2.1 Natural resource extraction

Natural resource extraction refers to the harvesting of natural products from non-cultivated sources, such as wood from natural forests or fish from lakes (Angelsen et al., Reference Angelsen, Jagger, Babigumira, Belcher, Hogarth, Bauch, Börner, Smith-Hall and Wunder2014). In this study, environmental income is defined as the difference between total extraction output at market price and extraction cost. Hence, it is the net revenue from extraction. According to previous research (Córdova et al., Reference Córdova, Wunder, Smith-Hall and Börner2013; López-Feldman, Reference López-Feldman2014; Wunder et al., Reference Wunder, Noack and Angelsen2018), households in resource-rich areas are particularly poor. They also spend more effort on extraction and have fewer alternatives in times of crisis. Dependence on natural resources is further exacerbated by the fact that these households use environmental products for subsistence rather than for sale (Nerfa et al., Reference Nerfa, Rhemtulla and Zerriffi2020). Thus, environmental income in our study includes both – consumed and sold environmental products.

Although households that extract natural resources are more likely to be income poor, empirical research suggests that natural resource extraction can reduce poverty and income inequality (López-Feldman et al., Reference López-Feldman, Mora and Taylor2007; Fonta and Ayuk, Reference Fonta and Ayuk2013; Chhetri et al., Reference Chhetri, Larsen and Smith-Hall2015). Without environmental income, some households may even be at risk of not meeting their basic needs and slipping into deeper poverty (López-Feldman et al., Reference López-Feldman, Mora and Taylor2007). Using Mexico as an example, López-Feldman et al. (Reference López-Feldman, Mora and Taylor2007) show that environmental income from non-timber forest products (NTFP) reduces poverty and inequality. Since not all households participate in extraction, environmental income is unequally distributed. Nevertheless, it has an equalizing effect on overall income inequality as it favors the poor. Using data from rural Nepal, Chhetri et al. (Reference Chhetri, Larsen and Smith-Hall2015) find that, unlike other sources of income such as remittances, environmental income reduces income inequality.

Nguyen et al. (Reference Nguyen, Do and Grote2018a) examine the differences between extracting and non-extracting households in the welfare impacts of extraction in rural Laos. They show that if extracting households ceased their extraction activity, they would experience a reduction in household income, consumption and food security. However, non-extracting households that are forced to extract would also have higher levels of food security from extraction, but lower household income and consumption due to the loss of income and consumption from their current activities. Although these results provide valuable insights, they cannot be extrapolated to other regions. In Vietnam, the rural population is characterized by its distinction between ethnic majority and minority (Dang, Reference Dang, Hall and Patrinos2012). It is likely that there are also ethnic differences in natural resource extraction and in the welfare impacts of extraction, driven by differences in socio-economic characteristics and the returns to these characteristics. This is the topic of the following subsection.

2.2 Ethnic minorities in Vietnam

Some ethnic minorities have lived in Vietnam for centuries, such as the Tay-Thai groups. Others have been in the country only since the last millennium, such as the Nung (Dang, Reference Dang, Hall and Patrinos2012). The 53 ethnic minorities differ in their culture, language and religion. Some of the groups are so small that they consist of only a few hundred people. The largest groups are the Tay, Thai, Muong, Khmer, and Hmong. The ethnic minority groups also differ in their socio-economic status. While some are better able to lift themselves out of poverty, others – particularly those based in the Central Highlands – still have very high rates of multidimensional poverty (United Nations Development Programme, 2018).

Previous research has identified three main factors driving ethnic poverty (Baulch et al., Reference Baulch, Chuyen, Haughton and Haughton2007): objective factors (e.g., infrastructure, climate), subjective factors (e.g., education, capacities), and institutional factors (e.g., politics, poor targeting). An important aspect is the remoteness of ethnic minorities – not only in terms of geographical location, but also in terms of language and cultural barriers. Van de Walle and Gunewardena (Reference Van de Walle and Gunewardena2001) point out that inequality is the result of different characteristics as well as different returns to these characteristics. Unequal returns to productive characteristics such as education can be interpreted as a structural component of inequality, which may include, for example, past discrimination or different cultural backgrounds. Ethnic inequality affects many areas of daily life. Adults belonging to the ethnic majority are better educated and have higher literacy rates (Dang, Reference Dang, Hall and Patrinos2012). At the same time, school enrolment rates of minorities remain lower. One of the reasons is that they have to travel longer distances to reach primary and secondary schools. This problem is exacerbated by an age bias among school children due to later enrollment, class repetition or school dropout (Dang, Reference Dang, Hall and Patrinos2012; Nguyen, Reference Nguyen2019). Furthermore, Vo et al. (Reference Vo, Van, Tran, Vu and Ho2021) report an ethnic wage gap. Ethnic minorities are also more dependent on agriculture and natural resources to make a living (Tuyen, Reference Tuyen2016). The majority, on the other hand, is more involved in off-farm activities. Nguyen (Reference Nguyen2017) emphasizes the lower health insurance coverage of poor people and the lower quality of health care in remote areas. Minorities living in these areas are particularly affected. In addition, ethnic minorities face difficulties in accessing formal credit (Nguyen et al., Reference Nguyen, Nguyen and Grote2020).

