Available via license: CC BY 4.0
Content may be subject to copyright.
Measuring Child Labor: Whom Should Be Asked,
and Why It Matters
Guilherme Lichand ( guilherme.lichand@econ.uzh.ch )
Department of Economics, University of Zurich https://orcid.org/0000-0002-9118-1745
Sharon Wolf
University of Pennsylvania
Article
Keywords:
Posted Date: March 21st, 2022
DOI: https://doi.org/10.21203/rs.3.rs-1474562/v1
License: This work is licensed under a Creative Commons Attribution 4.0 International License.
Read Full License
Measuring Child Labor: Whom Should Be Asked,
and Why It Matters
Guilherme Lichand and Sharon Wolf ∗
March 17, 2022
Abstract: Child labor is a pervasive practice; according to the International Labor Or-
ganization, there are 160 million child workers worldwide. That figure might, however,
greatly underestimate the extent of the issue, since child labor indicators are typically
based on surveys with parents – who have no incentive to truthfully disclose that their
children work. This, in turn, poses important challenges to the ability of governments and
international organizations to monitor and enforce children’s rights. This paper provides
first-hand evidence on this issue by combining survey data, based on independent reports
from primary school children and their parents in two cocoa-producing regions of Côte
d’Ivoire, with novel third-party data from costly certification of cocoa production in these
regions, partly based on satellite imagery. We show that adults dramatically under-report
child labor in our study sample by a factor of at least 60%. Calibrating a statistical model
to account for non-linearity in under-reporting and heterogeneity by local characteristics,
we predict the bias-adjusted share of children in employment for each country, and estimate
that child labor might affect over 373 million 7-14 year-old children worldwide (95% CI:
[∼336 million; ∼412 million]), nearly 3-fold its global prevalence according to the World
Development Indicators. In turn, we document that children self-reports provide accurate
regional and aggregate accounts of child labor. Last, evaluating the impacts of a campaign
to discourage child labor, we also show that parents’ reports not only underestimate its
prevalence, but can even lead to the wrong conclusions about whether and how policy
interventions affect child labor.
∗Lichand: Department of Economics, University of Zurich (email: guilherme.lichand@econ.uzh.ch);
Wolf: Graduate School of Education, University of Pennsylvania (email: wolfs@upenn.edu). Study un-
dertaken in partnership with Transforming Education in Cocoa Communities (TRECC) and generously
funded by the Jacobs Foundation Science of Learning initiative. Any views and opinions contained in
this paper are those of the authors and do not necessarily reflect the views or opinions of TRECC or
the Jacobs Foundation. We thank Maxime Deroubaix from ENVERITAS for kindly granting access to
their geo-referenced child labor data. Guilherme is a partner and chairman at Movva, the startup that
implements the nudges evaluated as part of this study. We acknowledge helpful comments from Andrew
Dillon and Megan Passey (International Cocoa Initiative), excellent research assistance by Laura Ogando
and Gabriel de Campos, and field supervision by Nicolo Tomaselli and Innovations for Poverty Action. All
remaining errors are ours.
1
STRUCTURED ABSTRACT:
Introduction: Child labor is a pervasive practice in agriculture, especially in West Africa,
where the global cocoa industry sources about half of its produce. According to 2020 data
from the ILO, 160 million children worldwide were active workers – 9.6% of all 5-17 year-
old children. Strikingly, that figure was over 2-fold in Sub-Saharan Africa, affecting 23.9%
of children. In Côte d’Ivoire, 15% of the cocoa industry employees that year were children.
This figure might, however, greatly underestimate the extent of the issue, since child labor
indicators are typically based on surveys with parents – who have no incentive to truthfully
disclose that their children work. To our knowledge, no study has, to this date, documented
the extent of under-reporting or how to circumvent it in official statistics.
Rationale: We compare survey data, based on independent reports from primary school
children and their parents in cocoa-producing regions of Côte d’Ivoire, with novel third-
party data, from costly certification of cocoa production in these regions. The certifier
leverages satellite imagery to collect prevalence data during the harvest season, when child
labor is very visible – preempting under-reporting. We also take advantage of a randomized
control trial that assigned some Ivorian parents to messages discouraging child labor in
cocoa fields to study whether the estimated impact of the intervention on child labor
depends on how the latter is measured.
Results: We show that adults substantially under-reported child labor, while children
self-reports accurately accounted for it, both within each region and in the aggregate. In
regions with subsequent third-party verification, 45.5% of children reported having worked
in cocoa plantations in the previous month, matching almost exactly the 44.4% prevalence
indicated by the certifier (p-value of the difference = 0.881). In contrast, only 16.2% of
parents in those regions admitted to children in employment – a nearly 2/3 reporting gap
(p= 0.000). Across regions, under-reporting ranged from 60% to 85% (p= 0.000 in each
case). Calibrating a statistical model to account for non-linearity in under-reporting and
heterogeneity by local characteristics, we predict the bias-adjusted share of children in
employment for each country, and estimate that child labor might affect over 373 million
7-14 year-old children worldwide (95% CI: [∼336 million; ∼412 million]), nearly 3-fold its
global prevalence according to the World Development Indicators. We also document that
basing child labor accounts on surveys with parents not only underestimated its prevalence,
but also changed the conclusions about the impacts of the intervention. Concretely, while
messages had no effect on children in employment according to children themselves (p=
0.390), they significantly increased children in employment according to parents (by 55.1%,
p= 0.033) – presumably because the intervention signaled willingness to support farmers
rather than to punish them, partially deterring social desirability bias.
Implications: This study provides first-hand evidence that usual child labor indicators,
based on adults’ survey responses, might dramatically underestimate child labor. Accord-
ing to our estimates, the actual number of child workers could be higher than 373 million
globally. Our results also show that conclusions about how child labor responds to different
interventions might even be reversed if appropriate measurement is not in place. Based
on our findings, children’s self-reports should be used instead as an inexpensive way to
accurately account for child labor.
1 Introduction
Child labor is a pervasive practice in agriculture, especially in West Africa – where the
global cocoa industry sources roughly half of its produce. 2020 data from the International
Labour Organization (ILO) documented that 160 million children worldwide were active
workers, – 9.6% of all 5-17 year-old children (ILO and UNICEF,2021). Strikingly, that
figure was over 2-fold in Sub-Saharan Africa, affecting 23.9% of children. In Côte d’Ivoire,
15% of the cocoa industry employees that year were children (Sadhu et al.,2020). Although
already extremely high, those numbers might have been even higher before: according to
Save the Children, child labor plummeted by 38% globally between 2000 and 2016, declining
from 246 to 152 million (Christian Science Monitor,2019).
Are these figures and trends, however, reliable? Child labor indicators are based on
two components: children in employment (the % of children who work at all) and work
conditions for those employed (from number of hours to occupational hazards). While
child workers are surveyed about their work conditions directly, in the ILO methodology,
it is always parents who are surveyed about whether children work or not. This gives rise
to a challenge: parents1have no incentive to truthfully disclose children in employment.
After all, they might fear that admitting to child labor – actively discouraged since decades
by several NGOs and multilateral organizations – could adversely impact their livelihoods,
either directly (e.g., if they face legal action from child protective services) or indirectly
(e.g., if the companies sourcing from them are punished for child labor in their supply
chain, which could trickle down to lower prices or even to discontinuation of their income).
Because, according to the ILO methodology, child labor is contained in children in employ-
ment (International Labour Organization,2017), if parents in fact under-report children
in employment, then child labor will necessarily be under-estimated in official statistics.
In fact, there are striking differences between children’s and adults’ reports in the few
settings where both are asked independently about children in employment. According to
NORC (a research institute based at the University of Chicago), which surveys children
directly about the number of hours worked in cocoa fields, 38% of 5-17 year olds in Côte
d’Ivoire reported to have worked in 2018-19 (Sadhu et al.,2020); in contrast, the ILO
figure for 2016 – based on adult reports for children employment – was only 23% (ILO
and UNICEF,2021). Even worse, different sources tell very different stories about child
labor’s recent trends. While ILO data indicate a 38% decrease in child labor worldwide
since 2000 (Christian Science Monitor,2019), NORC data record a nearly 65% increase in
child labor since 2008-09 (Sadhu et al.,2020).
While these differences are suggestive that official statistics might be biased, it is not
clear that each measure captures exactly the same phenomenon, due to differences in
geographical coverage, timing of data collection, and survey methodologies. Moreover,
even if they were directly comparable, it is also not clear that children’s reports are free of
biases either, in the absence of objective measurement. To our knowledge, no study has,
to this date, documented the extent of under-reporting in the context of child labor, or
advanced a methodology to circumvent it in official statistics.
This paper combines typical survey data used to measure child labor, based on inde-
pendent reports from primary school children and their parents in two cocoa-producing
regions of Côte d’Ivoire, with novel third-party data from costly certification of cocoa pro-
duction in these regions. The certifier leverages satellite imagery to collect prevalence data
during the harvest season, when child labor is evident – preempting under-reporting. We
show that adults indeed dramatically under-report children in employment. In turn, rel-
1Throughout the text, we use the term ‘parent’ to refer to the child’s primary adult caregiver.
1
atively inexpensive surveys yield reliable prevalence data based on children’s self-reports.
Calibrating a statistical model to account for non-linearity in under-reporting and hetero-
geneity by local characteristics, we predict the bias-adjusted prevalence of child labor for
each country and globally. We also document that basing child labor accounts on surveys
with parents not only underestimates its prevalence, but can also bias evaluations of how
it responds to policy interventions.
