What Determines Learning among Kinh and ... - Wiley Online Library

bs_bs_banner

Asia & the Pacific Policy Studies, vol. 2, no. 3, pp. 494–516 doi: 10.1002/app5.102

Original Article What Determines Learning among Kinh and Ethnic Minority Students in Vietnam? An Analysis of the Round 2 Young Lives Data Paul Glewwe, Qihui Chen and Bhagyashree Katare*

Abstract An analysis of the Young Lives data collected in 2006, involving a younger cohort (aged 5) and an older cohort (aged 12), yields three important findings regarding the Kinh–ethnic minority gaps in mathematics and reading skills in Vietnam. First, large disparities exist even before children start primary school. Second, language may play an important role: Vietnamese-speaking ethnic minority children scored much higher than their non-Vietnamesespeaking counterparts, even though tests could be taken in any language the child chooses. Third, Blinder–Oaxaca decompositions indicate that higher parental education among Kinh children explains about one third of the gap for both cohorts. For the older cohort, Kinh households’ higher income explains 0.2–0.3 standard deviations (SDs) of the gap (1.3–1.5 SDs). More time in school, less time spent working, and better nutritional status each explain about 0.1 SDs of the mathematics * Glewwe: Department of Applied Economics, University of Minnesota, St Paul, MN 55108, USA; Chen: Department of Applied Economics, College of Economics and Management, China Agricultural University, Beijing 100 083, China; Katare: Department of Agricultural Economics, Purdue University, West Lafayette, IN 47907, USA. Corresponding author: Glewwe, email ⬍[email protected]⬎. We thank Laura Camfield, Stefan Dercon, Helen Pinnock, Caitlin Porter and Chi Truong for helpful comments on earlier drafts of this article.

score gap; Kinh children’s more years of schooling explains about 0.3 SDs of the Peabody Picture Vocabulary Test score gap. Key words: cognitive skills, ethnic minority, Blinder-Oaxaca decomposition, Vietnam, education 1. Introduction Vietnam is one of the poorest countries in south-east Asia, but since the late 1980s it has enjoyed a high rate of economic growth, which is almost certainly due to its Doi Moi policies, which in effect replaced Vietnam’s planned economy with a market economy. Like many formerly socialist countries, Vietnam has long had relatively high levels of education compared to other low income countries. Its primary school gross enrolment rate has been close to 100 per cent since the early 1990s, and its secondary school gross enrolment rate increased from about 32 per cent in the early 1990s to 57 per cent in the late 1990s and about 85 per cent in 2006 (Glewwe 2004; General Statistics Office 2007). While education is often measured in terms of years of schooling completed, the benefits of education are ultimately determined by the skills individuals acquire in school. Thus, it is important to understand what individual, family and school characteristics lead to the acquisition of academic skills. Unfortunately, thorough studies of the determinants of aca-

© 2015 The Authors. Asia and the Pacific Policy Studies published by Crawford School of Public Policy at The Australian National University and Wiley Publishing Asia Pty Ltd. This is an open access article under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs License, which permits use and distribution in any medium, provided the original work is properly cited, the use is non-commercial and no modifications or adaptations are made.

Glewwe et al.: Young Lives Data Analysis demic skills in developing countries are somewhat rare owing to lack of data on both academic skills and their determinants. This article examines the acquisition of mathematics and reading skills in Vietnam, using the Round 2 Young Lives household survey data collected in 2006. While Vietnam has long had unusually good performance in education as measured by years of education completed, its relatively short school day may limit the skills students acquire in school. Moreover, ethnic minority students have much lower outcomes, in terms of both years of schooling and test scores, than ethnic Vietnamese (Kinh). Thus, this article focuses on explaining the Kinh–ethnic minority learning gap in Vietnam. The rest of this article proceeds as follows. Section 2 briefly describes Vietnam’s education system, and reviews several recent studies on the determinants of school enrolment and student learning in Vietnam. Section 3 describes the data. Section 4 presents the methodological framework underlying the analysis. The results for the younger cohort (aged 5 in 2006) and older cohort (aged 12 in 2006) are given in Sections 5 and 6, respectively. The final section draws conclusions and presents some suggestions for future research. 2. Literature Review Vietnam’s education system has three levels: primary, secondary and tertiary. Primary education (Grades1–5) is for children aged 6–10. Secondary education consists of lower secondary education (Grades 6–9) for children aged 11–14, and upper secondary education (Grades 10–12) for children aged 15–17. Various types of tertiary education, ranging from university degree programs to a wide variety of technical training, are available for the population aged 18 years and older. The vast majority of Vietnam’s schools are public (governmentoperated) schools. Relative to its low income level, Vietnam has achieved remarkable success in terms of basic education outcomes. While its gross domestic product per capita in 2004 (US$502) was less than one half the average of East

495

Asian and Pacific countries and a quarter of the average of middle-income countries, it has similar literacy rates to those two groups of countries (Dang 2007). Its primary school completion rate, 92 per cent, is even slightly higher than the average for the abovementioned groups of countries; gross enrollment rates in Vietnam were 90 per cent, 76 per cent and 16 per cent at the primary, secondary and tertiary levels, respectively, in 2006 (World Bank 2008). Several recent studies have examined the educational performance of ethnic minority children in Vietnam.1 Dang (2003), while not focusing on test scores, estimates the determinants of years of schooling in rural Vietnam. Using data from the 1997–1998 Vietnam Living Standards Survey, he finds significant Kinh–ethnic minority gaps in years of schooling. His regressions include separate ethnic minority dummy variables for the Northern Uplands, Central Highlands and Mekong Delta regions (where most ethnic minorities in Vietnam reside), as well as a general ethnic minority dummy for ‘all other regions’ in Vietnam. He finds that while ethnic minority children in the Northern Uplands tend to have more years of schooling than Kinh children (significant at the 10 per cent level), ethnic minority children in the Central Highlands and the Mekong Delta, as well as in ‘all other regions’ of Vietnam, had one to two fewer years of schooling than Kinh children after controlling for the aforementioned variables. These negative impacts are especially strong in the Central Highlands and regions other than the three regions just mentioned. Note that since many ethnic minority children live in remote areas, the distance-to-the-nearest-town effect will have an additional negative impact on their years of schooling. Similarly, the qualifications of teachers in their schools are likely to be lower, further reducing their years of schooling. 1. There are also several more recent papers on student performance in Vietnam that do not emphasise Kinh– ethnic minority differences. To avoid making this article overly long, and to focus on issues pertaining to ethnic minorities, we do not review those papers here.

© 2015 The Authors. Asia and the Pacific Policy Studies published by Crawford School of Public Policy at The Australian National University and Wiley Publishing Asia Pty Ltd

496

Asia & the Pacific Policy Studies

More recent papers have examined gaps in test scores between Kinh and ethnic minority children in Vietnam. Himaz (2009) examined mathematics and reading (vocabulary) skills using the Round 2 Young Lives data for the older cohort children (aged 12 when tested in 2006). She regressed scores on the mathematics test and the Peabody Picture Vocabulary Test (PPVT) on a set of child and household characteristics and a dummy variable for the mother belonging to an ethnic minority.2 She found that mothers’ and fathers’ education had strong positive effects for both tests, as did child nutrition (measured by the heightfor-age z-score). Household size and having an ethnic minority mother had strong negative effects on both tests. The negative effect of household size probably reflects the fact that the household wealth variable is not in per capita terms; as discussed below, once household consumption is expressed in per capita terms there is no additional explanatory power of household size. In general, virtually all of Himaz’s results are consistent with many other studies of the determinants of test scores in developing countries (see Glewwe & Kremer 2006 for a review). They confirm the importance of parental education and child nutrition, as well as negative outcomes for ethnic minority children even after controlling for these two important determinants. A completely different data source on test scores was used by the World Bank, which conducted two reading and mathematics studies with Vietnam’s Ministry of Education and Training in 2001 (World Bank 2004) and 2007 (World Bank 2011). Both studies focus on children in the last year of primary school (Grade 5). The general findings are as one would expect: children with better-educated parents and from wealthier households have higher scores on the two tests. Also, ethnic

2. The child and mother ethnic minority status variables are extremely highly correlated; of the 989 observations that have both variables, there are only 26 cases where the child was from an ethnic minority while the mother was not, or the mother was from an ethnic minority but the child was not.

