School quality signals and attendance in rural Guatemala

Economics of Education Review 30 (2011) 1445–1455

Contents lists available at ScienceDirect

Economics of Education Review journal homepage: www.elsevier.com/locate/econedurev

School quality signals and attendance in rural Guatemala Jeffery H. Marshall a,∗ a

Instituto de Investigación y Evaluación Educativas y Sociales, Universidad Pedagógica Nacional Francisco Morazán (UPNFM), Tegucigalpa, Honduras

a r t i c l e

i n f o

Article history: Received 23 October 2010 Received in revised form 12 July 2011 Accepted 13 July 2011 JEL classification: D1 I21 I28 Keywords: School attendance Household decision-making School quality Economics of education Guatemala

a b s t r a c t This study analyzes school dropout in rural Guatemala using event history data and unusually detailed data on schools and teachers. Significant results for language of instruction, teacher education and fighting between students demonstrate the importance of accounting for school context influences on an outcome that has, historically, been analyzed mainly as a function of family background. Less support is found for the contention that dropout is lower in schools that are better at maximizing student achievement. Finally, using interaction analysis some of the school effects vary significantly by student gender and ethnicity. The various linkages between school features and dropout highlight the complicated reality of identifying the kinds of features of schools that are valued by poor families. © 2011 Elsevier Ltd. All rights reserved.

For more than 30 years researchers have used household survey data to analyze variation in grade attainment and school enrollment rates in developing countries (Buchmann & Hannum, 2001; Chernichovsky, 1985). These studies have relied heavily on family background and community characteristics as predictors of attendance, and the most consistent explanation for why some young children are not in school is simply that their families are too poor to send them. This “poverty explanation” has helped justify a range of price-reducing interventions around the globe, including abolishing school fees (World Bank, 2009), building more schools, offering free meals in school, and providing families with targeted cash or in-kind transfers (Filmer & Shady, 2008). The effectiveness of these efforts to make school more affordable for poor people is difficult to judge. One concern

∗ Correspondence address: 4610 Ironstone Lane, West Lafayette, IN 47906, USA. Tel.: +1 765 838 8096. E-mail address: [email protected] 0272-7757/$ – see front matter © 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.econedurev.2011.07.011

with the historical research emphasis on family background is that the direct effect of socioeconomic status will likely be overstated if the distribution of school features is itself correlated with social class background, which seems likely. As a result, policymakers may feel pressure to make schooling more affordable for poor people instead of considering ways to make schools more responsive to their needs. This is not to say that abolishing fees or providing scholarships are ineffective: the evidence from scholarship programs like Progresa in Mexico (Shultz, 2004), and recent initiatives to abolish school fees in Africa (UNICEF, 2010), reinforces the potential for addressing the price of schooling. But a lingering concern is that researchers and policymakers may be missing some important school effects that help explain why some children are not in school, including the possibility that the school is not providing much in the way of learning. Underlying this discussion is the central role of information when households are weighing expected benefits against costs, as in the forward-thinking human capital model of school attendance (Becker, 1967). For the

1446

J.H. Marshall / Economics of Education Review 30 (2011) 1445–1455

average developing country household the challenges of evaluating the payoffs to schooling are considerable, beginning first with the actual learning that takes place inside the school walls. We know little about how households evaluate school quality and learning when deciding on attendance, and for poor parents with low levels of education the informational challenges would appear to be greatest. Researchers have been including more and more features of schools in large-sample analyses of attendance in developing countries (Lloyd, Mete, & Grant, 2009). For example, Case and Deaton (1999) find that school dropout is more likely in South Africa when class sizes are larger, while Fuller, Singer and Keiley (1995) show that desertion in Botswana is less likely when parental perceptions of school quality are higher. These kinds of results strongly suggest some form of interaction between the household and the local school, which in turn points to a number of important questions that require more attention. How do households decide which school features are important? How is this information collected? And how are these processes themselves mediated by variables like SES? With better data and more complete explanatory frameworks researchers can continue to fill in these gaps. Accordingly, in this paper I bridge together elements of the family background and school effects research genres through analysis of event history data from rural Guatemala. The data include up to four years of information for roughly 850 children who were in first grade in 1999, and the year-specific multinomial dependent variable measures grade passing, failure and dropout. The independent variable set includes extensive information on student and family background together with timevarying measures of schools and teachers. The main finding is that primary school dropout in rural Guatemala is associated with school features such as fighting between students and the teacher’s ethnicity. Additionally, through interaction analysis I find that the household’s apparent responsiveness to certain features of schools varies significantly by the child’s gender and family ethnicity. These linkages go beyond most studies of attendance and grade attainment in developing areas. They also help extend the analysis of the family and the ways in which parents manage their children’s schooling to especially poor contexts where little is known about these processes. The paper proceeds as follows. Section 1 adapts Glewwe’s (2002) model of school attendance to a rural developing country setting, with an emphasis on the role of school quality signals. Section 2 discusses the Guatemalan context and data. Section 3 presents the empirical framework for testing the main hypotheses. Section 4 presents and discusses the results, and includes the concluding remarks. 1. Conceptual framework Formal models of school attendance commonly begin with a two-period utility function where households (or more specifically parents) maximize present and (discounted) future consumption. The depiction here borrows heavily from Glewwe (2002) general model, and applies it to a poor rural developing country context. This begins with

a simple utility function with form U = C1 + ıC2 . Current consumption is equal to total (non child) family income Y1 the child’s school investment (price p multiplied by total time devoted to school S) and the income generated by the child (Y1c ) in their non-school time. Future consumption is simply Y2 plus the k fraction of the child’s earnings in the post-schooling (i.e. adult) period that accrues to the household. Substituting these into the utility function gives: U = (Y1 + ıY2 ) − pS + (1 − S)Y1c + ık(Y2c )

(1)

The arguments in (1) are well known. School attendance comes with a current cost in the form of direct expenses (fees, etc.) and foregone child labor and provides future benefits via the child’s (adult) earnings. This is a simplification of the decision making process, for several reasons,1 but the function in (1) provides a useful framework for understanding the school attendance decision in places like rural Guatemala. To solve for optimal S the payoffs to schooling in the second (adult) period are given by Y2c = H, which demonstrates that households need to evaluate both the child’s human capital (H) as well as other factors () that determine the utility of these skills on the labor market (Glewwe’s, 2002).2 Total H is given by H = f(Q)g(S) where individual ability() interacts with both the quantity (S) and quality (Q) of schooling. Glewwe (2002) gives Q and S functional forms [(Q)ˇ (S) ] which makes it possible to show that the marginal utility of increasing school attendance is given by: ∂U = −p − (Y1c ) + ık(Q )ˇ (S)−1 ∂S

(2)

and with (2) solve for the optimum S:

 ∗

S =

(ık(Q )ˇ ) (p + Y1c )

1/1− (3)

