Language, mixed communes and infrastructure sources of inequality and ethnic minorities in vietnam (2)

World Development Vol xx, pp xxx–xxx, 2017 0305-750X/Ó 2017 Elsevier Ltd All rights reserved www.elsevier.com/locate/worlddev http://dx.doi.org/10.1016/j.worlddev.2017.03.004 Language, Mixed Communes, and Infrastructure: Sources of Inequality and Ethnic Minorities in Vietnam HOA-THI-MINH NGUYEN a, TOM KOMPAS b,c, TREVOR BREUSCH a and MICHAEL B WARD d,* a Crawford School of Public Policy, Australian National University, Canberra, Australia b Australian Centre for Biosecurity and Environmental Economics, Crawford School of Public Policy, Australian National University, Canberra, Australia c Centre of Excellence in Biosecurity Risk Analysis, University of Melbourne, Melbourne, Australia d Department of Economics, Monash Business School, Monash University, Melbourne, Australia Summary — This paper re-examines the sources of inequality in Vietnam, a transitional economy with large reductions in poverty from recent and dramatic economic growth, but vastly unequal gains across ethnic groups Using a decomposition approach to disentangle factor endowments and returns by ethnic group, we draw four key conclusions First, removing language barriers would significantly reduce inequality among ethnic groups, narrowing the ethnic gap, and especially so through enhancing the gains earned by minorities from education Second, variations in returns to education exist in favor of the majority in mixed communes, suggesting that either the special needs of minority children have not been adequately addressed in the classroom, or unequal treatment in favor of the majority exists in the labor market Third, in contrast to recent literature, there is no difference in the benefits drawn from enhanced infrastructure at the commune level across ethnic groups Finally, we find little evidence to support the established views that the ethnic gap is attributed largely to differences in the returns to endowments Overall, our research highlights the importance of considering language barriers and the availability of infrastructure for ethnic inequality Ó 2017 Elsevier Ltd All rights reserved Key words — ethnic inequality, language, infrastructure, education, rural development, Vietnam, Asia INTRODUCTION The Government of Vietnam has a number of policies and programs in place to help the minority group These policies and programs are based on two approaches, those that target communes and those that target households As an example of the former, Program 135 largely finances local infrastructure improvement (e.g., the provision of roads, power, and water) in communes faced with extreme difficulties, often in remote and mountainous areas with large minority populations For the latter, the Hunger Eradication and Poverty Reduction Program (HEPR), targets poor households (largely the minority), by providing access to credit, exemption from education fees, and support for health care, among other benefits In spite of these policies and programs, progress in raising the living standard of the minority has been much slower than that for the majority This paper examines what drives the gap in the living standard between the majority and minority groups, measured by differences in household expenditures per person In particular, we investigate the role of language barriers and how they may hinder minority households from taking advantage of their acquired skills and attributes; whether commune infrastructure, a key instrument used by the Vietnamese Government to narrow the ethnic gap, works for or against Inequality in wealth and income is often the source of tension between large disenfranchised groups of relatively poor minorities and the majority population Failure to address this inequality may lead to ethnic conflict, resulting in poor economic performance and political instability (Easterly & Levine, 1997) Although ethnic inequality is not a characteristic of transitional economies alone, such concerns tend to predominate in these countries due to high but unequally shared growth in incomes, substantial differences in initial endowments and dramatically changing institutions and economic conditions that often quickly leave the poor behind Vietnam offers a useful case study in this regard In the transition to a market-based economy, Vietnam has experienced remarkable success in economic growth and poverty reduction GDP per capita in 2008, for example, was three times larger than that in 1986, when Vietnam first made a landmark commitment to economic reform (General Statistic Office, 2000, 2009) Between 1993, when the first household expenditure survey was conducted, and 2006, the poverty rate among the population as a whole fell from 58% to 16% Nonetheless, the gains from growth have not been shared proportionately among different groups of people For example, while the poverty rate of the Kinh and Chinese (defined as the ‘‘majority” in this paper) fell from 54% in 1993 to 10% in 2006, for other ethnic minorities as a whole (defined as the ‘‘minority”), it decreased more modestly, from 86% to 52% over the same period of time (World Bank, 2007) Moreover, in 2006, the minority group accounted for 44% of the poor and 59% of those classified as ‘‘hungry” in Vietnam, despite representing only 14% of the country’s population (World Bank, 2007) The gap in expenditure between the two groups has also widened over time (Baulch, Nguyen, Nguyen, & Pham, 2010; Baulch, Pham, & Reilly, 2012) * Special thanks to Martin Rama for helpful discussions on earlier versions of this work and on ethnic inequality in Vietnam in general Comments and advice from reviewers, Henrik Hansen, Dung Doan, Yuk Chu Liu, Nguyen Thang and seminar participants at the Crawford School of Public Policy, the 2nd Vietnam Economics Workshop at the Australian National University, the 8th Australasian Development Economics Workshop at Monash University, and the 12th Econometric Society Australasian Meetings in Melbourne are greatly appreciated Final revision accepted: March 5, 2017 Please cite this article in press as: Nguyen, H -T M et al Language, Mixed Communes, and Infrastructure: Sources of Inequality and Ethnic Minorities in Vietnam, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.03.004 WORLD DEVELOPMENT the minority; and to what extent preferential treatment and differences in endowments, as opposed to returns to endowments, explain the ethnic gap in expenditures Using household survey data in Vietnam, conducted in 2006, our paper contributes to the literature in two important ways First, it points to language as potentially an important factor explaining the ethnic gap in expenditures Traditionally, language has yet to be explicitly controlled for in econometric models examining ethnic inequality in Vietnam One reason for this could be the small sample size of household data with language variables in the first and early national surveys conducted in Vietnam But perhaps a more fundamental reason is that most studies are restricted to an Oaxaca–Blinder (OB) decomposition framework Here, the differences in average outcomes in the two groups are decomposed into differences in the average level of each characteristic and differences in the returns to these characteristics between groups As language barriers are almost entirely restricted to the minority, inclusion of this variable in the model for only one group is essentially a problem of a lack of common support, violating a key identification assumption in an OB framework (Fortin, Lemieux, & Firpo, 2011) In qualitative analyses, on the other hand, this is never the case Language barriers are seen as key and have been highlighted as a major constraint preventing the minority from taking advantage of government policies and programs (see Tran, 2004; Vasavakul, 2003; Vietnam Academy of Social Sciences, 2009; World Bank, 2009a, among others) Our findings quantitatively corroborate this claim Second, we use an econometric technique that allows us to consistently estimate the difference in household responses to both household-level and commune-level observed variables between the two groups Specifically, we use the moment conditions underlying fixed effects estimators to define instruments for some variables This allows us to combine the mathematical logic of fixed-effects techniques at differing spatial levels along with conventional instruments in an overarching instrumental variables framework To this end, our estimation method differs substantially from existing economic studies on ethnic inequality in Vietnam, which either consistently estimate the household response to household-specific variables only, or estimate its response to both household and commune variables but with significant bias The objective of measuring the household response to both householdspecific and commune-specific observed variables is complicated by two potential concerns The first is the existence of common commune-specific unobserved characteristics, such as local customs, practices, and land and school quality, which are likely correlated with household-specific characteristics Failure to control for this correlation in the use of an OLS estimator (e.g., Baulch et al., 2010, 2012), for example, causes potential bias in estimating returns to those household-specific characteristics (Baltagi, 2005; Hsiao, 2003) This bias can be eliminated using least-squares dummy-variable (LSDV) or fixed effects (FE) estimators as used in Hoang, Pham, Tran, and Hansen (2007) and Van de Walle and Gunewardena (2001) However, this way of eliminating the bias comes at the expense of the ability to estimate household responses to commune-specific observed attributes such as geographical characteristics and infrastructure The second concern involves measuring the household response to infrastructure There are at least two reasons for the need to consider infrastructure in quantifying the expenditure gap by ethnicity in Vietnam The first is that the minority tends to live in more remote areas, characterized by difficult terrain, poor roads, no power, and limited access to markets, making it difficult to isolate the effect of ethnicity itself on the expenditure gap The second reason is that majority households living in the same remote and impoverished areas are doing increasingly well, to the point of being difficult to distinguish from their counterparts in the low-land communes (Swinkels & Turk, 2006) This second concern comes from the way infrastructure is provided and how decisions are made on its provision in Vietnam On the surface, there might be a two-way relationship between infrastructure and household income, hence expenditure But in Vietnam, infrastructure is delivered at the discretion of the Government in a one-party system Furthermore, the Government has invested in both growth and disadvantaged areas, making the link between household expenditure to infrastructure, if any, weak However, depending on how these often decentralized government decisions are made, it is still possible that there is a high correlation between the availability of infrastructure and unobserved factors at different government levels If so, ignoring this possible correlation using an OLS estimator could lead to potential bias in estimating returns to infrastructure Recent literature using OLS estimator indeed suggests that majority households benefit more from local investment and government poverty reduction programs than minority households, thereby further exaggerating the gap in expenditure between the two groups (Pham, Le, & Nguyen, 2011, p 3) Our results find no evidence to support this view The paper is organized as follows Some background is provided in Section 2, including detail on ethnicity and the key government programs designed to assist the minority Data and variables are described in Section Section provides the model specification and estimation method and Section presents results Section concludes, highlighting policy implications and scope for further research ETHNICITY, MIGRATION, AND PROGRAMS AFFECTING THE MINORITY There are 54 officially classified ethnic groups living in Vietnam, including the Kinh and Chinese (Bui, 1999) The Kinh group accounts for about 86 per cent of the population In spite of their diversity, ethnic minorities are usually grouped based on the place where they live For instance, there is a tendency to lump together the ethnic groups in the Northern Mountains, in the Central Highlands, and in the lowlands The largest proportion (about 12 million people) lives in the first two areas The group in the lowlands comprises mainly the Chinese in Ho Chi Minh City, the Khmer in the Mekong Delta, and the Cham in the Southern Coast The biggest groups after the Kinh are the Tay (1.2 million people), the Thai (one million), and the Khmer (one million) The smallest groups, including the Si La, the Pu Pep, the Ro Man, the Brau, and the O Du, include less than one thousand people each (Huynh, Duong, & Bui, 2002) Ethnic minorities have been affected by the consequences of Doi Moi, the process of economic reform and trade liberalization initiated in 1986, much the same as the rest of the population These economic reforms have resulted in an unambiguous improvement in living standards in Vietnam A combination of better incentives, improved access to markets and government support was critical to this success But ethnic minorities have also been affected, not always positively, by specific public policies and programs Some of those policies and programs have had a dramatic impact on their livelihoods, both before and after Doi Moi Please cite this article in press as: Nguyen, H -T M et al Language, Mixed Communes, and Infrastructure: Sources of Inequality and Ethnic Minorities in Vietnam, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.03.004 LANGUAGE, MIXED COMMUNES, AND INFRASTRUCTURE: The first one is the establishment of new economic zones in the areas that had been traditionally inhabited by ethnic minorities, in the form of state agricultural and forest