WPS6904
Policy Research Working Paper 6904
Child Labor and Learning
Patrick M. Emerson
Vladimir Ponczek
André Portela Souza
The World Bank
Development Economics Vice Presidency
Office of the Chief Economist
June 2014
Policy Research Working Paper 6904
Abstract
This paper uses a unique micro panel dataset of Brazilian working on learning outcomes in math and Portuguese
students to investigate the impact of working while in is found. The effects of child work range from 3 to 8
school on learning outcomes. The potential endogeneity percent of a standard deviation decline in test score,
is addressed through the use of difference-in-difference which represents a loss of about a quarter to a half of a
and instrumental variable estimators. A negative effect of year of learning on average.
This paper is a product of the Office of the Chief Economist, Development Economics Vice Presidency. It is part of a larger
effort by the World Bank to provide open access to its research and make a contribution to development policy discussions
around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The authors
may be contacted at patrick.emerson@oregonstate.edu, vladimir.ponczek@fgv.br, and andre.portela.souza@fgv.br.
The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development
issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the
names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those
of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and
its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent.
Produced by the Research Support Team
Child Labor and Learning⇤
e Portela Souza§
Patrick M. Emerson† Vladimir Ponczek‡ Andr´
JEL classiﬁcation : J13,I21. Keywords : Child Labor; Learning; Proﬁciency; Education.
⇤
Acknowledgements: For useful comments and advice, we thank seminar participants at the World Bank
and S˜ao Paulo School of Economics. For outstanding research assistance we thank Eduardo Tillman. We
thank the Funda¸ ao de Amparo `
c˜ ao Paulo (FAPESP) for ﬁnancial support.
a Pesquisa do Estado de S˜
†
ao Paulo School of Economics, FGV patrick.emerson@oregonstate.edu.
Oregon State University, IZA & S˜
phone: 541-737-1479. Fax: 541-737-5917
‡
ao Paulo School of Economics, FGV & CNPq. vladimir.ponczek@fgv.br
S˜
§
ao Paulo School of Economics, FGV, & CNPq. andre.portela.souza@fgv.br
S˜
1 Introduction
Though the global trend is downward, the incidence of child labor remains very high in
developing countries. For example, in 2000 the International Labour Organization estimated
that 246 million, or almost 16 percent, of the world’s children between the ages of 5 and
17 were child laborers. By 2008 this had fallen to 215 million - a dramatic decline but
still representative of 13.6 percent of the world’s children. In response, many governments
have proposed or implemented policies designed to reduce the incidence either in their own
countries, through labor laws that restrict or prohibit children working, or in other countries,
through policies such as restricting the importation of goods that use children in some part
of their production.
Though the goal of protecting children from harmful work is a recognized aim of the
rights-based approach to child labor, and while the worst-forms of child labor are prima facie
intolerable, there is much to be learned about what types of activities and their intensities
constitute harmful work. In order to make e↵ective and responsible policy it is important to
understand the consequences working as a child on participants and the interaction of child
work with other activities such as schooling. Doing so will add clarity to questions such as
how young is too young to work, how much is too much work at given ages and does working
interfere with human capital accumulation. This study contributes to this understudying
by demonstrating that working while attending school has a large negative and signiﬁcant
impact on the learning progress of schoolchildren in Brazil. This ﬁnding is also consistent
with most theoretical treatments of child labor that stress the trade o↵ between work and
human capital accumulation.
While much is known about the incidence and the determinants of child labor, surprisingly
little is known about the consequences of child work on participants. Important policy
2
questions such as how young is too young to work, are some work activities better or worse
than others, does work impair the health of children and does combining work and school
hinder learning remain largely unanswered. This paper seeks to provide an answer the last
of these questions and contributes to the literature by assessing the impact of working while
ao Paulo municipal
in school on learning as measured by the proﬁciency of students in the S˜
school system through their performance on standardized exams.
Perhaps the main reason for the dearth of received evidence on the impact of child work
is the inherent di culty in uncovering causal linkages between the activities of children
and their subsequent outcomes. The e↵ects of confounding and unobserved variables are
persistent problems that are di cult to overcome. In this study, we utilize a unique panel
ao Paulo
data set of the standardized test scores of Brazilian children in the metropolis of S˜
that allow us to explore the causal link between the work activity of children and its e↵ect
on their exam performance.
We ﬁnd that, controlling for individual time-invariant unobservable characteristics, work-
ing while remaining in school has a depressing e↵ect on their proﬁciency test scores. The
magnitude of these e↵ects range from 3% of a standard deviation in test scores to 8% which
represents from one quarter to one half of a year of lost learning. Additionally, we ﬁnd that
the magnitude of the negative impact increases with a students ability and that there are
both lingering and cumulative negative e↵ects from working while in school. These results
provide valuable information to policy makers who wish to understand what types of child
work to target for elimination and the e↵ect of child labor on human capital accumulation.
As these are short-run e↵ects, it is likely that the long-run impact of the will be quite a
bit larger. Additionally these results apply only to those who remain in school, factoring
in drop-out and grade repetition would certainly increase the magnitude of the impact of
3
working while in school.
The results are robust to idiosyncratic preferences and we perform robustness checks to
show that the results are not due to idiosyncratic trends or shocks at the household level.
We implement placebo tests and control for idiosyncratic economic shocks at the household
level. Moreover, we implement a LATE strategy using the minimum legal age to entry in the
labor market to induce an exogenous variation in child labor status. The robustness check
results are consistent with our main ﬁndings. Finally, we ﬁnd possible channels through
which child labor can impact learning as participants in the labor market are more likely to
report that they miss school days, turn homework in late and complete homework while in
school rather than at home.
In recent years, a large body of theoretical and empirical research has emerged that has
studied the economics of child labor (see Edmonds, 2008; Edmonds and Pavcnik, 2005; Basu
and Tzannatos, 2003; Basu, 1999, for extensive literature reviews). Much of the received
empirical work has focused on the determinants of child labor while questions of the con-
sequences of child labor have been largely contained in the theoretical literature (see e.g.
Edmonds, 2008; Edmonds and Pavcnik, 2005; Basu and Tzannatos, 2003; Basu, 1999; Emer-
ortner, 2004;
son and Knabb, 2012, 2007, 2006; Horowitz and Wang, 2004; Ejrnæs and P¨
Basu, 2002; Dessy and Pallage, 2001; Baland and Robinson, 2000; Dessy, 2000; Basu and
Van, 1998). The theoretical literature has long emphasized the trade-o↵ between child labor
and human capital accumulation to justify policy interventions assuming depressing impacts
from child labor. As noted above, the empirical foundations to support these assumptions
1
are weak.
Though evidence of the e↵ects of child work on participants is still relatively scarce,
1
(See also Ponczek and Souza, 2012; Emerson and Souza, 2003; Ilahi et al., 2001; Psacharopoulos and
Patrinos, 1997)
4
there is a new and growing literature that has begun to ﬁll in the lacuna. Beegle et al.
(2009), studies a ﬁve year panel of school children in Vietnam and ﬁnds that child labor
has negative consequences on school participation and educational attainment. Orazem and
Lee (2010) uses Brazilian household data to examine the impact of working as children on
self-reported health outcomes and ﬁnds negative impacts of child labor. Emerson and Souza
(2011) examines retrospective data from Brazil and ﬁnds that child work before 13-14 years
old negatively a↵ects adult incomes, but that this a↵ect turns positive after these ages.
Interestingly, this study ﬁnds that the e↵ect of child work on earnings remain even when
controls for years of education are included, raising the possibility that child labor may
a↵ect the learning of those workers who remain in school, which provides a motivation for
the current study.
Research on the consequences of child work on education have mainly focused on atten-
dance rather than learning (e.g. Ravallion and Wodon, 2000; Assaad et al., 2001; Canals-
Cerda and Ridao-Cano, 2004; Beegle et al., 2008, 2006; Assaad et al., 2005) and have found
modest negative e↵ects of child work. But as Emerson and Souza (2011) suggest, school
enrollment may not be the only important measure, especially in countries like Brazil where
combining work and school are common.
The direction of the expected impact of child work on learning is unclear. Working
requires time and energy that could hamper a student’s ability to learn, but some work
activities could involve tasks that are either directly related to learning (like reading, writing
and math) or indirectly related but still involve use of these skills. If a work activity involves
learning-by-doing or is otherwise positively correlated to learning the skills tested in school
(in our case math and Portuguese), work could, in fact, have a positive impact. In the end
the true nature of the relationship between work and learning, if the two are substitutes or
5
complements, is an empirical issue.2 Understanding this relationship is extremely important
as previous research has shown a very strong connection between educational proﬁciency
and adult income and economic growth, and that proﬁciency is a stronger determinant than
completed years of schooling [see, e.g. Hanushek and Zhang (2009) and Hanushek and Kimko
(2000)].