The motivation for this paper lies primarily in understanding how school characteristics affect the household’s school attendance cost–benefit calculus. The function in (3) highlights two main areas where actions by policymakers can affect household schooling decisions. First, the household’s sensitivity to the price p of schooling is consistent with a very large empirical literature linking household income with school attendance. This in turn supports price-reducing interventions that are fairly straightforward in implementation, meaning that the policy can be fully

1 For example, the household is reduced to a single unitary actor (Alderman, Chiappori, Haddad, Hoddintot, & Kanbur, 1995), no attempt is made to deal with inter-generational dynamics (Socias, 2004), and credit markets are assumed to be non-existent (Brown & Park, 2001). I also only consider the case for a single child and parental valuation of education is restricted to its effect on the child’s earnings with no consumption value or tastes for educated children (Schultz, 1963). 2 These extensions are based on two additional simplifying assumptions. First, the child’s wages in the first (i.e. non-adult) period are fixed and unaffected by his/her schooling. This is not unreasonable for young children in places like rural Guatemala where subsistence farming is prevalent and access to wage labor is limited for children under the age of 16 (INE, 2000). Second, human capital only comes from school attendance or school-related activities outside of school (homework). This rules out on-the-job learning or other activities that increase the child’s skills.

J.H. Marshall / Economics of Education Review 30 (2011) 1445–1455

enacted in a relatively short period of time. These interventions also tend to provide very clear signals: even poorly educated parents can calculate the effect on p of abolishing school fees, or providing scholarships. The second area where policymaker actions can impact household behavior in Eq. (3)—via school quality Q—is somewhat more complicated. First, the choice of form for total human capital (H) results in a function where the optimal schooling investment is positively affected by school quality Q. In an intuitive sense this depiction of school quality and attendance as complements is reasonable, although other plausible forms for total human capital would predict very different relationships between S and Q.3 Nevertheless, there is still the challenge for households to operationalize what school quality means in terms of what the local school is offering. For example, developing country parents generally cannot rely on published results of standardized tests to evaluate local schools.4 This leaves more direct observation of school quality in the form of the child’s actual learning, or the teaching and learning environment itself. The empirical evidence on these kinds of quality linkages with attendance outcomes in the developing world is somewhat thin, especially in comparison with the evidence base for household resource effects. Hanushek and Lavy (1994) find that dropout is less likely in urban Egyptian schools that have higher levels of value added on achievement tests. Bedi and Marshall (1999, 2002) find that students attend more days during the year when their predicted achievement is higher in Honduras. These are attention-grabbing results because they suggest some very well informed marginal consumers (Hamilton & Guin, 2005). But questions remain about the ability of the average household in the developing world, especially in rural areas, to effectively monitor the learning that takes place inside the school walls. For example, even in industrialized contexts there is substantial evidence of information asymmetries where poor parents have limited access to objective measures of school quality (for US research see Lee, Croninger, & Smith, 1996; Schneider, Teske, Marschall, & Roch, 1998). The more widely tested hypothesis in this literature is that households are basing perceptions of Q on more readily visible features of schools. This includes learning environment indicators such as teacher experience and education levels (Abraha et al., 1991; Birdsall, 1985), the availability of books and learning guides (Edwards, Fuller, & Parandekar, 1996), the school climate as measured by problems between students or with staff (Lloyd, el Tawila, Clark,

3 For example, the total school investment (in terms of attendance) will actually be lower in high quality schools if parents only want their children to attain a basic level of literacy and numeracy. But if parents are not seeking a fixed income target for their children, and instead adopt the same maximizing rules as individuals in the basic human capital model, then higher quality schools should stimulate household investments in education (Birdsall, 1985, pp. 69–70; Mincer, 1974). 4 They do receive updates on their child’s progress in the form of progress reports and final grades. This is potentially useful information about how the child is progressing, but it may be limited if the grading scheme is based more on norms and relative standing than absolute standards (Marshall, 2003).

1447

& Mensch, 2003), class sizes (Case & Deaton, 1999) and perceptions of overall quality (Fuller et al., 1995). Physical characteristics include non-leaking roofs (Glewwe & Jacoby, 1994), electricity and telephones (Bedi & Marshall, 2002), and school furniture (Ilon & Moock, 1991). School access indicators refer to and the number of days the school is open during the year (Marshall, 2009) and the distance to schools at different levels (Lavy, 1996). There is little in the way of theory to guide this empirical work. For example, are families more likely to value schools with better physical characteristics or more educated teachers? Also, in the absence of data on student learning and a convincing econometric identification scheme it is impossible to know if households are responding directly to these observable quality signals, variation in actual learning that is correlated with these same features (Bedi & Marshall, 2002), or some other variable that is correlated with all of the above. Nevertheless, taken together these findings help fill in the poverty explanation for dropout and low attainment in the developing world because poor children are less likely to have access to schools with the critical features that families appear to value. 2. Analytical framework 2.1. Guatemala context Guatemala provides an excellent context for analyzing the effects of schools on attendance outcomes like dropout. Like most countries school attendance is compulsory for young people, but physical isolation and poverty make enforcement impossible. Primary school dropout rates are among the highest in Latin America. UNESCO (2005) estimates that only about half of the children who enroll in school make it to the end of the primary cycle. Poverty is also widespread, with average education levels that are among the lowest in the region (Marshall & Calderón, 2006). In rural areas many adults never attended school, and children are engaged in work at home and in the fields at a young age (Marshall, 2011). Given the widespread poverty it is easy to attribute low levels of grade attainment to socioeconomic background influences (World Bank, 2004). But there is also a systemic component. A recent sector study highlights the structural constraints, as per pupil spending at the primary level (in 2006) was roughly 400 US dollars per year (PREAL, 2008). Most teachers are graduates of normal schools that provide three years of high school-level instruction. In rural areas few primary school teachers have completed postsecondary education, although during the time of this data collection in-service training opportunities were expanding. Textbooks have also become more widely available. In the more isolated rural areas primary schools have only become available in the last decade, in part because of official recognition of bilingual education (in 1996) and the implementation of the Programa Nacional de Autogestión para el Desarrollo Educativo (PRONADE) community school program (DiGropello, 2005). High dropout rates, widespread poverty and a changing institutional landscape provide the basic ingredients for testing the conceptual framework. But there is also

