IMPACT EVALUATION REPORT ON INTERVENTIONS ... - ILO

IMPACT EVALUATION REPORT ON

INTERVENTIONS ON STARTING-UP SMALL ENTERPRISES AND REINTEGRATION OF CHILDREN TO SCHOOL OF THE

PROJECT “ELIMINATING CHILD LABOUR IN EL SALVADOR THROUGH ECONOMIC

EMPOWERMENT AND SOCIAL INCLUSION”

INTERNATIONAL LABOUR ORGANIZATION (ILO)

Copyright © International Labour Organization 2017

First published 2017

Publications of the International Labour Office enjoy copyright under Protocol 2 of the Universal Copyright Convention.

Nevertheless, short excerpts from them may be reproduced without authorization, on condition that the source is indicated.

For rights of reproduction or translation, application should be made to ILO Publications (Rights and Licensing), International

Labour Office, CH-1211 Geneva 22, Switzerland, or by email: [email protected]. The International Labour Office welcomes such

applications.

Libraries, institutions and other users registered with a reproduction rights organization may make copies in accordance with

the licences issued to them for this purpose. Visit www.ifrro.org to find the reproduction rights organization in your country.

ISBN:

The designations employed in ILO publications, which are in conformity with United Nations practice, and the presentation of

material therein do not imply the expression of any opinion whatsoever on the part of the International Labour Office concerning

the legal status of any country, area or territory or of its authorities, or concerning the delimitation of its frontiers.

The responsibility for opinions expressed in signed articles, studies and other contributions rests solely with their authors, and

publication does not constitute an endorsement by the International Labour Office of the opinions expressed in them.

Reference to names of firms and commercial products and processes does not imply their endorsement by the International

Labour Office, and any failure to mention a particular firm, commercial product or process is not a sign of disapproval.

ILO publications and digital products can be obtained through major booksellers and digital distribution platforms, or ordered

directly from [email protected]. For more information, visit our website: www.ilo.org/publns or contact [email protected].

Cover photo credits: ©Jonathan Levinson/IRIN

Design by the International Training Centre of the ILO, Turin – Italy

Acknowledgements

iii

TABLE OF CONTENTS

1. Introduction 1

2. Child labour in El Salvador 3

3. Project description 5

Overall project 5

Interventions selected for the impact evaluation 7

Theory of change 7

4. Evaluating the impact of the household level interventions 11

Selection of the communities and beneficiaries 11

Evaluation strategy 13

5. Project implementation 15

Intervention take-up 15

Interventions provided 15

Timing 176

6. Data collection 19

7. Sample attrition 21

8. Methodology 27

9. Results 31

Impact on households 32Adult labour market outcomes 32

Household income and expenditures 34

Household decision-making 35

Impact on children 37Children’s education and work 37

Parent’s attitudes 41

iv

IMPACT EVALUATION REPORT ONinterventions on starting-up small enterprises and reintegration of children to school of the Project “Eliminating Child Labour in El Salvador through Economic Empowerment and Social Inclusion”

10. Conclusions 45

11. References 47

12. Annexes 49

Annex 1. Variable discontinuity: Scatter plots 49

Annex 2. Estimates of the effect of interventions: scatter plots 54

Annex 3. RD Estimates of the effect of interventions: Full results 57

Annex 4. Impact evaluation design 71

Annex 5. Baseline suvey report 98

Annex 6. Interim report on follow-up survey (Spanish) 127

Annex 7. Follow-up survey household questionnaire 210

List of tablesTable 1. Follow-up survey, children's employment and child labour, full sample of

children aged 5-17 years 4

Table 2. Basic information on selected municipalities 12

Table 3. Program take-up among selected beneficiaries 15

Table 4. Distribution of intention-to-treat and control households by municipality, baseline and follow-up survey 21

Table 5. Estimated discontinuities of the probability of being re-interviewed at the follow-up survey 22

Table 6. Estimated discontinuities in baseline covariates at the threshold in the wealth index: households interviewed in follow-up survey 24

Table 7. Estimated discontinuities in baseline outcome variables at the threshold in the wealth index: members of households interviewed in follow-up survey 25

Table 8. Variable definitions 31

Table 9. Impact on adult female labour market outcomes 33

Table 10. Impact on adult male labour market outcomes 33

Table 11. Female labour income, household adult labour income per capita and household expenditure per capita 34

Table 12. RD estimates of the effects of interventions on women decision-making within household 36

Table 13. RD estimates of the effects of interventions on children’s school attendance, regularity of school attendance and school expenditures, children aged 5-15 years in the baseline survey 37

Table 14. Comparison of attendance from school survey and self-reported school attendance, children aged 5-17 in the follow-up survey 38

Table 15. RD estimates of the effects of interventions on children’s involvement in employment and household chores, children aged 5-15 years in the baseline survey 39

v

Table of contents

Table 16. RD estimates of the effects of interventions on child time use, children aged 5-15 years in the baseline survey 40

Table 17. RD estimates of the effects of interventions on child involvement in own or house-hold business, working hours, work in hazardous conditions and child labour, children aged 5-15 years in the baseline survey 41

Table 18. RD estimates of the effects of interventions on the attitudes of the household head towards children’s employment 42

Table 19. RD estimates of the effects of interventions on the attitudes of the household head towards children’s schooling 43

Table A1. RD Estimates of the effect of interventions on child outcome variables, children aged 5-15 in the baseline survey 57

Table A2. RD Estimates of the effect of interventions on child outcome variables, children aged 7-15 in the follow-up survey 56

Table A3. RD Estimates of the effect of interventions on child outcome variables, children aged 5-15 years in the baseline survey 61

Table A4. RD Estimates of the effect of interventions on child outcome variables, children aged 7-15 in the follow-up survey 64

Table A5. RD Estimates of the effect of interventions on employment of eligible females and adult males from eligible households 65

Table A6. RD Estimates of the effect of interventions on the attitudes of the household head towards children's employment and schooling 66

Table A7. RD Estimates of the effect of interventions on employment of eligible females and adult males from eligible households 67

Table A8. RD Estimates of the effect of interventions on the attitudes of the household head towards children's employment and schooling 69

List of figuresFigure 1. Percentage of children in child labour, 5-17 years age range, 2001, 2007 and 2013 3

Figure 2. Strategic framework for the project “Eliminating Child Labour in El Salvador through Economic Empowerment and Social Inclusion” 6

Figure 3. Theory of change 9

Figure 4. Frequency distribution of treated households by specific intervention received 16

Figure 5. Number of households treated by year 17

Figure 6. Probability that household is interviewed in the follow-up survey 22

Figure 7. Density of the wealth index 23

Figure 8. AIC comparison for estimations of equation (1) for the preferred cross-validation criterion and with different orders of polynomial 28

Figure A1. Household level wealth indicators 49

Figure A2. Outcome variables: eligible women 51

Figure A3. Outcome variables children: aged 5-17 years 52

Figure A4. Impact of interventions: children aged 5-15 years at the time of the baseline survey 54

Figure A5. Impact of interventions: eligible females 56

2SLS Two-Stage least squares regression

AIC Akaike information criterion

CVC cross-validation criterion

DEICO Desarrollo, Investigación y Consultoría S.A. de C.V.

EHPM Encuesta Nacional de Hogares de Propósitos Múltiples

ILO International Labour Organization

IPEC International Programme for the Eradication of Child Labour

ITT intent-to-treat effect

OLS Ordinary Least Squares regression

RD Regression Discontinuity

TOT treatment-on-the treated effect

UCW Understanding Children's Work Programme

USDOL United States Department of Labor

LIST OF ACRONYMS AND ABBREVIATIONS:

vi

vii

EXECUTIVE SUMMARY

Child labour remains an important issue in El Salvador.1 According to the country’s national household survey, 8.5 per cent of all 5 to 17 year-olds, that is 144,200 children, are engaged in child labour. Even though this proportion declined between 2001 and 2007 (from 9.8 to 8.2 per cent), it rose again in the following six years. Most of these child labourers are boys in rural areas. Moreover, three out of five child labourers are engaged in hazardous activities, including cutting sugar cane, fishing, harvesting coffee, and street vending.

This policy concern is being addressed by the International Labour Organization (ILO) project entitled “Eliminating Child Labour in El Salvador through Economic Empowerment and Social Inclusion”, which aims to contribute to the elimination of child labour, particularly in its worst forms throughout the country. Funded by the United States Department of Labor (USDOL), the project sets six main objectives, including mainstreaming child labour strategies in in-stitutions targeting the poor, developing law enforcement and protection mechanisms, and enhancing national research, monitoring and impact evaluation capacity.

Given the complexity and size of the project, the impact evaluation presented here focuses on a specific intervention at the household level: a support package offered to help the moth-ers of child labourers to start a small business by providing vocational training, business-plan development assistance, and a monetary starting kit. The training comprised 29 different business activities, being patisserie, bakery, and tailoring the most common. In addition to the increased income (and reduced child labour reliance) sought by the intervention, there are female empowerment effects that are expected to be used for the protection of children from child labour and ensuring their schooling.

A selection of five municipalities with 1) high child labour rates, b) low crime rates, and c) a minimum of 19,000 inhabitants, was made. These are Tacuba, Izalco, San Luis la Herradu-ra, Tecoluca, and Santiago Nonualco. Within the five municipalities, a group of 37 cantons and 2,098 households in these cantons were selected for the evaluation. The poorest 1,098 households were selected into the intervention and the other 1,000 constituted the control group. Taking advantage of this selection strategy, the evaluation used a regression discon-tinuity design.

The impact evaluation found significant impacts on adults in several aspects. First, a large in-crease (40 per cent) was found in women’s employment, while men’s work remained equal. Second, even though no effects of the intervention on household income and expenditures were observed, the decision-making role of women within the household was enhanced. Im-pacts were positive on decisions of five topics: purchases of food, child school attendance, business investments, loans, and expenses to improve housing.

1 A child is considered to be in child labour if he or she falls into one of the following categories: aged 5-13 in employment, aged 14-15 working in hazardous conditions or employed for more than 34 hours a week, or aged 16-17 working in hazardous conditions.

viii


In addition to the mentioned effect, the evaluation found that child work and education, as well as parents’ attitudes, changed because of the programme. On the one hand, school attendance was increased and both the proportion of children in work only and the number of hours worked by children were reduced. Hazardous work or the child labour headcount re-mained unchanged. On the other hand, in terms of parents’ attitudes, substantial reductions in the proportions of parents thinking that working children are stronger and healthier and of those thinking that education is not important for success, were observed.

The impact of the programme was substantially positive for children and households. The project was effective at substituting children’s work time for classroom time, rather than taking children permanently out of work. Having said this, the intervention succeeded in im-proving the conditions and prospects of child labourers. Whether such improvements will be sustained in the long run is a question that remains open. Additionally, the project increased the participation of women in their own business and their intra-household decision-making role. The design of the evaluation does not allow us to ensure if these changes in women’s involvement in employment and their empowerment were behind the reductions in child labour.

1

The ILO project entitled “Eliminating Child Labour in El Salvador through Economic Empowerment and Social Inclusion” aimed at helping to eradicate child labour in El Salvador. The project supported a wide range of interventions at three levels: macro (national policies and institutional framework), meso (target municipalities and schools) and micro (child labourers’ households).

A set of household-level interventions (International Programme for the Eradication of Child Labour), implemented in five municipalities in El Salvador, is the object of the impact evaluation discussed here. The impact evaluation is one of the first to identify the causal effect of an IPEC project using a quasi-experimental approach with a valid counterfactual.

As described in more detail below, the targeted set of interventions consisted of providing support to mothers of child labourers to start a small enterprise as well as of “flexible education interventions” to help their children return to school at the appropriate level in case they have dropped out.

Similar sets of household-level livelihood and social mobility interventions can be found in other IP-EC-supported country-level projects and United States Department of Labour (USDOL) programmes in a number of countries. By identifying the causal effect of this set of interventions, therefore, the impact evaluation provides a unique insight into the functioning and effectiveness of an important dimension of IPEC’s and USDOL’s broader approach to addressing child labour.

In April 2012, in collaboration with the ILO El Salvador, drafted an impact evaluation design (UCW, 2012; attached as Annex 4 to this Report). As part of this design, UCW proposed to allocate the set of interventions to eligible households on the basis of a wealth index. This procedure ensures a fair distri-bution of the interventions to those households who are most in need. At the same time, as explained by UCW (2012), this allocation procedure can be exploited in a regression discontinuity framework to identify the causal effect of the set of interventions on outcomes of interest, such as education, child labour, participation in work of adult household members, and multiple other wellbeing indicators.

The International Labour Organization (ILO) conducted a baseline survey in the five municipalities in April and May of 2012. Subsequently, UCW (again in collaboration with the ILO in El Salvador) used these data to generate the wealth index utilized to assign the intervention. This process is described in more detail in the baseline survey report released in 2013 (UCW, 2013a; attached as Annex 5 to this Report). The follow-up survey was undertaken in July and August of 2015.

The current Report discusses the results of the follow-up survey and of the impact evaluation more broadly. It is structured as follows. Following this introduction, Section 2 provides a brief background overview of the child labour situation in El Salvador and how it has changed over time. Section 3 briefly discusses the project at large, the interventions selected for the impact evaluation and the underlying theory of change. Section 4 explains the selection process and the evaluation strategy. Section 5 looks at project implementation, specifically at intervention take-up, services provided, and the timing of imple-

1. INTRODUCTION

2


mentation. Section 6 describes data collection and Section 7 addresses the issue of sample at-trition between the baseline and follow-up surveys. Section 8 describes the methodology for the impact evaluation. Section 9 discusses the results and Section 10 presents the conclusions.

3

2. CHILD LABOUR IN EL SALVADOR

Figure 1. Percentage of children in child labour, 5-17 years age range, 2001, 2007 and 2013

Source: UCW calculations based on El Salvador Encuesta Nacional de Hogares de Propósitos Múltiples, 2001, 2007, and 2013.

Official surveys show that child labour in El Salvador have declined in the last years but it is still a pressing issue. Estimates based on the 2013 Encuesta Nacional de Hogares de Propósitos Múltiples (EHPM, 2013) show that 8.5% of all 5-17 year-olds, 144 200 children in absolute terms, are engaged in child labour. By age range, child labour affects almost six percent of children aged 5-14 years and 16% of adolescents aged 15-17 years. A comparison between estimates based on the 2001 and 2007 rounds of the same survey suggests that progress against child labour over the last 12 years has been limited and uneven. As illustrated in Figure 1, there was a decline from 9.8% in 2001 to 8.5% in 2013, but this overall decline masked a slight rise in child labour from 2007 to 2013.

The child labour population in El Salvador is male-dominated and concentrated in rural locations. The share of boys in total child labourers (13%) is more than three times that of girls (4%), while child labour incidence in rural areas (12%) is more than twice as high as incidence in urban areas. Over half of all working children are found in the agriculture sector; the commerce and services sectors ac-

4


count for most of the remainder. The majority of children’s jobs are characterised by informal, non-remunerated work arrangements, and are typically undertaken within the family unit.

Much of the work performed by children is hazardous in nature, adding to the urgency of elimination efforts. For the 5-14 age group, three out of five child labourers are found in haz-ardous sectors / occupations or are working excessive hours. Examples of work in which dan-gers are encountered include cutting of sugar cane, fishing (particularly mollusc extraction), harvesting coffee and street vending. Excessive hours, heavy loads, and a lack of protective gear are among the aggravating factors. Additional forms of hazardous labour in which a lesser number of children are involved include recycling of garbage, fireworks production, and domestic work in third-party households.

Table 1 reports the incidence of children’s employment and child labour for the full sample (i.e. for both the treatment and control group) of households considered in this study. As it is easy to see, the incidence of child labour in the targeted households is much higher than the national average.

Table 1. Follow-up survey, children's employment and child labour, full sample of children aged 5-17 years

FREQUENCY PERCENT

Employment 658 21.4

Child labour 622 20.3

Total 3,081

5

�Overall project The ILO project entitled “Eliminating Child Labour in El Salvador through Economic Empowerment and Social Inclusion” attempts to eradicate child labour in El Salvador. To achieve this goal, the project set the following six objectives:

1. By the end of the project, child labour elimination strategies will have been adopted by national and local institutions that implement poverty reduction, decent work, and social protection pro-grammes for the rural and urban poor;

2. By the end of the project, child labour law enforcement and protection mechanisms, particularly for the worst forms of child labour, will be developed and fully operational;

3. By the end of the project, national capacity to conduct child labour research, monitoring, and im-pact evaluation will be enhanced and pilot interventions documented;

4. By the end of the project, municipal capacity to prevent and eliminate child labour, in particular the worst forms, will have been strengthened;

5. By the end of the project, viable improved livelihood alternatives will have been implemented to reduce family reliance on child labour in target municipalities; and

6. By the end of the project, inclusive education models will have been implemented in selected schools of target municipalities to prevent and reduce child labour and improve education indica-tors.

Figure 2 provides a bird’s eye view of the project’s strategic framework and of how the project’s six strategic objectives tie in with El Salvador’s wider national strategic agenda to eradicate child labour. As shown, the project involves support to a large number of outcomes and associated interventions in order to achieve the six strategic objectives. Some of these interventions are implemented at the national level, some in selected municipalities, some in selected schools, and some are targeted at selected households.

3. PROJECT DESCRIPTION

6


Figure 2. Strategic framework for the project “Eliminating Child Labour in El Salvador through Economic Empowerment and Social Inclusion”

Source: UCW calculations based on El Salvador Encuesta Nacional de Hogares de Propositòs Múltiples, 2001, 2007, and 2013.

7

3. Project description

�Interventions selected for the impact evaluationGiven the complexity of the project and its multilevel structure, it was decided to limit the focus of the impact evaluation to a subset of interventions at the household level, selected in consultation with the Government of El Salvador and the US Department of Labor. This set consisted of the following two components:

1. A support package offered to help the mothers of child labourers start a small enterprise. This support consists of three steps: vocational training of choice;1 business training and preparation of a business plan; and a starting kit based on the needs identified in the business plan to kick-start the enterprise (value between US$ 100 and US$ 300).

2. If the children living in the households of the eligible women were not attending school according to the baseline information, then they are offered training to help them enter school at the appropriate level for their age. Children are invited to participate in this in-tervention by means of a sensitization campaign.2

However, as discussed below, only a minority of children took up the second component and, therefore, the impact we attempt to measure here is mainly relative to the support package for mothers.

Similar sets of livelihood and social mobility interventions can be found in other IPEC coun-try-level projects and USDOL programmes in a number of countries, and therefore their impact on child labour is also of broader interest.

�Theory of changeBroadly speaking, the program considered here belongs to a set of interventions that aim to increase productive employment opportunities for the rural poor in low- and lower mid-dle-income countries, and are widely implemented around the world (Cho and Honorati, 2014; Banerjee et al., 2015). These programs typically address household capital or skills constraints that can limit the possibility of poor households to cover setup costs and develop potentially profitable farm and non-farm economic activities (Eswaran and Kotwal, 1986). Some programs are targeted explicitly to women, as women are more likely than men to face obstacles in accessing credit and labour markets. In the case of the program considered here, the anticipated improvement in the livelihood of the household is also expected to have a positive effect on the investment in children’s human capital.

As discussed in De Hoop and Rosati (2013), on which this section draws, evidence of the impact of providing physical capital and skills training on children’s time use is scarce and results are varied. Banerjee et al. (2011) find limited effects of the Indian THP (Targeting the Hard-core Poor) program on children’s school attendance and labour supply. Bandiera et al. (2013), however, find that a similar program in Bangladesh increased children's work in self-employment. Karlan and Valdivia (2011) find that business training in Peru lowered

1 In each canton a limited number of vocational training courses is offered in areas such as baking bread, sewing clothes and others. The courses on offer are determined on the basis of an analysis of the local labour market.

2 Out of school children are not obliged to participate and the mother’s participation in the vocational training does not depend on the participation of the child.

8


children's participation in work and increased their participation in school, although these effects are not statistically significant.

Del Carpio and Loayza (2012) study the effects of a conditional cash transfer program com-plemented with a productive investment grant in Nicaragua. Their study focuses on a dif-ferent program than the one we analyse in this paper, as well as on a different (although not very dissimilar) region. The authors show that the intervention contributed to reducing overall child participation in household chores and agricultural work, but increased child participa-tion in commerce and retail. This is consistent with results in Del Carpio and Macours (2010) on the same intervention, which indicate that the productive investment grant added to a cash transfer reinforced existing specialization in non-agricultural activities and domestic work for girls, but that overall child labour did not decrease.

Most recently, De Hoop and Rosati (2015) analysed the impact on education and child labour of a program aimed at increasing women economic participation in Nicaragua. The results point to the importance of women’s empowerment in determining the impact of the program on child labour and children’s schooling.

The project considered here should result in new economic opportunities for participating women, influencing, in turn, decisions concerning children’s work and schooling through three basic channels.

First, the increased income available to the household through women’s new economic op-portunities should make the households less reliant on the income and productivity of their child members, and leave them in a better position to afford the direct and indirect costs of their children’s schooling.

The second channel relates to women’s empowerment. The new income opportunities re-sulting from the training and start-up support are likely to increase women’s control of house-hold resources and their bargaining power within the household. Insofar as women attribute more importance than men to children’s education, the increased women empowerment generated by the project should contribute to the reduction of child labour and the increase of school attendance.

The third channel relates to the impact of expanded household business activities in possibly increasing children’s work: household business activities may increase children’s produc-tivity in economic activities or lead to children substituting adults in performing household chores. As long as capital and children’s work are gross substitutes, the increased economic activity at the household level is likely to increase the household benefits of children’s work or of children’s performance of household chores instead of adults.

This third impact channel could limit or counterbalance the positive effects of the project on child labour and education. We cannot, therefore, discount a priori the possibility that new production opportunities will also create new work opportunities for children.

However, as the female beneficiaries are drawn from poorest households, where the income effect of new productive opportunities is likely to be greatest, it is reasonable to think that the net impact of the project will be a reduction in child labour and an increase in schooling. The positive change will be supported by training to out-of-school children aimed at facilitating their school (re)entry. The theory of change is depicted in very basic terms in Figure 3 below.

9

3. Project description

Figure 3. Theory of change

Inco

me e

ffect

s stem

ming

from

star

t-up

activ

ities m

ade p

ossib

le by

train

ing

and s

tart-u

p kits

.

Subs

titutio

n effe

cts ste

mming

from

ex

pand

ed ho

useh

old pr

oduc

tive

activ

ities .

New

oppo

rtunit

ies ar

e cre

ated f

or

child

ren’s

labo

ur as

a re

sult o

f exp

ande

d ho

useh

old pr

oduc

tive a

ctivit

ies .

Grea

ter in

come

secu

rity m

akes

ho

useh

old le

ss re

liant

on in

come

or

prod

uctiv

ity of

child

ren’s

labo

ur an

d be

tter a

ble to

affor

d the

inve

stmen

t in

their s

choo

ling.

Scho

ol (re

)entr

y fur

ther

facilit

ated b

y tra

ining

for o

ut-of-

scho

ol ch

ildre

n.

Ince

ntive

s cre

ated

favo

urin

g ch

ildre

n’s s

choo

ling

as

oppo

sed

to th

eir

invo

lvem

ent i

n wo

rk.

Ince

ntive

s cre

ated

favou

ring

child

ren’s

invo

lveme

nt in

work

as op

pose

d to t

heir

scho

oling

.

Fem

ale em

powe

rmen

t effe

cts

stemm

ing fr

om w

omen

’s gr

eater

co

ntrol

of ho

useh

old re

sour

ces a

nd

grea

ter ba

rgain

ing po

wer w

ithin

the

hous

ehold

.

Wom

en us

e the

ir gre

ater c

ontro

l of

reso

urce

s and

grea

ter ba

rgain

ing po

wer

to pr

otect

their c

hildr

en fr

om ch

ild

labou

r and

to en

sure

their

scho

oling

.

Elig

ible

mot

hers

are p

rovid

ed th

e vo

catio

nal t

rain

ing

of ch

oice

.

Elig

ible

mot

hers

are p

rovid

ed w

ith b

usin

ess

train

ing

and

star

t up

kits b

ased

on

the

need

s ide

ntifi

ed in

their

bus

ines

s pl

ans.

Out-o

f-sch

ool c

hild

ren

of el

igib

le m

othe

rs ar

e pro

vided

train

ing

to h

elp th

em

(re)e

nter

scho

ol at

the a

ppro

priat

e lev

el

for t

heir

age.

DR

IVER

S O

F C

HA

NG

EPR

OJE

CT

OU

TPU

TSIM

PAC

TS

11

�Selection of the communities and beneficiariesThe project team started by selecting a group of 10 municipalities to participate in the interventions. In order to select these municipalities, the project team first produced a shortlist. To appear on the shortlist, municipalities had to:

� Have a high child labour rate. In order to determine child labour rates, the project team used the 2009 school census which contains information on participation in employment for all children attending school (either basic education, “bachillerato general” or “bachillerato técnico”).3 Municipalities in which the involvement in employment among children in school exceeded 6.2% were considered to be municipalities with a high child labour rate.

� Be reasonably safe to ensure that it is possible to implement the necessary fieldwork. Municipalities were considered to be safe if they had a homicide and extortion rate of 128 per 100,000 inhabitants or lower according to the national police.

� Have more than 19,000 inhabitants to make certain that an overall target of 5,000 households with child labourers could be reached by all household level interventions.

A shortlist of 21 municipalities resulted from the application of these criteria, from which a total of five municipalities (i.e., Izalco, San Luis Herradura, Santiago Nonualco, Tacuba, Tecoluca) within reason-able geographic proximity of each other was eventually selected by the project team for participation in the interventions that were the subject of the current impact evaluation.

In order to ensure that the impact evaluation did not capture the effects of other interventions included in the overall project, only the interventions that were the focus of the impact evaluation were imple-mented in the selected 10 municipalities.

