Child Labor and the Transition Between School and Work [1 ed.] 9780857240019, 9780857240002

Citation preview

CHILD LABOR AND THE TRANSITION BETWEEN SCHOOL AND WORK

RESEARCH IN LABOR ECONOMICS Series Editor: Solomon W. Polachek IZA Co-Editor: Konstantinos Tatsiramos Volume 23: Volume 24:

Volume 25: Volume 26:

Volume 27:

Volume 28:

Volume 29:

Volume 30:

Accounting for Worker Well-Being Edited by Solomon W. Polachek The Economics of Immigration and Social Diversity Edited by Solomon W. Polachek, Carmel Chiswick and Hillel Rapoport Micro-Simulation in Action Edited by Olivier Bargain Aspects of Worker Well-Being Edited by Solomon W. Polachek and Olivier Bargain Immigration: Trends, Consequences and Prospects for The United States Edited by Barry R. Chiswick Work, Earnings and Other Aspects of the Employement Relation Edited by Solomon W. Polachek and Konstantinos Tatsiramos Ethnicity and Labor Market Outcomes Edited by Amelie F. Constant, Konstantinos Tatsiramos and Klaus F. Zimmermann Jobs, Training, and Worker Well-Being Edited by Solomon W. Polachek and Konstantinos Tatsiramos

RESEARCH IN LABOR ECONOMICS VOLUME 31

CHILD LABOR AND THE TRANSITION BETWEEN SCHOOL AND WORK EDITED BY

RANDALL K. Q. AKEE Tufts University, Boston, MA, USA and IZA, Germany

ERIC V. EDMONDS Dartmouth College, Hanover, NH, USA and IZA, Germany

KONSTANTINOS TATSIRAMOS IZA, Germany

United Kingdom – North America – Japan India – Malaysia – China

Emerald Group Publishing Limited Howard House, Wagon Lane, Bingley BD16 1WA, UK First edition 2010 Copyright r 2010 Emerald Group Publishing Limited Reprints and permission service Contact: [email protected] No part of this book may be reproduced, stored in a retrieval system, transmitted in any form or by any means electronic, mechanical, photocopying, recording or otherwise without either the prior written permission of the publisher or a licence permitting restricted copying issued in the UK by The Copyright Licensing Agency and in the USA by The Copyright Clearance Center. No responsibility is accepted for the accuracy of information contained in the text, illustrations or advertisements. The opinions expressed in these chapters are not necessarily those of the Editor or the publisher. British Library Cataloguing in Publication Data A catalogue record for this book is available from the British Library ISBN: 978-0-85724-000-2 ISSN: 0147-9121 (Series)

Awarded in recognition of Emerald’s production department’s adherence to quality systems and processes when preparing scholarly journals for print

CONTENTS LIST OF CONTRIBUTORS

vii

PREFACE

ix

SELECTION INTO WORST FORMS OF CHILD LABOR Eric V. Edmonds

1

HOUSEHOLD POVERTY AND CHILD LABOR DECISIONS IN MALAWI Levison S. Chiwaula

33

HOW MUCH WORK IS TOO MUCH? EFFECTS OF CHILD WORK HOURS ON SCHOOLING – THE CASE OF EGYPT Ragui Assaad, Deborah Levison and Hai-Anh Dang

53

LIFETIME HEALTH CONSEQUENCES OF CHILD LABOR IN BRAZIL Chanyoung Lee and Peter F. Orazem

99

MEASURING CHILD LABOR: COMPARISONS BETWEEN HOURS DATA AND SUBJECTIVE MEASURES Andrew Dillon ALLOCATION OF CHILDREN’S TIME ALONG GENDER LINES: WORK, SCHOOL, AND DOMESTIC WORK IN BRAZIL Diana I. Kruger, Matias Berthelon and Rodrigo R. Soares

v

135

161

vi

CONTENTS

THE IMPACT OF CONDITIONAL CASH TRANSFER PROGRAMS ON HOUSEHOLD WORK DECISIONS IN BRAZIL Andrea R. Ferro, Ana Lu´cia Kassouf and Deborah Levison

193

INTRA-HOUSEHOLD TIME ALLOCATION IN RURAL MEXICO: EVIDENCE FROM A RANDOMIZED EXPERIMENT Marta Rubio-Codina

219

LEVELING THE INTRA-HOUSEHOLD PLAYING FIELD: COMPENSATION AND SPECIALIZATION IN CHILD LABOR ALLOCATION Ximena V. Del Carpio and Karen Macours

259

ADULT RETURNS TO SCHOOLING AND CHILDREN’S SCHOOL ENROLLMENT: THEORY AND EVIDENCE FROM SOUTH AFRICA Sarah Donovan and Kenneth A. Swinnerton

297

LOCAL LABOR DEMAND AND CHILD WORK Marco Manacorda and Furio Camillo Rosati

321

LIST OF CONTRIBUTORS Ragui Assaad

Humphrey Institute of Public Affairs, University of Minnesota, Minneapolis, MN, USA

Matias Berthelon

School of Business Administration, Pontificia Universidad Cato´lica de Valparaı´ so, Valparaı´ so, Chile

Levison S. Chiwaula

Department of Economics, University of Malawi, Zomba, Malawi

Hai-Anh Dang

World Bank, Washington DC, USA

Ximena V. Del Carpio

World Bank, Washington DC, USA

Andrew Dillon

International Food Policy Research Institute, Washington DC, USA

Sarah Donovan

Bureau of International Labor Affairs, U.S. Department of Labor, Washington DC, USA

Eric V. Edmonds

Department of Economics, Dartmouth College, Hanover, NH, USA

Andrea R. Ferro

Universidade Federal de Sa˜o Carlos, Sorocaba, SP, Brazil

Ana Lu´cia Kassouf

Department of Economics, ESALQUniversity of Sa˜o Paulo, Piracicaba, SP, Brazil

Diana I. Kruger

School of Business Administration, Pontificia Universidad Cato´lica de Valparaı´ so, Valparaı´ so, Chile

Chanyoung Lee

Samsung Economic Research Institute, Seoul, Korea vii

viii

LIST OF CONTRIBUTORS

Deborah Levison

Humphrey Institute of Public Affairs, University of Minnesota, Minneapolis, MN, USA

Karen Macours

Johns Hopkins University, Baltimore, MD, USA

Marco Manacorda

Department of Economics, Queen Mary, University of London, London, UK

Peter F. Orazem

Department of Economics, Iowa State University, Ames, IA, USA

Furio Camillo Rosati

Department of Economics, Universita` degli Studi di Roma ‘‘Tor Vergata’’, Rome, Italy

Marta Rubio-Codina

Center for the Evaluation of Development Policies, The Institute for Fiscal Studies, London, UK

Rodrigo R. Soares

Deparment of Economics, Pontifical Catholic University of Rio de Janeiro, Rio de Janeiro, Brazil

Kenneth A. Swinnerton

Bureau of International Labor Affairs, U.S. Department of Labor, Washington DC, USA

PREFACE There are an estimated 190.7 million economically active children in the world today.1 Most of these children are living in poor countries. Sixty-four percent live in Asia where nearly 1 in 5 children work. Sub-Saharan Africa’s population is much smaller, but more than 1 in 4 children are economically active. These statistics do not include the hundreds of millions more that provide unpaid household services to their families. Economists since Adam Smith have been interested in understanding why children work, and with the widespread availability of nationally representative multi-purpose household survey data, empirical research on child labor has proliferated in the past 20 years. Many of the types of activities studied in the economics literature on child labor do not correspond to the horrific images that motivate popular concern about the topic, but the analytical case for distinguishing between work in a family enterprise and images that make better photographs is far from obvious. This volume collects recent advances in the empirical literature that aims to understand why children work and what the consequences of that work are for children. In the 1st chapter, Eric Edmonds investigates the determinants of child labor participation in some of the worst forms of child labor – rag picking and portering in Nepal. The two kinds of employment are considered especially distasteful because of the hazardous conditions associated with the first endeavor and the extremely arduous and physically taxing nature of the latter occupation. Utilizing data from a survey of porters and ragpickers and the national census, the author explores the correlates of employment in these worst forms of child labor. Paternal disability is a strong predictor of having a child employed in a worst form of child labor. The results also support the notion that employment in these worst forms of child labor are observed only when other options do not exist. The presence of employment opportunities within the child’s own household is associated with a diminished risk of entry into worst forms. These findings are consistent with a model where there are negative amenities associated with these jobs that induce the poor and those with the fewest alternative earnings options to select into these worst forms of child labor. ix

x

PREFACE

An important determinant of child labor is household poverty. Levison Chiwaula, in the 2nd chapter, examines the role of household poverty and assets on child labor in Malawi, one of the poorest countries in the world located in Southern Africa with about 1.4 million children between ages 5 and 14 years engaged in some form of employment. Utilizing an extensive household survey, the study provides evidence that child labor emerges from household poverty. The incidence and intensity of child labor is higher in domestic work than in unpaid market work for this sample. The author finds a significant relationship between household poverty and child labor in unpaid market work, but no relationship is found between poverty and domestic work. These findings suggest that household poverty results in child labor activities that can directly affect household income. In activities that do not directly generate income such as domestic work, child labor is related to other factors. The policy implications from these findings are that anti-poverty programs that serve to increase household income may diminish the occurrence of the market work component of child labor in Malawi. When child labor is not the ‘‘worst form,’’ is there some amount of child labor that is not problematic? In the 3rd chapter of this volume, Ragui Assaad, Deborah Levison, and Hai-Ahn Dang investigate this question for Egypt using the Egypt Labor Market Survey (ELMS-1998). The authors find that effect of working on school attendance is actually fairly weak for boys up to 14 hours of work per week. Whereas girls tend to work more than boys, and there is a large negative effect of work on schooling. The authors find, however, there is almost no effect when girls are working less than 10 hours per week. These findings suggest that policies aimed at restricting total hours of labor force participation may not always work as parents may view child labor as a training program itself. Therefore, these apprenticeship-type work activities may be viewed as a favorable means to permanent employment in the future. Any attempts to curb these activities may be met with resistance; the better compromise may be to encourage the firms to structure their employment in shifts so that the children can both participate in the training and attend school. For girls, who are more likely to be working in the household, improvements in basic resources for communities such as access to running water, garbage collection, and cooking fuels will significantly reduce the burden of household chores and work activities. The effects of participation in child labor can manifest itself in numerous ways such as reduced educational attainment and health consequences. Chanyoung Lee and Peter Orazem, in the 4th chapter, examine whether

Preface

xi

children who begin working at young ages experience increased incidence of illness or physical disability as adults. Using the 1998 Brazilian Household Survey (PNAD), the authors find evidence that individuals who participated in child labor in their youth (and were therefore not in school) had higher incidences of physical problems such as back problems, arthritis, and reduction in stamina or strength in adulthood. These findings imply that delaying entry into child labor while increasing time in school significantly lowers the probability of early onset of physical ailments. They also indicate that there may be multiple ways in which child labor could adversely affect individuals over time – reducing human capital through a reduction in schooling as well as reducing an individual’s overall health in the long run. Definitions of child labor differ depending on the country, region, or industries studied. Additionally, even when there are generally agreed definitions, objective measures of child labor are difficult to elicit in surveys. In the 5th chapter, Andrew Dillon examines a subjective measure of child labor as an alternative to time use hourly data for eliciting the distribution of children’s time between work, school, and leisure. The methodological contribution of this chapter is that it provides some evidence for alternative methods for eliciting survey responses that are notoriously difficult to acquire. Precise estimates of the total hours spent in school or work by a child are often underestimated by parents. The author finds that by allowing the child to provide a subjective measure of their time use results in larger marginal effects of child’s age, parental education, and school availability on child labor probabilities than using standard measures. These results draw our attention to the potential issues of bias from existing survey methodologies. Households face a trade-off regarding the decision between child work and schooling. This decision involves the allocation of time and tasks among children and across genders. This allocation depends to a large extent on the financial constraints faced by the households. The next four chapters deal with these issues. In the first of these chapters, Diana Kruger, Matias Berthelon, and Rodrigo Soares explore the household decision to send children of different genders to school and work using the 2001–2003 Brazilian household surveys (PNAD). Their model predicts that girls are less likely than boys to be engaged in market work with similar characteristics. However, when household and domestic chores are considered work, then girls are much more likely to be participating than boys. These results indicate that the household composition and the number of adults in the household can have a significant impact on whether children attend school or are assigned to domestic or market work. This research further indicates

xii

PREFACE

the degree of gender bias present in household decision-making with regard to schooling and work choices for household children. Addressing these differential rates of schooling and working probabilities are important issues for developing country development strategies. Interventions such as conditional cash transfer programs can affect the household allocation of resources and have been used to increase the health and educational outcomes for children in developing countries. These investments in human capital have obvious benefits to the long-run outcomes for children from these affected households in terms of higher levels of education and overall health. In the next chapter, Andrea Ferro, Ana Lu´cia Kassouf, and Deborah Levison investigate the impact of the Brazilian Bolsa Escola conditional cash transfer program on children’s and parents’ labor status. They find that the effect of conditional cash transfers has not only increased children’s school enrollment, but it has also increased both mothers’ and fathers’ labor force participation. The authors find that to maintain eligibility for the program, households adhere to the conditions of the cash transfer program – sending their children to school. However, to make up the short fall in household income, adults must increase their labor force participation. These results are particularly encouraging because the conditional cash transfer program does not deter parents from working and it increases the education of the household children. The conditional cash transfer program does not appear to increase dependency on the government transfers as the parents from treated households actually appear to work more than those in the untreated households. A similar study conducted on the Oportunidades program in Mexico finds similar results for children with regard to school attendance and market work. On average, children from treated households work less and attend school more than those from untreated households. In this chapter, Marta Rubio-Codina finds, however, that there is little difference in the allocation of the adult household members’ time allocation with respect to market work. Importantly, she also concludes that this program does not decrease the incentives for labor force participation for adults from treated households. The author does find some evidence that adult women tend to work more on unpaid work activities after treatment. Presumably, these unpaid activities were formerly conducted by the household children. In the next chapter, Ximena Del Carpio and Karen Macours examine how the conditional cash transfer program Atencio´n a Crisis in Nicaragua differentially affects child labor by gender and age. The authors find that the conditional cash transfer has a differential effect for older boys. Before the program, older boys tended to work more than their younger siblings, and

Preface

xiii

after the program, their labor participation decreases. A secondary program that provides a productive business grant for women tends to increase the specialization of girls in nonagricultural and domestic-type activities; however, there is no change in the total hours of labor for household girls. The results indicate that policies aimed at alleviating household budget constraints through cash transfer programs may have differential effects depending on how the household or society allocates labor resources. Specialization across household activities and labor force participation means that children of different ages and genders in the household will be differentially affected by conditional cash transfers. Decisions to invest in human capital are related to the returns that can be realized in labor markets. In labor markets where the returns to education are low, there is little incentive to invest in education. Therefore, working in the current period may actually have higher returns for children. On the contrary, when there are large returns to education, which is often observed in developing countries, there should be a large incentive to invest in education. In the next chapter, Sarah Donovan and Kenneth Swinnerton investigate how the returns to education in the adult labor market affect children’s school enrollment. They show, in a theoretical model that they develop, that the relationship between labor market returns and educational investment is ambiguous when households face budget or liquidity constraints. The ambiguity results from a negative liquidity effect that works against a parent’s incentive to substitute more of a child’s time away from work and into schooling when there is an improvement in the returns to education. Using the South Africa Integrated Household Survey (SAIHS) around the end of Apartheid, the authors find evidence for liquidity constraints in their sample of the African South African households. Despite the presence of liquidity constraints, they also find a significant correlation between a South African region’s adult rates of return to education and its enrollment rate of 8- to 15-year-old children. In addition to differences in labor market returns, changes in labor demand can also affect household decisions about children’s time use. In the last chapter, Marco Manacorda and Furio Rosati examine how local labor demand affects the working and schooling decisions for children aged 10–15 years in Brazil. Using the Brazilian household surveys (PNAD) between 1981 and 2002, they find that schooling is essentially unchanged no matter the local labor market conditions. On the contrary, they do find evidence that child work is pro-cyclical – children do end up working more when local labor market demand increases. As there is little evidence for changes in school enrollment, this suggests that the additional work time crowds out

xiv

PREFACE

leisure or other nonschool activities. Additionally, the authors find that having older brothers, who may work, tends to reduce employment levels for younger urban boys. This finding suggests that the older brothers are able to compensate for their younger brothers’ school and leisure time. As with past volumes, we aimed to focus on important issues and to maintain the highest levels of scholarship. We encourage readers who have prepared manuscripts that meet these stringent standards to submit them to Research in Labor Economics (RLE) through the IZA website (http:// www.iza.org/rle) for possible inclusion in future volumes. We thank all the referees for insightful editorial advice in preparing this volume.

NOTE 1. These estimates come from the Understanding Children’s Work Project (http:// ucw-project.org/) at the time of writing, February 2010.

Randall K. Q. Akee Eric V. Edmonds Konstantinos Tatsiramos Editors

SELECTION INTO WORST FORMS OF CHILD LABOR Eric V. Edmonds ABSTRACT Little is known about why children participate in activities that are labeled worst forms of child labor (WFCL). Case–control approaches common in medicine are adapted to consider the correlates of participation in worst forms in the context of two WFCL in Nepal: portering and ragpicking. Paternal disability is a strong predictor of entry into each of the worst forms, and the presence of productive assets within the child’s home reduces the risk a child is observed in a worst form. We argue that our findings are consistent with a model where there are negative amenities associated with these jobs that induce the poor and those with the fewest alternative earnings options to select into these WFCL in Nepal.

