Introduction Summaries The central theme
The central theme of this dissertation is the enduring effects of early life experiences. It
is a collection of three essays on the effect of parental income and labor supply decisions on
children’s health and their labor market outcomes after becoming adults. The second
chapter investigates the long-run effect of childhood poverty on labor market outcomes.
The third chapter examines the child health effect of the minimum wage introduced by the
1966 Fair Labor Standards Act. The fourth chapter explores health inequality resulting
from maternal work intensity differences.
Chapter 2 is motivated by the two overlooked issues in studying the long-run effect of
childhood poverty. First, the proportion of the total effect of childhood poverty on adult
labor market outcomes that passes through mediator covariates has barely been studied.
Second, many studies considered childhood as a single period and neglected the differences
in stages of child development (infancy, early childhood, mid-childhood, and late childhood)
for acquiring specific skills. Separate analyses for each stage enable us to identify the most
critical childhood period.
Chapter2aimstobridgethesegapsusingthePanelStudyofIncomeDynamics(PSID) data. The
study employs propensity score matching (PSM) and average causal mediation effect
(ACME) methods to provide empirical evidence for the overlooked issues. The total effect
estimation results show that persistent childhood poverty significantly decreases average
annual labor income (by $11,252) and hours worked (by 182.1 hours), whereas it increases
the 2
number of weeks unemployed per annum (by 1.96 weeks). The annual labor income causal
mediation estimate shows that 42.25 percent of the total effect of childhood poverty passes
indirectly through the negative influence of childhood poverty on education and childhood
health. Similar figures for hours worked and the number of weeks unemployed are 22.57
and 21.73 percent, respectively. The PSM and ACME analyses for stages of child
development revealed that early childhood is the most critical period.
Inchapter3,themotivationistoinvestigatetheimpactoftheminimumwagepolicybeyond its
effect on the individuals currently participating in the labor market. Prior literature on
minimumwagemainlyfocusesonexaminingtheeffectofminimumwagechangeonemployment
and earnings of labor market participants. However, the impact transcends the labor
market outcome performances of participants. Due to its effect on parents’ income, time
allocation, and health investment capacity, minimum wage change also impacts child health.
Using the PSID’s intergenerational survey data, chapter 3 examines the effect of the
minimum wage introduced by the 1966 Fair Labor Standards Act on child health. To obtain
estimates, I use a difference-in-differences research design. The treatment group is
comprised of individuals whose parents were working in the industries covered by the
1966 minimum wage extension. The control group includes individuals whose parents
worked in the industries covered by the 1938 Fair Labor Act. The findings show that the
1966 minimum wage extension significantly and positively affects child health. The results
from estimation using the overall sample and
cohortandstatefixedeffectshowthattheminimumwageintroductionincreasestheproportion
of children who have a healthy childhood. The positive child health impact of the minimum
wageextensionisnothomogeneousacrossallstates. TheFairLaborStandardsActwasalarger
shock for states with no minimum wage before the Act. The estimation results, which
consider treated individuals only from states which had no minimum wage before the Act,
show that the child health improvement impact of the minimum wage policy is stronger
than the effect based on the overall sample. The effect is also heterogeneous for females and
males. Therefore,
whenimplementingpoliciesdesignedtoenhancechildren’shealththroughincomeimprovement
programs for parents, it is essential to consider the intervention’s varying impact on the
health outcomes of male and female children. Additionally, it is crucial to note that the effect
of the federal minimum wage policy on child health outcomes is more significant in states
without a state minimum wage policy.
Thelaborforceparticipationofworkingmothershasincreaseddramaticallyoverthelastsix
decades (the U.S. BLS, 2022). Chapter 4 is motivated by this growing trend of mothers’ labor
forceparticipationanditspotentialimpactonthehealthofupcominggenerations. Thechapter
examines the effect of maternal work intensity, measured by average annual hours worked,
on child health. The empirical analysis is for both the entire childhood period, from birth to
17 years of age, and explores how the effect of maternal work intensity varies across
different stages of child development. The study employs the PSID’s intergenerational
survey data and instrumental variable (IV) probit regression estimation technique using
state-level women’s labor force participation as an instrumental variable. The results show
that maternal hours
workedhasapositiveandsignificanteffectonthelikelihoodofhavingahealthychildhood. The
fixed effect estimation that accounts for the heterogeneity across stages of child
development (infancy,earlychildhood,mid-
childhood,andlatechildhood)showsthatmaternalhoursworked has a negative and significant
effect on childhood health for the first two stages (infancy and
earlychildhood)ofchilddevelopmentandapositiveandsignificanteffectforthelatechildhood
period. Policies like maternal leave, childcare subsidies, and welfare for work exemption of
mothers of infants, which aim to improve the well-being of children through mothers’
payoffs from labor market involvement, should consider the heterogeneity of the effect
across stages of child development.
Chapter 5 summarizes the findings of the three chapters that examine different
childhood circumstances and their long-term effect and draws conclusions.
CHAPTER 2
CHILDHOOD POVERTY AND LABOR MARKET
OUTCOMES: THE MEDIATING ROLE OF HEALTH AND
EDUCATION
2.1 Introduction
Millions of Americans live on incomes below the federal poverty line, even though the
U.S. is a relatively wealthy country by international standards. In 2019, the U.S. Census
Bureau reported that 38.1 million people (or 11.8 percent of the total population) were
living in poverty. Moreover, wealth deprivation is unevenly distributed across different age
groups, with the most severe effect manifested in the lowest age group. The same report
stated that 16.2 percent of people under the age of 18, (or 1 child out of 6) live in poverty,
whereas for the adult counterpart, the ratio is 1 adult out of 9. Hence, poverty analysis
requires special consideration for children. Furthermore, childhood poverty calls for special
deliberation since its effect extends to adulthood and manifests in different forms:
educational attainment, health status, labor force participation, and earnings. This study
aims to investigate the effect of childhood poverty on late-life labor market outcomes.
Prior studies show the negative effect of childhood poverty as children mature into
adulthood. Many children with disadvantaged backgrounds struggle academically and do
not complete high school (Chaudry & Wimer, 2016; Dahl & Lochner, 2012; Duncan et al.,
1998, 2011; Kubilius & Corwith, 2018; Levy & Duncan, 2000). These children only obtain
spotty employment and low earnings as young adults (Gregg et al., 1999; Lesner, 2018;
Ratcliffe & McKernan, 2010, 2012). As adults, they experience deteriorating physical and
mental health due to their parents’ cumulative underinvestment in their healthcare (Currie
& Almond, 2011; Duncan et al., 1998; Levy & Duncan, 2000; Ziol-Guest et al., 2012). The
combined negative consequences of childhood poverty lead to low-income adulthood and a
higher average probability of remaining poor (Bellani & Bia, 2019; Wagmiller & Adulman,
2009).
In these studies, however, two critical issues have been overlooked. First, researchers
barely studied the role of mediator covariates (health and education) and what proportion
of the total effect of childhood poverty on adult outcomes passes through these mediator
covariates.
Second,thesestudiesfocusonoverallchildhoodpovertyasasingleperiodbutneglecttheeffects of
childhood poverty that occurs at different stages of childhood development.
The California Department of Education (2000) reported that children’s physical and
mental development have their own critical periods. This gives some insight into the fact
that poverty at different stages of child development (infancy, early childhood, mid-
childhood, and late childhood) may have different immediate and long-term impacts. The
poverty that occurs during early childhood may not have a similar effect as poverty that
occurs during late childhood. For example, Conti and Heckman (2012) showed that
parental investment in children’s education to learn a second language has better efficiency
before age 12. Nevertheless, no study, at least in the major literature, has separately
analyzed the heterogeneous effects of childhood poverty, which occurs at different stages of
child development, on adult outcomes.
Furthermore, children with a similar type of economically disadvantaged background
may achieve different levels of success during adulthood. Ratcliffe (2015) found that not all
impoverished children have poor young adult outcomes. This indicates that mediator
complementary factors that occur between infancy and late adulthood play a significant
role
6
in success later in life. The studies surveyed in this paper show separately the negative
effect of childhood poverty on health, education, and different adult outcomes. However, no
study, at least in the major economic literature, deals with the transmission mechanism that
maps the interaction of childhood poverty with mediator factors and adult labor market
outcomes by considering education and health as interrelated mediator factors.
The overall purpose of this study is to evaluate the impact of childhood poverty on labor
market outcomes in a way that bridges the identified knowledge gaps. In addition to the
childhood poverty effect analysis that treats childhood as a single period, poverty
incidences that occurred at different stages of child development are studied separately. For
both types of analyses, the transmission mechanism from childhood poverty to labor
market outcomes through mediator variables are traced and quantified into direct and
indirect effects.
Thestudyusesanestimationstrategythatcombinespropensityscorematching(PSM)and
averagecausalmediationeffect(ACME)analysisusingthePSIDdataset. Thepropensityscore
matching estimates the effect of childhood poverty on an individual’s health and
educational achievement. These outcome variables are used as mediator variables while the
second method gaugestheeffectofchildhoodpovertyonlabormarketoutcomesperannum:
laborincome,hours worked, and the average number of weeks unemployed. This method
enables us to decompose the total effect of childhood poverty on labor market outcomes
into direct and indirect effects. The method also allows dependency between the two
mediator factors.
Propensity score matching estimation results show that childhood poverty has a
significant and negative impact on children’s health and educational attainment. The
probability of having a healthy childhood is significantly lower (by 19.62 percent) for those
who experienced persistent childhood poverty than those who did not. Similarly, the
7
probability of earning a college degree is significantly lower (by 12.45 percent) for those
individuals who had persistent childhood poverty experiences. The same type of
comparison regarding years of schooling outcome variables provides a consistent result.
The study also traces the transmission mechanism of childhood poverty to adult labor
market outcomes. The average causal mediation analysis results showed that experiencing
persistent childhood poverty significantly lowers annual income and hours worked,
whereas it increases the number of weeks unemployed per annum. The total effect
estimation results are annual labor income ($11,252), annual hours worked (182.1 hours),
and the number of weeks unemployed per annum (1.96 weeks) differences on average
between impoverished and nonimpoverished groups. For all three outcome variables, a
significant portion of the total effect of persistent childhood poverty passes indirectly
through the two mediator variables: education andchildhoodhealth.
Outofthetotaldeterioratingeffectofchildhoodpovertyonannuallabor income, 42.25 percent of
it indirectly passes through poverty’s negative influence on health and educational
attainment. The remaining proportion is the direct effect of childhood poverty on labor
income. Similarly, hours worked and the number of weeks unemployed receives 22.57 and
21.73 percent, respectively, of the total effect of childhood poverty indirectly through its
negative effect on education and childhood health.
The gender disaggregation analysis of the ACME indicates that gender plays a significant
role in abating or magnifying the negative effect of childhood poverty on adult labor market
outcomes. The average annual labor income gap between women who did not experience
persistent childhood poverty and those who experienced it is significantly larger than a
similar comparison for men. On the other hand, the average hours worked and the number
of weeks unemployed gaps between men who did not experience persistent childhood
poverty and those who did is significantly larger than a similar comparison for women.
8
Similar PSM estimation and ACME of poverty on labor market outcomes have been
studied separately for the four stages of child development (infancy, early childhood, mid-
childhood, and late childhood). The results show that childhood poverty during early
childhood has the largest and most significant deterioration effect on overall childhood
health capital and educational attainment. The likelihood of having a healthy childhood is
significantly lower (by 16.04 percent) for those who experienced persistent poverty during
their early childhood compared to those who did not. Similarly, the probability of earning a
college degree is significantly lower (by 11.19 percent) for those individuals who
experience persistent poverty during early childhood. Hence, the early childhood stage is
critical in terms of the long-run effects of poverty on health and educational attainment.
Furthermore, for all stages of child development, except infancy, persistent childhood
poverty significantly and negatively impacts adult hours worked. The sensitivity analysis
using ACME’s sensitivity parameters and 125 percent of the federal poverty line reinforces
the above PSM and ACME
estimations result.
From the findings and methods of estimations used to measure the effect of childhood
poverty on the mediator and adult outcomes, this study contributes twofold. First, due to
the time lag between the cause (childhood poverty) and its effect (adult labor outcomes),
this study focuses on mediating roles played by health and education. This estimation
procedure enables us to decompose the total effect of childhood poverty on outcome
variables into indirect effects through mediator variables (health and education) and direct
effects. Second, a separate analysis of childhood poverty and its effect is conducted at four
major stages of child development: infancy, early childhood, mid-childhood, and late
childhood. Unlike the estimation procedure, which treats the whole of childhood years
9
(birth to 18 years old) as a single period, the separate analyses enable us to identify the
critical stage of child development that affect adult outcomes the most.
The rest of this chapter is structured as follows: Section 2 reviews the major literature.
Section 3 presents the theoretical framework that relates childhood poverty to mediator
variables (health and education) and labor market outcome variables (labor income, hours
worked, and the number of weeks unemployed). Section 4 describes the methodology of
the study: data, descriptive statistics, and identification strategy. Section 5 provides
empirical results and sensitivity analyses for assumption relaxation and redefinition of the
poverty line.
Section 6 discusses the major findings and draws conclusions.
2.2 Literature Review
Poverty hinders the investment capacity of a household to meet the individual and
collective demands of its members. This results in underinvestment in household essentials
that in turn leads to substandard housing, homelessness, inadequate nutrition and food
insecurity, inadequate childcare, lack of access to health care, unsafe neighborhoods, and
under-resourced schools. These disadvantages deteriorate the well-being of the household
members. In particular, the negative effect of poverty on the well-being of children has both
immediate as well as long-standing consequences and also transcends specific
consequences in determining success during adulthood. This section presents the reviews
of major studies in the literature that illustrate the short-run and long-run effects of
childhood poverty.
10
Poverty in early life has a long-lasting, negative impact on children that cause them to
perform less in late-life success indicators: educational attainment, health capital, and labor
market outcomes (Duncan et al., 2012; Duncan & Magnuson, 2013; Ratcliffe, 2015; Rizky et
al., 2019). Many studies have focused on its monetary aspect while dealing with childhood
poverty. For example, Lesner (2018) defined childhood poverty as living in a household that
earns 50 percent or less of the population’s median income. The monetary aspect of
childhood poverty has its own theoretical and empirical foundations; for example, Rizky et
al. (2019) explained that parental background, mainly income, plays a crucial role in
determining children’s conditions relative to the poverty threshold. Mayer (1997) pointed
out that parental income may influence children’s outcomes further through two main
competing hypotheses: the investment theory and the good parent theory. The first theory
emphasizes parents who invest time and money in their children through education, health,
or a good home environment. The second theory posits that low income induces severe
parental stress and hence poor parenting.
The empirical findings also reinforce the aforementioned theoretical link between
poverty anditseffectonchildren’seducationalattainmentandhealthproduction.
Asfarasitseffecton
childeducationisconcerned,lowparentalincomenegativelyinfluencestestscoresandduration
ofschooling(Dahl&Lochner,2012;Duncanetal.,2011;Levy&Duncan,2000). Lesner’s(2018)
findingsalsoshowedthesamenegativeimpactofchildhoodpoverty. Heemployedanintegrated
databaseforlabormarketresearch(IDA)providedbyStatisticsDenmarkfortheperiod1980to
2011 and demonstrated that individuals who experience childhood poverty are likely to
end up with a lower GPA. Furthermore, one additional year of childhood poverty reduces
the duration of schooling for the individual (by about 2 months), and she or he ends up with
education attainment that fits lower average earnings and lower labor market attachment.
11
Similarly,BellaniandBia(2019)usededucationasanintermediatevariabletoshowthelongru
n link between childhood poverty and adult outcomes. To this end, they employed the data
fromtheintergenerationaltransmissionofthe2011EUSurveyofIncomeandLivingConditions
(SILC) data. Their findings showed that being poor during childhood significantly decreases
one’s income level in adulthood and increases the average probability of being poor.
Moreover, theresultsrevealedthesignificantroleofeducationinthis
intergenerationaltransmission. The effect of childhood poverty is an average decrease of 11
percentage points in the probability of completing at least secondary education, which
indicates children who did not experience poverty are, on average, 1.8 times more likely to
have at least a secondary education degree.
The estimates are generally very close between the two methods Bellani and Bia
employed: propensity score matching and entropy balancing methods with the least
squares (probit) regression, both in magnitude and significance. The indirect effect
measures, conveyed through the educational level, suggested that a significant portion of
the average total effect is attributable to a decrease in the probability of graduating from
high school. Then, they concluded that growing up poor induces a lower level of education
that accounts for more than 30 percent of the total effect of childhood poverty on adult
income and the risk of poverty. Their findings are particularly relevant for Mediterranean
and Central and Eastern European countries.
Literature that employed the U.S. data for their empirical analyses also showed
consistent results. Ratcliffe (2015) used the PSID data and showed that persistently poor
children are 13 percent less likely to complete high school by age 20, 29 percent less likely
to enroll in post-secondary education by age 25, and 43 percent less likely to complete a
four-year college degreebyage25.
Persistentlypoorchildrenarealsolesslikely(by37percent)tobeconsistently
12
employedasyoungadultsthantheirever-poorandnon-persistentlypoorcounterparts. Duncan
andMagnuson(2013)usedasimilardatasetandvariousoutcomevariablestoshowthelong-run
effect of childhood poverty. Based on their three income groups categorization, which used
the official poverty line of $22,000 for a family of four, a child from a household with an
income belowtheofficialpovertyline,onaverage,getsschoolingonlyfor11.8years.
Ontheotherhand, children from households with an income between one and two times the
poverty line and more than twice the poverty line have average schooling years of 12.7 and
14 years, respectively.
Other empirical literature that attempted to causally estimate the link between parental
income and children’s outcomes suggested that income has short- and medium-term
effects, specifically on educational attainment (Dahl & Lochner, 2012; Ratcliff, 2015).
Furthermore, some studies investigated during which specific childhood age or age range
does experiencing poverty have more influence on adult outcomes. Interestingly, these
timing effect analyses of childhood poverty have mixed results. Studies carried out using the
U.S. data showed that family income matters most in the early years for the child’s
educational achievement (Duncan et al., 2012; Duncan & Magnuson, 2013; Ziol-Guest et al.,
2012), which is contrary to Lesner’s
(2018) findings based on the data from Denmark.
Besides the negative impact of childhood poverty on children’s educational attainment
discussed earlier, theoretical and empirical literature also showed its negative impact on
children’s health production. Ziol-Guest et al. (2012) explained that low income during
prenatal and childhood periods influences later disease processes, especially immune-
related processes, that may be plausibly connected through several potential pathways. The
fetal origins hypothesis posits a biological programming process, maternal diet and
smoking, for example, have known effects on neonatal development (Noakes, 2006;
13
Roseboom et al., 2001; Strauss, 1997). They also showed that low caloric intake during
pregnancy is associated with increases in chronic health conditions in the later life of
infants, such as coronary heart disease, hypertension, and obesity. Furthermore, low
parental income during the postnatal period also harms childhood and adult health capital.
The negative poverty-related adversity is known to impede parents’ abilities to engage in
warm and sensitive interactions with their children (Shonkoff & Phillips, 2000), and low
income in early childhood has been linked to poor mental health in adulthood (Currie &
Almond, 2011).
Ziol-Guestetal. (2012),basedonthePSIDdata,showedthatadulthealthdifferedmarkedly
depending on a child’s family income during pregnancy and infancy. Their findings revealed
thatthereportofdiagnosisofadultarthritisbychildrenfromafamilythatearnslessthan$25k per
annum (10.6 percent) is significantly larger than that of non-poor children (4.6 percent).
The hypertension report of the two groups is also significantly different: 19.0 percent and
11.2 percent, respectively. Although increases in childhood income between the prenatal
period and age 2 are associated with reductions in hypertension, arthritis, and activities of
daily living(ADLs)limitations,theseassociationsarestatisticallysignificantonlyamonglow-
income children. Specifically, a $5,000 increase in household income among low-income
children in any of the years between the prenatal period and age 2 (a total of 4 years) is
associated with a 1.1 percentage point reduction in the proportion of years that
hypertension was reported between the ages of 30 and 41 years old. It also showed that
there is a 1.3 percentage point reduction in
theproportionofyearsthatarthritiswasreportedanda0.02pointreductionintheADLindex.