September 2015

minority children are found to have lower scores, but neither study provides a detailed analysis of the reasons for this. Two very recent papers have examined the gap in per capita household expenditures between Kinh and ethnic minority households in Vietnam, which are of interest since those gaps can lead to gaps in student learning. Baulch et al. (2010) examined ethnic inequality in household expenditure per capita, using data from the Vietnam Living Standards Surveys of 1993 and 1998 and the Vietnam Household Living Standard Surveys (VHLSS) of 2002, 2004 and 2006. The study adopts the Blinder–Oaxaca decomposition method to identify the sources of the ethnic gap in per capita expenditure, examining both differences in characteristics and differences in returns to those characteristics. Their (mean and quantile) decomposition results show that at least a half of the ethnic gap in per capita expenditure can be attributed to the lower returns to characteristics received by ethnic minority households. Nguyen et al. (2014) also examine ethnic inequality in household expenditures per capita, using data from the 2006 VHLSS. More specifically, they investigate: (i) how language barriers (inability to speak Vietnamese) may hinder ethnic minority households from taking advantage of their acquired skills and attributes; (ii) whether commune infrastructure works for or against ethnic minorities; and (iii) the extent to which differences in endowments, as opposed to returns to endowments, explain the ethnic gap in per capita household expenditures. Applying the Blinder–Oaxaca decomposition technique (using instrumental variables to consistently estimate the impacts of household and commune level variables), they find the following: first, removing language barriers significantly reduces ethnic inequality through enhancing ethnic minorities’ gains from education; second, returns to education favour the Kinh–Chinese majority in mixed communes, which suggests that the special needs of minority students are not adequately addressed, or that there exists unequal treatment in the labour market. Third, except for


Glewwe et al.: Young Lives Data Analysis hard-surfaced roads, there is little difference in the benefits generated from enhanced infrastructure at the commune level across ethnic groups. Finally, in contrast to other studies (including Baulch et al. 2010), as much as 49–66 per cent of the ethnic gap in per capita household expenditures is due to differences in endowments rather than differences in the returns to endowments. Finally, three recent qualitative studies have examined education and gaps between Kinh and ethnic minority children in Vietnam. Truong (2009) conducted a qualitative study based partially on the Young Lives data. Without attempting to estimate the determinants of test scores, Truong documents large Kinh–ethnic minority gaps in school enrolment and dropout rates and in test scores. This article focuses on Bao Ly commune in Lao Cai province, which has a high H’mong population, and on Ea Mua commune in Phu Yen province, which has a sizeable Cham population. The H’mong children in Bao Ly commune are much more likely to drop out of school (by age 12), and score far lower on reading, writing and mathematics tests, compared to Kinh children in that commune. Second, a study by the World Bank (2009) argues that lack of bilingual instruction and lack of ethnic minority instructors in schools with ethnic minorities are barriers to learning among ethnic minority students. Third, a report by the Vietnam Academy of Social Sciences (2009) points out that ‘Ethnic minority people who are not fluent in the Vietnamese language are 1.9 times more likely to be poor than ethnic minority people who are fluent in Vietnamese, and 7.8 times more likely to be poor than the Kinh/Chinese people’. This article is the first to examine the determinants of test scores and the Kinh–ethnic minority gaps in test scores in detail using the Young Lives data from Vietnam. More importantly, it is the first article to apply the Blinder– Oaxaca decomposition to understand test score gaps in Vietnam using any data source. These gaps are arguably one of the most important sources of socio-economic inequality in Vietnam today.

497

3. Data This article analyses the Round 2 Young Lives data from Vietnam, which were collected in 2006. The Round 1 survey in 2002 collected data on 2,000 children aged 6–17 months (younger cohort) and 1,000 children aged 7.5– 8.5 years (older cohort). Strictly speaking, this sample is not representative of Vietnam as a whole. Instead, five of Vietnam’s 63 provinces were selected to be ‘representative’ of most of Vietnam’s regions. However, comparisons with the nationally representative Vietnam Household Living Standards Survey suggest that the Young Lives sample is broadly representative of Vietnam as a whole (see Table A1 in Appendix 1). Section 3.1 explains the sampling procedure, and following it, section 3.2 describes the data, focusing on the tests administered in round 2. For more details on the sampling procedure, see Tran et al. (2003). 3.1 Sample Vietnam can be divided into eight socioeconomic regions: North-West, North-East, Red River Delta, North Central Coast, South Central Coast, South-East, Central Highlands, and Mekong River Delta. To ensure that the Young Lives sample included a major urban centre, a new ‘region’ (‘Cities’) was created consisting of all major urban provinces (Hanoi, Ho Chi Minh City, Da Nang, Hai Phong and Ba Ria-Vung Tau). Of these nine ‘regions’, five (North-East, Red River Delta, Cities, South Central Coast, and Mekong River Delta) were chosen as ‘representative’ of Vietnam in the sense that they: (i) include regions in the northern, central and southern areas of Vietnam; (ii) include urban, rural and mountainous areas; (iii) are relatively poor; and (iv) reflect some unique factors of Vietnam, such as areas prone to natural disasters and areas heavily affected by past wars. From each of these five regions, a ‘typical’ province was chosen after consultation with both government and international experts. The five selected provinces (see Figure 1) were: Lao Cai (North-East region), Hung Yen (Red River Delta), Da Nang (Cities), Phu Yen


498


Figure 1 The Five Provinces Featured in the Young Lives Survey

(South Central Coast) and Ben Tre (Mekong River Delta). Within each province, at least four communes (‘sentinel sites’) were chosen, giving greater weight to poor communes.3 More specifically, all communes in the province were ranked by poverty level: poor, average, better off and rich. Two communes were selected from the poor group, one from the average group, and one from the above average group (combining ‘better off’ and ‘rich’). The selection of communes from each group was not random; the criteria considered include: (i) whether the commune represents common provincial/regional features; (ii) 3. Vietnam is divided into about 65 provinces. Each province is divided into districts, and each district is further divided into communes. These are all political units. Overall, Vietnam has over 10,000 communes.

September 2015

whether there was commitment from the local government for the research; (iii) feasibility in terms of research logistics; and (iv) population size. If a selected commune had a population of less than 6,000 persons, a ‘similar’ commune at the same poverty level was selected to assure that 100 Younger Cohort children could be found in that ‘sentinel site’. Following this procedure, 31 communes were finally selected: seven in Lao Cai, six in Hung Yen, four in Da Nang, six in Phu Yen, and eight in Ben Tre. In the second round of data collection in 2006, the number of communes increased to 34 because one commune in Da Nang province was split into two between 2002 and 2006, and another in Da Nang was split into three. In each selected commune, lists were compiled of all children born between 1 January 1993 and 31 December 1994 and between 1 January 2000 and 31 December 2001. These 2-year ranges were used due to uncertainty about the exact starting date of data collection. Random sampling was then applied in each commune to select 100 children of 6.0–17.9 months and 50 children aged 7.5–8.5 years at the time of the fieldwork (July–November 2002). The refusal rate was only 1.2 per cent (=36/3000) and replacement sampling was used in the case of refusals. For reasons including child deaths, refusals and households’ moving away from Vietnam, 30 of the 2,000 younger cohort children and 10 of the 1,000 older cohort children in round 1 did not participate in round 2, leaving 1,970 younger cohort children and 990 older cohort children in the round 2 sample. The main reasons for the sample attrition in the younger cohort are that the child died (11 cases) and that the household could not be found (13), with the remaining cases being refusal or moving away from Vietnam. At least 82 households, and perhaps as many as 95, of the 1,970 households that were re-interviewed moved within Vietnam from 2002 to 2006, and all were found and re-interviewed. (There are 82 households that clearly moved, and another 13 for which missing data make it unclear whether they moved.) The main reasons for sample attrition in the older cohort are that the


Glewwe et al.: Young Lives Data Analysis child died (two cases), the household refused to participate (two cases), or the household could not be found (three cases). Finally, of the 990 older cohort households re-interviewed in 2006, 14 moved within Vietnam and were found and re-interviewed, and another household was interviewed, but missing data make it impossible to determine whether that household moved. 3.2 Tests The cognitive skills tests administered to the children varied by cohort and the round of the survey. In round 1, the younger cohort was too young to take any tests. In round 2, they took two tests: the Cognitive Development Assessment test of basic quantitative skills (CDA-Q) and the PPVT test. For more information on these tests, see Cueto et al. (2009). The former test is the quantitative sub-test of the CDA test developed by the International Evaluation Association (IEA). For each of its 15 items, a child is shown a picture and asked a question (e.g. ‘Look at the bowls of eggs. Point to the bowl that has the most eggs.’) and then chose the best answer from three or four choices. Each correct answer scores 1 point; an incorrect or blank answer scores 0. The total CDA-Q score is the number of correct answers.4 The PPVT mainly tests vocabulary acquisition of children aged 2.5 years and older. It consists of 17 sets ranked in order of difficulty. Each set includes 12 items, each item containing a picture. The child is asked to say the name of the object or activity in the picture. An initial, ‘basal’ set is given based on a child’s age and is selected to be very easy for the child. If the child correctly answers 11–12 items in the basal set, the next more difficult set is administered. The child is given increasingly difficult sets until reaching a set that is too difficult (i.e. unable to correctly answer at least five items in the set). The PPVT score is the number of correctly answered items out of all sets taken by the child, plus the number of 4. One of these 15 questions was found to have poor statistical properties based on Item Response Theory, and was dropped from the calculation of the total score.

