Getting to Grade 10 in Vietnam: Does an Employment Boom Discourage Schooling?

Blue-collar employment growth increases schooling opportunities by raising incomes, but also reduces incentives for some students to advance beyond compulsory education. These contradictory influences may help to explain relatively slow and uneven growth of progression to upper-secondary schooling in Vietnam, which has experienced a foreign investment boom in mainly low-skill manufacturing industries. We use data on participation rates and scores in an upper-secondary school entrance exam to analyze variation due to demand-side and supply-side factors. The data come from less advanced provinces and so illuminate the challenges of deepening educational development at the extensive margin, especially among ethnic minority populations.

Keywords: Skill premium; economic development; test scores; school–work transition; Vietnam

1. Introduction

Vietnam has made tremendous progress in increasing school enrolment and achievement. In about one generation, it has achieved near-universal rates of primary and lower secondary school enrolment, up from about 86% and 72% respectively in the early 1990s; upper secondary enrolment has increased from 27% to more than 75% during the same interval (Rolleston et al. [45]; Dang and Glewwe [14]). Average years of schooling in the adult population are higher than in countries such as the Republic of Korea and Thailand at comparable income levels (Figure 1). Despite these aggregate gains, however, significant educational achievement and access gaps remain, especially between wealthy and poor populations – a distinction that is increasingly strongly correlated with ethnic majority and ethnic minority populations. Over these categories, we see that educational flows (i.e. the production of new graduates at each grade level) are quite unequal. Disparities widen rapidly after Grade 9 (ages 14–15). From age 15 to 18, the overall net enrolment rate declines from over 90% to a little more than 50%, but the fall is much faster for children from poor and ethnic minority households. Moreover, the available data suggest that schooling achievement gaps are closing very slowly, if at all.

PHOTO (COLOR): Figure 1. Average Years of Schooling in Vietnam, the Republic of Korea, and Thailand. Note: Data points at 5-year intervals from 1960–2010 (Vietnam from 1985–2010). Source: Authors' calculations based on data from data.worldbank.org and barrolee.com, accessed 1 January 2020.

Vietnam, like many low- and middle-income countries, offers few opportunities for adult education. For most teenagers, their last year of schooling is the highest credential they will earn and exerts a strong influence over their lifetime earnings. This provides the motivation for our study. The decision to terminate schooling has implications both for individual earning power and for human capital accumulation in the economy as a whole.[1]

Demand for schooling is subject to income effects and relative price effects. Average household incomes have risen, and this typically corresponds to increased demand for education as household financial constraints diminish. However, GDP growth has been led by low-skill, labor-intensive industries whose expansion reduces the skill premium and raises the opportunity cost of schooling for children approaching working age. The income and relative price effects are likely to have contradictory influences on schooling decisions. Moreover, these two effects may apply with different forces for different groups in the population, leading to heterogeneous patterns in educational decision-making.

Our goal is to understand the reasons for the variation in educational attainment across the population and to consider welfare and policy implications. Our central question is: how much variation in school achievement and progress is due to income growth, how much to predetermined household or community factors such as ethnicity, and how much to 'pull' from the low-skill labor market? In addition, is there selection into test-taking based on ability? Our purpose in asking these questions is to understand and inform educational policy, especially the goal of educational deepening.

To answer this question, we use a unique data set of individual test-takers and scores from the Vietnam 10th grade entrance examination (henceforth, G10 exam), a national exam to determine access to upper secondary school (Grades 10–12). We merge these with data from other sources to obtain district-level controls for economic and social conditions and labor market activity. The data come from a subset of provinces that are poorer and more remote than the national average. As such – and also because the variable cost of schooling rises with grade – they present an opportunity to study challenges to educational development at the extensive margin of economic development.

With district-level test-taking rates as the dependent variable, we find, as expected, that higher household income or socio-economic status (SES) is associated with higher participation rates. However, the labor market plays a significant and negative role. Our measure of labor market activity, the district-level share of local employment in FDI firms, has a significantly negative effect on test-taking rates. As long as Vietnam continues to attract investments that take advantage of abundant blue-collar labor, the tension between income effects and relative price effects will persist among poorer families. Ethnic minority households are especially susceptible since they have much lower initial wealth, and their economic gains from the Vietnamese economic miracle have been relatively small (Kozel [28]; Baulch et al. [8]). Finally, our model and results highlight interactions among schooling policies and other development policy instruments. Initiatives to increase schooling access by lowering costs, or less directly by addressing sources of household poverty, are likely to fall short of their goals if the labor market signals that low-skill employment is an attractive option, or if the quality of schooling offered is low.

2. Globalization, growth, and the labor market

Vietnam's experience with globalization has been a defining feature of its recent economic history. Exports of goods and services grew from 7% of GDP in 1986 to more than 100% in 2018; their real value expanded at an average annual rate of more than 15%, three times faster than the low middle-income country average. Annual foreign direct investment (FDI) inflows during that period averaged 6% of GDP – more than double the average for lower middle-income countries (World Bank [52]). FDI inflows have expanded by more than 9% annually since 2000.

It follows that globalization – that is, increasing integration with world markets for products, services, and factors – has been a major driver of change in the domestic labor market. Joining the global market alters relative prices and raises returns on investments in export-oriented industries. New capital (and other productive attributes bundled with it) raises labor productivity. Those changes in turn help to determine the industries and occupations in which job growth is fastest. Aggregate income increases, but the distribution of gains within the domestic economy is conditional on many factors, notably household assets (including labor and skills) and labor mobility in response to changing sectoral structure.

Trade theory provides a useful lens through which to view these trends. Vietnam has specialized in industries making intensive use of relatively low-skilled labor, often with complementary inputs of foreign investment. Its import-substituting industries are far more capital- and technology-intensive than its exporters. This pattern of specialization has fueled a generation-long boom in non-farm job creation and GDP growth.

One implication of this pattern of investment and trade is that demand for blue-collar labor has risen rapidly relative to demand for skills.[2] The link from globalization to this structural change can be appreciated in stylized form with the help of a Lerner diagram (Deardorff [15]). Assume that the economy is endowed with skills (H) and blue-collar labor (L), and uses these as inputs in the production of two goods of differing skill intensity. In Figure 2 below, production technologies for the two goods are represented by unit value isoquants Qi = 1/pi, each showing combinations of H and L compatible with sectoral production of output worth $1. In equilibrium, since factors are mobile between sectors, there is a single relative price of skills at which factor markets clear; this is shown as a unit isocost line with slope w = wH/wL (its intercepts are 1/wH and 1/wL). When factor markets clear, then the point on each unit isoquant with slope equal to w is the cost-minimizing input ratio hi = Hi/Li in that industry. At the tangency, employment in each industry is found by measuring the corresponding distance along each axis. Output in each industry is measured by the length of a ray from the origin through the tangency point.

PHOTO (COLOR): Figure 2. Effect of a Rise in P1/P2 on the Skill Premium.

Without loss of generality, choose units of the skill-intensive good such that its price (p2) is equal to one. The autarky relative price of the less skill-intensive good (p1) is lower than its world price; therefore globalization increases this price. As a result, Q1 moves closer to the origin. The original factor price ratio w is now no longer compatible with market clearing. To re-establish factor market equilibrium, the relative price of L, the factor used intensively in industry 1, must rise. The new factor price w' reflects this shift. In other words, a rise in the relative price of the labor-intensive good has caused the skill premium to decline, since w' < w.

Under full employment, both output and the quantity of each factor employed in 1 increase, while in 2 they must decline. In Figure 3, point E shows the economy's total endowments of H and L. The allocation of each factor to each sector is found by vector addition using the cost-minimizing factor proportions measured by hij, where i = 1,2 denotes sector, j = 0 denotes the initial equilibrium, and j = 1 the equilibrium after the price change. A relative product price change that raises demand for low-skilled labor increases the skill-intensity of production in both sectors (another stylized fact in globalizing economies), but if the total endowment shown at E remains unchanged, then sector 1's share in employment of both factors must also increase in proportion to its output. Conversely, employment shares in sector 2 must fall. As drawn, for example, the share of L employed in sector 1 increases as a result of the price change, from 0L10/0LT to 0L11/0LT. The reader can verify that an analogous change occurs in the allocation of H.

PHOTO (COLOR): Figure 3. Output and Employment Effects of the Price Rise.

Of course, real-world adjustment may involve more conditions – notably, differential elasticities of factor supply, non-traded goods, adjustment times and costs, and transitions between informal and formal employment (Winters, McCulloch, and McKay [50]). In addition, we know that along with the structural change depicted here, globalization also increases aggregate income due to gains from specialization and trade. Together, these results bring the schooling decision problem into focus. On one hand, income growth induces greater schooling by helping to relax credit constraints and other economic factors limiting educational spending by households. On the other hand, the lower skill premium and faster job growth in blue-collar occupations raises the opportunity cost of schooling. Predictably, these two forces will be felt with different intensity by different groups within the economy, depending on the distribution of gains from trade and on how each household calculates the expected net benefits of additional schooling.