Incrementsinincomedidnothavestatisticallysignificantassociationswithadulthypertension
and ADLs for low or higher-income households when the ages of children were between 3
and 5 or 6 and 15. The findings also showed that there are no significant associations
14
between early childhood income and mental health. These results contradict the findings of
Currie and Almond (2011). Also, empirical findings based on the US data showed that
parental income has an effect on mental and physical health as well as the Intelligence
Quotient (IQ) of children (Duncan et al., 1998; Levy & Duncan, 2000).
There is also literature that provides additional evidence on the effects of childhood
poverty from a different perspective than the comparison of health outcomes of poor and
non-poor groups. The studies demonstrated that improvement in parental income reduces
the detrimental health impact of child poverty. For example, Ziol-Guest et al.’s (2012)
estimates implied that raising the average income of low-income children by $5,000 per
annum over the entire 4-year interval is associated with about 5 percentage point
reductions in the risks of adult arthritis and hypertension. A $20,000 income increase
reduced the risk of hypertension (by one-quarter) and arthritis (by one-half) given the 19
percent and 11 percent rates, respectively, among those adults who were poor in early
childhood.
In addition to the intermediate impact on education and health discussed above,
empirical findingsalsoshowedthatchildpovertyhasalong-
rangeimpactonlabormarketoutcomeslater in life. Lesner (2018) found out that one
additional year of childhood poverty decreases the disposable income of the adult
individual by 6.4 percent. He further argued that childhood
povertyoccurrenceatdifferentageshasdifferentimpactsacrosstheboardandthelargesteffect
ofchildhoodpovertyonadultearnings(12.4percent)wasforpovertyexperiencedbetweenages
13 to 15.
The heterogeneity considered by Lesner (2018) refers to the age at which the child is
exposed to poverty. This can be referred to as ”cause heterogeneity”. Other researchers
consider childhood poverty’s heterogeneous effect from the viewpoint of its effect on adult
15
wage distribution. This can be referred to as ”effect heterogeneity”. Cho and Heshmati
(2015) employed the Korean Labor Income Panel Study (KLIPS) data to demonstrate the
latter. The findings from quantile regression analysis showed that the log wage differential
between poor and non-poor groups is observed at the overall distribution of wages from 8
percent to 21 percent while the OLS estimates indicate 16 percent. The magnitude of the
differential is more significant at the first quantile than the last quantile, which suggests a
left-skewed wage distribution with high dispersion of the lower-class group compared with
the middle-class group.
Consistently, analyses that did not account for any type of the above heterogeneity also
showed how childhood poverty negatively impacts adult labor market outcomes. Rizky et
al.’s (2019) instrumental variable estimation using Indonesian Family Life Survey (IFLS)
data showed that a child who lived in a poor family when aged between eight and 17 years
old suffers from an 87 percent earnings penalty on an hourly wage of 2014 relative to a
child who did not grow up in a poor family. Similarly, after taking the mediators into
account, a child who lived in a poor family in 2000 earned 85–92 percent less hourly wage
than one who did not live in a poor family. This result is consistent with the previous effect
that does not take the mediators into account (87 percent). Among the mediators, only
cognitive and mathematics skills are positively associated with an hourly wage. A one-
standard-deviation increase in cognitive or mathematics scores implies an increase of 6
percent and 12 percent hourly wage, respectively. This result indicates that the effect of
mathematics skills on labor market outcomes is worth about twice that of the cognitive skill
effect. However, they did not find any evidence that shows receiving various government
transfer programs mediates the effect of growing up poor on adulthood earnings. This
result is not consistent with Ziol-Guest et al.’s (2012) findings.
16
Instrumental variable estimation was considered because of the endogeneity of poverty
status in the year 2000 to explain the hourly wage of 2014. The researcher constructed an
instrumentforpovertyusingflowsintheagriculturalsectorfrom1997to2000. Thisinstrument
isadaptedfromBartik(1991)andisdefinedas“shift-share”or“enclave-based”. Twoconditions
mustbemetfortheshift-shareintheagriculturalsectorduringthecrisistobeavalidinstrument.
The first condition is it must be relevant or have a statistical relationship with the
endogenous poverty status. Second, it must not have a direct causal relationship with the
dependent variables or be correlated with the residuals.
The Instrumental Variable (IV) estimation result is consistent with that of OLS. Based on
the OLS estimation, a child who lived in a poor family in 2000 earns 19 percent less than a
child who did not live in a poor family. The coefficient is considerably higher when the 2SLS
is used for the estimation, in which a child who lived in a poor family in 2000 earns 87
percent less than one who did not.
Similar results were found in studies carried out based on European data. Bellani and
Bia’s (2019) study showed that being poor during childhood significantly decreases the
level of adult income and increases the average probability of being poor. Moreover, the
results revealed a significantroleofeducationininter-generationaltransmission.
Theresultsshowedasubstantial decrease (an average loss of around 764 Euro) in
equivalized income during adulthood due to exposure to childhood poverty for 28 EU
countries. This amount is half of the average gross monthly earnings of people with less
than a secondary degree in 2010. Bellani and Bia also found a significant increase in the
probability of falling into adulthood poverty for those who grew up poor. Average poor
children are 1.5 times more at risk of poverty in adulthood than their non-poor
counterparts.
17
The results from other counties are in line with studies that base their analyses on the
U.S. data. Ziol-Guest et al. (2012) found out that children in low-income families between
the prenatal period and age two had lower annual earnings as adults ($21,600) than
children from high-income families ($53,400). For similar groups, annual work hours are
1,460 and 1,877 and hourly earnings are $13.60 and $26.50 per hour during adulthood,
respectively. Similarly, Duncan and Magnuson (2013) showed that children who are raised
in the three categories of households have average annual earnings of $17,900, $26,800,
and $39,700, respectively, as adults.
Comparableannualworkhoursforadultsfromthethreegroupsofhouseholdsare1,512, 1,839,
and 1,963, respectively. A similar study by Ratcliff (2015) showed that only 35.4 percent
ofpersistentlypoorchildren were consistentlyemployedfromages25 to30. Thisfigureismuch
smaller than the same results for not persistently poor and never poor children, which are
63.6 percent and 70.3 percent, respectively.
The empirical analyses of the long-run impact of childhood poverty face many
challenges. First, isolating its causal impact on children’s well-being is difficult irrespective
of the timing of its occurrence. Because poverty is associated with other experiences of
disadvantage (such as poor schools, unfavorable homes, and neighboring environments, or
being raised by a single parent), it is difficult to know for certain whether it is childhood
poverty per se that matters or other related experiences. This indicates identifying causal
inferences from child poverty to late-life outcome variables is complicated. The literature
surveyed for this research employed differenttechniquestodealwiththisproblem.
AccordingtoDuncanandMagnuson(2013), the best method for identifying the extent to
which income matters would be an experiment that compares families who receive some
additional money to similar parents who do not receive such money. Studies took advantage
of an increase in the maximum earned income tax credit by more than $2,000 between
18
1993 and 1997 for poor families with more than two children. This generous increase in tax
benefits enabled researchers to compare the school achievement of children in otherwise
similar—and even the same—working families before and after the tax credit increase. The
authors found that improvements in low-income children’s achievement in middle
childhood coincided with the policy change made in income tax.
Bellani and Bia (2019) employed propensity score matching and entropy balancing after
controlling the confounding childhood factors (other than income poverty) that determine
the wellbeing of the child later in life to measure the long-run impact of childhood poverty.
Other literature employed instrumental variable estimation (Rizky et al., 2019) and the
quantile regression and decomposition method (Cho & Heshmati, 2015) to deal with the
matter. Despite all these efforts, Currie’s (2009) inter-generational research based on a
survey of various countries’ empirical studies questioned findings that reported a causal
effect of childhood poverty on adult outcomes. Rather, she argued that there is a strong and
exceedingly robust correlation between various measures of parental background or socio-
economic status (SES) and child health, though it is difficult to prove that the relationships
are causal. She also concluded that child health insults may have a large effect on future
earnings and/or employment probabilities by influencing noncognitive skills through
effects on adult health, though they have relatively little effect on the completion of
education.
The second concern of the identification process of the long-run effect of childhood
poverty is the heterogeneity account of the analysis. Lesner (2018) used siblings to identify
the heterogeneous effect. His results showed that age differences among siblings and the
timing of parental poverty allow for the identification of heterogeneous age effects of
childhood poverty. He concluded that parental income seems to matter more for small
children, while Northern European studies (Humlum, 2011; Jenkins & Schluter, 2002)
19
showed that the impact is largest for teenaged children. Rizky et al. (2019) showed that
being poor before the age of seven years had a larger direct effect on child cognitive
function than the combined indirect effect of poverty from age 7 to 14 years and school
attendance/home environment at age 7 to 14 years.
Studies in the U.S. found that child poverty seems to matter more at the earlier ages of a
child’s life. Ziol-Guest et al.’s (2012) findings showed that there is a much stronger effect of
income in the first years of life (between the prenatal period and 2 years) than later in
childhood. This finding is consistent with hypotheses claiming that the early years
represent a sensitive period in which social processes become embedded in biology and
epigenetic modifications (Hertzman & Boyce, 2010). Similarly, Duncan and Magnuson
(2013) argued that early childhood poverty matters the most. Based on emerging evidence
from both human and animal studies, they further discussed the critical importance of early
childhood for brain development and establishing the neural functions as well as structures
that shape future cognitive, social, emotional, and health outcomes. This implies that the
occurrence of poverty is not a single responsible factor for children’s outcomes; its timing is
also relevant. The poverty that occurs in the early ages of a child’s life is more detrimental
for some outcomes later in life, particularly for those related to achievement skills and
cognitive development.
Though factors like health and education play mediating roles to carry on poverty’s
effect
fromearlylifeexposureandlaterlifeoutcomevariables,verylittleofthesurveyedliteraturehas
triedtoquantifythetransmissionmechanism. BellaniandBia’s(2019)analysisgaveeducation a
mediating role to play in their long-run empirical inquiries. Similarly, Currie (2009), in her
study based on a survey of the literature, identified health as playing a mediating role.
20
However, no study in the major economics literature assigned a mediating role for the two
variables simultaneously.
Generally, the surveyed literature consistently showed how economically disadvantaged
children perform less in many of the indicators considered in the analysis. Despite some
sort of consistency in the conclusion of the negative effect of childhood poverty, the
researchers’ methods of dealing with the research question and their strengths and
weaknesses are different. The dissimilarities are manifested in terms of data, empirical
identification, and choice of outcome variables.
In most of the surveyed literature childhood is considered a single period. However,
experiencing poverty at different stages of child development does not have a homogeneous
effect on child development and adult labor market outcomes. The California Department of
Education (2000) reported that both the physical and mental development of a child have
their critical periods. Hence, in addition to the aggregate childhood poverty investigation,
this study analyzed childhood poverty for different stages of child development and their
effect on adult outcomes.
Scholars have different opinions regarding the number of stages of child development.
This study follows Lally and Valentine-French’s (2017) definition and categorization of
child development stages. According to their classification, there are four main stages of
child development: a “preoperational” or infancy stage (from birth through 2 years old),
early childhood (from two years old through the age of 6), a “concrete operational”/ mid-
childhood stage (from 6 years old to 12 years old), and late childhood (from 12 years old to
18 years old).
This study accounts for a separate analysis of each stage of child development in
addition to the persistent childhood poverty that covers more than one stage. Both analyses
gauge the
21
totaleffectofchildhoodpovertyonadultoutcomeswithhealthandeducationplayingmediating
roles. The analyses enable us to decompose the total effect into direct effect and indirect
effect channeled through child health and education. The inclusion of all these factors in a
single analysiswillcontributetotheliteraturethatdealswiththelong-
runeffectofchildhoodpoverty.
Dueemphasisisgiventoboththedataandmethodstodealwiththelimitationsofthesurveyed
literature. The data used for this study is the PSID, as it has wider coverage both in terms of
socio-economic variables and inter-generational time span. The combined empirical
models of average causal mediation effect analysis together with propensity score matching
estimation address most of the limitations raised in the surveyed literature.
2.3 Theoretical Model
Parental investment is referred to as any expenditure (time, energy, resources, etc.) that
parentsincurtobenefitanoffspring(Wang,2015). Agrowingbodyofresearchineconomicsand
developmental psychology has argued that attributes shaped by parental investment during
childhood have a significant role in determining adult outcomes. At least 50 percent of the
variability of lifetime earnings across individuals results from attributes of persons
determined by age 18 (Cunha et al., 2005; Huggett et al., 2011; Keane & Wolpin, 1997). This
implies that thefinancialwell-beingofparents,
whichisamajordeterminantforinvestmentinchildren, has an effect on the human capital
development of children and their adult outcomes.
Hence, poverty that occurs during childhood, when foundational cognitive and non-
cognitive skills are being formed, is likely to have more severe and long-term consequences
than poverty that occurs later. This assertion is in line with the argument that claims that
22
human capital production is a cumulative process subject to critical investment periods
(Cunha & Heckman, 2007; Kautz et al., 2014).
Inthisstudy,thehumancapitalmeasuresofachildnotonlyembraceeducationalattainment
but also accumulated health capital. Cunha and Heckman (2007) explained that although
thousandsofarticlesandbooksfollowedonhumancapitaldimensionsofeducationandtraining,
there have been fewer discussions of health as human capital. However, health and
education are considered to be the most critical components of human capital (Grossman,
2000; Schultz, 1961). Investing in them makes individuals more productive and there are
several important differences between them as well (Galama & van Kippersluis, 2015).
Hence, the theoretical framework focuses on how childhood poverty impacts the
production of both children’s health and educational attainment which in turn affects the
accumulation of human capital and adult labor market outcomes.
The theoretical model used in this study follows Galama and van Kippersluis’s (2015)
and
Meghir,Nix,andAttanasio’s(2015)procedureastheyconsideredbothhealthandeducationas
components in the formation of human capital. In their terms, these two components of
human capital are loosely referred to as “health capital” and “skill capital”. This study also
treats health as a form of human capital distinct from the component of human capital
accumulated through education and training.
Thetheoreticalframeworkfocusesonasingleagentwhogoesthroughtwogroupsofperiods:
childhood and adulthood. For the former, the child has no decision-making role, whereas, in
the latter period, the child becomes a decision-maker and makes an optimal decision based
on a utility function. Hence, during childhood, children’s human capital (health and
education capital) is accumulated based on decisions made by parents at the household
level.
23
Accumulated human capital at the beginning of adulthood (H) is mostly contributed by
investment in a child’s education (e) and health (h). Here, e is broadly defined to represent
both cognitive and non-cognitive skills. Hence, a human capital function is defined as:
H = H(e,h) (2.1)
The cognitive and noncognitive development, as well as health capital accumulation
from investment in education and health, are produced throughout childhood. The process
is governed by production functions that define how parental investment determines
children’s skills and health capital.
e = G(Ie) (2.2)
h = F(Ih) (2.3)
where Ie and Ih are parental investments for the production of child education and health for
thewholechildhoodperiod. Inthissegmentofthetheoreticalframework, thechildhoodperiod
(from birth up to age 18) is treated as a single period. Both functions are assumed to be
strictly increasing and strictly concave. Parental investment in children’s health takes place
through expenditures on medical care and time investments (e.g., exercise). Parents invest
in children’s skill development through outlays for schooling and on-the-job training. In
addition, parents also make time investments in children to teach them skills that are
incorporated as part of human capital. However, the parents’ time input analysis is not
included in this study.
Skill capital e (equation 2.2) and health capital h (equation 2.3) can be improved
through investments in skill capital (Ie) and health (Ih), respectively, that depend on parental
labor and non-labor income. Parental income is defined as:
A = wN + y(2.4)
24
where w is wage rate, N is hours worked and y is non-labor income.
Productions of skill and health capital, which use market goods and services as inputs,
are functions of the household’s intertemporal income. They are denoted by Xe(A) and
Xh(A), respectively. In addition, parents’ own time inputs for a child’s skill and health capital
improvement, τe and τh, are considered as components of the two types of investment in
children. In this study, the time inputs are not empirically analyzed due to data limitations.
The two types of investment can be expressed as:
Ie = Ie[Xe(A),τe] and Ih = Ih[Xh(A),τh].(2.5)
The skill-capital (e) and health-capital (h) production processes are assumed to be
increasing and strictly concave with respect to the investment inputs. This concavity
implies:
0; 0;
0; and 0;
This assumption of diminishing returns to investment (concavity) addresses the
degeneracy of the solution for investment that plagues the health-capital literature as a
result of the common assumption of constant returns to scale (Galama & Van Kippersluis,
2013; Galama, 2015 ).
Equations (2.1) through (2.5) show how parental investment decisions play a role in
linking children’s human capital accumulation with the household’s income. This paves the
way to show the relationship between childhood poverty and human capital development.
Therefore,
growingupinahouseholdwithalowerincome(A)hampershumancapitaldevelopmentthrough
low investment in the children’s education and health (Ih and Ih, respectively). Hence, the
25
incidenceofhouseholdpoverty(lowerA)hasadeterioratingeffectonchildren’shealthoutcomes
i.e., there is a positive relationship between household income and children’s health capital
accumulation:
0 (2.6)
Similarly, the negative effect of poverty incidence (lower A) on child educational outcome is
thus given by:
0 (2.7)
Consequently, accumulated human capital at the beginning of adulthood (H) is positively
affected by households’ income as:
0 (2.8)
which has an effect on the determination of labor market outcomes.
The early influential literature on children’s human capital accumulation, for example,
Becker and Tomes (1979; 1986), collapsed childhood into a single period and implicitly
assumed that investment at all ages of the child is the perfect substitute. This is similar to
the discussion made so far based on equation (2.1) through equation (2.8). This assumption
misses an important feature of the skill development process since all stages of child
development are not equally critical for acquiring specific types of skills. In this study, I
assume that different stages of child development play a different role in the process of
child human capital accumulation. Based on this assumption, a similar theoretical
foundation for the relationship between parental income and children’s human capital
accumulation at different stages of child development is analyzed and presented as follows.
26
Investments in health and education at different stages of child development are
independent and treated
separately.
) (2.9) ) (2.10)
where subscripts 1, 2, 3, and 4 indicate infancy, early childhood, mid childhood and late
childhood stages of child development. The parental income at each stage of child
development is positively related to both health and education capital accumulation.
>0 (2.11)
and are positive because of the strictly increasing and strictly concave
assumptions. If for all t′ = t, t is a sensitive period. A sensitive period exists
when the investment has a higher payoff in that period than in any of the others (but the
payoff in other periods is not necessarily zero). For example Conti and Heckman (2012)
showed that learning a second language is easier before age 12.
Similarly, the negative effect of poverty incidence (lower A) on child educational
outcome for the three stages of child development that exclude infancy is thus given below.
The infancy stage of child development is excluded while measuring parental income’s
effect on investment in education.
0 (2.12)
Consequently, accumulated human capital at the beginning of adulthood (H) is positively
affected by household income as:
27
0 (2.13)
The framework that shows the effect of childhood poverty on human capital
accumulation during childhood in its entirety (equation 2.8) or at different stages of child
development (equation 2.13) enables us to link childhood poverty and labor market
outcomes as discussed in the next section.
Therefore, the effect of childhood poverty on each mediator outcome variable, health
and educational capital, is stated in the following two hypotheses, which are derived from
the above four equations (2.6, 2.7, 2.11, and 2.12).
Hypothesis 1: Other things remain constant, childhood poverty reduces available
resources allotted to purchase goods and services which are necessary to make an
investment in child health. Hence, childhood poverty deteriorates the health capital of a
child.
Hypothesis 2: Similarly, the incidence of childhood poverty reduces the educational capital
of a child, ceteris paribus.
In the second (adulthood) period, children join the labor market and start making their
own labor supply decisions. Hokayem and Ziliak’s (2014) life cycle labor supply framework
is employed to trace the effects of a child’s cumulative capital of health and education on
labor supply decisions. This step enables us to link optimal labor supply decisions with the
previous section analysis to determine the effect of childhood poverty on labor market
outcomes that pass through the intermediaries: health and educational capitals.
The labor supply decision is a result of an agent’s lifetime utility maximization process.
