model. * denote significant at 5% level; ** significant at 1% level Models estimated with robust standard errors assuming interdependence within schools, independence across. All regressions include controls for gender, grade, age, past achievement, missing values and a constant, and are based on a sample of 3326 children. Base category: Female, in grades 1-3 or not enrolled.

For definition of these variables, see Table 3 or discussion in text.

24

Con’t, Table 2C. Math Achievement Family Background aspiration index father’s educ in yrs mother’s educ in yrs

(1)

(2)

(3)

(4)

0.890 (0.02)* 0.202 (0.00)** 0.069 (0.19) 0.169 (0.29)

0.855 (0.01)** 0.205 (0.00)** 0.079 (0.14) 0.160 (0.32)

0.707 (0.05) 0.074 (0.27) -0.024 (0.70) -0.318 (0.06)

0.699 (0.05) 0.071 (0.29) -0.029 (0.64) -0.321 (0.06)

(0.02)* -1.925 (0.02)* -0.052 (0.00)**

-2.465 (0.02)* -1.919 (0.02)* -0.057 (0.00)**

-4.388 (0.00)** -1.450 (0.04)* -0.037 (0.00)**

-4.239 (0.00)** -1.483 (0.03)* -0.039 (0.00)**

-0.295 (0.02)* 0.361 (0.09) -0.006 (0.28)

-0.306 (0.01)* 0.433 (0.06) -0.008 (0.17)

log(per HH member expenditure) School Attended Characteristics -2.566 private indicator if school streaming student per teacher

Teacher Characteristics (School Avg) -0.304 -0.332 Age Experience Experience2

(0.02)* 0.587 (0.01)** -0.013 (0.02)*

Indicator variables1 Mover

0.950 (0.14) -1.414 (0.02)* 1.737 (0.45) -0.784 (0.34) 0.165 (0.84)

Not attending nearest school No choice but to transfer Transferred schools Transfer*Mover Adj. R-squared

(0.01)** 0.677 (0.00)** -0.015 (0.01)**

0.63

0.63

0.899 (0.14) -0.914 (0.14) -0.135 (0.93) 1.527 (0.06) -0.585 (0.54)

0.12

0.11

Notes: 1. 2. 3. 4. 5. 6.

Robust p-values are reported in parentheses. Columns (1)-(2) using the LEVELS specification; (3)-(4) using the GAIN model. * denote significant at 5% level; ** significant at 1% level Models estimated with robust standard errors assuming interdependence within schools, independence across. All regressions include controls for gender, grade, age, past achievement, missing values and a constant, and are based on a sample of 3326 children. Base category: Female, in grades 1-3 or not enrolled.

25

TABLE 3. DESCRIPTION OF MATCHING GROUPS SAMPLE ENROLLED IN AY 1996/97

ENROLLED IN PRIMARY GRADES IN AY 1996/97 ENROLLED IN AY 1996/97

ENROLLED IN PRIMARY GRADES IN AY 1996/97 ENROLLED IN AY’s 1994/95, 1995/96, 1996/97 AND HAD A CHOICE TO 28 TRANSFER ENROLLED IN AY’s 1994/95, 1995/96, 1996/97, HAD A CHOICE TO TRANSFER1 AND IN HIGH SCHOOL IN 1996 ENROLLED IN AY’s 1994/95, 1995/96, 1996/97 AND HAD A CHOICE TO TRANSFER1 ENROLLED IN AY’s 1994/95, 1995/96, 1996/97, HAD A CHOICE TO TRANSFER1 AND IN HIGH SCHOOL IN 1996

28

TREATED MOVER=1 if barangay of school attended in AY 1996/97 is the same as residence barangay (n = 1547) MOVER=1 if barangay of school attended in AY 1996/97 is the same as residence barangay (n = 976) NOT NEAR=1 if school attended in AY 1996/97 is NOT the nearest (n = 1841) NOT NEAR=1 if school attended in AY 1996/97 is NOT the nearest (n = 1082) TRANSFER=1 if transferred schools anytime between AY 1994/95 & 1996/97 (n = 228) TRANSFER=1 if transferred schools anytime between AY 1994/95 & 1996/97 (n = 94) TRANSFER*MOVER=1 if transferred to a school outside the barangay of residence anytime between AY 1994/95 & 1996/97 (n = 164) TRANSFER*MOVER=1 if transferred to a school outside the barangay of residence anytime between AY 1994/95 & 1996/97 (n = 97)

CONTROL NON-MOVER (n = 1452)

NON-MOVER (n = 1237)

NEAR (n = 1158)

NEAR (n = 1131)

NON-TRANSFERS (n = 2027)

NON-TRANSFERS (n = 83)

NON-TRANSFER*MOVER (n = 2091)

NON-TRANSFER*MOVER (n = 90)

That is, student’s grade level in 1994 is not the highest offered at his/her school in 1994, nor is his/her grade level in 1995 not the highest offered at his/her school in 1995.

26

TABLE 4. SELECTED CHARACTERISTICS OF MATCHING GROUPS (See text for description of some variables.) 4A. SELECTED CHARACTERISTICS OF MOVERS vs. NON-MOVERS non-mover mover non-mover mover Index child Characteristics of schools in barangay of residence 0.589 (0.492) 0.698 (0.459) Age no of schools in brgy 11.843 (1.284) 12.050 (1.173) 2.713 (1.729) 1.821 (1.315) Male no of private schools in brgy 0.521 (0.5) 0.495 (0.5) 0.529 (0.88) 0.622 (0.698) Grade enrolled in 1996/97 no established before 1960 0.982 (0.665) 0.747 (0.612) grade1 mean entrance fee (Phil. Peso) 0.008 (0.091) 0.008 (0.088) 19.758 (72.669) 146.646 (252.291) grade2 mean amt parents contribute to PTA 0.008 (0.091) 0.008 (0.088) 11.964 (13.039) 34.562 (49.739) (0.334) (0.269) (25.72) grade3 mean % parents contribute to PTA 0.127 0.078 66.089 78.596 (23.141) (0.384) (0.34) (17.385) grade4 avg % passing NEAT 0.179 0.133 70.280 81.749 (16.405) (0.386) (0.351) (36.814) grade5 avg NEAT score 0.183 0.144 89.778 90.201 (29.074) (0.469) (0.432) (8.889) grade6 avg student/teacher 0.326 0.249 36.375 31.559 (9.134) (0.355) (0.483) (4.648) grade > 6 avg teacher experience 0.148 0.370 12.722 11.075 (5.319) Test scores avg teacher age 36.599 (5.792) 34.244 (7.069) IQ test score Characteristics of schools in barangay of attendance 65.690 (11.82) 68.530 (11.93) 1994 Math score no of schools in brgy 15.150 (14.01) 18.640 (14.82) 2.713 (1.729) 2.860 (1.314) 1994 English score no of private schools in brgy 12.320 (12.65) 15.740 (14.32) 0.529 (0.88) 0.830 (0.949) Household characteristics no established before 1960 0.982 (0.665) 1.360 (0.903) urban mean entrance fee (Phil. Peso) 0.650 (0.48) 0.780 (0.41) 19.758 (72.669) 49.884 (154.656) =1 if changed brgy resided 94-96 mean amt parents contribute to PTA 0.024 (0.152) 0.034 (0.182) 11.964 (13.039) 11.125 (17.348) (3.732) (3.914) (25.72) father's education (yrs) mean % parents contribute to PTA 7.170 8.492 66.089 51.339 (21.687) (3.467) (3.913) (17.385) mother's education (yrs) avg % passing NEAT 6.837 8.173 70.280 70.763 (18.168) (1.23) (1.33) (36.814) log(per capita hhold expenditure) avg NEAT score 7.240 7.490 89.778 106.895 (39.227) (8.889) avg student/teacher 36.375 29.304 (10.225) (4.648) Characteristics of school attended in 1996/97 avg teacher experience 12.722 13.297 (5.998) distance in meters avg teacher age 629.557 (1178.266) 1391.230 (2256.264) 36.599 (5.792) 32.824 (11.328) private Characteristics of nearest school in barangay of residence 0.068 (0.253) 0.180 (0.384) established before 1960 distance in meters 569.348 (1140.704) 0.640 (0.48) 0.785 (0.411) 764.229 (1803.067) amount parents contribute to PTA private 12.920 (14.93) 15.720 (26.55) 0.052 (0.222) 0.461 (0.499) % of parents making contributions % passing NEAT 74.400 (31.38) 70.290 (25.37) 75.300 (15.79) 85.560 (14.26) student/teacher NEAT average 41.040 (10.558) 38.540 (18.818) 78.370 (2.35) 81.610 (4.58) (954.945) (1859.616) total enrollment Characteristics of nearest school if nearest is not in barangay of residence 1213.500 2602.951 % passing NEAT =1 if nearest is NOT in brgy resid 76.010 (16.1) 81.440 (12.01) 0.135 (0.342) 0.324 (0.468) (2.23) (2.71) (332.636) average NEAT distance in meters 78.420 79.740 83.321 179.854 (333.764) (12.439) (10.222) (0.167) % teachers w/in-service training private 87.717 93.310 0.029 0.059 (0.236) (6.554) (6.94) (26.56) avg teacher experience % passing NEAT 14.711 18.014 10.120 26.450 (39.03) avg teacher age NEAT average 41.634 (6.847) 44.515 (8.347) 10.570 (26.86) 25.820 (37.35) Note: Means estimated over enrolled sample. Standard errors in parentheses. Means across schools within barangays are weighted by enrollment share in each school.

