Investigating the “Wealth Paradox” on Child Labor:
A Case Study of rural areas in Vietnam
Preliminary version
Vu Minh Hien
1CIFREM, University of Trento May 2012
Abstract
The study aims to investigate the effect of farm land on child labor along with consumption expenditure. The paper uses data in rural areas and for children from 10 to 14 years old of the Vietnam household living standard survey in 2008. The specifications are constructed in the three comparable models: Tobit, Heckman, and double-hurdle to identify the better econometric approach. The results support the hypothesis that child labor increases in land-rich but among poor income households, then decreases among non-poor income households. The findings add another factor to challenge the presumption that child labor appears in poorest households.
Key words: Wealth paradox, child labor, poverty, farm land, rural areas, Vietnam.
1 I am grateful for comments of Prof. Christopher L. Gilbert and Prof. Gabriella Berloffa.
1. Introduction
During the past two decades, studies on child labor have been growing in both theoretical and empirical research. The main issues are concentrated on addressing the three relevant areas including the causes, consequences and policies of child labor. In particular, there are five major trends in analyzing the cause of child labor: poverty, credit market imperfections, land and labor market imperfections, parental characteristics, and macroeconomic factors (Fors, 2010).
Understanding child labor plays an important role in designing the appropriate policies for governing and eliminating the incidence of child labor. This paper aims to focus on the first issue which explores the influence of factors inducing child labor. Among these factors, poverty has been traced widely and found to be the main cause of child labor. The effect of poverty
correlated with imperfect markets, household characteristics, and other external factors which generate a complicated mechanism affecting child labor. Mostly findings come out with positive relationship between poverty and child labor. However, empirical evidence has shown mixed results since there are differences not only of sample characteristics but also of definitions or methodologies. Besides, the decision to employ children and household incomes are joint outcomes of a single decision making process that creates the causal relationship between poverty and child labor. Other characteristics of household might also affect household income that lead to problem of collinearity and omitted variables (Edmonds & Schady, 2010). .
According to ILO statistics, in the age group 5 to 14 years, the total child population in economic activity was estimated at 176 millions, which accounts for 14.5 percent in 2008 (Diallo,
Hagemann, Etienne, Gurbuzer, & Mehran, 2010). Categorized into four groups of economic activity: agriculture, industry, services, and others, the data shows that agriculture is the largest sector of all child labor in the 5 to 17 years old group at 60 percent. The services, industry sectors follow with 26 percent, 7 percent, and the rest is not defined group at 7.5 percent.
Meanwhile, by employment status, two thirds of child labor are unpaid family workers (64 percent for boys versus 73 percent for girls). Paid employment and self-employment account respectively for 21 and 5 percent of all child laborers in the same age group. Hence, the data implicitly presents the fact that the majority of child labor gets involved in agricultural work which is managed by their families. Land is the main source of wealth in agricultural sector and used by households in different forms of ownership. In this context, landholding has been taken into consideration as another factor that causes child labor. Land has a relationship to household income, so the hypothesis that child labor is found in the poorest households is worth revisiting.
Regarding the concept of child labor, there is no consistent definition in the literature. Based on the ILO classification, working children include three types: children in employment, child laborers, and children in hazardous work. The categories of children in employment is the broadest range, containing children in all types of paid and non paid activities. The category of child laborers is more restrictive, excluding children over 14 years old in hazardous work and including children over 11 engaging in light work (i.e. work that does not influence children’s ability to attend school and vocational training). The last one is children involved in work that might harm their safety, moral development, and physical and mental health (Diallo et al., 2010).
2. Literature review
Considered as the main cause of child labor, poverty have been substantially focused in the majority of the studies comparing to other determinants. The fundamental framework of Basu and Van (1998) has emerged to catch the significant attention in economic literature of child labor. The authors propose the assumption that subsistence poverty causes child labor. Basu (2000) extends the model to analyze the effect of a minimum adult wage on child labor. The study demonstrates that child labor incidence may fall or rise as the adult wage is raised by minimum law. Numerous empirical studies show the relationship between living standards and child labor. Krueger (1996) illustrates obvious trend of cross-country sample that employment of young children is common in low-income countries and uncommon in high-income countries.
Kambhampati and Rajan (2005) find the relation in India that an increase in parents’ wage reduces child labor, but the effects of the mother’s wage and father’s wage are different.
Beegle, Dehejia and Gatti (2006) view crop shocks are significant relative to household wealth and find a strong effect of an unexpectedly poor harvest on an increase in the level of child labor.
Duryea, Lam, and Levison (2007) explore that the unemployment shock of household male head increases the probability that a child enters the labor force in urban Brazil.
The market failures which have a close correlation with poverty might be a transmission effect to child labor. In aspect of theoretical framework, Ranjan (1999) shows how poverty in
combination with credit constraints can increase child labor in developing countries. The author adapts the two-period model in which a household decides to send their children to work or school in the first period. In the cases of no credit market, there is a threshold income above which households will send children to school and below which households will send children to work. In the cases of access to the international credit market, all households decide to send children to school as long as the return on education is higher than the interest rate on the market. Baland and Robinson (2000) develop a two-period model to present the existence of child labor equilibrium despite the fact of socially inefficiency. When parents use child labor as a substitute for transferring income from children to parents, it can be Pareto inefficient. Child labor is inefficient when household is poor that parents could not leave bequests to their children, or when bequests are positive and transferring income from the future to the present is stopped in an imperfect credit market. In addition, empirical studies also demonstrate the significant link of an imperfect credit market with child labor. Dehejia and Gatti (2002), Beegle et al. (2003) find a negative effect in the extent of access to the credit market on child labor through cross-country sample and in Tanzania, respectively. Edmonds (2006) finds support for importance of liquidity constraints for child work in economic activities. The sample is examined in South Africa if the elderly men receive full pension. Guarcello et al. (2010) analyze the data in Guatemala to show the substantial impact of credit constraints on lowering school enrolment and increasing child labor.
Theoretical and empirical studies show plentiful mechanisms of poverty causing child labor.