Table 2 provides an overview of the children’s involvement in employment, total population and the crime rate (homicide and extortion per 100,000 inhabitants) in the five municipalities.

3 Basic education in El Salvador consists of 9 years of education divided into three cycles of 3 years each. Pupils are supposed to enter basic education when they are 6 years old and finish when they are 15 years old. Basic education is followed by a 2 year “bachillerato” degree.

4. EVALUATING THE IMPACT OF THE HOUSEHOLD LEVEL INTERVENTIONS

12


Table 2. Basic information on selected municipalities

MUNICIPALITY URBAN OR RURAL

CHILDREN’S EMPLOYMENT

(%)(a)

POPULATION(b) CRIMERATE(c)

Tacuba Rural 15.4 30,718 68

Izalco Urban 8.4 74,085 115

San Luis la Herradura

Rural 7.9 21,388 122

Tecoluca Rural 7.9 25,344 55

Santiago Nonualco Rural 6.3 41,287 63

Notes: (a) Children’s employment is the percentage of working children identified in the 2010 school census; (b) Population estimates based on 2007 population census; and (c) Crime rates (homicide and extortion per 100,000) based on 2010 national police statistics.

Within each of the selected municipalities, a group of cantons was selected into the project. The selection of cantons took place on the basis of the information available from the 2010 school census and the 2007 population census. Cantons were selected if they had high numbers of child labourers according to the school census and had high child labour ratios.4 However, explicit thresholds for the selection cantons were not established. In total, the pro-ject team selected 37 cantons across the five municipalities.

Within each of the selected cantons, eligible households were identified on the basis of the data collected at baseline. Baseline data collection consisted of two parts: (1) a short-listing questionnaire administered to all households in 37 cantons, and (2) a more in-depth base-line questionnaire administered to all households identified as having working children in the listing exercise. The project team succeeded in tracking and administering the in-depth baseline questionnaire to 3,064 (93%) of the 3,294 households with working children iden-tified in the shortlisting.

Households were defined as being eligible for participation if (1) they had one or more work-ing child aged 5-17 years and (2) they had one or more literate woman aged 18 to 65 years that was not employed as a salaried worker. Of the 3,064 households with working children interviewed at baseline, 2,098 (68%) had a mother or another woman who was eligible ac-cording to these criteria.

The beneficiary households were selected out of the eligible households on the basis of a wealth index constructed using information provided by households during baseline data collection (see UCW 2012 and UCW 2013a for details). The poorest 1,098 households (i.e., the 1,098 households with the lowest wealth index) were selected into the program, while the remaining 1,000 households served as the control group. Note that the selection on the basis of a wealth index not only ensured a fair distribution of the interventions to those households that were most in need but also ensured that the causal effect of the set of interventions could be identified in a regression discontinuity framework.

4 As suggested by the number of child labourers in the school census divided by the total canton population according to the 2007 census.

13

4. Evaluating the impact of the household level interventions

�Evaluation strategyThe impact evaluation is aimed at identifying the causal effect (i.e., the impact) on child labour of the set of interventions on education, child labour, and participation in labour of adult household members (especially women, as they are the key target group of the inter-ventions). To identify the causal effect, the impact evaluation must estimate a counterfac-tual outcome: the education and (child) labour outcomes that would have been observed among project beneficiaries in the absence of the project. This counterfactual outcome is unobserved by definition. However, a valid counterfactual can be established by identifying a comparison group that is similar to the group of beneficiaries at the start of the project in all respects, with the exception that they do not participate in the intervention.

As explained in detail in UCW (2012), to identify a credible comparison group, the impact evaluation exploits the fact that the set of interventions described above is assigned on the basis of a wealth index using the so-called regression discontinuity methodology.

The intuition behind the regression discontinuity methodology is as follows. All households with working children and a mother who can participate in the intervention are ranked from poorest to richest on the basis of a wealth index generated using baseline data collected by the project, as described above. The poorest households (i.e., the households in the lower half of the wealth distribution) are assigned to the intervention group. The richest households (i.e. the households in the upper half of the wealth distribution) do not receive any inter-ventions and constitute our control group. While these two groups are obviously not similar before the implementation of the interventions, they may be expected to be virtually identical close to the point in the wealth index that separates the comparatively poor intervention households and the comparatively rich control households. By comparing intervention and control households close to this threshold, we can identify the (local) impact of the program.5

5 For a heuristic discussion see Lee and Lemieux (2010).

15

�Intervention take-upMonitoring information collected at the time of treatment indicates that there was a total of 680 treated households out of the 1,098 households initially selected as beneficiaries for the support packages to mothers described above. A total of 418 selected households, therefore, did not accept to participate in the program resulting in the take-up rate of about 62%. As shown in Table 3, Santiago Nonualco municipality was host to the largest number of treated households (206), followed by Tacuba (181), Luis La Herradura (128), Izalco (94) and Tecoluca (71).

Table 3. Program take-up among selected beneficiaries

PROGRAM TAKE-UP AMONG SELECTED BENEFICIARIES

Municipality

Households initially selected as beneficiaries

Participated in the program

Did not participate in the program Total

Tacuba 181 105 286

Izalco 94 36 130

Santiago Nonualco 206 139 345


128 69 197

Tecoluca 71 69 140

Total 680 418 1,098

Only a minority of children took up the education package. This is because of the relatively high level of school attendance. Therefore the rest of this discussion focuses on the package for the mothers.

�Interventions providedInformation on the specific interventions received by the treated households, also collected as part of monitoring at the time of treatment, is reported in Figure 4. As shown, the project involved the provision of 29 training-related interventions, distributed unequally across the 680 beneficiary house-holds. Patisserie and bakery were the most common, benefiting 80 and 79 households, respectively, followed by tailoring (74 households), training in business (72 households) and chicken breeding (57

5. PROJECT IMPLEMENTATION

16


households). At the other end of the frequency distribution, fewer than five households were provided with orientation training, such as rapid productive assessment.

Figure 4. Frequency distribution of treated households by specific intervention received

17

5. Project implementation

�TimingImplementation began in 2012 when a total of 162 households were treated. Implementation peaked the following year at 355 households, slowing after that to 123 households in 2014 and 38 households in 2015 (Figure 5).6

Figure 5. Number of households treated by year

Notes: (a) Information was missing for two households.

6 Information on the date of treatment was missing for two households.

19

Baseline data were collected in the five participating municipalities at the end of April and in May of 2012 (see UCW 2012 and UCW 2013a for further details). Follow-up data collection was undertaken primarily in July and August of 2015.7 The follow-up survey consisted of an in-depth questionnaire ad-ministered to the same households identified as having working children in the initial listing exercise. The follow-up survey questionnaire consisted of the following sections:

1. Identity and address of the respondent;

2. Characteristics of the household members;

3. Education status of household members (current school participation, highest level of education attended, ability to read and write);

4. Economic activities, vocational/professional training, other income, and other activities of all house-hold members aged five years or more;

5. Characteristics of the house inhabited by the household and access to durable goods;

6. Health of household members aged five to 17 years;

7. Businesses or enterprises of household members;

8. Participation in public or private social protection programs;

9. Respondent's attitudes towards child labour and education;

10. Household decision-making; and

11. Awareness of the IPEC Eliminating Child Labour in El Salvador through Economic Empowerment and Social Inclusion program.

The follow-up survey questionnaire is included as Annex 7 to the Report. Additionally, OIT and DEICO (2015), attached as Annex 6 to the Report, describes the follow-up data collection and discusses the main descriptive results of the follow-up survey (in Spanish).

7 In El Salvador the academic year runs from January to November. Thus, the baseline and follow-up surveys were fielded during the academic year and not during school holidays. Therefore, school holidays are not a factor in explaining differences in school attendance and child labour. Both the baseline and followup surveys were fielded during the rainy season, therefore also excluding seasonal factors in explaining the observed differences between treated and non-treated households.

6. DATA COLLECTION

21

Sample attrition from the baseline to the follow-up survey was unfortunately substantial. Of the 2,098 eligible households interviewed at baseline, only 1,496 participated in the follow-up survey, for an overall attrition rate of 28.7%. One of the reasons for the high attrition rate was the worsening security situation in the country and in the concerned municipalities that made it difficult, and at times impos-sible, to access all of the original households.

Breaking the attrition figure down further, the follow-up survey covered 794 of the initial 1,098 select-ed beneficiaries (attrition rate of 27.7%) and 702 of the initial 1,000 control households (attrition rate of 29.8%).8 Attrition rates for the control and intention-to-treat households decomposed by munici-pality are reported in Table 4. The municipalities of Izalco and San Luis la Herradura had the highest attrition rates, mainly due to the worsening security situation in these locations.

Table 4. Distribution of intention-to-treat and control households by municipality, baseline and follow-up survey

MUNICIPALITY CONTROL INTENTION-TO-TREAT TOTAL

Baseline (no.)

Follow-up (no.)

Attrition (%)

Baseline (no.)

Follow-up (no.)

Attrition (%)

Baseline (no.)

Follow-up (no.)

Attrition (%)

Tacuba 139 122 12.2 286 225 21.3 425 347 18.4

Izalco 109 45 58.7 130 56 56.9 239 101 57.7

Santiago Nonualco 410 296 27.8 345 273 20.9 755 569 24.6


149 88 40.9 197 138 29.9 346 226 34.7

Tecoluca 193 151 21.8 140 102 27.1 333 253 24.0

Total 1,000 702 29.8 1,098 794 27.7 2,098 1,496 28.7

The relatively high attrition rate observed might invalidate the inference if it is not independent of treat-ment status and if generates an unbalanced sample in terms of baseline characteristics. Note that, as shown in the baseline report, the impact evaluation sample was well balanced at baseline. In order to exclude these possibilities, we carried out a set of tests that confirm the independence of attrition from treatment and balance in terms of baseline characteristics.

First of all, we test whether attrition is independent of treatment status. For all 2,098 eligible house-holds, we generate a dummy variable that takes the value 1 if the household was interviewed at fol-

8 The follow-up survey covered 522 of the 680 treated households (i.e., of the households that the monitoring data indicated actually received the treatment).

7. SAMPLE ATTRITION

22


low-up and 0 otherwise. As illustrated graphically in Figure 6, and shown formally in Table 5, the results of this test indicate that the probability of being interviewed at follow-up is not significantly different in the values that are closer to the threshold.

Figure 6. Probability that household is interviewed in the follow-up survey

Table 5. Estimated discontinuities of the probability of being re-interviewed at the follow-up survey

DISCONTINUITY STD. ERROR P-VALUE

(1) (2) (3)

Household is interviewed at the follow-up survey 0.015 0.056 0.786

Notes: Column (1) displays estimated discontinuities in the probability of being re-interviewed at the follow-up survey displayed in the stub column at the threshold separating the intervention households from the control households (based on the estimation procedure of Austin Nichols). Columns (2) and (3) respectively show the standard errors and P-values of the estimated coefficients.

In the baseline survey report (UCW 2013) we showed both graphically and formally that the households in the proximity of the wealth threshold for selection in the treatment (our forcing variable) were similar in terms of wealth characteristics. Moreover, we showed that the den-sity of the wealth index was continuous at the threshold.

In light of the relatively high attrition rate, we now repeat our graphical and formal tests for non-attritees (i.e., for the households re-interviewed at follow-up).

Following McCrary (2008), we examine whether the density of the wealth index is continuous at the threshold. The density of the wealth index not being continuous is an indication that there could be systematic differences between the intervention and the control group. Fig-ure 7, which was generated using McCrary’s sub-routine for Stata, depicts the density of the wealth index. The density of the wealth index is still reasonably continuous for the group of households also interviewed at followup. Most importantly, the confidence intervals indicate that the estimated density is not significantly different just above and below the threshold.

Prob

abili

ty

Wealth index

23

7. Sample attrition

Figure 7. Density of the wealth index

We then test for the absence of discontinuity in the baseline household characteristics (which will be used as covariates) and in the baseline outcome variables for children, eligible females, and adult males. To test whether there are significant discontinuities at the thresh-old score, we run local linear regressions within an optimal bandwidth around the threshold score (making use of a Stata sub-routine written by Austin Nichols).

The scatter plots presented in Annex 1 (Figure A1 to Figure A4) provide a first indication that non-attrited households just above and below the threshold remain similar in terms of baseline. The scatter plots suggest that the relationship between the wealth index and the covariates is roughly continuous at the threshold for selection into the project. The Lowess regressions for the intervention and control groups hit the threshold roughly at the same value of the wealth indicator.

The formal test also finds no evidence of discontinuities in the household level covariates. Column (1) of Table 6 shows the estimated discontinuity based on the regression disconti-nuity estimation procedure of Austin Nichols. Columns (2) and (3) show the standard errors and P-values, respectively, of the estimated coefficients. The only significant discontinuities in wealth indicators are found on the surface of land owned by the household. There are also no discontinuities in other important covariates, such as “female headed household” and “literate household head” (Annex 1 Figure A1.j and Figure A1.k).9

9 These variables are not included in the wealth index because they are endogenous (i.e. they are both an indicator of wealth but also a potential cause of household wealth).

24


Table 6. Estimated discontinuities in baseline covariates at the threshold in the wealth index: households interviewed in follow-up survey

HOUSEHOLD-LEVEL INDICATORS DISCONTINUITY STD. ERROR

P-VALUE MEAN FULL

SAMPLE

(1) (2) (3)

TV 0.007 0.049 0.889 0.725

Fridge -0.008 0.066 0.904 0.352

Oven -0.034 0.021 0.103 0.023

Access to electricity -0.013 0.047 0.782 0.737

Bedrooms per capita 0.039 0.028 0.159 0.182

No sanitary facilities -0.005 0.014 0.709 0.051

Floor made of dirt -0.057 0.061 0.354 0.375

Roof made of metal sheets 0.058 0.066 0.379 0.477

Wall made of adobe 0.054 0.047 0.251 0.161

Land surface (in manzanas) 0.359 0.176 0.042 0.513

Female headed household -0.014 0.066 0.836 0.311

Household head is literate -0.013 0.049 0.786 0.818

Notes: Column (1) displays estimated discontinuities in the household indicators displayed in the stub column at the threshold separating the intervention households from the control households (based on the estimation procedure of Austin Nichols). Columns (2) and (3) respectively show the standard errors and P-values of the estimated coefficients.

In terms of the outcome variables, the non-attrited households also look reasonably similar. As shown in Table 7 and Annex Figure A2 to Figure A4, there does not appear to be any significant discontinuity.

25

7. Sample attrition

Table 7. Estimated discontinuities in baseline outcome variables at the threshold in the wealth index: members of households interviewed in follow-up survey

DISCONTINUITY STD. ERROR

P-VALUE MEAN FULL

SAMPLE

(1) (2) (3)

Outcome variables eligible women

Working -0.076 0.072 0.289 0.446

Work in own or household business (self-employed or unpaid family)

-0.025 0.065 0.702 0.297

Monthly wage (conditional on working) -10.885 17.242 0.528 96.767

Average weekly working hours (conditional on working)

8.573 4.085 0.036 36.810

Outcome variables men aged 18+

Working 0.013 0.041 0.748 0.897

Work in own or household business (self-employed or unpaid family)

0.062 0.066 0.349 0.481

Outcome variables children 5-17

In school -0.027 0.032 0.388 0.859

Literate -0.004 0.031 0.910 0.878

Working 0.010 0.040 0.796 0.602

Work exclusively 0.007 0.027 0.792 0.093

Study exclusively -0.028 0.039 0.481 0.351

Work and study 0.002 0.042 0.972 0.508

Idle 0.022 0.019 0.252 0.048

In hazardous work 0.028 0.043 0.515 0.551

In chores 0.022 0.028 0.435 0.866

Weekly hours worked (conditional on working)

8.537 1.828 0.000 17.040

Weekly hours in chores (conditional on being in chores)

3.119 1.033 0.003 12.994

Notes: Column (1) displays estimated discontinuities in outcome variables displayed in the stub column at the threshold separating the in-tervention households from the control households (based on the estimation procedure of Austin Nichols). Columns (2) and (3) respectively show the standard errors and P-values of the estimated coefficients.

27

In the previous sections, we have seen how households close to the threshold for program assignment have similar characteristics at baseline. Also, attrition between the baseline and thefollow-up surveys appears to be random with respect to program assignment, and households in the followup survey continue to have similar baseline characteristics close to the threshold for assignment.

We can, therefore, proceed with Regression Discontinuity (RD) estimates to assess the impact of the program.

As we have seen, the compliance was far from perfect: only 680 out of 1,098 selected beneficiary households took up the program, a take-up rate of about 60%. Moreover, only 522 of 680 treated households were re-interviewed in the follow-up survey due to attrition. For these reasons, we esti-mate both the intent-to-treat and the treatment-on-the-treated effects, as they are likely to differ in our setting. We use the sharp RD design to estimate the intent-to-treat effect (ITT). The treatment-on-the treated effect (TOT) is estimated by applying the fuzzy RD design, using the treatment assignment as an instrument for actual program participation.

To measure the intent-to-treat effect of the program we estimate the following sharp regression dis-continuity equation:

Yij = α + βDj + ∑k≥1 γk (Xj – c)k+ ∑k≥1 δk Dj (Xj – c)k + ϑZi + εi (1)

where c – h≤ X ≤ c + h ;

Yij is the outcome of interest for individual i from household j; α is the intercept; Dj is a dummy tak-ing the value 1 if a household was selected for the interventions (i.e. had e wealth index below the implicit threshold); ∑k≥1 γk (Xj – c)k is a polynomial of order k that captures the relationship between the outcome of interest and the distance of a household's wealth index Xj to the threshold c, while the term ∑k≥1δk Dj (Xj – c)k allows for a different slope below and above the threshold. Zi is a vector of the baseline covariates, εi is the error term, and h is a bandwidth. The standard errors are clustered at the municipality level.

The estimated coefficient β identifies the average local ITT effect of households being selected for the interventions.

The fuzzy RD design can be described by the following two equations system:

Yij=α + βTj+∑k≥1γk (Xj – c)k+∑k≥1δk Dj(Xj – c)k+ϑZi+εi (2)

Tj= σ + τDj + ∑k≥1 ρk (Xj – c)k + ∑k≥1 φk Dj (Xj – c)k + θZi + ϵi (3)

8. METHODOLOGY

28


where Tj is the treatment dummy variable taking the value 1 if a household actually received the treatment; the other variables are as described in the equation (1).

Choice of the optimal bandwidth and order of polynomial

As is well known, the choice of the optimal bandwidth h is the crucial issue for the RD design, since there is always a tradeoff between precision and bias of the estimates. For each out-come, we select the optimal bandwidth using the cross-validation criterion (CVC) proposed by Imbens and Lemieux (2008):

CVCY(h) =(1/N) ∑ N (Yi – Ŷ( Xi ))2,

with the preferred bandwidth given by:

hCV = arg min CVY (h)

In particular, we estimate the CVC starting with a bandwidth of 0.2 and proceeding with steps of 0.2. For the fuzzy RD design, we use the same bandwidth selected for the sharp RD estimates.

We choose the order of polynomial of ∑k≥1 γk (Xj – c)k using the Akaike information criterion (AIC). The AIC does not vary in any relevant way with the order of polynomial (Figure 8) at the optimal bandwidth. Therefore, also for the sake of simplicity and to improve the efficiency we use the first order polynomial.

Figure 9. AIC comparison for estimations of equation (1) for the preferred cross-validation criterion and with different orders of polynomial

i=1

opt

h

29

8. Methodology

The vector Z of baseline covariates for child outcomes includes child age dummies; sex of child; sex, literacy and education level of household head; number of children aged 5-14 years in household; and household wealth characteristics (ownership of TV, refrigerator, oven; access to electricity and sanitary facilities; whether the walls of dwelling are made of adobe; whether the floor of dwelling is made of dirt; whether the roof of dwelling is made of metal sheets; bedrooms per capita; and land surface). The vector Z of baseline covariates for adult outcomes includes a second order age polynomial, education level; number of children aged 5-14 years in household; and household wealth characteristics as described above for children’s outcomes.

31

We now turn to the impact of the program on eligible woman, men from eligible households, and, most importantly, on children. Table 9 to Table 19 show the results of the impact es-timates for both the treatment-on-the-treated (TOT) and the intent-to-treat (ITT) groups; 10 for each outcome variable the tables also report the optimal bandwidth on which the esti-mates are based. Standard errors are clustered at the municipality level. These estimates are based on the optimal bandwidth, a first order polynomial, and include the set of covariates described above. The set of complete results with and without covariates and for different bandwidths can be found in Annex 3. Table 8 presents the definitions of the variables used in the estimates.

Table 8. Variable definitions

School attendance

A child is defined as attending school if he/she is in attendance at any regular accredited educational institution or program for organized learning.

School attendance regularity

A child’s attendance regularity reflects whether school days were missed during the past week.

School expenditures

School expenditures comprise monthly expenditures for enrolment, books, uniform, footwear, other materials, parent fee, tuition, and transport.

Literacy

An individual is defined as literate if he/she can read and write.

Employment

An individual is defined as employed if he/she worked in the past week or in the past 12 months. In particular, if she/he 1) worked in past week at least 1 hour; or 2) has a job but did not work during the past week for some reason (sickness, maternity leave, change of shift, and etc.); or 3) in the past week participated in any cash or in-kind remunerated activities; and/or 4) in the past 12 months worked in a non agricultural enterprise; and/or 5) in the past 12 months was involved in any farming activities; and/or 6) in the past 12 months worked in an enterprise owned by himself/herself.

Household chores

A child is defined as performing household chores if he/she performs at least one of the following activities: cleaning the house; fetching wood; fetching water; taking care of small/old/sick members of the household; cooking; washing dishes; repair work; and others.

Own or household business

An individual is defined as working in own or household business if he/she works as an employer, self-employed or unpaid family worker in the main or secondary activities.

10 Also shown are the standard errors and bandwidths on which the estimates are based and the mean of outcome variable in the full sample.

9. RESULTS

32


Work in hazardous conditions

A child is defined as working in hazardous conditions if his or her work involves: work in dusty, smoky, noisy environment; work in extreme hot or cold; work with dangerous tools, chemical pesticides; work during the night or early morning; carrying heavy loads; work in rivers, lakes or under water.

Child labour

A child is defined as involved in child labour if he or she falls into one of the following categories: aged 5-13 years in employment; aged 14-15 years working in hazardous conditions or employed for more than 34 hours a week; or aged 16-17 working in hazardous conditions.

The estimated effects of the interventions are, not surprisingly, larger for the treatment-on-the-treated group, as it reflects those that actually participated in the program. The results for the intent-to-treat group differ in terms of magnitude, but otherwise are largely consistent with the TOT. Note that most of our findings are robust to the exclusion of covariates, to vari-ation of bandwidths. The cases in which the results are not robust are mentioned in the text.

�Impact on households

Adult labour market outcomes

The program had the expected impact on the employment of eligible women. Among the women actually taking part in the program, employment increased by 40 percentage points as a result of the programme (Table 9a).11 This rise in women’s employment was largely driv-en by involvement in own or household business, which increased by 44 percentage points among those who took up the program.12

11 ITT effect 23 percentage points (Table 9b).

12 ITT effect 25 percentage points (Table 9b).

33

9. Results

Table 9. Impact on adult female labour market outcomes

(A) TREATMENT-ON-THE-TREATED: 2SLS ESTIMATES

Employment Own/household business(1)

Impact estimate 0.403***

(0.123)

0.439**

(0.201)

Bandwidth 0.6 0.6

Num. observations 295 295

Mean full sample 0.54 0.39

(B) INTENT-TO-TREAT: OLS ESTIMATES



(0.084)

0.250**

(0.108)

Bandwidth 0.6 0.6



Notes: all estimates are based on a first order polynomial and include the covariates described in section 8. Standard errors (in parentheses) are clustered at the municipality level. (1) Work as self-employed or unpaid family worker in the main or secondary activities.

* Statistical significance at 10%; ** Statistical significance at 5%; *** Statistical significance at 1%

The increase in women self-employment did not generate a reduction in other forms of eco-nomic activities. There is no significant reduction in the involvement of women in paid work as a consequence of the program. As shown in Annex 3 the impact on female paid employ-ment is very small and not significant.

This increase in female employment does not appear to spill over to other members of the family. Adult males from eligible households did not increase their participation in economic activities or their involvement in own household business. While the TOT estimates do in fact indicate a positive effect on adult male employment, this effect is only marginally significant and also not robust to changes in the bandwidth (Table 10). This result should, therefore, be considered with care.

Table 10. Impact on adult male labour market outcomes



Impact estimate 0.063*

(0.037)

-0.026

(0.082)

Bandwidth 0.6 Full sample

34


Num. observations 331 1,591




Impact estimate 0.036

(0.023)

-0.017

(0.053)

Bandwidth 0.6 full sample

Num. observations 331 1,591


Notes: all estimates are based on a first order polynomial and include the covariates described in section 8. Standard errors (in parentheses) are clustered at the municipality level. (1) Work as employer, self-employed or unpaid family worker in the main or secondary activities.* Statistical significance at 10%; ** Statistical significance at 5%; *** Statistical significance at 1%

Household income and expenditures

Notwithstanding the substantial increase in women participation in economic activity, and in their own business in particular, it is not possible to identify a significant impact on household income and expenditures.

The point estimates indicate a large increase in the labour income of treated women: almost 50 per cent for the TOT estimates (Table 11b). The coefficients are, however, not significant. This might be due to the small sample size at the optimal bandwidth and/or to measurement errors that are known to affect income reporting, especially for non-wage employment. How-ever, as we have no way to test for this, the positive coefficient must be taken at most as a possible indication of direction of impact.

Similarly, the program shows a positive, but small and non-significant, impact on household per capita labour income and household per capita expenditures.

Table 11. Female labour income, household adult labour income per capita and household expenditure per capita

RD ESTIMATES OF THE EFFECTS OF INTERVENTIONS ON LABOUR INCOME OF ELIGIBLE FEMALES; HOUSEHOLD ADULT LABOUR INCOME P.C. AND HOUSEHOLD

TOTAL EXPENDITURE P.C.