27

4B. SELECTED CHARACTERISTICS OF NEAR GROUP near Index child 0.517 (0.5) Age 11.687 (1.33) Male 0.537 (0.5) Grade enrolled in 1996/97 grade1 0.012 (0.11) grade2 0.012 (0.11) grade3 0.145 (0.35) grade4 0.213 (0.41) grade5 0.205 (0.4) grade6 0.363 (0.48) grade > 6 0.023 (0.15) Test scores IQ test score 64.720 (11.64) 1994 Math score 13.760 (13.66) 1994 English score 10.670 (11.77) Household characteristics urban 0.570 (0.5) =1 if changed brgy resided 94-96 0.011 (0.1) father's education (yrs) 6.929 (3.68) mother's education (yrs) 6.479 (3.37) log(per capita hhold expenditure) 7.150 (1.23) Characteristics of school attended in 1996/97 distance in meters 450.796 (323.95) private 0.019 (0.14) established before 1960 0.609 (0.49) policy of streaming 0.562 (0.5) amount parents contribute to PTA 12.290 (15.11) % of parents making contributions 69.050 (31.75) student/teacher 42.374 (18.65) total enrollment 1244.221 (1086.62) % passing NEAT 75.190 (16.25) average NEAT 78.230 (2.16) % teachers w/in-service training 87.811 (12.78) avg teacher experience 14.876 (6.82) avg teacher age 42.181 (6.95)

not-near 0.726 (0.45) 12.116 (1.13) 0.489 (0.5) 0.005 0.005 0.075 0.119 0.136 0.238 0.413

(0.07) (0.07) (0.26) (0.32) (0.34) (0.43) (0.49)

68.690 (11.91) 18.960 (14.72) 16.230 (14.29) 0.810 0.041 8.448 8.202 7.500 1460.448 0.211 0.795 0.743 15.950 75.710 37.628 2402.066 81.650 79.790 92.529 17.451 43.654

(0.39) (0.2) (3.89) (3.85) (1.31) (2365.35) (0.41) (0.4) (0.44) (25.24) (25.51) (9.98) (1781.07) (12.06)

near Selected characteristics of nearest school distance in meters 450.796 (323.95) private 0.019 (0.14) established before 1960 0.609 (0.49) offers grades 1-6 only 0.871 (0.33) % passing NEAT 75.190 (16.25) average NEAT 78.230 (2.16) % teachers w/in-service training 87.811 (12.78) avg teacher experience 14.876 (6.82) avg teacher age 42.181 (6.95)

(2.65) (10.35) (6.8) (8.28)

Note: Means estimated over enrolled sample. Standard errors in parentheses.

28

not-near 432.739 0.421 0.498 0.590 82.880 80.820 86.410 13.004 38.838

(382.24) (0.49) (0.5) (0.49) (15.41) (4.5) (20.53) (7.37) (8.51)

4C. SELECTED CHARACTERISTICS OF TRANSFERS non-transfer transfer Index child 0.536 (0.5) 0.667 (0.47) Age 11.733 (1.29) 11.904 (1.29) (0.5) Male 0.509 0.566 (0.5) Grade enrolled in 1996/97 grade1 0.002 (0.05) N/A grade2 0.002 (0.05) N/A grade3 0.128 (0.33) 0.114 (0.32) grade4 0.208 (0.41) 0.118 (0.32) grade5 0.210 (0.41) 0.154 (0.36) grade6 0.391 (0.49) 0.180 (0.38) grade > 6 0.041 (0.2) 0.412 (0.49) Test scores IQ test score 65.800 (11.74) 69.260 (12.12) (13.79) 1994 Math score 15.230 18.440 (14.66) (12.88) 1994 English score 12.280 16.960 (14.62) Household characteristics urban 0.710 (0.46) 0.880 (0.33) =1 if changed brgy resided 94-96 0.020 (0.14) 0.133 (0.34) father's education (yrs) 7.624 (3.87) 9.235 (4.04) mother's education (yrs) 7.235 (3.65) 8.980 (4.13) log(per capita hhold expenditure) 7.280 (1.26) 7.730 (1.34) non-transfer transfer Characteristics of school attended in 1996/97 Characteristics of school attended in 1994/95 distance in meters 750.020 (1024.3) 2879.836 (4574.7) distance in meters 750.020 (1024.3) 1287.675 (2042.43) private 0.073 (0.26) 0.229 (0.42) private 0.073 (0.26) 0.164 (0.37) established before 1960 0.713 (0.45) 0.643 (0.48) established before 1960 0.713 (0.45) 0.724 (0.45) (0.47) (0.48) (0.47) policy of streaming 0.657 0.637 policy of streaming 0.657 0.650 (0.48) (148.75) (245.76) (15.86) entrance fee 28.850 87.680 student/teacher 40.574 36.559 (6.91) amount parents contribute to PTA 13.370 (21.46) 15.900 (19.71) % passing NEAT 78.340 (14.26) 81.210 (12.99) % of parents making contributions 72.610 (28.68) 70.360 (30.98) average NEAT 78.970 (2.45) 79.620 (2.72) student/teacher 40.574 (15.86) 37.645 (8.46) % teachers w/in-service training 90.277 (11.69) 93.260 (12.87) total enrollment 1938.386 (1610.83) 1724.841 (1728.74) avg teacher experience 16.805 (6.53) 18.333 (7.11) % passing NEAT 78.340 (14.26) 78.490 (15.89) avg teacher age 43.732 (7.05) 44.877 (8.38) average NEAT 78.970 (2.45) 79.100 (2.69) % teachers w/in-service training 90.277 (11.69) 90.036 (11.76) avg teacher experience 16.805 (6.53) 15.201 (7.88) avg teacher age 43.732 (7.05) 41.403 (8.73) Note: Means estimated over continuously-enrolled sample and students whose grade level in 1994 is not the highest offered at his/her school in 1994, nor is his/her grade level in 1995 not the highest offered at his/her school in 1995. Standard errors in parentheses.

29

4D. SELECTED CHARACTERISTICS OF TRANSFER*MOVER non-transfer*mover

transfer*mover

index

0.535 (0.5)

0.732 (0.44)

age

11.726 (1.3)

12.055 (1.13)

male

0.509 (0.5)

0.579 (0.5)

grade1

0.002 (0.05)

0.000 (0.)

grade2

0.002 (0.05)

0.000 (0.)

grade3

0.131 (0.34)

0.073 (0.26)

grade4

0.208 (0.41)

0.085 (0.28)

grade5

0.209 (0.41)

0.146 (0.35)

grade6

0.387 (0.49)

0.140 (0.35)

0.044 (0.2)

0.530 (0.5)

Grade enrolled in 1996/97

grade > 6

non-transfer*mover

transfer*mover

Characteristics of schools in barangay of residence

Test scores IQ test score

65.820 (11.75)

70.320 (12.01)

no of schools in brgy

1994 Math score

15.190 (13.78)

20.160 (14.88)

no of private schools in brgy

1994 English score

12.370 (12.87)

17.590 (15.39)

no established before 1960 mean entrance fee (Phil. Peso)

Household characteristics

2.298 (1.59)

2.009 (1.34)

0.546 (0.8)

0.698 (0.82)

0.876 (0.64)

0.897 (0.57)

72.669 (184.42)

159.582 (240.67)

urban

0.710 (0.45)

0.910 (0.29)

mean amt parents contribute to PTA

21.788 (35.84)

51.106 (62.99)

=1 if changed brgy resided 94-96

0.027 (0.16)

0.084 (0.28)

mean % parents contribute to PTA

70.830 (25.75)

76.456 (26.25)

avg % passing NEAT

74.938 (18.17)

80.622 (18.31)

avg NEAT score

87.039 (31.69)

102.777 (35.74)

father's education (yrs)

7.636 (3.88)

9.678 (3.99)

mother's education (yrs)

7.252 (3.66)

9.413 (4.1)

log(per capita hhold expenditure)

7.280 (1.26)

7.900 (1.29)

Characteristics of school attended in 1996/97

avg student/teacher

34.549 (9.44)

33.170 (8.88)

avg teacher experience

11.954 (4.98)

12.107 (5.63)

avg teacher age

35.603 (6.42)

36.014 (7.38)

792.608 (1209.89)

3312.197 (5045.65)

private

0.074 (0.26)

0.313 (0.47)

established before 1960

0.710 (0.45)

0.656 (0.48)

policy of streaming

0.656 (0.48)

0.635 (0.48)

no of schools in brgy

2.637 (1.57)

3.168 (1.35)

offers grades 1-6 only

0.823 (0.38)

0.667 (0.47)

no of private schools in brgy

0.582 (0.91)

1.027 (1.02)

28.580 (147.17)

130.790 (300.67)

amount parents contribute to PTA

13.290 (21.19)

19.080 (23.94)

mean entrance fee (Phil. Peso)

% of parents making contributions

72.600 (28.84)

69.150 (28.99)

mean amt parents contribute to PTA

12.044 (15.55)

11.421 (17.67)

student/teacher

40.458 (15.67)

38.309 (9.58)

mean % parents contribute to PTA

62.279 (24.82)

50.031 (23.86)

total enrollment

1921.831 (1600.96)

1947.146 (2000.77)

78.230 (14.42)

81.260 (13.11)

distance in meters

entrance fee

% passing NEAT

78.960 (2.45)

79.610 (2.77)

90.257 (11.71)

90.334 (11.34)

avg teacher experience

16.740 (6.56)

15.628 (8.13)

avg teacher age

43.648 (7.08)

41.758 (9.29)

average NEAT % teachers w/in-service training

Characteristics of schools in barangay of attendance

no established before 1960

1.091 (0.75)

1.549 (1.03)

26.556 (101.02)

58.882 (159.9)

avg % passing NEAT

69.563 (18.07)

avg NEAT score

93.839 (37.61)

71.630 (14.46) 112.421 (41.7)

avg student/teacher

34.788 (9.25)

32.432 (13.71)

avg teacher experience

13.453 (5.29)

11.391 (5.63)

avg teacher age

36.697 (7.55)

28.863 (11.06)

Note: Means estimated over continuously-enrolled sample and students whose grade level in 1994 is not the highest offered at his/her school in 1994, nor is his/her grade level in 1995 not the highest offered at his/her school in 1995. Standard errors in parentheses. Means across schools within barangays are weighted by enrollment share in each school.