The effect has been found prevalently to be positive. However, there are studies that present an ambiguous or negative effect. Bhalotra and Heady (2003) point out the fact that ratio of working child is higher and schooling child is lower in land-rich households than land-poor households.
They refer to this as wealth paradox and challenge the assumption that poverty is the primarily cause of child labor. Land is considered a productive asset that compels wealth and substitute effects on child labor. The household with larger farm land tends to create more income resulting in wealth effect, while the assumptions of imperfect labor and land market, and
imperfect credit market lead to substitute effect. In the case of an imperfect capital market, the richer landholding households have more access to credit for needs of financing child enrolment when they offer land as collateral. In the case of an imperfect labor (or land) market, households are unable to hire laborers on the market. This generates the incentive to employ their children.
Another characteristic of the labor market in rural areas is that agricultural work relies on certain periods and location. For instance, land holders might lack labors in harvesting season.
Although the households could find laborers and send children to school, they may face the problem of moral hazard (Foster & Rosenzweig, 1994). Consequently, the effect of land is dynamic on child labor and depends on which dominates, the wealth effect or the substitute effect. Bhalotra and Heady model the effects and obtain estimates from data in rural areas in Pakistan and Ghana. The findings show that farm size has positive effect on girl labor in both countries, but disappears for boys conditional on other covariates. Supporting evidence has also been found in rural India (Congdon Fors, 2007) and Burkina Faso (Dumas, 2007).
Bar and Basu (2009) use the overlapping generation model to investigate the effect of
landholding on child labor. Under the assumption of imperfect labor market, the authors reveal that the landholding has a positive effect on child labor in the short run but negative in the long run. Yet, in the short run, there exists a critical level below which landholding will cause a decrease in child labor. Basu, Das, and Dutta (2010) propose a similar model that exhibits the invert-U curve relationship between land size and child labor. The framework has been
supported by the significant relationship of the data set in Northern India. The existence of the turning point shows that land size increases child hours worked if under such a level of land size, and then decline if land size continues to increase above that level.
Taking the case study of Vietnam, the child labor has been documented in relation to living standards, enrolment decision, and globalization after its transition. Edmonds (2001) use the nonparametric decomposition to document panel data of the national surveys in 1992-93 and 1997-98. The result show that an increase in expenditure can explain 80 percent of the decline in child labor in the households move out of poverty. Edmonds and Turk (2002) further explore the significant decrease of child labor and the heterogeneity across households in the incidence of participation, level of participation, and categories of works by both nonparametric and
parametric methods. The study finds a strong correlation between living standard improvement and a decline in child labor. The levels of reduction in categories of business household work and outside household work are higher than farm household work, traditional work. The ethnic minority and recent migrants children have experienced less decrease. The paper suggest that the progress of child labor reduction should be enclosed in the anti-poverty program.
Rosati and Tzannatos (2006) analyze the decision between child labor and school attendance, and determinants of child labor also in these two surveys. The study reports the non linear relationship of income to child labor supply by using altruistic model and applying multinominal logit equation. Income has a negative effect on child labor supply and the magnitude of impact declines when income increases. The impact of income tends to diminish in the richer
households. With the same sample, Edmonds and Pavcnik (2002) show evidence that Vietnam’s efforts to become a significant player in global rice markets are linked directly to a decrease in child labor. The increase of rice price contributes for nearly half of the decline in child labor during that period. The decline is found largest in group secondary school girls’ age and the increase school enrolment also occurs for this group.
The recent report of a interagency project about child labor has documented the overview of child labor in Vietnam from 1992 until 2006: the extent of child labor and schooling, its determinants, and its consequences on health and education (UCW Project, 2009). The report selects the survey in 2006 and applies bivariate probit model to identify the determinants of child labor and schooling. The estimated results show the expected and significant effects of explanatory variables including: children age, household income, education of household head, place of residence, access to water, labor market demand and supply, household perceptions of life quality, school quality.
3. Descriptive data
After the economic transition in1986, Vietnam conducted two important household surveys:
Multi-Purpose Household Survey (MPHS) and Vietnam Living Standard Survey (VLSS) during the 1990s. The two surveys are overlapped that motivated General Statistics of Vietnam to merge MPHS and VLSS to become a new Vietnam Household Living Standard Survey (VHLSS). The VHLSS has been implemented biennially during the period 2000-2010 with technical
assistant from UNDP and the World Bank. This paper uses the 2008 VHLSS, which contains more information about land. There are two modules of questionnaire: short household questionnaire (including income information); and long household questionnaire (including income and consumption expenditure information) in all rounds of the VHLSS. In order to get data of consumption expenditure in detail, we use the long questionnaire.
The survey is sampled by three stages: communes/wards at the first stage, census enumeration areas (EAs) at the second stage, and households at the third stage. At the first stage, the sample is selected from the master frame designed for the VHLSSs in the period 2000-2010 which
included 3,063 communes/wards from 1999 Population Census. At the second stage, wards and communes are partitioned into EA and three EAs in communes/wards selected. Only one EA constitutes for each round of survey and the two others are for the sequential rotated rounds. At the third stage, a sample of households is selected systematically with 20 households in rural EAs and 10 households in urban EAs. This is technically a three-stage design (including the selection of households), but it is operationally equivalent to a two-stage design as only one EA is selected within each commune for a specific survey. The sample is rotated 50 percent from one round to the next round of the VHLSS based on the units construct the master sample.
Therefore, 50 percent areas of the 2006 VHLSS are retained, and the other 50 areas percent are newly selected from the sample frame, which were not in 2004 and 2006 VHLSS (General Statistics Office of Vietnam, 2008).
3.1. Basic information of the survey
The communes/wards and EAs are randomly selected with probability proportional to size.