1. INTRODUCTION Popular horror of the prevalence and persistence of worst forms of child labor (WFCL) in developing countries is nearly universal. One hundred and sixty countries have ratified International Labour Organization (ILO) Convention 182 ‘‘Concerning the Prohibition and Immediate Action for the Elimination of the Worst Forms of Child Labor.’’ Commensurate with this

Child Labor and the Transition Between School and Work Research in Labor Economics, Volume 31, 1–31 Copyright r 2010 by Emerald Group Publishing Limited All rights of reproduction in any form reserved ISSN: 0147-9121/doi:10.1108/S0147-9121(2010)0000031004

1

2

ERIC V. EDMONDS

public attention to worst forms is a literature within economics seeking to understand why children are engaged in WFCL. Policy tends to view worst forms as evidence of victimization. Children are often not free to choose their own time allocation, and one argument for the persistence and prevalence of worst forms is that they reflect parental neglect, indifference to the child’s welfare, or coercion. Empirical evidence on the determinants of selection into the WFCL is scarce (Edmonds, 2007, is a review), because worst forms are difficult to capture with randomized sampling. Most of our understanding comes from research that interviews children engaged in a specific activity about their working conditions and why they participate in the work. Children often respond that they are working because either they or their family need the money. However, the fact that children work in worst forms for income does not itself answer the question of why children are working in these activities. Children work in plenty of activities for income, many of which would not be considered hazardous or a worst form of child labor. More generally, it is impossible to understand why children are involved in some activity without talking to children who are not involved in that activity. To design policy aimed at preventing child involvement in WFCL, policy needs to know what factors are associated with entry to WFCL and whether these correlates of entry differ from correlates of entry into other types of work. This study argues for analyzing the correlates of participation in worst forms by pooling nationally representative data and survey data from children in worst forms. The present study applies the simplest of approaches from the statistical literature on inference in contaminated samples (where sampling probabilities are correlated with treatment status) to consider the correlates of participation in WFCL. This is not a causal approach, and inference is limited to child background characteristics that are collected in the survey of children in a worst form and the nationally representative data. This study examines what child background characteristics make it more likely that a child is observed as a short route porter or ragpicker in Nepal. Both types of work have been defined as a worst form of child labor in Nepal by the Nepali government and explicitly targeted for eradication. Survey data from children engaged in each activity are combined with estimates of the incidence of each in the population and with nationally representative data from Nepal’s population census to show how this combination of data can be used to infer the correlates of selection into these WFCL in Nepal. Each activity is considered separately, and findings for children observed in worst forms are compared to results from analyzing selection into regular wage work.

Selection into Worst Forms of Child Labor

3

Some striking patterns are observable in the data. Paternal disability appears to be strongly correlated with child participation in the worst forms considered herein. The presence of employment opportunities within the child’s own household is associated with a diminished risk of entry into worst forms. Such patterns appear in the worst forms examined herein, but these correlates of entry into worst forms do not predict positively participation in wage work. Given that children rarely work for wages in Nepal, the association between participation and household employment opportunities suggests that children are more likely to be observed in a worst form when the return to the worst form is large relative to other options available to the child. These findings are consistent with Dessy and Pallage’s (2005) model of compensating wage differentials in WFCL. The negative amenities associated with the WFCL are compensated, so that there is sorting into worst forms along the marginal utility of income. Section 2 of the study discusses the concept of a worst form of child labor and reviews the existing theoretical literature on why children participate in worst forms. Section 3 describes the present methods for studying entry into worst forms. Section 4 details the data used in this analysis, and Section 5 presents the findings. Section 6 concludes with a discussion of the lessons of the empirical findings and considers the implications of this study’s weaknesses for future studies of entry into WFCL.

2. THEORY – ARE WORST FORMS DIFFERENT? ILO Convention C182 on the Worst Forms of Child Labor asks signatory countries to clarify the definition of WFCL in the signatory’s country and to develop specific plans for their eradication. Article 3 of C182 contains several guidelines for what types of activities are to be considered for persons under the age of 18. These include all forms of slavery and ‘‘practices similar to slavery.’’ This later clause is noted to include the sale and trafficking of children, debt bondage, serfdom, and forced or compulsory labor including for the purposes of armed conflict. Children in prostitution, pornography, the production or processing of drugs are also noted as being in ‘‘worst forms’’ of child labor. Article 3 (d) is the most ambiguous part of the convention as it allows worst forms to include ‘‘work which, by its nature or the circumstances in which it is carried out, is likely to harm the health, safety, or morals of children.’’ Article 4 of the convention is explicit that it is up to individual countries to define what types of work are considered ‘‘worst forms’’ of child labor under this clause. Activities labeled ‘‘worst forms’’

4

ERIC V. EDMONDS

under Article 3(d) of C182 are often labeled as ‘‘Hazardous forms of child labor.’’ The companion recommendation document for C182, R190 Worst Forms of Child Labor Recommendation, suggests that these hazardous forms of child labor include: (a) work which exposes children to physical, psychological, or sexual abuse; (b) work underground, under water, at dangerous heights, or in confined spaces; (c) work with dangerous machinery, equipment and tools, or which involves the handling or transport of heavy loads; (d) work in an unhealthy environment which may, for example, expose children to hazardous substances, agents or processes, or to temperature, noise levels, or vibrations damaging to their health; (e) work under particularly difficult conditions such as work for long hours or during the night or work where the child is unreasonably confined to the premises or the employer. (R190, Section II.3.a–e)

The ILO’s Statistical Information Monitoring Programme on Child Labour (SIMPOC) estimates that a total of 8.4 million children are involved in child trafficking, in forced or bonded labor, are soldiers, are prostitutes or involved in pornography, or participate in illicit activities (ILO, 2002). Sixtyeight percent of these children are in bonded or forced labor. The same SIMPOC study calculates that 170.5 million children are engaged in activities that have been labeled hazardous in their home country. Altogether then, SIMPOC estimates that 178.9 million children are engaged in WFCL. A simple analytical model based on Dessy and Pallage (2005) and Rogers and Swinnerton (2008) will help fix ideas in our discussion of the determinants of entry into worst forms. The first issue that any framework of child time allocation must address is the question of who is making decisions. Agency in work decisions is important when discussing WFCL. This study does not inform about agency issues; hence, the present discussion of why children participate in worst forms is framed around an agent making an informed decision about job type without clarifying who the relevant agent might be. A child participates in a worst form of child labor when the decision-making agent’s utility is higher than when the child does not:     (1) u y c ; c þ ec  u y 0 ; 0 þ e 0 c is an indicator for whether the child participates in the given worst form, yc is the agent’s income when the child participates in the worst form, and y0 is the agent’s income when the child does not. ec and e0 are stochastic, mean zero, error terms that reflect some randomness in the agent’s decisions. Let the decision-maker’s utility when the child does not participate in the worst form be represented by an indirect utility function:     u y0 ; 0 ¼ v y 0 ; p

Selection into Worst Forms of Child Labor

5

The agent’s relevant income when the child participates in a worst form is the agent’s income absent the child’s participation plus the net economic gain from having the child in the worst form: y c ¼ y 0 þ w

(2)

An alternative to participating in a worst form (reflected in y0) is participation in other types of work. w is the premium the worst form pays above those other forms of work. For analytical clarity, treat the disutility from having the child involved in a worst form as additively separable from the utility owing to the other decisions the agent makes with their income:       (3) u y c ; c ¼ u y 0 þ w ; c ¼ v y 0 þ w  ; p  t The disutility of participation in a worst form t is known with certainty. Rogers and Swinnerton (2008) emphasize that uncertainty and poor information may be important in explaining why some children end up in worst forms. It is possible to conceptualize t as a parameter that reflects the risk perceived by the agent and thereby interpret this set-up within their model. The functional form assumption on preferences in Eq. (3) is equivalent to assuming that the disutility that the agent gets from having a child in a worst form is independent from their income (or price) level. Poor families and rich families are made equally worse off by having a child pick through garbage. This does not imply that the marginal utility from having a child in a worst form, relative to not, will be the same for poor and rich families as the marginal utility associated with the child’s net economic contribution in the worst form will differ between poor and rich. The incidence of children involved in worst forms is then:       PrðC ¼ 1Þ ¼ Pr v y0 þ w ; p  t þ ew  v y0 ; p þ e0      ¼ Pr e0  ew  v y0 þ w ; p  t  v y0 ; p ð4Þ Define u ¼ e0  ew . u has a cdf F(u) and strictly positive density f(u). Thus      PrðC ¼ 1Þ ¼ F v y0 þ w ; p  t  v y0 ; p We totally differentiate to organize the determinants of different risks of being observed in a worst form of child labor:     @vw @v0 @vw @v0 @vw  dPrðC ¼ 1Þ ¼ f ðuÞ  dy0 þ  dp  dt þ dw @y @y @p @p @y (5)

6

ERIC V. EDMONDS

    With diminishing marginal utility of income, @vw =@y o @v0 =@y . Declines in income opportunities open to the child absent participation in the worst forms tend to push children toward participation in worst forms. The amount of the push depends on the curvature of the indirect utility function. Higher net income available in the worst form also pulls children toward the activity. The extent of the pull depends on the marginal utility of income. Hence, poorer families are more likely to select into worst forms, because they are poorer. The agent’s disutility from the activity is also an influence as are prices. Testing between this model of entry into worst forms and the victimization model often posited in policy work depends on a comparison of the determinants of entry into other types of child labor. The poor will always select into work, because their marginal utility of income is higher. The victimization model implies that children arbitrarily enter into worst forms because of indifferent or uncaring adult agents. This implies t ¼ 0. In equilibrium, then worst forms should not pay more than other forms of work w ¼ 0, and the determinants of participation in worst forms should look like any other type of work. Thus, the comparison of the determinants of entry into worst forms and other forms of work is central to gauging the appropriateness of the compensating wage differential model.

3. METHODOLOGY – ESTIMATING THE CORRELATES OF SELECTION Why are children engaged in WFCL? Empirically, this is a hard question to answer, because WFCL are relatively rare. The probability that random sampling captures children engaged in any given WFCL is typically low. Hence, data collection can be prohibitively costly, and statistical power is always a concern. Researchers have had to turn to other data sources. The most common approach is inherently qualitative. Researchers find children engaged in a worst form and interview them to find out about their circumstances. Sometimes, these interviews are unstructured, but often researchers follow a survey questionnaire that can permit quantitative analysis. It is impossible to learn about why children are in worst forms from only interviewing children in worst forms. Consider some factor D that influences selection into activity C. The researcher is interested in knowing how factor D increases the probability that a child with other characteristics X enters into activity C. When D is discrete, this is PðC ¼ 1jD ¼ 1; X Þ

Selection into Worst Forms of Child Labor

7

PðC ¼ 1jD ¼ 0; X Þ. Neither probability can be computed in the set of C ¼ 1. Put another way; let us say a child in an interview remarks that they are engaged in activity C because of factor D (‘‘I am a ragpicker, because my dad lost his job’’). There may be lots of children who experience factor D that do not select into C (lots of children have parents become unemployed without becoming ragpickers), but without data on children not in C, there is no way to compute the increased chance of engaging in C with a change in D. The problem of drawing inference about rare events is not unique to WFCL. Most observational inference in medicine is made under precisely these circumstances, and this study applies these approaches to rare events from epidemiology to the study of selection into WFCL. These techniques do not appear to have been applied to the analysis of selection into worst forms before. The present discussion draws heavily from papers such as Prentice and Pyke (1979), Lancaster and Imbens (1996), King and Zeng (2001), and Manski (2001). Let C i be an indicator that child i is involved in the given worst form of interest. Di is the covariate of interest. In the present discussion, Di is binary, but the discussion generalizes to when Di takes more than two values. Our interest is in estimating the impact of Di on the probability that child i is involved in the given worst form. Later attention will be placed on estimating this probability conditional on other confounding variables that are correlated with both C i and Di . There are three main outcomes of potential interest. 1. Absolute risk. How likely is an individual with Di to be involved in the given worst form: pi ¼ PrðC ¼ 1jDi Þ

(6)

2. Relative risk. How much more likely is a child with D ¼ 1 to be observed in activity Y than a child with D ¼ 0: R¼

PrðC ¼ 1jD ¼ 1Þ PrðC ¼ 1jD ¼ 0Þ

(7)

3. Attributable risk. How much does an individual’s risk of engaging in Y increase with a change in D from 0 to 1: A ¼ PrðC ¼ 1jD ¼ 1Þ  PrðC ¼ 1jD ¼ 0Þ

(8)

Each of these outcomes is potentially of considerable interest to researchers and policy. For example, absolute risk is of interest to assess how likely a child

8

ERIC V. EDMONDS

with a given characteristics is to be in a WFCL. An index of vulnerability to WFCL would be constructed entirely by combining measures of absolute risk. Researchers interested in how participation in WFCL differs with variation in observable characteristics will be most concerned with relative or attributable risk. Relative risk is the most straightforward to estimate. However, relative risk can often be misleading to interpret in the context of low probability events. For example, suppose that the probability of observing a child in a WFCL is extremely low when a certain characteristic D is not present (e.g., 0.00001) and suppose the probability is higher when the characteristic is present (e.g., 0.0001) but still so small as to not be substantive. Estimates of relative risk in this case would be very large (10) even though the probabilities are so small as to not be substantive. Hence, at a minimum, relative risk should never be considered without attention to the baseline absolute risk. In contrast, attributable risk gives a direct measure of how much a child’s risk of being involved in a WFCL changes with an observed characteristic. Consequently, it is the outcome of interest most often. Estimating absolute, attributable, or relative risk requires data on both cases (subjects where C ¼ 1) and controls (C ¼ 0). When data on both cases and controls can be collected in a single randomized survey, standard cohort comparison techniques are available. However, typically the incidence of most forms of WFCL is such that a survey would need to be extremely large in scale to recover engaged children using random sampling. Thus, a more common situation is to have separately collected data on children not engaged in the WFCL (the control data) and data on children engaged in the activity (the case data). The case data do not need to be obtained through randomized sampling, but estimating absolute or attributable risk requires knowledge of the probability a child engages in the WFCL. This is most easily assessed if the case data collection is designed, in part, to estimate this parameter. Moreover, whatever sampling procedure generates the case data, sampling must be independent of the covariates D of interest except in as much as D is correlated with selection into the case data. Put another way, the data generation process can generate bias if it is correlated with covariates of interest for reasons other than that the covariates are correlated with selection into worst forms. Ideally, the survey instrument used to collect data on the case and control populations will be identical. In practice, it is rare that similar case and control data exist. It will only be possible to compute any of the risk parameters of interest for covariates that appear in both the case and the control data. Moreover, a common problem is that even when there are similar questions, the case and control data will be answered by different people. Often case data

Selection into Worst Forms of Child Labor

9

are collected by interviewing children while most household surveys and censuses (typical sources of control data) interview household heads or their spouses. Biases from differences in respondents can be as substantive as biases from different framing of questions, and these dissimilarities make it very challenging to assess the risk parameters of interest. The classic case–control approach makes the rare events assumption to estimate relative risk. That is, the case and control data are pooled, and it is assumed that the probability of observing a case individual tends to zero (conditional on observed characteristics) in the limit. This assumption allows the researcher to interpret the odds ratio from a logit of participation in the WFCL on observable characteristics as an estimate of relative risk. The appeal of this approach is that it is possible to estimate relative risk without identifying absolute risk in the population. However, the rare events assumption is problematic. The existence of case data implies that the probability of observing a case is not zero, and the rare events assumption implies that attributable risk is zero. Knowledge of the probability of observing a case in the population substantially improves estimation. Let l denote the incidence of the worst form in the population, and let C be the fraction of the case–control pooled data that is from the case data. To estimate parameters such as absolute risk and attributable risk, the constant from the logit needs to be corrected to  Specifically, the regression’s intercept reflect the difference between l and C. in the logit b0 needs to be adjusted as:     1l C b0  ln l 1  C

(9)

This result is attributable to Manski and Lerman (1977) or Prentice and Pyke (1979). The intuition behind this adjustment is that in general the ratio of case to control observations in the pooled data will not correspond to the ratio expected in the population. Hence, predicted probabilities from the pooled regression would not reflect the true, underlying population prevalence of the worst form. Rescaling the intercept as in Eq. (9) assures that the predicted probabilities match what could be observed if all of the population data existed. Sometimes researchers will not have an estimate of the incidence rate in the population. Theoretical work such as Manski (2001) considers cases where there is no prior information about the range of plausible values of l. However, at a minimum, researchers will have some idea of a plausible range of values for incidence in the population. Let lL and lH indicate the

10

ERIC V. EDMONDS

lower and upper values of the plausible range of l. King and Zeng (2001) suggest computing bounds on possible values of absolute and relative risk by estimating at both lL and lH. Because absolute and relative risk are positive monotone functions of l, computing either at the lower and upper values of l defines bounds on the range of possible absolute and relative risks. Attributable risk is more difficult, because it is not a positive monotone function of l. Define A(lk) as the estimate of attributable risk associated with an estimated incidence of lk. King and Zeng (2001) suggest checking for whether lL and lH are in a monotone region of attributable risk by evaluating whether attributable risk appears to have the same derivative with respect to l at its both high and low values. This can be checked by verifying that the signs of A(lLþe)  A(lL) and A(lHþe)  A(lH) are the same. When lL and lH are in a monotone region of A, then bounds can be calculated as:

 A 2 min½AðlL Þ; AðlH Þ; max½AðlL Þ; AðlH Þ

(10)

Sometimes, population prevalence rates will be in a non-monotone region of attributable risk. In this case, King and Zeng (2001) show that bounds on attributable risk are given by:

 A 2 min½AðlL Þ; Aðl0 Þ; AðlH Þ; max½AðlL Þ; Aðl0 Þ; AðlH Þ where Aðl0 Þ ¼

(11)

pffiffiffiffi  pffiffiffiffi  o1 o þ 1 and o is the odds ratio: o¼

PrðC ¼ 1jD ¼ 1ÞPrðC ¼ 0jD ¼ 0Þ PrðC ¼ 0jD ¼ 1ÞPrðC ¼ 1jD ¼ 0Þ

(12)

In what follows, econometric work focuses on presenting estimates of attributable risk. Attributable risk is the focus of this study, because the primary aim of the empirical work is to isolate indicators associated with increased risk of participating in a worst form. Hence, the size of the increased risk associated with a given factor of interest is important. In the present case, attributable risk is computed using prior correction (Eq. (9)) to compute absolute risks and in turn compute the difference. All empirical work is implemented using the regression code available freely from Tomz, King, and Zeng (2003).

