This is to certify that the
dissertation entitled

AN EVALUATION OF ARGENTINA’S JEFES YJEFAS DE
HOGAR PUBLIC EMPLOYMENT PROGRAM

presented by

Randall C Juras

has been accepted towards fulfillment
of the requirements for the

Ph.D. degree in Economics

 

 

 

/ / Major Professor’s Signature
3] 1/ 6:, 20/ 0

r

 

Date

MSU is an Affirmative Action/Equal Opportunity Employer

 

 

— —.—--~—.---.--c---a-n--u-o---.-o-o-o---o-u—--—.-.-— -

un—o-.--.-.--u-u—----.- -

 

PLACE IN RETURN BOX to remove this checkout from your record.
TO AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

DATE DUE DATE DUE DATE DUE

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

5/08 KzlProleccaPrelelRCIDateDueindd

AN EVALUATION OF ARGENTINA’S JEFES YJEFAS DE HOGAR PUBLIC
EMPLOYMENT PROGRAM

By

Randall C Juras

A DISSERTATION
Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of
DOCTOR OF PHILOSOPHY
Economics

2010

ABSTRACT

AN EVALUATION OF ARGENTINA’S JEFES YJEFAS DE HOGAR PUBLIC
EMPLOYMENT PROGRAM

BY

Randall C Juras

This dissertation concerns the use of public employment as an anti-
poverty measure following Argentina’s 2001-2002 economic collapse. All three
chapters focus on Argentina’s Jefes yJefas de Hagar social safety net
employment program.

The first chapter addresses the program’s unexpectedly high proportion of
partnered female applicants, which had caused the government some concern
insofar as many of these workers had not previously been counted among the
unemployed, and their participation thus did little to lower the official
unemployment rate. I estimate a structurally derived multinomial probit model of
application to Jefas, and find that individuals’ opportunity costs of time were
strongly correlated with the potential to earn income in market employment, with
the result that the optimal enrollee in a household was often a female “secondary
earner,” who had inferior access to the labor market. I also show that public
schooling likely played an important role in enabling women to enroll in Jefas.

The second chapter addresses how changes to the design of Jefas would
have affected the gender composition and poverty levels of applicants. In
particular, I use the reservation wages estimated in the previous chapter to

estimate the effect of two hypothetical changes: removal of the “one person per

household” rule, and perfect enforcement of the requirement that applicants be
primary earners who lost their jobs during the crisis. Counterfactuals are based
on weighted averages of characteristics, using probabilities predicted by a probit
model of participation. I find that in terms of the overall goal of the heads of
household program, which was to reduce poverty given a limited budget, it does
not appear that any of the potential changes outlined in this paper would
significantly improve performance, nor would allowing enrollment by more than
one member of a household have significantly changed applicants’ gender
composition.

The third chapter evaluates Jefes’ impact on child labor and school
attendance, using a comparison group of children whose parents applied to the
program but did not receive benefits. I develop both parametric and
semiparametric estimates of the program’s impact, which require different
assumptions, and reveal different information about the relationship between the
treatment and outcomes of interest. Using cross-sectional and longitudinal data, I
find that children age 10-14 whose parents enrolled in the Jefas program and
received benefits were less likely to report working by around 1.3 percentage
points, and more likely to report attending school by approximately 2.0
percentage points, compared with similar children whose parents were not

enrolled.

DEDICATION

To Anna, for somehow making
graduate school fun.

iv

ACKNOWLEDGMENTS

I was very fortunate to have an excellent, generous, and patient advisor,
Gary Solon, to guide me through the process of writing this dissertation. l
benefited greatly from his invaluable comments and challenging questions, and it
would be an understatement to say that this work would have looked much
different without him. I would also like to thank the members of my committee,
Stacy Dickert-Conlin, Stephen Woodbury, and Songqing Jin, for their
contributions and support. I couldn’t have asked for a better group of scholars to
guide me through my formative years as a researcher.

This dissertation would not have been possible without the help of my
friends and family, especially my wife and parents, who supported me
intellectually and emotionally not only during graduate school but also throughout
the entire rest of my life. I owe you one.

Many thanks to my fellow graduate students, particularly attendees at the
weekly meeting—Eric, Lizzy, Cris, Jeff, Stacy, Andrew R., Josh, Sarah and many
others—for your moral support. I also would never have made it this far had it not
been for my cycling, running, and camping buddies, including Floyd, Steve,
Charlotte, Adrianne, Brian and Cristina, Sean, Andrew C., and all the folks in
GLH3. Apologies to anyone I haven’t named; it doesn’t mean I don‘t like you. It’s

just that it really does take a village to raise a new PhD.

TABLE OF CONTENTS

List of Tables ................................................................................................
List of Figures ...............................................................................................

Chapter One ..

Structural Estimation of a Model of Workfare Enrollment: An Analysis of

Argentina’s Heads of Household Program ...................................................
1.1 Introduction ........................................................................................
1.2 Institutional Details of the Jefes Program ..........................................
1.3 Model of Participation Decision .........................................................
1.4 Model Estimation ...............................................................................
1 .5 Data ...................................................................................................
1 .6 Empirical Results ...............................................................................
1 .7 Analysis .............................................................................................
1 .8 Conclusion .........................................................................................
Appendix to Chapter One .......................................................................
References ..............................................................................................

Chapter Two
Design Changes and Estimated Profiles of Heads of Household
Beneficiaries .................................................................................................
2.1 Introduction .......................................................................................
2.2 Background on Evaluating Antipoverty Programs ............................
2.3 General Methodology .......................................................................
2.3.1 Using Current Income to Measure Well Being ........................
2.3.2 Participation of Singles ...........................................................
2.3.3 Attributes of Applicants ...........................................................
2.4 Allowing Enrollment by All Individuals ..............................................
2.5 Limiting Enrollment to Those Who Lost Their Jobs ..........................
2.6 Discussion ........................................................................................
2.7 Conclusion ........................................................................................
References ..............................................................................................

Chapter Three

The Impact of Jefes yJefas de Hagar on Children’s Work and School

Attendance ...................................................................................................
3.1 Introduction .......................................................................................
3.2 Child Work, Schooling, and Jefas in Argentina .................................
3.3 Semiparametric and Parametric Estimation of Treatment Effects

vi

viii

46
46
49
51
52
55
57
58
61
64
67
83

85
85
87
89

Chapter Three (continued)
3.4 Estimation of Treatment Effects for the Longitudinal Sub-sample
3.5 Data and Descriptive Statistics .........................................................
3.6 Results and Discussion .....................................................................
3.7 Conclusion ........................................................................................
Appendix to Chapter Three ......................................................................
References ..............................................................................................

vii

93
94
97
101
113
119

Table 1.1
Table 1.2

Table 1.3

Table 1.4

Table 1.5
Table 1.6

Table 1.7

Table 1.8
Table 1.9

Table 2.1
Table 2.2
Table 2.3
Table 2.4

Table 2.5

Table 2.6
Table 2.7

Table 2.8

LIST OF TABLES

Percent of All Beneficiaries Complying with Requirements .....
Data Definitions .......................................................................

Probit Estimation of the Probability of Participation,
Conditional on Applying to Jefas .............................................

Factors Influencing Date of Enrollment for Those Enrolled .....

Summary Statistics of Independent Variables by Applicant
Status .......................................................................................

Summary Statistics of Independent Variables by Gender of
Applicant ..................................................................................

Men’s and Women’s Reservation Wage Equations .................

Sensitivity Analysis—Main Regression with Only Enrolled
Applicants and Non-Applicants ................................................

Men’s and Women’s Reservation Wage Equations using
2002 Data ................................................................................

Tobit Regression Used for lmputing Income for Couples ........
Tobit Regression Used for lmputing Income for Singles ..........
Linear Regression Used for lmputing Income for Couples ......
Linear Regression Used for lmputing Income for Singles ........

Single Men’s and Single Women’s Reservation Wage
Equations .................................................................................

Comparison of Estimated and Actual Applicants .....................
Comparison of Applicants in Households Headed by Couples

Comparison of Applicants in Households Headed by Single
Men and Women .....................................................................

viii

31

32

33

34

35

36

37

41

42

72

73

74

75

76

77

79

8O

Table 2.9

Table 3.1

Table 3.2

Table 3.3

Table 3.4

Table 3.5

Table 3.6

Table 3.7

Table 3.8

Table 3.9

Table 3.10

Table 3.11

Table 3.12

Comparison of Primary Earners Who Lost Jobs (Couples) .....

Descriptive Statistics of Children of Applicants and
Participants ..............................................................................

Probit Regression for Calculating the Propensity Score ..........

Propensity-Matched Estimate of the Average Impact of
Participation on Work and School Attendance ........................

Probit Regressions for Work and School Attendance ..............

Probit Estimate of the Average Impact of Participation on
Work and School Attendance (ATT) ........................................

Propensity-Matched Estimate of the Average Impact of
Participation on Work and School Attendance Using the
Longitudinal Sub-sample .........................................................

Comparison of Point Estimates of Treatment Effect ................

Sensitivity Analysis: Propensity-Matched Estimate with Only
One Child Per Household ........................................................

Sensitivity Analysis: Propensity-Matched Estimate of the
Average Impact on Work and School Attendance for Children
Age 10—1 6 ................................................................................

Propensity-Matched Difference-in-Difference Estimate of the
Average Impact of Participation on Work and School
Attendance ...............................................................................

Difference-in-Difference Probit Regressions for Work and
School Attendance ...................................................................

Probit Difference-in-Difference Estimate of the Average
Impact of Participation on Work and School Attendance .........

ix

81

104

105

106

107

108

109

110

111

112

116

117

118

LIST OF FIGURES

Figure 1.1 Estimated Effect of Increasing Man’s Income on the

Probabilities of Participation ...................................................... 29
Figure 1.2 Estimated Effect of Increasing Woman’s Income on

Probabilities of Participation ...................................................... 30
Figure 2.1 Kernel Density Comparison of Applicants with Primary Job

Losers ....................................................................................... 69
Figure 2.2 Kernel Density Estimate—Primary Losers w/ Children ............. 70
Figure 2.3 Kernel Density Comparison of Job Losers ................................. 71

Figure 3.1 Overlapping Support in the Distribution of the Propensity
Score ........................................................................... 103

CHAPTER ONE

Structural Estimation of a Model of Workfare Enrollment:
An Analysis of Argentina’s Heads of Household Program

1.1 Introduction

Beginning in December 2001, Argentina experienced a serious economic
crisis accompanied by rapidly falling real wages and crushing poverty. At the
height of the crisis, the poverty rate roughly doubled from 30 percent to around
58 percent of the population, with 27 percent unable to meet even the most basic
needs. To help households cope with the effects of the crisis, the government
instituted a large-scale public employment program called Jefes yJefas de Hagar
Desaaupadas (Unemployed Heads of Household), or Jefas, that provided a cash
transfer in exchange for twenty hours per week of work and was modeled after
an existing workfare program in Argentina. Although lacking a specific expiration
date, Jefas was designed to provide temporary support for impoverished workers
with children and was never intended to provide long-term assistance (World
Bank, 2002). Program rules limited enrollment to one individual per household.

Jefas was notable for the high rate of take-up among partnered women, a

majority of whom had never before been active labor market participants and a

majority of whose male partners were employed.‘ Aside from taking officials by
surprise (Galasso and Ravallion, 2004; World Bank, 2003), this phenomenon
created some political difficulties for the govemment because the “new” workers
did little to lower the official unemployment rate. At the height of enrollment in
2002, women accounted for over 70% of the nearly 2 million program
beneficiaries, and more than 700,000 women entered the labor force specifically
to enroll in the program. Women with domestic partners—married or unmarried—
were responsible for the bulk of this phenomenon, accounting for 87% of the new
labor force entrants enrolled in the program. Viewed another way, nearly 60% of
partnered women who enrolled in Jefas entered the labor force to do so, whereas
single women who enrolled were largely drawn from unemployment. Why was
there such a surge in demand for workfare employment from partnered women?
The variables typically thought to have contributed to the observed long-run
increase in female labor force participation rates did not change suddenly as a
result of the crisis, so they cannot be responsible for the behavior of these
workers. Nor was there a corresponding increase in female labor force
participation outside the program. Rather, the workfare enrollment decision, in
the face of exogenous shocks to household income and employment, must have
been governed by a separate set of considerations. The goal of this study is to
elucidate the factors that led so many partnered women to apply for Jefas

employment, as a step towards understanding household labor supply responses

 

' I say ‘partnered’ rather than ‘married’ because formal marriage is relatively uncommon
in the demographic of interest; these relationships are typically equivalent to marriage for
practical purposes.

to economic downturns, and the wide range of women’s take-up rates in public
employment programs in general.

In this paper, I develop a model of participation in which the probability
that an individual enrolls in workfare depends on that individual’s opportunity cost
of time, or reservation wage, as well as the reservation wage of his or her
partner. Reservation wages are in turn functions of market wage potential,
household financial resources, and the value of time spent in home production.
Because enrollment in Jefas is limited to one individual per household, family
members must compare the values of their reservation wages. In partnered
households, the final enrollment decision depends on the relative values of the
man’s and woman’s reservation wages, and the wage offered by the government
for Jefas employment. I use likelihood maximization techniques to generate
parameter estimates.

I find that higher potential market wages in non-Jefas employment were
associated with lower probabilities of participation in the program. Presumably
the rigid, mandatory work requirement associated with Jefas served as a
deterrent for individuals who had market employment possibilities—which was,
after all, its purpose. At the same time, Jefas was attractive to other family
members in those same households who did not have the potential to earn
significant income, and who may even have been unemployable at prevailing
market wages. I also find that having more school-age children increased a

woman’s probability of applying to the program, which seems inconsistent with

previous studies showing that fertility decreases women’s labor force
participation. I attribute this phenomenon to the wide availability of low-cost, high-
quality public schooling.

The paper is structured as follows: I describe institutional details of the
Jefas program in Section 2. Section 3 models the household’s workfare
enrollment decision. Section 4 details the econometric specification. Section 5
describes the data. Section 6 presents empirical results. Section 7 analyzes the

results in terms of the added-worker effect, and Section 8 concludes.

1.2 Institutional Details of the Jefas Program

Recognizing the severity of the 2001 -2002 economic crisis, the Argentine
government instituted the emergency social safety net program Jefas in April
2002, with partial financing from the World Bank in the form of a $600 million
loan. Jefas was closely modeled after an existing workfare program that predated
the crisis (Plan Trabajar) but Jefas imposed more restrictions on enrollment.

The stated purpose of Jefas was to provide direct income support for
families with dependents who had lost their main source of earnings due to the
crisis, and an explicit secondary goal was to reduce the official unemployment
rate by classifying participants as employed. This program was the government’s
primary safety net response to the crisis. Most other safety net programs were
eliminated or reduced in order to shift funding to Jefas at the height of

participation in 2002, government spending on Jefas was about 1% of GDP.

As originally implemented, Jefas transferred AR$150 per month (about
US$50) to participants, who were enrolled on a first-come, first-serve basis
subject to eligibility. Four eligibility criteria required beneficiaries to:

- be unemployed;
- be the head of a household;
- live in a household with at least one minor below the age of 18, a pregnant
woman, or a handicapped individual of any age;
- work (or participate in training or education activities) for 4-6 hours a day
in exchange for the payment.
Notably, the “head of household" requirement limited enrollment to one individual
per household, which was an attempt to curtail costs while spreading benefits to
as many families as possible.

Jefas was overseen at the national level by the Ministry of Labor and
Social Security, but the Ministry only provided the general guidelines of the
program. Administration took place at the local level, with municipalities,
churches, and organizations of the unemployed (piqueteras) enrolling
participants and later organizing workfare activities. This was meant to ensure
that projects were well targeted to the specific needs of the communities, and the
tasks were organized by non-governmental organizations and non-profits that
already operated within these localities. In theory, this should also have screened

out individuals whom local administrators knew to be ineligible.2 By registering

 

2 Increasingly, this sort of decentralized, or community—based targeting is a precondition
for World Bank financing. Ravallion (2006) reviews the literature on income transfer

workers in a national database and assigning them social security numbers, it
was also intended to help formalize the labor market.

To prove eligibility, participants were required to show documentation to
program administrators that verified unemployment status and legal guardianship
of a minor. These documents were then cross-checked against Social Security
Administration lists. Verifying unemployment proved not to be feasible, however,
since 50% of economic activity in Argentina is estimated to take place in the
informal economy, and only employment in formal sector jobs is verifiable. Thus,
this requirement appears to have been rarely enforced in practice. Administrators
were also not able to check whether an applicant was really the head of a
household, and in fact the concept was not even well defined, so this requirement
was also not enforced.3

In contrast, there is ample evidence that households believed the work
requirement would be binding at the time they registered. The requirement was

heavily publicized in advance of enrollment, and a program evaluation by the

 

programs, and Canning and Kevane (1999) discuss community-based targeting
mechanisms. In the case of Jefas, this has caused some problems. There have been
widespread complaints of rent-capture and favoritism by the piqueteras, who administer
about 10% of the workfare activities. Participants also complain that if not for Jefas, the
government would have to hire them to perform the same work anyway, and pay them a
higher wage.

3 While not formally defined, it is my impression from reading program documents that
the government intended to give benefits to primary wage earners who had become
unemployed. “Head” status was entirely self-reported. Nonetheless, the name of the
program, “Heads of Household,” reflects the centrality of this concept to Jefas’ designers.

Ministry of Labor reports that the majority of beneficiaries (93% of women and
81.6% of men) fulfilled the conditional work requirement (MTESS, 2004).4

Other than the work requirement and the dependent—child requirement, it
appears that in practice the eligibility requirements were not enforced (which, in
some cases, would have been nearly impossible to do). Table 1.1 gives some
perspective regarding the lack of enforcement. Thus, in practice, Jefas was
effectively a cash-for-work public employment program targeted at families with
children. Due to lax enforcement, it allowed one household member to
participate, regardless of official employment status or head of household status.

Besley and Coate (1992) establish that self-targeting to individuals through
a work requirement can be efficient if the wage is set below the market wage for
unskilled labor in the informal sector, so that only individuals with a very low
reservation wage have an incentive to participate, and Ravallion and Datt (1995)
show that such a scheme can be cost-effective in practice. The Jefas wage was
set to meet this criterion: Jefas required that beneficiaries participate in an
eligible work or job-training activity no less than 4 hours per day, 5 days per
week, and the monthly transfer of AR$150 was set below the expected full-time

wage rate for households in the bottom decile of the income distribution.5

 

4 The World Bank (2006) reports that while initially high, compliance had fallen to about
69% by the beginning of 2005, and is expected to deteriorate further. For the purposes of
studying enrollment, what matters is the expectation that the requirement is binding.

5 The World Bank calculates the expected hourly wage rate (averaging formal and
informal work) in the first income decile as AR$O.97 and in the second decile as
AR$2.01. This corresponds to monthly incomes of AR$180 and AR$370, respectively. In
my data, only 8% of men and 12.6% of women who work at least 20 hours per week earn

Jefas initially experienced rapid enrollment growth, increasing in size from
574,000 participants in May 2002 to 1,857,000 in Dec. 2002, with aid reaching an
estimated 10 million family members in participating households. Mostly as a
response to budget constraints in light of the rapid growth, registration unofficially
closed in Sept. 2002 and officially closed in May 2003 with close to 2 million
enrolled,6 and enrollment steadily decreased to about 1.4 million in 2006. The
growth in the program exceeded what both the Bank and the government had

expected based on estimates of the target population (World Bank, 2006).

1.3 Model of Participation Decision

There is a reasonably large literature on targeting through workfare. Early
contributors include Akerlof (1978) and Nichols and Zeckhauser (1982), who say
that the imposition of “ordeals” on welfare claimants may improve targeting. The
first detailed theoretical analysis of the screening argument for workfare was
Besley and Coate (1992). In their model, developed to explore the incentive case
for workfare, an individual decides between spending time working as part of the
government program and using that time for other purposes, such as leisure,
home production, or wage labor. A prediction of their model is that, once the

wage offered as part of the benefits package has been set, an individual will

 

less than AR$150 /month. Mean income for men is AR$6OO /tnonth and for women is
AR$450 /month.