30

Con’t, 4D. non-transfer*mover

transfer*mover

non-transfer*mover

775.493 (1089.6)

1155.327 (1945.79)

private

0.075 (0.26)

0.178 (0.38)

established before 1960

0.708 (0.45)

0.803 (0.4)

policy of streaming

0.656 (0.48)

0.658 (0.48)

offers grades 1-6 only

0.818 (0.39)

0.375 (0.49)

student/teacher

40.513 (15.67)

35.709 (6.65)

% passing NEAT

78.300 (14.37)

82.870 (10.28)

distance in meters

78.980 (2.47)

79.810 (2.57)

90.335 (11.74)

93.687 (12.67)

avg teacher experience

16.808 (6.53)

19.032 (7.16)

avg teacher age

43.724 (7.06)

45.562 (8.63)

average NEAT % teachers w/in-service training

transfer*mover

Characteristics of nearest school in barangay of residence

Characteristics of school attended in 1994/95

distance in meters 608.609 (1172.3) private % passing NEAT NEAT average

1634.177 (4143.8)

0.229 (0.42)

0.460 (0.5)

79.580 (16.27)

87.380 (12.85)

79.740 (3.84)

82.510 (5.12)

Characteristics of nearest school if nearest is not in barangay of residence =1 if nearest is NOT in brgy resid

0.206 (0.4)

distance in meters 119.657 (350.66) private

0.354 (0.48) 155.626 (267.01)

0.036 (0.19)

0.106 (0.31)

% passing NEAT

16.260 (32.79)

30.000 (41.4)

NEAT average

16.300 (32.06)

28.590 (38.87)

Note: Means estimated over continuously-enrolled sample and students whose grade level in 1994 is not the highest offered at his/her school in 1994, nor is his/her grade level in 1995 not the highest offered at his/her school in 1995. Standard errors in parentheses. Means across schools within barangays are weighted by enrollment share in each school.

31

TABLE 5. PROBITS TABLE 5A. Determinants of attending a school outside the barangay of residence in school year 1996/97 (MOVER) (1) (2) Other Household Characteristics urban 0.154 0.139 (0.052)** (0.031)** dummy=1 if changed barangay -0.210 -0.240 resided between 1994 & 1996 (0.094)* (0.06)** Characteristics of schools in residence barangay no. of schools in brgy -0.112 -0.097 (0.032)** (0.037)** no. of private schools in brgy 0.034 0.089 (0.056) (0.063) no. established before 1960 -0.212 -0.287 (0.052)** (0.060)** mean amount parents contribute 0.005 0.005 to PTA (0.001)** (0.001)** mean % parents contribute to PTA 0.004 0.004 (0.001)** (0.001)** avg teacher experience 0.025 0.028 (0.012)* (0.013)* avg student/teacher -0.005 -0.009 (0.003) (0.004)* avg teacher age -0.021 -0.027 (0.008)* (0.009)** avg % passing NEAT -0.007 -0.004 (0.002)** (0.002) Characteristics of closest school in residence barangay distance in meters 1.99e-05 1.59e-05 (1.1e-5) (1.2e-05) private 0.288 0.439 (0.064) (0.11)** % passing NEAT 0.003 -0.002 (0.002) (0.002) Characteristics of closest school if closest is not in residence barangay dummy=1 if nearest school 0.145 0.321 is NOT in brgy of residence (0.065)* (0.244) distance in meters 1.12e-04 8.42e-05 (1.26e-04) (1.23e-04) private -0.057 -0.032 (0.085) (0.095) % passing NEAT 0.003 0.002 (0.002) (0.002) Observations 2999 1887 Pseudo R2 0.493 0.48 Notes: 1. 2. 3. 4. 5.

Robust standard errors are reported in parentheses, assuming interdependence within barangay of residence, independence across. Coefficients reported are derivatives (dF/dX). Column (1) estimated over ALL enrolled children; (2) over children enrolled in primary grades only. * denote significant at 5% level; ** significant at 1% level. All regressions include controls for gender, grade, age, past achievement, father’s and mother’s education in years, log(household expenditure per capita), missing values and a constant.

32

TABLE 5B. Determinants of attending a school that is NOT the nearest by distance in school year 1996/97 (NOT-NEAR) Other Household Characteristics urban dummy=1 if changed barangay resided between 1994 & 1996 Characteristics of nearest school distance in meters private % passing NEAT established before 1960 established in the 1960s offers grades 1-6 only (offers grade1-6)*(grade in 1996 > 6) Observations Pseudo-R2

(1)

(2)

0.228 (0.092)* 0.222 (0.039)**

0.265 (0.113)* 0.372 (0.053)**

2.6e-04 (7.11e-05)** 0.446 (0.053)** -0.003 (0.002) -0.009 (0.033) -0.011 (0.075) -0.023 (0.086) 0.332 (0.063)** 2999 0.469

3.59e-04 (9.8e-05)** 0.679 (0.042)** -0.005 (0.003) -0.032 (0.028) -0.028 (0.069) 0.007 (0.078) 2212 0.39

Notes: 1. 2. 3. 4. 5.

Robust standard errors are reported in parentheses, assuming interdependence within barangay of residence, independence across. Coefficients reported are derivatives (dF/dX). Column (1) estimated over ALL enrolled children; (2) over children enrolled in primary grades only. * denote significant at 5% level; ** significant at 1% level. All regressions include controls for gender, grade, age, past achievement, father’s and mother’s education in years, log(household expenditure per capita), missing values and a constant.

33

TABLE 5C. Determinants of transferring schools between SY 1994/95 & 1996/97. (TRANSFER) (1) (2) Other household characteristics urban 0.027 0.723 (0.024) (0.053)** dummy=1 if changed barangay 0.345 0.101 resided between 1994 & 1996 (0.068)** (0.197) Characteristics of school attended in SY 1994/95 distance in meters 1.45e-05 -1.38e-05 (4.03e-06)** (7.31e-05) % passing NEAT -0.001 -0.004 (0.001) (0.012) established before 1960 -0.047 -0.229 (0.019)* (0.179) established in 1960s -0.035 0.080 (0.01)* (0.178) established after 1985 0.072 0.485 (0.053) (0.046)** private -0.044 -0.958 (0.01)* (0.037)** student/teacher -0.002 -0.007 (0.001) (0.006) offers grades 1-6 only -0.048 (0.027)* Observations 2580 164 Pseudo-R2 0.282 0.58 Notes: 1. 2. 3. 4. 5.

Robust standard errors are reported in parentheses, assuming interdependence within barangay of residence, independence across. Coefficients reported are derivatives (dF/dX). Column (1) estimated over ALL enrolled children; (2) over children enrolled in secondary level only. * denote significant at 5% level; ** significant at 1% level. All regressions include controls for gender, grade, age, past achievement, father’s and mother’s education in years, log(household expenditure per capita), missing values and a constant.

34

TABLE 5D. Determinants of transferring to a school outside the barangay of residence between SY 1994/95 & 1996/97 (TRANSFER*MOVER) (1)

(2)

Other household characteristics urban

0.012 (0.01) dummy=1 if changed barangay 0.042 resided between 1994 & 1996 (0.032) Characteristics of school attended in SY 1994/95 distance in meters 5.446e-06 (2.33e-06)* % passing NEAT -3.64e-05 (4.01e-04) private -0.019 (0.009) established before 1960 -0.014 (0.016) established in 1960s -0.022 (0.007)* established after 1985 0.017 (0.007)* student/teacher -0.001 (0.001)

dropped (predicts 0 perfectly) 0.271 (0.148) -4.79e-05 (7.33e-05) -0.053 (0.024)* -0.757 (0.193)* 0.580 (0.28) 0.402 (0.216) 0.444 (0.057)** -0.019 (0.011)

Characteristics of schools in residence barangay no. of schools in brgy 2.3e-05 (0.007) no. of private schools in brgy 0.003 (0.011) no. established before 1960 -0.020 (0.011) mean amount parents contribute 2.67e-04 to PTA (1.39e-04)* mean % parents contribute to PTA 6.48e-05 (2.17e-04) avg teacher experience 0.001 (0.002) avg student/teacher 2.94e-04 (0.001) avg teacher age 3.0e-04 (0.002) avg % passing NEAT -1.36e-04 (3.7e-04) Characteristics of closest school in residence barangay distance in meters 3.98e-06 (1.82e-06)* private -0.006 (0.012) % passing NEAT 3.47e-04 (0.001) Characteristics of closest school if closest is not in residence barangay dummy=1 if nearest school 0.028 is NOT in brgy of residence (0.036) distance in meters -1.11e-05 (1.02e-05) private -0.008 (0.013) % passing NEAT 3.53e-05 (3.28e-04) Observations 2337 Pseudo-R2 0.376

-0.329 (0.147)* 0.356 (0.232) 0.344 (0.261) -0.004 (0.003) -0.003 (0.005) -0.016 (0.066) -0.038 (0.026) 0.059 (0.043) -0.014 (0.012) 1.23e-04 (1.09e-04) -0.486 (0.217)* 0.038 (0.010)** -0.646 (0.768) -4.5e-04 (1.8e-04)* -0.641 (0.109)** 0.010 (0.014) 147 0.57

Notes: 1. Robust standard errors are reported in parentheses, assuming interdependence within barangay of residence. 2. Coefficients reported are derivatives (dF/dX). 3. Column (1) estimated over ALL enrolled children; (2) over children enrolled in secondary level only. 4. * denote significant at 5% level; ** significant at 1% level. 5. All regressions include controls for gender, grade, age, past achievement, father’s and mother’s education in years, log(household expenditure per capita), missing values and a constant.