Selection into Worst Forms of Child Labor

11

4. DATA There are an estimated 127,143 children engaged in the WFCL in Nepal (ILO, 2001). There are approximately 8 million children below the age of 16 in Nepal, and the ILO estimates that 1.5 percent of these children work in these WFCL. This study uses a survey of short-route porters, ragpickers, and the population census to consider the correlates of selection into WFCL in Nepal for children age 10–14.

4.1. Porters For many areas of Nepal, porters are critical for transporting consumer goods, getting business output for market, and delivering construction materials to remote areas. Porters are typically classified as long-route and short-route porters, and the two types of porters appear to be somewhat segmented. This study focuses on short-route porters. Short-route porters are typically contracted at spot markets in local markets and bus parks. Portering is considered a worst form of child labor, because children often carry heavy loads, across difficult terrain, for long hours. In the data used in this study, short-route porters report working approximately 10 hours per day for 6 days a week on average (KC, Adhikari, Subedi, & Gurun, 2001a). Two-thirds of short-route porters report averaging roughly 10 routes per day that range in weight from 10 to 50 kilograms (although one has to be suspect about self-reported load weights). Sixty percent report not wearing protective gear such as boots, gloves, or pads on the head. The short-route porter survey (SRP) was conducted in urban areas of Nepal as that is where short-route portering is concentrated. The SRP sampled work sites: markets and bus parks. Out of an estimated 423 market centers and bus stops, a random sample of porters was interviewed in 97 randomly selected market centers and 15 randomly selected bus parks. When appropriately weighted, the SRP suggests a total of 5,087 short-route child porters aged 6–17 in Nepal in 2003. A total of 30 of these are below the age of 10, and most are aged 14 or more. In the present study, we focus on short-route porters aged 10–14. There are an estimated 1,404 short-haul porters aged 10–14 in urban Nepal in 2003. Column 1 of Tables 1–4 provide summary statistics on the 164 porters aged 10–14 captured in the SRP. Table 1 summarizes child characteristics. Table 2 describes the child’s background (information about where the child’s family lives). Table 3 contains information on the child’s father. Table 4 details the

12

ERIC V. EDMONDS

Table 1. Mean/SE

Number of observations Estimated population size Age Female Ethnicity High-status Hindu caste Tharu Newar Dalit Muslim Other Native language Nepali Tharu Other In school Can read and write Completed some school Completed Std. 5 Completed post primary

Child Characteristics in Porters Survey, Ragpickers Survey, and Census. Short-Route Porters Survey

Ragpickers Survey

(1)

2001 Population and Housing Census Wage work

Home enterprise work

Not work

(2)

(3)

(4)

(5)

164

372

6,900

25,390

297,506

1,404

974

63,143

254,290

2,592,568

13.0 (0.107) 0.282 (0.058)

12.0 (0.096) 0.198 (0.043)

12.4 (0.020) 0.370 (0.008)

12.3 (0.010) 0.612 (0.004)

11.8 (0.003) 0.470 (0.001)

0.192 (0.036) 0.126 (0.042) 0.013 (0.008) 0.285 (0.068) 0.037 (0.023) 0.348 (0.058)

0.163 (0.049) 0.006 (0.003) 0.026 (0.014) 0.339 (0.104) 0.080 (0.059) 0.384 (0.071)

0.094 (0.005) 0.151 (0.007) 0.027 (0.003) 0.302 (0.009) 0.100 (0.007) 0.325 (0.008)

0.253 (0.005) 0.062 (0.003) 0.025 (0.002) 0.202 (0.004) 0.047 (0.003) 0.411 (0.006)

0.351 (0.003) 0.076 (0.002) 0.057 (0.002) 0.145 (0.002) 0.034 (0.001) 0.336 (0.002)

0.588 (0.067) 0.109 (0.042) 0.303 (0.059) 0.190 (0.054) 0.687 (0.046) 0.869 (0.056) 0.159 (0.041) 0.085 (0.034)

0.394 (0.090) 0.007 (0.004) 0.599 (0.089) 0.100 (0.026) 0.450 (0.068) 0.944 (0.019) 0.067 (0.020) 0.011 (0.008)

0.222 (0.008) 0.133 (0.007) 0.644 (0.010) 0.159 (0.006) 0.272 (0.008) 0.185 (0.007) 0.063 (0.004) 0.025 (0.002)

0.484 (0.006) 0.052 (0.003) 0.464 (0.006) 0.271 (0.005) 0.381 (0.005) 0.291 (0.005) 0.105 (0.003) 0.049 (0.002)

0.520 (0.003) 0.058 (0.002) 0.422 (0.003) 0.864 (0.001) 0.875 (0.001) 0.822 (0.002) 0.346 (0.002) 0.191 (0.002)

Note: Sample restricted to children aged 10–14.

13

Selection into Worst Forms of Child Labor

Table 2. Mean/SE

Background Characteristics in Porters Survey, Ragpickers Survey, and Census. Short-Route Porters Survey

Belt Hill Terai Region East Central West Mid-West Far-West Household background Owns farmland

Ragpickers Survey

2001 Population and Housing Census Wage work

Home enterprise work

Not work

(1)

(2)

(3)

(4)

(5)

0.503 (0.086) 0.497 (0.086)

0.540 (0.190) 0.460 (0.190)

0.191 (0.011) 0.789 (0.011)

0.495 (0.007) 0.368 (0.006)

0.462 (0.005) 0.478 (0.005)

0.131 (0.054) 0.300 (0.066) 0.363 (0.087) 0.172 (0.078) 0.034 (0.018)

0.165 (0.120) 0.616 (0.175) 0.177 (0.122) N/A 0.041 (0.044)

0.312 (0.011) 0.406 (0.012) 0.107 (0.006) 0.112 (0.007) 0.063 (0.006)

0.192 (0.005) 0.306 (0.006) 0.155 (0.005) 0.188 (0.005) 0.158 (0.005)

0.230 (0.004) 0.335 (0.005) 0.223 (0.004) 0.117 (0.003) 0.094 (0.003)

0.671 (0.075)

0.397 (0.077)

0.508 (0.010)

0.934 (0.002)

0.821 (0.005)

Note: Sample restricted to children aged 10–14.

information on the child’s mother. All information in the SRP is collected from interviewing the child porter. Means and standard errors are reported. Both are corrected for sample design and weighted to be nationally representative. Porters are mostly boys. Less than one in five attend school. Less than 70 percent say they are literate. They come from the hills and the Terai and from the Central and Western areas. Their parents are relatively old, illiterate, and much more likely to be disabled than other populations. 4.2. Ragpickers Ragpickers collect rags and other used goods to be recycled and reused. As an activity, ragpicking is primarily an urban activity. Adult and child ragpickers collect plastics, polyethylene, bottles, metals, and tins from dumping sites,

14

ERIC V. EDMONDS

Table 3. Mean/SE

Reports characteristics Age Can read and write Completed some school Completed Std. 5 Completed post primary Disabled Not work Owns small business Works for wages Employed in agriculture

Paternal Characteristics in Porters Survey, Ragpickers Survey, and Census. Short-Route Porters Survey

Ragpickers 2001 Population and Housing Census Survey Wage Home Not work work enterprise work

(1)

(2)

(3)

(4)

(5)

0.847 (0.044) 48.779 (1.696) 0.263 (0.054) N/Aa

0.871 (0.020) 44.110 (0.735) 0.299 (0.043) 0.166 (0.028) 0.105 (0.022) 0.082 (0.018) 0.035 (0.013) 0.087 (0.017) 0.095 (0.042) 0.762 (0.057) 0.088 (0.024)

0.910 (0.004) 43.855 (0.196) 0.290 (0.011) 0.179 (0.010) 0.148 (0.009) 0.124 (0.009) 0.001 (0.001) 0.057 (0.004) 0.108 (0.006) 0.571 (0.009) 0.629 (0.011)

0.906 (0.002) 45.175 (0.099) 0.335 (0.004) 0.162 (0.003) 0.101 (0.003) 0.068 (0.002) 0.002 (0.000) 0.036 (0.002) 0.059 (0.002) 0.095 (0.003) 0.862 (0.003)

0.888 (0.001) 44.717 (0.040) 0.573 (0.003) 0.334 (0.003) 0.269 (0.003) 0.214 (0.003) 0.001 (0.000) 0.066 (0.001) 0.109 (0.002) 0.225 (0.003) 0.682 (0.005)

0.342 (0.136) 0.251 (0.120) 0.005 (0.004) 0.059 (0.023) 0.053 (0.021) 0.425 (0.065) 0.545 (0.067)

Note: Sample restricted to children aged 10–14. All children report parent completing at least grade 1.

a

streets, river banks, etc. These collected materials are sold to junkyards and shops that in turn sell these materials to suppliers for recycling. Ragpicking is nearly universally viewed as a worst form because of the extremely hazardous work environment (KC, Gurung, Adhikari, & Subedi, 2001b). The ragpickers survey (RAG) was conducted in urban areas of Nepal. The original survey design was to sample sites where ragpickers worked. However, researchers found it difficult to interview children in dumping areas, garbage disposal and refuse areas, slums, and river banks and faced additional difficulties associated with the mobility of ragpickers. Thus, while

15

Selection into Worst Forms of Child Labor

Table 4.

Maternal Characteristics in Porters Survey, Ragpickers Survey, and Census.

Mean/SE

Short-Route Porters Survey

Ragpickers Survey

2001 Population and Housing Census Wage work

Reports characteristics Age Can read and write Completed some school Completed Std. 5 Completed post primary Disabled Not work Owns small business Works for wages Employed in agriculture

Home enterprise Not work work

(1)

(2)

(3)

(4)

(5)

0.842 (0.036) 40.512 (1.872) 0.065 (0.025) N/Aa

0.828 (0.026) 36.914 (0.786) 0.127 (0.036) 0.088 (0.024) 0.042 (0.017) 0.021 (0.012) 0.013 (0.007) 0.413 (0.050) 0.032 (0.013) 0.496 (0.074) 0.076 (0.023)

0.919 (0.004) 39.468 (0.172) 0.151 (0.010) 0.088 (0.008) 0.079 (0.007) 0.072 (0.007) 0.007 (0.002) 0.376 (0.010) 0.098 (0.006) 0.350 (0.009) 0.477 (0.011)

0.928 (0.002) 40.337 (0.084) 0.075 (0.003) 0.026 (0.001) 0.017 (0.001) 0.010 (0.001) 0.006 (0.001) 0.161 (0.004) 0.098 (0.003) 0.027 (0.002) 0.780 (0.004)

0.953 (0.001) 39.484 (0.035) 0.232 (0.003) 0.117 (0.003) 0.090 (0.002) 0.068 (0.002) 0.004 (0.000) 0.366 (0.003) 0.102 (0.001) 0.053 (0.001) 0.556 (0.004)

0.042 (0.051) 0.042 (0.051) 0.018 (0.017) 0.284 (0.048) 0.007 (0.004) 0.301 (0.057) 0.503 (0.071)

Note: Sample restricted to children aged 10–14. a All children report parent completing at least grade 1.

the survey was being fielded, enumerators abandoned the original sample frame and interviewed children in junkyard shops or locations where they spend their leisure time (Mhukherjee, 2003). The non-random nature of the survey and this disconnect between sample design and survey implementation creates an unknowable array of problems for inference and makes it impossible to know whether estimates of the incidence of ragpicking from these data are accurate. If one is willing to treat the RAG data as if it were based on random sampling of job sites, it is possible to make inferences about the scope of ragpicking in Nepal. That

16

ERIC V. EDMONDS

said, the survey suggests that there are 3,695 child ragpickers aged 6–18 in Nepal in 2002; 974 of these ragpickers are aged 10–14. Column 2 of Tables 1–4 provides summary statistics for the 372 children aged 10–14 interviewed in the RAG survey. They are mostly boys. One in ten is in school. Less than half say they can read and write. They come from the central hills and Terai. Their parents are more likely to be literate than Porters, but their parents are also far more likely to be disabled.

4.3. The Census The population and housing census of 2001 is used for the control sample. It precedes RAG by a year and SRP by two, but it contains considerable overlap in content. In using the earlier sample as a control, the analysis herein is based on the assumption that there are not substantive changes over time in any of the characteristics that are the focus of the analysis. For example, if the civil war in Nepal lead to a substantive rise in disability between 2001 and 2002, then one might misattribute the time trend in disability in Nepal as a correlate of selection into ragpicking. The proximity in time between the census and the two targeted survey mitigates this concern, but it would be ideal to have the control data simultaneous with the targeted surveys. There are several substantive issues that arise from using the Census as a control sample beyond timing. First, it is only possible to study children aged 10–14. Information on economic activities is not collected for children below age 10. Education data are incomplete on children above age 14 in the census because of an odd skip pattern in the questionnaire. Second, it is only possible to discern familial relationships for children of the household head. This introduces non-random selection bias into the control sample by eliminating children who live in households where a parent is not codified as the head. This could be a problem for inference if whether a parent is coded as a household head in the census is correlated with selection into portering or ragpicking and other observable household characteristics. Parental death is one potential concern. A child who has experienced a parental death is less likely to have a parent coded as the household head (mechanically, they have one fewer parent who could be a household head). If parental death is associated with other background characteristics and selection into a worst form, any estimates of attributable risk associated with the background characteristics

Selection into Worst Forms of Child Labor

17

could be severely biased. Unfortunately, there appears to be no obvious solution to this problem. Third, it is not possible to identify porters and ragpickers in the census. Occupation and Industry of employment are only reported at the one-digit level. It is technically possible that our analysis is biased by miscoding some ‘‘case’’ observations as ‘‘control.’’ This would diminish our ability to identify correlates of entry into these worst forms. Given that the public use microdata is a (11 percent) subsample of the entire census and that portering and ragpicking are rare, it is unclear how substantive this bias is likely to be. Fourth, the census data are collected from the head of the household. Thus, the respondent in the census is different than the respondent in the SRP and RAG, who are children aged 10–14. It is possible that this difference in respondents creates a source of bias. For example, in Tables 3 and 4, all porters respond that their parents have some school, whereas only 18 percent of wage working children have parents with some school. This may reflect differences in interviewer instructions, the exact wording of the questionnaire, or differences in the respondent. Ultimately, it is impossible to discern when differences in data between surveys reflect substantive differences versus any of these other issues. Columns 3–5 of Tables 1–4 provide summary statistics for children aged 10–14 of the household head in the public use census micro-data. Census children are trifurcated into children involved primarily in wage work, primary in home enterprise work, or no form of work. This latter category includes children who work in domestic service in their own home, children who are inactive, and children who are primarily students. For each category of activity, means and standard errors are reported. Both are corrected for sample design and weighted to be nationally representative. The analysis in this study compares the correlates of entry into worst forms with the correlates of entry into wage work. The choice of wage work for benchmarking is driven by the fact that, like portering and ragpicking, the activity takes place outside of the child’s home. Other research on child labor has suggested that work inside and outside of the home can be treated very differently (e.g., Edmonds, 2008). Hence, it is more interesting to compare findings on ragpickers and porters to wage workers rather than home enterprise workers, for example. In fact, although magnitudes differ substantively, the flavor of the basic comparisons would be similar if all working children were used as a benchmark rather than just wage workers. The correlates of selection into worst forms looks distinct compared to both wage working children and children in family businesses or farms.