enterprises, as well as new economic villages; a process that reached its peak during 1960–75 This was followed by a planned migration program to massively move Kinh people into these zones As a result, this program affected access to land by ethnic minorities and the ecosystems on which their livelihoods depended The local indigenous people, which accounted for the bulk of the population in the Central Highlands in 1975 ultimately represented only 26% of the population in early 2000s (Luu, 2010) As much as 60% local indigenous households were reported as having no production land as of 2002, while most of the fertile land was in hand of ‘‘immigrants” (Luu, 2010) The second program is the so-called ‘‘sedentarization program” launched in 1968 with an aim to provide support for agricultural production and livelihoods in the mountainous regions It did so by facilitating fixed settlement and cultivation in combination with providing technical training, capacity building, technology transfer, and raising market awareness By 1998, when activities of the sedentarization program were merged into Program 135, which will be discussed later, about 3.8 million people had been resettled The sedentarization program created a challenge for the minority by forcing them to give up their traditional way of cultivation and adopt a new methods apparently more suitable to the lowlands (Jamieson, 1996; Rambo, 1997) In combination with the migration program for the Kinh, this program brought disparity in the way the two groups were treated: majority households received more support from the Government to migrate into minority areas while little support was available to ethnic minorities The World Bank (2009a, p 27), for example, noted that ‘‘some investment programs for the highlands, particularly in the Central Highlands, initially focused on bringing in Kinh migrants to set up services and work opportunities with the State, rather than hiring or promoting local ethnic minorities.” Finally, Vietnam’s national targeted programs (NTP) including HEPR and Program 135 have served to promote balanced growth by investing in disadvantaged areas These two programs were launched in 1998, after Vietnam concentrated first on high-growth areas The HEPR, with a household-targeting approach, aims at eliminating chronic hunger and reducing poverty rates by providing access to credit, exemption from education fees, and support for health care, among other benefits, to entitled households (see Nguyen & Baulch, 2007) On other hand, Program 135, with a commune-targeting approach, largely supports infrastructure development in communes faced with extreme difficulties These communes were selected based on: (i) remoteness (e.g., border communes or more than 20 km distant from a development center); (ii) lack of infrastructure (e.g., no access by car, and with little electricity supply, clean water, schools and health centers); and (iii) having poverty rates exceeding 60%, illiteracy rates exceeding 60% or prevalent health problems (World Bank, 2005) While both NTP are considered as successful in targeting poor households and poor communes, there has been some considerable leakage The poverty rate, one of the selection criteria in both programs, is subject to manipulation since it is defined using a bottom-up approach Evidence suggests that the reported poverty rate is often influenced by ‘‘the willingness of local and provincial authorities to showcase their performance, or simply to match the resources available for targeted programs with the alleged number of eligible beneficiaries” (Nguyen & Rama, 2007) For the Program 135, due to the program’s large coverage, with one in every four communes in Vietnam as its beneficiaries, 45% of non-poor households, on average, live in beneficiary communes (World Bank, 2005) DATA AND VARIABLES (a) Data Our estimates rely on the Vietnam Household Living Standard Survey in 2006 (VHLSS 2006) This is a multi-stage stratified household survey on expenditure, among other indicators, representative of the whole country, for both urban and rural areas Furthermore, the sampling ensured that surveyed households came from all 64 provinces and cities directly under the Government (hereafter referred to as provinces) and the number of households surveyed in each province was in proportion to its population Overall, three households were randomly selected from a single census enumeration area in each commune, making up 9,189 households, living in 3,063 communes, 630 districts, and 64 provinces Since the survey does not cover all the different ethnic groups in full detail, we are forced to simplify the ethnic divisions we consider This lack of information is compounded by the fact that both the poor and ethnic minority households tend to live in rural areas, where we are interested in the determinants of poverty for ethnic minorities alone To focus attention on the questions of interest, and avoid undue confounding of factors, we follow previous research in limiting our sample to rural households, which covers most ethnic minority groups, and then group households into a majority group, combining Kinh and Chinese households, and a minority group, which includes the remaining 52 officially recognized ethnicities One important characteristic of many communes is the presence of majority and minority households within the same commune However, precise information on actual ethnic breakdown in Vietnam’s rural communes is not directly available in the survey As a result, we use information on households surveyed per commune, together with an ethnicity indicator, to construct a proxy of the overall ethnic composition of the commune Expenditure data from VHLSS 2006 also does not allow us to get a good proxy of ethnic composition, since there are only three households surveyed per commune We overcome this by supplementing with income data, which includes more households per commune Based on these constructions, we describe communes as either mixed in the sense they include both the majority and the minority, or non-mixed, when all households sampled in a commune belong to the same ethnic category In total, the rural sample used for this paper has 5,391 majority households and 1,168 minority households in 2,187 communes (in 566 districts, 64 provinces) Using information from income data, 1,611 communes (in 434 districts, 58 provinces) have observations only from the majority group and 243 communes (in 134 districts, 36 provinces) have observations only from the minority On the other hand, 559 majority households and 439 minority households are found in 253 (163) and 214 (134) mixed communes (mixed districts), respectively, in 49 provinces Please cite this article in press as: Nguyen, H -T M et al Language, Mixed Communes, and Infrastructure: Sources of Inequality and Ethnic Minorities in Vietnam, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.03.004 WORLD DEVELOPMENT (b) Variables and summary statistics The key dependent variable we model is real household expenditure per person It is measured in thousand Vietnamese Dong, using January 2006 prices to correct for different inflation rates due to different survey times in different regions of the country It might be argued that income per person, rather than expenditure, is a better indicator of economic welfare, and thus the welfare gap between ethnic groups Nevertheless, income is difficult to measure in a developing country like Vietnam, particularly for the rural and minority groups we are studying In rural areas, there are often no welldeveloped input markets to compute net income from farming and household activities, and no reliable measures of ‘‘ownincome” for household-managed and operated farms, making it difficult to distinguish between revenue and costs (Che, Kompas, & Vousden, 2006), not to mention the potentially large amount of ‘‘informal” or unreported income The extent of own and informal income might differ systematically between social groups, therefore further distorting measured income as an indicator of group welfare Another advantage of using expenditure as a welfare measure is that consumption tends to be smoothed in response to income fluctuations over relatively a long period of time (Deaton, 1997), as thus may be a better indicator of economic welfare than a snapshot of income which is possibly highly transient The explanatory variables we use are characteristics and endowments at both the household and commune levels These include the demographic characteristics of the household and endowments of the household in human and physical capital At the commune level, endowments as infrastructure are similarly distinguished from geographical characteristics Summary statistics for all variables are shown in Table The characteristics of the majority and minority ethnic groups are often very different, and there clearly exist substantial gaps in endowments between the two groups Of particular interest in this study, along with differences in language, infrastructure, and years of schooling, real expenditures per person among the minority are more than a third lower than those for the majority group The mean difference of the two groups is statistically different at the 1% level for all of the variables in Table 1, except for irrigated land, perennial land, other land, and existence of a primary school, and a health care center (i) Household-level variables Household characteristics of interest include household size, proportions of children in the 0–6 and 7–16 year brackets, proportions of male and female adults, household structures (describing whether a household consists of two generations with fewer than three children, two generations with three or more children, and three generations and other household structures), and some household-head specific variables such as age and gender Table suggests that the minority is more likely to have larger families, live in three-generation households, and be headed by a man, while majority households tend to have a higher proportion of members over 16 Household characteristics are expected to have an impact on household expenditure per capita Economies of scale and complementarity in consumption and production are all factors evident in literature (Deaton, 1980; Lazear & Michael, 1988), suggesting the need to control for household size and its composition while other explanatory variables are being considered A pattern of strong negative correlation between household size and consumption per person is found in many household surveys, especially in developing countries (Lipton & Ravallion, 1995) Increasing returns in household produc- tion due to specialization and complementarity of skills could also be expected in agricultural and household businesses, while differences in the age and gender structure could influence consumption patterns across households Furthermore, as child labor is not unusual in Vietnam (Edmonds & Pavcnik, 2005), having an additional household member beyond years old in the household may reduce the total ‘‘time cost” per person for cooking and cleaning, and hence free time for other members to work Household human capital in our data consists of two variables The first variable is the years of schooling of the most educated member Schooling has been widely shown to have strongly positive impacts on income, and hence on expenditures The previous literature on ethnic inequality in Vietnam measured schooling as the highest level of attainment, and used indicator dummy variables to show the impact of different education levels on household living standards We believe that schooling years as a (nearly) continuous variable will allow us to measure incremental returns to education in a better way In our data, the majority group dominates with an average of 9.6 years of schooling of the most educated member, significantly more than the minority group with about eight years The second variable on household human capital of interest is fluency in Vietnamese In the survey, this is simply indicated by whether an interpreter was required to complete the survey Language barriers are expected to influence the household living standard of the minority For example, the minority often reports on their lack of knowledge of government policies and programs due to their inability to read or hear about these measures, let alone to request government services they are entitled to (World Bank, 2009a) Language barriers prevent minority women, in particular, from using government health services, even though they possess health care coverage cards (Tran, 2004) This barrier also results in a higher likelihood of minority children beginning school late, repeating school, or dropping out, compared with their majority counterparts (World Bank, 2009a) Lack of language skills, which leads to lack of confidence and limits social networking, also hinders the minority from sharing information, and accessing off-farm employment outside their specific group (World Bank, 2009a; Vietnam Academy of Social Sciences, 2009) In our sample, the minority group has a significant language barrier, as shown in Table 1, with 28% of such households requiring the needed of language interpretation for the survey interview The primary measure of household physical capital in the survey is land We include different types of land: irrigated annual, non-irrigated annual, perennial, forests, watersurfaced land, and other land Land disaggregation by type is needed to control for heterogeneity in land productivity of various types of land, crops that they can produce, and land tenures, as well as their usefulness as collateral for loans (Kompas, Che, Nguyen, & Nguyen, 2012) In general, the quantity of land area is markedly in favor of minority households, except for the water-surface land where holdings by majority households are nearly four times as great as minority households The distinction is important because of great variability in land quality It is generally the case that the more fertile and higher market-valued land is in the deltas and coastal areas (ii) Commune-level variables In our model, the commune-level variables include commune characteristics and infrastructure Commune characteristics capture inherent location effects There are three types of communes in the data set: coastal and delta land, hilly land, Please cite this article in press as: Nguyen, H -T M et al Language, Mixed Communes, and Infrastructure: Sources of Inequality and Ethnic Minorities in Vietnam, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.03.004 LANGUAGE, MIXED COMMUNES, AND INFRASTRUCTURE: Table Descriptive statistics Variables Majority mean Minority mean Difference mean 5,151 2,916 2; 235a 4.10 5.14 À1:04a 0.07 0.19 0.34 0.39 0.12 0.24 0.31 0.33 À0:05a À0:05a 0:03a 0:06a 0.54 0.21 0.13 0.12 0.22 49.94 0.36 0.30 0.17 0.17 0.11 43.95 0:18a À0:09a À0:04a À0:05a 0:11a 5:98a 0.00 9.58 0.28 8.01 À0:28a 1:57a 2.73 0.48 1.42 0.61 0.39 0.03 3.24 4.93 1.48 5.54 0.11 0.07 À0:50a À4:45a À0:06 À4:93a 0:28a À0:04 Commune characteristics Proportion of households living in communes located in Rural coastal and delta land Rural hilly land Rural mountains Distance to city (1,000 km) 0.71 0.08 0.20 0.14 0.10 0.01 0.89 0.25 0:62a 0:07a À0:69a À0:11a Commune infrastructure Proportion of households living in communes that have Power Car road Primary school Health care center Village daily market Local enterprise 0.99 0.39 0.50 0.98 0.36 0.67 0.95 0.14 0.52 0.98 0.11 0.40 0:04a 0:25a À0:02c 0.00 0:25a 0:27a Number Number Number Number 5,391 1,864 474 63 1,168 457 221 50 Real per capita expenditure in Jan 06 prices (VND000s) Household characteristics Household size Proportion of household member(s) Aged from to Aged from to 16 Males, aged over 16 Females, aged over 16 Proportion of households that consist of Parent(s) only/Parent(s) with fewer than children Parent(s) and three or more children Three-generation households Other household structures Headed by female (yes = 1) Age of household head Household Human Capital Require interpretation Years of schooling of the most educated member Household physical capital: land area (1,000 m2) Irrigated annual land Non-irrigated annual land Perennial land Forestry land Water-surface land Other land of of of of households communes districts provinces (a) p < 0.01, (b) p < 0.05, (c) p < 0.10 and mountain land Distance from the commune to the nearest city is also a key identifier Geography and distance to a city has been widely shown to be important in development (e.g., Audretsch & Feldman, 1996) There is sharp contrast between the proportions of households living in various types of communes: 71% of majority households are concentrated in rural coastal and delta land areas, while as much as 89% of minority households reside in rural low- or high-mountain areas Likewise, minority households live almost twice as far from the city compared to majority households All of this suggests the need to control for these commune characteristics in model specification Commune infrastructure reflects the existence of basic infrastructure, trade facilities, and off-farm employment opportunities We include four measures of basic infrastruc- ture which cover access to power, car-suitable roads, a primary school, and a health care center Each of these variables has a plausible causal relationship with household income and expenditure For example, one might expect having access to power would enhance household production and consumption and thus positively influence household expenditure The availability of a car-suitable road greatly assists the transportation of goods and services, and reduces travel costs of dwellers both in time and money Finally, the existence of a primary school and a health care center would likely help increase community-wide education and health, thereby enhancing living standards We measure ‘‘trade facilities” of households in the commune by the existence of a daily village market Clearly, a daily village market is likely to help facilitate trade and the exchange Please cite this article in press as: Nguyen, H -T M et al Language, Mixed Communes, and Infrastructure: Sources of Inequality and Ethnic Minorities in Vietnam, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.03.004 WORLD DEVELOPMENT of products, which helps promote household production and consumption Finally, off-farm employment is often seen as one important channel out of poverty in rural areas Its importance is quantified by the existence of local enterprises in the commune which usually indicates the presence of a highly developed local market economy For all items in commune infrastructure, minority households are often far less well off than their majority counterparts Exceptions are access to a primary school where the difference by ethnicity is in favor of the minority, though marginal, and access to a health care center which is equal between the two groups Furthermore, many more majority households live in communes with local enterprise(s) One limitation of our data is that variables on commune endowments not measure quality, or differences in quality, which are likely to vary across regions If anything, we suspect the data we have on infrastructure understates the differences between the two groups MODEL SPECIFICATION AND ESTIMATION METHOD Household expenditure per capita is assumed to be a function of household-specific and commune-specific characteristics and endowments The dependent variable appears in the equation as its logarithm, representing constant proportional effects of explanatory variables on per capita household expenditure Household size as an explanatory variable also appears in log form We are interested in exploring how the two ethnic groups differ in their responses to their household-specific characteristics and endowments, which also vary on average across communes and districts, as well as to communespecific characteristics and infrastructure, which are variable only between communes but across districts Our basic specification for each group is: ln Ejil ẳ X 0jil b ỵ Z 0il c ỵ ẵgl ỵ ail ỵ jil ; for j ¼ 1; ; M i ; i ¼ 1; ; N l ; l ¼ 1; ; L ð1Þ where Ejil is the per capita household expenditure for household j living in commune i of district l; X jil is a K  vector of household-specific explanatory variables, which include household characteristics and household human and physical capital, while Z il is a G  vector of commune-specific explanatory variables, which include commune characteristics and commune infrastructure These variables are described in detail in the previous section We note explicitly that X jil varies over household j, commune i, and district l, while Z i varies over only commune i and district l The dimensioned parameter vectors b and c, the district-specific intercept gl , the commune-specific intercept ail , and the error term jil are all unobserved There are M i households in each of the i ¼ 1; ; N l communes in each of the l ¼ 1; ; L districts, making LNM observations in P P total, where LNM ¼ Ll¼1 Ni¼1 M i The composite error term in Eqn (1), formed by the districtspecific intercept gl , the commune-specific intercept ail , and the error term jil in the square brackets, captures both the variation in coefficients across districts, communes and the usual idiosyncratic error term If this composite error term was uncorrelated with each of the explanatory variables, OLS would provide a consistent estimator Furthermore, with the assumptions of homoskedasticity and independence of effects, a random-effects (RE) estimator would be consistent and efficient However, such an absence of correlation is unlikely The first concern is correlation of the commune-specific unobserved effect ail with commune averages of the household-level explanatory variable X jil In our application, when households are sampled by communes, they share common commune-level effects, thereby forming a panel-like data structure These commune-level effects partially reflect local practices and customs as well as the unobserved qualities of institutions and endowments such as schools and land It is likely that these terms are correlated with the commune averages of household characteristics and endowments, such as family size, educational attainment, and land ownership For example, a commune with a school of good (or bad) quality would witness their households having more (or fewer) years of schooling on average Such unobserved school quality is included in ail , making it correlated with variable household years of schooling in X jil The second concern involves measuring the household response to infrastructure which stems from the way infrastructure is provided and how decisions are made on its provision in Vietnam On the surface, there might be a two-way relationship between infrastructure availability and household income, hence expenditure That is, infrastructure is more likely provided to wealthy areas (and often with high population density) for economic efficiency, while household income increases due to infrastructure availability But in Vietnam, the causal effect from household living standards to infrastructure delivery is not very likely for a few reasons First, infrastructure delivery is determined by the Government in a market economy with socialist orientation, not by profit-oriented companies Second, the impact of household expenditure on infrastructure delivery, if any, would cancel each other out on average because the Government has invested in both high-growth and disadvantaged areas Finally, Program 135, a key program to promote balanced growth by investing in infrastructure development in poor communes, has considerable leakage and, as discussed, a very wide coverage across communes in Vietnam, with an average of 45% of non-poor households in each beneficiary commune While the link from household expenditure to infrastructure is probably weak, there is likely much heterogeneity in the way in which decisions are made on infrastructure provision in Vietnam High decentralization gives provincial governments decisive power in resource allocation within their provinces For the small- and medium-scaled infrastructure of our interest in this paper, provincial governments allocate resources and delegate authority to their subordinate government levels at the district level for implementation in line with the province’s socio-economic development plans and strategies (World Bank, 2009b) In Program 135, in particular, district governments can receive central level funding directly and make final decisions on the list of infrastructure investment projects to be carried out in participating communes (World Bank, 2009b) In the context of our model, the way infrastructure is delivered in Vietnam suggests a possible high correlation of districtspecific unobserved effects gl with commune-level explanatory variables on infrastructure in Z il The district-level unobserved effects may reflect their geographical features (aside from distance and regions captured in commune characteristics), political and leadership qualities of local government, and other socio-economic characteristics that are correlated with the amount of resources a district receives for infrastructure development They also could well capture the heterogeneity across districts in project selection and implementation which in its turn affects the availability of infrastructure at its subordinate level, the commune Please cite this article in press as: Nguyen, H -T M et al Language, Mixed Communes, and Infrastructure: Sources of Inequality and Ethnic Minorities in Vietnam, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.03.004 LANGUAGE, MIXED COMMUNES, AND INFRASTRUCTURE: When the commune and district effects are correlated with explanatory variables, OLS (and RE) produces inconsistent estimates for the coefficients b and c The standard remedy is to use a FE estimator, which is equivalent to including a dummy variable for each commune and district Unfortunately, under the FE approach, coefficients c of communelevel variables Z are not identified In a model with a random intercept, Hausman and Taylor (1981) observe that the FE estimator can be viewed as an instrumental variables (IV) procedure, where the instruments are deviations in the X explanatory variables around their group means (the ‘‘within-group” variation) This interpretation suggests that the coefficients b and c can then be estimated, provided sufficient additional instruments to separately identify them are available as shown by Breusch, Ward, Nguyen, and Kompas, 2011; Breusch, Ward, Nguyen, and Kompas, 2011 To describe our IV estimator, some notation is helpful We can stack the data with all households in all communes and districts, so the model for LNM observations can be written compactly as: Y ẳ X b ỵ Z ịc1 ỵ Z ịc2 ỵ  2ị where Z and Z are vectors of commune characteristic and infrastructure, respectively It is useful to define the LNM  LNM ‘‘within” operator QC as the matrix that converts any variable of data, such as Y = lnE and the columns of X, into the deviations from its commune-level means Likewise, the LNM  LNM ‘‘within” operator QD as the matrix that converts any variable of data, such as the columns of Z, into the deviations from its districtlevel means In this notation, the instruments for the household-level variables X are QC X , their within-commune deviations and the instruments for the commune infrastructure Z are QD Z , their within-district deviations Commune characteristics including geography and distance to the city Z are instrumented by themselves since they can clearly be seen as exogenous The IV estimator has the following moment conditions: ½QC X ; Z ; QD Z Š0 ½Y À X b À Z c1 Z c2 ẳ 0; 3ị which decompose as: ẵQC X Y X bị ẳ 0; ½Z ; QD Z Š0 ½Y À X b À Z c1 À Z c2 Š ¼ ð4Þ The first of these conditions in (4) shows that the estimator of b is the same as fixed-effects In particular, b will be consistent in situations when there exists correlation between the commune-specific random intercept ail and a household-level explanatory variable X jil The second condition in (4) assumes that Z is exogenous When this