2 Background
2.1 How child labor is measured
2.1.1 ILO methodology
The International Labour Organization (ILO) follows the Convention on the Rights of the
Child, the ILO Minimum Age for Admission to Employment Convention (No. 138), and
the ILO Worst Forms of Child Labour Convention (No. 182) to define child labor (ILO
and UNICEF,2021). According to these conventions, whether children in employment
characterizes child labor depends on the child’s age, the number of hours dedicated to
work, and the work conditions. For children less than 11 years old, any employment
characterizes child labor. For those between 12 to 14 years old, 15 or more weekly work
hours or hazardous work conditions characterize child labor. Last, for those between 15
to 17 years old, child labor applies in case of 43 or more weekly work hours or hazardous
work conditions. Figure 8in the Supplementary Materials illustrates all conditions used
by ILO to define child labor for children of different age groups.
Statistics on children in employment, number of hours and work conditions come from
different surveys around the globe. ILO does not collect the data itself, but rather harmo-
nizes data from these different sources to compute the prevalence of child labor according
to its methodology. Child labor is computed based on surveys with adults and children.
ILO’s methodology uses the adult questionnaire to compute children in employment, and
the children’s questionnaire only to assess hazardous work conditions (“[a]s in the previ-
ous rounds, the current round of the Global Estimates of Child Labor uses data obtained
from the adult questionnaire, except for conditions of work, where the information from
the child questionnaire is deemed to be more reliable” ;International Labour Organization,
2017, p. 59).
All leading international organizations follow ILO’s methodology closely or strictly.
For example, UNICEF adapted its Multiple Indicator Cluster surveys after 2013 to match
ILO guidelines. The World Bank also tracks child labor following the same methodology,
only for 7-14 year-olds rather than 5-17 year-olds. As such, all official statistics on child
labor depend crucially on the accuracy of parents’ reports. After all, for under-11-year-old
children, child labor equates to children in employment, and for those 12-17 years old, child
labor is a subset of children in employment.
2.1.2 NORC methodology
NORC, a research institution at the University of Chicago, has tracke child labor in the
cocoa industry for Ghana and Côte d’Ivoire since 2015, building on Tulane University’s
work in the region dating back to 2008. It reports statistics associated with child labor, for
cocoa production in particular and for agricultural activities more broadly (Sadhu et al.,
2020). NORC defines child labor based on the number of work hours and work conditions
for children of different age groups, following ILO’s methodology exactly (see Figure 9in
2
Supplementary Materials). Different from ILO, however, NORC surveys children about
the number of hours worked directly, and defines child labor based on children’s self-
reports (“[u]sing the responses of children relating to engagement in cocoa production, we
generated estimates of children’s engagement in child labor and in hazardous child labor
in cocoa production-related activities”; Sadhu et al.,2020, p.61).
Comparing the different data sources – which differ according to the reporting sources
used for computing children in employment – is telling. 2016 ILO data for Côte d’Ivoire
(https://ilostat.ilo.org/topics/child-labour/) indicated that 17.5% of 5-17 year-
old children engaged in economic activity and household chores. In contrast, NORC data
indicated that, as recently as 2018-19, 64% of all 5-17 year-olds in cocoa-producing regions
of Côte d’Ivoire worked in the past 7 days, and 78% in the past 12 months.
While differences are striking, the NORC surveys cover different geographies and years
than those used to compute official statistics, making it hard to attribute the gaps to
under-reporting by adults in the ILO, UNICEF and World Bank data. Moreover, even
if one would accept that parents report children in employment to a lesser extent than
children themselves (see, for instance, Dillon,2010;Galdo et al.,2020) or that parents’
reports are prone to social desirability biases (Jouvin,2021), it could be that children’s
self-reports are similarly unreliable. Without hard data to provide insight into the actual
prevalence of children in employment free of reporting biases, one simply cannot tell.
2.1.3 ENVERITAS methodology
ENVERITAS is a non-profit startup that certifies coffee (and, more recently, cocoa) com-
panies by verifying farming practices in their supply chain – from chemical usage to child
labor (see, for instance, Tran et al.,2021). The certifier has specialized in reaching small
and remote farmers, who often cannot be reached or even located by traditional certifiers.
To do that, ENVERITAS relies on cloud-free satellite imagery, combined with machine
learning models to identify specific crops (Enveritas,2020).
In Côte d’Ivoire, the certifier has applied this methodology to cocoa farmers to identify
constraints to quality education and early childhood development – including child labor
– in cocoa-growing communities (TRECC,2021). ENVERITAS monitors cocoa harvests
in the country with satellite imagery, surveying farmers specifically during harvest season.
With harvesting activities ongoing at the time of the survey, the potential presence of child
labor becomes apparent, making it much harder for parents to falsely deny it where it is
prevalent. Since 2020, their methodology has been aligned with the ILO’s definitions.
2.2 Background for this study
2.2.1 Study sample
Our study takes place in the cocoa-producing regions of Aboisso and Bouafle in Côte
d’Ivoire. Along with Ghana, the country hosts almost 2/3 of the world’s cocoa production.
This has been linked to one of the highest incidences of child labor worldwide, with nearly
1.6 million children employed in cocoa fields (https://foodtank.com/news/2021/02/norc-
report).
Primary education in Cote d’Ivoire is organized in three cycles: CP (grades 1 and 2),
CE (grades 3 and 4) and CM (grades 5 and 6). Children are assessed at the end of each
cycle (CP2, CE2 and CM2). Our study focuses on those final years of the first two cycles,
comprising 2nd and fourth graders. In our sample, the typical number of students is 49 per
classroom at CP2 and 46 at CE2. In our sample, average grade repetition rates across all
3
primary grades are over 15%, and yearly dropout rates average 4.7% (Lichand and Wolf,
2022).
Table 2provides descriptive statistics for our study sample. Almost all (92%) of partic-
ipating children are 5-11 years old – for whom any form of employment is considered child
labor according to ILO guidelines. Half of children in our sample are girls and live in rural
areas (defined according to their parents’ income source), and slightly over half of them are
enrolled in CP2. Nearly a quarter (22%) of households in our sample are extremely poor
– at baseline, they made at most a little over 1 USD a day. For only 18% of households,
income from all sources was more than 6 USD a day at baseline.
2.2.2 Campaign to discourage child labor
Lichand and Wolf (2022) evaluated an educational program (Eduq+, powered by the Brazil-
ian edtech Movva) that delivered nudges directly to parents’ and teachers’ mobile phones.
Nudges were organized in thematic sequences – comprised of four messages –, with two
messages delivered each week. Content was catered to students’ age group. Messages tried
to encourage parents to participate more actively in their children’s school life. There
was some emphasis on showing up to school, especially in the context of parent-teacher
meetings, and on discouraging harmful practices like corporal punishment as a disciplining
strategy. Several sequences explicitly discouraged child labor in cocoa fields, describing
how it might detract from child development and learning. The intervention was imple-
mented over the entire 2018-19 school year. See Lichand and Wolf (2022) for additional
details of the intervention, including sample nudge sequences.
In the experiment, nudges to parents were cross-randomized with nudges to teachers,
aimed at increasing their attendance and time-on-task while teaching. For the purposes of
this paper, we focus on differences between the treatment condition that had only parents
nudged and the control group. Lichand and Wolf (2022) and Wolf and Lichand (2022)
compile results on the impacts of the intervention on educational outcomes, and Wolf and
Lichand (2022) discusses the full range of effects of the different treatment arms on different
measures of child employment (above and beyond those used in the context of this study).
3 Data and outcomes
3.1 Third-party data
We use child labor data from ENVERITAS collected in early 2020, during the harvest
season in Aboisso and in two sub-regions of Bouafle (Bouafle 2 and Tiapoum Adiake).
ENVERITAS’s sampling frame relied on geographical units of 10,000 farmers, identified
via satellite imagery. They randomly drew 125 farmers to be surveyed in each unit. 8,150
households were approached by ENVERITAS, 7,402 of which were successfully surveyed.
The certifier also surveyed schools in regions outside of our study sample; see Figure 1.
We do not use data in these other regions, or collected prior to 2020 (before the survey
instrument was consistent with ILO’s methodology regarding the definition of child labor2).
2Which considers three main aspects: number of hours, age, and activities.
4
Figure 1: Geographical coverage of ENVERITAS data
Surveyed participants into these defined areas were asked “Do any of your children
between 6 and 16 years old help you work on the cocoa farm?”. Children were not surveyed
directly by ENVERITAS; thus far, they have only piloted surveys with children in Tonkpi,
a region outside our study sample. Nevertheless, as the data collection happens during the
harvest season in cocoa household areas, it becomes more difficult for parents to under-
report during interviews if their children are engaging in work activities.
3.2 Survey data
We rely on data collected in the context of a broader project that includes companion
papers Lichand and Wolf (2022), Wolf and Lichand (2022) and Finch et al. (2022). All
details of the experimental design and pre-analysis plans were pre-registered at the AEA
RCT Registry on October 31, 2018 (AEARCTR-0003385).
Our study comprises 198 CP2 and CE2 classrooms (second and fourth grades, respec-
tively) across 99 Ivorian public schools in the cocoa-producing regions of Aboisso and
Bouafle. Within each school, we randomly drew 13 CP2 students and 12 CE2 students
to be surveyed at baseline (at the beginning of the school year, in October 2018) and end
line (at the end of the school year, in June 2019). Importantly, children and parents were
surveyed independently. Enumerators ensured that this was the case, especially since we
also tested children’s numeracy and literacy skills as part of the broader project, which
required them to sit by themselves – only accompanied by our survey team.
We also conducted an additional follow-up survey at the beginning of the following
school year (in October 2019). In this follow-up, we surveyed all teachers and only one
parent per classroom, but no children. This follow-up data focused on collecting additional
information about work conditions for children in employment, but it also provides a
measure of children in employment according to parents already after the growing season
(Yoroba et al.,2019) and much closer to the timeline of the ENVERITAS data collection.