1448

J.H. Marshall / Economics of Education Review 30 (2011) 1445–1455

the cultural context to consider. Guatemala is one of the few places in Latin America where indigenous people have retained their dress and language; it is also unusual (for the region) in that girls are less likely than boys to be enrolled in school (Marshall & Calderón, 2006). A sizeable anthropological literature has analyzed the interaction between gender, ethnicity and household behavior in rural Guatemala (Bossen, 1984; Ehlers, 1990). The data used here do not allow for such precise qualitative detail. Nevertheless, it seems likely that household perceptions about schooling—and by extension the signaling power of schools—will vary by ethnicity and gender. First, parents may have doubts about the girl’s ability to continue her schooling beyond the primary level, in part because of marriage pressures which can reduce the share of her adult income that accrues to the household (k in Eq. (3)). Cultural factors related to patriarchal norms also downplay the importance of education for girls (Bossen, 1984). For indigenous families there may be related concerns about the future utility of schooling ( in Eq. (3)) based on discrimination and/or a perceived lack of access to wage-paying jobs and more developed labor markets in urban areas (Kochar, 2004). By reducing expected returns these kinds of concerns make schooling investments for female and indigenous households more price sensitive. Finally, for Mayan-speaking families the task of evaluating school quality may be complicated by language issues, which could in turn result in a modified set of school quality signals that they are most responsive to. 2.2. Data and sample The data were collected during the 2002 school year in 55 rural primary schools. I supervised bilingual research teams who made multiple visits to each school. All 55 schools were originally part of a nationally representative sample used in 2001 by the Programa Nacional de Evaluación del Rendimiento Escolar (PRONERE) to measure student achievement (De Baessa, 2002). For this study three states (or “departamentos”) were chosen to cover the three main community types in rural Guatemala: largely indigenous Alta Verapaz in the northern highlands, largely ladino (non-indigenous) Escuintla along the southern coast, and Chimaltenango in the central highlands where both indigenous and ladino populations reside. The averages for 2001 test scores and parental education in these 55 schools are very similar to the averages in the 400 plus school PRONERE sample. This does not mean the results are generalizable to all of rural Guatemala. But the schools were not drawn solely from a single region in what is a diverse country linguistically and ethnically. In each school first grade enrollment sheets from the 1999 school year were used to select up to 20 children at random, and schooling histories were then filled in using school records on a year-by-year basis. Each child therefore has a maximum of four rows of data corresponding to each individual schooling spell (i.e. school year) they were enrolled in between 1999 and 2002. Table 1 summarizes the variables used in the analysis. In all about 3000 schooling episodes (or spells) are available, meaning that the child can be matched with his/her family

background, teacher and school data for that year. The multinomial dependent variable is specific to each school year, and has three possible outcomes. In about 70 percent of the spells the student passed and moved on to the next grade. Another 22 percent of the episodes ended with the student failing the grade, but remaining enrolled. Finally, about eight percent of the spells ended with the student leaving school, either during the school year or in between school years. This eight percent average is somewhat misleading: among the children that have complete data (in Table 1) the dropout rate is roughly 15 percent, which corresponds to a smaller percentage of the total spells overall. One of the contributions of this study is to analyze these outcomes as a function of an unusually complete set of independent variables, many of which are specific to the individual schooling period (denoted by boldface in Table 1). The variable categories cover some critical features of life and schooling in rural Guatemala, beginning with family background information obtained through individual interviews with children. School access indicators include enrollment fees, days offered in the school and the physical proximity of post-primary schooling. The teacher variables cover the basics like experience and education. These are augmented with information that is important in rural Guatemala, such as ethnicity and the number of consecutive years they have been working with student cohorts. The frequency children report fighting with others at school is measured individually as well as for the school average; the latter provides a more certain school-wide indicator of climate. School characteristics include class size, school size and a value-added measure of achievement (also see Hanushek & Lavy, 1994) that is described in more detail below. 2.3. Methodology In linear form the theoretical function summarized in Section 1 (Eq. (3)) is given by: ln S = a0 + a1 ln Q + a2 ln(p + Y1c )

(4)

where a1 = ˇ/1 −  and a2 = Q − 1/1 − . In this study I estimate a modified version of (4) using event history data. The empirical strategy builds on the grade level transitions framework for analyzing stratification (Cameron & Heckman, 1998; Mare, 1981). The main difference is that the multinomial dependent variable is year-specific and includes passing, failing and dropping out, as opposed to a dichotomous indicator for completing each level of schooling (i.e. primary, secondary). The use of schooling spells divided into individual years not only allows for an additional outcome to be analyzed (grade failure), but this approach also better captures the year-by-year variation in teacher and school characteristics compared with traditional survival analysis. The grade failure episodes—while not the focus of the analysis—are important in their own right, as evidenced by the high rate of grade failure in the sample (Table 1) and the strong link between grade failure and eventual dropout established in the literature (McCombs et al., 2009). The actual estimation uses multinomial regression with the logit link together with random effects for individual

J.H. Marshall / Economics of Education Review 30 (2011) 1445–1455

1449

Table 1 Variable definitions, means and standard deviations. Variable:

Definition

Outcome variable: Passed grade

Percent of total spells that end in following outcomes: 1 = Student passed grade and remained enrolled (excluded category) 2 = Student failed grade and remained enrolled 3 = Student dropped out of school during year, or in between years Percent of total spells: Student enrolled in grade 1 (excluded category) Student enrolled in grade 2 Student enrolled in grade 3 Student enrolled in grade 4

Failed grade Dropout Grade controls: Grade 1 Grade 2 Grade 3 Grade 4 Child–family characteristics Age Gender (female) Maya speaker Father missing Mother missing Mother works Total siblings Other adults in home Average parental education SES factor Books in home Attended preschool Child-reported fighting: None (excluded category) Some Many School access: Distance to middle school Days offered Enrollment fee Teacher characteristics Teacher experience Teacher post normal education

Teacher uses Mayan language Teacher years with cohort Teacher uses homework with applications School characteristics: Average value added residual Class size Total enrollment School average fighting

Mean

Std. Dev.

70.2

–

21.8 8.1

– –

41.5 29.2 19.7 9.5

– – – –

9.7 46.4 71.3 7.2 3.7 22.4 4.9 0.10 1.8 −0.06 6.0 66.9 1.8 0.68 0.26 0.06

1.8 – – – – 2.2 0.4 2.0 1.00 11.1 – 0.7 – – –

Kilometers (from primary school) to nearest middle school Percentage of official calendar days worked during 2002 school year Official enrollment fee (Quetzales) charged by school at beginning of year ($1USD = 7.5 Q)

6.0 80.2

6.2 7.2

8.9

8.2

Teacher total experience (in years) Teacher education measured as 1 = teacher has continued their education past Normal school (does not mean they have a degree); 0 = teacher has Normal School (High School) education Teacher reports giving at least part of class in Mayan language (%) Consecutive years teacher has worked with same cohort (in addition to current year) Teacher reports assigning homework with applications (%)

10.9 0.15

5.8 –

41.3 0.44

– 0.8

60.0

–

−0.2

(4.0)

Student’s age in years Student is female (%) Student speaks Q’eqchi or Kaqchikel Mayan language in home (%) Father is not in home (%) Mother is not in home (%) Mother works in sewing, selling food (%) Total number of siblings in family Number of other adults reported living in home Average reported parental education for child’s parents Principal component factor of household possessions-services Number of books in household Child reports attending preschool or kindergarten (%) “Do other students fight with you?”: “None” “Do other students fight with you?”: “Some” “Do other students fight with you?”: “Many”