6 Some additional registration appears to have occurred in the year following the
unofficial end of registration, which did not carry legal weight. Beneficiary numbers
increased significantly just before the presidential election of 2003, and registration was
legally closed in May 2003 (World Bank, 2006).

accept the government’s offer if and only if his or her opportunity cost of time is
below that wage offer. Then, modeling an individual ’3 opportunity cost of time as
a function of observable characteristics, with some error, it is possible to predict

the probability that an individual enrolls as:

(1) pr (individual enrolls) = pr (opportunity cost 5 government’s wage offer)

An analysis of the Jefas program must differ slightly from this model because
Jefas allows enrollment of only one person in any household. Thus, the
individual’s decision will necessarily be affected by the incentives of other family
members who may wish to enroll, especially the individual’s partner, and this
process must be explicitly modeled. Assuming that household-level decision
making involves only the household head and his or her partner (and that all
households contain such a pair), the final decision on whether to enroll will take
place only after negotiation between those partners. 7 Nonetheless, it seems
reasonable that the same basic principle applies—individuals compare their utility
from enrollment with the alternative of non-enrollment, and wish to enroll only if
they consider enrollment the more valuable option. Simultaneously, partners
must also have some negotiating mechanism by which they compare their

relative opportunity costs. At the conclusion of negotiations, an individual will

 

7 Any framework for infra-household decision-making that implies a Pareto efficient
outcome would be sufficient for the purposes of this model. For examples of potential
bargaining frameworks see Chiappori (1992). The purpose of this research is not to test
any one such framework

apply to enroll if and only if two conditions are met: (a) that individual’s
opportunity cost of time is judged to be the loweriof the two, and (b) the value
associated with enrollment in Jefas exceeds the opportunity cost of time.
Hereafter, I refer to the lowest pecuniary wage offer that would satisfy condition

(b) for an individual as that lndividual’s reservation wage.

There are three observable outcomes of the application decision: yo, y1, y2
6 {0,1} which correspond to the events that both decline the government’s offer

(yo = 1), the man applies (y1 = 1), and the woman applies (y2 = 1), respectively.
Denote the Jefas wage offer W J, which is a scalar, the man’s reservation wage

WM, and the woman’s reservation wage WF. The probability of observing each

outcome can be fully described as:

(2) PFIYO=ll=PrleSWM§DfiWJSWFl
(3) pr(y1=1)=pr(WMsWFam1WMsWJ)
(4) PT(Y2=1)=Pr(WFSWMafldWF-"WJ)

where pr (yo = 1) + pr (y1= 1) + pr (ya = 1) = 1. In other words, the individual will

apply to enroll if his/her reservation wage is lower than the Jefas wage offer and
if his/her reservation wage is also low relative to his/her partner.

In practice, this model allows for estimation not only of the man’s and
woman’s enrollment probability, but also of the determinants of the man’s and

woman’s reservation wages. I allow the reservation wage to depend primarily on

10

three components: individuals’ market wage opportunities and working
conditions, the value of their time spent in home production, and the value of
other financial resources (assets) available to help weather the crisis.

Unlike much of the existing literature, I do not incorporate the assumption
that the reservation wage is equal to the market wage for employed individuals.
This assumption would be troublesome in the context of this study for two
reasons: First, it implies that work hours are flexible for those working, so that
individuals may adjust them to equate the shadow price of time with the market
wage. This was clearly not possible for many in the Argentine economy in 2002,
when employers were restricting hours and cutting employment in the face of the
recession. Second, it implies that individuals will take a competing job offer if and
only if the wage rate is higher than their current wage rate, which rules out the
idea of compensating differentials. Workfare employment in particular is likely to
differ from other market employment in a number of dimensions, including the
perceived temporary nature of the program, the fixed hours, potential job-training,
and/or potentially better working conditions mandated for government
employment. To mitigate these potential problems, I allow the reservation wage

to differ from the observed market wage for employed individuals.

11

1.4 Model Estimation

I estimate equations (2), (3), and (4) by specifying the functional form of
the two reservation wage equations for any given household 1' as follows
(suppressing the i subscript for clarity):

(5) log W = aim 8M
(6) log WI.- = 5,3)“ 6F
where X is a vector of individual and household characteristics (described below),

and eMand s): are jointly normally distributed according to:

trill) 2°2le

The likelihood for household i can be written as

 

[i = (PTO’1 =1))y1(Pr(y2 = 1))y2 (pr(y0 = 1))l’y I'y 2

where pr(yo = 1) + pr(y1 = 1) + pr(y2 = 1) = 1. The likelihood function for all N

observations is

N
L-na.

i-I

With small w’s signifying (log W,- / o), and defining [3M -0611, m) and pp = ()3; m)

the probabilities of the events corresponding to (2), (3), and (4) are:

Prtyo = 1) = ¢2(l3M’X- WJ, fiF’X- WJI P)

12

fi}X-fi},4X

w -3 X’___|_1-_p
J M VZ—Zp 1lz-zp

 

Pf()’1 =1) " (D2

flMX-flFXI l-p

w - 'X;——————— .
J flF Jz-zp 2-2pl

(1)2 is the bivariate standard normal cumulative distribution function. The resulting

 

prO’z =1) 3 (1)2

 

estimator is the maximum likelihood estimator of a trinomial probit with correlated
errors, and is consistent, asymptotically efficient, and asymptotically normally
distributed.8 Estimates of 6,, and 8.: were obtained using State’s <ml> command
for maximizing user-generated likelihood functions. Predicted probabilities of

each outcome were also generated for each household.

1.5 Data

Prior studies of labor force participation, including Heckman (1974),
Gronau (1973), Dooley (1982), and Cebula and Coombs (2008), guide the
selection of independent variables, which I describe in Table 1.2. In general,
these characteristics will affect reservation wages either through the expected
market wage, the value of home production, or access to financial resources. A
higher market wage opportunity, or higher-valued home productivity, will
positively affect the reservation wage if either must be foregone to participate.
Access to financial resources is expected to increase the reservation wage

through the income effect, assuming that leisure is a normal good.

 

8 The appendix gives the full model specification.

l3

Data come from Argentina’s Permanent Household Survey (EPH), which
the National Institute for Census and Statistics (INDEC) conducts twice per year,
in May and October. The EPH collects information from households in large
urban areas that account for 70% of the Argentine population; rural areas are not
sampled. Each household is interviewed in four consecutive surveys, so that at
any given time up to a half of the interviewed households will be linked as a panel
to the corresponding survey from the previous year, depending on attrition. This
study uses data from October 2001, two months before the economic collapse,
and October 2002, nearly a year after the crisis and after the bulk ofenrollment in
Jefas had occurred.

An advantage of the EPH is its commitment to anonymity, which means
that respondents may report information that would have resulted in
disqualification from Jefas without fear of reprisal. In fact, about 45% of partnered
female Jefas beneficiaries report being the “spouse of the household head” in the
EPH survey. Likewise, about 5% of beneficiaries are in a household having more
than one beneficiary. Both of these situations are in clear violation of program
rules. Households that report having more than one beneficiary are excluded
from the data, because that situation is incompatible with the empirical model.

In the EPH, individuals within a household are separately interviewed, but
a common household identification number links their answers. Each person’s
relation to the household head and civil status (single, married, partnered,

widowed, or separated/divorced) are reported. Because the focus of this study is

14

on partnered couples, households are included in the sample only if there is a
male-female cohabitating couple identified as married or partnered, and complete
information on both individuals. This leaves 10,678 households, of which 4,435
are linked as a panel to the 2001 data.

In the October 2002 survey there was also a special module on Jefas
participation, for the purposes of program evaluation, which was administered to
both Jefas participants and individuals who had applied as of the survey date but
were still on the waiting list.9 In the empirical analysis that follows, both of these
groups—enrolled and non-enrolled applicants—are pooled together into a group
of “applicants,” for the reason that all applicants, including those not yet enrolled,
should have self-selected into the program based on program requirements and
their opportunity costs of time in a similar manner. That is, I assume that all
applicants, at the time they applied to Jefas, anticipated accepting the
government’s offer of enrollment if and when it came. There are two reasons that
this assumption might prove troublesome, and applicants still on the waiting list
might differ from enrollees. First, if applying was not costly, individuals who were
unsure of their future reservation wages may have signed up, knowing that if the
state of nature were good at the time of enrollment they could simply decline the
government’s offer at that time. Second, enrollment was on a first-come, first-
serve basis, and the neediest households, or those first affected by the crisis,

might have been motivated to apply for program benefits at the earliest dates.

 

9 Galasso and Ravallion (2004) use the group of Jefas applicants as a control group in
their analysis of foregone income and of the program’s effect on the overall
unemployment rate.

15

Those households would then be more likely to appear in the group of enrollees.
To test whether applicants on the waiting list differ in observable ways from
enrollees, I estimate a probit regression of enrollment on household
characteristics, conditional on application (Table 1.3). Individuals who were
unemployed prior to the crisis, construction workers, public employees, and those
from large households appear more likely to receive benefits, which is consistent
with anecdotal evidence that the construction industry was one of the hardest hit
sectors immediately following the economic collapse, and that early in the
program, municipalities fraudulently enrolled their employees as a budget-cutting
strategy. For enrollees (but not applicants still on the waiting list), the data also
include the date of enrollment. Because enrollment was by and large first-come,
first-serve, I can test whether similar characteristics are correlated with the date
of application for program enrollees. A regression of enrollment date on
household and individual characteristics, conditional on enrollment (Table 1.4),
shows that women who had an unemployed spouse were likely to enroll earlier,
as were female domestic workers (who have a flexible schedule and are usually
paid under the table) and men who have more dependent children. Nonetheless,
enrolled and non-enrolled applicants look quite similar compared with the general
population, and I will combine the two groups—hereafter referred to simply as

applicants—for this study.lo Additionally, in approximately 0.5% of the original

 

‘0 In addition, Galasso and Ravallion (2004) conduct a more thorough analysis and
conclude that non-enrolled applicants are a suitable control group for enrollees in their
study of foregone income and of the program’s effect on the overall unemployment rate. I
perform a sensitivity analysis by excluding non-enrolled applicants from the data; the

16

sample of applicant households, the applicant was an individual other than the
head or spouse. These households have been dropped from the sample.

Table 1.5 contrasts descriptive statistics of non-applicants with the group
of applicants. These statistics paint a compelling, if rudimentary, picture of factors
influencing the enrollment decision. Applicants are on average younger, lower- 1
income, have more children, and have fewer assets than non-applicant couples.
Table 1.6 compares the descriptive statistics of male applicants with female
applicants. While there are fewer differences between these two groups, a few
intriguing patterns emerge. Specifically, men are more likely to have worked in
2001, worked more hours, and earned more income in households with a female
applicant. Likewise, women worked more often and, conditional on working,
earned more income in households with a male applicant. Finally, women in the
latter households were more likely to have had non-labor income. While this
begins to elucidate some of the important factors involved in the application
decision, the maximum likelihood analysis will provide additional insight into the

factors that drive the gender of the enrollee.

1.6 Empirical Results

Factors used to predict the enrollment decision include variables

correlated with individuals’ potential market wages, the value of home production,

 

results are presented in the appendix in Table 1.8, and are qualitatively similar to those
from the main regression.

l7

and household assets available to help weather the crisis. Because the potential
post—enrollment market wage is not observable, especially for enrollees who
receive benefits in exchange for working, I include several factors that I expect to
be correlated with it among the vector of characteristics used to explain
reservation wage. These factors include applicants’ 2001 monthly earned
income, education, and age. The need for data on pre-crisis monthly income
limits the sample to the 4298 households linked as a panel, including 653
applicants.‘1 A possible concern with this approach is that applicants might have
experienced larger declines in real income as a result of the economic collapse
than did non-applicants. If that were the case, the model would underestimate the
magnitude of the effect that income has on the household’s decision. Results
from the maximum likelihood estimation are presented in Table 1.7. Coefficients
represent the estimated effect of the explanatory variables on individuals’
reservation wages; these are then fed back into the model to generate the
predicted probabilities of each observable outcome for each household.
Variables associated with an individual ’3 market wage prospects are
estimated to have the biggest effect on the application and enrollment decision,

while variables related to the partner’s market wage contribute to a lesser

 

11 As a robustness check, I estimate an identical'model using individuals’ actual 2002
monthly income and hours in place of 2001 income. Applicants who are actually enrolled
are not included in this regression. Results are presented in the appendix in Table 1.9.
Despite endogenously truncating the sample by excluding participants, the estimates are
similar to those from the main regression. For common observations, the predicted
probabilities of men and women applying to Jefas (y) = 1 and yz = 1) generated by the
two models have correlations of .75 and .78, respectively. As would be expected if
applicants experienced larger drops in income than non-applicants, the estimated effect of
income on enrollment is relatively larger in the model using actual 2002 income.

18

degree. The most important predictor, in terms of both statistical and practical
significance, is monthly labor income from the previous year’s survey. Evaluated
at values of the independent variables typical of an applicant,” and holding hours
constant, an increase in a man’s 2001 monthly income from zero to 400 pesos
decreases the estimated probability that he applies to the program by roughly two
thirds, from 0.163 to 0.059, with the biggest marginal change at income = 0
(Flgure 1.1).l3 Simultaneously varying 2001 weekly hours from 0 to 40 makes
little difference, giving estimated probabilities of 0.175 and 0.056. I take this as
evidence that reservation wages are influenced by an individual’s total income
earning ability, rather than that individual’s hourly wage. For the same increase in
the man’s income, the predicted probability that the woman applies declines by
less than one third, from 0.193 to 0.152.

Likewise, an increase in the woman’s income from zero to 400
pesos/month decreases her predicted application probability from 0.210 to
0.063—a 70% drop—while the probability of her partner’s application remains
relatively constant, even slightly increasing, as shown in Figure 1.2. This occurs
because while the man’s reservation wage increases, it is far more likely that the
woman’s reservation wage rises above his than that his rises above the

threshold.

 

‘2 To describe the “typical applicant,” I use the mean values of explanatory variables for
the pool of enrolled and non-enrolled applicants in the panel sample.

'3 Median 2001 male income in the sample is about 565 pesos/month, and median female
income is 425 pesos/month for those who earn income.

19

Education, which has long been shown to be positively correlated with
market wages, has a similar effect. An increase in the man’s education from
completed primary to completed secondary education (7 years to 12 years)
results in a small but significant 16% decrease in his predicted probability of
applying to Jefas, from 0.102 to 0.086, while the same increase in a woman’s
education decreases her application probability from 0.189 to 0.176, aeteris
paribus. Flexible working conditions, which may increase applicants’ ability to
retain current employment while working for Jefas benefits, should be expected
to negatively affect the opportunity cost of time and lead to increased
participation. This may be the case: a typical man who reported being self-
employed in 2001 is estimated to be three percentage points more likely to have
applied to the program. On the other hand, the estimated effect of a woman’s
status as a domestic worker indicates that the causal interpretation may not be
so clear—women who were employed as domestic workers are significantly less
likely to be participants, which may indicate that they chose domestic
employment because they value flexible work hours. Age variables are estimated
to increase reservation wages, consistent with previous studies that use them to
proxy for labor market experience.

The estimated coefficient on the weekly hours variable is more difficult to
interpret. Holding income constant, an increase in the number of work hours
implies a lower hourly wage, which I would expect to be correlated with a lower

reservation wage. On the other hand, moving from a very small number of hours

20

to a large number may result in less flexibility, and a higher reservation wage, as
discussed above. The estimated coefficient on an individual’s weekly hours is
positive and significant for each individual—implying a higher reservation wage—
but has little effect on the predicted probabilities of application. For example, an
increase from 10 to 30 in an individual’s weekly work hours, holding income
constant, changes the predicted probability of that individual’s participation by
less than 10%. I interpret this to mean that the hourly wage rate is less important
than total income in the decision to apply for Jefas public employment. This is
consistent with evidence that employers curtailed hours in the face of the crisis,
with the result that employees could not increase their work hours to earn extra
income.

Controlling for the market wage variables above that appear central to the
enrollment decision, several other factors were estimated to have influenced
participation to a lesser, but still noteworthy, extent. These include non-earned
income and assets, and demographic variables that contribute to the value of
home production. These are each briefly discussed below.

Non-earned income and assets should theoretically be important
contributors to the work decision through the income effect. Indeed, the man’s
and the woman’s non-earned incomes (for example, investments and pensions)
are both significant variables in the woman’s reservation wage equation, with
both leading to a decreased probability of applying to Jefas. The marginal effect

of unearned income on participation appears large in magnitude—in fact, as

21

large as the estimated effect of earned income. Yet it seems not to have played a
large practical role in the overall rate of participation, since the variation in non-
earned income in the population of interest is quite small. The man’s non-earned
income is similarly significant in his reservation wage equation. The data include
three measures of assets: roams per person, which is a metric sometimes used
by the World Bank; lives in shantytawn, an indicator of poverty that should be
negatively correlated with unobserved assets; and lives rent-free (that is, with a
friend or family member), which should also be negatively correlated with
financial resources. 01 these, both roams per person and lives rent-free are
practically and statistically significant in the man’s reservation wage equation. An
increase of one standard deviation centered around the mean in the number of
rooms per person is associated with a 27% decrease in the predicted probability
that a man applies (from 0.115 to 0.091); living rent-free, which describes about
7% of the population, is associated with a 58% increase in the same (from 0.093
to 0.147). Both estimates suggest that assets have a strong upward effect on the
man’s reservation wage. Neither is significant in the woman’s estimated
reservation wage equation.

Variables related to home production are in theory likely to be more
important factors for women than for men, since women often have a
comparative advantage in those activities. The presence of children or elderly
individuals in the household is identified in the data. Having more dependent

children is often found to increase the value of an individual’s time spent in home

22

production because children require significant care from adults, while the elderly
could affect home production positively or negatively, depending on the amount
of help they provide or care they require. For women, the estimated coefficient an
elderly, a dummy variable indicating the presence of an adult over the age of 65,
is significant and positive, suggesting that the elderly require more care than they
provide. Having an elderly family member decreases the predicted probability
that a woman applies to Jefas (again, at values of the other explanatory variables
typical for an applicant), by 28%, from 0.192 to 0.138. It does not change the
prediction for the man.

Contrary to what is commonly seen in the literature, I estimate that having
more school-age children decreases, rather than increases, the woman’s
reservation wage. For a household of four with two children, adding an additional
child age 6-10 (with corresponding increase in household size to 5) raises the
predicted probability that a woman applies for Jefas employment from 0.177 to
0.208. Similarly, an additional child age 11-15 raises the probability to 0.202. This
is likely a function of the specific population being considered.” Families
debating enrollment in Jefas are by and large extremely poor, with many unable
to meet even the most basic household financial needs. To the degree that
additional children increase the overall cost of living for a given family unit, they
may increase the need for both adult members of the household to seek

temporary employment for the family to meet its subsistence needs. Additionally,

 

'4 This is not the first study to find such an effect. Kamitewoko and Jin (2004) see a
similar effect among poor women in the Congo, which they attribute to childcare
provision by older children.

23

Lee and Cho (2005) note that high-quality publicly funded education is an
important area of public policy in Argentina. Free public schools are widely
available and mandatory for children between the ages of 6 and 14, which may
facilitate the labor force participation of mothers who have children in the
appropriate age range. While imprecisely estimated, the coefficients on child age
groups below and above the typical schooling ages reinforce this interpretation,
since they are actually estimated to decrease the probability of a woman’s
participation by an insignificant amount. The estimates of children’s effect on the
man’s reservation wage are not significant for any age groups. Finally, formal
marriage is correlated with an increase in both men’s and women’s reservation

wages, which is consistent with previous studies.