499

items in all sets that were easier than the basal set, under the assumption that the child would have correctly answered all items in these easier sets.5 In round 2, the CDA-Q scores are missing for 64 of the 1,970 younger cohort children for unknown reasons, leaving 1,906 children with CDA-Q scores. The PPVT scores are missing for 223 children, mainly because the test was not administered properly to most of these children, leaving 1,747 children with PPVT scores. The reasons for the missing test scores are as follows. First, the basal set of questions was too difficult for 97 of the younger cohort children. Second, for 104 children, the test was stopped ‘too early’, that is before the child had reached the set that was too difficult for him or her. If either of these two mistakes in test administration occurs, the test is considered invalid and no score is contained in the data. This leaves 1,769 children (=1,970—201), but another 22 also had missing PPVT scores, leaving a sample of 1,747 with PPVT scores. Of these 22, 18 suffered from physical or mental disabilities that precluded them from taking the test, two were prevented from taking the test due to insufficient lighting, and one could not take the test due to a vision or hearing problem. Next, consider the tests taken by the older cohort, who were 8 years old in round 1 (2002) and 12 years old in round 2 (2006). Very simple reading and writing tests were administered to these children in both rounds, and a very simple mathematics test was administered in round 1. Most of the children received the highest scores on these tests, leaving little variation to be explained, so these tests will not be analysed in this article. The older cohort also took a mathematics test in round 2, which consists of 10 questions (five multiple-choice and five open-ended questions) chosen from the Trends in International Mathematics and Science Study developed by the IEA in 2003, focusing on number 5. As in the CDA-Q test, some items were found to have poor statistical properties in the Vietnam context and thus were excluded from the calculation of the total PPVT scores.


500


sense and basic mathematical skills. One question was excluded from the analysis owing to potential gender bias, yielding a maximum score of 9. The older cohort also took the same PPVT test taken by the younger cohort in round 2. Of the 990 older cohort children tested in round 2, mathematics scores are missing for nine cases; for two children the test conditions were not appropriate, but no reason is specified for the other seven. PPVT scores are missing for 45 children, for the following reasons. First, one child did not take the test because of a serious mental or physical disability. Second, for 11 children the PPVT test was stopped too early, and for 31 children the basal set was not correctly set during test administration. Finally, the test conditions were inadequate (insufficient light) for two children. Note that 14 children did not have their tests stopped at the right point, but such continuing the test when one should stop still allows one to obtain the correct score (the one that would have been recorded if the test had been stopped at the correct point), and so these observations still have valid test scores. 4. Methodological Issues6 Education increases individuals’ well-being primarily through their acquisition of cognitive (e.g. literacy and numeracy) and noncognitive (e.g. social and organisational) skills. This skill acquisition process can be studied using the education production function framework, which provides crucial guidance on how to use education data to estimate the causal determinants of acquired skills. How cognitive skills are ‘produced’ depends on many factors (or ‘inputs’), including child, household, and school variables. The relationship between skills learned (A) and these factors can be described in a cognitive skills production function:

A = a (Q, C, H, I )

(1)

6. For a more detailed discussion of the issues discussed in this section, see Glewwe and Lambert (2010).

September 2015

where Q is the set of school and teacher characteristics, C is child characteristics and H is household characteristics that affect learning; I is educational ‘inputs’ that households contribute, such as children’s years of schooling and daily attendance, and purchases of textbooks and other educational materials. Unfortunately, production functions are often difficult to estimate. To see why, consider a simple linear specification of equation (1):

A = β0 + βQ1Q1 + βQ 2 Q 2 + … + βC1C1 + βC 2 C2 + … + βH1H1 + βH 2 H 2 (1′) + … + β I1 I1 + β I 2 I 2 + … + u A where each variable in Q, C, H and I is shown explicitly. Assuming linearity is not very restrictive if one adds squared and interaction terms to the variables in (1). The ‘error term’ (uA) accounts for all variables in (1) that are not in the data, as well as for measurement errors in A and in the right-hand-side variables in (1′). By standard econometric theory, the causal impacts of the observed variables in (1′) on learning (the βs) can be consistently estimated by using ordinary least-squares (OLS) only if uA is uncorrelated with ALL of the observed ‘explanatory’ variables. Unfortunately, uA is likely to be correlated with those variables, and so will cause biases in OLS estimates of the βs in (1′) for three reasons.7 4.1 Omitted Variables No dataset contains all the explanatory variables in equation (1). Difficult-to-observe variables include: teachers’ motivation and head teachers’ management skills (Q), children’s ability and motivation (C), parents’ willingness and capacity (H) to help, and their time spent helping (I) with children’s schoolwork. These variables, if unobserved, are probably correlated with some observed variables in (1′), causing omitted-variable biases. For 7. A fourth problem, selection and attrition bias, is not discussed here as it seems unlikely to be a problem with the Young Lives data, which tests all children whether or not they are currently in school. See Glewwe and Lambert (2010) for an explanation of this type of bias.


Glewwe et al.: Young Lives Data Analysis example, ‘high-quality’ schools are usually better in many dimensions, both observed and unobserved. This produces positive correlation between uA and observed school quality variables, causing overestimation of the impacts of those variables. Omitted-variable bias can also lead to underestimation. For example, high school quality may lead parents to reduce time spent helping their children, generating negative correlation between school quality and uA (e.g. working through some unobserved parental efforts). Omitted-variable bias affects estimates of the βs not only for observed variables that are correlated with uA but also for observed variables that are uncorrelated with u A. 4.2 Endogenous Program Placement School quality could also be correlated with uA if governments improve schools with unobserved education problems (Pitt et al. 1993). Governments may also raise school quality in areas with good education outcomes, if those areas have political influence (World Bank 2001). The former causes underestimation of school quality impacts on learning, while the latter causes overestimation. 4.3 Measurement Error Even the best survey data collected in developing countries contain many errors. Data on child, household, and school characteristics may be inaccurate or out of date. Because measurement error is the difference between the true and observed values of a variable, it causes uA to be correlated with the observed variable. Random measurement error typically causes underestimation of true impacts, while non-random errors could cause either underestimation or overestimation. This article addresses each of these estimation problems. Regarding the first problem, omitted-variable bias, the best approach is to collect data on as many of the explanatory variables in equation (1′) as possible. Given the richness of the Young Lives data, many, and perhaps even most, of the variables in (1′)

501

are available. In addition, in regressions the impacts of all school quality variables can be controlled for by using commune fixed effects. Strictly speaking, this works only if the school quality variables do not interact with the other variables that determine test scores, which is unlikely to hold. Nevertheless, the ability of these dummies to fully control for the impacts that do not include interaction terms should remove the most serious estimation problems due to unobserved school quality variables. A related point is that community fixed effects work best if all children enroll in the school located in their commune and there is only one school per commune (or there are multiple schools of similar quality). While children generally enroll only in schools in their commune, about half of the communes have more than one school (see Appendix 1 Table A2). Yet in most of the communes with multiple primary schools, those schools were probably very similar, since they had names such as ‘An Hoa Primary School No. 1’ and ‘An Hoa Primary School No. 2’, and only three older cohort students were attending private primary schools. We tried to obtain the names of primary schools attended for older cohort children (which would allow us to use school fixed effects instead of commune fixed effects), but we have only the current school attended, which for most older cohort children is the lower secondary school. The second problem, endogenous program placement, seems unlikely to be a problem in the Vietnamese context. District governments are responsible for financing primary and preschool education in Vietnam, and historically there has been little attempt by the central Government to provide additional resources to poorer districts (World Bank 1997). The amount of money transferred by provincial or district governments to schools is a set amount per pupil, or more recently a set amount per school-aged child, without reference to academic performance (World Bank 2003; Nordic Consulting Group 2008). Although in recent years the Government has instituted some programs that are intended to provide more resources to poorer districts and communes, the amount of funds allocated by these initia-


502


tives is very small (less than 1 percent of total school expenditure) and the share of primary school expenditure received from these sources appears to be uncorrelated with district-level poverty rates (Nordic Consulting Group 2008), which suggests little reason to worry about bias from endogenous program placement. Finally, regressions using school/ commune fixed effects will avoid bias due to any omitted school variables, as explained in the previous paragraph. The third estimation problem, measurement error bias, can be addressed using instrumental variables as long as the measurement errors in the instruments are uncorrelated with those in the explanatory variables. The variable most likely to suffer from measurement error is household expenditure per capita, which will be instrumented using a household wealth index. Another variable that could suffer from measurement error is the height-for-age z-score, an indicator of early childhood malnutrition. The measurement error arises from the fact that there is variation in height, and weight, even among healthy children, which implies that this z-score is a noisy measure of early childhood nutritional status. In estimates not reported (but available upon request), we used each child’s weight-for-age z-score as an instrument for his/her height-for-age z-score. This had little effect on the results, and in any case it is unlikely to be a valid instrument because the same ‘noise’ in height-for-age also affects weight-for-age (among healthy children some are naturally taller, and they will usually also be heavier). Given that the main aim of this article is to examine the underlying causes of the Kinh– ethnic minority learning differences, the decomposition proposed by Blinder (1973) and Oaxaca (1973) is adopted to explore these differences. More specifically, consider estimating equation (1′) separately for the Kinh (subscripted k) and ethnic minority (subscripted m) populations:8 8. In Vietnam, the Chinese ethnic minority is not considered disadvantaged, and shares many cultural similarities with the Kinh, so in virtually all studies of ethnic minorities in Vietnam, the Chinese are grouped together with the Kinh and together they are considered the ‘ethnic major-