Average residual for Spanish and Mathematics based on test score analysis for 2001–2002 school year in grades 3–4 Number of students teacher has in classroom Total enrollment of students in schools School average for student-reported fighting (scale: 1: none, 2: some, 3: many)

32.7 218.4 1.8

8.0 131.6 0.3

Source: Author data, 2003. Notes: Variables in bold are time-varying and are specific to each year between 1999 and 2002. Averages correspond to pooled data.

children and schools that account for additional heterogeneity (see Cameron & Heckman, 1998; Lillard & Willis, 1994; Rabe-Hesketh, Skrondal, & Pickles, 2004). The random effects extension is made possible by pooling the schooling spells—upwards of four per child—and adding grade intercepts. The resulting model is a form of hazard (or survival) analysis based on discrete time (see Singer & Willett, 2003). The complete model takes the form:   (G)it + ˇQ (Q )n + (i , n , εit ) Sitn = ˇX (X)it + ˇG

(5)

where multinomial school attendance outcome S for child i in time period t and school n is modeled as a vector of

individual and family background characteristics (X), a vector of grade controls for the current year of enrollment (G), and a vector of school quality features (Q). The error structure accounts for individual child heterogeneity ( i , fixed across time periods), a school-specific random effect ( s ), and the remaining error for each schooling spell (εit ). One challenge when modeling school progress is accounting for the effects of previous grade failure and the non-random attrition that inevitably takes place over time. In the present case the children are all beginning in grade one in 1999, and can reach as high as grade four in 2002. Including the grade intercepts with pooled data means that

1450

J.H. Marshall / Economics of Education Review 30 (2011) 1445–1455

the parameter estimates are weighted averages across the four grades. But the grade intercepts also help control for past performance: for example, the children who made it to grade four in 2002 have not repeated at all. The controls for age and especially the individual child random effects—which are obtained by considering the child’s performance during the four year period compared against similar children—also capture previous performance and ability. An alternative approach would be to estimate a version of Eq. (5) separately by grade including controls for the number of times the child has repeated that specific grade (or all grades). This extension is indeed incorporated and commented on (briefly) below. But it is not the preferred approach because it loses the random effects extension gained by pooling, the control for previous repetition is collinear with the grade controls, and the variation in coefficients by grade levels—while interesting—is not as central to the present research task compared with studies that look at attendance over a longer time frame and in different levels of schooling. Finally, it is important to note that the Q vector does not capture all plausible school quality signals that may affect household perceptions of the benefits of keeping their children in school. But the range of information available is unusual, including the school’s value-added achievement. This is taken from a regression using a 2001–2002 cohort of grade three and four students who participated in achievement testing in two consecutive years (see PRONERE description above, and Marshall, 2009). The production function work is based on the same vector of family and student controls (X), and each of the 55 schools (minus one) is included as a dummy variable. The coefficient for each school therefore provides a basic measure of the school’s value-added to achievement for a single cohort during the 2001–2002 period. The main results are based on four variations of Eq. (5) that begin with baseline estimations that are added to with different measures for school quality Q. The main specifications are also re-estimated incorporating interaction terms for gender, ethnicity and family background. The interaction work is more exploratory, although the discussion so far has introduced some specific hypotheses: for example, the girl’s schooling is expected to be more sensitive to costs, while schools with indigenous teachers should be better at retaining indigenous students. The purpose of this work is to consider how school quality signals vary by child and family characteristics, which in turn may provide some clues about how households process information about schooling investments. 3. Results 3.1. Multinomial logit results Table 2 summarizes the results from the multinomial logit models. The t-statistics (in parenthesis) are based on robust standard errors that correct for the nesting of students by school. Dummy variables for state of residence are also included (although not presented) to control for regional differences. Each outcome is interpreted in relation to the excluded category where the child passed the

grade that year. The coefficients are odds ratios for a single unit change in the independent variable. The results for the grade controls and family background are very consistent across all four specifications. In the baseline models the random effects for children and schools are each very significant (bottom of Table 2). However, once the grade intercepts are added there is no significant residual child heterogeneity (estimation 2). This result confirms that these controls are related to individual child ability and help capture the effects of previous performance. The significant random effects at the school level also disappear when the full complement of school and teacher variables is added (estimation 4). This means that these variables are capturing meaningful variation across schools, although it does not mean they are capturing all variation in school-level factors that affect grade failure or dropout. As expected, the odds (or hazard) of grade failure and dropout decline steadily by grade. This result provides some insight into the relationship between previous grade failure and future performance, but it also needs to be understood as a feature of a research design that constricts the schooling spells to a four year period where each child began in grade one in 1999. In grade four the odds of failing (relative to passing) are only about 35 percent as likely as they are in grade one, while the odds of dropout are only 20 percent as likely (compared with grade one). These parameters do not represent the “true” effect of being in grade four versus grade one since there are likely to be other (noncohort) children in grade four who have repeated grades in previous years. Nevertheless, the results clearly show how a select group of children separate themselves from the others by moving from grade to grade without repeating. For the majority of the children—more than 70 percent of the 3000 schooling spells that are analyzed here correspond to grades one and two—the reality instead is that they fail at least one time during this sequence, and in some cases the same grade is repeated over and over. When controlling family background (estimations 1–2) and school features (estimations 3–4) girls are significantly more likely than boys to leave school before completing grade four (odds ratio about 1.5). For children who speak a Mayan language at home the point estimates are also consistently positive, but not statistically significant. Both are returned to below in the interaction analysis. Older students are less likely to fail the grade, ceteris paribus, but each year of age significantly increases the odds of dropout. As expected, grade failure and dropout are less likely in families with higher levels of capital (see parental education, SES, and books in home). Also, grade failure and dropout are each more likely in households where the mother is reported to be missing or working. Missing fathers—a more prevalent problem—are only significantly associated with grade failure. For Estimation 3 the school-specific value-added measure of student achievement is added to the child–family background characteristics. The odds of dropout are indeed lower in schools that are associated with more achievement gains, but do not approach significance. The divergence with the Hanushek and Lavy (1994) finding from urban Egypt could be attributable to lower levels of

Table 2 Random effects multinomial logit estimates of grade failure and dropout 1999–2002 (odds ratios with t-statistics in parentheses). Independent variable:

Baseline estimations: (1)

(3) With school value-added

(4) With school-teacher variables

(2) Dropout

Failure

Dropout

Failure

Dropout

Failure

Dropout

– – –

– – –

0.59*** (−3.76) 0.40*** (−5.55) 0.36*** (−4.34)

0.97 (−0.11) 0.54* (−1.91) 0.22*** (−3.15)

0.59*** (−3.79) 0.40*** (−5.57) 0.36*** (−4.33)

