WPS6476
Policy Research Working Paper 6476
Can Conditional Cash Transfers
Compensate for a Father’s Absence?
Emla Fitzsimons
Alice Mesnard
The World Bank
Development Economics Vice Presidency
Partnerships, Capacity Building Unit
June 2013
Policy Research Working Paper 6476
Abstract
This paper investigates how the permanent departure this finding strongly suggests that the channel through
of the father from a household affects children’s school which the father’s departure most affects children is by
enrollment and work participation in rural Colombia. reducing the income of very poor households, which
The results indicate that the permanent departure of tightens their liquidity constraints. This finding also
the father decreases children’s school enrollment by highlights the important safety-net role played by welfare
approximately 5 percentage points and increases child programs with respect to disadvantaged households,
labor by 3 percentage points. This paper explores the particularly because these households are unlikely to have
rollout of a conditional-cash-transfer program during the formal or informal mechanisms with which to insure
period of study and shows that this program counteracts themselves against such vagaries.
these adverse effects. When coupled with other evidence,
This paper is a product of the Partnerships, Capacity Building Unit, Development Economics Vice Presidency. It is part
of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy
discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org.
The authors may be contacted at emla_f@ifs.org.uk and Alice.Mesnard.1@city.ac.uk.
The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development
issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the
names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those
of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and
its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent.
Produced by the Research Support Team
Can Conditional Cash Transfers Compensate for a Father’s
Absence?
Emla Fitzsimons and Alice Mesnard
JEL classification codes: I20, J12, J22, O16
Keywords: child labor; schooling; permanent absence; income loss; credit and insurance
market failures; conditional cash transfer; safety net.
Sector Board: Social Protection (SOCPT)
Emla Fitzsimons (corresponding author) is Co-Director of the Centre for the Evaluation
of Development Policy (EDePo) at the Institute for Fiscal Studies; her email address is
emla_f@ifs.org.uk. Alice Mesnard is Reader at City University, Senior Research
Economist at the Institute for Fiscal Studies, and Research Affiliate of the CEPR; her
email address is Alice.Mesnard.1@city.ac.uk. The authors are grateful to participants at
the IFS, CMPO, World Bank, and 3ie-LIDC Seminar Series and to participants in the
NEUDC conference for comments. They thank Orazio Attanasio, Rajeev Dehejia, Ian
Preston, Dominique van de Walle, and Marcos Vera-Hernández for useful comments.
They also thank the Editor and two anonymous referees for very helpful suggestions.
Funding from ESRC, grant number RES-000-22-1742, is gratefully acknowledged. All
errors are the responsibility of the authors.
A major disruption to family life can have serious consequences for children. A
particularly traumatic event is the father's permanent departure from the household. There
are at least three different channels through which this event can affect children’s human
capital accumulation, particularly their school and work participation rates (for further
discussion of the following points, see Case, Paxson, and Ableidinger [2004] and Gertler,
Levine, and Ames [2004]). First, this event is likely to involve substantial income loss
that may particularly affect school choices coupled with credit and insurance market
failures. Second, the balance of decision-making power in the household may change,
with the preferences of the remaining adults gaining increased importance, which may
have important consequences for children. Third, the loss of a parent may have
significant emotional and psychological consequences for children. The importance of
the first and third channels was highlighted in a World Bank Development Outreach
report (Bell, Bruhns, and Gersbach 2006):
If parents sicken and die while their children are still young, then all the means
needed to raise the children so that they can become productive and capable
citizens will be greatly reduced. The affected families’ lifetime income will
shrink, and hence also the means to finance the children’s education, whether in
the form of school fees or taxes. On a parent’s death, moreover, the children will
lose the love, knowledge and guidance which complement formal education.
Some countries, particularly in Africa, have established policies to provide
education and healthcare support to children who have lost one or both parents. These
policies may have been implemented in response to the increase in HIV-associated
mortality, which has resulted in millions of children losing parents to AIDS. However, a
2
father who is absent from the household while a child is young is a pervasive global
phenomenon, although there is surprisingly little evidence that shows how children are
affected by the long-term departure of one or more parents and that recommends policies
to protect children against such adversities. In this paper, we first investigate how the
father’s departure from the household that results in his permanent absence affects
children’s school enrollment and work participation in Colombia. 1 We are interested in
the effects on children’s school and work participation because of the importance of these
institutions for human capital accumulation. Moreover, child labor affects family income
and current poverty, which is the reason we may expect it to increase to compensate for
income reductions. We then explore the rollout of a conditional cash transfer (CCT)
program, Familias en Acción (the purpose of which is to increase schooling for children
from poor backgrounds), and we examine the extent to which the effects of the CCT
program mitigate the adverse effects that are caused by the father's permanent absence.
The departure of the father from the family household is a relatively rare
occurrence among households in rural Colombia. To focus on more permanent reductions
in income, which are more difficult to insure against than transitory departures, we
consider only departures due to death and divorce, which we can be confident are
permanent. A central concern is that divorce or widowhood is not exogenous with respect
to other determinants of child outcomes (see van de Walle [2011] for related selection
issues). Previous work has attempted to exploit exogenous variation to overcome this
problem, for instance, in divorce laws (Gruber 2004) and child sex composition (Dahl
and Moretti 2008). In this paper, we provide several pieces of evidence that, taken
together, build confidence in the quasi-random nature of the father’s departure. First, we
3
show that important observable characteristics of households (before the departure
occurred) in which the father did and did not subsequently depart are similar. Although
reassuring, the concern remains that unobserved heterogeneities might distinguish these
types of households. Thus, we address time-invariant unobserved heterogeneity by
allowing for household-fixed effects over a three-year period with a panel of households.
Consistent with the related literature (see, for example, De Janvry et al. [2006]), our
empirical method assumes common trends across both types of households. Note that this
is conditional on a set of covariates, including transitory income shocks, which makes our
approach more credible. We assess the plausibility of this assumption of common trends
by examining predeparture trends in children’s schooling and per capita income across
households in which the father does and does not subsequently depart. We are reassured
by the fact that these trends do not differ significantly. To address endogeneity concerns
caused by potential correlation with time-varying shocks, we determine whether divorce
is correlated with recent significant time-varying shocks, including crop losses, business
losses, and illnesses. We find that it is not correlated in our sample. Finally, to build more
confidence in the quasi-random nature of the departure, we perform a falsification
exercise by determining whether current child activities are correlated with future
departure of the father (that is, if departure is quasi-random, then future departure should
not lead to a significant effect on current activities). Reassuringly, we find no evidence
that it does.
In the setting we consider in Colombia, our main finding is that the father’s
permanent absence from the household adversely affects the schooling of both boys and
girls and increases their participation in paid and unpaid work. These findings are
4
particularly pronounced for poorer households, which are likely to face more severe
liquidity constraints, and are consistent with the father’s absence affecting activities
directly through the associated income reduction. A second key finding of the paper is
that the CCT program Familias en Acción helps protect children against the vagaries of
the event by protecting their schooling and offsetting increased child labor after the
father’s departure. Because the CCT program effectively acts as a safety net, it appears
that the main impact of the father’s departure is felt through its associated income loss.
The paper is structured as follows. In section I, we provide a brief overview of the
related literature. Section II describes the data that we use in this research. We discuss
identification issues in section III and present the empirical methodology and main results
in section IV. Section V considers whether the CCT program introduced into the
environment we consider has cushioned poor households in our sample against these
effects, and section VI concludes.
<>I. RELATED LITERATURE
Our work fits into several strands of the literature. First, it is related to the
growing literature in developing countries on parental death and children’s education.
This literature investigates the importance of different channels in explaining the
observed effects (Case, Paxson, and Ableidinger 2004; Gertler, Levine, and Ames 2004;
Yamano and Jayne 2005; Beegle, De Weerdt, and Dercon 2006, 2010; Evans and Miguel
2007; van de Walle 2011). The literature generally finds adverse effects on schooling,
particularly on primary school participation. However, this literature generally does not
consider effects on child labor; clearly, this is an important economic activity among
5
children in developing countries and one that may be particularly responsive to an event
that induces a substantial income reduction. We consider the permanent departure of the
father through either death or divorce, which are the two events in our data that
consistently result in substantial income reduction for households. Although the channels
through which these events might affect outcomes may differ, 2 our empirical work
suggests that income reduction is the main driver of the observed effects.
