Trade-off between Child Labour and Schooling in Bangladesh: Role of Parental Education

Full text

1
DEPARTMENT OF ECONOMICS
ISSN 1441-5429
DISCUSSION PAPER 21/11
Trade-off between Child Labour and Schooling in Bangladesh:
Role of Parental Education
*
Salma Ahmed
†
Abstract
The paper examines whether there is any trade-off between child labour hours and child schooling
outcomes. By drawing on Bangladesh National Child Labour Survey data, we find that children’s
work, even in limited amounts, does adversely affect child human capital. This is reflected in
reduced school attendance and age-adjusted school attendance rates. We find that parents do not
have identical preferences towards boys’ and girls’ schooling decisions. While both, educated
mother and father shifts the trade-off towards girls’ schooling as opposed to market work, the
differential impact of mother’s education on girls is significantly larger. These conclusions persist
even after allowing for sample selection into child’s work. Our results intensify the call for better
enforcement of compulsory schooling for children.
Keywords: Child labour, education, Bangladesh
JEL codes: J13, J22, J24, O12
*
We have benefitted from comments by Pushkar Maitra, Glen Harrison and participants at the Monday workshop in the
Department of Economics, Monash University, participants at the Australasian Development Economics Workshop
2011 and participants at the 40
th
Australian Conference of Economists 2011. The usual caveat applies.
†
Email: Salma.Ahmed@monash.edu
© 2011 Salma Ahmed
All rights reserved. No part of this paper may be reproduced in any form, or stored in a retrieval system, without the prior written
permission of the author.

2
1 Introduction
Child labour is not a new phenomenon in a low income economy where children are a current
economic resource for poor parents. Poverty is considered as the major driving force behind
child labour in low income countries (Maitra and Ray 2002; Ersado 2005). In under-
developed economies, where labour markets are usually quite imperfect; the poor households
want to send their children to work in order to escape extreme poverty. Such a view
underlines, for example, the Luxury Axiom
4
of Basu and Van (1998) and corroborates the
belief that, in developing economies, in the absence of a credit market, households,
particularly in rural areas, react to temporary income shortfalls by increasing their
dependence on child labour earnings (see also Bardhan and Udry 1999). Others argue that
factors such as „bequest constraint‟
5
of parents play a greater role in sending children to work
(Baland and Robinson 1998).
A key concern about child labour is whether the work activities of children hamper
their school performance. This is an important question from a policy perspective because if
working during school has a harmful effect on academic performance, it might be reasonable
to reinforce laws that eliminate practice of child labour. However, policy measures to curtail
child labour may not be justified in certain occasions when child‟s poor schooling prospects
results from low school quality or lack of access to school. Elimination of child labour would
also do little to improving school performance if parents send their least motivated children to
work. Therefore, disentangling the direction of causality is crucial to implementing right
policies.
The principal aim of this paper is to investigate whether hours of work are really a
trade-off with schooling outcomes of children using nationally representative unit record
dataset from Bangladesh National Child Labour Survey. In doing so, we examine the impact
of child labour hours on school attendance and age-adjusted school outcome variable – Grade
for Age (GAGE). Ceteris paribus, child workers who spend longer hours on work activities
will have little time for school attendance and studying. Exhaustion from longer hours of
work could also prevent the children from being attentive inside and outside classrooms with
4
A family will send the child to the labour market only if the family‟s income from non-child labour sources
drops very low.
5
The „bequest constraint‟ refers to the parents‟ understanding of possible future financial benefit that might
impact on their present decisions concerning their child schooling.

3
implications for their educational performance.
6
A few studies analyse the connection
between the number of hours worked by children and their schooling (Akabayashi and
Psacharopoulos 1999; Rosati and Rossi 2003; Ray and Lancaster 2005) and conclude that
there is a trade-off between schooling and waged labour. Parikh and Sadoulet (2005),
however, argue that since child work is responsive to opportunities of work, school
attendance and labour are not necessarily incompatible, concluding there to be no real trade-
off. Still others argue there is a trade-off between the quality of school and child labour due to
children‟s “dual commitment” at work and in schooling (see, for example, Heady 2003;
Orazem and Gunnarsson 2004).
While many existing studies find evidence of trade-off between child labour and
schooling, only a limited number of studies directly attempt to investigate this, especially in
the Bangladesh context (Ravallion and Wodon 2000; Arends-Kuenning and Amin 2004).
Existing studies on child labour in Bangladesh have explored mainly whether child work is a
deterrent or complements to school attendance or enrolments (see for example Amin, et al.
2004). Apart from the literature on the potentially negative effect of child labour on school
attendance and performance, there has been quite a different strand of literature which
document that girls are more likely than boys tend to combine schooling with work in rural
areas (Khanam 2008).
Although most work on child labour viewed the household as having only one set of
preferences, there is now an extensive literature that presents some evidence that male and
female household heads may have different preferences for the outcomes for their children.
The resolution of the preference difference of the male and female household heads may
depend on the relative bargaining power of each individual, and this power may depend on (i)
control over assets, both current and those brought into marriage (ii) unearned income or
transfer payments and welfare receipts (iii) access to social and interpersonal networks and
(iv) attitudinal attributes. Indeed, a recent paper by Reggio (2011) provides empirical
evidence that an increase in mother‟s bargaining power proxied by access to credit,
significantly reduces girls working hours. Other studies have shown that when the women‟s
power rise measured by levels of education, child labour initially fall but beyond a point it
will tend to rise again (Basu K. and Ray 2001; Basu K. 2006). None of these studies,
6
Of course, a number of other school related factors may also have a role to play.

4
however, tests the impact of balance of power on the investment in education of child
workers. Exception is Ridao-Cano (2001).
7
The present paper is different from these earlier studies in several aspects. First, we
analyse the trade-off between child labour hours and child schooling outcomes. In doing so,
we contend that parental attitudes and preferences may affect work-schooling trade-off and
used fathers‟ and mothers‟ levels of education as an indicator of attitudes and parental
preferences. Arguably, if parental education positively influences parental preferences for
children‟s education, then an increase in parental education resulting in more schooling and
less child labour, even in poor households. Alternatively, parental education may increase the
efficiency or effectiveness of the time spent interacting with children (e.g., directly helping
with school work), and more educated parents may thus forgo some time spent working in
order to make greater time investments in their children‟s human capital which in turn lead to
less work for children. However, it is often posited that more educated parents in poor
households without access to credit may face a trade-off between education and current
consumption; this does not necessarily mean that children of more educated parents are more
likely to go school. Indeed, depending on circumstances, caring parents might insist on their
children working, and on using the additional income to improve children‟s nutrition rather
than increasing expenditure on education. Yet, another possibility is that even during income
shocks (e.g., unemployment and natural disasters), a household with educated parents is less
likely to pull a child out of school, practice child labour or both because educated parents
have safety nets (e.g., insurance).
Second, we employ an instrumental variable (IV) estimation strategy as we presume
the endogeneity of hours worked of children in the structural equation of schooling outcomes.
In designing empirical framework, we do not assume that schooling and work decisions of
children are independent (e.g., Maitra and Ray 2002; Ersado 2005). We also do not want to
assume any sequential process in the decision making process as we believe it is not
necessarily a sequential choice.
8
In addition, we do not treat schooling and working
7
Using data from rural Bangladesh, the author analyses the determinants of child labour (e.g., participation in
farm work) and schooling (participation in school) and concludes that mothers have a higher preference for
child schooling than fathers. This difference is mainly revealed through the relative bargaining power of
mothers proxied by access to credit from group based credit programs. Conversely, mother‟s and father‟s access
to credit have no significant effect on child work.
8
This approach has some attractive modelling features but it necessitates rather strong assumption about the
sequencing of decisions. In particular, at the first stage parents have to first decide whether to send their children
to school only or to engage in other activities, then whether to combine school with other activities and finally

5
possibilities as two interdependent choices which explicitly accounts for the fact that
disturbances between the two outcomes are correlated. The correlation of unobservables,
however, from the two outcomes caution one from measuring the effect of child employment
on schooling outcomes since some portion of the relationship may be driven by exterior
factors (For example, such unobserved characteristics may be in form of the perceived
improvement of income opportunities in case of children‟s school attendance; perceived
availability of schooling relative to the urge of sending children into the labour force or
simply the parent‟s desire to send their children to school, an affect which may not be
captured by any of the available variables).
Finally, we explore an important issue that have had received very little attention in
most existing studies that estimate the effect of working hours on child outcomes
(Akabayashi and Psacharopoulos 1999; Ray and Lancaster 2005). Since we restrict our
analysis on child‟s working hours, this may introduce the well-known problem of sample
selection bias. Hours of work are only available for working children and children who work
different hours might have unobserved characteristics that could be correlated with the
unobservables in the outcome equation (here it is school), causing our estimates to be biased
and inconsistent. To account for sample selection bias, we use the Heckman sample selection
model. In addition, a double-hurdle model is employed for comparison purposes. We have
also calculated a likelihood ratio statistic to compare the sample selection and double-hurdle
estimates.
Specifically, in this paper we examine the following questions:
1. Is there any trade-off between work and schooling?
2. How does the relative parental education affect work-schooling trade-off?
3. How are the results affected if we explicitly take selection into child employment
into account?
The paper finds a trade-off between work and schooling. The working hours
adversely affect child schooling from the very first hour of work but weakens as labour hours
increase. This finding complements Ray and Lancaster (2005) which find that a small
increase in child labour is detrimental to child learning. The results of this paper indicate that
whether to send their children to work only. Such decision structures are artificially imposed and not appealing
on a priori grounds. The approach is applied in a number of studies to test the robustness of the results from the
sequential probit model (see, for example, Ersado 2005).

6
parental education plays an important role in influencing child‟s work-schooling trade-off.
While both, mothers and fathers schooling shifts the work-schooling trade-off in favour of
education, mother‟s educational attainment has relatively stronger marginal effects on work-
schooling trade-off than father‟s education. The paper also provides strong statistical
evidence of gender bias against a female child. Both mother and father show a significant
preference for educating a female child. The same incentive effect is not found for a male
child thus suggesting more market work for a male child. Finally, sample selection into child
employment has a significant impact on the work-schooling trade-off and the results suggest
that not controlling for sample selection is likely to bias our estimated results.
2 Background
This paper‟s focus upon child labour in a South Asian nation is motivated by the facts that the
majority of the Asian child workers come from South Asia (Ray 2004). As regards
Bangladesh, a National Child Labour Survey (NCLS) conducted in 2002-03 by the
Bangladesh Bureau of Statistics (BBS) under the auspices of the ILO sponsored International
Program on the Elimination of Child Labour (IPEC), found that about 5 million (14 percent)
of the total 35 million children between ages 5-14 were economically active
9
, of this total 3.5
million (71 percent) were boys and 1.5 million (29 percent) were girls. Except India, this is a
strikingly high rate – especially in comparison with other countries in the South Asia region
(see Table 1).
10
Official statistics had also shown that the total working children population
between ages 5-17 was about 7.9 million, of which 5.8 million (73 percent) were boys and
2.1 million (27 percent) were girls.
Table 1: Estimates of economically active children aged 5-14 in the South Asia
Region, by gender
Source: Ray (2004).
a
Sri Lankan Child Activity Status (SIMPOC 1999).
9
A person who works one or more hours for pay or profit or working in a family farm or found not working but
had a job or business from which he/she is temporarily absent during the last week of the survey.
10
Precise data on the number of Bangladeshi children in one of the worst forms of child labour are not available
because children engaged in prostitution, or drug trafficking are usually hidden and cannot be easily reached
through traditional surveys.