18

ERIC V. EDMONDS

5. MAIN FINDINGS 5.1. Porters A comparison of the descriptive statistics of porters to wage works in Tables 1–4 highlights many similarities and differences. Table 1 summarizes child characteristics. Short-route porters differ from wage workers in that they are more likely to be high status, more likely to speak Nepali, and less likely to be Muslim. These differences likely reflect that short-route porters are only interviewed in urban areas and tend to be from those some areas. That is, most short-route porters in the survey are not in-migrants to urban areas. Thus, populations of rural origin (such as Muslims or non-Nepali speaking populations) are less present in the SRP. The reported completed schooling of porters is also higher than other wage workers. Background characteristics are in Table 2. Porters are more likely to be from hill areas than wage workers. It is not surprising that porters are more prominent in hill areas, less prominent in plains, as the road infrastructure around the Terai’s mid-sized cities are generally better than in the hill areas. Moreover, porters are more active in the western development region of Nepal than are wage-workers. This likely reflects the fact that short haul porters often work around bus stations and larger markets, which are more prevalent in the central and west regions of Nepal than elsewhere. Compared to wage workers, porters seem to come from relatively disadvantaged backgrounds. Literacy among both fathers and mothers is lower for porters than wage workers. Porters report higher levels of paternal schooling completion than paternal literacy; hence, either there is data error on the coding of paternal education or there are lots of illiterate fathers of porters who have completed primary school. The maternal education data are consistent with the observed lower maternal literacy rates for porters than wage workers. Porters are also more likely to report both a father or a mother who is disabled and to report a mother who is working. Paternal and maternal disability and maternal wage work stand out as strong predictors of selection into portering. The first column of Table 5 contains estimates of attributable risk for each listed row characteristic separately. Attributable risk is computed as described in the methodology section, following King and Zeng (2001). Specifically, the census and SRP data are pooled. An indicator that a child is a porter is regressed (using a logit) on age, gender, ethnicity, language, belt, development region, and, in column 1, the variable indicated by the row. The logit is estimated using prior correction (Eq. (9)) with a bias correction for small samples.

0.00036

0.00039 0.00006 0.00016 0.00024 0.00010 0.00043 0.00031 0.00009 0.00016 0.00039 0.00051 0.00034

0.00009

0.00024 0.00118 0.00000 0.00012 0.00030 0.00018

0.00020 0.00143 0.00006 0.00025 0.00127 0.00010

0.00009 0.00773 0.00006 0.00015 0.00257 0.00005

0.00012 0.00503 0.00025 0.00002 0.00065 0.00003

0.00008

0.00003

Upper

0.00010 0.00054 0.00001 0.00014 0.00042 0.00005

0.00012 0.00065 0.00010 0.00009 0.00003 0.00017

0.00009

Estimate

Attributable risk

0.00016 0.00007 0.00012 0.00025 0.00011 0.00024

0.00023 0.00000 0.00018 0.00017 0.00011 0.00038

0.00028

Lower

0.00004 0.00236 0.00010 0.00007 0.00105 0.00004

0.00004 0.00315 0.00003 0.00002 0.00006 0.00004

0.00005

Upper

95% confidence interval

Conditional

Notes: All regressions include controls for child age, gender, ethnicity, language, belt, and development region. All standard errors corrected for clustering at the block level (primary sampling unit). Estimates computed using King and Zeng’s relogit code with prior correction (http:// gking.harvard.edu/stats.shtml#relogit). Each estimate of attributable risk in the ‘‘unconditional’’ column is from a separate regression. Each estimate in the ‘‘conditional’’ column is from one regression, including all the listed covariates. All estimates assume an incidence of shortroute porters of 0.05 percent. Attributable risks are computed for a change in the row variable from 0 to 1 at the mean of all other covariates except all ‘‘conditional’’ estimates are computed at father and mother reports characteristics ¼ 1.

0.00078

Lower

95% confidence interval

0.00033

Estimate

Attributable risk

Unconditional

Attributable Risk Estimates for Background Characteristics in Short Route Porters Survey.

Household background Owns farmland Father characteristics Reports paternal characteristics Can read and write Disabled Not working Owns small business Works for wages Employed in agriculture Mother characteristics Can read and write Disabled Not working Owns small business Works for wages Employed in agriculture

Table 5.

Selection into Worst Forms of Child Labor 19

20

ERIC V. EDMONDS

Attributable risk is then computed by estimating the differences in absolute risk level as computed with Eq. (8). Standard errors are corrected for clustering owing to sample design. In column 4, the conditional attributable risk estimates are computed by including all the listed controls in the logit and holding all listed observable characteristics constant (at their mean) except for the variable specified by the row. Paternal disability is the largest predictor of selection into portering. Paternal disability raises the probability a child is observed portering by more than a tenth of a percent. Maternal disability has a similar positive association with portering. Another strong indicator factor associated with an elevated risk of being a porter is having a mother working for wages. Female wage work is relatively rare in Nepal, so that this observation might reflect something about the geographic location of the control population relative to the portering population. It is also consistent with the idea that women only enter the labor market when the family’s marginal utility of income is very high. Hence, the wage work observation might be consistent with a view that poverty is critical in explaining selection into portering. The attributable risk estimates in Table 5 are not causal estimates of how selection into portering will be affected by changes in any of the listed observable characteristics. Rather, they describe how the likelihood of observing a child porter varies with changes in maternal or paternal characteristics. Table 6 contains estimates of attributable risk for becoming a porter associated with changing several of the covariates from Table 5 simultaneously. For example, a mother who is disabled and not working raises the probability a child is observed as a porter by nearly 0.2 percentage points for a landless household (nearly double the risk observed in a household with land). In general, landless households are more likely to be observed sending children to porter in the context of a paternal or maternal disability or if both mother and father are observed working. A comparison of attributable risk estimates in Table 6 to that observed for wage work in Table 7 is illustrative. The patterns observed with disability and literacy are similar for portering and other types of wage work. The main difference with portering is that the presence of selfemployment in the household lowers the risk of portering (while it raises the risk of observing a child in wage work). Hence, the portering data at least contain some suggestion that the availability of employment within the household may be associated with a diminished risk of seeking work in portering outside of the family. It is important to note, however, that the magnitudes of the observed changes in attributable risk with home enterprises are very small.

0.0001 0.0003 0.0003 0.0003 0.0002 0.0003

0.0001 0.0001

0.0001 0.0001 0.0002 0.0002 0.0001 0.0002

0.0000 0.0003

0.0000 0.0005

0.0002 0.0003

0.0000 0.0001

0.0001 0.0010

0.0003

0.0001

0.0002 0.0003

0.0001

0.0000 0.0001 0.0001

0.0006 0.0016

Estimate

0.0017 0.0043

Upper

Attributable risk

0.0002 0.0001

0.0004 0.0006

0.0006

0.0005 0.0007

0.0003

0.0001 0.0001

Lower

0.0001 0.0015

0.0000 0.0001

0.0001

0.0001 0.0002

0.0000

0.0029 0.0102

Upper

95% confidence interval

Landless

Notes: Attributable risks computed using results from the ‘‘conditional regression’’ results in Table 5. The first columns compute probabilities for households with mean probability of holding land. The second column computes probabilities for household without landholdings. a Change in probability that child is a porter if father moves from not disabled and mean work to disabled and no work (any category). b Same as footnote (a) for mother. c Change in probability that child is a porter if dad is literate with average schooling and mom moves from illiterate to literate (with no schooling). d Change in probability that child is a porter if illiterate mom and dad shifts to a illiterate mom with literate dad (no schooling). e Change in probability that a child is a porter if illiterate mom and dad shifts to literate mom and dad (no schooling). f Change in probability that child is porter if household moves from no self-employment to mom self-employment. g Same as footnote (f ) only for father. h Change in probability that child is a porter if household moves from no self-employment to both mom and dad in self-employment. i Change in probability that child is a porter if household moves from no wage work to father wage work. j Same as footnote (i) except mom and dad in wage work.

0.0000 0.0000

Lower

95% confidence interval

0.0004 0.0008

Estimate

Attributable risk

At Mean Landholding Rate

Attributable Risk Estimates for Various Scenarios in Short-Route Porters Survey.

Disability Dad is disabled and cannot worka Mom is disabled and cannot workb Literacy Literate dad (average schooling) & illiterate mom (no schooling) to literate momc Illiterate mom and dad to literate dad (no schooling)d Illiterate mom and dad to literate mom and dad (no schooling)e Home enterprises Household without any self-employment to mom selfemploymentf Household without self-employment to dad self–employmentg Household without self-employment to mom and dad selfemploymenth Wage labor Household with no wage work to dadi Household with no wage work to mom and dadj

Table 6. Selection into Worst Forms of Child Labor 21

0.002 0.007 0.009

0.000 0.005 0.006

0.001 0.006 0.008

0.001 0.007 0.009

0.012

0.009

0.002

0.014

0.011

0.002

0.010 0.013

0.002

0.000 0.005 0.006

0.003 0.021

Estimate

0.012 0.024

Upper

Attributable risk

0.010

0.008

0.000

0.012 0.015

0.003

0.007 0.008

Lower

0.020

0.015

0.004

0.008 0.010

0.000

0.021 0.041

Upper

95% confidence interval

Landless

Notes: The first columns compute probabilities for households with mean landholdings. The second column computes probabilities for household without landholdings. a Change in probability that child is a wage worker if father moves from not disabled and mean work to disabled and no work. b Same as footnaote (a) for mother. c Change in probability that child is a wage worker if dad is literate with average schooling and mom moves from illiterate to literate (with no schooling). d Change in probability that child is a wage worker if illiterate mom and dad shifts to a illiterate mom with literate dad (no schooling). e Change in probability that a child is a wage worker if illiterate mom and dad shifts to literate mom and dad (no schooling). f Change in probability that child is wage worker if household moves from no self-employment to mom self-employment. g Same as footnote (f ) only for father. h Change in probability that child is a wage worker if household moves from no self-employment to both mom and dad in self-employment.

0.004 0.004

Lower

95% confidence interval

0.002 0.013

Estimate

Attributable risk

Average Land Holdings

Attributable Risk Estimates for Various Scenarios, Census Wage Workers.

Disability Dad is disabled and cannot worka Mom is disabled and cannot workb Literacy Literate dad (average schooling) and illiterate mom (no schooling) to literate momc Illiterate mom and dad to literate dad (no schooling)d Illiterate mom and dad to literate mom and dad (no schooling)e Home enterprises Household without any self-employment to mom selfemploymentf Household without self-employment to dad selfemploymentg Household without self-employment to mom and dad self-employmenth

Table 7.

22 ERIC V. EDMONDS

Selection into Worst Forms of Child Labor

23

5.2. Ragpickers Ragpickers also appear distinct from wage workers. Among the child characteristics described in Table 1, ragpickers tend to be younger than wage workers, and they are much less likely to be ethnic Tharu. The fact that ragpickers are younger is consistent with a role for employment opportunities in selection into ragpicking as young children have fewer formal wage earning opportunities. Ragpickers also appear to be relatively more educated although it seems likely that this difference with the census might reflect biases owing to who responds to the questionnaire. Ragpickers are less likely to be higher caste than the general population, and less likely to be Tharu. The low incidence of Tharu ragpickers is interesting. Two possible explanations seem obvious. First, ragpicking may be more common in places where the Tharu are less prevalent. Second, desperate Tharu may have better options than ragpicking. Bonded labor is common in the Tharu population, and one interpretation of their lower incidence of ragpicking is that the disamenities associated with accepting bondage are not as bad as those associated with a life of ragpicking. In Table 2, ragpickers appear more likely to be from hill areas than are wage workers and are more likely to be from central Nepal. The concentration of ragpickers is consistent with the location of the large recycling centers that are especially prevalent in the Kathmandu Valley (central-hill). However, this is also where trash is especially concentrated because of the population density. Hence, one should not infer that the presence of the recycling industry is the reason why there are ragpickers in the Valley. Of course, if there was no market for their output, it seems unlikely children would pick through trash except to help meet basic needs. Ragpickers are also less likely to come from households that own farmland. This observation is consistent with the view that a lack of alternative income-generating strategies may play an important role in selection into ragpickers. To some extent, this seems obvious as it is hard to imagine that picking through trash and debris is ever someone’s first choice for income. However, it is easy to over interpret this correlation between farmland and ragpicking. Children working for wages are less likely to own farmland than children who work in family enterprises (like farms). Moreover, a lack of land may be correlated with fewer at home employment opportunities, but it may also be correlated with a lack of income. Several parental background characteristics in Tables 3 and 4 suggest that selection into ragpicking is correlated with having a relatively disadvantaged background. Maternal literacy is lower than wage workers and both

24

ERIC V. EDMONDS

mothers and fathers of ragpickers are less likely to have some post primary education. Moreover, parental disability is a strong correlate of ragpicking (as has been observed with porters as well). Four percent of ragpickers have a disabled father, and 1 percent of ragpickers have a disabled mother. In contrast, less than one-tenth of 1 percent of the general population has a disabled father. Also, ragpickers are less likely to have a parent who owns a small business or is employed in agriculture. Although 63 percent of children in wage work have a father who works in agriculture, less than 9 percent of ragpickers do. Forty-eight percent of wage-earning children have a mother in agriculture. Less than 8 percent of ragpickers have a mother engaged in agriculture. It is impossible to discern whether this reflects the employment opportunities open to the children, the family’s disadvantaged background, or something transitory in the child’s family’s economic environment. However, the differences in the means are not present in other activities. Table 8 provides estimates of attributable risk by observable background and family characteristic. It is constructed identically to Table 5. Paternal disability stands out as the largest predictor of selection into ragpicking. Less than three-hundredths of a percent of children aged 10–14 are engaged in ragpicking, but paternal disability raises the probability that a child is observed in ragpicking by nearly two-tenths of a percent. Although no other characteristic is as strong a predictor as paternal disability, the observation that the child’s family’s employment background is an important risk factor persists in the attributable risk estimates. Either owning agricultural land or maternal or paternal work in agriculture substantially lowers the odds of observing a child in ragpicking. This may reflect differences in location rather than the household’s employment opportunities, but the fact that maternal self-employment also is associated with a diminished risk of observing a child as a ragpicker suggests that at least some part of why these are risk factors may owe to employment opportunities. Estimates of changes in attributable risk are generally uninformative in the conditional specification. The one exception is with regard to paternal disability, because that is such a large predictor of selection into ragpicking. In Table 9, observing a disabled father significantly increases the risk that a child is observed ragpicking, and this increased risk of ragpicking is larger for the landless than for children who come from families with land. The larger magnitudes estimated for landless families are consistent with the descriptive data, which also suggest a link between selection into ragpicking and employment opportunities. However, in general, there are few observable characteristics other than paternal disability, which can predict

0.00031 0.00006 0.00008 0.00052 0.00000 0.00004 0.00030 0.00023 0.00003 0.00001 0.00003 0.00007 0.00021

0.00002 0.00004 0.00154 0.00002 0.00000 0.00016 0.00012 0.00001 0.00016 0.00001 0.00004 0.00011

Lower

0.00000 0.00002 0.00000 0.00001 0.00002

0.00001 0.00024 0.00000 0.00000 0.00003

0.00000 0.00002 0.00376 0.00006 0.00009 0.00007 0.00005 0.00002 0.00057 0.00008 0.00001 0.00005

0.00000

Estimate

Attributable risk

0.00008

Upper

95% confidence interval

0.00017

Estimate

Attributable risk

Unconditional

0.00001 0.00000 0.00000 0.00002 0.00005

0.00002 0.00006 0.00002 0.00001 0.00009

0.00001

Lower

0.00000 0.00015 0.00001 0.00000 0.00000

0.00000 0.00064 0.00000 0.00001 0.00001

0.00000

Upper

95% confidence interval

Conditional

Attributable Risk Estimates for Background Characteristics in Ragpickers Survey.

Notes: All regressions include controls for child age, gender, ethnicity, language, belt, and development region. All standard errors corrected for clustering at the block level (primary sampling unit). Estimates computed using King and Zeng’s relogit code with prior correction (http:// gking.harvard.edu/stats.shtml#relogit). Each estimate of attributable risk in the ‘‘unconditional’’ column is from a separate regression. Each estimate in the ‘‘conditional’’ column is from one regression, including all the listed covariates. All estimates assume an incidence rate of ragpicking of 0.03 percent. Attributable risks are computed for a change in the row variable from 0 to 1 at the mean of all other covariates except all ‘‘conditional’’ estimates are computed at father and mother reports characteristics ¼ 1.

Household background Owns farmland Father characteristics Reports characteristics Can read and write Disabled Not working Owns small business Employed in agriculture Mother characteristics Reports characteristics Can read and write Disabled Not working Owns small business Employed in agriculture

Table 8.

Selection into Worst Forms of Child Labor 25

0.00001 0.00003 0.00004

0.00002 0.00001 0.00002

0.00000 0.00001

0.00000 0.00001 0.00001

0.00000 0.00000 0.00000

0.00000 0.00005

0.00002 0.00017

0.00001 0.00000

0.00001 0.00008

0.00000 0.00001

0.00001

0.00001 0.00001

0.00000 0.00000

0.00000

0.00000

0.00075 0.00014

Estimate

0.00000

0.00121 0.00061

Upper

Attributable risk

0.00000 0.00002

0.00001 0.00002

0.00002

0.00004 0.00005

0.00001

0.00017 0.00001

Lower

0.00002 0.00024

0.00001 0.00000

0.00000

0.00000 0.00000

0.00000

0.00208 0.00064

Upper

95% confidence interval

Landless

Notes: Attributable risks computed using results from the ‘‘conditional regression’’ results in Table 8. The first columns compute probabilities for households with mean probability of holding land. The second column computes probabilities for household without landholdings. a Change in probability that child is a ragpicker if father moves from not disabled and mean work to disabled and no work (any catagory). b Same as footnote (a) for mother. c Change in probability that child is a ragpicker if dad is literate with average schooling and mom moves from illiterate to literate (with no schooling). d Change in probability that child is a ragpicker if illiterate mom and dad shifts to a illiterate mom with literate dad (no schooling). e Change in probability that a child is a ragpicker if illiterate mom and dad shifts to literate mom and dad (no schooling). f Change in probability that child is ragpicker if household moves from no self-employment to mom self-employment. g Same as footnote (f ) only for father. h Change in probability that child is a ragpicker if household moves from no self-employment to both mom and dad in self-employment. i Change in probability that child is a ragpicker if household moves from no wage work to father wage work. j Same as footnote (i) except mom and dad in wage work.