The dynamics of the globalization–human capital nexus were first explored by Findlay and Kierzkowski ([19]) in a model that integrated Heckscher-Ohlin specialization with models of human capital accumulation due to Mincer ([33]), Schultz ([46]), and Becker ([9]). Their model used exogenous changes in relative product prices to drive educational decisions by individuals, taking account not only of skill-specific expected lifetime earnings but also of the direct and implicit (opportunity) costs of schooling. An insight from the Findlay-Kierzkowski model is that the process may be self-reinforcing. Over time, a rise in the relative price of the less skill-intensive good (which contributes to a lower skill premium) induces a change in the composition of the labor force, and this in turn further increases the relative size of the less skill-intensive industry through Rybczinksi effects (Findlay and Kierzkowski [19], 968–969).[3] This can be seen in Figure 3. Faster growth of L relative to H moves the endowment point E further to the right than upward. A lower relative price of skills reduces investments in skill acquisition, other things equal.

One important qualification to this heuristic account is that the economy also produces and consumes non-tradable services. Demand for services is typically income-elastic, and so grows at least as rapidly as the aggregate economy. Therefore, income growth will increase services' prices. Whether this second-round effect raises the skill premium or lowers it will depend on the relative skill intensity of services production. In lower-income countries, services are dominated by activities such as construction, sales, food and hospitality, and personal services, and these are largely supplied by small, often family-run firms with low levels of capitalization. In this case, a rise in services demand will further raise relative demand for blue-collar labor.[4]

Finally, it may be realistic to suppose that the supply of low-skilled labor is elastic due to underemployment in a 'backstop' sector such as agriculture. World Bank data show employment in Vietnamese industry expanding from 12% of the labor force in 1991 to 26% in 2019, and that of services rising from 19% to 35% over the same interval. Agriculture's share in total employment fell from 69% to 39%, and its share in GDP decreased from 40% to 15%, suggesting a substantial reallocation of low-skilled workers. Suppose for simplicity that the production of services uses only labor (or a combination of labor plus the innate entrepreneurial flair of owner-operators). Then, an elastic supply of low-skilled labor means that the output of services can increase in response to growing demand without significant price rises. In this case, the total net increase in low-skilled labor demand is the sum of that in the expanding tradable sector 1, minus that in the contracting sector 2, plus that in the expanding non-traded services sectors.

3. Related studies

Vietnam's experience with globalization may be typical of that among lower middle-income countries. Rapid employment growth concentrated in low-skilled occupations poses a particular challenge when adolescents must choose between continuing with school or entering the wage labor market. These choices have lifetime implications for the current generation – lower skilled workers earn much less over their working careers – but also for overall economic growth and for the distribution of gains from growth in future generations.

It is by now well known that variation in cognitive and non-cognitive abilities in adolescence or even later in life originates very early, even in utero (Almond and Currie [3]). Maternal, household, and environmental conditions all play a potentially long-lasting role in young children's intellectual development and this in turn affects a child's performance upon entry to the formal education system (Almond and Currie [4]). Thereafter, there is potential for a persistent widening of educational achievement gaps as more advanced and better-motivated children, and those whose home conditions are more conducive to learning, both learn better and capture more educational resources, such as the attention and encouragement of teachers. Thus, variation in socio-economic status, an indicator for a range of health and nutritional variables that contribute to early life and childhood development, is an important predictor of later-life outcomes.

In adolescence and beyond, standardized tests provide a transparently comparable metric of educational attainment. Test scores are complementary with other frequently used but indirect measures such as total schooling years and schooling for age (Ray and Lancaster [43]), with lower measurement errors. Additionally, test scores provide a direct measure of knowledge acquisition or cognitive capacity, something that can only be very broadly inferred from data on schooling duration or grade progression since those also depend on factors such as school availability and quality (Hanushek and Woessmann [25]). High-school test scores are robust predictors of labor market outcomes (Neal and Johnson [35]). Finally, in the Vietnam case selection into test-taking is an important indicator in its own right, since it signals intent to continue with education beyond 9th grade.

3.1. Relevant trends in Vietnam

In Vietnam, growth has been concentrated in export-oriented processing industries such as garments, footwear, and electronics. Employment and skill premium trends are broadly consistent with the Stolper – Samuelson conclusion of the previous section – that growth has raised relative demand for less-skilled workers and lowered the skill premium, other things being equal. In this section, we summarize evidence from studies of these trends.

An important early study found evidence of a strong income effect from domestic market liberalization, leading to the withdrawal of children from the family/farm labor force and a preference for increased schooling, even among low-income households (Edmonds and Pavcnik [18]). However, this study uses data from the earliest years of Vietnam's reform era. Large-scale labor demand growth by foreign-invested firms and domestic non-farm private sector firms began in earnest only after about the year 2000.

The Vietnam–US Bilateral Trade Agreement (USBTA) of 2001 brought about a sudden and substantial lowering of US tariffs on Vietnam's exports. Using a measure of provincial exposure to export-increasing trade policy changes, Fukase ([21]) found that the effect of the USBTA on relative wages in export-exposed provinces significantly mitigated a generalized national rise in the skill premium due to ongoing domestic economic reforms. McCaig and Pavcnik ([30]) found that response to the treaty included large labor market effects, notably an increase in formal employment by younger workers in locations where trade-related industries were expanding rapidly.

Internal migration responses to globalization in Vietnam have been substantial (Coxhead, Nguyen, and Vu [12]). Rapid expansion of low-skilled jobs appears to have raised the opportunity cost of schooling: in particular, globalization-related job creation, proxied by the local intensity of jobs in foreign-invested firms, is seen to have had a significantly negative effect on high school attendance (Coxhead and Shrestha [13]).[5] In short, the aggregate evidence points to a clear tension between the positive schooling effects of income growth and the negative effects of a booming low-skilled job market.[6]

Structural change in labor demand translates into differential changes in demand for skills, and changes in the skill premium reflect these (Katz and Murphy [27]). During Vietnam's reform era, returns to education rose rapidly with the relaxation of the command economy wage grid. However, recent studies have shown that returns to schooling peaked around the mid-2000s (Doan and Gibson [16]; Phan and Coxhead [39]; Doan, Tuyen, and Quan [17]; McGuinness et al. [31]). According to Doan, Tuyen, and Quan ([17]), the return to an additional year of schooling in 2014 was only 5.4%, three percentage points lower that its peak (8.5%) in 2008, and much lower than world averages reported in the survey by Psacharopoulos and Patrinos ([41]). The returns to upper secondary schooling in particular are very low, with wages of Grade 12 graduates barely above those to Grade 9 graduates.[7] The differences are especially small for cohorts that are close in age to the current generation of school-leavers.

Some studies have attributed declining returns to schooling to the rising supply of high school and college graduates (e.g. Doan, Tuyen, and Quan [17]). This trend is no doubt influential. However, comparable countries have experienced rapid growth in educational attainment without declining returns. China, whose experience Vietnam's closely resembles, is one (Heckman and Li [26]). From 1995 to 2014, college admissions in China increased more than sevenfold (Li et al. [29]) yet returns to college education increased (Gao and Smyth [22]). The most likely explanation is capital–skill complementarity – the growth of skilled labor raises returns on capital investments, and the resulting increase in capital stocks raises skilled labor productivity, thus increasing returns to education. This complementarity favors the growth of industries that are intensive in capital and skills, a progression that is evident in China's recent history but much less visible thus far in Vietnam. Tertiary enrolments in Vietnam increased from 1998 to 2013 by a factor of 2.5 (Doan, Tuyen, and Quan [17]): a substantial gain, but far smaller than in China. So, if returns to schooling have fallen in Vietnam despite a boom in capital investments, then either the type of new investment was not conducive to capital–skill complementarities and/or its impacts were diminished by other factors.

If skill premia are small, then it also follows that the economics of decision-making over the transition from school to work will be dominated by two near-term considerations: the financial cost of schooling, and the opportunity cost of time spent in school and thus not in the labor force. These items will be especially influential for households confronting credit constraints or other circumstances that cause them to discount the future more heavily.

3.2. Schooling trends and policy targets

The Government of Vietnam's education strategy for 2011–2020 includes, among others, the following targets: 'By 2020, the rate of primary school students and lower secondary school students of eligible age will reach 99% and 95% respectively; 80% of young people will reach education of upper secondary school level or equivalent at eligible age'. The first two goals have been reached. The third − high school enrolments − remains unfulfilled (Asian Development Bank [6]).