We are aware of three previous studies that have examined the impact of child work
on student proﬁciency. The ﬁrst paper, closely related to the present study is Bezerra et
al. (2009), which uses cross-sectional Brazilian data to test the impact of working on the
performance on similar exams. The authors ﬁnd that working has a negative impact on the
performance of participants. Another closely related study is Dumas (2012) which exploits
retrospective data from Senegal to examine the e↵ects of child work on the test scores of
Senegalese children and ﬁnds some evidence of positive e↵ects of child work. Finally Gun-
narsson et al. (2006) uses data from nine Latin American countries (including Brazil) and
ﬁnds negative and signiﬁcant impacts of working on student test scores. All three studies are
constrained by the inherent di culties in overcoming the potential endogeneity of child la-
bor and all three implement instrumental variables strategies in an attempt to overcome the
problem. In all three cases, the challenge of ﬁnding sources of variation that are correlated
with the decision to work but uncorrelated with the unexplained variation in school perfor-
mance is severe, leading to questions of the validity of the instruments themselves and thus
the results. In our case the use of time-series data presents a huge advantage in our ability
to control for both the endogeneity of child labor and the presence of other unobservables
(e.g. parental preferences) that are potentially correlated with both the decision to work and
the aptitude for, and attitude toward, school. We are also able to explore the lingering and
2
We use the terms ’learning’ and ’proﬁciency’ synonymously, but for both we are referring speciﬁcally to
only those aspects of learning that are measured by standardized tests in Portuguese and Math.
6
cumulative impacts of child labor as well as explore the heterogeneous e↵ects on student of
di↵erent ages and abilities. In addition, through the time-use information available to us,
we are able to explore the potential channels through which child labor may interact with
the process of learning and we are thus able to shed light on the mechanisms involved.3
This paper proceeds as follows: In section 2 we describe the data used in the study. In
ao
section 3 we describe the general child labor and educational environment in the City of S˜
Paulo. In section 4 we explain the empirical strategy, how we identify our model and the
robustness checks we employ. In section 5 we present and discuss the results of the empirical
investigation. In section 6 we summarize the paper and discuss the policy implications of
the results of the empirical investigation.
2 Data, Sample Selection and Descriptive Statistics
ao Paulo started an evaluation system for students enrolled in municipal
In 2007, the City of S˜
schools involving a set of proﬁciency exams in mathematics and Portuguese. These exams
were accompanied by a questionnaire that was given to each student taking the exams as well
as an additional questionnaire that was given to the parents of the students taking the exams
about the socio-economic characteristics of the family, although the parents questionnaire
was not administered in 2008.4 The exam is called the Prova S˜ ao Paulo Exam)
ao Paulo (S˜
and was implemented annually until 2013 when it was discontinued. The microdata from
the 2007 to 2010 exams are available and used in this paper. In 2007 all students in the
even grades (2nd , 4th , 6th and 8th ) took the exam. In 2008, all students in the 2nd , 4th and 6th
3
Also related are studies from cross sectional data such as Heady (2003) which ﬁnds large and negative
impacts of child work on educational achievement and Akabayashi and Psacharopolous (1999) which examines
the correlation between work and subjective measures of ability, and ﬁnds a negative impact from child work.
4
Questionnaires were also given to the principals which asked them to answer questions about themselves,
their school, the teachers and supervisors, and the student population, but it was not administered in 2009.
7
grades took the exam, and randomly selected students in 3rd , 5th , 7th and 8th grades took the
exam (one class per grade per school was randomly selected to take the exam). From 2009
and on, all students in the even grades and randomly selected students in the odd grades
(35 students per grade per school) took the exam.5 In each year, around 500,000 students in
ao Paulo Exam was structured
8,000 classes in 500 schools took the exam. Importantly, the S˜
based on the Item Response Theory (IRT) so that the results are comparable across grades
and across years6 .
We have information from the students’ questionnaires for all years on the students’
working status. Only students in the ﬁfth grade and above answer the question about child
labor, which restricts the population of analysis.7 The precise wording of the question about
working status is as follows (translation ours):
“During school days, do you work?
(A) Yes, outside of the house; (B) Yes, at home helping with the chores; (C) No, I only
study.”
Respondents were limited to one answer only. We set the indicator variable M arketLabor
equal to 1 whenever a student answered ”A”, and to 0 otherwise so we are comparing those
that work outside of the home to those who work at home on chores and those who do not
work at all.
5
All students in the ﬁfth grade who scored below 150 points in the previous year are also included in
sample.
6
There are common items in the tests over time and across grades that can be used as linking items
in constructing comparable scores. Please, check http://portalsme.prefeitura.sp.gov.br/Projetos/
nucleo/AnonimoSistema/MenuTexto.aspx?MenuID=44&MenuIDAberto=38formoredetails.
7
In 2008, students in the ﬁfth grade answered a version of the questionnaire that was distributed to the
younger students in second to fourth grade. In this questionnaire, there is no question about the child labor
status. Together with the fact that ﬁfth graders are not randomly selected, we decided to exclude them from
our sample.
8
From the parent’s questionnaire we collected information about the father’s employment
status for 2007, 2009 and 2010.
The student questionnaire also asks students about their studying habits such as whether
they miss classes, hand in homework late, prepare for exams in advance, and complete home-
work at school. From the responses to these questions we created four indicator variables.
We construct three di↵erent samples for this study. The ﬁrst includes all observations of
students in the 6th , 7th and 8th grades in 2007, 2008, 2009 and 2010. This sample constitutes
an unbalanced panel of 473,051 observations of 313,297 students (158,180 boys and 155,117
girls).8 We call this the “full sample”. The second sample encompasses all students that
were in 6th grade in 2007 or in 2008 and were found two years later from the ﬁrst observation.
Therefore, we have exactly two observations for each student. This balanced panel has 48,009
boys and 48,161 girls. We call this the “paired sample”. The third sample also encompasses
all that were in 6th grade in 2007 or in 2008, but we include only those who were also found in
next two consecutive years after the ﬁrst observation. Depending on the exercise, we use the
ﬁrst and third observations or the ﬁrst and second observation for each students. Therefore,
we have exactly two observations for each student. The sample contains observations on
6,563 boys and 6,630 girls and we call this the “3 period sample.”9
Tables 1 to 3 presents some descriptive statistics of the three samples separately for boys
and girls. The incidence of child labor in the full sample is around 12% for boys and 6% for
girls. The average age is 14 years old and, on average, boys outperform girls in math and
the reverse occurs in Portuguese.
8
We also dropped 994 students aged nine years old or below. We believe those are measurement error,
since the regular age for the 6th grade in Brazil is 12 years old. Nevertheless, none of the deleted students
appear more than once in the sample, therefore the trimming does not change the estimate of the parameter
of interest in the ﬁxed e↵ect speciﬁcation.
9
No student was age trimmed in the paired and 3 period samples
9
Since we use ﬁxed e↵ect estimators, it is necessary for identiﬁcation to have transitions
into and out of market labor. Table 4 shows the transition matrix of the market labor
variable. We can see that for both boys and girls we have a su cient number of students
transiting in and out of working status. For instance, 4,500 boys and 2,300 girls change their
status from not working to working in one year.
ao Paulo
It is worth noting here that we observe only those who are enrolled in the S˜
municipal school system and who remain in the school system. There is no forced repetition
ao Paulo municipal schools for the 6th , 7th and 8th grades, but there can be
of grades in S˜
drop-out, movement to state or private schools or movement out of the area. As drop-out
and delay are likely two other important e↵ects of child labor, it is important to note that
we are not examining these ancillary e↵ects of child labor.
3 ao Paulo
Child Labor and Students in the City of S˜
ao
This section describes the incidence of child labor and school attendance in the City of S˜
Paulo.10 The data used in this study come from the 2010 Demographic Census collected by
the Brazilian Census Bureau (IBGE). It contains information on socio-demographic charac-
teristics, fertility, migration, and time allocation for all individuals sampled. It is a sample
of the entire population representative at the municipality level. Most important for the
present study, it contains information about school attendance, labor force participation
and occupation in the reference month of the survey (July). The information for labor
market outcomes are available for individuals aged 10 years old and above.
