Economics Working Papers (2002–2016) Economics

8-1-2003

Child labor and school achievement in Latin

America
Victoria Gunnarsson
International Monetary Fund, [email protected]

Peter F. Orazem
Iowa State University, [email protected]

Mario A. Sánchez
Inter-American Development Bank, [email protected]

Follow this and additional works at: http://lib.dr.iastate.edu/econ_las_workingpapers

Part of the Economics Commons

Recommended Citation
Gunnarsson, Victoria; Orazem, Peter F.; and Sánchez, Mario A., "Child labor and school achievement in Latin America" (2003).
Economics Working Papers (2002–2016). 208.
http://lib.dr.iastate.edu/econ_las_workingpapers/208

This Working Paper is brought to you for free and open access by the Economics at Iowa State University Digital Repository. It has been accepted for
inclusion in Economics Working Papers (2002–2016) by an authorized administrator of Iowa State University Digital Repository. For more
information, please contact [email protected].
Child labor and school achievement in Latin America
Abstract
Child labor's effect on academic achievement is estimated, using unique data on 3rd and 4th graders in 9 Latin
American countries. Cross-country variation in truancy regulations provides an exogenous shift in the ages of
children normally in these grades, providing exogenous variation in opportunity cost of child time. Least-
squares estimates of the impact of child labor on test scores are biased downward, but corrected estimates are
still negative and statistically significant. Children working one standard deviation above the mean have
average scores that are 16% lower on mathematics exams and 11% lower on language exams, consistent with
estimates of the adverse impact of child labor on returns to schooling.

Keywords
child labor, Latin America, academic achievement, schooling, education

Disciplines
Economics

This working paper is available at Iowa State University Digital Repository: http://lib.dr.iastate.edu/econ_las_workingpapers/208
 Child Labor and School Achievement in Latin America

Victoria Gunnarsson, Peter F. Orazem, and Mario A. Sánchez

Abstract

Child labor’s effect on academic achievement is estimated, using unique data on 3rd and 4th

graders in 9 Latin American countries. Cross-country variation in truancy regulations provides

an exogenous shift in the ages of children normally in these grades, providing exogenous

variation in opportunity cost of child time. Least-squares estimates of the impact of child labor

on test scores are biased downward, but corrected estimates are still negative and statistically

significant. Children working one standard deviation above the mean have average scores that

are 16% lower on mathematics exams and 11% lower on language exams, consistent with

estimates of the adverse impact of child labor on returns to schooling.

_______________

Corresponding author: Peter F. Orazem, Department of Economics, 267 Heady Hall, Iowa State

University, Ames, IA 50011-1070. Phone: 515-294-8656. Fax: 515-294-0221. E-mail:

[email protected].
 Child Labor and School Achievement in Latin America

Victoria Gunnarsson, Peter F. Orazem, and Mario A. Sánchez

Abstract: Child labor’s effect on academic achievement is estimated, using unique data on 3rd

and 4th graders in 9 Latin American countries. Cross-country variation in truancy regulations

provides an exogenous shift in the ages of children normally in these grades, providing

exogenous variation in opportunity cost of child time. Least-squares estimates of the impact of

child labor on test scores are biased downward, but corrected estimates are still negative and

statistically significant. Children working one standard deviation above the mean have average

scores that are 16% lower on mathematics exams and 11% lower on language exams, consistent

with estimates of the adverse impact of child labor on returns to schooling.

About one of every eight children in the world is engaged in market work. Despite

general acceptance that child labor is harmful and despite international accords aimed at its

eradication, progress on lowering the incidence of child labor has been slow. While often

associated with poverty, child labor has persisted in some countries that have experienced

substantial improvements in living standards. For example, Latin America, with several

countries in the middle or middle-upper income categories, still has child labor participation rates

that are similar to the world average.

Countries have adopted various policies to combat child labor. Most have opted for legal

prohibitions, but these are only as effective as the enforcement. As many child labor

relationships are in informal settings within family enterprises, enforcement is often difficult.

Several countries, particularly in Latin America, have initiated programs that offer households an

income transfer in exchange for the household keeping their children in school and/or out of the

labor market.

1
 Presumably, governments invest resources to lower child time in the labor market in

anticipation that the child will devote more time to acquisition of human capital. The

government’s return will come from higher average earnings and reduced outlays for poverty

alleviation when the child matures. However, despite a huge acceleration in the research on

child labor, there is surprisingly little evidence that relates child labor to schooling outcomes in

developing countries.1 In fact, most children who work are also in school, suggesting that

perhaps child labor does not lower schooling attainment. Additionally, studies that examine the

impact of child labor on test scores have often found negligible effects, although most of these

are in developed country contexts. More recently, Heady (2003), and Rosati and Rossi (2003)

have found some evidence that child labor lowers primary school test scores in developing

countries.

This study builds on these last two papers by examining the linkage between child labor

and school achievement in 9 countries in Latin America. The current study benefits from more

detailed data sets that allow controls for child, household, school, and community variables, and

it makes use of an empirical strategy that controls for the likely endogeneity of child labor. Our

results are very consistent: in all 9 countries, child labor lowers performance on tests of language

and mathematics proficiency, even when controlling for school and household attributes and for

the joint causality between child labor and school outcomes. To the extent that lower cognitive

attainment translates to lower future earnings, as argued by Glewwe (2002), these results suggest

that there is a payoff in the form of higher future earnings from investing in lowering the

incidence of child labor.

2
 I. LITERATURE REVIEW

Most studies that analyze the relationship between time at work and school attainment

have focused on high school or college students in developed countries.2 These studies have

generally found little evidence that part-time work combined with schooling hurts school

achievement. When adverse effects are found, they are only apparent at relatively high work

hours. Important exceptions include recent papers by Tyler (2003) and Stinebrickner and

Stinebrickner (2003) that found that after controlling for likely endogeneity of child labor,

working while in school led to much larger implied declines in high school math scores and in

college G.P.A.s than had been found previously. Post and Pong (2000) also found a negative

association between work and test scores in samples of 8th graders in many of the 23 countries

they studied.3

There are several reasons why the experience of older working students may not extend

to the experience of young children working in developing countries. Young children may be

less physically able to combine work with school, so that working children may be too tired to

learn efficiently in school or to study afterwards. Children who are tired are also more prone to

illness or injury that can retard academic development. It is possible that working at a young age

disrupts the attainment of basic skills more than it disrupts the acquisition of applied skills for

older students. School and work, which may be complementary activities once a student has

mastered literacy and numeracy, may not be compatible before those basic skills are mastered.