0.97 (−0.13) 0.54** (−1.92) 0.22*** (−3.18)

0.59*** (−3.91) 0.40*** (−5.50) 0.33*** (−4.79)

1.01 (0.05) 0.56* (−1.83) 0.24*** (−3.05)

Child-family characteristics: Student age Student female Maya speaker Father missing Mother missing Average parental education SES factor Books in home

0.78*** (−6.60) 0.92 (−0.83) 1.37 (1.41) 1.59** (2.07) 2.25*** (3.42) 0.91 (−0.59) 0.82*** (−4.23) 0.99 (−1.22)

1.42*** (4.78) 1.49** (2.06) 2.16 (1.31) 1.73 (1.29) 5.19*** (3.35) 0.64*** (−2.61) 0.76*** (−3.14) 0.92*** (−2.57)

0.91*** (−2.46) 0.93 (−0.86) 1.19 (0.91) 1.45** (2.03) 1.70*** (2.89) 0.95 (−0.36) 0.85*** (−4.19) 0.99 (−1.18)

1.60*** (6.35) 1.51** (2.31) 1.87 (1.09) 1.48 (0.91) 4.06*** (2.95) 0.66*** (−2.68) 0.79*** (−2.71) 0.92*** (−2.57)

0.91** (−2.42) 0.93 (−0.86) 1.19 (0.93) 1.44** (2.04) 1.70*** (2.89) 0.95 (−0.39) 0.85*** (−4.15) 0.99 (−1.19)

1.60*** (6.37) 1.50** (2.26) 1.95 (1.19) 1.46 (0.89) 4.11*** (2.95) 0.65*** (−2.78) 0.79*** (−2.71) 0.92*** (−2.59)

0.91** (−2.39) 0.90 (−1.16) 1.05 (0.26) 1.36* (1.68) 1.76*** (3.23) 0.99 (−0.04) 0.86*** (−4.21) 0.99 (−0.73)

1.57*** (6.47) 1.39* (1.78) 2.33 (1.36) 1.43 (0.82) 4.93*** (3.20) 0.58*** (−2.86) 0.76*** (−2.88) 0.92*** (−2.59)

Child-reported fightingb : Some Many Attended preschool Family sells part of harvest

1.66*** (4.40) 2.99*** (4.20) 0.74** (−2.19) 0.76** (−2.41)

2.11*** (3.34) 5.67*** (3.29) 0.67 (−1.47) 0.54** (−2.21)

1.52*** (4.62) 2.38*** (4.18) 0.82* (−1.72) 0.82** (−2.08)

1.85*** (2.89) 4.18*** (2.94) 0.75 (−1.14) 0.58** (−1.95)

1.52*** (4.66) 2.38*** (4.19) 0.82* (−1.72) 0.82** (−2.10)

1.87*** (2.90) 4.30*** (3.01) 0.76 (−1.14) 0.57** (−2.02)

1.54*** (4.86) 2.39*** (4.35) 0.84 (−1.55) 0.79** (−2.19)

1.79** (2.29) 3.69*** (2.55) 0.71 (−1.43) 0.54** (−2.30)

–

–

–

–

–

–

1.03*** (3.30) 1.02 (0.37) 1.01 (0.40)

1.02* (1.61) 0.94 (−0.49) 1.05*** (3.04)

School access: Distance to middle school Days offered Enrollment fee

–

–

–

–

–

–

Teacher characteristics: Teacher post normal education Teacher experience Teacher years with cohort Teacher uses Mayan language Teacher uses homework with applications Teacher frequency of exam-based evaluations

– – – – – –

– – – – – –

– – – – – –

– – – – – –

– – – – –

– – – – –

0.95 (−0.43) 0.98 (−1.45) 0.93 (−1.29) 0.80* (−1.69) 0.63*** (−2.53) 0.64*** (−2.80)

0.53** (−2.11) 1.01 (0.53) 0.92 (−0.77) 0.48*** (−3.23) 0.33*** (−2.69) 2.77*** (2.56)

School characteristics: School valued-added achievement Class size School average fighting (scale measure) Total enrollment Child heterogeneity  i (t-stat) School heterogeneity  s (t-stat) Sample size

– – – – 0.54*** (3.65) 0.17*** (2.77) 3007

– – – –

– – – – 0.01 (0.07) 0.11** (2.43) 3007

– – – –

0.99 (−0.54) – – – 0.001 (0.01) 0.11** (2.44) 3007

0.98 (−0.92) – – –

– 1.01 (1.39) 0.93 (−0.24) 0.99 (−0.57) 0.001 (0.02) 0.03 (0.83) 3007

– 1.04*** (4.55) 3.98** (1.97) 0.99 (−1.24)

1451

Source: Author data, 2003. Notes: Estimation based on Generalized Linear Latent and Mixed Models (GLLAMM), see Rabe-Hesketh et al. (2004). Time-varying variables identified with boldface. Additional variable results are not presented due to lack of significance, including: total siblings, other adults in home, mother works, and controls for state (“departamento”) of residence (see Table 1). a Excluded category for grade intercepts is grade 1. b Excluded category for Child-Reported Fighting is “None”. *** p ≤ 0.01 (two tailed tests). ** p ≤ 0.05 (two tailed tests). * p ≤ 0.10 (two tailed tests).

J.H. Marshall / Economics of Education Review 30 (2011) 1445–1455

Failure Grade intercepts: Grade 2 Grade 3 Grade 4a

1452

J.H. Marshall / Economics of Education Review 30 (2011) 1445–1455

education for parents in this sample (less than two years), who may have fewer means of reliably measuring school quality on the basis of learning. But this is an imperfect test given the fact that the value-added measure comes from grades three and four in the 2001–2002 period, while the attendance sequence for this sample spans grades one through four between 1999 and 2002. To address this limitation a separate analysis included the value-added measure in the multinomial logit model only for the 2001 and 2002 school years (combined and separately): these results are also insignificant. The results in Table 2 do provide support for the idea that the decision to retain children in school is affected by visible (to parents and families) school features (estimation 4). This is the central hypothesis in this study, and one of the advantages is that this idea can be tested with a wide range of school and teacher variables. First, children who report more frequent episodes of fighting with other children are significantly more likely to fail grades and leave school. For the individual measure this is true for those who report fighting with “Some” and “Many” other children (“None” is reference category). The individuallevel measure is included as a child background control; for example, it may be capturing behavioral or learning problems (also see Table 1). But the significant point estimate for the school average scale for the fighting variable (estimation 4) suggests a more far-reaching school climate problem. In the absence of enforceable compulsory attendance laws mistreatment is a potentially serious problem for families that have limited resources to afford schooling. During the fieldwork one father recounted how he had withdrawn his daughter from school “before the other students killed her.” At the very least the results highlight the importance of identifying at-risk children, many of whom may be affected by domestic violence issues in the home (Mayorga Salas, 1996; UNICEF, 2009). The school access measures provide some additional clues about the factors that influence household decision-making. A standard deviation increase in the school enrollment fee increases the odds of dropout by about 1.7 times. These fees—less than $1.25 on average (Table 1)5 —may seem negligible. But in rural areas where subsistence farming is prevalent the poorest parents may be hard pressed to come up with cash, or they may have doubts about whether or not the fees are really going to improve quality. The distance to the nearest middle school is a very significant predictor of grade failure, and moderately significant predictor of leaving school. When parents have concerns about access constraints for post-primary schooling then the “option value” associated with primary schooling—and the primary school diploma—necessarily decreases (Carnoy, 2001). This is an important reminder that physical accessshould not be considered solely on the basis of the proximity of the current level of schooling (Lavy, 1996).