10
ao Paulo is the name of both the City and the State in which it resides and there are both state and
S˜
ao Paulo
municipal schools inside the City of S˜
10
3.1 Child Laborers
Table 5 below presents the ﬁgures for the time allocation of individuals aged 10 to 17 living
ao Paulo City in 2010 for male and female individuals, separately. There are around 1.4
in S˜
million individuals in this category and the majority attend school. In fact, 85.4% (86.6%) of
boys (girls) attend school only; and 6.2% (5.4%) of the boys (girls) divide their time between
school and work. Thus around 92% (93%) of boys (girls) attend school. Conversely, 2.5%
(1.7%) of boys (girls) work only; and 5.9% (6.3%) of boys (girls) neither work in the labor
market nor attend school. Summing up those that work only and those that work and attend
school, the incidence of child and adolescent work among boys is 8.7% and among girls it
is 7%. Note that of all boys working in the labor market, 71% attend school, and for girls
the ﬁgure is 76%. On the other hand, of all boys attending school, 6.8% work in the labor
market, and of all girls attending school, 5.8% work in the labor market.
3.2 Students
Our data encompass students enrolled in the 6th , 7th and 8th grades at S˜
ao Paulo municipal
schools. According to the 2010 Educational Census from the Brazilian Ministry of Education,
ao Paulo in 2010.
there are around 580,000 students enrolled in these grades in the City of S˜
Table 6 shows their distribution across grades and school systems. Of all of them, 49.6%
are enrolled in the public municipal schools, 31.5% are enrolled in public state schools and
18.8% are enrolled in private schools and these proportions are similar for all grades. Since
our data are of municipal school children only, we observe roughly half of the 6th , 7th and 8th
ao Paulo City, a population of around 290,000 students.
grade students in S˜
The IBGE Demographic Census has information about the type of school system in
which the student is enrolled as well. It classiﬁes schools as public or private but does
11
not distinguish between municipal and state public schools. Table 7 below presents the
distribution of 6th , 7th and 8th grade students across public (municipal and state) and private
schools. These ﬁgures are presented for students who work and who do not work separately.
ao
According to the census, there are around 540,000 individuals living in the City of S˜
Paulo in 2010 that reported that they attend 6th , 7th and 8th grades. Of all of them, 82%
attend public schools. Of all middle school students, around 11% work in the labor market.
However, these ﬁgures are sharply di↵erent between public and private school students.
Among public school students, 12.8% work in the labor market, whereas among private
school students, 2.4% work in the labor market. As expected, the incidence of child and
adolescent work increase with grade. Among 6th , 7th and 8th graders in public schools, the
incidences of working the labor market are 7%, 12.6%, and 20.4%, respectively.11
If the proportion of child workers among 6th , 7th and 8th grade students is similar between
municipal and state school students, then there are roughly 37,000 municipal middle school
ao Paulo in 2010.
students working in the labor market of the City of S˜
What do these working students do? Table 8 below presents the occupational distribution
of the working public school students in 6th , 7th and 8th grade in the City of S˜
ao Paulo
according to the 2010 Demographic Census.
Most students who work, work in the service sector. Indeed, 26.1% of them work as
domestic servants, street vendors, car washers, and others; 25.4% work as service and retail
vendors; 14.8% are in o ce work; and 10.2% are in military service occupations.
11
In our pooled sample the averages (across all three grades) for the market labor variable are 12% for
boys and 6% for girls.
12
4 Empirical Strategy
The main challenge in estimating the impact of child labor on learning is overcoming the
potential endogeneity of child labor. The decision about the child’s time allocation could
be made based on unobservable characteristics of the individual that also determine her
proﬁciency. It is very likely that ability is correlated with proﬁciency and the parents’
perception of the value of education which determines the child’s time allocation. In this
case, a simple OLS estimator for child labor and proﬁciency will be biased. Depending on
the correlation between the unobservables and time allocation decisions and between the
unobservables and proﬁciency, the OLS estimator could be upward or downward biased.
In the above example, if there is a positive relationship between ability and the perception
of the value of education, meaning parents with high ability children are more likely to
prioritize schooling, we would expect that a naive approach would overestimate the actual
impact of child labor on proﬁciency. On the other hand, one can imagine that more able
children have better opportunities in the labor market. In this case, the OLS estimator
would underestimate the e↵ect of child labor on proﬁciency.
Therefore controlling for such unobservable characteristics is essential to consistently
ao
estimate the impact of child labor on proﬁciency. The longitudinal dataset of Prova S˜
Paulo allows us to control for unobservable characteristics that are ﬁxed over time.
Our ‘benchmark’ strategy is, therefore, a ﬁxed e↵ect estimator, we run the following
regression separately for boys and girls and for math and Portuguese:
2
Tigt = 0 + 1 M arket Laborigt + 2 Ageigt + 3 Ageigt + ✓i + t + g + ✏igt (1)
where Tigt is the math or Portuguese language proﬁciency test score of student i in grade g ,
13
and year t. M arket Laborigt is an indicator variable that assumes 1 if the student i, in grade
g is working at year t. ✓i is the individual ﬁxed e↵ect, t is a time-speciﬁc e↵ect, and g is
the grade ﬁxed e↵ect; ✏isct is the error term with school clustered variance-covariance matrix.
In this case, 1 is the parameter of interest. Note that by including both and age and grade
ﬁxed e↵ects we are estimating the impact of working on students who are the same age and
in the same grade, thus we are deliberately netting out the potential e↵ects of drop out and
delay as mentioned previously.
This strategy is consistent even if there are unobservable attributes that simultaneously
determine child labor and proﬁciency as long as those characteristics are constant over time.12
We run the benchmark speciﬁcation using both full and paired samples.
The identiﬁcation strategy in our panel structure requires some individuals to transit in
and out of the labor market. In the above estimations we implicitly assume that the e↵ects
of these transits are the same for all individuals regardless of age, ability, and whether they
are entering into child labor or exiting out of child labor. It is likely that these e↵ects are not
the same - that there is heterogeneity based on age ability and the direction of the transit
- and we therefore conduct four tests of the possibility of heterogeneous e↵ects using the
paired sample.
First, we test if younger students su↵er more from working than older students.13 Second,
we test if students with di↵erent ability levels, as measured by ﬁrst year test scores, su↵er
di↵erential impacts of working while studying.
To conduct these two tests we estimate the two speciﬁcations below:
12
Indeed, we need only that the variation of the unobservable attributes is not jointly correlated with the
child labor and proﬁciency variation.
13
Emerson and Souza (2011) have shown the presence of heterogeneous e↵ects of child labor on adult labor
market outcomes depending the age the child enter in the labor market in Brazil.
14
2
Tigt = 0 + 1 M arketLaborigt + 2 M arketLaborigt ⇥Ageig 1 + 3 Ageigt + 4 Ageigt +✓i + t + g +✏igt
(2)
2
Tigt = 0 + 1 M arketLaborigt + 2 M arketLaborigt ⇥ Tig 1 + 3 Ageigt + 4 Ageigt + ✓i + t + g + ✏igt
(3)
where M arketLaborigt ⇥ Ageig1 and M arketLaborigt ⇥ Tig1 are interaction terms between
child labor status and age and test score at the ﬁrst year the student is observed in the sample,
respectively. A negative coe cient estimate associated with the interaction between child
labor and age suggests that child labor e↵ects younger students more than older students;
while a negative coe cient associated with the interaction between test score indicates that
higher scoring students are more harmed by working.
In order to account for the possibility of heterogeneous transition e↵ects, we estimate
speciﬁcation (1) separately for those entering and those exiting the labor market using the
paired sample. Thus our third test of heterogeneous e↵ects is on those who enter into child
labor compared to those who never worked, and our fourth test is on those who exit out
of child labor compared to those who work in both periods. Notice that this is a di↵erent
comparison to the benchmark test which compares those who transit into or out of market
labor to those who not change their status (both not working in all periods and working in
all periods of observation).
Next, we ask whether child labor has cumulative and lingering e↵ects. To test for these,
we analyze whether the impact of work on learning depends on the length of time spent
working. For these exercises, we use the 3 period sample restricted to students that were
15
not working in the ﬁrst observation t.
To check for the presence of cumulative e↵ects we compare the evolution of test scores
between t and t + 2 among three groups of students. Students in the comparison group have
not worked in all three periods. Students of group 1 work at period t + 2 only. Students of
group 2 started working in t + 1 and remain working in t + 2. Therefore, we run the following
speciﬁcation:
1t 2t 2
Tigt = 0 + 1 M arketLaborigt + 2 M arketLaborigt + 3 Ageigt + 4 Ageigt + ✓i + t + g + ✏igt (4)
1t
where M arket Laborigt indicates whether the student has been in the labor market for
2t
one year (started worked in t + 2) and M arket Laborigt indicates whether the student has
been in the labor market for two years (started worked in t + 1 and remains working in t + 2).
We then test if 1 = 2.
To test for the presence of lingering e↵ects we compare the evolution in scores from t
to t + 2 between two groups of students: a comparison group (worked in no periods), and
a treatment group of students who have only worked in t + 1 and have stopped working in
t + 2.