35

TABLE 6. OLS, IV, and QUANTILE REGRESSIONS for select matching groups A. Using ALL MOVERS as treatment (LEVELS MODEL) ENGLISH (1) (2) (3) (4) OLS IV q10 q50 Household Characteristics 0.287 0.428 0.313 0.200 father’s educ in yrs mother’s educ in yrs aspiration index

(0.00)** 0.182 (0.00)** 1.191 (0.00)** 0.396 (0.00)**

(0.00)** 0.516 (0.07) 1.140 (0.04)* 1.160 (0.20)

(0.01)** 0.081 (0.55) 0.695 (0.27) 0.248 (0.39)

(0.01)** 0.147 (0.07) 1.153 (0.00)** 0.301 (0.14)

0.399 (0.00)** 0.251 (0.04)* 1.289 (0.02)* 0.816 (0.00)**

(0.57) -0.514 (0.35) -0.025 (0.10)

-31.207 (0.23) 0.356 (0.97) -0.130 (0.66)

-3.798 (0.11) -0.036 (0.96) -0.067 (0.20)

-1.143 (0.32) -0.687 (0.21) -0.036 (0.10)

0.169 (0.94) -1.186 (0.16) -0.022 (0.26)

0.197 (0.44) 0.001 (0.93) -0.327 (0.04)* 3232 0.32

0.538 (0.01)** -0.007 (0.13) -0.422 (0.00)** 3232 0.43

0.788 (0.00)** -0.018 (0.02)* -0.327 (0.03)* 3232 0.43

log(per HH member expenditure) School Attended Characteristics -0.547 private indicator if school streaming student per teacher

Teacher Characteristics (School Avg) 0.583 1.075 Experience Experience2 Age

(5) q90

(0.00)** -0.009 (0.09) -0.412 (0.00)** 3326 0.65

(0.46) -0.014 (0.69) -0.644 (0.39) 3326 0.45

Observations R-sqd / Pseudo R2 Notes: 1. Column (1) report OLS estimates; (2) report IV estimates using characteristics of the

2.

nearest school as instruments for the characteristics of the school actually attended by the student (including distance and performance on the NEAT of the nearest school); (3) – (5) report quantile regression estimates at the 10th, 50th, and 90th percentile, respectively, using propensity score matched sample. Standard errors for cols (3)-(5) were estimated via bootstrap and p-values are reported in parentheses above.

MATH

(1) OLS Household Characteristics 0.202 father’s educ in yrs mother’s educ in yrs aspiration index

(2) IV

(3) q10

(4) q50

(5) q90

(0.00)** 0.069 (0.19) 0.890 (0.02)* 0.169 (0.29)

0.389 (0.02)* 0.458 (0.18) 0.966 (0.11) 0.802 (0.48)

0.230 (0.02)* 0.027 (0.79) -0.354 (0.44) -0.032 (0.89)

0.172 (0.01)** 0.041 (0.59) 1.003 (0.01)** 0.212 (0.34)

0.275 (0.01)* 0.089 (0.42) 1.328 (0.02)* 0.286 (0.28)

(0.02)* -1.925 (0.02)* -0.052 (0.00)**

-35.034 (0.29) -9.174 (0.43) -0.227 (0.51)

-1.747 (0.23) -0.158 (0.82) -0.031 (0.02)*

-3.619 (0.00)** -2.064 (0.00)** -0.069 (0.00)**

0.311 (0.83) -3.038 (0.00)** -0.043 (0.04)*

0.253 (0.34) -0.001 (0.91) -0.389 (0.03)* 3232 0.36

0.466 (0.00)** -0.014 (0.01)** -0.218 (0.02)* 3232 0.43

0.501 (0.06) -0.019 (0.00)** 0.003 (0.99) 3232 0.36

log(per HH member expenditure) School Attended Characteristics -2.566 private indicator if school streaming student per teacher

Teacher Characteristics (School Avg) 0.587 1.544 Experience Experience2 Age Observations R-sqd / Pseudo R2 3. 4. 5.

(0.01)** -0.013 (0.02)* -0.304 (0.02)* 3326 0.63

(0.40) -0.014 (0.73) -0.938 (0.39) 3326 0.39

* denote significant at 5% level; ** significant at 1% level. All regressions include controls for gender, grade, age, past achievement, missing values and a constant, and are estimated on propensity score samples (see text for discussion of how each sample was generated). Base category: Female, in grades 1-3 or not enrolled.

B. Using ALL MOVERS as treatment (GAIN MODEL) ENGLISH

(1) OLS Household Characteristics 0.117 father’s educ in yrs mother’s educ in yrs aspiration index

(2) IV

(3) q10

(4) q50

(5) q90

(0.09) -0.025 (0.68) 0.592 (0.06) -0.046 (0.76)

0.078 (0.69) -0.122 (0.71) 0.601 (0.19) -0.415 (0.67)

0.190 (0.05)* -0.215 (0.05) -0.123 (0.82) -0.274 (0.23)

0.086 (0.29) 0.060 (0.50) 0.449 (0.22) 0.040 (0.85)

0.163 (0.25) 0.070 (0.62) 0.906 (0.17) 0.176 (0.66)

(0.00)** -0.216 (0.64) -0.026 (0.03)*

6.286 (0.81) -5.549 (0.45) -0.076 (0.70)

-4.848 (0.01)** 1.515 (0.02)* -0.001 (0.95)

-4.519 (0.00)** -0.263 (0.66) -0.048 (0.07)

-0.384 (0.89) -1.656 (0.18) -0.028 (0.27)

-0.263 (0.27) 0.018 (0.01)** -0.388 (0.01)* 3232 0.11

0.342 (0.04)* -0.006 (0.25) -0.229 (0.08) 3232 0.05

1.142 (0.00)** -0.029 (0.00)** -0.267 (0.20) 3232 0.06

log(per HH member expenditure) School Attended Characteristics -2.845 private indicator if school streaming student per teacher

Teacher Characteristics (School Avg) 0.286 1.217 Experience Experience2 Age Observations R-sqd / Pseudo R2

(0.09) -0.003 (0.50) -0.253 (0.03)* 3326 0.12

(0.32) -0.027 (0.35) -0.321 (0.64) 3326 0.06

MATH

(1) OLS Household Characteristics 0.074 father’s educ in yrs mother’s educ in yrs aspiration index

(2) IV

(3) q10

(4) q50

(5) q90

(0.27) -0.024 (0.70) 0.707 (0.05) -0.318 (0.06)

0.110 (0.62) 0.050 (0.90) 0.584 (0.23) -0.297 (0.79)

0.092 (0.44) -0.076 (0.50) 0.360 (0.45) -0.594 (0.03)*

0.034 (0.66) -0.032 (0.67) 0.694 (0.03)* -0.410 (0.02)*

0.169 (0.32) 0.005 (0.98) 2.228 (0.00)** 0.023 (0.96)

(0.00)** -1.450 (0.04)* -0.037 (0.00)**

-9.814 (0.75) -7.417 (0.37) -0.243 (0.29)

-2.907 (0.16) -0.774 (0.31) -0.050 (0.21)

-3.917 (0.01)* -1.399 (0.01)* -0.034 (0.04)*

-7.977 (0.00)** -3.393 (0.01)** -0.048 (0.01)**

0.323 (0.31) -0.001 (0.95) -0.370 (0.13) 3232 0.13

0.265 (0.20) -0.008 (0.15) -0.147 (0.18) 3232 0.05

0.271 (0.44) 0.005 (0.65) -0.553 (0.01)* 3232 0.07

log(per HH member expenditure) School Attended Characteristics -4.388 private indicator if school streaming student per teacher

Teacher Characteristics (School Avg) 0.361 1.100 Experience Experience2 Age Observations R-sqd / Pseudo R2

(0.09) -0.006 (0.28) -0.295 (0.02)* 3326 0.12

(0.38) -0.011 (0.69) -0.711 (0.34) 3326 0.05

Notes: 1.

2.

Column (1) report OLS estimates; (2) report IV estimates using characteristics of the nearest school as instruments for the characteristics of the school actually attended by the student (including distance and performance on the NEAT of the nearest school); (3) – (5) report quantile regression estimates at the 10th, 50th, and 90th percentile, respectively, using propensity score matched sample. Standard errors for cols (3)-(5) were estimated via bootstrap and p-values are reported in parentheses above.

3. 4. 5.

* denote significant at 5% level; ** significant at 1% level. All regressions include controls for gender, grade, age, missing values and a constant, and are estimated on propensity score samples (see text for discussion of how each sample was generated). Base category: Female, in grades 1-3 or not enrolled.