More generally, our work is related to the literature that considers the causal
effects of family disruption on child outcomes, which has concentrated on the United
States and other developed countries. Although this literature has mainly focused on the
permanent absence of a parent resulting from death or divorce, recent work has also
highlighted the negative effects of parental migration on children’s schooling—
particularly the migration of the father because it is males who migrate in most contexts
(see Antman 2013 for a survey; Giannelli and Mangiavacchi 2010; Lahaie et al. 2009). A
key finding throughout this literature is that the absence of a parent is likely to be
correlated with unobserved factors, which may also explain the poorer outcomes of the
children. Several methods have been used to account for the influence of such factors.
For example, some studies have used sibling-difference (household fixed effects) models
(Case, Lin, and McLanahan 2001; Ermisch and Francesconi 2001; Ginther and Pollak
2004; Gennetian 2005) that account for the fixed and unobservable endowments that are
shared by siblings from the same family or the same mother, whereas others have
compared children’s outcomes before and after the divorce of their parents (Cherlin et al.
1991; Painter and Levine, 2000), which assumes that preexisting disadvantages of the
family or the child are captured by child fixed effects. Finally, quasi-experimental studies
6
have either considered parental death as an exogenous source of parental absence
(Biblarz and Gottainer 2000; Corak 2001; Lang and Zagorsky, 2001) or exploited
exogenous variation in separation rates based on differences in divorce laws across states
(Gruber 2004) or over time (Piketty 2003).
Our work also dovetails with the literature that considers the relationship between
children’s work participation and negative income shocks in developing countries, such
as labor market shocks (Parker and Skoufias 2006) or crop losses (Jacoby and Skoufias
1997; Dehejia and Gatti 2005; Beegle, Dehejia, and Gatti 2006; Dammert 2007; Duryea,
Lam, and Levison 2007; Gubert and Robilliard 2008; Guarcello, Mealli, and Rosati
2010). As described in this strand of the literature, our results are consistent with the
presence of credit and insurance market failures in rural Colombia.
The second part of the paper, which provides evidence that the CCT programs
attenuated the negative income effects entailed by the permanent absence of the father on
children’s activities, is consistent with the growing literature on the role of CCTs as
safety nets. Indeed, CCT programs are a rapidly growing part of social safety-net
policies, and there is evidence that they provide households with protection against both
systemic and idiosyncratic short-term shocks. For instance, De Janvry et al. (2006) show
that the Mexican PROGRESA program fully protected children’s schooling from shocks
caused by unemployment and illness of the head of the household and by natural disasters
in the community. Maluccio (2005) shows that the Nicaraguan Red de Protección Social
protected families’ total expenses (including food expenses) and children’s school
attendance against the effects of the Central American coffee crisis in 2000–2001. More
recently, Gitter, Manley, and Barham (2011) offer evidence that CCT programs have
7
mitigated the effects of negative shocks on physical development in early childhood. Our
results are consistent with these papers, suggesting that CCT programs do provide a
safety net against income losses for poor households. A distinctive feature of our work is
that we consider income losses that are likely to be permanent and that are therefore even
more difficult to insure against than transitory reductions in income.
<>II. DATA
In this section, we discuss the data used in the paper and present key descriptive
statistics relating to our sample.
<>Background
We use three years of panel data from a survey of households and individuals in
rural Colombia. These data were collected to evaluate the large-scale welfare program
Familias en Acción, which has been in place in several rural areas of Colombia since
2002 and has since expanded to cover urban areas. The program aims to alleviate poverty
by fostering human capital accumulation among the poorest households through
conditional subsidies for investments in education, nutrition, and health.
The first wave of data collection for the evaluation of the program occurred in
2002, when approximately 11,500 households were interviewed. We refer to this as the
baseline survey. A year later, after the program began, a second wave of data was
collected, and a third wave was collected in 2006. We refer to these as the first and
second follow-up surveys, respectively. In this paper, we estimate the effects of the
father’s permanent absence on children’s outcomes. 3 The socioeconomic data are rich
and reflect face-to-face interviews that lasted 3.5 hours on average.
8
<>Descriptive Statistics
We follow the school and work status of children in households with at least one
child aged 7–14 years at the baseline across the first follow-up survey (1 year later) and
the second follow-up (3.5 years after the baseline) until they are 17 years of age, at most.
Because we consider the effects of the father’s departure since the baseline, we restrict
the sample to households in which both parents are present at the baseline. 4
<>Outcomes. We consider two outcomes: school enrollment, which relates to
whether the child is enrolled in school at the time of the survey, and work participation,
which includes all types of paid and unpaid economic activities and includes looking for
work as a main activity. 5 Table 1 shows the proportions of our sample enrolled in school
and participating in work by age and gender (table 1). The table indicates that school
participation rates are high among children aged 7–11 years, which corresponds to
primary school. 6 The first substantial drop in school enrollment is observed at age 12,
which is the transition from primary to secondary school. Another point worth noting is
that the school enrollment of females is higher than that of males. Engagement in work is
approximately twice as high for males as for females and is very low for both before the
age of 12 (participation in work is not recorded for individuals under 10 years of age).
<>Permanent absence of the father. To capture a potentially important
disruption to family life and long-term reduction in income, we focus on the departure of
the father from the household since the baseline. In particular, we focus on departure that
results in the father’s subsequent permanent absence from the household. 7 Divorce and
death are the two primary reasons for permanent absences that are identifiable from the
data. Because these are relatively rare events, we pool them to improve statistical
9
precision. 8 There may be some concern that these events result in different levels of
transfers to the household. However, in our data, we found that the amount of transfers
received by the household is very similar in magnitude after death and divorce. This is
shown in table S1 in the appendix.
To measure the incidence of divorce, we combine information on the marital
status of the child’s mother at times t – 1 and t and the status of the father at time t. In
particular, if the mother's marital status at time t is divorced and her status at time t – 1 is
married and if the father’s status at time t is “no longer in the household�, we consider
this a divorce. Deaths, in contrast, are coded directly into the survey. The departure of the
father as the result of death or divorce occurred in 5.6 percent of our sample of
households (that is, those with at least one 7- to 14-year-old at baseline). Divorce
accounts for 82 percent of such departures, and death accounts for 18 percent. 9
The average age of fathers who leave the household is 43 years at the baseline,
which results in a substantial income reduction because 90 percent of fathers were
working at the baseline. To provide an idea of the extent of the income loss associated
with departure, we compared total household labor earnings across households with and
without an absent father. Total labor earnings are approximately 22 percent lower after
controlling for household composition (number of male adults, number of female adults,
number of children aged 0–6 years, and number of children aged 7–17 years). 10 We also
compared total household consumption, a more direct indicator of the welfare of the
households in our sample, and found that it was lower by approximately 13 percent in
households in which the father subsequently departed than in households in which he did
10
not depart, after controlling for household composition as above. Both differences were
statistically significant at the 1 percent level.
Regardless of whether such events can be fully anticipated, it is unlikely that the
households in our sample have mechanisms to fully insure against the income losses they
entail, particularly because these households are in rural municipalities in which credit
and insurance markets are thin (Edmonds 2006). Our data confirm that although
monetary transfers and in-kind transfers increase significantly after the departure of the
father, the magnitude is small compared to the income loss the departure entails. 11 Under
these conditions, we expect paternal absence to affect decisions about sending children to
school or work. Furthermore, paternal absence is likely to have a number of other
important repercussions (see, for example, Gertler, Levine, and Ames [2004] for a
discussion). First, the father is likely to have been one of the key decision makers in the
household, and his departure may bring about changes in bargaining power and decision
making within the household that may affect decisions about children’s education and
work. Second, the father may be an important figurehead for children. Although we
cannot disentangle these channels in the available data, in anticipation of our results, we
note that the evidence is strongly consistent with income loss as the key factor that affects
children’s activities.
<>III. IDENTIFICATION
Two issues that arise in identification are related to the potential endogeneity of
parental absence and attrition from the sample over time. In this section, we discuss each
of these issues in turn.
11
<>Endogeneity
An important concern about paternal absence, and one that has received
considerable attention in the related literature (see, for example, Gruber [2004]), is that it
may not be exogenous to the outcomes of interest (that is, children’s work and
schooling). For instance, couples may divorce because they have different preferences for
investment in children, in which case we may be identifying the effects of preferences
rather than divorce per se. 12 Although it is reassuring that the predeparture (that is,
baseline) observed characteristics of households that do and do not experience the
subsequent departure of the father are similar (the chief difference mainly relates to the
education level of the head of household, as shown in table 2), it is important to address
endogeneity concerns.