Conflicting trends and population heterogeneity present the designers of education policy with a particular challenge, since it is no longer the case that 'one size fits all' in measures to promote schooling. This heterogeneity is especially important in Vietnam, where the pace of economic growth is rapid but is not uniform across regions or sub-populations. For example, Figure 4 shows the proportion of children in school, by level of schooling, for the top and bottom quintiles of households by expenditure over a decade-long interval, 2006−2016. In the top quintile, school persistence and grade progression were already high in 2006 and improved even more, especially at upper secondary and tertiary levels. Among households in the bottom expenditure quintile, the change was far more modest. This differential response may be due to (i) greater income effects among wealthier households; (ii) labor demand growth (and especially, growth in demand for blue-collar labor) causing a higher increase in the opportunity cost of schooling; (iii) institutional differences in the rates of return to secondary and tertiary schooling (Phan and Coxhead [39]); or (iv) differences in ethnic composition, since ethnic minority groups are heavily over-represented in the lowest income quintile (Kozel [28]). Figure 5 confirms that, by comparison with the kinh majority, educational progress among Vietnam's ethnic minority children has been solid up to about age 14, but sparse thereafter.

PHOTO (COLOR): Figure 4. Enrolment Rates by Age and School Type (1st and 5th Quintiles), 2006 and 2016. Note: Frac = fraction. Source: Authors' calculations based on data from VHLSS 2006 and 2016.

PHOTO (COLOR): Figure 5. Enrolment Rates by Age, School Type, and Ethnicity, 2006 and 2016. Note: Frac = fraction. Source Authors' calculations based on data from VHLSS 2006 and 2016.

Improving the supply of educational resources – schools, teachers, books and equipment, and services such as curriculum design – is fundamental and is the primary policy domain of public sector agencies with educational mandates. However, schooling demand beyond the age of compulsory education is also subject to influence from labor markets, credit markets, and more. Employment, returns to skills, and incomes may all rise in the course of economic growth, but each of these changes will have a distinct (and not necessarily positive) influence on incentives to remain in school (Becker [9]). Moreover, each of these variables may have different effects on schooling when individuals are heterogeneous along some exogenous or differentially constrained dimension such as ethnicity, credit access, or labor mobility.

4. A model of schooling decisions

Our empirical focus is on the watershed choice by children at the threshold of working age to continue or terminate schooling. Our marker is the decision to take the 10th grade entrance exam, and in the foregoing survey we have identified household, labor market, and school system-based reasons for variation in the test-taking rate. The model developed below identifies, in stylized form, key elements of the test-taking decision. A feature of the model is that it shows how schooling costs, poverty (as indicated by credit constraints), and the skill premium jointly condition the likelihood of test-taking conditional on a child's inherent cognitive ability. A second feature is that the model reveals some gray areas in which the quality of the schooling experience (discussed further below) may also be influential at the margin.

Suppose there is a mass of the student population such that their inherent cognitive ability

Graph

$a$ is distributed uniformly in

Graph

$\left[0,1\right]$ . This population is composed of two groups. One is the poor, who discount the return to education. Their discount factor is

Graph

$\text{β}\in \left[0,1\right]$ , where a lower value reflects a higher discount rate, or a binding credit constraint on parents.[8] The other group is the rich, who do not discount the return to education. Their discount factor is 1. The share of the poor in the population is

Graph

$x$ and that of the rich is

Graph

$1-x$ .

A student completing lower secondary school faces a choice between joining the labor market or proceeding to high school and later, perhaps, tertiary education. In order to continue in school they must take and pass the G10 exam. This model shows combinations of ability, financial capacity, schooling costs and quality, and labor market conditions that determine this choice.

We denote the level of education that accumulates until lower secondary school as

Graph

$b\left(1+\delta a\right),$ where b ≥ 0 is basic learning and

Graph

$\delta \ge 0$ indicates capacity to augment cognitive ability. We suppose that

Graph

$\delta$ captures some combination of a student's motivation to study and the nature of their interaction with the school system – that is, the quality of schooling offered, as well as attention to individual students by teachers. All students accumulate

Graph

$b$ by attending school, but students who experience a better quality of interaction with the school system will gain more, conditional on their inherent ability. However, additional educational gains up to Grade 9 do not have any significant return in the labor market; the earnings of a student who finishes school in Grade 9 will depend only on their inherent ability

Graph

$a$ .

If a student chooses to keep studying, they will face a cost

Graph

$c$ . On the other hand, the return to upper secondary education (Grades 10-12) is

Graph

$r\left(1+\delta a\right)$ . For students in the rich group, the outcome of proceeding to high school is

Graph

$a+\left(b+r\right)\left(1+\delta a\right)-c$ . We assume that

Graph

$b+r\ge c$ . With this assumption, all children from the rich group will take the entrance exam with the intent of continuing to upper secondary school.

For the poor group, continuing to upper secondary school yields lifetime earnings of

Graph

$\text{β}a+\text{β}\left(b+r\right)\left(1+\delta a\right)-c$ . They will attempt the exam if

Graph

$\text{β}\left(b+r\right)-c\ge a\left(1-\text{β}-\text{β}\delta \left(b+r\right)\right)$ , or after rearranging:

Graph

$y=\left[\text{β}\left(b+r\right)-c\right]+a\text{β}\left(1+\delta \left(b+r\right)\right)\ge a$ (1)

The value of continued schooling increases with the discount factor and returns to education, declines with schooling costs, and increases with the quality of the school experience. To keep notation clean, denote unconditional returns to schooling (the intercept in (1)) by

Graph

$R=\text{β}\left(b+r\right)-c,$ and the marginal additional return conditional on a by

Graph

$G=\text{β}(1+\delta (b+r).$ Figure 6 shows how the schooling decisions of poorer students are governed by combinations of these values. In panel 6a, unconditional returns to schooling are positive, so

Graph

$R>0.$ Case 1 in the figure depicts lifetime earnings when the values of

Graph

$\text{β},\delta$ and (b + r) are such that G = 1. For example, one set of values satisfying this would be a discount rate of 10% (so

Graph

$\text{β}$ = 0.91) and a return to education (b + r) of 10%, with

Graph

$\delta =1.$ Case 2 illustrates an earnings path when G > 1. So long as R > 0 and G ≥ 1, all children take the entrance exam. Case 3 shows, however, that if G < 1, it is possible that not all children take the exam due to some combination of high discount rates, low market returns to education and/or poor quality of schooling. In this case there is negative selection into upper secondary schooling: children of ability greater than

Graph

$\overline{a}$ will judge that the additional time spent in school will not compensate them for earnings forgone in the unskilled job market.

PHOTO (COLOR): Figure 6. (a) High school decisions when unconditional net returns are positive. Note: R=β(b+r)−c is discounted unconditional return to upper secondary schooling. G=β(1+δ(b+r) as described in text. In case 3, there is negative selection into test-taking. Students of ability a>a¯ choose to drop out rather than continue schooling. The low return to additional schooling is insufficient given their relatively high potential earnings in unskilled work. (b) High school decisions when unconditional net returns are negative. Note: R and G as defined above. In case 6, there is positive selection into test-taking: The high return to additional schooling causes students of ability a>a¯¯ choose to continue rather than drop out.

In panel 6b, unconditional returns are negative. Children facing R < 0 will choose to drop out after Grade 9 for any value of G less than 1 (Case 5). However, if the value of

Graph

$G$ is sufficiently high due to some combination of high discount factor, high market returns to schooling and high quality of schooling, then the possibility exists that a positively selected subset of children with ability

Graph

$a\ge \overline{\overline{a}}$ will take the entrance exam, while those with ability

Graph

$a<\overline{\overline{a}}$ drop out (Case 6).

These contrasting cases highlight the threat to continued schooling posed by a labor market in which the skill premium is low. As the value of (b + r) diminishes, it has the same unconditional effect as higher schooling costs or a higher discount rate. When unconditional returns are negative, a lower value of (b + r) also offsets additional gains from school quality, so moving the cutoff for test-taking to the right along the ability scale. The model highlights the 'last mile' nature of the problem of raising the rate of progression to upper secondary school in areas or among populations where credit constraints bind, school costs are high, and/or school quality is low. In such areas, lower labor market returns to schooling may be quite influential at the margin. Our econometric analysis, in the next section, includes a test for selection, that is whether policy or other shocks that influence the test-taking rate might result in higher or lower average test scores. Appendix A provides a brief formal statement of the test for selection.