1.7 Analysis

Taken together, these estimates begin to explain the high take-up of Jefas
among partnered women with no prior labor market experience. The work
requirement apparently served as a deterrent to the enrollment of primary
workers, even in low-income households, who had better market opportunities
than Jefas employment. At the same time, it did not serve as much of a deterrent
to other family members in those same households, for whom market
opportunities were relatively poor. Thus, in low-income households that saw a fall
in the real wage of the primary eamer—as many did during the 2001 crisis—and

households whose primary worker was unemployed but had a reasonable

24

expectation of market success, a secondary worker could have applied for Jefas
employment in an attempt to maintain household consumption. Evidence in this
paper suggests that this was especially true in households with no assets
available to help weather the crisis. High take-up by women can be explained by
the observation that women are typically secondary workers in Argentine
households, as they are in much of the developing world.

This behavioral response appears to be consistent with the added-worker
hypothesis, but seems surprising in view of that literature, which seldom finds
evidence of an effect. The added-worker effect is often defined as the increase in
the labor supply of one household member as the consequence of the
unemployment of another member, and is usually viewed as an insurance
mechanism. Available evidence from the added-worker literature suggests that
the effect is small; presumably the fact that the primary worker becomes
unemployed sends a signal of poor job prospects to the secondary worker, who
becomes discouraged from seeking employment. In fact, the literature has found
virtually no evidence of an added-worker effect in the developing world. Bardhan
(1984) finds that the male unemployment rate in India has a strong negative
effect on female labor supply, and concludes that a job search discouragement
effect outweighs the income effect related to the unemployment of men in the
household. Serneels (2004) investigates the added-worker effect in middle-class
Ethiopian households using both actual labor supply and a measure of desired

labor supply, and also finds that there is no added-worker effect. He concludes

25

that households have other ways to cope with unemployment, including using
savings and selling assets.

A possible reconciliation with this existing literature is that, as a public
program not concerned with the skill of its employees, Jefas provided a
guaranteed outlet for otherwise-unemployable family members’ desired labor
supply responses during the Argentine crisis, and thus the desired added-worker
effect was made visible. The absence of skill or job-experience related screening
would certainly have been attractive to inexperienced workers who could not
otherwise have found employment at the effective minimum wage. This suggests
that an added-worker type response may be visible during arises when
employment opportunities are available to absorb the desired labor supply of
inexperienced secondary workers. It appears that this response can occur as a
result of declining income, even in households that have not experienced a job
loss. Further work in this area is needed to lead to a fuller understanding of the

mechanisms at play.

1.8 Conclusion

As workfare programs become an important component of poverty
alleviation schemes in the developing world, policymakers designing them will
wish to know more about the incentives they create for potential labor market
participants. This paper contributes to that understanding by explaining why

secondary workers in low-income households were drawn to workfare

26

employment following Argentina’s 2001 economic collapse. I find that Jefas
workfare discouraged participation by active labor market participants but was
appealing to other family members who wished to increase their labor supply but
had few market opportunities.

Specifically, I estimate that the correlates of household members’
expected market wages were particularly important predictors of the workfare
application decision. Greater income-earning potential led to lower predicted
application probabilities for an individual, while a partner’s potential income
contributed to a lesser extent. Individuals in households with few assets are also
shown to have been more likely to apply for workfare employment. I estimate that
having more school-age children increased a woman’s probability of applying to
the program, which may be due to the provision of free, high-quality public
schooling by the government. The implications of this finding for child welfare and
women’s labor force participation may warrant further inquiry.

This research demonstrates that unskilled, low-wage, public employment
is attractive to secondary workers with few other labor market opportunities. Such
employment may be useful for households that wish to smooth income during
economic downturns in developing countries, even in cases when the primary
earner does not suffer unemployment. A possible policy implication is that
restricting workfare participation to the unemployed may be misguided if the goal
is to transfer income to needy households, as long as a low-wage work

component is a strict requirement of continued program participation. I explore

27

this question further in the second chapter of my dissertation. An important
caveat is that this prescription is unlikely to apply in OECD countries, where
households have other forms of insurance against unemployment and

underemployment.

28

Figure 1.1 Estimated effect of increasing man’s income on the probabilities of
application

 

h!—

.15

.1
1

Participation Probability

 

 

I l l I I

2 3.
Man's 2001 Monthly Income.

 

—l— Pr (Man Participates). —o— Pr (Woman Participates)

 

 

 

 

29

Figure 1.2 Estimated effect of increasing woman’s income on probabilities of

application

 

.15

.1
I

Participation Probability

 

 

 

I T l

1. 2 3. Jr. 5
Woman's 2001 Monthly Income.

 

l—O— Pr (woman participates). ——I— Pr (man participatesfl

 

 

30

 

Table 1.1 Percent of All Beneficiaries Complying with
Refinements

Child under 18 95.0%
Complies with work requirement 89.3%
Unemployed 62.1%
Head of household 50.3%
No household member employed 43.0%

 

Source: ’Encuesta Permanente de Hogares data, collected
by INDEC, except compliance with work requirement,
which was estimated In MTESS (2002). "Head of
Household" status is self-identified In the survey.
"Unemployed" means the beneficiary did not report being
out of the labor force in the previous survey.

31

Table 1.2 Data Deflnltlons
Variable
Income & Market Wage

Description

Earns income
Earned income amt

Educaflon
Domestic worker

Self-employed

2001 weekly work hours

Household characteristics

Age
Size
Children
Elderly
Married
Eligible

Assets

Has non-labor income
Non-labor Income amt

Shantytown
Rent-free

Rooms per person

Region Dummies

1 If individual earns labor income, 0 otherwise
(men and women)

Amount of earned income per month, measured as
pesos/ 100 (men and women)

Years of formal education (men and women)

1 if individual is a domestic worker, 0 otherwise
(women only)

1 if individual reports income from self-
employment, 0 otherwise (men and women)
Number of hours worked for pay during survey
reference week

Individual's age In years (men and women)
Number of individuals living in the household
Number of children in the household

1 if person age >65 In h'hld, 0 otherwise
Couple is formally married

1 If couple has children or cares for a disabled
person, 0 otherwise

1 if individual has non-labor income, 0 otherwise
(men and women)

Amount of non-labor income per month,
pesos/100 (men and women)

1 if h'hld lives in a Shantytown, 0 otherwise

1 If h'hld lives with friend or relative rent-free, 0
otherwise

Number of rooms divided by number of persons in
h'hld

Four variables indicating geographic regions of the
country, Northwest, Northeast, Cuyo, Pampeana.
The dummy takes a value of 1 if the household is
located in the corresponding region.

All data are from the Encuesta Permanente de Hogares (EPH), INDEC Argentina

32

Table 1.3 Probit Estimation of the Probability of
Participation, Conditional on Applyinlto Jefas

 

Coeff. t-stat
Age 18-24 0.772 3.33
Age 25-29 0.522 2.64
Age 30-39 0.464 2.93
Age 40-49 0.494 3.37
Male -0.47 -3.1
Head 0.225 1.59
Single -0.202 -1.14
Married 0.043 0.41
Shantytown -0.32 -0.15
Apartment -0.263 -1.37
Rooms -0.074 -1.67
Bathroom 0.021 0.14
Renter -0.637 -3.06
Rent-free -0. 177 -1.3
Masonry -0.054 -0.36
Childratio 0.196 0.73
Elderly -0.088 -0.57
Size 0.052 2.11
Employed 2001 0.066 0.47
Unemployed 2001 0.566 3.57
Public Employee 0.667 3.11
Teacher -0.499 -0.91
Construction worker 0.482 2.38
Domestic worker 0.137 0.67
Street vendor 0.633 2.08
Migrant -0.533 -2.24
Illiterate 0.144 0.56
Observations 903
Pseudo R2 0.1026

 

Note: dependent variable =1 If individual participated in
Jefas in October 2002. Panel data from Oct. 2001 and
Oct. 2002 EPH.

33

Table 1.4 Factors Influencing Date of Enrollment for Those Enrolled
Dependent Variable: Month of Application,Ja nuary - October (1-10)

 

 

 

 

Men Women Combined
Man's 2001 earned income 0.0101 -0.0139 -0.0004
[0.1185] [0.0524] [0.0458]
Woman's 2001 earned income -0.3445 -0.2784 -0.2686
[0.1514]** [0.1962] [0.1097]**
Partner unemployed in 2002 0.5098 -0.5072 -0.3188
[0.6120] [0.2511]** [0.2323]
Man's education -0.0591 -0.0365 -0.043
[0.0531] [0.0429] [0.0333]
Woman's education —0.0159 0.0203 0.0272
[0.0553] [0.0423] [0.0331]
Woman is domestic worker -0.2415 -3.8471 -0.1005
[0.4769] [0.7659]*** [0.4080]
Man is self-employed 0.0401 -0.0827 -0.0131
[0.3612] [0.2226] [0.1801]
Woman is self-employed 0.2718 0.517 0.3468
[0.6942] [0.3584] [0.2925]
Man's 2001 non-earned income -0.1512 0.1077 0.0359
[0.2146] [0.2505] [0.1506]
Woman's 2001 non-earned income 0.0905 0.2486 -0.0538
[0.2011] [0.2717] [0.1430]
Rooms per person 0.0621 0.0404 0.0308
[0.6069] [0.0269] [0.0270]
Lives in a Shantytown 1.1048 -0.1285 0.4434
[0.5410]** [0.8215] [0.5080]
Lives rent free 0.2567 -0.0247 0.0545
[0.4318] [0.3300] [0.2415]
Household size 0.1524 0.1051 0.1416
[0.1663] [0.1034] [0.0773]*
Number of children -0.466 -0.0942 -0.209
[0.2089]** [0.1311] [0.1041]**
Elderly in household 0.0169 -0.1578 -0.0604
[0.6280] [0.3997] [0.3356]
Married -0.2627 0.1692 0.0025
[0.3428] [0.2465] [0.1934]
Man's age -0.0331 0.0212 0.0117
[0.0297] [0.0177] [0.0148]
Woman's age -0.0015 -0.0218 -0.0195
[0.0350] [0.0205] [0.0169]
Eligible 1.066 -1.5019 0.3136
[0.8831] [0.6094]** [0.5912]
Constant 9.5056 5.3225 6.7041
[1.4688]*** [1.2936]*** L1.1972]***
Observations 130 281 411
R-squared 0.17 0.07 0.05

 

Robust standard errors in brackets

* significant at 10%; ** significant at 5%; *** significant at 1%

Note: Negative coefficients indicate earlier enrollment date.

34

.NOON .uUO Dcm HOON .uUO .UmOZH ~memOOI OD OHCOCMELOQ mumwzucm ”mu—MD

 

 
 

 

 

time- 85 BS 85 Ed 823:9 m2:
Iimnm- 85 No.0 Ed mod $8355 5 32..
1.1.8.2 -.o mg 35 SR mom menses
13.3.3 Rd one. 35 mad... 8.... MS:
1.13.3 85 2.0 85 35 2.8a
1:13 SS 35 No.0 35 8:5:
11.8.2 85 35 85 $5 828 .8 382
time- 85 mm.~ BS oo.m 8 A E 5.226 .6 been:
.133me mod N2 83 om.m 2a 20528:
$5. 86 2.6 86 2.0 :83 86356.28 2 amass
13.8.? 85 mg mod and 38$ 86352.8 a cm:
$5. 85 mod 35 mod .98; 6385.8 a case;
1.1%.: Ed who 9.6 SK 283 8:838 menses
2.3.2 85 and 25 «.3 E83 838:8 9%:
l...8.~ meom 8&2 BR R49 3558 E Em 058:. 282-8: menses
3.69m 85 one 85 Be 6685 3.22.8: mm; 5525
like 3.9 333 m3: 863 8238a S “Em 3.85 282-5: 9%:
1.3: 85 Rd 85 85 .6685 33.8: m2 52
1.84.4 Rd 2.9:. $4 33 zoom .8635» .5 etc; .o ago; 583 Manse;
1.185 mg 83¢ 85 3.9 32K 69635 E 0.3; Lo 2:2 >28; 9%:
1.63 8.3 8.84 on: $42 3.8 .8635 5 EB 6585 8&8 Memees
1.1.83 85 mg No.0 Rd 9:85 9:8 Sec;
3.3.2 8.2 3.68 Re 86% zoom 8636.... .5 Em 6585 8&8 MS:
182. SS 85 SS NR... 9:85 2:8 cm:
"D Em .um com: Eu .3 com: ozmtafifimcu
Scout... 5. 8.0530: 3.2330:
8:9.an “28:38:02 Emu=qa< =<

 

magnum Emo=aa< >9 moEmtm> ucmucaamncu Co mozmzmum tmEEnm man ofinh

35

.ucmozaam cm m. Eon; Co 9.0 “mag um .3303 new new; mm 29.8 2.25m... new 29: 5.2. mEocmmsoc 89:2: .302
.38 .30 can Sam .80 .UmozH .3800: an 35553... 3335 ”Son.

 

 

 

*8; mod Ed mod Ed 8:58. 82..

1.de 8d edd Nod mod $9385 5 82..

*8; ad 3.8 dud 8.8 88 58an

m: Ed 8d... odd 8.? m8 .982

3.8.5 8d 8d Nod Ed E85

dwd- Nod 8d mod mmd 8:82

mod- mod mmd Nod vmd 8&8 :8 2:8,:

3.? mod Ed 2d 8d 8 A 5 8.5.8 ed .852

8.? 3d 8d 3d ~m.m 8...". 5288..

8d- 8d 8d Nod 8d :83 8638....8 m. 8562.

3... mod Ed 8d 8d :85 8.63858 5 8:

8d- 3d 8.5 Rd 85 8.58:8 98an

Rd mad 8.5 Rd 8.5 8:838 982

mm; dvdm 2.3; 8.8 dodmm $258.. .5 “Em 2:85 .8515: 5852...

1.83 8d .dd Nod 8d 8585 .8215: 8.. 8an

Rd 8.8 8de 8.8 8.28 3258.. .5 #8 9:85 .8556: 5:2...

and 8d mod Nod 8d 2:85 .5878: 8: 8...

em.” 84 88 8d mddm .88 .8258“. 5 0.3; Lo 2:9. 58; 985.25

13d- mo; dev NZ Rdv deN 6258.. .5 3:25 .0 .28: 20.8; 58:

5.3.5: 3.: .8.va 8.2 vde :8." .8258 .5 8:85 858 58an

182 Nod m~d mod 8d 8585 E8 8an

3.83- 3.3 8.3m 8.: R58 .88 .8258 .5 q.585 .858 582

El- Nod and mod Ed 8585 2:8 82

.1.» km .5 com: Em .um :82 Sumtafiflmcu
ltfiulsmmmfl 58:92 .58.. 58.52 8.3,.

8:0th 5:5 29.60 5.2. 2300

 

Emo=aa< Co .6956 >2 maBmth “concede-US .6 mozmzmum EmEEam 0.." 0:3...

36

Table 1.7 Men's and Women's Reservation Wage Eguations

 

 

 

 

Men Women
Man's 2001 earned Income 0.2265 0.1295
[0.0311]*** [0.0206]***
Woman's 2001 earned Income 0.0317 0.2476
[0.0394] [0.0513]***
Man's 2001 weekly hours 0.0046 0.0029
[0.0024]* [0.0020]
Woman's 2001 weekly hours 0.0037 0.0056
[0.0035] [0.0032]*
Man's education 0.0474 0.0864
[0.0182]*** [0.0160]***
Woman's education 0.0439 0.0205
[0.0185]** [0.0160]
Woman Is domestic worker (2001) 0.0043 0.6296
[0.1878] [0.2031]***
Man Is self-employed (2001) -0.4794 -0.4231
[0.1108]*** [0.0964]***
Woman is self-employed (2001) 0.0951 -0.1893
[0.1998] [0.1735]
Man's 2001 non-earned Income 0.2873 0.2359
[0.0652]*** [0.0579]***
Women's 2001 non-eamed Income -0.0521 0.2193
[0.0443] [0.1036]**
Rooms per person 0.3848 0.0842
[0.2015]* [0.0748]
Lives In a Shantytown -0.3019 0.2957
[0.2481] [0.2467]
Lives rent free -0.3718 -0.0496
[0.1559]** [0.1449]
Household size 0.0292 -0.0262
[0.0524] [0.0427]
Number of children age 0-5 -0.057 0.0733
[0.0865] [0.0748]
Number of children age 6-10 -0.1 178 -0.1447
[0.0802] [0.0682]**
Number of children age 11-15 -0.0887 -0.1 132
[0.0822] [0.0685]*
Number of children age 16-18 0.1708 0.0513
[0.1213] [0.0965]
Elderly In household 0.2228 0.3154
[0.1771] [0.1597]**
Married 0.1567 0.1733
[0.1208] [0.10461“
Man's age 0.0109 0.0135
[0.0092] [0.0079]*
Woman's age 0.0088 0.0242
[0.0097] [0.0085]***
Constant 0.3804 0.1495
[0.5237] [0.4820]
Observations 4298 4298

 

* significant at 10%; ** significant at 5%; *** significant at 1%.

Asymptotic (Inverse, negative Hessian) standard errors are In brackets.

Notes: Data from Oct. 2001 and Oct. 2002 EPH; both equations Include region and
eligibility dummies. Rho vas estimated to be 0.366 with standard error 0.153.

37

APPENDIX TO CHAPTER ONE

Letting lower-case w’s represent log wages throughout, the two reservation wage

equations for each household i are of the form
III *I III
WM = fiMX + a M

w; = fipx + .;.
and serve as latent variables in the model. The Jefas wage offer is represented

by wf. The errors 5M and a): follow a joint normal distribution with mean zero,

fully characterized by

(3}

To allow for identification of the covariance parameter p I set 0M = 0):. (For a

2
GM pOMZOFD,
poMa'F 0F

 

 

 

more complete explanation of the requirements for identification, see Train

(2003) or Bunch (1990).) Thus,

8;! 0) 02 p02
5;. 0 ’ p02 02
Normalizing the reservation wage equations and defining pM - (3;, /0),

6,, = (312/0). w,- = (wf/o), and 2,: (sf/o) the joint density can then be written

(1_p2)y2 2

¢2 = Texp —m(aM + a}, -2psMsF) .

and the probability associated with condition (2) in the paper is

38

Pr(yo=1) =pr(W15WM,WJsWF)
=pr(EM2WJ-l3M’X, EFZWJ-AF’X)
= pr(sM sfiMX-wJ , 8F sfiFX—wj)

5M X'WJ IBEX-WI
= f f ¢2(zl .zz)dzldzz

= dbl/84X — w], p'FX — wJ l p)

where <I>2 is the bivariate standard normal cumulative distribution function. The
probability associated with condition (3) in the paper is

prlyi =1) =pr(WMsWJ,WMsWF)

= Pf (8MS WJ - fiM’X, 5M' EFS .BF’X- fiM’X)

8
5M TEF

with bivariate covariance matrix
1 0 1 p 1 l l l— p
l —l p l 0 —1 1 - p 2 — 2p

. 'X— ' x _.
wJ—fiMXfL—fl—l 1 p
1/2-2p 2-2p

The probability associated with condition (4) is likewise

:1 4er
PFX-PMX

 

 

thus

 

pr(y1=1) =(D2

 

PTO’2=1) =PTIWF5WJ,WFSWM)

= pr (EFS WJ - fiF'X, €F- EMS .BM'X- flF’X)

39

YJ-flbx
PMX‘PFX

 

 

8
5F ' EM

with associated bivariate covariance matrix
0 l l p O —l l l-p
—1 1 p 1 1 1 l—p 2-2p

fiirX-fiivX, 1-.. l

- ' X,————— _____
w’ W 1l2-2p 1/2-2,o
The resulting likelihood function

L - “Pr(m - 1) HPIM = 1) ”Pill? =1)

yo-l y1-1 y2-1

thus

13702:” =‘I’2

 

is a trinomial probit with correlated errors. The maximum likelihood estimator is
consistent, asymptotically efficient, and asymptotically normally distributed.