0.00011 0.00001

Lower

95% confidence interval

0.00046 0.00009

Estimate

Attributable risk

At mean landholding rate

Attributable Risk Estimates for Various Scenarios in Ragpickers Survey.

Disability Dad is disabled and cannot worka Mom is disabled and cannot workb Literacy Literate dad (average schooling) and illiterate mom (no schooling) to literate momc Illiterate mom and dad to literate dad (no schooling)d Illiterate mom and dad to literate mom and dad (no schooling)e Home enterprises Household without any self-employment to mom selfemploymentf Household without self-employment to dad self-employmentg Household without self-employment to mom and dad selfemploymenth Wage labor Household with no wage work to dadi Household with no wage work to mom and dadj

Table 9. 26 ERIC V. EDMONDS

0.00003 0.00010 0.00013 0.00007 0.00004 0.00008 0.00001 0.00008

0.00000 0.00001 0.00001 0.00001 0.00000 0.00001 0.00007 0.00080

0.00000 0.00002

0.00022 0.00011 0.00020

0.00039 0.00047

0.00009

0.00017 0.00001

Lower value

0.00022 0.00238

0.00000 0.00001 0.00000

0.00000 0.00000

0.00000

0.01995 0.00621

Upper value

95% Confidence Intervals for Bounds

Notes: Incidence rates bounded between 0.3 percent and 0.03 percent. Attributable risks computed using results from the ‘‘conditional regression’’ results in Table 8 assuming an incidence of 0.03 percent and unreported regressions assuming an incidence of 0.3 percent. a Change in probability that child is a ragpicker if father moves from not disabled and mean work to disabled and no work (any catagory). b Same as footnote (a) for mother. c Change in probability that child is a ragpicker if dad is literate with average schooling and mom moves from illiterate to literate (with no schooling). d Change in probability that child is a ragpicker if illiterate mom and dad shifts to a illiterate mom with literate dad (no schooling). e Change in probability that a child is a ragpicker if illiterate mom and dad shifts to literate mom and dad (no schooling). f Change in probability that child is ragpicker if household moves from no self-employment to mom self-employment. g Same as footnote (f ) only for father. h Change in probability that child is a ragpicker if household moves from no self-employment to both mom and dad in self-employment. i Change in probability that child is a ragpicker if household moves from no wage work to father wage work. j Same as footnote (i) except mom and dad in wage work.

0.00075 0.00014

Lower

0.00742 0.00136

Upper

Estimated Bounds

Bounds on Attributable Risk Estimates for Various Scenarios in Ragpickers Survey, Landless Households.

Disability Dad is disabled and cannot worka Mom is disabled and cannot workb Literacy Literate dad (average schooling) and illiterate mom (no schooling) to literate momc Illiterate mom and dad to literate dad (no schooling)d Illiterate mom and dad to literate mom and dad (no schooling)e Home enterprises Household without any self-employment to mom self-employmentf Household without self-employment to dad self-employmentg Household without self-employment to mom and dad self-employmenth Wage labor Household with no wage work to dadi Household with no wage work to mom and dadj

Table 10. Selection into Worst Forms of Child Labor 27

28

ERIC V. EDMONDS

a risk of ragpicking. This suggests that most of the determinants of selection into ragpicking are outside the scope of the available data. Another important reason why the attributable risk of ragpicking is so small is that ragpicking is estimated to be extremely rare (less than three hundredths of a percent of children aged 10–14). Section 3 discussed how to estimate bounds on attributable risk when the incidence of a worst form is uncertain. Table 10 implements this methodology. The incidence of ragpicking is assumed to vary between 0.03 percent and 0.3 percent. Thus, the estimates from Table 9 are used for one bound and attributable risks are recalculated assuming an incidence of three-tenths of a percent to form the other bound. The data pass the test for positive monotonicity suggested in Section 3. Table 10 contains bounds on attributable risk for landless households. Contrasting Tables 9 and 10 highlights how important estimates of baseline incidence are for computing attributable risk. In very low probability events, it is a challenge to capture covariates that substantially increase the risk of the child entering the worst form simply because the event itself is rare. In general, the patterns recovered by the bounds estimates in Table 10 suggest risk factors for entry into ragpicking that are similar to that observed for portering and different with regards to self employment from what was observed in Table 7 for wage work.

6. CONCLUSION This study illustrates an approach to study the correlates of participation in a worst form of child labor. Survey data on the background characteristics of children engaged in WFCL are combined with nationally representative data on those same background characteristics. With this combination of data, it is possible to calculate what characteristics are associated with an increased risk of engaging in a worst form of child labor. When combined with data on the incidence of the worst form in the population, it is possible to compute how large of an increased risk of involvement in a worst form can be attributed to variation in a characteristic. This simple, descriptive comparison sheds some light on how the background characteristics of children engaged in portering and ragpicking in Nepal differ from the general population of children in Nepal. Are worst forms different than other more common forms of employment from the perspective of the agent who makes decisions about child time allocation? The data are consistent with the view that worst forms are different. Most theoretical treatments of entry into worst forms posit that

Selection into Worst Forms of Child Labor

29

children are more likely to enter worst forms when their alternative employment opportunities are limited. A child is more likely to participate in a worst form when the net economic return is larger. The data suggest that children are more likely to be involved in wage work when there is a family enterprise. This could reflect a causal impact of the family’s business, or it might reveal that family’s are more apt to own businesses in locations with more active labor markets. However, children are less likely to engage in work as ragpickers and porters when there is a family business at home. This association could reflect something about the impact of a family enterprise on entry into worst forms through the value of child time in the family business or the enterprise’s correlation with family incomes. Alternatively, the association between family enterprises and entry into worst forms might owe to an association between family enterprises and the overall local labor market (as speculated with regard to wage work). Most porters and ragpickers are working in the same geographic location as their parents. Thus, if omitted labor market characteristics were driving the finding that home enterprises are associated with a reduced risk of participation in a worst form, it is surprising that the ragpicker and porter patterns would differ from that observed for wage work. The idea that the association between home enterprises and entry is driven by either the potential economic contribution of the child to its household or the household living standards are more compelling. There are some further associations in the data that are consistent with the idea that the child’s employment opportunities in their household cast an important influence on entry into worst forms. Households with porters and ragpickers are less likely to own agricultural land, although this association is not particularly robust for these two populations. Portering is most prevalent in areas where there is the most need for porters as ragpickers are most prevalent in areas where there is trash and a recycling industry. Maternal wage work also seems to predict portering, and self-employment is negatively correlated with ragpicking. However, all these characteristics predict only a small amount of the observed prevalence of each worst form. Parental, especially paternal, disability stands out as a strong predictor of observing a child in a worst forms in Nepal. Relative to wage-working children, porters are five times and ragpickers are four times more likely to report that their father is disabled. This association between paternal disability and entry into worst forms could reflect that children are more vulnerable to victimization when their father is disabled, but their fathers are still living, and there is little correlation between paternal and maternal disability in the data. Paternal disability also does not appear to be strongly associated with some particular source location for the child; it is not likely

30

ERIC V. EDMONDS

to be capturing omitted geographic factors. Moreover, the magnitudes are so much larger than what is observed for any individual measure of parental self-employment or other household economic activity, it seems likely that paternal disability reflects more than an association between paternal disability and employment opportunities open to the child within its own household (which are conditioned on in the empirical work). It seems most plausible that the strong association between paternal disability and entry into worst forms reflects that paternal disability is strongly correlated with the child’s family being substantially poorer. If this interpretation is correct, then the data support Dessy and Pallage’s (2005) model of partially compensated wage differentials for WFCL. The methodology used to assess the correlates of selection into worst forms is general, but its data requirements are not trivial. Namely, four conditions must be met: 1. The type of work that qualifies as a worst form is explicitly identified 2. Reasonable estimates of the incidence of that worst form exist in the population 3. There are individual level data on background characteristics of children engaged in the worst form available 4. There are nationally representative data on the same set of background characteristics available for the general population. Unfortunately, the data on children in worst forms and the representative data used in this study are not perfectly consistent in how they collect information, and there is limited information that is in common in the targeted surveys and the nationally representative data. This problem is easily resolved if future survey work on children in worst forms would merely be attentive to existing data resources, and design their survey work to be in part consistent with nationally representative data. Even better, of course, would be to integrate target surveys into a broader national survey program and to combine that effort with scientific evaluation of interventions aimed at children engaged in WFCL.

ACKNOWLEDGMENTS I am grateful to Salil Sharma and Maheshwor Shrestha for exceptional research assistance. My time on this project was funded by the International Child Labor Program of the Bureau of International Labor Affairs, U.S. Department of Labor, and portions of this chapter appear in the ICLP

Selection into Worst Forms of Child Labor

31

report: ‘‘Alternative Income Generation and Entry into Worst Forms of Child Labor.’’ This chapter has benefited greatly from the comments of Alessandro Cigno, Deb DeGraff, Ken Swinnerton, and Sarah Gromly as well as to seminar attendants at the U.S. Department of Labor, The World Bank, the IZA/WB Conference on Employment in Development, the University of Minnesota Development Conference, and the Understanding Children’s Work/University of Paris I child labor conference.

REFERENCES Dessy, S., & Pallage, S. (2005). A theory of the worst forms of child labour. Economic Journal, 500(1), 68–87. Edmonds, E. (2007). Child labor. In: T. P. Shultz & J. Strauss (Eds), Handbook of development economics (Vol. 4, pp. 3607–3709). North-Holland: Amsterdam. Edmonds, E. (2008). Defining child labour: A review of the definitions of child labour in policy research. Geneva: International Labour Organization-IPEC. International Labour Organization. (2001). The time bound program in Nepal. Kathmandu, Nepal: International Labour Organization. International Labour Organization. (2002). Every child counts: New global estimates on child labour. Geneva: International Labour Organization-IPEC. KC, B. K., Adhikari, K. P., Subedi, G., & Gurun, Y. B. (2001a). Situation of child porters: A rapid assessment. Investigating the Worst Forms of Child Labour Series, No.6. ILO-IPEC, Geneva. KC, B. K., Gurung, Y. B., Adhikari, K. P., & Subedi, G. (2001b). Nepal, situation of child ragpickers: A rapid assessment. Investigating the Worst Forms of Child Labour, No.4. ILO-IPEC, Geneva. King, G., & Zeng, L. (2001). Logistic regression in rare events data. Political Analysis, 9(2), 137–163. Lancaster, T., & Imbens, G. (1996). Case control studies with contaminated controls. Journal of Econometrics, 71(1), 145–160. Manski, C. (2001). Nonparametric identification under response-based sampling. In: C. Hsiao, K. Morimune & J. Powel (Eds), Nonlinear statistical inference: Essays in honor of Takeshi Amemiya (pp. 241–258). New York: Cambridge University Press. Manski, C., & Lerman, R. (1977). The estimation of choice probabilities from choice-based samples. Econometrica, 45(8), 1977–1988. Mhukherjee, S. (2003). Child ragpickers in Nepal: A report on the 2002–2003 baseline survey. Kathmandu, Nepal: International Labour Organization. Prentice, R. L., & Pyke, R. (1979). Logistic disease incidence models and case-control studies. Biometrika, 66(3), 403–411. Rogers, C., & Swinnerton, K. (2008). A theory of exploitative child labor. Oxford Economic Papers, 60(1), 20–41. Tomz, M., King, G., & Zeng, L. (2003). ReLogit: Rare events logistic regression. Journal of Statistical Software, 8(2), 1–27.

HOUSEHOLD POVERTY AND CHILD LABOR DECISIONS IN MALAWI Levison S. Chiwaula ABSTRACT The positive relationship between household poverty and child labor decisions need not to be generalised across different types of works and geographical regions. This chapter studies this relationship using the 2004 Malawi Integrated Household Survey data. The study attempts to identify the influence of exogenous change in household consumption on child labor decisions by using consumer durable goods as an instrument. These findings show that child labor was most prevalent and intensive in domestic work, but significant negative relationships between household consumption and child labor supply are only found in unpaid market work. These findings support both poverty reduction and awareness campaigns as child labor eradication strategies. Promotion of non-labor intensive income sources also seems to be an attractive policy option.

Child Labor and the Transition Between School and Work Research in Labor Economics, Volume 31, 33–51 Copyright r 2010 by Emerald Group Publishing Limited All rights of reproduction in any form reserved ISSN: 0147-9121/doi:10.1108/S0147-9121(2010)0000031005

33

34

LEVISON S. CHIWAULA

1. INTRODUCTION Child labor research in recent years has focused on finding the determinants of child labor supply, with more emphasis on the role of poverty, fertility and liquidity constraints (Belletinni, Ceroni, & Ottavianow, 2005). Among the determinants that have been studied so far, poverty has emerged as the major determinant, and a positive relationship is generally expected. There are different theoretical explanations to the relationship between poverty and child labor. According to Edmonds (2008), the explanations for this negative connection between family incomes (positive connection between poverty) and child labor include (1) child labor being a ‘bad’ in parental preferences so that as incomes improve, the family chooses to have children work less; (2) the value of the marginal contribution of the child’s income decreases as family incomes increase due to diminishing marginal utility of income and (3) higher family incomes may facilitate the purchase of substitutes for child labor that lower the return to child labor within the household such as a washboard, fertilizer spreader or a combine. Although the positive relationship between child labor and poverty has been well explained theoretically, empirical research has not been conclusive. Positive relationships between poverty and child labor supply have been established in many cases (Blunch & Verner, 2000; Ray, 2000; Edmonds, 2005; Okupkpara & Odurukwe, 2006; Edmonds, 2006; Edmonds & Schady, 2008), and this has been rejected in some cases (Nielsen, 1998; Ray, 2000). Some studies have not established any relationship. Edmonds (2005) argued that insignificant relationships can be found if there are wrong assumptions such as assuming a linear relationship when the actual relationship is non-linear. More strange results have been found between child labor and household wealth in the form of land holding size. A positive relationship has been found in some studies (Bhalotra & Heady, 2003; Dumas, 2007), but the recent work by Basu, Das, and Dutta (2009) has explained this theoretically and empirically by showing that this is due to labor market imperfections. The conflicting research findings on the relationship between child labor and poverty may be due to a number of differences in these studies, such as definitional differences, methodological differences and spatial differences. Different results can be obtained when child labor is defined differently in different studies. Different definitions of child labor relate to types of work that are considered as child labor, the amount of time a child is involved in an activity and the definition of the child itself. For example, Edmonds (2005) found negative relationship between household expenditure and child labor in work outside the household and work in agriculture, but positive

Household Poverty and Child Labor Decisions in Malawi

35

relationship between child labor in family business and household expenditure implying that child labor in different activities relates to household poverty differently. Even when the definitions of child labor have been standardised, different results can still be obtained if different methodologies have been used. Most studies use parametric methods, but some studies, such as by Edmonds (2005), use non-parametric methods. Use of non-parametric methods allowed Edmonds (2005) to explore non-linear relationships, but the approach ignored some econometric issues such as endogeneity. Additionally, there is a cultural component associated with child labor decisions, and this means that the significance of poverty in explaining child labor decisions in different countries and cultural settings will differ. This brings the need for country-specific child labor studies and also studies that broaden the definition of child labor. Defining a child as a person who is at most 14 years old (see NSO, 2002), this chapter explores the relationship between child labor decisions and household poverty in Malawi. Malawi is one of the poorest countries in the world located in Southern Africa where it is estimated that about 1.4 million children between ages 5 and 14 years work (NSO, 2002). The country is highly agrarian and rural with about 80% of the population living in rural areas and depending on agriculture. The rest of the chapter progresses as follows. Section 2 presents data issues and variable description, while Section 3 presents the empirical strategy. Empirical results are presented in Section 4, and Section 5 concludes the chapter.

2. DATA Data used is from the 2004 Malawi Integrated Household Survey that was collected by the Malawi National Statistical Office from March 2004 to April 2005. The survey collected data from a nationally representative sample of 11,280 households, and it was designed to cover a wide array of subject matter, with primary objective of providing a complete and integrated data set to better understand the socioeconomic status of the population in Malawi (MEPD, NSO, and World Bank, 2005). A sample of children aged between 5 and 14 years was extracted from this sample. Children who were employed as house servants in the interviewed households were excluded from the sample because the poverty of the households where they came from, which was not captured in the study, is more important for them. The final sample consisted of 7,930 children.