5. Data and estimation

5.1. Data

This project uses a unique data set with significant complementarities to existing data on schooling and labor market transitions in Vietnam. Students completing Grade 9 who wish to continue to upper secondary (and perhaps beyond) must take the G10 exam. This exam covers two core subjects, mathematics and literature, and may include other subjects at the discretion of provincial education authorities.[9] High schools use these results in their admissions process, setting minimum scores based on a variety of factors including school capacity and quality. We have access to individual G10 exam scores, plus limited other basic information about test-takers, for four provinces and several years in each province. We do not have household data for the test-takers. We merge these records at district level with data from the Vietnam Household Living Standards Survey (VHLSS), censuses, and labor force surveys to provide additional information on local incomes and distribution, population structure, and labor demand.

A test such as the G10 exam provides a reasonably clear indication of the candidate's intention to continue their studies, since there is no reason for taking the test other than to continue in school. As noted, however, an important qualification is that not all children who are eligible to take the G10 exam actually choose to do so.

This data set is unique, but it is not the only vehicle for studying educational investment and attainment in Vietnam. A recent study using longitudinal survey of rural households (Vietnam Access to Resources Household Survey, VARHS) to evaluate schooling decisions in the context of labor market trends found no relationship between local labor markets and schooling decisions (Mergoupis, Phan, and Sessions [32]). However, VARHS is designed to study rural asset markets, and its indicator of labor market activity is wage offers in agricultural tasks such as harvesting. Agriculture, while still a large employer, is no longer the primary source of jobs for school-leavers, and wage offers for seasonal work such as harvesting do not convey a complete labor market signal.

Vietnam is also one of four countries in the global Young Lives survey, a longitudinal study of child poverty with a strong focus on education. Young Lives surveys collect detailed data on a panel of children who have been followed since 2002, enabling in-depth studies linking individual and family conditions to educational outcomes (Rolleston et al. [45]). Rolleston and Iyer ([44]), for example, used Young Lives data to confirm 'meritocratic progression' by finding that good test scores in Grade 5 predict progression to Grade 10. Nguyen ([37]) examined the outcomes for children 'left behind' by parents who migrate for work. Ethnicity accounts for a large share of the income and education gap in Vietnam, and for this reason is much studied (Baulch et al. [8]; Truong [48]; Glewwe, Chen, and Katare [24]; Arouri, Ben-Youssef, and Nguyen [5]; Nguyen [38]); the most detailed studies in this literature also rely on Young Lives data. One consistent conclusion is that while individual and household conditions matter for all children's educational decisions and outcomes, external factors – including parental education and economic conditions, peer and school effects, and language barriers – are of greater importance to children from minority backgrounds (Glewwe, Chen, and Katare [24]; Nguyen [38]). This conclusion, if robust in nationally representative data, suggests that measures aimed at improving education for ethnic minority children should focus more on their circumstances and less on providing specialized schools or curricula, as at present.

The data set we use is in some respects complementary to Young Lives. Extant Young Lives studies take advantage of very detailed information on individuals and households, but do not explore 'macro' labor market influences on school decisions. By merging with macro datasets we are able to exploit variation in local labor markets and economic conditions in addition to community-level characteristics, to identify demand-side effects.

5.2. Estimation

We want to understand why some children choose to take the G10 test while others do not. Conditional on taking the test, we also seek to explain variation in the scores obtained.

The G10 test data cannot be matched with individual records from other socio-economic surveys, but we can identify the lower-secondary school attended as well as the upper secondary school where they take the exam – that is, their intended Grade 10 school. This is sufficient for multi-level analysis after merging with data from the VHLSS and labor force surveys at district level.

We have individual test score records for several years in four largely rural provinces (Dong Thap, Ninh Thuan, Thanh Hoa, and Lao Cai). Our analytical methods are constrained by the paucity of complementary information on individual test-takers. Therefore, in this paper, we use individual observations aggregated to district level in each province.

For the test-taking rate, we regress the measured rate on the variable of primary interest, a district-level measure of labor market activity, as well as controls for district-level socio-economic status along with fixed effects for province and birth cohort and a dummy variable for the district in which the provincial capital city is located. Defining the test-taking rate in district i of province j by birth cohort t as

Graph

$T_{ijt}$ , our estimating model is:

Graph

$T_{ijt}=\alpha +{\theta }_1{L}_{ijt}+{\theta }_2{S}_{ijt}+\mu P_{ij}+{\gamma }_t+{\phi }_j+{ϵ}_{ijt},$ (2)

in which

Graph

$L_{ijt}$ is the labor market variable,

Graph

$S_{ijt}$ captures SES,

Graph

$P_{ij}$ is a dummy for the provincial capital district, and

Graph

${\gamma }_t$ and

Graph

${\phi }_i$ are fixed effects for province and cohort respectively. The variable of primary interest is

Graph

${\theta }_1$ , the effect of the labor market on the test-taking rate.

We construct the dependent variable as follows. We have test score records only for children who have taken the G10 exam, and we want to know what fraction of the eligible population they make up. We use the 2014 Intercensal Survey of Population and Housing to compute the size of each cohort of test-taking age in our data set. We then use a count of test participants as the numerator, and the census count as the denominator to calculate the G10 participation rate, or the fraction of children taking the test in a given year.[10]Table 1 summarizes the calculated district-level participation rates.[11] Provincial averages range from 42% to 55%, and girls take the test at a much higher rate than boys in two provinces while in a third, Lao Cai, their rate is much lower. Within provinces, district-level averages display a very wide range in Lao Cai, and smaller yet still notable ranges in the other provinces. Lao Cai is remarkable in that the range of district average test-taking rates is 21%–74% for boys, and 11%–73% for girls.

Table 1. District average test-taking rates in sample provinces.

Province (No. districts)/cohort	Sample	Mean	S.D.	Min	Max
Lao Cai (9)	All	0.358	0.195	0.124	0.681
Birth cohorts: 1999, 2002	Male	0.359	0.189	0.123	0.725
	Female	0.371	0.233	0.104	0.744
Thanh Hoa (26)	All	0.779	0.149	0.403	1.325
Birth cohorts: 2002–2003	Male	0.711	0.198	0.416	1.348
	Female	0.866	0.227	0.268	1.396
Ninh Thuan (7)	All	0.598	0.175	0.382	1.013
Birth cohorts: 2000–2002	Male	0.521	0.239	0.219	1.235
	Female	0.704	0.167	0.450	1.031
Dong Thap (12)	All	0.569	0.086	0.380	0.749
Birth cohorts: 2000–2002	Male	0.530	0.110	0.306	0.885
	Female	0.621	0.103	0.415	0.813

1 Note: Figures show district average test-taking rates. Birth cohort refers to the birth year of the eligible population.

• 2 Source: Authors' calculations from original data provided by Vietnam Ministry of Education and Training.
• 3 S.D.: standard deviation.

It is important to note that our data are drawn from relatively poor provinces and as such are not nationally representative (Figure 7). Rather, they embody characteristics analogous to the last mile problem of delivery costs. The cost of delivering educational services is convex both in distance and in level of schooling. At the extensive margin of development, people live in more remote locations, are more widely dispersed, and are typically less wealthy, and more dependent on agriculture. These populations may also be less well connected in terms of services, information flows (internet and telecommunications backbone), and exposure to migrant networks. Looking over schooling levels, the per-student cost of supplying upper-secondary education in the developing world is typically several times greater than for primary education (UNESCO [49]). Moreover, the task of keeping children in school beyond legal working age is more challenging than that at lower grades due to opportunity cost. From the student's point of view, the direct and implicit costs of getting an education of a given quality are higher than for populations in regional centres and major cities.

PHOTO (COLOR): Figure 7. Provincial Average Per Capita Income as Percentage of National Average, 2014. Note: Not all province names included due to space constraints. Source: General Statistics Office of Vietnam ([23]).

5.2.1 Socio-economic status

Because the number of district-year observations is small, we can use only a few variables to capture local economic conditions. At district level, we seek to control for household SES. The maintained assumption in this multi-level model is that variation within a location is less than that between locations, so a district-wide measure of economic well-being is a meaningful statistic for individual test-takers. We use district-level income and poverty data merged from the VHLSS,[12] in addition to location and cohort fixed effects as shown above. As an alternative to SES we also include a measure of the share of district population counted as belonging to ethnic minority populations. The correlation between this and other SES variables is very high, with absolute values about 0.9. This correlation captures a widening inter-ethnic disparity in household income growth rates during Vietnam's boom years. In 2016, ethnic minority groups (i.e. those other than Kinh and Hoa (ethnic Chinese)) made up about 15% of Vietnam's population but accounted for 75% of poor households (World Bank [51]).