The sample comprises 1,285 CP2 students and 1,190 CE2 students surveyed at baseline
along with their primary parents, in addition to their 198 teachers. We were able to
track all teachers, 1,157 CP2 students (90.0%) and 1,086 CE2 students (91.3%) at end
line. We assigned replacement households in case the ones drawn could not be tracked by
enumerators.
Table 1compiles our survey questions related to children in employment in each wave.
As indicated, we asked parents and children the same question about children in employ-
5
ment in cocoa fields at both baseline and end line. This aligned with the data collected
by ENVERITAS and NORC, and follows ILO’s methodology for measuring children in
employment. Even though we also surveyed parents and children about other forms of
employment (e.g., domestic work and construction), we do not analyze these data in this
paper to ensure comparability across the different data sources. These additional measures
of children in employment are described and analyzed in Wolf and Lichand (2022).
4 Methods
4.1 Assessing the accuracy of different reporting sources
We assess the reliability of reported children in employment according to parents and
children by comparing it to the ENVERITAS data, both within each region and in the
aggregate, considering the three geographical units for which both data overlap.
We report p-values from tests of differences in proportions of children in employment
according to each source, considering equal population variances when comparing parents’
and children’s reports (through an Ordinary Least Squares regression, given the paired
design), and unequal population variances when comparing any of them to ENVERITAS
data – since the latter was collected from a different sample. We cluster standard errors
at the regional level.
4.2 Predicting the bias-adjusted prevalence of child labor
Treating ENVERITAS data as the ground truth for the prevalence of children in em-
ployment in each region, we can estimate the relationship between adult reports (the ILO
standard measure) and this ground truth not only to compute the extent of under-reporting
in our data, but also to predict what accurate reports would have been in geographical
units without ENVERITAS data. Here we show how estimating this bias-adjustment fac-
tor for children in employment allows us to compute the bias-adjusted prevalence of child
labor for each country, and globally.
According to the ILO definition:
CLi=CEixH Ci,(1)
where CLiis the share of child labor in country i,CEiis the share of children in employ-
ment in country i, and HCiis the share of children in employment in country iwho work
in hazardous conditions (= 1 for children under 12, and = % of those working long hours
and/or under heavy or dangerous work conditions, otherwise).
Once we determine that children’s self-reports are accurate and, as such, can be used
as the ground truth, in our data we observe both CEj, the share of children in employment
reported by parents in classroom j, and CE∗
j, the ground truth for children in employment
in classroom j. We are interested in predicting C E∗
ifor all countries i∈I. To do that, we
first estimate:
CE∗
j=f(CEj, Xj) + εj,(2)
where f(CEj, Xj)is a function of children in employment reported by parents in classroom
jand other classroom characteristics Xj, and εjis an error term.
We then use our estimates to compute:
ˆ
CE∗
i= ˆαixCEi,(3)
6
where ˆαi=ˆ
f(CEi,Xi)
CEi≥1.
Those same estimates can be used to recover CL∗
i, the ground truth for child labor in
country i. This is thanks to the fact that H Ciis free of bias (since = 1 for 5-11 year olds,
and elicited directly from 12-14 year-old children), and to the multiplicative nature of its
definition:
CL∗
i=CE∗
ixHCi= ˆαixCEixH Ci= ˆαixC Li(4)
Concretely, we estimate a quadratic polynomial, allowing the extent of under-reporting
to vary with both the local level of children in employment reported by parents and with
the local urbanization rate:
CEc
j=
2
X
k=0
βkCEp
jk+
2
X
k=0
γkCEp
jkxurbanj+εj,(5)
where CEc
jis % child labor according to children in classroom j,CEp
j, that according to
parents in classroom j, and urbanjis the % of students from urban areas in classroom j.
We only estimate heterogeneity with respect to children in employment reported by
adults and urbanization rates because these are available both in our data and in the
World Development Indicators. Other variables, like annual per capita income, are not
available in our data.
Last, we compute ˆαi=max n1,ˆ
CE c
i
CEio, where ˆ
CEciis the predicted value of CE cfor
country iusing equation (5). As the formula indicates, we constrain the estimated bias-
adjustment factor (as well as its confidence interval) to be greater or equal to 1. In practice,
we do this simply by censoring values that violate the constraint by imposing ˆ
CEc
i=CEi.
This affects only 3 out of 97 countries with child labor data in the World Development
Indicators.
World Development Indicators track child labor for 97 countries (those where the issue
is considered to be relevant by the data collection organizations, such as UNICEF country
offices), focusing on 7-14 year-old children. With the bias-adjustment factor that we predict
for each of these countries using the procedure described above, we compute country-level
bias-adjusted prevalence of child labor and number of child workers (and accompanying
95% CIs) using data from the most recent year available for each country in the World
Development Indicators. To arrive at the bias-adjusted prevalence figures, we have to
rely, in addition, on World Development Indicators data on the number of 0-14 year old
children by country, and subtract these from the UNICEF figures for under-5 children
by country. For both population counts, we use 2020 data – assuming that child labor
indicators remained constant since their most recent measurement. The global prevalence
of child workers is obtained simply by summing over country-level figures.
4.3 Assessing the sensitivity of evaluation results to different reporting
sources for children in employment
To document whether different reporting sources might lead to bias in evaluating the
impacts of interventions to discourage child labor, we contrast effect sizes of nudges to
parents on the prevalence of children in employment in cocoa fields based on children’s
self-reports or adult reports.
In Supplementary Materials, we also contrast the relationship between the classroom-
level prevalence of child labor and average standardized test scores comparing, within
7
each treatment cell of the experiment, the slope of linear associations between the two
outcomes measures when child labor is based on children’s self-reports and when it is
based on parents’ indirect reports.
5 Results
5.1 Descriptive statistics based on children’s self-reports
We start by documenting the aggregate prevalence of children in employment, and that
by student characteristics, according to children’s self-reports at baseline. Figure 2shows
that 38.1% of children reported to have worked at least one hour in cocoa fields over
the previous month. As a point of comparison, by the end of the school year (closer to
harvest season), this figure was up to 50% (see Figure 10 in Supplementary Materials). The
prevalence of child labor was only slightly higher among fourth graders than that among
second graders (39.6% vs. 36.8% at baseline, and 51% vs. 48% at end line). Boys were
nearly 50% more likely to work in cocoa fields at baseline than girls (44.4% vs. 31.2%).
The baseline prevalence of children in employment was higher within the bottom income
bracket (38.9%), but not low even within the highest income bracket (23.9%). Naturally,
child employment in cocoa fields was much higher in rural areas (52.4% vs. 23.9%). For
an account of adult reports about children’s work conditions, collected at the follow-up
surveys, see Figure 10.
Figure 2: Share of students who worked for at least one hour in cocoa plantations over the
last month, according to children
Notes: Figure 2shows the share of students who report to have worked in cocoa plantations in the last
month for one hour or more at baseline, in response to the following question: “In the last month, have
you engaged in one or more of the following activities, for one hour or more? Work in a cocoa plantation”,
as described in Table 1. The first bar comprises the whole sample in Lichand and Wolf (2022); additional
bars consider the indicated sub-samples. “Poorest” comprises households with monthly income reported
8
by parents below 10,000 CFA (∼19 USD), while “Richest” comprises those with monthly income reported
by parents above 200,000 CFA (∼372 USD). Rural and urban areas are defined according to parents’
main occupation (agricultural or plantation activities are assigned to the former). Samples sizes are the
following: (i) Total: 2,475; (ii) CP2: 1,285; (iii) CE2: 1,190; (iv) Boys: 1,237; (v) Girls: 1,238; (vi) Poorest:
139; (vii) Richest: 163; (viii) Urban: 1,234; (ix) Rural: 1,237.
9
Figure 3turns to correlations between classroom-level prevalence of child labor and
educational outcomes, documented through binscatter plots. Panel A illustrates the rela-
tionship between the summary measure of test scores (which averages over numeracy and
literacy standardized test scores, following Kling et al.,2007) and the % children in em-
ployment, both measured at baseline. Panel B turns to student dropouts over the course of
the school year, available at the classroom level from administrative data from the Ministry
of Education.
Figure 3: Baseline correlation between child labor and educational outcomes
Panel A: Correlation between test scores and child labor
Panel B: Correlation between student dropouts and child labor
Notes: Panel A reports a bin-scatter plot of baseline standardized test scores as a function of baseline child
labor in cocoa fields, reported by children. Standardized test scores are a summary measure of numeracy
and literacy test scores (averaging across each component, normalized by their mean and standard deviation
in the control group), following Kling et al. (2007). Children in employment stands for the baseline share
of students who report to have worked in cocoa plantations in the last month for one hour or more, in
response to the following question: “In the last month, have you engaged in one or more of the following
10
activities, for one hour or more? Work in a cocoa plantation”, as described in Table 1. Test scores and
the prevalence of child labor are averaged at the classroom level. Panel B reports a bin-scatter plot of
student dropout rate as a function of baseline child labor in cocoa fields, reported by children, based on
administrative data (see Lichand and Wolf,2022). Student dropouts and the prevalence of child labor
are averaged at classroom level. Because student dropouts are defined at the end line, Panel B restricts
observations to the control group of the intervention.
Consistent with common sense, classrooms with a higher share of children in employ-
ment at baseline feature lower test scores by the beginning of the school year, and higher
dropout rates over the course of the school year. Estimating a linear relationship between
the variables in each case suggests that moving from 0% to 20% children in employment
is associated with about 0.08 s.d. lower test scores – what children tend to learn in one
school quarter, and the magnitude of effect sizes of many educational interventions, such
as nudges to parents evaluated in this setting (Lichand and Wolf,2022). Similarly, moving
from 0% to 40% children in employment is associated with roughly doubling dropout rates.