In our empirical work, we address these concerns in two ways. First, we control
for time-invariant unobserved confounding factors through household fixed effects.
Accordingly, our identification strategy relies on the assumption that the time trends are
the same in households in which the father does and does not depart. We examine the
plausibility of this assumption in detail below. Second, to address the concern that there
may be time-varying factors correlated with the father’s departure and child outcomes,
we assess the relationship between divorce and observed time-varying shocks, including
crop losses, business losses, and illnesses. We then control for such time-varying shocks
in the analysis to improve the conditional exogeneity of paternal departure.
To examine the plausibility of the common trends assumption, we investigate
trends in two key variables. First, we examine whether trends in children’s schooling
were the same in both types of households before the father departed. We have two
12
periods of school enrollment data before the departure, the baseline (2002) and the year
before (collected retrospectively at the baseline). We cannot reject the possibility that
schooling trends are the same in both types of households, as indicated by the statistically
insignificant coefficient of the interaction between the type of household (“absence�) and
the year dummy in the upper panel of table 3. As a second check for common trends, we
compare trends in household per capita income in both types of households before the
father departed. We have three periods of income data before the departure, all collected
retrospectively at the baseline. The evolution of per capita household labor income in the
years 1999, 2000, and 2001 is similar across both types of households prior to the father's
departure (lower panel of table 3). This finding gives us no reason to believe that they
would have differed if departure had not occurred.
It is also notable that the signs of the estimates in table 3 point to a positive
selection of families with absent fathers in terms of schooling and income trends
(although the selection is not statistically significant). 13 In this worst-case scenario, we
would underestimate the magnitude of the “true� impacts of departure. To anticipate our
main results, the detrimental impacts of the permanent absence of the father on school
enrollment and work would be even more pronounced than what we find in our study.
As an additional exercise to build confidence in the quasi-random nature of the
father’s absence, we determine whether current child activities are correlated with the
future\e absence of the father; that is, future absence should not lead to a significant effect
on current activities if departure is random. To do so, we regress current children’s
activities (schooling/work at time t for t = 1,2) on future absence (at time t + 1).
13
Reassuringly, we find extremely small and statistically insignificant correlations between
these variables (0.009 for schooling, –0.009 for work, both with P values of 0.5).
Although all of the evidence above is reassuring, it does not address concerns that
there may be unobserved time-varying shocks that affect both the father’s permanent
departure and the children’s outcomes. For instance, a temporary shock to income, such
as a crop/business loss or illness, may affect the quality of the marital relationship and the
likelihood of divorce in addition to affecting children’s outcomes. We can gauge the
importance of this type of shock to some extent because households report the most
important shocks in the year prior to the survey, including crop loss, illness, and business
loss. When we check whether such shocks in period t – 1 are correlated with divorce in
period t, we find no correlation (table 4), which is reassuring. When we control for the
point estimates of the main coefficients of interest with and without such shocks in our
empirical work, we note that they are very similar.
Taken together, the above evidence helps to build confidence in the quasi-random
nature of the father’s absence. We reiterate that we control for time-invariant unobserved
household-level characteristics and time-varying observed characteristics (including
shocks) throughout the empirical analysis.
<>Attrition
Overall, approximately 5 percent of households left the sample between the
baseline survey and the first follow-up, and an additional 8.5 percent of households left
between the first and second follow-ups (3.5 years after the baseline). 14 Although this
attrition rate is relatively low, 15 it is a concern if the reason for leaving the sample is
related to the behavior being modeled, as might be the case if households from which the
14
father departs are more likely to drop out of the sample. To address this concern, we
compare the baseline characteristics of households that did and did not subsequently
leave the sample (table 5). As expected, households that own a house or that do not live
at high altitudes are significantly less likely to abandon the study than those who do not
own a house or who live at high altitudes. Other than these factors, attrition is not
systematically related to any of the variables considered in the table. Although this is
reassuring, potential selection biases on the basis of unobserved characteristics cannot be
ruled out, for which we account in our empirical work. The methods we use to correct for
this are discussed in section IV. All results presented take into account this possible
selection problem, although it makes little difference to the effects that we estimate.
<>IV. EFFECTS OF PERMANENT ABSENCE OF THE FATHER ON SCHOOLING AND WORK
In this section, we present the empirical specification used to estimate the effects of the
permanent absence of the father on children’s schooling and work outcomes. We then
present our empirical findings.
<>Main Specification
To estimate the effects of the permanent absence of the father on children’s
school and work participation, we estimate the following model:
(1) ′ α 3 + I ′jt −1α 4 + f j + δ t + uijt
yijt = α1 + α 2 D jt + X ijt
,
where i denotes the child; j denotes the household; t denotes time, where t = 1 denotes
baseline, t = 2 denotes the first follow-up, and t = 3 denotes the second follow-up; yijt is a
discrete indicator for participation in school or work; and Djt is an indicator that takes the
value one if the father is absent from the household permanently and zero otherwise.
15
Note that Dj1 = 0, by definition. 16 If the father departed the household between the
baseline and the first follow-up, then Dj2 = 1 and Dj3 = 1; if the father departed between
the first and second follow-ups, then Dj2 = 0 and Dj3 = 1. Xijt is a vector of observed time-
varying child and household characteristics, including a cubic in the age of the child and
in the number of siblings in different age categories (0–6, 7–12, 13–17, 18+); Ijt – 1 is a
vector of time-varying shocks that occurred in the year prior to the survey, including
dummies for crop losses, business losses, and illnesses; fj is a household fixed effect
capturing the effects of unobserved time-invariant household characteristics; δt is a
survey-round dummy; and uijt is an error term that we assume to be independent and
identically distributed. The coefficient of interest is α2, which is the effect of the absence
of the father on the outcome (school or work participation).
We estimate equation (1) using a linear probability model and cluster the standard
errors at the municipality level to adjust for potential correlations of household decisions
within the same municipalities. Although the dependent variable is discrete, the main
advantage of the linear model over discrete choice models in our case is that it is
considerably easier to incorporate fixed effects. Most of the explanatory variables are
discrete and take on only a few values in our application, which strengthens the case for
the linear probability model (Wooldridge 2002, chapter 15). Although a potential
limitation of the linear probability model is that it may yield predicted probabilities
outside the unit interval, this is not a concern in our case because less than 3 percent of
predictions lie outside the unit interval. We verify the robustness of our results to this
linear specification by estimating a fixed-effects logit model (Honoré 2002). Although
the estimates are less precisely estimated because they are based on the subset of children
16
who changed their activity over time, they indicate the same patterns of coefficients that
are discussed in the main text on the basis of linear probability models. These are shown
in table S2 in the appendix.
As discussed in section III, an important issue is that our variable of interest, the
father’s permanent absence, may be correlated with unobserved household characteristics
that have a direct effect on children’s schooling and work. To determine the effects of
unobserved characteristics that are fixed over time and that may lead to spurious
correlations between the father’s permanent absence and children’s outcomes, we use a
household fixed-effects model. We also control for important time-varying shocks to
mitigate concerns that shocks may determine both paternal absence and the child
outcomes.
A second issue, which is also discussed in section III, is that if nonrandom
attrition is present, it will yield inconsistent parameter estimates. We use a standard
correction in a two-step sample selection model (Heckman 1979) and estimate the
probability that the individual will not leave the survey by using a probit model:
(2) ′ −1β3 + η j + δ t + vijt
Pr(Sijt = 1) = β1 + β 2 Z jt −1 + X ijt
,
where Sijt takes the value one if child i from household j does not leave the survey
between wave t – 1 and wave t and zero otherwise; Zjt – 1 are the instruments used for
identification, discussed below; Xijt – 1 are individual and household characteristics at
wave t – 1; δt is a survey round dummy; ηi is a household-level fixed effect, which may
be correlated with fj in equation (1); and νijt is an error term.
17
The instrument set Zjt – 1 includes characteristics of the previous interview,
including the date (day of the month) of the interview and whether the survey respondent
was the household head or spouse. Both may affect the overall experience of the
interview and the respondent’s willingness to be reinterviewed, but they are unlikely to
affect the outcomes of interest because they relate to the previous interview, which
occurred at least a year earlier. 17 The estimates from equation (2) are shown in table S3 in
the appendix. The instruments are jointly statistically significant at the 1 percent level.