Past research on the consequences of child labor on schooling in developing countries has

concentrated on the impact of child labor on school enrollment or attendance. Here the evidence

is mixed. Patrinos and Psacharopoulos (1997) and Ravallion and Wodon (2000) found that child

labor and school enrollment were not mutually exclusive activities and could even be

complementary activities. However, Rosenzweig and Evenson (1977) and Levy (1985) found

3
evidence that stronger child labor markets lowered school enrollment. There is stronger evidence

that child labor lowers time spent in human capital production, even if it does not lower

enrollment per se. Psacharopoulos (1997) and Sedlacek et al. (2005) reported that child labor

lowered years of school completed and Akabayashi and Psacharopoulos (1999) discovered that

child labor lowered study time.

Nevertheless, school enrollment and attendance are not ideal measures of the potential

harm of child labor on learning because they are merely indicators of the time input into

schooling and not the learning outcomes. Even if child labor lowers time in school, it may not

hinder human capital production if children can use their limited time in school efficiently. This

is particularly true if the schools are of such poor quality that not much learning occurs in the

first place. On the other hand, the common finding that most working children are enrolled in

school may miss the adverse consequences of child labor on learning if child labor is not

complementary with the learning process at the lower grades. A more accurate assessment of the

impact of child labor on human capital production requires measures of learning outcomes, such

as test scores rather than time in school, to determine whether child labor limits or enhances

human capital production. Moreover, evidence suggests that cognitive skills, rather than years of

schooling, are the fundamental determinants of adult wages in developing countries (Glewwe

1996; Moll 1998). Therefore, identifying the impact of child labor on school achievement will

yield more direct implications for child labor’s longer term impacts on earnings and poverty

status later in the child’s life.

Direct evidence of child labor on primary school achievement is quite rare. Heady (2003)

found that child work had little effect on school attendance but had a substantial effect on

learning achievement in reading and mathematics in Ghana. Rosati and Rossi (2003) report that

in Pakistan and Nicaragua, rising hours of child labor is associated with poorer test scores. Both

4
of these studies have weaknesses related to data limitations. Heady treated child labor as

exogenous, but it is plausible that parents send their children to work in part because of poor

academic performance. Rosati and Rossi had no information on teacher or school characteristics,

although these are likely to be correlated with the strength of local child labor markets.

This study makes several important contributions to existing knowledge of the impact of

child labor on schooling outcomes in developing countries. First, it shows how child labor affects

test scores in 9 developing countries, greatly expanding the scope of existing research. Because

the same exam was given in all countries, we can illustrate how the effect of child labor on

cognitive achievement varies across countries that differ greatly in child labor incidence, per

capita income, and school quality. Because the countries also differ in the regulation and

enforcement of child labor laws, we can utilize cross-country variation in schooling ages and

truancy laws to provide plausible instruments for endogenous child labor. Finally, because the

data set includes a wealth of information on parent, family, community, and school attributes, we

can estimate the impact of child labor on schooling outcomes, holding fixed other inputs

commonly assumed to explain variation in schooling outcomes across children. The results are

very consistent. Child labor lowers student achievement in every country. The conclusions are

robust to alternative estimation procedures and specifications. We conclude that child labor has a

significant opportunity cost in the form of foregone human capital production, a cost that may

not be apparent when only looking at enrollment rates for working children.

5
 II. EMPIRICAL MODEL

Ben Porath (1967) laid out the classic model of human capital investments over the life

cycle. There are diminishing marginal returns to time in school because of concavity in the

human capital production process and because the opportunity cost of allocating time to further

skill acquisition increases as skills are accumulated. In addition, finite life spans limit the length

of time to capture returns from schooling as age increases, further decreasing the marginal

returns to time in school as age rises. All of these factors suggest that time invested in human

capital production will decrease as an individual ages. However, early in life, children may

specialize in schooling if the present value of the return is sufficiently high relative to its current

marginal cost.4

We are interested in the tradeoff parents face in deciding whether the child should

specialize in schooling or should divide time between school and work. By age t, the child has

completed Et years of schooling. In addition, the child has matured for t years. The opportunity

cost of child school time is assumed to rise with Et and t, and is also a function of local labor

market conditions Zt. The returns to time in school will depend on how much the child is

expected to learn, Qt. A vector of observable parent, home, school and community variables, Ht

may affect tastes for child labor as well as affecting the productivity of child time in school

through Qt. The child’s labor supply function will be of the form

(1) Ct = c(Et, t, Z t, Qt, Ht, ε t)

where ε t is a random error.

The human capital production process is assumed to depend on past human capital

accumulations, current factors that would make the child’s time in school more productive and

the time spent in school. Letting Qt be an observable measure of cognitive skills produced in

school, the human capital production process will be of the form

6
(2) Qt = q(Et, t, C t, H t, η t )

where η t is a component of cognitive ability that the parents can observe but not the

econometrician.

Because the decision on whether or how much the child works is based in part on the

parents’ knowledge of η t , and because student outcomes are influenced by child labor,

Var( ε t ,η t ) ≠ 0 , and ordinary least squares estimation of (2) will be biased. Short of a

randomized experiment that assigns children into working and not working groups, the best

candidate to resolve the problem will be to find variables that shift the probability that a child

works but do not directly affect child learning in school. We will rely on variables that alter the

local labor market for child labor, Zt, to provide exogenous shifts in the child labor equation in

estimating equation (2).

Factors Shifting the Probability of Child Labor

We require elements of the vector Zt that alter the local labor market for children but do

not affect test scores. Because the probability of working rises with age, factors that alter the age

at which a child would normally be in a given grade will also affect the probability that the child

will be working. In Latin America, the age at which children are expected to start school varies

across countries from 5 to 7 years of age. The age until which children must remain in school

varies from 12 to 16 years of age. As a consequence, children must attend school as few as five

years in Honduras to as many as 10 years in Peru.

These differences in laws regulating child labor and school attendance alter the age at

which children would normally enter grades 3 and 4, and thus vary the opportunity costs between

countries of being in those grades. Children starting school earlier will be younger at grade 3

and will be more likely to attend school full time without working. Third and fourth graders in

7
countries with the lowest working ages are more likely to appear legal, even if they are under 12

years of age. Therefore, children in countries with low truancy ages will be more likely to be

working while attending school.

An alternative measure of the opportunity cost of attending school would be the local

market wage for children. Because most child labor is unpaid work for family enterprises,

however, market wages would not adequately capture the value of time outside school even if

such information were available. Instead, we utilize the presumed upward relationship between

the marginal productivity of child labor and the child’s age which we assume is driven largely by

physical stature.5 Interactions between measures of a country’s school starting age or truancy

age and child age are used to capture exogenous variation across countries in the probability that

third and fourth graders work. These shifts in the net return to time in school provide the needed

exogenous shift in C.6

Within countries, the largest source of variation in demand for child labor occurs across

rural and urban areas. There are more uses for child labor in rural markets, and so rural child

labor force participation rates exceed that for urban children in all the countries in this study. We

capture that source of variation with interactions between child age and a dummy variable

indicating rural residence for boys and girls.