September 2015

A k = β0 k + β k ′ x k + u Ak

(2)

A m = β0 m + βm ′ x m + u Am

(Kinh) (3)

(ethnic minority) where all the coefficients in equation (1′) except the constant term have been incorporated into the β vectors, and all of the explanatory variables have been incorporated into the x vectors. Note that the impacts of the x variables on test scores (the βs) can differ across the two ethnic groups. Since the mean values of the error terms (uAk and uAm) in equations (2) and (3) equal zero, the following relationships hold, where ‘bars’ indicate mean values of variables for the respective ethnic group:

A k = β0 k + β k ′ x k

(2′)

A m = β0 m + βm ′ xm

(3′)

Thus, the between-group difference in the mean test scores can be expressed as:

A k – A m = (β0 k – β0 m ) + (β k ′ x k – βm ′ xm ) (4) The term (β k ′ x k − βm ′ xm ) can be decomposed into two parts, one reflecting the between-group difference in the means of the x variables and the other reflecting the betweengroup difference in the coefficients:

A k – A m = (β0 k – β0 m ) + β k ′ ( x k – xm ) (5) + (β k – βm ) ′ xm This decomposition can also be done in another, analogous, way, which multiplies the between-group differences in the means by βm and multiplies the between-group differences in the βs by x k :

ity’. This article follows this classification, although in practice it makes no difference because only one of the 2,000 younger cohort children is Chinese, and none of the 1,000 older cohort children is Chinese.


Glewwe et al.: Young Lives Data Analysis

A k – A m = (β0 k – β0 m ) + βm ′ ( x k – xm ) + (β k – βm ) ′ x k

(6)

In most cases, these two decompositions should give similar results. Both will be shown below. To see how to interpret these decompositions, consider equation (5). The first term, β0k − β0m, shows the differences in test scores unaccounted for by mean differences in the x variables, nor by the differences in their impacts (the βs), for some reference value for x (averaged over all groups). Many studies implicitly set the reference value of x to zero, but this could be misleading if some x variables never attain a value of zero. For example, consider the case of only one explanatory variable for a test score, namely years in school. Suppose that years of school has a stronger effect on Kinh children than on ethnic minority children, perhaps because the former attend higher quality schools. If all children have had 4 or more years of schooling, fitting separate regression lines for both Kinh and ethnic minority children will lead to a line with a higher slope for Kinh children, which if extended back to the point where years of schooling equals 0 could lead to an intercept for Kinh children that is lower than that for ethnic minority children, that is, β0k ⬍ β0m, which could be mistakenly interpreted as ‘discrimination’ against Kinh children even if the regression line for the Kinh lies above that of the ethnic minority children for all values of years of schooling in the population. The point here is that interpretation of β0k − β0m must be done very cautiously. This article sets the reference value as the lowest value for each x variable; thus β0k − β0m is the difference in the test scores of a child for whom all the x variables take their lowest value. This should be kept in mind when interpreting the contribution of β0k − β0m to A k − A m . The second term in equation (5), β k ′ ( x k − xm ) , is the sum, over all the x variables, of the contributions of the Kinh–ethnic minority differences in the mean values of the x variables to explaining the Kinh–ethnic minority differences in the mean test scores. For example, as will be seen below, on

503

average, Kinh children have better-educated parents than do ethnic minority children, and educated parents have a direct, positive impact on test scores. In fact, the contributions of each of the x variables can also be obtained separately from the regression estimates, and indeed are shown below. Finally, the last term in equation (5), (β k − βm ) ′ xm , is the sum, over all the x variables, of the contributions to the Kinh–ethnic minority gap in mean test scores caused by the Kinh–ethnic minority differences in the impacts of the x variables on test scores. For example, it may be that parental education has a higher impact on test scores for the Kinh than for ethnic minorities, perhaps due to higher school quality among the Kinh population. Again, while this term sums up these effects for all the x variables, the impact for each x variable can be identified in the regression (shown below). Recall that this article controls for differences in school quality by controlling for commune fixed effects. Note that, while in most of the communes (22 or 23 out of 31, depending on the cohort) are ‘segregated’ (either all Kinh or all ethnic minority), the effects of the school variables could vary by ethnic group in the (8 or 9 out of 31) ‘mixed’ communes. Thus, in those communes alone the regressions include an interaction term between the commune dummy and the ethnic minority dummy to capture this (potential) differential. A final issue with the use of commune fixed effects is that there is no longer a constant term for either ethnic group. In effect, the constant term for the Kinh is the (weighted) average of the commune dummy variables for the communes in which the Kinh population resides, and the same for the ethnic minority population (in which case the interaction term between the ethnic minority dummy and the commune dummy in the ‘mixed’ communes must be added to the commune dummy). The averages of these dummy variables are shown in the results below. Note that they represent differences in school quality as well as the (β0k − β0m) term; these two differences cannot be distinguished in estimation using commune fixed effects.


504


5. Analysis for the Younger Cohort This section examines the test scores of the younger cohort who were 5 years old when tested in 2006.The analysis first compares the test scores of Kinh and ethnic minority children, first for the entire sample and then for the subsample of ‘mixed’ communes. It then presents regression estimates of the determinants of test scores, and uses them to explain the Kinh–ethnic minority gaps in test scores. Table 1 presents means and standard deviations (SDs) of the CDA-Q and PPVT scores, first for the full sample and then separately for the two ethnic groups. Panel A shows statistics for all students, while panel B is limited to the nine ‘mixed’ communes. Beginning with panel A, the Kinh–ethnic minority gap in the CDA-Q (mathematics) score is quite large: the mean score of the Kinh is 10.2 (out of 14) while that of ethnic minorities is 7.4. The difference of 2.8 points is equivalent to 1.1 SDs of the distribution of test scores. The mean PPVT scores were 39.4 and

September 2015

23.5 for the Kinh and ethnic minorities, respectively. The gap, 15.9 points, is equivalent to 0.9 SDs of the distribution of test scores. It is also informative to distinguish between ethnic minority children whose ‘mother tongue’ is Vietnamese and those whose mother tongue is not. (The ethnic groups least likely to speak Vietnamese as their mother tongue are the H’mong (3 per cent) and the Dao (24 per cent), while those most likely to speak Vietnamese are the Cham (97 per cent) and the Ba Na (64 per cent).) The former group (about one third of ethnic-minority children) scored much higher than the latter group. More specifically, the mean CDA-Q score for the former group (8.3) is about 0.5 SDs larger than that for the latter group (6.9). The difference on the PPVT test is even higher: the former group scored (30.1) about 0.6 SDs higher than the latter group (20.1). As explained above, part of the Kinh–ethnic minority gap in the test scores may reflect the fact that the two ethnic groups live in different communities and attend different schools. To

Table 1 Mean Test Scores for Ethnic Majority and Ethnic Minority Children (Younger Cohort, 5 Years Old) Student type A: All communes Full sample Kinh Ethnic minority Ethnic minority (speaks Vietnamese) Ethnic minority (speaks other language) B: Mixed communes Full sample Kinh Ethnic minority Ethnic minority (speaks Vietnamese) Ethnic minority (speaks other language)

Variable (raw score)

Mean

Standard deviation

Number of observations

CDA-Q PPVT CDA-Q PPVT CDA-Q PPVT CDA-Q PPVT CDA-Q PPVT

9.79 37.0 10.20 39.4 7.36 23.5 8.26 30.1 6.88 20.1

2.51 18.2 2.29 18.0 2.34 12.2 2.09 12.1 2.32 10.7

1,906 1,747 1,631 1,480 275 267† 95 92 180 174

CDA-Q PPVT CDA-Q PPVT CDA-Q PPVT CDA-Q PPVT CDA-Q PPVT

8.94 32.1 9.78 37.0 7.86 25.8 7.90 29.8 7.83 23.2

2.35 14.8 1.97 14.6 2.36 12.6 2.23 13.3 2.45 11.5

369 356 209 202 160 154 61 60 99 94

†Small discrepancies in some of these figures are due to missing data on whether one ethnic minority child speaks Vietnamese or another language. CDA-Q, Cognitive Development Assessment test-quantitative; PPVT, Peabody Picture Vocabulary Test. © 2015 The Authors. Asia and the Pacific Policy Studies published by Crawford School of Public Policy at The Australian National University and Wiley Publishing Asia Pty Ltd

505 Notes: Per capita expenditure is instrumented, using a household wealth index. Significance is based on robust standard errors, clustered at the commune level. Sample size is 1,815. Significance at the 10%, 5% and 1% levels is denoted by *, ** and ***, respectively.