Based on the ratio of population between urban and rural, the sample size in rural areas is around three times larger than in urban areas. Out of total 9,189 households, the urban and rural
households account for 26 and 74 percent, correspondingly. Household size of total sample is 4.16 members and it is a slightly larger in rural areas than urban areas. While the education level of the head of household shows a noticeable gap, 2.69 in urban areas and 1.85 in rural areas. The education level is assigned by numbers such as: 0 is without education, 1 is elementary school, 2 is secondary school, 3 is high school, etc. Another suspicious factor is female headship that could affect the decision on household expenditure and children’s investment (Joshi, 2004).
Data in table 1 presents more significant disparities between urban and rural areas. The per capita income at current price is nearly double between urban and rural. The gap is slightly lower in per capita total expenditure and living expenditure. The living expenditure includes self-sufficient food and non-food such as: house, clothes, health care, education, recreation, transportation, etc;
while total expenditure adds other expenses such as: gift, contribution, etc. Hence, these numbers are reasonable since the per capita living expense contains basic expenditure and appears to be the least unequal.
The poverty rate is measured by percentage of people having average income or expenditure per capita under the poverty line. There are two methods to compute poverty in Vietnam. The first method based on international standards was developed by the General Statistics Office (GSO) with support from World Bank. The common poverty line is used for both areas and has changed over rounds of the survey. The second method replies on the poverty line of the Ministry of Labor, Invalids and Social Affairs, which is set in period 2006-2010 for urban and rural areas.
The first poverty line is higher resulting in higher poverty rate, 14.5 and 13.4 in the whole country. Due to the different standards between urban and rural areas of the second poverty line, the gap between the two areas is lower.
As the leading agricultural exporter in several products in Vietnam, the labor force and
production in this sector are prevalent, particularly in rural area. The data shows that 88 percent of households involve in agricultural, forestry and aquacultural production activities in rural areas. This rate is 28 percent in urban areas. These farming activities heavily rely on land, which is a basic factor of farming production. Therefore, almost farming households use land for their production activities. This is shown that the percentage of households managing land in farming activities is nearly the same as the percentage of farming households.
In the survey, approximately 40 percent of households used more than one piece of land. Pieces of land were distinguished under the management of the household for cultivation at the time of sampling. These pieces were assigned in 8 categories: annual crops, perennial crops, forestry, water surface, grass field, resident, shifting cultivation and others. The annual crop land is used for growing of plants that have the time period from cultivating to harvesting not exceeded one year. The perennial crop land is used for cultivation of plants with growing cycle more than 1 year from plating to harvesting. Forestry land is land with natural forest or planted forest that has forest standards. Water surface land is used for growing aquaculture products. Grass land is used to grow grass for cattle growing. Pond or garden adjacent to residential land is the area that lies or surrounds residential land area of household. Shifting cultivation land refers to the area that is managed by mountainous households by clearing trees or firing forestry for cultivation some crops. Other agricultural lands include areas to build glasshouse and other kinds of development purpose for farming activities. The majority of those piecies is annual crops land, which accounts as much as 74 percent, follwing by residential land and perennial land: 12 percent and 8 percent, respectively. There are 79 percent of farm households have used annual crop land.
According to the law on land amended in 2001 by goverment, the quota on allocation of agriculural land varies by diffrent types of land. The quota on allocation to each household or individual of land for planting annual crops, land for aquaculture and land for salt production shall be no more than 7.4 acres (or 3 hectares) of each type of land. The quota on allocation of land for perennial crops shall be no more than 24.7 acres (or 10 hectares) in plains and no more
than 74 acres (or 30 hectares) in midlands and in mountain regions. However, the average area of agricultural land per household in survey is 2.2 acres which reflects the small scale in production activities of households.
In the sample, the child labor under 10 years old appears in a small fraction. Thus, this paper focuses on group of children from 10 to 14 years old. The ratio between households having child workers and households having children from 10 to 14 years old is estimated at 13 percent. Out of 4,034 households, there are 338 households with one working child, and 56 households with more than one working child. As expected, this number is remarkable higher in rural areas than in urban areas. In addition, calculated child labor by individuals, out of 12 percent of working children, there are as high as 88 percent children participate in agriculture (including 83 percent work for their household), following by industry and services, calculated at around 10 percent and nearly 2 percent, respectively.
Table 1: Basic information of the survey
Characteristics Urban Rural Total
Number of household in sample 2,352 6,837 9,189
Household size 4.07 4.19 4.16
Education of head 2.69 1.85 2.18
Household w/t female head (%) 36.65 20.3 24.49
Per capita income ( 1,000 VND/month) 1,607 827 1,027
Per capita total expenditure2 1,215 646 792
Per capita living expenditure3 1,065 564 693
Poverty rate (by World Bank & GSO)4 3.3 18.7 14.5
Poverty rate (by Government)5 6.7 16.1 13.4
Households involve in farm activities (%) 27.51 87.95 72.48
Households use land in farm activities (%) 25.59 87.38 71.49
Land area used in farm-activities6 (acres) 1.62 2.28 2.22
Households with child work (from 10-14) 4.62 15.46 12.99
Incidence of child work (from 10-14)7 4.13 14.28 12.12
Agriculture 88.18
Industry 10.17
Services 1.65
Source: Calculation from the 2008 VHLSS and Statistical Yearbook of Vietnam in 2008
2 Total expenditure includes expenses on food, non-food (house, clothes, health care, education, recreation,
transportation…) other expenses (gift, contribution, etc.), excluding production cost, production tax, savings, loans, debt payment, and others.
3 Living expenditure includes expenses on food, non-food only.
4 Poverty rates have been calculated by monthly average expenditure per capita according to the General poverty lines by GSO and WB as 280 thousands VND (2008 Statistical Yearbook of Vietnam, 2009)
5 Poverty rates have been measured by monthly average income per capita according to the latest standard of the Government for the period 2006 - 2010 with different standards as follows: 260 thousands VND for urban; 200 thousands VND for rural (2008 Statistical Yearbook of Vietnam, 2009)
6 Land area is calculated on the information about managed/operated land for agriculture and forestry or water surface for aquaculture
7 Out of 3,762 children from 10 to 14 years old, 456 children participate in at least one economic activities,
including 20 children participate in more than one activity in different sectors. Only the most time-consuming work that child laborers participate in is calculated.