Thetotalincomeoftheagentcomprisesnon-laborincome(yc)andlaborearnings(wN), where w
is the before-tax wage rate. The superscript c in the non-labor income of children is added
28
to differentiate it from parents’ non-labor income. Income can be spent on nonmedical and
medical consumption with a normalized price of 1. The resulting budget constraint is:
C = wN + yc.(2.14)
FollowingShaw’s(1989)specification,Idefinedtheobservedwage(w)astheproductofadult
human capital stock as expressed in equation (2.1), and rental rate on human capital (R), w
= R ∗H(e,h), where e is a composite function of investment in child education. The
investment in child education depends on goods and services purchased from the market
using parents’ intertemporal income, stock of health and educational capital, and parental
background. h is also a composite function defined as similar to e. The human capital that
continues to accumulate during adulthood is an extension of human capital development
during childhood. The rental rate, R, is the market price for the services of a unit of human
capital. It is a positive market-clearing price at which the aggregate supply and aggregate
demand for human capital services are in equilibrium. Hence, the wage rate is a function of
accumulated human capital which is determined by parents’ intertemporal income during
childhood and a person’s labor market decisions during adulthood.
The individual’s endowment of time at each period is normalized to 1. Since I denoted N
as the amount of labor it supplies, leisure becomes 1 − N. It is also assumed that leisure is
an argument in the utility function, more leisure leading to more utility or work has a
disutility effect on an individual’s welfare.
Anindividualchoosesleisuretime(1−N)andconsumption(C)tomaximizeutilitydefined
over leisure (L) and consumption(C), U(L,C). Hence, the utility maximization problem is:
(2.15)
29
The maximum level of utility for the agent is attained if both the decision variables, L and C
are optimally allocated.
The process of finding optimal values of the decision variables is presented as follows.
The Lagrangian of equation (2.15) is written as:
(2.16)
Then, the first-order conditions are:
= 0 (2.17)
= 0 (2.18)
(2.19)
where
Plugging the expression from equation (2.18) into equation (2.17) gives us:
UL − wUC = 0 (2.20)
The first-order equation for leisure has been rewritten to more readily examine the
effect of humancapitalinvestmentontheoptimallabor-leisurechoice.
Toanalyzethiseffect,decompose condition (20) into two parts:
30
1. UL, which denotes the gain in current utility due to an increase in the number of
leisure hours.
2. −wUC,whichistheutilitylossduetoadecreaseinconsumptionthatarisesfromadecrease in
earnings. It is because leisure trades off hours worked.
The condition in equation (2.20), UL = wUC, is the traditional optimal condition wherein
agents choose the optimal combination of consumption and leisure time which sets the
ratio of the marginal substitution between consumption and leisure equal to relative prices
at each time period.
The optimal decision explained above can be linked to the incidence of childhood
poverty through human capital formation. From equations (2.8) and (2.13) together with
Shaw’s (1989) wage equation, the incidence of childhood poverty reduces human capital
stock and consequently the wage rate. The reduction of the wage rate makes the slope of
the budget line flatter than the original. This leads to a low level of labor supply (higher
level of leisure) decision. On the other hand, higher household intertemporal income
improves human capital stock and wage rate. The slope of the budget line becomes steeper
and that leads to greater hours worked (low level of leisure).
Hence, the effect of the incidence of childhood poverty on adult labor outcomes can be
expressed as follows:
Otherthingsremainingconstant,theincidenceofchildhoodpovertydeteriorateshumancapital
stock (equations 2.8 and 2.13), which in effect decreases the wage rate of an individual. This
is derived from the presumption that higher income leads to better health and better
education.
Better health and education, in turn, contribute positively to human capital accumulation (H
= H(e,h)) that enhances wage rate; i.e.,
31
.
Hypothesis 3: Childhoodpovertynegativelyaffectshoursworkedthroughchannelsofhuman
capital development and wage rate, ceteris paribus, for lower income individuals; i.e.,
0.
An increase in w has opposite signed income and substitution effects whereas the net effect
on hours worked is ambiguous. However, for lower-income individuals, the substitution
effect is larger than the income effect, which leads to a positive effect of wage rate on hours
worked. In this case, childhood poverty negatively affects hours worked during adulthood.
Hypothesis 4: Childhood poverty has a negative effect on labor earnings (E) through both
wage rate and hours worked, ceteris paribus. This hypothesis emanates from the
hypothesized relationship established for the incidence of childhood poverty with both
wage rate and hours worked as:
.
2.4 Empirical Methods
2.4.1 Data and Descriptive Statistics
The PSID began interviewing a national probability sample of families in 1968 to assess
PresidentLyndonJohnson’swaronpoverty. Theoriginal1968PSIDsamplewasdrawnfromtwo
32
independentsamples: anover-sampleof1,872low-incomefamiliesfromtheSurveyofEconomic
Opportunity (the “SEO sample”) and a nationally representative sample of 2,930 families
designed by the Survey Research Center at the University of Michigan (the “SRC sample”).
The oversampling of families who were poor in the late 1960s resulted in a sizable
subsample of African Americans. These two samples combined to constitute a national
probability sample of U.S. families in 1968 (Institute for Social Research, University of
Michigan, 2019). These families were re-interviewed each year through 1997 when
interviewing became biennial. All persons in PSID families in 1968 are followed in
subsequent waves. In addition, anyone born to or adopted by PSID sample members is also
followed. When children become adults and leavetheirparents’homes,
theybecometheirownPSIDfamilyunitandareinterviewedineach wave (Johnson & Schoeni,
2011). This method of sampling “split offs” has been found to be an important procedure
for yielding a nationally representative sample (Fitzgerald et al., 1998).
The latest PSID data has 41 panel waves from 1968 to 2019 and has a total of 82,576
individuals. Thisstudyreliesonextracteddatathathasbothchildhoodandadultlabormarket
outcome information for the same individual. Specifically, I chose PSID sample members
born in 1968 and later in order to have full information about their childhood experiences
and adult outcomes.
However,theinclusionofallindividualswithrangesofadultyearsisnotappropriate
sincetheannuallaborincomeofanagentfluctuatesfordifferentstagesofone’slifecycle. Instead, I
limited the sample to the observation of adults who were born in 1968 or later and within
the age ranges of 35 to 54 years old. This age range was chosen based on the U.S. Bureau of
Labor
33
Statistics (2020) data, which indicates the stated age range is a plateau of annual labor
income. Typically, prior to age 35 average labor incomes are rising and after age 54 labor
incomes are
declining.
Hence, the data used in this study comprises both childhood and adult covariates. For
childhood years, I included children’s information together with parents’ background
information to measure childhood poverty status. The covariates here include household
income, family size, number of adults in the household, race, parental education, parents’
age when the child was born, children’s education, children’s health status, and parental
marital status. In addition, the labor market information and other covariates of the same
individuals when they became adults are extracted. These covariates include education,
health status, income from the labor market, hours worked and the number of weeks
unemployed per year. This allows us to link individuals’ childhood experiences with their
adult labor market outcomes. Limiting the sample as explained above yields 2,503
individuals of whom 52.06 percent (1,303) are female. The sample’s decomposition in
terms of race indicates that 67.76 percent of respondents are White, 30.84 percent are
Black and 1.40 percent are of other races. Because the average values of labor market
outcome variables are considered across different waves, the number of observations and
the number of individuals are the same.
In the PSID data, in each interview year, family annual income is collected for the prior
calendar year. This data in combination with the family size is used to construct the poverty
status of each household in which the sampled children were living. Then, the poverty
status of thehouseholdisanalyzedbasedontheofficialdefinitionofpoverty.
Undertheofficialdefinition, a family is poor if its gross annual money income is below the US
federal poverty level. The
34
strengthofusingthisofficialpovertymeasureisthatitallowsforstraightforwardpovertystatus
comparisonsovertimeamonghouseholds. Iusedweightedaveragepovertythresholdsprovided
by the U.S. Bureau of the Census and combined it with family annual income and family size
data of the PSID to compute household poverty status. For example, in 1975, the federal
poverty threshold for a family of three was $4,293. Since the household poverty analysis is
directly cascaded to childhood poverty analysis, a child who lives in a household that earns
less than $4,293 and has a family size of three or more is considered as a poor child.
The childhood poverty analysis is done for the whole childhood age range and for the
four stages of child development separately. For both types of descriptive analyses, three
types of livelihood status of a child, viz., persistently poor, ever poor, and never poor, are
identified. A persistently poor child lived in a poor family for at least half of his or her
childhood years. The ever-poor livelihood status indicates that a child lived in a poor
household for at least one year during his or her childhood years but is not persistently
poor. The never-poor livelihood status indicates that a child lived in a household that never
experienced poverty during his or her childhood years. Similar criteria were employed to
categorize the livelihood status of children in the four stages of child development. For
example, a child is considered persistently poor during early childhood if he or she lived in
a poor household for half or more of his or her early childhood years. Similarly, a child is
considered ever poor during early childhood years if he or she lived at least one year in
poverty but is not persistently poor. A similar categorization applies to the other stages of
child development.
Following children from birth through age 18 reveals that 13.9 percent of children are
persistently poor, meaning that they spent at least half their childhood years in a household
that earns an income below the federal poverty threshold. As Table 2.1 shows, from the
remaining proportion 24.73 percent of the children experienced poverty at least one year
35
during their childhood and the remaining 61.37 percent never experience poverty during
their childhood. This figure is not similar across different races. Children of color perform
much worse than average. While 30.57 percent of Black children were persistently poor,
only 6.1 percent of White children and 20 percent of other children were persistently poor.
The proportion of children who never experienced poverty is largest for White children and
smallest for Black children.
Table 2.1: Percentage of Childhood Poverty
Persistently Poor Ever Poor Never Experienced Poverty
White 6.1 22.46 71.34
Black 30.57 29.53 39.90
Others 20.00 28.57 51.43
Total 13.90 24.73 61.37
Notes: Persistently poor children are poor at least half the years from birth through age 18. ”Others”
includes Hispanic, Asian American and Pacific Islander, and Native American children. I am unable to
separately examine these groups because of sample size limitations.
Since experiencing poverty at different stages of child development has different effects,
the decomposition of childhood poverty across stages of a child’s development is
important. As Table 2.2 shows, during the infancy period, 152 (6.07 percent) of children
were persistently poor. This figure too is not evenly distributed among different racial
groups. For example,
14.64percentofBlackchildrenarepersistentlypoorwhereasthesamefigureforWhitechildren is
2.3 percent. Out of the total individuals who had experienced persistent poverty during
their infancy stage, 102 (67 percent) of them had experienced persistent poverty in their
entire childhood.
Table 2.2: Childhood Poverty at Different Stages of Child Development
Persistently Poor Ever Poor Never Poor
Infancy White 2.30 7.02 90.68
36
Black 14.64 18.91 66.45
Others 0 11.43 88.57
Total 6.07 10.75 83.18
Early Childhood White 8.73 7.13 84.14
Black 34.59 12.82 52.59
Others 8.57 25.71 65.71
Total 16.70 9.15 74.15
Mid Childhood White 7.43 12.74 79.83
Black 32.64 19.30 48.06
Others 25.71 17.14 57.14
Total 15.46 14.82 69.72
Late Childhood White 11.20 13.27 75.53
Black 38.86 15.80 45.34
Others 25.71 11.43 62.86
Total 19.94 14.02 66.04
During their early childhood, 418 (16.70 percent) of the children were persistently poor.
Similar figures during their mid-childhood and late childhood are 387 (15.46 percent) and
499 (19.94 percent), respectively. This decomposed analysis also reveals unevenly
distributed economic well-being of children of different races as shown in Table 2.2. Out of
the total individuals who had experienced persistent poverty during their early childhood
period, 286 (68.4percent)ofthemhadexperiencedpersistentpovertyintheirentirechildhood.
Thesimilar figure for mid-childhood and late childhood stages of child development are 317
(81.9 percent) and 300 (60.12 percent), respectively.
Table 2.3 presents definitions of all analytical variables used. The variables are extracted
from the PSID survey research center data set. The data provides information about
individual characteristics: race, age, gender, and childhood health condition. The second
group of variables provides information about parental characteristics: parents’
educational level, their age when the respondents were born, mothers’ marital status while
37
raising the children, their Home Ownership status, and the streams of the family income
that help to determine the childhood poverty status of respondents. Furthermore, the third
group of variables includes major dependent variables of labor market outcomes together
with health and education-related mediator outcomes. The labor market outcome variables
are labor income, hours worked, and the number of weeks unemployed in a year.
InthePSIDdata,oneofthemediatorvariables,education,ismeasuredbyyearsofschooling.
They take values ranging from 0 (no schooling) to 17, where 12 indicates that an individual
has completed high school, 16 indicates that an individual has completed a college degree,
and 17 indicates that the individual has done some postgraduate work (beyond college).
The data is designed such that the years of schooling are the same as the number of years
necessary to complete a certain level of education. For instance, if students take five or even
six years to complete college study, their years of schooling are still reported as 16.
Table 2.3: Definitions of Major Variables
Variable Definition
Persistently poor =1 if the respondent experienced childhood poverty for 9 years
or more, 0 otherwise
Mother’s age Age of mother at birth
Father’s age Age of father when the child was born
Mother HS graduate =1 if the mother has completed high school, 0 otherwise
Father HS graduate =1 if the father has completed high school, 0 otherwise
Mother college =1 if the mother has at least one year of college education,
0 otherwise
Father college =1 if the father has at least one year of college education,
0 otherwise
Mother dropout =1 if the mother has not completed high school, 0 otherwise
Father dropout =1 if the father has not completed high school, 0 otherwise
Single mother =1 if the respondent is raised by single mother, 0 otherwise
Home Ownership =1 if parents own their own house, 0 otherwise
Family size Respondent’s family size when she or he was a child
Number of Adults Number of adults in the family
38
Male =1 if the respondent is male, 0 if the respondent is female.
Healthy childhood =1 if the respondent has reported an excellent or very good or
good childhood health status, 0 otherwise
White =1 if the respondent is White, 0 otherwise.
Black =1 if the respondent is Black, 0 otherwise
Other =1 if the respondent is other than White or Black, 0 otherwise
Years of schooling Respondent’s number of years of schooling
School dropouts =1 if the respondent has not completed high school, 0 otherwise
Highschool graduate =1 if the respondent has completed high school, 0 otherwise
College =1 if the respondent has completed college education, 0
otherwise.
Average labor income Respondent’s average labor income per annum
Average hours worked Respondent’s average worked hours per annum
Average unemployment Respondent’s average number of weeks stayed unemployed
weeks per annum
The average annual values of labor market variables (labor income, hours worked,
number of weeks unemployed) are manipulated by taking the average of the annual values
when the individual is in the age range between 35 to 54 years old. In addition, the variable
expressed in monetary terms, labor income, is adjusted for inflation using the conversion
rate from the consumer price index (CPI) of the U.S. Bureau of Labor Statistics (2020). Table
2.4 presents descriptive statistics of the major variables.
Table 2.4: Summary Statistics of Major Variables
Mean SD Min Max N
Persistently poor 0.139 0.346 0 1 2503
Mother’s age 26.452 7.06 12 69 2503
Father’s age 29.05 8.046 15 81 2503
Mother HS graduate 0.312 0.463 0 1 2503
Father HS graduate 0.286 0.452 0 1 2503
Mother college 0.491 0.500 0 1 2503
Father college 0.467 0.499 0 1 2503
Mother dropout 0.036 0.186 0 1 2503
Father dropout 0.067 0.25 0 1 2503
39
Single mother 0.17 0.377 0 1 2503
Home Ownership status 0.618 0.486 0 1 2503
Family size 4.454 1.265 2 11 2503
Number of Adult 2.083 0.458 1 6 2503
Male 0.479 0.480 0 1 2503
Healthy childhood 0.469 0.499 0 1 2503
White 0.678 0.467 0 1 2503
Black 0.308 0.462 0 1 2503
Other 0.014 0. 117 0 1 2503
Years of schooling 14.886 1.939 8 17 2503
School dropouts 0.022 0.145 0 1 2503
Highschool graduate 0.479 0.500 0 1 2503
College 0.500 0.500 0 1 2503
Average labor income 63,650.68 93,636.35 20 1,876,069 2503
Average hours worked 1,982.352 645.424 7 5,460 2503
Average unemployment
weeks
2.051 5.734 0 52 2503
Table 2.5 describes the comparison of the mean values of the major variables based on
the childhood poverty status of individuals. Parental characteristics comparisons showed
that thosewhogrewuppoorperformedlowerinbetterwell-
beingindicatorvariableslikeeducational achievement and Home Ownership status. More
than 50 percent of parents of individuals who grew up non-poor did have at least one year
of a college education. However, only 23.6 percent of mothers and 19.5 percent of fathers of
individuals who grew up poor did have at least one year of a college education. The group
also performs lower from the perspectives of childhood health status, college education,
dropout rate, and adult labor market outcomes.
Table 2.5: Test on Major Variables Means’ Differences by Poverty Status
Variable Non-Poor Poor Difference T-stat
40
2.4.2 Empirical Models
This section explains two identification methods that estimate the effect of childhood
poverty on both mediator and labor market outcomes. The first method uses propensity
scores to estimate the propensity of a child to be raised in a poor household and the effect
on mediator outcomes (child’s health and education), which is equivalent to the average
treatment effect on the treated. The second method is the average causal mediation effect,
which deals with the total effect of childhood poverty on labor market outcomes (hours
worked, labor income, and the number of weeks unemployed) and decomposes the indirect
effect that comes through mediator outcomes (educational achievement and health
41
production during childhood). There is some sort of consensus in the literature that
growing up in a poor family is associated with a higher tendency of falling below the
poverty threshold in adulthood. However, the key contentious question for policy is
whether this association is truly causal in the sense that poverty in childhood per se
influences adult outcomes or whether it is driven by other factors that are correlated with
both childhood poverty and those outcomes. Moreover, it is relevant for policy makers to
examine plausible causal channels through which growing up in a poor household affects
the individual’s economic and social status as an adult.
Because of the complexity of the transmission mechanism that links childhood poverty
with adult labor market outcomes (hours worked, labor income, and the number of weeks
unemployed), meticulously designed statistical techniques are needed to trace the long-run
effect of child poverty on adult outcomes. Mediator causal analysis plays this role, as it
allows inter-linkage among initial causes, intermediaries, and final outcomes. It is employed
as a framework to quantify the effect of experiencing financial difficulties during childhood
on adult labor market outcomes. The framework also introduces individuals’ educational
capital and childhood health covariates as mediator variables. The following subsection
discusses a framework employed to analyze the impact of childhood poverty on mediator
variables, followed by the main empirical method of analysis that gauges the causal
mediation effect of childhood poverty on major outcome variables.
2.4.2.1 Average Treatment Effect on the Treated
ThepotentialoutcomesapproachforcausalinferencefollowsaframeworkfromRobin(1974,
1978). Consider a set of n individuals with subscript i: i = 1,2,..,n. The treatment in this
42
studyiswhetherthechildwasgrowingupinapoorhouseholdornot,Ti = 1(treatedorgrowing up
in a poor household) and Ti = 0 (control or growing up in a non-poor household).
Thetreatedandnon-treatedgroupsareidentifiedusingthecombinationofthePSIDsurvey
data and annual poverty-line data from the U.S. Bureau of the Census. A child who lived in
poverty for more than half of his or her childhood ages (i.e., 9 or more years) is considered
as a persistently poor child. The group consisting of these persistently poor children is
considered as a treated group. The other groups are included in the control group. Similar
definitions are
consideredtoanalyzetheimpactsofchildhoodpovertyatdifferentstagesofchilddevelopment.
The treated group consists of individuals who spent half or more years of the specific stages
of child development in poverty. The rest are the control group.
For each individual i, we observe a vector of pre-treatment covariates, Xi and vector of
outcome variables (childhood health status and educational attainment) associated with
the treatment and they are denoted by Yi(1) for being a poor child and Yi(0) for not being a
poor child. Then, the average treatment on the treated (ATT) is defined as:
ATT(X) = E[Yi(1) − Yi(0)|Ti = 1,Xi]
= EXi|Ti=1[E[Yi(1) − Yi(0)|Ti = 1,Xi]] (2.21)
= EXi|Ti=1[E[Yi(1)|Ti = 1,Xi]] − EXi|Ti=1[E[Yi(0)|Ti = 1,Xi]]
The second term of the last equation, which indicates the outcome of the treated individual
if she or he would not be treated, is not observable but represented by the counterfactual.
Theoutcomevariablesarechildhoodhealthstatusderivedfromanindividual’sreportabout
her or his childhood health. A healthy childhood dummy variable is created with a value of
1 for those who reported that they had good, very good, and excellent status of childhood
43
health and 0 for the rest. The other outcome variable is the educational attainment
extracted from the years of schooling variable of an individual.