We will illustrate how these elements of Zt affect the probability of engaging in child

labor in figures 1-3 discussed below.

Factors Affecting School Outcomes

8
 Estimation of equation (2) follows the educational production function literature in that Q

is measured by test scores that are explained by variables characterizing the student’s parents,

household, teacher, school and community (Hanushek 1995). Measures used include most of

those that have been found to be important in developing country settings (Hanushek 1995;

Kremer 1995).

Estimates of educational production functions are subject to numerous biases.7 Among

the most commonly discussed is the lack of adequate control for the student's innate ability.8

Many studies have attempted to correct for the problem by using two test scores taken at

different times. If ability has an additive effect on school achievement, the difference between

the two output measures will be purged of the ability effect. The data for the current study only

includes tests taken at one point in time, so the differencing option is not available. However,

there are reasons why undifferenced data may yield satisfactory or even preferred estimates to

the differenced data. As Glewwe (2002) argues, if measures of Ht vary slowly over time, the

value of the differenced measure of achievement is minimal. This is more likely to be true at the

earliest stages of schooling where there is less variation in curriculum, educational materials or

teacher training. Furthermore, the use of parental attributes such as education and income should

partially control for inherited ability. Finally, if there is considerable measurement error in

estimates of Qt, the level of Qt may be measured more reliably than the change in Qt. In any

event, the results of the production function estimation in this study should be interpreted as

cumulative as of grade 3 or 4 rather than the additional learning obtained in that grade.

9
 III. DATA

In 1997, the Latin-American Laboratory of Quality of Education (LLECE) carried out the

First Comparative International Study on Language, Mathematics and Associated Factors for 3rd

and 4th graders in Latin America. LLECE collected data initially in 13 countries, but the required

information for our regression analysis was only available for 9: Argentina, Bolivia, Brazil,

Chile, Colombia, Honduras, Paraguay, Peru, and the Dominican Republic.9

The data set is composed of a stratified sample designed to insure sufficient observations

of public, private, rural, urban and metropolitan students in each country. Data were collected on

40 children from each of 100 schools in each country for a total of 4,000 observations per

country. Half of the students were in the 3rd grade and half in the 4th grade. For budgetary

reasons, LLECE had to use a priori geographic exclusions to limit the transportation and time

costs of data collection. Very small schools with too few 3rd and 4th graders and schools in

remote, difficult to access, or sparsely inhabited regions were excluded. Because of the cost of

translating exams, schools with bilingual or indigenous language instruction were also

excluded.10 As the excluded schools would cater to relatively more disadvantaged populations,

our results should be viewed as applying to school populations that are less rural, from more

majority ethnic groups and somewhat more advantaged than average for all Latin American

children.

Test Scores

Survey instruments consisted of tests administered to the sample of children of the

sampled schools, and self-applied questionnaires to school principals (Pr), to the teachers (T) and

parents (or legal guardians) (P) of the tested children, and to the children themselves (C). In

addition, surveyors collected information on the socioeconomic characteristics of the community

10
(S). A description of the variables used in the Latin America analysis can be found in Table 1.

Summary statistics are reported in Table 2.11

{Tables 1 and 2 about here}

All children were tested in mathematics, and all were tested in Spanish except the

Brazilian children who were tested in Portuguese. It should be noted that the tests and

questionnaires were given only to children who attend school, so no information was obtained on

children who are not in school. Therefore the results can only be applied to enrolled children. If

working children who perform most poorly in school drop out to work full time, our estimate of

the consequences of child labor on schooling outcomes may miss some of those most harmed by

child labor while including children who can work and still perform well in school. However,

95% of children aged 9-11 are enrolled in Latin America, so the bias is likely to be modest.12 In

other settings where primary enrollment rates are much lower, the bias could be substantial.

Child Labor

Child labor is measured by each child’s response to a question asking whether s/he is

engaged in work outside the home.13 Our concentration on paid work outside the home avoids

some definitional problems related to distinguishing between unpaid work for home enterprise

from household chores. However, it is also apparent in our application that child labor in the

home does not have the same apparent negative consequences on student achievement as does

work outside the home.

Table 3 reports child labor participation rates and average test scores for children by

whether they work inside and/or outside the home. The first two columns give the average

language and mathematics test scores for children reporting they worked outside the home often,

sometimes or never for nine of the countries.14 The percentage increase or decrease in average

11
test scores for the groups working less intensively relative to the average for those who work

often are reported in parentheses. Across all 9 countries and two achievement tests, 18 cases in

all, the pattern never varies. Children who work only some of the time outperform those who

often work. Children who almost never work outperform those who work sometimes or often.

The differences are almost always statistically significant. The advantage is large for children

who almost never work over those who often work, averaging 22% on the mathematics exam

and 27% on the language exam. The test advantage for occasional child laborers is smaller but

still significant at 8.4% for mathematics and 9% for languages.

{Table 3 about here}

Children were asked a similar question regarding how intensively they worked inside the

home. It seems that working inside the home is less costly in terms of human capital

development in schools. Taking the average across all countries, those who work often inside

the home have average test scores only 7% lower than those who almost never work inside the

home, and only 4% lower than those who sometimes work in the home. The test score gaps for

those working outside the home were considerably larger. Furthermore, in only 3 of the 9

countries were average test scores significantly higher for children almost never working in the

home relative to those often working in the home. In 3 other countries, those often working in

the home had higher average test scores than did those rarely working in the home.

Nevertheless, there is a more basic reason that we do not analyze the implications of

working inside the home on student achievement: over 95% reported working in the home

sometime or often with nearly identical incidence of reported home work for girls and boys and

for urban and rural children. This lack of meaningful variation in child work in the home means

that the pattern of test scores against home work intensity is unlikely to be reliable. In fact,

empirical models we attempted could not distinguish statistically between children who did or

12
did not work in the home—everyone was predicted to participate in household labor. It is

possible that work in the home is damaging to schooling outcomes, but our data lack sufficient

variation in measured household work to capture the effect. For these reasons, we concentrate

our analysis on child labor outside the home.

Exogenous Variables

We rely on the presumed positive relationship between age and the value of child time

working outside the home to identify the child labor equation. This relationship varies across

urban and rural areas and between boys and girls. It also appears to shift as children reach 10

years of age or more. We allow this effect with a spline defined as follows. A dummy variable,

d10, takes the value of one for children under 10 and zero otherwise. For children aged 10 and

over, the age effect is allowed to captured by interactions between (1-d10) and age.