3.2. Land, labor, and product markets
In 1986, after the transition, the Government decentralized agricultural land from collective system and assigned to individuals and households to a period up to 15 years. In 1993, Vietnam has implemented a new land law that allows land-used rights could be transferred, exchanged, leased, mortgaged and inherited. This important reform was expected to bring benefit for land holder, particular in agricultural sector. First, the security land usage are enhanced that might affect to agricultural investment decisions. Second, the access to credit could be facilitated when land serves as collateral. Based on the two rounds of the survey VLSS 1992-93 and 1997-98, the additional land rights are found to lead an increase in long-term investment but to be irrelevant to credit access (Do & Iyer, 2004). In this sample, the raw data also reveals the similar results. The ratios of long-term land-use certificate possession in two groups, households with and without loans, are equal at around 77 percent.
Noticing a fact in Vietnam that the procedure of issuing land-use certificate to farmers is time- consuming and proceeded at unequal pace cross regions. For instance, after three years new law taken in effect, 72 percent of farmers in Mekong delta sample reported possession of their
certificate while the Red River delta was only 8 percent. The empirical evidence suggests that the certificate contributes rather small in absence of the appropriate conditions and constitutions (Hare, 2008). Hence, the land-use certificate seems to be inadequate to represent the actual land property rights and could cause the bias among regions. The possession of land could be
alternatively distinguished by the information which is provided from question: how does household get the land. Relied on the security level of land, these categories are divided by two groups: long-term and short term land. The long-term land includes: long term used allocation, signing contract, gift or inheritance, purchase, proclaiming, bartered. In other words, land users in this group are the land owners. The short-term land includes: auction, rent, borrow. It makes sense that the land-use certificate in long-term group is two times higher than in short-term group, around 76 percent and 39 percent.
Panel A of table 2 shows the distribution of land market. The vast majority of land used in farm activities is long-term land, consisting of 94 percent. As the assumption in study of Basu et al.
(2010), land is considered immobile. The authors agree that farmers still sell land and move but these are rare and harmless. The high ratio of long-term land in the sample could be also
acceptable to use that assumption.
When land distribution is measured by areas, the gap between urban and rural groups even larger than by household’s usage, roughly 93 percent in rural and 7 percent in urban. That is because the average land area of households is fairly uneven between the two areas. Furthermore, land is investigated whether each plot is managed by households themselves or for other purposes, including: rental, lend without payment, exchange for other plots, fallowing, and others. Around 93 percent of land area in rural group is operated by households, whereas this rate is around 84 percent in urban group. Comparing in types of land, the larger land tends to be managed by households if it is short-term land, as large as 99 percent. Due to the small fraction of short-term land and land in urban group, these lower or higher rates are nearly ineffective to this rate in categories of long-term land and land in rural, which is measured at around 92 percent.
During the last two decades, the rapid urbanization and industrialization have been witnessed as primary attributes of economic development process. Consequently, laborers move out of agricultural sector to participate in industry and services sectors. Despite of structural changing
in labor force, labors working in agriculture still constitute the largest part. From the sample, 55 percent of labors work in agriculture, forestry and aquaculture as the most time-consuming job, and 65 percent of labors engages in any of these activities for their households. Like other developing countries, labor market is imperfect due to lack of labor incentives, labor mobility and information. In particularly, labor market in rural areas might be characterized by seasonal labor demand, high transaction cost. That salient feature leads to the situation that workers in family enterprises or unpaid family laborers are pervasive in agriculture and non-agriculture as well.
The labor market in Vietnam is not the exception. Panel B displays the labor usage of households involved in farm production or service. The working information is retrieved from the questions that ask individuals whether they work for salary, for their household’s farming work, or for their household’s non-farm work. The household members might participate more than one of these types of jobs. Household belongs to working category in the panel B if there is at least one member participate in household production or service in agriculture, forestry, fishery. The information on hiring outside labor is computed indirectly from the expenditure of household’s production, which includes an item of hiring outside labor. As predicted, more than 80 percent of households employed only their members, whereas merely less than 1 percent of household hire only outside labors. Households both using members and hiring labor make up around 10 percent. There are around 4 percent of farm households without family workers and hiring expense. This could be the case that households rent out land, share land and production, or exchange product for labor cost. This distribution is similar in rural and urban areas. Therefore, the labor market in agricultural sector could be seen as imperfect.
In the situation of missing or imperfect markets, the farm households in developing countries typically are dependant to agricultural products as the consumer and producer. There is no separability between consumption and production. In the absence of labor market, the labor allocation in production and consumption might circularly affect each others. Panel C presents that majority of farm households utilizes output for both purposes: sale in the market and households consumption. This rate is higher in rural areas than in urban areas: at 91 percent and 72 percent, respectively. As the result, ratios of households retaining or selling all output are lower in rural areas than urban areas.
Table 2: Land, labor, and products distribution of households in agricultural sector Panel A: Land distribution (by area)
Managed Managed & rent Total
Long term 95.52 4.48 92.83
Short term 98.18 1.82 6.17
Others 88.67 11.33 1.00
Urban 84.7 15.3 7
Rural 93.2 6.8 93
Panel B: Labor distribution
Work only Work & Hire/Hire None Total
Obs. % Obs. % Obs. % Obs. %
Urban 534 78.88 63 9.3 80 11.82 677 10.07
Rural 5,106 84.42 727 12.2 215 3.55 6,048 89.93
Panel C: Product distribution Retained &
Sold Sold all Retained all Total
Obs. % Obs. % Obs. % Obs. %
Urban 488 72.08 82 12.11 107 15.81 677 10.07
Rural 5,515 91.22 209 3.46 322 5.33 6,046 89.93
3.3. Child labor participation
As the above descriptive data, the salient point is that a substantial share of households and child laborers involves in agricultural activities and resides in rural areas. Thus, this study focuses on the sample in rural households. Table 3 summarizes the distribution of child labor in relation with other main factors that might influence to child labor. Although almost of child laborers engage in agricultural activities, we examine child laborers employed in two categories of work:
all type of work and agricultural work. Out of 2,962 children 10 to 14 years of age, 423 child workers occurs in the category of all kind of work, including 373 child workers in the category of agricultural work. The children might occasionally work such as in the period of harvest, summer holiday, ect. The questionnaire records the number of hours worked per day, the number of days worked per week, and the number of weeks worked during last 12 months. Based on these numbers, working hours are calculated on average per week in the last 12 months. The data in sample presents that children employed in agriculture spend one hour less than children in all sample, around 11 hours and 12 hours, respectively. This implies that child labor in non-farm group spends more time than in farm group, although the fraction is much smaller.