5.2.2 Labor market variables

Our measure of labor market activity is the share of foreign-invested firms in district-level employment. In Vietnam, employment by foreign-invested firms is a small but rapidly rising share of total employment. In general, however, foreign investment leads indirectly to the creation of many more jobs than in the industry or sector to which the investment is directed. In the United States, Moretti ([34]) found large effects of greenfield investment in tradable sectors on jobs in non-tradable sectors, with 1.6–2.9 jobs created in the latter for each one created in the former. In China, Fu and Balasubramanyam ([20]) found that both export orientation and FDI induce increases in total employment. In Vietnam, the Labor Force Survey shows that more than 90% of young school-leavers aged 15 and 16 find employment in household businesses. However, these businesses, most of which produce a variety of services (e.g. construction, trade and transport, retail, and personal services such as hairdressing and household help) flourish mainly in areas experiencing large injections of new investment.

Broad-based labor demand growth may both increase incomes and raise the opportunity cost of schooling. Coxhead and Shrestha ([13]) pursued a quantity-based approach using FDI employment shares. Another approach to labor demand is through the influence of skill premia. A higher skill premium indicates a higher relative return to investment in education and should in principle be positively associated with test participation. However, this is conditional on other factors, most obviously SES. If highly educated workers are more geographically mobile than the less educated, then the value of a local skill premium may reflect labor immobility – a function of poverty, credit constraints, and other characteristics – as much as it does the returns to education per se. For this reason, the use of the skill premium to signal labor market activity should be accompanied by a rich set of other controls. In their absence, the skill premium is likely to be strongly correlated with other indicators of socio-economic status – and this is the case with our Vietnam data in their present form. Accordingly, as stated above we use the percentage of workers employed in FDI industries as an indicator of labor market activity that is only weakly correlated with socio-economic status.[13] We expect that relatively greater labor demand, other things being equal, should be negatively associated with test participation.

5.2.3 G10 test scores

Individual test scores are a composite of results for components of the G10 exam and sum to 50. Figure 8 summarizes individual test score distributions over some relevant characteristics. There is wide variation within provinces by ethnicity and location as well as (in some) by gender. These distributions also highlight differences between provinces in locational and gender-based disparities (Dong Thap Province has a negligible minority population).

PHOTO (COLOR): Figure 8. G10 Test Score Distributions by Gender, Location, and Ethnicity. Source: Authors' calculations from Grade 10 test score data.

We are interested in estimating the relationship between the test-taking rate and the average score in the 10th grade entrance exam. Our reduced-form econometric model is:

Graph

$y_{ijt}={\rho }_0+{\rho }_1{T}_{ijt}+{\rho }_2{S}_{ijt}+\mu P_{ij}+{\phi }_j+{\gamma }_t+{ϵ}_{ijt}$ (3)

In this equation,

Graph

$y_{ijt}$ is the average test score at district

Graph

$i$ of province

Graph

$j$ in cohort

Graph

$t$ and

Graph

$T_{ijt}$ is the test-taking rate. As performance in the test depends on socioeconomic status, the control variable

Graph

$S_{ijt}$ is one of the three SES variables defined above (per capita expenditures, poverty rate, ethnic minority population share), and

Graph

$P_{ij}$ is the provincial capital dummy as before. We also control for province and cohort fixed effects,

Graph

${\phi }_j$ and

Graph

${\gamma }_t$ , respectively, since the test can differ by province and year. With these fixed effects and other control variables, the source of variation that identifies the parameter of interest,

Graph

${\rho }_1$ , comes from differences in the test-taking rate across districts within each province and cohort.

If the OLS estimate of

Graph

${\rho }_1$ is positive, then if a district experienced improvements in conditions governing the test-taking decision – such as lower school costs, or better-quality school experiences – the additional students who would then take the test would perform marginally better than the current pool of test-takers in the same district. On the other hand, a negative estimate of

Graph

${\rho }_1$ would indicate that a higher test-taking rate is associated with lower average score. These were possible outcomes discussed in the model developed earlier.

However, the OLS estimate of

Graph

${\rho }_1$ could still be biased due to private educational investments, which are unobservable in our dataset. Because a higher test-taking rate is associated with higher living standards and better schools, OLS would overestimate the impact of higher test-taking rates on test scores. OLS may further overestimate impacts if better market access leads to lower test-taking rates, as in Aggarwal ([2]). Alternatively, better market access might induce more educational investment (Adukia, Asher, and Novosad [1]), which could raise the test-taking rate. OLS might underestimate the impact if the change in test-taking rate is proportionally higher than the change in test scores.

To address the endogeneity issue, we propose using the size of the foreign firm presence, measured by the share of the labor force that works for FDI firms, as an instrumental variable. Since the sample consists of districts in mainly rural provinces, the presence of foreign firms is plausibly exogenous since it is unlikely that firms will make location decisions based on the educational attainment of the local labor force. Using Equation (3) as the second stage, the first stage of this model is:

Graph

$T_{ijt}={\alpha }_0+{\alpha }_1{F}_{ijt}+{\alpha }_2{S}_{ijt}+\mu P_{ij}+{\varphi }_j^{{}^{\mathrm{\prime }}}+{\gamma }_t^{{}^{\mathrm{\prime }}}+{ϵ}_{ijt}.$ (4)

Foreign firm presence could increase the test-taking rate by raising household incomes, but it might also lead to a fall in the skill premium and thus reduce the test-taking rate. The presence of foreign firms should not impact test scores directly, but should affect them indirectly through test-taking rates. It is, however, possible that FDI firms are more likely to locate in high market access locations, which would provide a separate association with test scores. To address this possibility, we control for socioeconomic variables at district level, and include the province capital dummy variable.

6. Results

6.1. Test participation rates

We first examine the correlations of test participation and measures of socio-economic status. Table 2 shows summary statistics for variables to be used in the estimation.

Table 2. Summary statistics of variables used in estimation.

Variable	Obs	Mean	S.D.	Min	Max
Ln per capita expenditure	120	9.89	0.33	9.07	10.66
Poverty headcount (%)	120	0.25	0.21	0.01	0.82
Poverty gap	120	0.06	0.07	0.00	0.31
Poverty severity	120	0.02	0.03	0.00	0.14
Ethnic minority population (%)	120	0.30	0.37	0.00	1.00
FDI employment share (%)	120	0.40	0.75	0.00	3.40
Test participation rate	120	0.62	0.21	0.12	1.33
Test score %	120	44.14	9.98	15.63	66.90

• 4 Sources: As described in text.
• 5 FDI: foreign direct investment; Obs: observations; S.D.: standard deviation.

We next turn to the main hypothesis test – that of a separate labor market effect on the propensity to take the G10 test. Table 3 reports the results of the OLS regressions based on Equation (2), with district-year average test participation rates as the dependent variable. The variable of primary interest is the percentage of the district labor force employed by FDI enterprises. The SES proxy variables are also defined at district-year level; most education policies target communities rather than individuals, so the case for the multi-level approach seems defensible. The table uses log of per capita expenditure, poverty headcount, poverty gap, and the squared poverty gap squared (poverty severity). Any one of these district-level measures (they are all very highly correlated), together with province-level fixed effects, explains roughly 60% of the variation in the participation rate.

Table 3. Grade 10 exam participation rate: OLS estimates.

Variable	(1)	(2)	(3)	(4)	(5)
FDI employees (%)	−3.968**	−4.373**	−4.167**	−3.942**	−3.777**
	(1.829)	(1.793)	(1.773)	(1.770)	(1.888)
Log of p.c. exp.	29.186***
	(4.630)
Pov. headcount	1.174	7.080	8.855*	9.791**	10.741**
	(5.037)	(4.587)	(4.504)	(4.496)	(4.753)
Pov. gap		−50.466***
		(7.341)
Pov. severity			−145.335***
			(20.743)
Ethnic min. share				−320.869***
				(46.245)
Provincial capital					−24.418***
					(4.369)
Constant	−242.827***	63.171***	60.561***	58.665***	56.086***
	(44.511)	(5.823)	(5.515)	(5.379)	(5.770)
N	120	120	120	120	120
F	22.346	24.025	24.436	24.224	20.447
r2	0.672	0.688	0.692	0.690	0.652

• 6 Notes: Standard errors in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01.
• 7 Dependent variable is district average fraction of children recorded as taking the exam, by birth cohort.
• 8 Source: Authors' calculations from original data provided by Vietnam Ministry of Education and Training.

The labor market variable is significant throughout, with a consistent value of −0.041 ± 0.003. This is an economically meaningful value, with an average marginal effect similar in magnitude to that of per capita expenditure. At the mean, a one standard deviation increase in the percentage of workers employed in FDI enterprises is associated with a reduction in the district test participation rate of about 4 percentage points. A higher rate of FDI employment is associated with a lower propensity to seek to continue to upper secondary school by children of working age. Employment growth raises the opportunity cost of schooling, and there is clear evidence that this substitution effect is important – even in Vietnam's relatively remote, agricultural, and poor provinces.