Overall, ethnic minorities are forced to rely on natural resources. However, minorities who have lived in the remote mountainous areas of Central Vietnam for centuries have adapted to these living conditions (World Bank, 2009; Tan et al., Reference Tan, Van Chuong, Linh, Tung, Dinh and Tuyet2023). They have built up a unique body of knowledge that is well-adjusted to the environment. It includes expertise in agroforestry, traditional medicine as well as resource and natural disaster management. Each ethnic group has its own customs and practices that can affect the use of natural resources. Although there are interactions among ethnic groups, indigenous minorities have an advantage in knowledge compared to the ethnic majority and to immigrated minority groups. Traditional knowledge improves climate change adaption and provides a basis for self-sufficiency and self-determination for many ethnic minorities, as they are less dependent on external factors and are already familiar with indigenous practices (Tan et al., Reference Tan, Van Chuong, Linh, Tung, Dinh and Tuyet2023). Accordingly, this could be beneficial in terms of ethnic inequality and welfare.

3.1 Study site and data collection

For our analysis, we use household and village level data from the TVSEP project (TVSEP, 2016). This long-term project has collected data from six rural provinces in Thailand and Vietnam since 2007 to study income and poverty dynamics. The TVSEP data are representative of poor rural regions because of the sampling procedure which includes the following three steps (Hardeweg et al., Reference Hardeweg, Klasen, Waibel, Klasen and Waibel2013). In the first step, provinces were selected based on their low average per capita income, high dependence on agriculture, and poor infrastructure. In the second step, two villages per sub-district were chosen with probability proportional to the population size of the sub-district. In the final step, a fixed sample size of ten households per village was randomly selected from a list of households with equal probability.

For our analysis, we focus on the dataset from Dak Lak province because it is inhabited by a high proportion of ethnic minority people and includes large forest areas (World Bank, 2009; Pham et al., Reference Pham, Le, Tuyet, Pham, Tran, Tran, Tran, Nguyen and Nguyen2021). The province is located in the Central Highlands of Vietnam and close to neighboring Cambodia (figure 1). Our analysis sample includes the 2010, 2013, and 2016 panel waves. In total, we refer to 2,053 household observations from 76 villages in rural communes (appendix table A1). Villages are marked as small dots in figure 1.

Figure 1. The studied province Dak Lak in Vietnam.

Shape source: Humanitarian Data Exchange (2023).

Due to government resettlement programs, many people from the Kinh majority migrated from the lowland regions to the Central Highlands (World Bank, 2009). The government wanted to promote economic development in these rural areas, which had previously been predominantly inhabited by indigenous minorities (e.g., Ede, Mnong) (table A2). Although most migrants belong to the Kinh majority, ethnic minorities (e.g., Tay, San chai) from other parts of the country also settled in the Central Highlands. Economic opportunities for the migrants arose from financial support and increasing opportunities for coffee cultivation. Today, the economy in Dak Lak is dominated by the production of coffee, which, along with rice, is Vietnam's most important export commodity (Ho et al., Reference Ho, Hoang, Wilson and Nguyen2018; Byrareddy et al., Reference Byrareddy, Kouadio, Kath, Mushtaq, Rafiei, Scobie and Stone2020). However, the changing population composition has increased inequality in Dak Lak, particularly between the Kinh majority and ethnic minorities (World Bank, 2009). These tensions erupt in conflicts over land and other natural resources (Baulch et al., Reference Baulch, Chuyen, Haughton and Haughton2007; World Bank, 2009).

The TVSEP data contain a large amount of information on household and village characteristics. For the following analysis, we use the questionnaire sections that ask about the socio-demographic structure of households, sources of income, assets, and consumption behavior. The data collected consider the last 12 months before the date of the survey.