4O

Table 1.8 Sensitivity Analysis - Main Regression with Only Enrolled Applicants

and Non-applicants

 

 

 

Men Women
Man's 2001 earned income 0.2462 0.1281
[0.0413]*** [0.0245]***
Woman's 2001 earned income 0.1287 0.1895
[0.0675]* [0.0621]***
Man's 2001 weekly hours 0.0037 0.0038
[0.0029] [0.0024]
Woman's 2001 weekly hours 0.0029 0.0103
[0.0045]** [0.0041]
Man's education 0.058 0.0945
[0.0222]*** [0.0190]***
Woman's education 0.0335 0.0121
[0.0226] [0.0191]
Woman is domestic worker (2001) 0.0715 1.8992
[0.2420] [0.4592]***
Man is self-employed (2001) -0.5702 -0.5066
[0.1338]*** [0.1124]***
Woman is self-employed (2001) -0.1299 -0.444
[0.2355] [0.1966]**
Man's 2001 non-eamed Income 0.2587 0.2949
[0.0813]*** [0.0878]***
Woman's 2001 non-earned income -0.0212 0.3236
[0.0745] [0.1649]**
Rooms per person 0.221 0.059
[0.2498] [0.0792]
Lives in a Shantytown -0.2711 0.4406
[0.2910] [0.2971]
Lives rent free -0.3209 0.0399
[0.1920]* [0.1742]
Household size 0.0072 0.0177
[0.0611] [0.0520]
Number of children age 0-5 -0.0436 0.0242
[0.1033] [0.0893]
Number of children age 6-10 -0.0832 -0.2008
[0.0973] [0.0815]**
Number of children age 11-15 -0.1158 -0.1646
[0.0970] [0.0810]**
Number of children age 16-18 0.1682 -0.1289
[0.1437] [0.1124]
Elderly in household 0.325 0.1655
[0.2272] [0.1934]
Married 0.1702 0.2371
[0.1453] [0.1216]*
Man's age 0.0174 0.0103
[0.0117] [0.0093]
Woman's age 0.0044 0.0317
[0.0125] [0.0102]***
Constant 1.1562 1.1539
[0.6897]* [0.7295]
Observations 4071 4071

 

* significant at 10%; ** significant {5%; *** significant at 1%.
Asymptotic (inverse, negative Hessian) standard errors are in brackets.

Notes: Data from Oct. 2001 and Oct. 2002 EPH; both equations include region
and eligibility dummies. Rho was estimated to be 0.366 with standard error

0.153.

Table 1.9 Men's and Women's Reservation Wage Equations Using 2002 Data

 

 

 

Men Women
Man's 2002 earned income 0.24 0.1459
[0.0304]*** [0.0198]***
Woman's 2002 earned income -0.0192 0.3435
[0.0271] [0.0622]***
Man's 2002 weekly work hours 0.006 0.0002
[0.0022]*** [0.0018]
Woman's 2002 weekly work hours 0.0023 0.0057
[0.0030] [0.0032]*
Man's education 0.0288 0.0505
[0.0153]* [0.0135]***
Woman's education 0.0618 0.0298
[0.0157]*** [0.0139]**
Woman is domestic worker -0.2235 -0.3936
[0.1479] [0.1516]***
Man is self-employed -0.4751 -0.1566
[0.0986]*** [0.0873]*
Woman is self-employed -0.1151 -0.5569
[0.1709] [0.1644]***
Man's 2002 non-eamed income 0.2902 0.181
[0.0576]*** [0.0434]***
Woman's 2002 non-earned income -0.0015 0.8118
[0.0571] [0.1651]***
Rooms per person 0.4221 0.3583
[0.1622]*** [0.1430]M
Lives In a Shantytown -0.2929 -0.1781
[0.2173] [0.2081]
Lives rent free -0.2124 -0.2112
[0.1330] [0.1149]*
Household size 0.0453 -0.0075
[0.0489] [0.0418]
Number of children age 0-5 -0.0442 0.0304
[0.0787] [0.0680]
Number of children age 6-10 -0.1123 -0.0904
[0.0724] [0.0618]
Number of children age 11-15 -0.0681 -0.0425
[0.0747] [0.0627]
Number of children age 16-18 0.0446 -0.0213
[0.1053] [0.0844]
Elderly in household -0.2041 0.4097
[0.1499] [0.1474]***
Married 0.1116 0.2181
[0.1043] [0.0912]**
Man’s age 0.0039 0.0062
[0.0075] [0.0069]
Woman's age 0.0185 0.0141
[0.0078]** [0.0072]:
Constant 0.1057 0.3819
[0.4125] [0.3862]
Observations 9715 9715

 

* significant at 10%; ** significant at 5%; *** significant at 1%.

Asymptotic (inverse, negative Hessian) standard errors are in brackets.

Notes: Data from Oct. 2002 EPH; both equations include region and eligibility dummies.

Rho was estimated to be 0.368, with a standard error of 0.138

42

REFERENCES

Akerlof, G. (1978), “T he Economics of 'Tagging’ as Applied to the Optimal
Income Tax, Welfare Programs, and Manpower Planning.” American Economic
Review, Vol. 68, No. 1, pp. 8-19.

Bardhan, P. (1984), Lari Lab ran Rur iP v :E in D v I
Economics. New York, Columbia University Press.

Besley, T. and S. Coate (1992), “Workfare versus Welfare: Incentive Arguments
for Work Requirements in Poverty-Alleviation Programs.” American Economic
Review, Vol. 82, No. 1, pp. 249-261.

Bunch, OS. (1991), “Estimability in the Multinomial Probit Model.” Transportation
Research 8, Vol. 258, No. 1, pp. 1-12.

Cebula, R. and C. Coombs (2008), “Recent Evidence on Factors Influencing the
Female Labor Force Participation Rate.” Journal of Labor Research, Vol. 29, No.

Chiappori, P.(1992), “Collective Labor Supply and Welfare.” Journal of Political
Economy, Vol. 100, No. 3, pp 437-67.

Canning, J. and M. Kevane (2002), "Community Based Targeting Mechanisms
for Social Safety Nets: A Critical Review." Warfd Development, Vol. 30, No. 3, pp.
375-395.

Cuff, K. (2000), “Optimality of Workfare with Heterogeneous Preferences.” The
Canadian Journal of Economics, Vol. 33, No. 1, pp. 149-174.

Dooley, M. (1982), “Labor Supply and Fertility of Married Women: An Analysis
with Grouped and Individual Data from the 1970 US. Census.” Journal of Human
Resources, Vol. 17, No.4, pp. 499-532.

Galasso, E. and M. Ravallion (2004), “Social Protection in a Crisis: Argentina’s
Plan Jefes yJefas.” World Bank Policy Research Working Paper 3165. Available
at SSRN: http://ssrn.com/abstract=636584.

Graham, C. (1994), Safem Nets, Pglitics and the Poor: Transitions to Market
Emnomies. Washington, DC: The Brookings Institution Press.

Gronau, R. (1973), “The Effect of Children on the Housewife’s Value of Time."

Journal of Political Economy, Vol. 81, No. 2, Part 2: New Economic Approaches
to Fertility, pp. 8168-8199.

43

Heckman, J. (1974), “Shadow Prices, Market Wages, and Labor Supply.”
Econometrica, Vol.42, No. 4, pp. 679-694.

Kamitewoko, E. and X. Jin (2005), “Labour Force Participation of Married Women
in China and Congo.” Journal of Zhejiang University SCIENCE, Vol. GA, No. 4,
pp. 350-354.

Lee, K. W. and K. Cho (2005), “Female Labour Force Participation during
Economic Crises in Argentina and the Republic of Korea.” International Labour
Review, Vol. 144, No. 4, pp. 423-449.

MTESS (2004). “Insercién laboral de Ios beneficiaries del Programa Jefes de
Hagar.” Ministry of Labour, Gobierno de la Republica Argentina, Buenos Aires.

MTESS (2007). “lnsercion Laboral de los beneficiaries del PJH en el empleo
registrado”. Direccién general de estudios y formulaeidn de politicas de empleo,
Subsecretaria de Empleo, Buenos Aires.

Newman, J., S. Jorgensen and M. Pradhan (1991), “How Did Workers Benefit
from Bolivia’s Emergency Social Fund?” World Bank Economic Review, Vol. 5,
No. 2, pp. 367-393.

Nichols, A. L. and R. J. Zeckhauser (1982), “Targeting Transfers through
Restrictions on Recipients.” American Economic Review, Vol. 72, No. 3, pp. 372-
377. ’

Ravallion, M. and G. Datt (1995), "Is Targeting through a Work Requirement
Efficient?" P II n in n th P r: Th n Evi n . Ed. D. van de
Walle and K. Nead, Baltimore: Johns Hopkins Press.

Ravallion, M.- (2006), “Targeted Transfers in Poor Countries: Revisiting the
Tradeoffs and Policy Options.” Understanding Pavefly. Ed. A. Banerjee, R.
Benabou, and D. Mookherjee. Oxford: Oxford Scholarship Online Monographs,
pp. 203-231.

Sen, A. K. (1979), “Utilitarianism and Welfarism.” Journal of Philosophy, Vol. 76,
No. 9, PP. 463-89.

Serneels, P. (2004), “The Added Worker Effect and lntrahousehold Aspects of
Unemployment.” Development and Camp Systems 0409014, EconWPA.

Subbarao, K., A. Bonnerjee, J. Braithwaite, S. Carvalho, K. Ezemenari, C.

Graham, and A. Thompson (1997), f N Pr r m n P v
Washington: World Bank.

44

World Bank (2002), Report #PID10834, May 21 , 2002.

World Bank (2005), “Argentina: Seeking Sustained Economic Growth with Social
Equity.” Report No. 32553

World Bank (2006), “Project Appraisal Document on a Proposed Loan in the

Amount of US$350 Million to the Argentine Republic for a Heads of Household
Transition Project.” World Bank Report No. 32463-AR, Feb. 21, 2006.

45

CHAPTER TWO

Design Changes and the Resulting Profiles of
Heads of Household Beneficiaries

2.1 Introduction

In the first chapter of the dissertation, I developed a model to describe
participation in Argentina’s Heads of Household public employment program
based on individuals’ reservation wages, which I estimated as functions of
household and individual characteristics. I took the program’s design as given
and made certain assumptions about how program requirements were enforced
or not enforced. In this chapter, I estimate how changing the design and
implementation of Jefas would have affected the gender composition and “poverty
level of beneficiaries. In particular, I estimate the effect of two hypothetical
changes: removal of the “one person per household” rule, and perfect
enforcement of the requirement that applicants be primary earners who lost their
jobs during the crisis.

While Jefas was originally intended as a way to combat poverty during the
economic crisis, its stated goal was to assist primary earners who had lost their
jobs as a result of the downturn. Much of the subsequent criticism of the program

concerned the failure to achieve this stated goal, in light of the gender and

46

employment history of many participants. To the degree that lifting the somewhat
unusual “one person per household” requirement would have changed the
gender composition of applicants, it might have alleviated some of this political
distress and re-focused the public’s attention on the goal of poverty alleviation.
Moreover, the World Bank has stated that it views the difficulty of enforcing the
originally specified eligibility criteria—especially the dual requirements that
participants be ‘heads of household’ and ‘unemployed’—as a serious
shortcoming of program implementation (World Bank, 2006). At the same time,
the Bank acknowledges that Jefas compares well to other social assistance
programs in Latin America in terms of the share of benefits going to the poor, and
coverage of the poor. The Bank takes this as evidence that “the difficulties in
enforcing these two eligibility criteria did not adversely affect the performance of
the Program in reaching or targeting to the poor.” Yet, this seems naive at best.
Had the government taken steps to remedy these issues, the program might
have done a much better—or much worse—job of reaching the poor, and there is
no reason to expect other socioeconomic attributes of beneficiaries to have
remained unchanged. While imperfect information in the hands of program
administrators seriously limited the ability to target benefits to heads of household
or the unemployed, it is worth assessing whether that would have even been a
desirable goal. Public employment programs like Jefas are a growing component

of social assistance in developing countries (Subbarao et al., 1997), and

47

information on the optimal design should be helpful to policymakers going
forward.

In this study, I find that by allowing any individual to access the program,
rather than only one member of each household, the number of male applicants
would have increased substantially, by around 46%, while the number of female
applicants would have increased by only around 16% assuming the work
requirement were still enforced. However, the fraction of applicants who were
male would only have increased from 0.34 to 0.39 and the poverty level of
applicants would not have dramatically changed. This finding indicates that the
“one person per household” requirement was not responsible for the high
proportion of female applicants as compared with other programs, and so an
explanation of that phenomenon must be found elsewhere. On the other hand,
restricting application to primary earners who lost their jobs, while continuing to
enforce the work requirement, would have led to a dramatically lower-income, but
much smaller, group of applicants, 70% of whom would have been men. The
small size of the group of applicants would have rendered Jefas largely
ineffective as a social assistance program. Enforcement of the work requirement
is key to the low income level of this group—overall, primary earners who lost
their jobs are wealthier than actual applicants to the Heads of Household
program, but still primarily male. These results suggest that, if the goal of Jefas

was to reduce poverty, the government should not worry about failure to target

48

benefits to primary earners who lost their jobs, because doing so would not have
improved the program ’3 performance.

The remainder of this chapter is structured as follows. First, I briefly
describe the current thinking on evaluating workfare programs. Then, I describe
how I assess the well being of households, discuss the general methodology i
use, and outline the program changes that I study. Finally, I discuss the empirical

results, and make recommendations.

2.2 Background on Evaluating Antipoverty Programs

How should a workfare program be evaluated? Two often contradictory
considerations—coverage of the poor and exclusion of the non-poor—are the most
commonly used measures of the performance of antipoverty programs.‘ Perfect
coverage of the poor implies that no member of the poor population should be
excluded, and could be achieved by transferring income to every individual in
society without regard to income. However, the cost of such a program would
often be quite large. Perfect exclusion of the non-poor implies rigid screening to
ensure that all participants are eligible and could, in theory, be achieved by not
transferring income to anyone. Aside from being costly or difficult to implement,

the effectiveness of such a program is sometimes questionable: Many poor

 

' Ravallion and Datt (1995) give a much more comprehensive summary of current
thinking on program evaluation. Many of their recommendations are incorporated in this
study.

49

individuals may be excluded or deterred from applying, resulting in a too-small
group of beneficiaries to result in much social benefit.

Apart from these most commonly used considerations, the literature also
identifies a number of secondary measures of targeting performance. For
instance, many evaluations focus on cost-effectiveness; that is, the cost to the
government for a given impact on poverty. These sorts of evaluations often take
into account the dollar amount of the transfer, as well as administrative costs, to
be weighed against the reduction in the poverty rate or the income gains of
participants. In these analyses, it is important to take into account the foregone
income of participants, which is not always trivial. The appropriate measure of
“income gains” should then be the income net of participation costs. For the Jefas
program, Galasso and Ravallion (2004) have estimated the foregone incomes of
participants, and note that they appear positive but quite low. This is consistent
with the observation that many participants were drawn from labor market
inactivity. Another issue with workfare is that often its proponents implicitly do not
value leisure - that is, they define poverty as too little income as opposed to too
little utility. Sen (1979) points out that “a previously idle person who takes up
workfare employment may be only slightly better off in terms of utility (allowing for
the loss of leisure), but much better off in terms of income.” For this reason,
some welfarist analyses of workfare, for example Cuff (2000), specify the
poverty-alleviation objective in utility terms. As other authors have noted, this sort

of perspective appears to carry little weight with policymakers. and if anything,

50

society often appears to attach positive value to employment. For that reason, in
this analysis I primarily evaluate design changes in terms of coverage of the poor
and exclusion of the non-poor, but I will address these secondary issues as

appropriate.

2.3 General Methodology

In this chapter I will examine two hypothetical changes to the program:
removing the requirement that only one person per household is eligible to enroll,
and enforcing the requirement that applicants be primary earners who lost their
jobs during the crisis. Why these two changes? The one-person-per-household
requirement is not a feature of all public employment programs, including
Argentina’s Trabajar II, and may have been partly responsible for the large
number of female applicants. Then, the fact that there were so many applicants
who were not unemployed primary earners (primarily women) came as a surprise
to the government and was viewed negatively by policymakers. Yet, this does not
imply that it interfered with the program’s original aim of supplementing the
incomes of the poor.

In the preceding chapter, I made several assumptions about the way the
program was implemented, and l briefly outline them again here in the interest of
clarity. First, I assumed that applicants expected the work requirement to be
binding, so that their reservation wages were determined by the decrease in

household utility from working 20 hours per week. I will retain that assumption in

51

this chapter, except where I explicitly indicate othenivise. Second, while it appears
that applicants were for the most part required to have children or a disabled
dependent, l modeled “eligibility" in that sense as a component of the reservation
wage, rather than a binding requirement. Unlike the requirement that applicants
be unemployed, which was an attempt to identify the poor in the absence of
satisfactory data, I perceive the requirement that applicants have children as a
political decision about who was most deserving of aid, and so I will not estimate

the effect of changing that requirement.

2.3.1 Using Current Income to Measure Well Being

A key issue in evaluating the success of an anti-poverty program at
reaching the poor or excluding the non-poor is assessing the counterfactual well
being of affected households. According to the World Bank, when programs like
Jefes are implemented during economic crises, “typically, the main aim of
workfare is to raise the current incomes of poor families hurt by the crisis” (Jalan
and Ravallion, 1999, p.2). Because household wealth—another characteristic of
interest-is so difficult to measure given the data at hand, I follow the World Bank
in using current per-capita household income to measure the well being of
households. For non-participant households, I use observed 2002 per-capita
household income. In households with a Jefas participant, current income is
endogenous, so I impute values of per-capita income for those households as

discussed below.

52

In the terminology of Little and Rubin (2002), the values of household per-
capita income for participant households in the data set are missing for
“nonignorable” reasons—that is, the reason the data are missing is correlated
with participation in the program, and participants were likely to have experienced
larger drops in income than non-participants. Thus, estimates of per-capita
income for participant households, derived from relationships observed in other
households, are likely to be biased upwards. Fortunately, in this case the
missingness can be made “accessible” (Graham and Donaldson, 1993) by
comparing participants to applicants not yet receiving benefits. If applicants are
an acceptable control group, then the household income of participants can be
thought of as “missing at random” (MAR) in comparison with the group of
applicants.2 When data are MAR, missing data can be thought of as depending
on known values of the other variables, and accounting for the values that
“cause” the missing data will produce unbiased results in an analysis.

In this case, I impute the missing values using a parametric regression
method, in which a regression model is fitted on the group of applicants, and then
the fitted model is used to impute missing values based on the observed values
of the other variables. Variables are chosen either because they are correlated
with the missing variable (eg. 2001 per-capita household income, which is
correlated with 2002 per-capita income), the reason for missingness (e.g. losing

one’s job, which is correlated with participation), or both. As opposed to using,

 

2 In the third chapter, I more thoroughly make the case that applicants can indeed be used
as a control group for participants.

53

say, mean values, this method maintains at least the part of the original variability
of the missing data that can be predicted from observables. I use both tobit and
linear regression models to fit the data. Imputed values from the linear model are
trimmed so that they are strictly positive, by assigning all negative values a value
of zero. Using the tobit model, income is predicted as the “unconditional”
expectation E(yl x) = Pr(y> 0 l x) - E(yl x, y> 0). Because the tobit model
already restricts imputed values to a plausible range, I prefer it a priori to the
linear model. Alternately, I considered using 2001 per-capita household income
as a proxy for 2002 per-capita income for all households including non-
participants due to its simplicity. Although the direction of bias using this measure
is clear, I will argue later that it is a particularly bad proxy for those who lost their
jobs, which is a primary group of interest in this study. Although I present
descriptive results for all three measures, l primarily analyze and discuss results
in terms of observed (for non-participants) or tobit-imputed (for participants) 2002
per-capita income.