(a) Intent-to-treat: OLS estimates

Labour income of eligible females

Household adult labour income

p.c.

Household total expenditure p.c.


(24.493)

2.171

(6.011)

2.383

(2.981)

35

9. Results

Bandwidth 0.8 Full sample Full sample

Num. observations 384 1,463 1,496

Mean full sample 72.60 57.81 30.47

(b) Treatment-on-the-treated: 2SLS estimates

Labour income

of eligible females

Household adult labour income p.c.

Household total expenditure p.c.


(40.715)

3.383

(9.395)

3.704

(4.644)

Bandwidth 0.8 Full sample Full sample

Num. observations 384 1,463 1,496


Notes: all estimates are based on a first order polynomial and include variables displayed in section 8 as controls. Standard errors (in paren-theses) are clustered at the municipality level. * Statistical significance at 10%; ** Statistical significance at 5%; *** Statistical significance at 1%

Household decision-making

The program apparently affected the “balance of power” within the household. The follow-up survey contains a section of decision-making within the household. In particular, it is asked which individual contributes to the decisions regarding a broad set of activities ranging from purchase of food to education, investments, and loans. We have estimated the impact of the project on the probability that a beneficiary woman contributed to the decisions in the set of categories detailed below. As it can be seen in Table 12, the project has substantially increased the beneficiary women's decision-making role within the household with respect to the control group. The impact appears positive for all the categories, but large and significant only for a subset of them. Of particular relevance is the large increase in the role of women with respect to children’s school attendance and on household business investments.

36


Table 12. RD estimates of the effects of interventions on women decision-making within the household

(A) IN

TE

NT-

TO

-TR

EA

T: O

LS

ES

TIM

AT

ES

1.

Purc

hase

s:

Food

2.

Purc

hase

s:

Child

clo

thes

3.

Purc

hase

s:

Ele

ctr

ical

(house

hold

) applia

nces

4.

Child

sc

hool

att

endance

5.

Vis

its

to

hosp

ital

when

child

ren a

re

sick

6.

Invest

ments

to

fam

ily

busi

ness

7.

Fam

ily

pla

nnin

g

8.

Loans

9.

Part

icip

ation

in t

he

com

munity

activitie

s

10

.E

xpense

s to

im

pro

ve

housi

ng

Impa

ct e

stim

ate

0.04

8**

(0.0

22)

0.08

8

(0.0

64)

0.06

5

(0.0

52)

0.11

4*

(0.0

66)

0.10

4

(0.0

82)

0.17

9**

(0.0

84)

0.11

9

(0.1

15)

0.11

0**

(0.0

54)

0.04

8

(0.0

73)

0.04

7**

(0.0

22)

Ban

dwid

th1.

60.

81.

61.

01.

01.

00.

61.

61.

81.

6

Num

. obs

erva

tions

726

365

718

438

447

428

247

648

776

719

Mea

n fu

ll sa

mpl

e0.

840.

830.

740.

830.

840.

710.

710.

630.

750.

72

(B) T

RE

AT

ME

NT-

ON

-TH

E T

RE

AT

ED

: 2

SL

S E

ST

IMA

TE

S

1.

Purc

hase

s:

Food

2.

Purc

hase

s:

Child

clo

thes

3.

Purc

hase

s:

Ele

ctr

ical

(house

hold

) applia

nces

4.

Child

sc

hool

att

endance

5.

Vis

its

to

hosp

ital

when

child

ren a

re

sick

6.

Invest

ments

to

fam

ily

busi

ness

7.

Fam

ily

pla

nnin

g

8.

Loans

9.

Part

icip

ation

in t

he

com

munity

activitie

s

10

.E

xpense

s to

im

pro

ve

housi

ng

Impa

ct e

stim

ate

0.07

6*

(0.0

40)

0.15

0

(0.1

05)

0.10

4

(0.0

89)

0.17

8*

(0.1

04)

0.16

6

(0.1

33)

0.30

0*

(0.1

62)

0.18

3

(0.1

89)

0.17

8**

(0.0

72)

0.08

0

(0.1

27)

0.07

6***

(0.0

29)

Ban

dwid

th1.

60.

81.

61.

01.

01.

00.

61.

61.

81.

6

Num

. obs

erva

tions

726

365

718

438

447

428

247

648

776

719

Mea

n fu

ll sa

mpl

e0.

840.

830.

740.

830.

840.

710.

710.

630.

750.

72

Not

es: a

ll es

timat

es a

re b

ased

on

a fir

st o

rder

pol

ynom

ial a

nd in

clud

e va

riabl

es d

ispl

ayed

in ta

ble

5 as

con

trol

s. S

tand

ard

erro

rs (

in p

aren

thes

es)

are

clus

tere

d at

the

mun

icip

ality

leve

l. *

Stat

istic

al s

igni

fican

ce a

t 10%

; **

Sta

tistic

al s

igni

fican

ce a

t 5%

; ***

Sta

tistic

al s

igni

fican

ce a

t 1%

37

9. Results

�Impact on childrenThe effects of the project on household behavior set the stage for analyzing the impact on children and understanding the reasons for the observed changes, if any.

Children’s education and work

The program had a large impact on children’s school attendance. As reported in Table 13a, the probability of school attendance increased by 7.6 percentage points as a result of the program,13 a substantial increase given the already high level of school attendance at baseline. In fact, if taken at face value, the program appears to have brought about universal school attendance to children in the treated households.

The design of the evaluation does not permit the precise identification of the mechanisms that determined the observed change. It is likely, however, that mothers who were provided livelihood opportunities were in a better position to afford the direct and indirect costs as-sociated with sending their children to school. At the same time, training and sensitization efforts likely helped smooth children’s school (re)entry. These effects, however, may have been partially offset by the increase in the productivity of children’s work due to the increased level of own business activity. The fact that there was also no change in terms of children’s involvement in the household business (Table 17) is an indication that the program did not generate additional demand for children’s work as a result of the increased level of household economic activity.

Table 13. RD estimates of the effects of interventions on children’s school attendance, regularity of school attendance and school expenditures, children aged 5-15 years in the baseline survey


School attendanceRegularity of

school attendance(a)

School expenditures(b)


(0.026)

-0.019

(0.026)

4.868

(3.185)

Bandwidth 1.8 4.2 Full sample

Num. observations 1,308 1,581 2,025



School attendanceRegularity of

school attendance(a)

School expenditures(b)


(0.015)

-0.013

(0.017)

3.191

(2.166)

13 The increase is 2.5 percentage points according to the ITT estimates, Table 13b.

38


Bandwidth 1.8 4.2 Full sample

Num. observations 1,308 1,581 2,025


Notes: (a) Did a child miss any school day during last week? (b) School expenditures include monthly expenditures for enrollment, books, uniform, footwear, other materials, parent fee, tuition and transport. All estimates are based on a first order polynomial and include covar-iates described in section 8. Standard errors (in parentheses) are clustered at the municipality level. * Statistical significance at 10%; ** Statistical significance at 5%; *** Statistical significance at 1%

The regularity of school attendance did not change significantly as a result of the interven-tions (Table 13), indicating that once the children began to attend school they behaved more or less like other students. The fact that ITT regularity of school attendance did not decrease seems to indicate that the children who started to attend school but did not stop working (see discussion below) made their work compatible with education, at least as reflected by the regularity of their school attendance.

We could also check self-reported information on attendance and regularity of attendance against information collected directly at the school level.

The school survey asks school teachers about the regularity of students' school attendance. The survey covers 637 children aged 5-17 years from eligible households. Children's' school attendance from school module almost perfectly coincides with self-reported school attend-ance. The majority of 637 children (603 children) attend school always or regularly according to the school survey. Of the 603 children reported to attend school, 592 did self-report to at-tend school. The data from the school survey appear to confirm the reliability of self-reported information on school attendance.

Table 14. Comparison of attendance from school survey and self-reported school attendance, children aged 5-17 in the follow-up survey

Attendance from school survey

Self-reported school attendance Total

Not attend Attend

Always 10 530 540

Regularly 1 62 63

Sometimes 0 15 15

Infrequently 0 7 7

Never 4 8 12

Total 15 622 637

The impact on school expenditures is positive and large, and robust across bandwidths. This impact is not, however, statistically significant (Table 13). There is hence an indication that the program might also have induced households to increase expenditures on their chil-dren’s education, but this effect cannot be identified precisely.

Children’s involvement in employment and household chores did not change as a result of the program: the coefficient is positive but small and the standard error is much larger than the coefficients themselves (Table 15).

39

9. Results

Table 15. RD estimates of the effects of interventions on children’s involvement in employment and household chores, children aged 5-15 years in the baseline survey


Employment Household chores


(0.084)

0.182

(0.242)

Bandwidth 1.0 0.4




Employment Household chores


(0.055)

0.100

(0.128)

Bandwidth 1.0 0.4



Notes: all estimates are based on a first order polynomial and include covariates described in section 8. Standard errors (in parentheses) are clustered at the municipality level. * Statistical significance at 10%; ** Statistical significance at 5%; *** Statistical significance at 1%

The picture changes somewhat, however, when we look more closely at how the program affected the division of children’s time between work and schooling. The share of children working only (i.e., without also attending school), one of the worst-off categories of child workers, declined significantly. As reported in Table 16a, the share of children working only from treated households declined by 3.8 percentage points as a result of the program.14 Again, while we cannot identify with certainty the specific mechanism of this change, it is worth noting that it was this group of out-of-school children that was specifically targeted by the program with special training and sensitization efforts aimed at encouraging their school (re)entry. The decline in the working-only group was mirrored by an apparent rise in those working and attending school (although this result is not robust across all bandwidths). This suggests that families responded to the program by sending their children to school without fully withdrawing them from work at the same time.

14 The ITT effect was 2.5 percentage points, Table 16b.

40


Table 16. RD estimates of the effects of interventions on child time use, children aged 5-15 years in the baseline survey


Work exclusively Study exclusively Work and study Idle

Impact estimate -0.038***

(0.014)

-0.006

(0.103)

0.121

(0.116)

-0.034

(0.045)

Bandwidth 2.6 1.0 1.0 1.0

Num. observations 1,760 801 801 801

Mean full sample 0.07 0.66 0.17 0.09


Work exclusively Study exclusively Work and study Idle

Impact estimate -0.025***

(0.008)

-0.004

(0.069)

0.081

(0.076)

-0.022

(0.029)

Bandwidth 2.6 1.0 1.0 1.0

Num. observations 1,760 801 801 801

Mean full sample 0.07 0.66 0.17 0.09

Notes: all estimates are based on a first order polynomial and include covariates described in section 8. Standard errors (in parentheses) are clustered at the municipality level. * Statistical significance at 10%; ** Statistical significance at 5%; *** Statistical significance at 1%

The program interventions resulted in a notable reduction in the time children spent in em-ployment. Children worked an estimated 13.6 fewer hours each week due to the program (Table 17a),15 a decline of almost one-half, leaving them more time for their studies and leisure, and reducing the length of their exposure to any workplace hazards. At the same time, families did not fully withdraw their children from work when sending them to school in response to the program. It nonetheless appears that more classroom time served to crowd out time that might have otherwise been spent working.

The reduction in the working hours was not accompanied by a shift in the nature of the work performed by children. Of particular importance in this regard is the fact that the program interventions did not result in a significant change in work in hazardous conditions or in work constituting child labour in accordance with national legislation (Table 17). We tried to iden-tify whether the large reduction in working hours also translated into a reduction of children whose working hours exceeded the hours threshold for light work. However, given the small number of observations for this group and the consequently low power, we could not identify any significant effect. There was also no change in terms of children’s involvement in the household business.

In conclusion, the key impact of the program appears to have been a substitution of work time with classroom time, rather than a reduction in the incidence of work or a change in its characteristics.

15 The ITT effect was 7 hours, Table 17b.

41

9. Results

Table 17. RD estimates of the effects of interventions on child involvement in own or household business, working hours, work in hazardous conditions and child labour, children aged 5-15 years in the baseline survey


Own/household business(1)

Weekly working hours

cond. on working

Work in hazardous

conditions(2)Child labour(3)


(0.122)

-13.593***

(3.327)

0.033

(0.064)

0.030

(0.062)

Bandwidth 1.0 1.0 1.0 1.0

Num. observations

801 195 793 794

Mean full sample 0.20 24.8 0.22 0.23


Own/household business(1)

Weekly working hours

cond. on working

Work in hazardous

conditions(2)Child labour(3)


(0.079)

-6.979**

(3.227)

0.022

(0.042)

0.020

(0.041)

Bandwidth 1.0 1.0 1.0 1.0

Num. observations

801 195 793 794

Mean full sample 0.20 24.8 0.22 0.23

Notes: all estimates are based on a first order polynomial and include covariates described in section 8. Standard errors (in parentheses) are clustered at the municipality level. (1) Work as self-employed or unpaid family worker in the main or secondary activities. (2) Work in hazardous conditions includes: work in dusty, smoky, noisy environment; work in extreme hot or cold; work with dangerous tools, chemical pesticides; work during the night or early morning; carrying heavy loads; work in rivers, lakes or under water. (3) Child labor includes all children aged 5-13 years in employment; children aged 14-15 years working in hazardous conditions or employed for more than 34 hours a week; and children aged 16-17 working in hazardous conditions. * Statistical significance at 10%; ** Statistical significance at 5%; *** Statistical significance at 1%

Parents' attitudes

The program appears to have had some impact on parents' attitudes towards child labour employment and schooling. As reported in Table 18a, treated households are significantly less likely to agree with statements favourable to child labour such as “work keeps children out of trouble” and “working children are stronger and healthier”. Similarly, for schooling, treated households are significantly less likely to agree with statements downplaying the im-portance of schooling such as “education is only one of the factors that influence the success of a person” and “education is not important to success in life” (Table 19a).

42


Table 18. RD estimates of the effects of interventions on the attitudes of the household head towards children’s employment


My children work because it will give them more opportunities in

the future

Work keeps children out of trouble/gives

them direction

Working chil-dren are more responsible

than non-work-ing children

Working children are stronger and healthier than non-working

children


(0.211)

-0.137**

(0.067)

-0.007

(0.056)

-0.729**

(0.298)

Bandwidth 1.8 1.2 3.4 0.2

Num. observations

838 592 1,262 96

Mean full sample 0.62 0.58 0.85 0.62


My children work because it will give them more opportunities in

the future

Work keeps children out of trouble/gives

them direction

Working chil-dren are more responsible

than non-work-ing children

Working children are stronger and healthier than non-working

children


(0.127)

-0.084**

(0.039)

-0.005

(0.036)

-0.403**

(0.197)

Bandwidth 1.8 1.2 3.4 0.2

Num. observations

838 592 1,262 96

Mean full sample 0.62 0.58 0.85 0.62

Notes: all estimates are based on a first order polynomial and include variables displayed in Table 6 as controls. Standard errors (in parenthe-ses) are clustered at the municipality level. * Statistical significance at 10%; ** Statistical significance at 5%; *** Statistical significance at 1%

43

9. Results

Table 19. RD estimates of the effects of interventions on the attitudes of the household head towards children’s schooling


Education is the only way to succeed in life

Education is only one of the

factors that influence the success of a

person

Ensure that my children go to college is far from

my economic possibilities

Education is not important to success in

life


(0.087)

-0.110**

(0.045)

0.014

(0.043)

-0.281***

(0.088)

Bandwidth 1.2 0.4 1.0 0.4

Num. observations

592 184 493 184

Mean full sample 0.76 0.84 0.87 0.56


Education is the only way to succeed in life

Education is only one of the

factors that influence the success of a

person

Ensure that my children go to college is far from

my economic possibilities

Education is not important to success in

life


(0.150)

-0.181**

(0.072)

0.028

(0.075)

-0.465***

(0.174)

Bandwidth 1.2 0.4 1.0 0.4

Num. observations

592 184 493 184

Mean full sample 0.76 0.84 0.87 0.56

Notes: all estimates are based on a first order polynomial and include variables displayed in Table 6 as controls. Standard errors (in parenthe-ses) are clustered at the municipality level. * Statistical significance at 10%; ** Statistical significance at 5%; *** Statistical significance at 1%

45

There are no impact evaluations of programs targeting child labour through support to wom-en’s businesses that we can use to frame the results presented here. However, some limited evidence exists on the impact of programs providing business training and seed capital to households. This evidence is reviewed in detail in De Hoop and Rosati (2013), from which the discussion below is taken.

Banerjee et al. (2011) study the effects of a program in India targeting women in the poorest of the poor households aimed at lifting them out of poverty by improving their income-gener-ating capacity. They find that the program increased the time spent at school by children, but did not reduce the time spent working. A comparable program (implemented by the same NGO) in Bangladesh (Bandera et al., 2013) resulted in a small but statistically significant increase in the working hours of children in economic activities and household business in particular.

Evidence from Nicaragua’s Results-Based Initiative, which provided business training and start-up capital to selected women in poor rural communities (De Hoop, Rosati and Vakis, 2015), points to an increase in children’s school attendance without any significant change in the involvement of children in economic activities.

In the light of this evidence, the impact of the El Salvador program was substantially positive, although the IE design does not allow us to identify the mechanisms through which it oper-ated.16 The program analysed in this Report produced substantial changes in household be-haviour. The participation of women in economic activities, and in own business in particular, increased substantially, but without generating relevant changes in the labour supply of other members of the households.

Children appear to have benefited substantially from the program. Their school attendance increased significantly and household expenditures on education showed signs (albeit not robust) of an increase. Children were not driven to work in the newly created or expanded household businesses, not an unusual outcome in this kind of programs, but they did not stop working either. Rather, they experienced a marked reduction (by about one-half) in their weekly working hours, making more than enough room for a regular school attendance.

The data do not allow, because of lack of power, the determination of whether the reduction in working hours also resulted in a reduction of the number of children working excessive hours. Even if we cannot identify a reduction in the child labour headcount, it is clear that the program succeeded in improving the current conditions and the future perspectives of

16 This is a common situation when there is more than one possible channel through which a program can affect the relevant outcomes. The random or as good as random (as in the present case) variation allows to identify the reduced formed effect, but not the structural parameters. A much more complex, and often unfeasible, design would be required to identify the structural parameters and hence the relative importance of the different channels.

10. CONCLUSIONS

46


child labourers. Whether such improvements will be sustained in the long run is, of course, a question that remains open, but goes beyond the aim of the present impact evaluation.

With only one instrument, namely the allocation of the program based on the wealth index, it is not possible to identify the exact mechanism that was behind the observed results. However, the evidence shows that the increased involvement in economic activity of women did result in a substantial increase in their decision-making capability within the household and in a possible, albeit not well identified, increase in household income. These two effects appear to have dominated any possible increase in child labour due to a higher demand for children’s time generated by the expanded household business activities.

47

Bandiera, O., R. Burgess, N. Das, S. Gulesci, I. Rasul, M. Sulaiman. 2013. “Can Basic Entrepreneurship Transform the Lives of the Poor?” Working Paper.

Banerjee, A., Duflo, E., Goldberg, N., Karlan, D., Osei, R., Parienté, J. Shapiro, B. Thuysbaert and Udry, C. (2015). A multifaceted program causes lasting progress for the very poor: Evidence from six countries. Science, 348(6236).

Banerjee, A., E. Duflo, R. Chattopadhyay, and J. Shapiro. 2011. “Targeting the Hard-Core Poor: An Impact Assessment.” Working paper.

Cho, Y., and Honorati, M. (2014). Entrepreneurship programs in developing countries: A meta regression analysis. Labour Economics, 28.

De Hoop, J and F. Rosati (2013), The complex effects of public policy on child labour, UCW Working Paper

De Hoop, J., Premand, P, Rosati, F. and R. Vakis. 2015, The Effects of Promoting Women’s Productive Capacity on Children's Work and Schooling.

Eswaran, M and A. Kotwal (1986), “Access to Capital and Agrarian Production Organisation”, Economic Journal, 96 (382), 482-498.

Imbens G. and T. Lemieux. 2008. “Regression Discontinuity Designs: A Guide to Practice.” Journal of Econometrics, 142 (2), 615-635.

Lee, D. S. and T. Lemieux. 2010. “Regression Discontinuity Designs in Economics.” Journal of Economic Literature, 48, 281-355.

Manacorda M. (2009) "Criteria and Guidelines fro the Evaluation of the Impact of Child Labour Interventions" UCW Working Paper

McCrary, J. 2008. “Manipulation of the Running Variable in the Regression Discontinuity Design: A Density Test.” Journal of Econometrics, 142 (2), 698-714.

UCW. 2012. "Impact Evaluation Design: Eliminating Child Labour in El Salvador through Economic Empowerment and Social Inclusion".

UCW. 2013a. "Baseline report: Eliminating Child Labour in El Salvador through Economic Empowerment and Social Inclusion".

UCW. 2013b. "Entendiendo los resultados del trabajo infantil y el empleo juvenil en El Salvador".

11. REFERENCES

48


OIT and DEICO. 2015. “Ejecución de Encuestas de resultados e Impacto del Proyecto: Elim-inación del trabajo infantil en El Salvador a través del empoderamiento económico y la in-clusión social”, www.deicoasociados.com.

49

�Annex 1. Variable discontinuity: Scatter plots

Figure A1. Household-level wealth indicators

(a) Household owns a TV (b) Household owns a fridge

(c) Household owns an oven (d) Access to electricity

(e) Number of bedrooms per capita (f) Dwelling has no sanitary facilities

12. ANNEXES

50


Figure A1.Cont’d

(g) Floor made of dirt (h) Roof made of metal sheets

(i) Walls made of adobe (j) Female-headed household

(k) Household head literate

51

Annexes

Figure A2. Outcome variables: eligible women

(a) Eligible women in work (b) Eligible woman in own or household business

(c) Monthly wage of eligible women (d) Weekly hours worked by eligible women

Figure A3. Outcome variables: men aged 18+

(a) Adult males in work (b) Adult males in own or household business

52


Figure A3. Outcome variables children: aged 5-17 years

(a) Children aged 5-17 years in school (b) Children aged 5-17 years literate

(c) Children aged 5-17 years in work (d) Children aged 5-17 years in work exclusively

(e) Children aged 5-17 years in school exclusively (f) Children aged 5-17 years in work and in school simultaneously

53

Annexes

Figure A4.Cont’d

(g) Children aged 5-17 years in neither work nor school (h) Children aged 5-17 years in hazardous work

(j) Children aged 5-17 years in chores(i) Weekly hours worked by children aged 5-17 years (conditional on working)

(k) Weekly hours worked in chores by children aged 5-17 years (conditional on working in chores)

54


�Annex 2. Estimates of the effect of interventions: scatter plots

Figure A4. Impact of interventions: children aged 5-15 years at the time of the baseline survey

(a) Impact on school attendance (b) Impact on employment

(c) Impact on being in employment exclusively (d) Impact on being in school exclusively

(e) Impact on being in employment and school simultaneously (f) Impact on being neither in work nor school

55

Annexes

Table A5.Cont’d

(g) Impact on working in hazardous conditions (h) Impact on weekly working hours

(i) Impact on involvement in child labour (j) Impact on being involved in household chores

(k) Impact on participation in own/household business (l) Impact on school expenditures

(m) Impact on occurrence of missed school days during last week

Notes: Dots present local averages at a bin size of 0.1 and the lines represent linearly fitted regressions.

56


Figure A5. Impact of interventions: eligible females

(a) Impact on the participation in employment (b) Impact on participation in own/household business

Notes: Dots present local averages at a bin size of 0.1 and the lines represent linearly fitted regressions.

Figure A7. Impact of interventions: adult males from eligible households

(a) Impact on the participation in employment (b) Impact on the participation in own/household business

Notes: proportion of adult males worked as self-employed or unpaid family workers in year 2015 as a function of the intervention forcing vari-able. Dots present local averages at a bin size of 0.1 and the lines represent linearly fitted regressions.

57

Annexes

Tabl

e A

1.