condition is satisfied, the estimators of c1 and c2 will be consistent even if there exists correlation between the district-specific random intercept gl and a commune-level explanatory variable Z ji One of the key concerns in our research is the difference in returns to characteristics and endowments between groups To allow for this possible difference, we add interactions between the covariates and group dummy-variables to the regression specified in Eqn (1) applied to the whole sample ESTIMATION RESULTS We first check if our estimation method is appropriate, and then examine the effect of language, mixed communes, and infrastructure, with a focus on differences in returns to those attributes between the two ethnic groups Finally, after various robustness checks, we decompose the ethnic gap (a) Checking the estimation method In the previous section, we discussed two possible channels of endogeneity and reasons why they could happen Here is just a summary The first channel of endogeneity is the correlation of the commune-specific unobserved effect ail with commune averages of the household-level explanatory variable X jil A good example would be that communes with schools of good quality increase the quantity of education households obtain on average Such an ‘‘omitted” variable on school quality is included in ail and correlated with the commune average part of the household-level variable years of schooling in X jil The second channel is the correlation of district-specific unobserved effects gl with commune-level explanatory variables on infrastructure in Z il A good example would be that political and leadership qualities of a district government would affect the amount of resources the district can attract for infrastructure, which in its turn affects the average availability of infrastructure facilities at its subordinate level, the commune With these two channels of endogeneity in mind, we proposed the use of within-commune deviations QC X for household-level variables X and within-district deviations QD Z for commune-level infrastructure A valid instrument needs to be exogenous and relevant Our instruments are exogenous under these two channels of endogeneity because ail and gl are correlated with the average part in X jil and Z il Removing the average part in X jil and Z il or their ‘‘betweenvariation” eliminates these sources of endogeneity The second condition on instrument relevance is straightforward in this case Since the instruments QC X and QD Z are formed from the ‘‘within-variation” of X and Z, respectively, they must be strongly correlated with each other For the rest of this sub-section, we check if data confirm our concerns over the two channels of endogeneity, and test the relevance of instruments The condition of exogeneity cannot be tested, but we believe our arguments based on the economic intuition above and in the previous section are sufficient In the sub-section on robustness checks, we discuss other possible sources of endogeneity, try different instruments where possible, and discuss the consequences if no valid instruments are available to address any particular source of endogeneity We begin by testing the need for using within-commune deviations QC X for household-level variables X and withindistrict deviations QD Z for commune-level infrastructure The Hausman test, which compares coefficients estimated using RE and FE estimators, rejects the null hypothesis that a RE estimator is consistent (v224 ¼ 110:88), suggesting the need for a FE estimator This test result confirms our concern about the correlation between commune-level unobserved effects and (commune-level averages of) household-level characteristics and endowments as well as the correlation between district-level unobserved effects and (district-level averages of) commune-level infrastructure, thereby lending credence to our use of within-commune deviations QC X and within-district deviations QD Z as instruments Unlike simple FE estimation, where the nature of the instruments obviates the need to adjust for the error correlation that is due to the panel structure of the data, in this broader IV estimation the error correlation should be taken into account We conduct this IV estimation in Stata, version 11, using the command ivregress 2sls with option vce(robust) to correct for heteroskedasticity The F-statistic for the first stage regression Please cite this article in press as: Nguyen, H -T M et al Language, Mixed Communes, and Infrastructure: Sources of Inequality and Ethnic Minorities in Vietnam, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.03.004 WORLD DEVELOPMENT Table Difference in determinants of household expenditure by ethnicity Household characteristics Household size (log) Aged from to 16 Males, aged over 16 Females, aged over 16 Parent(s) and three/more children Three-generation households Other household structures Headed by female (yes = 1) Age of household head Age of household head (squared/100) Household human capital Require interpretation Years of schooling Household physical capital Irrigated annual land (1,000 m2) Non-irrigated annual land (1,000 m2) Perennial land (1,000 m2) Forestry land (1,000 m2) Water-surface land (1,000 m2) Other land (1,000 m2) Commune characteristics Rural mountains Rural hilly land Distance to city (1,000 km) Commune infrastructure Power Car-suitable road Primary school Health care center Village daily market Local enterprise Majority Group (1) Pooled (2) Majority (3) Minority (4) Difference À0:332a ðÀ11:43Þ 0:159a (2.70) 0:484a (6.98) 0:427a (5.78) À0:060a ðÀ2:93Þ À0:056a ðÀ2:26Þ 0.000 (0.01) À0:071a ðÀ3:49Þa 0:011a (2.90) À0:012a ðÀ3:57Þa À0:328a ðÀ10:30Þ 0:150b (2.22) 0:474a (6.07) 0:423a (5.13) À0:068a ðÀ2:89Þ À0:052c ðÀ1:89Þ 0.013 (0.44) À0:069a ðÀ3:14Þ 0:014a (3.26) À0:015a ðÀ3:85Þ À0:395a ðÀ5:40Þ 0:292b (2.39) 0:569a (3.58) 0:499a (2.72) À0:030 ðÀ0:69Þ À0:060 ðÀ1:06Þ À0:026 ðÀ0:45Þ À0:089 ðÀ1:60Þ À0:002 ðÀ0:20Þ À0:000 ðÀ0:04Þ 0.067 (0.84) À0.142 (À1.01) À0.095 (À0.54) À0.076 (À0.38) À0.038 (À0.77) 0.008 (0.12) 0.039 (0.60) 0.020 (0.32) 0:016c (1.67) À0:015 ðÀ1:60Þ À0:139b ðÀ2:06Þ 0:054a (20.55) 0:058a (20.04) À0:123b ðÀ2:00Þ 0:035a (5.33) 0:023a (3.24) 0:009a (5.01) 0:006a (3.23) 0:009a (5.97) 0:001b (2.27) 0:012a (4.12) 0.006 (0.60) 0:010a (6.61) 0.004 (1.44) 0:008a (5.37) 0:001c (1.92) 0:011a (3.91) 0.007 (0.68) 0:007c (1.69) 0:009a (4.23) 0:013a (4.22) 0:001c (1.79) 0:022a (4.47) À0.003 (À0.14) 0.003 (0.69) À0.004 (À1.14) À0.005 ðÀ1:35Þ 0.000 (0.10) À0:011c (À1.92) 0.010 (0.38) À0:057a ðÀ3:56Þ 0.014 (0.60) À0:310a ðÀ6:27Þ À0:028 ðÀ1:63Þ 0.015 (0.61) À0:370a ðÀ7:05Þ À0:192a ðÀ4:33Þ À0:006 ðÀ0:06Þ À0:234b (À2.13) 0:164a (3.44) 0.021 (0.20) À0.136 (À1.12) 0:195a (2.71) 0:030b (2.02) 0.000 (0.01) 0:129a (2.63) 0:040a (2.65) 0:053a (3.64) 0:247a ð9:60Þ À0:041 ðÀ0:07Þ 0:029c (1.85) À0:001 ðÀ0:05Þ 0:141b (2.50) 0:038b (2.45) 0:044a (2.78) 0:201a (2.93) 0.031 (0.77) 0.024 (0.76) À0:021 ðÀ0:23Þ 0.073 (1.35) 0:097a (2.88) À0:242 ðÀ0:43Þ À0:002 ðÀ0:05Þ À0:025 ðÀ0:71Þ 0.162 (1.51) À0:035 ðÀ0:62Þ À0:053 ðÀ1:41Þ (continued on next page) Please cite this article in press as: Nguyen, H -T M et al Language, Mixed Communes, and Infrastructure: Sources of Inequality and Ethnic Minorities in Vietnam, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.03.004 LANGUAGE, MIXED COMMUNES, AND INFRASTRUCTURE: Table (continued) Constant Number Number Number Number of of of of households communes districts provinces (1) Pooled (2) Majority (3) Minority (4) Difference 7:176a (57.14) 7:537a (13.13) 7:826a (33.56) À0:289 (À0.47) 6559 2187 566 64 5391 1864 474 63 1168 457 221 50 6559 2187 566 64 Log of real per capita expenditure as dependent variable t statistics are in brackets (a) p < 0.01, (b) p < 0.05, (c) p < 0.10 The regression omits the proportion of members aged 0–6; one- or two-generation households that consist of no more than two children; commune geographical types as coastal/ delta; other dummy variables have yes = and no = for the whole rural sample is 35.00 It is well above the critical value of 21.42, identified by Stock and Yogo (2005) for a 5% maximal IV relative bias It thus appears to exclude a possible problem with weak instruments Finally, we perform a Chow test for whether the coefficients estimated over the majority are equal to the coefficients estimated over the minority group The tests soundly reject the null hypothesis that they are equal between groups (v226 ¼ 69:51) (b) The effect of language Table reports results of estimating Eqn (2) by an IV estimator with the instruments described in Eqn (3) We find a negative and significant effect of the presence of a language barrier on household expenditure Being language ‘‘incompet ent”—requiring interpretation during the survey interview— is associated with an approximately 13% fall in household expenditure in the pooled model This negative association is not sensitive to the choice of instruments for language ability, which we will discuss in detail in our robustness checks Furthermore, this relationship remains almost the same in the regression for minority households only, suggesting that even among minority households, where people can communicate in their own minority language, not being fluent in Vietnamese imposes a substantial disadvantage In Table 2, column (4), highlighting differences between groups, and where language differences are excluded, we see little evidence of a significant difference in returns to characteristics and endowments An exception is in years of schooling and whether or not the household lives in a mountainous area This is in spite of substantial difference in intercepts, as shown in column (1), where being a majority household is associated with a roughly 22% increase in expenditure To further investigate the effect of language in household returns to attributes, we compare returns between households with and without language ability (i.e., those that not or require interpretation) In Table 3, for brevity, we present only variables with statistically significant differences in returns While we observe statistically significant differences in returns to some infrastructure facilities between the two language groups, the most pronounced finding is that the difference in returns to education between them is highly statistically significant This difference is much larger than that between the two ethnic groups: 0.042 per year of schooling as compared with 0.023 Language barriers thus seem to contribute to a widening of the ethnic gap, at least through the channel of differences in household gains from education Does the ethnic gap dissipate among language competent households? In the regression for language competent households, we measure this gap using an indicator of ‘‘belonging to the majority group” By doing so, we assume that the returns to attributes are the same among language competent households Results are shown in Table 3, where we find the ethnic gap — the estimated coefficient of the indicator ‘‘Majority Group”—remains almost the same as that in the pooled regression This result implies that being competent in Vietnamese is not enough to eliminate the ethnic gap in expenditures To further disentangle the ethnic gap among language ‘‘competent” households, we narrow our sample to households that speak the majority language, and we relax our assumption that returns to attributes by ethnicity are the same In Table 4, for brevity, we again only present variables with statistically significant differences between the two ethnic groups In terms of returns per year of schooling, the results still favor the majority, albeit now at much smaller values Specifically, the difference between groups falls to 0.013 per year of schooling, and is statistically insignificant This difference is less than two thirds of the difference between the two ethnic groups in the whole rural sample On the other hand, the ethnic difference in returns to perennial land still favors minority households and becomes almost twice as much as that in the whole rural sample Working on perennial land is a comparative advantage of minority households, given their indigenous knowledge Having language skills seems to further enhance this advantage Results in Table 4, thus suggest, once again, that the ethnic gap could be narrowed by removing the language barrier through enhanced returns to education and (as well) a less restricted use of perennial land for the minority Before ending this subsection, we add a caveat to the interpretation of the coefficient estimate of the language barrier variable Since the formulation of this variable – requiring interpretation during the survey interview or not—is relatively crude, and does not reflect the whole spectrum of language incompetence of those being interviewed, we might over- or under-estimate the effect of this variable Despite this possible lack of precision in the coefficient estimate, our results are qualitatively correct, and provide a call for removing language barriers to reduce the ethnic disparity (c) The effect of mixed communes Table extends the results in Table by separating language ‘‘competent” households by both ethnicity and whether they live in a mixed or non-mixed commune The idea is to see if living in the same commune helps further reduce the ethnic gap after controlling for language In this case, we find that the difference in returns to education between majority and minority households in non-mixed communes is very small and statistically insignificant, or 0.006 per year of schooling (p-value = 0.478) On the contrary, the difference is much larger, at 0.021 per year of schooling between majority and Please cite this article in press as: Nguyen, H -T M et al Language, Mixed Communes, and Infrastructure: Sources of Inequality and Ethnic Minorities in Vietnam, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.03.004 10 WORLD DEVELOPMENT Table Difference in determinants of household expenditure by language ability Minority group (yes = 1) Household human capital Require interpretation (yes = 1) Years of schooling Commune characteristics Rural mountains