If the time variation of the relevant unobservables is correlated with the child labor and
proﬁciency variations, the ﬁxed-e↵ect estimator is inconsistent.14 Therefore, we perform
several robustness checks in order to validate our identiﬁcation assumption. The robustness
checks use information from students that appear at least three years in our sample. We
compare students with the same working history in the ﬁrst two years we observe them, but
14
Ideally, we would observe the same individual’s proﬁciency at the same time: working and not working.
In this hypothetical situation, the mean proﬁciency di↵erential would be a clear and immediate indicator of
the impact of child labor on proﬁciency.
16
with a di↵erent working status in the third year. The idea is that the outcome in the second
year should not be impacted by a future (third year) working event. If this were so, then
the assumption of identical trend for treatment and comparison groups would be invalid.
Therefore, we estimate speciﬁcation (1) for the students that: (i) are in the 3 Period sample
(i.e. appear in three consecutive years); and (ii) have the same working status history in the
ﬁrst two years in the sample.
Though our empirical strategy allows us to control for all time-invariant individual and
family characteristics there may still be some time-variant characteristics that are important.
For example, it is likely that some transitions into and out of child labor are correlated with
idiosyncratic transitory shocks at the household level and, if so, there may be a direct e↵ect
of such shocks on learning.15 In this case our estimates would be biased upward as we
would attribute to child labor the direct e↵ect of the shock itself. We believe that it is
likely that most of the impact of a shock that causes a child to enter the labor market is
through the interference that the labor itself has on the time allocation of the student leading
to less studying, fatigue, etc. Direct e↵ects from, for example, the additional stress of an
employment or health shock in the household, would likely be small relative to the disruption
the child experiences from working itself.
Nevertheless, we are able to perform a robustness check to see if such transitory shocks
are biasing our estimates by including a control for the employment status of the father.
Since we do not have this information from 2008 we estimate our benchmark model on a
sample from 2007, 2009 and 2010 for observations that include the father’s employment
status with and without the father’s employment status included as a control.
Moreover, other sources of shocks may be both correlated to child labor transitions and
15
For instance Duryea et al. (2007) ﬁnds that household economic shocks are important determinants of
transit into and out of child labor in metropolitan Brazil.
17
changes in learning. For instance, it is possible that changes of the parents marital status
or health problems of a family member may bias our results. Therefore, we implement
an IV strategy to attempt to address this potential problem. Speciﬁcally, we explore the
transitions in child labor status induced by the change in age that turn the individual eligible
to work. In Brazil the minimum legal age to participate in the labor market is sixteen years
old. Therefore, we will estimate a Local Average Treatment E↵ect (LATE) as explained by
Imbens and Angrist (1994).
Finally, if child labor does have a signiﬁcant e↵ect on learning it would be useful to
understand the nature of this e↵ect. To do so we identify some channels through which
working could a↵ect student performance that are related to the time use and study habits
of the students. Speciﬁcally, we use the full sample and run speciﬁcation (1) to test the
impact of child labor on four di↵erent outcomes: missing classes, preparing for exams in
advance, completing homework at school, and turning in homework late.
All regressions use robust standard errors.
5 Results
5.1 Benchmark model
ao Paulo city
To assess the impact of working while in school on the learning outcomes of S˜
school children we begin by estimating the benchmark model on both the full pooled sample
(an unbalanced panel) and the paired sample (a balanced panel).
Table 9 presents the results of these regressions. The ﬁrst four columns present the pooled
sample estimation results for the math and Portuguese test scores for boys and girls sepa-
18
rately. The second four columns present the same estimation results for the paired sample.
In all eight regressions the coe cient estimate on the child labor dummy variable is negative
and signiﬁcant at the one percent level. The point estimates range from -1.278 (boys math)
to -3.951 (girls Portuguese). The results suggest that working while in school negatively
impacts the students’ performance on standardized exams in both math and Portuguese.
For math for boys and girls and for Portuguese for boys the coe cient estimates translate to
around 3% to 3.5% of a standard deviation decrease in test scores. For Portuguese for girls
the coe cient estimates translate to a 6.7% of a standard deviation decrease in the pooled
sample and an 8.1% decrease in the paired sample. We can interpret those coe cients with
the average proﬁciency gain a student obtain of one extra year of schooling. The average
annual increase is 11 points in math and 12 in Portuguese which suggests the impact is a
loss of from around 10% to 40% of a year of learning.
The students’ age coe cients are all positive and signiﬁcant except for boys and Por-
tuguese suggesting that the older a child is (in a given grade) the better he or she generally
does on standardized tests with the exception of boys and Portuguese. The squared age
coe cient estimates are all negative and signiﬁcant but very small suggesting that the age
e↵ect is slightly non-linear but not enough to turn the net e↵ect negative.
5.2 Heterogeneity
Tables 10 through 13 present the results of the tests of heterogeneity, all of which use the
paired sample.
19
5.2.1 Marginal impacts of age and proﬁciency
Table 10 presents the results of the estimation with the child labor indicator variable inter-
acted with age at the ﬁrst observation. The results from this regression suggest that, after
controlling for grade and year, the e↵ect of child labor does not signiﬁcantly change with the
student’s age.
Table 11 presents the results of the estimation with the child labor indicator variable
interacted with test score. In contrast to age, the e↵ect of child labor does vary depending
on test score: students are more negatively impacted by child labor the higher their initial
test score. This may be perhaps because better students are more prone to study at home
where child labor might interfere more or perhaps fatigue has a larger marginal impact on
better students.16
5.2.2 Isolating transitions into and out of child labor
Tables 12 and 13 present coe cient estimates on the paired sample that attempt to isolate
the e↵ect of a child starting to work and a child stopping working. In order to accomplish
this, the estimation presented in Table 12, ’Ins’, selects all those who did not work in the
ﬁrst period of observation (t) and compares those that continued without working in the
next window of observation (t + 2) to those were working in the next window of observation
(t + 2). Table 13, ’Outs’, does the opposite: it considers all those that were observed to
be working in the ﬁrst period (t) and compares those that remained working in the next
period of observation (t + 2) to those that were no longer working in the next period of
observation. The coe cient estimates for the child labor variable presented in Table 12 for
the ’ins’ are all negative and signiﬁcant at the one percent level, similar to Table 9, but
16
The negative e↵ect dominates for those whose test scores are above: 171.1 for boys math; 187.9 for girls
math; 173.1 for boys Portuguese; 113.8 for girls Portuguese. All well below their respective means.
20
the point estimates are larger.17 The marginal impact of these new estimates now range
from 6.2% (girls math) to 10.6% (girls Portuguese) of a standard deviation decline in test
score. Interpreting these in terms of average annual increases in test scores reveals that the
impact of starting to work is equivalent to roughly one half to a whole year of learning loss.
The coe cient estimates for the child labor variable presented in Table 13 for the ’outs’ are
all negative and signiﬁcant at the one percent level, similar to Table 9, but again the point
estimates are larger. The marginal impact of the estimates now range from 6.7% (boys math)
to 19.0% (girls Portuguese) of a standard deviation decline in test score of roughly one half
to almost two years of learning loss for those that remain in the labor market compared to
those who exit.
The ‘In’ and ‘Out’ coe cient estimates have similar magnitudes and we cannot reject
that they are statistically equal to each other. Therefore there is no evidence of heterogeneity
between moment into or out of the labor market. This could suggest that transitions into and
out of the labor market may be due to individual idiosyncratic shocks that are orthogonal
to proﬁciency.
5.3 Exposure and lingering e↵ects of child labor
We now turn to two questions that demand the use of the third ’three period’ sample. Recall
that this sample takes all of the children we observe in 6th grade in either 2007 or in 2008
and whom we observe in the next two consecutive years (regardless of progression through
the grades) and we compare scores in t and t + 2. Table 14 presents estimates of ’exposure
e↵ects’ and takes all children who are not working in the ﬁrst observation year and compares
17
It is important to understand that these estimates are from an entirely di↵erent model and are thus not
directly comparable. Whereas the baseline model imposes homogeneous e↵ects and estimates the average
impact, this model estimates the impact of working conditional on not working in period t.
21
those who only work in the third year (group 1) to those who work in the second and
third years (group 2) to see if consecutive years of exposure have increasing or decreasing
marginal e↵ects on the students’ test scores. For boys math scores the coe cient estimates
on both the indicator variable for one year working and two years working are negative and
signiﬁcant. Interestingly, the two year indicator variable coe cient is almost double the one
year estimate, suggesting that the e↵ect is essentially linear. Each year of working leads
to about a 3.1 point drop in test scores. For the other regressions the e↵ects could not be
separately identiﬁed perhaps due to the fairly small sample size we are now working with.