7
The widespread prevalence of child labour in Bangladesh, despite the government‟s
programs and laws prohibiting work by children, suggests that additional policy measures to
curb child labour are warranted.
11
A view held that there is a lack of harmony among laws
that uniformly prohibit the employment of children or set a minimum age for employment.
12
This is because these laws, however, focus mainly on the employment of children in the
factory, shop and establishment sectors ignoring the employment of children in the
agricultural sector, which absorbs about 56 percent of the total child labour force. In addition,
informal sector and household employment are exempted from these laws. Thus more than 80
percent of the economic activity of children falls outside the protection of the labour code. In
addition, Bangladesh signed a Memorandum of Understanding (MOU) in 1995 undertaken by
the ILO and the UNICEF to eliminate child labour in particular from garments industry. As
reported by Rahman, et al. (1999), this approach did not lead to a decline in child labour
among these children nor to a commensurate increase in their schooling.
13
A second MOU
was undertaken by the same parties in 2000 to reaffirm the agreements of the first MOU and
to develop a long-term and sustainable response to monitoring child labour in garment
industry (Khanam 2006).
Bangladesh also adopts provision of school subsidies to improve schooling so that it
will attract and retain them. The innovative program of this type is Food for Education (FEE),
which was introduced in 1993 and made available to rural children.
14
Ravallion and Woden
(2000) find that the FEE program has been successful to increase school enrolment (from
approximately 75 to 90 percent) but did not have much of an effect on child labour. Another
educational incentive program that encouraged girls to increase their secondary schooling
11
There are 25 special laws and ordinances in Bangladesh to protect and improve the status of children in
Bangladesh (Khanam 2006). The Employment of Children Act of 1938 prohibits children as young as 12 years
from being employed in leather tanning and in the production of carpets, cement, matches, and fireworks,
among other items. According to this law (as amended in 1974) the minimum employment age for work in
factories is14 years; for work in mines and railways, the minimum age is 15 years. The Factory Act of 1965 also
prohibits the employment of children below the age of 14 in any factory. The law further adds that young
workers (that is children and adolescents) are only allowed to work a maximum of 5 hours per day and only
between the hours of 7am and 7pm. The Children‟s Act of 1974 prohibits the employment of children less than
16 years of age in begging, and the exploitation of children in brothels (Khanam 2006).
12
Under the current law the legal minimum age for employment varies, according to sector, between 12 and 16
(Khanam 2006).
13
They reported that of 61,000 terminated from garment industries by 1996, only 1464 were placed in schools
by September 1996. Many of these children found jobs that were not poorly paid, but were also more dangerous
than garment work. These observations are consistent with Basu (1999) who suggest that sanctions would make
the child workers worse off. See also Udry (2003) for more discussion on this issue.
14
The main feature of the program is to provide a free monthly food ration contingent on the family being
judged as poor and having at least one primary-school-age (at least 6 years old) child attending school that
month.

8
(that is grade 6 to grade 10) (and delay marriage) was found to be effective in increasing
secondary school attendance for girls (see Arends-Kuenning and Amin 2004, for a survey).
No research thus far attempts to shed light on whether this particular subsidy for girls causes
less child labour.
As previously mentioned, this paper‟s empirical analysis is based upon the individual
level data from the 2002-03 National Child Labour (NCLS). The NCLS considers a child
(age 5-17) to be employed if he or she worked at least one hour during the reference week.
15
However, the survey does not consider child participation in domestic work to be labour. To
enable our empirical analysis, we define child labourers aged between 5 and 17 as children
working at least one hour during the reference week as a paid (wage) employee (paid in cash
or in kind) was self-employed
16
or worked as an unpaid employee (e.g., work on the family
farm or in the family business) related to the household head.
17
This is especially important as
globally only a relatively small fraction of children works for wages. Also, we follow the
definition of work similar to the NCLS, that is, exclude domestic work. For the estimation of
child labour, 5 years may be considered extreme. But it is not unusual in case of Bangladesh,
particularly in rural areas. On the other hand, although official enrolment age in Bangladesh
is 6 years, there are some children who start school at age 5 years. We also consider the
propensity of late enrolment, which is very common, especially in rural areas.
We use two measures of children‟s schooling – school attendance and grade for age,
though it is often posited that a more accurate assessment of the impact of child labour on
human capital development should focus on measures of learning outcomes, such as test
scores, rather than school enrolment or attendance. We depart from this practice for two main
reasons: (a) test scores are not available for children in the dataset considered here, and (b)
the reading, language and mathematical skills, which the test scores measure does not always
reflect the complete picture of learning achievements, especially in the context of a
developing country like Bangladesh where enrolling all school-aged children in school is still
a major development challenge for policy makers. In this survey, each child was asked
whether he/she attending school (full-time/part-time) at the time of the survey, whereas age-
15
Week preceding to the day of the survey.
16
A self-employed or own account worker is officially defined as a person who works for his/her own farm or
non-farm enterprise for profit or family gain.
17
NCLS 2002 classified children as sons and daughters if they are the son or daughter of the head of the
household or spouse. The father is called the head of the household if the head is identified as male and the
mother is called the spouse (if listed as the opposite sex), and the mother is called the head if the head is
identified as female and the father is called the spouse (if listed as the opposite sex).

9
adjusted measure of educational attainment () is defined as follows (Psacharopoulos
and Yang 1991):
 



  

 
where  is the highest grade of formal schooling attained by child i,  is child age,  is the
usual school entry age. All those with a score under 100 are considered as being below
normal progress in the school system because of grade repetition or late entry.
18
3 Data and descriptive statistics
The empirical analysis is based on the data drawn from the Bangladesh National Child
Labour Survey, conducted by the Bangladesh Bureau of Statistics (BBS) in 2002-03. This
survey has been designed in the context of the commitments made by the Government of
Bangladesh, following the ratification of the ILO Worst Forms of Child Labour Convention
(No. 182) 1999. The NCLS 2002 is designed to provide reliable estimates of child labour at
national, urban and rural levels, as well as by region. The survey covered the child population
aged 5 – 17 years living in the households, while children living in the streets or institutions
such as prisoners, orphanages or welfare centres are excluded. The sample size and the
coverage of the survey are such that it could furnish reliable key estimates by some
administrative units such as divisions and regions of the country. NCLS 2002 is undertaken
using Integrated Multipurpose Sample (IMPS) design, covering about 40000 households.
The analysis is performed upon a sample of 14062, 5-17 years old children drawn
from the survey‟s urban and rural respondents. In this sample, 9404 (67 percent) are males
and 4658 (33 percent) are females. Out of this sample, 2801 males and 1439 females reside in
urban areas, while 6603 males and 3219 females reside in rural areas. Of this sample, 8900
children are actively participating in the labour force, consisting of 6750 males and 2150
females. Of this sample of working children, 2508 children reside in urban areas while 6392
children reside in rural areas, 76 percent of both the urban and rural samples are male. Tables
2, 3, 4 and 5 present descriptive statistics regarding child employment, education, and
employment status by urban-rural residence and gender.
18
Ray and Lancaster (2005) employ the “schooling for age” (SAGE) variable that measures schooling
attainment relative to age. It is given by SAGE = Years of schooling/(Age-E) x100 where E represents the usual
school entry age in the country. SAGE could not be calculated in the present application because NCLS does not
report “years of schooling” as a continuous variable.

10
Table 2 presents definitions and descriptive statistics of the independent variables in
our analysis by child work status. Child-specific characteristics include child‟s age,
education, and working hours (per week). Descriptive statistics conditional on work status
suggest that working children are on average older and are generally combine school with
work than their non-working counterparts and the difference is generally statistically
significant at conventional levels.
19
In addition, a child gender dummy is included to capture
the gender disparities in education and work that may arise due to differences in parental
preferences. Indeed, the statistics indicate that there is a negative relationship between labour
supply and female child. It may be surmised that, the demand for female child labour is high
at home. At the household level, household composition and household assets are included.
Household composition includes number of adult males and females aged higher than 17
years, which may be negatively related to pressures upon the individual child. There is some
differential between working and non-working children, and the difference is statistically
significant at the 1 percent level. Since children must often care their younger siblings, the
number of younger children aged between 0 and 4 years is also included. It is evident from
the statistics that, as compared to children who do not work, children who supply labour
come from families with no-accommodation facilities, smaller land holdings and higher
number of school aged children and the difference is generally statistically significant. We
also included a set of parental characteristics that may influence parental decisions with
regard to child work. The statistics suggest that an improvement in parental education will
reduce child labour supply. This has important policy implications. Interestingly, there is
little difference between working and non-working children in the effect of father‟s
education, but the difference is never statistically significant.
The remaining measure includes a set of community variables that may influence
household decisions. These may include variables that may influence demand for child
labour. Hence, residence of the household is included as a regressor. The descriptive statistics
suggest that the rate of incidence of labour varies by urban and rural areas, but the evidence
confirms that the number of children is high and worthy of policy concern. Besides location,
a policy measure to reduce child labour is improvement in school quality. While we don‟t
have information on measures with direct bearing on student achievement, we explore
potential school input effects by including a set of dummy variables that capture the quality
of school. These are: presence of formal school administered by the government and the
19
We computed standard t-TEST.

11
NGO school run by the non-government organisations. On average, about 59 percent of
children who work go to the formal school, while the corresponding number for the non-
working children is about 21 percent. The difference is statistically significant at
conventional levels.
Table 3 presents the incidence of child labour force participation and school
attendance for children aged 5 to17. In all areas, the child participation rate in the labour
market increases with age, though not monotonically. In case of child schooling, the
attendance rate peaks around 12 years in urban and rural areas, and then falls. The gender
picture is similar in both urban and rural areas with respect to child labour with males
registering a higher participation than females. However, the situation differs sharply with
respect to child schooling with a more even gender imbalance in the attendance rate between
males and females in the later age groups of 12-17 years, the participation of females in
schooling falls than that of males. There are several possible reasons for this drop off. Girls
are separated away from male contact at an early age (on the basis of religion). Since there
are few primary schools, and even fewer secondary schools reserved for girls, young females
have to leave school on reaching adolescence. Another possible explanation is that it is
customary for girls to marry early, which tend to further curtail schooling.
Interestingly, the school attendance rates of rural children in almost all age groups are
consistently larger than their urban counterparts, with the former registering figures about 60
percent for males and 50 percent for females around 12 years and falls off sharply beyond 14
years. In case of urban areas, the attendance rate rarely goes above 50 percent and falls off
sharply beyond 14 years.
Table 4 shows employment status of children. In rural areas 52 percent of working
males were unpaid family workers and 40 percent are paid workers; in urban areas, however,
about 48 percent of working males were paid workers and 22 percent were unpaid family
workers. Similar patterns are not found for working female children. A large proportion of
female child workers work without pay in family related activities, and it is relatively high in
rural areas.
These patterns suggest that opportunities for child labourers are quite different in rural
and urban areas. In rural areas, children are more likely to engage in agricultural activities
and become unpaid “family helpers”, especially female children; in urban areas, children are

12
more likely to find opportunities for some paid work. The gender difference in employment
status among child labourers is also significant in Bangladesh. Young females are more likely
than young males to be unpaid family worker in both urban and rural areas. This may imply
that male children are increasingly entering the formal wage labour market rather than
working as unpaid family workers, and thus allowing female children to substitute into the
unpaid family related activities.
Table 4: The employment status of working children aged 5-17 in urban and rural areas,
by gender
___________________________________________________________________________
Urban Rural
___________________________ ________________________________
Male Female Male Female
_____________ _____________ ____________ ____________________
No. (%) No. (%) No. (%) No. (%)
__________________________________________________________________________________________
Paid employed 918 48.01 165 27.68 1947 40.24 216 13.90
Self-employed 164 8.58 27 4.53 363 7.50 39 2.51
Unpaid family worker 830 43.41 404 67.79 2528 52.25 1299 83.59
Total 1912 100.00 596 100.00 4838 100.00 1554 100.00
__________________________________________________________________________________________
Table 5 shows that industries that employ child labour are quite different for males
and females in urban and rural areas. Although the agricultural sector is large, males are more
likely to work in wholesale and retail industries in urban areas and females are more likely to
work in the agricultural industries in rural areas. In urban areas 44 percent of working males
were employed in the wholesale and retail industry, followed by 23 percent in the
manufacturing industry and 22 percent in agricultural industry, 42 percent of working
females were employed in the manufacturing industry, followed by 18 per cent in wholesale
and retail. In rural areas, 63 per cent of working males are employed in agriculture, followed
by 21 percent in the wholesale and retail industry; 71 percent of working females were
employed in agriculture, followed by 15 percent in manufacturing.

