Poverty and other determinants of child labor in Bangladesh.

1. Introduction

Americans have demonstrated strong disapproval of child labor through their own history and in current movements to boycott imports from countries employing child labor. (1) Wasserman (2000) demonstrated the parallels between the historical pattern of the decline in child labor in the United States and situations today in developing countries. She relied heavily on the notion that the greater the extent of poverty in a country, the greater the amount of child labor. She noted that this relationship holds within a country over time as well as across countries at a given point in time. Though poverty may be a determinant of child labor, it cannot be examined in the absence of cultural and social factors, such as education, culture, and urbanization.

Why should we care if poverty is a determinant of child labor in a developing country such as Bangladesh? As researchers have noted, boycotting exports of goods produced by children may actually worsen the welfare and well-being of those children and their families, first by lowering their living standards and second by pushing children into dangerous work endeavors, such as begging and prostitution (Bissell and Sobhan 1996; Wasserman 2000). Measures such as the Child Labor Deterrence Act (the Harkin Bill), which would ban imports of products that children help produce, may be well meaning (Harkin 1999) but may not be in the best interests of those they are attempting to help (Bissell and Sobhan 1996; Rahman, Khanam, and Absar 1999). (2)

In this research we examine poverty (as measured by income) and other child and family characteristics as determinants of child labor in Bangladesh. If we can understand the dynamics of poverty, education, and child labor, we may be in a position to create more effective ways of eliminating child labor through education and economic development efforts. First, however, we must understand the nature of child labor in Bangladesh and establish the connections that exist among child labor, poverty, education, and other socioeconomic factors.

2. Background: Child Labor in Bangladesh

Why Do Children Work?

Child labor is pervasive in Bangladesh. According to reported statistics, as many as 19% of children ages 5 through 14 are in the labor force (Rahman, Khanam, and Absar 1999). As these authors state, "Child labor is deeply rooted in poverty and social customs" (p. 999). Thus, one reason for such high labor force participation is that working children are from impoverished families (Basu and Van 1998; Basu 1999).

There are also social and cultural explanations for widespread child labor in Bangladesh. Delap (2001) concluded from a survey of families in three Dhaka slums that purely economic explanations for child labor are not adequate to explain the phenomenon. One cultural factor motivating families to send their children to work is a fear that the children will be idle if they do not work. Delap reported that the majority of parents in a survey of the urban poor indicated that the income of working children was not critical to the family's survival, but the children worked because "it was improper for children, especially older children, to do nothing" (Delap 2001, p. 15). Idleness was deemed especially harmful to poor urban boys, whom parents feared would become involved in criminal activities.

Not only is idleness to be avoided, but children's work is also viewed favorably as a means of preparing young people for work as adults. They can begin to learn the skills of farming or a trade. Rahman, Khanam, and Absar (1999) offered the example of employment in an engineering workshop that provides an opportunity for boys to learn an employable skill. For many girls, serving as maids in households prepares them for their future as wives and mothers.

Schooling does not necessarily limit child's work as it does in industrialized countries. Ravallion and Wodon (2000) examined the work and schooling choices of rural Bangladeshi children to determine whether an enrollment subsidy plan was effective in limiting child labor. Because rural schools are open only part of the year and only part of the day, the authors contended that the time children spend in school would not necessarily limit time they would have spent working. The enrollment subsidy program did increase schooling of those offered the plan, but child labor did not decrease comparably. Rahman, Khanam, and Absar (1999) underscored these findings in noting that, for many children displaced from garment factories, school was not an option. Instead, these children found work in the informal sector of the economy where jobs on the street are less secure, more poorly paid, and sometimes dangerous.

What Work Do Children Do?

The answer to this question depends on the age and gender of the child and whether the child lives in a rural or urban area. Table 1 shows the distribution of younger (ages 5-11) and older (ages 12-14) urban and rural boys and girls by general occupational category, based on answers to the 1995-96 Household Expenditure Survey (HES) of Bangladesh, the data source for our empirical study. These two age-groups are chosen for three reasons. First, by 12 years of age, girls have entered puberty, and secluding them from the public is important. Second, age 12 is a point where children may make a decision to continue with their education. Third, an examination of our empirical data shows a substantial increase in the proportion of children working at age 12. The reported percentages are calculated on the basis of the number of working children who indicated an occupation.

Some occupations shown in Table 1 are mostly urban (e.g., transport and communications workers), while some are mostly or completely rural (such as agricultural occupations). Likewise, some are mostly or completely boys' jobs (salesmen/businessman, transport and communications worker, and day laborer). No major occupational categories shown represent completely girls' jobs, though specific employment within these occupations may be gender specific. Most urban boys are in the category of salesman/businessman. These boys work as shop assistants, street vendors, and tea and food vendors. Urban boys may also work as assistants in match, biscuit, shrimp, cigarette, and salt processing factories as well as in tanneries. Some boys also work in garment, hosiery, and small engineering factories (Rahman, Khanam, and Absar 1999). They also work in smaller numbers as porters, cycle rickshaw pullers, ticket sellers (transport workers), servants, and day laborers. Additional employment for boys in the informal sector of the urban economy includes stone crushing and firewood collection. Rural boys' jobs are largely agricultural, with some rural boys selling items and some younger rural boys working as servants.

Delap (2001) emphasized the importance of the practice of seclusion, or purdah, in explaining girls' jobs in Bangladesh. Urban girls aged 12 or older are restricted to jobs in which they are secluded from men. Table 1 shows two major occupations for girls: working in production, especially in garment factories where workers are primarily women or younger boys, and working as maids and domestic servants. Rural girls primarily work as domestic servants and maids. Some girls, especially older girls, work in the agricultural sector, and a few rural girls work as vendors or in production jobs.

Is Poverty a Determinant of Child Labor in Developing Countries?

In examining factors that affect child labor in Bangladesh, we place special emphasis on poverty. Our interest in the role of poverty is attributed to Basu and Van's seminal article in which they proposed a "luxury axiom" of child labor stating, "A family will send the children to the labor market only if the family's income from non-child-labor sources drops very low" (Basu and Van 1998, p. 416). In other words, a child's leisure or nonwork is a luxury that these poor families cannot afford. Thus, the "luxury good" is a child at home rather than in the workforce. As family income increases, "consumption" of this luxury good would increase. This translates into a lower probability that a child would work for families above some minimum poverty level.

There are many challenges that researchers face in attempting to test the luxury axiom of child labor. First is to interpret what Basu and Van mean when they say that a "poor" household cannot afford to consume the luxury good of children's leisure. The luxury axiom itself defines poor in terms of "the family's income" (Basu and Van 1998, p. 416). Yet elsewhere, the authors expand their concept of poverty when they say, "More generally, all we want is to give primacy to the household or family wealth as a determinant of child labor" (p. 415). Thus, we must consider both income and wealth as dimensions of poverty.