The countries included in our data differ in legal regulations governing the age at which

children enter school and when they can legally exit. Information on compulsory schooling laws

for each country was obtained from the UNESCO (2002). These laws were allowed to shift the

age-child labor relationship beyond age 10, using interaction terms of the form AGE*(1-

d10)*LAGE where LAGE is the legal age of school entry or school exit.15

The child’s value of time in school will depend on how much the child can learn. This

will depend on the availability of home attributes that are complementary with child time in

school such as books and parental education; and on the quality of the school. Most of these

measures are self-explanatory. However, some of the school variables merit some comment. The

measure of the classroom environment, inadequacy, is a weighted average of several measures of

poor school infrastructure and supplies. Teachers were asked the extent to which they judged

classroom lighting, classroom temperature, classroom hygiene, classroom security, classroom

13
acoustics, language textbooks, mathematics textbooks and other textbooks to be inadequate. The

weighted sum of the responses is used as the aggregate index of school shortages, where the

weights were taken as the first principal component from a factor analysis of the teachers’

responses. The number of Spanish or Portuguese speaking students is included as a measure of

the cost of providing schooling services. As the number of nonnative speakers of the language

of instruction increases, resources must be diverted to second-language instruction, potentially

limiting school productivity.

IV. ECONOMETRIC STRATEGY

The results in Table 3 suggest a strong negative effect of child market labor on school

achievement, but the effect may be in the reverse direction–poor schooling outcomes leading to

child labor. The direction of this bias is difficult to predict. The most plausible is that poor

school performers are sent to work so that the least squares coefficient on child labor will be

biased downward. However, Both Tyler (2003) and Stinebrickner and Stinebrickner (2003)

found biases in the opposite direction for older students, so the better students were more likely

to work. Measurement error in the self-reported incidence of child labor could also bias the

estimated coefficient of child labor on schooling outcomes. The cumulative direction of these

sources of bias cannot be established, but both simultaneity and measurement error can be

handled by the use of plausible instruments that alter the probability of engaging in child labor

without directly affecting test scores.

The first step in the estimation process is to predict child labor. Our categorical measure

of child market work includes 0 (almost never work); 1 (sometime work); and 2 (often work).

We use an ordered probit specification to estimate equation (1), using child, parent, school, and

community variables to explain variation in market work. Predicted child labor from (1) is used

14
as the measure of C in estimating equation (2). This two-stage estimation leads to consistent, but

inefficient estimates of the parameters of the achievement equation. To correct for the

inefficiency in the estimators we utilize a bootstrapping method in which 100 samples with

replacement are drawn from the original data, subjected to the ordered probit estimation and then

inserted into the second stage achievement equation in order to simulate the sampling variation

in the estimates. The bootstrap standard errors are reported for the test score equations.

V. DETERMINANTS OF CHILD LABOR

Estimates from the probit child labor supply equation are reported in Table 4. These

estimates are needed to identify the effect of child labor on test scores, but also have interest in

their own right. The estimation makes use of the dependent variables reported in Table 3 except

that data for Venezuela and Mexico had to be dropped because child age was not reported.

Because the two samples are not identical, we report separate estimates for the samples of

children taking the mathematics and language exams. The coefficients on the age interacted

variables differ somewhat across the two samples, but the overall relationship between age and

child labor is similar between the two samples. The other coefficient estimates are very similar

across the two samples.

{Table 4 about here}

Boys are more likely than girls to work outside the home, and rural boys and girls work

more than their urban counterparts, who in turn work more than their metropolitan counterparts.

Children of more educated parents and children who have access to more books in the home are

less likely to work outside the home. School quality also affects the incidence of child labor.

Schools with inadequate supplies encourage child labor. Children in schools with more non-

Spanish or non-Portuguese language speakers among their peers are also more likely to work

15
outside the home. Schools that offer more classes in Spanish/Portuguese and mathematics per

week also lower the incidence of child labor. In general, these results suggest that better

schooling inputs in the home or at school lower the incidence of child labor. The exception is

that attending preschool does not have a significant effect on child labor in this sample.

The joint test of the null hypothesis that the instrumental variables have no effect on child

labor is easily rejected. Variation in truancy laws across countries and in the market for boys

within countries do shift the probability that children work. We illustrate the impact of these

laws on the average incidence of child labor in Figures 1 and 2. Figure 1 illustrates how raising

the school starting age affects child labor outside the home by age. The effect was disabled

below age 10. As the school starting age rises from age 5 to 7, the probability of child labor rises

about 6 percentage points for a ten year old, all else equal. The effect increases to 10 percentage

points by age 14. Figure 2 shows how child market labor changes with the school leaving age.

As the truancy age rises from 12 to 16, the probability of child labor falls by 8.5 percentage

points for a ten year old. By age 14, the effect rises to 11.5 percentage points by age 14. These

results suggest that truancy laws do have an effect on child labor on average.

{Figures 1 and 2 about here}

Figure 3 illustrates how regional variation in the market for child labor shifts child labor

supply for boys and girls. The dummy variable spline effectively fixes child labor intensity for

children under ten. After age ten, child labor intensity rises for both boys and girls. In each

market, boys work more than girls.16 The higher market labor force participation for boys is

consistent with the presumption that the marginal product of child labor is higher for boys than

girls. However, rural girls have higher labor force participation than metropolitain boys.

{Figure 3 about here}

16
 VI. CHILD LABOR AND SCHOOL ACHIEVEMENT

Table 5 reports the results from estimating equation (2) both with and without controls

for the endogeneity of child labor. In the specification in Table 5, when child labor is treated as

exogenous, it takes the values of 0 (almost never work); 1 (sometime work); or 2 (often work).

When treated as endogenous, child labor is a continuous variable with domain over the real line

taken as the fitted values from the ordered probit estimation in Table 4. The rest of the

regressors are the child, household, parent and school variables used as regressors in Table 4.17

{Table 5 about here}

The impact of child labor on test scores is negative and significant whether or not child

labor is treated as exogenous or endogenous. 18 Because of the difference in the scale of the

measured child labor across the two specifications, it is difficult to directly compare the

magnitude of the implied effect of child labor on test scores. We compare the results in two

ways. First, we compute the implied effect of a one standard deviation increase above the mean

in child labor in each of the equations. These beta coefficients are reported in brackets below the

child labor coefficients. When treated as exogenous, a one standard deviation increase in child

labor causes both mathematics and language tests scores to fall by about .2 standard deviations.

In other words, children working one standard deviation above the mean score on average 8%

lower on mathematics exams and 6% lower on language exams than do otherwise identical

children working at the mean level. When controlling for endogeneity, the effect increases to .4

standard deviation (16%) drop in the mathematics exam and a .3 standard deviation (11%) drop

in the language exam. This finding that the magnitude of the child labor effect on academic

achievement rises after controlling for endogeneity is consistent with results reported by Tyler

(2003) and Stinebrickner and Stinebrickner (2003) for older U.S. students.