13
Table 5: Employment of children aged 5-17 in urban and rural areas, by gender and industry
__________________________________________________________________________________
Urban Rural
___________________________ ________________________________
Male Female Male Female
_____________ _____________ ____________ ____________________
No. (%) No. (%) No. (%) No. (%)
_________________________________________________________________________________
Agriculture 411 21.50 173 29.03 3059 63.23 1099 70.72
Manufacturing 446 23.33 252 42.28 489 10.11 237 15.25
Construction 103 5.39 30 5.03 150 3.10 26 1.67
Wholesale and retail 848 44.35 108 18.12 1017 21.02 136 8.75
Services 104 5.44 33 5.54 123 2.54 179 3.60
Total 1912 100.00 596 100.00 4838 100.00 1554 100.00
__________________________________________________________________________________________
Figure 1 shows the interaction between work hours and school attendance, which is
important for the purposes of this paper because of the direct effects of working hours on
school outcomes. The more children have to work, the more tired they will be when in school
and the less time they will have for study. Consequently, work may have an adverse effect on
learning while in school, even if it does not have a large effect on enrolment. In Bangladesh,
there appears to be an hour‟s threshold beyond which work is strongly associated with
reduced school attendance. Work hours appear to have a relatively small impact on school
attendance up to 15-29 hours cohort, but attendance falls of dramatically when children work
greater than 29 hours per week. The decline is more gradual for girls than for boys, perhaps
because some kinds of chores and subsistence work are more compatible with school than
male children‟s labour force work.
4 An analytical framework
In general, a unitary model (or common preference model) is typically used to analyse the
economic contribution of children.
20
Here a household maximises the joint utility function of
all its members. Quantity and quality of children, consumption of leisure and other market
goods all enter the household utility function.
There is also an extensive literature that suggests that male and female heads of the
household may have different utilities, reservation utilities, and budget constraints and
therefore may make different decisions (see, for example, McElroy and Horney 1981). In this
20
Unitary models assume that either all household members share the same preference function or that a single
decision maker takes all decisions. In any case, the upshot is that the household behaves as if it were a single or
a unitary agent (Becker 1981).

14
context, we allowed that parental preferences may not be conjugal and differences may exist
in parental preferences regarding child‟s work and schooling due to differences in attitudinal
attributes, but where, for simplicity, all children are treated equally. In the absence of direct
information on parental preferences as well as non-earned income of each spouse, we shall, in
this paper, consider differential role of father‟s and mother‟s education. We propose that
women‟s education increases women‟s bargaining power in household decisions and if the
mother obtains higher marginal disutility from child labour than the father, an increase in the
bargaining power would lead to a reduction in the amount of her child‟s work and increase in
child schooling. However, this preference will not be reflected in household outcomes if
mothers cannot participate in household decision making on equal terms.
To fix ideas, assuming a household comprised of father and mother and some number
of children, who can be sons and daughters in particular point in time. Fertility is assumed
exogenous. In general, each parent is considered altruistic in that they care about the
consumption of each member of the household and the quality (educational attainment) of
their children. All decisions are made by altruistic parents, and children are treated as
recipients.
21
The parents allocate the child‟s total endowment of time between school
attendance and work in the labour market. Leisure is left out of the analysis for
simplification. Household income must meet the cost of household consumption and
schooling. Household‟s income is generated by a typical household production function with
decreasing returns. Assume it is a function of parent‟s non-labour income, parent‟s labour
income and the child labour earnings. We shall maintain the strong assumption that non-
labour income is exogenous and therefore ignore the fact that current unearned income
probably reflects past labour supply decisions. While primary education is almost free in
Bangladesh, schooling costs can be significant in terms of costs on transportation, school
uniform, utensils etc. especially for the poor. Parental investment for a given level of
schooling would depend on a vector of individual, household and community characteristics,
household income, parental education and the child labour supply. Child labour supply may
be depending on a vector individual, household and community characteristics, household
income and parental education.
This simplified framework form the basis of our empirical analysis discussed in the
following section.
21
In line with the literature, it is assumed that children do not bargain with their parents because they do not
have a valid fall back option (Bhalotra 2007).

15
5 Empirical model 
Let 


denote „desired‟ educational attainment of child . The equation for 


is written as
follows:



 

 



 




 



 



 




 


 



 




 


+ 



 



 

  

 

(1)
where 

represents child ‟s   weekly working hours in the previous week and
its quadratic term 


captures non-linear effects of hours worked, 

is the dummy for female
child. 

and 

denote fathers and mothers education measured by highest grade attained.
As discussed in section 4, father‟s and mother‟s education may affect investments in male
and female child differently, the gender of the sampled child is interacted with each parent‟s
education 

  . We also include a dummy variable taking the value of one if the
mother‟s education level is higher than that of father and zero otherwise 

. The mothers‟
education relative to fathers‟ is assumed to influence her ability to negotiate in defence of her
preferences with respect to her children‟s schooling.  is a vector of exogenous child,
household and community characteristics and

the random factor.
The basic setup of equation (1) is applied to estimating equations for two dependant
variables: school attendance and the age-adjusted school attendance – grade-for-age ().
The equations are estimated in two specifications, the first dependent variable is estimated by
probit and the second dependent variable is estimated using tobit because GAGE has
observed zero values for more than 35 percent of children.
22
Let assume a child is attending
school only if the parents‟ desired educational investment is positive. That is,


 



 


where 

  if child  is attending school, and 

  otherwise.
One potential concern of estimating equation  is that child labour hours and child
labour hours (squared) may not be exogenous. Specifically, some unobserved factors that
determine school attainment may also explain hours worked, causing our probit or tobit
22
We observed zero values due to censoring.

16
estimates to be biased and inconsistent.
23
To avoid this endogeneity problem, we use two-
stage instrumental variable (IV) strategy. There is however one another problem. Note that
household‟s income is most likely to be endogenous, since whether children work or not is
likely to influence parents‟ reservation wages and the labour market participation of the
parents (Wahba 2006). Another potential problem with using household income is that it
might not always reflect household welfare in developing countries like Bangladesh where
subsistence agriculture is common, and households consume what they produce. To avoid
endogeneity of this variable, we include occupation status of the father to proxy wealth.
The reduced-form work equation is written as follows:


 





 

 



 



 



 





 



 






 



 

  

(2)
where 

is the instrumental variables and 

the random factor.
The instrumental variable (IV) procedure can be described as follows: In the first
stage, we estimate equation (2) by OLS and obtain the residual (. We follow the same
procedure when the child labour hours (squared) is the dependent variable. In the second
stage, we included predicted residuals from first stage regressions into equation (1) as
additional regressors. The significant coefficients on residuals imply that the null hypothesis
of exogeneity of child labour hours is rejected. Equation (1) can now be written as follows:



 

 



 




 




 





 



 



 



 






 



 

  



 



 

  

 

(3)
where 


and 



are predicted residuals from the OLS estimates of child labour hours and
child labour hours (squared) of equation (2).
6 Results and analysis
We now turn to regression results. To examine whether there is any trade-off between
working hours and studying and to test the impact of parental education on the trade-off
between working and studying by gender an IV approach is implemented to take account of
23
The unobserved determinants could be person specific or individual attributes – such as “motivation” or
“energy” – that might drive certain children to both work more and study more.

17
possible endogeneity bias between educational outcomes and child labour hours.
24
To solve
the endogeneity problem, we relied on instruments such as a set of industry dummies
(agriculture, manufacturing, wholesale and retail and service) where the child works.
Industrial sector is a possible determinant of child working hours and, controlling for all else,
there seems no reason for child schooling to be influenced by child‟s sector of employment.
There is widespread evidence that children working hours differ significantly across sector of
employment (Guarcello, et al. 2004) because some sectors require more hours of work than
the other sectors. For example, using the Cambodian Child Labour Survey (CCLS),
Guarcello, et al. (2004) find that children aged 12-14 working long hours (e.g., 40 hours or
more per week) outside the agriculture sector. Guarcello, et al. (2004) also observe a similar
picture in Brazil, where around one in three working children in commerce, services and
manufacturing sectors put in a work week of at least 40 hours. Therefore, although working
hours vary somewhat by sectors of employment, the general feature is that working in
different sectors require an elements of physical activity and do not provide time that can be
used for studying. Thus, there is no obvious reason that a child in any particular sector would
be at a disadvantage because he/she finishes work especially exhausted or because he/she
receives less study time than the non-working children. Thus, it seems quite possible to
believe that working in different sectors do not have a direct effect on school attendance.
The relevance of the instruments is checked through empirical tests. The relevant test
lends strong credence to our use of industry dummies as instruments for both “child labour
hours” and “child labour hours (squared)” in the estimation of all school regressions (with a
p-value <0.000).
25
Over-identification tests are also conducted to evaluate whether the
proposed instruments can sensibly be excluded from both “school attendance” and “grade-
for-age (GAGE)” equations. The Hansen J-statistic concludes that all instruments can
legitimately be excluded in the estimation of both “school attendance” and “grade-for-age”
equations (results are provided in Tables 6 and 8). We estimate two sets of regressions: an IV
probit equation characterising the school attendance and an IV tobit equation characterising
grade-for-age (GAGE). The full set of coefficient estimates for child labour hours and child
labour hours (squared) are presented in the appendix, Table A1.
24
Hausman tests reveal that the null hypothesis of exogeneity of child labours hours and its square term is
rejected in the estimation of both “school attendance” and “grade-for-age” equations (results are provided in
Tables 6 and 8).
25
To understand whether these instruments are strong, we perform -test examining that coefficients on
instruments are strong. Under null hypothesis, the instruments are uncorrelated with the error term in the
estimation of both “school attendance” and “grade-for-age” equations (results are provided in Table A1).

18
School attendance
Table 6 presents the results of school attendance using the IV-probit for all children.
26
We
also present the marginal effects in Table 7 as they are more easily interpreted.
27
The
marginal effects of each variable on the probabilities of attending school presented in Table 7
show the following.
In column (1) we include child‟s age, its square term and the dummy for female child
but exclude education of the mother and father and other covariates as explanatory variables
in the regression. There is a significant age effect – age is positive and statistically
significant. This implies that probabilities of school attendance increases with age of the
child. The square of the age of the child show that there is a non-linearity in the age effect
(the square of the age term turns out to be negative and statistically significant). Being a
„female child‟ reduce the chance of being in school. This is a common scenario in
Bangladesh and in other South Asian countries that there is a significant gender differential in
the probability of school attendance and the bias is generally in favour of male children. The
marginal effects (evaluated at the mean of the independent variables) show that being a
female child is associated with an almost 17 percentage points decrease in the probability of
school attendance.
The coefficient estimates for „child labour hours‟ and „child labour hours (squared)‟
are statistically significant, but are of opposite signs at conventional levels. The negative
magnitude of the estimated coefficients of the work hours variable support the proposition
that work hours adversely affect the probability of the child attending school from the very
first hour of work. However, the estimated positive coefficients of “child labour hours
(squared)” suggest that the adverse marginal impact of child labour hours on the schooling
variable weakens as the labour hours increase. This clearly indicates that the hours worked by
children has a U-shaped relationship with schooling outcomes. The result is consistent with
evidence from Belize, Cambodia and Panama showing that child labour has a significant U-
shaped relationship with child education (Ray and Lancaster 2005). The marginal effects with
respect to hours worked suggest that beyond 7 hours a day (43 hours per week) the marginal
26
Given space constraints we do not present actual probit estimates of the determinants of child‟s school
attendance. These are however available on request.
27
The marginal effects for dummy variables are computed as


as  changes from zero to unity, holding other explanatory
variables at their sample means.