5 In the parental interviews it was noted that informal fees were sometimes charged for things like exams and materials. But since the parental samples were not representative by school it was not possible to construct a reliable school-wide measure for these charges.

A standard deviation increase in class size (about 8 more students) predicts higher odds of dropout by about 1.7 times. Class sizes in this sample average about 32 students, which is not large by regional standards. In a separate analysis (not presented) a series of dummies were used to test whether households are particularly sensitive to very large class sizes (above 50). The results found no significant effects, so the linear measure is used instead. Interestingly, the consistently negative effects of class size in the school attendance literature—albeit with only a handful of studies (see Behrman & Knowles, 1999; Case & Deaton, 1999; Drèze & Kingdon, 1999)—stands in contrast to the fairly inconsistent effects of class size in the international test score literature (Fuller & Clarke, 1994). Nevertheless, there is some potential for reverse-causation, especially depending on the grade sequence that is being analyzed. In the present case this is less of an issue because the sequence spans grades one through four, so it does not seem likely that the class size effect is a product of schools retaining more children in early grades, with resulting large classes and high rates of dropout later on. Furthermore, the inclusion of a school-specific random effect helps control for un-measured aspects of the school’s propensity to retain (or pass) students. There is evidence that remaining in school is related to the work and qualifications of the teacher. For example, children are about half as likely to drop out, ceteris paribus, when the teacher has some amount of post-high school education. At the time of this data collection only about 15 percent of the teachers had continued their education beyond this level, so some caution is required in attributing this result to teacher effectiveness. The most significant teacher effect is related to ethnicity. The odds of grade failure and dropout are significantly reduced when the student is studying with a teacher who uses a Mayan language in class. This is an important result given the number of children in the sample who speak an indigenous language (71 percent). It suggests a cultural matching effect similar to what Dee (2001) finds between African American students and teachers in the United States, although this needs to be explored further in the interaction analysis. The insignificant predictors of grade failure and dropout merit some discussion as well. Measures of physical conditions in the school (e.g. lighting), the availability of services and learning materials, and distance to the primary school are also available in the dataset. They were included in preliminary analyses, but due to a lack of significance and some collinearity with the total enrollment measure they were dropped. The same is true for a series of indicators for teaching methodology covered in the teacher questionnaire. These were also insignificant and moderately correlated with the teacher variables that remain. Perhaps the most important exclusion is a measure of parental involvement in the school. In a preliminary phase of the data collection we piloted a battery of questions to parents about their views on school quality and specific actions they had taken vis-à-vis school personnel. The results showed that most were reluctant to enter into discussions with school personnel about specific topics related to the school, so this aspect of the study was not continued in much detail. What is available is an indicator

J.H. Marshall / Economics of Education Review 30 (2011) 1445–1455

from the teacher questionnaire about parental involvement in education in general, as well as an average for the frequency of school–parent meetings based on a nonrandom sample of parents in each school (see note 5). Neither was a significant predictor of dropout, and because of some correlation with parental education they were also not included in the final model.

3.2. Interaction analysis The interaction analysis helps extend the results in Table 2 by identifying specific features of schools that certain households appear to be particularly responsive too. Table 3 summarizes the main results, focusing only on the dropout outcome (and excluding grade failure). For each interaction grouping variable two estimations are summarized. The first tests for interaction with the school value-added measure (columns 1, 3 and 5). The second includes a larger group of interactions based on the more complete specification with multiple teacher and school variables (columns 2, 4 and 6; also see estimation 4 in Table 2). As noted before the quantity of potential interaction means there is some element of exploration in this section. To conserve space only the interaction terms are presented without the main effects (although these are discussed). For girls there is limited evidence suggesting greater sensitivity to the costs of schooling. Older girls are not significantly more likely to leave school than older boys, and the girl’s schooling is actually significantly less affected by enrollment fees. However, for girls the dropout outcome is significantly more likely when the distance to the nearest middle school is greater (Lavy, 1996). This apparent sensitivity to the availability of middle schooling may be related to concerns about commuting or leaving the

1453

village for girls. It is also possible that patriarchal norms are stronger in households located in more isolated communities, and as a result there may be less demand for schooling for girls. The other main findings are for ethnicity. First, indigenous children are significantly less likely to drop out of school when the teacher reports using a Mayan language in class. This result is somewhat sensitive to model specification. But it is an important result in this context, and provides support for bilingual education programs and efforts to recruit teachers from indigenous communities to work with similar children. There is also some significant interaction between ethnicity and the school’s value-added indicator of achievement. When controlling for interaction the main effect for the school’s value-added achievement on dropout is significant and negative. This is an important divergence from the main results presented in Table 2 (estimation 3), and more consistent with Hanushek and Lavy (1994) findings from urban Egypt. The positive interaction in Table 3 means that both ladino and indigenous children are less likely to drop out of schools with higher achievement, but ladino children are even less likely to do so. This suggests that the ladino families are more responsive to school quality related to learning gains, but given the limitations of the value-added measure (see previous section) this result needs to be treated with a lot of caution.

3.3. Sensitivity analysis Event history data allow for a range of statistical specifications. My favored approach takes advantage of recent developments in econometric software by accounting for both individual student and school heterogeneity (RabeHesketh et al., 2004). Nevertheless, in order to assess the stability of the main findings I carried out a series

Table 3 Interaction results for the student dropout outcome with gender, ethnicity and books in home (odds ratios with t-statistics in parentheses).