The central assumption of the propensity score matching is unconfoundedness. The
“assignment to treatment” is unconfounded given the set of observable pre-treatment
characteristics: Yi(0)⊥Ti|Xi, where, within each cell defined by X, the treatment assignment
is random and the outcome of controls is used to estimate the counterfactual outcome of
the treated in case of no treatment. Some confounding factors affect parental income status
(childhood poverty or the treatment variable) and childhood health and education. This
makes the two variables (treatment and outcomes) interdependent. Parental education,
wealth indicators variables like homeownership status of the household and other
households, and parents’ characteristics influence the income-generating power of parents
(i.e., childhood poverty) and parental investment capacity on health and education
(outcome variable). For example, highly educated parents have less chance of being poor
and have a high investment capacity in children’s health and education. If all such covariates
are controlled, treatment
(poverty status) becomes exogenous and independent of the outcome variable. Hence, the
uncounfoudedness assumption will be satisfied. This study considers the exhaustive list of
pretreatment covariates to satisfy the criteria.
Thepre-treatmentconfoundersarethemother’sageatbirth,thefather’sagewhenthechild
wasborn,theeducationalstatusofthemother(dummyvariablesofschooldropouts,highschool
completed, andatleastoneyearofcollegeeducation), educationalstatusofthefather, whether
the child is raised by a single mother or not, homeownership status of the household, family
size, number of adults in the family and race. These covariates affect the income-generating
and asset-building capacity of a household, then in effect determine households’ potential
to invest in the production of a child’s health and education. The treatment variable,
44
childhood poverty, is endogenous before matching since it is not independent of these
factors, which can affect the outcome variables (child health and educational attainment). It
becomes exogenous after matching or controlled for the pretreatment covariates.
Let p(X) be the probability of growing up in a poor household given the set of covariates
X : p(X) = Pr(T = 1|X = x) = E[T|X = x]. Rosenbaum and Rubin (1983) showed that, if
the potential outcome Yi(0) is independent of treatment assignment conditional on X, it is
also independent of treatment conditional on p(X): Yi(0)⊥Ti|p(Xi). Thus, for a given
propensity score value, exposure to treatment can be considered as random and thus poor
and non-poor children should be observationally equivalent on average. This is also
checked by pstest that evaluates balance among covariates after the propensity score
matching.
Formally, given the population of units i, if we know the propensity score p(Xi), the
average effect of being poor on those exposed to childhood poverty (the average treatment
effect on treated [ATT]) can be rewritten as:
τ= E[Yi(1) − Yi(0)|Ti = 1]
= Ep(Xi)|Ti=1[E[Yi(1) − Yi(0)|Ti = 1,p(Xi)]] (2.22)
= Ep(Xi)|Ti=1[E[Yi(1)|Ti = 1,p(Xi)]] − Ep(Xi)|Ti=1[E[Yi(0)|Ti = 1,p(Xi)]]
When Ti = 1, then Yi(1) = Yi, whereas Yi(0) is never observed.
To estimate the ATT, the first step is to estimate the propensity score of growing up poor
using a probit model, ˆp(Xi); then the unobserved potential outcomes are estimated based
on the propensity score (PS) matching. The propensity score matching technique has been
chosen for this study because of the following advantages. Guo et al. (2020) argued that the
greatest advantage of the propensity score is its reduction in dimensions, which solves the
problem of insufficient sample cases in exact matching. Second, matching does not require
45
functionalformassumptionsfortherelationshipbetweentheexpectedoutcomesandthevalues
of characteristics; i.e., the relationship is left unspecified and can have a quite general form
(Bryson et al., 2002).
Thenearestneighbormatchingtechniquewithreplacementandcaliperof0.001isemployed
fortheaboveprocedure. Letl denotetheindexoftheunitintheoppositetreatmentgroupthat is
closest to unit i :
|pˆ(Xl) − pˆ(Xi)| ≤ |pˆ(Xj) − pˆ(Xi)| ∀j,Tj = Ti(2.23)
This procedure enables us to select a control group from non-treated individuals (in this
case non-poor as a child) who are very similar to treated individuals in terms of their
probability of being poor during their childhood. Then the average treatment effect on
treated estimate τis
estimatedby ˆ (0)]withYˆ
i(0) = Yl(0),Yˆ
i(1) = Yi andn1 isthenumber
ofindividualswhogrewuppoor. Theseestimatesoftheaveragetreatmenteffectonthetreated
enable us to test the first two hypotheses discussed under the theoretical model.
A similar propensity score matching estimation procedure is employed for the four
stages of child development. The treated groups comprise individuals who lived half or
more years of their life at that particular stage in poverty. The rest were considered the
control group. The mediatoroutcomevariablesandthepre-
treatmentcovariatesarethesamefortheinfancystage of child development. However, the
other stages of child development incorporate the poverty statuses of the previous stages
as part of the pre-treatment covariates.
The discussion we made so far enables us to measure the effect of childhood poverty on
mediator variables. However, its effect transcends to labor market outcomes through its
46
effect on mediator outcomes (health and education) and other channels. The next section
deals with the estimation method, which quantifies the total effect of childhood poverty on
labor market outcomes and its decomposition into direct effect and indirect effect which
passes through the mediator outcomes.
2.4.2.2 Average Causal Mediation Effect
Hypotheticalinterventionsthatimproveordeterioratethehealthcapitaland/oreducational
attainment of an individual can affect adult outcomes. Childhood poverty is one of these
types of shocks that deteriorate the health and educational capital of a person and
consequently affect labor market outcomes. In addition to measuring the total impact of
childhood poverty on mediator outcomes using the method explained in the previous
section, in this study, the
aimistodecomposethecausaleffectofchildhoodpovertyonadultearnings, hoursworked, and
the number of weeks unemployed into an indirect effect, transmitted through health status
and educational attainment, and a direct effect (which includes all other potential
mechanisms).
This study follows a framework developed by Imai and Yamamoto (2013) to estimate
the total effect and its decomposition. Unlike prior literature on the topic, their method
considers a possible causal relationship between mediator variables. In this study, I
assumed that child health affects educational achievement and the assumption is in line
with Gracy et al.’s (2017) findings. Poor children have more prevalence of asthma, vision
problems, hearing loss, dental pain, and persistent hunger which cause absenteeism and
lack of concentration while attending school.
Thesenegativehealthconditionsinturnleadtonegativeschooloutcomes. Thenegative
47
educational outcomes are manifested in the form of grade repetition, lower academic
scores, disengagement with school, and attendance problems.
The reason for the unidirectional effect between health and education is because
parents’ investment decisions on child health influence a child’s health outcome that can
also affect the child’s educational performance. However, since children are not decision-
makers in the household, their educational attainment cannot influence the health
investment decision of the household and the child’s health outcome. Figure 2.1 depicts the
interaction between the two mediator variables (H and E) together with the childhood
poverty (treatment (T)) and outcome variables (Y ). In both panels, the treatment variable T
could affect outcome Y in three ways: through the main mediator of interest E, through the
set of alternative mediator H, and directly.
Figure 2.1: Treatment, mediators and outcomes (adapted from Imai & Yamamoto, 2013)
In this study, the interaction among the variables is similar to the one depicted in Panel
(b), whereT (childhoodpoverty)affectsoutcomeY (laborincome,
hoursworkedandthenumberof weeks unemployed), mediators E (educational attainment)
and H (childhood health status). ThemediatorH
confoundstherelationshipbetweenthemediatorofinterestE andtheoutcome Y . Panel (a) is
not considered in this study because it implicitly assumes the absence of a causal
relationship between the corresponding mediators. Unlike Panel (a), for which Imai and
Yamamoto (2013) developed a different methodology, Panel (b) is an alternative scenario,
48
where H can affect Y either directly or indirectly through E, thereby allowing for the
potential causal relationship between E and H. The quantity of interest for the proposed
method is the average causal mediation effect among the treated (ACMET) with respect to
E, which is represented by the dashed arrows connecting T and Y through E.
The Setup and Assumptions
The potential values of educational attainment (E) is affected by the health status of a child
as explained above. That is, Ei(t,h) denotes the potential value of educational attainment for
unit i when the childhood poverty status is t = 0,1, and the alternative mediator H
(childhood health status) takes the value of h = 0,1. Then, the observed value of the
mediator for this unit is given by Ei = Ei(Ti,Hi(Ti)). Similarly, Yi(t,E(H(t))) denotes the
potential outcome under the treatment status t and the mediator value e derived from h
value of H, where the observed outcome Yi equals Yi = (Ti,Ei(Ti,Hi(Ti))).
Under this setting, for each unit, the causal mediation (with respect to E) and direct
effects are defined as:
δi(t) ≡ Yi(t,Ei(1,Hi(1)),Hi(t)) − Yi(t,Ei(0,Hi(0)),Hi(t)) (2.24)
ζi(t) ≡ Yi(1,Ei(t,Hi(t)),Hi(1)) − Yi(0,Ei(t,Hi(t)),Hi(0)) (2.25)
For t = 0,1, where δi(t) corresponds to the causal effect of childhood poverty on the labor
market outcomes that transmits through the mediator of interest (educational attainment).
Thus, for annual labor income, δi(t) represents the difference between the two potential
labor incomes for subject i, who actually receives the treatment (persistently poor during
childhood) andthosewhoareinthecontrolgroup(non-poorduringchildhood).
Thisisacausalmediation effect or indirect effect of education, which is affected by childhood
poverty and health status, on annual labor income. The indirect effect on the other labor
market outcomes, hours worked and the number of weeks unemployed can also be
explained in a similar fashion.
49
On the other hand, ζi(t) represents the rest of childhood poverty effect (denoted by the
solid arrow at the bottom of Panel (b) in Figure 2 and the combination of the arrows that go
from T to Y through H but not through E). Hence, ζi(1) ≡ Yi(1,Ei(1,Hi(1)),Hi(1)) −
Yi(0,Ei(1,Hi(1)),Hi(0))representsthedifferenceinannuallaborincomeundertreatment(the
individual who experienced persistent childhood poverty) and control (who did not
experience persistent childhood poverty), holding the level of education constant at the
level that would be realized under treatment.
Thus, as expected, the sum of these two effects equals the total treatment effect,
τi ≡ Yi(1,Ei(1,Hi(1)),Hi(1)) − Yi(0,Ei(0,Hi(0)),Hi(0)) (2.26)
τi = δi(t) + ζi(1 − t)for t = 0,1 (2.27)
Instead of the individual-level effects which are discussed so far, we are interested in
estimating average effects, i.e., ACME or δ¯
i(t) ≡ E(δi(t)), average direct effect or ζ¯
i(t) ≡
E(ζi(t)) and average total effects or ¯ τ≡ E(τi).
The following weaker version of the sequential ignorability (SI) and interaction of
treatment and mediator assumptions are needed in order to identify ACME under the
scenario depicted in Panel (b) of Figure 2 .
Assumption 1: Sequential Ignorability with Multiple Causally Dependent
Mediators
The following three conditional independence statements are assumed:
Yi(t,e,h),Ei(t,h),Hi(t)⊥Ti|Xi = x
Yi(t,e,h),Ei(t,h)⊥Hi|Ti = t,Xi = x(2.28)
50
Yi(t,e,h)⊥Ei|Hi(t) = h,Ti = t,Xi = xfor any t,e,h,x.
Sequential ignorability implies that the treatment assignment is essentially random after
adjusting for observed pre-treatment covariates and the assignment of mediator values is
also essentially random once both observed treatment and the same set of observed pre-
treatment covariates are adjusted for (Imai et al., 2011, pp. 863–864).
This study uses different covariates (factors) that affect the treatment, mediators and
outcome variables to guarantee the sequential exogeneity of treatment and mediators.
Factors that determine parents’ income-generating ability and factors that determine the
per capita share of household members from the total income are considered as the pre-
treatment confounders. Hence, treatment (childhood poverty) is assumed to be random or
ignorable given the confounders that include mother’s age at birth, father’s age when the
child was born, educational status of mother (dummy variables of high school completed,
college degree, and dropout), the same dummy variables for the educational status of
fathers, whether the child is raised by a single mother or not, home ownership status of the
household, family size and race.
The second part of assumption 1 implies that the observed mediator (child health) is
ignorable or random, given the actual treatment status (childhood poverty) and the stated
pretreatment confounders. Similarly, educational attainment is ignorable or random, given
the actual treatment status, pre-treatment confounders, and child health. Hence, exogeneity
is assumed for the treatment T, the alternative mediators H, and the mediator of interest E.
The sensitivity analysis also analyzes the robustness of the results after relaxing this
assumption.
Assumption 1 is insufficient for identifying the ACME in the presence of causally
dependent multiple mediators. However, Robins (2003) showed that if the extra
51
assumption of no treatment-mediator interaction effect holds on top of the sequential
ignorability assumption, the ACMEs are identifiable.
Assumption 2: (No Interaction Between Treatment and Mediator)
For every unit i, we assume the following equality,
Yi(1,e,Hi(1)) − Yi(0,e,Hi(0)) = Yi(1,e′,Hi(1)) − Yi(0,e′,Hi(0)), for any e,e′.
Theproblemwiththisassumptionis,itisunlikelytobecredibleinmostapplicationsbecause it
must hold for every unit. In our context, it is difficult to hold this assumption since
children’s health capital and educational attainment have interaction with childhood
poverty. Heckman and Pinto (2015) also argued that even though the Imai and Yamamoto
(2013) approach is basedonweakerassumptionsthanthePearl(2001)solution,
whichisbasedonlackofvariation of unobserved inputs (mediator in our context), their
assumptions are nonetheless still quite strong.
To overcome this limitation, Imai and Yamamoto (2013) introduced the following
methodology that relaxes the no-interaction assumption. Even though this remedy relaxes
the second assumption, homogeneous interaction between treatment and mediator is still
needed. But, the sensitivity analysis discussed below addresses the robustness of ACME for
the possibility of heterogeneous interaction.
ACMEs are identified using two unobserved quantities that are used as sensitivity
parameters. The first parameter is the correlation between the mediator of interest Mi(t)
and the individual-level treatment-mediator interaction effect κi i.e., ρt = Corr(Mi(t,Wi(t)),κi).
The second one is the standard deviation (SD) of the individual-level coefficient for the
treatment-mediator interaction, i.e., σ= pV (κi). When ρt = 0, we can identify the ACME
regardless of the value of σ. However, when ρt, is not equal to zero, we must specify both ρt
and σin order to estimate the ACME under assumption 1.
52
Imai et al.(2010a, 2010b) used coefficients of determination as an alternative
parameterization to ease interpretation of the parameters. Specifically, the proportion of
the unexplained or original variance of the outcome explained by incorporating the
heterogeneity in the treatment-mediator interaction has been used to substitute the
function played by ρt.
Thus, the sensitivity parameter represents how important it would be to incorporate the
interaction heterogeneity in the regression model to explain the variation in the outcome
variable. Formally, these parameters are defined as:
and (2.29)
for the proportion of unexplained variance and that of the original variance explained by
the heterogeneity of the treatment-mediator interaction effects, respectively. Then, it is
possible to directly relate these quantities to the ACME through the following one-to-one
relationship between σand each of these coefficients of determination as:
(2.30)
This implies that σis bounded from above by because 0
< R2 ∗< 1. The
sensitivity to the interaction heterogeneity can be assessed by studying how the ACME
varies depending on the values of R2 ∗and R˜2. This can also be done by calculating the ratio
of σto its upper bound.
Yin and Hong (2019) also showed that, assuming the generalized SI and linear structural
equation model (LSEM) for all measured and unmeasured variables, the general average
direct effect(ADE)andgeneralACMEcanbeidentifiedthroughtwolinearregressionequations.
53
This identification is applicable even when multiple causally-dependent mediators are
unmeasured. Theidentificationimplies ADEandACME canbe easilycalculatedusing
thecoefficientsof the two linear regression equations.
ACME = E(δi(t)) ≡ E[Yi(t,Mi(1,Wi(1),Wi(t))) − Yi(t,Mi(0,Wi(0),Wi(t)))] (2.31) ADE = E(ζi(t))
≡ E[Yi(1,Mi(t,Wi(t),Wi(1))) − Yi(0,Mi(t,Wi(t),Wi(0)))] (2.32)
Hence, Average Treatment Effect (ATE) can be decomposed into ADE and ACME as:
ATE = ADE(0)+ ACME(1) = ADE(1)+ ACME(0).
This simply means that ATE is equivalent to ¯ τ≡ E(τi), which is the same as the estimator of
equation (2.27).
Generally, the estimation procedure of ACME, ADE and ATE follows the following steps
as explained in Imai et al. (2010a).
(a) Fit the models for the observed outcome and mediator variables.
(b) Simulate the model parameters from their sampling distribution.
(c) Repeat the following three steps:
(i) simulate the potential values of the mediator,
(ii) simulate the potential outcomes given the simulated values of the mediator and
(iii) compute the causal mediation effects.
(d) Compute summary statistics such as point estimates and confidence intervals.
Then, conduct a sensitivity analysis to gauge the robustness of the potential violation of the
homogeneous interaction assumption by examining how the location and width of the
bounds varyas σchanges.
Theresultsfromthewholeprocedureprovideuswithinformationtotestthe
hypothesesaboutthelabormarketoutcomesdiscussedunderthetheoreticalmodel(hypothesis
3 and hypothesis 4).
54
A similar average causal mediation analysis procedure was employed for the four stages
of child development. The treated groups comprise individuals who lived half or more years
of their life in that particular stage of poverty. The labor market outcome variables and the
pre-treatment covariates are the same for the infancy stages of child development.
However, the other stages of child development incorporated the poverty statuses of the
previous stages as part of the pre-treatment covariates.
2.5 Empirical Results
2.5.1 Effect of Childhood Poverty on Adult Outcomes
The first step to measuring the impact of growing up poor on adult outcomes is to
estimate a probit model for an individual’s propensity score of being poor during childhood
years. This estimation gives the probability of a child growing up in a poor household given
the observed pretreatment characteristics. The estimated propensity score indicates that
within the same value of propensity score, growing up poor or not does not depend on the
values of the pretreatment characteristics. This implies that those who grew up poor and
those who did not should be similar on average conditional on observable pretreatment
characteristics.
Table 2.6 shows the probit estimate of the propensity score, which measures the effect
of parental and family-level characteristics on the probability of being poor during
childhood. The result shows that the propensity of an individual being poor during his or
her childhood is
negativelyandsignificantlyaffectedbyparentaleducationalattainmentdependingonwhether
parents own their own residence, the number of adults in the family, the fathers’ age at
0.0224Mother’sage ∗∗
∗
(0.0084)
-0.0134Father’sage ∗
(0.00741)
-0.191MotherHSgraduate ∗
∗
(0.0945)
FatherHSgraduate -0.0510
(0.0926)
Mothercollege -0.506 ∗∗
∗
(0.0979)
Fathercollege -0.345 ∗∗
∗
(0.0991)
Singlemother 0.304 ∗∗
∗
(0.0879)
HomeOwnershipstatus -0.671 ∗∗
∗
(0.0759)
Familysize 0.199 ∗∗
∗
(0.0288)
NumberofAdults -0.246 ∗∗
∗
(0.0736)
White -0.560 ∗
∗
(0.270)
Black -0.0396
(0.271)
Constant -0.718 ∗
∗
(0.323)
χ2507(12)
55
birth, andbeingWhite.
Ontheotherhand,theprobabilityoffacingchildhoodpovertyissignificantly
pronouncedifasinglemotherraisesachild,hasalarge-sizedfamily,andolderageofthemother at
the time the child was born.
An important requirement for propensity score matching identification is the presence
of common support, which ensures that for each treated unit (or a group comprised of
individuals who grew up poor) there are control units (non-poor) with the same
observables. Based on the psmatch2 matching with 0.001 caliper distance, 83 observations
from the treated group are outside the common support of propensity score distributions
for the two groups as it is shown in Figure 2.2.
Table 2.6: Propensity Score Estimation Dependent
Variable: Persistently Poor
0.251PseudoR-square
2503N
Standarderrorsareinparentheses. ∗p < ,0.10 ∗
∗p < 0.05, ∗∗
∗p < 0.01
56
Figure 2.2: Common support of propensity score distribution for poor and non-poor
After the propensity score matching, the Kernel density of the control group changed
significantly and became comparable with the same distribution of the treated group.
Figure 2.3 shows the two groups’ propensity scores distributions before and after the
matching.