We use these estimates to construct the inverse Mills ratio, which is appended to the set
of control variables in equation (1). The selection correction term is not statistically
significant at conventional levels in most cases, and the estimates change very little when
it is included in equation (1). Nonetheless, all reported results take into account this
selection correction.
<>Results
We next turn to the estimates from our equation of interest, equation (1), which
are shown in table 6. Because we observe work (schooling) for children aged 10 years (7
years) and above (see section II), we include an additional column containing estimates
for schooling for the subsample aged 10 years and above to enable meaningful
comparisons between the estimates for work and schooling outcomes. 18 We observe in
column 3 that the permanent absence of the father from the household significantly
increases participation in work by approximately 3 percentage points. Notably, we
observe in column 2 that the increase in work comes entirely from schooling (and not
leisure) because the absence of the father has a significant negative effect on school
enrollment of nearly 5 percentage points. 19 Note that the effects on schooling for the full
18
sample, which are shown in column 1, are similar to those for the restricted sample. The
estimated effects are not significantly different by gender (columns 4–6).
An important reason why the negative effects on schooling and positive effects on
work may be expected, which is discussed in section II, is that households in which the
father leaves permanently incur a substantial income reduction. To investigate the extent
to which the income loss associated with the absence of the father underlies the estimated
impacts, we interact it with the education of the household head (at the baseline, that is,
predeparture), which is a proxy for household income. On the one hand, households with
relatively less educated heads have less to lose from a departure through an “income
effect.� 20 On the other hand, the relatively less well off are more likely to face credit
constraints and insurance market failures and are more likely to have fewer formal ways
to mitigate the effects of income loss; thus, they are likely to suffer more from the
father’s absence. Accordingly, the interaction effect may run in both directions. We test it
empirically in columns 7–9. We observe that the detrimental effects of the father’s
absence on schooling and child labor are driven by relatively less educated households.
This finding highlights the importance of liquidity constraints for these households,
which dominates the effect of the loss of relatively less income with the father’s
departure.
Finally, we determine whether the effects vary depending on the reason for the
father’s absence by allowing the effects of death and divorce to be different. A caveat is
that the incidence of death is very low, affecting just 1 percent of our sample of
households (compared with 4.6 percent for divorce), which results in imprecise
estimations of its effects. The results (available upon request) show that the effects appear
19
to be driven primarily by divorce, although we cannot reject that the coefficient estimates
are statistically the same. In this study, we continue to pool these events because we are
interested in events that induce permanent income reductions. Another reason for
combining these events is to maintain statistical power given their rarity.
<>V. DO CCTS HELP PROTECT CHILDREN?
In this section, we investigate whether the effects of the father’s permanent
absence on children’s outcomes differ depending on whether a CCT program is in effect.
We begin by describing the CCT program. We determine whether the absence of the
father due to divorce is affected by the CCTs and find no evidence that it is. We then
estimate whether CCTs mitigate the adverse effects of the father’s permanent absence.
<>The CCT Program
To evaluate the impacts of the Familias en Acción CCT program, a representative
stratified sample of municipalities was selected. Strata were defined in terms of region
and by using an index of infrastructure relating to health and education. Municipalities
from the same strata that were excluded from receiving CCTs (but that were similar to
eligible municipalities in terms of population, area, and a quality-of-life index) were
chosen as controls. 21 A total of 122 municipalities were chosen for the evaluation. Of
these, 70 that were eligible for CCTs received the CCTs, which were phased in during the
period that we are considering; 26 received CCTs by the time of the baseline survey
(“early-treat�); 31 received CCTs by the first follow-up (“mid-treat�); and 13 received
CCTs by the second follow-up (“late-treat�). The final evaluation sample comprised
approximately 100 households that were randomly selected in each of the 122
20
municipalities. Attanasio et al. (2010) provide an evaluation of the program’s main
impacts.
<>The CCT Program and Divorce
Before studying the interaction between CCTs and the permanent absence of the
father, we address the potential concern that the absence of the father—particularly in the
case of divorce—may be affected by CCTs. Indeed, there is direct evidence of positive
effects of the PROGRESA CCTs on the divorce rate in Mexico (Bobonis 2011) and
indirect evidence that the Familias en Acción CCTs may have increased the bargaining
power of women (Attanasio, Battistin, and Mesnard 2012). Therefore, women receiving
CCTs might be expected to transit more readily out of relationships.
To estimate the effect of CCTs on divorce, we use data from the first and second
follow-ups only (because there is no variation in the outcome of divorce at the baseline;
see note 3). 22 We estimate the following regression at the household level on our sample
of households:
(3) y jt = α 0 + α1T jt + X ′jtα 2 + I ′jt −1α 3 + f j + δ t + u jt
,
pooling t = 2 and t = 3, where yjt is a dummy variable indicating whether the parents
living in household j divorced between periods t – 1 and t and Tjt is an indicator equal to
one if household j lives in a municipality that is receiving CCTs at time t and zero
otherwise. Note that Tjt = 1 for {(early-treat = 1 or mid-treat = 1) and t = 2,3} and for
{late-treat = 1 and t = 3}. Xjt are time-varying measures of the composition of children in
the household in period t; Ijt – 1 is a vector of dummies indicating whether the household
21
experienced a crop loss, business loss, or illness in period t – 1; fj is a household fixed
effect; δt is a survey round dummy; and ujt is an error term.
We cannot reject the hypothesis that the cash transfers have no statistically
significant effect on divorce (table 7). 23 It is thus unlikely that such an effect underlies the
results we discuss next, which show that receiving CCTs compensates for a father’s absence.
<>Interaction Effects
There is a growing body of literature on the safety net role played by CCTs in the
presence of income shocks. To our knowledge, however, no work has examined the role
of CCTs in alleviating the risk entailed by permanent loss of income. To investigate
whether the effects of the father’s permanent absence on children’s outcomes differ
depending on whether CCTs are in place, we augment equation (1) to include an
interaction between our variable of interest, father’s permanent absence, and receiving
CCTs. We thus estimate the following model:
(4) ′ α 5 + I ′jt −1α 6 + f j + δ t + uijt
yijt = α 0 + α1 D jt + α 2 D jt * T jt + α 3T jt + X ijt
,
where Tjt is equal to one if household j lives in a municipality that is receiving CCTs at
time t and zero otherwise; all other notations are as defined in equation (1). As before, Tjt
reflects the gradual rollout of the program, such that Tjt = 1 for {early-treat = 1 and t =
1,2,3}, {mid-treat = 1 and t = 2,3}, and {late-treat = 1 and t = 3}.
In equation (4), the coefficient of interest, α2, measures the extent to which
receiving CCTs mitigates the effect of the permanent absence of the father, α1. 24 Note
that the above specification also implicitly controls for preprogram differences in
22
outcomes across municipalities that are and are not eligible for CCTs (through fixed
effects), which is potentially important because of the quasi-experimental setting.
In municipalities that are not receiving CCTs, the permanent absence of the father
reduces the likelihood of school enrollment and increases the likelihood of child labor,
particularly among relatively less educated households (the left-hand columns of table 8).
This is indicated by the coefficient α1 (displayed in the first row), which estimates the
effect of departure in the absence of CCTs. The second row, α2, indicates that when
CCTs are in place, these adverse effects are offset (as shown by α1 + α2, which is close to
zero and not significantly different from zero, as indicated by the P values of the test). 25
Finally, as a robustness check, we restrict the comparison to households living in
municipalities eligible for the CCTs that fall within the common support (that is, the
region in which treated individuals have a counterpart in the group of controls according
to the propensity score). Consistent with Attanasio et al. (2010), we do this by matching
treatment and control observations using kernel-weighted propensity score matching and
by imposing common support by dropping the 10 percent of the treatment observations in
which the propensity score density of the control observations is the lowest. The results
are qualitatively similar and are shown in table S4 of the appendix.
The fact that the welfare program provides insurance to protect very poor children
from the adverse consequences of a father’s permanent absence is perhaps not surprising
to the extent that the CCTs received represent a sizeable share of income for these
households (more than 20 percent of their monthly total consumption on average; see
Mesnard [2009]), and the drop in household labor earnings entailed by the father’s
departure is of a similar magnitude. Moreover, the welfare program is in place on a
23
permanent basis, which ensures that the insurance it provides will continue as long as the
child is enrolled in school. Notably, this result is somewhat different from that of De
Janvry et al. (2006), who show that PROGRESA did not prevent children from working
more following shocks related to unemployment and illness of the household head, in
addition to natural disasters in the community, although it fully protected their schooling.