19
impact changes direction; the child labour hours impact positively on school attendance for
all children.
28
While column (1) establishes the negative link between working hours and school
attendance, it does not account for other important covariates which might to some extent
involve in low school attendance. In column (2), therefore, other covariates are included as
regressors. The marginal effects further confirm the U-shaped relationship between child
schooling and child labour hours, and the effect is now remarkably robust, namely 6.8 hours a
day. The coefficient for a female child is lower but significant. We note that presence of
school children between 5 and 17 reduces the probability of school attendance. This might be
a reflection of household budget constraint. However, there is strong evidence that presence
of adult female members in the household increases the probability of child school
attendance. This effect might reflect the larger decision making power of females among
adults in the household. We also note that father‟s occupation reduces the probability of
attending school if father involved in agricultural activities (the reference category is non-
agricultural activities). There are two routes whereby father‟s agricultural activities may
affect his child‟s schooling. First, a person with agricultural activities may be resource
constrained compared to those involved in non-agricultural activities and may decide that it is
not worthwhile to send his children to school at all. Second, agricultural activities in
Bangladesh is characterised by seasonal variation and therefore it is not uncommon for
families to engage in non-agricultural activities to supplement household income.
Involvement in non-agricultural activities by adult members increases the demand for child
labour in activities where child and adult labour are substitutes. As expected, the probability
of child school attendance increases in urban areas than in rural areas.
Next we turn to the analysis of how parental education influences the nature of trade-
off between working hours and school attendance. In column (3), we use the first
comprehensive indicator of mother status – if mother more educated than father. In this
specification, we do not include father‟s schooling to avoid possible correlation with
mother‟s schooling, if there is assortative mating. In general the results confirm that if
mother education is higher than that of father it promotes child welfare by reducing children‟s
hours of work and by increasing the school attainment. The effect, however, is not stronger
28
These figures are obtained by calculating 






where 



are the estimated coefficients of “work hours”
and “ work hours square” in the relevant regression. See equation (3) above.

20
for a female child as indicated by the interaction of the female dummy with mother‟s
education.
In column (4) other indicator of mother status, the level of schooling of the mother
(measured by highest grade attained) is included to check the robustness of our results
presented in column (3). We observe that child labour hours and schooling retain a U-shaped
relationship. Mother‟s schooling shows a greater positive effect on the schooling of male and
female children but has a differentially higher effect on the female child shown by the
positive coefficient for the interaction of the female dummy with the mother‟s schooling. The
result further corroborates the findings that children‟s school attainment is linked to the
educational attainment of the parent of the same sex as the child (see, for example, Emerson
and Souza 2002).
Column (5) includes father‟s schooling (measured by highest grade attained) to
account for differences in father‟s and mother‟s education in child outcomes. In sum, the
results suggest that father‟s education has a greater positive impact on his daughter‟s school
attendance than on a son‟s and shifts the trade-off towards females‟ schooling as opposed to
market work. This is not the effect one might expect in South Asian context where fathers
tend to favour sons relative to daughters. The positive and significant coefficient of
interaction term between female child and father‟s schooling might reflect the effect of
scholarship programs in Bangladesh designed to keep girls in school. By contrast, the effect
of mother‟s education though increases the probability of school attendance for all children;
the preference for educating a female child is virtually disappear when controlled with
father‟s education. Indeed, the chi-square test for the parents‟ schooling coefficients show
that effects of father‟s and mother‟s schooling differ significantly from each other, while
taking into account interactions with child gender the effect of father‟s schooling does not
differ significantly from mother‟s schooling (see Table 10). We check the robustness of these
results in column (6) by including additional covariates, namely household composition,
household assets and community characteristics. If the results remain unchanged even with
the inclusion of these variables, it enhances robustness of the results shown in column (5).
Importantly, in column (6), the same pattern emerges as in column (5). Two
observations are noticeable from these results: The estimates strengthen the negative impact
of child labour on schooling when all explanatory variables are controlled. It is the female
child‟s schooling relative to the male child‟s schooling that reacts to father‟s education and

21
thus show a strong substitution effect on girl‟s work-schooling trade-off, while the
substituting nature of father‟s education on the work-schooling trade-off is less pronounced
for the male child.
Grade-for-age
Further evidence on the adverse impact of child labour on child schooling is shown in Table
8, which presents the IV-tobit estimates of grade-for-age.
29
To better understand the results,
we limit our discussion to marginal effects presented in Table 9. In column (1) the marginal
effects of IV estimates further confirm the U-shaped relationship between child schooling and
child labour hours. The turning point at which child work starts to impact negatively on
education is 6.7 hours a day (40.4 hours per week). As expected from the analysis in the last
section, being a female child is associated with a lower grade attainment compared to their
male counterparts. In column (2) we include other covariates but exclude indicators for
mother‟s and father‟s education to check whether the conclusions drawn in column (1) is
remain unchanged. Surprisingly, being a female child does not affect grade attainment and
child labour hours do not have any significant effect on child schooling. Another useful piece
of evidence that the results provide relate to the household land-holdings. The results clearly
show the incentive effects that arise from large land-holdings. While income increases with
land ownership, for any given level of income, children are more likely to go to school and
less likely to do market work.
A further robustness check for whether child labour hours have negative impact on
child schooling would be inclusion of parental education in the regression. In column (3), the
analysis repeats the estimates presented in column (1), but with additional variable „mother
more educated than father‟. The results are similar to those drawn in column (3) in Table 7,
that is, if mother‟s level of education is higher than that of father‟s, it raises grade attainment
for both male and female children and therefore shifts the work-schooling trade-off more in
favour of schooling. The analysis in column (4) includes mother‟s education represented by
highest grade and shows a positive and significant effect of mother‟s education on grade
attainment of her children. However, the effect is stronger for educating a female child. In
column (5), interestingly, inclusion of father‟s education shows that there is no significant
change on the coefficients for mother‟s education (measured by highest grade) though the
29
Actual tobit estimates of the determinants of GAGE are available on request.

22
magnitude of the coefficient is now smaller.
30
Also, the magnitude of the impact of father‟s
education is significantly smaller than that of mother‟s. Father‟s education however doesn‟t
reveal any bias towards a female child.
The final robustness check includes all covariates and presents the results in column
(6). The qualitative results regarding parental education remain in most cases except that
father‟s preference for investing in a female child is now significant.
7 Selection into employment?
As our estimates refer only to the subsample of children working in economic activity, this
could generate a selection bias in the estimates. Children who participate may have
unobserved characteristics that are correlated with the unobservables in the outcome equation
(here it is school).
31
We hence address this selection bias by using the Heckman two-step
sample selection model. The first step involves the estimation of a selection equation and the
second step involves the estimation of main equation (school outcome). The selection should
be on hours worked, as we could imagine that children working different hours might share
different characteristics. The selection equation should then be based on a tobit model.
However, the selection variable would be either 0 for non-workers and positive for the other
observations. One simple way to estimate the model is to reformulate the tobit model as a
probit model, selecting on a variable defined as 1 for working children and 0 for non-working
children. This is the approach followed in the paper.
32
Hence, the selection equation is:


 




 

 


 

(4)
where 

and 

are the latent and observed working hours of child  respectively. We
observe child work participation 

  if 

 

 

 . The vector 

denotes the
aggregate form of the explanatory variables and 





. We compute the inverse
Mills ratio  from a selection equation (4), which is included as an additional regressor in
equation (3):
30
However, these findings must be treated with some caution due to the effect of ability bias or the effect of
assortative mating on the intergenerational transmission of schooling.
31
Working children, for example, who are in school are often found to have lower test scores than non-working
children. However, it may be wrong to say that working leads to poor test outcomes, based simply on that
information. It may be that children with less scholastic aptitude “select” into the state of working, while more
intellectually capable children remain in school.
32
Theoretically this approach may sacrifice some efficiency by discarding information on the dependent
variable. However, this is not necessary true in a finite sample (Green 1997).

23



 

 



 




 




 





 



 



 



 






 



 

  



 



 

  

 



 

(5)
Although the tobit model has been used to estimate GAGE in equation (3), it is no
longer reliable in sample selection framework. The Heckman model assumes that the errors
of the participation and schooling outcomes equations are correlated and the participation
decision dominates the schooling outcomes. Dominance implies that observed zero values for
GAGE are the result of participation decision only and that once the first hurdle (that is,
participation) is passed censoring is no longer relevant. This implies that only individuals
with positive values for GAGE are included in the GAGE equation, as zeros are not
generated by the participation decision. This is typical Heckman‟s generalised sample
selection model (Jones 1989). In our case, we simplify the Heckman model by assuming that
the participation and schooling outcomes equations are independent. In this case, the model
reduces a probit for participation (equation 4) and an ordinary least squares using the
subsample when GAGE is greater than zero.
As is well known, the Heckman model requires an exclusion restriction, in the form of
one or more variables that appear in the participation equation but not in the schooling
outcomes equation.
33
We include ownership of dwelling (accommodation) and the sex of
household head as the exclusion restrictions. These variables have a well-established effect
on the decision of sending a child to work, but there is very little reason to expect that these
variables will have an effect on the probability of school outcomes, which is market
determined (and beyond the control of any individual).
How important is the selection effect?
The results from the second-stage of Heckman selection model of school attendance and
grade-for-age are shown in Tables 11 and 13. Marginal effects of the corresponding
estimations for school attendance are also reported (see Table 12). The signs of the estimated
coefficient on the inverse Mills ratio are of interest here. The results suggest that any estimate
33
In principle, instruments are not needed. The Heckman sample selection model can be identified by non-
linearity of the inverse Mills ratio. But estimates of the Heckman model often lead to substantial collinearity
between predicted inverse Mills ratio term and the remaining covariates in the outcome equation (school) when
a common set of covariates was entered into probit and outcome equations.

24
of the determinants of either the probability to go to school or age-adjusted grade attainment
are indeed different from those determining probability of child work participation.
34
In general, IV estimates are largely similar to the Heckman estimates. We find that
with the inclusion of Heckman correction term (inverse Mills ratio), the magnitude of
coefficients changes (as well as their standard errors), but the overall pattern are not
dramatically different. The pattern of significant coefficients is largely unaffected.
We next turn to the marginal effects of the probability of school attendance adjusted
for sample selection bias (Table 12). In particular, we note that U-shaped relationship
between child labour and child‟s schooling remains even after correcting for sample selection
bias. As expected from the analysis of column 3 of Table 7, higher level of schooling of the
mother compared to father has a positive and significant effect on school attendance for all
children (see column 3 of Table 12). In column 4, we exclude the variable “mother more
educated than father” but include the level of schooling of the mother measured by highest
grade attained. The magnitude of the impact of mother‟s education on the female child is
smaller but continues to be significant. As before, the impact of mother‟s education continues
to be significant when we included father‟s education in the regression (see column 5).
Father‟s education still shows significant preference for educating a female child. Finally the
model is estimated with all covariates to ensure that the results are not being biased by the
potential sample selection (column 6). When we compare these results with those presented
in column 6 of Table 7, although the results are qualitatively similar, the magnitude of the
impact of father‟s education on school attendance of a female child is now smaller.
Turning to the OLS estimates of age-adjusted school attendance corrected for sample
selection bias (Table 13), we reveal interesting and important differences from the IV
estimates presented in Table 8. First, the OLS estimates show that, adjusted for sample
selection bias, if mother is more educated than father, it doesn‟t have any impact on her
child‟s education (column 3). Next, the impact of father‟s education turns out to be negative
and statistically insignificant when we included it in the regression with mother‟s education
(column 5) and continues to be statistically insignificant when the model is estimated with all
34
It is worth noting that inverse Mills ratio is always positive and statistically significant in all equations. It
implies that there are unmeasured effects which tend both towards making an individual more likely to
participate in work yet towards making that individual‟s levels of school attainment higher. This is probably
because children who are in work also attending school part-time but not full-time.