A second challenge is to determine what level of income or wealth triggers sending a child into the workforce or removing a child from work. Basu and Van state that "a poor household cannot afford to consume this good [children's nonwork] but it does so as soon as the household income rises sufficiently" (Basu and Van 1998, p. 415). How low must income fall before a child works? Must the family exhaust all assets and have insufficient income to meet current period needs in order to send their child to work? How would the presence of a social safety net in a country affect the income requirements of families and their decisions concerning child labor?

Given the multidimensional aspect of poverty, a third challenge in examining Basu and Van's luxury axiom is to acquire both income and asset data, preferably over time. Finding such data for developing countries is particularly difficult given severely limited data sources. Our information sources for Bangladesh limit us to income data for members of the household and do not provide adequate information on family assets or income from social programs. Defining poverty as low family income provides an initial approach to examining Basu and Van's idea that poverty is a determinant of child labor, but looking at income alone does not capture other dimensions of family poverty. Thus, establishing a negative link between income and child labor provides a necessary but not sufficient condition for the luxury axiom to hold.

Challenges in testing the luxury axiom have not deterred researchers from investigating the role of family poverty in child labor. Although Delap emphasized the importance of looking at social and cultural determinants of child labor in her 2001 study of children in Dhaka's slums, she did find an association between lower household income and children's labor force participation. Household poverty was the most frequent reason given for children to have engaged in income-generating work. In an appendix to her paper, Delap reported that lower household income was associated with a higher probability that a child would work.

Ray (2000a, b) provided empirical tests of the connection between family income and children's work in Peru and Pakistan. First, he used logistic regression to calculate the probability that a child would be working, based on a set of independent variables (Ray 2000a). In his second paper (Ray 2000b), he estimated a child labor supply function. In both studies, poverty was defined as family income below the poverty line. He showed that the luxury axiom, when applied to the probability of working, was rejected for Pakistan and showed weak support for Peru (Ray 2000a). He also found a significant positive relationship between a child's hours worked and poverty for Pakistan but not for Peru (Ray 2000b).

A third study looked directly at the influence of poverty on child labor for children in an African country. Grootaert (1999) used a sequential choice model to find the determinants of child labor. Using the Cote d'Ivoire Living Standards Survey 1985-1988, he separated the data into two groups, urban and rural, and then conducted a multivariate analysis of the determinants of child labor force participation. He controlled for child characteristics, parent characteristics, and household characteristics. Grootaert included a dummy variable indicating whether household income fell in the lowest quintile. He argued that inclusion of this variable captured the constraints faced by the poorest households in terms of access to credit and insurance. Grootaert found some support for the idea that financial constraints affecting the poorest households were among the most important variables in determining child work and schooling outcomes for a sample of urban children. For rural children, household poverty rarely influenced the decision to work.

When we examine empirical studies of child labor determinants for developing countries, we find some support for the importance of poverty. Given mixed results from the literature, the purpose of our research is to deepen the understanding of the connection between poverty and child labor by developing a model that includes economic as well as social and cultural information for different groups of children in Bangladesh. We contribute to the existing literature in several ways. We improve on the income variables used by Grootaert (1999), Ray (2000a, b), and Delap (2001) by dividing nonchild family income into quintiles. Unlike studies that focus on either urban or rural children, we examine both. Finally, we develop separate models for younger and older children to see if family poverty and other variables have stronger effects on younger children. (3)

3. Data, Variables, and Models

The Data

We use logistic regression to estimate whether a child works, based on information about the child and his or her family. Our dependent variable is binary; either the child works or the child does not work. Data are drawn from the 1995-96 Household Expenditure Survey of Bangladesh, conducted periodically by the Bangladesh Bureau of Statistics (BBS). These data were released for public use in 1998. A two-stage stratified random sampling technique was followed, under the framework of the Integrated Multipurpose Sample (IMPS) design which was developed on the basis of the Population and Housing Census of 1991. The primary sampling unit used in the HES is the household. Although there is an enormous amount of individual (household member) data available from the survey, the focus of the survey was on the household and data on all household members were provided in the household survey. A total of 371 communities were chosen for the survey to be nationally representative with twenty households randomly selected from each of these communities. This procedure produced data for a total of 7420 households.

The survey provides information on an individual's location within Bangladesh, his or her economic activity, occupation, industry, age, and educational background as well as household size, household income, (4) household expenditure, consumption, changes in wealth, and health and sanitation. Our sample consists of 11,373 children aged 5 through 14 who live in 5394 households. We apply our empirical analyses to 11,282 of these children (5) and then to sub-samples of boys and girls from urban and rural areas.

Defining Child Labor

In order to examine the determinants of child labor in Bangladesh, we developed a definition of child labor based on the response to the question concerning "activity" of the child during the previous week. The choices in answering this question were as follows: (i) employed and worked the previous week, (ii) employed but did not work the previous week, (iii) did household work, (iv) searched for a job but did not find a job, (v) did not search for a job and had no interest in finding a job, (vi) was a student, (vii) was retired, or (viii) other.

As indicated earlier, the dependent variable in our logistic regression model is binary (see Table 2 for a summary of variables used in the model and descriptive statistics for these variables). We define our dependent variable as Work; this variable takes on a value of I if a child works and a value of 0 otherwise. The estimated value of Work is the probability that a child will work, P(Work). Our definition of child labor takes cultural and societal norms into account and, at the same time, recognizes the nature of the data obtained in the survey. Though the "activity" of an individual pertains to the previous week, the variables we will use to estimate the work status (a child worked or did not work) are defined on an annual basis. Under our definition, a child is considered to work if either he or she was employed the previous week (activities i and ii), whether income was earned during the year or not, or if he or she earned income during the year regardless of his or her activity status for the previous week. In other words, earning income over the past year becomes the criterion for being classified as working for those who did not report that they were employed during the previous week. (6)

A total of 1064 children are considered to be child laborers under this definition; this amounts to 9% of those in the sample. This figure is somewhat lower than reported in other studies (Rahman 1997; Ravallion and Wodon 2000) because our definition requires that a child who did not work during the previous week would have had to earn income during the previous year to be counted as child labor. Thus, our definition may underestimate the extent of child labor in Bangladesh. First, we purposely omit children who were not paid for doing household chores; some may consider this child labor, but we do not in this study. We exclude these children because the definition of "labor" in the term "child labor" generally means that a child works on a regular basis for pay or is unpaid but produces output that will be sold in the market (Bachman 2000). The only children claiming the activity of household work who would be included in our study are those who earned income during the year; there was only one such child. We cannot determine from our data source whether this child's income was from household work or from other jobs this child held earlier during the prior year; thus, we retain this child in our sample.