The final column of Table 3 uses the ethnic minority fraction in the district population in place of SES controls. The well-known correspondence between ethnic minority status and poverty in Vietnam (World Bank [51]) is immediately on display. Children in districts with higher ethnic minority concentrations are less likely to take the G10 exam. The fact that minority status is so highly correlated with measures of SES or poverty that they cannot both be included in a single regression model speaks eloquently to the differential economic conditions of this subset of the Vietnamese population.

6.2. Test scores

The model developed earlier showed that when unconditional returns to schooling are negative, it is possible (under some parameter value combinations) that the test participation rate may simply reflect selection in which higher-ability children opt to acquire more skills while lower-ability children choose to take blue-collar jobs. If this is the only (or the dominant) explanation for test participation, then policy measures to encourage G10 test-taking and progression to upper secondary school may simply set lower-ability children up for failure by encouraging them to enter an upper-secondary system from which they are unlikely to successfully graduate. On the other hand, the model in section 4 also raised the possibility that children of higher cognitive ability but limited economic means are selected out due to economic constraints. If so, then expanding the test participation rate may make a positive contribution to skills accumulation both at the individual and aggregate levels. With test scores from children who do take the exam, we can make an initial exploration of these issues.

Table 4 reports OLS estimates of the test score model in Equation (3). If SES variables are excluded (as in columns 1–2) then the estimated coefficient of the test-taking rate is positive, indicating that measures to increase the test-taking rate will increase average test scores. However, after we control for SES or ethnic minority population share (Columns 3–5) there is no evidence of selection on ability.

Table 4. Grade 10 entrance examination scores: OLS estimates.

Variable	(1)	(2)	(3)	(4)	(5)
Test-taking rate	0.271***	0.220***	0.018	−0.009	0.027
	(0.048)	(0.046)	(0.036)	(0.034)	(0.031)
Log of p.c. exp.			22.511***
			(1.943)
Pov. headcount				−39.123***
				(2.978)
Ethnic min. share					−21.668***
					(1.541)
Provincial capital		10.233***	−0.011	4.531***	6.592***
		(2.421)	(1.853)	(1.574)	(1.473)
Constant	22.76***	23.41***	−185.32***	51.70***	47.14***
	(3.221)	(3.006)	(18.130)	(2.857)	(2.467)
Province fixed effects	Y	Y	Y	Y	Y
Cohort fixed effects	Y	Y	Y	Y	Y
N	120	120	120	120	120
R2	0.470	0.544	0.796	0.824	0.838

• 9 Notes: Standard errors in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. Dependent variable is district average score on 10th grade entrance examination, by birth cohort. Bonus points unrelated to exam performance are excluded.
• 10 Source: Authors' calculations from original data provided by Vietnam Ministry of Education and Training.

The propensity to take the G10 exam is endogenous, so to account for non-random selection into the sample for which scores are available, we propose the instrumental variables model described in Equations (3) and (4) above. In the first stage, we instrument the exam participation rate using the fraction of FDI workers and SES variables. In the second stage, the instrumented participation rate enters as a control in the test score equation with FDI employment as the excluded variable as discussed previously. Table 5 shows these estimates for three models using different SES variables.

Table 5. Grade 10 entrance examination scores: IV estimates.

	(1)	(2)	(3)
Variable	First stage Test rate (%)	Second stage Score	First stage Test rate (%)	Second stage Score	First stage Test rate (%)	Second stage Score
Test participation		0.001		0.062		0.062
		(0.167)		(0.145)		(0.158)
FDI employment (%)	−3.968**		−4.373**		−3.777**
	(1.829)		(1.793)		(1.888)
Ln p.c. exp.	29.186***	22.967***
	(4.630)	(4.721)
Headcount poverty			−50.466***	−35.886***
			(7.341)	(7.012)
Ethnic min. share					−24.418***	−20.906***
					(4.369)	(3.713)
Prov. capital	1.174	−0.040	7.080	4.289***	10.741**	6.341***
	(5.037)	(1.789)	(4.587)	(1.603)	(4.753)	(1.804)
Constant	−242.83***	−189.08***	63.17***	47.47***	56.09***	45.30***
	(44.511)	(39.704)	(5.823)	(8.805)	(5.770)	(8.566)
Province FE	Y	Y	Y	Y	Y	Y
Cohort FE	Y	Y	Y	Y	Y	Y
Observations	120	120	120	120	120	120
R-squared	0.672	0.795	0.688	0.816	0.652	0.836
IV F-statistic	4.705		5.948		4.004

• 11 Note: Estimation method is LIML. Standard errors in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1. Dependent variable (second stage) is district average score on 10th grade entrance examination, by birth cohort. Bonus points unrelated to exam performance are excluded.
• 12 Source: Authors' calculations from original data provided by Vietnam Ministry of Education and Training.

The data set is not large, so the reported estimates should be treated with caution. F-statistics for the excluded instrument are below the conventional size (10) to reject weak instruments. For this reason, we use the limited information maximum likelihood (LIML) estimator in preference to two-stage least squares, as LIML is less susceptible to bias in small samples with weak instruments (Pischke [40]). The second-stage estimates are very similar to their OLS counterparts reported in Table 4.

Tables 4 and 5 both indicate that in the current data set, there is no evidence for selection on ability into test-taking. The fraction of children in a district taking the G10 exam has no statistical influence on test scores in the same district and cohort. It follows that measures to increase the test-taking rate can be expected to draw in children from a range of abilities that match that of the test-taking sample. However, whether this result is robust in larger samples or richer datasets remains a task for future research.

Finally, we note once again the special circumstances of ethnic minority status. Higher ethnic minority population share results in a significantly lower propensity to take the G10 exam. Table 5 shows that this effect persists in the second-stage result. The very high correlation of ethnic minority population share with district-level poverty and (inversely with) per capita expenditures indicates that Vietnam's ethnic minority children remain in the most distant part of the 'last mile' problem in schooling.

The two-stage model is not free of problems, mostly those related to the small size of the data set. We cannot be completely confident that the exclusion restriction holds for the FDI employment variable. This concern can only be alleviated by obtaining more data and exploiting greater variation through time and across observational units.

6.3. Discussion

Despite data limitations, the estimation results in Tables 4 and 5 provide useful information in their own right and create optimism that richer results could emerge from a larger data set. We conclude this section with a brief discussion of some other threats and opportunities to be considered and, if possible, addressed in future research.

Omitted variables: We have minimal information on individual children, so we are required to assume that individual differences in G10 exam participation and scores are randomly distributed after controlling for known characteristics: ethnicity, age, and local socio-economic and labor market conditions.

Endogenous selection into lower-secondary schools: Some children may attend schools outside their hometown to take advantage of better educational opportunities (this can be seen in the summary statistics, where the maximum value of the test participation rate is above 100% in some districts). If this is widespread, then estimates of community-level influences over outcome variables will exhibit bias. The test score data sets record whether a child is a migrant to the area where they attend school, so there is some capacity to control for this form of selection.

Endogenous investment in schools and educational services: wealthier communities naturally lobby to win high-quality schools and teachers. If this is widespread and unobserved in the data, then we will overestimate the effects of community characteristics. Conversely, if there is deliberate 'catch-up' investment in deprived areas (and if that investment is effective), then we may underestimate the influence of community characteristics. In both cases, information on specific investments will help. In future extensions of this work, we hope that access to school-level data will help minimize this source of bias.

Gender-differentiated effects: While some FDI industries such as ready-made garment production are known to be female-intensive, there is no evidence of a strong gender bias in overall employment shocks inclusive of service sectors such as construction and transportation. However, there may be gendered differences in youth labor supply. These might be economic (for example, boys may be more likely to anticipate taking over a family farm) or cultural (for example, norms of female mobility vary over ethnic identity and religious affiliation). Our data provide no information on these household-level characteristics. However, the test-taking rate and test score models can be estimated separately for boys and girls. Doing so reduces the precision of our estimates slightly but leaves point estimates essentially unchanged (results available from the authors on request). Deeper analysis of gender differences remains an interesting and potentially important area for future research.

7. Conclusions and policy discussion

Vietnam's economy has emerged into middle income after a generation of very rapid growth. That growth has been fueled by globalization, in which the country has exploited its comparative advantage in low-skilled labor-intensive activities. The pattern of growth is not unambiguously positive. While higher per capita incomes promote increased schooling, the labor market offers instant rewards to school-leavers from a very young age. Moreover, the pattern of income gains and job market opportunities appears to be unevenly distributed; for poorer individuals, labor market effects might dominate income effects.

In this study, we examine newly available data on schooling – labor market interactions from several less-privileged provinces within Vietnam. These data are drawn from provinces at the extensive margin of educational and economic development. Their populations are poorer and ethnic minority groups are more prevalent than in the nation as a whole. In locations such as these, the marginal cost of providing additional educational opportunities, especially at upper secondary level, is likely to be high. This, in turn, is a factor inhibiting both poverty reduction and intergenerational mobility, and as such places an additional burden on policy.