Traced the distribution of child labor by household expenditure, the sample is divided into three groups: under 50th percentile, from 50th to 80th percentile, and the rest 20th percentile. In the lowest group of expenditure, the child labor rate is considerably higher than in the two other groups of richer households. In term of hours of work, children in poorer households tend to be employed longer in category of farm work, but less time in category of all work.
To observe the relationship between child labor and land size, we split again the per-capita managed land of households by percentiles of children from 10 to 14 years of age. The
distribution of land by percentiles partly reflects the fact that agricultural sector in Vietnam are characterized by small-scale farm households. Therefore, we relatively divide into 3 unequal groups by two breaking points at 50 percent and 80 percent; which correspond to the average land area at 0.23 and 0.68 acre. The raw data apparently suggests that the proportion of child
labor increases when land size is larger. Children also work longer time in the larger group of land size. Child labor incidence appears at lowest rate in group of household without land comparing to groups with land, but they spend more working hours. This evidence proves the above implication of the average hours worked in the two categories of work.
Despite of missing labor market, the data suggests potential impact of labor usage of household to child labor. Child labor in the group of using household members is higher than in the group of households including outside labor. On the contrary, the hours worked increases slightly in the group of households hiring laborers. A small number of child laborers belongs to group of household without farming production, although their work is faming activity. That is because they work for other households which are excluded in the sample.
Concerning to the usage of agricultural products, households are measured by percentage of the revenue yielding from sold output. As the majority of households using their products for both consumption and market, the child labor appears to be decreased when the share of sold products increase. However, the data suggests the converted sign to working hours. Child laborers spend more time in the households selling more.
Based on the different effects of main factors to the possibility of being child laborer and the hours worked, the model specifications could be appropriately identified to evaluate these two dimensions of child labor. The raw data somehow reflects the more clear impact to possibility of being child laborers but ambiguous impact to hours worked.
Table 3: Child labor in rural areas cross percentiles and categories of variables
Farm work All type
No. obs. % Hours/week No. obs. % Hours/week
Total 373 12.59 10.8 423 14.28 11.94
Living Expenditure (VND)
≤ 50 % (420,000) 273 18.43 11.16 300 20.26 11.86
50-80 % (629,000) 69 7.77 9.90 79 8.9 11.41
> 80 % 31 5.23 9.65 44 7.42 13.38
Land size (acres)
≤ 50 % (0.23 ) 105 8.05 8.12 134 10.28 10.15
50-80 % (0.68) 136 17.17 10.38 143 18.06 11.03
> 80 % 115 21.99 13.08 118 22.56 13.07
Non-managed land 17 4.96 15.35 28 8.16 20.34
Labor allocation
Work only 331 14.21 10.72 363 15.58 11.36
Work & Hire 34 11.04 10.74 39 12.66 12.1
Non-farm hh 8 2.99 13.08 21 7.84 21.62
Product usage (% sold revenue)
≤ 50 % (60%) 260 19.17 9.85 275 20.28 10.42
50-80% (81%) 74 9.7 12.51 85 11.14 13.28
>80 % 25 5.46 13.86 34 7.42 14.34
Retained all 7 5.83 12.89 11 9.17 16.4
Non-farm hh 7 2.64 15.36 18 6.79 21.51
4. Theoretical framework
The theoretical framework of this study is based on the model in Basu et al. (2010). The
economy is assumed to be one in which households treat non-work of children as a luxury good.
This assumption is the luxury axiom in study of Basu and Van (1998) which implies poverty is the cause of child labor. Landholding area is a proxy of household wealth, and suppose each household has k units of land. When the economy has a perfect labor market, if k rises, child labor will fall. On the contrary, child labor will increase in the context of an imperfect labor market, since poor households want to send their children to the labor market but there is no employment availability for them. On the other hand, poor households might are unable to hire labor on the market for their farm, which push them to employ their children. The marginal product of labor rises with farm size and the incentive for child labor is higher among the larger land owners. Nonetheless, the incidence of child labor increases until a level of landholding threshold is reached, then starts to decrease, since households are wealthy enough to not require children working.
The utility function of household:
(1) ( )
where x is total household consumption, is the amount of children’s work (i.e. take value 0 if children do not work and value 1 if children work with all effort. Let assume adults always work regardless of the wage or leisure.
By taking the quasi-linear of utility function:
(2) ( )
where ( ) , ( ) , for all x; and c is a positive real number. To verify child labor is luxury good, we can double the household income and the non-labor child, 1-e, increase more than double.
4.1. The perfect labor market
The perfect labor market implies that the supply and demand of labor in households are always equalized, and with the same wage, w. The adult equivalence correction of child wage does not make a difference, so assume wages of adults and children are equal, w.
The household profit from landholding is ( ) . The adult always provides 1 unit of labor, and e units of child labor, so the total household profit is:
(3) ( ) Hence, the household’s problem is:
(4) ( ( ) ) The first order condition is:
(5) ( ( ) ) From (3) we have:
(6)
( )
Differentiate (6) with respect to k:
(7)
( )
As the result from (7), , if land holding k rises, supply of child labor e will decrease.
4.2. The imperfect labor market
This is opposite context in which there is no labor market could solve supply and demand labors.