Second, among those in our study who reported working the previous week, only 14% reported receiving income during the year. This tells us that there might be other children who worked during the year (but not during the previous week) who would not have earned cash income. (7) The main point is that these unpaid workers who did not report working during the previous week would not be included as child labor in our study because we cannot distinguish them from those who did not work at all during the previous year (see Ravallion and Wodon 2000 for an alternative measure of child labor).

Factors Influencing Child Labor

What factors determine whether a child works? Based on the work of Basu and Van (1998) as well as others, income (or, more precisely, poverty) is a key factor. One approach is to use the natural log of income as an independent variable since the negative influence of higher income on a child's likelihood of working would diminish as income increases. Although we estimated such models, we have not included the results in our tables because the log-of-income variable does not capture the fundamental idea that poverty is a determinant of child labor. (8)

Instead of using the log of income, we define five dummy variables as follows: Poorest takes on a value of 1 if the nonchild family income is in the lowest 20% of the household incomes in the sample of 11,282 children and a value of 0 otherwise. Similarly, the dummy variable Poor equals 1 for a child in the second-lowest income quintile, Middle equals 1 for a child in the third-lowest income quintile, Rich equals 1 for a child in the fourth-lowest income quintile, and Richest equals 1 for a child in the highest income quintile. The omitted category is Richest; thus, a positive sign on a dummy income category variable indicates that a child in this income category is more likely to work than a child in the omitted category (Richest); that is, the probability of work (our dependent variable) is higher for this income category than for the Richest category. We realize that poverty may exist at a low level of income; thus, we expect that some higher-income-category variables may have coefficients that are not statistically different from zero.

Table 2 shows additional determinants of child labor included in our model. Age indicates the age of the child in years; the average age is 9.45 years. We expect families to be more likely to send older children to work; thus, the coefficient on Age is expected to be positive. We also account for age by dividing children into two groups. Younger children are ages 5 though 11, and older children are ages 12 though 14. We make this distinction because the motivation for families to send children to work may be quite different for younger as compared to older children.

The variable Boy takes on a value of 1 if the child is a boy and 0 if the child is a gift. Table 2 shows that 51% of the children in our study are boys. Families are more likely to send boys to work; thus, we expect a positive sign on Boy. According to Delap (2001, p. 11), "household decision-makers are reluctant to send girls out to work and will do so when all other household members ... are working and extra income is still required." The variable Eduyr is the number of years of education completed by the child. (9) The average years of schooling for children in our study is 2.3 years. We would expect a negative sign on the coefficient of Eduyr since those children with more education at any given age are more likely to still be in school and, thus, less likely to be working. Urban takes on a value of 1 for a child living in an urban area and 0 if the child lives in a rural area. Only 25% of children in our study are from urban areas. This is generally consistent with the overall population distribution in Bangladesh, which is 85% rural (Rahman 1997). We expect a positive sign on Urban because child work outside the home is more prevalent in the informal sector in urban areas of Bangladesh (Rahman 1997).

The variable HHSize measures the household size as the number of people in the household; the mean household size is 6.56 individuals. We expect a positive sign on the coefficient of HHSize since the larger the family, the more mouths to feed and the greater the family's need to have children work for income. The variable Headmale takes on the value 1 if the head of the household is a man. Ninety-two percent of children in our study come from male-headed households. For 88.1% of children in male-headed households, the head is the child's father; in the remaining cases, the head is a grandfather, uncle, brother, or other male relative. For 87.7% of children in female-headed households, the head is the child's mother; in the remaining cases, the head is a grandmother, aunt, sister, or other female relative. We expect that male-headed households are less likely to have children working because of social status and because of greater income stability than in female-headed households. Thus, the expected sign on Headmale is negative. Finally, the variable MHeadedu represents the household head's education in years for a male household head, and FHeadedu represents the household bead's education in years for a female household head. Studies have shown that the more educated the parents, the more likely the children will attend school rather than work (Grootaert 1999; Ray 2000b). Hossain (1996) suggested that poorly educated parents may not see the eventual benefits of education as compared with the immediate advantages of income from child labor. Thus, we expect negative signs on the coefficients of MHeadedu and FHeadedu.

The model that we estimate is shown here:

P(Work) = [b.sub.0] + [b.sub.1] Poorest + [b.sub.2] Poor + [b.sub.3] Middle + [b.sub.4] Rich + [b.sub.5] Age + [b.sub.6] Boy + [b.sub.7] Eduyr + [b.sub.8] Urban + [b.sub.9] HHSize + [b.sub.10] Headmale + [b.sub.11] MHeadedu + [b.sub.12] FHeadedu + error.

This model is estimated for the full sample of 11,282 children (first all children and then younger and older children) and then for eight specific demographic groups described in the next section.

4. Method and Results

Logistic Regression

We apply logistic regression techniques to determine the way in which the previously discussed factors influence the probability that a child will work. Because the dependent variable is binary, ordinary least squares (OLS) estimates are not ideal. Instead, a logit model is used to estimate work status equations. (10)

We report estimated coefficients and their standard errors as well as marginal effects (partial derivatives) of independent variables in the logistic model. The marginal effect of the probability of a particular independent variable is calculated as [delta]P(y = 1)/[delta]x = [beta]P(1 - P), where x is the independent variable, [beta] is the logit estimate, P is the probability that y equals 1, and (1 - P) represents the probability that y is 0 (Maddala 1988; Liao 1994; Allison 1999). Some children in our sample are in the same households and therefore do not constitute independent observations. Because of this, standard errors of the coefficients have been corrected for clustering.

Regression Results for the Full Sample

Results of the logistic regression analysis for the full sample of children are presented in Table 3, which reports the coefficient of each variable, its standard error and statistical significance, and the marginal effect of a one-unit change in each independent variable on the probability that a child will work. (11) To test the overall statistical reliability of each model, we calculate the log likelihood and subject it to a chi-square test. As Table 3 indicates, the chi-square value is statistically significant for all three estimated equations.

The estimated models for the full sample of children in the study show statistically significant (beyond the .10 level) coefficients and expected signs for all child and household variables. Of particular interest are the positive signs on the coefficients for Age, Boy, and Urban. The older the child, the greater the probability the child works; each additional year adds .02 to the probability of working. Moreover, boys are more likely to work than girls; being a boy increases the likelihood of working by almost .04. Similarly, being from an urban area adds .03 to the probability a child will work.

It is apparent from Table 3 that some factors have different effects on the probability of working for younger versus older children. Because of these differences and because of the influence of age and gender on the probability that a child will work, we estimate separate equations for younger and older boys and gifts in urban and rural areas. Rather than providing additional discussion of income as well as child and family characteristics for the full sample, we present detailed discussions for each of the eight equations we have estimated for demographic subgroups.

Work Behavior of Urban Children

Table 4 provides descriptive statistics for four groups of urban children. The mean value of the variable Work shows the probability that a child will work. For younger urban boys the probability is 7%, and for older urban boys it is 25%. Only 6% of younger urban girls work, while the figure is 11% for older urban girls. Mean values for other variables are shown in Table 4.