The statistical pstest numerical result reinforces the balance among covariates after the
propensity score matching as graphically illustrated in Figure 2.3. Contrary to the
unmatched
data,whichhasasignificantmeandifferenceforallcovariates,thematcheddatahassignificantly
57
non-differentiable covariates. The overall chi-square measurement also showed that there
are
Figure 2.3: Propensity score density distributions before and after matching
no significant covariates value differences between the treatment and control groups after
the matching has been done. The mean bias after matching became very small (4.3) as
compared to the mean bias before matching (50.7) as shown in Table 2.7.
58
2.5.2 Childhood Poverty and Mediator Outcomes
The mediator outcomes considered for the analysis in this section are childhood health
status and educational attainment. The result presented in Table 2.8 shows that childhood
poverty significantly and negatively affects the likelihood of having a healthy childhood.
This Table 2.7: Covariates Comparison of the Two Groups Before and After Matching
χ2
4.29
+ Mean Difference of Treated (Poor=348) and control (Non-poor=2155) ∗p <
0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
impliesthatpersistentchildhoodpovertyreducestheprobabilityofhavingahealthychildhood by
0.1962 points (or 19.62 percent), other things being equal.
Regarding education outcomes, growing up poor has significant and negative effects on
years of schooling and earning a college degree and above. Other things being equal,
persistent childhood poverty reduces the probability of having a college degree by 0.1245
MeanDifference +
Variables Unmatched Matched
1.531Mother’sage ∗∗
∗-0.400
1.853Father’sage ∗∗
∗-0.415
MotherHSgraduate 0.051 ∗0.000
0.058FatherHSgraduate ∗
∗-0.019
-0.297Mothercollege ∗∗
∗0.015
-0.316Fathercollege ∗∗
∗0.038
0.252Singlemother ∗∗
∗-0.042
-0.381HomeOwnershipstatus ∗∗
∗-0.011
0.995Familysize ∗∗
∗0.085
0.0584NumberofAdult ∗
∗-0.045
-0.437White ∗∗
∗0.000
0.429Black ∗∗
∗-0.004
LR
50.7 4.3MeanBias
p-value χ 20.000 0.978
59
points (or 12.45 percent). This result aligns with the non-linear evaluation of the impact of
growing up poor on years of schooling. The results in Table 2.8 show that persistent
childhood poverty reduces years of schooling by a magnitude of 0.826 years.
Table 2.8: Average Treatment Effect of Childhood Poverty on Health and Education
Healthy Childhood (%) Years of Schooling College Degree (%)
ATT -19.62∗∗∗ -0. 826∗∗∗ -12.45∗∗∗
(0.048) (0.186) (0.045)
Standard errors in parentheses. ∗p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
The effect of childhood poverty is not confined to its effect on mediator variables. The
longrun effect of childhood poverty on adult labor market outcomes that passes through
these two mediator variables and the transmission mechanism that flows among them will
be discussed in the next section.
2.5.3 Effect of Childhood Poverty and Causal Mediating Roles of Health
and Education
The effects of childhood poverty on children’s health capital and education attainment
transcend to other adult labor market outcomes. This section aims to gauge the total effect
of childhood poverty on labor market outcomes and the proportion of the effect that passes
through the two mediator variables discussed earlier.
The study employs the causal mediation analysis procedure to determine the total effect
of childhood poverty on outcome variables and its decomposition to direct and indirect
effects. In addition, the proportion of the total effect explained by the mediating role of
health and education was also measured. For all ACME estimations, 1000 simulations of
potential values of the mediator and corresponding potential values of outcomes given the
60
simulated values of the mediator have been manipulated. Then, the causal mediation effects
are calculated based on the given potential values. This helps to estimate the standard
errors from which the significance level of ACME and ADE estimates are drawn. The
simulation type employed here is the Imai et al. (2010a) default simulation of the quasi-
Bayesian Monte Carlo method based on normal approximation.
Table 2.9 presents ACME estimates for both the treated (poor) and control (non-poor)
groups together with the weighted average ACME and the total effect of childhood poverty
for three labor market outcomes: average annual values of labor income, hours worked, and
the number of weeks unemployed. Childhood poverty significantly and negatively impacts
average annual labor income, amounting to a total of $11,252 difference between the
treated (poor duringchildhood)andcontrol(non-poorduringchildhood)groups. And,
outofthistotallabor income impact of childhood poverty, 42.25 percent of its proportion is
transmitted through its negative effect on child health and education.
Table 2.9: ACME and Total Effect of Childhood Poverty on Labor Market Outcomes
Labor Income Hours Weeks
Worked Unemployed
†95% confidence intervals. ∗p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
61
Similarly, growing up poor negatively impacts other labor market outcomes. It has a
significant and negative effect on average annual hours worked and a significant and
positive impact on the annual average number of weeks an individual stays unemployed in
a year. The figures are 182.1 hours and 1.96 weeks, respectively.
From the estimates for average annual hours worked and the average number of weeks
unemployed, the indirect effect or ACME of mediator variables (health and education)
constitutes a significant portion of the total effect of childhood poverty, both in magnitude
and significance. Out of the total effect of childhood poverty on annual hours worked, 41.1
hours reduction or 22.57 percent passes through its negative influence on a child’s health
and education. This magnitude is even larger when the treatment effect is calculated only
on the poor (treated) group, 65 hours or 35.7 percent. The ACME and total effect estimation
of the number of weeks unemployed show that a significant portion of childhood poverty’s
effect indirectly passes through its influence on child health and education before it affects
the outcome variable. It constitutes 21.73 percent of the total effect and is highly significant.
This estimation is more pronounced when calculated for the counterfactual constructed
from those who grew up poor (32.45 percentage points). Figure 2.4 demonstrates these
quantitative
results.
62
Figure 2.4: ACME and total effect on labor market outcomes with 95% CI
2.5.4 Gender Disaggregated Analysis of Impact of Childhood
Poverty
Women and men’s labor market experiences differ in earnings and labor force
participation (the U.S. Bureau of Labor Statistics, 2017). Hence, this section deals with the
gender disaggregation of the impact of childhood poverty on labor market outcomes. The
study has 1,303 women and 1,200 men, of which 217 women and 131 men experienced
persistent poverty during childhood.
63
The disaggregation is done in two possible combinations of gender and poverty
statuses: poor women vs. non-poor women and poor men vs. non-poor men. The estimates
for the two comparison groups presented in Table 2.10 show that individuals who
experienced childhood poverty performed less. Experiencing childhood poverty has a
negative effect on labor income, hours worked and the number of weeks unemployed. The
total labor income effect ($12,820) of childhood poverty is larger for women (i.e. when
comparing persistently poor women with non
persistentlypoorwomen)thanmen(persistentlypoormenvsnonpersistentlypoormen). Out of
the total effect of childhood poverty ($12,820), 39.78% of the effect passes indirectly
through the two mediator outcomes.
On the other hand, the total effects of childhood poverty on the other two labor market
outcomes are larger for men. The total hours worked difference between not persistently
poor men and persistently poor men is 264.64 hours. Similarly, the number of weeks
unemployed difference between men who experienced persistent poverty during their
childhood and those who were not is 2.45 weeks. A similar comparison number of weeks
unemployed per annum
betweenwomenwhoexperiencedpersistentpovertyduringtheirchildhoodandthosewhowere
not is 1.86 weeks.
Table 2.10: Gender Disaggregated Impact of Childhood Poverty
Compariso
n
Effec
t
Labor Hours Weeks
Groups Income Worke
d
Unemploye
d
% of
ACM
E
107.0
1 18.73
64
% of
ACM
E 91.79 2.76 26.75
∗p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
2.5.5 Poverty at Different Stages of Child Development and Its
Effect
Childhood poverty considered in the previous section is persistent poverty in which a
child spends more than half of his or her childhood ages in poverty. It crosses more than
one stage of child development. Hence, the focus of the analysis in this section is to measure
the effect of childhood poverty confined to a single stage of child development. This enables
us to identify which stage is most critical for having a long-run effect. The previous stages of
child poverty
statusesareconsideredpretreatmentcovariateswhenestimatingtheeffectofchildhoodpoverty
for each stage of child development.
As the estimation results in Table 2.11 show, for all health and education-related
mediator outcomes, poverty during early childhood has the largest and significant negative
effect compared to poverty at any other stage of child development. The PSM estimation
results showed that the probability of having a healthy childhood and attaining a college
degree is significantly lower (by 16.04 percent and 11.19 percent, respectively) for those
who experienced persistent poverty during early childhood as compared to those who did
not.
Hence, the early childhood stage is critical regarding the long-run effects of poverty on
health and educational attainment.
Table 2.11: Effect of Childhood Poverty at Different Stages of Child Development
65
Healthy Childhood Years of Schooling College Degree
Standard errors in parentheses ∗p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
The causal mediation analysis in Table 2.12 also shows that poverty at all stages of child
development,exceptinfancy,hasasignificantreductioneffectonhoursworkedwithmagnitudes
of 117.5, 178, and 92.1 hours, respectively. Out of these total effects of childhood poverty,
35.74, 27.47, and 25.84 percent, respectively, channeled through its negative effect on
health and education (or the proportions are ACME or indirect effect). However, there is no
strong evidence that supports childhood poverty at these stages impacts adult labor
income. On the other hand, poverty during the middle childhood stage of child development
has a significant total effect on the number of weeks unemployed (1.18 weeks).
There is no strong evidence for the impact of poverty during infancy on adult hours
worked and the number of weeks unemployed. Similarly, there is no strong evidence for the
impact of poverty during early childhood and late childhood on the number of weeks
unemployed. Table 2.12 shows ACME estimation results for childhood poverty that lasts for
only a single stage of child development.
Table 2.12: ACME for Different Stages of Child Development
Labor Income Hours Worked Weeks Unemployed
66
†95% confidence intervals are in square brackets.
2.5.6 Sensitivity Analysis
The causal mediation estimation carried out in the previous section falls under the two
untestable assumptions: sequential ignorability and homogeneous treatment-mediator
interaction assumptions. This section deals with the sensitivity of the estimation results
when these assumptions are relaxed. One of the important parameters to make the
sensitivity analysis is ρ. It represents the correlation across the two error terms of outcome
and mediator equations. If the sequential ignorability assumption holds, all relevant pre-
treatment confounders have been conditioned on, thus ρequals zero. The nonzero values of
ρimply departures from the sequential ignorability assumption and that some hidden
confounder is biasing the ACME estimate.
The key concern in this study is an unmeasured confounder that affects both mediator
and outcome variables. For instance, a child’s ability can affect both his educational
67
achievement and labor market outcomes later in life. Any confounding of this type will be
reflected in the data-generating process as a correlation between the two error terms.
Ignoring this and estimating the two models separately will lead to a biased estimate of the
ACME. Thus, ρcan
serveasasensitivityparametersincelargerextremevaluesofρrepresentmoredeparturesfrom
the sequential ignorability assumption. Because the true parameter value of this coefficient
is unknown and its interpretation is relatively difficult, the coefficients of determinations of
the two models play equivalent roles. The first coefficient of determination is the
proportion of the
totalvarianceoflabormarketoutcomesthatwouldbeexplainedifweconsidertheheterogeneity
of the treatment–mediator interaction (childhood poverty and educational attainment) in
the regressionmodel(R2t).
Thesecondcoefficientofdeterminationistheproportionofunexplained variance explained by
an additional term for interaction heterogeneity (R2*).
For example, if a confounder is important in determining the level of education and the
labor market outcome measures, then the models excluding the confounder will have a
much smaller value of R2 compared to a model including the confounder. On the other hand,
if the confounder is unimportant, R2 will not be very different whether including or
excluding the variable. Thus, this relative change in R2 can be used as a sensitivity
parameter to check the robustness of ACME estimates.
The other parameter used for the sensitivity analysis is the standard deviation (SD) of
the individual level coefficient for the treatment-mediator interaction (childhood poverty
and educational achievement). Other things remaining equal, a low value of this upper
bound indicates a more robust estimate of the ACME since it leaves less room for an
unobserved confounder to bias the result.
68
Table 2.13 presents ACME’s sensitivity parameter estimates of the weighted average of
the treated (poor) and control (non-poor) groups. The other details of sensitivity analysis
for the treated and control groups’ estimates are annexed in Appendix A1.
Table 2.13: Sensitivity Parameter Estimates of ACME on Treated
σ(b)†R2*(b) R2t(b)
Labor Income 3,130 0.03 0.0288
Hours Worked 25.262 0.04 0.0396
Number weeks Unemployed 0.272 0.06 0.0569
† bs in the columns name refer to bounds of the sensitivity parameters.
The Labor income sensitivity parameter estimates show that ACME does not become
positive until σbecomes greater than 3,130, or 41.4 percent of its largest possible value
given the data (7,550). This implies that the ACME estimate still provides some support for
the important mechanism that links childhood poverty with labor market outcomes even
after we allow a certain degree of violation of the no interaction assumption. The
corresponding coefficients of determination estimates are 3 and 2.88 percent, respectively.
The estimated values are too small and they require unobserved confounders that are not
in the model would need to explain 97 percent or more of the remaining variation in
education (main mediator) and labor income (outcome) for the ACME to lose its statistical
significance.
The Hours Worked sensitivity parameter estimates show that ACME does not become
positiveuntil σbecomesgreaterthan25.262, or25percentofitslargestpossiblevalue(101.05).
The corresponding coefficients of determination are 4 and 3.96 percent, respectively. These
estimates are very small and require confounders that are not incorporated in the model
would need to explain 96 percent or more of the remaining variation in education (main
mediator) and hours worked (outcome) for the ACME to lose its statistical significance. This
69
is a strong indication of the robustness of ACME estimation. It provides a great deal of room
to relax both homogeneous interactions as well as sequential ignorability assumptions.
Hence, there is strong evidence for the negative causal mediation impact of poverty on
hours worked through its negative effect on childhood health and level of education.
A similar conclusion also can be drawn about the Number of Weeks Unemployed. It has
a sigma bound of 0.271, which is 24.49 percent of 1.12 (the largest possible value) and the
correspondingcoefficientsofdeterminationare6and5.96percent,respectively. Thisleavesless
room for unobserved confounders to bias the estimation. Hence, there is robust evidence
about the positive indirect effect (ACME) of childhood poverty on the number of weeks
unemployed which passes through its negative effect on health and education.
Figure 2.5 graphically shows the quantitative analyses discussed so far. The broken lines
indicate ACME estimates and the solid lines are sensitivity analysis parameters. The gray
area is the 95 percent confidence interval. The graphical details of the sensitivity analysis
for total, direct and indirect effects are annexed in Appendix A1 through A4.
70
Figure 2.5: ACME sensitivity analysis plots
The other sensitivity analysis considered in the study is done after altering the poverty
line. The 125 percent of the federal poverty line value is taken to differentiate the treated
(those who grew up in a household that had an annual family income below the threshold)
and control (who had above the threshold) groups. The propensity score matching estimate
of persistent
childhoodpovertyshowedthatchildhoodpovertysignificantlyandnegativelyaffectschildhood
healthandeducationalattainment(mediatoroutcomes). Table2.14showsthattheprobability of
having a healthy childhood for the group who grew up in poverty is 20 percent smaller than
the same probability for the matched control group (those who grew up non-poor).
Similarly, the likelihood of individuals in the treated group with a college degree is 15.45
percent less than the likelihood for those in the control group.
Table 2.14: Average Treatment Effect of Childhood Poverty on Health and Education
71
Healthy Childhood(%) Years of Schooling College Degree(%)
ATT -20∗∗∗ -0. 694∗∗∗ -15.45∗∗∗
(0.042) (0.159) (0.041)
Standard errors are in parentheses. ∗p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
The ACME estimates also show that higher values of the poverty line reinforce the
longrun impact of childhood poverty on labor market outcomes. As Table 2.15 presents,
growing up poor significantly and negatively impacts average annual labor income, which
amounts to a total of $14,672 difference between the treated (poor during childhood) and
control (nonpoor during childhood) groups. Out of this total labor income impact of
childhood poverty, 18.7 percent of its proportion is transmitted through its negative effect
on child health and education. Similarly, growing up poor has a negative impact on other
labor market outcomes. It has a significant and negative effect on average annual hours
worked and a significant and positive impact on the annual average number of weeks an
individual stays unemployed in a year. The figures are 159 hours and 1.99 weeks,
respectively.
Table 2.15: ACME and Total Effect of Childhood Poverty on Labor Market Outcomes
Effects Labor Hours Weeks
Income Worked Unemployed
A similar estimation of the effects of childhood poverty, using 125 percent of the federal
poverty line, is computed for the four stages of child development and presented in Table
2.16. The results show that all health and education-related mediator outcomes are
72
negatively affected by poverty during each stage of child development. Particularly,
childhood poverty
duringearlyandmiddlechildhoodstageshasconsistentlysignificantnegativeeffectscompared
to poverty that occurs in other stages of child development.
Table 2.16: Effect of Childhood Poverty at Different Stages of Child Development
Years of
Healthy Childhood Schooling College Degree
The causal mediation analysis results in Table 2.17 also show that poverty at all stages
of child development, with the exception of infancy, have a significant reduction effect on
hours worked with magnitudes of 90.5, 132.5 and 72.4 hours, respectively, in their
sequential order.
There is also strong evidence supporting the conclusion that poverty during the first three
stages of child development has a negative and significant impact on annual labor income.
Furthermore, poverty during the early and middle childhood stages of child development
has a positive and significant total effect of number of weeks unemployed (1.19 and 0.785
weeks,
respectively).
Table 2.17: ACME for Different Stages of Child Development
73
Labor Hours Weeks
Income Worked Unemployed
ACME-
3,340∗∗∗ -29.8∗∗∗
2.6 Discussion and Conclusion
UsingPSID’sintergenerationalsurveydata,thisstudyevaluatesthemediatorandlong-run
effects of growing up poor. Propensity score matching and ACME estimations were
employed to measure the interaction among childhood poverty, mediator variables, and
labor market outcomes. The propensity score matching estimates show that childhood
poverty significantly and negatively affects health status. This finding is consistent with
some of the literature reviewed in section 2 (Currie & Almond, 2011; Duncan et al., 1998;
Levy & Duncan, 2000;
Ziol-Guest et al., 2012).
Similarly, the findings of the effects of childhood poverty on years of education is
negative and significant. Furthermore, the propensity score matching estimates show
strong evidence that supports childhood poverty has a negative effect on higher-level
74
educational attainment. The probability of earning a college degree among those who grew
up poor is 12.45 percentage points less than the probability of earning a college degree
among those who grew up nonpoor. This result matches the findings of prior studies
(Chaudry & Wimer, 2016; Dahl & Lochner, 2012; Kubilius & Corwith, 2018). The estimated
results for linear and quadratic values of years of schooling reinforce the negative effect of
childhood poverty on educational attainment. The potential explanation for the mixed
result of childhood poverty on lower and higher levels of education may be the U.S.
education policy. The policy enables children to enroll in school without direct financial
outlay, but attending higher levels of education for those with disadvantaged economic
backgrounds is challenging.
The ACME estimation results showed that persistent childhood poverty has a strong and
negative total effect on annual labor income, annual hours worked, and the number of
weeks unemployed per annum. From the total effect of childhood poverty on the labor
market outcomes, a significant portion passes indirectly through the two mediator
variables: education and child health. Individuals who grew up poor have a lower average
annual labor income than those who did not grow up poor, which amounts to $11,252.
More than 42 percent of the total effect on annual labor income is channeled through the
negative effect of poverty on childhood health and educational attainment. Similarly, the
total effects on average hours worked and the number of weeks unemployed are 182.1
hours and 1.964 weeks, respectively, and 22.57 and 21.73 percent of the totals are
channeled through the negative effect of poverty on childhood health and educational
attainment.
Thesesignificantandadverseeffectsofchildhoodpovertyonemploymentandearningalign
with some of the literature reviewed for this study (Gregg et al., 1999; Lesner, 2018;
Ratcliffe
75
& McKernan 2010, 2012). Using data from various European countries, Bellani and Bia
(2019) showedthatchildhoodpovertyleadstolow-
incomeadulthoodandahigheraverageprobability of being poor. But, they used a single
dummy mediator in their analysis.
The gender disaggregation analysis based on the two possible combinations of gender
and poverty statuses (poor women vs. non-poor women and poor men vs. non-poor men)
showed annual labor income differences ($12,820) is larger for comparison between
women. On the other hand, the total effects of childhood poverty on both the other two
labor market outcomes are larger for men. The total hours worked difference between
persistently poor men and those whoarenotis264.64hours.
Similarly,thenumberofweeksunemployeddifferencebetweenmen
whoexperiencedpersistentpovertyduringtheirchildhoodandthosewhodidnotis2.45weeks.