0.069 0.113 0.174 0.000 0.066 0.001 0.027 0.0 0.025 −1.184 0.0 0.0 0.0 0.480 0.0 0.0 0.0 0.0 0.911 0.113 0.174 0.000 0.009 0.001 0.027 0.000 0.025 −2.027 0.0 0.0 0.0 0.535 0.0 0.0 0.0 0.0 1.943 8.37 7.72 0.49 15.28 3.977 5.541 6.739 17.62 Log (per capita expenditure) Father’s education Mother’s education Girl Child’s age Height-for-age z-score Log (educ. expenditure) Time spent in nursery Time spent in pre-school Average const. (segregated) Average const. (mixed)

0.087 0.022** 0.031*** 0.015 0.042*** 0.001 0.009 0.000 0.004 −1.226 −0.868

1.128*** 0.022** 0.031*** 0.015 0.006 0.001 0.009 0.000 0.004 −2.545 −2.325

−1.043*** 0.0 0.0 0.0 −0.035** 0.0 0.0 0.0 0.0

1.135 3.24 2.11 0.463 13.71 2.863 2.501 0.555 11.48

(β k − β m ) ′ xm β m ′ ( xk − xm )

(β k − β m ) ′ xk xm xk βk − βm βm βk Variables

Table 2 Regression Estimates for CDA-Q Test, Younger Cohort

‘control for’ differences in communities and schools, panel B of Table 1 limits the comparison across ethnic groups to the nine ‘mixed’ communes. This does reduce the gaps somewhat. The difference in the CDA-Q score (1.9) is 32 per cent smaller than the gap when comparing all communes (2.8). Similarly, the difference in the PPVT score (11.2) is 30 per cent smaller than the gap when comparing all communes (15.9). Nevertheless, there are still large Kinh–ethnic minority gaps even in the ‘mixed’ communes. Note also that the gap between Vietnamese-speaking and nonVietnamese-speaking ethnic minorities is much smaller than for the full ethnic minority sample; indeed for the CDA-Q test there is almost no difference. To understand better the nature of the gaps, we next present regressions that attempt to explain the Kinh–ethnic minority gaps in the CDA-Q (Table 2) and PPVT scores (Table 3). The test scores are standardised to have an SD of 1, for ease of interpretations of the coefficients. Recall that the Blinder–Oaxaca decomposition divides the overall Kinh–ethnic minority gaps in mean test scores into three parts: (i) the differences attributable to the differences in the means of the xs; (ii) the differences due to different impacts of the xs (the βs) on test scores; and (iii) differences in the constant terms (the β0s). In fact, the younger cohort includes only 283 ethnic minority children, so the precision of the estimated βs for ethnic minorities is often low, making these estimates not significantly different from those for the Kinh children. In this case, allowing separate estimates for the two ethnic groups can produce large apparent ‘explanations’ of the test scores gaps that are, in fact, not statistically significant. Thus, whenever the difference βk − βm was not statistically significant for a given variable, the two associated βs were constrained to be equal. This should also increase the precision of estimates of the impact of differences in the means of the x variables on the Kinh–ethnic minority test scores gaps. The first and second columns of Table 2 show the estimates of βk in equation (5) and

β k ′ ( xk − xm )



Notes: Per capita expenditure is instrumented, using a household wealth index. Significance is based on robust standard errors, clustered at the commune level. Sample size is 1,668. Significance at the 10%, 5% and 1% levels is denoted by *, ** and ***, respectively.

−0.621 0.0 0.0 0.0 0.493 0.0 0.0 0.0 0.0 0.718 0.103 0.185 −0.001 0.029 0.056 0.024 0.011 0.023 −1.063 0.0 0.0 0.0 0.550 0.0 0.0 0.0 0.0 1.943 8.37 7.72 0.49 15.28 3.977 5.541 6.739 17.62 Log (per capita expenditure) Father’s education Mother’s education Girl Child’s age Height-for-age z-score Log (educ. expenditure) Time spent in nursery Time spent in pre-school Avg. const. (segregated) Avg. const. (mixed)

0.341*** 0.020** 0.033*** −0.041 0.054*** 0.050 0.008 −0.002 0.004 −2.094 −1.913

0.888*** 0.020** 0.033*** −0.041 0.018** 0.050 0.008 −0.002 0.004 −2.086 −2.408

−0.547** 0.0 0.0 0.0 0.036*** 0.0 0.0 0.0 0.0

1.135 3.24 2.11 0.463 13.71 2.863 2.501 0.555 11.48

(β k − β m ) ′ xm β m ′ ( xk − xm )

(β k − β m ) ′ xk xm xk βk − βm Variables

βk

βm

Table 3 Regression Estimates for PPVT Test, Younger Cohort

0.276 0.103 0.185 −0.001 0.085 0.056 0.024 −0.012 0.024

Asia & the Pacific Policy Studies β k ′ ( xk − xm )

506

September 2015

those of βm in equation (6), respectively.9 The third column shows the difference in the two estimates. (The actual regression coefficients are those in columns 1 and 3 of Table 2, and statistical significance is shown for those two columns; column 2 is calculated as the difference, and calculating its statistical significance is straightforward.) Somewhat surprisingly, household per capita expenditure has no significant effect on Kinh children’s CDA-Q scores, but it has a large and statistically significant effect among ethnic minority children. This difference reflects partly the fact that ethnic minorities are poorer, and income effects may be stronger for the poor. Both father’s and mother’s education have strong impacts on the test scores of both groups of children, although the betweengroup differences in their impacts were not statistically significant. Child age had a significant impact on the CDA-Q test for Kinh children but not for ethnic minority children. Presumably, this reflects the fact that older children are, ceteris paribus, more mature and thus have acquired more skills; recall that almost none of these children have started primary school so there is no variation in the sense that older children have been at school longer. It is not clear why this ‘maturity’ effect does not show up very strongly among ethnic minority children. For all of the remaining variables—child sex, height-for-age z-scores, educational expenditure on the child, months spent in a community nursery (crèche) from birth to 36 months of age, and months spent in a preschool since 36 months of age—neither the coefficient nor the Kinh–ethnic minority difference in the coefficients is statistically significant. The last two lines in Table 2 show the average constant term (average community fixed effect) for the two ethnic groups, separately for segregated (all Kinh or all ethnic minority) and mixed communes. Two main lessons can be drawn from these figures. First, 9. These are instrumental variable estimates; to minimise measurement error bias, the (log of) per capita expenditure variable is instrumented by an index of household wealth.


Glewwe et al.: Young Lives Data Analysis for both ethnic groups the constant terms are fairly similar whether they live in segregated or mixed communes; in particular, there is no large advantage for ethnic minority children and no disadvantage for Kinh children from living in a mixed commune. Second, there is a Kinh–ethnic minority gap of 1.3–1.5 SDs of a test score, even after controlling for all other variables. This is rather surprising given that the unadjusted difference in Table 1 (about 1.1 SDs) was smaller. However, as discussed above, these differences in the (average) β0 terms are difficult to interpret and may simply reflect ‘adjustments’ in the constant term to accommodate differences in the ‘slope’ parameters that are statistically significant across the two groups. The Blinder–Oaxaca decomposition can be used to see how much of the observed Kinh– ethnic minority gap in the mean CDA-Q scores is explained by the differences in the means of the x variables (shown in columns 4 and 5 of Table 2) and how much is explained by differences in the impacts of those variables. Note that any variable that has no significant explanatory power for either group will have little role to play. The role played by per capita expenditure is the strongest. Since Kinh children live in wealthier households, they can gain (using the ethnic minority coefficient) about 0.9 SDs of a test score, although this effect almost disappears when using the small Kinh coefficient (0.1 SDs). More consistent across the two decompositions, the difference in those two coefficients is highly favorable to ethnic minority children; they benefit much more than Kinh children from an increase in household income, adding 1.2–2.0 SDs to their test scores. Overall (combining the two separate parts of the decomposition), rather than explaining why ethnic minorities have lower CDA-Q scores than do Kinh children, household income in effect increases the gap by about 1.1 SDs. Higher fathers’ and mothers’ education among Kinh children together explain about 0.3 SDs of the gap in the CDA-Q scores. Intuitively, parental education raises test scores, and Kinh children’s parents are much more

507

educated than the parents of ethnic minority children; their fathers have, on average, 5 more years of education and their mothers have almost 6 more years of education. Child age also explains a sizeable part of the Kinh–ethnic minority gap; Kinh children ‘mature’ in some way that ethnic minority children do not, which boosts their scores by about 0.5 SDs. But exactly what is happening here is far from clear; it is unlikely to be a ‘biological’ maturation so presumably it may reflect something about Kinh culture. On the other hand one cannot rule out that this difference could be spurious; there are nine explanatory variables in Table 2, and even if all the differences in the parameters between Kinh and ethnic minority children were zero, there is a 1 out of 20 chance that any given difference will be significant at the 5 per cent, and this may be such a case. Overall, while the regression coefficients usually have the expected signs, it is not clear that the Blinder–Oaxaca decomposition is providing clear insights into the causes of the differences in the CDA-Q scores among the younger cohort. Perhaps the main lesson is that it matters little whether Kinh or ethnic minority children live in segregated communities or mixed communities. Next, consider the Kinh–ethnic minority gap in the PPVT scores. The estimates in Table 3 are analogous to those for the CDA-Q test (Table 2). The first notable result is that, for Kinh children, household expenditure per capita has a strong and statistically significant impact, in contrast with the small and insignificant impact it has on the CDA-Q scores. The impact is even stronger for ethnic minority children, and the difference in these impacts is quite large and statistically significant. In addition, both mothers’ and fathers’ education have significantly positive impacts on both ethnic groups’ PPVT scores, as they did on the CDA-Q scores (although the betweengroup differences in the coefficients are again statistically insignificant). As with the CDA-Q test, child age has a strong positive impact on PPVT scores for both groups, although the impact on ethnic minority children is much smaller than on Kinh children, the difference