36

LEVISON S. CHIWAULA

The dependent variables were generated from the module on time use and labor of the Integrated Household Survey questionnaire. Household members were asked the question: ‘How many hours in the last seven days did you do (insert name of activity)?’ Recalls for household chores were for the previous 24 hours because these are normally done on daily basis and a 24-hour recall would give a more precise estimate. The dependent variable was categorised into (1) total child labor supply; (2) child labor supply in domestic work that includes cooking, cleaning, washing, collecting water and collecting firewood and (3) child labor supply in unpaid market work that includes work in agriculture, livestock, fishing activities and family business. Although child labor in paid employment may be interesting to be looked at separately, the incidence of child labor in this type of work was very low in the sample and I decided not to estimate separate models for this activity. Explanatory variables included control variables for child, household and community characteristics. Child age and its square, sex of the child and relationship of the child to the household head (whether the child is a biological child of the household head or not) were used to control for child characteristics. Number of male and female adult (more than 14 years) members, number of underfive-aged children, age of the household head, age of the mother and sex of the household head were used to control for household demographic characteristics that also reflect the intrahousehold labor supply situation. Additionally, dummy variables for the highest educational attainment for the household head (no education and primary education) were used to control for household level characteristics. Community level characteristics were controlled for by the inclusion of dummy variables of two of the three major administrative/political regions (south and north), distance to the nearest primary school and number of school days in the previous two weeks. Number of school days in the previous two weeks was also used to assess the competition on child time between schooling and child labor. Considering the fact that some of the activities that are done by the children are seasonal and that the households were interviewed at different times of the year, I included a dummy variable that controlled for seasonality of the activities, which took the value 1 if the household was interviewed during the rainy season and 0 otherwise. The descriptive statistics of the variables used are presented in Table 1. The descriptive statistics show that about 54% of the children worked at least in one activity during the survey period. It is also shown that incidence of child labor is almost double in domestic work than in unpaid market work, while incidence in paid employment is just in 3% of the sample.

37

Household Poverty and Child Labor Decisions in Malawi

Table 1.

Descriptive Statistics.

Variable

Child Child Child Child

labor labor labor labor

Mean

incidence incidence incidence incidence

Unconditional Unconditional Unconditional Unconditional Conditional Conditional Conditional Conditional

child child child child

child child child child

in in in in

all work activities (0/1) paid employment (0/1) domestic work (0/1) unpaid market work(0/1)

labor labor labor labor

labor labor labor labor

supply supply supply supply

supply supply supply supply

in in in in

Consumption expenditure (MK) Land holding (ha) Child age (years) Child sex (0/1) Own child (0/1) Household size Male adults Female adults Infants Head sex (0/1) Head age (years) Mother age (years) Mother uneducated (0/1) Head uneducated (0/1) Head primary education (0/1) Distance to primary school (km) School days (days) Southern region (0/1) North region (0/1) Season (0/1) Rural (0/1) N

in in in in

all work activities (hours) paid employment (hours) domestic work (hours) unpaid market work (hours)

all work activities (hours) paid employment (hours) domestic work (hours) unpaid market work(hours)

Standard Deviation

0.54 0.03 0.43 0.26

0.50 0.17 0.49 0.44

7.70 0.33 4.85 2.52

11.88 2.95 8.59 6.21

14.38 11.83 11.34 9.60

12.94 13.18 9.95 8.88

17735.37 0.45 9.19 0.50 0.76 6.78 1.44 1.54 1.27 0.76 45.05 28.81 0.38 0.50 0.11 1.55 7.48 0.43 0.17 0.48 0.90

21683.21 1.28 2.85 0.50 0.43 2.57 1.05 0.85 1.02 0.43 13.79 18.71 0.49 0.50 0.31 2.86 3.65 0.50 0.37 0.50 0.30 7,930

On average, the total unconditional child labor supply is about 7.70 hours in week, and children were observed to work for more hours in domestic work than in other types of work. When I considered children who are working only, that is, conditional child labor supply, the statistics show that child

38

LEVISON S. CHIWAULA

labor supply in different activities rose by more than 100%. This shows that a great difference exists between working children and non-working children. I find that the average total child labor hours in a week rose to 14.38. The highest conditional child labor hours worked was observed in paid employment and the lowest was observed in unpaid market work. This suggests that the children who are working in paid employment, although few, are the most exploited in terms of number of hours worked. Descriptive statistics for key explanatory variables show that the average per capita consumption expenditure per year was just MK 17,735.37. This is slightly above the consumption poverty line derived from the same data set by the Malawi Government and the World Bank.

3. EMPIRICAL STRATEGY 3.1. Empirical Models The dependent variables are number of child labor hours in different activities, which contain a large proportion of observations with zero hours worked. Standard linear regression models based on ordinary least square (OLS) estimation technique would yield biased as well as inconsistent parameters for such data types. Tobit models (Tobin, 1958) have been widely used for data that have such characteristics. The tobit models assume that all zeros are attributable to standard corner solutions. Negative values of the dependent variables are assumed to exist but are considered to be unobservable and bunched at zero (Keelan, Newman, & Henchion, 2008). Defining H as the latent variable for the hours of child labor supplied, the Tobit model is specified as follows: H ij ¼ x0i b þ i , H ij ¼ 0 H ij ¼ H ij

if H ij  0 if H ij 40

ð1Þ

i  Nð0; s2 Þ where Hij denotes number of child labor hours for child i in activity j; x0i denotes a vector of explanatory variables; b is a vector of parameters that were estimated and ei is the normally distributed error term that is assumed to have a zero mean and constant variance. The likelihood function used for

Household Poverty and Child Labor Decisions in Malawi

the estimation of tobit models is given as follows:  0  Y    Y xb H ij  x0i b 1F i s1 f Lðb; s2 Þ ¼ s s H ¼0 H ¼1 ij

39

(2)

ij

where F denotes the standard normal cumulative distribution function and f denotes the standard normal density function. To obtain consistent estimates from the maximisation of this likelihood function, there is need to correctly specify the model. Specification analysis of Tobit model is an important consideration because, unlike the standard regression model, either heteroskedasticity or non-normality can render the maximum likelihood parameter estimates inconsistent (Reynolds & Shonkwiler, 1991; Keelan et al., 2008). I relaxed the homoskedasticity assumption by modelling the variance as follows: si ¼ expðzi dÞ

(3)

where zi is a vector of explanatory variables that are a subset of xi and d is a vector of coefficients. Transformation of the dependent variable is one of the options for fixing the non-normality problem. I used the inverse hyperbolic sine (IHS) transformation, I(.), to transform the dependent variable. This transformation has been used in many studies (Reynolds & Shonkwiler, 1991; Sinning, 2007; Keelan et al., 2008) and it takes the form as follows: IðH ij Þ ¼ g1 ðgH ij þ ðg2 H 2ij þ 1Þ1=2

(4)

where g is a parameter that can be estimated from the data. The IHS is symmetric about 0 in g, the limit of the I(Hij) as g-0, and for relatively large values of g, the transformation behaves logarithmically (Burbidge, Magee, & Robb, 1988). I assumed that g ¼ 1 because it is mostly assumed so in empirical analysis (Sinning, 2007). The IHS transformation is continuously defined over positive, zero and negative values unlike the Box– Cox transformation, T(.), which does not result in normal distribution of the residuals because the transformation is not defined for H ij o0 (Maddala, 1983; Reynolds & Shonkwiler, 1991). The likelihood function of the heteroskedastic IHS tobit model is therefore given as follows:  0  Y    1=2 Y xij b IðH ij Þ  x0ij b  2 1 2 2 1 þ g H ij 1F si f Lðb; s Þ ¼ sij sij H ¼0 H ¼1 ij

ij

(5)

40

LEVISON S. CHIWAULA

The maximum likelihood estimates of this log-likelihood function were obtained in Stata. This was achieved by incorporating the IHS transformation to the Stata command written to estimate heteroskedastic tobit models.1 To assess the relationship between incidence of child labor and poverty, I used heteroskedastic probit models. Probit estimations were straightforward by using the hetprob command in Stata. Same explanatory variables used in the tobit models were used in the probit models.

3.2. Model Estimation One of the estimation challenges in this study was the endogeneity of household per capita consumption expenditure that was used as an indicator of household poverty. If this problem is not controlled, the use of standard regression techniques to investigate this relationship would likely overestimate the effect of household consumption expenditure on child labor. Empirical estimations need to identify the exogenous variations in household consumption expenditure and relate this with child labor. There are two approaches researchers mostly use to handle this problem in child labor studies. Firstly, some studies address the problem by relating child labor to variation in income that excludes the child’s income (Ray, 2000; Duryea & Arends-Kuenning, 2003). While this technique addresses the mechanical source of endogeneity, it does not deal with the joint nature of child time allocation and family living standards (Edmonds, 2008). Additionally, this is only possible when we are considering child labor in paid work. In cases where child labor includes working in unpaid work such as household chores or work in family enterprises, this approach is not adequate because it will not value the indirect income the child is contributing to household income by freeing adults’ time to other paid work. This approach is also not possible in this case because I used household consumption expenditure and not income as an indicator of household living standards. The second approach is more appropriate because it handles the broader endogeneity problem by employing the instrumental variables (IVs) estimation technique (see Bhalotra & Heady, 2003; Ersado, 2005). The implementation of this approach however depends on the availability of good instruments that are correlated to the endogenous variable but not the dependent variable. This study used the monetary value of consumer durable goods2 as an instrument for consumption expenditure. Consumer durable goods are used as an instrument for consumption expenditure firstly because it is assumed that these

Household Poverty and Child Labor Decisions in Malawi

41

goods are not directly used for household production activities, which mean that these goods do not affect the shadow wages of child labor thereby having no relationship with child labor decisions. Secondly, I look at consumer durable goods as a proxy for household permanent income, which is expected to be highly correlated with consumption expenditure. As such, consumer durable goods are expected to be highly correlated with consumption expenditure. Another argument for using consumer durable goods is that most of the consumer durable goods owned by the households were not purchased during the survey period such that none of the income from child labor was used to purchase these goods. In this case, the instrument would break any form of reverse causality between child labor supply and household consumption, which is one of the potential causes of endogeneity between child labor and poverty. There are however some cases where the value of consumer durable goods may not be good IVs for consumption expenditure. For example, some of the consumer durable goods such as tape recorders, televisions and VCRs may make leisure more valuable (i.e. lowering shadow wages of child labor). Additionally, these types of consumer durable goods can make working at home more attractive. These cases would make consumer durable goods directly related to child labor decisions. However, I feel that this case may be less prevalent in this sample because the sample is predominantly rural and poor where these types of goods are rarely possessed. Another case where consumer durable goods may not prove to be a good instrument can be in instances when past shocks affect the holdings of consumer durable goods and child time allocation because transitory income shocks have been found to lead to an increase in child labor (see Beegle, Dehejia, & Gatti, 2006). The joint effects of past shocks on consumer durable goods holdings and child time allocation will result in the correlation between the instrument and the dependent variable. The IV approach of estimating the models followed a two-step procedure that was proposed by Smith and Blundell (1986) and can be used with both tobit and probit models (see Smith & Blundell, 1986; Rivers & Vuong, 1988; Wooldridge, 2002). In the first stage, a reduced form equation of the per capita consumption is regressed on all the exogenous variables and the IV using OLS from which the residuals are obtained. The second step estimates the standard probit or tobit models of the child labor supply equations in which the residuals from the first step are included as one of the explanatory variables. The usual t-statistic on the residual term in the second step models provides a simple test of the null that consumption expenditure is exogenous. Significance of the parameter estimates on the residual term in

42

LEVISON S. CHIWAULA

the child labor supply equations confirms and controls the presence of endogeneity. The Smith–Blundell approach is valid without any distributional assumptions on the reduced form of consumption expenditure (Wooldridge, 2002).

4. EMPIRICAL RESULTS I first present the results of the reduced form regression of the per capita consumption expenditure on all explanatory variables plus the IV. This is the first step of the two-step procedure. The probit and tobit models share the same first step results. The results are presented in Table 2. The results show that the IV is highly correlated with the endogenous variable with a very high t-statistic. I further checked the correlations between the dependent variables and the instrument by including the instrument in standard probit and tobit models, and I found that the instrument is not significantly correlated to any of the dependent variables. This suggests that the instrument is good enough to identify the variation in consumption expenditure that is exogenous to child labor supply decisions because it is meeting the exclusion restriction. The residuals from this regression were included in the child labor supply models to test and control for endogeneity of consumption expenditure. I also tested for the normality of error terms in the IV tobit models using the conditional moment test for normality that was derived by Skeels and Vella (1999) using an already available Stata command that implements the bootstrap method described by Drukker (2002). The results of the normality test are presented in Table 3. The results of the normality test reject the normality assumption in all the estimated models. I therefore applied the IHS transformation, Eq. (4), to the dependent variables to fix this problem. The log-likelihood function for the tobit model presented by Eq. (5) was then used to estimate the tobit models. When estimating, the standard errors were adjusted for clustered sampling, but standard errors in the probit models were not adjusted because the Stata command used to estimate the heteroskedastic probit did not allow the cluster sampling option. A total of 81 enumeration areas that were the primary sampling units (PSU) during the Integrated Household Survey were used as clustering units. The transformation of the dependent variables in the tobit model and the implementation of the two-step estimation procedures complicate the interpretation of the parameter estimates and the computational of the

43

Household Poverty and Child Labor Decisions in Malawi

Table 2.

Reduced Form Regression of Per Capita Consumption Expenditure.

Variables

Coefficient

t-Statistic

Land holding Land holding squared Child age Child age squared Child sex Own child Household size Male adults Female adults Infants Head sex Head age Mother age Mother no education Head no education Head primary education Primary school School days Southern region North region Season Rural Consumption asset Constant

2,624.32 44.58 305.29 18.23 604.49 1,512.87 1,475.37 2,667.14 1,495.84 111.98 4,741.25 127.72 9.49 2,884.19 3,646.58 2,844.70 152.15 70.27 6,770.01 3,839.56 2,393.56 20,861.31 0.08 49,754.05

7.43 6.66 0.54 0.62 1.44 2.74 8.74 8.73 4.30 0.38 5.79 6.39 0.49 5.88 7.79 3.79 2.03 1.19 14.31 6.12 5.62 28.72 27.32 16.91

R2 F-statistic N

0.26 122.11 7,930

Significance at 1%. Significance at 5%. Significance at 10%.

Table 3. Child labor Supply Domestic work Unpaid market work Total labor Significance at 1%. Significance at 5%. Significance at 10%.

Conditional Moment Test for Normality. Log-Likelihood

Chi-Square Statistic

15,959.12 10,644.09 19,991.12

146.43 82.38 223.98

44

LEVISON S. CHIWAULA

marginal effects and their standard errors (see Reynolds & Shonkwiler, 1991; Wooldridge, 2002). The second step standard errors and related statistics are incorrect because they fail to account for the fact that unobservable regressors have been estimated in calculating second step coefficients and standard errors (Murphy & Topel, 1985), and these are supposed to be corrected. However, Smith and Blundell (1986) stated that under weak exogeneity hypothesis, which is this case as can be seen from estimation results below, the asymptotic covariance matrices collapses to standard tobit covariance matrices, which imply that the estimators from the two-step tobit provides required statistics. I therefore interpreted the test statistics as they are in the second step models, and I did not compute marginal effects. As the result, I only interpreted the direction and significance level of the parameters. Results of the total child labor supply models are presented in Table 4. The table presents results of an IHS heteroskedastic tobit model and a heteroskedastic probit model. When homoskedasticity was assumed, the estimation results had larger log-likelihoods (21,312.77 for the tobit model and 4,513.90 for probit model), which implies that relaxing the homoskedasticity assumption improves the models. The log-likelihood in this model is higher than the one I found when I was testing for normality (see Table 3) because the standard errors in those equations were not adjusted to clustered sampling. If the IV tobit command is used together with the clustering option, the log-likelihood is much higher than the estimate in this model. This therefore suggests that the specification I used is better than a number of possible specifications. The 10% significance level of the residual term suggests that consumption expenditure is weakly exogenous in the tobit model and exogenous in the probit model. When exogeneity of consumption expenditure was assumed, the parameter estimate on consumption expenditure was insignificant. However, the current specification that corrected for endogeneity shows that increase in consumption expenditure significantly reduces total child labor supply. However, the results show that consumption expenditure does not significantly influence the probability that children will work at least in one type of work. Disaggregating child labor to household domestic work and unpaid market work, I find slightly different results, which are presented in Table 5. Estimation results suggest that consumption expenditure is exogenous in these models. However, dropping the residual term from the model resulted in some changes in both the standard errors and the parameters. In general, the z-statistics for the parameter estimates on consumption expenditure

Significance at 1%. Significance at 5%. Significance at 10%.

Log-likelihood Wald chi-square N

Consumption Land holding Land holding squared Child age Child age squared Child sex Own child Household size Male adults Female adults Infants Head sex Head age Mother age Mother no education Head no education Head primary education Primary school School days Southern region North region Season Rural Residual Constant

Variables

0.0000 0.0643 0.0279 0.4898 0.0139 0.4927 0.0400 0.0361 0.0157 0.0286 0.0517 0.0456 0.0051 0.0013 0.0272 0.0712 0.0618 0.0052 0.0351 0.2723 0.2274 0.0808 0.0755 0.0000 2.4124

0.92 0.65 0.79 7.36 4.66 4.67 0.90 1.91 0.53 0.98 1.99 0.63 2.48 0.83 0.60 1.50 0.96 0.86 4.26 2.93 2.86 1.87 0.49 1.01 5.83

z-statistic

Probit

0.22

1.69

0.0044

0.0174

4507.53 69.97 7,930

1.99 2.35

z-statistic

0.0000 0.1142

Coefficient

Heteroskedasticity

0.0000 0.0447 0.0012 1.4742 0.0508 0.8985 0.0542 0.0307 0.0421 0.1047 0.0840 0.0746 0.0077 0.0032 0.0545 0.1422 0.1110 0.0045 0.0630 0.2759 0.3541 0.0742 0.2478 0.0000 7.5052

Coefficient z-statistic 3.04 0.78 1.02 16.38 12.53 14.60 0.73 1.12 0.91 2.03 1.77 0.80 2.41 1.13 0.66 2.04 1.11 0.27 4.63 2.14 2.33 0.62 1.02 1.86 11.47

Tobit

21165.23 1076.20 7,930

1.4403

0.0015

0.0808

0.0000 0.0052

Coefficient

20.60

0.21

19.33

4.24 0.71

z-statistic

Heteroskedasticity

Probit and Tobit Estimation of Total Child labor Supply Models.