While these associations are not causal, they help understand the centrality of the issue
for governments and international organizations monitoring children’s rights. This is why
accurate measurement is key.
5.2 Validating survey data with costly certification data
We now turn to the comparisons between independent reports by parents and children
in our end-line survey data, and ENVERITAS data. Figure 4presents the prevalence
of children in employment according to each reporting source, along with p-values for
pair-wise statistical tests of differences in proportions. Panel A considers the aggregate
prevalence figures for the regions where both data sources overlap, and Panel B documents
comparisons within each region.
11
Figure 4: Validation of child labor measures using third-party data
Panel A: Share of students who worked in cocoa plantations,
by reporting source
Panel B: Share of students who worked in cocoa plantations, by
source and region
Notes: Bars show the share of children who worked at least an hour in cocoa fields over the previous
month, according to children (in red), parents (in blue) and ENVERITAS (in green). Panel A reports
the average prevalence across all regions for which survey data overlaps with ENVERITAS data. Panel
B breaks down prevalence by region. Children answered the following question at end line: “In the last
12
month, have you engaged in one or more of the following activities, for one hour or more? Work in a cocoa
plantation”, as described in Table 1. Parents answered the following question at end line: “I will now ask
you some questions about activities that your children might have recently performed. In the last month,
has any of your children engaged in one or more of the following activities, for one hour or more? Work in
a cocoa plantation”, as described in Table 1. In the survey conducted by ENVERITAS during the harvest
season (identified through satellite imagery), farmers answered the following question: “Do any of your
children between 6 and 16 years old help you work on the cocoa farm?”. For the measures reported by
children and parents, observations are restricted to the control group of the intervention. P-values from
tests of proportions with unequal population variances (when children’s or parents’ reports are compared
to ENVERITAS data; accounting for each source’s intracluster correlation computed at the regional level,
in Panel A), and from tests of proportions with equal population variances (when comparing children’s
and parents’ reports) through Ordinary Least Squares regressions (clustering standard errors at regional
level, in Panel A).
If anything, in the absence of under-reporting, parents should answer affirmatively more
often than children in our survey, because our survey asked them about whether any of
their children worked in cocoa fields over the previous month – not necessarily the child
whom we surveyed in the context of the study. This is, however, strictly at odds with what
we find.
Panel A documents that, in regions with subsequent third-party verification, 45.5% of
children reported to have worked in cocoa plantations in the previous month, matching
almost exactly the 44.4% prevalence indicated by the certifier (p-value of the difference
= 0.881). In contrast, only 16.2% of parents in those regions admitted to employing
children – a nearly 2/3 reporting gap (p= 0.000). Panel B shows that, across regions,
under-reporting by adults was striking, ranging from 60% to 85% (p= 0.000 in each case).
Children’s self-reports, in turn, range from 75.6% to 134% of ENVERITAS figures, and are
not statistically different from ENVERITAS data in two out of three regions (p= 0.488 in
Aboisso, p= 0.016 in Bouafle 2, and p= 0.120 in Triapoum Adiake).
5.2.1 Follow-up survey
While children were not surveyed in the follow-up wave (October 2020), it is useful to
compare these data to our end-line survey to gauge whether adult reports of children in
employment converged to children’s self-reports as harvest season was approaching – closer
to when ENVERITAS collected their data. Figure 10 shows that this was not the case.
At the follow-up wave, only 28% of parents admitted that children worked at all in cocoa
fields during the school year. This figure was still only about half that reported by children
at end line when it comes to employment in cocoa fields over the previous month (50%),
even though parents were reporting on any work during the entirety of the previous school
year.
Figure 11 further documents that teacher reports of child labor throughout the school
year were about the same as parents’ reports, suggesting that accurate measurement can-
not be achieved by merely surveying other adults – either because of imperfect information,
or because they are also sensitive to social desirability biases. Figure 12 presents suggestive
evidence that, in effect, all adults are affected by such biases. When asked to report on
whether children worked 11+ hours a week (which characterizes child labor for all children
under-15), teachers were more conservative than parents if the question focused on em-
ployment during the school year, but less so if it focused on school holidays. Concretely,
teachers identified child labor for only 2.8% of students during the school year (8%, ac-
cording to parents) but for 12% of students during school holidays (10%, according to
13
parents).
5.3 Bias-adjusted prevalence of child labor
Having documented that children’ self-reports accurately account for children in employ-
ment, we can use these data to estimate the relationship between adult reports and the
ground truth. Table 3documents the results. Even though some coefficients are imprecisely
estimated (as we only have 198 classroom-level observations), as anticipated, the relation-
ship between adult reports and the ground truth is non-linear, and highly dependent on
the share of urban population.
Figure 5illustrates the data and regression results through a binscatter plot. For al-
most all classrooms, adult reports are higher than children’s self-reports (only 6 out of 198
observations are below the 45-degree line). The red line documents the quadratic rela-
tionship for mostly-rural classrooms (those with less than 20% of children living in urban
areas), and the blue line, that for mostly-urban classrooms (with over 80% of children living
win urban areas). No mostly-urban classroom has a prevalence of children in employment
greater than 20%. Within that lower range, however, under-reporting can be very high;
for instance, the estimated relationship predicts that if parents in these classrooms report
a 10% prevalence of children in employment, the bias-adjusted prevalence is actually closer
to 30%. For mostly-rural classrooms, under-reporting is predicted to be even the more
striking the lower the prevalence in adult reports is. In these classrooms, a 10% reported
prevalence would correspond to a nearly 45% bias-adjusted prevalence of children in em-
ployment. Naturally, in both cases, there is less room for under-reporting as prevalence in
adult reports increases.
Figure 5: Correlation between parents’ and children’s answer according to urbanization
level
14
Notes: Data is aggregated at classroom level: average child labor according to children and parents
were correlated according to composition on the classroom urbanization. “Urban >80%” represents the
quadratic fit between children’s and parent’s answer on child labor to classrooms that have more than
80% of their students in the urban category, while “Urban <20%” represents the quadratic fit between
children’s and parent’s answer on child labor to classrooms that have less than 20% of their students in the
urban category, i.e., majority of students are from rural areas. Children answered the following question
at baseline: “In the last month, have you engaged in one or more of the following activities, for one hour
or more? Work in a cocoa plantation”, as described in Table 1. Parents answered the following question
at baseline: “I will now ask you some questions about activities that your children might have recently
performed. In the last month, has any of your children engaged in one or more of the following activities,
for one hour or more? Work in a cocoa plantation”, as described in Table 1. Rural and urban areas are
defined according to parents’ main occupation (agricultural or plantation activities are assigned to the
former).
Based on these estimates, Table 4displays bias-adjusted shares and the prevalence
of child workers by each country for which the World Development Indicators track this
statistic, as well as global figures. As the table shows, while WDI statistics account for
∼136 mi child workers among 7-14 year-old children, we estimate that this figure is actu-
ally ∼373.5 mi (95% CI: ∼336mi, ∼412mi) once under-reporting is accounted for. Figure
6illustrates how correcting for under-reporting with our estimates shifts the global distri-
bution of child labor (for the countries where it is measured) to the right. Because of the
non-linear nature of under-reporting, the lower-end of the distribution is disproportionately
affected by the bias adjustment – and especially so for countries with lower urbanization
rates. Importantly, the histogram does not feature countries that do not even track child
labor, presumably because its current prevalence is zero or very close to it (Table 4’s notes
list the countries without prevalence data in the World Development Indicators).
Figure 6: Global prevalence distribution (subset of countries where child labor is measured)
Notes: Using data from World Development Indicators available at https://databank.worldbank.org
/source/world- development- indicators# we estimate a world bias-adjusted distribution to children in
employment based on the findings from Figure 5. Statistics extracted were “Children in employment (% of
children ages 7-14)” and “Rural population (% of total population)” to all available countries. Prediction
15
based on the single-equation regression CLc=α+β1C Lp+β2U rban +β3CLp
2+β4CLp∗U rban+β5C Lp
2∗
Urban +ε, which: CLcis child labor according to children, CLpis child labor according to parents, and
Urban the percentage of students living in urban areas. Results are displayed in Table 3. We impose a
restriction in which adjusted-bias estimate cannot be smaller than actual child labor statistic, and if this
is the case we equalize the bias-adjusted estimate to the actual statistic. Additionally, to estimate the
bias-adjusted prevalence we have to rely on the WDI estimates for the number of 0-14 year old children by
country, and subtract from it the UNICEF estimates for under-5 children by country. For both population
variables, we use the 2020 data, assuming the WDI child labor indicators remained constant since the last
measurement. Specific country estimates can be find at Table 4. Several countries had missing information
and were excluded from the analysis as indicated in Table 4.
The countries most affect by the bias adjustment (in % change) are India (from 1.7%
child labor in the World Development Indicators to 36.3% according to our estimates),
Romania (from 1.4% to 29.4%), Costa Rica (from 1.3% to 20%), Lesotho (from 2.6% to
39%) and Jordan (from 1.2% to 16.1%). On average, child labor figures increase by 321%
due to our bias adjustment procedure.
Do these predictions make sense? While it is impossible to verify them in each case,
precisely because there is no systematic data on self-reports of children employment, we
discuss the cases of Ghana and Brazil, both of which have data on children in employment
according to both children and adults.