Child–family characteristics: Student female Student AGE School–teacher characteristics: Average value added residual Teacher uses Mayan language Teacher post normal education Teacher years with cohort School average fighting Total enrollment Distance to middleschool Class size Enrollment fee Child heterogeneity  i (t-stat) School heterogeneity  s (t-stat) Sample size

Student female with:

Maya speaker with:

Books in home with:

(1)

(2)

(3)

(4)

(5)

(6)

– –

– 1.14 (0.94)

– –

2.76 (1.46) 0.70** (−2.21)

– –

1.03 (0.35) 1.01 (0.43)

1.02 (0.38) – – – – – – – – 0.01 (0.10) 0.11** (2.51) 3007

– 1.15 (0.37) 0.61** (−2.36) 1.02 (0.05) 1.92 (0.62) 1.01*** (3.58) 1.05* (1.85) 0.97 (−1.20) 0.96* (−1.65) 0.01 (0.13) 0.02 (0.98) 3007

1.20* (1.78) – – – – – – – – 0.01 (0.11) 0.08** (2.28) 3007

– 0.11** (−2.23) 1.15 (0.29) 1.09 (0.28) 8.55 (1.45) 1.01 (1.00) 1.15 (0.57) 1.04 (0.85) 1.06 (1.31) 0.01 (0.15) 0.01 (0.08) 3007

0.99 (−0.61) – – – – – – – – 0.01 (0.09) 0.10** (2.48) 3007

– 0.93 (1.39) 1.01 (1.14) 0.89*** (−2.91) 1.04 (0.49) 1.01* (1.85) 1.01** (2.02) 1.01 (1.28) 0.99 (−0.22) 0.01 (0.11) 0.10** (2.39) 3007

Source: Author data, 2003. Notes: Full results available on request. Each column refers to a separate estimation of a multinomial logit model, however these results only refer to the Dropout outcome. All estimations based on Generalized Linear Latent and Mixed Models (GLLAMM), see Rabe-Hesketh et al. (2004). Time-varying variables identified with boldface. See text and details at bottom of Table 2 for more information. *** p ≤ 0.01 (two tailed tests). ** p ≤ 0.05 (two tailed tests). * p ≤ 0.10 (two tailed tests).

1454

J.H. Marshall / Economics of Education Review 30 (2011) 1445–1455

of separate analyses. First, I estimated the multinomial logit separately by grade, and then with pooled data using interaction terms between the grade controls and independent variables. The results showed that most of the main findings from Table 2 were significant in at least one grade, but the variation across grades is not particularly systematic (like a decreasing effect of fighting across grades, or an increasing significance of teacher education). Furthermore, less than 30 percent of the schooling spells correspond to grades three and four, so by pooling and including intercepts I do avoid some problems associated with smaller sample sizes in these higher grades. Second, a Hausman test was used to assess the extent to which the Independence of Irrelevant Alternatives (IIA) assumption is valid with this particular multinomial outcome. The results showed no significant variation in the estimates for dropout with and without the grade failure outcome. As a further test the multinomial probit and binary logistic options were also employed, the latter based on a dependent variable of dropout versus all other spells. Finally, the main specifications were re-estimated using different sample weighting schemes to account for non-response in the child survey component of the data collection, and non-proportional sampling within schools. In general, the results for teacher ethnicity, school fees, class size, and the individual measure of student fighting are consistent (and significant) across these different models. The results for teacher education, the school average for fighting and the distance to the nearest middle school are significant in most cases. The results for these extensions are available upon request. 4. Discussion and conclusion In this study I analyzed school dropout and grade failure in rural Guatemala using detailed event history data over a four-year period. The significant effects for SES, family structure, and the child’s age are consistent with the “poverty explanation” most commonly attributed to young people doing poorly in school (and dropping out). However, it is not the case that family background characteristics—including the child’s previous performance—were the only predictors of grade failure and dropout. The statistical analysis also identified significant school effects related to teacher ethnicity, access, and problems between students. These results help extend a growing literature linking features of schools with attendance outcomes in developing countries by bringing together a number of variables that have been rarely analyzed previously. The four year event history sequence beginning in grade one is also an important extension on most previous studies, especially compared with those that regress grade attainment or current enrollment on a fixed group of predictors. Finally, the interaction analysis shed some additional light on how households respond to specific school features; among these results the significantly lower dropout probabilities for indigenous children studying with Mayan-language speaking teachers stands out. Overall, the policy implications are mixed. On the one hand the strong impact of family background on

attendance in the first four years of the primary cycle reinforces the need for policies that improve access to schooling for poor people in developing countries, including the use of targeted subsidies and scholarships. For older children especially these price-reducing interventions are likely to have the biggest impact, especially as the growing supply of primary school graduates in places like rural Guatemala increases the demand for middle schooling (and beyond). Nevertheless, it is a mistake to assume that government policy can only affect attendance via the price of schooling, as largely implied by the vast number of grade attainment studies in developing countries. Policymakers can also make schools more responsive to their user’s needs, beginning at an early age. This last point has not received enough attention in policy circles due to data limitations and a tendency to conceptualize school quality—and school effects—on the basis of test score analyses. More research is required to understand the underlying mechanisms that link features of schools with the family’s decision to discontinue schooling for children. This means digging deeper to understand how developing country households articulate their interest in schools and schooling, how this articulation is itself affected by the larger opportunity structure as well as negotiations and bargaining between family members, and how information is collected to make decisions.

Acknowledgements Very useful comments were provided by Miguel Socias, Martin Carnoy, Susanna Loeb, Eric Hanushek, David Suarez, Jo Boaler, Nancy Tuma, Arjun Bedi and two anonymous referees. All remaining errors are of course my own. In Guatemala, Yetilu de Baessa and her research team at the PRONERE evaluation project provided crucial logistical assistance with the data collection. Partial funding was provided by the Spencer Foundation.

References Abraha, S., Assefa, B., Tesfaye, D., Fuller, B., Holloway, S., & King, E. (1991). What factors shape girls school performance? Evidence from Ethiopia. International Journal of Educational Development, 11, 107– 118. Alderman, H., Chiappori, P., Haddad, L., Hoddintot, J., & Kanbur, R. (1995). Unitary versus collective models of the household: Time to shift the burden of proof? The World Bank Research Observer, 10, 1–19. Bedi, A. S., & Marshall, J. H. (2002). Primary school attendance in Honduras. Journal of Development Economics, 69, 129–153. Bedi, A. S., & Marshall, J. H. (1999). School attendance and student achievement: Evidence from rural Honduras. Economic Development and Cultural Change, 47, 657–682. Becker, G. S. 1967. Human capital and the personal distribution of income: An analytical approach. Woytinsky Lecture no. 1, Ann Arbor: University of Michigan. Behrman, J. R., & Knowles, J. C. (1999). Household income and child schooling in Vietnam. World Bank Economic Review, 13, 211–256. Birdsall, N. (1985). Public inputs and child schooling in Brazil. Journal of Development Economics, 18, 67–86. Bossen, L. (1984). The redivision of labor: Women and economic choice in four Guatemalan communities. Albany: State University of New York Press. Brown, P., & Park, A. (2001). Poverty and education in rural China. Economics of Education Review, 21, 523–541.