Variables used to fit the model and their coefficients are shown for the
tobit model in Table 2.1 for households headed by couples and Table 2.2 for
households headed by singles, and for the linear model in Table 2.3 for couples
and Table 2.4 for singles. Variables include previous (2001) income and
employment, household attributes, and a dummy variable indicating whether

each individual lost their job during the crisis. (The construction of this variable is

54

explained in detail in a later section of the paper.) All data come from the
Encuesta Permanente de Hogares, described in detail in the previous chapter.
On a cautionary note, a possible reason for skepticism regarding imputed
per-capita household income is that for some participants, imputed 2002 income
is higher than observed 2002 income, which includes the Jefes transfer payment.
This is the case for 33% of participant households using tobit-imputed income
and for 39% of participant households using linear-regression-imputed income. It
is possible that for some households income would have been higher in the
absence of the program because, for example, participants might have found
higher-paying jobs. However, that seems unlikely to be the case for such a large
proportion of participants, especially in light of the fact that so many came from
out of the labor force, and probably would not have been employed at all if not for
the program. Nonetheless, even if the fitted model does a poor job predicting
income for any given individual, the analysis that follows will be valid as long as
imputed income is representative of actual income on average for relevant

subsets of the group of participants.

2.3.2 Participation of Singles

The first chapter was primarily concerned with households headed by
couples, because the intent was to explain the large number of partnered women
applying for benefits. While those households continue to be the main focus of

analysis in this chapter, there may be important changes involving single-headed

55

households as well—for example, design changes may increase enrollment by
single parents at the expense of couples. In order to expand the scope of the
analysis to include single heads of household, I must estimate probabilities of
participation for that group. Singles’ application decision follows the model
proposed by Besley and Coate, where an individual applies for workfare if that
individual’s opportunity cost of time is lower than the wage offered by the
government. Following the notation outlined in the first chapter, person i’s log

reservation wage is given by:
W: 3 fifX T E:

where w,-' is the log of the reservation wage and in this case 81. follows a normal
distribution with standard deviation 0. As in the first chapter, define W] as the log

of the government’s wage offer. Dividing through by a so that p,- = (p; m) and

w = (w; /a) and rearranging, the probability of participation can be written

pr(y =l)=-<I>(wJ -fi,'-X)

where <1) is the standard normal CDF. Coefficients are estimated using data from

the Oct. 2002 Encuesta Permanente de Hogares, and are presented separately

56

for men and women in Table 2.5. Probabilities of participation are then generated

from the fitted model for each individual using Stata’s <predict> command.

2.3.3 Attributes of Applicants

For households headed by both couples and singles, attributes Awof the

group of applicants estimated under the various scenarios that I study are

computed as a weighted average of the attributes A,of each individual who has a

positive probability of applying to the program, as follows, with y,- = 1 in the event

that an individual participates:

211M)? '1)‘ At)
W 2, Pr(yi '1)

 

The number of applicants n is estimated from the N observations in the sample

as

N
n - Ema.- =1)

i-l

57

Both the weighted average of attributes and the number of participants are
computed separately for men and women because the probability of participation
is computed separately.

Unsurprisingly, using the estimated probabilities of participation and
weighting procedure to compute the attributes of applicants to the Heads of
Household program as it was implemented leads to values that are similar to
those of actual observed applicants, with the exception of per-capita household
income, which was not used to fit the regression model in chapter 1, and which is
over-estimated in all cases. It is these estimated values, given in Table 2.6, that I
will compare with the estimated values from program design changes to
determine the magnitude of the effect. I refer to this estimated group of
participants as the comparison group. To the degree that the unobserved
characteristics causing bias are similar among various groups, using estimated
participants instead of observed participants should make comparisons more

accurate.

2.4 Allowing Enrollment by All Individuals

As a way to address budget concerns in light of a 65% poverty rate,
program designers limited participation to one individual per household. In this
section I present the methodology for estimating the effects of removing that

requirement, and opening participation to any interested individual.

58

Although I no longer restrict participation to one person per household, I
will retain the assumption that the work requirement was, in fact, a requirement.
That is, individuals are weighing the benefit of enrollment against the cost of
working 20 hours per week, with reservation wages defined as in the previous
chapter. However, I will now make one additional strong assumption. In the first
chapter, the reservation wage of an individual was conditional on that individual’s
partner not participating in the program. Now, I assume that an individual’s
reservation wage is unaffected by their partner’s participation. I believe that

formally, this is equivalent to saying that within a household, the amount the

government must pay to employ both individuals, W3, is the sum of the amount it

would have to pay the man, WM, and the woman, WF. This is unlikely to be

literally true, and depends on the complementarities of home production and
leisure among members of the household, as well as the impact of earnings on
the amount of time devoted to home production. I will argue that this condition is
likely to represent a lower bound on the joint reservation wages of both
individuals within a household, and thus the estimated change in participation will
represent an upper bound on the increase in the number of applicants, for the
following reasons. Recent empirical evidence suggests that husbands and wives
are substitutes in home production and that income is at most an imperfect
substitute for home production (Leeds and von Allmen, 2004). That is, one
member of the couple taking work increases the marginal productivity of the other

at home. Through this first effect, enrollment of one family member should serve

59

to increase the reservation wage of the “second enroller.” Contributing to this
upward pressure on the second enroller’s reservation wage is the income effect,
through which the demand for leisure of the second enroller should increase due
to the increase in family income from participation of the first enroller. However,
other research suggests that spousal leisure is complementary, which drives the
second earner’s reservation wage in the opposite direction. That is, when one
spouse takes work, it makes the other spouse’s leisure time less valuable. it is
not obvious in theory which effect should dominate in these circumstances,
implying that the reservation wage of the second earner is imprecisely estimated.
Nonetheless, it seems reasonable that for poor households during an economic
crisis, the estimated increase in participation represents an upper bound,
because home production is of more importance to the household than is
complementary leisure time.

In the absence of the restriction that only one person per household can
participate, enrollment should proceed as it does for single household heads,
which is according to Besley and Coate’s (1992) basic model of workfare
participation. An individual will apply for benefits if and only if their opportunity

cost of time is lower than the wage offered by the government. With a reservation

wage estimated as W,-, the probability that an individual applies to the program is

pr (y = 1) = pr (W,-< W J), and the probability of declining to apply is pr 0» = 0) =
1 - pr (y = 1), where y = 1 in the event that an individual applies for benefits. Log

reservation wages w,- = wm, wf are estimated as

60

where the coefficients 6, = (Bi/o) were estimated from the full model in the

previous chapter and error terms are assumed to be normally distributed. The

probability that any individual applies for benefits can then be estimated as
pr(y = 1) = <p(w, -p}x)

This leads to an estimated probability of participation for the man in each
household, and for the woman in each household. For single men and women,
the estimated probabilities are unchanged from those found using the fitted
model in section 2.3.2 above.

Characteristics of the group of expected participants can then be
calculated as a weighted average. Results are presented in columns 2 and 5 of
Table 2.7 for couples and will be discussed in section 2.6. There is no change for

single men and women.

2.5 Limiting Enrollment to Those who Lost Their Jobs

As Jefas’was originally conceived, program benefits were to be available

only to the primary earner in a household, and only if that person had suffered a

61

job loss due to the economic crisis (World Bank, 2002). In this section, I outline a
method for estimating the characteristics of beneficiaries assuming the
government had been able to successfully target benefits only to this group, while
either enforcing or not enforcing the work requirement. Although this requirement
would have proved impossible to enforce in practice, it is worth assessing
whether it would even have been a desirable goal in light of the government’s
and Bank’s focus on it as a measure of targeting success.

I will estimate the characteristics of primary earners who lost their jobs
both under the assumption that the work requirement continued to be enforced,
and assuming it had not. Assuming enforcement of the work requirement
necessitates generating probabilities of participation for each primary earner, for
which I use the estimated probabilities from section 2.3.4. In this case, the
assumption that reservation wages are unaffected by a partner’s participation is
less critical, because most households have only one primary earner.

I identify the primary earner in each household as the individual (head or
spouse) who earned the most money in 2001. In households with no reported
earnings, an individual is counted as a primary earner if they were unemployed in
2001 and their partner was inactive. In cases where the primary earner is
unclear, both individuals are assigned primary status. Using this methodology,
men are the primary earner in 65.52% of households, and women in 15.84%.

The primary earner is unclear in the remaining 18.64% of households.

62

The next task is to identify which individuals lost their jobs during the crisis
(as there is no way to identify which lost their jobs due to the crisis.) This requires
two pieces of information: whether an individual had a job prior to the crisis, and
whether they were jobless following the crisis. On the first count, I assume that
an individual had a job prior to the crisis it that individual reported (1) being
employed in 2001 or (2) earning significant positive income (more than 100
pesos monthly) “from working.” (I do this because of the way the employment
variable is constructed in the survey - if you didn’t work the previous week, you
might not be counted as employed.)

On the second count, I take a relatively comprehensive approach to
identifying job loss: I count individuals who reported being without employment
(unemployed or inactive) in October 2002 as jobless, as well as individuals who
reported being employed as of October 2002 but who have held their job for less
than 5 months, meaning they would have been jobless at some point during the
Jefas application window. Even though Jefas participants are all listed as
“employed” in the EPH, I count them as employed only if their reported income is
greater than 250 pesos per month. This will almost certainly over-count
joblessness among Jefas participants, as some with very low incomes might
have had low-wage employment prior to enrolling, yet after enrollment would be
earning the 150 peso minimum. Galasso and Ravallion (2004) estimate that
participants suffer very little foregone income, and many participants came from

inactivity, so the degree of over-counting among participants is likely small.

63

Using these criteria, there are 667 male and 344 female primary earners
who lost their jobs, including 59.6% of male participants and 20.9% of female
participants. As in the previous section, I estimate characteristics of participants
with the work requirement enforced as a weighted average, using the
probabilities determined from the model in section 2.3.4 above, which assumes
that any individual can participate. Results are given in columns 3 and 6 of Table
2.7 for couples and columns 3 and 6 of Table 2.8 for single men and women.
Table 2.9 compares the characteristics of primary job-losers with and without the

work requirement.

2.6 Discussion

I estimate that lifting the “one person per household" requirement would
have increased the number of male applicants in households headed by a couple
by 46%, while increasing the number of female applicants in those households by
just 16% (Table 2.7). However, this would only have changed the fraction of
applicants who were male from .34 to .39 in households headed by a
male/female pair, while leaving the number and gender ratio of single applicants
unchanged.

Additional applicants come from households that are poorer than the
average applicant household in terms of per-capita income, but are otherwise
remarkably similar. Given an unlimited budget, this would imply better coverage

of poor individuals. However, since the budget was limited (with many applicants

64

turned away), coverage of the poor could not have improved, and the distribution
of benefits might have ended up more uneven at the household level, with some
households receiving benefits from two participants while others were turned
away. With first-come, first-serve enrollment, there would be no way to ensure
that those households enjoying additional benefits would be among the worse off
in terms of income. Given these considerations, if the goal is to give a minimum
income to as many poor households as possible, the results from this study
indicate that goal would be best achieved by maintaining eligibility of only one
person per household.

On the other hand, assuming the work requirement were still enforced,
limiting application to primary earners who lost their jobs would have resulted in a
much poorer group of applicants, both among households headed by couples
and by single earners (Tables 2.7 and 2.8, columns 3 and 6). This group of
applicants would also have been overwhelmingly male, with men comprising 70%
of likely participants in households with a man and woman.

Limiting enrollment in this way would have been far superior in terms of
exclusion of the non-poor, as measured by per-capita income. However, it would
come at a non-trivial cost in terms of covering of the poor, with only 33% as many
men and 6% as many women likely to have applied for benefits, implying a far
weaker safety net. The cost of the program under these circumstances, however,

would have been fairly low.

65

If the work requirement had not been enforced, and benefits had been
given to all primary earners who lost their jobs, the number of male applicants
would have increased by 277%, and the number of female applicants would have
decreased by 50% compared to the group of actual participants (Table 2.9). This
would have resulted in a group of applicants that was 63% male. Per-capita
income would have been somewhat higher in households with a male applicant,
while female applicants would have been far better off. On average, applicants
would have been less poor.3 Kernel density estimates of the per-capita incomes
of this group of applicants, compared with actual applicants, are given in Figure
2.1. Due to the huge increase in the number of applicants, coverage of the poor
would certainly have been good, but exclusion of the non-poor would have
suffered and the monetary cost of the program would have been unsustainable.

Another way for the government to reduce the budget would have been to
give benefits to primary earners who lost their jobs, but only if they had children.
The characteristics of this group are given in the last two columns of Table 2.9
for couples, again assuming the work requirement was not enforced. The size of
this group is more comparable with, though still larger than, the group of actual
applicants, and still heavily male. Household per-capita income is lower for men,
but higher for women, compared with the group of actual applicants, and higher

on average (116.59 pesos/month compared with 81.76 pesos/month for actual

 

3 Note that 2001 per-capita income is much higher than observed/imputed 2002 per-
capita income for job losers, as would be expected since these are people who, by virtue
of losing their jobs, experienced a substantial decline in income. Observed/imputed
income is probably the better measure in this case.

66

applicants). Kernel density estimates of the per-capita incomes of this group of
applicants, compared with actual applicants, are given in Figure 2.2.

Figure 2.3 compares the per-capita incomes of primary earners who lost
their jobs depending on whether or not they have children. As can be seen, those
who have children appear worse off. In neither of these scenarios do participants
have to work, which may be a benefit in terms of household utility. However, in
light of participants’ self-reported satisfaction with the work requirement, this
should not be a foregone conclusion. According to a survey conducted by
Tchemeva and Wray (2005), more female participants reported satisfaction with
the program due to the opportunity to “do something” outside the home than for
any other reason. When asked whether they would prefer to receive equivalent
welfare payments to stay home, the vast majority of women said “no.” Thus, it is
far from certain the increase in household utility from dropping the work

requirement would be very high.

2.7 Conclusion

In terms of the overall goal of the heads of household program, which was
to reduce poverty given a limited budget, it does not appear that any of the
potential changes outlined in this paper would significantly improve performance.
In light of the limited budget, coverage of the poor is best achieved, or at least not
significantly harmed, by maintaining the “one-person” requirement, even though

the fraction of male participants could be increased somewhat by dropping it.

67

Increasing the number of male applicants is not, however, a meaningful goal in
terms of reducing poverty, and the slight change in the gender ratio of applicants
would probably have negligible political benefits. Exclusion of the non-poor would
be harmed by targeting benefits to primary earners who lost their jobs, whether or
not they had children, unless the work requirement were maintained. In that case
coverage of the poor would suffer. Absent purely political considerations, the
government should not be too concerned about the failure to target benefits to
primary earners who lost their jobs. Compared with the alternatives covered here,
the Heads of Household program was effective in achieving its original goal of

targeting income to the poor.

68

Figure 2.1 Kernel Density Comparison of Applicants with Primary Job Losers

 

Density

.004

 

.l

.008 .01

.006

.002

 

 

 

o. 160. 260. 360,
Household Per-capita Income

 

----- Newly Unemployed Primary Earners.
Actual Jefas Applicants

kernel = epanechnikov, bandwidth = 20.0000.

 

 

 

69

 

Figure 2.2 Kernel Density Estimate — Primary Job Losers w/ Children

 

 

 

 

o 160. 260. 360,
Household Per-capita Income.

 

—-—-- Primary Job Losers w/Children
Actual Jefas Applicants

 

 

kernel = epanechnikov, bandwidth = 20.0000.

 

 

 

70

Figure 2.3 Kernel Density Comparison of Job Losers

 

 

co /'\
o _ -- -
°- /’ \ \,
./ \ \.
al ./ .\
' / \\
5' \\
8g / \\
CI .1 \\
'\
\‘-\.
.\.\ ———— \‘
\\\§\~_-§‘\
~ ..... E...\
o -i I 7 I I
0. 100 200. 300

Household Per-capita Income.

 

--—-- Primary Job Losers With Children
----- All Primary Job Losers

kernel = epanechnikov, bandwidth = 20.0000.

 

 

 

 

 

 

71

Table 2.1 Tobit Regression Used for lmputing Income for Couples

 

Dependent variable is 2002 household income per capita

 

 

 

 

Coefficient Std. Error
Household 2001 income per-capita 0.37 [0.07]***
Number of earners in hhld in 2001 -3.71 [4.26]
Household size 0.97 [4.27]
Man employed in 2001 -25.42 [12.78]**
Man unemployed in 2001 -32.57 [1433]"
Woman employed in 2001 10.59 [10.25]
Woman unemployed in 2001 -22.63 [16.95]
Man last job during crisis -17.34 [7.94]**
Woman last job during crisis -13.32 [13.64]
Man's education (years) -0.68 [1.28]
Woman's education (years) 1.15 [1.20]
woman is domestic worker -2.93 [11.04]
Man self-employed In 2001 -14.54 [7.15]**
Woman self-employed in 2001 —38.35 [17.67]**
Rooms per person 21.11 [14.50]
Lives in a Shantytown -1.12 [15.93]
Number of children age 0-5 -8.68 [6.17]
Number of children age 6-10 7.17 [5.17]
Number of children age 11-15 -8.27 [5.73]
Number of children age 16-18 -5.39 [9.06]
Elderly in household 7.47 [12.07]
Couple is formally married 18.47 [8.31]**
Man's age 0.59 [0.57]
Woman’s age -0.37 [0.60]
Constant 61.4 [34.89]*
Observations 227
Number of left-censored observations 23

 

* significant at 10%; ** significant at 5%; *** significant at 1%.
Standard errors are in brackets.
Regression includes region dummies

72

I’m

Table 2.2 Tobit Regression Used for Imputing Income for Singles

Dependent variable is 2002 household income per capita

 

 

 

Coefficient Std. Error
2001 Household income per capita 0.25 [0.08]***
Number of earners in household (2001) 6.81 [8.49]
Household size -1.61 [7.10]
Employed In 2001 27.43 [18.06]
Unemployed in 2001 27.05 [16.72]
Weekly work hours in 2001 -0.02 [0.08]
Last job during crisis 28.20 [18.62]
Education (years) -2.61 [1.91]
Self-employed in 200.1 -4.82 [17.56]
Rooms per person 6.27 [8.78]
Lives in a Shantytown -5.81 ' [27.86]
Number of children age 0-5 -1.01 [12.99]
Number of children age 6-10 1.18 [11.15]
Number of children age 11-15 2.17 [11.28]
Number of children age 16-18 -11.38 [12.27]
Elderly in household 16.50 [16.28]
Age -1.00 [0.62]
Constant 94.95 [42.52]**
Observations 96
Number of left-censored observations 10

 

* significant at 10%; ** significant at 5%; *** significant at 1%.

Standard errors are in brackets.

Regression includes region dummies, and a dummy for single men.

73

 

Table 2.3 Linear Regression Used for lmputing Income for Couples

 

Dependent variable is 2002 household income per capita

 

 

 

 

Coefficient Std. Error
Household 2001 income per-capita 0.37 [0.07]***
Number of earners in hhld In 2001 -4.27 [4.20]
Household size 1.35 [4.17]
Man employed in 2001 -24.09 [12.60]*
Man unemployed in 2001 -3o.71 [14.08]** 1*
Woman employed in 2001 9.70 [10.03]
Woman unemployed in 2001 -16.53 [16.19]
Man last job during crisis -13.79 [7.72]* 5
Woman last job during crisis -11.54 [13.21] ,’
Man's education (years) -0.40 [1.25]
Woman's education (years) 1.27 [1.17]
woman is domestic worker -1.69 [10.69]
Man self-employed in 2001 -13.24 [7.00]*
Woman self-employed in 2001 -30.38 [16.81]*
Rooms per person 18.33 [14.09]
Lives in a Shantytown 0.71 [15.37]
Number of children age 0-5 -8.25 [6.03]
Number of children age 6-10 5.57 [5.03]
Number of children age 11-15 -7.86 [5.58]
Number of children age 16-18 -5.78 [8.83]
Elderly in household 9.60 [11.71]
Couple is formally married 16.53 [8.11]**
Man's age 0.42 [0.56]
Woman's age -0.28 [0.58]
Constant 65.16 [34.18]*
Observations 227
R-squared 0.44

 

* significant at 10%; ** significant at 5%; *** significant at 1%.