RD

Est

imat

es o

f th

e ef

fect

of

inte

rven

tion

s on

chi

ld o

utco

me

vari

able

s, c

hild

ren

aged

5-1

5 in

the

bas

elin

e su

rvey

Inte

nt-t

o-tr

eat:

OLS

est

imat

es

BA

ND

WID

THO

PT

IMA

LS

MA

LL

(1

.0)

MID

DL

E(3

.0)

FU

LL

SA

MP

LE

ME

AN

F

UL

L

SA

MP

LE

Em

ploy

No

covari

ate

sW

ith

Covari

ate

sN

o

covari

ate

sW

ith

Covari

ate

sN

o

covari

ate

sW

ith

Covari

ate

sN

o

covari

ate

sW

ith

Covari

ate

s

0.25

Pol

ynom

ial o

f ord

er 1

0.05

9

(0.0

58)

0.02

6

(0.0

55)

-0.0

14

(0.0

69)

0.05

1

(0.0

57)

0.04

8

(0.0

46)

0.01

4

(0.0

53)

0.01

6

(0.0

30)

Num

obs

erva

tions

(ban

dwid

th)

1,43

4

(2.0

)

801

(1.0

)

801

(1.0

)

opt

1,89

8

(3.0

)

1,89

8

(3.0

)

2,43

02,

430

Att

end

Pol

ynom

ial o

f ord

er 1

0.03

9*

(0.0

21)

0.04

6***

(0.0

15)

0.06

5***

(0.0

21)

0.07

7***

(0.0

22)

0.03

9

(0.0

38)

0.03

8

(0.0

31)

0.02

2

(0.0

27)

0.00

9

(0.0

21)

0.83

Num

obs

erva

tions

(ban

dwid

th)

1,30

8

(1.8

)

1,30

8

(1.8

)

801

(1.0

)

801

(1.0

)

1,89

8

(3.0

)

1,89

8

(3.0

)

2,43

02,

430

Wor

k on

ly

Pol

ynom

ial o

f ord

er 1

-0.0

28**

*

(0.0

09)

-0.0

25**

*

(0.0

08)

-0.0

59**

*

(0.0

22)

-0.0

54*

(0.0

31)

-0.0

23**

(0.0

12)

-0.0

22**

*

(0.0

08)

-0.0

33

(0.0

22)

-0.0

28**

(0.0

11)

0.07

Num

obs

erva

tions

(ban

dwid

th)

1,76

0

(2.6

)

1,76

0

(2.6

)

801

(1.0

)

801

(1.0

)

1,89

8

(3.0

)

1,89

8

(3.0

)

2,43

02,

430

Stu

dy o

nly

Pol

ynom

ial o

f ord

er 1

-0.0

59

(0.0

79)

-0.0

04

(0.0

69)

0.02

0

(0.0

79)

-0.0

36

(0.0

79)

-0.0

32

(0.0

64)

-0.0

26

(0.0

58)

-0.0

35

(0.0

38)

0.66

Num

obs

erva

tions

(ban

dwid

th)

1,30

8

(1.8

)

801

(1.0

)

801

(1.0

)

opt

1,89

8

(3.0

)

1,89

8

(3.0

)

2,43

02,

430

Wor

k an

d S

tudy

Pol

ynom

ial o

f ord

er 1

0.09

3*

(0.0

49)

0.08

1

(0.0

76)

0.04

5

(0.0

86)

0.07

5

(0.0

46)

0.07

1*

(0.0

38)

0.04

7

(0.0

36)

0.04

5*

(0.0

26)

0.17

Num

obs

erva

tions

(ban

dwid

th)

1,43

4

(2.0

)

801

(1.0

)

801

(1.0

)

opt

1,89

8

(3.0

)

1,89

8

(3.0

)

2,43

02,

430

�Annex 3. RD Estimates of the effect of interventions: Full results

58


BA

ND

WID

THO

PT

IMA

LS

MA

LL

(1

.0)

MID

DL

E(3

.0)

FU

LL

SA

MP

LE

ME

AN

F

UL

L

SA

MP

LE

Idle

Pol

ynom

ial o

f ord

er 1

-0.0

06

(0.0

30)

-0.0

22

(0.0

29)

-0.0

15

(0.0

28)

-0.0

16

(0.0

26)

0.01

1

(0.0

12)

0.01

9

(0.0

14)

Num

obs

erva

tions

(ban

dwid

th)

801

(1.0

)

801

(1.0

)

opt

opt

1,89

8

(3.0

)

1,89

8

(3.0

)

2,43

02,

430

Wee

kly

wor

king

hou

rs

cond

. on

w

orki

ng

Pol

ynom

ial o

f ord

er 1

-9.9

71**

*

(1.9

38)

-6.9

79**

(3.2

27)

-9.6

34**

(4.7

27)

-9.5

17**

*

(2.9

72)

-7.9

80**

*

(3.2

27)

-8.0

17**

*

(2.5

51)

-6.9

44**

*

(2.5

77)

Num

obs

erva

tions

(ban

dwid

th)

225

(1.2

)

195

(1.0

)

195

(1.0

)

opt

430

(3.0

)

430

(3.0

)

574

574

Wor

k in

ha

zard

ous

cond

itio

ns(1

)

Pol

ynom

ial o

f ord

er 1

0.04

0

(0.0

54)

0.02

2

(0.0

42)

-0.0

17

(0.0

61)

0.03

7

(0.0

52)

0.03

8

(0.0

41)

-0.0

08

(0.0

51)

-0.0

05

(0.0

28)

Num

obs

erva

tions

(ban

dwid

th)

1,82

5

(2.8

)

793

(1.0

)

793

(1.0

)

opt

1,88

3

(3.0

)

1,88

3

(3.0

)

2,41

32,

413

Chi

ld

labo

ur(2

)

Pol

ynom

ial o

f ord

er 1

0.04

3

(0.0

54)

0.02

0

(0.0

41)

-0.0

15

(0.0

61)

0.03

6

(0.0

52)

0.03

6

(0.0

42)

0.00

1

(0.0

50)

0.00

5

(0.0

27)

Num

obs

erva

tions

(ban

dwid

th)

1,42

5

(2.0

)

794

(1.0

)

794

(1.0

)

opt

1,88

7

(3.0

)

1,88

7

(3.0

)

2,41

92,

419

In o

wn

or

hous

ehol

d bu

sine

ss

as m

ain

or

seco

ndar

y ac

tivi

ty(3

)

Pol

ynom

ial o

f or

der

10.

058

(0.0

59)

0.04

6

(0.0

79)

-0.0

05

(0.0

91)

0.04

4

(0.0

69)

0.04

0

(0.0

57)

0.01

3

(0.0

59)

0.01

1

(0.0

41)

Num

obs

erva

tions

(ban

dwid

th)

1,43

4

(2.0

)

801

(1.0

)

801

(1.0

)

opt

1,89

8

(3.0

)

1,89

8

(3.0

)

2,43

02,

430

In h

ouse

hold

ch

ores

Pol

ynom

ial o

f or

der

10.

120

(0.1

45)

0.10

0

(0.1

28)

0.12

0

(0.1

45)

0.13

4

(0.1

44)

0.00

0

(0.0

91)

0.01

8

(0.1

00)

-0.0

16

(0.0

89)

0.00

6

(0.0

88)

Num

obs

erva

tions

(ban

dwid

th)

1.0

(868

)

339

(0.4

)

1.0

(868

)

868

(1.0

)

2,06

5

(3.0

)

2,06

5

(3.0

)

2,62

32,

623

59

Annexes

BA

ND

WID

THO

PT

IMA

LS

MA

LL

(1

.0)

MID

DL

E(3

.0)

FU

LL

SA

MP

LE

ME

AN

F

UL

L

SA

MP

LE

Chi

ld

atte

ndan

ce

regu

lari

ty

duri

ng la

st

wee

k

Pol

ynom

ial o

f or

der

1-0

.012

(0.0

23)

-0.0

13

(0.0

17)

0.01

6

(0.0

32)

0.04

0

(0.0

44)

-0.0

10

(0.0

25)

-0.0

12

(0.0

25)

-0.0

25

(0.0

26)

-0.0

19

(0.0

14)

Num

obs

erva

tions

(ban

dwid

th)

1,58

1(4

.2)

1,58

1(4

.2)

598

(1.0

)

598

(1.0

)

1,39

3

(3.0

)

1,39

3

(3.0

)

1,76

81,

768

Mon

thly

sc

hool

ex

pend

itur

es

Pol

ynom

ial o

f or

der

11.

712

(1.7

79)

3.19

1

(2.1

66)

4.86

4

(3.6

73)

4.90

4

(6.1

00)

3.69

9

(2.6

92)

4.12

9

(3.2

30)

Num

obs

erva

tions

(ban

dwid

th)

2,02

5

(ful

l sam

ple)

2,02

5

(ful

l sam

ple)

681

(1.0

)

681

(1.0

)

1,59

9

(3.0

)

1,59

9

(3.0

)

opt

opt

Not

es:

(1) W

ork

in h

azar

dous

con

ditio

ns in

clud

es: w

ork

in d

usty

, sm

oky,

noi

sy e

nviro

nmen

t; w

ork

in e

xtre

me

hot o

col

d; w

ork

with

dan

gero

us to

ols,

che

mic

al p

estic

ides

; wor

k du

ring

nigh

t or e

arly

mor

ning

; car

ryin

g he

avy

load

s; w

ork

in r

iver

s, la

kes

or u

nder

wat

er. (

2) C

hild

labo

r co

mpr

ises

all

child

ren

aged

5-1

3 ye

ars

in e

mpl

oym

ent;

child

ren

aged

14-

15 y

ears

wor

king

in h

azar

dous

con

ditio

ns o

r em

ploy

ed fo

r m

ore

than

34

hour

s a

wee

k; a

nd c

hild

ren

aged

16-

17 w

orki

ng in

haz

ardo

us c

ondi

tions

. (3

) W

ork

as e

mpl

oyer

s, s

elf-

empl

oyed

or

unpa

id f

amily

wor

ker

in t

he m

ain

or s

econ

dary

act

iviti

es.

Stan

dard

err

ors

(in p

aren

thes

es)

are

clus

tere

d at

the

mun

icip

ality

leve

l. *

Stat

istic

al s

igni

fican

ce a

t 10%

; **

Stat

istic

al s

igni

fican

ce a

t 5%

; ***

Sta

tistic

al s

igni

fican

ce a

t 1%

60


Tabl

e A

2.

RD

Est

imat

es o

f th

e ef

fect

of

inte

rven

tion

s on

chi

ld o

utco

me

vari

able

s, c

hild

ren

aged

7-1

5 in

the

fol

low

-up

surv

ey

Inte

nt-t

o-tr

eat:

OLS

est

imat

es

BA

ND

WID

THO

PT

IMA

LS

MA

LL

(1

.0)

MID

DL

E(3

.0)

FU

LL

SA

MP

LE

ME

AN

F

UL

L

SA

MP

LE

Chi

ld la

bour

bas

ed

on t

he w

ork

for

exce

ssiv

e ho

urs

(1)

No

covari

ate

sW

ith

Covari

ate

sN

o

covari

ate

sW

ith

Covari

ate

sN

o

covari

ate

sW

ith

Covari

ate

sN

o

covari

ate

sW

ith

Covari

ate

s

0.11

Pol

ynom

ial o

f ord

er 1

0.03

5

(0.0

44)

0.05

0

(0.0

37)

0.03

5

(0.0

58)

0.04

0

(0.0

39)

0.04

5

(0.0

36)

0.02

5

(0.0

19)

0.03

0

(0.0

19)

Num

obs

erva

tions

(ban

dwid

th)

1,13

8

(2.2

)

5891

(1.0

)

589

(1.0

)

opt

1,40

9

(3.0

)

1,40

9

(3.0

)

1,81

81,

818

Not

es:

Stan

dard

err

ors

(in p

aren

thes

es)

are

clus

tere

d at

the

mun

icip

ality

leve

l. (1

) C

hild

labo

r ba

sed

on w

ork

for

exce

ssiv

e ho

urs

com

pris

es a

ll ch

ildre

n ag

ed 5

-13

year

s in

em

ploy

men

t; ch

ildre

n ag

ed 1

4-15

yea

rs e

mpl

oyed

for

mor

e th

an 3

4 ho

urs

a w

eek.

* S

tatis

tical

sig

nific

ance

at 1

0%; *

* St

atis

tical

sig

nific

ance

at 5

%; *

** S

tatis

tical

sig

nific

ance

at 1

%

61

Annexes

Tabl

e A

3.

RD

Est

imat

es o

f th

e ef

fect

of

inte

rven

tion

s on

chi

ld o

utco

me

vari

able

s, c

hild

ren

aged

5-1

5 y

ears

in t

he b

asel

ine

surv

ey

Trea

tmen

t-on

-the

-tre

ated

: 2

SLS

est

imat

es

BA

ND

WID

TH

OP

TIM

AL

SM

AL

L

(1.0

)M

IDD

LE

(3.0

)F

UL

L S

AM

PL

EM

EA

N

FU

LL

S

AM

PL

E

Em

ploy

No

covariate

sW

ith

Covari

ate

sN

o

covari

ate

sW

ith

Covari

ate

sN

o

covari

ate

sW

ith

Covari

ate

sN

o

covari

ate

sW

ith

Covari

ate

s

0.25

Pol

ynom

ial o

f ord

er 1

0.09

6

(0.0

94)

0.03

9

(0.0

84)

-0.0

21

(0.1

00)

0.07

9

(0.0

87)

0.07

4

(0.0

70)

0.02

2

(0.0

79)

0.02

5

(0.0

44)

Num

obs

erva

tions

(ban

dwid

th)

1434

(2.0

)

801

(1.0

)

801

(1.0

)

opt

1,89

8

(3.0

)

1,89

8

(3.0

)

2,43

02,

430

Att

end

Pol

ynom

ial o

f ord

er 1

0.06

4*

(0.0

36)

0.07

6***

(0.0

26)

0.09

5***

(0.0

35)

0.11

5***

(0.0

40)

0.05

9

(0.0

61)

0.05

8

(0.0

50)

0.03

3

(0.0

42)

0.01

4

(0.0

33)

0.83

Num

obs

erva

tions

(ban

dwid

th)

1,30

8

(1.8

)

1,30

8

(1.8

)

801

(1.0

)

801

(1.0

)

1,89

8

(3.0

)

1,89

8

(3.0

)

2,43

02,

430

Wor

k on

ly

Pol

ynom

ial o

f ord

er 1

-0.0

44**

*

(0.0

17)

-0.0

38**

*

(0.0

14)

-0.0

86**

*

(0.0

30)

-0.0

81*

(0.0

46)

-0.0

36*

(0.0

19)

-0.0

34**

*

(0.0

12)

-0.0

50

(0.0

36)

-0.0

42**

(0.0

19)

0.07

Num

obs

erva

tions

(ban

dwid

th)

1,76

1

(2.6

)

1,76

0

(2.6

)

801

(1.0

)

801

(1.0

)

1,89

8

(3.0

)

1,89

8

(3.0

)

2,43

02,

430

Stu

dy o

nly

Pol

ynom

ial o

f ord

er 1

-0.0

97

(0.1

29)

-0.0

06

(0.1

03)

0.02

9

(0.1

16)

-0.0

55

(0.1

20)

-0.0

49

(0.0

96)

-0.0

39

(0.0

86)

-0.0

53

(0.0

56)

0.66

Num

obs

erva

tions

(ban

dwid

th)

1,30

8

(1.8

)

801

(1.0

)

801

(1.0

)

opt

1,89

8

(3.0

)

1,89

8

(3.0

)

2,43

02,

430

Wor

k an

d S

tudy

Pol

ynom

ial o

f ord

er 1

0.15

2*

(0.0

82)

0.12

1

(0.1

16)

0.06

6

(0.1

27)

0.11

4

(0.0

72)

0.10

8*

(0.0

59)

0.07

2

(0.0

52)

0.06

7*

(0.0

38)

0.17

Num

obs

erva

tions

(ban

dwid

th)

1434

(2.0

)

801

(1.0

)

801

(1.0

)

opt

1,89

8

(3.0

)

1,89

8

(3.0

)

2,43

02,

430

62


BA

ND

WID

TH

OP

TIM

AL

SM

AL

L

(1.0

)M

IDD

LE

(3.0

)F

UL

L S

AM

PL

EM

EA

N

FU

LL

S

AM

PL

E

Idle

Pol

ynom

ial o

f ord

er 1

-0.0

08

(0.0

44)

-0.0

34

(0.0

45)

-0.0

23

(0.0

44)

-0.0

24

(0.0

41)

0.01

7

(0.0

18)

0.02

8

(0.0

21)

0.09

Num

obs

erva

tions

(ban

dwid

th)

801

(1.0

)

801

(1.0

)

opt

opt

1,89

8

(3.0

)

1,89

8

(3.0

)

2,43

02,

430

Wee

kly

wor

king

ho

urs

cond

. on

w

orki

ng

Pol

ynom

ial o

f ord

er 1

-19.

501*

**

(4.3

86)

-13.

593*

**

(3.3

27)

-17.

784*

*

(7.3

95)

-16.

314*

**

(4.7

80)

-12.

947*

**

(4.1

19)

-12.

551*

**

(4.7

76)

-10.

298*

*

(4.3

84)

24.8

Num

obs

erva

tions

(ban

dwid

th)

225

(1.2

)

195

(1.0

)

195

(1.0

)

opt

430

(3.0

)

430

(3.0

)

574

Wor

k in

haz

ardo

us

cond

itio

ns

Pol

ynom

ial o

f ord

er 1

0.06

2

(0.0

84)

0.03

3

(0.0

64)

-0.0

25

(0.0

89)

0.05

7

(0.0

82)

0.05

7

(0.0

64)

-0.0

12

(0.0

78)

-0.0

08

(0.0

42)

0.22

Pol

ynom

ial o

f ord

er 3

0.04

4

(0.1

22)

0.23

5

(0.1

44)

0.11

6**

(0.0

49)

0.03

4

(0.1

28)

0.04

4

(0.0

85)

0.03

6

(0.1

46)

0.03

8

(0.1

28)

Num

obs

erva

tions

(ban

dwid

th)

1,82

5

(2.8

)

793

(1.0

)

793

(1.0

)

opt

1,88

3

(3.0

)

1,88

3

(3.0

)

2,41

3

Chi

ld la

bour

Pol

ynom

ial o

f ord

er 1

0.07

1

(0.0

90)

0.03

0

(0.0

62)

-0.0

21

(0.0

89)

0.05

5

(0.0

81)

0.05

5

(0.0

65)

0.00

2

(0.0

75)

0.00

8

(0.0

41)

0.23

Num

obs

erva

tions

(ban

dwid

th)

1,42

5

(2.0

)

794

(1.0

)

794

(1.0

)

opt

1,88

7

(3.0

)

1,88

7

(3.0

)

2,41

9

In o

wn

or h

ouse

hold

bu

sine

ss a

s m

ain

or

seco

ndar

y ac

tivi

ty

Pol

ynom

ial o

f ord

er 1

0.09

5

(0.0

98)

0.06

8

(0.1

22)

-0.0

07

(0.1

32)

0.06

8

(0.1

06)

0.06

1

(0.0

86)

0.01

9

(0.0

89)

0.01

6

(0.0

61)

0.20

Num

obs

erva

tions

(ban

dwid

th)

1,43

4

(2.0

)

801

(1.0

)

801

(1.0

)

opt

1,89

8

(3.0

)

1,89

8

(3.0

)

2,43

02,

430

In h

ouse

hold

cho

res

Pol

ynom

ial o

f ord

er 1

0.17

9

(0.2

36)

0.18

2

(0.2

42)

0.20

3

(0.2

42)

0.00

1

(0.1

39)

0.02

8

(0.1

50)

-0.0

24

(0.1

34)

0.01

0

(0.1

30)

0.54

Num

obs

erva

tions

(ban

dwid

th)

868

(1.0

)

339

(0.4

)

opt

868

(1.0

)

2,06

5

(3.0

)

2,06

5

(3.0

)

2,62

32,

623

63

Annexes

BA

ND

WID

TH

OP

TIM

AL

SM

AL

L

(1.0

)M

IDD

LE

(3.0

)F

UL

L S

AM

PL

EM

EA

N

FU

LL

S

AM

PL

E

Chi

ld a

tten

danc

e re

gula

rity

dur

ing

last

w

eek

Pol

ynom

ial o

f ord

er 1

-0.0

19

(0.0

34)

-0.0

19

(0.0

26)

0.02

1

(0.0

44)

0.05

5

(0.0

60)

-0.0

16

(0.0

38)

-0.0

18

(0.0

38)

-0.0

39

(0.0

38)

-0.0

29

(0.0

21)

0.84

Num

obs

erva

tions

(ban

dwid

th)

1,58

1

(4.2

)

1,58

1

(4.2

)

598

(1.0

)

598

(1.0

)

1,39

3

(3.0

)

1,39

3

(3.0

)

1,76

81,

768

Mon

thly

sch

ool

expe

ndit

ures

Pol

ynom

ial o

f ord

er 1

2.65

5

(2.7

71)

4.86

8

(3.1

85)

7.01

5

(5.2

71)

7.20

2

(8.7

04)

5.70

7

(4.0

08)

6.30

3

(4.6

97)

9.32

Num

obs

erva

tions

(ban

dwid

th)

2,02

5

(full

sam

ple)

2,02

5

(full

sam

ple)

681

(1.0

)

681

(1.0

)

1,59

9

(3.0

)

1,59

9

(3.0

)

opt

opt

Not

es: (

1) W

ork

in h

azar

dous

con

ditio

ns in

clud

es: w

ork

in d

usty

, sm

oky,

noi

sy e

nviro

nmen

t; w

ork

in e

xtre

me

hot o

col

d; w

ork

with

dan

gero

us to

ols,

che

mic

al p

estic

ides

; wor

k du

ring

nigh

t or

early

mor

ning

; car

ryin

g he

avy

load

s; w

ork

in r

iver

s, la

kes

or u

nder

wat

er. (

2) C

hild

labo

r co

mpr

ises

all

child

ren

aged

5-1

3 ye

ars

in e

mpl

oym

ent;

child

ren

aged

14-

15 y

ears

wor

king

in h

azar

dous

con

ditio

ns o

r em

ploy

ed fo

r m

ore

than

34

hour

s a

wee

k; a

nd c

hild

ren

aged

16-

17

wor

king

in h

azar

dous

con

ditio

ns.

(3)

Wor

k as

em

ploy

ers,

sel

f-em

ploy

ed o

r un

paid

fam

ily w

orke

r in

the

mai

n or

sec

onda

ry a

ctiv

ities

. St

anda

rd e

rror

s (in

par

enth

eses

) ar

e cl

uste

red

at t

he m

unic

ipal

ity le

vel.

* St

atis

tical

sig

nific

ance

at

10%

; **

Stat

istic

al s

igni

fican

ce a

t 5%

; ***

Sta

tistic

al s

igni

fican

ce a

t 1%

64


Tabl

e A

4.

RD

Est

imat

es o

f th

e ef

fect

of

inte

rven

tion

s on

chi

ld o

utco

me

vari

able

s, c

hild

ren

aged

7-1

5 in

the

fol

low

-up

surv

ey

Trea

tmen

t-on

-the

-tre

ated

: 2

SLS

est

imat

es

BA

ND

WID

TH

OP

TIM

AL

SM

AL

L

(1.0

)M

IDD

LE

(3.0

)F

UL

L S

AM

PL

EM

EA

N

FU

LL

S

AM

PL

E

Chi

ld la

bour

bas

ed

on th

e w

ork

for

exce

ssiv

e ho

urs

(1)

No

covari

ate

sW

ith

Covari

ate

sN

o

covari

ate

sW

ith

Covari

ate

sN

o

covari

ate

sW

ith

Covari

ate

sN

o

covari

ate

sW

ith

Covari

ate

s

0.11

Pol

ynom

ial o

f ord

er 1

0.05

5

(0.0

74)

0.07

1

(0.0

65)

0.04

9

(0.0

87)

0.06

0

(0.0

61)

0.06

7

(0.0

57)

0.03

9

(0.0

29)

0.04

4

(0.0

30)

Num

obs

erva

tions

(ban

dwid

th)

1,13

8

(2.2

)

589

(1.0

)

589

(1.0

)

opt

1,40

9

(3.0

)

1,40

9

(3.0

)

1,81

81,

818

Not

es: (

1) C

hild

labo

r ba

sed

on w

ork

for

exce

ssiv

e ho

urs

com

pris

es a

ll ch

ildre

n ag

ed 5

-13

year

s in

em

ploy

men

t; ch

ildre

n ag

ed 1

4-15

yea

rs e

mpl

oyed

for

mor

e th

an 3

4 ho

urs

a w

eek.

Sta

ndar

d er

rors

(in

par

enth

eses

) ar

e cl

uste

red

at

the

mun

icip

ality

leve

l. *

Stat

istic

al s

igni

fican

ce a

t 10%

; **

Stat

istic

al s

igni

fican

ce a

t 5%

; ***

Sta

tistic

al s

igni

fican

ce a

t 1%

65

Annexes

Tabl

e A

5.

R

D E

stim

ates

of

the

effe

ct o

f in

terv

enti

ons

on e

mpl

oym

ent

of e

ligib

le f

emal

es a

nd a

dult

mal

es f

rom

elig

ible

hou

seho

lds

Inte

nt-t

o-tr

eat:

OLS

est

imat

es

BA

ND

WID

TH

OP

TIM

AL

SM

AL

L

(1.0

)M

IDD

LE

(3.0

)F

UL

L S

AM

PL

EM

EA

N

FU

LL

S

AM

PL

E

No

covari

ate

sW

ith

Covari

ate

sN

o

covari

ate

sW

ith

Covari

ate

sN

o

covari

ate

sW

ith

Covari

ate

sN

o

covari

ate

sW

ith

Covari

ate

s

Fem

ale

in

empl

oym

ent

Pol

ynom

ial o

f or

der

10.

234*

**

(0.0

87)

0.23

0***

(0.0

84)

0.24

0***

(0.0

74)

0.21

4***

(0.0

83)

0.24

3***

(0.0

56)

0.27

3***

(0.0

65)

0.16

4***

(0.0

33)

0.19

4***

(0.0

39)

0.54

Num

obs

erva

tions

(ban

dwid

th)

295

(0.6

)

295

(0.6

)

470

(1.0

)

470

(1.0

)

869

(2.0

)

869

(2.0

)

1,39

91,

399

Fem

ale

in o

wn

or

hous

ehol

d bu

sine

ss(1

)

Pol

ynom

ial o

f or

der

10.

320*

**

(0.0

93)

0.25

0**

(0.1

08)

0.28

2***

(0.0

93)

0.22

8***

(0.0

80)

0.26

8***

(0.0

24)

0.27

6***

(0.0

19)

0.18

6***

(0.0

39)

0.20

2***

(0.0

33)

0.39

Num

obs

erva

tions

(ban

dwid

th)

384

(0.8

)

295

(0.6

)

470

(1.0

)

470

(1.0

)

869

(2.0

)

869

(2.0

)

1,39

91,

399

Fem

ales

in p

aid

empl

oym

ent

as m

ain

acti

vity

Pol

ynom

ial o

f or

der

1-0

.027

(0.0

36)

-0.0

13

(0.0

33)

-0.0

35

(0.0

94)

0.00

1

(0.1

02)

-0.0

39

(0.0

44)

-0.0

15

(0.0

50)

0.12

Num

obs

erva

tions

(ban

dwid

th)

1,39

9

(full

sam

ple)

1,39

9

(full

sam

ple)

470

(1.0

)

470

(1.0

)

869

(2.0

)

869

(2.0

)

opt

opt

Adu

lt m

ale

in

empl

oym

ent

Pol

ynom

ial o

f or

der

10.