Each household has to manage themselves by creating work opportunities for their members. A household has the production function f as:
(8) ( )
where q is output, k is landholding are, e and 1 are child and adult labor supply, respectively. The assumption on f function are ; and . The last assumption implies that the larger land increases the productivity of labor.
There is no labor market, so the household consumes what it produces, The maximization problem of household is:
(9) ( ( )) The first order condition is:
(10)
Taking the differentials with respect to k and e to get result:
(11)
The denominator is always negative since and . Therefore, the sign of ⁄ depends on the sign of numerator. As the above hypothesis, an increase of wealth could result in a rise or fall of child labor. The first term of numerator is negative while the second term is positive. The relative magnitude of these two terms determines the sign of (11). If the first term of numerator is relatively larger than the second one, the relation of child labor and
landholding will have negative sign. The first term reach large when either or both and are large. Base on production theory and empirical evidence from Sen (1962), Deolalikar (1981) the small farms are likely to have higher productivity but lower labor productivity, namely high and low. There is no clear shift of first term on numerator when land area changes since labor productivity and land productivity move to opposite directions. Then the effect depends on the stronger effect among those. However, it is expected that the small farm has higher marginal productivity of output , and the extend of increase of marginal productivity when landholding area smaller depends on the degree of concavity of φ, in which increase positive effect of second term on child labor. Hence, we test the hypothesis that there might has a positive relationship between child labor and land size.
5. Econometric specifications
The paper’s main objective is to examine the effect of used land’s size on child labor. Since children without being employed account for a larger fraction of sample, the majority values of dependent variable are equal to zero. The linear regression OLS produces bias and inconsistent estimates when a significant ratio of dependent variable’s value is equal to zero. The Tobit model is traditional approach to deal with data with many zeros, which estimates the relationship
between non-negative dependent variable and an independent variable (Tobin, 1958). The marginal effect of independent variable is treated conditional on limited fraction of dependent variable which takes positive values as in this case.
There is an argument pointing out the inadequacy of Tobit model in solving that problem. In Tobit model, the choice of being censored (participation) and expected value conditional on un- censored (level of participation) are determined by the same factors. The model considers dependent variable to be censored at zero but ignores the source of zeros, in which could be
caused by deliberate household’s decision or certain circumstances (e.g. financial conditions, characteristics of demographic (Newman et al, 2003; Martinez-Espineira, 2006).
Heckman (1979) proposed the two-stage estimation procedure to deals with zero observation.
The author pointed out that an estimation on selected subsample results in selection bias. The first stage is Probit estimation and the second is censoring estimation on selected subsample. In other words, the first stage estimates the probability of observed positive outcome or participated decision. The second stage estimates the level of participation conditional on observed position values. The model assumes that the two stages are affected by different sets of independent variables, which is contrast to Tobit model. Another extended mark in Heckit model is that all zero observations are assumed to be derived from respondent’s deliberate choices.
Cragg (1971) proposed double-hurdle model, which is developed from Tobit model and Heckit model. The double-hurdle and Heckit models are similar in building two stages of decision. The first hurdle indicates the participation decision and the second hurdle refers to the level of participation decision. Both models are allowed to be affected by distinct explanatory variables.
However, the Heckit model assumes there is no zero observation in second stage of decision, whereas the double-hurdle permits the potential zero values in the second hurdle appearing from deliberate choice or random circumstances.
In this paper, the first stage or hurdle is a decision on whether child is employed, and the second stage is decision on how many hours child engages in. According to Heckit model assumption, all the observed observation are positive in second stage. However, in double-hurdle, there are zero observations which have potential positive hours worked. They report zero hours worked may be due to imperfect labor market, or no land in their households.
Both Heckit and double-hurdle models have been used widely in previous empirical studies, mainly in consumption decision and labor supply. The two-stage Heckman procedure is used in demand analysis of fish in Cheng and Capps (1988), habit of consuming seventeen goods in Heien and Durham (1991), etc. The Tobit model has been applied in labor supply by Blundell and Meghir (1987), Blundell et al (1987, 1998); in household consumption by Deaton and Irish (1984), Yen and Jones (1997), Burton et al (1996); or in loan default analysis by Moffatt (2005), land investment by Bekele and Mekenon (2010), etc.
5.1. Standard Tobit model The model is defined as below
(1) ( ) (2) {
where is latent unobserved endogenous variable presents preferred hours of work; and is corresponding observed variable measuring actual hours worked; and are vectors of independent variables and their parameters, respectively; is a homoskedastic and normally distributed error term. The equation (2) implies that the observed number of hours are positive continuous if the positive number of hours are desired, and zero otherwise. Due to the non- negative values of hours worked, dependent variable is censored at zero. This means that the observed zero on the dependent variable can be either “true” zero (i.e. individual deliberate
choice) or censored zero (i.e. data collection methods, certain circumstances). Using maximum likelihood method, the likelihood function of standard Tobit is:
(3) ∑ ( ) ∑ ( )
Where “0” denotes the zero observations (hours worked is zero) in the sample and “+”
indicates the positive observations (hours worked is positive); ( ) and ( ) denotes standard normal cumulative distribution function and standard normal probability density function, respectively.
5.2. Generalized Tobit or Heckit
As the argument above, Heckman (1979) propose the two-stage estimation method to correct selection bias. The first step estimates the participation decision and the second for level of participation decision. According to Heckman (1979) and Flood and Gråsjö (1998), the Tobit model is modified as:
The participation decision:
(4) ( ) (5) {
The level of participation decision:
(6) ( ) (7) {
In this model, and are vectors of explanatory variables in two stages of decision. Hence, the model assumes that the decisions of participation and level of participation are affected by separated sets of factors. As in Tobit model, and are corresponding vectors of parameters;
is a latent variable that denotes binary censoring; is the observed value representing the participation decision. The actual observed numbers of hours worked equals to unobserved latent value when the positive hours work is reported, otherwise it takes the value zero. The error terms
and are assumed to be independently distributed. This assumption implies that there is no relationship between two stages of decision.