Standard errors are in brackets.

Regression includes region dummies

Table 2.4 Linear Regression Used for lmputing Income for Singles

Dependent variable is 2002 household income per capita

 

 

 

Coefficient Std. Error
2001 Household income per capita 0.25 [0.09]***
Number of earners in household (2001) 2.53 [8.53]
Household size -0.38 [7.14]
Employed in 2001 20.87 [18.41]
Unemployed in 2001 23.42 [17.08]
Weekly work hours in 2001 0.00 [0.08]
Last job during crisis 28.23 [19.29]
Education (years) -2.23 [1.94]
Self-employed in 2001 -2.94 [18.14]
Rooms per person 7.06 [9.03]
Lives in a Shantytown -6.00 [28.91]
Number of children age 0-5 -0.32 [13.28]
Number of children age 6-10 -2.82 [11.47]
Number of children age 11-15 2.07 [11.22]
Number of children age 16-18 -11.63 [12.47]
Elderly in household 12.19 [16.76]
Age -0.80 [0.64]
Constant 94.18 [43.94]**
Observations 96
R-squared 0.48

 

* significant at 10%; ** significant at 5%; *** significant at 1%.

Standard errors are in brackets.

Regression includes region dummies, and a dummy for single men.

75

Table 2.5 Single Men's and Single Women’s Reservation Wage Equations

 

 

 

 

Men Women
2001 monthly earned income 0.3124 0.1278
[0.0728]*** [0.0345]***
2001 weekly hours of work 0.0003 -0.001
[0.0031] [0.0009]
Education 0.0703 0.0629
[0.0242]*** [0.0170]***
Woman is domestic worker -0.1472
[0.1460]
Self-employed (2001) 0.0677 -0.313
[0.2128] [0.1640]*
2001 monthly non-earned income 0.5704 0.1794
[0.1359]*** [0.0488]***
Rooms per person 0.0629 0.1473
[0.0954] [0.1122]
Lives in a Shantytown 0.1591 -0.0145
[0.4117] [0.3562]
Lives rent free 0.1565 -0.084
[0.2154] [0.1734]
Household size 0.5358 0.0934
[0.1206]*** [0.0580]
Number of children age 0-5 -1.0092 -0.1229
[0.2623]*** [0.1003]
Number of children age 6-10 -0.875 -0.1855
[0.2243]*** [0.0913]**
Number of children age 11-15 -0.5857 -0.2219
[0.2875]** [0.0877]**
Number of children age 16-18 -0.7319- -0.2473
[0.3627]** [0.1152]**
Elderly in household -0.6591 -0.0277
[0.2651]** [0.1762]
Age 0.0399 0.027
[0.0084]*** [0.0059]***
Constant -3.1906 -0.1011
[0.6208]*** [0.4914]
Observations 1173 1597

 

* significant at 10%; ** significant at 5%; *** significant at 1%.

Standard errors are in brackets.

Notes: Estimates are conditional on having at least one child or disabled individual in
the household, using 2001 EPH data. Coefficients were estimated in two separate
regressions. A positive sign on the coefficient should be thought of as increasing the
reservation wage, which corresponds to a decrease in the probability of applying.

76

mucmeoEma .8 6339:. m.

mEouc. Eocamao: 8.38-33 .302

 

 

mod mod 3d 2d 50:88: 5 2.25m
8R BR 8.3 8.3. 8m :26
Rd Rd Ed Rd 2:85 83.8 :58...
8.2: Rdfl RdR «SR 2,58: 5
8.8 8.2. 8.8 8de :85 9:85 8:.8 :26
8.8 Rdm Rd: Rd. «>58: 5
Gd m; 8.8 SR :85 98: >38; :26
8d 8d 8d 8d 8...:2
end end Rd Rd 2-3 8m
odd odd 8d 8d 2.: 8m
8d 8d 8d Rd 8-: 8m
8d 8d Rd Rd md 8:
. 8R 8R :3 RR 8.2.8 .3 .2352
8d 8d 8d 8d 5.3 5 82..
8d vmd 8d 8d :88: .8 2:8:
8R 8R :2 RR 858:8
8.2. Rdn 3R 8.? :85 2:85 5:: 338.8.
8.8 8.? 8.8 8.? 38:5 9:85 5:: 838.8:
3.8 2.2 8.8 on? 3.85 2:85 2:: 338-8:
8.. 8.. RR RR 98538 .o .8532
8m: 8m: :de :va 25.9.88 .3 .8532
332 835.3: .88.: 835:8

can>> 8.65.53 caz 6265.53

 

3:83.? 332 can omumEzmm 50 8335.00 RN 0.9:...

77

 

3000.03.00 .8 0300:... m. 088:. 06:00:02 3.000960 .302

 

 

 

3.0 00.0 v~.0 v~.0
00.3 00.2» 00.9 v0.m¢
00.0 ~00 00.0 00.0
mm.~0H mm.m0H 3.03 00.03
00.00 Samoa 00.00 00.00
mfimm H00.» 00.00 3.0:.
00.0H 00.5 n0.m~ mimm
0v.0 0Y0 3.0 00.0
00.0 00.0 3.0 3.0
00.0 N00 0N.0 0~.0
00.0 00.0 0.0.0 n~.0
m0.~ V00 H00 00.0
.00 00.0 00.0 00.0
~00 00.0 am; vmé
H0.\. 00.0 00.0 00.0
0.0.00 N040 00.00 N000
mfimn 3.00 #00 3‘03
H000 N000 H000 $.03
03 00.03 00 00.~0
00$ 003 mm: mm:
.0306. 0308.000 _030< 0305.me
c0Eo>> 205m. :02 2050

06:00:05 c. 2.020

000 :30

0E8... 050.00 00500....
020.8: 5

3003 0:60... 000.00 :30
0258: .5

:83 0.3: 500; :26
00:02

3.3 000

3-: 000

070 000

MIC mam

00.2.00 .6 .00502

0___> E 00>:

000.00 .00 05.03.

00:00:00

:83 0:.85 5:: 3.08.0:
C0005 0:60... 2:: 3.000960
0.80 0585 5:: 358-.....
3:00.600 .6 .0952
0003020000 .6 .00832

 

3:0u__00< .030< 0:0 030E300 .6 0050960 9035.503 0.N 030...

78

 

.E0.00.0 0:0 3 052000 .6 3:60:80 020.000 0 0:30:

2030.205 .0 00.05.300.05 0:. .6 000.0>0 03:902. 0 0.0 00:_0> 03.0000 .I0m ~00~ 0:0 H00~ E0... 0.0 0.00 .0302

 

 

79

8d 8d 8d .502 :58...
8d 8d 8d 2d 2d 2.0 50:88: 5 3.050
8.8 8.08 8.8 8.8 8.5. 8.5. 08 :30
8d 8d 8d 8d 8d 8.0 :90 5:85 05:.8 :58...
83.0: 8.00: 8.8: :88 8.88 ..~.8~ 020.8: ..
3...... :8 8.8 80.8 88: 0...de :83 0:85 .8500 :30
8.8 8.8 8.8 8.8 8.5. 8.5. 0258: 0.
8.8 8.0 08.0 0:8 008 8.8 0.8: >583 :30
Ed 8d 8.0 0...: 00.: 8d 8.8:
00.: 8d 8d ..~d 8d Rd 070: 08
8d {.0 8d 8d 8d 00.: 07: 08
8d 8d 8d 8d 8d 8d 8-0 08
8d 8d 8d 3.: 8d Ed 0.: 08
8R 8.0 8R 8.0 8R 8.. 8.5.8 .o .0082
8d 8d 8d 8d 8d 8d 5..., 5 83.
00.: 8d ..0d Gd 80d 00.: 80.0: .0: 058,:
8.: 0... 8.: 00.0 8.: 8.. 8.08500
8.8 8.00 8.8 8.00 8.00 8.8 :83 0:.85 3.08:8 5::
8.8 8.8 8.8 8.8.. :08 8.0: 0:050 2:85 3.08.80 5::
8.8 0H,... 2.0. 8.3. 00.8 05.8 0.80 0:.85 0580...... 5::
8 0D. 8.. 00 88 RR 3:05.80 0 05.8..
08 83. 83. 80 83. 83. 0:505:88 .0 .882
00a 00:
.8. .0500 8.05.80: 0.0.8.800 8.8 .0500 8.8.38... 60.02.80.
>.0E_.0 .0:t00 02 630.0 >.0E_.0 .0:..00 02 _030<

 

 

3:00:00... 20:50

ml|8|0:8.. < 0.:Iz

2000.8:0 2050.600. 5.030
00. :00 0 00000: 006:0000: :. 3:00: < .0 30:0 :60 EN 030...

 

 

 

.E0.mo.0 00. 00 00.2000 .0 3500080 020.000 0 9.300

0.000.200. .0 00.05.300.000 00. .0 0m0.0>0 0000902. 0 0.0 m0:_0> 000.0000 .I0m Noon 000 Sam 5.0.. 0.0 0000 .0302

 

 

 

 

 

80

0~d 00.0 00.0 .50... 8.000..
00.0 050 050 050 ..~d ..~d 50:00.00: 5 2.050
00.8 00.5. 00.5. 00.8 00.0.. 00.0.. 000 :30
8d 8.0 ~0d 8d 8d 8d 30.. 0585 050.00 5.000..
5.5.. ~08. ~08. 2.8. 8.00. 8.00. 020.000 ..
8R. 8.8. 8.8. 00.8. 00.8 00.8 3003 0:.85 00:.00 :30
~58 8.0.. 8.0.. 050.. 00.0.. 00.0.. 020.000 ..
-.- 8.2 8.2 8.8 8.8 00.8 3003 0.00: 5003 :30
8.0 00.0 00.0 8.0 00.0 00.0 070. 000
00.0 00.0 00.0 050 050 2d 07: 000
8d ~00 ~00 Ed 8d 8.0 2-0 000
00.0 00.0 8.0 00.0 Rd Rd 0-0 000
.0.~ v0.~ ..0.~ 00.. 00.0 00.0 :0.5.:0 .0 .00502
00.0 00.0 00.0 00.0 ..0.0 00.0 0__.> 5 00>:
00.0 00.0 00.0 00.. v~5 ..~.H 80.00 .00 03.00:
~0~ 8.0 8.. ~...0 8.0 8.0 8.000000
0.....0 ~05. ~05. 8.8 8.8 ~08 300$ 0:.85 00000-80 5:...
85.0 3.8 :8 8.8 00.8. 00.85 0005.. 058:. 0:08.80 55..
00.00 ~08 ~08 ~58 5.0.8. 00.02 9.03. 0:.85 0:08.80 5:...
5.d.. 00.0... 00.0... 00.... 8.~0 8.~0 3:05.000 0. 00:05.
00. 80. 80. 00. MR. 8: 25.020000 .0 .00502

:2 :2
.00.. .0: .00 5.05.000... 8000:5000. .03 .0: .00 8.05.0000 500006.000.
>.0E_.0 .05.00 02 _030< >.0E_.0 .05.00 02 _030<

 

 

00:00:03. 0.05.00. 00:0u__00< 202

8080.00 000800300. 5.0.5

 

 

 

0.0.00.0 000 00 9.3.000 .0 30.00090 020.000 0 9.300 0.030.200. .0 00.05.080.000
00. .0 000.0>0 000090.... 0 0.0 00:.0> 000.0000 .I0m meow 000 Sam E0... 0000 ”00002

 

00.0 00.0 “0.0... 5.000..
00.0 2.0 00.0 050 50:00.5: 5 0.050
00.8 0.0.00 ~08 00.5. 000 :30
8.0 00.0 Rd 00.0 30.. 0:.85 050.00 00000..
005.0. .0.~.~ 00.00. ..~.0m~ 020.000 ..
3...... 0~...0~ 00.8 00.00. :83 05005 000.00 :30
8.8 00.8 8.00 8.5. 020.000 ._
00.8 0500 00.0 ..0.0~ 0.50: 05003 :30
00.0 8.0 00.0 00.0 000.0...
00.0 v~d 00.0 Rd 070. 000
00.0 00.0 00.0 00.0 07: 000
50.0 00.0 00.0 ~0d 050 000
00.0 3.. 8.0 00.0 0-0 000
8.~ 8.0 8.~ 00.~ 00.5.:0 .0 .0050:
00.0 00.0 00.0 00.0 0___> 5 00>...
00.0 00.0 ..0.0 00.0 80.00 .00 00.00:
00.0 00.0 8.0 50.0 00.000000
00.00 00.8 00.00 00.00 :08. 0500:. 05000.00 5::
00.8 050.. 0~.00 8.00 0005.. 0500:. 05009.00 5::
8.00 8...... 0500 00.00 3.00.. 058:. 00.000000 5::
0~ 00 00.. ~H~ 0.005.000 .0 00:00..
00~ 80 8~.. 8~.. 0:0..020000 .0 .00502
000.02, 00... 000.02, 00... .0. 0000.00.00

 

81

0008000000 5.02.
000 000.. .00.0m >.0E..0

0000.85.00”. 0003
000005.00. .032

 

 

 

5.0.00.0 00... 3 9.3.000 .0 35.5390 025000 0 9.300 0.000.200. .0 02.000.00.000
00. .0 000.0>0 00.0902. 0 0.0 00:.0> 000.0000 ....0m Noom 000 Sam 0.0.. 0.00 .0302

 

82

00.0 00.0 .50... 05.00...
050 050 80 0~d 50:00.00: 5 0.050
.55. 00.... ~00. 00.... 000 :30
8.0 00.0 00.0 00.0 0.0.. 05005 05:.00 :0000...
....00~ 00.000 0.00~ 808 025000 5
00d..~ 00.000 ..0.~m~ 00.000 ..00~. 0500:. 00:.00 :30
0...... ~05. 00.8 8.0.. 025000 ..
05~m .~.~.. 00.50. .0.~.. 0.00: 05003 :30
~00 00.0 00.0 00.0 000.0...
5.d 00.0 00.0 0~.0 0.0. 000
00.0 00.0 00.0 00.0 0.-.. 000
00.0 00.0 ~50 ~0d 0.0 000
00.0 .0d 98 .0.0 0-0 000
00.~ 00.~ 00.. 0..~ :0.5.:0 .0 .00502
00.0 00.0 00.0 00.0 0...> 5 00>...
~0d ..0.0 00.0 00.0 000.00 .00 0500.0
.0.0 .0.0 00.0 00.0 05.00000
8.8. 00.00 00.~... .000. ..00~. 0500:. 05000.00 5::
00.00 05.0 00.8. -.00 0005.. 0500:. 05000.00 5::
05.0 05~0 0~.0~. 00.00 .503. 0500:. 05000.00 5::
0... 00.. 00~ 000 0.0005000 5. 00:05.
0... 00.. 00~ 000 0000020000 .0 .0050:
50503 cm: cmEoZ’ :02 ”m. “Ewan-.00.?—

 

 

00.0__0U 00... 0 5.0.5 02
000 000-. .00.0m. 3.00.00

000000.300... 5.0.5 02
00... .00.. .00-Em. >.0E..0

 

 

REFERENCES

Besley, T. and S. Coate ( 1992), “Workfare versus Welfare: Incentive Arguments
for Work Requirements in Poverty-Alleviation Programs.” American Economic
Review, Vol. 82, No. 1, pp. 249-261.

Cuff, K. (2000), “Optimality of Workfare with Heterogeneous Preferences.” The
Canadian Journal of Economics, Vol. 33, No. 1, pp. 149-174.

Galasso, E. and M. Ravallion (2004), “Social Protection in a Crisis: Argentina’s
Plan Jefes yJefas.” World Bank Policy Research Working Paper 3165.

Graham, J.W. and SJ. Donaldson, (1993), “Evaluating Interventions with I
Differential Attrition: The Importance of Nonresponse Mechanisms and Use of
Follow-up Data.” Journal of Applied Psychology, Vol. 78, No. 1, pp. 119-128.

 

Jalan, J. and M. Ravallion (1999), “Income Gains to the Poor from Workfare.”
World Bank Policy Research Working Paper #2149.

Leeds, M.A. and P. von Allmen (2004), “Spousal Complementarities in Home
Production.” American Journal of Economics and Sociology, Vol. 63, No. 4, pp.
795-812.

Little, R.J.A. and DB. Rubin (2002), tti i lAnl i wihMi in D t .New
York: John Wiley & Sons.

Ravallion, M. and G. Datt (1995), "Is Targeting through a Work Requirement
Efficient?" in Dominique van de Walle and Kimberly Nead (eds) Public Spending
and the Poor: Theory and Evidence, Baltimore: Johns Hopkins Press.

Sen, AK. (1979), “Utilitarianism and Welfarism.” Journal of Philosophy, Vol. 76,
No. 9, pp. 463-89.

Subbarao, K., A. Bonnerjee, K. Ezemenari, J. Braithwaite, C. Graham, S.
Carvalho and A. Thompson (1997), “Safety Net Programs and Poverty.”
World Bank Publication.

Tchemeva, P. and LR. Wray (2005), "Gender and the Job Guarantee: The
impact of Argentina's Jefes Program on Female Heads of Poor Households."
CEPS Working Paper #50.

World Bank (2002), Report #PlD10834, May 21 , 2002.

83

World Bank (2006), “Project Appraisal Document on a Proposed Loan in the
Amount of US$350 Million to the Argentine Republic for a Heads of Household
Transition Project.” World Bank Report No. 32463-AR, Feb. 21, 2006.

84

CHAPTER THREE

The Impact of Jefes yJefas de Hogar on Children’s
Work and School Attendance

“As long as there is family poverty, there will be child labor. "

-UN|CEF

3.1 Introduction

Many organizations that deal with child labor, including the lntemational
Labor Affairs Bureau (ILAB) at the US. Department of Labor, argue that outright
bans on child labor may be counterproductive. Rather, they propose that helping
families alleviate poverty is one of the best ways to address the issue of child
labor in developing countries. Conditional Cash Transfer (CCT) programs are
designed to address such issues explicitly, for example by requiring participants’
children to attend school as a condition of receiving benefits. The Department of
Labor believes public employment programs like Jefes yJefas de Hogar in
Argentina, despite their lack of conditionalities, may also have a significant
impact on the incidence of child labor in those countries by decreasing
households’ need for additional income. This view is consistent with theoretical

work done by Basu and Van (1998), who argue that the primary cause of child

85

 

labor is parental poverty, and empirical work by Edmonds (2005), who finds that
economic status improvements can explain declines in child labor.

At the same time, child labor is rarely included as an outcome variable in
impact evaluations of non-COT development programs, and the effect of
Argentina’s Jefas program on child labor has not yet been evaluated (Sipos and
Lyon, 2009). This study directly evaluates Jefes’ impact on child labor and school
attendance, using a comparison group of children whose parents applied to the
program but did not receive benefits because they applied after what turned out
to be the cutoff date. I develop both semiparametric and parametric estimates of
the program ’3 impact, because they require different assumptions, and reveal
different information about the relationship between the treatment and outcomes
of interest. Using cross-sectional and panel data, I find that children age 10-14
whose parents enrolled in the Jefas program and received benefits were 59
percent (1.3 percentage points) less likely to report working, and 38 percent (2
percentage points) less likely to forgo schooling, compared with similar children
whose parents were not enrolled.