Commune infrastructure Car-suitable road Primary school Health care center Local enterprise Number Number Number Number of of of of households communes districts provinces (1) Pooled (2) Not require interpretation (3) Require interpretation (4) Difference 0:247a ð9:60Þ À0:244a ðÀ12:54Þ À0:139b ðÀ2:06Þ 0:054a (20.55) 0:056a (20.88) 0.014 (1.05) 0:042a (3.15) À0:057a ðÀ3:56Þ À0:055a ðÀ3:39Þ À0:377a ðÀ2:81Þ 0:322b (2.39) 0:030b (2.02) 0.000 (0.01) 0:129a (2.63) 0:053a (3.64) 0:031b (2.07) À0.006 ðÀ0:42Þ 0:141a (2.75) 0:051a (3.45) À0.100 (À1.31) 0:158b (2.17) À0:143 ðÀ0:90Þ 0:208a (2.60) 0:131c (1.69) À0:163b ðÀ2:21Þ 0:284c (1.70) À0:158c ðÀ1:93Þ 6,559 2,187 566 64 6,228 2,099 543 64 331 136 78 27 6,559 2,187 566 64 Pool regression and notes are the same as those in Table Table Difference in determinants of household expenditure by ethnicity among households not requiring interpretation Household characteristics Proportion of household member aged from to 16 Age of household head Household human capital Years of schooling Household physical capital Perennial land (1,000 m2) Commune characteristics Rural mountains Distance to city (1,000 km) Minority Group (yes = 1) Number Number Number Number of of of of households communes districts provinces (1) Not require interpretation (2) Majority (3) Minority (4) Difference 0:192a (3.12) 0:011a (2.94) 0:152b (2.23) 0:014a (3.28) 0:420a (2.84) À0:004 ðÀ0:44Þ À0:268c ðÀ1:65Þ 0:018c (1.68) 0:056a (20.88) 0:058a (20.05) 0:045a (5.91) 0.013 (1.61) 0:009a (6.09) 0:008a (5.34) 0:017a (4.96) À0:009b (À2.38) À0:055a ðÀ3:39Þ À0:337a ðÀ7:03Þ À0:244a ðÀ12:54Þ À0:028 ðÀ1:63Þ À0:369a ðÀ7:04Þ À0:205a ðÀ4:31Þ À0:141 ðÀ1:17Þ 0:177a (3.49) À0:228c ðÀ1:74Þ 6,228 2,099 543 64 5,391 1,864 474 63 837 356 182 47 6,228 2,099 543 64 Regression for households not requiring interpretation is the same as that in Table Other notes are the same as those in Table minority households in mixed communes While the significance level of the latter is only 15%, it is possibly large enough for the small sample of mixed communes with the use of within-variance as instruments This larger difference in returns to education among language ‘‘competent” households in mixed communes is initially surprising In the absence of language barriers, one would expect to see that living in the same commune with Please cite this article in press as: Nguyen, H -T M et al Language, Mixed Communes, and Infrastructure: Sources of Inequality and Ethnic Minorities in Vietnam, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.03.004 LANGUAGE, MIXED COMMUNES, AND INFRASTRUCTURE: 11 Table Difference in determinants of household expenditure by ethnicity among households not requiring interpretation in mixed and non-mixed communes Non-mixed communes Majority Years of schooling Perennial land (1,000 m2) Number Number Number Number of of of of households communes districts provinces a Mixed communes Minority a Majority a Minority 0:058 (19.87) 0:008a (4.73) 0:053 (5.88) 0.012 (1.42) 0:060 (6.81) 0:009a (2.97) 0:039a (3.83) 0:018a (4.24) 4,832 1,611 434 58 450 163 100 32 559 254 163 49 387 193 119 44 Information based on ethnic mix is from income-level data, thereby causing differences in the number of mixed communes for the majority and the minority Differences in returns to years of schooling between the majority and minority in mixed communes is significant (p = 0.146); between majority and minority in non-mixed communes is insignificant (p = 0.478) Differences in returns to perennial land between majority and minority in mixed communes is significant (p = 0.096); between majority and minority in non-mixed communes is insignificant (p = 0.588) Other notes are the same as those for Table the majority group would generate comparable returns to education for the minority In mixed communes, one might also expect little difference in education quality—minority and majority students are in the same classrooms—and that the availability of employment in the commune to both groups should be comparable This wider disparity in returns to education suggests at least two possibilities First, education quality may not be uniform to majority and minority groups in the same classroom This will often be the case where special needs of minority children are not addressed in mixed classrooms, the way they might be in those with minority children only In mixed classrooms, in other words, minorities may be ‘‘left behind” compared to their majority counterparts This will be especially the case if minority children enter the classroom with differential skills and prior differences in home-based learning Second, the disparity in returns to education may suggest unequal treatment in labor markets within mixed communes This is likely rooted in the government policy which has largely favored Kinh movement into minority areas, rather than the reverse (World Bank, 2004; Huynh et al., 2002) As indicated in Section 2, the labor market in mixed communes was distorted by migration patterns (World Bank, 2009a), which generally favored the majority at the expense of minority groups in terms of concessions for migration and the establishment of preferential employment by local governments for the majority If the majority is favored in hiring in mixed communes—and differences between individuals by ethnicity are commonly known or identified at the point of the job application—this may explain the difference in returns and expenditure patterns Indeed, the results show that not only are returns to education of minority households in mixed communes much lower than those of the majority households, but that they are also much lower than those of their minority counterparts living in non-mixed communes (d) The effect of infrastructure In Table 2, we find no statistically significant difference in returns to commune infrastructure between the two ethnic groups This result is not sensitive when we further control possible province-specific effects in our estimation, or affected by possible correlation between infrastructure and communespecific unobserved effects, both of which we will discuss later in our robustness checks Nonetheless, most of infrastructure facilities have a statistically positive impact on the majority while few on the minority group It can be seen that the patterns in the returns to infrastructure are relatively divergent between the two groups This divergence is in contrast to the similarity in trend in returns to household physical and human capital Understanding this divergence sheds light on what mechanisms outside the household could give rise to the lower living standard of the minority There are a number of results to highlight in this regard First, our results show that having access to power is the only item among the basic infrastructure categories (i.e., power, road, primary schools and health care center), largely funded by the government, which benefits the minority The statistically insignificant impact for the majority is possibly due to little variation within this group (99% of majority households are already inside the network), which also explains why the estimated coefficients for accessing power in the pool and the minority regressions are similar This result, in fact, points more to the difference in access to power, rather than the difference in returns to access to power between the two groups Indeed, although for the whole country, 98% households have access to the national grid, the remaining 2% households are typically among minority households who live in poor and remote communes (Nguyen, Phung, Phung, Vu, & Westbrook, 2013; Wells-Dang, 2012) Our results suggest that the (minority) household expenditure would increase by about 18% with access to electricity, which is comparable to recent findings in Khandker, Barnes, and Samad (2014), who used a panel data for the years of 2001 and 2005 to analyze the welfare impact of rural electrification in Vietnam Second, it is clear from the statistical evidence that the existence of a car-suitable road and a daily market does not help the minority, but supports the majority This result is broadly consistent with the standard view in qualitative research that the majority does better than the minority in a market context and in market places, and that the majority dominates overall in trading systems (World Bank, 2009a; Wells-Dang, 2012) Barriers that the minority often face in the market place include lack of market information, language difficulties and calculation skills, and a culture of community reciprocity which prevents many minorities from doing business in an otherwise or strictly entrepreneurial fashion These barriers together with the attachment to their kinship also result in the minority’s reluctancy to take full use of physical connectivity to the world outside their villages (see Tran, 2004; World Bank, 2009a; Vietnam Academy of Social Sciences, 2009) Please cite this article in press as: Nguyen, H -T M et al Language, Mixed Communes, and Infrastructure: Sources of Inequality and Ethnic Minorities in Vietnam, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.03.004 12 WORLD DEVELOPMENT Third, comparing results from Tables and 3, what is interesting to see is that while both majority and minority households at large not benefit from primary schools, the language incompetent households among the minority This result corroborates anecdotal evidence in qualitative research indicating that there has been a change in views about education among minorities since the early 2000s brought about by improved living standards and more exchange with outside communities (Le, 2006) As such, the increase in enrollment rates is reportedly higher among minority households, especially the poor ones, than those among the majority (Nguyen et al., 2013) and the younger generation is becoming more educated than their parents (Wells-Dang, 2012) Fourth, living in a commune with a health care center has the largest statistically positive effect among all infrastructure facilities on the majority, but no statistical impact on the minority This result is comparable to the findings in a recent review by Ma˚lqvist, Hoa, Liem, Thorson, and Thomsen (2013) on the presence of severe inequity in health along ethnic lines, which derives from both external and internal group dynamics On the former, there is a lack of cultural adaptation and sensitivity in government policies and programs, coupled with language barriers, bad attitudes, and occasional discrimination from health staff toward minority persons (Toan, Trong, Hojer, & Persson, 2002) On the latter, customs, traditional practices, and beliefs have greatly influenced the minority’s use of the health care system (Castel, 2009; Rheinlander, Samuelsen, Dalsgaard, & Konradsen, 2011) Finally, in terms of off-farm employment, minority households appear to benefit more than majority counterparts from the presence of local enterprises Indeed, the presence of local enterprises has the largest positive impact on language incompetent minority households (Table 3) Lack of social networks, poor infrastructure, their isolated nature, attachment to their kinship, and limited mobility of many minority households are all likely key drivers for this result (e) Robustness checks Before turning to decomposition of the ethnic gap, which relies on estimated results, we subject our estimates to a few robustness checks The approach here is threefold, concentrating on three concerns First, we check if there is sensitivity in the choice of instrument for a key variable, namely ‘‘require interpretation.” Second, we further control possible province-specific effects by adding a dummy variable for each province to see if the results are robust Finally, we exclude outliers from our sample, in particular, households with remittances (often cash-transfers from overseas relatives) and households who are classified as poor in the survey year, or the year prior For the first concern, its clear the the ability of a household to speak Vietnamese may be affected by factors other than the commune-average ability to speak Vietnamese, which we have controlled for using within-commune variation as an instrument It may also depend on how assimilated to the Kinh group is that household and/or the specific ethnicity (out of the 52 officially recognized ones in the minority group) it belongs to At the household level, the degree of assimilation may, for example, come from intermarriage with a Kinh partner (at present or in the past) Unfortunately, we not have information on ethnicity of any household member rather than the household head, so we are not being able to assess this effect For a specific ethnicity, on the other hand, there is evidence that the better off a household is, the more assimilated it tends to be to the Kinh majority (Baulch, Chuyen, Haughton, & Haughton, 2007) To this end, the average household expenditure per capita of each of the 52 officially recognized ethnicities in the minority group may correlate with that ethnicity’s propensity to speak Vietnamese We address this concern using within-ethnicity variation to speak Vietnamese as an instrument for the variable ‘‘require interpretation”, checking to see if our main results are sensitive to this choice of instrument The idea of this instrument is to avoid possible endogeneity caused by between-ethnicity variation to speak Vietnamese which might correlate with the dependent variable, namely the average household expenditure per capita Results are presented in Table 6, columns (4)–(5) For brevity, only results in pooled and regression differences are presented As we might expect, we now find a smaller negative impact from the presence of a language barrier Not being able to speaking Vietnamese is now associated with a 8% reduction in expenditure as compared to 13% using the within-commune variation as an instrument Nonetheless, other variables in our regression remain extremely stable in their contribution to the ethnic gap regardless of which instrument was used The second concern stems from the fact that the provincial