25
covariates (column 6). Hence, the sample selection correction does lead to substantive
differences in inference.
Comparison with double-hurdle model
In this subsection we compare the Heckman selection model with the double-hurdle model.
In Heckman selection model, we estimate GAGE equation by only using positive
observations, but it could lead biased results. In order to solve this problem one need to
analyse the underlying reasons for observed zero values. One potential reason could be that
many of the respondents are not attending school during the particular survey period (that is
their highest grade should be zero); we therefore consider the possibility of corner solution.
This implies that zero values are the result of participation decision and the survey
respondents may have zero values for GAGE. The reason motivates us to use a different
approach to estimate the GAGE equation. The approach can be analysed by a double-hurdle
framework. Thus in order to observe positive values for GAGE one need to pass two hurdles:
the participation hurdle and the schooling outcome hurdle (that is, GAGE). In estimating the
double-hurdle model, it is common to assume that errors terms from equations (4) and (5) are
independent (Cragg 1971). The model is essentially then a two-step procedure with a probit
for probability of participation in the first stage and truncated normal regression in the second
stage. In contrast to Heckman selection model, the double-hurdle model does not require the
exclusion restrictions. Yet a specification issue in double-hurdle models concerns the choice
of the regressors to be included in participation and schooling outcome equations. Indeed,
Aristei, et al. (2008) argued that the inclusion of the same set of regressors in each hurdle
makes it difficult to identify the parameters of the model correctly and exclusion restrictions
must be imposed. The exclusion restrictions in our analysis will be the same as above –
ownership of dwelling (accommodation) and the sex of household head.
Generally we identify substantive differences between the estimates from these two
modelling approaches in several cases (see Table A3). First, the double-hurdle estimates
show that the coefficients for a female child are negative but never statistically significant. In
the Heckman selection model, the coefficients for a female child are negative and statistically
significantly different from the coefficients for a male child (the reference group). Second, in
the double-hurdle estimates in Table A3, the adverse effect of working hours on child‟s grade
attainment turn out to be statistically insignificant in certain occasions. In the Heckman
selection estimates in Table 13, the U-shaped relationship between child schooling and child

26
labour hours remain in nearly all specifications. Third, in the double-hurdle model, the
presence of school aged children between 5 and 17 years old in the household significantly
reduces the probability of school attendance, whereas the effect of the presence of school
aged children is never statistically significant in the Heckman model. Finally, in the double-
hurdle model, the effect of mother‟s education (measured by highest grade) turns out to be
insignificant when controlled with father‟s education (measured by highest grade). In
addition, father‟s education now shows a positive effect on his child‟s grade attainment but
continues to be statistically insignificant. However, the effect of father‟s education now
becomes statistically significant when the model is estimated with all regressors (column 6 of
Table A3). In the Heckman model, the effect of mother‟s education (measured by highest
grade) continues to be significant when controlled with father‟s education (measured by
highest grade), while the effect of father‟s education is negative and statistically insignificant.
Finally, we compute a likelihood ratio (LR) statistics to test whether the double-
hurdle model is significantly different from the Heckman selection model. In doing so, we
treat the likelihood function from the Heckman selection model as the restricted likelihood
while the likelihood function from the double-hurdle model as an unrestricted one. The test
statistic is 266.02, which is chi-squared with one degree of freedom. So we can reject the null
hypothesis that the restricted model is true. The double-hurdle model is therefore more
appropriate than the Heckman selection model.
8 Robustness and alternative specifications
This section discusses robustness of the results to the definition of child labour. It also
considers how the results change if the analysis is conducted separately for urban and rural
areas. The choice made in defining child labour was justified in section 2. The equations were
re-estimated restricting the sample to wage (paid) work. Wage work involves longer hours
than other sorts of work (see Figure 2), and virtually rules out school attendance. In contrast,
other forms of child work may be more compatible with schooling. Isolating wage work
therefore allows us to control the confounding influence of working hours on children‟s
schooling.
Results are slightly different when restricting the sample to child wage workers. We
do obtain significant results with the number of hours worked for school attendance outcomes
(see Table A4). Conversely, a different pattern emerges with the tobit regressions for GAGE

27
with the number of hours worked as we now obtain statistically insignificant effect of
working hours on child‟s grade attainment (Table A5). When we look at the results for
parental characteristics we find that the effect of mother‟s education (measured by highest
grade) continues to be significant with and without control for father‟s education in both
school attendance and GAGE regressions, while the effect of father‟s education though
significant in GAGE equations, it turns out to be statistically insignificant in school
attendance equations. Interestingly, mother‟s education does not show any preference for
educating a female child in all specifications under consideration, while father‟s education
though show a significant preference for educating a female child in school attendance
equation, the effect is now become statistically insignificant when controlled with mother‟s
education in GAGE equations.
There also exist some differences when comparing urban and rural locations, which
may be due to the pattern of activities available to the children in the two areas. Looking first
at the results for school attendance (Table A6), we find that child labour hours and school
attendance retain a U-shaped relationship in both areas. Results are reversed with the GAGE
regressions (see Table A7), as the adverse effect of working hours on grade attainment is to
some extent significant among the rural subsample. Thus working hours seem more
detrimental for children living in rural areas. Turning next to parental education, we reveal
that both mother‟s and father‟s education show a positive and significant effect on the school
attendance and grade attainment of male and female children in rural areas, while a similar
effect is not found in urban areas specially for child‟s school attendance. It is important to
note that mother‟s education (measured by highest grade) does not show any bias towards a
female child in both school attendance and GAGE regressions particularly in urban areas,
while for father‟s education the opposite is true except for the grade attainment. Conversely,
in rural areas mother‟s education (measured by highest grade) does show a significant
preference for educating a female child in all regressions under consideration but virtually
disappear when father‟s education and other covariates are controlled in the school
attendance regressions.
9 Concluding comments and policy implications
The paper investigates the trade-off between child labour hours and child schooling outcomes
in Bangladesh. Controlling for a large number of covariates and correcting for all sources of
endogeneity bias by instrumenting child working hours with a set of industry dummies, there

28
is evidence that children‟s work, even in limited amount, does affect child education,
reflected in reduced school attendance and age-adjusted school attendance rates.
We find that parental characteristics crucially their human capital affect work-
schooling trade-off across gender. Mother‟s education show a significant preference in
investing in female child‟s education while more educated mother compared to father don‟t
reveal any preferences toward a female child. A similar impact is found for age-adjusted
school attainment. However, the effect of mother‟s education virtually disappears when
father‟s education and other covariates are controlled particularly in the estimation of school
attendance. We find that father‟s education do affect daughter‟s school attendance relative to
sons. The interpretation of the results is unchanged when the model is estimated with full set
of covariates. The magnitude of the impact of father‟s education is larger in age-adjusted
grade attainment and shows significant preference for educating a female child. After
correcting for potential source of selection bias, the qualitative results remain for the school
attendance regression but moderately change for age-adjusted grade attainment regression.
The results have strong policy implications. We find that both mother‟s and father‟s
education have a positive impact on school attendance of a female child than of a male child
though the marginal impact of mother‟s education on a female child‟s schooling is
considerably larger. The findings therefore consistent with our previous proposition that
increases in women‟s schooling and consequent increases in bargaining power have greater
beneficial impact on children compared with increases in men‟s schooling. One other result is
worth noting for policy implication. Parent‟s schooling is more sensibly react to male child‟s
labour rather than schooling. Although the Bangladesh government adopt a variety of
initiatives to ease the problem of child labour, the results strongly suggest that, there is
limited relevance of these policies to supplement household income which ultimately results
in higher incidence of child labour particularly among young males. We therefore urge for
combination of policies such as stronger enforcement of compulsory schooling for children
and employment generation scheme for adults to reduce child labour.

29
Table 2: Descriptive statistics by child work status
*** p<0.01, ** p<0.05, * p<0.1
Note: Informal school: government informal activities. t-TEST for difference (working – non-working children).

30
Table 3: Participation and attendance rates of children aged 5-17 in urban and rural areas, by gender
____________________________________________________________________________________________________________________
Age Urban Rural Urban Rural
_______________________ ________________________________ _____________________________ ___________________________
Male Female Male Female Male Female Male Female
__________ _____________ ______________ _________________ ____________ _____________ ___________ _______________
No Attd. No Attd. No Attd. No Attd. No Part. No Part. No Part. No Part.
(%) (%) (%) (%) (%) (%) (%) (%)
___________________________________________________________________________________________________________________________________________
5 292 0.00 246 0.81 565 0.53 582 1.72 292 1.03 246 0.81 565 1.42 582 2.41
6 191 0.52 125 1.60 281 2.85 233 3.00 191 2.09 125 1.60 281 3.20 233 4.29
7 102 7.84 106 6.60 205 5.37 201 10.95 102 10.78 106 3.77 205 7.32 201 6.47
8 78 10.26 70 15.71 193 22.28 117 13.68 78 20.51 70 11.43 193 30.05 117 13.68
9 57 14.04 40 22.50 134 18.66 96 25.00 57 42.11 40 27.50 134 48.51 96 29.17
10 133 12.03 69 23.19 304 17.43 154 16.23 133 75.94 69 30.43 304 75.00 154 33.12
11 83 22.89 31 29.03 188 29.79 85 8.24 83 79.52 31 61.29 188 77.13 85 65.88
12 404 49.01 143 41.26 1112 59.08 412 49.27 404 91.09 143 70.63 1112 91.10 412 82.04
13 271 49.82 92 32.16 662 54.08 297 45.79 271 95.20 92 89.13 662 95.62 297 89.90
14 351 38.18 179 36.31 909 44.44 392 32.40 351 91.74 179 82.12 909 93.95 392 88.27
15 312 11.54 137 12.41 810 11.48 274 3.65 312 90.38 137 56.93 810 89.63 274 64.96
16 287 13.24 124 8.06 685 13.14 211 8.06 287 85.37 124 64.52 685 87.74 211 65.40
17 240 15.00 77 6.49 555 13.69 165 9.09 240 88.33 77 53.25 555 87.03 165 60.00
Total 2801 22.74 1439 16.82 6603 28.43 3219 19.23 2801 68.26 1439 41.42 6603 73.27 3219 48.28
____________________________________________________________________________________________________________________
Note: School attendance rate refers to the number of 5-17 years old children attending school expressed as a percentage of total children in this age group.

31
Figure 1: Work hours and school attendance of children aged 5-17, by gender
Figure 2: Distribution of weekly work hours of children aged 5-17, by work status
0
20 40
60 80
Percentage
1-14 15-29 30-35 36-42 43-50 50+
male female
0
20 40
60 80
Percentage
1-14 15-29 30-35 36-42 43-50 50+
weekly work hours
paid employee
self-employed
unpaid family worker

32
Table 6: IV probit results of school attendance
_______________________________________________________________________________________________________________________
Continued on next page
Variables (1) (2) (3) (4) (5) (6)
Child age 0.5874*** -0.1723** 0.5855*** 0.5763*** 0.5750*** -0.1667**
(0.0574) (0.0817) (0.0574) (0.0577) (0.0579) (0.0827)
Child age (squared) -0.0202*** 0.0027 -0.0199*** -0.0199*** -0.0201*** 0.0024
(0.0027) (0.0029) (0.0027) (0.0026) (0.0026) (0.0029)
Female child -0.5796*** -0.5085*** -0.5957*** -0.6262*** -0.6549*** -0.5854***
(0.0629) (0.0884) (0.0640) (0.0726) (0.0795) (0.1075)
Child labour hours -0.3512*** -0.3728*** -0.3602*** -0.3479*** -0.3402*** -0.3688***
(0.0681) (0.0876) (0.0685) (0.0728) (0.0745) (0.0906)
Child labour hours (squared) 0.0040*** 0.0043*** 0.0041*** 0.0039*** 0.0038*** 0.0043***
(0.0009) (0.0011) (0.0009) (0.0009) (0.0010) (0.0012)
Residual_child labour hours 0.2423*** 0.2713*** 0.2514*** 0.2433*** 0.2368*** 0.2689***
(0.0682) (0.0876) (0.0686) (0.0728) (0.0746) (0.0906)
Residual_child labour hours (squared) -0.0030*** -0.0035*** -0.0031*** -0.0030*** -0.0029*** -0.0034***
(0.0009) (0.0011) (0.0009) (0.0009) (0.0010) (0.0012)
Number of Children 0-4 years -0.0138 -0.0100
(0.0283) (0.0283)
Number of School Children 5-17 years -0.0440** -0.0412**
(0.0182) (0.0182)
Number of adult males higher than 17 years -0.0287 -0.0267
(0.0259) (0.0261)
Number of adult females higher than 17 years 0.0935*** 0.0900***
(0.0333) (0.0335)
Occupation of father -0.1510*** -0.1764***
(0.0503) (0.0508)
Household uses piped water -0.0421 -0.0668
(0.1350) (0.1378)
Household has a television -0.0444 -0.0693
(0.0839) (0.0802)
Household has a radio 0.0053 -0.0156
(0.0463) (0.0469)
Household has a bicycle 0.1950*** 0.1835***
(0.0492) (0.0491)