17
 Another way to compare the two sets of estimates can be found in Figures 4-5 which

trace out the predicted mathematics and language test scores at each decile of the reported and

predicted child labor distributions. At the break points of the exogenous measure (i.e. going

from child labor level 0 to level 1 at the 40th percentile and from level 1 to level 2 at the 74th

percentile) the predicted test scores using the reported and corrected measures. However, the

relationship is steeper at the upper and lower tails of the distribution of predicted child labor,

particularly for the mathematics test. The implication is that by restricting the range of child

labor to three discrete levels, the impact of child labor on test scores in the first two columns of

Table 5 is understated.

{Figures 4 and 5 about here}

Glewwe’s (2002) review of the human capital literature in developing countries argued

that cognitive ability as measured by test scores is strongly tied to later earnings as an adult. We

would therefore expect that returns to schooling for those who worked as children should be

lower than for those who did not work, all else equal. Consistent with that expectation, Ilahi et

al. (2003) found that, holding constant years of schooling completed, Brazilian adults who

worked as children received 4 to 11 percent lower returns per year of schooling completed. Our

estimates suggest that child labor outside the home reduces achievement per year of schooling

attended by 11 to 16%. Because many of the third and fourth graders in our sample will repeat

the grade, our estimates are an upper-bound measure of the lost human capital per year

completed, and so our results correspond closely in magnitude to their estimates of adverse

impacts of child labor on earnings.

Most of the other variables have similar effects across the two sets of estimates in Table

5. There are two main exceptions. The adverse effects of being a boy or being in a rural school

disappear in the instrumented equations. Gender and rural residence are closely tied to the

18
incidence of child labor. It is likely that the negative effects of being male and being in a rural

area on test scores is related to the indirect effect of these variables on the higher probability that

male and rural children work.

Parental education and availability of books in the home lose influence on test scores

after controlling for the endogeneity in child labor. School attributes also becomes less

important in explaining test scores. Again, these factors had strong negative effects on child

labor, and so part of their positive effect on school outcomes presumably works through their

impact on child school attendance and reduced time at work. The literature on the extent to

which school quality can explain variation in school achievement has emphasized the large

variation in coefficients for the same school inputs across studies and country settings (Hanushek

and Luque 2003). Our results suggest that one reason for the uncertain impact of school

attributes may be that school quality is more important in affecting child school attendance and

child labor than in directly affecting test scores.

VII. CONCLUSIONS

Working outside the home lowers average school achievement in samples of 3rd and 4th

graders in each of the 9 Latin American countries studied. Child labor is shown to have

significant adverse effects on mathematics and language test scores using various specifications

correcting for possible endogeneity and measurement error in self-reported child labor intensity.

Children who work even occasionally score an average of 7 percent lower on language exams

and 7.5 percent lower on mathematics exams. There is some evidence that working more

intensely lowers achievement more, but these results are more speculative in that empirical

models were unable to distinguish clearly between working “sometime” versus working “often”.

19
 These adverse effects of child labor on cognitive ability are consistent in magnitude with

estimated adverse effects of child labor on earnings as an adult. Thus, it is plausible that child

labor serves as a mechanism for intergenerational transmission of poverty, consistent with

empirical evidence presented by Emerson and Souza (2003) and the theoretical models of

poverty traps advanced by Basu (2000), Basu and Van (1998), and Baland and Robinson (2000).

Such large effects suggest that efforts to combat child labor may have substantial payoffs

in the form of increased future earnings or lower poverty rates once children become adults.

How to combat child labor is less clear. Our child labor supply equations suggest that truancy

laws appear to have some effect in lowering the incidence of child labor. However, most of the

variation in child labor occurs within countries and not across countries, so policies must address

local child labor market and poverty conditions as well as national circumstances in combating

child labor. Policies that alter the attractiveness of child labor or bolster household income, such

as income transfer programs that condition receipt on child enrollment or reduced child labor, are

likely candidates. Recent experience on such programs in Brazil, Honduras, Mexico and

Nicaragua would appear to support further development and expansion of such conditional

transfer programs.

20
 NOTES
Victoria Gunnarsson is a research officer in the Fiscal Affairs Department at the International Monetary Fund; her e-
mail address is [email protected]. Peter F. Orazem is a university professor at Iowa State University; his e-mail
address is [email protected]. Mario A. Sánchez is a social development specialist at the InterAmerican Development
Bank; his e-mail address is [email protected]. We thank Wallace Huffman, Robert Mazur, three referees, the
editor, and seminar participants at Iowa State and Minnesota for numerous comments and suggestions. The
findings, interpretations and conclusions are the authors’ own and should not be attributed to the World Bank or the
InterAmerican Development Bank, their Boards of Directors or any of their member countries.

1 Two excellent recent reviews of the recent literature include Basu and Tzannatos (2003) and Edmonds
and Pavcnik (2005).

2 D’Amico (1984); Ehrenberg and Sherman (1987); Howard (1998); Lillydahl (1990); Singh (1998); Stern
(1997); and Singh and Ozturk (2000).

3 This study included several developing countries including Colombia, Iran, South Africa, Thailand, and
the Philippines which had the largest estimated negative effects of child labor on school achievement. However, the
estimates do not control for school attributes or possible joint causality between school achievement and child labor.

4 The main predictions are not altered if leisure is added to the model. It will still be optimal to invest more
intensively in human capital early in life and decreasing investment intensity with age. In addition, because the cost
of leisure is the value of work time, individuals will consume the least leisure when wages are highest. In our
application, children will consume less leisure as they age, and so older children will still be expected to work more
than younger children. Heckman (1976) presents a detailed model of human capital investment, leisure demand and
consumption over the life cycle. Huffman and Orazem (2005) present a much simplified model that generates the
predictions discussed in the text.

5 Rosenzweig (1980) found that in a sample of adults, wages for day labor in India were primarily driven
by stature and not by acquired education. Wage patterns reported by Ray (2000) for boys and girls in Pakistan and
Peru suggest rising opportunity costs of child time as age increases.

6 Angrist and Krueger (1991) use variation in compulsory school starting ages across states to instrument
for endogenous time in school in their analysis of returns to schooling using U.S. Census data. Tyler (2003) uses
variation in state child labor laws to instrument for child labor in his study of U.S. high school tests scores. We
began with a large number of interactions, but the resulting variables were highly collinear, and so we used a
parsimonious subset of the fuller specification.

7 See Glewwe (2002) for a comprehensive review of the problems associated with estimating educational
production functions.

8 Ability bias has also been the subject of numerous papers estimating returns to schooling. The consensus
is that the bias is small (Card 1999). If earnings and cognitive skills are closely tied, as argued by Glewwe (2002),
then the role of ability bias should be small in educational production estimates also.