government has decisive power in allocating public resources within their provinces, thereby possibly affecting the availability of infrastructure locally We address this concern by checking if our estimates are sensitive when province-specific effects are further controlled for in the model Results are presented in Table 6, columns (6)–(7), where a dummy for each province was added to the model estimating the main results As can be seen, all coefficient estimates remain robust, except the ones for distance to cities and the indicator of being in rural hilly land These two variables switch in sign and the impact of distance becomes statistically insignificant, which might be well expected since dummies for provinces, which captures geography among other provincial characteristics, and geography itself tend to be strongly correlated with each other One might also suspect a possible correlation between commune-specific unobserved effects and infrastructure We cannot rule out this possibility, but we also not have much ground to establish a specific channel for it While our estimation approach could not address the endogeneity caused by this possible correlation, the overall impact on our results should be negligible for two reasons First, we are interested in the difference in returns to infrastructure of the two groups Second, the correlation between commune-specific unobserved effects and infrastructure, if any, should be similar across groups since infrastructure is provided by the government, not determined by the household, which leads to similar possible bias in the estimated coefficients for the two groups To this end, this possible bias will cancel each other out when we take the differences in the estimated coefficients of infrastructure for the two groups For the third concern, its also clear that households classified as poor are likely to receive support from various government policies, such as the HEPR program, as discussed in Section 2, and this will likely effect expenditures Furthermore, given remnants of the war in Vietnam as well as the Vietnamese culture of reciprocity, some households may receive overseas and/or domestic remittances This external support, regardless of whether it originates from relatives or through the government, would substantially alter expenditures and the way a household gains from its endowments In Table 6, we exclude each group of outliers in turn In columns (8)–(9), for the restricted sample of households not receiving remittances, only a negligible ethnic difference is observed in returns to household capital as compared with Please cite this article in press as: Nguyen, H -T M et al Language, Mixed Communes, and Infrastructure: Sources of Inequality and Ethnic Minorities in Vietnam, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.03.004 LANGUAGE, MIXED COMMUNES, AND INFRASTRUCTURE: 13 Table Robustness checks (1) Full sample H-hold characteristics Household size (log) Aged from to 16 Males aged over 16 years Female aged over 16 years Parent(s) and three or more children Three-generation households Other household structures Headed by female Age of household head Age of household head, squared/100 H-hold human capital Require interpretation Years of schooling H-hold physical capital Irrigated crop land (1,000 m2) Non-Irrigated crop land (1,000 m2) Perennial land (1,000 m2) Forestry land (1,000 m2) Water-surface land (1,000 m2) Other land (1,000 m2) Commune characteristics Rural mountains Rural hilly lands Distance to city (1,000 km) Commune infrastructure Power Car-suitable road Primary school Health care center (2) ‘‘Require Interpretation” instrumented by propensity to speak Vietnamese (3) Control for province-specific effects (4) Exclude households with remittances (5) Exclude poor households Pooled Diff-ce Pooled Diff-ce Pooled Diff-ce Pooled Diff-ce Pooled Diff-ce À0:332a ðÀ11:43Þ 0:159a ð2:70Þ 0:484a ð6:98Þ 0:427a ð5:78Þ À0:060a ðÀ2:93Þ À0:056b ðÀ2:26Þ 0:000 ð0:01Þ À0:071a ðÀ3:49Þ 0:011a ð2:90Þ À0:012a ðÀ3:57Þ 0:067 ð0:84Þ À0:142 ðÀ1:01Þ À0:095 ðÀ0:54Þ À0:076 ðÀ0:38Þ À0:038 ðÀ0:77Þ 0:008 ð0:12Þ 0:039 ð0:60Þ 0:020 ð0:32Þ 0:016c ð1:67Þ À0:015 ðÀ1:60Þ À0:332a ðÀ11:44Þ 0:160a ð2:72Þ 0:486a ð7:01Þ 0:428a ð5:80Þ À0:060a ðÀ2:91Þ À0:055b ðÀ2:25Þ 0:000 ð0:01Þ À0:071a ðÀ3:48Þ 0:011a ð2:90Þ À0:012a ðÀ3:57Þ 0:067 ð0:85Þ À0:142 ðÀ1:01Þ À0:097 ðÀ0:55Þ À0:076 ðÀ0:38Þ À0:038 ðÀ0:77Þ 0:007 ð0:12Þ 0:040 ð0:60Þ 0:019 ð0:32Þ 0:015c ð1:67Þ À0:014 ðÀ1:61Þ À0:332a ðÀ12:40Þ 0:159a ð2:91Þ 0:485a (7.49) 0:427a ð6:24Þ À0:060a ðÀ3:17Þ À0:056b ðÀ2:45Þ 0:000 ð0:02Þ À0:071a ðÀ3:77Þ 0:011a ð3:08Þ À0:012a ðÀ3:80Þ 0:081 ð1:04Þ À0:122 ðÀ0:90Þ À0:075 ðÀ0:43Þ À0:014 ðÀ0:07Þ À0:036 ðÀ0:75Þ 0:006 ð0:11Þ 0:034 ð0:53Þ 0:012 ð0:21Þ 0:014 ð1:51Þ À0:013 ðÀ1:45Þ À0:320a ðÀ10:95Þ 0:153b ð2:56Þ 0:483a ð6:83Þ 0:440a ð5:82Þ À0:062a ðÀ3:00Þ À0:069a ðÀ2:78Þ 0:004 ð0:14Þ À0:068a ðÀ3:21Þ 0:012a ð3:37Þ À0:014a ðÀ4:12Þ 0:085 ð1:06Þ À0:144 ðÀ1:03Þ À0:061 ðÀ0:34Þ À0:032 ðÀ0:15Þ À0:031 ðÀ0:63Þ À0:007 ðÀ0:11Þ 0:044 (0.66) 0.004 (0.07) 0.015 (1.63) À0:014 ðÀ1:55Þ À0:346a ðÀ11:11Þ 0:124c (1.95) 0:391a (5.27) 0:384a (4.87) À0:059a ðÀ2:67Þ À0:049c ðÀ1:92Þ À0:010 ðÀ0:34Þ À0:060a ðÀ2:79Þ 0:011a (2.67) À0:012a ðÀ3:17Þ 0.137 (1.37) À0:235 ðÀ1:47Þ À0:165 ðÀ0:80Þ À0:217 ðÀ1:00Þ À0:042 ðÀ0:72Þ À0:013 ðÀ0:18Þ 0.078 (0.98) À0:016 ðÀ0:22Þ 0.018 (1.64) À0:017 ðÀ1:57Þ À0:139b ðÀ2:06Þ 0:054a ð20:55Þ 0:023a ð3:24Þ À0:082b ðÀ2:26Þ 0:054a ð20:55Þ 0:023a ð3:24Þ À0:138b ðÀ2:26Þ 0:054a ð22:31Þ 0:022a ð3:31Þ À0:136b ðÀ1:98Þ 0:054a ð20:37Þ 0:019a ð2:71Þ À0:110 ðÀ1:14Þ 0:049a ð17:42Þ 0:027a (3:24) 0:009a ð5:01Þ 0:006a ð3:23Þ 0:009a ð5:97Þ 0:001b ð2:27Þ 0:012a ð4:12Þ 0:006 ð0:60Þ 0:003 ð0:69Þ À0:004 ðÀ1:14Þ À0:005 ðÀ1:35Þ 0:000 ð0:10Þ À0:011c ðÀ1:92Þ 0:010 ð0:38Þ 0:009a ð5:02Þ 0:006a (3.21) 0:009a (5.96) 0:001b (2.29) 0:012a (4.12) 0.006 (0.60) 0.003 (0.69) À0.004 (À1.17) À0.005 (À1.36) 0.000 (0.11) À0:011c (À1.92) 0.011 (0.38) 0:009a ð5:03Þ 0:006a ð3:28Þ 0:009a ð6:66Þ 0:001b ð2:31Þ 0:012a ð4:13Þ 0:006 ð0:61Þ 0:003 ð0:77Þ À0:004 ðÀ1:11Þ À0:005 ðÀ1:55Þ 0:000 ð0:01Þ À0:011c ðÀ1:94Þ 0:013 ð0:36Þ 0:008a (4.85) 0:006a (3.26) 0:009a (5.83) 0:001c (1.68) 0:011a (3.69) 0.006 (0.59) 0.003 (0.71) À0:003 ðÀ0:83Þ À0:004 ðÀ1:20Þ À0:000 ðÀ0:06Þ À0:007 ðÀ0:25Þ 0.010 (0.38) 0:007a (4.48) 0:004b (2.21) 0:008a (5.54) 0:001c (1.92) 0:011a (4.06) 0.000 (0.01) 0.005 (1.46) À0:002 ðÀ0:71Þ À0:007c ðÀ2:01Þ 0.001 (0.61) À0:012b ðÀ2:44Þ À0:078b ðÀ2:57Þ À0:057b ðÀ3:56Þ 0:014 ð0:60Þ À0:310a ðÀ6:27Þ 0:164a ð3:44Þ 0:021 ð0:20Þ À0:136 ðÀ1:12Þ À0:057a ðÀ3:56Þ 0.014 (0.59) À0:328a ðÀ7:16Þ 0:166a (3.49) 0.021 (0.19) À0:120 ðÀ1:11Þ À0:070a ðÀ3:65Þ À0:075 ðÀ3:02Þ 0:087 ð10:4Þ 0.062 (1.12) 0.075 (0.65) 0.159 (1.20) À0:038b ðÀ2:35Þ 0.032 (1.31) À0:320a ðÀ6:39Þ 0:165a (3.44) 0.005 (0.05) À0:174 ðÀ1:42Þ À0:025 ðÀ1:47Þ 0.029 (1.13) À0:249a ðÀ4:47Þ 0:097c (1.71) 0.102 (0.58) À0.008 ðÀ0:05Þ 0:195a ð2:71Þ 0:030b ð2:02Þ 0:000 ð0:01Þ 0:129a ð2:63Þ À0:242 ðÀ0:43Þ À0:002 ðÀ0:05Þ À0:025 ðÀ0:71Þ 0:162 ð1:51Þ 0:192a (2.69) 0:030b (2.02) À0:000 ðÀ0:03Þ 0:129a (2.62) À0.238 (À0.42) À0.001 (À0.03) À0.025 (À0.70) 0.166 (1.54) 0:197a ð2:92Þ 0:031a ð2:28Þ À0:001 ðÀ0:04Þ 0:135a ð2:90Þ À0:207 ðÀ0:45Þ À0:036 ðÀ0:82Þ À0:016 ðÀ0:46Þ 0:154 ð1:45Þ 0:181b (2.50) 0:029c ð1:95Þ À0.002 (À0.15) 0:094c (1.94) À0:232 0:269a 0.492 ðÀ0:41Þ (3.12) (0.83) 0.011 0:028c À0.057 (0.26) (1.82) (À1.09) À0:013 0.001 À0.020 ðÀ0:36Þ (0.10) (À0.52) 0.164 0:152a 0:264b (1.54) (2.72) (2.31) (continued on next page) Please cite this article in press as: Nguyen, H -T M et al Language, Mixed Communes, and Infrastructure: Sources of Inequality and Ethnic Minorities in Vietnam, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.03.004 14 WORLD DEVELOPMENT Table (continued) (1) Full sample Pooled a Village daily market Local enterprise Majority group Constant Number Number Number Number of of of of households communes districts provinces (2) ‘‘Require Interpretation” instrumented by propensity to speak Vietnamese Diff-ce Pooled a Diff-ce (3) Control for province-specific effects Pooled a Diff-ce (4) Exclude households with remittances Pooled a Diff-ce (5) Exclude poor households Pooled b Diff-ce 0:040 ð2:65Þ 0:053a ð3:64Þ 0:247a ð9:60Þ 7:176a ð57:14Þ À0:035 ðÀ0:62Þ À0:053 ðÀ1:41Þ À0:289 ðÀ0:47Þ 0:040 (2.70) 0:052a (3.62) 0:261a ð12:66Þ 7:168a (57.13) À0.037 (À0.65) À0.052 (À1.40) À0.301 (À0.49) 0:040 ð2:89Þ 0:053a (3.94) 0:253a ð10:09Þ 7:234a ð55:83Þ À0:033 ðÀ0:60Þ À0:027 ðÀ0:71Þ À0:305 ðÀ0:58Þ 0:042 (2.76) 0:046a (3.20) 0:255a (9.73) 7:153a (56.59) À0:017 ðÀ0:30Þ À0:046 ðÀ1:25Þ À0.322 (À0.52) 0:039 (2.50) 0:054a (3.52) 0:232a (7.80) 7:251a (50.58) À0.052 (À0.83) À0:087b (À2.13) À1:199c (À1.81) 6,559 2,187 566 64 6,559 2,187 566 64 6,559 2,187 566 64 6,559 2,187 566 64 6,559 2,187 566 64 6,559 2,187 566 64 6,234 2,180 566 64 6,234 2,180 566 64 5,365 2,120 552 64 5,365 2,120 552 64 Log of real per capita expenditure as dependent variable t-statistics are in brackets (a) p < 0.01, (b) p < 0.05, (c) p < 0.10 The regression results omit the proportion of members aged 0–6; one- or two-generation households that consist of no more than two children; commune geographical types as coastal/ delta; other dummy variables having yes = and no = that estimated for the full sample In particular, this ethnic difference falls slightly, undoubtedly since many more majority households, rather than minority ones that receive remittances, are removed from the sample Taking out households classified as poor in column (10)– (11) leaves coefficients relatively stable While language barriers become statistically insignificant in this subsample of ‘‘better off” households, disparity in returns to education increases, though marginally Non-poor minority households tend to be able to make more use of their land and local off-farm opportunities in this subsample (f) Decomposition of the ethnic gap In this section, we employ a Blinder–Oaxaca decomposition (Blinder, 1973; Oaxaca, 1973), commonly applied in labor economics, to investigate how household and commune characteristics and endowments contribute to differences between the majority and minority groups This technique has been applied in previous research on ethnic inequality in Vietnam (Baulch et al., 2007, 2010, 2012; Hoang et al., 2007; Van de Walle & Gunewardena, 2001) The basic idea is to split (in this case) the average log expenditure gap between the majority and minority populations into the gap caused by the difference in endowments (the regressors) and the difference in returns to endowments (the regression parameters), or ^m ỵ Z m^cm ỵ ^ ^e ỵ Z e^ce ỵ ^ am ị ðX e b ae Þ Yb m À Yb e ¼ ðX m b ^ ¼ ðX m À X e ịbm ỵ Z m Z e ị^cm |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} Difference in Endownment ^ b ^ ị ỵ Z ^c ^c ị ỵ ^ ^ị ỵ X e b am À a |fflfflfflfflfflfflfflfflmfflfflfflfflfflfflfflfflfflfflfflefflfflfflfflfflfflfflfflfflfflfflfflefflfflfflffl{mzfflfflfflfflfflfflfflfflfflefflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffleffl} ð5Þ Difference in Returns to Endowment ^ ỵ Z Z ị^c ẳ X m X e Þb |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffleffl{zfflfflfflfflfflfflfflmfflfflfflfflfflfflfflfflfflfflfflefflfflfflfflffl}e Difference in Endownment ^ Àb ^ ị ỵ Z ^cm ^ce ị ỵ ^ am ^ aị ỵ X m b |fflfflfflfflfflfflfflfflfflmfflfflfflfflfflfflfflfflfflfflfflefflfflfflfflfflfflfflfflfflfflfflfflmfflfflfflffl{zffl fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffle} Difference in Returns to Endowment ð6Þ where the subscripts m and e stand for majority and minority groups, respectively The left-hand side is the total difference in expenditures, or the ethnic gap, while the sum of the first two terms on the right-hand side of Eqns (5) and (6) is the difference in endowments, and the remainder of the expression is the difference in returns to endowments Arrangement in Eqn (5) assumes that in the absence of unequal treatment, the household expenditure distribution that prevails relates to that of the majority This assumption is sensible since the majority accounts for as much as 86% of the entire population in Vietnam Under this assumption, the coefficients of the majority are the non-discriminatory coefficients in the OB decomposition, or what we call the majority as the reference group (Oaxaca, 1973) On the other hand, the expression in Eqn (6) is based on a belief, which has not been underpinned by existing literature, that there is no discrimination against the minority but only (positive) discrimination of the majority Here, the minority group is the reference group The choice of the reference group would lead to an index number problem That is, the decomposition result would be unstable depending on this choice (Oaxaca, 1973) To address this problem, Oaxaca and Ransom (1994), among others, suggest the use of a weighted sum of the coefficients estimated separately for each group, or equivalently, the coefficients from a pooled model over both groups In our case, however, the coefficients of the pooled regression are very similar to those of the majority one (Tables 2–4), and so will be the decomposition results using these two sets of coefficients Therefore, we mainly discuss the decomposition results obtained using the majority as the reference group specified in Eqn (5) These results are presented in Table 7, alongside those using Eqn (6) as done in previous studies Two further issues are also worth mentioning First, controlling language barriers in the model, which are restricted almost entirely to the majority, leads to the problem of a lack of common support, violating a key identification assumption of the OB framework (Fortin et al., 2011) To address this problem, we restrict our attention to a sub-sample of language competent households Having no better information on Please cite this article in press as: Nguyen, H -T M et al Language, Mixed Communes, and Infrastructure: Sources of Inequality and Ethnic Minorities in Vietnam, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.03.004 LANGUAGE, MIXED COMMUNES, AND INFRASTRUCTURE: 15 Table Decomposition of ethnic gap Reference group Estimation method Not require interpretation, whole rural sample Not require interpretation, mixed