The remainder of the chapter is structured as follows. I describe relevant
details of child work and school attendance in Argentina in section 3.2. Section
3.3 outlines the methodology for estimating treatment effects using cross-
sectional data, and section 3.4 does the same for panel data. I describe the data
and descriptive statistics in section 3.5. In section 3.6, l present and describe

results. Section 3.7 concludes.

86

3.2 Child Work, Schooling, and Jefas in Argentina

Argentina’s 2001-2002 economic crisis caused a substantial decline in the
real incomes of many workers, and a doubling of poverty and extreme poverty.
According to McKenzie (2004), most workers were unable to draw from savings,
and unable to increase their work hours, as a way to mitigate the effects of the
crisis. One possible labor market response, outlined in the first chapter, would be
for affected families to send other individuals, including children, into the
workforce to compensate for the decline in household income. Indeed, while
estimates of child work in Argentina vary considerably, news sources at the time
of the crisis reported an increase in child work (Palacios, 2003) and
corresponding decline in school attendance (Hennigan, 2003).' Likewise,
UNICEF (2003) reported a 600 percent increase in child work between 1995 and
2003, with 40 percent of those children who worked abandoning school. ILO-
lPEC (2002) reported an increase in child work between 1997 and 2002, which it
attributes to the economic crisis. To my knowledge, there have been no formal
economic studies of this phenomenon.

Urban child labor is a visible problem in Argentina, with children working in
jobs including trash recycling, street sales, begging, selling trinkets on the

subway, and shoe shining (CONAEI'I, 2009). Seeing young children engaged in

 

I In 2001, the ILO estimated that 2.2 percent of urban children ages 10 to 14 years in
Argentina were working, and in 2002, the Ministry of Labor estimated that 7.1 percent of
children ages 5 to 14 were working, including in rural areas. In 2002, a UNICEF
representative reported that in urban areas 6 of every 10 children ages 13 to 17 were
working rather than studying.

87

difficult labor, often late into the night, is a daily occurrence in Buenos Aires, and
even tourists will encounter the problem. This happens even though it is not
legal: According to ILAB (2009), Argentine law prohibits employment for “minors
who have not completed compulsory education, which normally ends at age 15,
and with few exceptions bars employment of children under the age of 14
outright. Children age 14-18 require special permission from administrative
authorities to work. In addition, children ages 14 to 18 are prohibited from
working more than 6 hours per day, and must present medical certificates
attesting to their ability to perform such work. Children under the age of 18 are
prohibited from working between the hours of 8pm and 6am in work that could
endanger their safety, health, or moral integrity.” It appears that these laws are at
best loosely enforced, at least outside the formal sector.

One of the goals of the Jefas program was to help children, and the
program was explicitly targeted to families with minor dependents, because
children were perceived as being among the most vulnerable during the crisis. In
the media, many of the complaints about the program stemmed from the large
number of mothers participating instead of staying at home with their children,
which was perceived to harm those children. However, it is possible that by
allowing the mother to earn income, a goal she may have been unable to achieve
in the market, the program may have significantly alleviated the need for a child

to fill that role.

88

 

The primary goal of this study is to assess the effect of the Jefas program
on the incidence of child work. However, a child who is not working may engage
in any number of other activities, including substituting for the foregone home
production of parents who enroll in Jefas. For that reason, I will evaluate whether
there is a corresponding increase in school attendance. Argentina has one of the
best public education systems in Latin America, and education is free and

compulsory for 10 years, beginning at age 5.2 According to Argentina’s Census 3

 

and Statistics Institute (INDEC, 2001), in 2000 more than 90 percent of children UL
who enrolled in primary school in Argentina reached grade five and 79 percent

completed primary education, which comprises grades 1-9 (ages 6-14).

3.3 Semiparametric and Parametric Estimation of
Treatment Effects

This study addresses the impact of the Jefas program on child work and
school attendance. It is common practice in the evaluation literature to define the
“impact” of a social assistance program as the difference between some measure
of the outcome of interest with the program and its counterfactual value for
participants in the absence of the program, where the estimate of the
counterfactual is based on a matched comparison group of non-participants.
Such estimators are semiparametric in the sense that only assignment to the

treatment group is modeled parametrically. Estimating treatment effects in this

 

2 See Ministerio de Cultura y Educacion de la Nacion, “Argentine Education in the
Society of Knowledge.” Available at
http://www.zona.lacarabela.com/zona98/EASC/eng/home.html

89

manner, outlined by Rosenbaum and Rubin (1983), is appealing in cases where
(1) assignment to the treatment group can be made plausibly random conditional
on observed covariates, and (2) there is sufficient overlap in the conditional
probabilities of treatment between the treated and comparison groups. In this
study, I use as the treated group the sample of children with a parent who is
enrolled in the Jefas program and is receiving benefits. I will refer to members of
this group as “participants.” The comparison group is the sample of children with
a parent who had applied to the program, but was not receiving benefits, at the
time of the October 2002 survey. I will call members of this group “applicants."
Applicants are similar to participants in that, although they are not receiving
benefits, they have indicated a preference for program participation. Galasso and
Ravallion (2004) argue that applicants and participants are sufficiently similar that
assignment to the treatment group can be treated as random conditional on
observed characteristics. I will show that the children of applicants and
participants can be treated in the same manner.

In chapter 1, l outlined some differences between participants and
applicants—for example, that participants were more likely to have come from
larger households or to have been employed in construction prior to the crisis
than were applicants (see Table 1.3). To control for this sort of observable
heterogeneity, I construct a counterfactual outcome from the group of applicants
using their propensity scores, or probabilities of participation conditional on

observed covariates. I estimate propensity scores parametrically by fitting a

90

 

probit model of participation (i.e. assignment to the treatment group), and using
the fitted model to predict the propensity scores.

Once the propensity score for each individual has been estimated, I
estimate counterfactual outcomes of working and school attendance for children
of participants in the absence of the program from the outcomes for children of
applicants not yet receiving the program by taking weighted averages over
outcomes for individuals in the latter group, who are observationally similar to the
participants in terms of their propensity scores. There are N children of

participants indexed i: 1,...,Nand Pchildren of applicants indexed j = 1,...,P. Let
YiK be the outcome of interest for individual i in state K, where K = T for
participants (the treated group) and K: C for applicants (the control group).

Applying weights Wij to calculate the counterfactual for each participant, the

estimate of the mean impact is

IN
76.2

(III

P
T C
[1" ‘ 2“”:in
j-l

 

where for each i
P
EWU =1
j-l
Consistent with much of the impact evaluation literature, I use local linear
weights, which are constructed using all individuals in the control group and
which have been found to perform better at the boundaries of the propensity

score. Local linear matching is competitive with other estimators in terms of bias

91

 

so long as there is good common support, and achieves lower variance than
many estimators because more information is used (Busso, et al. 2009).
Matching is implemented in Stata using <psmatch2> with a default bandwidth of
0.8 (Leuven and Sianesi, 2003). Standard errors are bootstrapped with 100
repetitions.

Because I believe conditions (1) and (2) above are likely to be satisfied
(which I will verify later in the paper) the propensity-matched estimate of the
average treatment effect on the treated (ATT) will be my preferred estimate of the
Jefas treatment effect. However, for comparison with the semiparametric
approach outlined above, I find it useful to also parametrically estimate the
program’s impact on the outcomes of interest using a nonlinear model. Whereas
semiparametric estimation requires only that assignment into the treatment group
be random conditional on observable characteristics, traditional parametric
estimation of treatment effects requires assumptions about the full relationship
between the observed characteristics and outcomes of interest. If such
assumptions can be reasonably made, parametric regression estimates can yield
additional insight, unavailable in the context of semiparametric estimation, into
the possible causal effect of covariates other than program participation on the
outcomes. In this study I estimate probit models of the effect of participation on
child work and school attendance, in which concerns about unobserved
differences governing selection into the group of participants is mitigated by using

as a comparison the group of children whose parents are applicants, and whose

92

 

parents have thus indicated a preference for participation, as above. If selection
into the treatment group is sufficiently random (as will be suggested by the low
explanatory power of the probit used to compute the propensity score), then this
analysis should be appropriate. The primary coefficient of interest in the probit
models is the one indicating the effect of having a parent who is a participant,
and the standard error on that coefficient is a rough approximation of the
statistical certainty of the average treatment effect on the treated. To construct an
estimate of ATT, I use the fitted probit to predict the probability of the outcome
variable for each treated observation, and find the average. To estimate the
counterfactual outcome, l recalculate the predictions with the treatment dummy

set to zero, and again take the average. The difference in averages is the
estimate of ATT. To illustrate, let xK be a dummy variable representing treatment.

After fitting the model to estimate 300.3,. on the whole sample, ATT is estimated

using only the sample of participants as

N
711-2[¢(fl0 + 31sz + + 3k-1xiJc-1 + 3") ADM) + .81in + + ék'lxi’k-l )]

i-l

3.4 Estimation of Treatment Effects for the Longitudinal
Sub-sample

One possible source of concern about the cross-sectional estimates of the
average treatment effect described above is that there might be pre-existing

(before treatment) differences in the average reported child work and school

93

attendance rates between the treatment and control groups, which would bias the
estimate of the treatment effects. This could happen if selection into the group of
participants was caused by some unobservable trait which is also correlated with
work or school attendance—for instance, if children of participants were working
more than children of applicants before the crisis, because the families of
eventual participants were worse off, applied first, and were thus more likely to

have been selected into the program.

 

For this reason, I re-estimate semiparametric estimates of ATT using the

10.

sub-sample of children for whom there is panel data from before the crisis.
Weighted averages of all covariates are computed again for this group, using the
propensity scores that were previously computed using the entire sample. I then
compare the weighted averages of the pre-crisis values of the dependent
variables to verify that, after weighting, there are no significant differences in
those values between the treatment and control groups that would be of

concern.3

3.5 Data and Descriptive Statistics

As in the first chapter, I use data from the October 2001 and October 2002
rounds of the Encuesta Permanente de Hogares (EPH), which covers urban

areas accounting for 70% of the Argentine population. These data identify Jefas

 

3 A common way to address pre-treatment differences in the dependent variables due to
unobservables is to construct difference-in-difference (DD) estimates of ATT. I outline
this approach, and construct DD estimates of ATT for both outcome variables as a
robustness check, in the appendix to this chapter.

94

participants, as well as individuals who applied to the program but were not yet
receiving benefits, whom I call “applicants.” In the first chapter of the dissertation,
I showed that the observed characteristics of participants and applicants are
remarkably similar, which I believe makes the group of applicants a suitable
control group for program participants. This was the approach followed by
Galasso and Ravallion (2004) in their analysis of foregone incomes.

In this paper, the individuals of interest are children under the age of 15,
each having a parent who is either a participant (the “treatment group”) or
applicant (the “control group”). Because data are unavailable in the EPH for
children under the age of 10, the sample will be children age 10-14.4 Descriptive
statistics for this group, which includes 2,307 individual children in 1,536
households, are given in Table 3.1, where they are divided into two mlumns
depending on the application status of the parent. In the 2001 cross-section,
there are 1706 children with parents who are participants, and 601 with parents
who are applicants. Despite their similarities, there are some significant
differences in the explanatory characteristics of participants and applicants,
which motivates propensity matching to control for these observed differences.

The dependent variables that I use are dummy variables indicating
whether the child reports working (during the survey week), and whether the child
reports attending school. Both characteristics differ significantly between

participants and applicants. Children of participants appear about half as likely to

 

4 The results are sensitive to inclusion of older children. Propensity-matched estimates of
ATT using children up to age 16, which are smaller in magnitude and statistically
insignificant, are given in Table 3.9.

95

work (0.9 percent of the children of participants vs. 2.0 percent of the children of
applicants), and somewhat more likely to attend school. Other observed
characteristics are similar, as would be expected if assignment to the treatment
group were truly random, yet there are some significant differences. Participant
households are larger and have more children, for example. For that reason, I will
use propensity matching to balance the distribution of observables between the
two groups. With respect to published estimates of child work in Argentina’s
urban areas, it also appears that child work is under-reported in the EPH. This
will be a problem only to the degree that under-reporting varies between children
of applicants and children of participants.5

To find the estimates of the average treatment effect on the treated for the
longitudinal sample, I use the 2001-2002 panel of children who are age 10-14 in
2002, which includes data on 1034 individuals. Because subsequent interviews
are conducted at the same residence, rather for the same family, there is no
guarantee that panel observations will be of the same individual child in 2001 and
2002. For that reason, I verify age and sex for consistency, dropping
observations for which sex changes between 2001 and 2002, or for which age
does not increase by one year. This leaves 980 individuals in the panel data set.
Among the non-participant households observed in this data, the overall
percentage of children working tripled from 0.84% in 2001 to 2.52% in 2002,

while the percentage attending school fell from 98.32% to 94.54%. The observed

 

5 The most likely reason for this to occur would be if, despite the survey’s guarantee of
anonymity, children of participants or their parents (they answer the survey together)
were less likely to report child work for fear of being dropped from the program.

96

 

overall increase in working and decline in school attendance is consistent with
anecdotal accounts that some families sent children to work as a way to cope
with lost income during the crisis. There are, additionally, no statistically
significant pre-treatment differences in either child work or school attendance
between the treatment and control groups, although the children of participants
were 0.2 percentage points more likely to have reported working in 2001 and 0.2 P?
percentage points more likely to have attended school. The observation that

there were no significant pre-treatment differences helps allay concerns about z,

 

such differences causing bias in the estimate of ATT.

3.6 Results and Discussion

The probit model used to generate propensity scores for participants and
applicants is given in Table 3.2. Certain variables are significant in this
regression—children from larger households, as well as households with a
female applicant or in which the household head is an applicant, do appear
somewhat more likely to be in the group of participants, which is consistent with
the descriptive statistics. Nonetheless, the probit has relatively low explanatory
power, which is consistent with both the historical reason for applicants not
receiving benefits due to an arbitrary cutoff date, and with the clear ex ante
similarity of observable characteristics between applicants and participants. With
such similar treatment and control groups, it is not surprising that there is a large

region of overlapping support, shown graphically in Figure 3.1. There is also no

97

 

mass near the corners, which implies consistency of the propensity-matched
estimator. Because assignment to the treatment group appears to have been
made plausibly random, and there is good overlap in the conditional probabilities
of treatment between applicants and participants, I consider semiparametric
estimation based on the propensity score to be an appropriate, and my preferred,
method of estimating ATT.

The propensity-matched estimates of the average effect of participation on

 

the reported incidence of work and school attendance using cross-sectional data,
along with weighted averages of the explanatory variables, are shown in Table
3.3. A parent’s participation in Jefas is estimated to decrease the probability that
a child works by 1.3 percentage points, from 2.2% to 0.9%, and increase the
probability of school attendance by 2.0 percentage points, from 94.7% to 96.7%.
These estimates are significant at a 90% confidence level, with p values of
around 0.06. Also, note that weighting based on the propensity score seems to
have evened out the distribution of the explanatory variables between treatment
and control groups. Matching quality can be assessed using two-sample t-tests
to assess whether there are significant differences between covariate means for
both groups. By this standard, only the number of children and household size
remain unbalanced in the weighted sample.

Probit regressions used to parametrically estimate ATT for work and

school attendance, for comparison with the propensity-matched estimates, are

98

shown in Table 3.4}3 Point estimates of the average effects of participation on the
outcome variables (ATT) are given in Table 3.5, and are similar to the
propensity-matched estimates both in magnitude and significance. A parent’s
participation in Jefas is estimated to cause a 61% (1.4 percentage point)
reduction in the probability that a child reports working, from 2.3% to 0.9%.
Similarly, the estimated probability that a child attends school increases by 1.9
percentage points, from 94.8% to 96.7%.

Note that the estimated effect of participation varies significantly with age.
For the children most likely to work and least likely to attend school, those age
14, the estimated effect of participation is to decrease the probability of working
by 2 percentage points, from 5.6% to 3.6%, and to increase the probability of
attending school by 4.3 percentage points, from 88.1% to 92.4%.7

The probit analysis also suggests that the sex of the child and the wealth
of the household are significantly correlated with reported work and school
attendance for this group of children, independent of participation. Male children
are more likely to work and less likely to attend school. Greater household

wealth, measured by the number of rooms per person, is—not surprisingly-

 

6 Because some households contain more than one child, I report standard errors robust to
clustering at the household level. Accounting for clustering appears to have little effect
on the standard errors, which increases confidence in the original assumptions. I know of
no way to account for clustering in the propensity-matched estimates, and so I re-estimate
those results for the cross-section and longitudinal samples using only one randomly
selected child per household, and present the results in Table 3.8. Estimates are
generally consistent with the main results in Table 3.3.

Note that the estimated 5 .6% rate of child work for the counterfactual group is still low
for this age group, compared with some of the published estimates reported in section
3.2.

99

associated with a lower probability of working and a higher probability of
attending school. This is consistent with the proposed explanation for why
participation should help alleviate child labor: increasing the income of a family
should decrease the incidence of child labor because families use child work to
compensate for inadequate adult income or financial resources.
Propensity-matched estimates of the effect of participation on the reported
incidence of work and school attendance using longitudinal data are shown in
Table 3.6, along with weighted means of the covariates for the longitudinal sub-
sample, and lend confidence to the cross-sectional estimates. As can be seen,
after weighting on the propensity score there remain no significant differences in
pre-treatment reported work or school attendance between the treatment and
control groups, which reduces concern about selection into the treatment group
based on unobservable differences that are correlated with the outcomes.
Because a difference-in-difference (DD) analysis is intended to account for pre-
treatment differences in the outcome variable between treatment and control
groups that would bias the results, this finding of no significant pre-treatment
differences means that a DD analysis should be unnecessary in this case. Point
estimates of ATT using this group are similar to the cross-sectional estimates,
with a decrease in work of 1.9 percentage points and an increase in school
attendance of 2.4 percentage points, although the of the effect on school

attendance is no longer significant.

100

There are around 2 million participants in the heads of household
program, each having, on average, 0.896 children between ages 10-14, or about
1,792,800 children in this age group. Taken at face value, the point estimates,
which are summarized in Table 3.7, suggest a reduction in the number of such
children working of between 23,300 and 34,000, and an increase in the number
attending school of between 34,000 and 43,000. The propensity-matched
estimates indicate a reduction in working of 23,300 and an increase in school
attendance of 35,800. Given that the base rate of child work is likely to be under-
reported, these estimates are probably a lower bound on the number of affected

children.

3.7 Conclusion

In this study I show some evidence to suggest that parents’ participation in
the Jefas public employment program in Argentina helped alleviate child labor
and increase school attendance for children age 10-14. I find semiparametric
estimates of the average treatment effect on the treated (A‘l'l'), using applicants
to the program as a control group. Using cross-sectional data from 2002, I
estimate that participation decreased the reported incidence of child work by 1.3
percentage points, and increased school attendance by 2.0 percentage points.
The latter result suggests that children derive real benefits from their parents’
participation in the program. Parametric estimates of ATT confirm these results.

Longitudinal estimates of the treatment effect, using panel data from 2001-2002,

101

 

allay concerns that the treatment effect may be due to pre-existing differences
between the treatment and control groups. Back of the envelope calculations
show that the program may have caused a reduction in the number of children
working of around 23,300, and an increase in the number attending school of
around 35,800, which is likely an underestimate of the number of affected

children.