C. Using ALL NOT-NEAR as treatment (LEVELS MODEL) ENGLISH

(1) OLS Household Characteristics 0.287 father’s educ in yrs mother’s educ in yrs aspiration index

(2) IV

(3) q10

(4) q50

(5) q90

(0.00)** 0.182 (0.00)** 1.191 (0.00)** 0.396 (0.00)**

0.428 (0.00)** 0.516 (0.07) 1.140 (0.04)* 1.160 (0.20)

0.273 (0.00)** 0.121 (0.35) 1.039 (0.06) 0.243 (0.36)

0.195 (0.04)* 0.109 (0.17) 1.338 (0.00)** 0.348 (0.10)

0.390 (0.00)** 0.321 (0.01)* 1.125 (0.13) 0.768 (0.00)**

(0.57) -0.514 (0.35) -0.025 (0.10)

-31.207 (0.23) 0.356 (0.97) -0.130 (0.66)

-4.068 (0.09) 0.363 (0.66) -0.075 (0.13)

-0.188 (0.88) -0.745 (0.24) -0.040 (0.05)

-1.281 (0.53) -1.596 (0.04)* -0.024 (0.24)

0.299 (0.22) 4.0e-04 (0.94) -0.427 (0.02)* 3167 0.32

0.649 (0.00)** -0.010 (0.05)* -0.403 (0.00)** 3167 0.43

0.740 (0.00)** -0.015 (0.04)* -0.409 (0.00)** 3167 0.43

log(per HH member expenditure) School Attended Characteristics -0.547 private indicator if school streaming student per teacher

Teacher Characteristics (School Avg) 0.583 1.075 Experience Experience2 Age Observations R-sqd / Pseudo R2

(0.00)** -0.009 (0.09) -0.412 (0.00)** 3326 0.65

(0.46) -0.014 (0.69) -0.644 (0.39) 3326 0.45

MATH

(1) OLS Household Characteristics 0.202 father’s educ in yrs mother’s educ in yrs aspiration index

(2) IV

(3) q10

(4) q50

(5) q90

(0.00)** 0.069 (0.19) 0.890 (0.02)* 0.169 (0.29)

0.389 (0.02)* 0.458 (0.18) 0.966 (0.11) 0.802 (0.48)

0.229 (0.05)* 0.021 (0.84) -0.041 (0.95) -0.061 (0.78)

0.173 (0.01)* 0.052 (0.56) 0.983 (0.02)* 0.164 (0.42)

0.210 (0.08) 0.146 (0.19) 1.801 (0.00)** 0.225 (0.39)

(0.02)* -1.925 (0.02)* -0.052 (0.00)**

-35.034 (0.29) -9.174 (0.43) -0.227 (0.51)

-2.041 (0.15) -0.438 (0.54) -0.036 (0.01)**

-3.239 (0.02)* -2.106 (0.00)** -0.067 (0.00)**

0.701 (0.66) -3.282 (0.00)** -0.059 (0.02)*

0.462 (0.09) -0.005 (0.47) -0.458 (0.00)** 3167 0.36

0.530 (0.00)** -0.016 (0.00)** -0.223 (0.04)* 3167 0.43

0.592 (0.03)* -0.016 (0.02)* -0.105 (0.53) 3167 0.36

log(per HH member expenditure) School Attended Characteristics -2.566 private indicator if school streaming student per teacher

Teacher Characteristics (School Avg) 0.587 1.544 Experience Experience2 Age Observations R-sqd / Pseudo R2

(0.01)** -0.013 (0.02)* -0.304 (0.02)* 3326 0.63

(0.40) -0.014 (0.73) -0.938 (0.39) 3326 0.39

Notes: 1.

2.

Column (1) report OLS estimates; (2) report IV estimates using characteristics of the nearest school as instruments for the characteristics of the school actually attended by the student (including distance and performance on the NEAT of the nearest school); (3) – (5) report quantile regression estimates at the 10th, 50th, and 90th percentile, respectively, using propensity score matched sample. Standard errors for cols (3)-(5) were estimated via bootstrap and p-values are reported in parentheses above.

3. 4. 5.

* denote significant at 5% level; ** significant at 1% level. All regressions include controls for gender, grade, age, past achievement, missing values and a constant, and are estimated on propensity score samples (see text for discussion of how each sample was generated). Base category: Female, in grades 1-3 or not enrolled.

D. Using ALL NOT-NEAR as treatment (GAIN MODEL) ENGLISH

(1) OLS Household Characteristics 0.117 father’s educ in yrs mother’s educ in yrs aspiration index

(2) IV

(3) q10

(4) q50

(5) q90

(0.09) -0.025 (0.68) 0.592 (0.06) -0.046 (0.76)

0.078 (0.69) -0.122 (0.71) 0.601 (0.19) -0.415 (0.67)

0.176 (0.16) -0.166 (0.12) 0.249 (0.65) -0.170 (0.59)

0.115 (0.16) 0.097 (0.24) 0.505 (0.16) 0.020 (0.91)

0.055 (0.72) 0.055 (0.72) 0.840 (0.32) 0.401 (0.22)

(0.00)** -0.216 (0.64) -0.026 (0.03)*

6.286 (0.81) -5.549 (0.45) -0.076 (0.70)

-4.444 (0.02)* 1.459 (0.03)* -0.004 (0.90)

-3.634 (0.02)* -0.317 (0.63) -0.056 (0.01)**

-1.702 (0.50) -2.301 (0.04)* -0.033 (0.09)

-0.199 (0.37) 0.016 (0.03)* -0.362 (0.03)* 3167 0.11

0.404 (0.03)* -0.008 (0.11) -0.195 (0.15) 3167 0.05

1.377 (0.00)** -0.036 (0.00)** -0.311 (0.14) 3167 0.06

log(per HH member expenditure) School Attended Characteristics -2.845 private indicator if school streaming student per teacher

Teacher Characteristics (School Avg) 0.286 1.217 Experience Experience2 Age Observations R-sqd / Pseudo R2

(0.09) -0.003 (0.50) -0.253 (0.03)* 3326 0.12

(0.32) -0.027 (0.35) -0.321 (0.64) 3326 0.06

MATH

(1) OLS Household Characteristics 0.074 father’s educ in yrs mother’s educ in yrs aspiration index

(2) IV

(3) q10

(4) q50

(5) q90

(0.27) -0.024 (0.70) 0.707 (0.05) -0.318 (0.06)

0.110 (0.62) 0.050 (0.90) 0.584 (0.23) -0.297 (0.79)

0.070 (0.54) -0.128 (0.33) 0.316 (0.61) -0.538 (0.04)*

0.058 (0.49) -0.012 (0.91) 0.652 (0.08) -0.577 (0.00)**

0.123 (0.48) -0.025 (0.89) 2.065 (0.01)* -0.025 (0.96)

(0.00)** -1.450 (0.04)* -0.037 (0.00)**

-9.814 (0.75) -7.417 (0.37) -0.243 (0.29)

-2.599 (0.27) -1.222 (0.21) -0.045 (0.28)

-3.443 (0.01)** -1.400 (0.03)* -0.044 (0.01)**

-6.325 (0.01)** -2.961 (0.02)* -0.051 (0.01)**

0.438 (0.28) -0.003 (0.83) -0.393 (0.10) 3167 0.12

0.323 (0.08) -0.007 (0.16) -0.201 (0.08) 3167 0.05

0.545 (0.13) -0.003 (0.80) -0.495 (0.04)* 3167 0.07

log(per HH member expenditure) School Attended Characteristics -4.388 private indicator if school streaming student per teacher

Teacher Characteristics (School Avg) 0.361 1.100 Experience Experience2 Age Observations R-sqd / Pseudo R2

(0.09) -0.006 (0.28) -0.295 (0.02)* 3326 0.12

(0.38) -0.011 (0.69) -0.711 (0.34) 3326 0.05

Notes: 1.

2.

Column (1) report OLS estimates; (2) report IV estimates using characteristics of the nearest school as instruments for the characteristics of the school actually attended by the student (including distance and performance on the NEAT of the nearest school); (3) – (5) report quantile regression estimates at the 10th, 50th, and 90th percentile, respectively, using propensity score matched sample. Standard errors for cols (3)-(5) were estimated via bootstrap and p-values are reported in parentheses above.

3. 4. 5.

* denote significant at 5% level; ** significant at 1% level. All regressions include controls for gender, grade, age, past achievement, missing values and a constant, and are estimated on propensity score samples (see text for discussion of how each sample was generated). Base category: Female, in grades 1-3 or not enrolled.

FIGURE 1. Diagram illustrating the identification strategy: Suppose child A and child B are observably similar. We then use the outcome of child A to represent the outcome of child B had child B not chosen the “treatment”: to “move” by attending a school outside one’s barangay, to attend a school that is not the nearest (“not-near”), and to “transfer” schools between periods. A. Movers At time t, A

B

Brgy Lahug

Brgy Tisa

B. Not-near

At time t,

school J A B

school K

C. Transfers Between t-1 and t, A

B

school J

school K

D. Transfers*Movers Between t-1 and t,

J A

K

B

Brgy Lahug

Brgy Tisa

40

FIGURE 2. HISTOGRAMS OF PROPENSITY SCORES (Propensity Score Samples. See text for discussion.) Panel 2A. Selection: MOVERS (Enrolled Sample)

(Enrolled in Primary)

mover==0

mover==1

mover==0

500.00

0.00 0.02

0.22

0.41

0.61

0.80

1.00

0.02

0.22

0.41

0.61

0.80

1.00

Frequency

500.00

Frequency

mover==1

1000. 00

1000. 00

0.00 0.02

0.21

0.41

0.61

Pr(mover)

0.80

1.00

0.21

0.02

0.41

0.61

0.80 1.00

Pr(mover)

Histograms by mover

Histograms by mover

Panel 2B. Selection: NOT ATTENDING NEAREST (Enrolled Sample)

(Enrolled in Primary) notnear==1

notnear==0 1500. 00

1000. 00

1000. 00

500.00

500.00

0.00 0.08

0.26

0.45

0.63

0.82

1.00

0.08

Pr(notnear)

0.26

Histograms by notnear

0.45

0.63

0.82

1.00

Frequency

Frequency

notnear==0 1500. 00

notnear==1

0.00 0.04

0.23

0.43

0.62

0.81

1.00

0.04

Pr(notnear)

0.23

Histograms by notnear

0.43

0.62

0.81

1.00

Panel 2C. Selection: TRANSFERRED SCHOOLS (Enrolled Sample)

(Enrolled in HS)

transf2==0

transf2==1

transf2==0

2000. 00

transf2==1

15.00

1500. 00

10.00

1000. 00 5.00

0.00 0.01

0.20

0.39

0.58

0.77

0.95

0.01

0.20

0.39

0.58

0.77

0.95

Frequency

Frequency

500.00

0.00 0.04

Pr(transf2)