Taken together, our results indicate the existence of credit and insurance market
imperfections with adverse implications for children, who play an important role in
cushioning the household against income losses entailed by the departure of fathers.
Although the psychological effects of a departing parent cannot be ruled out, we believe
that they are of secondary importance to the income loss channel. In particular, we have
no reason to believe that psychological impacts would be stronger among less educated
populations, and they do not lend themselves to explaining why CCTs would help to
mitigate such effects.
<>VI. CONCLUSION
This paper has investigated the link between the permanent absence of the father
from the household and the school enrollment and work participation of children in rural
Colombia. We find that the absence of the father decreases schooling by approximately 5
percentage points and increases participation in work by approximately 3 percentage
points. We provide evidence that these effects are mainly driven by households with
relatively less educated heads, which are the very poorest of the indigent households in
our sample. We show that receiving CCTs offsets these adverse consequences and offers
children a form of insurance when the father leaves the household permanently. This
24
finding also suggests that the income reduction associated with paternal absence, which
tightens the liquidity constraints of already very poor households, is the main mechanism
at play.
Our results have a number of important policy implications. First, they suggest
that credit and insurance market failures are potentially important in the context of rural
Colombia and can contribute to reducing the human capital accumulation of children.
Second, an event such as the permanent departure of the father has potentially important
consequences for the schooling and employment of children. This is particularly true for
households with relatively low levels of education that are vulnerable to permanent
income losses because of insurance market failures. Third, these adverse effects can be
offset by well-designed CCT programs targeted at very poor households, which, in the
case of Colombia, represent more than 20 percent of total household consumption on
average and are in place as long as the child is enrolled in school.
The last finding is the first of its type and offers an important agenda for future
work. An important question is whether the same mechanism holds for investments other
than schooling (such as children’s health and nutrition) and in other contexts and
environments. Another question is whether this mechanism should be taken into account
in the design of safety nets and in targeting them at single parents. Such a policy might
have the unintended consequence of promoting single parenthood. A final thought
involves the particular relevance of these findings for sub-Saharan Africa, which has
witnessed a dramatic rise in orphanhood as the result of the prevalence of HIV/AIDS,
with estimates suggesting that 12 percent of all children are orphaned (UNICEF 2006).
25
Families and communities have shared this burden, and it might be time to institute
government support to help households cope.
REFERENCES
Antman, Francisca M. 2013. “The Impact of Migration on Family Left Behind.�
Forthcoming in The International Handbook on the Economics of Migration,
eds. Amelie Constant and Klaus F. Zimmermann, Chapter 16. Edward Elgar.
Attanasio, Orazio, Emla Fitzsimons, Ana Gomez, Maria I. Gutiérrez, Costas Meghir, and
Alice Mesnard. 2010. “Child Education and Work Choices in the Presence of a
Conditional Cash Transfer Program in Rural Colombia.� Economic Development
and Cultural Change 58 (2): 181–210.
———, Erich Battistin, and Alice Mesnard. 2012. “Food and Cash Transfers: Evidence
from Colombia.� Economic Journal 122 (559): 92–124.
Beegle, Kathleen, Rajeev Dehejia, and Roberta Gatti. 2006. “Child Labor and
Agricultural Shocks.� Journal of Development Economics 81 (1): 80–96.
———, Joachim De Weerdt, and Stefan Dercon. 2006. “Orphanhood and the Long-Run
Impact on Children.� American Journal of Agricultural Economics 88 (5): 1266–
72.
———, Joachim De Weerdt, and Stefan Dercon. 2010. “Orphanhood and Human Capital
Destruction: Is There Persistence into Adulthood?� Demography, 47 (1): 163–80.
26
Bell, Clive, Ramona Bruhns, and Hans Gersbach. 2006. “Economic Growth, Education,
and AIDS in Kenya: A Long-Run Analysis.� World Bank Policy Research
Working Paper 4025. Washington, DC: World Bank.
Biblarz, Timothy J., and Greg Gottainer. 2000. “ Family Structure and Children’ s
Success: A Comparison of Widowed and Divorced Single-Mother Families.�
Journal of Marriage and Family 62 (2): 533–48.
Bobonis, Gustavo J. 2011. “The Impact of Conditional Cash Transfers on Marriage and
Divorce.� Economic Development and Cultural Change 59 (2): 281–312.
Case, Anne, I-Fen Lin, and Sara McLanahan. 2001. “Educational Attainment of Siblings
in Stepfamilies.� Evolution and Human Behavior 22 (4): 269–89.
———, Christina Paxson, and Joseph Ableidinger. 2004. “Orphans in Africa: Parental
Death, Poverty and School Enrolment.� Demography 41 (3): 483–508.
Cherlin, Andrew J., Frank F. Furstenberg, P. Lindsay Chase-Lansdale, Kathleen E.
Kiernan, Philip K. Robins, Donna R. Morrison, and Julien O. Teitler. 1991.
“Longitudinal Studies of Effects of Divorce on Children in Great Britain and the
United States.� Science 252 (5011): 1386–9.
Corak, Miles 2001. “ Death and Divorce: The Long-Term Consequences of Parental Loss
on Adolescents.� Journal of Labor Economics 19 (3): 682–715.
Dahl, Gordon, and Enrico Moretti. 2008. “The Demand for Sons.� Review of Economic
Studies 75 (4): 1085–120.
Dammert, Ana. 2007. “Child Labor and Schooling Response to Changes in Cocoa
Production in Rural Peru.� IZA Discussion Paper 2869. Bonn: IZA.
27
Dehejia, Rajeev, and Roberta Gatti. 2005. “Child Labor and Agricultural Shocks.�
Journal of Development Economics 81 (1): 80–96.
De Janvry, Alain, Frederico Finan, Elizabeth Sadoulet, and Renos Vakis. 2006. “Can
Conditional Cash Transfer Programs Serve as Safety-Nets in Keeping Children at
School and from Working when Exposed to Shocks?� Journal of Development
Economics 79 (2): 349–73.
Duryea, Suzanne, David Lam, and Deborah Levison. 2007. “Effects of Economic Shocks
on Children’s Employment and Schooling in Brazil.� Journal of Development
Economics 84 (1): 188–214.
Edmonds, Eric. 2006. “Child Labor and Schooling Responses to Anticipated Income in
South Africa.� Journal of Development Economics 81 (2): 386–414.
Ermisch, John F., and Marco Francesconi. 2001. “Family Structure and Children’s
Achievements.� Journal of Population Economics 14 (2): 249–70.
Evans, David, and Edward Miguel. 2007. “Orphans and Schooling in Africa: A
Longitudinal Analysis.� Demography 44 (1): 35–57.
Fiszbein, Ariel, Norbert Schady, Francesco Ferreira, Margaret Grosh, Niall Keleher,
Pedro Olinto, and Emmanuel Skoufias. 2009. Conditional Cash Transfers:
Reducing Present and Future Poverty. Washington, DC: World Bank.
Gennetian, Lisa. 2005. “One or Two Parents? Half or Step Siblings? The Effect of
Family Composition on Young Children.� Journal of Population Economics 18
(3): 415–36.
28
Gertler, Paul, David Levine, and Minnie Ames. 2004. “Schooling and Parental Death.�
Review of Economics and Statistics 86 (1): 211–25.
Giannelli, Gianna, and Lucia Mangiavacchi. 2010. “Children’s Schooling and Parental
Migration: Empirical Evidence on the ‘Left-Behind’ Generation in Albania.�
Labour 24 (S1): 76–92.
Ginther, Donna, and Robert Pollak. 2004. “ Family Structure and Children’ s Educational
Outcomes: Blended Families, Stylized Facts, and Descriptive Regressions.�
Demography 41 (4): 671–96.
Gitter, Seth, James Manley, and Brad Barham. 2011. “The Coffee Crisis, Early
Childhood Development, and Conditional Cash Transfers.� IDB Working Paper
245. Washington, D.C.: Inter-American Development Bank.
Gruber, Jonathan. 2004. “Is Making Divorce Easier Bad for Children? The Long-Run
Implications of Unilateral Divorce.� Journal of Labor Economics 22 (4): 799–
833.
Guarcello, Lorenzo, Fabrizia Mealli, and Furio Rosati. 2010. “Household Vulnerability
and Child Labor: The Effects of Shocks, Credit Rationing, and Insurance.�
Journal of Population Economics 23 (1): 169-198.