Logistic regression results for the four groups of urban children are shown in Table 5. Differences are apparent between older and younger children and between boys and girls. Coefficients of child and family variables are statistically significant and with expected signs for older urban boys. The older the child, the greater the probability that he will work; moreover, each added year increases the probability of working by over .05. Having more years of education (given one's age) lowers the probability of working; each additional year of education decreases that probability by about .04.

Household size is a positive and statistically significant determinant of whether an older urban boy works. The probability that an urban family will send an older boy to work increases by .07 for a one-person increase in family size. Having a male head of household decreases an older urban boy's likelihood of working by .23, a much larger effect than for any other urban subgroup of children. When the male household head has a higher level of education, the probability that an older urban boy will work decreases by .01. The education level of a female household head has a negative effect of .06 on an older urban boy's likelihood of working, confirming our expectation that more educated parents choose not to send their children to work. Results for younger urban boys are similar except that neither the child's age, education, nor the male household head's education is a statistically significant determinant of child labor.

Of primary concern in this study is whether poverty (i.e., low nonchild family income) plays a role in the decision to send a child to work. Our models support the link between family poverty and child labor. For younger urban boys, the positive coefficient on Poorest is statistically significant; thus, boys from the poorest income quintile are more likely to work than are those from the richest income quintile. Boys in families from all other income quintiles are no more likely to be sent to work than are boys in families from the richest quintile. For older urban boys, the coefficient is positive and statistically significant on Poorest and on Middle, again showing the influence of low income on child labor. For older urban boys, being in the poorest income quintile increases the probability of working by almost .26 compared with older boys in the richest income quintile. This is the most important predictor of work status for older boys followed closely by Headmale. For younger urban boys, the marginal effect of Poorest is only about .08, whereas the marginal effect of Headmale decreases the probability of work by .17. In summary, we cannot reject the hypothesis that family poverty influences the child labor decision for both younger and older urban boys. In addition, being in a male-headed household has a strong negative impact on the probability that urban boys will work.

Logistical regression results for younger urban girls, those ages 5 through 11, show statistical significance and expected signs on Age but not on Eduyr. However, most family variables are statistically significant with expected signs; only female household head's educational level does not influence the decision to send a young urban girl to work. The probability that young urban girls will work decreases (by .03) for male-headed households, but the effect is smaller than for young urban boys (.17). A young urban girl's probability of working decreases by a negligible amount (.004) as the male household head's educational level increases. For these young urban girls, the effect of poverty on child labor is evident from the positive signs and significance levels for Poorest and Poor. Being in the poorest income quintile, for example, increases the probability that a young girl will work by about .07 as compared to a young girl in the richest income quintile; this marginal effect is about the same as for young urban boys.

By contrast, for older urban girls (ages 12-14), most child and family variables have no impact on the girl's employment status. Only Headmale and MHeadedu have a statistically significant and negative influence on the probability that an older urban girl will work, and neither of these factors decreases the probability of working by very much. Being older, more educated, from a larger household, or from a household whose female head is more educated do not increase the likelihood that an older urban girl will work. Family poverty, however, has a strong positive marginal effect (of about .23 for those in the poorest income quintile) on older urban girls' probability of working, an effect that is similar to that for older urban boys and stronger than the effect for younger urban girls. Poorest is the variable with the largest marginal effect on older urban girls' work status. We conclude that family poverty motivates older urban girls to work, but many of the child and family characteristics that influence the probability of working for children in other urban demographic groups do not affect older urban girls.

From our earlier observation of urban girls' occupations (see Table 1), we see that a high proportion of older girls worked as maids (12) or as production workers, primarily in the garment industry. Girls working in factories may be earning money for dowries that will be required when they marry. Although paying dowries is illegal in Bangladesh, it is still widely practiced. Also, urban girls may be in families who have moved from rural areas and continue to send money earned in the city back to their extended families in rural areas. Finally, when urban mothers work outside the home, they often take their daughters to work with them, or they find appropriate employment for their daughters so that the daughters are not left home alone. All these are offered as explanations of the motivations of urban girls to work that go beyond the child and family explanatory variables in our model.

Work Behavior of Rural Children

Four models are estimated for rural children. Descriptive statistics are shown in Table 6, which reveals a range of employment rates: 7% of younger rural boys, 29% of older rural boys, 5% of younger rural girls, and only 6% of older rural girls. Female seclusion is especially important for older rural girls; thus, we see lower percentages of girls at work than in urban areas. Though average years of education is slightly higher for younger rural boys than for younger girls, the reverse is true for older rural children.

Table 7 reports logistical regression results for younger and older rural boys and girls. Signs and significance levels of the coefficients of child and family variables are as expected for younger and older rural boys, with one exception: being from a male-headed household does not affect the probability that an older rural boy will work. The marginal effects of Age and Eduyr are more pronounced for older than for younger rural boys. An additional year of age, for example, adds .10 to the probability that an older rural boy will work.

Poverty is a determinant of child labor for both younger and older rural boys. For younger rural boys in each of the four lowest income quintiles and for older rural boys in the poorest and poor income quintiles, the probability of working is higher than for those in the richest quintile. Most families are compelled by their poverty to send even young boys to work, largely in farm labor jobs. For both younger and older rural boys, Poorest has the largest marginal effect of any of the variables on work status (.19 and .12, respectively). Headmale is second in importance for younger rural boys, indicated by a decrease of. 12 on the probability of working.

For both younger and older rural girls, age and years of education do not influence the probability that the girl will work; moreover, male household bead's educational level is not a factor for younger rural girls. Otherwise, child and family characteristics are statistically significant with expected signs. Headmale has a particularly strong marginal effect for both younger and older rural girls, decreasing the probability of working by. 14 for young rural girls and. 12 for older rural girls. For both age-groups of rural girls, poverty is a very important determinant of child labor. The impact on the probability of working of being in the Poorest income quintile is quantitatively similar for both younger and older rural girls, an increase of. 12 for younger girls and. 14 for older girls.

5. Summary and Conclusions

In this study, we develop a logistic regression model and estimate it using data for Bangladesh. We define income quintiles as a means of measuring family poverty and add child and family characteristics to our model. We estimate the likelihood that a child will work, using separate models for younger and older boys and girls in urban and rural areas. Our results support the notion that a family's poverty affects the probability that a child will work; keeping children away from work is a luxury these families cannot afford. In most models, being in a male-headed household is the second most important determinant of a child's work status.

Older urban girls show the greatest divergence from our expectations. For these girls, aside from poverty, only two variables, being from a male-headed household and a male household head's educational level, produced expected results. We conclude that these girls are motivated by factors not captured in our model.