However, Heckman (1979) assumes that the two errors term are correlated and the the first stage dominates the second one. Thus the error terms follow the bivariate normal distribution:
(8)
( ) ( ) (
)
Where is correlation coefficient of the errors terms. The domination assumption means that if the child who reports the positive hours worked is the desired purpose of their parents. In other words, the participation decision is deliberated choice. Then the model is estimated by Probit for the participation decision and standard OLS for the positive hours worked. The log-likelihood function for this approach is:
(9) ∑ ( ) ∑ ( √ ( ) ( ) If the error terms are independent, ρ=0, the equation (9) is simplified as:
(10) ∑ ( ) ∑ ( ) ( )
5.3. The double-hurdle
The model also extends the standard Tobit and Heckit models to overcome the zero hours
worked in level of working decision. The model is similar as in two stages of Heckit, but there is a slight modification in the equation (7) as following:
(11) {
The equation (11) implies that the actual observed of hours worked can be either censored at zero or data processing, and other circumstances. Thus, the zero hours worked could be caused by the fact that there is no land in the household. In other aspect, the households would send their children to work if there exits the perfect labor market. Assuming the error terms are independent, the log-likelihood function of double-hurdle is expressed as:
(12) ∑ ( ) ( ) ∑ ( ) ( )
The first term demonstrates for the observations with zero values. It implies that the zero observations are affected by both participation and level of participation decisions. It is contrast to Heckit model which indicates that all zero observations are only from participation decision.
The different point is shown by the additional term in equation (12), ( ), which
contributes for the effect of possible zero values in second stage decision. The second term in equation (12) expresses the conditional distribution and density function of censoring rule and observed positive values.
Under the assumption of independence between two error terms, the log-likelihood function of the double-hurdle is the summation log-likelihood of Probit model and truncated regression model (McDowell, 2003; Aristei et al, 2007). The model can be estimated by maximizing two component separately as there was unavailable software for double-hurdle before (John, 1989;
McDowell, 2003). However, this paper used the user-written program of Burke (2009) in Stata.
6. Model specifications 6.1. Instrument variables
Land is the main explanatory variable. The area is managed or operated by the household for agricultural, forestry, and aquacultural activities. As mentioned above, the total area of land includes long term and short term land, which is expected to be endougenous. The inherent land is clearly considered to be exogenous element and the appropriate instrument for land used.
The relationship of child labor and household consumption is causal effect that is resulted from jointly decision making process. The economic situation of household might lead the children working or not and children outcome also contributes to household income. Hence, consumption expenditure is expected to cause the potential endogeneity problem. This issue has been solved in previous studies by two solutions. First, the children’s income is excluded from household income to control the effect of child labor to consumption’s decision (Ray, 2000; Duryea and Arends-Kuenning, 2003). This approach solves the problem on the source of endogeneity and still leaves the problem on relationship of child time allocation and household living standard (Edmond, 2008). Moreover, the technique is appropriate only for the case children with paid job
that make their outcomes be observed. In fact, almost child labors are non-remunerated which including works for household farm, business; or paid by in-kind, meals, clothes, ect. The income of child work indirectly contributes to household income by substituting for adult labor employed in other paid jobs. Due to these disadvantages, instrumental variable solution has been conducted to address the endogeneity of household income (Bhalotra and Heady, 2003; Ersado, 2005). The procedure was initially proposed by Smith and Brunell (1986). The study provides the asymptotically optimal test of weak exogeneity for simultaneous equation in limit dependent variable model. Let consider the equation of hours worked as above but distinguish between exogenous and suspect endogenous explanatory variables:
(14) ( )
Where is censored dependent variable, is vector of exogenous variables and is vector if of endogenous variables. The auxiliary equation regresses the suspected endogenous variable by exogenous and instrumental variables as following:
(15) ( )
Where Z is vector of instrument variables and could include other exogenous variables in main equation. Both error terms are assumed jointly normal distribution. Expressing
conditional on as:
(16)
Substitution for in equation (14) to become the conditional model:
(17)
Where is residual estimated by OLS equation (16). Under the null hypothesis, the residual has no power explanation. Under the alternative hypothesis, the suspected endogenous variables are expressed as linear projections of a set of instruments, and the residuals from those first-stage regressions are added to the model.
6.2. Description of variables
The model contains the variables of characteristics of child, household and region. The ratio of gender is fairly similar in sample, although the empirical evidence reports mix results on gender.
The existence of gender bias might be derived from parental gender bias, labor market
discrimination, types of the work, difference on return of human capital. Edmonds and Pavcnik (2005) examine the data for thirty six countries and find that child labor for girls is higher than for boys if including market work and domestic work. But when the data exclude domestic work, the child labor for girls is less than for boys. According to report of Diallo et al. (2010) in 2008, there were 176 million boys in economic activity compared to 130 million girls. It is expected that child work is affected by availability of labor and household size. The larger household and more adult labors are expected to decrease child labor. The education level of parents or head affects to decision of child labor. The increase of parental education has direct positive effect to their children education and job opportunities resulting in decrease child work. The gender of head represents for bargaining power in household. Table 4 shows the gap of child work, especially in ratio of participation cross regions. Therefore, the regional factor is included to capture different demographic characteristics, conditions and economic levels.