3.2 Econometric specifications

3.2.1 Identifying ethnic differences in the determinants of natural resource extraction

The Oaxaca-Blinder decomposition investigates mean outcome differences between groups. It was originally developed by Oaxaca (Reference Oaxaca1973) and Blinder (Reference Blinder1973). The technique is commonly employed in labor economics to study wage differentials, but can also be used for many other applications (Jann, Reference Jann2008). In our analysis, we want to examine ethnic differences in natural resource extraction. Therefore, we apply the Oaxaca-Blinder two-fold decomposition to several indicators of extraction, namely environmental income, environmental income per capita, and extraction participation. The mean values ${Y_{eth}}$ for these indicators are determined as follows:

(1)\begin{equation}{Y_{eth}} = {X_{eth}}{\alpha _{eth}} + {\varepsilon _{eth}}\;eth \in ({\textrm{Majority,}\;\textrm{Minority}} ),\end{equation}

where X is a vector containing the predictors and $\alpha$ includes the corresponding coefficients. The formula differentiates between the mean values of ethnic majority and minority households. Since the error term $\varepsilon$ is assumed to be uncorrelated with X and to have a zero mean, the differences in mean outcome can be written as:

(2)\begin{align}& {Y_{\textrm{maj}}} - {Y_{\textrm{min}}} = {X_{\textrm{maj}}}{\alpha _{\textrm{maj}}} - {X_{\textrm{min}}}{\alpha _{\textrm{min}}}\end{align}

(3)\begin{align}& \begin{array}{@{}l} {Y_{\textrm{maj}}} - {Y_{\textrm{min}}} = {\alpha _{\textrm{maj}}}({{X_{\textrm{maj}}} - {X_{\textrm{min}}}} )+ {X_{\textrm{min}}}({{\alpha_{\textrm{maj}}} - {\alpha_{\textrm{min}}}} ),\\ ({\textrm{characteristics}} )\;({\textrm{structure}} )\end{array}\end{align}

where the first component on the right-hand side of equation (3) represents the differences in characteristics weighted by the majority coefficients. The second component on the right-hand side reflects the structural differences weighted by the minority predictors. It reflects the expected change in outcome assuming majority coefficients in the characteristics component and assuming minority predictors in the structural component. Majority and minority weights in the components are interchangeable, since there can be negative discrimination against minorities or positive discrimination against the majority (Oaxaca, Reference Oaxaca1973; Cotton, Reference Cotton1988).

Natural resource extraction is influenced by a number of different predictors X (table A3). These include socio-demographic characteristics such as the average household age, average education, size, dependency ratio, and the gender of the household head. Previous studies have shown that these factors are important determinants of extraction (Angelsen et al., Reference Angelsen, Jagger, Babigumira, Belcher, Hogarth, Bauch, Börner, Smith-Hall and Wunder2014; Nguyen et al., Reference Nguyen, Do, Bühler, Hartje and Grote2015, Reference Nguyen, Do and Grote2018a) and a source of ethnic inequality (Dang, Reference Dang, Hall and Patrinos2012; Nguyen, Reference Nguyen2017, Reference Nguyen2019; Vo et al., Reference Vo, Van, Tran, Vu and Ho2021). We use averages to better represent the diversity of household members. Furthermore, a large study from Sunderland et al. (Reference Sunderland, Achdiawan, Angelsen, Babigumira, Ickowitz, Paumgarten, Reyes-García and Shively2014) indicates that the responsibility for extraction is not limited to the household head. Asset valueFootnote ¹ includes the household's long-term accumulated wealth, which enables investment in extraction (Nerfa et al., Reference Nerfa, Rhemtulla and Zerriffi2020), but also reveals past discrimination (Van de Walle and Gunewardena, Reference Van de Walle and Gunewardena2001). Farmland size reflects the household's involvement in its own farming. Agriculture and natural resource extraction have been shown to be closely linked (Nguyen et al., Reference Nguyen, Do, Parvathi, Wossink and Grote2018b). The distance between the household's home to the extracting ground relates to the effort required to carry out extraction activities (Kabubo-Mariara, Reference Kabubo-Mariara2013; Nguyen et al., Reference Nguyen, Do and Grote2018a). Off-farm employment is considered as a non-farm component. The calculation of variance inflation factors (VIF) indicates that there is no multicollinearity between independent variables (table A4). Robust standard errors are bootstrapped with 1,000 replications and clustered at the village level.