044

(0.0

38)

0.03

6

(0.0

23)

0.00

3

(0.0

48)

0.01

1

(0.0

34)

-0.0

02

(0.0

29)

-0.0

06

(0.0

21)

0.01

9

(0.0

16)

0.01

4

(0.0

16)

0.93

Num

obs

erva

tions

(ban

dwid

th)

331

(0.6

)

331

(0.6

)

528

(1.0

)

470

(1.0

)

967

(2.0

)

967

(2.0

)

1,59

11,

591

Adu

lt m

ale

in o

wn

or

hous

ehol

d bu

sine

ss

(1)

Pol

ynom

ial o

f or

der

10.

038

(0.0

65)

-0.0

17

(0.0

53)

0.06

2

(0.0

92)

0.06

9

(0.0

72)

0.07

1

(0.0

86)

0.01

5

(0.0

79)

0.60

Num

obs

erva

tions

(ban

dwid

th)

1,59

1

(full

sam

ple)

1,59

1

(full

sam

ple)

528

(1.0

)

528

(1.0

)

967

(2.0

)

967

(2.0

)

opt

opt

Not

es:

(1)

Wor

k as

em

ploy

er, s

elf-

empl

oyed

or

unpa

id fa

mily

wor

ker

in th

e m

ain

or s

econ

dary

act

iviti

es. S

tand

ard

erro

rs (

in p

aren

thes

es)

are

clus

tere

d at

the

mun

icip

ality

leve

l. *

Stat

istic

al s

igni

fican

ce a

t 10%

; **

Stat

istic

al s

igni

fican

ce

at 5

%; *

** S

tatis

tical

sig

nific

ance

at 1

%

66


Tabl

e A

6.

RD

Est

imat

es o

f th

e ef

fect

of

inte

rven

tion

s on

em

ploy

men

t of

elig

ible

fem

ales

and

adu

lt m

ales

fro

m e

ligib

le h

ouse

hold

s

Trea

tmen

t-on

-the

-tre

ated

: 2

SLS

est

imat

es

BA

ND

WID

TH

OP

TIM

AL

SM

AL

L

(1.0

)M

IDD

LE

(2.0

)F

UL

L S

AM

PL

EM

EA

N

FU

LL

S

AM

PL

E

Fem

ale

in e

mpl

oym

ent

No

covari

ate

sW

ith

Covari

ate

sN

o

covari

ate

sW

ith

Covari

ate

sN

o

covari

ate

sW

ith

Covari

ate

sN

o

covari

ate

sW

ith

Covari

ate

s

0.54

Pol

ynom

ial o

f or

der

10.

399*

**

(0.1

25)

0.40

3***

(0

.123

)0.

381*

**

(0.0

74)

0.33

1***

(0

.091

)0.

395*

**

(0.1

01)

0.44

8***

(0

.107

)0.

252*

**

(0.0

55)

0.30

0***

(0

.058

)

Num

ob

serv

atio

ns

(ban

dwid

th)

295

(0.6

)

295

(0.6

)47

0 (1

.0)

470

(1.0

)86

9 (2

.0)

869

(2.0

)1,

399

1,39

9

Fem

ale

in o

wn

or

hous

ehol

d bu

sine

ss (1

)

Pol

ynom

ial o

f or

der

10.

531*

**

(0.1

36)

0.43

9**

(0.2

01)

0.44

6***

(0

.123

)0.

352*

**

(0.1

21)

0.43

5***

(0

.038

)0.

454*

**

(0.0

30)

0.28

7***

(0

.046

)0.

311*

**

(0.0

46)

0.39

Num

ob

serv

atio

ns

(ban

dwid

th)

384

(0.8

)

295

(0.6

)47

0 (1

.0)

470

(1.0

)86

9 (2

.0)

869

(2.0

)1,

399

1,39

9

Fem

ales

in p

aid

empl

oym

ent

as m

ain

acti

vity

Pol

ynom

ial o

f or

der

1-0

.041

(0.0

53)

-0.0

19 (

0.05

1)-0

.055

(0

.151

)0.

002

(0.1

57)

-0.0

63

(0.0

69)

-0.0

24

(0.0

81)

0.12

Num

ob

serv

atio

ns

(ban

dwid

th)

1,39

9

(full

sam

ple)

1,39

9 (fu

ll sa

mpl

e)47

0 (1

.0)

470

(1.0

)86

9 (2

.0)

869

(2.0

)1,

399

opt

1,39

9 op

t

Adu

lt m

ale

in

empl

oym

ent

Pol

ynom

ial o

f or

der

10.

073

(0.0

60)

0.06

3* (

0.03

7)0.

005

(0.0

76)

0.01

7 (0

.053

)-0

.004

(0

.047

)-0

.010

(0

.033

)0.

030

(0.0

23)

0.02

1 (0

.024

)

0.93

Num

ob

serv

atio

ns

(ban

dwid

th)

331

(

0.6)

331

(0.6

)52

8 (1

.0)

528

(1.0

)96

7 (2

.0)

967

(2.0

)1,

591

1,59

1

Adu

lt m

ale

in o

wn

or

hous

ehol

d bu

sine

ss (1

)

Pol

ynom

ial o

f or

der

10.

058

(0.0

98)

-0.0

26 (

0.08

2)0.

098

(0.1

45)

0.10

9 (0

.115

)0.

116

(0.1

37)

0.02

5

(0.1

27)

0.60

Num

ob

serv

atio

ns

(ban

dwid

th)

1,59

1

opt

1,59

1 op

t52

8 (1

.0)

528

(1.0

)96

7 (2

.0)

967

(2.0

)1,

591

opt

1,59

1 op

t

Not

es:

(1) W

ork

as e

mpl

oyer

, se

lf-e

mpl

oyed

or

unpa

id f

amily

wor

ker

in t

he m

ain

or s

econ

dary

act

ivit

ies.

S

tand

ard

erro

rs (

in p

aren

thes

es)

are

clus

tere

d at

the

mun

icip

alit

y le

vel.

*

Sta

tist

ical

sig

nific

ance

at

10

%; **

Sta

tist

ical

sig

nific

ance

at

5%

; **

* S

tati

stic

al s

igni

fican

ce a

t 1

%

67

Annexes

Tabl

e A

7.

RD

Est

imat

es o

f th

e ef

fect

of

inte

rven

tion

s on

the

att

itud

es o

f th

e ho

useh

old

head

tow

ards

chi

ldre

n's

empl

oym

ent

and

scho

olin

g

Inte

nt-t

o-tr

eat:

OLS

est

imat

es

BA

ND

WID

TH

OP

TIM

AL

SM

AL

L

(1.0

)M

IDD

LE

(2.0

)F

UL

L S

AM

PL

EM

EA

N

FU

LL

S

AM

PL

E

My

child

ren

wor

k,

beca

use

it w

ill

give

the

m g

reat

er

oppo

rtun

itie

s in

the

fu

ture

No

covari

ate

sW

ith

Covari

ate

sN

o

covari

ate

sW

ith

Covari

ate

sN

o

covari

ate

sW

ith

Covari

ate

sN

o

covari

ate

sW

ith

Covari

ate

s

0.62

Pol

ynom

ial o

f or

der

10.

008

(0.1

23)

0.00

7

(0.1

27)

0.03

3

(0.1

14)

0.06

9

(0.1

25)

0.00

7

(0.1

11)

0.00

3

(0.1

13)

0.02

2

(0.0

47)

0.01

0

(0.0

42)

Num

ob

serv

atio

ns

(ban

dwid

th)

838

(1.8

)

838

(1.8

)

493

(1.0

)

493

(1.0

)

917

(2.0

)

917

(2.0

)

1,49

51,

495

Wor

k ke

eps

child

ren

out

of t

roub

le/g

ives

th

em d

irec

tion

Pol

ynom

ial o

f or

der

1-0

.084

*

(0.0

48)

-0.0

84**

(0.0

39)

-0.0

45

(0.0

43)

-0.0

22

(0.0

46)

-0.0

27

(0.0

27)

-0.0

31

(0.0

26)

-0.0

27

(0.0

55)

-0.0

23

(0.0

55)

0.58

Num

ob

serv

atio

ns

(ban

dwid

th)

592

(1.2

)

592

(1.2

)

493

(1.0

)

493

(1.0

)

917

(2.0

)

917

(2.0

)

1,49

51,

495

Wor

king

chi

ldre

n ar

e m

ore

resp

onsi

ble

than

no

n-w

orki

ng c

hild

ren

Pol

ynom

ial o

f or

der

1-0

.002

(0.0

30)

-0.0

05

(0.0

36)

-0.0

11

(0.0

65)

0.00

6

(0.0

77)

-0.0

02

(0.0

39)

0.00

3

(0.0

38)

-0.0

04

(0.0

25)

0.01

3

(0.0

26)

0.85

Num

ob

serv

atio

ns

(ban

dwid

th)

1,37

3

(4.4

)

1,26

2

(3.4

)

493

(1.0

)

493

(1.0

)

917

(2.0

)

917

(2.0

)

1,49

51,

495

Wor

king

chi

ldre

n ar

e st

rong

er a

nd h

ealt

hier

th

an n

on-w

orki

ng

child

ren

Pol

ynom

ial o

f or

der

1-0

.339

**

(0.1

53)

-0.4

03**

(0.1

97)

-0.1

16*

(0.0

61)

-0.1

21**

(0.0

61)

-0.0

65

(0.0

62)

-0.0

76

(0.0

55)

-0.0

18

(0.0

58)

-0.0

32

(0.0

53)

0.62

Num

ob

serv

atio

ns

(ban

dwid

th)

96

(0.2

)

96

(0.2

)

493

(1.0

)

493

(1.0

)

917

(2.0

)

917

(2.0

)

1,49

51,

495

68


BA

ND

WID

TH

OP

TIM

AL

SM

AL

L

(1.0

)M

IDD

LE

(2.0

)F

UL

L S

AM

PL

EM

EA

N

FU

LL

S

AM

PL

E

Edu

cati

on is

the

onl

y w

ay t

o su

ccee

d in

life

Pol

ynom

ial o

f or

der

10.

071

(0.0

91)

0.06

6

(0.0

87)

0.06

0

(0.0

96)

0.05

4

(0.0

94)

0.07

9

(0.0

50)

0.06

7

(0.0

49)

0.01

8

(0.0

23)

0.01

6

(0.0

25)

0.76

Num

ob

serv

atio

ns

(ban

dwid

th)

592

(1.2

)

592

(1.2

)

493

(1.0

)

493

(1.0

)

917

(2.0

)

917

(2.0

)

1,49

51,

495

Edu

cati

on is

onl

y on

e of

the

fac

tors

tha

t in

fluen

ce t

he s

ucce

ss

of a

per

son

Pol

ynom

ial o

f or

der

1-0

.105

**

(0.0

41)

-0.1

10**

(0.0

45)

-0.0

11

(0.0

36)

-0.0

16

(0.0

37)

-0.0

16

(0.0

21)

-0.0

15

(0.0

16)

-0.0

07

(0.0

26)

-0.0

03

(0.0

31)

0.84

Num

ob

serv

atio

ns

(ban

dwid

th)

184

(0.4

)

184

(0.4

)

493

(1.0

)

493

(1.0

)

917

(2.0

)

917

(2.0

)

1,49

51,

495

Ens

ure

that

my

child

ren

go t

o co

llege

is

far

fro

m m

y ec

onom

ic p

ossi

bilit

ies

Pol

ynom

ial o

f or

der

10.

014

(0.0

43)

0.01

8

(0.0

47)

0.01

1

(0.0

42)

0.01

6

(0.0

45)

-0.0

26

(0.0

24)

-0.0

10

(0.0

18)

0.87

Num

ob

serv

atio

ns

(ban

dwid

th)

493

(1.0

)

493

(1.0

)

opt

opt

917

(2.0

)

917

(2.0

)

1,49

51,

495

Edu

cati

on is

not

im

port

ant

to s

ucce

ss

in li

fe

Pol

ynom

ial o

f or

der

1-0

.223

**

(0.0

96)

-0.2

81**

*

(0.0

88)

0.01

9

(0.0

70)

0.00

8

(0.0

85)

-0.0

54**

(0.0

24)

-0.0

57*

(0.0

32)

0.01

7

(0.0

26)

0.00

3

(0.0

28)

0.56

Num

ob

serv

atio

ns

(ban

dwid

th)

184

(0.4

)

184

(0.4

)

493

(1.0

)

493

(1.0

)

917

(2.0

)

917

(2.0

)

1,49

51,

495

Not

es:

Sta

ndar

d er

rors

(in

par

enth

eses

) ar

e cl

uste

red

at t

he m

unic

ipal

ity

leve

l. *

Sta

tist

ical

sig

nific

ance

at

10

%;

** S

tati

stic

al s

igni

fican

ce a

t 5

%;

***

Sta

tist

ical

sig

nific

ance

at

1%

69

Annexes

Tabl

e A

8.

RD

Est

imat

es o

f th

e ef

fect

of

inte

rven

tion

s on

the

att

itud

es o

f th

e ho

useh

old

head

tow

ards

chi

ldre

n's

empl

oym

ent

and

scho

olin

g

Trea

tmen

t-on

-the

-tre

ated

: 2

SLS

est

imat

es

BA

ND

WID

TH

OP

TIM

AL

SM

AL

L

(1.0

)M

IDD

LE

(2.0

)F

UL

L S

AM

PL

EM

EA

N

FU

LL

S

AM

PL

E

My

child

ren

wor

k, b

ecau

se

it w

ill g

ive

them

gre

ater

op

port

unit

ies

in

the

futu

re

No

covari

ate

sW

ith

Covari

ate

sN

o

covari

ate

sW

ith

Covari

ate

sN

o

covari

ate

sW

ith

Covari

ate

sN

o

covari

ate

sW

ith

Covari

ate

s

0.62

Pol

ynom

ial o

f or

der

10.

013

(0.2

08)

0.01

1

(0.2

11)

0.05

3

(0.1

85)

0.10

6

(0.2

02)

0.01

2

(0.1

85)

0.00

5

(0.1

87)

0.03

4

(0.0

75)

0.01

6

(0.0

67)

Num

obs

erva

tions

(ban

dwid

th)

838

(1.8

)

838

(1.8

)

493

(1.0

)

493

(1.0

)

917

(2.0

)

917

(2.0

)

1,49

51,

495

Wor

k ke

eps

child

ren

out

of

trou

ble/

give

s th

em d

irec

tion

Pol

ynom

ial o

f or

der

1-0

.139

*

(0.0

81)

-0.1

37**

(0.0

67)

-0.0

72

(0.0

65)

-0.0

33

(0.0

70)

-0.0

44

(0.0

45)

-0.0

51

(0.0

45)

-0.0

42

(0.0

87)

-0.0

35

(0.0

86)

0.58

Num

obs

erva

tions

(ban

dwid

th)

592

(1.2

)

592

(1.2

)

493

(1.0

)

493

(1.0

)

917

(2.0

)

917

(2.0

)

1,49

51,

495

Wor

king

chi

ldre

n ar

e m

ore

resp

onsi

ble

than

no

n-w

orki

ng

child

ren

Pol

ynom

ial o

f or

der

1-0

.003

(0.0

45)

-0.0

07

(0.0

56)

-0.0

18

(0.1

04)

0.00

9

(0.1

19)

-0.0

03

(0.0

64)

0.00

5

(0.0

62)

-0.0

06

(0.0

39)

0.02

0

(0.0

41)

0.85

Num

obs

erva

tions

(ban

dwid

th)

1,37

3

(4.4

)

1,26

2

(3.4

)

493

(1.0

)

493

(1.0

)

917

(2.0

)

917

(2.0

)

1,49

51,

495

Wor

king

chi

ldre

n ar

e st

rong

er a

nd

heal

thie

r th

an

non-

wor

king

ch

ildre

n

Pol

ynom

ial o

f or

der

1-0

.630

**

(0.2

49)

-0.7

29**

(0.2

98)

-0.1

85*

(0.1

03)

-0.1

88*

(0.0

99)

-0.1

08

(0.0

96)

-0.1

25

(0.0

83)

-0.0

28

(0.0

89)

-0.0

49

(0.0

81)

0.62

Num

obs

erva

tions

(ban

dwid

th)

96

(0.2

)

96

(0.2

)

493

(1.0

)

493

(1.0

)

917

(2.0

)

917

(2.0

)

1,49

51,

495

70


BA

ND

WID

TH

OP

TIM

AL

SM

AL

L

(1.0

)M

IDD

LE

(2.0

)F

UL

L S

AM

PL

EM

EA

N

FU

LL

S

AM

PL

E

Edu

cati

on is

th

e on

ly w

ay t

o su

ccee

d in

life

Pol

ynom

ial o

f or

der

10.

118

(0.1

58)

0.10

7

(0.1

50)

0.09

7

(0.1

60)

0.08

3

(0.1

52)

0.13

2

(0.0

89)

0.11

1

(0.0

86)

0.02

7

(0.0

37)

0.02

4

(0.0

39)

0.76

Num

obs

erva

tions

(ban

dwid

th)

592

(1.2

)

592

(1.2

)

493

(1.0

)

493

(1.0

)

917

(2.0

)

917

(2.0

)

1,49

51,

495

Edu

cati

on is

on

ly o

ne o

f th

e fa

ctor

s th

at

influ

ence

the

su

cces

s of

a

pers

on

Pol

ynom

ial o

f or

der

1-0

.172

**

(0.0

72)

-0.1

81**

(0.0

72)

-0.0

18

(0.0

55)

-0.0

25

(0.0

54)

-0.0

26

(0.0

35)

-0.0

24

(0.0

26)

-0.0

11

(0.0

40)

-0.0

04

(0.0

49)

0.84

Num

obs

erva

tions

(ban

dwid

th)

184

(0.4

)

184

(0.4

)

493

(1.0

)

493

(1.0

)

917

(2.0

)

917

(2.0

)

1,49

51,

495

Ens

ure

that

m

y ch

ildre

n go

to

colle

ge

is f

ar f

rom

m

y ec

onom

ic

poss

ibili

ties

Pol

ynom

ial o

f or

der

10.

022

(0.0

71)

0.02

8

(0.0

75)

0.01

9

(0.0

70)

0.02

7

(0.0

74)

-0.0

41

(0.0

36)

-0.0

16

(0.0

29)

0.87

Num

ob

serv

atio

ns

(ban

dwid

th)

493

(1.0

)

493

(1.0

)

opt

opt

917

(2.0

)

917

(2.0

)

1,49

51,

495

Edu

cati

on is

no

t im

port

ant

to

succ

ess

in li

fe

Pol

ynom

ial o

f or

der

1-0

.366

**

(0.1

85)

-0.4

65**

*

(0.1

74)

0.03

0

(0.1

09)

0.01

2

(0.1

31)

-0.0

91**

(0.0

43)

-0.0

94*

(0.0

56)

0.02

6

(0.0

39)

0.00

5

(0.0

43)

0.56

Num

ob

serv

atio

ns

(ban

dwid

th)

184

(0.4

)

184

(0.4

)

493

(1.0

)

493

(1.0

)

917

(2.0

)

917

(2.0

)

1,49

51,

495

Not

es:

Sta

ndar

d er

rors

(in

par

enth

eses

) ar

e cl

uste

red

at t

he m

unic

ipal

ity

leve

l. *

Sta

tist

ical

sig

nific

ance

at

10

%;

** S

tati

stic

al s

igni

fican

ce a

t 5

%;

***

Sta

tist

ical

sig

nific

ance

at

1%

71

Annexes

�Annex 4. Impact evaluation design

Acknowledgements

The design of the impact evaluation has been carried out in close collaboration with the project team and the government counterparts. The government formally agreed to carry out an impact evaluation of some interventions of the project on the basis of a vulnerability index. The project team has continuously supported the development of the IE design with its suggestions and assessments of constraints. It has also carried out the preparatory fieldwork with great care and insured the support of the different stakeholders.

Funding for this project was provided by the United States Department of Labor. This doc-ument does not necessarily reflect the views or policies of the United States Department of Labor, nor does mention of trade names, commercial products, or organizations imply en-dorsement by the United States Government.

1. Outline

The most recent information on child labor in El Salvador shows that around 1 out of 10 children in the country (177,070 boys, girls and adolescents) worked in 2010. The ILO project called “Eliminating Child Labour in El Salvador through Economic Empowerment and Social Inclusion” attempts to eradicate child labor in El Salvador. To achieve this objective the project will implement a large range of interventions at three levels: macro (national policies and institutional framework), meso (target municipalities and schools) and micro (child laborers’ households).

One set of household level interventions, which is to be implemented in 5 municipalities in El Salvador, will be the object of an impact evaluation. This set consists of education interven-tions for child laborers (and their siblings), vocational training for their mothers, and an ex-tensive child labor monitoring system administered by local municipalities. Initial discussions considered the possibility to evaluate only the impact of the vocational training of mothers of child laborers. However, as explained in this document, the setup of the project does not allow for the identification of the separate impact of this intervention or other individual com-ponents of the project.

The project will allocate this set of interventions to beneficiaries within the 5 municipalities on the basis of a vulnerability index. Households with high vulnerability scores will be offered the interventions, households with lower scores will not. The impact evaluation will exploit this allocation procedure in a regression discontinuity framework to identify the causal effect of the set of interventions on education, child labor, participation in labor of adult household members, and multiple other wellbeing indicators. By identifying the causal effect of this set of interventions, the impact evaluation can provide a unique insight into the functioning and effectiveness of the ILO’s approach of integrated child labor programs.1

1 For a discussion of several issues relating to the impact evaluation of ILO/IPEC child labour programs see UCW (2009a and 2009b) and Manacorda (2009)

72


This document provides a detailed description of the design of the impact evaluation. In section 2 it starts by sketching the background of the project and providing detailed information on its objectives and the interventions that will be implemented to reach these objectives. Section 3 discusses how the interventions will be allocated to beneficiaries. Section 4 explains how these allocation mechanisms can be exploited to measure the impact of the interventions. Section 5 discusses the data sources on which the impact evaluation will be based. Section 6 provides information on some technical issues, some issues to be determined, and a few possible exten-sions of the basic impact evaluation. Section 7 gives an overview of the limitations of the (design of) the impact evaluation. Finally, section 8 provides a timeline of the project.

2. The Project

2.1 Background2

2.1.1 El Salvador

The territory of the Republic of El Salvador covers an area of 21,040.79 km. The country is divided in 14 departments which in turn comprehend 262 municipalities. In 2009, El Salva-dor’s population was estimated at 6,150,9533. Sixty three percent of the population is urban and 53 % is female. Population growth rate is 1.69% (2007 est.) and fertility rate is 3.08 chil-dren per woman. Population growth is compensated by El Salvador’s high rate of emigration, a continuous trend during the last three decades.

In 2009, El Salvador’s Human Development Index of 0.747, ranked it in 106th position out of 182 countries reviewed. The Adult literacy rate (ages 15 and above) is 82%. Life expectancy at birth (2008 est.) is 72.06 years; but while it is 75.84 years for women it is just 68.45 years for men. This difference is due to the high rate of violent death for men, largely due to the social violence that has been affecting the country for the past two decades, after the end of the armed conflict.

According to the World Bank4 (WB), in 2009, the gross national income per capita was U$ 3,370. However, this gross figure does not reflect the wide inequalities in income distribution (and opportunities) among different departments and municipalities, which have shown no improvement during the recent past.

The Salvadoran economy is characterized by a large informal sector. A study5 supported by the Economic Policy Institute found that the informal sector represents 69% of Salvadoran economy and a large portion of its labor market. A significant number of informal workers, 62%, earn less than the minimum wage and 21% have no formal education. The Salvador-an economy is heavily dependent on the United States, both for exports and as a source of remittances from Salvadoran workers living there. As a result, following the global financial

2 This section is based on the Project Document, ILO (2010).

3 Ministerio de Economía – General Directorate of Statistics and Census (DIGESTYC), Encuesta de Hogares de Propósitos Múltiples (EHPP) 2009; Delgado, July 2010.

4 http://siteresources.worldbank.org/DATASTATISTICS/Resources/GNIPC.pdf

5 Tony Avirgan, L. Josh Bivens & Sarah Gammage, eds. (2005), Good Jobs, Bad Jobs, No Jobs: Labor Markets and Informal Work in Egypt, El Salvador, India, Russia, and South Africa; GPN/EPI, Washington D.C.

73

Annexes

crisis, in 2009 the Salvadoran economy contracted by 3.5% (while in 2008 it had grown 2.4%), as both exports and remittances dropped sharply.

Agriculture activities are the main source of employment and account for a large part of na-tional exports. There is a clear difference between the culture of commercial products and those for the household’s own consumption. The first ones involve large plots where sugar cane, corn or coffee are produced with modern techniques and supplies, while the second involve small household plots with limited productivity, where beans, corn and rice are pro-duced. During harvest time, entire households with small children move to the coffee zones and from one finca to another in order to make a living.

El Salvador faces serious public security challenges. Violence associated to gang-crimes runs high. In 2006 the country had the highest violent crime rate in the Latin American Region: 65 homicides per 100,000 inhabitants (the average for the region is 25 per 100,000 and for the world: 9 per 100,000). In 2008, El Salvador was the country with the second highest crime rate in the region.

El Salvador is home to multiple social protection programs. Two important programs are PATI and Red Solidaria. PATI is a government-run public works program, which provides train-ing and temporary employment opportunities in community projects. Red Solidaria is also a government-run program. It provides cash transfers conditional on adherence to several schooling and health related requirements to the poorest population of El Salvador. In addi-tion, it attempts to strengthen basic services such as education and health and it supports small-scale projects that aim to increase the productivity of poor agricultural producers.

2.1.2 Child Labor in El Salvador

The most recent information available on the number of child laborers in El Salvador6 shows that around 1 out of 10 children (e.g., 177,070 boys, girls and adolescents) worked in 2010. Those child labor figures represent a slight reduction (6.3%) with regards to the 2009 child labor estimates. Approximately 73% of child laborers are boys. Most child laborers (62%) live in rural areas and more than half of child laborers work in agricultural activities. A majority of child laborers (61%) work with their family and receive no pay. Twenty-two percent of work-ing children work in temporary jobs.