For Ghana, our adjustment leads estimates of child labor in 2012 to increase from
28.8% in the World Development Indicators to 46.7%. This is much closer to the 55%
figure in the NORC data, reported by children within cocoa-producing regions. For Brazil,
our adjustment leads estimates of child labor in 2015 to increase from 2.5% in the World
Development Indicators to 19.2%. The Brazilian Basic Education Evaluation System (Sis-
tema de Avaliação da Educação Básica, SAEB) includes questions about employment and
time use in its nationally representative student survey. In 2019, 15% of fifth-graders re-
ported working outside of home (https://novo.qedu.org.br/questionarios-saeb/al
unos-5ano/7-brasil?dependencia_id=5). If we add the 2% of 6-17 year-old children
who were not enrolled in school in 2019 (https://biblioteca.ibge.gov.br/visualiza
cao/livros/liv101736_informativo.pdf), assuming all of those work, that figure would
reach 17% – which is very close to our estimate.
5.4 How adult reports can also bias policy evaluations
Last, we present evidence that relying on parents’ reports for children in employment
can not only lead to inaccurate estimates about child labor levels, but also, to potentially
incorrect conclusions about the effects of interventions to discourage it. Figure 7documents
the prevalence of children in employment in cocoa fields at end line, separately by the
treatment and control groups, and by reporting source. As the figure shows, while an
intervention evaluated through a randomized control trial using messages to parents to
discourage child labor had no effect on children in employment in cocoa fields according to
children’s self-reports (p=0.390), they significantly increased it according to parents, from
19% to 29.4% – a striking 55.1% increase (p=0.033). In line with our previous discussion,
that could be explained by treatment effects on social expectations: if the intervention
signaled willingness to support farmers rather than to punish them, it could have partially
deterred under-reporting.
16
Figure 7: Share of students who worked in cocoa plantations, by source and treatment
status
Notes: Bars show the share of students who worked in cocoa plantations according to children (in red)
or parents (in blue). At end line, children and parents answered whether the child worked one hour or
more in cocoa plantation over the last month, as described in Table 1. The LHS bar in each set restricts
the sample to the control group, and the RHS bar in each set restricts the sample to parents who received
behavioral nudges via text or audio messages in schools where teachers were not assigned to the intervention
(see Lichand and Wolf,2022). P-values computed controlling for students’ baseline characteristics, grade
fixed-effects, and re-weighting observations based on the predicted probability of successfully tracking each
student at end line (also estimated based on student and household characteristics at baseline). Baseline
characteristics include the child’s gender, baseline standardized test scores (for numeracy and literacy), an
index of baseline parental engagement, an index of student effort, an index of baseline child labor, an index
of the child’s socio-emotional skills, measures of child’s executive functions (working memory and visual
attention), and measures of child’s impulsivity, self-esteem, and mindset; see Lichand and Wolf (2022) for
details on each of those measures.
Figure 13 in the Supplementary Materials also shows that, depending on reporting
sources, the correlation between child labor and educational outcomes at end line also
differs across treatment cells.
17
6 Discussion
Within the social sciences, an active literature discusses the potential merits and drawbacks
of child labor, from both the household and the societal perspectives (e.g., Basu and
Van,1998;Edmonds and Theoharides,2021). On the one hand, child labor might be
critical for families’ livelihoods, especially in rural settings where production is labor-
intensive, and challenges in enforcing land property rights might make it hard to hire
external workers. On the other hand, child labor might detract from children’s future adult
productivity, by competing with effort at school, or even leading children to drop out to
help supporting their families’ livelihoods (Shah and Steinberg,2017). The pandemic made
this discussion more alive than ever, with dropout risk skyrocketing among middle- and
high-school students in developing countries (Lichand et al.,2021), among other reasons
due to the economic crunch and subsequent surge in demand for child labor (World Bank,
2017).
The 1989 Convention on the Rights of the Child establishes that every child has the
right to universal education and protection against economic exploitation (UN General
Assembly,1989). As such, data on the prevalence of child labor across space and over
time is critical for governments and international organizations committed to ensuring
children’s rights. In the cocoa industry, particularly intensive in child labor given its low
rate of mechanization, data issues have posed important challenges to monitoring and
enforcement by international organizations and policymakers over the years. On the one
hand, chocolate companies – which committed to ending “at least the worst forms” of child
labor since the early 2000s (https://www.cocoainitiative.org/knowledge-hub/resour
ces/harkin-engel-protocol) – claim to have made progress. On the other hand, human
rights activists claim that these data are plagued by quality issues, from strategic timing
(asking farmers about child labor outside of the harvest season, which is when children
tend to work on the cocoa fields) to social desirability biases.
Recent certification efforts try to monitor farmers in more targeted ways, in an attempt
to limit under-reporting. ENVERITAS, in particular, monitors cocoa harvests with the
help of satellite imagery and surveys farmers only around this period. While clearly su-
perior, the challenge with third-party certification data such as those from ENVERITAS
is that it involves costly surveillance, limiting stakeholders’ ability to learn timely about
how child labor evolves across space, particularly in response to interventions that try
to mitigate it. Similarly, technologies like GPS trackers (Dillon et al.,2017) have shown
promise insofar as providing objective measures of child labor, but have only been piloted
at relatively small scale due to their costs and implementation challenges.
Our main contributions are two-fold. First, we document first-hand that adult surveys
indeed under-report the prevalence of child labor, and the extent of under-reporting. We
achieved that breakthrough thanks to the simultaneous availability of parents’ reports of
children in employment and an objective measure of the extent to which children in fact
work in cocoa fields – from costly certification surveys triggered by satellite imagery. Other
papers rely solely on comparisons (1) across different parents’ reports (under different con-
ditions for social desirability bias; Jouvin,2021), or (2) between adults’ and children’s re-
ports (without objective measurement; Dillon,2010;Galdo et al.,2020), or (3) between an
objective measure of children in employment and children’s self-reports (without parent’s
reports of children in employment; Dillon et al.,2017). Concretely, the latter compares
data from GPS trackers (an objective indication of whether children work in cocoa fields
in Côte d’Ivoire) with data from surveys and activity diaries. The study also documents
that children accurately report the number of hours they work. But it actively refrained
18
from having enumerators ask adults about child labor, whenever possible (“[a]lthough enu-
merators were instructed to ask questions about labor done to each household member,
sometimes not everyone was available for interviewing. Therefore, in a few cases, we had
the household head or his spouse responding on behalf of other members, which may lead
to bias”, p.53). The absence of parents’ reports of children in employment prevents them
from (1) documenting whether parents under-report children in employment and, if so, by
how much, and (2) estimating the relationship between child labor as reported in official
statistics and the ground truth in order to implement the bias-adjustment procedure.
Our second contribution is precisely to predict the bias-adjusted prevalence of child
labor for each of the 97 countries tracked in the World Development Indicators. While
the child labor figures in the context of our study closely match the statistics based on
surveys with children and adults in Ghana and Brazil – two of the few contexts where both
are available –, statistics on children in employment based on children’s self-reports are
available almost nowhere else. Our findings are extremely consequential, as they indicate
that official statistics under-estimate the global prevalence of child labor by nearly 2/3. For
some countries, the extent of under-reporting is tragic: in Brazil, child labor is likely almost
7-fold that in official statistics; in India, over 20-fold. Our estimates can also be used to
update bias-adjusted country-level and global figures as new data becomes available over
time.
Our finding that ILO statistics not only misrepresent the prevalence of child labor,
but also mischaracterize its trends (especially where interventions have been put in place),
raises critical concerns. While the general sentiment of the literature on child labor is that
substantial progress has been achieved in recent decades (“[a]n important lesson from all
the literature reviewed herein is that child labor can change dramatically and quickly in
countries as a result of changes in the economic and policy environment.”; Edmonds and
Theoharides,2021, pp. 27-28) –, our results call that sentiment into question.
In effect, while Save the Children reports (based on ILO data) a 38% decrease in child
labor worldwide since 2000 (National Geographic,2015), NORC data, based on surveys
with children in cocoa-producing countries, record a nearly 65% increase in child labor
since 2008-09 (Sadhu et al.,2020). The concerning increasing trend in the latter is more
likely to reflect the ground truth in face of our findings. As such, despite official statistics,
advocacy bodies, philanthropic organizations and citizens at large should revisit recent
enthusiasm when it comes to Sustainable Development Goal 8 (which includes child labor
in target 8.7). More fundamentally, governments and international organizations cannot
ensure children’s rights if child labor cannot be accurately tracked, and interventions cannot
be properly evaluated even when appropriate research designs are in place.
All in all, governments and international organizations should urgently revise their
methodology, ensuring that children are asked independently of adults not only about work
conditions to assess hazardous child labor, but also about the number of hours worked,
and that these self-reports are used to compute child labor indicators instead. While there
might be technical challenges in interviewing under-12 year-old children, NORC and Tulane
University have done this successfully in Côte d’Ivoire and Ghana since 2008-09 – a model
that could be replicated by other agencies moving forward.
A limitation of our findings is that we extrapolate from under-reporting of child labor
within the cocoa industry in Côte d’Ivoire to worldwide predictions of child labor across
all industries. While half of the world’s cocoa is sourced in the country – making it a very
relevant benchmark –, if under-reporting is less important in other industries or regions
(e.g., because of variations in stigma associated with child labor; Dillon et al.,2012), then
our estimates will overstate the global prevalence of child labor. The fact that we allow
19
predictions to vary with the degree of urbanization and that of children in employment
according to adults should, however, at least partly alleviate these concerns, since exposure
to different economic activities is a lot higher in urban areas. Although we are quite limited
in the range of variables that we can use to generate conditional predictions (since each
variable must be available both in our data and in the World Development Indicators), the
ones that we include are highly correlated with other variables that presumably affect both
the extent of child labor and the degree of under-reporting, such as income – which we
unfortunately only measured in the month prior to each survey, and without the required
additional information to input average annual per capita income for each classroom in
our data.