33
Table 6 (continued): IV probit results of school attendance
*** p<0.01, ** p<0.05, * p<0.1
Note: Standard errors in parenthesis.
Variables (1) (2) (3) (4) (5) (6)
Formal school 2.6106*** 2.5905***
(0.1345) (0.1329)
NGO school 2.1430*** 2.1600***
(0.1804) (0.1813)
Own marginal land if acre of land less than 0.5 0.0314 0.0552
(0.0698) (0.0692)
Own large land if acre of land higher than 2 0.0204 -0.0024
(0.0686) (0.0685)
Urban 0.1244* 0.1341**
(0.0668) (0.0666)
Mother more educated than father 0.1970**
(0.0893)
Mother more educated than father x Female 0.1806
(0.1709)
Mother's education (highest grade) 0.1585*** 0.1007** 0.0445
(0.0501) (0.0401) (0.0393)
Mother's education x Female 0.1337** 0.0526 0.0350
(0.0558) (0.0594) (0.0697)
Father's education (highest grade) 0.0818*** 0.0180
(0.0271) (0.0243)
Father's education x Female 0.0842** 0.1056**
(0.0386) (0.0443)
Constant 1.1254 5.2092*** 1.2318 1.1624 1.0794 5.1391***
(0.8389) (0.8126) (0.8412) (0.8575) (0.8728) (0.8396)
Hausman test of endogeneity: c2(2) 15.55 9.69 17.03 19.85 20.73 8.95
R>c2 (p = 0.0004)
(p = 0.0078)
(p = 0.0002)
(p = 0.0000)
(p = 0.0000)
(p = 0.0114)
Hansen J-statistic (Test of overidentifying restrictions) 257.46 8.25 246.85 217.30 164.41 7.60
(p = 0.0000)
(p = 0.0161)
(p = 0.0000)
(p = 0.0000)
(p = 0.0000)
(p = 0.0223)
Observations 8,900 8,900 8,900 8,900 8,900 8,900

34
Table 7: Marginal effects of school attendance
____________________________________________________________________________________________________________________
Continued on next page
Variables (1) (2) (3) (4) (5) (6)
Child age 0.2010*** -0.0378** 0.2003*** 0.1973*** 0.1967*** -0.0367**
(0.0196) (0.0175) (0.0196) (0.0198) (0.0198) (0.0178)
Child age (squared) -0.0069*** 0.0006 -0.0068*** -0.0068*** -0.0069*** 0.0005
(0.0009) (0.0006) (0.0009) (0.0009) (0.0009) (0.0006)
Female child -0.1792*** -0.0963*** -0.1835*** -0.1919*** -0.1994*** -0.1088***
(0.0172) (0.0146) (0.0173) (0.0194) (0.0209) (0.0170)
Child labour hours -0.1202*** -0.0818*** -0.1232*** -0.1191*** -0.1164*** -0.0812***
(0.0233) (0.0194) (0.0234) (0.0249) (0.0254) (0.0201)
Child labour hours (squared) 0.0014*** 0.0010*** 0.0014*** 0.0013*** 0.0013*** 0.0009***
(0.0003) (0.0002) (0.0003) (0.0003) (0.0003) (0.0003)
Residual_child labour hours 0.0829*** 0.0595*** 0.0860*** 0.0833*** 0.0810*** 0.0592***
(0.0233) (0.0193) (0.0234) (0.0249) (0.0255) (0.0200)
Residual_child labour hours (squared) -0.0010*** -0.0008*** -0.0011*** -0.0010*** -0.0010*** -0.0008***
(0.0003) (0.0002) (0.0003) (0.0003) (0.0003) (0.0003)
Number of children 0-4 years -0.0030 -0.0022
(0.0062) (0.0062)
Number of school children 5-17 years -0.0096** -0.0091**
(0.0040) (0.0040)
Number of adult males higher than 17 years -0.0063 -0.0059
(0.0057) (0.0058)
Number of adult females higher than 17 years 0.0205*** 0.0198***
(0.0073) (0.0074)
Occupation of father -0.0331*** -0.0388***
(0.0111) (0.0112)
Household uses piped water -0.0090 -0.0142
(0.0283) (0.0283)
Household has a television -0.0096 -0.0148
(0.0178) (0.0167)

35
Table 7 (continued): Marginal effects of school attendance
*** p<0.01, ** p<0.05, * p<0.1
Note: Standard errors in parenthesis.
Variables (1) (2) (3) (4) (5) (6)
Household has a radio 0.0012 -0.0034
(0.0102) (0.0102)
Household has a bicycle 0.0454*** 0.0427***
(0.0122) (0.0121)
Formal school 0.4848*** 0.4825***
(0.0200) (0.0197)
NGO school 0.7156*** 0.7196***
(0.0421) (0.0415)
Own marginal land if acre of land less than 0.5 0.0069 0.0121
(0.0152) (0.0150)
Own large land if acre of land higher than 2 0.0045 -0.0005
(0.0153) (0.0150)
Urban 0.0273* 0.0295**
(0.0147) (0.0147)
Mother more educated than father 0.0704**
(0.0332)
Mother more educated than father x Female 0.0645
(0.0635)
Mother's education (highest grade) 0.0542*** 0.0344** 0.0098
(0.0172) (0.0138) (0.0087)
Mother's education x Female 0.0458** 0.0180 0.0077
(0.0191) (0.0203) (0.0153)
Father's education (highest grade) 0.0280*** 0.0040
(0.0093) (0.0054)
Father's education x Female 0.0288** 0.0232**
(0.0132) (0.0098)
Observations 8,900 8,900 8,900 8,900 8,900 8,900

36
Table 8: IV tobit results of grade-for-age (GAGE)
___________________________________________________________________________________________________________________________________
Continued on next page
Variables (1) (2) (3) (4) (5) (6)
Child age 6.8397*** 1.0532*** 6.7652*** 6.2226*** 6.0699*** 1.0162**
(0.5446) (0.4017) (0.5426) (0.5297) (0.5250) (0.4013)
Child age (squared) -0.2276*** -0.0444*** -0.2241*** -0.2085*** -0.2054*** -0.0427***
(0.0238) (0.0149) (0.0237) (0.0225) (0.0223) (0.0147)
Female child -2.1515*** -0.2783 -2.2658*** -2.4606*** -2.4676*** -0.7363
(0.5115) (0.3988) (0.5147) (0.5568) (0.5962) (0.4689)
Child labour hours -1.4242*** -0.5444 -1.4426*** -1.1211** -1.0182** -0.5333
(0.4930) (0.3684) (0.4913) (0.5042) (0.5094) (0.3749)
Child labour hours (squared) 0.0175*** 0.0066 0.0177*** 0.0134** 0.0120* 0.0065
(0.0063) (0.0048) (0.0063) (0.0065) (0.0066) (0.0048)
Residual_child labour hours 0.8567* 0.4875 0.8814* 0.6356 0.5577 0.4866
(0.4941) (0.3691) (0.4924) (0.5053) (0.5105) (0.3756)
Residual_child labour hours (squared) -0.0126** -0.0063 -0.0128** -0.0091 -0.0080 -0.0063
(0.0063) (0.0048) (0.0063) (0.0065) (0.0066) (0.0048)
Number of children 0-4 years 0.0629 0.1081
(0.1346) (0.1336)
Number of school children 5-17 years -0.2870*** -0.2511***
(0.0868) (0.0857)
Number of adult males higher than 17 years -0.0216 0.0136
(0.1310) (0.1307)
Number of adult females higher than 17 years 0.2681 0.2468
(0.1671) (0.1660)
Occupation of father 0.1891 0.0208
(0.2402) (0.2401)
Household uses piped water 0.2794 0.0252
(0.6498) (0.6457)
Household has a television 1.2718*** 0.8215**
(0.3975) (0.3800)
Household has a radio 0.2026 0.0629
(0.2191) (0.2186)
Household has a bicycle 0.2047 0.0585
(0.2307) (0.2276)

37
Table 8 (continued): IV tobit results of grade-for-age (GAGE)
*** p<0.01, ** p<0.05, * p<0.1
Note: Standard errors in parenthesis.
Variables (1) (2) (3) (4) (5) (6)
Formal school 19.9456*** 19.5910***
(0.4859) (0.4714)
NGO school -0.6676 -0.5936
(0.9997) (0.9947)
Own marginal land if acre of land less than 0.5 -0.6443** -0.4269
(0.3187) (0.3129)
Own large land if acre of land higher than 2 0.6239** 0.3435
(0.3081) (0.3029)
Urban 0.1739 0.2982
(0.3101) (0.3059)
Mother more educated than father 4.0506***
(0.7860)
Mother more educated than father x Female 2.3450
(1.6505)
Mother's education (highest grade) 3.2142*** 1.7673*** 0.6667***
(0.3650) (0.3100) (0.1937)
Mother's education x Female 1.7313*** 1.3975*** 0.8232**
(0.4664) (0.5224) (0.3673)
Father's education (highest grade) 1.6818*** 0.2746**
(0.2106) (0.1210)
Father's education x Female 0.2258 0.4248*
(0.3330) (0.2293)
Constant -15.2544** 3.1575 -14.7513** -15.5740** -16.1523** 2.9338
(6.4927) (3.7397) (6.4618) (6.3679) (6.3884) (3.7979)
Hausman test of endogeneity: c2(2) 9.69 0.89 8.77 3.84 2.90 0.84
R>c2 (p = 0.0001) (p = 0.4101) (p = 0.0002)
(p = 0.0215)
(p = 0.0551) (p = 0.4302)
Hansen J-statistic (Test of overidentifying restrictions) 16.11 8.78 15.19 11.74 10.18 5.41
(p = 0.0003) (p = 0.0123) (p = 0.0005)
(p = 0.0028)
(p = 0.0061) (p = 0.0669)
Observations 8,900 8,900 8,900 8,900 8,900 8,900

38
Table 9: Marginal effects of grade-for-age (GAGE)
_________________________________________________________________________________________________________________________________
Continued on next page
Variables (1) (2) (3) (4) (5) (6)
Child age 4.1965*** 0.7846*** 4.1561*** 3.8612*** 3.7826*** 0.7597**
(0.3351) (0.2993) (0.3344) (0.3297) (0.3281) (0.3001)
Child age (squared) -0.1396*** -0.0331*** -0.1376*** -0.1294*** -0.1280*** -0.0319***
(0.0146) (0.0111) (0.0146) (0.0140) (0.0139) (0.0110)
Female child -1.2936*** -0.2067 -1.3626*** -1.4914*** -1.5018*** -0.5461
(0.3014) (0.2953) (0.3032) (0.3299) (0.3546) (0.3451)
Child labour hours -0.8738*** -0.4056 -0.8863*** -0.6957** -0.6346** -0.3987
(0.3025) (0.2744) (0.3018) (0.3129) (0.3175) (0.2803)
Child labour hours (squared) 0.0108*** 0.0049 0.0109*** 0.0083** 0.0075* 0.0049
(0.0039) (0.0035) (0.0039) (0.0040) (0.0041) (0.0036)
Residual_child labour hours 0.5256* 0.3632 0.5414* 0.3944 0.3476 0.3638
(0.3032) (0.2750) (0.3025) (0.3135) (0.3181) (0.2808)
Residual_child labour hours (squared) -0.0077** -0.0047 -0.0078** -0.0056 -0.0050 -0.0047
(0.0039) (0.0036) (0.0039) (0.0040) (0.0041) (0.0036)
Number of children 0-4 years 0.0468 0.0808
(0.1003) (0.0998)
Number of school children 5-17 years -0.2138*** -0.1877***
(0.0646) (0.0641)
Number of adult males higher than 17 years -0.0161 0.0101
(0.0976) (0.0977)
Number of adult females higher than 17 years 0.1997 0.1845
(0.1245) (0.1241)
Occupation of father 0.1409 0.0155
(0.1790) (0.1795)
Household uses piped water 0.2093 0.0189
(0.4893) (0.4832)
Household has a television 0.9649*** 0.6215**
(0.3069) (0.2908)