Gubert, Flore, and Anne Sophie Robilliard. 2008. “Risk and Schooling Decisions in
Rural Madagascar: A Panel Data-Analysis.� Journal of African Economies 17 (2):
207–38.
Heckman, James. 1979 “Sample Selection Bias as a Specification Error� Econometrica:
Journal of the Econometric Society, 153–161.
29
Honoré, Bo E. 2002. “Nonlinear Models with Panel Data.� Portuguese Economic Journal
1: 163–179.
Jacoby, Hanan, and Emmanuel Skoufias. 1997. “Risk, Financial Markets, and Human
Capital in a Developing Country.� Review of Economic Studies 64 (3): 311–35.
Lahaie, Claudia, Jeffrey Hayes, Tinka Piper, and Jody Heymann. 2009. “Work and
Family Divided across Borders: The Impact of Parental Migration on Mexican
Children in Transnational Families.� Community, Work and Family 12 (3): 299–
312.
Lang, Kevin, and Jay Zagorsky. 2001, “ Does Growing Up with a Parent Absent Really
Hurt?� Journal of Human Resources 36 (2): 253–73.
Maluccio, John. 2005. “Coping with the ‘Coffee Crisis’ in Central America: The Role of
the Nicaraguan Red de Protección Social.� FCND Discussion Paper 188.
Washington, DC: International Food Policy Research Institute (IFPRI).
Mesnard, Alice. 2009. “Migration, Violence and Welfare Programs in Rural Colombia.�
IFS Working Paper 09/19. London: Institute for Fiscal Studies.
Painter, Gary, and David Levine. 2000. “Family Structure and Youths’ Outcomes: Which
Correlations are Causal?� Journal of Human Resources 35 (3): 524–49.
Parker, Susan W., and Emmanuel Skoufias. 2006. “Job Loss and Family Adjustments in
Work and Schooling during the Mexican Peso Crisis.� Journal of Population
Economics 19 (1): 163–81.
30
Piketty, Thomas. 2003. “ The Impact of Divorce on School Performance: Evidence from
France, 1968–2002.� CEPR Discussion Paper 4146. London: Centre for
Economic Policy Research.
Ravallion, Martin, and Quentin Wodon. 2000. “Does Child Labour Displace Schooling?
Evidence on Behavioural Responses to an Enrollment Subsidy.� Economic
Journal 110 (462): C158–75.
UNICEF. 2006. Africa’s Orphaned and Vulnerable Generations: Children Affected by
AIDS. New York, NY: UNICEF.
Van de Walle, Dominique. 2011. “Lasting Welfare Effects of Widowhood in a Poor
Country.� World Bank Policy Research Working Paper 5734. Washington, DC:
World Bank.
Wooldridge, Jeffrey. 2002. Econometric Analysis of Cross Section and Panel Data.
Cambridge, MA: MIT Press.
Yamano, Takashi, and Thomas S. Jayne. 2005. “Working-Age Adult Mortality and
Primary School Attendance in Rural Kenya.� Economic Development and
Cultural Change 53 (3): 619–54.
31
TABLE 1. School and Work Participation, by Age, Survey and Gender
Males Females
Age at Baseline First Second Baseline First Second
(+3.5 (+3.5
baseline (+1 yr) yrs) (+1 yr) yrs)
School School
enrolment enrolment
% %
7 0.904 0.928 0.963 0.922 0.953 0.970
8 0.935 0.951 0.933 0.961 0.959 0.947
9 0.952 0.943 0.895 0.966 0.960 0.918
10 0.932 0.907 0.813 0.958 0.950 0.867
11 0.917 0.884 0.764 0.935 0.901 0.835
12 0.856 0.782 0.675 0.897 0.859 0.786
13 0.791 0.755 0.577 0.832 0.791 0.633
14 0.660 0.620 0.457 0.740 0.728 0.536
N 6090 5726 5033 5589 5266 4482
Work participation Work participation
% %
10 0.021 0.019 0.083 0.010 0.006 0.045
11 0.031 0.050 0.132 0.012 0.025 0.057
12 0.057 0.093 0.209 0.029 0.042 0.117
13 0.109 0.148 0.284 0.057 0.086 0.143
14 0.213 0.285 0.371 0.091 0.169 0.203
N 3672 4233 5022 3265 3870 4480
Notes: Work includes full-time paid and unpaid activities and looking for work as a
main activity. Figures in bold (italics) denote ages corresponding to post-compulsory
schooling. Note that +1 yr (3.5 yrs) means 1 yr (3.5 yrs) after the baseline survey. N
denotes the number of individuals (aged 7-14 at baseline) present in the survey listed
at top of column. Schooling observed for children aged ≥7; work observed for children
aged ≥10 years.
32
TABLE 2. Comparison of Baseline (Predeparture) Characteristics across
Households That Do and Do Not Experience Subsequent Departure
Permanent absence of father
Characteristic, Baseline ↓ (D)
D=1 D=0 P value
Age of household head 42.88 42.21 0.182
Age of spouse 37.81 37.14 0.110
Education of head
None 0.282 0.230 0.015
Some (complete/incomplete primary) 0.535 0.638 0.000
High (incomplete secondary or more) 0.181 0.132 0.004
Education of spouse
None 0.199 0.195 0.846
Some (complete/incomplete primary) 0.633 0.661 0.238
High (Incomplete secondary or more) 0.168 0.144 0.171
Household composition
Avg. # of children ≤ 6 years 0.388 0.467 0.011
Avg. # of boys 7–11 years 0.727 0.738 0.775
Avg. # of girls 7–11 years 0.718 0.684 0.366
Avg. # of boys 12–17 years 0.635 0.641 0.890
Avg. # of girls 12–17 years 0.581 0.590 0.809
Avg. # of female adults 1.232 1.244 0.708
Avg. # of male adults 1.366 1.396 0.432
School enrollment rate of
7- to 14-year-olds in household 0.924 0.899 0.057
Household monthly consumption 421,286 441,994 0.085
Program area 0.700 0.682 0.453
Altitude 574.45 601.90 0.451
N 426 5,720
Source: Familias en Accion Evaluation Data.
Note: Sample consists of households in which both parents are present at the
baseline and there is at least one 7- to 14-year-old child. N denotes the number of
households at baseline. P values are based on standard errors clustered at the
municipality level. Figures in bold in column (4) indicate that the correlation of the
figures in columns (2) and (3) are significant at the 5 percent level or less.
33
TABLE 3. Common Trends: Schooling and Income
School enrollmenta
Year = 2002 0.0343**
(0.005)
Absence * Year = 2002 0.0238
(0.0143)
N1 11,679
Per capita incomeb
Year = 2000 0.498**
(0.111)
Year = 2001 1.1019**
(0.1432)
Absence * Year = 2000 −0.0845
(0.4641)
Absence * Year = 2001 0.6144
(0.6028)
N2 5,066
Source: Familias en Accion Evaluation Data.
1
Note: N is the number of children in the sample at the baseline with
nonmissing school enrollment data. N2 is the number of households in
the sample at baseline that report income retrospectively for 1999,
2000, and 2001. Standard errors, clustered at the municipality level, in
parentheses.
a
Dependent variable is school enrollment. Estimates are from a
household fixed effects model, controlled for quadratic in child age
and gender (female = 1). Reference year is 2001.
b
Dependent variable is per capita household labor income. Estimates
are from a household fixed-effects model. Reference year is 1999.
34
TABLE 4. Correlation between Divorce and Shocks in Previous
Period
Divorce (t)
Crop loss (t − 1) −0.0033
(0.0045)
Business loss (t − 1) 0.0205
(0.0181)
Illness (t − 1) −0.0061
(0.0071)
P value for joint significance 0.55
N 5,796
Source: Familias en Accion Evaluation Data.
Note: Dependent variable is divorce. Reference year is 2001. Estimates
are from a household fixed-effects model pooling first and second
follow-ups, controlled for child composition and time dummies. N is the
number of households remaining in the sample by first follow-up.
Standard errors, clustered at the municipality level, are in parentheses.