Based on the results of our study, we recommend that researchers with access to data from other countries examine the role of poverty (in both its income and its wealth dimensions, if possible) and other socioeconomic factors to determine whether the poverty-child labor link holds internationally. If poverty is, indeed, a determinant of child labor in Bangladesh (and perhaps other countries) and if child labor is an undesirable social and economic condition, then policymakers must turn their attention to alleviating poverty. If other researchers find groups of children for whom poverty is not a determinant of child labor, alternative solutions need to be examined.

Industrialized countries' policies that ban imports of goods manufactured by children have come under heavy criticism since these policies have pushed children into more dangerous work activities. More promising than banning imports are efforts, already under way, to increase educational levels of Bangladeshi children. Emphasis on education could decrease child labor directly by increasing the time children devote to school and indirectly through investments in human capital that will improve productivity and family income. Future generations would be endowed with the attributes that we have found give families the luxury of keeping their young children out of the workplace, namely, greater parental education and higher family income.

Table 1. Percentages (a) of Working Children by Major Occupations (b)

                                                 Urban

                                  Boys               Girls
Occupation                       Younger   Older    Younger   Older

Farmworker                          0.00     3.81      0.00     0.00
Fisherman                           0.00     1.90      0.00     0.00
Forest and livestock worker         0.00     0.00      0.00     0.00
Servant/maid                       21.05    12.38     68.75    43.18
Salesman/businessman               52.63    29.53      0.00     0.00
Production worker                  15.79    21.90     25.00    43.18
Transport/communication worker      5.26     7.62      0.00     0.00
Day laborer                         5.26     2.86      0.00     0.00
Other                               0.01    20.00      6.25    13.64
Total                             100.00   100.00    100.00   100.00

                                                 Rural

                                  Boys               Girls
Occupation                       Younger   Older    Younger   Older

Farmworker                         38.75    55.40      6.67    29.63
Fisherman                           6.25     5.83      6.67     3.70
Forest and livestock worker        10.00     2.04      0.00     0.00
Servant/maid                       13.75     7.00     53.33    40.74
Salesman/businessman               12.50    12.83      6.67     3.70
Production worker                   5.00     7.00      6.67     3.70
Transport/communication worker      2.50     1.46      0.00     0.00
Day laborer                         5.00     1.46      0.00     0.00
Other                               6.25     6.98     19.99    18.53
Total                             100.00   100.00    100.00   100.00

(a) These percentages are based on the number of working children
who provided information about their current occupations.

(b) Occupational titles are from the 1995-1996 Household
Expenditure Survey.

Table 2. Variable Definitions and Descriptive
Statistics for the Full Sample

                                Definition

Dependent Variable
    Work                        1 if child works;
                                  0 otherwise
Independent variables
  Poorest income                1 if income is in the lowest
    < 15,800 taka                 quintile; 0 otherwise
  Poor income [greater than     1 if income is in the
    or equal to] 15,800,          second-lowest
    <28,600 taka                  quintile; 0 otherwise
  Middle income [greater than   1 if the income is in
    or equal to] 28,600,          the third quintile;
    <43,745 taka                  0 otherwise
  Rich income [greater than     1 if the income is in
    or equal to] 43,745,          the fourth-lowest
    <72,429 taka                  quintile; 0 otherwise
  Richest income                1 if the income is
    [greater than or              in the top quintile;
    equal to] 72,429 taka         0 otherwise
  Age                           Child's age in years
  Boy                           1 if boy; 0 otherwise
  Eduyr                         Child's education in years
  Urban                         1 if urban; 0 otherwise
  HHSize                        Household size
  Headmale                      1 if head of household is male;
                                  0 otherwise
  MHeadedu                      Male head's years of education
  FHeadedu                      Female head's years of education
N = 11,282

                                    Mean
                                 (Standard
                                 Deviation)

Dependent Variable              0.09  (0.29)
    Work

Independent variables           0.20  (0.40)
  Poorest income
    < 15,800 taka
  Poor income [greater than     0.20  (0.40)
    or equal to] 15,800,
    <28,600 taka
  Middle income [greater than   0.20  (0.40)
    or equal to] 28,600,
    <43,745 taka
  Rich income [greater than     0.20  (0.40)
    or equal to] 43,745,
    <72,429 taka
  Richest income                0.20  (0.40)
    [greater than or
    equal to] 72,429 taka
  Age                           9.45  (2.71)
  Boy                           0.51  (0.50)
  Eduyr                         2.30  (2.54)
  Urban                         0.25  (0.43)
  HHSize                        6.56  (2.45)
  Headmale                      0.92  (0.27)
  MHeadedu                      4.79  (4.60)
  FHeadedu                      3.22  (1.57)
N = 11,282

Table 3. Logistic Regression Results for the Full Sample

                Full Sample
                 Logit (a)                  Younger Children
Variables      Coefficient        ME (b)  Logit Coefficient     ME

Intercept         -6.664 (0.291)          -5.999 *** (0.441)
Poorest        1.312 *** (0.168)   0.102   1.900 *** (0.277)   0.128
Poor           0.929 *** (0.161)   0.064   1.369 *** (0.266)   0.079
Middle         0.777 *** (0.163)   0.051   1.181 *** (0.268)   0.065
Rich           0.457 *** (0.158)   0.028   0.780 *** (0.257)   0.038
Age            0.328 *** (0.015)   0.017   0.208 *** (0.029)   0.008
Boy            0.791 *** (0.086)   0.043       0.062 (0.115)   0.002
Eduy          -0.154 *** (0.016)  -0.008      -0.041 (0.030)  -0.002
Urban          0.449 *** (0.102)   0.027    0.322 ** (0.140)   0.013
HHSize         0.330 *** (0.027)   0.018   0.456 *** (0.038)   0.017
Headmale      -1.128 *** (0.132)  -0.092  -1.641 *** (0.158)  -0.123
MHeadedu      -0.070 *** (0.011)  -0.004  -0.064 *** (0.016)  -0.002
FHeadedu      -0.280 *** (0.044)  -0.015  -0.354 *** (0.058)  -0.014
Log
  likelihood            -2898.91                    -1589.27
Chi-square            946.85 ***                  357.63 ***
N                         11,282                        7986

                    Older Children
Variables          Logit Coefficient       ME

Intercept            -7.301 *** (0.822)
Poorest               0.969 *** (0.213)   0.124
Poor                  0.648 *** (0.204)   0.076
Middle                0.551 *** (0.202)   0.063
Rich                      0.301 (0.192)   0.032
Age                   0.376 *** (0.059)   0.038
Boy                   1.627 *** (0.136)   0.164
Eduy                 -0.206 *** (0.021)  -0.021
Urban                 0.643 *** (0.132)   0.072
HHSize                0.192 *** (0.038)   0.019
Headmale             -0.530 *** (0.198)  -0.062
MHeadedu             -0.083 *** (0.015)  -0.008
FHeadedu             -0.199 *** (0.059)  -0.020
Log
  likelihood                  -1217.75
Chi-square                  445.70 ***
N                                 3296

(a) Logit coefficients are reported with standard errors shown in
parentheses. Standard errors are adjusted for clustering.