An additional issue is that, if remote sensing through satellite imagery is imperfect (e.g.,
if the geographical coverage of ENVERITAS data is selective either when it comes to which
regions it covers or which farms it surveys within each region), and if regions/households
not covered differ systematically in the extent to which they under-report child labor, then
our estimates will be biased. While ENVERITAS relies on partnerships to access fresh
acquisitions of 50cm-resolution satellite data filtered for quality (with alleged maximum
cloud covers of 15%, “crucial for finding (...) farms in cloudy equatorial regions”; Enveritas,
2020), the extent to which it still leads to bias remains an open question for future research.
Another limitation is that our sample consists entirely of school children – all of them
were enrolled in school at the time of our surveys. This excludes many children who had
dropped out of school (or never enrolled in the first place), who are much more likely to
engage in child employment. It could be that, in their case, parents would be less sensitive
to social desirability biases, since their children are not enrolled in school to start with.
Having said that, there are many other potential sources of social pressure to under-report
children in employment – particularly economic pressures linked to restrictions to child
labor in global supply chains. While the extent to which our sample restrictions funda-
mentally challenge our conclusions is unclear, investigating how parents’ under-reporting
changes with children’s school enrollment remains an open question for future work.
Last, our analyses do not incorporate other sources of under-reporting that might
also affect children, from questionnaire design (e.g., whether children are asked inside our
outside the household; see Guarcello et al.,2010 and Dillon,2010 for a broader discussion
of different framing issues) to variations in the understanding of what exactly characterizes
‘work’ by children across different countries. As these additional sources of bias most likely
magnify the extent of under-reporting, our estimate of the global prevalence of child labor
can be thought of as a lower bound.
20
References
Basu, K. and Van, P. H. (1998). The economics of child labor, American economic review
pp. 412–427.
Christian Science Monitor (2019). Child labor in decline: Life gets better for 94 million
children. Last accessed February 3, 2022.
URL: https://www.csmonitor.com/World/Points-of-Progress/2019/0828/Child-labor-
in-decline-Life-gets-better-for-94-million-children
Dillon, A. (2010). Measuring child labor: Comparisons between hours data and subjective
measures, Research in Labor Economics (31): 135–169.
Dillon, A., Bardasi, E., Beegle, K. and Serneels, P. (2012). Explaining variation in child
labor statistics, Journal of Development Economics (98): 136–147.
Dillon, A., Hanicotte, H. and Valenza, M. (2017). Pilot data analysis & methodological
insights on labor measurement, Technical report, Innovations for Poverty Action.
Edmonds, E. V. and Theoharides, C. (2021). Child labor and economic development,
Handbook of Labor, Human Resources and Population Economics, Springer, pp. 1–29.
Enveritas (2020). Supply chain verification: Finding small-holder coffee farms. Last ac-
cessed 8 February 2020.
URL: https://storymaps.arcgis.com/stories/10ef0006e4d94b86b228748c7082cf85/print
Finch, J. E., Wolf, S. and Lichand, G. (2022). Efs and motivation uniquely predict chil-
dren’s academic development in civ.
Galdo, J., Dammer, A. and Abebaw, D. (2020). Gender bias in agricultural child la-
bor: Evidence from survey design experiments, Technical Report 13826, IZA Discussion
Paper.
Guarcello, K., Kovrova, I., Lyon, S., Manacorda, M. and Rosatti, F. (2010). Towards con-
sistency in child labourmeasurement: assessing the comparability of estimates generated
by different survey instruments, Technical Report 54, UCW Working Paper.
ILO and UNICEF (2021). Child labour: Global estimates 2020, trends and the road
forward.
International Labour Organization (2017). Methodology of the global estimates of child
labour, 2012-2016.
Jouvin, M. (2021). Addressing social desirability bias in child labor measurement : an ap-
plication to cocoa farms in côte d’ivoire, Technical Report 2021-08, Bordeaux Economics
Working Paper.
Kling, J., Liebman, J. and Katz, L. (2007). Experimental analysis of neighborhood effects,
Econometrica 75(1): 83–119.
Lichand, G., Doria, C., Leal, O. and Cossi, J. (2021). The impacts of remote learning in
secondary education during the pandemic in brazil.
Lichand, G. and Wolf, S. (2022). Arm-wrestling in the classroom: the non-monotonic
effects of monitoring teachers, University of Zurich, Department of Economics, Working
Paper (357).
21
National Geographic (2015). Children Behind Child Labor. Last accessed February 18,
2022.
URL: http:///www.nationalgeographic.com/culture/article/2015-09-26-data-points-
children-behind-child-labor
Sadhu, S., Kysia, K., Onyango, L., Zinnes, C., Lord, S., Monnard, A. and Arellano, I.
(2020). Norc final report: Assessing progress in reducing child labor in cocoa production
in cocoa growing areas of côte d’ivoire and ghana, Technical report.
Shah, M. and Steinberg, B. M. (2017). Drought of opportunities: Contemporaneous and
long-term impacts of rainfall shocks on human capital, Journal of Political Economy
125(2): 527–561.
Tran, H., Nathan, S., Ilmma, A., Burkiewicz, M. and Wisana, I. (2021). Identifying limiting
factors for feasible productivity improvement for smallholder farmers in coffee sector in
indonesia, Asian Journal of Agriculture 5(2): 53–60.
TRECC (2021). Transforming education in côte d’ivoire – reflections, lessons learned, next
steps. Last accessed March 4, 2022.
URL: https://www.cocoainitiative.org/sites/default/files/resources/TRECC-5-YEAR-
REPORT-4.pdf
UN General Assembly (1989). Convention on the rights of the child, United Nations, Treaty
Series 1577(3): 1–23.
Wolf, S. and Lichand, G. (2022). Nudging parents and teachers to improve learning and
reduce child labor in côte d’ivoire.
World Bank (2017). World development report 2018: Learning to realize education’s
promise, The World Bank.
Yoroba, F., Kouassi, B. K., Diawara, A., Yapo, L. A., Kouadio, K., Tiemoko, D. T., Koua-
dio, Y. K., Kone, I. D. and Assamoi, P. (2019). Evaluation of rainfall and temperature
conditions for a perennial crop in tropical wetland: a case study of cocoa in côte d’ivoire.
22
A Supp. Materials – Survey instruments
Table 1: Questions asked to participants during surveys
Questions Timeline Respondent
“In the last month, have you engaged in one or more of the following activities, for one hour or
more? Work in a cocoa plantation”
Baseline Children
“I will now ask you some questions about activities that your children might have recently
performed. In the last month, has any of your children engaged in one or more of the following
activities, for one hour or more? Work in a cocoa plantation”
Baseline Parents
“In the last month, have you engaged in one or more of the following activities, for one hour or
more? Work in a cocoa plantation”
End line Children
“I will now ask you some questions about activities that your children might have recently
performed. In the last month, has any of your children engaged in one or more of the following
activities, for one hour or more? Work in a cocoa plantation”
End line Parents
“Did (child name) work in cocoa fields more than 10 hours a week in a typical week last year?” Follow-up Parents
“Did (child name) work in cocoa fields more than 10 hours a week during holidays last year?” Follow-up Parents
“To the best of your knowledge in this list of students, can you point out the students who worked
in cocoa fields over 10 hours a week in a typical week last year?”
Follow-up Teachers
“To the best of your knowledge in this list of students, can you point out the students who worked
in cocoa fields over 10 hours a week in a typical week during vacation last year?”
Follow-up Teachers
“To the best of your knowledge in this list of students, can you point out the students who worked
at all in cocoa fields during the school year last year?”
Follow-up Teachers
Notes:
23
B Tables
Table 2: Descriptive statistics to baseline characteristics
Mean S.D. Obs
Child is a girl 0.50 0.50 2,475
Child age 2,150
Under 5 years old 0.00 0.04
5-11 years old 0.92 0.27
12-14 years old 0.07 0.26
15 years old and above 0.01 0.08
Enrolled in 1st primary cycle (CP2) 0.52 0.50 2,475
Rural household 0.50 0.50 2,471
Household monthly income (in 2015 USD) 2,177
Less than USD 19 0.06 0.24
USD 19-37 0.16 0.37
USD 37-55 0.16 0.36
USD 55-92 0.21 0.41
USD 92-185 0.23 0.42
USD 185-370 0.11 0.31
More than USD 370 0.07 0.26
Notes: CP2 is the second grade for the 1st primary cycle in Côte D’Ivoire education. Rural areas
are defined according to parents’ main occupation: agricultural or plantation activities are defined
as rural. Household monthly income was reported in CFCA and converted to 2015 USD.
24
Table 3: Relationship between children’s and parents’ reports at the classroom-level
(1)
Child labor according to parents 0.463
(0.294)
Urban -0.361***
(0.056)
(Child labor according to parents)2-0.069
(0.439)
(Child labor according to parents)*Urban 0.850
(0.633)
(Child labor according to parents)2*Urban -1.491
(1.270)
Constant 0.477***
(0.041)
Observations 198
Adjusted R20.570
Notes: * p < 0.10, ** p < 0.05, *** p < 0.01.
Estimation based on the following equation: CLc=α+β1C Lp+
β2Urban +β3CLp
2+β4CLp∗U rban +β5CLp
2∗Urban +ε, in
which: CLcis child labor according to children, CLpis child labor
according to parents, and Urban the percentage of students living
in urban areas. Results are for classroom-level averages.
Rural and urban areas are defined according to parents’ main oc-
cupation (agricultural or plantation activities are assigned to the
former). Child labor according to parents corresponds to the an-
swer on the following question “I will now ask you some questions
about activities that your children might have recently performed.
In the last month, has any of your children engaged in one or more
of the following activities, for one hour or more? Work in a cocoa
plantation”, as described in Table 1. Child labor according to chil-
dren stands for the baseline share of students who report to have
worked in cocoa plantations in the last month for one hour or more,
in response to the following question: “In the last month, have you
engaged in one or more of the following activities, for one hour or
more? Work in a cocoa plantation”, as described in Table 1. All
variables correspond to baseline answers.