102

 

 

Figure 3.1 Overlapping Stmport in the Distribution of the Propensity Score

 

 

 

 

.2 .11. .5. .5. 1,
Propensity Score.

|_ Untreated [:1 Treatedl

 

 

 

 

 

103

 

.Iam ~oo~ .30 SP: Ema .meon 3 amino mm; on; 29.3 m o>mc on; €73 can 5.6.29 2 29:8 ”mouoz

.Boxofin 5 2m mco32>on Emucflm

 

 

 

 

 

 

How OONH mco_um>._0mno Lo LUQEDZ
*and .32 2.3.3 #3 58.3 $3 Sc. 2233:
:93 m3? :93 note Emmi 23 .52. 228:2 8.th
L3: Rod 53 mono 8me mm; “.8; 228:9. magm
LR: ~36 5.2 cm: €me .83 29.8.8; 5 352.8% do .3532
*med 83 33.2 Red Head 2:... 2.2 c. Ev 56:5 so 285:2
L31 ~85- 5&2 Bed 23.3 mmvd 888 ha 253.
3.3%: m~o.o Emma v35 H83 mood £8.58
Home; 885. 5.3 $3. 53 $05 59355 m a. $2..
2288: as .6 859:2
$8.3 mad H83 owed 88.“ Bed .38; 228:2. 2 28.32
1:88; Rod- Home; man... So: and 22: m. EB=S<
Home ~mm.o 2.3.3 26% Emma 3.2 can 28%?
gaging. .6 859:2
$31 98.? row: 98.: 39...: «38.: a?
$21 Bod- 803 End H83 :3 was
56:5 so 833%
8.5.: 3:... ~32 Nmmd 52 $2 628. 8:32
3.82; «3.? 8e: oNod ammo; mood 9.82,
unumuu w:_m> .>w_u.um cam—z Smudm cmwz
diam - .32:
~65th Eamo__nq< 3522th

 

 

104

Table 3.2 Probit Regression for Calculating the Propensity Score

Dependent variable is a dummy indicating program participation

 

 

Coeff St.Err.
Lives in a Shantytown -0.04 [0.12]
Bathroom 0.12 [0.10]
Rooms per person -0.14 [0.15]
Number of children <18 -0.05 [0.03]*
Number of applicants in household 0.13 [0.07]*
Applicant age 0.01 [0.00]
Applicant is male -0.47 [0.09]***
Applicant is household head 0.33 [0.11]***
Single household head -0.12 [0.12]
Married household head -0.04 [0.07]
Child is male -0.02 [0.06]
Child's age -0.03 [0.02]
Household size 0.07 [0.03]***
Cuyo -0.71 [0.10]***
Noroeste -0.20 [0.07]***
Nordeste -0.13 [0.08]
Number of observations 2306
Number treated off support 0
Pseudo R2 0.04

 

* significant at 10%; ** significant at 5%; *** significant at 1%.

Standard errors are in brackets.

Notes: Sample is children age 10-14, who have a parent who has

applied to Jefas. Data from Oct. 2002 EPH.

105

.Iam Noon .60 Eat Ema .mmuoa 3 8:25 mm; oz; 293 m m>mz on; €73 mom c2273 m_ 29:3 ”9302

.3335 5 2m 225323 Emncmum

 

 

 

8: 82 mcozmtomno .6 .3532
53682 338.2 835 928 mm: 37. 2238:
$8.63.; $91 ¢~o.o- mote End 82 223:2 BEE).
$3.58.; :58 Sod 2:5 mm; 28; 29.3.5; 295
$8.83.; 83; ~86. memo ~93 one; 228.8; 2 28.33
38.58.; Home; «88. and 2m... 22. m. 28:22
H~m~zmmfi¢ 831 $~.o- 22m 29?. won £8.32
$8.63.; H-.8 god 8: ~o~.~ 223:2. 5 3:833 .o .3632
Eo~..~¢o.. itsem v2.8 Sam 33. 2.2 5 av 55:5 .3 22:52
88:28; 321 Sod- onto was 528 be mecca
P338; :9: God :35 ~85 52:58
38.58.; 821 Sod- Sod Sod czsbcmfi m 5 $2..
$3.53.; 2&3 mead- mmmd o3: om<
38:38.; $21 08.0. o~m.o 82 was
23.48.; Load e36 :55 Sad .858 $53
$8.59; L33 3...? -o.o 83. 9.3;
.3 #3 33m-“ c8: :82 :8:
2228 - ucmEumwbv m_ob:ou
mococota uncoumz 3592th

 

mucmucmfl< Begum ucm 3:25 cc (.2393th no woman: ounco>< 2: Lo oumEzmm nocoumz-fi_mcmao& n.n 03¢...

106

Table 3.4 Probit Regressions for Work and School Attendence

 

Attends
Works School
Parent is Jefes participant -0.48 0.24
[0.21]M [0.12]*
Lives In a Shantytown 0.32 -0.09
[0.27] [0.21]
Bathroom 0.08 0.15
[0.28] [0.17]
Rooms per person -1.22 1.1
[0.79] [0.42]***
Number of children <18 -0.03 -0.05
[0.07] [0.05]
Household size 0.05 0.02
[0.07] [0.04]
Number of applicants in household 0.33 -0.04
[0.20] [0.13]
Applicant age 0.01 -0.01
[0.01] [0.01]
Applicant Is male -0.66 0.15
[0.25]*** [0.20]
Applicant is household head 0.68 -0.11
[0.26]*** [0.22]
Single household head -0.53 0.01
[0.34] [0.24]
Married household head -0.17 0.25
[0.25] [0.14]*
Child is male 0.47 -0.36
[0.19]** [0.11]***
Child's age 0.33 -0.31
[0.06]*** [0.04]***
Constant -7.39 5.6
[0.86]*** [0.62]***
Pseudo R2 0.25 0.14
Number of observations 2306 2306

 

 

* significant at 10%; ** significant at 5%; *** significant at 1%.
Standard errors are in brackets, and are robust to clustering at the

household level.

Notes: Sample is children age 10-14, who have a parent who has applied to
Jefas. Data from Oct. 2002 EPH. Both regressions include region dummies.
Probit coefficients represent marginal effects.

 

1mm Noom 9.: Eat 8m mama .v.m
2an E “58a 9.: E0...— mozzfimnoca 83695 9.7.: #5. .6 $3553 qua "302

 

m3... 30.9. 8 u 2&2th Tam - S u “892:3 IE
omcmco .mEEmE
mvmd m~o.o 8 u 23.5.3 Tam
Road mood 2 u ESGEE IE
.006m 9:25 230
$55

 

Cb: oucmncofi< _oocum
Dam #83 co cozmafintma no Human: mmflw>< 05 Lo oumEzmm H.995 m.n 030...

108

 

.Iam 60.: 3mm. .mmhma 3 opium mm; on; ucoema m o>mc on; €73 omm 5.2.5 2 29:3 "mouoz

.mumxomcn 5 95 22333 Emncmum

 

 

 

 

109

050 Ohm mco_um>._mmno h.0 LUDEDZ
meémog **Smd 32 30.0 .93 SB 223.8:
38.58.; 25.8 Sad End mnmd 22 22322 8:52
”$9.30.; :93 mwod- 33 ~25 22 22082 226
88:39; 511 20.? 81¢ 2:3 22 223:2 m. 28.32
38.53.; :9: H86 32 33 22. 2 28.32
Hoomémtmg ”8.1 08.7 83% Bosm mum 28.32
Rodfion; Rod; mood- 23 £3 22252 5 328.3% .6 .2632
88:23 $an 82 amen 33. 22 5 av 5:26 20 22:52
88.68.; LEA; Rod- 83. Etc 822 .2 252
8238.; 5.3 ~86 Bad Sad 58.58
$8.58.; $31 Sod- v8.0 «.85 $8325 m a. $2..
$3.63.; 53; RS- Rod 8%: mm<
38.48.; :33 29o- Hmmd 32 222
$8.53.; R31 mood- Bad $3. 88 s 628 8282
$8.58; 5.3 mood- Sod Sad 88 s 2225
28.6.8.2 Sci «36 mead $66 .028 2282
38:39; L84; Sod- mmod 925 233
.3 $3 5.3 :82 :2: :8:
295.30 - ucoEumme m_obcou
8:88.20 nocoumz 3532th

 

oEEmmbam 3:633:04 9: 05m: 8:22.33
Begum Ucm x82, :0 cozmsgtma Lo fimaEH ommeo>< 9: no oumEzmm nocoumzLfificanQ ed 032.

 

 

 

 

Quad muod cued muod 02.0.5qu
m¢md ovmd nvmd Nmmd 3395::
momd nomd nomd nomd 033:...
cmcmnc ”.38: 3:32: 823% .0038 mucofi<
8:39:

mnodu v.36- nnodu fined- 00:95.50
omod mNod NNod oNod BEBE:
mood mood mood mood omummt.
922: ”—50.5 8:89.: 8233 9.33
3:39:

 

fimtm acmEummFr no mmumEzmm “Son. :0 :omtmoEoU EM 039..

110

.xam Eat 38 .2388; 2: 5?

383.5308 Em Eotm Emucfim .Eoncm. an 3623 6.03330: L3 2.5 mco 2.5 25.2... $353.. 5302

.5 am 28:29.0. 5:. “cam am 28533 1 £03 am 28533 *

 

 

 

 

 

V00 V00 mCODmEmmno u—O LQDEDZ

Foo..m~o.¢ 35.8 3...: $2 onmd .858 883
£8.58; :3; 3a.? ¢~o.o wood 9:25
.H.U o\omm umumnu mam) cme cmmz EQEansm _MC_U:u_mco:_
ommfi ommfi mcozmzmmgo Lo 355::

38:83 223.: «no... ~35 £2 .028 $55
H3953; L3,: 3...? o~o.o Sad 9:25
.3 £3 “33 c8: cum: cams 23mm .mcofimmamob Noam

cobcou - 2958.5 20.5.50
mucEmtE 8:322 mucmEuEmn.

 

2238... an. 26 80 >20 5;, mSmEEm 8552:2805 "29:92 Esamcmm ad 032.

111

.zam Noam 59: £8 .mcozsmae OS 5.; 832688
Em Eotw Emucmum =< .3355qu 5 8m Eotm Emucmum ”3oz

 

 

$89.28.; $30.58.; 3&8... 883:8 $3
$8.8 $898
336 . «wood. 6 u x _ 5m - f u x _ :8
8:89.06 8232).
385 $86 6 u x _ 5m
~33. “~85 c. u x _ com
_85m $5? 9:25 2:6

 

3-3 wm< cmLEEU L8 wucmocwfiq. .oocum 0cm x83 co 83c: mew><
2: ”6 333mm nmcsmzémcmaea "2392 333.23 ad 033

112

APPENDIX TO CHAPTER THREE

A common way to address pre-existing differences between treatment and
control outcomes based on unobservable differences is to construct a difference-
in-difference (DD) estimator to compare the outcomes of interest (child work and
school attendance) before and after the implementation of the Jefas program for
the treatment and control groups. Although nonparametric DD estimators cannot
be estimated consistently, they are common in the program evaluation literature.
In this appendix, I construct such DD estimates for comparison with the cross-

sectional estimates as a robustness check.

With outcomes Y,-,K for individuals 1' in time tand state K, the general form

of the DD estimate is (suppressing the i subscript for now):
DD - [3(Y1T)— 12116)] — [Eb/F) — 1%(YOC)]

If there are n individuals in the panel sample of participants and p individuals in
the panel of applicants, a matched double difference estimate of the average
treatment effect, which aligns the distribution of observables between participants
and applicants, can be implemented as

DD- ii Y5 4,}, - iWij-(chl 41%)

i-l j-l

Weights are found following the method in the previous section, using propensity

scores calculated from the cross-section of applicants and participants.

113

A difference-in-difference estimate of the average treatment effect can
also be found parametrically. Define a dummy variable 7', with T: 1indicating
assignment to the group of participants and T: 0 indicating assignment to the
group of applicants. The dummy variable findicates the time period with t: 1
during the period following implementation of the program (2002) and t: 0 before
program implementation (2001). The DD average treatment effect is found from
the regression

Yr =a+fiTi +Y1i+5(Ti"i)+£i
where the si are assumed to be normally distributed.

When the above equation is estimated as a linear regression, the
interaction term 6 can be thought of as the “true effect of treatment.” However,
because both outcomes of interest (working and attending school) are binary
variables, I run the regression as a probit model. In a non-linear DD regression,
interpretation of the coefficients is less straightforward than for the linear
regression. Puhani (2008) shows that the sign of the treatment effect in certain
non-linear DD regressions, including the probit model, is equal to the sign of the
coefficient on the interaction term. However, in any nonlinear model, the
coefficient on the interaction term may not be a reliable estimator of the true
interaction effect (see Ai and Norton, 2003). I therefore use the fitted model to
find the DD estimate of the average treatment effect in the manner suggested by

DeLeire (2004), by taking the discrete double difference of the standard normal

114

cumulative distribution function. Data used to find the DD estimate of ATT is the
same as for the longitudinal sample in the body of the paper.

Propensity-matched difference-in-difference estimates of ATT are given in
Table 3.10. Participation is estimated to decrease the probability that a child
reports working by 1 percentage point, and increase school attendance by 2.1
percentage points, which is consistent with the main cross-sectional estimates.
The reported incidence of work was estimated to slightly increase from 2001 to
2002 for the children of applicants but actually decrease for the children of
participants. Neither estimate is significantly at the 5% level, and I cannot reject
that the program had no effect on school attendance at any reasonable level of
significance.

Coefficients in the probit DD regressions are shown in Table 3.11. The
coefficient on the interaction term representing the treatment effect is only
marginally significant for working and not significant for school attendance. The
implied treatment effect on the treated, reported in Table 3.12, is to decrease the
reported incidence of child work by 3.0 percentage points and to increase school
attendance by 2.4 percentage points. Again, the sex of the child and wealth of
the household, using rooms per person as a proxy, are significantly correlated
with the outcome variables. Older male children are more likely to work and less
likely to attend school, while the opposite is true for children in wealthier

households.

115

.Iam 9: co 8.58
Noom ucm Hoom Eot Em Emu _ccma .mcoszcacc 8H 5.3 ucaambmuoon
8m Eotc Emocmum =< .3355ch c. Sm Eotc Emucmum ”muoz

 

 

58.53.; ”28:38.-“ 3285 888:8 $3
AmHNoc Awooog
336 33.? G n x _ skim - fi u x _ o>.im
8:8th mBJoU 3:33..
#8.? 885 G u x _ 338
$8.? 38.? f u x _ skim
Bofim 885 0:25 22c

 

coca—0:33 .0058 new v.33 co cozmggtmd no 63:: camco><
2: .6 @3833 cocoLoEoéToocccour—B ooanZLfificcaofi afin 2am...

116

Table 3.11 Difference-in-Difference Probit Regressions for Work and School

 

 

 

Attendence
Attends
Works School
After crisis 0.46 -0.39
[0.43] [0.28]
Parent is Jefes participant -0.12 0.05
[0.41] [0.27]
After*partlcipant -0.88 0.29
[0.51]* [0.32]
Lives in a Shantytown 0.58 0.52
[0.44] [0.46]
Bathroom -0.21 -0.02
[0.32] [0.20]
Rooms per person -2.71 1.81
[1.14]** [0.52]***
Number of children <18 0.16 -0.06
[0.13] [0.06]
Number of applicants in household 0.49 0.2
[0.18]*** [0.14]
Applicant age 0.05 -0.02
[0.02]** [0.01]**
Applicant is male -1.08 -0.32
[0.43]** [0.21]
Applicant is household head 0.63 0.27
[0.39] [0.23]
Single household head 0.54 -0.67
[0.50] [0.28]“
Married household head 0.95 -0.1
[0.32]*** [0.17]
Child is male 0.82 -0.26
[0.28]*** [0.14]*
Child's age 0.24 -0.23
[0.09]*** [0.05]***
Household size 0.03 0
[0.09] [0.05]
Constant -9.1 5.43
[1.68]*** [0.82]***
Pseudo R2 0.4 0.18
Number of observations 1960 1960

 

* significant at 10%; ** significant at 5%; *** significant at 1%.

Standard errors are in brackets.

Notes: Sample is panel of children age 10-14, who have a parent who has applied
to Jefas. Panel data from Oct. 2001 and Oct. 2002 EPH. Both regressions include

region dummies.

.Iam 05 “—0 mucae NOON .ucm Hoom Eat v.5 numb Benn. ”wuoz

 

 

cued owed- G u x _ £25m - r u x _ okim

mocccotfi 63:06 vacuum:

88. Rod 6 u x _ skim

688- mood- fi u x _ 33%.
Beam 8:32 9.82, 25

 

mocmucofia. _oo.._om ucm x33 co 5.392th no
”USE“ momcm>< c5 no BmEBmm oucwcctoéTmocccmtE ”.386 NH...” 039—.

118

REFERENCES

Ai, C. and EC. Norton (2003), “Interaction Terms in Logit and Probit Models.”
Economic Letters, Vol. 80, No. 1, pp. 123-129

Basu, K. and PH. Van (1998), “The Economics of Child Labor.” American
Economic Review, Vol. 88, No. 3, pp. 412-427.

Busso, M., J. DiNardo, and J. McCrary (2009), “Finite Sample Properties of
Semiparametric Estimators of Average Treatment Effects.” Working Paper,
available at http://www-persona|.umich.edu/~jdinardo/BDM2008_v11.pdf

Blundell, R., M. Costa Dias, C. Meghir, and J. Van Reenen (2002), “Evaluating
the Employment Impact of a Mandatory Job Search Program.” Journal of the
European Economic Association, Vol. 2, No. 4, pp. 569-606.

CONAETI (National Commission for the Eradication of Child Labor) (2009),
Trabajo infantil urbano, [online]; available from
http://www.conaeti.gov.ar/que_es/urbano.htm

DeLeire, T. (2004), “A Note on Calculating Difference-in-Differences Using Probit
Models versus Linear Probability Models.” Unpublished manuscript, Michigan
State University, East Lansing, MI.

Edmonds, E. (2005), “Does Child Labor Decline with Improving Economic
Status?” Journal of Human Resources, Vol. 40, No. 1, pp. 77-99.

Galasso, E. and M. Ravallion (2004), “Social Protection in a Crisis: Argentina’s
Plan Jefes yJefas.” World Bank Policy Research Working Paper 3165.

Hennigan, T. (2003), “Recession Pulls Children out of Argentina’s Classrooms.”
Christian Science Monitor (Buenos Aires), June 25, 2003; available from
http://www.csmonitor.com/2003/0625/p07s02-woam.html.

lLO-IPEC (2002), La OIT celebra el primer ’Dia mundia/ contra el trabajo infantil’,
press release, Buenos Aires, June 11, 2002

INDEC (2001), National Census, available online at
http ://www.indec.mecon.ar/censo2001 leampliada_index.asp ?mode=01

119

Leuven, E. and B. Sianesi (2003), “PSMATCH2: Stata Module to Perform Full
Mahalanobis and Propensity Score Matching, Common Support Graphing, and
Covariate Imbalance Testing.” http:l/Ideas.repec.org/c/boc/bocode/s432001.html.
Version 1.2.3.

McKenzie, David (2004), “Aggregate Shocks and Labor Market Responses:
Evidence from Argentina’s Financial Crisis.” Economic Development and Cultural
Change, Vol. 52, No. 4, pp. 719-58.

Palacios, C. (2003), “Crece la poblacion de chicos en las calles.” La Nacion
(Buenos Aires), July 9, 2003; available from http://www.lanacion.com.ar.

Puhani, P. (2008), “The Treatment Effect, the Cross Difference, and the
Interaction Term in Nonlinear ‘Difference-in-Differences’ Models.” IZA Discussion
Papers No. 3478, Institute for the Study of Labor (IZA).

Rosenbaum, P. and D. Rubin (1983), “The Central Role of the Propensity Score
in Observational Studies for Causal Effects.” Biometrika, Vol. 70, No. 1, pp. 41-
50.

Sipos, S. and S. Lyon (2009), “Policy options for addressing child labor.”
Presentation at the 4th European Forum on the Rights of the Child, 18 June 2009.

UNICEF (1997), “The State of the World’s Children 1997.”

UNICEF (2003), “Erradicacion del trabajo infantil.” See
http://www.dol.govfilab/media/reportsficlpltda2003/argentina.htm#__ftn202.

US. Department of Labor, ILAB (2008), “US. Department of Labor’s 2008

Findings on the Worst Forms of Child Labor.” Report required by the Trade and
Development Act of 2000. Available online at http://www.dol.govfilab/.

120