(b) Marginal effects (ME) are calculated as [delta]P(y = 1)/[delta]x =
[beta]P(1 - P) and are evaluated at the mean.

*** Significant at 0.01 level, ** significant at 0.05 level, and *
significant at 0.10 level.

Table 4. Variable Definitions and Descriptive Statistics for the Urban
Models (a)

                                                      Boys
                                               Younger       Older
Dependent variable
  Work                                       0.07 (0.24)   0.25 (0.43)

Independent variables
  Poorest (b) (1 if income                   0.23 (0.42)   0.19 (0.39)
    <18,610 taka; 0 otherwise)
  Poor (1 if income [greater than or         0.21 (0.41)   0.18 (0.38)
    equal to] 18,610, <36,000 taka;
    0 otherwise)
  Middle (1 if income [greater than or       0.21 (0.40)   0.18 (0.39)
    equal to] 36,600, <59,650 taka;
    0 otherwise)
  Rich (1 if income [greater than or         0.18 (0.39)   0.19 (0.40)
    equal to] 59,650, <105,600 taka;
    0 otherwise)
  Richest (1 if income [greater than or      0.18 (0.38)   0.26 (0.44)
    equal to] 105,600 taka; 0 otherwise)
  Age                                        8.10 (1.85)  12.89 (0.86)
  Eduyr                                      1.94 (1.90)   4.75 (3.34)
  HHSize                                     6.30 (2.38)   6.51 (2.47)
  Headmale                                   0.92 (0.27)   0.90 (0.30)
  MHeadedu                                   6.64 (5.04)   7.47 (5.10)
  FHeadedu                                   2.65 (1.48)   2.76 (1.57)
N                                               939           457

                                                      Girls
                                               Younger       Older

Dependent variable                           0.06 (0.23)   0.11 (0.31)
  Work

Independent variables                        0.23 (0.42)   0.13 (0.33)
  Poorest (b) (1 if income
    <18,610 taka; 0 otherwise)               0.19 (0.39)   0.17 (0.38)
  Poor (1 if income [greater than or
    equal to] 18,610, <36,000 taka;
    0 otherwise)                             0.20 (0.40)   0.23 (0.42)
  Middle (1 if income [greater than or
    equal to] 36,600, <59,650 taka;
    0 otherwise)                             0.20 (0.40)   0.24 (0.43)
  Rich (1 if income [greater than or
    equal to] 59,650, <105,600 taka;
    0 otherwise)                             0.18 (0.38)   0.23 (0.42)
  Richest (1 if income [greater than or
    equal to] 105,600 taka; 0 otherwise)     8.08 (1.86)  12.90 (0.84)
  Age                                        1.84 (1.97)   4.80 (3.39)
  Eduyr                                      6.49 (2.50)   6.83 (2.48)
  HHSize                                     0.92 (0.27)   0.91 (0.29)
  Headmale                                   6.88 (4.96)   7.92 (5.00)
  MHeadedu                                   3.78 (1.58)   3.96 (1.67)
  FHeadedu                                      940           450
N

(a) Data are means (standard deviations in parentheses).

(b) The income quintiles are on the basis of all urban households, so
the means in each subgroups will not necessarily be 20%.

Table 5. Logistic Regression Results for the Urban Models

                             Boys

                             Younger
Variables          Logit Coefficient (a)   ME (b)

Intercept          -3.987 *** (1.201)
Poorest               1.288 * (0.717)      0.077
Poor                    0.233 (0.693)      0.010
Middle                  0.299 (0.653)      0.013
Rich                    0.620 (0.623)      0.031
Age                     0.138 (0.089)      0.006
Eduyr                  -0.018 (0.091)     -0.001
HHSize              0.399 *** (0.100)      0.016
Headmale           -1.921 *** (0.420)     -0.171
MHeadedu               -0.061 (0.040)     -0.003
FHeadedu            -0.396 ** (0.181)     -0.016

Log likelihood                -194.85
Chi-square                  57.80 ***
N                                 939

                             Boys

                             Older
Variables          Logit Coefficient       ME

Intercept          -6.610 *** (2.057)
Poorest             1.490 *** (0.574)      0.264
Poor                    0.623 (0.591)      0.096
Middle                0.966 * (0.548)      0.158
Rich                    0.850 (0.531)      0.135
Age                 0.396 *** (0.153)      0.053
Eduyr              -0.312 *** (0.047)     -0.041
HHSize              0.549 *** (0.105)      0.074
Headmale           -1.257 *** (0.448)     -0.227
MHeadedu           -0.081 *** (0.031)     -0.011
FHeadedu           -0.450 *** (0.152)     -0.060

Log likelihood               -182.42
Chi-square                 126.39 ***
N                                457

                              Girls

                              Younger
Variables          Logit Coefficient       ME

Intercept          -6.705 *** (1.169)
Poorest               1.532 * (0.812)      0.069
Poor                 1.758 ** (0.810)      0.088
Middle                  0.668 (0.730)      0.023
Rich                    0.518 (0.728)      0.017
Age                 0.261 *** (0.076)      0.007
Eduyr                  -0.098 (0.086)     -0.003
HHSize              0.361 *** (0.096)      0.010
Headmale             -0.850 * (0.477)     -0.034
MHeadedu           -0.148 *** (0.040)     -0.004
FHeadedu               -0.054 (0.183)     -0.002

Log likelihood                 -171.5
Chi-square                  96.02 ***
N                                 940

                              Girls

                              Older
Variables          Logit Coefficient       ME

Intercept              -2.067 (2.709)
Poorest              2.082 ** (0.905)      0.225
Poor                    1.313 (0.873)      0.103
Middle                1.644 * (0.866)      0.135
Rich                    1.095 (0.916)      0.076
Age                     0.018 (0.201)      0.001
Eduyr                  -0.004 (0.057)     -0.0002
HHSize                  0.030 (0.146)      0.002
Headmale             -0.942 * (0.482)     -0.070
MHeadedu           -0.199 *** (0.041)     -0.010
FHeadedu                0.032 (0.186)      0.002

Log likelihood                -122.77
Chi-square                  43.60 ***
N                                 450

(a) Logit coefficients are reported with standard errors shown in
parentheses. Standard errors are adjusted for clustering.

(b) Marginal effects (ME) are calculated as [delta] P(y = 1)/[delta]x
= [beta]P(l -P) and are evaluated at the mean.