The results show significant differences in labor market outcome gaps for poor and non-
poor comparison for women and men. It indicates the significant role gender plays in
explaining the long-run effect of childhood poverty on labor market outcomes.
Similar propensity score matching and average causal mediation analysis of childhood
poverty were carried out for different stages of child development. The PSM results showed
that experiencing poverty during early childhood significantly deteriorates a child’s health
and education capital. The results indicate that the early childhood stage is a critical period
as far as the impact of financial difficulty on mediator outcome variables is concerned. This
finding matches with studies based on the US data (Duncan et al., 2012; Duncan &
Magnuson, 2013; Ziol-Guest et al., 2012) and is contrary to Lesner’s (2018) findings which
are based on data from Denmark. Furthermore, the ACME estimation results showed that
childhood poverty that occurs during the three childhood stages has a negative impact on
hours worked both directly and indirectly through health and education.
76
Evaluating the transmission mechanism of childhood poverty to adult labor market
outcomes is important, as it informs policymakers about how to tailor interventions that
positively influence children’s health outcomes and educational achievement that in turn
improve adult labor market outcomes. ACME estimates and the transmission mechanism of
effect from childhood poverty to adult labor market outcomes indicate that policy
intervention on child health and education has both employment and labor income
implications.
CHAPTER 3
MINIMUM WAGE AND CHILD HEALTH: EVIDENCE FROM
THE 1966 FAIR LABOR STANDARDS ACT
3.1 Introduction
Intheminimumwageliterature,itiscommontoexaminetheeffectofminimumwagechange
ontheemploymentandearningsoflabormarketparticipants. However, theimpacttranscends
beyond the labor market outcome performances of participants. Due to its effect on
parents’ income, time allocation, and health investment capacity, minimum wage change
also impacts child health. This study examines the effect of minimum wage increases on
child health using the 1966 Fair Labor Standards Act minimum wage increase as an
exogenous variation.
Studies show that an increase in minimum wage positively affects birth weight. Komro
et al.’s (2016) findings show that increased state minimum wages are associated with
reduced low birth weight births and reduced post-neonatal infant deaths. Similarly, Wehby,
Dave, and
Kaestner’s(2020)studyontheeffectofminimumwageoninfanthealthrevealedthatanincrease
in the minimum wage that causes a $1000 increase in annual household income is
associated with 8.5 gram (0.3 % relative to the mean) increase in birth weight and 0.2
percentage point decrease in low birth weight. However, the extended effects of a higher
minimum wage on post-birth child health have not been well-studied.
Wehby, Kaestner, et al.’s (2020) study is the pioneer study on the child health impact of
minimum wage change using data from the National Survey of Children’s Health (NSCH).
78
The results show that an increase in the minimum wage throughout childhood is associated
with a significant improvement in child health. They limit the sample to children aged 6 to
17 years since one of the primary outcomes they examined, missed school due to
illness/injury, is only measured for children aged six and older. The other outcome variable
considered in this study is the children’s general health status reported by parents. They
divided the sample into two cohorts: children aged 6 to 12 and 13 to 17. They find that an
increase in the minimum wage throughout childhood is associated with a significant
improvement in child health for both cohorts.
Wehby, Kaestner, et al. (2020) compare parental report health outcomes of children in
the same state “exposed” to continuous minimum wage change. Given the disparity
between parental reports and self-report health status of children (Baca et al., 2010; Gough
Kenyon et al., 2021), this study examines the effect of the minimum wage introduced by the
federal government on retrospective self-report of childhood health.
Using the Panel Study of Income Dynamics (PSID)’s intergenerational survey data, this
study examines the effect of the minimum wage introduced by the 1966 Fair Labor
Standards Act on child health. The effective date of the minimum wage hike was February
1967, and this study usesthe term1966 minimumwageextension throughoutthe paper.
Toobtain estimates, I use a difference-in-differences research design. The treatment group
comprises individuals whose parents were working in the industries covered by the 1966
minimum wage extension. The control group includes individuals whose parents worked in
the industries covered by the 1938 Fair Labor Act. The findings show that the 1966
minimum wage extension significantly and positively affects child health.
The results from estimation using the overall sample and cohort and state fixed effect
show that the minimum wage introduction increases the proportion of children who have a
healthy childhood by 7.16 percentage points. The positive child health impact of the
79
minimum wage extension is not homogeneous across all states. The Fair Labor Standards
Act was a larger shock for states with no state minimum wage before the Act. The
estimation results, which
considertreatedindividualsonlyfromstateswhichhadnominimumwagebeforetheAct, show
that the child health improvement impact of the minimum wage policy is stronger than the
effect based on the overall sample. The intervention increases the probability of having
healthy childhood by 17.9 percentage points, which is roughly 10 percentage points higher
than the effect estimated based on the overall sample.
In light of previous literature claims of the heterogeneous effect of policy interventions
on child health, this study also investigates whether the childhood health effect of the
minimum wageextensiondiffersbetweenmalesandfemales.
Boththeoverallsampleandstronglytreated sample estimation results show that the health
improvement effect of the policy is positive for bothfemaleandmalesubsamples. However,
themagnitudeislargerforthefemalesub-sample, especially for the strongly treated states
analysis estimation result. The intervention increases the proportion of children having
healthy childhood by 23.2 percentage points for females and 12.7 percentage points for
males.
The estimation results are robust for the reclassification of children’s self-reported
general health status and the analysis that only considers children who have employed
fathers in the treated industries and stay-at-home mothers as a treatment group.
Thecontributionofthisstudytotheexistingliteratureisthreefold. First,itprovidescausal
evidence on how the introduction of a minimum wage policy can affect childhood health
using self-reportedgeneralhealthstatusasanoutcomevariable.
Itisthefirstofitskindusingafederal
minimumwagepolicyasapolicyvariabletoshowtheintergenerationaleffectofminimumwage
80
policyonchildhealth. Hence,thestudycontributestotheliteraturebecausetheminimumwage
affects income. Thus, the findings that a higher minimum wage is associated with better
child health could also be interpreted as indirect evidence of a causal effect of income on
child health. Hence it will contribute to the literature which focuses on income-enhancing
interventions or policies that aim to improve child health. Second, the 1966 reform was a
large shock to treated industries in states that did not have a state minimum wage—in
these states, the wage floor moved from 0 to the prevailing federal minimum wage, at a high
level in the late 1960s. Hence, the findings in this study show the difference in effect when
national minimum wage policies were implemented in two groups of states (those which
have state minimum wage and those which have not before the 1966 minimum wage
extension). The study revealed that the child health improvement effect of the minimum
wage is stronger in states which had no minimum wage before the 1966 minimum wage
extension. Third, the study also shows that the child health improvement effect of the
minimum wage is larger for females than males. This finding adds evidence to the literature
that examines the heterogeneous effect of policy intervention on the health outcomes of
females as compared with males.
The remaining part of this chapter is structured as follows: Section 2 reviews the major
literature. Section 3 describes the methodology of the study: data, descriptive statistics, and
identification strategy. Section 4 provides empirical results and robustness checks that
analyze the consistency of the effect of minimum wage introduction when the
dichotomization criterion of a childhood health dummy variable is relaxed and the variation
of dual or single parent employment status is considered. Section 5 discusses the major
findings and concludes the chapter.
81
3.2 Literature Review
3.2.1 Effect of Minimum Wage on Labor Market Outcomes
Previous research focused extensively/heavily on the impact of the minimum wage on
labor market outcomes such as employment and earnings. The effect of minimum wage on
employment is mixed. Some claim with confidence that a $15 minimum wage will not result
in job loss (Reich et al., 2016). Others argue that a $15 minimum wage will lead to huge job
losses (Macpherson & Even, 2019). Literature prior to 1992 used state-level panel data to
examine the relationship and found a negative effect of minimum wage on employment
(Belman & Wolfson, 2014). Similarly, Neumark and Wascher (1992) draw on the framework
of the earlier aggregate time-series research on teenagers to study a panel of state-year
observations and found results consistent with the earlier studies.
On the contrary, Card and Krueger (1994) use the 1992 increase in New Jersey’s
minimum wagetoconstructaquasi-experiment. Thetreatmentgroupconsistsoffast-
foodestablishments in New Jersey, and the control group consists of similar establishments
in nearby Pennsylvania counties. In April 1992, the minimum wage in New Jersey rose from
$4.25 to $5.05, while that in Pennsylvania remained constant at $4.25. Using the difference-
in-differences method, they found no loss of employment in response to the minimum wage
increase. As a result, they
concluded,contrarytoNeumarkandWascher(1992),thattheminimumwageisnotnecessarily
bad for employment. However, based on both a quick examination along quasi-
experimental lines and a more careful regression analysis over a longer period, Neumark
and Wascher (2000) conclude, “Taken as a whole, it is our view that the BLS data on
82
employment at eating and drinking places neither confirm nor reject our findings from the
payroll data that the New
Jersey minimum-wage increase appears to have reduced fast-food employment in that
state. The BLS data do, however, provide complementary evidence that minimum-wage
increases reduce employment in the restaurant industry” (pp. 1389–1390).
Unlike the effect on employment, the effect of minimum wage on earnings is consistent
among the literature. Aaronson et al. (2012) using data from the Consumer Expenditure
Survey (CEX), Survey of Income and Program Participation (SIPP) and Current Population
Survey (CPS) analyzed the effect of minimum wage on earnings, spending and debt. They
discovered four significant findings: First, a $1 minimum wage hike increases household
income by roughly $250 and spending by approximately $700 per quarter (in 2005 dollars)
in the year followingtheminimumwagehike. Second,
themajorityofthisadditionalspendingcomesfrom a small number of households purchasing
debt-financed new vehicles. Third, total spending increases within one quarter of the
minimum wage increase and not prior, despite legislation typically passing 6 to 18 months
before enactment. Finally, high levels of durables spending, and debt accumulation persist
for several quarters after a minimum wage hike. These results are robust to changes in
sample selection criteria and covariates. Furthermore, they find that a minimum wage hike
has no income or spending effect on households with workers earning at least double the
minimum wage, providing further evidence that their estimates are not the result of
omitted variables.
Belman et al. (2015) surveyed empirical research to examine “Who is affected by the
minimum wage?”. They concluded that, while almost universal agreement exists that the
minimum wage raises earnings, evidence for a negative employment effect ranges from
mixed to nonexistent. David et al. (2016), using data from the Current Population Survey
83
Merged
OutgoingRotationGroup(CPSMORG),foreachyearandIVmethodfoundthattheminimum
wagereducesinequalityinthelowertailofthewagedistribution. Furthermore, theyarguethat
the decline in the real value of the minimum wage explains 30 to 40 % of the rise in lower
tail wage inequality in the 1980s.
Dube (2019), using the Current Population Survey (CPS) between 1984 and 2013 and
the classic two-way fixed effects model to a model of regional shocks and business cycle
heterogeneity, found similar results. He showed a clear, positive effect of minimum wages
on family incomes below the 20th quantile. The largest impact occurs at the 10th and 15th
quantiles, where estimates from most specifications are statistically significant, and the
long-run minimum wage elasticities for these family income quantiles range between 0.152
and 0.430 depending on control sets. In the preferred (most saturated) specification, the
family income elasticities with respect to the minimum wage are around 0.359 and 0.332
for the 10th and 15th quantiles, respectively, and diminish close to zero by the 30th quantile.
Since the conventional income definition used for official poverty calculations does not
include tax credits such as the Earned Income Tax Credit (EITC), or non-cash transfers such
as Supplemental Nutritional Assistance Program (SNAP), he also estimates the impact using
an expanded income definition. After accounting for tax credits and non-cash transfers, the
minimum wage effect on the level of family incomes (i.e., the semi-elasticities) are about
66% as large for the bottom 30 percent of the distribution. Overall, the evidence clearly
points to at least moderate-income gains for low-income families resulting from minimum
wage increases. Similarly, a Congressional Budget Office (2019) analysis concluded that an
increase in the minimum wage to $12 would increase wages for as many as 11 million
workers.
84
Overall, the evidence on the labor market effects of the minimum wage, at least over the
rangeofincreasesoccurringinthelast20to30years,indicatesthatahigherminimumwagewill
raise wages and income among many low-skilled persons, though some of these gains may
be partly offset by modest dis-employment effects (Wehby, Kaestner, et al., 2020). The
important implication of this literature for child health is that an increase in the minimum
wage and, in turn, family income may have improved child health.
3.2.2 Effect of Minimum Wage on Child Health
Thereisverylittleliteratureontheeffectofminimumwageonchildhealth,andthemajority of
it focuses on adult and workers’ health. The two recent studies, Komro et al. (2016) and
Wehby, Dave, and Kaestner (2020), examine how minimum wage affects infant health and
find that higher minimum wage is associated with small increases in birth weight.
Komroetal.,(2016)estimatetheeffectsofstate-levelminimumwagelawsusingadifferencein-
differences approach on rates of low birthweight (< 2500 g) and post neonatal mortality
(28–364 days) by state and month from 1980 through 2011. All models included state and
year fixed effects as well as state-specific covariates. They found evidence that a dollar
increase in state-level minimum wage, above the federal level, was associated with a 1% to
2% decrease in low birth weight births and a 4% decrease in post neonatal mortality.
Wehby, Dave, and Kaestner (2020), using data on the universe of births in the US over 24
yearsandadifference-in-differencesresearchdesign, findthatanincreaseintheminimumwage
is associated with an increase in birth weight driven by increased gestational length and
fetal growth rate. Their minimum wage estimates suggest that a $1000 change in annual
earnings in the two years prior to birth is associated with an 8.5 gram (0.3 %) increase in
85
birth weight. They also find an increase in prenatal care use and a decline in smoking
during pregnancy are some channels through which minimum wage can affect infant health.
Wehby, Kaestner, et al. (2020) is a pioneer study to examine the impact of minimum
wage on child health. They examine the effects of the minimum wage on child health using
data from the National Survey of Children’s Health (NSCH) and the difference-in-differences
empirical method. The difference-in-differences research design compares children in the
same state who were “exposed” to different minimum wages at specific periods of their
childhood while accounting for the state, birth cohort, and age at interview effects.
They found that an increase in the minimum wage throughout childhood is associated
with a significant improvement in child health. They limit the sample to children ages 6 to
17 years becauseoneoftheprimaryoutcomestheyexamined,missedschoolduetoillness/
injury,isonly measured for children aged 6 and older. The other outcome variable is parent-
reported general health status. They divided the sample into two cohorts: children aged 6
to 12 and 13 to 17. The estimation result of the effects of minimum wage on the health of
children aged 6 to 12 showed a $1 increase in the minimum wage during pregnancy is
associated with a 24% decrease in the likelihood of fair/poor rated health. For minimum
wage changes during ages 0-5 years, estimates are more consistently indicative of a
beneficial effect. A $1 increase in the minimum wage in each of these five years is associated
with a 0.11 (2.7%) improvement in general health
(onthefive-categoryscale)anda6.2percentagepoint(8.7%)increaseintheprobabilityofvery
good or excellent rated health. A $1 increase in the minimum wage at each of these ages is
also associated with a 3.8 percentage point (14%) decrease in the 3-question index
measure of poor health and a 0.57 (15.6%) decrease in missed school days. Note that these
are estimated
treatmenteffectsassociatedwithanaverageincreaseof$1intheminimumwageoverthechild’s
86
early life course, between the ages of 0 to 5, not that of a one-time increase at a single year
of age.
They also calculated the sum of the coefficients on the minimum wage variables across
all ages. These estimates measure the cumulative effect of a $1 change in the minimum
wage in each year of the child’s life: pregnancy, ages 0 to 5, and ages 6 to child’s current age,
not a one-year-only increase by $1. A $1 increase in the minimum wage throughout a child’s
life is associated with a 0.18 unit (4.4%) improvement in general health; a 7 percentage
point (10%) increaseintheprobabilityofverygoodorexcellentratedhealth;
an8.3percentagepoint(30%) decrease in the 3-question index of poor health; and 0.95
(26%) fewer missed school days. All of these estimates are statistically significant.
Similarestimationforeffectsofminimumwageonthehealthofchildrenaged13to17showed
that a $1 increase in the minimum wage in every year between ages 0 to 5 is associated
with an approximately 11% increase in the probability of excellent/very good rated health,
and a 90% decrease in the probability of poor/fair rated health. A $1 increase in the
minimum wage at these ages is also associated with an 8.2 percentage point (42%)
decrease in the 3-question index of poor health.
Similarly, a $1 increase in the minimum wage throughout childhood is associated with a
57% decrease in the 3-question index of poor health and a 42% decline in missed school
days. As noted before, this is the effect of a $1 increase in the average minimum wage over
the child’s life course and not a one-time (or one-year) increase by $1. Other estimates of
the cumulative effect of the minimum wage on child health, while insignificant, also suggest
improved health except for the case of BMI. Their study, however, was confined to a state-
level comparison of
childhealthresponsetocontinuouschangeinminimumwageusingparentalreportchildhealth
as an outcome variable. Given the disparity between parental reports and self-report health
87
statusofchildren(Bacaetal,2010;GoughKenyonetal.,2021),thisstudyexaminestheeffectof
minimum wage introduced by the federal government on retrospective self-report of
childhood health.
3.2.3 The 1966 Fair Labor Standards Act
In their historicalaccount of the Fair Labor StandardsAct (FLSA), Willis (1971) and Ruby
and Eisenbrey (2013) explained how the Act was established as well as its initial objectives
and achievements. The FLSA’s initial objectives were to improve job quality by abolishing
child labor and creating new jobs for millions of the nation’s unemployed population by
reducing overtime work. As the Great Depression endured, firms not only laid off hundreds
of thousands of workers, but also implemented significant wage rate cuts. Despite low
wages, or perhaps because of them, many workers (including children) continued to work
long hours in unjust conditions. During his first re-election campaign, President Roosevelt
publicly committed to eliminating child labor and improving labor standards for all
working Americans.
TheultimateversionoftheFairLaborStandardsAct,signedintolawbyPresidentRoosevelt on
June 25, 1938, established a 25-cent minimum wage (that would rise to 30 cents beginning
in October 1939), introduced a 44-hour maximum work week (that would first fall to 42
hours in October 1939 and would then fall to 40 hours in October 1940), and set the
general age of workforce entry at 16.
The Fair Labor Standards Act was an unequivocal success. Leiter (1962) reported that
the Act covered approximately 10,670,000 workers in September 1938, accounting for
roughly one-third of nonsupervisory wage and salary employees or one-quarter of the
88
employed civilian labor force in the United States. Seven months later, in April 1939, the
number of people covered was estimated to be 12,291,000; of these, 7,657,500, or more
than 60%, were employed inmanufacturing. Atthetime,
coverageappliedtoapproximately27%oftheemployedcivilian labor force. It also put adult
Americans back to work and guaranteed that they would be
treatedandcompensatedmorefairly. TheFairLaborStandardsActimprovedlaborstandards
and actual working conditions, a result that continues to improve the daily lives of millions
of working Americans.
The coverage of the Act, however, was incomplete: several sectors were excluded.
Derenoncourt and Montialoux (2021) explained that the 1938 FLSA covered about 54% of
the U.S. workforce in manufacturing, transportation and communication, wholesale trade,
finance, insurance, real estate, business and repair service, and public administration
sectors.
President Roosevelt intended to cover the economy as a whole but faced resistance in
Congress, particularly from Southern Democrats (Phelps, 1939). The law enacted in 1938
stipulates that only employees engaged in interstate commerce or the production of goods
for interstate commerce be covered (Daugherty, 1939).
Over time, a series of amendments to the 1938 FLSA extended the minimum wage to the
rest of the economy. Willis (1971) argued that when Congress passed the 1961
amendments to the Fair Labor Standards Act, it laid the groundwork for future extensions
of the Act’s coverage. The Fair Labor Standards Amendments of 1966 made changes to the
restrictions on the types of ”enterprises” as well as the dollar-volume test, and those
restrictions were eliminated or reduced. As a result, the Act’s coverage has been expanded
to include over eight million additional employees.
89
In this study, I focus on the 1966 FLSA amendments, which is an expansion of the federal
minimum wage to other sectors which were not incorporated in the previous amendments.