Common hazardous child labor in El Salvador includes cutting of sugar cane, fishing (par-ticularly mollusk extraction), harvesting coffee, other forms of agricultural activity, carrying weights in trades and street vending. Children in these activities work for long hours, under the rain or sun, with no gear protection. Additional forms of hazardous child labor in which a lesser number of children are involved include: recycling of garbage, fireworks production and domestic work.7 In addition to hazardous child labor, the other more common worst forms in El Salvador include commercial sexual exploitation and the use of children for illicit activities (e.g., drug trafficking).

6 Ministerio de Economía – DIGESTYC, “Midiendo el Trabajo Infantil en la Encuesta de Hogares de Propósitos Múltiples, EHPM 2010”. Presentation, June 2011.

7 El Salvador has not yet approved a list of activities to be considered as hazardous child labor, although local stakeholders are in the process of discussing this issue. The specific activities mentioned in this paragraph have been considered as hazardous child labor by other countries.

74


2.2 Project Objectives

The Eliminating Child Labour in El Salvador through Economic Empowerment and Social Inclusion project aims to achieve the following 6 objectives:

i. By the end of the project, child labor elimination strategies will have been adopted by national and local institutions that implement poverty reduction, decent work, and social protection programs for the rural and urban poor;

ii. By the end of the project, child labor law enforcement and protection mechanisms, par-ticularly for the worst forms of child labor, will be developed and fully operational;

iii. By the end of the project, national capacity to conduct child labor research, monitoring and impact evaluation will be enhanced and pilot interventions documented;

iv. By the end of the project, municipal capacity to prevent and eliminate child labor, in particular the worst forms, will have been strengthened;

v. By the end of the project, viable improved livelihood alternatives will have been imple-mented to reduce family reliance on child labor in target municipalities;

vi. By the end of the project, inclusive education models will have been implemented in selected schools in target municipalities to prevent and reduce child labor and improve education indicators.

2.3 Interventions Selected for Impact Evaluation

The ILO plans to implement a large number of interventions to achieve these 6 objectives. Some of these interventions will be implemented at the national level, some in selected municipalities, some in selected schools, and some will be targeted at selected households. One set of household level interventions will be the object of an impact evaluation. The set consists of a combination of vocational training for mothers, a group of “flexible” education interventions for their children, and a direct beneficiary monitoring system.

This set of interventions was selected primarily because it will be implemented on a suffi-ciently large scale to conduct statistical analysis. The other interventions implemented by the project are of interest too and would merit rigorous evaluation. However, such rigorous evaluation is not possible because these interventions are either administered to a smaller number of individuals, clustered in a small number of geographical areas, or administered at the national level. This sub-section describes the evaluated interventions in more detail.

2.3.1 Vocational Training of Mothers

Vocational training will be provided to 2000 selected mothers of child laborers within 5 se-lected municipalities. This vocational training intervention should help the mothers of child laborers either to enter the labor market or to start a small enterprise. The resulting additional income generated by these activities should reduce their household’s dependence on child labor.

75

Annexes

The project team is currently developing two main types of vocational training courses. One type will prepare participants to start a small enterprise and the other to enter the labor mar-ket. The women who participate in the intervention can choose one of these training courses according to their needs and preferences.

The vocational training program will be preceded by an orientation course offered by El Sal-vador’s Ministry of Labor. This orientation course will help the women who enter the program decide which type of vocational training course to enter. Upon completion of the vocational training program, the mothers will be provided with additional services either in the form of job placement or in the form of tools, supplies, and technical support to start a small business.

2.3.2 Monitoring

The ILO will monitor the school attendance and participation in child labor of the children of the mothers selected to participate in the vocational training program. In practice, this means that the ILO keeps a record of the interventions administered to individual beneficiaries. As part of keeping this record, the ILO, the implementing agencies, or monitors trained in the local community will regularly be in touch with the beneficiaries to check whether they participate in the interventions and whether this participation is accompanied by behavioral changes such as improved school attendance. Technically, the ILO does not perceive this so-called direct beneficiary monitoring and reporting system (DBMR) as an intervention, but as a part of its evaluation system. However, from the perspective of the impact evaluation it should be interpreted as an intervention, as it is likely to affect household decisions on edu-cation and children’s work decisions.8

The project team plans to train officials of the administration in each of the 5 municipalities to execute the monitoring work. Monitoring will consist of an interview with the head of the household on a 3 month basis. In addition, monitoring may include spot checks in schools and meetings with other people in the community to verify that the information given by the household is correct. The project team will encourage the local administration to expand this system of regular checks to all individuals in the community and to continue this process after the ILO project is over.

The goal is to provide referral services when children are found to be working or not attending school. The project team will work with the local administration to identify local organizations and opportunities to provide such services. The ILO will not fund these activities as such, but will provide training and technical support to make sure they get started.

2.3.3 Flexible Education Interventions

In addition, the ILO will provide so-called “flexible” education interventions (modalidades flexibles) to 12 to 17 year olds who failed to complete their basic education and who live in the households of the mothers who receive vocational training. These interventions will differ according to the needs of the 12 to 17 year olds who failed to complete their education. Flex-ible education can include the following interventions:

8 Moreover, as discussed in some more detail below, the DBMR may result in biased responses of program beneficiaries, in particular to questions that determine their eligibility status.

76


� Accelerated education in which the final grades of secondary school are finished in a shortened period of time;

� Weekend education;

� Online education;

� Evening education;

� Participation in a government exam (called Examen de Suficiencia), completion of which allows the individual to re-enter the regular education system at the appropriate level.

The project team is currently in the process of designing these interventions.

2.4 Interventions Not Selected for Impact Evaluation

The three direct-beneficiary interventions that will be the object of the impact evaluation are part of a bigger group of interventions to be implemented in these 5 communities. This sub-section describes these other interventions. A full description of all (other) interventions that will be implemented (at the national, community, school, and household level) as part of the project is available from the project team.

2.4.1 Inclusive Education

Faced with multiple shortcomings, El Salvador is currently redesigning its education system. The aim is to make sure that El Salvador’s schools start to provide “inclusive education” (escuela inclusiva a tiempo pleno). Inclusive education entails longer school days (typically school days now last for half a day) and an overhaul of the curriculum.9 The ILO will help the schools attended by the children of mothers who receive the vocational training intervention (as well as the schools of children in households that receive either the vocational training for youths or the agricultural productivity interventions) to implement inclusive education. The project team is currently in the process of designing these interventions.

2.4.2 Strengthening Capacity of Local Administration

Finally, the ILO will attempt to strengthen the capacity of the local administration to address child labor issues. As part of this process, the ILO will attempt to raise local officials’ aware-ness of child labor issues in the community and will help the municipality to establish a local committee that deals with these issues. It will provide technical support to help the commit-tee implement national policies with respect to child labor. The project team is currently in the process of designing these interventions.

9 See also the document called “De la escuela tradicional a una escuela inclusiva de tiempo pleno” of El Salvador’s Ministry of Education.

77

Annexes

3. Selection Procedures

This section describes how the project will assign the described interventions to municipali-ties, and schools and households within the selected municipalities. As will be explained in more detail below, this assignment procedure is crucial for the impact evaluation as it allows for the identification of a credible control group.

3.1 Selection of Target Municipalities

The project team started by selecting a group of 10 municipalities that participate in the in-terventions. In order to select these municipalities the project team first produced a shortlist. To appear on the shortlist, municipalities had to:

� Have a high child labor rate. In order to determine child labor rates, the project team used the 2009 school census which contains information on participation in work for all children attending school (either basic education, “bachillerato general” or “bachillerato técnico”).10 Municipalities in which the child work rates among children in school ex-ceeded 6.2% were considered to be municipalities with a high child labor rate.

� Be reasonably safe to ensure that it is possible to implement the necessary fieldwork. Municipalities were considered to be safe if they had a homicide and extortion rate of 128 per 100,000 inhabitants or lower according to the national police.

� Have more than 19,000 inhabitants to make certain that an overall target of 5000 house-holds with child laborers could be reached by all household level interventions.

The shortlist of 21 municipalities that resulted from these criteria can be found in appendix 1.

Subsequently, the project team made a map that displayed all of the shortlisted municipali-ties. Given that it was the initial goal of the project to end up with 5 to 12 municipalities, the project team then selected 12 municipalities that were reasonably close to each other. This procedure ensured that travel costs that would be incurred to setup and run the project could be minimized.

This group of 12 was supposed to be the final group of selected municipalities. However, a few changes to this selection took place at a later stage. First, the project team narrowed down the number of selected communities to 10 to reduce the scale and costs of the project. This was done by removing the 2 selected municipalities with the lowest child work rates according to the 2009 school census: San Antonio del Monte and Santiago de Maria. Then, the team made a few further changes as the 2010 school census became available. Specif-ically, the team removed 2 municipalities, Guaymango and San Pedro Masahuat, because they did not have high child work rates according to the 2010 school census. It replaced these 2 municipalities with 2 other municipalities that did have sufficiently high child work rates according to the 2010 school census and were geographically close to the other 8 se-lected communities: Santiago Nonualco and Santiago de Maria (which had previously been removed). A map of the selected municipalities can be found in appendix 2.

10 Basic education in El Salvador consists of 9 years of education divided into 3 cycles of 3 years each. Pupils are supposed to enter basic education when they are 6 years old and finish when they are 15 years old. Basic education is followed by a 2 year “bachillerato” degree.

78


Next the 10 final selected municipalities were divided into 3 groups for organizational rea-sons. The two selected municipalities with the highest child work rates according to the 2010 school census were grouped together. These two municipalities will participate in a first wave of interventions not described in this document. The 5 municipalities with the next highest child work rates were grouped together and so were the 3 municipalities with the lowest child work rates according to the 2010 school census. The latter two groups will respectively participate in a second wave of interventions (the interventions described above) and a third wave of interventions. In total the three waves of interventions should reach 5000 beneficiary households.

The impact evaluation will focus on the interventions that will be implemented in the second wave of the project (i.e. in the group of 5 municipalities with intermediate child work rates). Table 1 provides an overview of the child work rate, total population and the crime rate (hom-icide and extortion per 100,000 inhabitants) in the wave 2 municipalities.

Table 1. Basic Information on Selected Municipalities

MunicipalityUrban or rural

Child work rate Population

Crime rate

Tacuba Rural 15,4 30.718 68

Izalco Urban 8,4 74.085 115

San Luis la Herradura Rural 7,9 21.388 122

Tecoluca Rural 7,9 25.344 55

Santiago Nonualco Rural 6,3 41.287 63Child labor rate is the percentage of working children identified in the 2010 school census. Population estimates based on 2007 population census. Crime rates (homicide and extortion per 100,000) based on 2010 national police statistics.

3.2 Selection of Target Cantons

Within each of the selected wave 2 municipalities a group of cantons was selected into the project. Selection of cantons took place on the basis of the information available from the 2010 school census and the 2007 population census. Cantons were selected if they had high numbers of child laborers according to the school census and had high child labor ratios.11 However, explicit thresholds for the selection cantons were not established. In total, the pro-ject team selected 28 cantons in the 5 municipalities.

3.3 Selection of Direct Beneficiaries

Within the identified wave 2 cantons selection of beneficiary households will be conducted on the basis of a vulnerability index. This index will be established using information pro-vided by households during the baseline interview. Table 2 displays the vulnerability index proposed by the project team.

11 As suggested by the number of child laborers in the school census divided by the total canton population according to the 2007 census.

79

Annexes

Households with high vulnerability scores (i.e. vulnerability scores above a predetermined threshold) will receive the set of interventions that will be evaluated: vocational training for mothers, flexible education interventions for 12 to 17 year olds who failed to complete their basic education, and monitoring of child labor and school participation. Households with a vulnerability score below the threshold will not receive this set of interventions.

3.4 Selection of Target Schools

All the schools attended by households with working children will receive the inclusive edu-cation interventions described above.

Table 2. Proposed Vulnerability Index

Vulnerability score

The household head is not married 20

The household head is married 12

Wall material: Lamina, paja o palma o materiales de desecho 7

Wall material: Bahareque, adobe o madera 4

Wall material: Concreto mixto 1

Floor material: Tierra 6

Floor material: Ladrillo de cemento(o de barro), o de cemento

4

Roof material: Paja o Palma o materiales de desecho 7

Roof material: Loza de concreto, teja de barro o cemento, lámina de asbesto ( o metálica)

4

Ratio of unemployed 18 to 25 year olds to total number of 18 to 25 year olds in the household = 1

20

0.5 <= Ratio of unemployed 18 to 25 year olds to total number of 18 to 25 year olds in the household < 1

12

Ratio of unemployed 18 to 25 year olds to total number of 18 to 25 year olds in the household < 0.5

4

Ration of 5 to 17 year old children who work to total number of 5 to 17 year old children in the household = 1

20

0.5 < =Ratio of 5 to 17 year old children who work to total number of 5 to 17 year old children in the household < 1

12

Ratio of 5 to 17 year old children who work to total number of 5 to 17 year old children in the household < 0.5

4

Household located in extremely precarious urban area 20

Household located in precarious urban area 14

Household lives in other urban area 4

80


4. Estimation Strategy: Regression Discontinuity

The impact evaluation will examine the impact of the set of interventions described above on households’ wellbeing. That is, it will seek to establish which of the observed changes in wellbeing indicators among project beneficiaries can be directly attributed to the project. To identify the causal effect of this set of interventions, the impact evaluation must estimate a counterfactual outcome: the wellbeing indicators that would have been observed among project beneficiaries in absence of the project. This counterfactual outcome is unobserved by definition. However, a valid counterfactual can be established by identifying a comparison group that is similar to the group of beneficiaries at the start of the project in all respects, except that they do not participate in the intervention. 12

To identify a credible comparison group, the impact evaluation of the ILO project in El Salva-dor will exploit the fact that the set of interventions is assigned on the basis of a vulnerability index using the so-called regression discontinuity methodology. The intuition behind the regression discontinuity methodology is that households (and members of households) with a vulnerability score just below the threshold are similar to households with a vulnerability score just above the threshold. These households therefore serve as a valid control group to measure the impact of the set of interventions.13

4.1 Graphical Explanation

Figures 1 and 2 below provide a graphical explanation of the regression discontinuity meth-odology (using fictional data). Figure 1 shows the hypothesized relationship between house-holds’ vulnerability scores and child labor before the start of the interventions. Dots represent local child labor averages at the vulnerability scores and the lines are linearly fitted regres-sions. There is a clear positive relationship between vulnerability in child labor. However, there is no jump in child labor at the threshold (which is set at 50 in this graph).

Figure 2 shows the hypothesized relationship between vulnerability and child labor after the interventions have been implemented. The relationship is still positive. However, there is now a clear discontinuity in the probability that children are in child labor at the threshold score. This discontinuity represents the local causal effect of the interventions on child labor.

12 For detailed yet accessible introductions to impact evaluation and methods that can be used to establish a control group the reader is referred to Duflo et al. (2008), Gertler et al. (2011) and Khandker et al. (2010).

13 The project initially aimed to implement a randomized experiment to measure the impact of the interventions. However, the Government of El Salvador did not support such an approach and as a consequence it was decided to adopt this quasi-exper-imental approach.

81

Annexes

Figure 1. Hypothesized relationship between vulnerability scores and child labor before the start of the intervention

Figure 2. Hypothesized relationship between vulnerability scores and child labor after the start of the intervention

82


4.2 Estimating the Intent-to-Treat Effect

To gauge the size of this discontinuity, we will estimate the following sharp regression discon-tinuity equation:

∆Yhm = α+βD + ∑k≥1 γk (X - c)k + ∑k≥1 δk D(X - c)k + εhm , (1)

where ∆Yhm is the change in the outcome of interest for household h in municipality m, α is the intercept, D is a dummy taking the value 1 if a household was selected to receive the set of interventions (i.e. had a vulnerability score above the threshold), the term ∑k≥1 γk (X - c)k is a polynomial of order k that approximates the relationship between the outcome of inter-est and the distance of a the household’s vulnerability score X to the threshold score c. The term includes the dummy for selection into the interventions and thus allows for a different functional form of the polynomial above and below the threshold score. The error term ∑(k≥1) δk D(Xv - c)k captures all other determinants of the outcome of interest. The estimated coef-ficient D gives the average local effect of a households being offered the set of interventions (the intent-to-treat effect).

The regression discontinuity approach will yield consistent parametric estimates of the intent-to-treat effect if the specified polynomial correctly approximates the relationship between the distance of the village’s forcing variable to the cutoff scores (X-c) and outcome Yhm. Misspec-ification becomes more likely when observations further from the cutoff score are used. The estimation should therefore check for the robustness of the estimated results within multiple bandwidths around the cutoff scores.

Fixed effects will be incorporated to deal with potential differences between municipali-ties / cantons. Municipality, household, and child level controls can be included to increase the precision of the estimates. The optimal order of the polynomial can be determined based on the Akaike information criterion (AIC) following Lee and Lemieux (2010). The preferred bandwidth can be determined using a cross-validation criterion proposed by Imbens and Lemieux (2008).

4.3 Validity Checks

The regression discontinuity design provides a valid method to identify a control group if households cannot precisely manipulate their vulnerability index to self-select into the inter-ventions and if households just above and just below the vulnerability index are truly similar before the start of the interventions. In the case of the project in El Salvador, there is little reason to believe that the regression discontinuity design will not generate a valid control group. First, when people participate in the baseline interview, they do not know that they might be eligible to participate in an intervention, nor that eligibility will be determined on the basis of their answers during this interview. Second, the threshold scores will not yet have been determined at the time of the baseline interview. There is thus little room for precise manipulation of the vulnerability index. Finally, we know that in Red Solidaria (an integrated program that provides cash transfers, micro-credit, education, health, and nutritional ser-vices to poor households in El Salvador) that the regression discontinuity assumptions were satisfied which gives us more confidence in the methodology.

83

Annexes

There are several procedures that can be used to check whether the regression discontinuity design indeed establishes a credible control group. First, McCrary (2008) proposed a test that looks at the density of the forcing variable (in this case the vulnerability index) around the threshold. If people bunch just below or above the threshold, this can be a sign that individ-uals are able to manipulate the forcing variable. Second, it is possible to check for continuity in observed baseline covariates at the threshold to ensure that there are no (observed) dif-ferences between the control and the intervention group. Both of these checks will be done when the baseline data have been collected and the threshold has been determined.

5. Data Sources and Outcome Indicators

This section discusses the data sources that will be used in the impact evaluation. The main data sources will be a census and a baseline survey (to be administered before the beginning of the project) and a follow-up survey (to be administered towards the end of or shortly after the end of the project). To the extent possible we will substantiate the self reported data from these surveys with other data sources.

5.1 Census

A census will be administered to all households in the selected cantons. The census will gather information on some basic characteristics of household members: age, gender, school participation, highest level of schooling attained, and work in the week and year prior to the interview. The census instrument is available from the team in El Salvador or UCW on re-quest.

5.2 Baseline Survey

Subsequently, a more in-depth baseline survey will be administered to all households which, according to the census, have 5 to 17 year old working children. The baseline survey con-tains the following 7 sections:

1. A household roster that collects basic information on household members (age, gender, relationship to household head, and marital status) and on education (current school participation, highest level of education attended, ability to read and write);

2. A section on the characteristics of the house inhabited by the household and access to durable goods;

3. A section on economic activities of the adult household members;

4. A section on participation in credit schemes by the household members;

5. A section on participation in other social protection programs;

6. A section on chores and economic activities by minors living in the household;

84


7. A section on the respondent’s attitudes towards child labor.

The vulnerability scores for households (that determine whether they are eligible to partic-ipate in the set of interventions) will be calculated on the basis of this survey. The baseline survey instrument is available from the team in El Salvador or UCW on request.

5.3 Follow-up Survey

Both the census and baseline survey will be administered before the start of the interven-tions. A follow-up survey will be administered when the main interventions are completed (the monitoring work will still be going on). The follow-up survey will be an in-depth interview administered to the households that also participated in the baseline survey. The follow-up survey has not yet been designed. It will be similar to the baseline survey, but may be some-what more extensive.

The outcome variables measured in the baseline and follow-up survey will be used in the regression discontinuity estimations described earlier. The subsections below describe the main outcome indicators that will be investigated by the impact evaluation: children’s work, school participation, and adult work.

5.3.1 Children’s work

To measure the impact of the interventions on children’s work detailed information on the activities in which children are involved during the baseline and follow-up survey will be col-lected. These activities include both remunerated and non-remunerated economic activities (work in agriculture, work in non-agricultural enterprises etc.). They also include non-eco-nomic activities, such as household chores. Information will be collected both for the week prior to the interview and the year prior to the interview.

To better understand children’s activities and to be able to classify these activities as (worst forms) child labor the surveys will also establish the amount of time children spend in these activities, what remuneration the children received and who decides how the income earned by the child is spent. Moreover, information on the circumstances under which children work (such as work in dusty or noisy environments) and injuries suffered during work will be collected to identify hazardous work.

5.3.1 School Participation

In the baseline and follow-up survey, special attention will be paid to school participation. We will examine a series of other school related outcomes (such as the highest grade of school attended) and indicators of learning in school (such as ability to read and write and passing the grade attended by the pupil). We will examine both school participation in general and participation in the flexible education interventions implemented as part of the program.

85

Annexes

5.3.2 Adult Work

Given that the vocational training for women aims to let women enter the labor force or start a small business, we would expect to find substantial movement here. These movements may not only involve women who participate in training (as they become more involved in income generating activities), but also other adult members of the household. To investigate these issues, information on work will be collected for all adult members of the household. Similar to the section on children’s work, this section will investigate multiple aspects of adult labor, including sector, working hours, wages, and training. The survey also collects information on social security provided by the employer.

5.4 DBMR

The ILO will collaborate with local authorities to monitor the school attendance and participa-tion in child labor of the children of the mothers selected to participate in the vocational train-ing program on a three month basis. This monitoring work will result in a direct beneficiary monitoring database (typically called DBMR). The project team will set the DBMR up with unique identifying variables so that it can be linked to the census, baseline, and follow-up survey.

5.5 Project Monitoring

In addition, the impact evaluation will rely on information from the project’s monitoring sys-tem. For each household in the baseline survey, the project monitoring system will provide the following information:

i. Interventions offered to the household

ii. Whether the intervention was accepted by the household

iii. Whether the household participated regularly in the interventions

iv. Content of the interventions

v. Implementing agency

5.6 Triangulation

The monitoring work (DBMR) could potentially lead to substantial bias in answers to the follow-up survey. The households in the intervention group will have been subject to regu-lar checks on child labor and school participation. The members of these households may therefore be inclined to deny that children participate in work and confirm that children are in school, even when this is not correct. To come up with reliable measures of school participa-tion and children’s work the project team will explore the possibility to use additional sources of information to substantiate the survey data.

86


6. Technical Issues, Issues to Be Determined, and Possible Extensions

6.1 Power Calculations and Sample Size

To identify the impact of the household level interventions (vocational training for mothers, monitoring, and flexible education) the regression discontinuity design requires a sufficiently large group of households to be located above and below (and preferably close to) the es-tablished threshold in the vulnerability index. We conducted power calculations to determine what impact size we can expect to identify at reasonable levels of statistical significance and at different control group sizes. The calculations use the Jujutla pilot data as a starting point. These pilot data show that, in households with working children, 74% of the children in the data can be classified as working. Approximately 27% of all 18 to 45 year old women in these households have paid employment and the average monthly earnings of the 18-45 year old women in these households are US$ 28 (with a standard deviation of 56). 14 (The outcomes of our power calculations can be found in appendix 3.)

With a control group of 720 to 990 households, we find that the we will be able to detect a decrease in the proportion of working children of 10 percentage points and an increase in the proportion of working women by 10 percentage points with reasonable levels of statistical power (at significance levels of 5 and 10%).15 If the decrease in the proportion of working children or the increase in the proportion of working women is only 8 percentage points we can achieve a reasonable level of statistical power only with high sample sizes (990 or higher) and low levels of significance (10%). Similarly we are able to detect an increase in average monthly earnings of women of 12 US$ with reasonable levels of statistical power. But lower earnings increases are harder to detect, particularly for lower sample sizes and lower levels of statistical significance.

The minimum impact sizes that we can detect with the current regression discontinuity setup are comparatively large. Attanasio et al. (2009) estimate that a vocational training program previously implemented in Colombia increased the incidence of paid employment among women (54% in the control group) by 5 to 5.7 percentage points (intent to treat effect). It increased monthly salary earnings of women (166 Peso in the control group) by 30 to 34 Peso or 18 to 20% (intent to treat effect).16 It is likely that we will not be able to statistically establish a beneficial impact of the program in El Salvador, if the actual impact is comparable with that of the program evaluated by Attanasio et al. (2009).

Table 3 investigates the number of households with working children we currently expect to find in the selected cantons during the baseline survey. For each of the selected municipal-ities (and a pilot municipality called Jujutla), the first two columns show the total number of cantons and the number of selected cantons. The next two columns then show the number

14 The power calculations were conducted using the “sampsi” routine in Stata. Sampsi can be used for power calculations in randomized experiments. Cappelleri et al (1994) and Goldberger (1972) observed that the sample size for regression disconti-nuity designs should be approximately three times that of a randomized experiment to achieve equivalent statistical power and we follow this rule of thumb. The calculation assumes that the intervention group consists of 2000 households.

15 Power above 0.8 is typically deemed acceptable.

16 We have no information on the impact of this program on child labor, nor are we aware of any other studies that register the impact of vocational training of adult women on the labor outcomes of their children.

87

Annexes

of children in school and in work both in the entire municipality and in the selected cantons according to the 2010 school census.