Among districts in these provinces, there is considerable variation in the propensity to attempt the entrance examination to Grade 10 – the gateway grade to rise above blue-collar labor market status. This variation runs in predictable patterns across SES and ethnicity. We also find that a more active labor market plays a significant role in discouraging education beyond the working age. On the other hand, we find no evidence that the population of test-takers is selected in such a way as to raise or lower average test scores when other variables, such as schooling costs, are altered.

The idea of policies or programs to raise test participation rates in Vietnam is enshrined in the Education Law of 2019, which made education to Grade 9 free within the public school system. The intent of the law is clearly to promote further educational deepening, so progression to Grade 10 (and beyond) has become a focus of education policy. How to achieve educational deepening, at what cost, and with what aggregate or individual benefits, is a compelling question for policymakers. At present, education policy reforms are focused on supply-side innovations such as school construction, teacher quality, and curriculum reform, and on distributional equity in school access. These are important priorities. Curriculum reform, in particular, has the potential to alter the balance of costs and returns to upper secondary schooling – especially as the current curriculum at that level is tightly focused on academic work in preparation for the college entrance examination, rather than on preparation for labor market entry per se. Our findings, however, indicate that the task of expanding educational attainment extends well beyond the mandate of the Ministry of Education. Vietnam's dramatic entry into global markets, accompanied by huge expansion of employment in relatively low-skilled jobs, create substantial headwinds for educational development policy.

Looking further afield, Vietnam has some special features but is also broadly representative of developing countries that arrived late to globalization and now (or will soon) confront the need to move beyond production based on natural resources, low-skilled labor, and largely foreign capital. Indonesia, Bangladesh, Cambodia, and other countries now face similar human capital investment challenges to Vietnam. Lessons from one country will benefit others.

Acknowledgements

We are grateful to UNICEF-Hanoi and especially Ms. Le Anh Lan for framing conversations and access to data sources. We thank the editor and referees of this journal as well as Valerie Kozel and seminar participants at ERIA, IDE-JETRO (Tokyo), University of Economics Ho Chi Minh City, UNICEF-Hanoi, the Central Institute for Economic Management, and the University of Wisconsin–Madison for helpful comments on earlier drafts.

Disclosure statement

No potential conflict of interest was reported by the author(s).