25
Table 4: Bias-adjusted prevalence of child labor by country
Country Prevalence Children in employment
WDI* Bias-adjusted 95% CI WDI** Bias-adjusted 95% CI
Afghanistan 9.30% 44.25% [40.74%; 47.76%] 986,530 4,693,747 [4,321,554; 5,065,940]
Albania 5.50% 30.39% [27.55%; 33.24%] 17,771 98,207 [89,006; 107,409]
Algeria 7.50% 28.56% [24.73%; 32.40%] 634,267 2,415,586 [2,091,183; 2,739,989]
Angola 30.10% 44.92% [37.59%; 52.25%] 2,845,481 4,246,718 [3,553,766; 4,939,670]
Argentina 5.00% 20.28% [15.97%; 24.59%] 367,548 1,490,687 [1,173,990; 1,807,384]
Armenia 9.90% 33.73% [29.92%; 37.54%] 40,819 139,073 [123,382; 154,764]
Azerbaijan 5.20% 31.96% [29.19%; 34.74%] 80,518 494,936 [452,021; 537,851]
Bangladesh 5.00% 37.67% [34.29%; 41.04%] 1,486,694 11,200,297 [10,197,134; 12,203,461]
Belarus 2.30% 21.52% [18.50%; 24.54%] 24,590 230,114 [197,820; 262,409]
Benin 24.10% 46.68% [42.10%; 51.25%] 765,663 1,482,895 [1,337,477; 1,628,313]
Bolivia 13.90% 34.92% [29.56%; 40.28%] 325,343 817,310 [691,879; 942,740]
Bosnia And Herzegovina 10.60% 38.40% [35.21%; 41.59%] 36,406 131,887 [120,932; 142,842]
Brazil 2.50% 19.15% [15.70%; 22.60%] 738,609 5,658,290 [4,639,251; 6,677,329]
Burkina Faso 50.30% 59.71% [54.14%; 65.28%] 2,918,832 3,464,829 [3,141,741; 3,787,917]
Burundi 31.90% 58.44% [53.84%; 63.04%] 1,061,251 1,944,075 [1,791,090; 2,097,061]
Cambodia 11.50% 46.05% [42.63%; 49.48%] 390,003 1,561,788 [1,445,679; 1,677,896]
Cameroon 62.00% 62.00% [62.00%; 69.42%] 4,371,145 4,371,145 [4,371,145; 4,894,050]
Central African Republic 37.20% 53.34% [48.46%; 58.23%] 507,766 728,119 [661,435; 794,802]
Chad 55.90% 63.13% [57.30%; 68.96%] 2,630,895 2,971,168 [2,696,751; 3,245,585]
Chile 4.50% 21.15% [17.31%; 25.00%] 113,207 532,160 [435,428; 628,892]
Colombia 5.60% 24.34% [20.59%; 28.08%] 424,292 1,843,828 [1,560,046; 2,127,611]
Congo 31.50% 45.20% [37.58%; 52.83%] 458,429 657,876 [546,917; 768,835]
Costa Rica 1.30% 19.98% [16.85%; 23.11%] 9,269 142,448 [120,141; 164,756]
Côte D’Ivoire 36.50% 50.75% [44.80%; 56.71%] 2,488,650 3,460,540 [3,054,542; 3,866,538]
Democratic Republic Of The Congo 41.40% 53.58% [47.68%; 59.48%] 10,427,755 13,495,085 [12,008,345; 14,981,826]
Dominican Republic 13.20% 30.97% [24.54%; 37.41%] 260,514 611,306 [484,340; 738,272]
Continued on next page
26
Table 4 – continued from previous page
Country Prevalence Children in employment
WDI* Bias-adjusted 95% CI WDI** Bias-adjusted 95% CI
Ecuador 5.60% 29.83% [26.92%; 32.74%] 177,283 944,312 [852,158; 1,036,467]
Egypt 2.90% 34.56% [31.14%; 37.99%] 638,461 7,609,675 [6,855,500; 8,363,850]
El Salvador 7.80% 28.92% [25.02%; 32.83%] 89,600 332,263 [287,370; 377,156]
Eswatini 13.30% 47.08% [43.73%; 50.43%] 38,623 136,724 [127,000; 146,447]
Ethiopia 26.10% 54.07% [50.25%; 57.89%] 7,595,033 15,734,174 [14,621,871; 16,846,476]
Gabon 24.00% 36.49% [26.58%; 46.40%] 122,254 185,891 [135,405; 236,377]
Gambia 23.80% 43.08% [36.93%; 49.23%] 155,206 280,925 [240,815; 321,034]
Georgia 31.80% 47.34% [40.86%; 53.81%] 153,590 228,628 [197,365; 259,891]
Ghana 28.80% 46.67% [40.74%; 52.61%] 2,122,198 3,439,275 [3,001,906; 3,876,644]
Guatemala 7.20% 35.03% [32.12%; 37.94%] 256,030 1,245,536 [1,142,060; 1,349,013]
Guinea 38.10% 54.96% [50.49%; 59.44%] 1,353,895 1,953,067 [1,794,090; 2,112,044]
Guinea-Bissau 63.90% 63.90% [63.90%; 70.70%] 332,361 332,361 [332,361; 367,741]
Haiti 37.80% 49.75% [42.83%; 56.68%] 922,262 1,213,925 [1,045,063; 1,382,786]
Honduras 10.50% 35.64% [32.01%; 39.26%] 211,327 717,216 [644,301; 790,130]
India 1.70% 36.34% [32.10%; 40.58%] 4,150,339 88,717,408 [78,363,536; 99,071,272]
Indonesia 3.70% 30.59% [27.85%; 33.33%] 1,749,475 14,464,227 [13,169,662; 15,758,793]
Iraq 6.40% 28.43% [25.08%; 31.79%] 626,522 2,783,339 [2,454,825; 3,111,854]
Jamaica 6.20% 32.83% [29.98%; 35.68%] 28,562 151,237 [138,097; 164,377]
Jordan 1.20% 16.13% [12.35%; 19.91%] 27,529 369,988 [283,274; 456,701]
Kazakhstan 3.60% 30.16% [27.44%; 32.88%] 127,650 1,069,436 [972,975; 1,165,896]
Kenya 34.40% 55.93% [52.01%; 59.85%] 4,714,910 7,665,466 [7,127,975; 8,202,957]
Kyrgyzstan 41.10% 55.83% [51.05%; 60.60%] 572,145 777,132 [710,605; 843,660]
Lao Peoples Democratic Republic 8.60% 40.76% [37.64%; 43.87%] 131,351 622,479 [574,851; 670,107]
Lesotho 2.60% 39.01% [34.52%; 43.49%] 11,352 170,309 [150,719; 189,899]
Liberia 18.40% 42.67% [38.28%; 47.06%] 239,549 555,502 [498,353; 612,651]
Madagascar 26.00% 49.97% [46.06%; 53.87%] 1,816,029 3,490,005 [3,217,130; 3,762,880]
Malawi 48.90% 63.41% [58.80%; 68.02%] 2,591,877 3,360,969 [3,116,711; 3,605,227]
Continued on next page
27
Table 4 – continued from previous page
Country Prevalence Children in employment
WDI* Bias-adjusted 95% CI WDI** Bias-adjusted 95% CI
Mali 29.70% 50.27% [45.77%; 54.78%] 1,756,271 2,972,830 [2,706,471; 3,239,190]
Mauritania 14.50% 39.35% [35.22%; 43.48%] 167,507 454,571 [406,895; 502,248]
Mexico 5.60% 24.56% [20.86%; 28.26%] 1,251,663 5,488,671 [4,661,519; 6,315,823]
Moldova 29.00% 50.24% [45.87%; 54.61%] 61,872 107,188 [97,858; 116,518]
Mongolia 14.70% 35.90% [30.50%; 41.30%] 94,715 231,340 [196,548; 266,132]
Morocco 4.50% 29.04% [26.30%; 31.79%] 294,985 1,903,643 [1,723,704; 2,083,583]
Mozambique 27.40% 50.94% [47.05%; 54.84%] 2,360,515 4,388,726 [4,053,136; 4,724,317]
Nepal 42.80% 60.67% [56.41%; 64.93%] 2,434,135 3,450,397 [3,208,164; 3,692,631]
Nicaragua 47.70% 50.78% [47.70%; 61.08%] 618,806 658,806 [618,806; 792,349]
Niger 48.50% 63.53% [58.92%; 68.15%] 3,509,901 4,597,901 [4,263,632; 4,932,169]
Nigeria 35.10% 50.27% [44.45%; 56.09%] 19,552,870 28,003,914 [24,763,112; 31,244,716]
North Macedonia 19.80% 41.88% [36.67%; 47.08%] 44,755 94,657 [82,887; 106,426]
Pakistan 13.00% 43.33% [40.24%; 46.42%] 6,363,597 21,210,028 [19,697,984; 22,722,072]
Panama 5.60% 28.48% [25.43%; 31.53%] 42,235 214,768 [191,754; 237,783]
Paraguay 10.40% 34.45% [30.61%; 38.30%] 141,423 468,524 [416,227; 520,821]
Peru 22.60% 38.59% [30.54%; 46.64%] 1,199,711 2,048,397 [1,621,021; 2,475,773]
Philippines 9.00% 37.72% [34.72%; 40.73%] 2,007,458 8,413,881 [7,743,585; 9,084,178]
Portugal 3.60% 27.29% [24.59%; 29.99%] 33,994 257,674 [232,180; 283,167]
Romania 1.40% 29.38% [26.32%; 32.44%] 28,746 603,284 [540,380; 666,188]
Rwanda 5.90% 44.88% [40.24%; 49.51%] 190,439 1,448,471 [1,298,973; 1,597,969]
Senegal 25.10% 47.23% [42.61%; 51.85%] 1,133,642 2,133,227 [1,924,558; 2,341,896]
Serbia 17.90% 41.25% [36.51%; 45.99%] 115,024 265,071 [234,604; 295,538]
Sierra Leone 59.20% 59.20% [59.20%; 68.18%] 1,219,049 1,219,049 [1,219,049; 1,404,055]
Somalia 43.50% 53.89% [47.47%; 60.31%] 1,960,852 2,429,232 [2,139,731; 2,718,732]
South Sudan 45.60% 61.63% [57.27%; 65.99%] 1,331,118 1,799,117 [1,671,804; 1,926,430]
Sri Lanka 10.70% 47.18% [43.37%; 50.99%] 377,869 1,666,127 [1,531,428; 1,800,826]
Sudan 30.60% 52.72% [48.77%; 56.67%] 3,400,496 5,858,325 [5,419,412; 6,297,238]
Continued on next page
28
Table 4 – continued from previous page
Country Prevalence Children in employment
WDI* Bias-adjusted 95% CI WDI** Bias-adjusted 95% CI
Syrian Arab Republic 6.60% 33.42% [30.53%; 36.31%] 228,807 1,158,599 [1,058,570; 1,258,629]
Tajikistan 8.90% 43.57% [40.10%; 47.04%] 195,594 957,472 [881,208; 1,033,737]
Tanzania 34.70% 54.26% [50.12%; 58.40%] 5,648,430 8,832,135 [8,158,766; 9,505,505]
Thailand 15.10% 40.79% [36.88%; 44.70%] 1,201,603 3,245,957 [2,934,714; 3,557,200]
Timor-Leste 19.90% 48.76% [45.42%; 52.10%] 61,214 149,982 [139,705; 160,259]
Togo 35.20% 52.56% [47.80%; 57.33%] 754,750 1,127,033 [1,024,887; 1,229,179]
Trinidad And Tobago 3.40% 31.47% [28.63%; 34.32%] 6,554 60,667 [55,182; 66,153]