The number of households with working children is not necessarily equal to the number of children in school and in work according to the school census. A share of the working children may, for instance, not appear in the school census as they do not attend school. And multiple children identified as working and in school by the school census may live in the same household. In fact, in Jujutla (where the baseline has already been completed), the number of households with working children identified in the baseline survey turned out to be about 36% higher than the number of children in school and in work identified in the 2010 school census. Assuming this ratio is similar in other municipalities, we can estimate the number of households with working children in the selected and the remaining cantons of the wave 2 municipalities. These estimates are displayed in the last two columns of table 3.

Table 3 shows that we currently expect to find a total of about 3418 households with working children in the selected cantons. Approximately 500 households need to be discarded as the project will select a small group of cantons for (agricultural and youth) interventions not subject to impact evaluation.17 Given that it is the aim of the project to let 2000 households in the remaining cantons enter the vocational training for mothers we should expect to find a maximum of about 918 households that do not participate in any of the household level interventions and that could possibly enter the control group.

Table 3. Basic Information on Selected Cantons

MunicipalityTotal cantons

Cantons selected

Children in school and in work in the municipality

Children in school and in work in selected cantons

Expected number of households with working children in selected cantons

Expected number of households with working children in cantons not selected

Jujutla (pilot) 10 5 875 474 643 544Tacuba 14 5 1196 700 950 673Izalco 46 6 1644 701 951 1279San Luis la Herradura 10 4 571 356 483 292Tecoluca 26 4 640 345 468 400Santiago Nonualco 22 9 838 418 567 570Total 118 28 4889 2520 3418 3214Children in school and in work based on 2010 school census. Expected number of households with working children estimated using Jujutla as example.

In reality however, the final number of households in the control group is likely to be lower for multiple reasons. First, a percentage of the households that have been offered the inter-ventions will not accept the offer. A pilot that will be rolled out in Jujutla municipality will soon provide an indication of the take-up rate we should expect to encounter. If the take-up rate is relatively high, the household level interventions can be offered to exactly 2500 households. The wave 1 and wave 3 interventions can then be scaled up slightly to compensate any for

17 In principle all eligible households in these cantons will participate in either the agriculture and the youth intervention and there will be no control group in these cantons.

88


the households who do not take up the offer in wave 2. If the Jujutla data suggest that take-up rates are low, it may be necessary to offer the interventions to a larger number of wave 2 households, which would reduce the size of the control group.

Second, from the impact evaluation perspective, what matters is the number of households interviewed both at baseline and at follow-up. Usually, it is not possible to re-interview all households in the control group (and in the intervention group) towards the end of the pro-gram. Attrition may occur, for instance because households move to other municipalities or countries, because households fall apart etc. The survey firm hired to collect the data in Jujutla indicated that they expect an attrition rate of approximately 6%.

Third, in at least part of the analysis we probably cannot use control households that have children in schools that are not attended by children in the intervention households. The reason for this is that the so-called “inclusive education” interventions will be implemented only in those schools that are attended by children living in households that participate in the vocational training for mothers, vocational training for young adults, or agricultural produc-tivity interventions. Within those schools all pupils will be affected by the inclusive education interventions, even if they live in control households. Given that control households with children in these schools are equally affected by the inclusive education interventions, they serve as a (more) valid control group to measure the impact of the set of interventions that is the object of the impact evaluation.

Fourth, the baseline interviews in the selected cantons will take place during a busy agricul-tural season, when it is harder to find respondents.

These issues suggest that the number of cantons included may result in a control group that is too small to statistically identify meaningful effects of the program. UCW and the project team agreed to aim for a control group of 1000 households (after correction for the issues mentioned above). If the number of control observations appears to drop below 1000, it may be necessary to increase the number of cantons in which the interventions are implemented. Whether or not this increase in the number of cantons is necessary will be determined as soon as (i) the baseline data for the wave 2 cantons are available and (ii) we have a informa-tion on the take-up rate of the interventions implemented in the wave 1 communities.

A word of caution is necessary here. The project purposefully sampled cantons that are likely to have a high number of potential beneficiaries. Given that the number of potential benefi-ciaries per canton is much lower in the remaining areas (see Table 3) it will be expensive to increase the sample size. As a result only limited shortages in control observations can be dealt with.

6.2 Setting the Threshold Score

There are multiple possibilities when it comes to setting the threshold score. One option would be to use a fixed threshold score across municipalities and cantons. Another option would be to offer the interventions to a fixed percentage of households with child laborers, either at the municipality or at the canton level. In principle, the first option of a fixed thresh-old allows for a cleaner analysis in the impact evaluation, especially if there is still a control group in each canton. The final decision on this will be taken by the project team and UCW when the baseline data for the wave 2 municipalities has been collected.

89

Annexes

6.3 Smoothing the Vulnerability Index

The vulnerability index proposed by the project team is similar to the index used by other social protection programs such as PATI and Red Solidaria. However, this index does not appear to provide a smooth variable on the basis of which households can be selected. The histogram displayed in Figure 3 shows that there are large spikes in the density of the vulnerability index for households in the pilot municipality of Jujutla. These spikes may be problematic as they may lead to bunching of beneficiaries (with specific and endogenous characteristics) right above the threshold with few controls right below the threshold (or vice versa). UCW will use the wave 2 baseline data to propose a smoother vulnerability index.

Figure 3. Histogram of vulnerability index

0.0

5.1

.15

.2D

ensi

ty

50 60 70 80 90Vulnerability Index

6.4 DBMR in Control Group

Finally, as explained, the school attendance and participation in work of children in bene-ficiary households will be monitored. If possible, it would be interesting to implement this monitoring system also in control households. That way the impact evaluation would provide a purer insight into the effect of the vocational training of mothers. During our next meeting in El Salvador we will discuss in more detail the feasibility of such an approach.

7. Limitations

The impact evaluation of the “Eliminating Child Labour in El Salvador through Economic Em-powerment and Social Inclusion” project has multiple limitations. This subsection describes these limitations. First, the impact evaluation examines the causal effect of a combination of vocational training, flexible education, and monitoring interventions. However, it cannot iden-tify the impact of the individual components. Separate impact analysis could provide impor-

90


tant insights for the development of (the ILO’s) future household-level child labor programs. Unfortunately, it turns out that an impact evaluation that would disentangle the impact of one or more individual household level interventions is incompatible with the setup of the pro-ject. The household level interventions are administered as a package. When interventions overlap completely it is not possible to disentangle the impact of the individual components. There is no room to deviate from this procedure, because administering the interventions as a package is an integral part of the philosophy and design of the program.

Second, a series of interventions not part of the impact evaluation (such as capacity building among local administrations and school-level education interventions) is taking place in the background. These interventions do not reduce the internal validity of the impact evaluation, because they affect the households in the intervention and the control group equally. How-ever, they do imply that any impact identified should be interpreted as the interaction of the evaluated interventions with the background interventions. It cannot be guaranteed that a similar impact would be observed in the absence of the background interventions. Moreover, imagine that both the set of interventions evaluated and the interventions taking place in the background effectively eliminate child labor when implemented independently. In that case, the impact evaluation will (correctly) identify no impact of the evaluated interventions, because child labor has already been eliminated in both the intervention and control group by the interventions taking place in the background.

Third, the direct beneficiary monitoring system may be a source of bias in the impact eval-uation. Households in the intervention group will be interviewed about school participation and child labor on a 3 month basis. For multiple reasons, it is possible that these households start to adjust their answers to these questions in a way that is not consistent with actual changes in school participation and child labor. The households may, for instance, deny that children are in labor as they become aware that the program aims to eliminate child labor. The reliability of the measured impact on school participation and child labor will thus de-pend on possibilities to substantiate self-reported data with other sources of information. To see whether school participation improved, we might consider conducting independent spot checks in schools (if funds are available). For child labor we currently do not think there is a possibility to independently check the reliability of respondents answers. These issues will be discussed in more detail during the upcoming mission in El Salvador.

Fourth, there are some concerns about the statistical power of the current evaluation design. As shown in the appendix tables, our ability to detect statistically significant effects of the intervention drops quickly as the size of the control group shrinks. A low number of control observations could thus jeopardize the impact evaluation. We will more precise on of this issue once the baseline data are available and there is more information on the success of the interventions implemented in wave 1. If the control group turns out to be small, the team will try to increase the number of observations by moving into additional cantons. However, given that those cantons are likely to have a relatively low number of eligible households, this procedure will be successful only if the shortage in control observations is limited.

Finally, it should be noted that the regression discontinuity methodology only allows for the identification of a local treatment effect: the impact of the intervention for those households with a vulnerability score equal to the threshold. The average treatment effect for all indi-viduals participating in the intervention or heterogeneity of the treatment effects along the dimension of the vulnerability index cannot be identified.

91

Annexes

8. Planning and Timeline

The table below gives an overview of the timeline of the project.

Table 4. TimelineActivity Start FinishCensus and baseline April 2012 May 2012Vocational training women April 2012 October 2013Flexible education April 2012 October 2013Monitoring April 2012 November 2014Vocational training youth April 2012 October 2013Agricultural productivity April 2012 October 2013Inclusive education April 2012 October 2013Capacity building April 2012 December 2012Follow-up interview May 2014 May 2014

9. References

Attanasio, O. P., A. D. Kugler, and C. Meghir (2009). “Subsidizing Vocational Training for Disadvantaged Youth in Developing Countries: Evidence from a Randomized Trial.” IZA Discussion Paper 4251.

Cappelleri, J., R. Darlington, and W. Trochim (1994). “Power Analysis of Cutoff-Based Randomized Clinical Trials.” Evaluation Review, 18, 141-152.

Duflo, E., R. Glennerster, and M. Kremer (2008). “Using Randomization in Development Economics Research: A Toolkit”, Chapter 61 in Handbook of Development Economics, 4, 3895-3962.

Gertler, P. J., S. Martinez, P. Premand, L. B. Rawlings, and C. M. J. Vermeersch (2011). Impact Evaluation in Practice, The World Bank, Washington DC.

Goldberger, A. (1972). “Selection Bias in Evaluating Treatment Effects: Some Formal Illustrations.” Working Paper, Economics Department, University of Wisconsin.

ILO (2010). Project Document: Eliminating Child Labour in El Salvador through Economic Empowerment and Social Inclusion. The ILO, Geneva.

Imbens G. and T. Lemieux (Forthcoming). “Regression Discontinuity Designs: A Guide to Practice.” Journal of Econometrics, 142 (2), 615-635.

Khandker, S. R., G. B. Koolwal, H. A. Samad (2010). Handbook on Impact Evaluation: Quantitative Methods and Practices, The World Bank, Washington DC

Lee, D. S., and T. Lemieux (Forthcoming). “Regression Discontinuity Designs in Economics.” Journal of Economic Literature.

92


Manacorda, M. (2009). “Criteria and Guidelines for the Evaluation of the Impact of Child Labor Interventions.” UCW Working Paper.

UCW (2009a). “Technical Issues for the Impact Evaluation of Components of ILO / IPEC Projects.” UCW Working Paper.

UCW (2009b). “Technical Criteria for the Impact Evaluation of USDOL-Funded Child Labour Education Initiative Projects.” UCW Working Paper.

93

Annexes

Appendix 1: Shortlisted Communities

MunicipalityUrban or

ruralChild labor

ratePopulation Crime rate

JUAYUA URBAN 13,1 25.323 115ILOBASCO RURAL 11,7 68.976 88JUJUTLA RURAL 10,3 29.179 69NUEVA CONCEPCION RURAL 10,0 30.281 116APASTEPEQUE RURAL 9,9 19.493 77METAPAN RURAL 9,0 63.604 41GUAYMANGO RURAL 8,9 19.487 36SAN FRANCISCO MENENDEZ RURAL 8,9 44.085 29JIQUILISCO RURAL 8,8 49.777 80IZALCO URBAN 8,7 74.085 115TACUBA RURAL 8,6 30.718 68TECOLUCA RURAL 8,6 25.344 55COATEPEQUE RURAL 8,3 39.255 84SENSUNTEPEQUE RURAL 8,0 43.858 109EL TRANSITO RURAL 7,9 19.822 116SAN LUIS LA HERRADURA RURAL 7,9 21.388 122SAN PEDRO MASAHUAT URBAN 7,5 26.806 108AGUILARES URBAN 7,2 23.508 119SAN ANTONIO DEL MONTE URBAN 7,1 30.104 123SAN PABLO TACACHICO RURAL 6,9 21.751 51SANTIAGO DE MARIA URBAN 6,7 19.024 79

PANCHIMALCO RURAL 6,5 45.921 128

Child labor rates based on 2009 school census

Population rates based on 2007 population census

Crime rates based on 2010 national police statistics

94


Appendix 2: Map of Selected Communities

95

Annexes

Appendix 3: Power Calculations

Impact AlphaControl

householdsPower (beta)

(1) (2) (3) (4).1 .1 990 0.934.1 .1 900 0.9198.1 .1 810 0.9015.1 .1 720 0.878.1 .1 630 0.8474.1 .05 990 0.8848.1 .05 900 0.8638.1 .05 810 0.8378.1 .05 720 0.8056.1 .05 630 0.7652.08 .1 990 0.813.08 .1 900 0.79.08 .1 810 0.7628.08 .1 720 0.7305.08 .1 630 0.6919.08 .05 990 0.7193.08 .05 900 0.6912.08 .05 810 0.6588.08 .05 720 0.6215.08 .05 630 0.5784.06 .1 990 0.6028.06 .1 900 0.5785.06 .1 810 0.5514.06 .1 720 0.5209.06 .1 630 0.4867.06 .05 990 0.4806.06 .05 900 0.456.06 .05 810 0.4292.06 .05 720 0.3997.06 .05 630 0.3673

Proportion of working children

Estimated statistical power (column 4) for regression discontinuity estimates at different imact sizes (column 1), significance levels (column 2), and number of households in the control group (column 3). Estimates are triple those of randomized experiment estimates and assume intervention group of 2000 households.

96


Impact AlphaControl


(1) (2) (3) (4).1 .1 990 0.9305.1 .1 900 0.9158.1 .1 810 0.8971.1 .1 720 0.8730.1 .1 630 0.8418.1 .05 990 0.8793.1 .05 900 0.8578.1 .05 810 0.8313.1 .05 720 0.7985.1 .05 630 0.7577

.08 .1 990 0.8066

.08 .1 900 0.7833

.08 .1 810 0.7558

.08 .1 720 0.7233

.08 .1 630 0.6846

.08 .05 990 0.7111

.08 .05 900 0.6828

.08 .05 810 0.6503

.08 .05 720 0.6130

.08 .05 630 0.5699

.06 .1 990 0.5953

.06 .1 900 0.5711

.06 .1 810 0.5441

.06 .1 720 0.5139

.06 .1 630 0.4798

.06 .05 990 0.4727

.06 .05 900 0.4484

.06 .05 810 0.4218

.06 .05 720 0.3926

.06 .05 630 0.3606

Proportion of 18-45 year old women in paid work


97

Annexes

Impact AlphaControl


(1) (2) (3) (4)12 .1 990 0.938012 .1 900 0.924612 .1 810 0.907512 .1 720 0.885212 .1 630 0.856112 .05 990 0.889412 .05 900 0.869012 .05 810 0.843812 .05 720 0.812312 .05 630 0.77279 .1 990 0.77119 .1 900 0.74759 .1 810 0.72009 .1 720 0.68809 .1 630 0.65039 .05 990 0.66559 .05 900 0.63749 .05 810 0.60559 .05 720 0.56949 .05 630 0.52826 .1 990 0.47946 .1 900 0.45936 .1 810 0.43746 .1 720 0.41336 .1 630 0.38696 .05 990 0.35656 .05 900 0.33786 .05 810 0.31776 .05 720 0.29606 .05 630 0.2727


Monthly earnings of 18-45 year old women

98


�Annex 5. Baseline survey report

Acknowledgements

The design of the impact evaluation has been carried out in close collaboration with the project team and the government counterparts. The government formally agreed to carry out an impact evaluation of some interventions of the project on the basis of a wealth index. The project team has continuously supported the development of the IE design with its sugges-tions and assessments of constraints. It has also carried out the preparatory fieldwork with great care and insured the support of the different stakeholders.

Funding for this project was provided by the United States Department of Labor. This doc-ument does not necessarily reflect the views or policies of the United States Department of Labor, nor does mention of trade names, commercial products, or organizations imply en-dorsement by the United States Government.

1. Introduction

The ILO project called “Eliminating Child Labour in El Salvador through Economic Empow-erment and Social Inclusion” attempts to eradicate child labor in El Salvador. To achieve this objective the project implements a large range of interventions at three levels: macro (nation-al policies and institutional framework), meso (target municipalities and schools) and micro (child laborers’ households). One set of household level interventions, which is to be imple-mented in 5 municipalities in El Salvador, will be the object of an impact evaluation. This set of interventions consists of two main components: (i) a support package to help the mothers of child laborers start a small enterprise and (ii) so-called “flexible education interventions” which provide needs-based support to help the children of these women return to school at the appropriate level in case they have dropped out. A monitoring system will be setup to check whether beneficiary mothers and children indeed participate in the interventions and whether this participation is accompanied by behavioral changes such as improved school attendance and reduced child work (a description of this monitoring system can be found in Annex 1, in Spanish).

In April 2012, UCW (in collaboration with the ILO El Salvador) drafted an impact evaluation design (UCW 2012). As part of this design, UCW proposed to allocate the set of interven-tions to eligible households within the 5 municipalities on the basis of a wealth index. This procedure ensures a fair distribution of the interventions to those households who are most in need. At the same time, as explained in UCW (2012), this allocation procedure can be exploited in a regression discontinuity framework to identify the causal effect of the set of interventions on outcomes of interest, such as education, child labor, participation in labor of adult household members, and multiple other wellbeing indicators.18

The ILO recently conducted a baseline survey in the 5 municipalities where it will implement the set of interventions that will be the focus of the impact evaluation. Subsequently, UCW (again in collaboration with the ILO in El Salvador) used this data to generate the wealth index and assign the intervention. This document describes this process in more detail. In section 2 it starts by providing some background. Specifically, this section briefly discusses

18 The project initially aimed to implement a randomized experiment to measure the impact of the interventions. However, the Government of El Salvador did not support such an approach and as a consequence it was decided to adopt this quasi-exper-imental approach.

99

Annexes

the project at large, the interventions selected for the impact evaluation, and the basic idea behind the identification strategy of the impact evaluation. Section 3 proceeds by discussing the baseline survey and providing some descriptive statistics. Section 4 provides a detailed description of the procedure by which eligible households have been assigned to the inter-vention and control group. Section 5 examines whether this selection procedure was suc-cessful in the sense that (i) it assigned the poorest and most vulnerable households to the intervention group and (ii) it can be exploited to identify the impact of the interventions.

2 Background

2.1 The Project

As explained above, the ILO project called “Eliminating Child Labour in El Salvador through Economic Empowerment and Social Inclusion” attempts to eradicate child labor in El Salva-dor. To achieve this goal, the project set the following 6 objectives:

i. By the end of the project, child labor elimination strategies will have been adopted by national and local institutions that implement poverty reduction, decent work, and social protection programs for the rural and urban poor;

ii. By the end of the project, child labor law enforcement and protection mechanisms, par-ticularly for the worst forms of child labor, will be developed and fully operational;

iii. By the end of the project, national capacity to conduct child labor research, monitoring and impact evaluation will be enhanced and pilot interventions documented;

iv. By the end of the project, municipal capacity to prevent and eliminate child labor, in particular the worst forms, will have been strengthened;

v. By the end of the project, viable improved livelihood alternatives will have been imple-mented to reduce family reliance on child labor in target municipalities;

vi. By the end of the project, inclusive education models will have been implemented in selected schools in target municipalities to prevent and reduce child labor and improve education indicators.

The project’s strategic framework, outlined in ILO (2010), explains how these 6 objectives were identified. Figures 1 – 3 provide a bird’s eye view of the strategic framework. They re-spectively depict (i) the main factors that contribute to child labour in El Salvador, (ii) El Sal-vador’s wider national strategic agenda to eradicate child labour, and (iii) how this project’s 6 strategic objectives tie in with this agenda.

2.2 Interventions Selected for the Impact Evaluation

The ILO plans to implement a large number of interventions to achieve the 6 strategic objec-tives. Some of these interventions will be implemented at the national level, some in selected municipalities, some in selected schools, and some will be targeted at selected households.

100


One set of household level interventions targeted at households will be the object of an im-pact evaluation. This set consists of the following two interventions:

i. A support package that will be offered to help the mothers of child laborers start a small enterprise. This support consists of three steps: (i) vocational training of choice19, (ii) business training and preparation of a business plan, and (iii) a starting kit based on the needs identified in the business plan to kick-start the enterprise (value between US$ 100 and US$ 300).

ii. If the children living in the households of the eligible women do not attend school ac-cording to the baseline information, then they will be offered training to help them enter school at the appropriate level for their age. Kids will be invited to participate in this inter-vention by means of a sensitization campaign.20

2.3 Identifying the Impact of these Interventions

The impact evaluation aims to identify the causal effect (i.e. the impact) of this set of inter-ventions on education, child labor, and participation in labor of adult household members (especially women, as they are the key target group of the interventions). To identify the causal effect of this set of interventions, the impact evaluation must estimate a counterfactual outcome: the education and (child) labour outcomes that would have been observed among project beneficiaries in absence of the project. This counterfactual outcome is unobserved by definition. However, a valid counterfactual can be established by identifying a comparison group that is similar to the group of beneficiaries at the start of the project in all respects, except that they do not participate in the intervention.

As explained in detail in UCW (2012), to identify a credible comparison group, the impact evaluation will exploit the fact that the set of interventions described above is assigned on the basis of a wealth index. The idea is as follows. All households with working children and a mother who can participate in the intervention are ranked from poorest to richest on the basis of a wealth index generated using baseline data collected by the project. The poorest households (i.e. the households in the lower half of the wealth distribution) are assigned to the intervention group. The richest households (i.e. the households in the upper half of the wealth distribution) do not receive any interventions and constitute our control group. While these two groups are obviously not similar before the implementation of the interventions, they may be expected to be virtually identical close to the point in the wealth index that sep-arates the comparatively poor intervention households and the comparatively rich control households. By comparing intervention and control households close to this threshold, we can identify the (local) impact of the program.

The remainder of this document explains the implementation and validity of this procedure in detail. The discussion will assume that the reader is familiar with the basics of regression discontinuity methodology. UCW (2012) provides a graphical explanation of the regression discontinuity procedure for readers who are unfamiliar with the methodology. More in-depth discussion of the regression discontinuity methodology is provided by Imbens and Lemieux (2008) and Lee and Lemieux (2010).

19 In each canton a limited number of vocational training courses will be offered, in areas such as e.g. baking bread, sewing clothes etc. The courses on offer are determined on the basis of an analysis of the local labour market.

20 Out of school children are not obliged to participate and the mother’s participation in the vocational training does not depend on the participation of the child.

101

Annexes

3. Data

3.1 Baseline Data Collection

As mentioned above, the set of interventions that will be the focus of the impact evaluation will be implemented in 5 municipalities (Izalco, San Luiz Herradura, Santiago Nonualco, Tacuba, Tecoluca). Within these 5 municipalities, the ILO selected a group of 37 cantons into the project where they expected to find the highest number of child laborers. Annex 2 describes the procedures by which the municipalities and cantons were selected in detail (in Spanish). UCW (2012) provides a summary of this procedure (in English).

Baseline data were collected in these cantons in April and May of 2012. The baseline data collection consisted of two parts: (i) a short listing questionnaire administered to all house-holds in the 37 cantons and (ii) a more in-depth baseline questionnaire administered to all households identified as having working children in the listing exercise. Annex 3 to 7, contain the ILO-IPEC documentation describing the baseline data collection (including fieldwork, data consistency etc.), discuss the main results of the baseline data, and contain the admin-istered questionnaires for each of the 5 municipalities (in Spanish).

3.2 Identifying and Interviewing (Households with) Working Children

The working status of children (aged 5-17) was determined on the basis of the following 5 household roster questions in the listing questionnaire:

� In the past week, did […] work at least 1 hour?

� In the past week, did […] participate in any remunerated activities (cash or in-kind)? (Only asked if answer to previous question was “no”)

� In the past 12 months, did […] work in a non-agricultural enterprise?

� In the past 12 months, was […] involved in any farming activities?

� In the past 12 months, did […] work in an enterprise owned by himself/herself?

When the answer to any of these questions was “yes”, the child was considered to be a work-ing child. The census identified a total of 12,248 households with 16,597 children between the age of 5 and 17 (52,024 household members in total). A share of 35% of these children (5,856) was working according to the above definition. Working children were distributed across 3,294 households (i.e. 27% of all households had one or more working children).

After administering the listing questionnaires and processing the data, the survey team re-visited all households with one or more working children to administer the in-depth baseline survey. The baseline survey instrument contains sections that collect information on charac-teristics of the household members, educational attainment of household members, partici-pation in work by adults, participation in work by children, characteristics of the dwelling and durable goods owned by the household, loans taken out by the household, aid the household receives from other programs, and attitudes towards child labor. In total, the team managed

102


to track and administer this more comprehensive interview in 93% (3,064) of the 3,294 households with working children identified in the census.

3.3 Descriptive Statistics

We now proceed to provide some descriptive statistics based on the baseline survey data. We restrict the sample of households for which we provide descriptive statistics to those households that are, in principle, eligible to participate in the set of interventions that are the focus of the impact evaluation. Households are eligible to participate in these interventions if:

i. according to the listing data, they have one or more working children aged 5-17 (see definition above)

ii. according to the baseline data, they have one or more women fitting the following de-scription:

a. aged 18 to 65

b. knows how to read and write

c. not employed as a salaried worker

Of the 3064 households with working children interviewed at baseline 68% (2098) have a mother or another woman who is eligible according to these criteria.