1 Rankings from the recent World Economic Forum's Readiness for the Future of Production report (WEF [53]) highlighted the challenges that Vietnam faces in transitioning to globalization. On an indicator called 'drivers of production', Vietnam is ranked 13th on global trade and investment among 100 countries, between Australia and France. On human capital, however, it is ranked 70th, between Sri Lanka and Georgia, and on technology and innovation it is ranked 90th, between Paraguay and Cameroon. The human capital index notably includes low rankings for Vietnam on key components such as the mean years of schooling (rank: 74), quality of universities (75), quality of vocational training (80), and on-the-job training (74). 2 Demand for skills has also grown. But skilled jobs are concentrated in government and state-owned enterprises whose activity levels have grown far slower than private-sector industries. The latter are subject to crowding-out in domestic capital markets and have much lower capital and skill-intensity (Phan and Coxhead [39]). 3 Findlay and Kierzkowski ([19]) note that when all individuals are identical, the equilibrium difference in lifetime earnings between skilled and unskilled workers is zero: higher earnings of skilled workers are exactly offset by additional schooling costs that they incur. It follows that a relative price change that favours unskilled labor will induce more children to quit school earlier, other things equal. 4 If services production is of intermediate factor intensity relative to sectors 1 and 2, then an expansion in their demand must reduce the skill premium. Suppose that the endowment point E represents factors available for tradable production after demand for non-tradable services (S) has been satisfied, so E(H,L) = E(HT–HS, LT–LS). Then, increased output of S will reduce E, moving it down and to the left from its original location. The new skill premium must be lower than its original value. 5 For a comparable finding from Mexico see Atkin [7]. In a global panel of countries, Blanchard and Olney ([10]) also found a negative relationship between growth in less skill-intensive exports and educational attainment. 6 In another uniquely Vietnamese dimension, low overall returns to private sector employment of skilled workers have been exacerbated by capital constraints due to crowding out by state sector enterprises (Phan and Coxhead [39]). 7 These estimates are lower than the widely accepted world average of 10% (Psacharopoulos and Patrinos [41]), but are comparable with those from similar regional economies. Tangtipongtul ([47]) estimated an average return to education in Thailand of about 13%, but this is highly convex, with estimated returns to primary schooling only 1.8%, and general secondary school only 5%; the majority of the labor force is educated to these levels or below. For Indonesia, overall returns are estimated in the medium–high single digits (Purnastuti, Miller, and Salim [42]; Newhouse and Suryadarma [36]; Coxhead [11]). 8 A low value of this parameter will also indicate approximate effects of present bias or hyperbolic discounting on a student's schooling decision. 9 When test scores are the dependent variable, province fixed effects play an especially important role since the G10 exam is not nationally uniform. All exams are required to have math and literature sections, but individual provinces have the discretion to add other sections and to apply their own grading standards. A more accurate count of the eligible population would come from counts of school enrolment by the same cohort in years preceding the test, e.g. Grades 7–8. Obtaining these data will be a post-Covid research task. The provinces in our data set are poorer and more rural than the national average, so these test participation rates are also somewhat below the national average rate. The VHLSS covers only a sample of districts in each province. These have been expanded to a full set of district-level SES data using small-area estimation techniques by Nguyen ([37]). Several districts in our data report zero values for FDI employment. Seven district/year observations have very high values and the estimates are sensitive to these. To maintain the focus on relationships in the mass of the distribution rather than in its tails, we have dropped these. Supplemental data for this article can be accessed online at https://doi.org/10.1080/09645292.2022.2068138. Adukia, A., S. Asher, and P. Novosad. 2020. " Educational Investment Responses to Economic Opportunity: Evidence from Indian Road Construction." American Economic Journal: Applied Economics 12 (1): 348 – 376. doi: 10.1257/app.20180036 Aggarwal, S. 2018. " Do Rural Roads Create Pathways out of Poverty? Evidence from India." Journal of Development Economics 133 : 375 – 395. doi: 10.1016/j.jdeveco.2018.01.004 Almond, D., and J. Currie. 2011a. " Killing me Softly: The Fetal Origins Hypothesis'." Journal of Economic Perspectives 25 (3): 153 – 172. doi: 10.1257/jep.25.3.153 Almond, D., and J. Currie. 2011b. " Human Capital Development Before Age 5." In Handbook of Labor Economics Volume 4b, edited by O. Ashenfelter and D. Card, 1315 – 1486. Amsterdam : North-Holland. Arouri, M., A. Ben-Youssef, and C. V. Nguyen. 2019. " Ethnic and Racial Disparities in Children's Education: Comparative Evidence from Ethiopia, India, Peru, and Vietnam." Children and Youth Services Review 100 : 503 – 514. doi: 10.1016/j.childyouth.2019.03.031 Asian Development Bank. 2020. Viet Nam Secondary Education Sector Assessment and Road Map. Manila : ADB. Atkin, D. 2016. " Endogenous Skill Acquisition and Export Manufacturing in Mexico." American Economic Review 106 (8): 2046 – 2085. doi: 10.1257/aer.20120901 Baulch, B., T. M. N. Hoa, T. T. Phuong, and T. P. Hung. 2010. Ethnic Minority Poverty in Vietnam. Working Paper No. 169. Manchester: Chronic Poverty Research Centre. Becker, G. 1954. Human Capital: A Theoretical and Empirical Analysis with Special Reference to Education. New York : Columbia University Press. Blanchard, E. J., and W. W. Olney. 2017. " Globalization and Human Capital Investment: Export Composition Drives Educational Attainment." Journal of International Economics 106 : 165 – 183. doi: 10.1016/j.jinteco.2017.03.004 Coxhead, I. 2014. " Indonesia: Returns to Occupations, Education and Ability During a Resource Export Boom." In Trade, Development and Political Economy in East Asia: Essays in Honour of Hal Hill, edited by P.-C. Athukorala, A. A. Patunru, and B. Resosudarmo, 134 – 152. Singapore : Institute of Southeast Asian Studies. Coxhead, I., V. C. Nguyen, and H. L. Vu. 2019. " Internal Migration in Vietnam, 2000–2012." In Rural–Urban Migration in Vietnam, edited by A. Y. C. Liu and X. Meng, 67 – 96. Cham : Springer. Coxhead, I., and R. Shrestha. 2017. " Globalization and School–Work Choices in an Emerging Economy: Vietnam." Asian Economic Papers 16 (2): 28 – 45. doi: 10.1162/ASEP_a_00515 Dang, H.-A. H., and P. Glewwe. 2018. " Well Begun, but Aiming Higher: A Review of Vietnam's Education Trends in the Past 20 Years and Emerging Challenges." Journal of Development Studies 54 (7): 1171 – 1195. doi: 10.1080/00220388.2017.1380797 Deardorff, A. 2002. "Introduction to the Lerner Diagram." Accessed 31 January 2020. http://www-personal.umich.edu/~alandear/writings/Lerner.pdf. Doan, T., and J. Gibson. 2012. " Returns to Education in Vietnam During the Recent Transformation." International Journal of Education Economics and Development 3 (4): 314 – 329. doi: 10.1504/IJEED.2012.052322 Doan, T., Q. T. Tuyen, and L. Quan. 2018. " Lost in Transition? Declining Returns to Education in Vietnam." European Journal of Development Research 30 (2): 195 – 216. doi: 10.1057/s41287-017-0080-9 Edmonds, E., and N. Pavcnik. 2005. " The Effect of Trade Liberalization on Child Labor." Journal of International Economics 65 (2): 401 – 419. doi: 10.1016/j.jinteco.2004.04.001 Findlay, R., and H. Kierzkowski. 1983. " International Trade and Human Capital: A Simple General Equilibrium Model." Journal of Political Economy 91 (6): 957 – 978. doi: 10.1086/261195 Fu, X., and V. N. Balasubramanyam. 2005. " Exports, Foreign Direct Investment and Employment: The Case of China." The World Economy 28 (4): 607 – 625. doi: 10.1111/j.1467-9701.2005.00694.x Fukase, E. 2013. " Export Liberalization, Job Creation, and the Skill Premium: Evidence from the US–Vietnam Bilateral Trade Agreement." World Development 41 : 317 – 337. doi: 10.1016/j.worlddev.2012.05.035 Gao, W., and R. Smyth. 2015. " Education Expansion and Returns to Schooling in Urban China, 2001–2010: Evidence from Three Waves of the China Urban Labor Survey." Journal of the Asia Pacific Economy 20 (2): 178 – 201. doi: 10.1080/13547860.2014.970607 General Statistics Office of Vietnam. 2015. Statistical Yearbook of Vietnam 2015. Hanoi : GSO. Glewwe, P., Q. Chen, and B. Katare. 2015. " What Determines Learning among Kinh and Ethnic Minority Students in Vietnam? An Analysis of the Round 2 Young Lives Data." Asia & the Pacific Policy Studies 2 (3): 494 – 516. doi: 10.1002/app5.102 Hanushek, E. A., and L. Woessmann. 2012. " Do Better Schools Lead to More Growth? Cognitive Skills, Economic Outcomes, and Causation." Journal of Economic Growth 17 (4): 267 – 321. doi: 10.1007/s10887-012-9081-x Heckman, J., and X. Li. 2004. " Selection Bias, Comparative Advantage and Heterogeneous Returns to Education: Evidence from China in 2000." Pacific Economic Review 9 (3): 155 – 171. doi: 10.1111/j.1468-0106.2004.00242.x Katz, L. F., and K. M. Murphy. 1992. " Changes in Relative Wages, 1963-1987: Supply and Demand Factors." Quarterly Journal of Economics 107 (1): 35 – 78. doi: 10.2307/2118323 Kozel, V. 2014. Well Begun but Not Yet Done: Progress and Emerging Challenges for Poverty Reduction in Vietnam. Washington, DC : World Bank. Li, H., P. Loyalka, S. Rozelle, and B. Wu. 2017. " Human Capital and China's Future Growth." Journal of Economic Perspectives 31 (1): 25 – 48. doi: 10.1257/jep.31.1.25 McCaig, B., and N. Pavcnik. 2018. " Export Markets and Labor Allocation in a Low-Income Country." American Economic Review 108 (7): 1899 – 1941. doi: 10.1257/aer.20141096 McGuinness, S., E. Kelly, T. T. P. Pham, T. T. T. Ha, and A. Whelan. 2021. " Returns to Education in Vietnam: A Changing Landscape." World Development 138. doi: 10.1016/j.worlddev.2020.105205 Mergoupis, T., V. Phan, and J. G. Sessions. 2018. Puzzle me This? the Vietnamese Reverse Gender Education gap. UNU-WIDER Working Paper Series, No. 2018/116. Helsinki : United Nations University World Institute for Development Economics Research. Mincer, J. 1958. " Investment in Human Capital and Personal Income Distribution." Journal of Political Economy 66 : 281 – 302. Moretti, E. 2010. " Local Multipliers." American Economic Review 100 (2): 373 – 377. doi: 10.1257/aer.100.2.373 Neal, D. A., and W. R. Johnson. 1996. " The Role of Premarket Factors in Black–White Wage Differences." Journal of Political Economy 104 (5): 869 – 895. doi: 10.1086/262045 Newhouse, D., and D. Suryadarma. 2011. " The Value of Vocational Education: High School Type and Labor Market Outcomes in Indonesia." The World Bank Economic Review 25 (2): 296 – 322. doi: 10.1093/wber/lhr010 Nguyen, C. V. 2016. "Poverty and Inequality Estimates using the 2014 Vietnam Household Living Standard Survey and the 2014 Intercensal Population and Housing Survey." Unpublished. Nguyen, H. T. M. 2019. " Ethnic Gaps in Child Education Outcomes in Vietnam: An Investigation Using Young Lives Data." Education Economics 27 (1): 93 – 111. doi: 10.1080/09645292.2018.1444147 Phan, D., and I. Coxhead. 2013. " Long-run Costs of Piecemeal Reform: Wage Inequality and Returns to Education in Vietnam." Journal of Comparative Economics 41 (4): 1106 – 1122. doi: 10.1016/j.jce.2013.04.001 Pischke, S. 2018. "Weak Instruments (Presentation Notes)." Last accessed 25 November 2021. https://econ.lse.ac.uk/staff/spischke/ec533/Weak%20IV.pdf. Psacharopoulos, G., and H. Patrinos. 2004. " Returns to Investments in Education: A Further Update." Education Economics 12 (2): 111 – 134. doi: 10.1080/0964529042000239140 Purnastuti, L., P. Miller, and R. Salim. 2013. " Declining Rates of Return to Education: Evidence for Indonesia." Bulletin of Indonesian Economic Studies 49 (2): 213 – 236. doi: 10.1080/00074918.2013.809842 Ray, R., and G. Lancaster. 2005. " The Impact of Children's Work on Schooling: Multi-Country Evidence." International Labor Review 144 (2): 189 – 210. doi: 10.1111/j.1564-913X.2005.tb00565.x Rolleston, C., and P. Iyer. 2019. " Beyond the Basics: Access and Equity in the Expansion of Post-Compulsory Schooling in Vietnam." International Journal of Educational Development 66 : 223 – 233. doi: 10.1016/j.ijedudev.2018.09.002 Rolleston, C., Z. James, L. Pasquier-Doumer, N. T. M. T. Tran, and T. D. Le. 2013. Making Progress: Report of the Young Lives School Survey in Vietnam. Working Paper, No. 100. London : Young Lives, Oxford Department of International Development, University of Oxford. Schultz, T. W. 1961. " Investment in Human Capital." American Economic Review 51 (1): 1 – 17. Tangtipongtul, K. 2015. " Rates of Return to Schooling in Thailand." Asian Development Review 32 (2): 38 – 64. doi: 10.1162/ADEV_a_00051 Truong, H. C. 2011. Eliminating Inter-Ethnic Inequalities? Assessing Impacts of Education Policies on Ethnic Minority Children in Vietnam. Working Paper, No. 69. London : Young Lives, Oxford Department of International Development, University of Oxford. UNESCO. 2011. Global Education Digest 2011: Comparing Education Statistics Around the World – Focus on Secondary Education. Technical Report. Montreal : UNESCO Institute for Statistics. Winters, L. A., N. McCulloch, and A. McKay. 2004. " Trade Liberalization and Poverty: The Evidence so far." Journal of Economic Literature 42 (1): 72 – 115. doi: 10.1257/002205104773558056 World Bank. 2018. Climbing the Ladder: Poverty Reduction and Shared Prosperity in Vietnam, Vietnam Poverty and Shared Prosperity Update Report. Washington, DC : World Bank. World Bank. 2020. "World Bank Open Data." Last accessed 23 September 2020. www.data.worldbank.org. World Economic Forum, 2018. Readiness for the Future of Production Report 2018. Geneva : World Economic Forum.

By Ian Coxhead; Nguyen Dinh Tuan Vuong and Phong Nguyen

Reported by Author; Author; Author