3.2.2 Identifying ethnic differences in the impact of natural resource extraction on household consumption

Endogenous switching regressions estimate the impact on an outcome variable depending on two different states. To account for endogeneity and selection bias, this technique has been used in several studies to study consumption impacts (Ahmed and Mesfin, Reference Ahmed and Mesfin2017; Jaleta et al., Reference Jaleta, Kassie, Marenya, Yirga and Erenstein2018; Nguyen et al., Reference Nguyen, Do and Grote2018a). For instance, Ahmed and Mesfin (Reference Ahmed and Mesfin2017) find that membership in agricultural cooperatives improves consumption among smallholder farmers in Ethiopia. However, only Nguyen et al. (Reference Nguyen, Do and Grote2018a) investigate the impact of extraction on consumption and find that extraction contributes to household consumption in rural Laos. Our analysis further identifies ethnic differences in household consumption based on the decision to extract. It is a two-stage procedure with a selection and an outcome model. Selection is modeled as follows:

(4)\begin{align} I_j^\ast& = {Z_j}\beta + {\upsilon _j} {I_j} = 1\ \textrm{if }\ I_j^\ast > 0\nonumber\\ {I_j} & = 0\textrm{ otherwise}, \end{align}

where $I_j^\ast$ is a latent variable. Since households' extraction preferences depend on factors which are not fully observable, such as personal attitudes, $I_j^\ast$ approaches the realized extraction decision ${I_j}$ of household j. A household j decides to extract if $I_j^\ast > 0$, and not to extract otherwise. In order to improve the identification, a selection instrument ${Z_j}$ is included. $\beta$ contains the corresponding parameters. The error term is given as ${\upsilon _j}$.

In the second stage, we estimate the impact of the extraction status on household consumption. The outcome model is defined as follows:

(5a)\begin{align}& \textrm{Extraction}:{c_{1j}} = {Q_{1j}}{\gamma _1} + {\eta _{1j}}\ \textrm{if }\ {I_j} = 1\end{align}

(5b)\begin{align}& \textrm{No extraction}:{c_{2j}} = {Q_{2j}}{\gamma _2} + {\eta _{2j}}\ \textrm{if }\ {I_j} = 0,\end{align}

where ${c_{1j}}$ and ${c_{2j}}$ denote annual per capita household consumption, including food and non-food consumption, in the situation of extraction $({I_j} = 1)$ and no extraction $({I_j} = 0)$. ${Q_{1j}}$ and ${Q_{2j}}$ are the independent variables with ${\gamma _1}$ and ${\gamma _2}$ as the corresponding parameters. ${\eta _{1j}}$ and ${\eta _{2j}}$ capture the error terms. As we have already shown in equation (4), the extraction decision is endogenous: the selection into extraction and non-extraction is influenced by many observable and unobservable factors which simultaneously impact the level of consumption. This selection bias leads to erroneous estimates. Endogeneity can also arise from a reverse causal relationship between the extraction decision and consumption, i.e., not only does extraction affect consumption, but consumption can also influence the decision to extract. To solve the endogeneity problem, we need a selection instrument ${Z_j}$ (Wooldridge, Reference Wooldridge2010). Following Di Falco et al. (Reference Di Falco, Veronesi and Yesuf2011), this instrument is valid if it affects the decision to extract but not the level of household consumption. This exclusion restriction is satisfied by the distance from the household's home to the extracting ground (table A5). Since some households report no extraction, we imputed the variable for these households by replacing the missing values with the average distance at the village level. Households that are further away from the extracting ground have higher opportunity costs and are therefore less likely to extract (Kabubo-Mariara, Reference Kabubo-Mariara2013; Nguyen et al., Reference Nguyen, Do and Grote2018a).

We approximate the endogenous switching regressions with a full information maximum likelihood estimation, which is more efficient than two-stage least squares or maximum likelihood estimation (Lokshin and Sajaia, Reference Lokshin and Sajaia2004; Di Falco et al., Reference Di Falco, Veronesi and Yesuf2011). Standard errors are consistent because the model's binary and continuous equations are computed simultaneously. Furthermore, standard errors are bootstrapped with 1,000 replications and clustered at the village level.

After estimating the model parameters, we compare the expected consumption (6a) of households that decide to extract and (6b) of households that decide not to extract. At the same time, we estimate the counterfactual cases, i.e., the expected consumption (6c) of extracting households that do not extract and (6d) of non-extracting households that extract. These conditional expectations for household consumption are as follows:

(6a)\begin{align}& E({{c_{1j}}\textrm{|}{I_j} = 1} )= {Q_{1j}}{\gamma _1} + {\sigma _{1\upsilon }}\frac{{f({{Z_j}\beta } )}}{{F({{Z_j}\beta } )}}\end{align}