0.20

0.37

0.54

0.71

0.88

0.04

0.20

0.37

0.54

0.71 0.88

Pr(transf2)

Histograms by =1 if transfer in last 2 years

Histograms by =1 if transfer in last 2 years

Panel 2D. Selection: TRANSFERRED TO SCHOOL OUTSIDE BRGY RESIDED (Enrolled Sample)

(Enrolled in HS)

tranmov==0

tranmov==1

tranmov==0

2000. 00

tranmov==1

15.00

1500. 00

10.00

1000. 00 5.00

0.00 0.00

0.19

0.39

0.58

0.77 0.96

0.00

0.19

0.39

0.58

0.77 0.96

Frequency

Frequency

500.00

0.00 0.04

Pr(tranmov)

0.21

0.39

0.57

0.74 0.92

0.04

Pr(tranmov)

Histograms by tranmov

0.21

Histograms by tranmov

42

0.39

0.57

0.74 0.92

APPENDIX 1

Sample attrition and selection

Live Births in 33 Sample Barangays of Metro Cebu Of which: Twin Births Refusals Missed by Survey (discovered later) Birth Interview Too Late

3,289 27 97 58 22

(0.8 percent) (2.9 percent) (1.8 percent) (0.7 percent)

Live Births in Metro Cebu with Birth Interview Of which: Migrated Out of Metro Cebu by Age 2 Child Died by Age 2 Refusal (at later date)

3,085 318 156 50

(10.3 percent) (5.1 percent) (1.6 percent)

Still in Sample When Child is 8 Years Old Of which: Migrated Out/Could Not Find 29 Child Died

2,231 31 8

(1.4 percent) (0.4 percent)

30

Still in Sample When Child is 11 Years Old Of which: Never Enrolled in School Not Tested (e.g. refusal) Had Younger Sibling of School Age

2,192 9 (0.4 percent) 13 (0.6 percent) 1,261 (57.5 percent)

Still in Sample When Child is 13 Years Old Of which: Not Tested Had Younger Sibling of School Age

2,229 31 (1.4 percent) 1,155 (51.8 percent)

TOTAL OBSERVATIONS (INDEX AND SIBLING CHILDREN) WITH TEST SCORES 3,326 Of which:

Missing School Information Missing Teacher Information

905 (27.21 percent) 913 (27.45 percent)

29

31 children lost between 1991/92 and 1994/95 surveys is a net figure. 77 children interviewed in 199192 could not be located in 1994-95, but 46 children in the 1983-86 sample who were not found in 1991-92 were found in 1994-95.

30

A new sample of about 500 women were added to the original sample in the 1994/95 survey.

APPENDIX 2. GAIN vs. LEVELS Model This section shows that the levels estimator linearly controlling for a lagged dependent variable and the difference estimator may provide estimates that bracket the causal effect of interest. Guryan (1999) and Imbens, Liebman, and Eissa (1997) make similar points. As discussed in the text, the following LEVELS estimator, δL, in Υt = βΥt –1 + ΧδL + u where Υt are test scores and Χ include individual, household, and school characteristics, is equivalent to the GAIN estimator, δG, in: Υt - Υt –1 = ΧδG + v with the restriction that β = 1. (Note: We’ve suspended the time(t) and individual(i) subscript to simplify notation, i.e., X=Xit, u=uit, etc.) Hence if in the limit β = 1, then −1 p lim δ$ G = p lim δ$ L = δ + Qxx cov( X , u )

where δ are our parameters of interest and Qxx is a positive definite matrix equal to the plim of (X’X/n). If selection on school choice is “positive”, such that parents choose high inputs (Χ↑) for high ability or highly motivated children (u↑ or v↑), the estimators are then upward-biased. That is, plim δG–hat = plim δL–hat ≥ δ. Conversely, if selection on school choice is “negative”, such that parents choose high inputs (Χ↑) for children with less ability or less motivation (u↓), then plim δG–hat = plim δL–hat ≤ δ. Let us then consider the case when β ≠ 1. The levels estimator is given by:

~ ~

~

δ$ L = ( X ' X )−1 X '(Yt )

~ where X are the residuals from a regression of X on Yt −1. Then, ~ −1 p lim δ$ = Q ~~ cov( X , Y ) L

t

XX

~ ~ −1 $ δ + u) ~~ cov( X , βYt −1 + Xδ + X = QXX −1

~~ cov( X − γYt −1 , βYt −1 + u ) = δ + QXX

~ where X$ are the predicted values from a regression of X on Yt −1 , which are orthogonal to X , ~ ~ cov( X , Yt −1 ) X' X Y 'Y ~~ = p lim( γ = , QXX ), and QYt −1Yt −1 = p lim( t −1 t −1 ). QYt −1Yt −1 n n Expanding and rearranging, −1 p lim δ$ = δ + Q ~~ cov( u, X ) − [Q ~~ Q L

XX

XX

Yt −1Yt −1

]−1 cov( X , Yt −1 )cov( u, Yt −1 )

−1

−1 ~~ cov( u, X )[1 − {cov( u, X )QY Y } cov( X , Yt −1 )cov( u, Yt −1 )] = δ + QXX t −1 t −1

44

If we are willing to make the assumption that cov(u, Υt –1) = 0, that is zero serial correlation, then it is easy to see that δL is upward-biased when selection is positive and downward-biased when selection is negative. It is, however, difficult to justify this assumption as the test scores are only two years apart. Suppose then that we assume the structure of the serial correlation is such that cov(u, Υt –1) < 0, e.g. children who scored low in 1994 exert more effort to do well in 1996. Then if selection on school choice is “positive”, such that parents choose high inputs (Χ↑) for high ability or highly motivated children (u↑,Υt –1↑), the levels estimator is upward-biased. That is, plim δL–hat ≥ δ. Conversely, if selection on school choice is “negative”, such that parents choose high inputs (Χ↑) for children with less ability or less motivation (u↓,Υt –1↓), then plim δL–hat ≤ δ. If the serial correlation goes in the other direction, we would still have the same result as long as the term {cov(Χ,Υt –1) cov(u,Υt –1)}/{cov(Χ,u)QΥt –1,Υt –1} is between 0 and 1. 1

On the other hand, the gain estimator is given by:

δ$ G = ( X ' X )−1 X '(Yt − Yt −1 ) −1 p lim δ$ G = Q XX cov( Xδ + v t − v t −1 , X ) −1

= δ + Q XX cov( v t − v t −1 , X ) = δ − Q XX −1 cov( v t −1 , X ) Hence, δG is upward-biased when selection on lagged outcomes is negative and upward-biased when selection on lagged outcomes is positive. Therefore, when selection is positive, δG ≤ δ ≤ δL and when selection is negative, δG ≥ δ ≥ δL

1

Cov(vit , Xit) = 0 since one can think of the error term vit as vit = vi + vit where vi is an individual-specific time-invariant component which gets differenced out under the gain model. Hence any current unobservable component (vit) will not have been observed by parents to use as information in making schooling and academic input decisions (Xit) at the beginning of the period. On the other hand, parents may use information revealed from their children’s performance in the last period which is unobservable to the econometrician (vit) to update their decisions.

45

APPENDIX 3. QUANTILE REGRESSION ESTIMATES FOR MATCHING GROUPS A. MOVERS LEVELS ENGLISH MATH SAMPLE ALL PRIMARY ALL PRIMARY SE Coef. Coef. SE Coef. SE Coef. SE

q10 father's educ mother's educ aspiration ln(PCE) private streaming student/teacher teacher exper experience sqd avg teacher age

0.313 0.081 0.695 0.248 -3.798 -0.036 -0.067 0.197 0.001 -0.327

(0.128)

q50 father's educ mother's educ aspiration ln(PCE) private streaming student/teacher teacher exper experience sqd avg teacher age

0.200 0.147 1.153 0.301 -1.143 -0.687 -0.036 0.538 -0.007 -0.422

(0.083)

q90 father's educ mother's educ aspiration ln(PCE) private streaming student/teacher teacher exper experience sqd avg teacher age

0.399 0.251 1.289 0.816 0.169 -1.186 -0.022 0.788 -0.018 -0.327

(0.118)

N 3232 .10 Pseudo R2 0.322 .50 Pseudo R2 0.432 .90 Pseudo R2 0.426

(0.136) (0.646) (0.229) (2.463) (0.743) (0.05) (0.3) (0.007) (0.2)

(0.07) (0.369) (0.192) (1.338) (0.541) (0.022) (0.168) (0.005) (0.104)

(0.114) (0.603) (0.251) (2.17) (0.818) (0.019) (0.246) (0.007) (0.156)

0.189 0.056 2.000 0.081 -4.674 0.094 2.E-05 0.337 -0.001 -0.404

(0.14)

0.200 0.085 1.189 0.608 1.448 -0.243 -0.038 0.738 -0.015 -0.396

(0.1)

0.534 0.330 0.994 0.644 2.871 -0.772 -0.016 0.948 -0.024 -0.294

(0.13)

2052 0.246 0.382 0.412

(0.15) (0.69) (0.36) (3.42) (0.93) (0.04) (0.29) (0.01) (0.18)

(0.1) (0.44) (0.23) (1.67) (0.66) (0.02) (0.23) (0.01) (0.14)

(0.16) (0.87) (0.34) (4.01) (0.93) (0.02) (0.32) (0.01) (0.18)

0.230 0.027 -0.354 -0.032 -1.747 -0.158 -0.031 0.253 -0.001 -0.389

(0.12)

0.172 0.041 1.003 0.212 -3.619 -2.064 -0.069 0.466 -0.014 -0.218

(0.06)

0.275 0.089 1.328 0.286 0.311 -3.038 -0.043 0.501 -0.019 0.003

(0.11)

3232 0.358 0.434 0.363

(0.09) (0.49) (0.2) (1.44) (0.75) (0.02) (0.24) (0.01) (0.15)