39
Table 9 (continued): Marginal effects of grade-for-age (GAGE)
*** p<0.01, ** p<0.05, * p<0.1
Note: Standard errors in parenthesis.
Variables (1) (2) (3) (4) (5) (6)
Household has a radio 0.1512 0.0471
(0.1639) (0.1636)
Household has a bicycle 0.1529 0.0438
(0.1728) (0.1704)
Formal school 13.5135*** 13.3352***
(0.3016) (0.2930)
NGO school -0.4906 -0.4385
(0.7244) (0.7256)
Own marginal land if acre of land less than 0.5 -0.4812** -0.3197
(0.2387) (0.2348)
Own large land if acre of land higher than 2 0.4691** 0.2581
(0.2339) (0.2288)
Urban 0.1296 0.2229
(0.2310) (0.2287)
Mother more educated than father 2.6536***
(0.5467)
Mother more educated than father x Female 1.5005
(1.0983)
Mother's education (highest grade) 1.9945*** 1.1014*** 0.4984***
(0.2265) (0.1932) (0.1448)
Mother's education x Female 1.0743*** 0.8709*** 0.6154**
(0.2896) (0.3256) (0.2747)
Father's education (highest grade) 1.0481*** 0.2053**
(0.1312) (0.0905)
Father's education x Female 0.1407 0.3176*
(0.2075) (0.1715)
Observations 8,900 8,900 8,900 8,900 8,900 8,900

40
Table 10: Chi-square Tests Results
__________________________________________________________________________________
Chi-Square P>Chi-square
__________________________________________________________________________________
School attendance (column 5)
a
Tests on
Father‟s education=Mother‟s education 53.64 0.000
Tests on interaction terms with female child
dummy
Father‟s education=Mother‟s education 0.54 0.580
School attendance (column 6)
a
Tests on
Father‟s education=Mother‟s education 32.07 0.000
Tests on interaction terms with female child
dummy
Father‟s education=Mother‟s education 25.23 0.000
Grade-for-age (column 5)
b
Tests on
Father‟s education=Mother‟s education 12.98 0.000
Tests on interaction terms with female child
dummy
Father‟s education=Mother‟s education 111.76 0.000
Grade-for-age (column 6)
b
Tests on
Father‟s education=Mother‟s education 32.66 0.000
Tests on interaction terms with female child
dummy
Father‟s education=Mother‟s education 51.61 0.000
____________________________________________________________________________________________________
a
IV-probit results of school attendance,
b
IV-tobit results of grade-for-age (GAGE).

41
Table 11: Selectivity adjusted estimates of school attendance
______________________________________________________________________________________________________________________
Continued on next page
Variables (1) (2) (3) (4) (5) (6)
Child age 1.1104*** 1.3177*** 1.1446*** 1.1720*** 1.2030*** 1.3552***
(0.1938) (0.2411) (0.1937) (0.1946) (0.1941) (0.2400)
Child age (squared) -0.0387*** -0.0497*** -0.0398*** -0.0411*** -0.0424*** -0.0512***
(0.0071) (0.0085) (0.0071) (0.0071) (0.0071) (0.0085)
Female child -0.7735*** -1.0576*** -0.8018*** -0.8449*** -0.8811*** -1.1326***
(0.0931) (0.1219) (0.0935) (0.0996) (0.1038) (0.1350)
Child labour hours -0.3521*** -0.3899*** -0.3613*** -0.3489*** -0.3415*** -0.3866***
(0.0682) (0.0881) (0.0686) (0.0729) (0.0746) (0.0911)
Child labour hours (squared) 0.0040*** 0.0045*** 0.0041*** 0.0039*** 0.0038*** 0.0045***
(0.0009) (0.0011) (0.0009) (0.0009) (0.0010) (0.0012)
Residual_child labour hours 0.2429*** 0.2861*** 0.2521*** 0.2439*** 0.2377*** 0.2843***
(0.0683) (0.0881) (0.0686) (0.0729) (0.0747) (0.0911)
Residual_child labour hours (squared) -0.0030*** -0.0037*** -0.0031*** -0.0030*** -0.0029*** -0.0036***
(0.0009) (0.0011) (0.0009) (0.0009) (0.0010) (0.0012)
Number of children 0-4 years -0.0166 -0.0139
(0.0285) (0.0285)
Number of school Children 5-17 years -0.0111 -0.0079
(0.0190) (0.0189)
Number of adult males higher than 17 years -0.0542** -0.0521**
(0.0262) (0.0264)
Number of adult females higher than 17 years 0.0674** 0.0634*
(0.0336) (0.0338)
Occupation of father -0.3125*** -0.3384***
(0.0562) (0.0565)
Household uses piped water 0.0282 0.0082
(0.1353) (0.1382)
Household has a television -0.0434 -0.0636
(0.0843) (0.0806)
Household has a radio 0.0539 0.0355
(0.0471) (0.0478)
Household has a bicycle 0.1813*** 0.1715***
(0.0495) (0.0494)

42
Table 11 (continued): Selectivity adjusted estimates of school attendance
*** p<0.01, ** p<0.05, * p<0.1
Note: Standard errors in parenthesis.
Variables (1) (2) (3) (4) (5) (6)
Formal school 2.5666*** 0.0138
(0.1372) (0.0698)
NGO school 2.2482*** 0.0340
(0.1833) (0.0691)
Own marginal land if acre of land less than 0.5 -0.0036 2.5532***
(0.0703) (0.1357)
Own large land if acre of land higher than 2 0.0536 2.2710***
(0.0691) (0.1844)
Urban 0.1413** 0.1491**
(0.0671) (0.0669)
Inverse Mills Ratio 2.8588*** 8.1705*** 3.0579*** 3.2600*** 3.4373*** 8.3641***
(1.0095) (1.2467) (1.0096) (1.0144) (1.0116) (1.2411)
Mother more educated than father 0.1963**
(0.0894)
Mother more educated than father x Female 0.1409
(0.1717)
Mother's education (highest grade) 0.1523*** 0.0996** 0.0505
(0.0502) (0.0402) (0.0396)
Mother's education x Female 0.1239** 0.0544 0.0412
(0.0560) (0.0595) (0.0702)
Father's education (highest grade) 0.0757*** 0.0076
(0.0272) (0.0246)
Father's education x Female 0.0712* 0.0791*
(0.0388) (0.0446)
Constant -3.4078* -7.4386*** -3.6133** -4.0007** -4.3604** -7.7824***
(1.8094) (2.0914) (1.8097) (1.8234) (1.8260) (2.0909)
Hausman test of endogeneity: c2(2) 16.33 10.63 17.97 21.29 22.42 9.83
R>c2 (p = 0.0003)
(p = 0.0049)
(p = 0.0001) (p = 0.0000) (p = 0.0000) (p = 0.0073)
Hansen J-statistic (Test of overidentifying restrictions) 296.02 8.97 283.46 248.12 188.80 3.74
(p = 0.0000)
(p = 0.0113)
(p = 0.0000) (p = 0.0000) (p = 0.0000) (p = 0.1537)
Observations 8,900 8,900 8,900 8,900 8,900 8,900

43
Table 12: Marginal effects of selectivity adjusted estimates of school attendance
_____________________________________________________________________________________________________________________
Continued on next page
Variables (1) (2) (3) (4) (5) (6)
Child age 0.3796*** 0.2802*** 0.3912*** 0.4008*** 0.4112*** 0.2887***
(0.0662) (0.0520) (0.0661) (0.0665) (0.0663) (0.0519)
Child age (squared) -0.0132*** -0.0106*** -0.0136*** -0.0141*** -0.0145*** -0.0109***
(0.0024) (0.0018) (0.0024) (0.0024) (0.0024) (0.0018)
Female child -0.2295*** -0.1660*** -0.2363*** -0.2470*** -0.2554*** -0.1747***
(0.0230) (0.0148) (0.0228) (0.0237) (0.0243) (0.0159)
Child labour hours -0.1204*** -0.0829*** -0.1235*** -0.1193*** -0.1167*** -0.0824***
(0.0233) (0.0189) (0.0234) (0.0249) (0.0255) (0.0196)
Child labour hours (squared) 0.0014*** 0.0010*** 0.0014*** 0.0013*** 0.0013*** 0.0010***
(0.0003) (0.0002) (0.0003) (0.0003) (0.0003) (0.0003)
Residual_child labour hours 0.0830*** 0.0608*** 0.0862*** 0.0834*** 0.0812*** 0.0606***
(0.0233) (0.0188) (0.0234) (0.0249) (0.0255) (0.0195)
Residual_child labour hours (squared) -0.0010*** -0.0008*** -0.0011*** -0.0010*** -0.0010*** -0.0008***
(0.0003) (0.0002) (0.0003) (0.0003) (0.0003) (0.0003)
Number of children 0-4 years -0.0035 -0.0030
(0.0061) (0.0061)
Number of school children 5-17 years -0.0023 -0.0017
(0.0040) (0.0040)
Number of adult males higher than 17 years -0.0115** -0.0111**
(0.0056) (0.0056)
Number of adult females higher than 17 years 0.0143** 0.0135*
(0.0072) (0.0072)
Occupation of father -0.0664*** -0.0721***
(0.0120) (0.0121)
Household uses piped water 0.0061 0.0017
(0.0296) (0.0297)
Household has a television -0.0091 -0.0132
(0.0173) (0.0163)

44
Table 12(continued): Marginal effects of selectivity adjusted estimates of school attendance
*** p<0.01, ** p<0.05, * p<0.1
Note: Standard errors in parenthesis.
Variables (1) (2) (3) (4) (5) (6)
Household has a radio 0.0116 0.0076
(0.0103) (0.0104)
Household has a bicycle 0.0408*** 0.0386***
(0.0119) (0.0118)
Formal school 0.4665*** 0.0029
(0.0202) (0.0148)
NGO school 0.7389*** 0.0073
(0.0396) (0.0151)
Own marginal land if acre of land less than 0.5 -0.0008 0.4648***
(0.0150) (0.0199)
Own large land if acre of land higher than 2 0.0116 0.7438***
(0.0154) (0.0389)
Urban 0.0301** 0.0318**
(0.0143) (0.0143)
Inverse Mills Ratio 0.9774*** 1.7375*** 1.0450*** 1.1149*** 1.1750*** 1.7819***
(0.3449) (0.2626) (0.3448) (0.3467) (0.3456) (0.2619)
Mother more educated than father 0.0701**
(0.0332)
Mother more educated than father x Female 0.0499
(0.0628)
Mother's education (highest grade) 0.0521*** 0.0340** 0.0108
(0.0172) (0.0138) (0.0084)
Mother's education x Female 0.0424** 0.0186 0.0088
(0.0191) (0.0203) (0.0150)
Father's education (highest grade) 0.0259*** 0.0016
(0.0093) (0.0052)
Father's education x Female 0.0243* 0.0169*
(0.0133) (0.0095)
Observations 8,900 8,900 8,900 8,900 8,900 8,900

45
Table 13: Selectivity-adjusted estimates of grade-for-age (GAGE)
______________________________________________________________________________________________________________________
Continued on next page

46
Table 13 (continued): Selectivity-adjusted estimates of grade-for-age (GAGE)
*** p<0.01, ** p<0.05, * p<0.1
Note: Standard errors in parenthesis and are computed robustly to account for heteroskedasticity. NGO school dropped due to collinearity.