9 Costa Rica was included in the initial data collection but LLECE dropped their data due to consistency
problems. Cuba was excluded due to missing data on child labor. Mexico and Venezuela lacked required
information on child age.

10 For a detailed description of the a priori exclusions in each country, consult Table III.6 of the Technical
Bulletin of the LLECE.

11 For some reason, language scores were reported for two percent fewer students than were mathematics
scores. The missing scores appear to be due to random reporting errors as there were no large differences between

21
the sample means of the group taking the mathematics and language tests. We report the means from the sample
taking the mathematics exam.

12 Sedlacek et al. (2005) present data on enrollment by age for 18 Latin American countries. Even for the
poorest quintile of children, enrollments rates are over 90% for children aged 9-11.

13 As pointed out by a referee, it would be better to have information on hours of work rather than these
more vague measures of work intensity. Our instrumental variables procedure later is an attempt to correct for
biases due to measurement error in child labor.

14 We only report the averages for the subset of countries for which we had data on both language and test
scores and for which we could match responses for working inside and outside the home. We only had partial
information from Mexico and Venezuela, but we can report that the pattern of average test scores for children
working outside the home in Venezuela and Mexico were the same—children working more outside the home had
significantly lower average test scores. Data limitations prevented us from generating the corresponding average
test scores for children working inside the home for those two countries.

15 We report a more parsimonious specification than the one with all possible interaction terms. In
particular, separate coefficients on the dummy variable (1-d10) and their interactions with age, gender and rural
residence did not add to the explanatory power of the child labor equation.

16 We truncate ages below eight (0.4% of the sample) and above 15 (0.8% of the sample) as we do not
have sufficient observations at the lower and higher ages to generate reliable child labor supply trajectories.

17 We obtained similar estimates of the adverse effect of child labor on test scores when we used a school-
specific fixed effect to control for the impact of variation in school and community variables instead of the vector of
school and community variables.

18 The Davidson-MacKinnon (1993, pp. 237-240) variant of the Hausman test easily rejected the
assumption of exogeneity of child labor. The overidentification tests of the instruments failed to reject the null
hypothesis of exogeneity at the 10th percentile in the language test sample and at the 5th percentile for the
mathematics test sample.

22
 REFERENCES

Akabayashi, Hideo, and George Psacharopoulos. 1999. “The Trade-off between Child Labor and
Human Capital Formation: A Tanzanian Case Study.” Journal of Development Studies 35
(June): 120-40.
Angrist, Joshua D., and Alan B. Krueger. 1991. “Does Compulsory School Attendance Affect
Schooling and Earnings?” Quarterly Journal of Economics 106 (November): 979-1014.
Baland, Jean-Marie, and James A. Robinson. 2000. “Is Child Labor Inefficient?” Journal of
Political Economy 108 (August): 663-679.
Basu, Kaushik. 2000. Analytical Development Economics: The Less Developed Economy
Revisited. Cambridge, MA: The MIT Press.
Basu, Kaushik, and Pham Honang Van. 1998. “The Economics of Child Labor.” American
Economic Review 88 (June): 412-427.
Basu, Kaushik, and Zafiris Tzannatos. 2003. “The Global Child Labor Problem: What do we
know and what can we do?” The World Bank Economic Review 17: 147-173.
Card, David. 1999. “The Causal Effect of Education on Earnings.” In Ashenfelter and Card, eds.
Handbook of Labor Economics Vol. 3A. Amsterdam: Elsevier Science B.V.
D’Amico, Ronald. 1984. “Does Employment During High School Impair Academic Progress?”
Sociology of Education 3 (July): 152-164.
Davidson, Russell, and James G. MacKinnon. 1993. Estimation and Inference in Econometrics
Oxford: Oxford University Press.
Edmonds, Eric V., and Nina Pavcnik. 2005. “Child Labor in the Global Economy.” The Journal
of Economic Perspectives 19 (Winter): 199-220.
Ehrenberg, Ronald G., and Daniel R. Sherman. 1987. “Employment While in College, Academic
Achievement, and Postcollege Outcomes.” The Journal of Human Resources 22
(Winter): 1-23.
Glewwe, Paul. 1996. “The Relevance of Standard Estimates of Rates of Return to Schooling for
Educational Policy: A Critical Assessment.” Journal of Development Economics 51 (2):
267-290.
Glewwe, Paul. 2002. “Schools and Skills in Developing Countries: Educational Policies and
Socioeconomic Outcomes.” Journal of Economic Literature 50 (June): 436-482.
Hanushek, Eric A. 1995. “Interpreting Recent Research on Schooling in Developing Countries.”
The World Bank Research Observer 10 (August): 227-246.
Hanushek, Eric A., and Javier A. Luque. 2003. “Efficiency and Equity in Schools around the
World.” Economics of Education Review 22: 481-502.
Heady, Christopher. 2003. “What is the Effect of Child Labour on Learning Achievement?
Evidence from Ghana.” World Development 31 (February): 385-398.
Heckman, James J. 1976. “A Life-Cycle Model of Earnings, Learning, and Consumption”
Journal of Political Economy 84 (4) (August): S11-44.
Howard, Ian. 1998. “Does Part-Time Employment Affect A-level Grades Achieved?” PSSI
Forum N26 (October): 10-11.
Huffman, Wallace E., and Peter F. Orazem. 2005. “Agriculture and Human Capital in Economic
Growth: Farmers, Schooling and Nutrition.” In Evenson, R.E., P. Pingali, and T.P.
Schultz, eds. Handbook of Agricultural Economics, Vol. 3, North-Holland Publ.,
forthcoming.