Another question we seek to address using the three period sample is the question of
’lingering e↵ects.’ In Table 15 we present coe cient estimates of regressions where we again
start with those that initially do not work and compare those that remain not working for
all three years to those that work in the second year but do not work in the third. The
question is if having worked in the past continues to depress a child’s test score or do they
‘catch up?’ From the results in Table 15 we ﬁnd some evidence that the e↵ects do linger
for boys as for math and Portuguese the coe cient estimates are negative and signiﬁcant
(at the 10% level for math). For girls the point estimate for the math coe cient is negative
and of a similar magnitude to previous regressions but has a large standard error and is
not statistically signiﬁcant while the Portuguese coe cient estimate is both very small and
insigniﬁcant.
5.4 Robustness checks
5.4.1 Preexistent di↵erent trends
We next conduct a series of robustness checks. The ﬁrst set are described in Table 16:
using the three period sample we conduct a series of di↵erence-in-di↵erence estimates on the
22
relative test score increase over the ﬁrst two years for kids with identical work experiences
and condition on their third year (post evaluation) work experience. We expect to ﬁnd no
impact of working in the third year for those with identical work histories.18 Below we show
the four subsamples we use to conduct the robustness checks. Those who did not work in
year t or t + 1 (Treatment 001), those who worked in both year t and year t + 1 (Treatment
111), those who worked in year t but did not work in year t + 1 (Treatment 101)and those
that did not work in year t but did work in year t + 1 (Treatment 011). We will compare
the score trajectory between t and t + 1 for all subsamples.19
In each case we estimate our model for both boys and girls and for both math and
Portuguese for a total of 16 robustness tests. The results for boys are presented in Table
17 and the results for girls are presented in Table 18. In all cases but one the robustness
test yields the expected result: no di↵erence in the test score progression. In one case the
math scores for boys is negative and signiﬁcant. This could suggest that there could be
some selection on trends in scores: boys who observe their score progressing slowly select
into child labor or perhaps some correlation with unobserved prior work history.
5.4.2 Idiosyncratic shocks
The di↵erence in di↵erence estimators may be biased if there are time variant idiosyncratic
shocks correlated to both child labor transitions and variation in test scores such as an
unemployment shock that causes a child to begin working. To the extent that the e↵ects of a
shock on learning are through the physical, mental and time demands that are a consequence
18
We can only identify those with identical observed previous work histories, therefore we cannot completely
exclude the possibility that unobserved work histories are correlated with future work histories. This could
happen for children in households particularly sensitive to income shocks, for example.
19
Note that all 16 regressions are robustness checks of the baseline regressions but only some are relevant
to the ins and outs regressions.
23
of working itself, our estimates are unbiased. It is likely that direct e↵ects from the event
itself are small in comparison. We can test this supposition using the information of the time-
variant employment status of the father. As there are many missing observations and we do
not have information on the parents for 2008, the sample size has decreased considerably from
the benchmark case on the full pooled sample. For this reason we estimate the benchmark
regression both without and with the father’s employment status variable as a control using
this sample in order to gauge the extent of this potential bias.
Table 19 presents the result of the estimations of the benchmark model on the sample
of observations for which we have information on the employment status of the father both
including and excluding the father’s employment status as proxy for idiosyncratic economic
shocks to the household. We ﬁnd that for boys the point estimates for both math and
Portuguese are statistically signiﬁcant and larger than in the benchmark regression on the
pooled sample when father’s employment status is not included as a control. However,
there is almost no change in the size or signiﬁcance of the variable when we control for
father’s employment status suggesting the ﬁxed-e↵ect estimator is not biased upward. For
girls the change in sample size causes the point estimates for both math and Portuguese
to shrink and loose signiﬁcance relative to the benchmark case on the full sample, but
again, there is virtually no di↵erence between the coe cient estimates from the regressions
where father’s employment status is excluded and included. These results suggest that
our coe cient estimates are not biased due to correlation with time variant idiosyncratic
employment shocks to the household.20
20
We have information on the mother’s employment status as well and we ran regressions using mother’s
employment status and both mother’s and father’s employment status and found similar results, but using
mother’s employment status reduces sample size even further.
24
5.5 IV estimator
One may argue that these robustness checks are not su cient since unobservable individual
characteristics or other time variant idiosyncratic shocks beyond father´s unemployment may
be both correlated to labor market transitions and changes in test scores. For instance, a
family break up or health problems of a family member may not only send a child to the labor
maker but might also a↵ect learning. It is important to note that to the extent that such
shocks a↵ect learning through the disruptions caused by the child starting to work, there is
no bias concern as we control for child labor. It is only if there are e↵ects over-and-above
the work itself (perhaps stress) that there may be the potential of bias.
In order to further investigate the potential bias of our ﬁxed e↵ect estimators, we proceed
with an instrumental variable strategy. Brazilian law establishes a minimum legal age to
participate in the labor market.21 No ﬁrm can formally hire a worker younger than sixteen
years old and parents are also subject to legal punishments if an under age child works
in the labor market. Because of this Brazilian ﬁrms rarely hire formal workers under age
of sixteen. However, due to lax enforcement in the informal sector there remains much
employment of children younger than sixteen. Hence, our instrumental variable of the labor
market transition is a variable the indicates whether the individual is legally allowed to work
(older than 192 months old). We estimate the following system of equations:
M arket Laborigt = ↵0 + ↵1 M inimum Ageigt + ↵2 Ageigt + ⌘i + ⌫t + ⇡g + igt
Tigt = 0 + 1 M arket Laborigt + 2 Ageigt + ✓i + t + g + ✏igt
21
Constitutional Amendment No. 20 on December 16th, 1998 which increased the minimum legal age for
entry to labor market from 14 to 16.
25
Where M inimumAgeigt is the indicator variable that it is equal to one if the individual is
sixteen years or older at time t and zero otherwise. Note that we are controlling for individual
i, year t, and grade g ﬁxed e↵ects, and also for age measured in months. Therefore, our
identiﬁcation strategy comes from the variation on age from not allowed to allowed to work
over and above the direct e↵ect of age on the probability of working.
The identiﬁcation assumption is that there is no direct e↵ect of turning sixteen on learning
over and above the directly linear e↵ect of age controlled for in the regression. This is the
Local Average Treatment E↵ect (LATE) described by Imbens and Angrist (1994) and this is
the e↵ect of child labor on proﬁciency among a sub-population of those students induced to
work due to changing their legal status of work (i.e., turning sixteen years old). We perform
this exercise with the paired sample. In this sample, around 8.7% of the individuals move
from not allowed to allowed to work within two years.
The LATE estimator requires that the instrument is valid plus the monotonicity assump-
tion. The instrumental variable is correlated with the potential endogenous variable (this is
shown in Table 20 below). Also, it must not be correlated with the error term of the outcome
equation. This seems reasonable once we control for the individual ﬁxed e↵ect and age in
months. Finally, the monotonicity assumption looks reasonable as well. It requires that all
individuals change their probability to work in the same direction when they turn sixteen.
It does not seem likely that some individual will decrease their probability to work because
they turned sixteen.
Table 20 below presents the results for boys and girls separately. The ﬁrst stage regres-
sions show that boys and girls are more likely to work when they turn sixteen over and
above the linear age e↵ect. In fact, boys (girls) are around 6.5 (4.0) percentage points more
likely to work when they became sixteen years of age. The results of the second stage re-
26
gressions show strong negative impacts on proﬁciency. A working boy (girl) decreases his
proﬁciency in math by 94.7 (39.4) proﬁciency points. For Portuguese, the e↵ects are -91
and -137 proﬁciency points for boys and girls, respectively. The standard deviations of the
proﬁciency scores among individuals aged sixteen years or older ranges from 73.6 for boys in
Portuguese to 80.2 for boys in Math. The standard deviations for girls are 76.7 and 76.8 for
math and Portuguese, respectively. The results implies that the child labor e↵ect for this
subpopulation ranges from 0.5 (1.2) of a standard deviation for girls (boys) in math to 1.8
(1.2) of a standard deviation for girls (boys) in Portuguese.
The IV point-estimates are much larger than the di↵erence in di↵erence estimators. This
may suggest that the ﬁxed e↵ect estimators are attenuated (in absolute terms) due to mea-
surement error of the market labor variable. It is even more likely that the LATE e↵ects are
for a very particular subpopulation of those induced to work due to becoming able to work
legally. Because of the sixteen year old threshold, these are individuals who have already
experienced delay in their progression in school and who are, presumably, mostly going into
formal sector work.22 Thus there could be heterogenous e↵ects such that the impact of child
labor on this subpopulation is much stronger than for working students in general. Another
likely explanation is the classical problem of weak instruments. In this case, even though the
correlation is signiﬁcant it is relatively small as the law directly a↵ects very few children.