rural communes Majority group Majority group Minority group Minority group OLS FE IV OLS FE IV OLS FE IV OLS FE IV Total difference 0:445a (19.88) 0:445a (28.96) 0:445a (24.85) 0:445a (19.88) 0:445a (28.96) 0:445a (24.85) 0.351a (8.58) 0.351a (11.36) 0.351a (10.16) 0.351a (8.58) 0.351a (11.36) 0.351a (10.16) Endowments 0.200a (10.28) 0.075a (9.37) 0.047a (5.30) À0.016c (À1.66) 0.048a (6.22) 0.046a (3.69) 0.109a (7.64) 0.083a (10.80) 0.049a (6.52) À0.022b (À2.38) 0.179a (6.81) 0.083a (10.01) 0.049a (6.41) À0.022b (À2.04) 0.046 (1.16) 0.024a (4.12) 0.348a (9.79) 0.088a (5.61) 0.049a (4.84) À0.025b (À2.84) 0.159a (4.08) 0.077a (5.35) 0.104a (5.44) 0.089a (6.08) 0.039a (5.00) À0.024b (À2.50) 0.307a (9.06) 0.084a (5.16) 0.038a (4.42) À0.027b (À2.71) 0.144a (3.25) 0.068a (4.86) 0.162a (5.26) 0.042a (2.77) 0.064a (3.35) 0.005 (0.43) 0.025 (1.55) 0.027b (2.2) 0.098a (3.77) 0.035a (2.22) 0.061a (3.79) 0.002 (0.19) 0.132a (4.08) 0.036b (2.19) 0.059a (3.63) 0.002 (0.2) 0.021 (1.54) 0.015 (0.86) 0.247a (5.99) 0.064a (2.99) 0.058a (3.17) 0.01 (0.62) 0.074a (3.14) 0.041b (2.16) 0.121a (3.56) 0.064a (3.1) 0.049a (3.29) 0.008 (0.35) 0.214a (5.59) 0.061 (2.97)a 0.039a (2.63) 0.012 (0.55) 0.059b (2.58) 0.044a (2.42) 0.245a (10.50) 0.394c (1.81) À0.020 (À0.40) À0.008 (À0.45) 0.040 (0.86) 0.079 (1.05) À0.239 (À1.05) 0.336a (20.86) 0.430c (1.82) 0.099c (1.72) À0.001 (À0.06) À0.192 (À0.85) 0.266a (9.34) 0.467c (1.79) 0.114c (1.61) À0.013 (À0.60) 0.038 (0.93) À0.069 (À0.25) À0.270 (À0.68) 0.097a (2.65) 0.381c (1.77) À0.022 (À0.40) 0.001 (0.05) À0.071c (À1.88) 0.048 (0.65) À0.239 (À1.05) 0.342a (16.96) 0.424c (1.80) 0.109c (1.72) 0.001 (0.09) À0.192 (À0.85) 0.138a (3.99) 0.465c (1.79) 0.125c (1.61) À0.009 ðÀ0:60Þ À0.060c (À1.79) À0.113 (À0.39) À0.270 (À0.68) 0.188a (4.83) 0.805c (1.94) 0.05 (0.54) À0.019 (À0.83) 0.035 (0.43) 0.053 (0.44) À0.735c (À1.65) 0.252a (8.7) 0.909b (2.01) 0.106 (1.0) À0.011 (À0.35) À0.752c (À1.71) 0.219a (5.63) 0.951c (1.87) 0.168 (1.33) À0.033 (À0.89) 0.04 (0.49) À0.245 (À0.73) À0.664 (À1.04) 0.103b (2.3) 0.783c (1.89) 0.056 (0.54) À0.024 (À1.02) À0.015 (À0.2) 0.039 (0.32) À0.735c (À1.65) 0.23 (6.62) 0.88c (1.95) 0.119 (1.0) À0.017 (À0.49) À0.752c (À1.71) 0.136a (3.18) 0.927 (1.82)c 0.188 (1.33) À0.043 (À1.28) 0.002 (0.02) À0.273 (À0.79) À0.664 (À1.04) H-hold characteristics H-hold education H-hold land Com characteristics Com infrastructure Returns to end-ments H-hold characteristics H-hold education H-hold land Com Characteristics Com infrastructure Constant z statistis are in brackets (a) p < 0.01, (b) p < 0.05, (c) p < 0.10 language competency, discussed earlier, may not allow us to eliminate this problem entirely, but probably a large part of it The second issue is the ‘‘omitted group” problem which refers to the arbitrariness in detailed decomposition due to the choice of the omitted base category in categorical variables controlled in the model Since there is no general solution to this problem, we choose the commonly used approach in the literature of imposing a restriction so that the coefficients of all categorical groups reflect their deviations from their overall mean (Fortin et al., 2011) Decomposition results are obtained using a Stata package by Jann (2008) for all language competent households in the whole rural sample as well as in the mixed rural communes only The results are arranged in accordance with estimation results presented in Table 2, which include household characteristics, household education, household land, commune characteristics, and commune infrastructure We also present our IV-based decomposition results in comparison with those generated by OLS and FE estimates, which were previously applied in the literature Our IV-based decomposition results give little evidence to support the established view that the ethnic gap is largely explained by the differences in returns to endowments (Table 7, column 4) Furthermore, despite accounting for around 60% of the ethnic gap (columns and 10), and about 55% if the pooled regression coefficients are used as the reference, most of the detailed components in the difference in returns to endowment are statistically insignificant Key exceptions are differences in returns to household education and household characteristics, both statistically significant and ones that substantially widen the ethnic gap It is worth noting the contribution of the difference in constants is large, though not statistically significant, reducing the ethnic gap by as much as 61% On the other hand, almost all of the components in the difference in endowment are highly statistically significant Indeed, with an exception of land, they all contribute to widening the ethnic gap Our IV-based aggregate decomposition results are quite different from FE-based decomposition results (column 3) as well as established views in the literature In particular, the FE-based decomposition results suggest that differences in returns to endowments explain about 76% of the ethnic gap These results are very similar to findings in the literature using FE estimators on earlier household data (Baulch et al., 2007; Hoang et al., 2007; Van de Walle & Gunewardena, 2001), 10 and again, they are quite different compared to our IV-based decomposition results One possible reason for this difference between IV-based and FE-based decomposition results (using the same dataset VHLSS 2006) is that the FE estimator does not explicitly control for commune-specific observed attributes such as geography and infrastructure, and hence misses an important part of the story on inequality Since detailed decomposition and its standard errors are not reported in previous studies, we could not compare them with our results Nonetheless, it is important to emphasize that our results not only show little evidence of the dominant role of the difference Please cite this article in press as: Nguyen, H -T M et al Language, Mixed Communes, and Infrastructure: Sources of Inequality and Ethnic Minorities in Vietnam, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.03.004 16 WORLD DEVELOPMENT in the returns to endowment in the ethnic gap but also indicates the lack of statistical significance in most of its components OLS-based decomposition results, on the other hand, are closer to our IV-based decomposition ones because they can also estimate returns to commune-specific attributes However, the estimation bias of OLS causes a corresponding bias in the OLS-based decomposition results, which is especially evident in differences in returns to household education For example, IV-based (and FE-based) decomposition results suggest the principal driver in the difference in returns to endowments is the ethnic disparity in returns to education To put it differently, according to these two consistent estimators, even if the minority had the same level of education, they could never ‘‘catch up” with the majority since the minority has lower returns to those attributes On the contrary, OLS-based decomposition results show a statistically insignificant and much smaller difference in returns to education between the two groups The biased OLS-based decomposition results reported in Table (columns and 5) are also comparable to OLS-based decomposition findings in previous literature (Baulch et al., 2010, 2012) A sensible question raised by a reviewer is why the decomposition results are not stable using IV and OLS estimates versus FE estimates when switching the reference group This instability in IV-based and OLS-based decompositions stems from the difference in returns to commune characteristics (i.e., location) As seen in Eqns (5) and (6), this difference is calculated as the sum of products of group means of commune characteristics and the difference in returns to those characteristics between the two groups (i.e Z e ð^cm À ^ce Þ using Eqn (5) and Z m ð^cm À ^ce Þ using Eqn (6)) Depending on which group is the reference group, the group mean of the other group is used in the multiplication Since most minority households live in mountainous areas while majority households live in delta regions (Table 1), Z m substantially differs from Z e Meanwhile, the difference in coefficients between the two groups (^cm À ^ce ) is very tiny Altogether, they result in Z e ð^cm À ^ce Þ being wildly different from Z m ð^cm À ^ce Þ But this sizeable difference does not imply that there is anything wrong with the decomposition results The choice of which result is ‘‘correct” should be based on what would likely be the non-discriminatory coefficient estimates as explained previously Control variables in FE regression, on the other hand, belong to a subset of those controlled in IV and OLS regressions and are not largely different from one group to the other Therefore, FE-based decomposition results appear more stable than the IV-based and OLS-based ones One might wonder what we have learnt from using IV-based decomposition in addition to that from the FE-based decomposition? Is this additional information useful? To answer these questions, it is important to note that FE estimator controls for commune fixed effects in the sense that using the FE estimator to estimate the response of household-level variables X to the household expenditure is equivalent to using the Ordinary Least Squares (OLS) estimator but also controlling all commune dummies Estimates of these commune dummies are called the commune fixed effects, representing the effect of each commune over and above the effect of X The average of these commune fixed effects is the constant in a FE regression As it is useful to somehow demystify these effects, our IV-based decomposition aims to shed additional light on them by explicitly considering commune-level observables Given this aim, a couple of interesting insights are obtained from the IV-based decomposition results First, the difference in commune infrastructure widens the ethnic gap in the full rural sample (column 4) while the difference in returns to them seems to narrow it in the mixed rural communes (column 9) Second, both differences in location and returns to location enlarge the ethnic gap, suggesting the need to improve infrastructure and relax the restriction on mobility to narrow the ethnic gap In summary, our decomposition results support government investment in commune infrastructure to mitigate ethnic inequality and add a further call for improving education for the minority in both the number of years and (presumably) the quality of education as well as allowing for more mobility, in particular, among the minority to ease the ethnic disparity CONCLUSION This paper has examined what drives the gap in household expenditures between the two ethnic groups in Vietnam We have argued that language barriers contribute to worsening ethnic inequality and that there is no statistically significant difference in returns to commune infrastructure between the two ethnic groups We also show that there is little evidence to support the established view of an ethnic gap being largely explained by the difference in returns to endowment In addition, we demonstrate that while removing language barriers eases ethnic inequality, especially through the gains from schooling and, to a lesser extent, from land, the ethnic difference in returns to education are intensified in an environment where households from both groups are found to live Our results thus suggest that either the special educational needs of minority children are not being addressed in classrooms, or unequal treatment in favor of the majority exists in the labor market Despite some data limitations in our use of the language variable, the importance of language is a new and perhaps surprising addition to the understanding of the differences in expenditures among majority and minority households in Vietnam Much of this is potentially attributable to the quality of 2006 VHLSS data, and the use of a language indicator in our estimates, but the size of the effect in the estimation indicates its relative significance, and greatly conditions how we might interpret results from estimates using earlier versions of VHLSS data Vietnam is a highly dynamic country, and welfare data can easily become dated There is a clear need to keep updating the empirical evidence on this important topic In that sense, our work can be seen as an important case study, examining the situation at a particular place and time, with the hope that we can draw general inferences from this case and apply them accordingly Unfortunately, the lack of available information on land irrigation in VHLSS 2010 and afterward is a concern since it leaves out an important measure of the quality of land across data sets While the ethnic minority households tend to have much more land than their majority counterparts, their land is of much lower quality Not having this piece of information results in biased and inconsistent estimates of returns to land to both ethnicity groups With this mind, there are at least three issues worth exploring in future work First, it would be useful to further investigate the effects of social characteristics on household living standards Community and social effects, along with language skills and the ability to ‘‘social network”, are important indicators of differences in living standards and potential changes in the poverty rate Little in our current work accounts for network and community effects, except for the simple presence of language barriers and the co-location that goes with being in a Please cite this article in press as: Nguyen, H -T M et al Language, Mixed Communes, and Infrastructure: Sources of Inequality and Ethnic Minorities in Vietnam, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.03.004 LANGUAGE, MIXED COMMUNES, AND INFRASTRUCTURE: mixed commune Second, there is a need to investigate the role of labor mobility and barriers in the labor market that prevent more equalized returns from education These are partly bound up in language differences, but in many circumstances this may not be the case In this regard, our results would also benefit from measures of the differences