The 1966 FLSA amendments introduced the federal minimum wage (as of February 1,
1967) in the following sectors: agriculture, nursing homes, laundries, hotels, restaurants,
schools, hospitals, and recreation and entertainment services. The minimum wage enacted
in these sectors in 1967 ($1 in nominal terms) was initially below the federal minimum
wage but converged to the level of the federal minimum wage by 1971, except in agriculture
where convergence was only complete in 1977 . As a result, the ratio between the federal
minimum wage and the median wage continued to increase in the newly covered sectors
over time and reached 40% to
50% during the 1970s, a level close to the one seen in the industries that were covered in
1938.
DerenoncourtandMontialoux(2021)figuredoutthatthesectorscoveredinthe1966minimum
wage extension employed about 8 million workers in 1967, or 21% of the U.S. workforce. In
this study these sectors (industries) are considered treated industries whereas industries
that were covered in the 1938 amendment are considered as control industries. Hence, the
study compares childhood health status of children whose parents were working in the two
groups of industries before and after 1966 Fair Labor Standards Act amendment.
The 1966 reform was a large shock to treated industries in states that did not have a
state minimumwage—
inthesestates,thewagefloormovedfrom0totheprevailingfederalminimum wage, at a high
level in the late 1960s. Previous literature analyzed the effect of the shock on employment
and inequality. Bailey et al. (2021) investigate how the high nationwide minimum wage
mandated by the Act affected employment. Their results indicate that there is little
evidence of disemployment effects, neither overall nor for particular subgroups of the
90
population. Derenoncourt and Montialoux (2021) also found a consistent result. To date,
however, the effect of the Fair Labor Standards Act minimum wage enactment on children’s
healthhasnotbeeninvestigatedusingacausalresearchdesign. Hence,thisstudyaimstobridge
the gap and contributes to expansive literature on the economic effects of the minimum
wage. and contributes to the literature on the economic effects of the minimum wage.
3.3 Empirical Methods
3.3.1 Data
This study also uses the PSID data to assess the impact of the Fair Labor Standards Act’s
minimumwagepolicyonchildren’shealth. ThelatestPSIDdatahas41panelwavesfrom1968to
2019andhasatotalof82,576individuals. BecausethePSIDdatasetcontainsintergenerational
information on parental labor market outcomes and child health, it is ideal for studying the
child health impact of the 1966 Fair Labor Standards Act, which established a minimum
wage for industries that were not covered before. The effective date of the Act was February
1967
andthisstudyusestheterm”the1966minimumwageextension”torefertotheminimumwage
policy.
The PSID dataset was established in 1968 and for each year panel wave, it collects
parental labor market outcome information from a prior year and other household
information from the same year. In addition, it has retrospective parental labor market
outcomes information.
However,asthenumberofyearspriortothepolicyincreasesintheretrospectivedatacollection,
the sample size and information decrease significantly. To address this issue, the study was
91
confined to data from the six years preceding 1967 as a pre-policy period, and from post-
1966 minimum wage extension years as a post-policy period. Hence children who spent
their early childhood years in the six years prior to 1967 (children who were born between
1950 to 1961) were considered members of pre-policy treated and control groups as per
industries where their parents were working. Individuals born and raised after the
minimum wage extension fully converged (born between 1971 and 2001) are classified as
post-policy treatment and control groups based on the industries in which their parents
worked. This sample size adjustment does not jeopardize the sample design since empirical
evidence shows that household incomeenhancing exogenous policy interventions are more
effective in the early childhood period than
inthelateryears(Bragaetal.,2020;Wehby,Kaestner,etal.,2020). Inaddition,Iselectasample of
children in less educated families following Wehby, Kaestner, et al. (2020) recommendation
to focus on children most likely affected by the minimum wage. Hence, the sample in this
study excludes children whose parents’ educational attainment is a college degree and
above.
Limitingthesampleasexplainedaboveyields5,153individuals,ofwhom48.57%arefemale.
The treated individuals are those whose parents worked in the industries covered by the
1966 minimum wage extension. These industries are agriculture, personal service,
entertainment andrecreationservice,andprofessionalandrelatedservices.
Thecontrolgroupsareindividuals
whoseparentswereworkinginthebaseindustrieswhichwerecoveredbythe1938FLSA.These
industries include mining, manufacturing, transportation, communication, wholesale,
finance, insurance, real estate, repair service, and public administration.
To address different concerns, the sample size varies for the robustness checks: for
example,
92
tocheckthesensitivityoftheresultforadjustingthetreatmentgroup,whichincludesindividuals
living in a state with no state minimum wage.
Table3.1presentsdefinitionsoftheanalyticalvariablesusedfortheanalyses. Thevariables
are extracted from the PSID survey research center data set. The data provides information
about individual characteristics: race, gender, and childhood health condition. In addition,
the second group of variables provides information about parental characteristics: parents’
educational level, their Home Ownership status, and the household family size.
Childhood health is a retrospective self-evaluation using the standard 5-point scale
(excellent, very good, good, fair, or poor) of the general state of one’s health when one was
less than 17 years old. The children report their retrospective childhood health status after
they become adults because the PSID follows them in later panel waves. The healthy
childhood dummy variable is created with a value of 1 for excellent, very good or good
childhood health and 0 otherwise.
Table 3.1: Definitions of Major Variables
Variable Definition
Healthy childhood =1 if the respondent has reported an excellent, very good
or good childhood health status, 0 otherwise
Mother HS graduate =1 if the mother has completed high school, 0 otherwise
Father HS graduate =1 if the father has completed high school, 0 otherwise
Mother college dropout =1 if the mother has some years of college, 0 otherwise
Father college dropout =1 if the father has some years of college, 0 otherwise
Mother School dropout =1 if the mother has not completed high-school, 0 otherwise
Father School dropout =1 if the father has not completed high-school, 0 otherwise
Home Ownership =1 if parents own their own home, 0 otherwise
Family size Respondent’s family size
Male =1 if the respondent is male, 0 if the respondent is female.
White =1 if the respondent is White, 0 otherwise.
Black =1 if the respondent is Black, 0 otherwise
Other =1 if the respondent is other than White or Black, 0 otherwise
93
State The state where the child lived during early childhood
Year Year of birth of an individual
3.3.2 Descriptive Statistics of Major Variables
The criteria used to select the sample yield a total of 5,153 observations. 1296 of the
observationsareinthetreatmentgroupandoutofwhich554ofthemspenttheirearlychildhood in
the pre-policy period and the remaining 742 spent their childhood in the post-policy
period. Similarly, the control group has 1,127 and 2,730 individuals in the pre and post-
policy period, respectively. The sample does not include children with one parent who
works in the treated industriesandtheotherwhoworksinthecontrolindustries.
Inaddition,childrenwhohaveone parent working in either of the two groups of industries,
whereas the other one works outside of those industries are also excluded.
Table 3.2 presents the average value and standard deviations of the parental and
children’s variables of the control and treatment groups. Before the 1966 minimum wage
extension was implemented, the proportion of individuals who had a healthy childhood is
larger for the control group (42.8 % versus 34.1 %). However, after the 1966 minimum
wage extension was implemented, the proportion of individuals with a healthy childhood
became more balanced for the two groups.
Table 3.2: Summary Statistics for Major Variables
Variables Pre 1966 FLSA Post 1966 FLSA
Treated Control Treated Control
Mean SD Mean SD Mean SD Mean SD
Healthy Childhood 0.341 0.475 0.428 0.495 0.520 0.500 0.539 0.499
Mother HS graduate 0.271 0.445 0.423 0.494 0.280 0.449 0.449 0.498
Father HS graduate 0.152 0.359 0.275 0.447 0.379 0.485 0.500 0.500
Mother college dropout 0.038 0.191 0.047 0.212 0.063 0.244 0.093 0.291
Father college dropout 0.047 0.212 0.085 0.279 0.135 0.342 0.150 0.357
94
Mother School dropout 0.691 0.462 0.530 0.499 0.584 0.493 0.359 0.480
Father School dropout 0.801 0.399 0.640 0.480 0.375 0.484 0.240 0.427
Home Ownership 0.419 0.494 0.513 0.500 0.441 0.497 0.592 0.492
Family size 4.574 1.959 4.683 2.374 4.766 1.686 4.702 1.519
Male 0.520 0.500 0.488 0.500 0.569 0.496 0.509 0.500
White 0.202 0.402 0.570 0.495 0.336 0.473 0.538 0.499
Black 0.747 0.435 0.400 0.490 0.600 0.490 0.421 0.494
Other 0.051 0.219 0.030 0.171 0.065 0.246 0.041 0.198
N 554 1,127 742 2,730
3.3.3 Empirical Model
This section explains the identification methods that estimate the effect of minimum
wage policy on childhood health. The empirical method identifies the effect by comparing
the
selfreportedgeneralchildhoodhealthstatusofindividualswhospenttheirearlychildhoodperiod
s in the treated and control groups.
3.3.3.1 Difference-in-Differences Regression
Icomparehealthoutcomesforchildrenwhospenttheirfirstsixyearsasminorsinhouseholds
where parents worked in the treated industries versus in the control industries before and
after the 1966 minimum wage extension. To this end, the study employs a difference-in-
differences research design. This specification allows to compare the effect of an increase in
household incomes that resulted from the 1966 minimum wage extension on child health
outcomes for the treated individuals. Hence, the younger age cohort who were born after
the phase completion of the Fair Labor Standards Act function as the “after- treatment”
cases. The oldest cohort of children who were born and spent their early childhood period
before the implementation of the 1966 minimum wage extension function as the “before-
95
treatment” case. Hence, the study
focusesonmeasuringthetreatmenteffectofanincreaseinhouseholdincomeaftertheminimum
wage extension on children’s health outcomes.
(3.1)
In the equation above, Hi is the outcome variable of interest in terms of an individual’s
retrospective self-reported general childhood health status. The unit of observation, i, is the
child. Childhood health status, Hi, is a binary variable constructed from the Likert scale
retrospective self-reported childhood health status reported by children after they become
adults. The variable Hi takes a value of 1 if the respondent reports excellent, very good, or
good childhood health and 0 if they report it poor or fair. Alternative dichotomization is
based on Smith’s (2009) classification of five-category scale health status for excellent or
very good health versus less (good, fair or poor) to do a robustness check of the above
specification. Potsi is a dummy variable equal to 1 to indicate the individual was born and
spent her or his early childhood years after the 1966 minimum wage extension increment
fully converges (i.e., the child was born after 1971). Di is a dummy variable with a value
equal to 1, which indicates the treatment group comprising children whose parents were
working in the treated industry, and equal to 0, which indicates a control group comprising
children whose parents were working in the control industry. The vector X contains
household and individual characteristics that include the sex of the child, family size, race,
and parents’ education levels.
In this model, parameter β1 is the coefficient of the treatment variable, which captures
the estimated mean difference in H (childhood health) between the treatment and control
groups priortothe1966minimumwageextension.
96
Itrepresentswhatever”baseline”differencesexisted between the groups before the
intervention was applied to the treatment group. Parameter β2
capturestheexpectedmeanchangeinchildhoodhealthoutcomefrombeforetoaftertheonsetof
theinterventioneraamongthecontrolgroup. Itreflectsthepureeffectofthepassageoftimein the
absence of the actual intervention. Parameter β3 is the interaction term coefficient between
Di and Posti or the difference-in-differences (DID) estimator, which captures the expected
mean change in outcome from before to after in the two groups. It is the coefficient of
interest for this research and measures the effect of the minimum wage increase on
children’s general health status. The model also includes state-fixed effects (σs), and year
(cohort) fixed effects
(τt).
This difference-in-differences estimation requires fulfilling the parallel trend
assumption. A visual test of parallel trend assumption is not possible here due to single
period outcome variable information prior to the minimum wage policy implementation
period. However, the healthy childhood variable mean value is larger for the control group
until the minimum wage fully converges. This indicates the assumption holds for this
empirical analysis. The difference-in-differencesestimationisconductedinthenextsection.
Itcomparesthechildhood health status of individuals whose parents worked in the treated
industry throughout their early childhood years with children whose parents worked in the
control industries in those years while accounting for state and birth (cohort) fixed effects.
And the robust standard error is used to deal with the heteroscedasticity problem of the
linear probability model.
97
3.4 Empirical Results
3.4.1 Effect of Minimum Wage on Child Health
Table3.3demonstratestheresultsfromthedifference-in-differencesestimationoftheeffect
of the minimum wage policy of the 1966 Fair Labor Standards Act on children’s general
health status. This estimation’s main coefficient of interest is the interaction term (DID)
between the treatmentindicatordummyandthepost-policyperiodindicatordummy.
Thelinearprobability model estimation results reveal that the implementation of the 1966
minimum wage extension
causesapositiveandsignificanteffectonchildhealth(dummyvariableequalto1ifthechildhas
good, very good or excellent childhood health). In the absence of any controls or fixed
effects, the1966minimumwageextensionhadapositiveandsignificanteffectonchildhealth.
Itshows
thattheFairLaborStandardsActminimumwagepolicyinterventionincreasestheprobability of
having healthy childhood by 6.8 percentage points. The coefficients’ magnitudes slightly
decline and are still significant once we control for state and cohort fixed effects. Again, the
estimated coefficients remain comparable when we control for individual, year group
(cohort), andstatecharacteristics.
Theestimationresultspresentedincolumn4,whichcontrolindividual and household
characteristics with cohort and state fixed effect, show that the 1966 minimum
wageextensionincreasesthelikelihoodofhavingahealthychildhoodby7.16percentagepoints.
Table 3.3: Difference-in-Differences Estimation on Minimum Wage Effect on Child Health
(1) (2) (3) (4)
DID 0.0679** 0.0733** 0.0611* 0.0716**
98
(0.0324) (0.0320) (0.0313) (0.0310)
Treatment -0.0865*** -0.0646** -0.0663*** -0.0570**
(0.0250) (0.0253) (0.0253) (0.0254)
Post 0.111*** 0.0913*** 0.551*** 0.545***
(0.0176) (0.0176) (0.0246) (0.0251)
White 0.350*** 0.173***
(0.0255) (0.0283)
Black 0.351*** 0.207***
(0.0246) (0.0288)
Male -0.0415*** -
0.0444***
(0.0136) (0.0128)
Mother HS graduate 0.0178 0.0463***
(0.0154) (0.0147)
Father HS graduate 0.0911*** 0.0791***
(0.0161) (0.0151)
Mother some college 0.0296 0.0932***
(0.0292) (0.0291)
Father some college 0.0353 0.113***
(0.0242) (0.0243)
Family Size -0.0166*** -
0.0184***
(0.00373) (0.00358)
Constant 0.428*** 0.149*** 0.403*** 0.240*
(0.0147) (0.0351) (0.133) (0.139)
Cohort Fixed Effect NO NO YES YES
99
State Fixed Effect NO NO YES YES
N 5153 5153 5153 5153
Robust standard errors in parentheses
∗p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
Overall,theresultssofarshowsomeevidencethatchildhoodgeneralhealthstatusimproved
for children in the treated group i.e., children whose parents work in the industries
embraced by the 1966 minimum wage extension. The table also reports coefficients of the
variables capturing individual characteristics, the state where the child grew up and cohort
indicator year. This helps to identify whether these characteristics matter in explaining an
individual’s overall childhood health status.
3.4.2 The Effect of the Minimum Wage in Strongly Treated
States
Like they do today, in the 1960s some states had their own minimum wage laws (on top
of the federal minimum wage) while others did not. This section presents the effect of the
1966 minimum wage extension on child health by adjusting the treatment group into a
group that includes individuals only from strongly treated states. In the context of this
study, strongly treated indicates the treated individuals from states which had no state
minimum wage prior to the 1966 minimum wage extension. Because the federal minimum
wage was high in the late
1960s(muchhigherthantodayrelativetothemedianwage),the1966reformwasaparticularly
large shock in the strongly treated states. The estimation results presented in Table 3.4
100
show that the child health improvement impact of the minimum wage policy is stronger
than the effect estimated based on the overall sample discussed in the previous section.
Theestimationresultsincolumn4,whichtakecohortandtimefixedeffectintoaccountshow
thattheFairLaborStandardsActminimumwagepolicyinterventionincreasestheprobability of
children having healthy childhood by 17.9 percentage points. This is roughly 10 percentage
points higher than the overall estimation.
Table 3.4: Difference-in-Differences Estimation for Strongly Treated States
(1) (2) (3) (4)
DID 0.193*** 0.195*** 0.164*** 0.179***
(0.0486) (0.0483) (0.0468) (0.0466)
Treatment -0.109*** -0.109*** -0.109*** -0.113***
(0.0342) (0.0347) (0.0381) (0.0383)
Post 0.111*** 0.0886*** 0.541*** 0.537***
(0.0176) (0.0177) (0.0240) (0.0253)
White 0.344*** 0.186***
(0.0304) (0.0348)
Black 0.356*** 0.240***
(0.0303) (0.0365)
Male -0.0342** -
0.0372***
(0.0149) (0.0141)
Mother HS graduate 0.0194 0.0403**
(0.0169) (0.0162)
Father HS graduate 0.0953*** 0.0852***
(0.0175) (0.0165)
Mother some college 0.0208 0.0788**
101
(0.0314) (0.0315)
Father some college 0.0622** 0.136***
(0.0263) (0.0264)
Family Size -0.0211*** -
0.0231***
(0.00400) (0.00389
)
Constant 0.428*** 0.164*** 0.351** 0.183
(0.0147) (0.0395) (0.148) (0.159)
Cohort Fixed Effect NO NO YES YES
State Fixed Effect NO NO YES YES
N 4298 4298 4298 4298
Robust standard errors in parentheses
∗p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
3.4.3 Heterogeneous Effects
Empiricalevidenceongenderdifferencesinchildhoodhealthshowsthatboysperformedless
wellthangirlsinvarioushealthindicators. Gissleretal. (1999),intheirstudyonchildrenunder
the age of 7 years, show that boys had a 64 percent higher cumulative incidence of asthma,
a 43 percent higher cumulative incidence of intellectual disability, and a 22 percent higher
incidence of mortality. The healthcare-related indicators also showed poorer health for
boys, who had a
37percenthighermeanofhospitaldays,a28percenthigherriskforreceivingsocialbenefitsdue to
health problems and a 13 percent higher risk for long-term medication. There is also strong
evidence (Kelder et al., 1995; Mu¨ller et al., 2005; Perry et al., 1998; Plachta-Danielzik et al.,
2007) showing that policy interventions targeted at reducing the prevalence of obesity
among children are more effective for some groups of children than others: girls are more
102
responsive thanboys.
Inlightofthisclaim,Iconductanadditionalanalysisseparatelyformaleandfemale
subgroupsandinvestigatewhethertheminimumwagepolicyhassimilarheterogeneouseffects.
First, I investigate whether the childhood health effect of the Fair Labor Standards Act
differs between males and females for the overall sample. The results are presented in the
secondandthirdcolumnsofTable3.5. Theresultsshowthatthereissomeevidencesupporting
the hypothesis that claims female health outcome response for the intervention is higher
than male. It shows that the policy intervention effect is positive and marginally
insignificant for males. In contrast, the same estimation results for females show that the
Fair Labor Standards Act minimum wage policy intervention increases the probability of
having healthy childhood by 8.58 percentage points.
The second investigation checks similar heterogeneity of childhood health outcome
effect of the intervention for males and females using the subsample which includes
individuals from the strongly treated state as a treatment group. The results in columns 4
and 5 of Table 3.5 show that the effects correspond to the aforementioned claims. There is
strong evidence that the health improvement effect of the minimum wage policy is
significantly larger for females. The minimum wage increases the probability of having
healthy childhood by 23.2 percentage points for females. A similar estimation for male
counterparts shows a significant and positive effect of minimum wage on child health, but
the magnitude is smaller (12.7 percentage points).
Overall,heterogeneouseffectsconsiderationisessentialwhenweassesstheeffectofminimu
m wageandotherincome-enhancingpoliciesonchildhealth. Tosummarizetheresults,Ifindthat
child health is significantly improved for children who spent their early childhood in the
new minimum wage regime, more so for girls than boys.