*** Significant at 0.01 level, ** significant at 0.05 level, and *
significant at 0.10 level.

Table 6. Variable Definitions and Descriptive Statistics for the
Rural Models (a)

                                                        Boys
                                                 Younger        Older

Dependent variable
  Work                                       0.07 (0.26)    0.29 (0.45)

Independent variables
  Poorest (b) (1 if income
    <15,308 taka; 0 otherwise)               0.21 (0.41)    0.17 (0.38)
  Poor (1 if income [greater than or
    equal to] 15,308, <27,200 taka;
    0 otherwise)                             0.20 (0.40)    0.19 (0.39)
  Middle (1 if income [greater than or
    equal to] 27,200, <40,270 taka;
    0 otherwise)                             0.20 (0.40)    0.20 (0.40)
  Rich (1 if income [greater than or
    equal to] 40,270, <64,800 taka;
    0 otherwise)                             0.20 (0.40)    0.21 (0.41)
  Richest (1 if income [greater than or
    equal to] 64,800 taka; 0 otherwise)      0.19 (0.39)    0.23 (0.42)
Age                                          8.05 (1.81)   12.86 (0.86)
Eduyr                                        1.61 (1.72)    3.44 (3.18)
HHSize                                       6.54 (2.41)    6.59 (2.45)
Headmale                                     0.93 (0.26)    0.92 (0.27)
MHeadedu                                     3.92 (4.10)    4.37 (4.23)
FHeadedu                                     2.71 (1.40)    2.70 (1.40)
N                                                3072           1270

                                                        Girls
                                                 Younger        Older

Dependent variable
  Work                                       0.05 (0.21)    0.06 (0.23)

Independent variables
  Poorest (b) (1 if income
    <15,308 taka; 0 otherwise)               0.21 (0.41)    0.16 (0.36)
  Poor (1 if income [greater than or
    equal to] 15,308, <27,200 taka;
    0 otherwise)                             0.21 (0.41)    0.18 (0.38)
  Middle (1 if income [greater than or
    equal to] 27,200, <40,270 taka;
    0 otherwise)                             0.20 (0.40)    0.20 (0.40)
  Rich (1 if income [greater than or
    equal to] 40,270, <64,800 taka;
    0 otherwise)                             0.19 (0.40)    0.21 (0.41)
  Richest (1 if income [greater than or
    equal to] 64,800 taka; 0 otherwise)      0.18 (0.39)    0.25 (0.43)
Age                                          8.03 (1.81)   12.81 (0.83)
Eduyr                                        1.58 (1.76)    3.53 (3.24)
HHSize                                       6.55 (2.43)    6.81 (2.55)
Headmale                                     0.92 (0.27)    0.91 (0.29)
MHeadedu                                     3.83 (4.10)    4.50 (4.49)
FHeadedu                                     3.74 (1.48)    3.69 (1.52)
N                                                3035           1119

(a) Data are means (standard deviations in parentheses).

(b) The income quintiles are on the basis of all rural households, so
the means in each subgroups will not necessarily be 20%.

Table 7. Logistic Regression Results for the Rural Models

                             Boys
                            Younger
Variables        Logit Coefficient (a)   ME (b)

Intercept        -7.230 *** (0.623)
Poorest           2.266 *** (0.405)      0.193
Poor              1.638 *** (0.387)      0.117
Middle            1.321 *** (0.400)      0.086
Rich              1.304 *** (0.366)      0.085
Age               0.319 *** (0.042)      0.014
Eduyr             -0.113 ** (0.046)     -0.005
HHSize            0.436 *** (0.056)      0.019
Headmale         -1.519 *** (0.244)     -0.123
MHeadedu         -0.078 *** (0.023)     -0.003
FHeadedu         -0.244 *** (0.086)     -0.011

Log likelihood               -683.5
Chi-square               200.04 ***
N                              3072

xx                           Boys
                             Older
Variables        Logit Coefficient          ME

Intercept        -7.442 *** (1.125)
Poorest            0.623 ** (0.272)      0.124
Poor                0.506 * (0.259)      0.099
Middle                0.275 (0.275)      0.052
Rich                  0.124 (0.254)      0.023
Age               0.541 *** (0.082)      0.098
Eduyr            -0.282 *** (0.028)     -0.051
HHSize              0.082 * (0.049)      0.015
Headmale              0.284 (0.270)      0.048
MHeadedu         -0.062 *** (0.021)     -0.011
FHeadedu          -0.194 ** (0.083)     -0.035

Log likelihood               -632.52
Chi-square               202.38 ***
N                              1270

                             Girls
                             Younger
Variables        Logit Coefficient      ME

Intercept        -4.824 *** (0.729)
Poorest           2.112 *** (0.445)     0.121
Poor              1.447 *** (0.443)     0.066
Middle            1.484 *** (0.439)     0.070
Rich                  0.591 (0.441)     0.021
Age                   0.060 (0.055)     0.002
Eduyr                 0.075 (0.053)     0.002
HHSize            0.574 *** (0.059)     0.017
Headmale         -2.050 *** (0.253)    -0.144
MHeadedu             -0.008 (0.029)   -0.0002
FHeadedu         -0.604 *** (0.090)    -0.018

Log likelihood              -499.71
Chi-square               152.37 ***
N                              3035

                             Girls
                             Older
Variables        Logit Coefficient      ME

Intercept        -5.882 *** (2.078)
Poorest           1.970 *** (0.608)     0.138
Poor              1.734 *** (0.574)     0.107
Middle                0.893 (0.638)     0.040
Rich                  0.441 (0.552)     0.017
Age                   0.144 (0.155)     0.005
Eduyr                 0.048 (0.047)     0.002
HHSize            0.424 *** (0.088)     0.015
Headmale         -1.743 *** (0.414)    -0.123
MHeadedu          -0.088 ** (0.040)    -0.003
FHeadedu         -0.332 *** (0.129)    -0.012

Log likelihood              -209.75
Chi-square                55.20 ***
N                              1119

(a) Logit coefficients are reported with standard errors shown in
 parentheses. Standard errors are adjusted for clustering.

(b) Marginal effects (ME) are calculated as [delta]P(y = 1)/[delta]x =
[beta] P(1 - P) and are evaluated at the mean.

*** Significant at 0.01 level, ** significant at 0.05 level, and *
significant at 0.10 level.

We wish to thank Lisa Jepsen, Ken Brown, M. Imam Alam, Stan Lyle. Amanda Bieghler, and an anonymous referee for their helpful comments.

(1) Bachman (2000, p. 32) pointed out that the definition of child labor "is fraught with anomalies and contradictions, reflecting a tangle of international standards, national laws, cultural practices and social expectations." We define child labor later in the paper.

(2) The proposed bill was a great enough threat to Bangladesh's export industries that the Bangladesh Garment Manufacturers and Exporters Association signed a Memorandum of Understanding with the ILO and UNICEF in July 1995. As a result, employers in the Bangladeshi garment industry eliminated 50,000 jobs held by children (Rahman, Khanam, and Absar 1999). For more discussion of the Harkin Bill, see McClintock (2001).