Coefficient

Table 4. Household Poverty and Child Labor Decisions in Malawi 45

Consumption Land holding Land holding squared Child age Child age squared Child sex Own child Household size Male adults Female adults Infants Head sex Head age Mother age Mother no education Head no education

Variables

Table 5.

0.0000 0.0060 0.0001 0.2989 0.0112 0.3819 0.0021 0.0012 0.0155 0.0420 0.0027 0.0212 0.0014 0.0008 0.0370 0.0328

Coefficient 0.77 0.74 0.57 8.44 7.90 6.69 0.12 0.22 1.53 3.38 0.29 0.77 2.07 1.30 2.16 1.96

z-statistic

Probit

0.0000 0.0305 0.0003 1.4023 0.0491 1.8147 0.0386 0.0175 0.0675 0.2286 0.0104 0.0678 0.0069 0.0050 0.1684 0.1433 1.42 0.78 0.49 14.89 11.56 19.56 0.48 0.61 1.38 3.96 0.20 0.61 2.03 1.57 1.65 1.88

z-statistic

Tobit Coefficient

Domestic Work

0.0000 0.1442 0.0066 0.3554 0.0110 0.1064 0.0406 0.0002 0.0140 0.0161 0.0236 0.0435 0.0036 0.0011 0.0530 0.0153

Coefficient

Coefficient 0.0001 0.4452 0.0080 1.5634 0.0490 0.3835 0.1333 0.0435 0.0457 0.0015 0.0916 0.2360 0.0088 0.0028 0.2562 0.1263

1.81 2.66 0.80 8.23 5.63 3.56 1.18 0.01 0.46 0.68 1.39 0.64 1.95 1.00 1.27 0.34

1.83 2.31 1.90 7.30 5.58 3.94 1.00 0.53 0.29 0.01 1.20 0.79 1.09 0.61 1.12 0.53

z-statistic

Tobit

z-statistic

Probit

Unpaid Market Work

Probit and Tobit Estimation Results for Child labor Supply in Domestic Work and Unpaid Market Work.

46 LEVISON S. CHIWAULA

4438.18 96.67

0.0192 0.0043 0.0124 0.0589 0.1675 0.0392 0.0036 0.0000 1.4268

0.0909 0.0136 0.0514 0.1635 0.7102 0.2110 0.0562 0.0000 6.6878

17179.91 1370.60 7,930

0.78 1.73 4.76 2.84 5.05 2.62 0.10 0.27 8.33

0.71 0.73 3.45 1.19 3.90 1.63 0.19 0.91 9.18

Note: Results of the variance equation are not presented due to space limitations. Significance at 1%. Significance at 5%. Significance at 10%.

Log-likelihood Wald chi-square N

Head primary education Primary school School days Southern region North region Season Rural Residual Constant 3809.10 129.69

0.0073 0.0063 0.0186 0.2914 0.0896 0.2711 0.2946 0.0000 2.9313 0.0123 0.0413 0.0812 1.1442 0.0039 0.9680 0.6843 0.0001 11.2188

11178.10 667.08 7,930

0.14 1.40 4.03 3.00 1.84 4.30 1.38 1.69 5.56 0.05 1.06 3.62 2.95 0.01 4.34 0.56 1.62 3.60

Household Poverty and Child Labor Decisions in Malawi 47

48

LEVISON S. CHIWAULA

increases when the residual term is included in the models although they are not statistically significant. I therefore decided to maintain them in the models even though the exogeneity has been confirmed. As expected, the results from the models are different from one another and also from the total child labor supply models. For child labor in domestic work, the results show that consumption expenditure does not significantly influence the probability of working and the amount of child labor hours supplied. On the contrary, the estimates in the child labor supply in unpaid market work shows that consumption expenditure significantly reduces both the probability of working and the amount of child labor supply. This suggests that poverty is an important determinant of child labor in unpaid market work. Child labor decisions in domestic work do not necessarily respond to household poverty level. These results are similar to the findings of Edmonds and Schady (2008) that showed that child labor in market work declines, but domestic work increases with the increase in income. The general message here is that household poverty and child work in domestic work do not relate in the famous positive fashion. Apart from the parameters on poverty indicators which were the primary purpose of this study, I also find other results worth of a discussion. For example, land holding size is found to have a significant quadratic relationship with the probability and amount of child labor supply in unpaid market work. Although the parameter estimate on the square of land holding size is insignificant in the probit model, a joint significance test for land holding and its square shows that the two are jointly significant. This means that the two variables can be interpreted together although one is individually insignificant. Another important result is that on the number of school days in the previous two weeks, which has been found to be negatively correlated to child labor. Similarly, distance to the nearest primary school is negatively related to child labor in domestic work. These results imply that when the need for schooling increases, the probability and intensity of child labor reduces significantly. This concurs with Edmonds (2006) who found simultaneous increases in schooling and decreases in child labor hours as a response to cash transfers to the elderly in South Africa. Similarly, Edmonds, Pavcnik, and Topalova (2007) found that schooling costs play an important role in the relationship between poverty, child labor and schooling, which relates to the current result on distance to the nearest primary school. With these results, it can be speculated that parents prioritise child schooling to child labor as ‘good’. Parameter estimates on child sex dummy in the total child labor supply models suggest that girls work more than boys, but the estimates in the child labor supply models in

Household Poverty and Child Labor Decisions in Malawi

49

domestic work and unpaid market work show that significant gender differences exist in child labor decisions. Girls are found to work more in domestic work while boys work more in unpaid market work. These differences largely follow social norms in Malawi. The results also show that the number of female adults in the household reduces child labor while increases in number of infants increases child labor. This suggests that child labor decisions are also dependent on intrahousehold labor supply and demand status.

5. CONCLUSIONS The study was conducted to attempt to understand the relationship between child labor and household poverty in Malawi. The results confirm that child labor emerges from household poverty. Although the incidence and intensity of child labor is higher in domestic work than in unpaid market work, the study finds no significant relationship between child labor decisions in domestic work and poverty, but a significant relationship is found between household poverty and child labor in unpaid market work. From these findings, it can be concluded that household income poverty results in child labor activities that can directly generate income. In activities that do not directly generate income such as domestic work, child labor is mostly induced by other factors. In principle, child labor in domestic work is expected to be induced by more income up to a certain extent, because for the very poor households, there is no food to cook, any dishes and clothes to wash. A number of policy directions for Malawi and other developing countries can thus be drawn from these results. Generally, the findings from this study accept child labor eradication interventions that aim at reducing income poverty. In case of child labor in domestic work, interventions that aim at creating awareness seem to be more appropriate. It should still be made clear that child labor in domestic work in developing countries can be reduced but not eradicated because this seems not to be due to poverty. The positive linear relationship between child labor and land holding size that was found in earlier studies (Bhalotra & Heady, 2003; Dumas, 2007) and the quadratic relationship between child labor and land holding size that was found by Basu et al. (2009) and also in this study is important but not conclusive. According to Basu et al. (2009), the quadratic relationship shows that labor markets are imperfect and that increase in household assets increases working opportunities for children. I feel that there is still need for

50

LEVISON S. CHIWAULA

further research in this area because Basu et al. (2009)’s work does not empirically show that the relationship would be negative if labor markets are perfect.

NOTES 1. The program was supplied by Maarten L. Buis from Vrije Universiteit Amsterdam on Statalist. 2. Consumer durable goods that made up this instrument include beds, tables, chairs, air conditioners, radio, tape or CD player, and television and VCR.

ACKNOWLEDGMENTS This work was supported by an African Economic Research Consortium (AERC) research grant. Resource persons of the poverty group of the AERC research network are acknowledged for their valuable comments at different stages of this study. Comments from two anonymous referees and the editors are also greatly appreciated. However, the views expressed in this chapter do not reflect AERC’s views but mine.

REFERENCES Basu, K., Das, S., & Dutta, B. (2009). Child labour and household wealth: Theory and empirical evidence of an inverted U. Journal of Development Economics, 91(1), 8–14. Beegle, K., Dehejia, R. H., & Gatti, R. (2006). Child labour and agricultural shocks. Journal of Development Economics, 81, 80–96. Belletinni, G., Ceroni, B. C., & Ottavianow, G. I. P. (2005). Child labour and resistance to change. Economica, 72, 397–411. Bhalotra, S., & Heady, C. (2003). Child farm labour: The wealth paradox. The World Bank Economic Review, 17, 197–227. Blunch, N. H., & Verner, D. (2000). Revisiting the link between poverty and child labour: The Ghanaian experience. The World Bank Policy Research Working Paper no. 2488. The World Bank, Washington DC, 21pp. Burbidge, J., Magee, L., & Robb, L. (1988). Alternative transformations to handle extreme values of the dependent variable. Journal of American Statistical Association, 83, 123–127. Drukker, D. M. (2002). Bootstrapping a conditional moments test for normality after tobit estimation. Stata Journal, 2(2), 125–139. Dumas, C. (2007). Why do parents make their children work? A test of the poverty hypothesis in rural Burkina Faso. Oxford Economic Papers, 59, 301–329. Duryea, S., & Arends-Kuenning, M. (2003). School attendance, child labour and local labour market fluctuations in urban Brazil. World Development, 31, 1165–1178.

Household Poverty and Child Labor Decisions in Malawi

51

Edmonds, E. V. (2005). Does child labour decline with improving economic status? Journal of Human Resources, 40, 77–99. Edmonds, E. V. (2006). Child labour and schooling responses to anticipated income in South Africa. Journal of Development Economics, 81(2), 386–414. Edmonds, E. V. (2008). Child labour. In: T. P. Schultz & J. Strauss (Eds), Handbook of development economics (Vol. 4, pp. 3607–3709). Amsterdam: Elsevier, North-Holland. Edmonds, E. V., Pavcnik, N., & Topalova, P. (2007). Trade adjustment and human capital investments from Indian tariff reforms. NBER Working Paper no. 12884, 57pp. Edmonds, E. V., & Schady, N. (2008). Poverty alleviation and child labour. The World Bank Policy Research Working Paper no. 4702, 35pp. Ersado, L. (2005). Child labour and school decisions in urban and rural areas: Comparative evidence from Nepal, Peru, and Zimbabwe. World Development, 33, 455–480. Keelan, C., Newman, C., & Henchion, M. (2008). Quick-service expenditure in Ireland: Parametric vs. nonparametric analysis. Applied Economics, 40, 2659–2669. Maddala, G. S. (1983). Limited dependent and qualitative variables in econometrics. Cambridge: Cambridge University Press. MEPD, NSO, and World Bank. (2005). Second integrated household survey: An extract of findings. Ministry of Economic Planning and Development, National Statistical Office, and The World bank, Lilongwe, Malawi, 29pp. Murphy, K. M., & Topel, R. H. (1985). Estimation and inference in two-step econometric models. Journal of Business and Economic Statistics, 3(4), 370–379. Nielsen, H. S. (1998). Child labour and school attendance: Two joint decisions. CLS-WP 98-15, Centre for Labour Market and Social Research, Aarhus, Denmark, 42pp. NSO. (2002). Malawi child labour survey 2002: Report of analysis. Malawi National Statistical Office (NSO) and International Labour Organization, Zomba, Malawi, 112pp. Okupkpara, B. C., & Odurukwe, N. (2006). Incidence and determinants of child labour in Nigeria: Implications for poverty alleviation. AERC Research Paper no. 156. African Economic Research Consortium, Nairobi, Kenya, 51pp. Ray, R. (2000). Child labour, child schooling, and their Interaction with adult Labour: Empirical evidence for Peru and Pakistan. The World Bank Economic Review, 14(2), 347–367. Reynolds, A., & Shonkwiler, J. S. (1991). Testing and correcting for distributional misspecifications in the tobit model: An application of the information matrix test. Empirical Economics, 16, 313–323. Rivers, D., & Vuong, Q. H. (1988). Limited information estimators and exogeneity tests for simultaneous probit models. Journal of Econometrics, 39, 347–366. Sinning, M. (2007). Determinants of savings and remittances: Empirical evidence from immigrants to Germany. IZA Discussion Paper no. 2966, 32pp. Skeels, C. L., & Vella, F. (1999). A Monte Carlo investigation of the sampling behavior of conditional moment tests in tobit and probit models. Journal of Econometrics, 92, 275–294. Smith, R. J., & Blundell, R. W. (1986). An exogeneity test for a simultaneous tobit model with an application to labour supply. Econometrica, 54(3), 679–685. Tobin, J. (1958). Estimation of relationships for limited dependent variables. Econometrica, 26, 24–36. Wooldridge, J. M. (2002). Econometric analysis of cross section and panel data. Cambridge, MA: MIT Press.

HOW MUCH WORK IS TOO MUCH? EFFECTS OF CHILD WORK HOURS ON SCHOOLING – THE CASE OF EGYPT Ragui Assaad, Deborah Levison and Hai-Anh Dang ABSTRACT How much work is ‘‘too much’’ for children aged 10–14 in Egypt? Our narrow focus here is on ‘‘work that does not interfere with school attendance.’’ For girls, work includes time spent in household chores and subsistence activities. We estimate simultaneous hours of work and school attendance equations as a joint Tobit and Probit model, then conduct simulations. Substantial negative effects on attendance are observed above about 10 hours per week (girls) and 14 hours (boys). For girls, heavy household work appears causal, but for boys, it seems that poor schooling leads to boys’ dropout, then subsequent work.

INTRODUCTION How much work is ‘‘too much’’ for children? In this chapter, we examine how many hours of work can be undertaken before negative effects on school attendance are observed, using 1998 data from Egypt. Child Labor and the Transition Between School and Work Research in Labor Economics, Volume 31, 53–97 Copyright r 2010 by Emerald Group Publishing Limited All rights of reproduction in any form reserved ISSN: 0147-9121/doi:10.1108/S0147-9121(2010)0000031006

53

54

RAGUI ASSAAD ET AL.

There is general agreement that children should not be doing any work that is clearly harmful, hazardous, or morally objectionable, as evidenced by the rapid rate of ratification of the International Labour Organization’s (1999) Convention C182 on the ‘‘worst forms’’ of child labor. There is less agreement about work that is not so clearly problematic – and these disagreements can be found among policy makers, child advocates, and analysts. Some believe that children should not work at all, whereas others believe that work in moderation can be helpful in developing skills, confidence, and good habits. Part of the problem arises in the definition of ‘‘in moderation,’’ which is both subjective and context-specific. Our narrow interpretation of ‘‘work in moderation,’’ for the purposes of this chapter, is ‘‘work that does not interfere with school attendance.’’ Large proportions of adolescents in industrialized countries are employed in labor force work (e.g., Mortimer, 2003; White, 1994). Although many U.S. experts view work over 20 hours per week as deleterious for an adolescent’s education, Mortimer (2003) finds that a minority of youth benefit from longer work hours. She also documents positive aspects of parttime (o20 hours/week) work. In industrialized countries, although some children and youth work ‘‘under the table,’’ the majority appear to be in formal sector employment, and thus subject to labor regulations. In contrast, in developing countries, most of children’s work takes place outside the formal employment sector. Much of it is found in the informal economy and, for girls, in the home. The degree to which children’s work interferes with school attendance can vary greatly, depending on the institutional structure of the sector of work and also depending on the structure of the school day. In some areas, schools function in two or three shifts per day, and it may be possible for a child to fit a substantial amount of informal-sector or domestic work around a 4-hour school day. In a previous paper (Assaad, Levison, & Zibani, 2010), we jointly estimated Egyptian girls’ participation in work and attendance at school. We counted girls’ domestic chores among girls’ work,1 but girls who worked less than 14 hours per week were included in the ‘‘not working’’ group. That is, we were interested in whether substantial hours of work, broadly defined, affected schooling. The standard labor force definition whereby a person is employed if he/she worked at least one hour in the reference week did not seem useful in this case. Using this binary definition of work, we established that working for 14þ hours per week has a negative impact on schooling for girls, when girls’ work included time spent on household tasks. These results did not inform us, however, about the effect of an additional hour of work on the

How Much Work Is Too Much?