508


being statistically significant. Intuitively, children of better-educated parents learn more quickly, and as they get older (even before entering primary school), they learn more. For other variables, as seen in Table 2, none has a statistically significant effect for either ethnic group, and the differences were also statistically insignificant. Finally, the last two lines of Table 3 show average constant terms across communes for both ethnic groups, again separately for segregated and mixed communes. As before, it matters little whether either group of children lives in a segregated or mixed commune. More interestingly, the between-group differences are much smaller than in Table 2: only about 0.5 SDs of a test score for mixed communes, and no difference at all for segregated communes. This implies that most of the raw difference in PPVT scores (about 0.9 SDs) is explained by the regression results. Now turn to the Blinder–Oaxaca decompositions for the Kinh–ethnic minority gap in the PPVT scores, which was about 0.9 SDs of the distribution of the PPVT score. The two decompositions indicate that the difference in the mean of (log) per capita expenditures accounts for 0.3–0.7 SDs of this gap. But the coefficient is much higher for ethnic minorities, which implies that, for an average ethnic minority child, there is a 0.6–1.1 SD benefit. As before, rather than explaining the gap, the combined effect of per capita expenditure increases it by about 0.3 SDs. Similar to the case with the CDA-Q scores, differences in the means of parental education explain about 0.3 SDs of the Kinh–ethnic minority gap, after summing the effects of both parents. Child age has a much larger impact on Kinh children than on ethnic minority children, explaining a very large part, about 0.5 SDs, of that gap. As one would expect, none of the other variables provides any sizeable explanation of the gap, given that they are all statistically insignificant. 6. Analysis for the Older Cohort This section analyses the data for the older cohort children, who were 12 years old in

September 2015

2006. As in Section 5, the analysis first documents the Kinh–ethnic minority gaps in the test scores, first for the entire sample and then for mixed communes. It then presents estimates of a cognitive skills production function that attempt to explain the determinants of test scores and the gaps in test scores. Table 4 presents means and SDs of the mathematics and PPVT scores for the full sample and separately for the two ethnic groups. Panel A shows statistics for all students, while panel B limits the sample to the nine mixed communes. As with the younger cohort, the Kinh–ethnic minority gap is quite large: the difference in the mean mathematics score between the Kinh (7.8 out of 9) and ethnic minorities (5.3) is 2.5 points, equivalent to 1.3 SDs of the distribution of test scores. Similarly, the mean PPVT score for the Kinh (142.3) was much higher than that for ethnic minorities (104.3). This 38-point gap is equivalent to 1.5 SDs of the distribution of test scores. Moreover, as seen with the younger cohort, Vietnamese-speaking ethnic minority children do much better than their non-Vietnamese-speaking counterparts. Limiting the above comparison to the nine mixed communes reduces the Kinh–ethnic minority gaps to some extent. The gap in the mathematics test (1.8) is 28 per cent smaller than the gap when comparing all communes (2.5). Similarly, the gap in the PPVT score (25.2) is 34 per cent smaller than the gap when comparing all communes (38.0). Nevertheless, there are still large Kinh–ethnic minority gaps even among children in the same commune. Note also that in the mixed communes the gaps between the Vietnamese-speaking and nonVietnamese-speaking ethnic minority children are much smaller than was the case for all communes. To understand better the nature of the gaps, turn next to regression estimates and Blinder– Oaxaca decompositions for the Kinh–ethnic minority gaps in the mathematics (Table 5) and PPVT (Table 6) scores. Again, the test score has been standardised to have an SD of 1. The first and second columns of Table 5 show the estimates of βk in equation (5) and those of βm in equation (6), respectively, for the



509

Table 4 Mean Test Scores for Ethnic Majority and Ethnic Minority Children (Older Cohort, 12 Years Old) Student type

Variable (raw score)

A: All communes Full sample

Math (IEA) PPVT Kinh Math (IEA) PPVT Ethnic minority Math (IEA) PPVT Ethnic minority (speaks Vietnamese) Math (IEA) PPVT Ethnic minority (speaks other language) Math (IEA) PPVT B: Mixed communes Full sample Math (IEA) PPVT Kinh Math (IEA) PPVT Ethnic minority Math (IEA) PPVT Ethnic minority (speaks Vietnamese) Math (IEA) PPVT Ethnic minority (speaks other language) Math (IEA) PPVT

Mean

Standard deviation

Number of observations

7.44 137.6 7.75 142.3 5.28 104.3 6.27 119.8 4.18 86.8

1.92 26.1 1.51 18.8 2.78 41.5 2.31 31.6 2.85 45.0

981 945 855 827 126 118† 66 63 60 54

6.62 130.4 7.44 141.8 5.64 116.6 5.86 117.9 4.32 117.9

2.32 29.1 1.58 18.6 2.66 33.3 2.47 30.7 3.40 33.3

217 206 118 113 71 68 49 48 22 19

†Small discrepancies in some of these figures are due to missing data on whether one ethnic minority child spoke Vietnamese or another language. IEA, International Evaluation Association; PPVT, Peabody Picture Vocabulary Test.

mathematics score of the older cohort. The third column shows the few cases of statistically significant differences in the two estimates. Household per capita expenditure has a significantly positive impact on the mathematics scores of both ethnic groups (the difference between the two coefficients is statistically insignificant). Both mothers’ and fathers’ education also have significantly positive impacts on the mathematics scores of both groups. Turning to child characteristics, although there are no gender differences for Kinh children there is a strong gender difference for ethnic minorities: girls scored about 0.3 SDs higher than boys, ceteris paribus. Years of schooling has a strong and statistically significant impact on both groups of children, with a significantly larger impact on the ethnic minorities. Also, the hours per day spent in school has a strong and significantly positive impact on both ethnic groups (although the difference between these two impacts is statistically insignificant). Finally, hours spent

working has a negative impact, significant at the 10 per cent level for both groups. The last five variables in Table 5 measure different aspects of health and disability, and three are statistically significant for both ethnic groups (with no statistically significant different impacts by ethnic group). First, the heightfor-age z-score has a significantly positive impact. Second, children who have difficulty understanding what their parents are saying have much lower scores (0.7 SDs lower). Third, children who had had an injury or episode of illness in the last 4 years that was so severe that the parents thought they might die have significantly lower scores (0.2 SDs lower). Examination of the average constant terms again shows that, it matters little whether Kinh students are in a segregated or a mixed commune. However, for ethnic minority students, living in a mixed commune implies a drop of about 1.2 SDs in a test score relative to living in a segregated commune. Moreover, ethnic minority students seem to do worse in


0.269** 0.024** 0.024*** 0.014 −0.008 0.009 0.256*** 0.141*** 0.010 −0.048* 0.003 0.064** −0.023 −0.668*** −0.043 −0.244** −3.327 −3.474

βk

0.269** 0.024** 0.024*** 0.014 0.271* 0.009 0.358** 0.141*** 0.010 −0.048* 0.003 0.064** −0.023 −0.668*** −0.043 −0.244** −3.380 −4.582

βm 0.0 0.0 0.0 0.0 −0.279* 0.0 −0.102* 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

βk − βm 2.085 8.515 7.651 6.027 0.502 15.163 5.954 4.504 2.901 1.826 1.913 3.728 0.208 0.015 0.063 0.056

xk 1.384 2.902 1.619 2.905 0.503 13.669 5.133 4.000 1.579 3.495 0.291 2.721 0.007 0.031 0.086 0.055

xm 0.701 5.613 6.032 3.122 −0.001 1.494 0.821 0.504 1.322 −1.669 1.622 1.007 0.201 −0.016 −0.023 0.001

( xk − xm ) 0.0 0.0 0.0 0.0 −0.140 0.0 −0.607 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

(β k − β m ) ′ xk 0.189 0.135 0.145 0.044 −0.000 0.013 0.294 0.071 0.013 0.080 0.005 0.064 −0.005 0.011 0.001 −0.000

β m ′ ( xk − xm ) 0.0 0.0 0.0 0.0 −0.150 0.0 −0.688 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

(β k − β m ) ′ xm

0.189 0.135 0.145 0.044 0.000 0.013 0.210 0.071 0.013 0.080 0.005 0.064 −0.005 0.011 0.001 −0.000

β k ′ ( xk − xm )

Notes: Per capita expenditure is instrumented, using a household wealth index. Significance is based on robust standard errors, clustered at the commune level. Sample size is 893. Significance at the 10%, 5% and 1% levels is denoted by *, ** and ***, respectively.