35
TABLE 5. Comparison of Characteristics across Households That Do and Do Not
Abandon the Study at Any Time after Baseline
Did Did
Baseline Characteristics not abandon abandon P value difference
Age of head 42.1123 42.3617 0.5057
Age of spouse 37.1413 37.4586 0.2987
Head no education 0.2309 0.2497 0.2263
Spouse no education 0.1927 0.2120 0.1889
Head some education 0.6321 0.6208 0.5251
Spouse some education 0.6579 0.6655 0.6668
Head high education 0.1365 0.1260 0.4047
Spouse high education 0.1493 0.1225 0.0398
Treated area 0.6844 0.6756 0.6064
Altitude 577.69 726.43 0.0000
Crop loss at first survey 0.1339 0.1249 0.4708
0.0000
Owns house 0.6466 0.5321
N 5,289 857
Source: Familias en Accion Evaluation Data.
Note: Sample consists of households in which both parents are present at the baseline
and there is at least one 7- to 14-year-old child. N denotes the number of households at
baseline. P values are based on standard errors clustered at the municipality level. Figures
in bold in column (4) indicate that the correlation of the figures in columns (2) and (3) are
significant at the 5 percent level or less.
36
TABLE 6. Marginal Effects of the Father’s Absence on Children’s Schooling and Work
School School Work School School Work School School Work
Overall Restricted Overall Restricted Overall Restricted
Permanent Absence −0.0412* −0.0484* 0.0301+ −0.0422+ −0.0486+ 0.0361 −0.0556** −0.0641** 0.0360+
(0.0172) (0.0216) (0.0167) (0.0233) (0.0289) (0.0233) (0.0193) (0.0234) (0.0183)
Permanent Absence * Girl 0.0021 0.0005 −0.0122
(0.0241) (0.0275) (0.0245)
Permanent Absence * Head High Education 0.0742* 0.0851* −0.0319
(0.0313) (0.0367) (0.0295)
N 32,186 24,531 24,531 32,186 24,531 24,531 32,186 24,531 24,531
Source: Familias en Accion Evaluation Data.
Note: Marginal effects are presented from a fixed-effects linear-probability model (equation (1)), controlling for absence of father from
household for unknown reason; absence of both parents; time dummies; cubic in child age; sibling composition; dummies for crop losses,
illness, and business shocks; and the inverse mills ratio, as computed in equation (2) (see table S3). High education is one for incomplete
secondary or more at baseline, zero otherwise. N is the number of children in the sample pooled across three waves. We included 6,146
households in our initial sample, from which 426 fathers have subsequently departed. Schooling is observed for all children in sample ≥7 years
(overall sample); work is observed for children ≥10 years (restricted sample). Robust standard errors clustered at the municipality level are in
parentheses. + significant at 10 percent; *significant at 5 percent; ** significant at 1 percent.
37
TABLE 7. Marginal Effects of CCTs on
Divorce
Pr (Divorce = 1)
CCTs 0.0024
(0.0082)
Time = 2 −0.0286
(0.0039)
N 5,796
Source: Familias en Accion Evaluation Data.
Note: Marginal effects are presented from
equation from a fixed-effects linear-probability
model (equation (3)), pooling first and second
follow-ups and controlling for child
composition; dummies for crop losses, illness,
and business shocks; and the inverse mills
ratio, as computed in equation (2) (see table
S3). N is the number of households that did not
abandon the study by the first follow-up.
Robust standard errors clustered at the
municipality level are in parentheses.
38
TABLE 8. Cushioning Effects of CCTs: Marginal Effects on Schooling and
Work
Low Education All
School School Work School School Work
Overall Restricted Overall Restricted
−0.103*
Permanent Absence (α1) * −0.114* 0.0782* −0.0801** −0.0919* 0.0694*
(0.0341) (0.0437) (0.0355) (0.0284) (0.0376) (0.0300)
Permanent Absence
* CCTs (α2) 0.0818* 0.0833+ −0.0687+ 0.0566+ 0.0615 −0.0542+
(0.0381) (0.0472) (0.0356) (0.0330) (0.0420) (0.0307)
−0.0257*
CCTs (α3) 0.0143 0.0057 −0.0262** 0.0121 0.0053 *
(0.0123) (0.0148) (0.0096) (0.0112) (0.0134) (0.00902)
Time = 2 0.1080 0.1520 −0.2010* 0.0783 0.1180 −0.1500+
(0.0957) (0.105) (0.0847) (0.0872) (0.0964) (0.0773)
Time = 3 0.0261 0.0689 −0.148+ −0.0025 0.0347 −0.1010
(0.1010) (0.1100) (0.0889) (0.0927) (0.1020) (0.0809)
P value (α1 + α2) = 0 0.3498 0.2636 0.6224 0.2633 0.2379 0.3854
N 28,027 21,464 21,464 32,186 24,531 24,531
Source: Familias en Accion Evaluation Data.
Note: Marginal effects are presented from a fixed-effects linear-probability model (equation
(4)), controlling for absence of father for unknown reason; absence of both parents; cubic in
child age; sibling composition; dummies for crop losses, illness, and business shocks; and the
inverse mills ratio, as computed in equation (2) (see table S3). “CCTs� indicates whether the
household lived in a municipality that was receiving CCTs at the time of survey. N is the number
of children in the sample pooled across three waves. We included 6,146 households in our
initial sample, from which 426 fathers have subsequently departed. Schooling is observed for all
children in sample ≥7 years ('overall' sample); work is observed for children ≥10 years
('restricted' sample). Robust standard errors clustered at the municipality level are in
parentheses. + significant at 10 percent; *significant at 5 percent; ** significant at 1 percent.
39
APPENDIX
TABLE S1. Marginal effects of paternal death and divorce on transfers received
by household
Institutional Monetary In-Kind
Death -1,317 16,809* 22,596
(8680) (7512) (14788)
Divorce 1,512 17,704** 8,484
(4294) (4106) (7967)
N 6,069
Notes: N is number of households at baseline for which we observe
transfers. Complete data on transfers missing for 77 of the sample of
6146 households. Pools baseline, first and second follow-ups. We trim the
top 1% of outliers in each period. Each column represents a separate
regression. Also control for household fixed effects, absence of father for
unknown reason, absence of both parents, time dummies, household
child composition, dummies for crop, illness and business shocks, inverse
mills ratio computed as in equation (2) (see Table A3). Robust standard
errors clustered at municipality level in parentheses. Robust standard
errors clustered at municipality level in parentheses. + significant at 10%;
*significant at 5%; ** significant at 1%.
40
TABLE S2. Marginal effects of the father’s absence on children’s schooling and work
Estimates from Conditional Logit Model
School School Work School School Work School School Work
Overall Restricted Overall Restricted Overall Restricted
Permanent Absence -0.507* -0.596* 0.550+ -0.488* -0.579* 0.661+ -0.555* -0.662** 0.515
(0.223) (0.248) (0.309) (0.245) (0.277) (0.349) (0.231) (0.250) (0.317)
Permanent Absence * Girl -0.0423 -0.0367 -0.241
(0.267) (0.300) (0.376)
Permanent Absence * High Educated Head 0.555 0.711 0.471
(0.847) (0.855) (1.007)
1
N 4284 3895 2536 4284 3895 2604 4284 3895 2604
2
N 1829 1661 1064 1829 1661 1075 1829 1661 1075
Notes: Marginal effects from a conditional logit model reported (equation (1)) with household fixed effects. Additional controls include
control for absence of father for unknown reason, absence of both parents, time dummies, cubic in child age, sibling composition, dummies
for crop, illness, and business shocks, inverse mills ratio computed as in equation (2) (see Table A3). High Educated is equal to 1 if incomplete
secondary or more at baseline, 0 otherwise. N1 (N2) is the number of children (households) in the sample that (contain a child that) switch
outcome status at least once, pooled across three waves. Non-switcher children drop out of the conditional likelihood function. Schooling
observed for all children in sample, i.e. ≥7 ('overall' sample); work observed for children ≥10 ('restricted' sample). Robust standard errors
clustered at municipality level in parentheses. + significant at 10%; *significant at 5%; ** significant at 1%.