5.6 Channels
Finally we explore potential channels through which child labor could be causing the sup-
pression of test scores. Table 21 has the results of the OLS regression of the answers to
the four channels questions on the child labor indicator variable. For three of the variables
22
For example, in the formal sector there is a system of apprenticeships that are expected to lead to
permanent employment and thus may induce particular detachment from schoolwork.
27
there is a positive and signiﬁcant coe cient estimate for both boys and girls: missing classes,
completing homework at school, and turning in homework late. Boys (girls) are 6.4 (3.9)
percentage points more likely miss classes if they work while in school compared to students
who do not work, which translates to 29% (14%) more likely to miss classes. Similarly both
boys and girls are 2.2 percentage points more likely to complete their homework at school
(rather than at home) if they work, which translates to 8% (10%) more likely to complete
homework at home. Finally, boys (girls) are 3.1 (4.4) percentage points more likely to turn in
homework late if they work while in school, which translates to 5% (9%) more likely to turn
in homework late. These results suggest that the time burden of working while in school is
interfering with attendance and the careful and timely completion of assignments.
6 Discussion and Conclusion
Working while in school has negative and lasting consequences for children who participate
in the labor market compared to those who do not. This paper ﬁnds negative and signiﬁcant
impacts of working while in school on the math and Portuguese proﬁciency scores of children
ao Paulo municipal schools. The impacts are economically signiﬁcant. The lower
enrolled in S˜
bound of the average e↵ects of working while in school hovers around 3 percent of a standard
deviation in math scores for boys and 6% for girls, and 5% for Portuguese for boys and 7%
for girls. When we isolate the e↵ects for just those students who transition into child labor
we ﬁnd negative e↵ects of over 6% to over 10%. Extrapolating from the year to year average
gain in proﬁciency scores the average e↵ect of transitioning to work while in school the e↵ect
of working is equivalent to one quarter to an entire year of learning.
We show that these results are robust to idiosyncratic preferences and perform robustness
28
checks to rule out selection on idiosyncratic trends and shocks at the household level. We
ﬁnd some evidence that the e↵ect of working while in school is cumulative and that the
e↵ect lingers over time. We also ﬁnd evidence that the negative e↵ect of child labor while in
school operates through the interference in students’ study time allocation and habits such
as attending class, doing homework outside of class and turning in homework on time.
This is not to say that students that work are not optimizing. Their behavior could be
the consequence of an optimal decision including the cases where they have large discount
rates, are myopic about the future returns and simply prioritize current consumption as
some sociologists in Brazil have suggested. It is also possible that students lack information
about the true returns to learning in the adult labor market. Another explanation could
be that individuals are credit constrained and are unable to borrow against future earnings.
However, this individually rational (or boundedly rational) behavior of working while in
school is likely ine cient in the sense that the real cost to the individual exceeds the beneﬁts
particularly in developing countries like Brazil where returns to education are very high.
Whatever the reason as student learning is highly correlated with adult outcomes as well as
economic growth, working while in school is likely ine cient.
Though it is tempting to suggest that the policy prescription is to prohibit working for
students, one must proceed with caution. It is possible that without the ability to work
while in school these students would drop out of school entirely. In general the ﬁndings of
this paper show that while learning is impaired by working, learning still occurs even when
the child works at the same time. This suggests that working and going to school is better
than not going to school at all. We are also unable to comment on other e↵ects of working
while in school such as grade repetition and dropping out as we are able to study only those
that remain in school. Nevertheless, policy interventions that manage to keep kids in school
29
while curtailing their work activity have the potential of producing a dramatic improvement
in their academic achievement. As the channels results indicate, it appears the time cost of
work activity interferes with the process of learning, so policies that minimize the time spent
in work activities and promote study at home have the potential of signiﬁcantly improving
learning outcomes.
30
References
Akabayashi, Hideo and George Psacharopolous, “The Trade-o↵ Between Child La-
bor and Human Capital Formation: A Tanzanian Case Study.,” Journal of Development
Studies, June 1999, 35, 120–140.
Assaad, Ragui, Deborah Levison, and H. Dang, “How Much Work Is Too Much?
Thresholds in the E↵ect of Child Work on Schooling - The Case of Egypt,” 2005. mimeo.
, , and Nadia Zibani, “The E↵ect of Child Work on School Enrollment in Egypt.,”
Working Papers 0111, Economic Research Forum April 2001.
Baland, Jean-Marie and James A. Robinson, “Is Child Labor Ine cient?,” Journal of
Political Economy, August 2000, 108 (4), 663–679.
Basu, Kaushik, “Child Labor: Cause, Consequence, and Cure, with Remarks on Interna-
tional Labor Standards,” Journal of Economic Literature, September 1999, 37 (3), 1083–
1119.
, “A note on multiple general equilibria with child labor,” Economics Letters, February
2002, 74 (3), 301–308.
and Pham Hoang Van, “The Economics of Child Labor,” American Economic Review,
June 1998, 88 (3), 412–27.
and Zaﬁris Tzannatos, “The Global Child Labor Problem: What Do We Know and
What Can We Do?,” World Bank Economic Review, December 2003, 17 (2), 147–173.
31
Beegle, Kathleen, Rajeev Dehejia, and Roberta Gatti, “Why Should We Care About
Child Labor?: The Education, Labor Market, and Health Consequences of Child Labor,”
Journal of Human Resources, 2009, 44 (4).
, Rajeev H. Dehejia, and Roberta Gatti, “Child labor and agricultural shocks,”
Journal of Development Economics, October 2006, 81 (1), 80–96.
, , , and Sofya Krutikova, “The consequences of child labor : evidence from
longitudinal data in rural Tanzania,” Policy Research Working Paper Series 4677, The
World Bank July 2008.
arcio Eduardo, Ana L´
Bezerra, M´ ucia Kassouf, and Mary Arends-Kuenning, “The
impact of child labor and school quality on academic achievement in Brazil,” IZA Discus-
sion Paper 4062, 2009.
Canals-Cerda, Jose and Cristobal Ridao-Cano, “The dynamics of school and work in
rural Bangladesh,” 2004. World Bank Policy Research Working Paper 3330.
Dessy, Sylvain E., “A defense of compulsive measures against child labor,” Journal of
Development Economics, June 2000, 62 (1), 261–275.
and Stephane Pallage, “Child labor and coordination failures,” Journal of Development
Economics, August 2001, 65 (2), 469–476.
Dumas, Christelle, “Does Work Impede Child Learning? The Case of Senegal,” Economic
Development and Cultural Change, 2012, 60 (4), 773 – 793.
Duryea, Suzanne, David Lam, and Deborah Levison, “E↵ects of economic shocks
on children’s employment and schooling in Brazil,” Journal of Development Economics,
September 2007, 84 (1), 188–214.
32
Edmonds, Eric V., “Child Labor,” in T. Paul Schultz and John A. Strauss, eds., Hand-
book of Development Economics, Vol. 4 of Handbook of Development Economics, Elsevier,
December 2008, chapter 57, pp. 3607–3709.
and Nina Pavcnik, “Child Labor in the Global Economy,” Journal of Economic Per-
spectives, Winter 2005, 19 (1), 199–220.
ortner, “Birth order and the intrahousehold allocation of
Ejrnæs, Mette and Claus C P¨
time and education,” Review of Economics and Statistics, 2004, 86 (4), 1008–1019.
e Portela Souza, “Is There a Child Labor Trap? Inter-
Emerson, Patrick M. and Andr´
generational Persistence of Child Labor in Brazil,” Economic Development and Cultural
Change, January 2003, 51 (2), 375–98.
and , “Is Child Labor Harmful? The Impact of Working Earlier in Life on Adult
Earnings,” Economic Development and Cultural Change, 2011, 59 (2), 345 – 385.
and Shawn D. Knabb, “Opportunity, Inequality and the Intergenerational Transmis-
sion of Child Labour,” Economica, 08 2006, 73 (291), 413–434.
and , “Fiscal Policy, Expectation Traps, And Child Labor,” Economic Inquiry, 07
2007, 45 (3), 453–469.
and , “Self-Fulﬁlling Expectations, Child Labor and Economic Development,” Cana-
dian Journal of Economics, August 2012, 46 (3), 900–927.
Gunnarsson, Victoria, Peter F. Orazem, and Mario A. Sanchez, “Child Labor
and School Achievement in Latin America,” World Bank Economic Review, 2006, 20 (1),
31–54.
33
Hanushek, Eric A. and D. D. Kimko, “Schooling, Labor Force Quality and the Growth
of Nations,” American Economic Review, 2000, 90 (5), 1184–1208.
and L. Zhang, “Quality consistent estimates of international schooling and skill gradi-
ents,” Journal of Human Capital, 2009, 3 (2), 107–143.
Heady, Christopher, “The E↵ect of Child Labor on Learning Achievement,” World De-
velopment, February 2003, 31 (2), 385–398.