55

probability of school attendance, nor did they let us examine the possibility that there are thresholds – in terms of numbers of weekly hours worked – beyond which work interferes substantially with school attendance. In this chapter, we jointly estimate hours worked and school attendance using methods that will allow us to examine how the probability of attending school varies with the number of hours worked by 10–14-yearolds in Egypt. We use instrumental variables that explain the variation in work hours but that are assumed not to directly affect schooling. To identify the impact of market work on schooling, which is mostly relevant for boys, we use instruments that indicate the prevalence of child-friendly occupations in the local community, such as the prevalence of agriculture, craft, service, and trade occupations. Since most children are likely to work close to home, they will be most affected by labor demand conditions in their own village or neighborhood of residence. Although it is possible that a high prevalence of manual occupations in a community may depress the returns to education and therefore also affect schooling, we rule that out by arguing that households in Egypt are much more likely to make decisions on education on the basis of regional and national rather than local returns to education. To identify the impact of domestic work on schooling, which is mostly relevant for girls, we use instruments that proxy for the increase in the domestic work burden that results from poor access to basic public services such as piped water, sewage disposal, or garbage collection either at the household or at community levels. Again, it is possible that variables on the quality of water and sanitation services are correlated with potentially omitted variables relating to the availability and quality of schools at the local level and, therefore, should not be excluded from the schooling equation. To minimize this possibility, we include in both the work and the schooling equations local-level controls indicating levels of access to schools for the relevant age and sex groups and school quality indicators such as student–teacher ratios and the percentage of schools with multiple shifts. Finally, to empirically test our exclusion restrictions, we conduct overidentification tests to ensure that the exclusion of the instruments from the schooling equations is supported empirically.

LITERATURE REVIEW There is a vast literature on children’s education in developing countries, but until fairly recently only a small minority of researchers with interests in

56

RAGUI ASSAAD ET AL.

child schooling has recognized or been in a position to take account of the interrelationships between children’s multiple activities. Rosenzweig and Evenson (1977), Rosenzweig (1978), and DeTray (1983) were among the pioneers in this area, insofar as they recognized the potential trade-offs between children’s labor force participation and their school attendance. They were also among the first to analyze children’s hours of market work in conjunction with their schooling. The studies by Rosenzweig and Evenson (1977) and Rosenzweig (1978) are, however, focused on the economic value of children to their parents and the implications of the value of children for fertility. DeTray also writes for the discussion on the economic value of children of that time period but frames his results broadly; he specifically mentions, for example, possible effects of children’s household work on their time in school. The 1990s saw a boom in academic studies of child work, often analyzed jointly with the school attendance or enrollment of the children. (See Edmonds, 2008; Orazem & Gunnarsson, 2003; Bhalotra & Tzannatos, 2003 for recent reviews.) Still, most of these studies have been limited in two ways that the current analysis is not. First, most have focused on labor force participation, while neglecting measures of intensity of labor force work, such as hours worked. This study attempts to determine whether the negative effects of hours worked on schooling strengthens after a certain numbers of hours of work. Second, most used traditional definitions of ‘‘work’’ and thus focused on labor force participation.2 In contrast, we argue that the boundaries for productive activities as defined by the International Conference of Labour Statisticians may be appropriate for the purposes of defining gross domestic product but inappropriate for studies of child activities. A child who is cooking at a market stall is doing the same task as a child who is cooking at home. In both cases, cooking may be displacing other activities – such as playing or studying or attending school. Many children combine work and education, but at some point long work hours must interfere with learning, simply because of having only 24 hours in a day. Our concern is with the opportunity cost of working, if any, with respect to schooling. Non-labor force work is particularly important for girls, whose main work activities in Egypt are at home. To this end, we add hours spent in non-labor-force activities to those for labor force work, to the extent that our data allows. A number of papers have successfully used instrumental variables methods to identify the causal impact of child labor on educational outcomes in the context of developing countries, but in almost every case the focus has been on market work and usually on agricultural work. Beegle, Dehejia, and Gatti (2005) exploit rice prices and commodity disasters as a

How Much Work Is Too Much?

57

source of exogenous variation in child labor in Vietnam. Using panel data, they find significant negative impacts of child labor on school participation and educational attainment. Gunnarsson, Orazem, and Sa´nchez (2006) use cross-country variation in truancy regulations to identify the (negative) effect of children’s labor force work on test scores of third and fourth graders in nine Latin American countries. Bezerra, Kassouf, and Arends-Kuenning (2007) use city-level instruments to estimate the effect of children’s work – labor force work, domestic, or both – on school-based achievement tests. They found negative effects of both kinds of work, but noted that working up to two hours per day (14 hours per week) had little or no impact on school achievement. Ray and Lancaster (2004) instrument children’s work hours using data from Belize, Cambodia, Namibia, Panama, the Philippines, Portugal, and Sri Lanka but do not make a convincing case for their instruments: household access to water and electricity, as well as various household assets. It is useful to consider in greater detail two studies that examine child labor and schooling simultaneously while taking the intensity of child labor into account (Rosati & Rossi, 2001; Boozer & Suri, 2001). In a similar approach to ours, Rosati and Rossi (2001) jointly estimate a probit model for school enrollment and a Tobit model for work hours on data for Pakistan and Nicaragua. They include schooling status as an endogenous regressor in the hours equation to be able to predict hours conditional on the schooling status of the child. However, their identification strategy is somewhat suspect as they rely on the exclusion of parental schooling from the hours equation even though their levels of schooling can directly influence parents’ decisions to put children to work. Boozer and Suri (2001), in a paper on Ghana, resolve the difficult identification issue by relying on shocks in the amount of rainfall that induce exogenous variations in child labor and therefore subsequent fluctuations in the hours spent in school. They specify both schooling and labor as continuous variables measured in hours and estimate a log-linear specification of the structural schooling equation using ordinary least squares and two-stage least squares IV estimators. Their estimates of work and school hours are carried out at the level of the household rather than at the level of the individual child. To pursue their identification strategy, they rely on the fact that their crosssectional sample is spread out across several regions in both Northern and Southern Ghana and over 11 months of the year. They want to ensure that variations in rainfall only affect the marginal product of child labor and not the income of the household and, through income, the children’s schooling. To this end, they include month and region dummies in both schooling and

58

RAGUI ASSAAD ET AL.

work equations to capture seasonal and regional variations in rainfall, and they use month–region interactions as instruments for child labor. They assert that this strategy provides them with the long-term relationship between hours of work and hours of schooling. As an alternative identification strategy to capture the short-run relationship between child work and schooling, they use observed rainfall as an instrument and include month–region interactions and the month and region main effects in both the schooling and work equations. Under all the identification strategies they use, Boozer and Suri find that there is a significant negative effect of child labor on schooling and that the effect is stronger for boys than for girls, especially in the short-run. Their results indicate that the effects for girls are larger in the long-run than in the short-run, but this differential holds to a greater extent at the extensive margin of work (whether or not they work) than at the intensive margin (how many hours they work). We come back to issues of identification in the methodology section.

DATA AND DESCRIPTIVES Before presenting some details about children’s lives in Egypt – including information about the school system and kinds of work done by children – it is useful to describe the data used for this analysis. We will make use of these data in describing the Egyptian context, later.

ELMS-1998 The data for this study are obtained from the Egypt Labor Market Survey (ELMS-1998), which is a nationally representative household survey carried out on a sample of 5,000 households in late 1998.3 The ELMS-1998 survey instrument comprised a household questionnaire, an individual questionnaire, and a family enterprises questionnaire. The household questionnaire was administered to the head of each household or his/her spouse, and an individual questionnaire was administered for each member of the household aged 6 and above. The individual questionnaire included modules on parents’ characteristics, education, work status in a reference week and reference three months, unemployment, characteristics of employment, detailed work histories, and earnings from work for wage workers. If any of the members of the household reported being self-employed or an employer, the household also answered a family enterprises questionnaire.

How Much Work Is Too Much?

59

Details about children’s schooling were also collected and are discussed in more detail later. Completed questionnaires were obtained for 4,816 households and 23,997 individuals, of whom 5,003 were children aged 6–14 and another 1,801 were adolescents aged 15–17. All the results that we present about labor force work are based on a reference period of three months. Although labor statistics were also collected for a one-week reference period, we use the longer reference period because short reference periods can result in intermittent child workers being mis-identified as being out of the labor force (Levison, Hoek, Lam, & Duryea, 2007). For those children and youth who are usual participants in the labor market during the three-month reference period, we have information on hours worked. Labor market work hours refer to average weekly hours worked during the three-month reference period. Both boys and girls in Egypt engage in market work, but boys are involved in the labor force to a much greater extent than girls. Girls are often expected to work in subsistence agriculture and on domestic chores. The data include information on girls’ work in domestic tasks; a question asks, ‘‘On average, how many hours per day do you spend on all household chores?’’ Although it is likely that hours spent in chores, especially childcare, are underreported (Reynolds, 1991), we expect the relative orders of magnitude to be informative. Hours spent in subsistence agriculture was also collected for most girls, including time spent doing agricultural tasks and animal husbandry. Subsistence agriculture data was not, however, recorded for girls engaged in labor force work. Moreover, pre-tests of the ELMS-1998 that attempted to capture boys’ involvement in subsistence agriculture and on domestic chores were not successful – even if boys did those things, respondents were not willing to say so – thus questions about those activities were not asked for boys. In addition to the question about number of hours, girls were asked, ‘‘In which of the following chores do you spend most of your time?’’ Respondents are prompted to tell on what chores they spend the most time, the second-longest amount of time, and the third-longest amount of time. The choices are preparing food, running errands, cleaning and taking care of the house, fetching water for the household, doing laundry, looking after children, taking care of animals and poultry, and ‘‘I don’t do any of these chores.’’ We examined responses about chores for 6- to 14-year-old girls. ‘‘Running errands’’ was chosen as the activity taking the most time by the greatest percentage of girls and ‘‘cleaning and taking care of the house’’ was chosen by the greatest percentage of girls as the second-most

60

RAGUI ASSAAD ET AL.

time-consuming task. These were followed by food preparation and washing clothes, an activity that can take considerable amount of time when the household does not have ready access to a convenient source of water. The ordering of chores did not change very much when the answers were disaggregated by urban/rural location, but washing clothes and fetching water were of greater importance for rural girls. It is also likely that hours spent on child care are substantially underreported. Reynolds (1991) has shown that, in the context of rural Zimbabwe, only very young children consider child care to be a kind of work. Other children seem to consider it a fact of life rather than a task, and, in Reynolds’ study, adults completely overlooked the responsibility of children for younger children when reporting on children’s work.

THE EGYPTIAN CONTEXT Formal Schooling Children in Egypt are required, by law, to attend six years of primary education and three years of lower secondary education. From 1990 to 1999, however, this requirement was adjusted to allow the school system to absorb a large number of children: primary school was reduced to five years. All the children in our sample were required to take five years of primary and three years of lower secondary school. After the nine years of basic education, students take a government-mandated exam that determines whether they may continue their education, and if so, in what type of upper secondary school. There is a university-bound general track as well as several types of terminal secondary tracks. Because the law mandating years of schooling is not generally enforced, it is not uncommon to find youth who have dropped out before completing this requirement, which should happen at about age 15. All the 6- to 14-year-olds in our sample are supposed to be enrolled, with 14-year-olds in their last year of basic education if they have not repeated any grades. In fact, 11 percent of children aged 6–14 were out of school in 1998. Moreover, among sampled 14-year-olds in 1998, 7.2 percent of them had never attended school. These statistics have, however, improved over time. Assaad et al. (2010) describe substantial increases in enrollment, especially for rural girls, since 1988. Langsten and Hassan (2009) document high levels of failure to ever enroll in school, with poor girls in rural Upper Egypt being most disadvantaged. They attribute this to the out-of-pocket school-related

61

How Much Work Is Too Much?

costs and poor access to schools that impose time and transportation costs beyond the abilities of poor families, as well as the opportunity costs of a girl’s time in the home. The ELMS-98 includes information on the number of daily hours that children spend in school, if they are currently attending school. We analyzed the schooling information for 6- to 11-year-olds to minimize cohort effects. Since only 32 sampled children in this age group were out of school, we focus on those attending school. Overall, children aged 6–11 spend 5.9 hours per day in school on average (SD ¼ 0.8); there is almost no variation by region of the country or by sex, although boys and girls attend sexsegregated schools. Hours do vary by the number of shifts in children’s primary schools, which is related to the school type. There is additional information on the type of primary school attended and how many shifts it has, collected separately for children attending and not attending school. (If a child was not attending, the survey asks about the last school attended. If the child never attended, then the information was not collected.) For 6- to 11-year-olds in school, 59.5% attend primary schools with one shift, whereas 38.8% attend schools with two shifts. Less than 2% attend primary schools with three shifts. The great majority (90%) of sampled children attend public primary schools. Small percentages attend private schools, which do not (in our data) have shifts, or religious schools, which are equally divided between one and two shifts. Children attending public schools with one shift were, at the time of the survey, spending an average of 6.2 hours (SD ¼ 0.8) in school per day, whereas those who attended public schools with two shifts were spending 5.5 hours (SD ¼ 0.7) daily hours in school; these estimates do not change if we also include the small number of children in religious schools with one or two shifts. Again, these shift-specific estimates are surprisingly robust by region and by sex. Since most children attend public school and are unlikely to have a choice about which public school they may attend – either it has shifts or it does not – the lack of variation in daily hours within shift categories justifies our treatment of schooling as a limited dependent variable in the regression analysis that follows.

Work By law, children were permitted to engage in some types of labor force work starting at age 12 until 1996, when the minimum age was increased to 15, thus bringing it in line with mandatory basic education requirements. As in many low-income countries, law and practice regarding children’s

62

RAGUI ASSAAD ET AL.

work differ substantially. A sizeable minority of children work before age 15. Applying the percentages we estimate to Egypt’s population implies that approximately 318 thousand 10- to 14-year-olds do market work, whereas about 2 million do some kind of work, including domestic work. Table 1 presents a number of summary statistics on children’s participation in school and work, as well as hours worked, if any. The first row shows that a relatively high percentage of children attend school and that attendance is higher for boys than for girls, controlling for age group. The second row shows the proportions of children working. For girls, there are two definitions of ‘‘work,’’ one that includes only labor market work and Table 1. Weighted Proportions of Children Aged 6–14 Working and Attending School and Average Hours Worked Egypt, 1998. Boys Market worka Proportion attending school

Girls Market worka

Inclusive worka

0.928 (0.258) 0.033 (0.179)

0.857 (0.350) 0.016 (0.127)

0.857 (0.350) 0.422 (0.494)

Mutually exclusive categories (work W0 hours/week) Proportion in school only 0.910 (0.286) Proportion who only work 0.027 (0.162) Proportion who are both working 0.018 and in school (0.132) Proportion who are neither working 0.045 nor in school (0.207)

0.856 (0.351) 0.015 (0.121) 0.002 (0.041) 0.128 (0.334)

0.550 (0.498) 0.114 (0.318) 0.308 (0.462) 0.029 (0.167)

Proportion working (if work hours/ week Z1)

Number of observations Average hours worked/week, if work hours W0

2461 43.0 (21.9)

2388 47.4 (21.6)

2388 21.2 (16.0)

Number of observations Average hours worked/week for those who combine work and school, if work hours W0

110 24.9 (11.4)

35 – –

959 14.6 (8.7)

Number of observations

44

744

Source: Authors’ calculations from ELMS 1998. Notes: Standard deviations in parentheses. Inclusive work includes market work, subsistence agriculture work, and domestic work. – denotes fewer than 10 observations. a Market work includes only work for purposes of market exchange.

How Much Work Is Too Much?

63

the other (‘‘inclusive work’’) that includes labor market work, subsistence agriculture work, and household work. A small proportion of 6- to 14-year-old boys (3.3%) and girls (1.6%) are engaged in market work. This changes as children age: close to a quarter of 15- to 17-year-old boys work in the labor market, for example (not shown). In contrast, sizeable proportions of girls are engaged in housework, driving up the proportions under ‘‘inclusive’’ work. Over 42 percent of girls aged 6–14 are working by the inclusive definition. The next panel of Table 1 shows that few boys are able to combine work and school. Similarly, few girls are able to combine market work and school. The proportion combining work and school among girls is larger if we consider the inclusive definition of work. Then, over 30 percent of 6- to 14-year-old girls combine work and school. Children who perform market work do so for relatively long hours, averaging 43–47 hours per week depending on age and sex. Conditional on doing labor force work, girls who are employed work just as long hours as boys, if not longer. Girls’ inclusive definition of work includes more observations at the low end of the distribution, pulling the average down to 21 hours per week. It is therefore not surprising that market work seems to interfere more with schooling than inclusive work. Predictably, children who combine work and school work fewer hours per week on average, compared to all working children. Note in particular that those 6- to 14-year-olds engaged in market work averaged over 45 hours per week, whereas girls doing ‘‘inclusive work’’ averaged over 21 hours per week. All the means for hours worked have large standard deviations, and the distributions also show substantial variation in hours worked. Fig. 1 describes total work hours for boys and girls, by age groups, for those children with work hours greater than zero. Looking first at the two panels showing boys’ total hours of labor force work, we see that boys work a wide range of hours, and some boys work very long hours indeed – the data show heaping on numbers that suggest 10- and 12-hour workdays, for six or even seven days per week. The distributions for girls’ work hours, which combine hours of labor force work, subsistence work, and household work, are much smoother. Girls’ hours peak at 14–21 hours per week and then taper off, with very few girls working 49 or more hours per week. While the percentage of girls working very long hours is lower than the percentage of boys, a much higher percentage of girls are working non-zero hours compared to boys (Table 1), so the girls’ distribution represents a substantially greater number of children.

64

RAGUI ASSAAD ET AL. boys

girls

30

Percent

20

10

zero 1-