Tunisia 3.40% 26.01% [23.27%; 28.75%] 63,472 485,542 [434,349; 536,734]
Turkey 2.60% 23.00% [20.11%; 25.88%] 354,263 3,133,237 [2,739,836; 3,526,638]
Uganda 36.70% 57.50% [53.49%; 61.52%] 4,863,530 7,620,622 [7,089,122; 8,152,122]
Ukraine 5.00% 27.53% [24.56%; 30.51%] 247,168 1,361,123 [1,214,132; 1,508,114]
Uruguay 7.30% 21.69% [16.26%; 27.13%] 34,265 101,814 [76,299; 127,328]
Uzbekistan 5.10% 33.80% [30.95%; 36.66%] 327,769 2,172,478 [1,989,018; 2,355,938]
Venezuela 3.90% 20.32% [16.58%; 24.05%] 210,187 1,094,873 [893,632; 1,296,115]
Vietnam 10.90% 41.96% [38.91%; 45.00%] 1,600,637 6,161,152 [5,714,055; 6,608,249]
Yemen 16.10% 44.99% [41.72%; 48.25%] 1,202,201 3,359,186 [3,115,605; 3,602,766]
Zambia 34.40% 51.85% [46.96%; 56.73%] 1,770,271 2,668,075 [2,416,708; 2,919,443]
World 135,832,023 373,513,346 [335,964,180; 411,899,341]
Notes: * using the latest available WDI figures, following Edmonds and Theoharides (2021); ** using WDI and UNICEF population figures for 0-14 and under-5
children for 2020. Bias-adjusted prevalence of child labor based on Table 3’s estimates for country-specific adjustment factors. Countries with missing data were
excluded from the table. They are: Andorra, Antigua And Barbuda, Australia, Austria, Bahamas, Bahrain, Barbados, Belgium, Belize, Bhutan, Botswana, Brunei
Darussalam, Bulgaria, Cabo Verde, Canada, China, Comoros, Croatia, Cuba, Cyprus, Czech Republic, Democratic Peoples Republic Of Korea, Denmark, Djibouti,
Dominica, Equatorial Guinea, Eritrea, Estonia, Fiji, Finland, France, Germany, Greece, Grenada, Guyana, Hungary, Iceland, Iran, Ireland, Israel, Italy, Japan, Kiribati,
Kuwait, Latvia, Lebanon, Libya, Liechtenstein, Lithuania, Luxembourg, Malaysia, Maldives, Malta, Marshall Islands, Mauritius, Micronesia, Monaco, Montenegro,
Myanmar, Namibia, Netherlands, New Zealand, Norway, Oman, Palau, Papua New Guinea, Poland, Qatar, Republic Of Korea, Russian Federation, Saint Kitts And
Nevis, Saint Lucia, Saint Vincent And The Grenadines, Samoa, San Marino, Sao Tome And Principe, Saudi Arabia, Seychelles, Singapore, Slovakia, Slovenia, Solomon
Islands, South Africa, Spain, Suriname, Sweden, Switzerland, Tonga, Turkmenistan, Turks and Caicos, Tuvalu, UAE, UK, USA, Vanuatu, and Zimbabwe.
29
D Supp. Materials – Prevalence of children in employment
and hazardous work conditions measured in the follow-up
survey
Figure 10: Parents’ reports on children in employment and hazardous work conditions at
the follow-up wave (October 2019)
Notes: Bars show the share of students who parents have reported to be in the following conditions
(yes/no questions): 1) Help family: “Did (child name) work at all to help you around home, assist in a
family business or earn pocket money outside school hours under adult supervision during the school year
last year?”; 2) Help family 1 hour: “Did (child name) work at all to help you around home, assist in a
family business or earn pocket money outside school hours under adult supervision over 1 hour a week
during the school year last year?”; 3) Cocoa: “Did (child name) work at all in cocoa fields during the school
year last year?”; 4) Cocoa > 10 hrs: “Did (child name) work in cocoa fields more than 10 hours a week in
a typical week last year?”; 5) Cocoa > 10 hrs vacation: “Did (child name) work in cocoa fields more than
10 hours a week during vacation last year?”; 6) Dangerous work: “Was (child name) involved in activities
in cocoa fields such as clearing of forests and felling of trees, bush burning, manipulating agrochemicals
or using sharp tools during the school year last year?”; 7) Night work: “Did (child name) work between 7
p.m. and 7 a.m. during the school year last year?”; 8) Heavy work: “Was (child name) engaged in heavy
physical labor in a typical week last year?”; 9) Machete: “Did (child name) used a machete while woking
in the fields last year?” 10) Hurt: “Did (child name) get hurt at least once while woking in the fields last
year?”. All measures were collected at the follow-up surveys (Lichand and Wolf,2022). Across all bars,
the sample is restricted to the control group.
31
Figure 11: Share of students who worked in cocoa plantations during the school year,
according to different sources
Notes: Bars show the share of students who have worked in cocoa plantations according to children (in
red), parents (in blue), or teachers (in green). Children and parents answer at end line if the child engaged
in one hour or more in cocoa plantation activities over the last month, as described in Table 1. Teachers
answer at follow-up if each of their students worked in cocoa plantation at all over the last school year, as
described in Table 1. Across all bars, the sample is restricted to the control group. The first set of bars
comprises the whole sample in Lichand and Wolf (2022), while the additional ones split the sample by
primary grades (CP2 and CE2). Sample sizes are the following: (i) Total: 625 ; (ii) CP2: 323; (iii) CE2:
302.
32
E Supp. Materials – Social desirability bias
Figure 12: Share of students who worked in cocoa plantations at least 10 hours/week,
during the school year and during the school holidays
Notes: Bars show the share of students who worked in cocoa plantations during different periods and
according to two different sources. Parents (in red) and teachers (in blue) answered whether the child
worked 10 hours/week or more in cocoa fields during a typical week (LHS bars) and whether the child
worked 10 hours/week or more in cocoa fields during school holidays in the previous school year, as
described in Table 1. All measures were collected at the follow-up surveys (Lichand and Wolf,2022).
Across all bars, the sample is restricted to the control group. Sample sizes are the following: (i) Parents:
200; and (ii) Teachers: 2,500.
33
F Supp. Materials – How adult reports can also bias corre-
lations
Figure 13: Correlation between end-line test scores and child labor, by source and treat-
ment status
Panel A: Control group Panel B: Nudges to parents
Panel C: Nudge to parents and teachers
Notes: All panels show bin-scatter plots of end-line standardized test scores as a function of child labor in
cocoa fields reported at end line. Each panel estimates linear relationship between the variables according
to different measures of child labor; that using that reported by children is shown in black, and the one
using that reported by parents is shown in gray. Children answered the following question: “In the last
month, have you engaged in one or more of the following activities, for one hour or more? Work in a cocoa
plantation”, as described in Table 1. Parents answered the following question: “I will now ask you some
questions about activities that your children might have recently performed. In the last month, has any
of your children engaged in one or more of the following activities, for one hour or more? Work in a cocoa
plantation”, as described in Table 1. Standardized test scores are a summary measure of numeracy and
literacy test scores (averaging across each component, normalized by their mean and standard deviation in
the control group), following Kling et al. (2007). Panel A restricts the sample to the control group of the
intervention; Panel B, to parents who received behavioral nudges via text or or audio messages in schools
where teachers were not assigned to the intervention; and Panel C, to parents who received behavioral
nudges via text or audio messages in schools where teachers were also assigned to behavioral nudges via
text messages; see Lichand and Wolf (2022).
34