Horowitz, Andrew W. and Jian Wang, “Favorite son? Specialized child laborers and
students in poor LDC households,” Journal of Development Economics, April 2004, 73
(2), 631–642.
Ilahi, Nadeem, Peter Orazem, and Guilherme Sedlacek, “The Implications of Child
Labor for Adult Wages, Income and Poverty: Retrospective Evidence from Brazil.,” 2001.
Working Paper, Iowa State University.
Imbens, Guido W. and Joshua D. Angrist, “Identiﬁcation and Estimation of Local
Average Treatment E↵ects,” Econometrica, 1994, 62 (2), pp. 467–475.
Orazem, Peter F. and Chanyoung Lee, “Lifetime Health Consequences of Child Labor
in Brazil.,” Research in Labor Economics, 2010, 31, 90–133.
Ponczek, Vladimir and Andre Portela Souza, “New Evidence of the Causal E↵ect of
Family Size on Child Quality in a Developing Country,” Journal of Human Resources,
2012, 47 (1), 64–106.
Psacharopoulos, George and Harry Anthony Patrinos, “Family size, schooling and
child labor in Peru - An empirical analysis,” Journal of Population Economics, 1997, 10
(4), 387–405.
34
Ravallion, Martin and Quentin Wodon, “Does Child Labour Displace Schooling? Ev-
idence on Behavioural Responses to an Enrollment Subsidy,” Economic Journal, March
2000, 110 (462), C158–75.
35
Tables
Table 1: Full Sample - Descriptive Stats
Boys
Min Max Mean SD
Proﬁciency Math 0 411.87 219.00 43.63
Proﬁciency Portuguese 0 380.02 204.42 51.05
Market Labor 0 1 0.12 0.33
Chores 0 1 0.32 0.47
Age in Months 128 250 166.86 15.15
6th grade 0 1 0.53 0.50
7th grade 0 1 0.10 0.30
8th grade 0 1 0.37 0.48
Late Homework 0 1 0.58 0.49
Homework at School 0 1 0.27 0.45
Prepare for exam 0 1 0.70 0.46
Miss Classes 0 1 0.22 0.41
Father’s Unemployment 0 1 0.11 0.32
Girls
Min Max Mean SD
Proﬁciency Math 0 409.97 216.50 40.55
Proﬁciency Portuguese 0 381.03 218.60 49.20
Market Labor 0 1 0.06 0.23
Chores 0 1 0.56 0.50
Age in Months 119 249 164.64 14.45
6th grade 0 1 0.54 0.50
7th grade 0 1 0.10 0.29
8th grade 0 1 0.36 0.48
Late Homework 0 1 0.51 0.50
Homework at School 0 1 0.22 0.42
Prepare for exam 0 1 0.71 0.45
Miss Classes 0 1 0.27 0.44
Father’s Unemployment 0 1 0.12 0.33
36
Table 2: Paired Sample - Descriptive Stats
Boys
Min Max Mean SD
Proﬁciency Math 99.79 411.86 221.40 41.73
Proﬁciency Portuguese 0 380.0234 206.3126 49.98
Market Labor 0 1 0.11 0.32
Chores 0 1 0.32 0.46
Age in Months 128 241 167.35 14.75
6th grade 0 1 0.51 0.50
7th grade 0 1 0.0054 0.074
8th grade 0 1 0.48 0.49
Girls
Min Max Mean SD
Proﬁciency Math 108.13 409.97 218.96 38.63
Proﬁciency Portuguese 0 381.03 220.86 48.90
Market Labor 0 1 0.055 0.22
Chores 0 1 0.558999 0.49651
Age in Months 124 246 165.70 14.26
6th grade 0 1 0.51 0.50
7th grade 0 1 0.004 0.061
8th grade 0 1 0.48 0.49
Table 3: 3 Period Sample - Descriptive Stats
Boys
Min Max Mean SD
Proﬁciency Math 113.57 391.26 220.88 41.90
Proﬁciency Portuguese 0 352.60 206.61 48.39
Market Labor 0 1 0.07 0.26
Chores 0 1 0.34 0.47
Age in Months 135 232 170.94 16.27
6th grade 0 1 0.50 0.50
7th grade 0 1 0.01 0.08
8th grade 0 1 0.50 0.50
Girls
Min Max Mean SD
Proﬁciency Math 108.13 395.50 217.22 38.77
Proﬁciency Portuguese 0 380.02 218.75 46.93
Market Labor 0 1 0.04 0.19
Chores 0 1 0.58 0.49
Age in Months 131 267 169.18 15.42
6th grade 0 1 0.50 0.50
7th grade 0 1 0.00 0.07
8th grade 0 1 0.49 0.50
37
Table 4: Transition Matrix - Market Labor - Full Sample
Boys
t+1
Not Working Working Total
30,592 4,412 35,004
Not Working
87.40% 12.60%
t 2,734 1,690 4,424
Working
61.80% 38.20%
33,326 6,102 39,428
Total
84.52% 15.48%
Girls
t+1
Not Working Working Total
32,906 2,267 35,173
Not Working
93.55% 6.45%
t 1,251 485 1,736
Working
72.06% 27.94%
34,157 2,752 36,909
Total
92.54% 7.46%
ao Paulo 2010
Table 5: School Attendance and Child Labor - City of S˜
Number and Proportion of 10 to 17 Year Olds
Boys Girls Total
Only School 590.19 591.004 1,181,194
85.39% 86.63% 86.01%
Only Work 17.146 11.37 28.516
2.48% 1.67% 2.08%
School and Work 43.084 36.674 79.758
6.23% 5.38% 5.81%
No School and No Work 40.773 43.155 83.928
5.90% 6.33% 6.11%
Total 691.193 682.203 1,373,396
100% 100% 100%
Source: IBGE Demographic Census 2010.
ao Paulo 2010
Table 6: School Enrollment by Grade and School System - City of S˜
Municipal Schools State Schools Private Schools Total
6th Graders 94.532 64.243 37.893 196.7
48.07% 32.67% 19.27% 100%
7th Graders 92.125 60.03 36.572 188.7
48.81% 31.81% 19.38% 100%
8th Graders 100.742 58.46 34.698 193.9
51.96% 30.15% 17.89% 100%
Total 287.399 182.733 109.163 579.3
49.61% 31.54% 18.84% 100%
Source: INEP/MEC, Censo Escolar 2010.
38
ao Paulo 2010
Table 7: School Attendance by Grade and School System - City of S˜
Public Private Total
Not Work Work Not Work Work Not Work Work
6th Graders 132.838 9.529 33.448 346 166.286 9.875
93.31% 6.69% 98.98% 1.02% 94.39% 5.61%
7th Graders 117.233 12.628 28.962 658 146.195 13.286
90.28% 9.72% 97.78% 2.22% 91.67% 8.33%
8th Graders 132.808 34.124 34.742 1.404 167.55 35.528
79.56% 20.44% 96.12% 3.88% 82.51% 17.49%
Total 382.879 56.281 97.152 2.408 480.031 58.689
87.18% 12.82% 97.58% 2.42% 89.11% 10.89%
Source: IBGE Demographic Census 2010.
Table 8: Occupational Distribution (%) - 2010
Sao Paulo Public School Students: 6th, 7th, and 8th Graders
O ce Work 14.83%
Services and Retail Vendors 25.35%
Industry 23.53%
Domestic services, street vendors, car washers, and others 26.08%
Military Service Occupations 10.21%
Source: IBGE Demographic Census 2010.
39
Table 9: Benchmark Regressions
Full Sample Paired Sample
Math Portuguese Math Portuguese
Boys Girls Boys Girls Boys Girls Boys Girls
Market Labor -1.278*** -1.464*** -1.811*** -3.348*** -1.506*** -1.566*** -1.592*** -3.951***
(0.347) (0.458) (0.463) (0.630) (0.445) (0.568) (0.583) (0.774)
Age in Months 2.127*** 1.654*** 1.631** 1.686*** 2.350*** 1.381*** 1.346* 1.373***
(0.561) (0.347) (0.729) (0.470) (0.590) (0.392) (0.758) (0.522)
Squared Age in Months -0.006*** -0.005*** -0.003*** -0.005*** -0.006*** -0.005*** -0.002*** -0.005***
(0.000) (0.000) (0.001) (0.001) (0.000) (0.000) (0.001) (0.001)
F 1938.69 1449.57 334.15 426.14 1870.48 1371.13 301.01 382.21
N 191,494 190,067 186,035 185,860 80,260 82,485 78,109 80,693
40
Table 10: Interacting with age - ﬁrst observ.