(0.07) (0.41) (0.21) (1.36) (0.54) (0.02) (0.17) (0.01) (0.11)

(0.1) (0.63) (0.23) (1.85) (0.93) (0.02) (0.31) (0.01) (0.24)

0.155 -0.133 0.429 -0.010 -0.355 0.445 -0.035 0.671 -0.010 -0.489

(0.14)

0.175 0.083 1.140 0.175 -3.394 -1.825 -0.066 0.531 -0.018 -0.163

(0.08)

0.144 0.288 1.416 0.262 -3.025 -2.516 -0.030 0.360 -0.019 0.071

(0.12)

2052 0.288 0.391 0.371

(0.12) (0.61) (0.26) (2.07) (0.92) (0.01) (0.32) (0.01) (0.17)

(0.11) (0.4) (0.24) (1.76) (0.73) (0.03) (0.22) (0.01) (0.13)

(0.16) (0.77) (0.3) (2.18) (1.) (0.03) (0.36) (0.01) (0.25)

GAIN ENGLISH MATH ALL PRIMARY ALL PRIMARY SE Coef. Coef. SE Coef. SE Coef. SE 0.190 -0.215 -0.123 -0.274 -4.848 1.515 -0.001 -0.263 0.018 -0.388

(0.126)

0.086 0.060 0.449 0.040 -4.519 -0.263 -0.048 0.342 -0.006 -0.229

(0.081)

0.163 0.070 0.906 0.176 -0.384 -1.656 -0.028 1.142 -0.029 -0.267

(0.114)

3232 0.112 0.054 0.058

(0.114) (0.566) (0.239) (1.832) (0.751) (0.022) (0.234) (0.008) (0.176)

(0.082) (0.383) (0.195) (1.339) (0.637) (0.028) (0.182) (0.005) (0.138)

(0.134) (0.8) (0.371) (2.816) (1.277) (0.023) (0.356) (0.009) (0.205)

0.205 -0.480 0.444 -0.101 -7.476 1.451 -0.005 -0.283 0.015 -0.288

(0.19)

-0.005 0.014 0.462 0.585 -2.291 -0.052 -0.038 0.319 -0.008 -0.131

(0.09)

0.118 0.219 1.087 0.384 -2.204 -1.443 -0.022 0.972 -0.029 -0.216

(0.16)

2052 0.121 0.062 0.063

(0.14) (0.7) (0.32) (5.41) (1.12) (0.04) (0.36) (0.01) (0.19)

(0.1) (0.41) (0.23) (2.02) (0.72) (0.03) (0.25) (0.01) (0.15)

(0.18) (0.72) (0.46) (3.53) (1.27) (0.02) (0.44) (0.01) (0.22)

0.092 -0.076 0.360 -0.594 -2.907 -0.774 -0.050 0.323 -0.001 -0.370

(0.12)

0.034 -0.032 0.694 -0.410 -3.917 -1.399 -0.034 0.265 -0.008 -0.147

(0.08)

0.169 0.005 2.228 0.023 -7.977 -3.393 -0.048 0.271 0.005 -0.553

(0.14)

(0.12) (0.53) (0.27) (2.04) (0.95) (0.04) (0.33) (0.01) (0.24)

(0.08) (0.31) (0.16) (1.43) (0.6) (0.02) (0.23) (0.01) (0.12)

(0.18) (0.68) (0.46) (2.67) (1.28) (0.02) (0.33) (0.01) (0.24)

3232 0.126 0.052 0.072

Notes: 1. 2. 3.

Bootstrap standard errors are reported in parentheses. All regressions include controls for gender, grade, age, past achievement, missing values and a constant, and are estimated on propensity score samples (see text for discussion of how each sample was generated). Base category: Female, in grades 1-3 or not enrolled.

46

-0.002 -0.162 0.912 -0.674 -2.338 0.251 -0.067 0.394 -0.001 -0.368

(0.17)

0.042 -0.002 0.696 -0.439 -0.803 -1.343 -0.038 0.488 -0.013 -0.180

(0.1)

0.030 0.158 1.955 -0.086 -7.360 -3.500 -0.043 0.558 -0.011 -0.429

(0.15)

2052 0.121 0.071 0.104

(0.23) (0.78) (0.36) (2.75) (1.31) (0.04) (0.51) (0.01) (0.27)

(0.11) (0.51) (0.23) (2.25) (0.75) (0.02) (0.26) (0.01) (0.14)

(0.17) (0.8) (0.45) (3.04) (1.26) (0.02) (0.45) (0.01) (0.3)

B. NOT-NEAR LEVELS ENGLISH MATH SAMPLE ALL PRIMARY ALL PRIMARY Coef. SE Coef. SE Coef. SE Coef. SE

q10 father's educ mother's educ aspiration ln(PCE) private streaming student/teacher teacher exper experience sqd avg teacher age

0.273 0.121 1.039 0.243 -4.068 0.363 -0.075 0.299 4.E-04 -0.427

(0.102)

q50 father's educ mother's educ aspiration ln(PCE) private streaming student/teacher teacher exper experience sqd avg teacher age

0.195 0.109 1.338 0.348 -0.188 -0.745 -0.040 0.649 -0.010 -0.403

(0.069)

q90 father's educ mother's educ aspiration ln(PCE) private streaming student/teacher teacher exper experience sqd avg teacher age

0.390 0.321 1.125 0.768 -1.281 -1.596 -0.024 0.740 -0.015 -0.409

(0.123)

N 3167 .10 Pseudo R2 0.318 .50 Pseudo R2 0.431 .90 Pseudo R2 0.425

(0.108) (0.531) (0.21) (2.315) (0.768) (0.051) (0.247) (0.006) (0.199)

(0.077) (0.369) (0.192) (1.146) (0.566) (0.018) (0.184) (0.004) (0.12)

(0.126) (0.639) (0.277) (2.363) (0.927) (0.02) (0.265) (0.007) (0.143)

0.132 0.135 1.295 -0.031 -2.381 0.310 4.E-04 0.371 0.002 -0.468

(0.119)

0.080 0.142 1.134 0.542 1.750 -1.015 -0.035 0.767 -0.012 -0.421

(0.094)

0.351 0.270 1.210 0.856 -0.887 -2.136 -0.027 0.924 -0.021 -0.341

(0.141)

2385 0.235 0.379 0.406

(0.138) (0.655) (0.289) (3.066) (0.841) (0.053) (0.286) (0.008) (0.148)

(0.089) (0.37) (0.236) (1.608) (0.503) (0.021) (0.185) (0.005) (0.111)

(0.144) (0.692) (0.272) (3.859) (1.008) (0.018) (0.228) (0.007) (0.152)

0.229 0.021 -0.041 -0.061 -2.041 -0.438 -0.036 0.462 -0.005 -0.458

(0.104)

0.173 0.052 0.983 0.164 -3.239 -2.106 -0.067 0.530 -0.016 -0.223

(0.072)

0.210 0.146 1.801 0.225 0.701 -3.282 -0.059 0.592 -0.016 -0.105

(0.136)

3167 0.356 0.434 0.364

(0.101) (0.597) (0.26) (1.459) (0.792) (0.013) (0.273) (0.007) (0.146)

(0.081) (0.374) (0.208) (1.403) (0.588) (0.02) (0.193) (0.006) (0.121)

(0.107) (0.64) (0.245) (1.571) (0.847) (0.025) (0.298) (0.007) (0.206)

0.158 -0.121 0.725 0.032 -0.243 -0.729 -0.039 0.823 -0.015 -0.464

(0.131)

0.193 0.008 1.100 0.150 -2.507 -1.989 -0.068 0.543 -0.015 -0.247

(0.074)

0.116 0.186 1.714 0.318 -2.610 -3.165 -0.047 0.522 -0.014 -0.109

(0.149)

2385 0.279 0.379 0.362

(0.094) (0.536) (0.276) (2.185) (0.743) (0.012) (0.322) (0.009) (0.139)

(0.106) (0.425) (0.253) (1.821) (0.642) (0.02) (0.198) (0.006) (0.122)

(0.121) (0.685) (0.322) (1.971) (1.042) (0.023) (0.306) (0.007) (0.218)

GAIN ENGLISH MATH ALL PRIMARY ALL PRIMARY SE Coef. SE Coef. SE Coef. SE Coef. 0.176 -0.166 0.249 -0.170 -4.444 1.459 -0.004 -0.199 0.016 -0.362

(0.121)

0.115 0.097 0.505 0.020 -3.634 -0.317 -0.056 0.404 -0.008 -0.195

(0.077)

0.055 0.055 0.840 0.401 -1.702 -2.301 -0.033 1.377 -0.036 -0.311

(0.132)

3167 0.110 0.053 0.059

(0.123) (0.59) (0.279) (1.859) (0.73) (0.031) (0.266) (0.006) (0.176)

(0.083) (0.406) (0.207) (1.232) (0.65) (0.025) (0.21) (0.005) (0.13)

(0.155) (0.815) (0.387) (2.607) (1.131) (0.018) (0.316) (0.008) (0.198)

0.204 -0.389 0.461 -0.428 -4.918 1.202 -0.004 -0.008 0.015 -0.502

(0.149)

0.013 0.082 0.562 0.263 -1.575 -0.223 -0.051 0.467 -0.009 -0.200

(0.104)

-0.037 0.208 1.349 0.201 -0.573 -0.947 -0.021 1.057 -0.025 -0.319

(0.154)

2385 0.117 0.064 0.062

(0.126) (0.639) (0.29) (3.977) (0.826) (0.035) (0.291) (0.009) (0.196)

(0.115) (0.422) (0.216) (1.819) (0.642) (0.027) (0.246) (0.007) (0.117)

(0.205) (0.746) (0.475) (3.318) (1.429) (0.022) (0.351) (0.01) (0.211)

0.070 -0.128 0.316 -0.538 -2.599 -1.222 -0.045 0.438 -0.003 -0.393

(0.108)

0.058 -0.012 0.652 -0.577 -3.443 -1.400 -0.044 0.323 -0.007 -0.201

(0.068)

0.123 -0.025 2.065 -0.025 -6.325 -2.961 -0.051 0.545 -0.003 -0.495

(0.152)

(0.121) (0.585) (0.283) (2.337) (0.925) (0.038) (0.375) (0.01) (0.227)

(0.083) (0.354) (0.16) (1.459) (0.603) (0.017) (0.211) (0.005) (0.12)

(0.161) (0.897) (0.457) (2.258) (1.357) (0.021) (0.332) (0.01) (0.266)

-1.E-04 -0.324 0.745 -0.716 0.858 -1.469 -0.053 0.633 -0.009 -0.349

(0.146)

0.086 -0.009 0.545 -0.538 -1.098 -1.295 -0.044 0.373 -0.010 -0.157

(0.087)

0.145 0.031 2.377 -0.086 -7.464 -2.577 -0.046 0.038 0.013 -0.564

(0.144)

3167 0.120 0.050 0.072

Notes: 1. 2. 3.