47
References
Akabayashi, H. and Psacharopoulos, G. (1999),"The Trade-Off between Child Labour and
Human Capital Formation: A Tanzanian Case Study", Journal of Development
Studies,35(5):120-140.
Amin, S.,Quayes, M. S., and Rives, J. M. (2004),"Poverty and Other Determinants of Child
Labor in Bangladesh", Southern Economic Journal,70(4):876-892.
Arends-Kuenning, M. and Amin, S. (2004),"School Incentive Programs and Children's
Activities: The Case of Bangladesh", Comparative Education Review,48(3):295-317.
Aristei, D.,Perali, F., and Pieroni, L. (2008),"Cohort, Age and Time Effects in Alcohol
Consumption by Italian Households: A Double-Hurdle Approach", Empirical
Economics,35(1):29-61.
Baland, J. M. and Robinson, J. A. (1998). "A Model of Child Labor", CRED Serie
Recherche 206, University of Namer, Belgium
Bardhan, P. and Udry, C.(1999):Development Microeconomics, Oxford University Press.
Basu, K. (1999),"Child Labor: Cause, Consequence, and Cure, with Remarks on International
Labor Standards", Journal of Economic Literature,37(3):1083-1119.
--- (2006),"Gender and Say: A Model of Household Behaviour with Endogenously
Determined Balance of Power", The Economic Journal,116(511):558-580.
Basu, K. and Ray, R. (2001). "The Collective Model of the Household and an Unexpected
Implication for Child Labor: Hypothesis and an Empirical Test", Working Paper
2813, Washing D. C: The World Bank.
Basu, K. and Van, P. H. (1998),"The Economics of Child Labor", The American Economic
Review,88(3):412-427.
Becker, G.(1981):A Treatise on the Family, Cambridge, Harvard University Press.
Bhalotra, S. (2007),"Is Child Work Necessary?", Oxford Bulletin of Economics &
Statistics,69(1):29-55.
Cragg, J. G. (1971),"Some Statistical Models for Limited Dependent Variables with
Application to the Demand for Durable Goods", Econometrica,39(5):829-844.
Emerson, M. P. and Souza, A. (2002). "Bargaining over Sons and Daughters: Child Labor,
School Attendance and Intra-Household Gender Bias in Brazil", Working Paper No.
02-W13. Department of Economics, Vanderbilt University, NashVille, TN.
Ersado, L. (2005),"Child Labor and Schooling Decisions in Urban and Rural Areas:
Comparative Evidence from Nepal, Peru, and Zimbabwe", World
Development,33(3):455-480.
Green, W. H.(1997):Econometric Analysis, (3rd ), Prentice-Hall, Inc.
Guarcello, L.,Lyon, S., and Rosati, F. (2004). "Impact of Working Time on Children's
Health", Research report, Understanding Children's Work.
Heady, C. (2003),"The Effect of Child Labor on Learning Achievement", World
Development,31(2):385-398.
Jones, A. M. (1989),"A Double-Hurdle Model of Cigarette Consumption", Journal of Applied
Econometrics,4(1):23-39.
Khanam, R. (2006),"Child Labour in Bangladesh: Trends, Patterns and Policy Options
", Asian Profile,34(6):1-33.
--- (2008),"Child Labour and School Attendance: Evidence from Bangladesh", International
Journal of Social Economics,35(1/2):77-98.
Maitra, P. and Ray, R. (2002),"The Joint Estimation of Child Participation in Schooling and
Employment: Comparative Evidence from Three Continents", Oxford Development
Studies,30(1):41-62.

48
McElroy, M. B. and Horney, M. J. (1981),"Nash-Bargained Household Decisions: Toward a
Generalization of the Theory of Demand", International Economic Review,22(2):333-
349.
Orazem, F. P. and Gunnarsson, V. (2004). "Child Labour, School Attendance, and
Performance: A Review", Department of Economics, Working Paper No. 014001
IOWA State University
Parikh, A. and Sadoulet, E. (2005). The Effects of Parents' Occupation on Child Labor and
School Attendance in Brazil. Unpublished Manuscript, University of California at
Berkeley.
Psacharopoulos, G. and Yang, H. (1991),"Educational Attainment among Venezuelan
Youth: An Analysis of Its Determinants", International Journal of Educational
Development,11(4):289-294.
Rahman, M. M.,Khanam, R., and Absar, N. U. (1999),"Child Labor in Bangladesh: A Critical
Appraisal of Harkin's Bill and the Mou-Type Schooling Program", Journal of
Economic Issues,33(4):985-1003.
Ravallion, M. and Wodon, Q. (2000),"Does Child Labour Displace Schooling? Evidence on
Behavioural Responses to an Enrollment Subsidy", The Economic
Journal,110(462):C158-C175.
Ray, R. (2004),"Child Labour: A Survey of Selected Asian Countries", Asian-Pacific
Economic Literature,18(2):1-18.
Ray, R. and Lancaster, G. (2005),"The Impact of Children's Work on Schooling: Multi-
Country Evidence", International Labour Review,144(2):189-210.
Reggio, I. (2011),"The Influence of the Mother's Power on Her Child's Labor in Mexico",
Journal of Development Economics,96(1):95-105.
Ridao-Cano, C. (2001). "Child Labor and Schooling in a Low Income Rural Economy",
Population Aging Center Working Paper PAC 2001-0001, Institute of Behavioral
Science, University of Colorado at Boulder
Rosati, C. F. and Rossi, M. (2003),"Children's Working Hours and School Enrollment:
Evidence from Pakistan and Nicaragua", The World Bank Economic
Review,17(2):283 - 295.
Udry, C. (2003). "Child Labor", Center Discussion Paper No. 856 Economic Growth Center,
Yale University
Wahba, J. (2006),"The Influence of Market Wages and Parental History on Child Labour and
Schooling in Egypt", Journal of Population Economics,19(4):823-852.

49
Appendix A
Table A1: First stage OLS estimates of child labour hours
Continued on next page
Variables
Child labour
hours
Child labour
hours
2
Child
labour
hours
Child labour
hours
2
Child labour
hours
Child labour
hours
2
Child
labour
hours
Child labour
hours
2
Child labour
hours
Child labour
hours
2
Child
labour
hours
Child labour
hours
2
Child age -4.6797*** -364.2844*** -2.7165*** -244.8781*** -4.6535*** -362.5228*** -4.2318*** -335.8697*** -4.1576*** -331.1257*** -2.6349*** -239.3926***
(0.5685) (45.3490) (0.5571) (45.0121) (0.5682) (45.3341) (0.5627) (45.0852) (0.5611) (45.0095) (0.5543) (44.8632)
Child age (squared) 0.2537*** 18.2197*** 0.1881*** 14.2585*** 0.2527*** 18.1515*** 0.2361*** 17.0991*** 0.2337*** 16.9498*** 0.1838*** 13.9708***
(0.0217) (1.7345) (0.0213) (1.7176) (0.0217) (1.7340) (0.0215) (1.7246) (0.0215) (1.7216) (0.0212) (1.7121)
Female child -7.8234*** -558.0371*** -8.2657*** -587.3124*** -7.8609*** -560.1888*** -8.3044*** -587.3813*** -8.6962*** -612.1867*** -9.1009*** -639.4684***
(0.3671) (29.2830) (0.3561) (28.7780) (0.3745) (29.8793) (0.3869) (30.9968) (0.4037) (32.3822) (0.3951) (31.9826)
Number of children 0-4 years 0.6067*** 43.2885*** 0.5522*** 39.4679**
(0.2073) (16.7480) (0.2064) (16.7104)
Number of school Children 5-17 years 0.4070*** 25.0818** 0.3561*** 21.5989**
(0.1268) (10.2441) (0.1263) (10.2205)
Number of adult males higher than 17 years -0.2450 -14.9995 -0.2842 -17.6645
(0.2051) (16.5718) (0.2043) (16.5361)
Number of adult females higher than 17 years -0.3359 -24.7975 -0.3300 -24.3213
(0.2606) (21.0589) (0.2593) (20.9867)
Occupation of father -1.8453*** -138.8575*** -1.6757*** -127.0299***
(0.4075) (32.9315) (0.4063) (32.8828)
Household uses piped water 5.0204*** 424.7246*** 5.2520*** 440.8984***
(0.9486) (76.6511) (0.9445) (76.4504)
Household has a television -2.6219*** -151.1824*** -1.9084*** -102.5120***
(0.4691) (37.9036) (0.4732) (38.2988)
Household has a radio -0.0410 -14.9690 0.0424 -8.8819
(0.3508) (28.3495) (0.3497) (28.3014)
Household has a bicycle -2.1076*** -156.8816*** -1.9197*** -143.8875***
(0.3831) (30.9574) (0.3819) (30.9097)
Formal school 2.6759*** 168.0671*** 2.4040*** 149.1312***
(0.3803) (30.7317) (0.3807) (30.8143)
NGO school -0.4729 -10.3495 -0.1983 8.9358
(0.5064) (40.9204) (0.5056) (40.9233)
Own marginal land if acre of land less than 0.5 -5.2628*** -318.5831*** -4.7312*** -282.1031***
(0.3242) (26.1980) (0.3285) (26.5927)
Own large land if acre of land higher than 2 -6.5255*** -488.3792*** -6.6715*** -498.2775***
(1.5288) (123.5291) (1.5208) (123.0984)
Urban -1.2575*** -134.7278*** -1.4713*** -149.3000***
(0.3825) (30.9098) (0.3813) (30.8615)
(6)
(1)
(2)
(3)
(4)
(5)

50
Table A1 (continued): First stage OLS estimates of child labour hours
*** p<0.01, ** p<0.05, * p<0.1
Note: Standard errors in parenthesis and are computed robustly to account for heteroskedasticity.

51
Table A2: Double-Hurdle estimates of participation
_____________________________________________________________________________________________________________
Continued on next page

52
Table A2 (continued): Double-Hurdle estimates of participation
*** p<0.01, ** p<0.05, * p<0.1
Note: Standard errors in parenthesis and are computed robustly to account for heteroskedasticity.

53
Table A3: Double-Hurdle estimates of GAGE
____________________________________________________________________________________________________________
Continued on next page

54
Table A3 (continued): Double-Hurdle estimates of GAGE
*** p<0.01, ** p<0.05, * p<0.1
Note: Standard errors in parenthesis and are computed robustly to account for heteroskedasticity.

55
Table A4: Marginal effects of school attendance for child wage workers
*** p<0.01, ** p<0.05, * p<0.1
Note: Standard errors in parenthesis and are computed robustly to account for heteroskedasticity.

56
Table A5: Marginal effects of GAGE for child wage workers
*** p<0.01, ** p<0.05, * p<0.1
Note: Standard errors in parenthesis and are computed robustly to account for heteroskedasticity.

57
Table A6: Marginal effects of school attendance, by rural-urban location
*** p<0.01, ** p<0.05, * p<0.1
Note: Standard errors in parenthesis and are computed robustly to account for heteroskedasticity.

58
Table A6 (continued): Marginal effects of school attendance, by rural-urban location
*** p<0.01, ** p<0.05, * p<0.1
Note: Standard errors in parenthesis and are computed robustly to account for heteroskedasticity.

59
Table A7: Marginal effects of GAGE, by rural-urban location
*** p<0.01, ** p<0.05, * p<0.1
Note: Standard errors in parenthesis and are computed robustly to account for heteroskedasticity.

60
Table A7 (continued): Marginal effects of GAGE, by rural-urban location
*** p<0.01, ** p<0.05, * p<0.1
Note: Standard errors in parenthesis and are computed robustly to account for heteroskedasticity.