Math Portuguese
Boys Girls Boys Girls
Market Labor -5.098 -0.278 -7.167 6.879
(7.086) (9.341) (9.379) (12.786)
Age in Months 2.367*** 1.378*** 1.372* 1.348***
(0.591) (0.393) (0.759) (0.523)
Squared Age in Months -0.006*** -0.005*** -0.002*** -0.005***
(0.000) (0.000) (0.001) (0.001)
Market Labor⇥Age (1st Obs.) 0.023 -0.008 0.035 -0.069
(0.045) (0.060) (0.060) (0.082)
F 1662.64 1218.75 267.60 339.82
N 80,260 82,485 78,109 80,693
Table 11: Interacting with proﬁciency score - ﬁrst observ.
Math Portuguese
Boys Girls Boys Girls
Market Labor 8.393*** 20.479*** 20.265*** 5.121
(2.744) (3.577) (2.636) (3.688)
Age in Months 2.377*** 1.401*** 1.391* 1.380***
(0.590) (0.392) (0.757) (0.522)
Squared Age in Months -0.006*** -0.005*** -0.002*** -0.005***
(0.000) (0.000) (0.001) (0.001)
Market Labor⇥ Port. Score -0.117*** -0.045**
(0.014) (0.018)
Market Labor⇥ Math Score -0.049*** -0.109***
(0.013) (0.017)
F 1664.29 1224.03 276.18 340.47
N 77,007 79,623 75,825 78,773
41
Table 12: Ins - e↵ect of entering the labor market
Math Portuguese
Boys Girls Boys Girls
Market Labor -3.337*** -2.392*** -3.330*** -5.190***
(0.584) (0.695) (0.760) (0.944)
Age in Months 2.484*** 1.379*** 1.258 1.424***
(0.619) (0.396) (0.792) (0.527)
Squared Age in Months -0.006*** -0.005*** -0.002*** -0.005***
(0.000) (0.000) (0.001) (0.001)
F 1751.92 1331.87 284.25 372.13
N 62,662 69,405 61,237 68,118
Table 13: Outs - e↵ect of leaving the labor market
Math Portuguese
Boys Girls Boys Girls
Market Labor -2.765** -5.318** -3.784** -9.312***
(1.336) (2.164) (1.819) (2.906)
Age in Months -0.108 -12.921** 0.776 -14.027*
(1.975) (5.953) (2.612) (7.759)
Squared Age in Months -0.003** -0.004** -0.000 0.000
(0.001) (0.002) (0.002) (0.002)
F 126.04 47.43 20.71 15.44
N 5,908 2,515 5,725 2,442
Table 14: Exposure e↵ects
Math Portuguese
Boys Girls Boys Girls
One year in the labor market -3.240** -1.174 -3.341** -3.288
(1.389) (1.695) (1.644) (2.046)
Two years in the labor market -6.206*** -3.817 0.248 -6.950
(2.271) (3.632) (2.703) (4.340)
Age in Months 3.967** 1.239** 1.349 0.715
(1.940) (0.622) (2.265) (0.731)
Squared Age in Months -0.005*** -0.003*** -0.003** -0.003***
(0.001) (0.001) (0.001) (0.001)
F 251.90 182.68 74.66 93.72
N 13,154 13,287 12,815 12,995
Table 15: Lingering e↵ects
Math Portuguese
Boys Girls Boys Girls
In the market in the previous year -3.507* -3.869 -4.649** -0.431
(1.905) (2.446) (2.211) (2.896)
Age in Months 4.461* 1.139* 3.356 0.897
(2.578) (0.636) (2.979) (0.744)
Squared Age in Months -0.005*** -0.003*** -0.003** -0.004***
(0.001) (0.001) (0.001) (0.001)
F 250.33 193.17 79.74 102.55
N 11,235 12,300 10,954 12,041
42
Table 16: Robustness check samples
Treatment 001 Treatment 111
Periods t t+1 t+2 Periods t t+1 t+2
Treatment Not Working Not Working Working Treatment Working Working Working
Comparison Not Working Not Working Not Working Comparison Working Working Not Working
Treatment 101 Treatment 011
Periods t t+1 t+2 Periods t t+1 t+2
Treatment Working Not Working Working Treatment Not Working Working Working
Comparison Working Not Working Not Working Comparison Not Working Working Not Working
Table 17: Robustness checks - Boys
Treatment 001 Treatment 111 Treatment 101 Treatment 011
Math Portguese Math Portguese Math Portguese Math Portguese
Treatment 001 -2.846** -1.786
(1.350) (1.668)
Treatment 111 -2.978 1.991
(4.367) (5.896)
Treatment 101 -0.674 -3.313
(3.527) (4.323)
Treatment 011 -2.633 3.919
(2.539) (3.168)
Age in Months 2.279*** 2.760*** 4.545 -2.924 2.710 1.356 2.809 -0.054
(0.808) (1.011) (3.169) (4.578) (2.131) (2.617) (1.935) (2.424)
Squared Age in Months -0.006*** -0.009*** -0.010 -0.009 -0.008 -0.009 -0.008 -0.003
(0.002) (0.003) (0.008) (0.011) (0.006) (0.007) (0.005) (0.007)
F 76.46 29.91 3.75 1.46 4.20 2.65 6.70 2.36
N 11,829 11,514 500 482 947 915 1,236 1,187
Table 18: Robustness checks - Girls
Treatment 001 Treatment 111 Treatment 101 Treatment 011
Math Portguese Math Portguese Math Portguese Math Portguese
Treatment 001 -0.849 -2.116
(1.591) (1.997)
Treatment 111 2.653 2.262
(8.165) (10.173)
Treatment 101 4.326 -3.024
(5.197) (7.150)
Treatment 011 -2.506 1.152
(4.178) (6.054)
Age in Months 3.139*** 2.613*** 4.167 -7.944 5.495* -0.739 -0.422 4.044
(0.760) (0.972) (5.237) (6.538) (3.008) (4.958) (3.073) (4.569)
Squared Age in Months -0.009*** -0.009*** -0.012 0.022 -0.018** -0.003 0.003 -0.006
(0.002) (0.003) (0.015) (0.019) (0.008) (0.011) (0.008) (0.011)
F 38.64 45.57 1.33 0.48 3.55 0.58 2.04 0.82
N 12,672 12,411 166 162 460 441 548 535
43
Table 19: Robustness Check - Father’s Employment Status
Math Portuguese Math Portuguese
Boys Girls Boys Girls Boys Girls Boys Girls
Market Labor -2.232** -1.527 -2.443** -0.633 -2.217** -1.518 -2.443** -0.634
(0.949) (1.192) (1.131) (1.536) (0.950) (1.192) (1.131) (1.535)
Age in Months 1.137 3.059** -0.099 2.229 1.109 3.083** -0.100 2.226
(1.652) (1.341) (1.966) (1.734) (1.652) (1.341) (1.966) (1.734)
Squared Age in Months -0.004*** -0.006*** 0.001 -0.004** -0.004*** -0.006*** 0.001 -0.004**
(0.001) (0.001) (0.002) (0.002) (0.001) (0.001) (0.002) (0.002)
Father Unemployed -2.223** 0.953 -0.076 -0.100
(1.075) (0.979) (1.284) (1.253)
F 149.00 149.70 43.66 47.04 169.61 170.96 49.90 53.77
N 89,647 94,976 88,614 94,040 89,647 94,976 88,614 94,040
Table 20: IV - Minimum Age Labor Law
Math Portuguese
Boys Girls Boys Girls
First Stage
Allowed to work 0.065*** 0.040*** 0.066*** 0.037***
(0.008) (0.007) (0.008) (0.007)
Age in Months 0.002 -0.002 0.000 -0.002
(0.007) (0.003) (0.007) (0.003)
Second Stage
Market Labor -94.723*** -39.382*** -91.044*** -137.220***
(15.454) (14.167) (25.533) (38.725)
Age in Months 0.583 0.606 -0.441 -0.550
(0.865) (0.780) (0.472) (0.657)
N 80260 78109 82485 80693
44
Table 21: Channels
Miss Classes Prepare for exam Homework at school Late homework
Boys Girls Boys Girls Boys Girls Boys Girls
Market Labor 0.064*** 0.039*** 0.003 -0.003 0.022*** 0.022*** 0.031*** 0.044***
(0.006) (0.008) (0.006) (0.008) (0.006) (0.008) (0.007) (0.009)
Age in Months 0.003 0.012** 0.006 -0.005 0.008 0.006 0.008 0.013*
(0.009) (0.006) (0.009) (0.006) (0.010) (0.006) (0.011) (0.007)
Squared Age in Months -0.000*** -0.000*** 0.000 0.000*** -0.000** -0.000*** -0.000 -0.000*
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
F 125.09 794.06 1715.58 1701.32 185.65 156.31 130.36 97.16
N 188,051 187,216 186,432 186,591 187,017 186,965 186,487 186,528
45