23
Ilahi, Nadeem, Peter F., Orazem, and Guilherme Sedlacek. 2003. “How Does Working as a
Child Affect Wages, Income and Poverty as an Adult?” In Peter F. Orazem and
Guilherme Sedlacek eds. Eradicating Child Labor in Latin America in the 90s: The
Promise of Demand Side Interventions. Forthcoming.
Kremer, Michael R. 1995. “Research on Schooling: What We Know and What We Don't: A
Comment on Hanushek.” The World Bank Research Observer 10 (August): 247-254.
Levy, Victor. 1985. “Cropping Pattern, Mechanization, Child Labor, and Fertility Behavior in a
Farming Economy: Rural Egypt.” Economic Development and Cultural Change 33
(July): 777-791.
Lillydahl, Jane H. 1990. ”Academic Achievement and Part-Time Employment of High School
Students.” Journal of Economic Education 21 (Summer): 307-316.
Moll, Peter. 1998. “Primary Schooling, Cognitive Skills and Wages in South Africa.” Economica
65: 263-284.
Patrinos, Harry A., and George Psacharopoulos. 1997. “Family Size, Schooling and Child Labor
in Peru - An Empirical Analysis.” Journal of Population Economics 10 (October): 387-
405.
Post, David, and Suet-ling Pong. 2000. “Employment During Middle School: The Effects on
Academic Achievement in the U.S. and Abroad.” Educational Evaluation and Policy
Analysis 22 (Fall): 273-298.
Psacharopoulos, George. 1997. “Child Labor versus Educational Attainment: Some Evidence
from Latin America.” Journal of Population Economics 10: 337-386.
Ray, Ranjan. 2000. “Child Labor, Child Schooling and their Interaction with Adult Labor:
Empirical Evidence for Peru and Pakistan.” The World Bank Economic Review 13: 3-19.
Ravallion, Martin, and Quentin Wodon. 2000. “Does Child Labor Displace Schooling? Evidence
on Behavioral Responses to an Enrollment Subsidy.” Economic Journal 110: C158-
C175.
Rosati, Furio Camillo, and Mariacristina Rossi. 2003. “Children’s Working Hours and School
Enrollment: Evidence From Pakistan and Nicaragua.” The World Bank Economic Review
17 (2): 283-295.
Rosenzweig, Mark R. 1980. “Neoclassical Theory and the Optimizing Peasant: An Econometric
Analysis of Market Family Labor Supply in a Developing Country.” Quarterly Journal of
Economics 94 (February): 31-55.
Rosenzweig, Mark R., and Robert Evenson. 1977. “Fertility, Schooling and the Economic
Contribution of Children in Rural India: An Econometric Analysis.” Econometrica 45
(July): 1065-79.
Sedlacek, Guilherme, Susanne Duryea, Nadeem Ilahi, and Masaru Sasaki. 2005. “Child Labor,
Schooling, and Poverty in Latin America.” in Peter F. Orazem and Guilherme Sedlacek
eds. Child Labor and Schooling in Latin America. Forthcoming.
Singh, Kusum. 1998. ”Part-Time Employment in High School and its Effect on Academic
Achievement.” The Journal of Educational Research 91 (January/February): 131-139.
Singh, Kusum, and Mehmet Ozturk. 2000. “Effect of Part-time Work on High School
Mathematics and Science Course Taking.” Journal of Educational Research, 91(2): 67–
74.
Stern, David. 1997. “Learning and Earning: The Value of Working for Urban Students.” ERIC
Digest #128.

24
Stinebrickner, Ralph, and Todd R. Stinebrickner. 2003. “Working During School and Academic
Performance.” Journal of Labor Economics 21 (April): 473-491.
Tyler, John H. 2003. “Using State Child Labor Laws to Identify the Effect of School-to-Work
on High School Achievement.” Journal of Labor Economics 21 (April): 381-408.
UNESCO. 2002. Education Statistics 2001-Regional Report on Latin America.
(http://portal.unesco.org/uis). November, 2002.

25
TABLE 1. Variable Description

Endogenous variables

Math Score Mathematics test score (C)

Language Score Language test score (C)
Work Outside Index of how often student works outside the home (0-2) (C)
Often: Student reports that s/he often works outside the home (C)
Sometime: Student reports that s/he sometimes works outside the home (C)
Almost Never: Student reports that s/he almost never works outside the home (C)

Exogenous variables

Child
Age Student age (years) (C)
d10 Dummy if student is below 10 years old
Boy Dummy if student is a boy (C)
No Preschool Student did not attend preschool/kindergarten (C)

Parents/Household
Parent Education Average education of parent(s) or guardian(s) (P)
Books at Home Number of books in student’s home (P)

School
Spanish Enrollment Total number of Spanish (Portuguese) speaking students enrolled (Pr)
Inadequate Supply Index of school supply inadequacy (Pr)
Math/week Number of mathematics classes per week (Pr)
Spanish/week Number of Spanish (Portuguese) classes per week (Pr)

Community (Reference: Metropolitan area with 1M people or more)

Urban Dummy variable indicating if school is located in an urban area (2,500-1M people) (S)
Rural Dummy variable indicating if school is located in a rural area (less than 2,500 people) (S)

Instruments
Legal structure
Compulsory Start Compulsory school ending age in the country (U)
Compulsory End Compulsory school ending age in the country (U)

Sources: C: Child survey or test; P: Parent’s survey; T: Teacher’s survey; Pr: Principle’s survey; S: Survey
Designer’s observation; U: UNESCO estimate (2002).

26
TABLE 2. Summary Statistics

Variable N Mean Std. Dev. Min Max

Endogenous variables

Mathematics Score 20699 14.62 5.87 0 32

Language Score 20290 11.30 4.22 0 19
Work Outside 20699 0.86 0.79 0 2
Often 20699 0.25 0.43 0 1
Sometime 20699 0.36 0.48 0 1
Almost Never 20699 0.39 0.49 0 1
Exogenous variables

Child

Age 20699 9.95 1.59 6 18

d10 20699 0.46 0.50 0 1
Boy 20699 0.50 0.50 0 1
No Preschool 20699 0.25 0.43 0 1
Parents/Household

Parent Education 20699 1.66 1.62 0 6

Books at Home 20699 1.61 1.22 0 4
School

Spanish Enrollment 20699 439.51 548.82 0 452

Inadequacy 20699 3.68 2.73 0 7.93
Math/Week 20699 4.66 3.35 0 30
Community

Urban 20699 0.45 0.50 0 1

Rural 20699 0.35 0.48 0 1
Instruments

Compulsory Start 20699 5.94 0.74 5 7

Compulsory End 20699 13.74 1.13 12 16
Source: Author’s computations based on LLECE data as described in the text.