41
TABLE S3. Probability of not leaving the sample, marginal effects
Dep vble=1 if stay in sample, 0
otherwise
Female 0.0057+
(0.0031)
Time = 2 0.0367**
(0.0071)
Owns house 0.0544*
(0.0213)
Urban 0.0075
(0.0069)
Day of month 1 -0.0478+
(0.0288)
Day of month 2 -0.0385
(0.0318)
Day of month 3 -0.0583+
(0.0349)
Day of month 4 -0.0287
(0.0309)
Day of month 5 -0.0011
(0.0228)
Day of month 6 -0.0410
(0.0265)
Day of month 7 -0.0050
(0.0246)
Day of month 8 -0.0593+
(0.0331)
Day of month 9 -0.0466
(0.0351)
Day of month 10 -0.0130
(0.0263)
Day of month 11 -0.0132
(0.0227)
Day of month 12 -0.0771*
(0.0392)
Day of month 13 -0.0645+
(0.0378)
Day of month 14 -0.0581
(0.0356)
Day of month 15 -0.0514
(0.0374)
Day of month 16 -0.0826*
(0.0361)
42
Day of month 17 -0.0311
(0.0335)
Day of month 18 -0.0165
(0.0290)
Day of month 19 -0.0524
(0.0366)
Day of month 20 -0.0584
(0.0414)
Day of month 21 -0.0471
(0.0361)
Day of month 22 -0.0319
(0.0323)
Day of month 23 -0.0429
(0.0341)
Day of month 24 -0.0435
(0.0322)
Day of month 25 -0.0194
(0.0234)
Day of month 26 -0.0439
(0.0364)
Day of month 27 -0.0456
(0.0308)
Day of month 28 -0.0135
(0.0249)
Day of month 29 -0.0034
(0.0234)
Day of month 30 0.0078
(0.0224)
Respondent = head 0.0132
(0.0188)
Respondent = spouse 0.0597*
(0.0272)
p-value of joint significance of
instruments 0.0000
N 11679
43
Notes: N is the number of children in the sample at baseline with non-
missing school enrolment data. Day of month = dummy variables for day
baseline interview took place. Robust standard errors clustered at
municipality level in parentheses. + significant at 10%; *significant at 5%;
** significant at 1%.
44
TABLE S4. Marginal effects of the father’s absence on children’s
schooling and work, common support only
School Work
Overall Restricted
Permanent Absence (α1) -0.102** -0.106* 0.0682+
(0.0373) (0.0477) (0.0367)
Permanent Absence * CCTs1 (α2) 0.0881* 0.0829 -0.0658+
(0.0398) (0.0506) (0.0368)
CCTs1 (α3) 0.0145 0.006 -0.0224*
(0.0125) (0.0146) (0.0103)
Time = 2 0.121 0.17 -0.222*
(0.101) (0.11) (0.087)
Time = 3 0.0371 0.0856 -0.167+
(0.106) (0.114) (0.091)
N 24982 18972 18972
Notes: See notes to Table 8. Note further that we match treatment and control
observations using kernel-weighted propensity score matching, and impose
common support by dropping 10% of the treatment observations at which the
propensity score density of the control observations is the lowest.
45
NOTES
1 Note that departure of the mother is also an important issue and may have
different effects from those stressed in this paper. However, there is insufficient variation
in the data to allow us to examine this issue.
2 An absent but living father can visit and influence the children’s upbringing in a
way that a deceased father obviously cannot. However, relations with the absent parent’s
family may be very different in the two cases and may be more supportive in cases of the
early death of the father than in cases of acrimonious separation. Moreover, transfers
from the father or in-laws may compensate in different ways depending on the reason for
the departure. However, in our data, transfers are very similar regardless of whether
departure is caused by death or divorce, as shown in table S1 in the appendix.
3 We begin with a sample of households in which both parents are married; thus,
by definition, fathers are all present at baseline.
4 This sample selection criterion means that we retain 9,187 out of 11,502
households. The reason we do not keep single-parent households is that the departure of
the father (if present) in such households would raise additional issues that would be
difficult to disentangle.
5 School enrollment is defined on the basis of whether the child is registered at
school in the academic year corresponding to the survey. Work participation is equal to
one if the child’s main activity in the week before the survey is reported to be any work,
46
household chores (paid and unpaid), or looking for work. We note that our main results
are similar if we exclude unpaid household chores.
6 The school system in Colombia operates as follows. Compulsory education is
free and lasts for nine years; it consists of basic primary (educación básica primaria, five
years, ages 7 through 11) and basic secondary (educación básica secundaria, four years,
ages 12 through 15). The secondary school system includes the middle secondary cycle
(educación media, two years, ages 16 and 17). Successful completion of studies leads to
the Bachillerato. Students must pass an entrance examination for access to universities.
7 Note that absent fathers are not being “replaced� in households, at least in the
3.5-year span of our surveys. Whereas the number of male adults is lower by almost one
in households that experience departure, the number of female adults and children is the
same.
8 We also verified that, considered separately, they do not have significantly
different effects (see section IV).
9 In an additional 1 percent of households, both the father and mother have left
the household for an unknown, possibly temporary, reason. There is also a small
percentage (1.2 percent) of households in which the father has left for an unknown
reason, but the mother has remained in the household and reports being married, so we
assume that these are temporary departures. These are not the main variables of interest,
but we control for them throughout the analysis.
10 If we do not control for adult composition, the difference is larger,
approximately 34 percent, which we would expect because the departure of the father
47
decreases the number of adults in the household. We see this as a lower bound of the
magnitude of the departure effect in terms of total household adult earnings because it
includes labor supply responses, which are likely to cushion the potential adverse effects
on income. This figure excludes earnings from children to mitigate this problem.
11 Table S1 in the appendix shows that the total value of additional transfers
received by the households after paternal departure (institutional, monetary, and in-kind)
is less than 50,000 pesos per annum compared to an average monthly total household
consumption in excess of 420,000 pesos at the baseline. Nonetheless, these responses by
extended family or friends may contribute to explaining why household consumption
does not drop as much as household labor income, as noted above.
12 However, if this were the case, then we would expect to see a positive impact
on children, whereas we observe the contrary. It must be acknowledged that the departure
of the father because of death may not be a random event, although this is much less of a
concern.
13 In the lower panel of table 3, the coefficient associated with the interaction of
the absence of the father with the year of survey 2000 (which is negative) is very small in
magnitude and has a very large standard error; it is therefore of no concern.
14 Attrition at the individual level is extremely rare, less than 1 percent.
15 This is comparable to the attrition rate of 6 percent between the baseline and
follow-up surveys for the evaluation of the Bono de Desarrollo CCT program in Ecuador,
which is considered “low,� and just under the attrition rate of 15 percent over four years
in Nicaragua for the evaluation of the Red de Protección Social CCT program, which is
48
considered “reasonably low.� It is slightly higher than the rate for the PROGRESA
program, which was approximately 6 percent over the first three years of the program and
is considered “very low� (Fiszbein et al. 2009).
16 As discussed in section II, our sample is restricted to households in which the
father is present at the baseline. We only observe departures after the baseline.
17 Attrition in our sample is predominantly at the household level. Moreover,
very few households (3.7 percent in the entire sample) have migrated out of their village
of residence, and additional resources have been invested to track them (Mesnard 2009).
Attrition is thus mostly due to nonwillingness to answer.
18 We retain estimates for the full sample to improve statistical power.
19 This suggests that child labor and schooling are strong substitutes, in contrast
to the findings of Ravallion and Wodon (2000) indicating that increases in schooling in
Bangladesh following a welfare program are only partially due to decreased child labor.
20 Similarly, if paternal education is positively correlated with paternal quality as
a figurehead/role model, then one would expect the loss of a highly educated father to
involve the loss of a more positive impact on the child’s life.
21 To be eligible to qualify for the program, municipalities had to satisfy the
following four criteria: (i) have fewer than 100,000 inhabitants and not be a departmental
capital, (ii) have basic education and health infrastructure, (iii) have a bank, and (iv) have
relatively up-to-date welfare lists at the municipality administrative office. The
evaluation design was undertaken by a consortium led by the Institute for Fiscal Studies
and included the authors of this paper.
49
22 This also means that the identification of the effect of the CCTs on divorce
comes from the rollout of the program to late-treat areas at t = 3.
23 As an additional check, we compared the characteristics of households that
divorced across areas eligible for the CCTs and control areas and found them to be
similar for both types of areas.
24 Note that because of the gradual phasing in of the CCTs, the early-treat
municipalities do not contribute to identifying α3, the impact of the CCTs, because there
are no preprogram data for these municipalities. However, we retain them in the analysis
because they contribute to identifying α1 and α2.
25 The table also shows that the effect of the CCTs on children in our sample who
are not affected by paternal absence, given by α3, is to increase school enrollment and
reduce child labor. Although the effect on schooling is not significant, this is most likely
because the early-treat municipalities do not contribute to the identification of the CCT
effect, as opposed to the findings of Attanasio et al. (2010), which provide a general
analysis of the impacts of the program.
50