Table 4: Description of variables
No. obs. Mean Std. Dev. Min Max Dependent variables
Children in farm work
Participation 2962 0.13 0.33 0 1
Hours worked 2962 1.36 5.12 0 58.15
Children in all work
Participation 2962 0.14 0.35 0 1
Hours worked 2962 1.70 6.06 0 64.62
Independent variables
Child age 2962 12.15 1.4 10 14
Child gender 2962 0.5 0.5 0 1
Children under 10 years 2962 0.65 0.84 0 6
Members 15-19 years 2962 0.6 0.76 0 4
Education of head 2962 6.51 3.58 0 12
Female headship 2962 0.14 0.34 0 1
Expenditure (per capita) 2962 486 286 70 4146
Land (per capita) 2962 0.43 0.86 0 18.53
Land square (per capita) 2962 0.92 9.83 0 343.18
Sold revenue (percent) 2962 51.58 31.71 0 100
Hired outside labor (dummy) 2962 0.1 0.31 0 1
Regions
Red River Delta 2962 0.18 0.38 0 1
North East 2962 0.15 0.36 0 1
North West 2962 0.07 0.26 0 1
North Central Coast 2962 0.14 0.35 0 1
South Central Coast 2962 0.1 0.3 0 1
Central Highlands 2962 0.1 0.3 0 1
North East South 2962 0.1 0.29 0 1
Mekong River Delta 2962 0.17 0.37 0 1
7. Results
7.1. Model selection
The empirical estimation begins with the selection of the appropriate model. I first compared the Tobit specification with the double-hurdle specification using the likelihood ratio test. The log likelihood for Tobit model is the sum of the log likelihoods of the truncated regression model and probit model, the likelihood ratio (LR) statistic can be computed as:
LR = -2[logLT – (logLP + logLTR)]
Then the LR is compared with chi-square critical value.
Secondly, we test the double hurdle against Heckman. According to Vuong test (1989), it is based on the standard likelihood ratio statistics also, but using another transformation:
LR = (logLP + logLTR) - logLH W= (1/n)[LR]2 – [(1/n)LR]2 V = LR/(w*n)
Where V is test statistic and has the normal distribution. If this value is greater than critical value, then the double-hurdle is better than Heckman. Otherwise, the test could not discriminate between the two models.
The estimated coefficients across the three approaches are almost indifferent in term of sign. The level of significance also similar. In order to select the appropriated model, two steps of test is conducted. The first one is to examine Tobit against double-hurdle and the second one is to test the double-hurdle model against the Heckman model. Table 6 displays results that the Tobit model is strongly rejected and there is no difference between the double-hurdle and the Heckman model. The result are the same for farm work and all work sample. This implies the presence of the two stages of decision-making which households decide to employ their children and how many hours to work. Hence, the Heckman model and double-hurdle model are better specification to identify the impact factors at both stage of decision-making.
Table 5: Model selection test
Farm work sample Test value Decision
Standard Tobit Vs. Double-hurdle 38 (20) [0.000] Reject Tobit
Double-hurdle Vs. Heckman 0.000 No difference
All work sample
Standard Tobit Vs. Double-hurdle 66 (20) [0.000] Reject Tobit
Double-hurdle Vs. Heckman 0.000 No difference
Note: The column reports the LR and Vuong test statistics in rows 1 and 2, respectively, the degrees of freedom of the chi-square statistics (in round bracket), and the corresponding p-value (in squared bracket), respectively.
7.2. Discussion
Table 6 and table 7 (Appendix) keep on showing the results of the Tobit model versus the Heckman model and the double hurdle model for comparable observation. The marginal effects of explanatory variables are estimated for participation decision (probability of working) and level of participation decision (hours worked). Since the assumption in Heckman and double- hurdle models allow the different explanatory variables in two stages, several insignificant are excluded in the second stage. These excluded factors also are inline with the descriptive data.
The Smith-Blundell test reveals the rejection of exogeneity in participation stage but acceptance in participation level stage for both expenditure and land residuals in the Heckman model and the double-hurdle model. This means that the instrumental variables contribute to the variation of expenditure and land in participation decision, and jointly significant at 1 percent and 10 percent respectively. However, instruments variables are irrelevant for level participation decision in the Heckman and double-hurdle models. Therefore, the instrumental variables also contribute to explanation power on the probability of working.
Main variables
In general view, the results are similar between Heckman and the double-hurdle, and between farm work and all work sample, in aspect of sign and level of significance. First, the focused factors in this paper are land and expenditure. Both variables have strong significant effect on the probability of working. Expenditure causes negative influence, whereas land size has positive
impact. Children from households operating larger land but less expenditure are likely to be employed. In addition, quadratic land has significant negative effect to participation stage, except for the case of second stage in all work sample. In other words, land’s effect decreases when land size continuously increases. These effects of land and quadratic land clearly indicate the non linear relationship of land size to child work.
Nonetheless, expenditure and land cause no significant effect on hours worked. The level of participation in working depends on other factors, other than expenditure and land, that shows the evidence on two separate decision-making stages.
Other variables
Besides main variables of interest in the study, we explore other determinants that also affect child work. Tracking children’s characteristics, the data shows that the older children are significantly employed more than the younger ones. This effect does not appear in level of participation. Moreover, the possibility of working between boys and girls are the same, as descriptive data displyed in previous section.
In addition, structure of household provides informative clues for analysis source of labor within household which could affect to decision on child work. The groups of children in household are categorized by age to investigate whether there exists the supplement or substitute effects among children within household. The number of children under 9 years old has positive impact to the level of participation in farm work and all kind of work. In the sample, the children participate in working from 6 to 9 years old is a small fraction. Hence, the effect of this group could not be interpreted as the additional source for elder child work. A possible reason is that households have more children probably to get more incentives from children labor source. Or an alternative reason is that if the household has child work, they will be likely to let these other younger children work in future. The coefficients of children from 15 to 19 years old have positive effect on hours worked, but insignificant on farm work and significant on all kind of works. This result implies that children in this group participate in paid work and non-farm work more than
children from 10 to 14 years old. The estimation makes scene to show that the
children have more opportunities and abilities to work in other activities if they are older.
Regarding to the characteristics of household head, we test for dummy variable if household head as female, and the education of household head. The number female household head accounts for only 14 percent in the sample. This determinant causes significant impact to participation decision in both models of farm work sample, but only in Heckman model in all work sample. Whereas education of household head has expected negative effect on both stage of decision, in both models and samples. The impacts are strongly significant but magnitudes of effects are lower than other factors as land, expenditure and household structure. These values are remarkable if we classify education as no education, primary, secondary, and high school level.
The factor of labor distribution and production distribution are found to be irrelevant to child labor. As in the descriptive data and assumption from theoretical framework, the labor market is considered imperfect and the estimation reveal the insignificant effect in all three models. The usages of agricultural production also has no effect to both stages of decision.