Bootstrap standard errors are reported in parentheses. All regressions include controls for gender, grade, age, past achievement, missing values and a constant, and are estimated on propensity score samples (see text for discussion of how each sample was generated). Base category: Female, in grades 1-3 or not enrolled.

2385 0.126 0.066 0.083

(0.173) (0.669) (0.395) (2.831) (0.972) (0.044) (0.361) (0.01) (0.26)

(0.097) (0.458) (0.232) (1.632) (0.69) (0.016) (0.212) (0.006) (0.12)

(0.167) (0.714) (0.468) (2.373) (1.204) (0.018) (0.373) (0.011) (0.275)

C. TRANSFERS (HIGH SCHOOL PROPENSITY SAMPLE ONLY) LEVELS ENGLISH MATH SE SE Coef. Coef.

ENGLISH SE Coef.

q10 father's educ mother's educ aspiration ln(PCE) private streaming student/teacher teacher exper experience sqd avg teacher age

0.371 -1.056 -1.312 1.921 9.737 -0.864 0.378 -3.117 0.082 1.041

(0.625)

0.351 -1.377 0.064 2.177 13.857 -5.372 -0.082 -1.582 0.013 1.738

(0.721)

q50 father's educ mother's educ aspiration ln(PCE) private streaming student/teacher teacher exper experience sqd avg teacher age

0.539 -0.522 -0.574 2.586 19.928 -1.273 -0.020 -2.143 0.093 0.411

(0.525)

0.612 -2.115 -2.281 4.890 11.424 5.285 0.077 -2.628 0.092 0.603

(0.529)

q90 father's educ mother's educ aspiration ln(PCE) private streaming student/teacher teacher exper experience sqd avg teacher age

0.098 0.375 3.269 1.999 10.427 3.044 0.017 0.398 0.020 -0.596

(0.696)

0.931 -1.716 -8.608 3.404 13.680 10.906 1.269 0.864 -0.002 -0.344

(0.897)

N 75 .10 Pseudo R2 0.679 .50 Pseudo R2 0.594 .90 Pseudo R2 0.609

(0.712) (4.928) (1.776) (12.319) (6.288) (0.286) (2.399) (0.081) (1.572)

(0.803) (5.049) (2.166) (11.201) (8.602) (0.353) (2.108) (0.07) (1.371)

(0.738) (5.027) (2.008) (12.028) (7.731) (0.402) (2.556) (0.071) (1.852)

0.250 0.468 0.151 1.029 -1.108 1.976 0.298 -1.294 -0.029 1.274

(0.632)

0.781 0.309 -4.180 1.260 0.673 -4.318 -0.091 -1.590 0.029 0.559

(0.604)

2.395 -1.051 -2.987 1.385 -3.295 2.861 -0.241 0.433 -0.008 -0.347

(0.777)

75 0.659 0.557 0.522

(0.653) (4.963) (1.85) (15.251) (6.553) (0.293) (2.382) (0.087) (1.634)

(0.711) (5.322) (1.744) (11.982) (6.094) (0.308) (2.288) (0.082) (1.46)

(0.814) (6.565) (2.449) (14.762) (7.21) (0.451) (2.826) (0.084) (2.25)

GAIN

75 0.260 0.169 0.303

(0.574) (4.854) (1.71) (15.73) (6.5) (0.327) (3.021) (0.092) (1.897)

(0.636) (5.079) (2.897) (17.39) (8.201) (0.349) (2.687) (0.087) (1.689)

(1.024) (7.648) (3.871) (43.624) (19.487) (0.964) (8.332) (0.208) (4.504)

MATH SE Coef. 0.520 -0.250 -1.994 0.855 -11.864 -2.390 0.165 1.235 -0.083 0.628

(0.655)

0.266 -0.117 -0.316 -0.190 -3.752 -3.292 -0.010 -1.063 0.073 -0.872

(0.812)

0.556 1.030 11.627 -4.201 -6.389 -12.507 -0.876 3.183 -0.042 -1.990

(1.188)

(0.718) (4.487) (1.355) (15.44) (6.446) (0.306) (3.011) (0.097) (2.1)

(0.877) (6.119) (2.368) (14.022) (8.149) (0.392) (3.04) (0.084) (2.253)

(1.143) (10.16) (3.473) (20.456) (7.93) (0.566) (4.59) (0.123) (2.863)

75 0.173 0.152 0.347

Notes: 1. 2. 3.

Bootstrap standard errors are reported in parentheses. All regressions include controls for gender, grade, age, past achievement, missing values and a constant, and are estimated on propensity score samples (see text for discussion of how each sample was generated). Base category: Female, in grades 1-3 or not enrolled.

D. TRANSFER*MOVER (HIGH SCHOOL PROPENSITY SAMPLE ONLY) LEVELS ENGLISH MATH SE SE Coef. Coef.

GAIN ENGLISH MATH SE SE Coef. Coef.

q10 father's educ mother's educ aspiration ln(PCE) private streaming student/teacher teacher exper experience sqd avg teacher age

0.747 -1.598 2.389 0.265 21.765 -0.209 0.021 4.168 -0.029 -1.978

(0.527)

1.117 -1.691 -1.242 1.342 15.546 6.029 0.100 4.246 0.021 -3.150

(0.643)

q50 father's educ mother's educ aspiration ln(PCE) private streaming student/teacher teacher exper experience sqd avg teacher age

0.753 -0.444 2.068 1.526 27.635 -1.272 0.032 -0.422 0.045 0.174

(0.679)

1.028 -1.641 -3.898 -0.216 49.939 -3.067 0.278 -1.518 0.190 -0.823

(0.52)

q90 father's educ mother's educ aspiration ln(PCE) private streaming student/teacher teacher exper experience sqd avg teacher age

1.534 -0.143 -4.770 0.478 24.742 6.167 0.276 -0.270 0.071 -0.703

(0.61)

1.037 -0.841 -5.244 2.093 25.158 10.318 0.738 -0.823 0.106 -0.503

(0.762)

N 75 .10 Pseudo R2 0.672 .50 Pseudo R2 0.644 .90 Pseudo R2 0.662

(0.608) (4.845) (1.549) (17.203) (6.999) (0.309) (2.67) (0.106) (1.521)

(0.569) (3.83) (1.847) (12.209) (6.172) (0.219) (1.87) (0.058) (1.413)

(0.565) (3.416) (1.719) (9.423) (5.648) (0.243) (2.276) (0.065) (1.657)

0.807 -0.694 0.868 1.726 7.613 10.068 0.543 -0.542 0.003 0.187

(0.637)

1.042 0.407 -6.391 -1.040 7.983 5.181 0.211 1.023 -0.051 0.321

(0.561)

1.193 -0.802 4.346 -1.575 21.287 3.579 0.199 3.349 0.006 -1.964

(0.897)

75 0.661 0.602 0.529

(0.798) (4.799) (1.564) (26.685) (8.327) (0.423) (3.077) (0.121) (1.741)

(0.519) (6.478) (1.554) (19.614) (7.442) (0.343) (2.501) (0.088) (1.736)

(0.831) (8.475) (2.808) (21.201) (8.614) (0.421) (3.639) (0.124) (2.055)

75 0.318 0.280 0.475

(0.484) (4.133) (1.586) (18.939) (7.196) (0.407) (3.321) (0.111) (2.118)

(0.598) (4.332) (1.837) (17.95) (6.091) (0.312) (2.442) (0.084) (1.931)

(0.848) (5.407) (2.855) (15.262) (9.537) (0.403) (3.203) (0.084) (2.752)

0.369 -0.060 -2.905 1.011 -20.925 -5.467 0.054 0.623 -0.120 1.645

(0.743)

0.358 -0.415 0.368 -0.235 3.141 -0.378 0.415 1.116 0.053 -1.876

(0.706)

-0.632 -0.371 16.564 -2.646 16.476 -5.573 0.422 6.240 -0.011 -3.958

(1.217)

(0.669) (6.411) (1.809) (30.186) (8.471) (0.38) (3.762) (0.14) (2.182)

(0.873) (7.653) (2.884) (25.452) (14.017) (0.487) (4.31) (0.125) (2.563)

(1.156) (8.605) (3.099) (23.244) (12.375) (0.61) (4.476) (0.126) (2.96)

75 0.179 0.155 0.504

Notes: 1. 2. 3.

Bootstrap standard errors are reported in parentheses. All regressions include controls for gender, grade, age, past achievement, missing values and a constant, and are estimated on propensity score samples (see text for discussion of how each sample was generated). Base category: Female, in grades 1-3 or not enrolled.