27
TABLE 3. Unconditional Average Language and Mathematics Test Scores, By Country and Type and Level

of Child Labor for Children Aged 6-18 Years Olda

Working outside of the home Working in the home

Language Test Mathematics Test Language Test Mathematics Test

(Maximum Score = 19) (Maximum Score = 32) (Maximum Score = 19) (Maximum Score = 32)
Country
Argentina
Oftenb 12.0 15.7 13.9 17.9
Sometimec 13.0** d 17.2** 14.3* 18.6*
(8.3%) e (9.6%) (2.9%) (3.9%)
Almost Neverf 14.3** 18.4** 14.7* 19.9 **
(19.2%) (17.2%) (5.8%) (11.2%)
Bolivia
Often 9.7 14.2 11.2 15.9
Sometime 10.1* 14.7 11.2 16.0
(4.1%) (3.5%) (0.0 %) (0.6%)
Almost Never 11.3** 15.1** 11.8 17.2*
(16.5%) (6.3%) (5.4%) (8.2%)
Brazil
Often 11.2 14.4 13.0 16.9
Sometime 11.7* 15.5** 13.4** 18.0**
(4.3%) (7.6%) (3.1%) (6.5%)
Almost Never 13.5** 17.9** 13.0 17.5
(20.5%) (24.3%) (0.0%) (3.6%)
Chile
Often 11.3 13.3 13.4 16.7
Sometime 12.1** 14.8** 13.7 17.3*
(7.1%) (11.3%) (2.2%) (3.6%)
Almost Never 13.5** 16.1** 14.0 17.7*
(19.5%) (21.1%) (4.5%) (6.0%)
Colombia
Often 9.9 13.9 11.7 15.7
Sometime 11.1** 15.3** 12.2* 15.8
(12.1%) (10.1%) (4.3%) (0.6%)
Almost Never 12.4** 15.9** 12.2 16.1
(25.3%) (14.4%) (4.3%) (2.5%)
Dominican Rep.
Often 9.6 12.8 10.3 13.2
Sometime 9.6 13.2 10.8 13.8
(0.0%) (3.1%) (4.8%) (4.5%)
Almost Never 10.8** 13.2 10.2 12.4
(12.5%) (3.1%) (-1.0%) (-6.1%)
Honduras
Often 8.9 11.7 10.2 13.2
Sometime 9.4* 12.3** 10.0 12.7
(5.6%) (5.1%) (-2.0%) (-3.8%)
Almost Never 11.6** 14.5** 9.5 10.8
(30.3%) (23.9%) (-6.9%) (-10.6%)

28
 Working outside of the home Working in the home

Language Test Mathematics Test Language Test Mathematics Test

(Maximum Score = (Maximum Score = (Maximum Score = (Maximum Score =
19) 32) 19) 32)
Country
Paraguay
Often 10.2 12.9 12.5 16.4
Sometime 11.3** 14.9** 13.5** 17.9**
(10.8%) (15.5%) (8.0%) (9.1%)
Almost Never 12.1** 16.4** 11.1 14.8
(18.6 %) (27.1%) (-11.2 %) (-9.8%)
Peru
Often 8.7 11.0 10.6 12.7
Sometime 9.5** 11.2 11.0** 13.5**
(9.2%) (1.8%) (3.8%) (6.3%)
Almost Never 11.2** 12.9** 10.6 13.0
(28.7%) (17.3%) (0.0%) (2.4%)
All Countries
Often 9.9 13.1 11.7 15.4
Sometime 10.8** 14.2** 12.2** 16.1**
(9.0%) (8.4%) (4.3%) (4.5%)
Almost Never 12.6** 16.0** 12.5** 16.5**
(27.3%) (22.1%) (6.8%) (7.1%)
a
Simple mean test score over all children in the child labor group in the county. b Child often works outside the
home when not in school. c Child sometimes works outside the home when not in school. d Indicates that difference
in mean test score from the “often working” group is significant at the 0.05(*) or 0.01(**) level of significance. e
Percentage difference relative to children who often work outside the home when not in school. f Child never works
outside the home.
Source: Authors’ computations based on LLECE data.

29
TABLE 4. Ordered Probit Regression Results on Child Labor

Variable Mathematics Language

Exogenous Variables
Child
Age 0.048** -0.014
(0.009) (0.009)
Boy 0.291** 0.163**
(0.036) (0.037)
No Preschool -0.016 0.029
(0.019) (0.019)
Parents/Household
Parent Education -0.065** -0.046**
(0.007) (0.008)
Books at Home -0.080** -0.071**
(0.012) (0.012)
School
Spanish Enrollment/100 -0.004** -0.005**
(0.002) (0.002)
Inadequate Supply 0.062** 0.065**
(0.009) (0.009)
Math/Week (Spanish/Week) -0.014** -0.010**
(0.004) (0.003)
Community
Rural 0.350** 0.290**
(0.033) (0.034)
Urban 0.197** 0.121**
(0.033) (0.031)
Instruments
Boy*Rural -0.019 0.144**
(0.045) (0.045)
Boy*Urban -0.062 0.103**
(0.043) (0.044)
Age*Compulsory Start*(1-d10) 0.004** 0.002*
(0.001) (0.001)
Age*Compulsory End*(1-d10) -0.002** 0.000
(0.000) (0.001)
LL -21623.743 -21179.099
2
Pseudo-R 0.034 0.034
N 20699 20290
** indicates significance at the 0.05 confidence level. * indicates significance at the 0.10 confidence level.

Standard errors in parentheses. Regressions also include dummy variables that control for missing values.

Source: Authors’ computations based on LLECE data.

30
TABLE 5. Least Squares and Instrumental Variables Equations on Test Scores

Child Labor Exogenousa Child Labor Endogenousb

Variable Mathematics Language Mathematics Language
Work Outside -1.184** -1.087** -7.603** -3.980**
(0.051) (0.036) (1.248) (0.484)
Beta Coefficientc [-0.159] [-0.204] [-0.408] [-0.295]
Child
Age 0.097** 0.045** 0.309** 0.162**
(0.027) (0.019) (0.070) (0.024)
Boy 0.731** -0.165** 2.480** 0.679**
(0.079) (0.056) (0.358) (0.155)
No Preschool -0.256** -0.181** -0.376** -0.079**
(0.093) (0.066) (0.088) (0.040)
Parents/Household
Parent Education 0.327** 0.280** -0.107 0.134**
(0.036) (0.026) (0.106) (0.042)
Books at Home 0.735** 0.497** 0.196** 0.258**
(0.061) (0.042) (0.100) (0.037)
School
Spanish Enrollment/100 -0.046** 0.022** -0.079** 0.007
(0.008) (0.006) (0.010) (0.005)
Inadequate Supply -0.329** -0.357** 0.073 -0.140**
(0.046) (0.031) (0.096) (0.038)
Math/week (Spanish/week) 0.027 0.022** -0.073** -0.049**
(0.017) (0.006) (0.016) (0.012)
Community
Urban 0.730** 0.240** 1.847** 0.794**
(0.107) (0.076) (0.225) (0.117)
Rural -0.692** -0.893** 1.641** 0.275
(0.122) (0.087) (0.410) (0.202)
Constant 13.778** 10.657** 14.400** 8.045**
(0.446) (0.248) (0.453) (0.391)
R2 0.084 0.127 0.063 0.091
N 20699 20290 20699 20290
a
Standard errors in parentheses. b Bootstrap standard errors in parentheses. ** indicates significance at the 0.05
confidence level. * indicates significance at the 0.10 confidence level. c The beta coefficients indicates the number of
standard deviation the test score will change from a one standard deviation increase in child labor. Regressions also
include dummy variables controlling for missing values.
Source: Authors’ computations based on LLECE data.

31
32
33
34
35
36
Iowa State University does not discriminate on the basis of race, color, age,
religion, national origin, sexual orientation, gender identity, sex, marital
status, disability, or status as a U.S. veteran. Inquiries can be directed to
the Director of Equal Opportunity and Diversity, 3680 Beardshear Hall, (515)
294-7612.