(3) We thank an anonymous referee for suggesting separate models for younger and older children.

(4) The survey requested information for all household members on their total income and the relevant sources of such income. A separate section of the questionnaire collected detailed information about the household's income from agricultural production. The income measure is fairly reliable since the discrepancy is not substantial when matched with expenditure and savings data. However, it should be noted that income is traditionally underreported in most surveys conducted in developing countries.

(5) We omitted 91 children from our empirical analysis because they were reported as household members in households in which they were employed. Thus, the data on family characteristics, such as the household head's education, applied to the head of the household employing these children rather than to their own father or other relative. In these cases, the link between household characteristics and the decision to have the child work is obscured. Fifty-three of the omitted children are girls, 13 from rural areas and 40 from urban areas. Thirty-eight are boys, 23 from rural areas and 15 from urban areas.

(6) We initially also defined child labor more narrowly as children who reported that they worked the previous week (category 1) and who reported positive income for the past year. There were only 78 working children according to this definition, which represents only 0.7 of 1% of those in the sample. This small number indicates the narrowness of this definition in light of the cultural norms in Bangladesh.

(7) This does not mean that children who do not report income are not paid; rather, their payment could come in the form of food, lodging, or clothing. This would apply to servants and maids, for example. Agricultural workers who do not receive cash payments may he paid with crops, fruit, or vegetables.

(8) Results show that when the natural log of income is used instead of dummy income quintile variables, the coefficient of the log of income is negative and statistically different from 0 at the 5% level or better for all models reported in Tables 3, 5, and 7. Income is thus negatively related to the probability that a child will work, and the effect lessens as the income level increases.

(9) Reporting errors resulted in cases for which years of education exceeded age minus 4 (the earliest age at which formal education could begin). For these cases, we defined Eduyr as Age -4.

(10) Studenmund (2001, p. 436) outlined several reasons for not using an OLS model. First, the unrestricted OLS model can predict a negative value or a value greater than 1 as the probability of labor force participation. Second, the OLS estimates are less efficient than the logit estimates because of heteroscedasticity in the disturbances. Third, the standard errors of the OLS parameters are not consistent, whereas asymptotic efficiency and consistency are well established properties of the maximum likelihood estimates.

(11) Marginal effects can be interpreted as follows using Table 3. The marginal effect of Age on the probability of working for the full sample is 0.0174. This means that a one-year increase in age will add about .02 to the probability that a child will work. For dummy variables, the interpretation of the marginal value is as follows using the dummy variable Boy from the same equation in Table 3 (Allison 1999). The marginal effect is 0.0425, meaning that, on average, a boy's probability of working is 0.04 of a percentage point higher than a girl's. There are alternative approaches to interpreting the marginal effects of dummy variables (Liao 1994).

(12) Forty urban girls were among the 91 cases omitted because they were reported as members of households where they were employed. These are cases where younger or older urban girls were employed as maids or servants. Thus, the maids and servants remaining in our sample of urban girls are those reported as members of their own and not their employers' households.

References

Allison, Paul D. 1999. Logistic regression using the SAS system: Theory and application. Cary, NC: SAS Institute Inc.

Bachman, S. L. 2000. The political economy of child labor and its impacts on international business. Business Economics, July, pp. 30-41.

Bangladesh Bureau of Statistics. 1998. Household Expenditure Survey, 1995-96. Dhaka, Bangladesh: Bangladesh Bureau of Statistics.

Basu, Kaushik. 1999. Child labor: Cause, consequence, and cure, with remarks on international labor standards. Journal of Economic Literature 37:1083-119.

Basu, Kaushik, and Pham Hoang Van. 1998. The economics of child labor. American Economic Review 88:413-27.

Bissell, Susan, and Barbar Sobhan. 1996. Child labour and education programming in the garment industry of Bangladesh: Experiences and issues. Dhaka, Bangladesh: UNICEF Education Section, Occasional Papers.

Delap, Emily. 2001. Economic and cultural forces in the child labour debate: Evidence from urban Bangladesh. Journal of Development Studies 37:1-22.

Grootaert, Christiaan. 1999. Child labor in Cote d'Ivoire. In The policy analysis of child labor: A comparative study, edited by Christiaan Grootaert and Harry Anthony Patrinos. New York: St. Martin's Press, pp. 23-62.

Harkin, Tom. 1999. The United States should ban imports of products made by children. In Child labor and sweatshops, edited by Mary E. Williams. San Diego: Geenhaven Press, pp. 38-42.

Hossain, B. B. M. Jaber. 1996. Child labor in Dhaka City: A case study. M.A. thesis, Jahangirnager University, Dahka, Bangladesh.

Liao, Tim Futing. 1994. Interpreting probability models: Logit, probit, and other generalized linear models. Thousand Oaks, CA: Sage Publications.

Maddala, G. S. 1988. Introduction to econometrics. New York: Macmillan.

McClintock, Brent. 2001. Trade as if children mattered. International Journal of Social Economics 28:899-910.

Rahman, Mohammad Mafizur, Rasheda Khanam, and Nur Uddin Absan 1999. Child labor in Bangladesh: A critical appraisal of Harkin's bill and the MOU-type schooling program. Journal of Economic Issues 33:985-1003.

Rahman, Wahidur. 1997. Child labour situation in Bangladesh: A rapid assessment. Dhaka, Bangladesh: International Labour Organization.

Ravallion, Martin, and Quentin Wodon. 2000. Does child labour displace schooling? Evidence on behavioral responses to an enrollment subsidy. Economic Journal 110:C158-75.

Ray, Ranjan. 2000a. Analysis of child labour in Peru and Pakistan: A comparative study. Journal of Population Economics 13:3-19.

Ray, Ranjan. 2000b. Child labor, child schooling, and their interaction with adult labor: Empirical evidence for Peru and Pakistan. World Bank Economic Review 14:347-67.

Studenmund, A. H. 2001. Using econometrics: A practical guide. 4th edition. Boston: Addison-Wesley Longman.

Wasserman, Miriam. 2000. Eliminating child labor. Regional Review (Federal Reserve Bank of Boston), 2nd quarter, pp. 8-17.

Shahina Amin, * M. Shakil Quayes, ([dagger]) and Janet M. Rives ([double dagger])

* Department of Economics, University of Northern Iowa, Cedar Falls, IA 50614-0129, USA; E-mail Shahina.Amin@ uni.edu; corresponding author.

([dagger]) Department of General Business, Southeastern Louisiana University, Hammond, LA 70402, USA; E-mail [e-mail].

([double dagger]) Department of Economics. University of Northern Iowa, Cedar Falls. 1A 50614-0129, USA; E-mail Janet. [e-mail].

Received July 2002; accepted June 2003.