Column (1) of Table 1 provides descriptive statistics for these 2098 eligible households and their household members. We turn first to the children living in these households. In total there are 5535 children aged 5-17 in these households. Nearly 85% of these children is in school and 88% of them is literate. The prevalence rate of child work (same definition as in the listing questionnaire) is 60%. Working children work about 16 hours a week on average. More than half of the children is involved in some form of dangerous work: working in an en-vironment with dust, smoke, or noise; exposure to heat or cold; working with dangerous tools, chemicals, or pesticides; working at night or early morning; carrying heavy loads; working in or on water. The majority of children (86%) is involved in household chores and children who are involved in chores spend a substantial amount of their time on this activity (12.5 hours a week on average).21

The idea is that increasing the earning capacity of the mothers of these children (or another eligible woman in their household) will help move children out of work and into school. So, what is the current working status of mothers and other eligible women? Do these women work and how much do they contribute to family income? To answer these questions we rely on the answers to the following three questions:22

� In the past week, did […] work at least 1 hour?

21 Interestingly, by construction, all of the households that were selected for the in-depth baseline interview had one or more working children according to the listing exercise. However, in 270 of the 2098 households included in the impact evaluation, there is no working child according to the baseline survey.

22 In accordance with the eligibility criteria we exclude mothers who are employed as a salaried worker.

103

Annexes

� Although […] did not work, did he/she have permanent employment or an enterprise owned by himself/herself to which he/she will soon return? (Only asked if answer to pre-vious question was “no”)

� In the past week, did […] participate in any remunerated activities (cash or in-kind)? (Only asked if answer to previous two questions was “no”)

Currently, nearly half of the eligible women in our sample work (i.e. answer “yes” to one of the three questions above). On average eligible women who work, work over 34.5 hours a week and earn nearly US$100 a month.

Finally, we look at a few selected household level wealth indicators. We observe that the eligible households can be described as disadvantaged. More than a third of household dwellings has a dirt floor, nearly half of all household dwellings has a roof made of metal sheets, rooms are shared by 5 household members on average, and only 36% of households has a fridge. Monthly per capita household income (earned by adult laborers) and monthly per capita household expenditures are respectively US$41 and US$33. Access to electricity and ownership of a TV (both available in roughly three quarters of the households) is fairly common.

4. Beneficiary Selection

A share of the 2098 eligible households described above has been selected for participation in the set of interventions. To ensure that the interventions are administered to households most in need, the intervention has been assigned on the basis of a wealth index. Households with wealth below a pre-specified threshold score in the index are considered poor enough to participate in the interventions and have been selected. Households with a wealth index above the threshold are deemed too rich and will not receive an invitation to participate in the intervention.

Selection on the basis of a wealth index not only ensures a fair distribution of the interven-tions to those households who are most in need. It also ensures that the causal effect of the set of interventions can be identified in a regression discontinuity framework. The intuition behind the regression discontinuity methodology is that households (and members of house-holds) with a wealth score just above the threshold are similar to households (and members of households) with a wealth score just below the threshold. Households right above the threshold score therefore serve as a valid comparison group to measure the (local) impact of the interventions (see UCW 2012 for more technical details). In this subsection we provide a description of this selection procedure.

4.1 Generating the Wealth Index

We start by describing how the wealth index was generated for the 2098 eligible households. We first list the wealth indicators (which we obtained from the baseline survey) on the basis of which the wealth index was constructed:23

23 Some of these wealth indicators, such as water, light, and sanitation, contain an additional category called “other”. We have not included this category in the wealth index, because it is not immediately clear what this category entails.

104


� A binary indicator for land ownership;

� The amount of land owned by the household for cultivation (in manzanas);

� The number of bedrooms in the dwelling divided by number of household members;

� Binary indicators for ownership of the following assets: radio, tv, video, fridge, iron, oven, vehicle for own use, vehicle for company use, sewing machine, grill, shelves, tables, plow, yoke, pump, barn, tools, and boat;

� Binary indicators for materials of which the dwelling’s roof is made: concrete, tiles made of clay or cement, asbestos sheets, metal sheets, or straw, palm, or waste materials24;

� Binary indicators for materials of which the dwelling’s walls are made: concrete, adobe, wood, metal sheets, or straw, palm, or waste materials 25;

� Binary indicators for materials of which the dwelling’s floor is made: ceramic tiles, ce-ment tiles, clay tiles, cement, or earth;

� Indicators for main source of to light in the dwelling: electricity, Kerosene (gas), Candles;

� Binary indicators for main source of water in the dwelling: public network inside home, public network outside home, water tank, water truck, well (private or public), or a lake, river, or stream;

� Binary indicators for type of sanitary facilities in the dwelling: toilet connected to sewage system, toilet connected to septic tank, private latrine, public latrine, private latrine (extra hygienic), public latrine (extra hygienic), or no sanitary facilities;

� Binary indicators for house ownership: rented, owned and paid off, owned and still pay-ing off, living in someone else’s house (without pay), borrowed).

Following Filmer and Pritchett (2001), we use the first principal component of the described wealth indicators as our wealth index. Because a proper technical discussion of principal component analysis is beyond the scope of this report, we do not discuss the technical de-tails of the procedure here.26 Suffice it to say that principal component analysis allows us to reduce the dimensionality of the dataset such that we obtain one composite index (the first principal component) representing the highest possible level of variation in the assembled wealth indicators.27

24 We combined 2 categories because number of observations is low

25 combined 2 categories because number of observations is low

26 Readers who wish more detail can, among many other texts, refer to a tutorial (Smith, 2002) that can readily be found online.

27 An important advantage of the procedure is that it is not necessary to rely on arbitrary assumptions (such as assigning weights to individual wealth indicators) to construct the index.

105

Annexes

4.2 Selecting the Poorest Households

Subsequently, we ranked eligible households according to the wealth index. The poorest 1098 households (i.e. the 1098 households with the lowest wealth index) were selected as beneficiaries while the remaining 1000 households serve as our control group. We did not stratify by municipality nor did we apply any other form of stratification. This procedure thus generates an implicit overall threshold in the wealth index, separating the poorest 1098 households (the intervention group) from the richest 1000 households (the control group) in the 5 municipalities. All households with an eligible woman and a wealth index below this threshold are selected, while households with an eligible woman and a wealth index above this threshold are not selected.28

4.3 Selecting Direct Beneficiaries within the Poorest Households

Within the selected households the interventions will be provided to individual women and children. Here we briefly discuss how beneficiaries within households will be selected. First, we look at the vocational training intervention provided to eligible women. One woman per household can participate in this intervention. When the mother of the children in the house-hold resides in the same household and is eligible according to the criteria described above, she and only she can participate in the intervention. When there are multiple eligible mothers in the household, the household can decide which of these mothers can participate in the intervention. When the mother of the children in the household does not reside in that house-hold herself, the household may decide which of the other eligible women in the household can participate in the intervention.

Flexible education interventions will be available to all 12 to 17 year olds who failed to com-plete their basic education and who live in the households of the mothers who receive voca-tional training. These children do not necessarily have to be involved in labor according to the listing or baseline data.

5. Validity of the Selection Procedure

Next, we check whether this selection procedure has been successful. We start by checking whether the wealth index successfully predicts wealth and is thus a proper tool to select the poorest households as beneficiaries. Then, we explore whether this selection procedure gives us a proper mechanism to identify the causal effect of the interventions in a quantita-tive impact evaluation. Finally, we check how selected individuals are geographically spread. As we explain below, geographical bunching of intervention and control households might invalidate the regression discontinuity design.

5.1 Does the First Principal Component Indeed Predict wealth?

Because the principal component procedure described above is rather abstract it may not be entirely clear that the wealth index it generates indeed properly predicts household wealth. In this subsection we therefore briefly show that the wealth index (the first principal component

28 Section XXX describes how the intervention and control households are spread geographically across municipalities and cantons.

106


of the listed wealth indicators) does indeed predict wealth and thus serves as an appropriate tool to distribute the interventions to the poorest eligible households. For this purpose, Fig-ures 4 – 10 graphically show that individual wealth indicators (e.g. ownership of a tv, owner-ship of a fridge, access to electricity, rooms per capita, no access to sanitary facilities, floor made of dirt, and roof made of metal sheets) of households with a high composite wealth index compare favorably with the individual wealth indicators of households with a low com-posite wealth index.

In each graph the horizontal index represents the wealth index. We normalized this index, such that it takes the value 0 at the implicit threshold that separates the poorest 1098 house-holds (the intervention group) from the richest 1000 households (the control group). With the exception of the variable “rooms per capita”, the vertical axis displays the proportion of households with the wealth characteristic under consideration (e.g. proportion of households owning a tv, proportion owning a fridge etc.). The dots are local averages and the lines are Lowess regressions, which provide a non-parametric estimate of the relationship between the calculated wealth index and the wealth indicator under consideration.29 The Lowess regres-sion for the intervention group is given in red (i.e. below threshold) and the Lowess regression for the control group is given in green (i.e. above threshold).

We see a clear relationship between the wealth index and the displayed wealth indicators. The ownership of TVs and fridges increases rapidly with the composite household wealth index. The same holds for access to electricity and the number of rooms per capita. On the other hand, the probability that a household has no sanitary facilities, has a dirt floor, or has a roof made of metal sheets (all indicators of poverty) drops rapidly with household wealth. These graphs thus strongly suggest that the wealth index is a useful tool to select beneficiary households on the basis of (observable and hard to manipulate) indicators of household wealth and poverty.

5.2 Validity of the Identification Mechanism

As explained in detail in the impact evaluation design, the selection procedure described above can be exploited in a regression discontinuity framework to identify the (local) impact of the vocational training interventions. The regression discontinuity methodology is valid only if households cannot precisely manipulate their wealth index (i.e. self-select into the inter-ventions) and if households just above and just below the wealth index are truly similar before the start of the interventions. In the case of the project in El Salvador, there is little reason to believe that the regression discontinuity design will not generate a valid control group. First, when people participated in the baseline interview, they did not know that they might be eligi-ble to participate in an intervention, nor that eligibility will be determined on the basis of their answers during this interview. Second, the procedure on the basis of which beneficiaries would be selected (described above) had not yet been determined at the time of the baseline interview. There is thus little room for precise manipulation of the wealth index.

To further ensure that the selection procedure indeed resulted in a valid control group, we conduct a series of additional tests. First, following McCrary (2008), we examine whether the density of the wealth index is continuous at the threshold. When the density of the wealth in-dex is not continuous this is an indication that there could be systematic differences between

29 We use non-parametric (Lowess) regressions as a convenient tool to graphically approximate the relationship between the wealth index and the individual wealth indicators. Other parametric and non-parametric procedures are available to approxi-mate this relationship (see Lee and Lemieux, 2010).

107

Annexes

the intervention and the control group.30 Figure 11, which was generated using McCrary’s software, depicts the density of the wealth index. We note that the density is of the wealth index is reasonably continuous. Most importantly, the confidence intervals indicate that the estimated density is not significantly different just above and below the threshold.

Second we test both graphically and formally whether the intervention and control group (households just above and below the threshold) are similar at the threshold. Here, we look not only at wealth indicators, but also at outcome indicators. While the wealth indicators in-cluded in the wealth index may be expected to be continuous at the threshold in the wealth index more or less by construction (because they are determinants of this same index), this is not the case for outcome variables. Our formal test relies on a regression discontinuity proce-dure written for Stata by Austin Nichols, which runs local linear regressions within an optimal bandwidth around the threshold score to estimate whether there are significant disconti-nuities in the wealth and outcome variables at the threshold score. The optimal bandwidth is estimated using a cross-validation procedure proposed by Imbens and Lemieux (2008), which balances the precision of the estimates (which increases with the bandwidth) against the bias that may result from using too large a bandwidth.

Figures 4-10 provide a first indication that households just above and below the threshold were indeed similar prior to the intervention. The scatter plots suggest that the relationship between the wealth index and the wealth indicators (TV, fridge, electricity, rooms per capita, sanitary facilities, floor made of dirt, and roof made of metal sheets) is roughly continuous at the threshold for selection into the program. And the Lowess regressions for the intervention (red) and control (green) group hit the threshold roughly at the same value of the wealth indicator.

The formal test also finds no evidence of discontinuities in these household level wealth indicators. The results of this test are displayed in Table 2. Column (1) shows the estimated discontinuity based on the regression discontinuity estimation procedure of Austin Nichols. Columns (2) and (3) respectively show the standard errors and P-values of the estimated coefficients. As can be inferred from the small discontinuity coefficients (column 3) and the high P-values (column 5). The only significant discontinuities in wealth indicators we do find (not displayed in the table) are in the probability of having an oven, the surface of land owned by the household, and whether the walls of the dwelling are made of adobe. Impor-tantly, there are also no discontinuities in other important wealth indicators not included in the wealth index, such as “female headed household” and “literate household head” (Figure 12 and 13 respectively).31

In terms of outcome variables (variables that are potentially affected by the intervention) households also look reasonably similar. Monthly per capita household income earned by adults (Figure 14) and monthly per capita household expenditure (Figure 15) are not dif-ferent in the intervention and control group. Eligible women look similar in terms of the

30 To understand why, consider the following explanatory example provided by McCrary (2008): “A doctor plans to randomly assign heart patients to a statin and a placebo to study the effect of the statin on heart attack within ten years. The doctor randomly assigns patients to two different waiting rooms, A and B, and plans to give those in A the statin and those in B the placebo. If some of the patients learn of the planned treatment assignment mechanism, we would expect them to proceed to waiting room A. If the doctor fails to divine the patients’ contrivance and follows the original protocol, random assignment of patients to separate waiting rooms may be undone by patient sorting after random assignment.” Similar problems may arise in the regression discontinuity context, for instance (in our case), when households manipulate their answers to the questions on the basis of which we constructed the wealth index such that that they end up in the intervention group (i.e. below the threshold) even though they should not be.

31 These variables are not included in the wealth index because they are endogenous (i.e. they are both an indicator of wealth but also a potential cause of household wealth).

108


probability of participating in work (Figure 16). However, we also observe some significant discontinuities for eligible women in terms of income from work (conditional on working) (Figure 17) and hours worked (conditional on working) (Figure 18).

Outcome variables for children (5-17) are also similar in the intervention and control group. Children right above and right below the threshold are equally likely to be in school (Figure 19), to be literate (Figure 20), to work (Figure 21), to be involved in dangerous work (Figure 22), and to participate in chores (Figure 23). Similar to eligible women, we do observe sub-stantive differences in the intensive margin of work (conditional on working) (Figure 24) and chores (Figure 25). Children in the control group work longer hours and they spend more time on chores.

Nonetheless, overall the regression discontinuity results suggest that the households just above and below the threshold are indeed similar in terms of wealth characteristics. There are some important differences in the extensive margin of work and chores for eligible wom-en and children alike. In principle, these observed differences at baseline can be accom-modated through difference in differences analysis and hence they should not invalidate the regression discontinuity approach.

5.3 Geographical Spread of the Intervention

We finish by examining one final aspect of the selection procedure that could invalidate the regression discontinuity design. Because the selection procedure is not stratified, it is possi-ble that the beneficiaries are bunched in some geographical areas while control households are bunched in other geographical areas. Geographical bunching could be problematic for the impact evaluation as the regression discontinuity methodology may pick up geographical differences in outcome variables rather than an impact of the administered intervention itself.

Table 3 displays the distribution of eligible households across the intervention and control group by municipality. Table 4 displays the distribution of eligible households across the intervention and control group by canton (the lowest geographical level of aggregation). At the municipality level intervention and control households are fairly evenly distributed. The municipality of Tacuba contains the highest percentage of intervention households (67%). In other municipalities the percentage of treated households ranges from 42 to 57%. At the canton level we observe larger differences in the percentage of intervention and control households. Yet, with only one exception (the canton of El Puente, which contains only 6 eligible households) all cantons contain both intervention and control households.

Given that all geographical areas and (sub-areas) contain both control and treated house-holds the geographical spread of intervention and control households does not appear to be cause for concern. We can simply add municipality or canton dummies in our regression discontinuity estimates to deal with any geographical differences. Otherwise, no adjustments seem to be necessary.

109

Annexes

References

Filmer, D. and L. H. Pritchett. 2001. “Estimating wealth effects without expenditure data-or tears: an application to educational enrollments in states of India.” Demography 38 (1), 115–132.

ILO (2010). Project Document: Eliminating Child Labour in El Salvador through Economic Empowerment and Social Inclusion.

Imbens G. and T. Lemieux. 2008. “Regression Discontinuity Designs: A Guide to Practice.” Journal of Econometrics, 142 (2), 615-635.

Lee, D. S. and T. Lemieux. 2010. “Regression Discontinuity Designs in Economics.” Journal of Economic Literature, 48, 281-355.

McCrary, J. 2008. “Manipulation of the Running Variable in the Regression Discontinuity Design: A Density Test.” Journal of Econometrics, 142 (2), 698-714.

Smith, L. I. 2002. A Tutorial on Principal Component Analysis.

UCW. 2012. Impact Evaluation Design: Eliminating Child Labour in El Salvador through Economic Empowerment and Social Inclusion.

110


Appendix: Tables and Figures

(1)Household level wealth indicators

TV 0.72Fridge 0.36Access to electricity 0.73Bedrooms per capita 0.19No sanitary facilities 0.05Floor made of dirt 0.36Roof made of metal sheets 0.49Female headed household 0.65Household head is literate 0.83

Household level outcome indicatorsMonthly per capita adult labor income (US$) 41.30Monthly per capita expenditures (US$) 33.04

Outcome variables eligible womenWorking 0.48Monthly wage (conditional on working) 98.70Average weekly working hours (conditional on working) 34.66

Outcome variables children 5-17In school 0.85Literate 0.88Working 0.60In dangerous work 0.54In chores 0.86Weekly hours worked (conditional on working) 16.26Weekly hours in chores (conditional on being in chores) 12.60

Table 1. Descriptive statisticsAverage eligible

households

Column (1) shows the average values of the wealth indicators and outcome variables displayed in the stub column.

111

Annexes

Discontinuity Std. Error P-value(1) (2) (3)

Household level wealth indicatorsTV 0.028 0.043 0.517Fridge -0.067 0.053 0.204Access to electricity -0.011 0.038 0.763Bedrooms per capita 0.020 0.024 0.393No sanitary facilities 0.002 0.016 0.908Floor made of dirt -0.058 0.051 0.262Roof made of metal sheets 0.012 0.055 0.827Female headed household 0.054 0.058 0.357Household head is literate 0.005 0.040 0.891

Household level outcome indicatorsMonthly per capita adult labor income (US$) -5.661 4.687 0.227Monthly per capita expenditures (US$) 0.363 2.654 0.891

Outcome variables eligible womenWorking -0.039 0.059 0.507Monthly wage (conditional on working) -28.826 16.193 0.075Average weekly working hours (conditional on working) 9.412 3.627 0.009

Outcome variables children 5-17In school -0.027 0.029 0.344Literate 0.005 0.027 0.863Working 0.050 0.037 0.176In dangerous work 0.058 0.036 0.107In chores 0.030 0.025 0.233Weekly hours worked (conditional on working) 3.671 1.278 0.004Weekly hours in chores (conditional on being in chores) 2.434 0.872 0.005

Table 2. Estimated discontinuities in baseline covariates and outcome variables at the threshold in the wealth

Column (1) displays estimated disontinuities in the wealth indicators and outcome variables displayed in the stub column at the threshold separating the intervention households from the control households (based on the estimation procedure of Austin Nichols). Columns (2) and (3) respectively show the standard errors and P-values of the estimated coefficients.

Table 3. Distribution of treatment and control individuals by municipality% Control % Treatment Observations

Tacuba 33 67 425Izalco 46 54 239Santiago Nonualco 54 46 755San Luis La Herradura 43 57 346Tecoluca 58 42 333Total 48 52 2,098

112


Tabl

e 4.

Dist

ribut

ion

of tr

eatm

ent a

nd c

ontro

l indi

vidua

ls by

can

ton

Can

ton

% C

ontro

l%

Tre

atm

ent

Obs

erva

tions

Can

ton

% C

ontro

l%

Tre

atm

ent

Obs

erva

tions

AG

UA

FRI

A64

3633

SAN

AN

TON

IO A

BAJO

5545

40A

MU

LUN

CO

7129

49SA

N A

NTO

NIO

ARR

IBA

3070

92A

REA

URB

AN

A14

867

SAN

AN

TON

IO L

OS

BLA

NC

OS

5050

163

CRU

Z G

RAN

DE

3169

61SA

N C

ARL

OS

5347

116

CU

NTA

N13

888

SAN

ISID

RO47

5317

EL A

RCO

6535

31SA

N JO

SE L

OM

A64

3622

EL C

ARA

O67

3378

SAN

JOSE

OBR

AJI

TO63

3797

EL C

HA

GU

ITE

1783

65SA

N JU

AN

2377

78EL

PO

RRIL

LO60

4048

SAN

LU

IS JA

LPO

NG

UIT

A58

4211

8EL

PU

ENTE

100

06

SAN

RA

FAEL

4456

99EL

RO

DEO

2971

77SA

N R

AFA

EL T

ASA

JERA

3169

39G

UA

DA

LUPE

LA

ZO

RRA

2773

63SA

N S

EBA

STIA

N A

BAJO

4951

170

LA E

SPER

AN

ZA25

754

SAN

SEB

AST

IAN

ARR

IBA

3367

6LA

PU

ERTA

5149

43SA

N S

EBA

STIA

N E

L C

HIN

GO

4753

81LA

S G

UA

RUM

AS

6238

85SA

NTA

BA

RBA

RA40

605

MER

CA

DO

6436

22SA

NTA

CRU

Z47

5345

MO

NTE

HER

MO

SO3

9730

SAN

TA T

ERES

A58

4276

PIED

RAS

PAC

HA

S53

4710

7TA

PALS

HU

CU

T50

506

QU

EBRA

DA

ESP

AÑ

OLA

5545

11To

tal

4852

2098

113

Annexes

114


115

Annexes

116


Figure 4. Household owns a TV

-.5

0.5

1

-5 0 5Wealth index

Local average Lowess interventionLowess control

Figure 5. Household owns a fridge

0.2

.4.6

.81

-5 0 5Wealth index


117

Annexes

Figure 6. Household has access to electricity0

.2.4

.6.8

1

-5 0 5Wealth index


Figure 7. Number of bedrooms per capita

0.5

11.

52

-5 0 5Wealth index


118


Figure 8. Dwelling has no sanitary facilities

0.2

.4.6

.81

-5 0 5Wealth index


Figure 9. Floor made of dirt

0.2

.4.6

.81

-5 0 5Wealth index


119

Annexes

Figure 10. Roof made of metal sheets0

.2.4

.6.8

1

-5 0 5Wealth index


Figure 11. Density of the wealth index

0.0

5.1

.15

.2.2

5

-10 -5 0 5 10Wealth index

120


Figure 12. Female headed household

0.2

.4.6

.81

-5 0 5Wealth index


Figure 13. Household Head Literate

0.2

.4.6

.81

-5 0 5Wealth index


121

Annexes

Figure 14. Monthly per capita adult labor income0

5010

015

020

025

0

-5 0 5Wealth index


Figure 15. Monthly per capita household expenditure

050

100

150

-5 0 5Wealth index


122


Figure 16. Eligible women in work

0.2

.4.6

.81

-5 0 5Wealth index


Figure 17. Monthly wage of eligible women (conditional on working and restricted to local average wages below 500 for easy viewing)

010

020

030

040

050

0

-5 0 5Wealth index


123

Annexes

Figure 18. Weekly hours worked by eligible women (conditional on working)0

2040

60

-5 0 5Wealth index


Figure 19. Child (5-17) in school

.2.4

.6.8

1

-5 0 5Wealth index


124


Figure 20. Child (5-17) literate

.5.6

.7.8

.91

-5 0 5Wealth index


Figure 21. Child (5-17) in work

0.2

.4.6

.81

-5 0 5Wealth index


125

Annexes

Figure 22. Child (5-17) in dangerous work0

.2.4

.6.8

1

-5 0 5Wealth index


Figure 23. Child (5-17) in chores

0.2

.4.6

.81

-5 0 5Wealth index


126


Figure 24. Weekly hours worked by child (5-17) conditional on working

010

2030

4050

-5 0 5Wealth index


Figure 25. Weekly hours in chores by child (5-17) conditional on being in chores

010

2030

40

-5 0 5Wealth index


127

Annexes

�Annex 6. Interim report on follow-up survey (Spanish)

128

IMPACT EVALUATION REPORT ONSupport mothers of child labourers to start a small enterprise and for reintegration of their children to school intervention of the Project “Eliminating Child Labour in El Salvador through Economic Empowerment and Social Inclusion”

129

Annexes

130


131

Annexes

132


133

Annexes

134


135

Annexes

136


137

Annexes

138


139

Annexes

140


141

Annexes

142


143

Annexes

144


145

Annexes

146


147

Annexes

148


149

Annexes

150


151

Annexes

152


153

Annexes

154


155

Annexes

156


157

Annexes

158


159

Annexes

160


161

Annexes

162


163

Annexes

164


165

Annexes

166


167

Annexes

168


169

Annexes

170


171

Annexes

172


173

Annexes

174


175

Annexes

176


177

Annexes

178


179

Annexes

180


181

Annexes

182


183

Annexes

184


185

Annexes

186


187

Annexes

188


189

Annexes

190


191

Annexes

192


193

Annexes

194


195

Annexes

196


197

Annexes

198


199

Annexes

200


201

Annexes

202


203

Annexes

204


205

Annexes

206


207

Annexes

208


209

Annexes

210


�Annex 7. Follow-up survey household questionnaire

211

Annexes

212


213

Annexes

214


215

Annexes

216


217

Annexes

218


219

Annexes

220


221

Annexes

222


223

Annexes

224


225

Annexes

226


227

Annexes

228


229

Annexes

230


231

Annexes

232


233

Annexes

234


235

Annexes

236


237

Annexes

238


239

Annexes

240


241

Annexes

242


243

Annexes

244


245

Annexes

IMPACT EVALUATION REPORT ON INTERVENTIONS ... - ILO

Documents

Transcript of IMPACT EVALUATION REPORT ON INTERVENTIONS ... - ILO