Log (per capita exp.) Father’s education Mother’s education Log (educ. expend.) Girl Child’s age Years of schooling Hours in school/day Hours spent studying/day Hours spent working/day Extra math class (last 6 months) Height-for-age z-score Hearing problem Understands what parents say Long-term health problem Serious illness/injury (last 4 years) Average const. (segregated) Average const. (mixed)

Variables

Table 5 Regression Estimates for Mathematics (IEA) Test, Older Cohort

510 Asia & the Pacific Policy Studies September 2015


0.396*** 0.024*** 0.007 −0.030 −0.080 0.025*** 0.320*** 0.036 0.001 −0.011 0.021* 0.037 −0.618*** −0.191* −0.148* −0.104 −3.267 −3.370

βk

0.396*** 0.024*** 0.099 −0.030 −0.080 0.025*** 0.320*** 0.036 0.001 −0.011 0.021* 0.037 −0.618*** −0.191* −0.148* −0.104 −5.402 −2.673

βm 0.0 0.0 −0.092** 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

βk − βm 2.085 8.515 7.651 6.027 0.502 15.163 5.954 4.504 2.901 1.826 1.913 3.728 0.208 0.015 0.063 0.056

xk 1.384 2.902 1.619 2.905 0.503 13.669 5.133 4.000 1.579 3.495 0.291 2.721 0.007 0.031 0.086 0.055

xm 0.701 5.613 6.032 3.122 −0.001 1.494 0.821 0.504 1.322 −1.669 1.622 1.007 0.201 −0.016 −0.023 0.001

( xk − xm ) 0.0 0.0 −0.704 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

(β k − β m ) ′ xk 0.278 0.135 0.597 −0.094 0.000 0.037 0.263 0.018 0.001 0.018 0.034 0.037 −0.124 0.003 0.003 −0.000

β m ′ ( xk − xm ) 0.0 0.0 −0.149 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

(β k − β m ) ′ xm

0.278 0.135 0.042 −0.094 0.000 0.037 0.263 0.018 0.001 0.018 0.034 0.037 −0.124 0.003 0.003 −0.000

β k ′ ( xk − xm )

Notes: Per capita expenditure is instrumented, using a household wealth index. Significance is based on robust standard errors, clustered at the commune level. Sample size is 860. Significance at the 10%, 5% and 1% levels is denoted by *, ** and ***, respectively. PPVT, Peabody Picture Vocabulary Test.

Log (per capita exp.) Father’s education Mother’s education Log (educ. expend.) Girl Child’s age Years of schooling Hours in school/day Hours spent studying/day Hours spent working/day Extra math class (last 6 months) Height-for-age z-score Hearing problem Understands what parents say Long-term health problem Serious illness/injury (last 4 years) Average const. (segregated) Average const. (mixed)

Variables

Table 6 Regression Estimates for PPVT Test, Older Cohort

Glewwe et al.: Young Lives Data Analysis 511


512


mixed communes than do Kinh students, a loss of about 1.1 SDs. Yet caution is in order for the results pertaining to ethnic minorities living in a segregated commune because they are based on a single commune. Next, consider the decompositions for the Kinh–ethnic minority gaps in the test scores of the older cohort (last four columns of Table 5). The difference in (log of) per capita expenditure explains about 0.2 SDs of the gap. Differences in fathers’ and mothers’ education together explain another 0.3 SDs. The fact that ethnic minority girls do much better than boys actually increases the gap by about 0.15 SDs. Turn next to the time-related variables. First, although the fact that Kinh children have somewhat more years of schooling explains about 0.2–0.3 SDs of the gap, the larger impact of years of schooling for ethnic minority children ‘unexplains’ 0.6–0.7 SDs of that gap. On the other hand, about 0.1 SDs of the gap are explained by the fact that Kinh children spend about 30 more minutes per day in school than do ethnic minority children, and the fact that they work 1.7 fewer hours per day explains another 0.1 SDs of that gap. Finally, better nutrition (height-for-age) among Kinh children explained another 0.1 SDs. None of the other components of the decomposition is very large, and so none of them has noticeable explanatory power. Table 6 shows the regression and decomposition results for PPVT scores of the older cohort. For both groups of children, household per capita expenditure and fathers’ education (but not mothers’ education for Kinh) have strong impacts on learning. Girls do slightly worse, but child age and years of schooling have significantly positive impacts on the PPVT score. Only one of the health variables, a hearing problem, has a large and significantly negative effect, which is intuitive since hearing problems are presumably more important for reading skills than for mathematics skills. The average constant terms show a 0.7 SD difference between Kinh and ethnic minority children who live in mixed communes, but show no difference between Kinh children living in segregated communes and those in

September 2015

mixed communes. However, there is a very large gap between ethnic minority children living in segregated communes and those living in mixed communes; the former have test scores almost 2.7 SDs lower. Again, caution is needed as the sample contains only one segregated ethnic minority commune. Finally, turn to the decomposition results. Kinh children’s test scores are 0.3 SDs higher than ethnic minority children because, on average, they live in wealthier households. The positive impact of father’s education, in conjunction with the higher levels of that variable among Kinh children, explains another 0.1 SDs. However, given the much higher impact of mother’s education among ethnic minority children, overall, that variable ‘unexplains’ about 0.1 SDs of the test score gap. The significant impact of years of schooling, combined with a higher value for that variable for the Kinh children, explains almost 0.3 SDs of the gap. Finally, Kinh parents are more likely to report that their children have a hearing problem, which ‘unexplains’ about 0.1 SDs of that gap. 7. Conclusions Ethnic minority children in Vietnam have much lower mathematics and reading test scores than do ethnic Kinh children, both among a group of 5-year-old children and among another group of 12-year-old children tested in 2006. Given the importance of education in determining adults’ socio-economic success, and the generally lower incomes of ethnic minorities in Vietnam (Baulch et al. 2004), this suggests that today’s ethnic minority children will be poorer than today’s Kinh children when they reach adulthood. A major policy challenge for the Vietnamese government, and for donor agencies active in Vietnam, is to understand the causes of these disparities and then to formulate policies that can reduce them. This article has investigated the causes of these disparities, using the Round 2 Young Lives data from Vietnam. Although the findings raise many questions, suggesting the need for further research, some conclusions can be drawn based on these findings.


Glewwe et al.: Young Lives Data Analysis A first, rather obvious but perhaps overlooked, finding is that these disparities are already very large even before children start primary school, as was seen with the younger cohort data. It is possible that pre-school and even nursery factors play a role in generating these disparities, but it is difficult to pin this down with the data available. Indeed, time spent in nurseries and pre-schools had no explanatory power in determining the younger cohort’s test scores, so the problem may lie elsewhere. Second, language may play an important role. Tables 1 and 4 show that, in general, Vietnamese-speaking ethnic minority children had much higher scores than their nonVietnamese-speaking counterparts. Yet when comparisons are limited to children living in ‘mixed’ communes, these differences were smaller, which suggests that it is really differences between communes, not language itself, that matter. Perhaps one benefit of living in an ethnically mixed commune is that ethnic minority children interact with other children in Vietnamese, and so obtain much more facility in that language at an early age. Note that all tests were administered in whatever language the children wanted to take them in, so the poor performance of ethnic minority children on these tests is not simply due to being forced to take the test in Vietnamese. Clearly, much more research needs to be done on the role of mother tongue, and the mother tongue of children’s peers, to understand more fully why ethnic minority children fall behind. Third, the Blinder–Oaxaca decompositions offer at least a partial explanation of the Kinh– ethnic minority gap in test scores. For the younger cohort, the role played by household expenditure is puzzling because although

513

ethnic minority children live in poorer households, they are more successful in ‘converting’ what little income they have into higher test scores. In contrast, for the older cohort there is no ambiguity; Kinh and ethnic minority children are equally capable at ‘converting’ household income into higher test scores, and the higher per capita expenditure of Kinh households explains about 0.2–0.3 SDs of the gap (of 1.3–1.5 SDs) in test scores. Parental education also plays a role, usually explaining about 0.3 SDs of the gap for both the younger and the older cohorts (the one exception being the impact of mother’s education on the older cohort’s PPVT scores, which is difficult to interpret). None of the other variables offered much explanatory power for explaining the gap among the younger cohort. Among the older cohort, more time spent in school, less time spent working, and better nutritional status each explain about 0.1 SDs of the gap in the mathematics score, and more years of schooling among Kinh children explains about 0.3 SDs the gap for the PPVT score. Further progress on understanding the causes of the Kinh–ethnic minority gap in learning may require different data than those analysed in this article. Qualitative research along the lines of World Bank (2009) and Vietnam Academy of Social Sciences (2009) could be quite useful, and more quantitative analysis using a dataset with a larger number of ethnic minority students and more detailed school data could be desirable. The 2006 Vietnam Household Living Standards Survey is a promising dataset for further analysis of this learning gap. July 2015.