in labor market experience among households, as well as a further investigation of potential preferential treatment for the majority (in particular) in mixed communes Finally, minorities tend to live in the more remote and mountainous areas of Vietnam and it is clear that differences in location may exacerbate ethnic disparity A 17 relaxed policy on migration, especially for the minority would help ease this disparity On the other hand, while constraints derived from both external and internal group dynamics remain to be addressed for the minority to take full use of some basic infrastructure, increases in infrastructure availability likely contributes to enhancing the living standards of minorities and reduces the ethnic gap through providing more equalized levels of commune endowments As such, Government programs, such as Program 135, appear to be justified for this purpose It is worth investigating whether and to what extent any of this is true NOTES The language barrier here refers to the inability to speak Vietnamese According to Ethnologue (as quoted in World Bank (2009a)), Vietnam encompasses seven major language families but 102 distinct languages Vietnamese, the language of the Kinh, is the majority and official language Without estimating how language barriers affect ethnic inequality, Baulch et al (2010) suggest that not being able to speak Vietnamese substantially increases the minority’s likelihood of being poor For example, the work of Van de Walle and Gunewardena (2001) noted (but did not report) their attempt to use a language dummy in regressions for the minority group using the 1993 household data, and found no significant effects For cross-country comparisons, Grafton, Kompas, and Owen (2007) find that language barriers generate social barriers to communication and impede knowledge transfer and productivity In various country studies, Patrinos, Velez, and Psacharopoulos (1994) and Parker, Rubalcava, and Teruel (2005) report school inequality and language barriers for indigenous children, and Chiswick (1991), Chiswick and Miller (1995), Chiswick, Patrinos, and Hurst (2000) and Dustmann and Fabbri (2003), among others, show evidence of the importance of language skills in labor market participation and the earnings of immigrants The coefficient for being a majority household increases from 0.247 to 0.281 when the variable language is excluded from the pooled model specified in column 1, Table World Bank (2004) finds that students who always speak Vietnamese outside school or belong to the ethnic majority Kinh group are likely to have higher test scores than students who never speak Vietnamese outside school, or who belong to the ethnic minority group There might be a concern that household heterogeneity may affect educational attainment and returns to education Our use of withincommune deviations as an instrument for education does not correct for this possibility However, if this bias is systematically similar for both the majority and minority groups, noting that we are interested only in the ethnic differences in educational attainment and returns to education, its effect would be negligible We use information from income data which has a larger sample to construct the propensity to speak Vietnamese for each ethnicity which is the proportion of the households who can speak Vietnamese out of all the households belonging to that ethnicity Results using coefficients from a pooled model as the reference are available upon request 10 The contribution of the differences in returns to endowments to the ethnic gap was found to be 97% in mixed communes using household data in 1993 by Van de Walle and Gunewardena (2001, p 202); 71% in rural areas using household data in 1998 by Baulch et al (2007, p 1172); 77% in rural areas using household data in 1993 and 1998 and 82% using household data in 2004 by Hoang et al (2007, p 160) REFERENCES Audretsch, D., & Feldman, M (1996) R&D Spillovers and the geography of innovation and production The American Economic Review, 86(3), 630–640 Baltagi, B (2005) Econometric analysis of panel data New York: Wiley Baulch, B., Chuyen, T T K., Haughton, D., & Haughton, J (2007) Ethnic minority development in Vietnam The Journal of Development Studies, 43(7), 1151–1176 Baulch, B., Nguyen, H T M., Nguyen, P T T., Pham, H T (2010) Ethnic minority poverty in Vietnam UK Chronic Poverty Research Centre Working Paper No 169 Baulch, B., Pham, H T., & Reilly, B (2012) Decomposing the ethnic gap in rural Vietnam, 1993–2004 Oxford Development Studies, 40(1), 87–117 Blinder, A (1973) Wage discrimination: Reduced form and structural estimates Journal of Human Resources, 8, 436–455 Breusch, T., Ward, M., Nguyen, H T M., & Kompas, T (2011) On the fixed-effects vector decomposition Political Analysis, 19(2), 123–134 Breusch, T., Ward, M., Nguyen, H T M., & Kompas, T (2011) FEVD: Just IV or just mistaken Political Analysis, 19(2), 165–169 Bui, T (1999) 54 Dan Toc o Vietnam va Cac Ten Goi Khac (54 ethnic groups in Vietnam and other names) Thanh Nien Castel, P (2009) Vietnam health insurance: Use of health care services by the poor: efficiency and equity issues in the province of Kon Tum SSRN eLibrary Che, T N., Kompas, T., & Vousden, N (2006) Market reform, incentives and economic development in Vietnamese rice production Comparative Economic Studies, 48(2), 277–301 Chiswick, B (1991) Speaking, reading, and earnings among low-skilled immigrants Journal of Labor Economics, 13(2), 149–170 Chiswick, B., & Miller, P (1995) The endogeneity between language and earnings: International analyses Journal of Labor Economics, 9(2), 246–288 Chiswick, B., Patrinos, H., & Hurst, M (2000) Indigenous language skills and the labor market in a developing economy: Bolivia Economic Development and Cultural Change, 48(2), 349–367 Deaton, A (1997) The analysis of household surveys: a microeconometric approach to development policy Johns Hopkins University Press Deaton, A (1980) Economics and consumer behavior Cambridge University Press Dustmann, C., & Fabbri, F (2003) Language proficiency and labour market performance of immigrants in the UK The Economic Journal, 113(489), 695–717 Please cite this article in press as: Nguyen, H -T M et al Language, Mixed Communes, and Infrastructure: Sources of Inequality and Ethnic Minorities in Vietnam, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.03.004 18 WORLD DEVELOPMENT Easterly, W., & Levine, R (1997) Africa’s growth tragedy: Policies and ethnic divisions The Quarterly Journal of Economics, 112(4), 1203–1250 Edmonds, E., & Pavcnik, N (2005) The effect of trade liberalization on child labor Journal of International Economics, 65(2), 401–419 Fortin, N., Lemieux, T., & Firpo, S (2011) Decomposition methods in economic In Handbook of labor economics (4, pp 1–102) General Statistic Office (2000) Statistical yearbook of Vietnam 1999 Hanoi: Statistical Publishing House General Statistic Office (2009) Statistical yearbook of Vietnam 2008 Hanoi: Statistical Publishing House Grafton, Q., Kompas, T., & Owen, P (2007) Bridging the barriers: Knowledge connections, productivity and capital accumulation Journal of Productivity Analysis, 28(3), 219–231 Hausman, J., & Taylor, W (1981) Panel data and unobservable individual effects Econometrica, 49(6), 1377–1398 Hoang, H., Pham, G., Tran, M., & Hansen, H (2007) Ethnicity and poverty reduction In T Nguyen, & H Hansen (Eds.), Market policy and poverty reduction in Vietnam National Political Publisher Hsiao, C (2003) Analysis of panel data Cambridge University Press Huynh, T B., Duong, B H., & Bui, T C (2002) Indigenous peoples/ethnic minorities and poverty reduction Asian Development Bank Jamieson, N (1996) Ethnic minorities in Viet Nam: A country profile Hanoi, mimeo: Hanoi and World Bank Resident Mission Jann, B (2008) A Stata implementation of the Blinder–Oaxaca decomposition Stata Journal, 8(4), 453–479 Khandker, S R., Barnes, D F., & Samad, H A (2014) Welfare impacts of rural electrification: A panel data analysis from Vietnam Economic Development and Cultural Changes, 61(3), 659–692 Kompas, T., Che, N T., Nguyen, H T M., & Nguyen, H Q (2012) Productivity, net returns and efficiency: Land and market reform in Vietnamese rice production Land Economics, 88(3), 478–495 Lazear, E., & Michael, R (1988) Allocation of income within the household University of Chicago Press Le, T S (2006) Tac dong cua cac chuong trinh phat trien doi voi van hoa toc nguoi: Nghien cuu truong hop nguoi HMong tai huyen Sapa, tinh Lao Cai [Impacts of Development Programs on Ethnic Culture: Research on the Case of Hmong People in Sapa District, Lao Cai Province] (MA thesis) Department of History, Hanoi National University Lipton, M., & Ravallion, M (1995) Poverty and policy Handbook of Development Economics, 3, 2551–2657 Luu, V S (2010) Mot so diem nong chinh tri xa hoi dien hinh tai cac vung da dan toc o mien nui nhung nam gan day: Hien trang, van de, cac bai hoc kinh nghiem xu ly tinh huong (Some critical socio-political areas in mountainous multi-ethnic regions in recent years: situation, issues and lessons learnt in handling situations) Nha xuat ban chinh tri quoc gia Ma˚lqvist, M., Hoa, D T P., Liem, N T., Thorson, A., & Thomsen, S (2013) Ethnic minority health in Vietnam: A review exposing horizontal inequity Global health action (19803) Co-Action Publishing Nguyen, T T P & Baulch, B (2007) A review of ethnic minority policies and programs in Vietnam A product of the IDS-UoS-CAF Project on Ethnic Minority Development in Vietnam financed by the UK Economic and Social Research Council (RES-167-25-0157) and the Department for International Development Nguyen, C., Phung, T., Phung, T., Vu, N., & Westbrook, D (2013) The impact of a national poverty reduction program on ethnic minorities in Vietnam: The lens of baseline and end line surveys MPRA Paper No 50477 Nguyen, N N & Rama, M (2007) A comparison of quantitative and qualitative poverty targeting methods in Vietnam Q-Squared Working Paper No 32 Oaxaca, R (1973) Male–female wage differentials in urban labor markets International Economic Review, 14(3), 693–709 Oaxaca, Ronald L., & Ransom, Michael R (1994) On discrimination and the decomposition of wage differentials Journal of econometrics, 61(1), 5–21 Parker, S., Rubalcava, L., & Teruel, G (2005) Schooling inequality and language barriers Economic Development and Cultural Change, 54(1), 71–94 Patrinos, H., Velez, E., & Psacharopoulos, G (1994) Language, education, and earnings in Asuncion, Paraguay The Journal of Developing Areas, 29(1), 57–68 Pham, T H., Le, D T., & Nguyen, V C (2011) Poverty of ethnic minorities in Viet Nam: Situation and challenges in programme 135 phase II communes, 2006–07 Indochina Research & Consulting Rambo, A T (1997) 2011 Development trends in Viet Nam’s northern mountain region In D Donovan, A T Rambo, J Fox, L T Cuc, & T D Vien (Eds.) Development trends in Viet Nam’s northern mountain region An Overview and Analysis (Vol 1, pp 5–52) Hanoi: National Political Publishing House Rheinlander, T., Samuelsen, H., Dalsgaard, A., & Konradsen, F (2011) Perspectives on child diarrhoea management and health service use among ethnic minority caregivers in Vietnam BMC Public Health, 11 (1), 690 Stock, J., & Yogo, M (2005) Testing for weak instruments in linear IV regression In D Andrews, & J Stock (Eds.), Identification and inference for econometric models: Essays in Honor of Thomas Rothenberg Cambridge University Press Swinkels, R & Turk, C (2006) Explaining ethnic minority poverty in Vietnam: a summary of recent trends and current challenges Draft Background Paper for Committee of Ethnic Minorities/Ministry of Planning and Investment Meeting on Ethnic Minority Report, 28 September 2006 Toan, N V., Trong, L N., Hojer, B., & Persson, L A (2002) Public health services use in a mountainous area, Vietnam: Implications for health policy Scand J Public Health, 30, 86–93 Tran, T T T (2004) Vietnam’s rural transformation: Information, knowledge and diversification In D McCargo (Ed.), Rethinking Viet Nam London: Routledge Curzon Van de Walle, D., & Gunewardena, D (2001) Sources of ethnic inequality in Vietnam Journal of Development Economics, 65(1), 177–207 Vasavakul, T (2003) Language policy and ethnic relations in Vietnam In E Brown, & S Ganguly (Eds.), Fighting Words: Language Policy and Ethnic Relations in Asia Cambridge: MIT Press Vietnam Academy of Social Sciences (2009) Participatory Poverty Assessment 2008: Synthesis Report Hanoi: The Gioi Publisher Wells-Dang, A (2012) Ethnic minority development in Vietnam: What leads to success? Background paper for the 2012 Programmatic Poverty Assessment World Bank (2004) Vietnam reading and mathematics assessment study World Bank human development sector report Washington, DC: World Bank World Bank (2005) Governance Joint donor report to the vietnam consultative group meeting Hanoi: World Bank World Bank (2007) Social protection Joint donor report to the vietnam consultative group meeting Hanoi: World Bank World Bank (2009a) Country social analysis: Ethnicity and development in Vietnam Washington, DC: World Bank World Bank (2009b) Modern institutions Joint donor report to the vietnam consultative group meeting Hanoi: World Bank Available online at www.sciencedirect.com ScienceDirect Please cite this article in press as: Nguyen, H -T M et al Language, Mixed Communes, and Infrastructure: Sources of Inequality and Ethnic Minorities in Vietnam, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.03.004 ... coastal and delta land, hilly land, Please cite this article in press as: Nguyen, H -T M et al Language, Mixed Communes, and Infrastructure: Sources of Inequality and Ethnic Minorities in Vietnam, ... year of schooling between majority and Please cite this article in press as: Nguyen, H -T M et al Language, Mixed Communes, and Infrastructure: Sources of Inequality and Ethnic Minorities in Vietnam, ... of the dominant role of the difference Please cite this article in press as: Nguyen, H -T M et al Language, Mixed Communes, and Infrastructure: Sources of Inequality and Ethnic Minorities in Vietnam,