(6b)\begin{align}& E({{c_{2j}}\textrm{|}{I_j} = 0} )= {Q_{2j}}{\gamma _2} - {\sigma _{2\upsilon }}\frac{{f({{Z_j}\beta } )}}{{\{{1 - F({{Z_j}\beta } )} \}}}\end{align}

(6c)\begin{align}& E({{c_{2j}}\textrm{|}{I_j} = 1} )= {Q_{1j}}{\gamma _2} + {\sigma _{2\upsilon }}\frac{{f({{Z_j}\beta } )}}{{F({{Z_j}\beta } )}}\end{align}

(6d)\begin{align}& E({{c_{1j}}\textrm{|}{I_j} = 0} )= {Q_{2j}}{\gamma _1} - {\sigma _{1\upsilon }}\frac{{f({{Z_j}\beta } )}}{{\{{1 - F({{Z_j}\beta } )} \}}},\end{align}

where ${\sigma _{1\upsilon }}$ stands for the covariance between ${\upsilon _j}$ and ${\eta _{1j}}$; ${\sigma _{2\upsilon }}$ is the covariance between ${\upsilon _j}$ and ${\eta _{2j}}$; $f({\cdot} )\;$indicates a normal density distribution function and $F({\cdot} )$ a cumulative normal distribution function.

Since equations (6a) and (6b) reflect the actual cases, whereas equations (6c) and (6d) represent the counterfactual cases, the average treatment effects on the treated (ATT) and on the untreated (ATU) are calculated as follows:

(7a)\begin{align}& ATT = E({{c_{1j}}\textrm{|}{I_j} = 1} )- \; E({{c_{2j}}\textrm{|}{I_j} = 1} )= ({{\gamma_1} - {\gamma_2}} ){Q_{1j}} + ({{\sigma_{1\upsilon }} - {\sigma_{2\upsilon }}} )\frac{{f({{Z_j}\beta } )}}{{F({{Z_j}\beta } )}}\end{align}

(7b)\begin{align}& ATU = E({{c_{1j}}\textrm{|}{I_j} = 0} )- \; E({{c_{2j}}\textrm{|}{I_j} = 0} )= ({{\gamma_1} - {\gamma_2}} ){Q_{2j}} + ({{\sigma_{1\upsilon }} - {\sigma_{2\upsilon }}} )\frac{{f({{Z_j}\beta } )}}{{\{{1 - F({{Z_j}\beta } )} \}}}.\end{align}

To get the ATT, equation (6c) has to be subtracted from equation (6a). This is the difference between the actual expectation of extracting household's consumption and the counterfactual expected consumption if extracting households did not extract. For the ATU, equation (6b) has to be subtracted from equation (6d). Equally, this is the difference between the actual expectation of non-extracting household's consumption and the counterfactual expected consumption if non-extracting households did extract.

3.3 Descriptive statistics

Within the total sample, 1,292 household observations (63 per cent of the sample) belong to the ethnic majority, while 761 observations (37 per cent) belong to one of the minority groups (table A6). This ratio is similar to the results of a 2017 census in Dak Lak: the majority accounts for 67 per cent of the provincial population and, accordingly, 33 per cent belong to the minority groups (Ministry of Natural Resources and Environment Vietnam, 2019). It appears that minority households are significantly younger than majority households. They also have lower average education levels. Minority households are larger and have a higher dependency ratio, meaning there are fewer household members caring for dependent members. Minority households also have fewer valuable assets than majority households. In addition, they are less likely to be employed off-farm. This suggests that minority households may need to rely more heavily on other livelihood strategies such as farming and natural resource extraction. Therefore, they cultivate larger areas of farmland. Descriptive statistics further reveal that the majority households live closer to the extracting ground. Since they have migrated to Dak Lak, they have settled in the periphery of the villages, closer to the forest and water areas.

With respect to indicators on natural resource extraction, ethnic minority households have significantly higher environmental income for the household as a whole and per capita (table A7). Moreover, minorities are more likely to participate in extraction. Figure 2 shows the sum of environmental income by ethnic status and extracted products. Firewood and timber products are by far the most important environmental products (Nguyen et al., Reference Nguyen, Do and Grote2018a). The category mainly includes firewood, as timber harvesting is highly regulated (Pham et al., Reference Pham, Le, Tuyet, Pham, Tran, Tran, Tran, Nguyen and Nguyen2021). Animals (e.g., fish, small amphibians), fruits and vegetables, as well as other products (e.g., honey, mushrooms) are extracted less and are of lower value (Bierkamp et al., Reference Bierkamp, Nguyen and Grote2021).

Figure 2. Sum of environmental income from all household observations in 2010, 2013, and 2016 by ethnic status and extracted products.