J.H. Marshall / Economics of Education Review 30 (2011) 1445–1455 Buchmann, C., & Hannum, E. (2001). Education and Stratification in developing countries: A review of theories and research. Annual Review of Sociology, 27, 77–102. Cameron, S. V., & Heckman, J. J. (1998). Life cycle schooling and dynamic selection bias: Models and evidence for five cohorts of American males. The Journal of Political Economy, 106, 262– 333. Carnoy, M. (2001). Are educational reforms in Latin America working? A new look at understanding whether education is getting better. Washington DC: Inter-American Development Bank. Case, A., & Deaton, A. (1999). School inputs and educational outcomes in South Africa. The Quarterly Journal of Economics, 114, 1047–1084. Chernichovsky, D. (1985). Socioeconomic and demographic aspects of school enrollment and attendance in rural Botswana. Economic Development and Cultural Change, 33, 319–332. De Baessa, Y. (2002). Summary of 2001 PRONERE test application. Guatemala: Universidad del Valle. Dee, T. S. (2001). Teachers, race and student achievement in a randomized experiment. National Bureau of Economic Research working paper #8432. DiGropello, E. (2005). A comparative analysis of school-based management in Central America. World Bank Working Paper No. 72. Drèze, J. & Kingdon, G. G. (1999). School participation in rural India. Development Economics Discussion Paper Series No. 18. London School of Economics. Edwards, J. H. Y., Fuller, B., & Parandekar, S. (1996). Primary education efficiency in Honduras: What remains to be done? Tegucigalpa, Honduras: USAID. Ehlers, T. B. (1990). Silent looms: Women and production in a Guatemalan town. Boulder, CO: Westview Press. Filmer, D., & Shady, N. (2008). Getting girls into school: Evidence from a scholarship program in Cambodia. Economic Development and Cultural Change, 56, 581–617. Fuller, B., & Clarke, P. (1994). Raising school effects while ignoring culture? Local conditions and the influence of classroom tools, rules and pedagogy. Review of Educational Research, 64, 119–157. Fuller, B., Singer, J., & Keiley, M. (1995). Why do daughters leave school in Southern Africa? Family economy and mother’s commitments. Social Forces, 74, 657–680. Glewwe, P. (2002). Schools and skills in developing countries: Education policies and socioeconomic outcomes. Journal of Economic Literature, 90, 436–482. Glewwe, P., & Jacoby, H. (1994). Student achievement and schooling choice in low-income countries: Evidence from Ghana. Journal of Human Resources, 29(3), 843–864. Hamilton, L. S., & Guin, K. (2005). Understanding how families choose schools. In J. R. Betts, & T. Loveless (Eds.), Getting choice right: Ensuring equity and efficiency in education policy (pp. 40–60). Washington, DC: Brookings. Hanushek, E. A. & Lavy, V. (1994). School quality, achievement bias, and dropout behavior in Egypt. Working Paper no. 107. Washington DC: World Bank Living Standards Measurement Study. Ilon, L., & Moock, P. (1991). School attributes, household characteristics, and demand for schooling: A case study of rural Peru. International Review of Education, 37, 429–451. InstitutoNacional de Estadística (INE). (2000). 1998 Household survey data. Guatemala City, Guatemala. Kochar, A. (2004). Urban influences on rural schooling in India. Journal of Development Economics, 74, 113–136. Lavy, V. (1996). School supply constraints and children’s educational outcomes in rural Ghana. Journal of Development Economics, 51, 291–314.

1455

Lee, V. E., Croninger, R. G., & Smith, J. B. (1996). Equity and choice in Detroit. In B. Fuller, & R. F. Elmore (Eds.), Who chooses? Who loses? New York: Teachers College Press. Lillard, L. A., & Willis, R. J. (1994). Intergenerational educational mobility: Effects of family and state in Malaysia. Journal of Human Resources, 29(4), 1126–1166. Lloyd, C. B., Mete, C., & Grant, M. J. (2009). The implications of changing educational and family circumstances for children’s grade progression in rural Pakistan: 1997–2004. Economics of Education Review, 28, 152–160. Lloyd, C. B., el Tawila, S., Clark, W. H., & Mensch, B. S. (2003). The impact of educational quality on school exit in Egypt. Comparative Education Review, 47, 444–467. Mare, R. D. (1981). Change and stability in educational stratification. American Sociological Review, 46, 72–87. Marshall, J. H. (2003). Grade repetition in Honduran primary schools. International Journal of Educational Development, 23, 591–605. Marshall, J. H. (2009). School quality and learning gains in rural Guatemala. Economics of Education Review, 28(2), 207–216. Marshall, J. H. (2011). Child labor and schooling outcomes in rural Guatemala. Unpublished manuscript. Marshall, J. H. & Calderón, V. (2006). Social exclusion in education in Latin America and the Caribbean. Sustainable Development Department, Technical Paper Series EDU-122. Washington, DC: Inter-American Development Bank. Mayorga Salas, L. (1997). Violence and aggression in the schools of Colombia, El Salvador, Guatemala, Nicaragua and Peru. In T. Ohsako (Ed.), Violence at school: Global issues and interventions. Paris: UNESCO. McCombs, J. S., Kirby, S. N., & Mariano, L. T. (Eds.). (2009). Ending Social Promotion Without Leaving Children Behind: The Case of New York City. Santa Monica, CA: The RAND Corporation. Mincer, J. (1974). Schooling experience and earnings. New York: National Bureau of Economic Research. PREAL (2008). Education: A challenge of national urgency. Partnership for Educational Revitalization in the Americas and Center for National Economic Research (CIEN). Rabe-Hesketh, S., Skrondal, A., & Pickles, A. (2004).GLLAMM Manual. University of California at Berkeley, Division of Biostatistics Working Paper 160. Schneider, M., Teske, P., Marschall, M., & Roch, C. (1998). Shopping for schools: In the land of the blind, the one-eyed parent may be enough. American Journal of Political Science, 42, 769–793. Shultz, T. P. (2004). School subsidies for the poor: Evaluating the Mexican Progresa poverty program. Journal of Development Economics, 74(1), 199–250. Schultz, T. W. (1963). Human capital. Chicago: University of Chicago Press. Singer, J. D., & Willett, J. B. (2003). Applied Longitudinal Data Analysis: Methods for Studying Change and Event Occurrence. New York: Oxford University Press. Socias, M. (2004). Intergenerational education mobility: Measurement, theory and policy evaluation. School of Education, Stanford University. The World Bank. (2009). Abolishing School Fees. In Africa: Lessons from Ethiopia, Ghana, Kenya, Malawi and Mozambique. Washington, DC: The World Bank. UNESCO (2005). Education indicators. www.unesco.org. UNICEF (2009). Characteristics of a rights-based, child-friendly school. http://www.unicef.org/lifeskills/index 7260.html. Accessed 26.11.09. UNICEF. (2010). An analysis of out-of-school children: Ghana demographic and health surveys (GDHS) 2003–2008. Accra: UNICEF. World Bank. (2004). Poverty in Guatemala. Washington, DC: The World Bank.