514


September 2015

Appendix 1 Table A1 Comparison of Young Lives Data to Nationally Representative Data Younger cohort Variable Enrolled in school (%) Ethnic minority (%) Mother’s years of schooling Father’s years of schooling Height-for-age z-score (2002 VNHS) Stunted in 2002 (HAZ ⬍ −2) (%) Family ownership of durable goods (%): Motorbike TV Refrigerator Urban (%) Electricity (%) Water (%): Borewell Piped Other well Rain water Other Toilet (%): Flush Latrine Field/pond Other/none

Older cohort

Young Lives data†

2006 VHLSS data‡

Young Lives data

2006 VHLSS data

N/A 14.4 6.9 7.6 −0.53 10.0

N/A 20.1 6.8 7.6 −0.70 13.2

96.6 12.8 6.9 7.8 −1.42 27.8

94.0 18.3 7.0 7.8 −1.43 27.2

61.7 81.0 19.9 20.6 94.4

53.9 80.3 19.5 23.7 93.9

64.6 86.4 20.5 20.6 95.1

53.5 84.8 19.2 21.5 94.7

30.1 13.8 36.2 11.6 8.3

22.3 22.4 29.5 11.1 14.7

30.7 14.0 32.4 12.9 9.9

20.6 17.8 34.6 12.5 14.5

32.6 24.0 35.5 7.9

31.1 16.3 24.2 28.4

34.7 26.1 31.9 7.4

28.6 19.7 23.8 27.9

†Data refer to round 2, when the younger cohort children were about 5 years old and the older cohort were about 12 years old, except height-for-age, which is from round 1. ‡Vietnam Household Living Standards Survey. These data are for 5-year olds (younger cohort) or 12-year olds (older cohort) children. The height-for-age data are from the Vietnam Health Survey (VNHS), which was conducted in 2002.



515

Table A2 Communes in Young Lives Survey for Vietnam

Province

District

Commune (pseudonym)

Code

Phú Yên

Tuy Hòa Tuy An So’n Hòa Sông Ca`ˆ u So’n Hòa Sông Ca`ˆ u Bình ¯Da·i Bình ¯Da·i Châu Thành Châu Thành Bình ¯Da·i Bình ¯Da·i Châu Thành Châu Thành Bát Sát Ba´˘ c Hà Ba? o Tha´˘ ng Ba? o Tha´˘ ng Bát Sát Ba´˘ c Hà Ba? o Tha´˘ ng

Chu Se Tam Ky Van Lan 1 Tuy Duc 1 Van Lan 2 Tuy Duc 2 Ha Tinh 1 Ha Tinh 2 Ly Hoa 1 Duc Lap 1 Ben Hai 1 Ben Hai 2 Ly Hoa 2 Duc Lap 2 Lang Ho 1 Krong Buk 1 Play Kep Gian Son 1 Lang Ho 2 Krong Buk 2 Gian Son 2

VN001 VN002 VN003 VN004 VN005 VN006 VN007 VN008 VN009 VN010 VN011 VN012 VN013 VN014 VN015 VN016 VN017 VN018 VN019 VN020 VN021

Va˘n Giang Va˘n Giang Phù Cu`’ Phù Cu`’ Phù Cu`’ Phù Cu`’ Thanh Khê Ha? i Châu Thanh Khê Thanh Khê Thanh Khê Liên Chiểu Liên Chiểu

Na Hang Ha Giang Phu Thong 1 Cao Ky 1 Cao Ky 2 Phu Thong 2 Dai Tu Pho Lu Van Ban 1 Van Ban 2 Van Ban 3 Hania Lo 1 Hania Lo 2

VN022 VN023 VN024 VN025 VN026 VN027 VN028 VN029 VN032 VN033 VN034 VN035 VN036

Be´ˆ n Tre

Lào Cai

Hu’ng Yên

¯Dà Na˜˘ ng

Number of primary schools

Number of satellite schools

Poverty rate (%)

2 (1 and 2) 2 (1 and 2) 1 1 1 1 1 1 2 (A and B) 1 1 1 1 2 (A and B) 1 2 (A and B) 4(1,2,3 and 4) 2 (1 and 2) 1 1 3(1,2 and 3) 1

0 0 2 1 7 1 0 0 0 0 0 0 0 0 6 7 0 0 5 2 0 0

24.3 23.7 27.0 8.8 27.0 8.8 13.8 5.9 14.2 6.9 13.8 5.9 14.2 6.9 79.5 29.0 22.0 27.8 79.5 29.0 27.8 7.7

1 1 1 2 (A and B) 1 1 1 1 3 1 3(2 (AandB)) 2

0 0 0 0 0 0 0 0 0 0 0 0

7.7 26.6 24.7 24.7 26.6 2.0 6.3 3.5 3.5 3.5 6.3 6.3

Note: Commune VN030 was split into two communes between 2002 and 2006; the new communes are called VN035 and VN036. Similarly, VN031 was split into three communes: VN032, VN033 and VN034.

References Baulch B, Chuyen TTK, Haughton D, Haughton J (2004) Ethnic Minority Development in Vietnam: A Socioeconomic Perspective. In: Glewwe P, Agrawal N, Dollar D (eds) Economic Growth, Poverty, and Household Welfare in Vietnam, pp. 273– 310. World Bank, Washington, DC. Baulch B, Nguyen HTM, Phuong PTT, Pham HT. (2010) Ethnic Minority Poverty in Vietnam. Chronic Poverty Research Centre Working Paper no. 169, Manchester, UK. Blinder A (1973) Wage Discrimination: Reduced Form and Structural Estimates.

Journal of Human Resources 8.4, 436– 55. Cueto S, Leon J, Guerrero G, Muñoz I (2009) Psychometric Characteristics of Cognitive Development and Achievement Instruments in Round 2 of Young Lives, Technical Note 15, Oxford: Young Lives. Dang H-A (2003) An Examination of the Factors Determining the School Progress of Ethnic Majority and Ethnic Minority Children in Vietnam. MS dissertation, Department of Applied Economics, University of Minnesota, Minneapolis/St. Paul. Dang H-A (2007) The Determinants and Impacts of Private Tutoring Classes in


516


Vietnam. PhD thesis, Department of Applied Economics, University of Minnesota, Minneapolis/St. Paul. General Statistics Office (2007) Results of the Viet Nam Household Living Standards Survey 2006. Statistical Publishing House, Hanoi. Glewwe P (2004) An Investigation of the Determinants of School Progress and Academic Achievement in Vietnam. In: Glewwe P, Agrawal N, Dollar D (eds) Economic Growth, Poverty and Household Welfare in Vietnam, pp. 467–501. World Bank, Washington, DC. Glewwe P, Kremer M (2006) Schools, Teachers, and Education Outcomes in Developing Countries. In: Hanushek E, Welch F (eds) Handbook of the Economics of Education, pp. 945–1017. North Holland, Amsterdam. Glewwe P, Lambert S (2010) Education Production Functions: Evidence from Developing Countries. In: Brewer D, McEwan P (eds) Economics of Education, pp. 137–47. Academic Press, Oxford. Himaz R (2009) ‘What Drives Cognitive Achievement? An Analysis Based on the Young Lives Older Cohort Data’. unpublished paper. Nguyen HTM, Kompas T, Breusch T, Ward MB (2014) Language, Mixed Communes and Infrastructure: Sources of Inequality and Ethnic Minorities in Vietnam. Crawford School of Public Policy. The Australian National University, Crawford School Research Paper no. 12-07. Nordic Consulting Group (2008) ‘Public Expenditure Tracking Survey for Primary Education in Vietnam: Final Report’. Joint Study by Ministry of Finance, Ministry of Education and Training, Mekong Economics and Nordic Consulting Group, Oslo: Nordic Consulting Group. Oaxaca R (1973) Male-Female Wage Differentials in Urban Labor Markets. International Economic Review 14.2, 693– 709.

September 2015

Pitt M, Rosenzweig M, Gibbons DM (1993) The Determinants and Consequences of the Placement of Government Programs in Indonesia. World Bank Economic Review 7(3), 319–48. Tran T, Pham TL, Harpham T, et al. (2003) Young Lives Preliminary Country Report: Vietnam. Vietnam, viewed January 2009 . Truong HC (2009) ‘Aspects of Education Policies for Ethnic Minorities in Vietnam: Experience of “Young Lives” Children’. unpublished paper. Vietnam Academy of Social Sciences (2009) Participatory Poverty Assessment 2008: Synthesis Report. The Gioi Publisher, Hanoi. World Bank (1997) Vietnam: Education Financing. Country Social Analysis: Ethnicity and Development in Vietnam. World Bank, Washington DC. World Bank (2001) World Development Report 2000/2001: Attacking Poverty. World Bank , Washington, DC. World Bank (2003) Vietnam Development Report 2004: Poverty. Joint Donor Report to the Vietnam Consultative Group Meeting, Hanoi, 2–3 December 2003. World Bank (2004) Vietnam: Reading and Mathematics Assessment Study. East Asia and the Pacific Region. World Bank, Washington, DC. World Bank (2008) World Development Indicators, viewed March 2009 . World Bank (2009) Country Social Analysis: Ethnicity and Development in Vietnam. World Bank, Washington, DC. World Bank (2011) Vietnam: High Quality Education for All by 2020. East Asia and the Pacific Region (with Belgian Development Cooperation and UK Department for International Development. World Bank, Washington, DC.