103
Table 3.5: Heterogeneous Effect of Minimum Wage on Child Health
Overall Sample Strongly Treated
Male Female Male Female
DID 0.0602 0.0858* 0.127** 0.232***
(0.0419) (0.0466) (0.0610) (0.0720)
Treatment -0.0626* -0.0480 -0.0910* -0.120**
(0.0348) (0.0374) (0.0509) (0.0584)
Post 0.608*** 0.468*** 0.617*** 0.440***
(0.0336) (0.0378) (0.0290) (0.0428)
White 0.231*** 0.122*** 0.270*** 0.108**
(0.0366) (0.0439) (0.0434) (0.0542)
Black 0.236*** 0.186*** 0.295*** 0.194***
(0.0365) (0.0454) (0.0451) (0.0572)
Mother HS graduate 0.0188 0.0854*** 0.00858 0.0878***
(0.0205) (0.0218) (0.0226) (0.0240)
Father HS graduate 0.0605*** 0.0958*** 0.0509** 0.116***
(0.0210) (0.0222) (0.0230) (0.0241)
Mother some college 0.110*** 0.0848** 0.0914** 0.0793*
(0.0417) (0.0414) (0.0457) (0.0443)
Father some college 0.0949*** 0.121*** 0.105*** 0.157***
(0.0339) (0.0351) (0.0374) (0.0380)
Family Size -0.0144*** -0.0219*** -0.0181*** -
0.0267***
(0.00472) (0.00542) (0.00516) (0.00590)
Constant 0.192 0.0972 0.127 0.0123
(0.177) (0.142) (0.198) (0.158)
Cohort Fixed Effect YES YES YES YES
State Fixed Effect YES YES YES YES
104
N 2650 2503 2192 2106
Robust standard errors in parentheses
∗p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
3.4.4 Robustness Check
The empirical findings discussed in the previous section are robust for an alternative
definition of childhood health dummy variable and alternative consideration of parental
employment in the treated industries.
3.4.4.1 Alternative Definition of Childhood Health Status
Formanyofthesurveydata,includingthePSID,self-reportedhealthstatusisaLikertscale
variable ranging from very poor to excellent health status. Finnas et al. (2008)
demonstrated that the self-reported health status variable on the five-point Likert scale can
be dichotomized. However, there are certain methodological concerns that need to be
addressed. Their empirical findings, which were based on data from Finland, demonstrated
that the cut-off points for bad versus good self-reported health, as well as the decision
about where to position moderate self-reported health status, are not affected by age.
However, when the categorization of poor self-reportedhealth excludes moderateself-
reportedhealth, the covariateof marital statusand educational level were found to be highly
age dependent.
Similarly,Bourne’s(2009)findingsshowthatthecut-offofthedichotomizationhasdifferent
effects for males and females. When the cut-off point for men includes moderate health
status, the impact of covariates on self-reported illness is greater than when the cut-off
point is only poor or very poor health status. Embedded in this finding is the vast difference
that is created by merely changing the cut-off point from poor health status to moderate-to-
105
very poor health statusformales. However,thisdisparitydoesnotemergeforfemales.
Hence,thissectionchecks the responsiveness of the previous results when the cut-off of the
healthy childhood dummy variable varies. The dummy variable healthy childhood takes a
value 1 when the self-reported health status during childhood is reported as very good or
excellent and 0 otherwise.
ThefindingsinTable3.6showthatthealternativedefinitionofthedependentvariablesdoes
not affect the results. For the full sample, the estimation results in column 2 which takes
cohort and time-fixed effect into account show that the Fair Labor Standards Act minimum
wage policy intervention increases the probability of having a healthy childhood by 7.14
percentage points. This result is consistent with the previous results based on
dichotomization of the healthy childhood dummy variable which gets value 1 when the self-
reported childhood health status is excellent, very good or good and 0 otherwise.
For the strongly treated sample, the estimation results in column 3 show that the
minimum wage policy intervention increases the likelihood of having a healthy childhood
by 17.16 percentage points. These findings are also consistent with the previous results
which are estimated based on the overall sample.
Table 3.6: Estimation for Alternative Definition of Childhood Health Dummy
Main Model Strongly Treated
DID 0.0714** 0.176***
(0.0312) (0.0471)
Treatment -0.0606** -0.109***
(0.0255) (0.0384)
Post 0.552*** 0.544***
(0.0258) (0.0260)
White 0.170*** 0.184***
(0.0282) (0.0346)
106
Black 0.197*** 0.231***
(0.0287) (0.0364)
Male -0.0430*** -0.0359**
(0.0129) (0.0142)
Mother HS graduate 0.0498*** 0.0418**
(0.0149) (0.0164)
Father HS graduate 0.0803*** 0.0877***
(0.0153) (0.0167)
Mother some college 0.0922*** 0.0749**
(0.0294) (0.0318)
Father some college 0.116*** 0.139***
(0.0244) (0.0266)
Family Size -0.0196*** -0.0239***
(0.00359) (0.00390)
Constant 0.247* 0.189
(0.139) (0.159)
Cohort Fixed Effect YES YES
State Fixed Effect YES YES
N 5073 4237
Robust standard errors in parentheses
∗p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
3.4.4.2 One Parent Employment
Intheprevioussectionanalysis,thetreatmentgroupincludeschildrenwhohaveatleastone
parent working in the treated industry: i.e., includes those who have both parents working
in the treated industry or one of them is a stay-at-home parent and the other works in the
treated industries. The same categorization is applied for control group children. However,
107
the health
outcomesofchildrenwhenbothparentsareworkingversusonlyoneparentworksandtheother is
a stay-at-home parent are not similar.
From the standpoint of a child’s mental and physical health, dual-parent employment
has both benefits and drawbacks. With the income from double sources, parents find
themselves more able to make choices for their families when it comes to nutrition and
education. Studies
showthenegativesideofdualparentemployment,particularlyduetomothers’employment. A
handfulofstudieshavediscoveredsignificantlinksbetweenmaternalemploymentandchildren’s
weight,withmothersworkinglongerhoursorfull-timebeingassociatedwithanincreasedriskof
obesity in their children (Brown et al., 2010; Gaina et al., 2009; Hawkins et al., 2008;
Morrissey et al., 2011). Thus, factors in the family environment that promote healthful
eating include the healthfulness of foods available in the home (Haerens et al., 2008;
Hanson et al., 2005), the frequency of family meals (Hammons & Fiese, 2011), and parental
modeling of and support for children’s healthful eating (Patrick & Nicklas, 2005).
Since the child health outcomes differ for children from dual-parent-employment and
from one-parent-employment, special deliberation is needed to create comparability of the
treated and control groups. This section analyzes the sensitivity of the result when the
estimation is made for the sub-sample which includes employed father and stay-at-home
mother households for both the treatment and control groups.
The study sample has very limited households with stay-at-home fathers and working
mothers for both the treated and control group since the labor force participation rate of
mothers is lower than that of fathers. Hence, the robustness check is done by comparing the
results from the overall sample with that of a sub-sample that includes households with
fathers who do outside work and stay-at-home mothers for both the treated and control
108
groups. As the results in Table 3.7 clearly show the findings based on the new subsample
provide consistent and larger magnitude estimates as compared to the results which are
based on the overall sample. It shows that the Fair Labor Standards Act minimum wage
policy intervention increases the probability of having a healthy childhood by 9.23
percentage points, which is larger than the estimated effect using the overall sample.
Similarly, the estimation result using one-parent employment in a strongly treated state
show that the minimum wage extension brings an improvement in the likelihood of having
a healthy childhood by 23.1 percentage point.
Table 3.7: Estimation for Children with Employed Fathers and Stay-at-Home Mothers
Main Model Strongly Treated
DID 0.0923* 0.231***
(0.0503) (0.0733)
Treatment -0.0487 -0.141**
(0.0431) (0.0635)
Post 0.544*** 0.574***
(0.0481) (0.0461)
White 0.220*** 0.219***
(0.0366) (0.0439)
Black 0.278*** 0.275***
(0.0383) (0.0469)
Male -0.0309 -0.0177
(0.0194) (0.0212)
Mother HS graduate 0.0561** 0.0510*
(0.0250) (0.0267)
Father HS graduate 0.105*** 0.0968***
(0.0231) (0.0252)
Mother some college 0.168*** 0.154***
109
(0.0458) (0.0466)
Father some college 0.111*** 0.153***
(0.0352) (0.0383)
Family Size -0.0151*** -0.0173***
(0.00496) (0.00537)
Constant 0.122 0.108
(0.152) (0.182)
Cohort Fixed Effect YES YES
State Fixed Effect YES YES
N 2307 1964
Robust standard errors in parentheses
∗p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
3.5 Discussion and Conclusion
Theeffectofminimumwagepoliciesonlabormarketoutcomeshasbeenextensivelystudied
and continues to be a focus of interest for researchers. However, the minimum wage
policies’ effect beyond labor market outcome is less studied. Wehby, Kaestner, et al.’s (2020)
study is the only exception that investigates the child health effect of minimum wage using
state-wise
comparisonofchildhealthoutcomechangesresultingfromachangeintheminimumwage. This
study aims to bridge this gap and exploits the unique feature of the 1966 Fair Labor
Standards Act to evaluate the impact of minimum wage declaration on child health using a
nationally representative sample.
Using the PSID’s intergenerational survey data, this study examines the effect of the
1966 Fair Labor Standards Act minimum wage declaration on child health. The treated and
control groups comprise individuals whose parents were working in the treated and
110
control industries during their early childhood years before and after the full
implementation of the Act. To obtain estimates, I use a difference-in-differences research
design. The findings show that the 1966 Fair Labor Standards Act minimum wage
declaration significantly improves the general health status of children whose parents were
working in the treated industries.
The results from estimation using the overall sample and cohort and state fixed effect
show that the Fair Labor Standards Act minimum wage policy intervention increases the
likelihood of having a healthy childhood increase by 7.16 percentage points. This finding is
similar to a recentstudybyWehby,Kaestner,etal.
(2020)whichshowsthathigherminimumwageincreases during ages 0 to 5 has a significant
effect on child health. Another study based on an exogenous
increaseinhouseholdincomeduetoagovernmentincometaxcreditalsoshowedsimilarresults.
Braga et al.’s (2020) study revealed that the effect of Earned Income Tax Credit (EITC)
during childhood on self-reported general health (reporting excellent or very good health)
of young adults is largest for those who got the credit during their first five years of life.
The positive child health impact of the minimum wage is not homogeneous across all
states. The Fair Labor Standards Act delivered a particularly large shock for states with no
state minimumwagebeforetheAct.
Theestimationresultsforthechildhealthimprovementimpact of the minimum wage policy
are stronger for these states than the estimated results based on the overall sample. The
intervention increases the probability of having a healthy childhood by 17.9 percentage
points, which is roughly 10 percentage points higher than the effect based on the overall
sample. Other studies for different outcome variables provide similar magnitude
differences in the impact of the minimum wage for the two groups of states. Derenoncourt
and Montialoux’s (2021) study showed that the 1966 extension of the minimum wage can
111
explain more than 20 percent of the reduction in the racial earnings and income gap during
the Civil Rights Era. However, the effect is stronger for strongly treated states than states
which had a state minimum wage law before the extension.
In light of previous literature claims of the heterogenous effect of policy interventions
on child health, this study also investigates whether the childhood health effect of the Fair
Labor Standards Act differs between males and females for the overall sample and strongly
treated sub-sample. Both estimation results show that the health improvement effect of the
minimum wage extension is positive for both female and male sub-samples. However, the
magnitude is larger for the female sub-sample, especially for the estimation result of
strongly treated states.
Theinterventionincreasestheprobabilityofhavinghealthychildhoodby23.2percentagepoints
for females and only 12.7 percentage points for males.
These gender differences in health can be caused by biological, social or behavioral
factors (Sayer & Britt, 1996). Biological gender differences in health at a young age may be
inherent to human nature. However, social and behavioral factors may have a noticeable
gender-specific effect directly on health and through care-seeking patterns and the use of
healthcare services and disease-related benefits (Gissler et al., 1999).
Theestimationresultsarerobustforthealternativecategorizationofchildren’sself-reported
general health status. It is a dummy variable with a value of 1 when the child reports very
good and excellent childhood health and with a value of 0 otherwise. The Fair Labor
Standards Act minimum wage policy intervention increases the likelihood of having a
healthy childhood by 7.14 percentage points for the overall sample and 17.16 percentage
points for the strongly treated subsample. The result is also robust for the analysis, which
only considers children who have employed fathers and stay-at-home mothers.
112
Therefore, policies that aim to improve the health status of children through wage or
other
incomeenhancementprogramsforparentsshouldconsidertheheterogeneityoftheeffectofthe
intervention on child health outcomes for females and males. It is also important to
consider that the federal minimum wage policy child health effect is stronger in states
which do not have a state minimum wage policy.
CHAPTER 4
MATERNAL WORK INTENSITY AND CHILDHOOD
HEALTH: HETEROGENEITY OVER STAGES OF CHILD
DEVELOPMENT
4.1 Introduction
The labor force participation of working mothers has increased dramatically over the
last six decades. In 1960, 18.6% of married women with children under the age of six were
in the labor force, and the figure increased to 30.3% in 1970 (Desai et al., 1989). By 2000,
this figure had risen to 62.8% (U.S. Bureau of the Census, 2002). The figure has reached
more than 70 percent in recent years. The US Bureau of Labor Statistics (2022) reported
that the labor force participation rate for all mothers with children under age 18 was 71.2
percent in 2021, unchanged from 2020 but down from 72.3 percent in 2019. Mothers
whose youngest child was under age 6 remained less likely to participate in the labor force
in 2021 than mothers whose youngest child was aged 6 to 17 (65.6 percent and 75.5
percent, respectively).
Thisincreasingmaternallaborforceparticipationcallsforrigorousinvestigationofwhether
theiremploymentpositivelyornegativelyaffectschildhealthanddevelopment. Itisimportant
for policymakers to formulate legislation on labor market policies which particularly affect
mothers and for mothers to make optimal labor market decisions after the birth of a child.
Most of the prior literature on the effect of mothers’ labor market decisions on children’s
life circumstances are mainly focused on child development as an outcome, perhaps due to
the wider availability of objective measures such as academic performance (Bernal, 2008;
114
Blau & Grossberg, 1990; Desai et al., 1989; Kaestner & Corman, 1995; Reynolds et al., 2017;
Ruhm, 2004; Waldfogel et al., 2002).
Empirical research on the impact of maternal work intensity on children’s health and
education produces mixed results and most of them use OLS, which limits their ability to
draw causal inferences (Bernal & Keane, 2005). However, a few studies used causally
plausible identification methods (Reynolds et al., 2017). For instance, Blau and Grossberg
(1990) employ an instrumental variable (IV) approach and find that maternal employment
has a negative effect on cognitive development during the child’s first year of but a positive
effect in the second year. Behrman et al. (2017) also find significant improvements in
cognitive, language, and socio-emotional skills for children whose mothers worked for
higher fractions of their lives in the first two years. However, the effects are not significant
after controlling for socioeconomic and demographic variables. Their instrumental variable
and propensity score matching estimations results also show no significant effect.
In contrast, Hill et al. (2005) find small but negative and significant effects of maternal
employment on cognitive outcomes using propensity score matching. Similarly, James-
Burdumy (2005), using individual and family fixed effects estimation, reveal that maternal
hours and weeks worked in the first year of a child’s life negatively affected math test
scores. Bernal (2008) best modeled the simultaneity of the working and childcare decisions
and finds that full-time maternal employment and childcare during the first five years of life
were associated with a reduction in math, vocabulary, and reading test scores.
Similartothechilddevelopmenteffectofmaternallabormarketchoicetheeffectsofmaternal
work on child health are also mixed. Some studies find no significant effects (Baker &
Milligan, 2008; Gwozdz et al., 2013) and others show adverse effects. Morrill (2011) finds
that maternal employment increased the probability of adverse health events by nearly 200
115
percent, while Gennetian et al. (2010) find a modest adverse effect on children’s health
using instrumental variables estimation and experimental data.
Despite the numerous works on the topic, previous research on the impact of maternal
work intensity on child health has limitations for a couple of reasons. First, to the best of my
knowledge, none of the studies comprehensively examined the heterogenous impact of
maternal work intensity across all the stages of child development. Given its potential
positive effect on earnings and negative effect on maternal time allocated for childcare, it is
plausible that the same level of maternal work intensity affects children’s health at different
stages of development differently. Related to this is also the fact that the previous literature
observes mothers’ employment for a very short span of time. Only looking at a specific
period in a child’s development does not align with the understanding that child
development measures are cumulative and do not just reflect current inputs (Cunha et al.,
2006). Second, most of the prior literature employed the technique of OLS (Bernal & Keane,
2005) which ignores the potential simultaneity problem between maternal labor market
choice and child health (Bernal & Keane, 2011). Third, none of the previous studies used a
general self-reported measure of the health status of the entire childhood. A self-reported
measure of health is a good predictor of mortality, morbidity, or use of medical care (Appels
et al., 1996; Idler & Kasl, 1995; McCallum et al., 1994), strongly correlated with other health
outcomes (Wagstaff & van Doorslaer, 1994) and reliably reported by respondents about its
current status and how it changes over time (Vaillant & Wolff, 2012). Moreover, mixed
findings of the prior studies warrant a more comprehensive examination of the effect of
maternal work intensity on child
health.
This study intends to address limitations stated above and contribute to the current
literature on the topic. I use an intergenerational dataset and examine the effect of maternal
116
work intensity on child health for the full childhood period, from birth to 17 years of age.
Furthermore, I disaggregate childhood into multiple stages of development to investigate
the heterogenous effect of maternal work intensity across those stages. As a general
measure of child health, I use the child’s self-reported general childhood health status. The
outcome variable ‘healthy childhood’ dummy is constructed from the Likert scale self-
reported childhood health status reported by children after they become adults. The
dummy variable has a value of 1 for the respondent reporting excellent, very good, or good
childhood health, and 0 otherwise. To this end, the study employs the Panel Study of Income
Dynamics (PSID)’s intergenerational survey data to examine the causal relationship
between maternal work intensity, measured by average annual hours worked, and
childhood health. The average annual hours worked is calculated based on the history of
the mother’s work during the child’s entire childhood period in aggregate as well as during
each stage of child development. To address the potential endogeneity problem, I employ
the instrumental variable (IV) probit regression estimation technique using state-level
women’s labor force participation as an instrumental variable.
The result from instrumental variable probit regression estimation that accounts for
cohort fixed effect shows that maternal hours worked has a positive and significant effect
on the likelihood of having a healthy childhood. The marginal effect estimation of maternal
work intensity shows that each additional 1000 hours of maternal hours worked is
associated with a 0.6 percentage points increase in the likelihood of having a healthy
childhood. This indicates that the positive income effect of maternal employment on child
health outweighs the negative impact of employment due to time reduction from childcare.
The above IV probit estimation treats the entire childhood as a single period and assumes
the effects of maternal work intensity are similar across different ages. However, maternal
work intensity’s positive child health impact is not homogeneous across all stages of
117
childhood development. The fixed effect estimation that accounts for the heterogeneity
across stages of child development (infancy, early childhood, mid-childhood, and late
childhood) shows that maternal hours worked has a negative and significant effect on
childhood health for the first two stages (infancy and early childhood) of child
development. Contrary to the early stages of child development, maternal work intensity
during late childhood has a positive and significant effect on child health. The marginal
effect estimation shows that each additional 1000 hours of maternal hours worked is
associated with a 0.775 percentage points increase in the likelihood of having a healthy
childhood.
In light of the previous literature’s claims about the heterogeneous effect of parental
health investment on child health between boys and girls (Mu¨ller et al., 2005; Perry et al.,
1998; Kelder et al., 1995; Plachta-Danielzik et al., 2007), this study also investigates
whether the childhood health effect of maternal work intensity differs between males and
females. The separate estimation results of the IV probit show that maternal work intensity
significantly affects child health for both female and male subsamples. However, the
magnitude of the
maternalworkintensityeffectonchildhealthisslightlylargerforthefemalesub-sampleduring
the infancy stage of child development. For the later stages of child development, the effect
is slightly larger for the male sub-sample.
The estimation results are robust for the alternative definition of the self-reported
general health status dummy variable with a value of 1 if the respondent reports excellent
or very good childhood health status and 0 otherwise. Ordered probit estimation is also
conducted using the PSID’s 5-point scale (excellent, very good, good, fair, or poor) of
childhood health status as a dependent variable to check the robustness of the estimation
based on the binary dependent variable. The extended ordered probit (eoprobit), which
118
accounts for the endogeneity problem, estimation results show consistent findings.
Maternal work intensity is significantly and negatively related to poor childhood health and
positively and significantly related to excellent childhood health. Furthermore, the study
investigates the differential effect of part-time and full-time working mothers on child
health. A mother is categorized as full-time worker if she worked for an average of 1700
hours per year or more (national average as per OECD.stat, 2022). The estimation result
implies that the likelihood of having a heathy childhood for children whose mothers
working full-time during the childhood years is significantly larger than that of children
whose mothers were working part-time.
The remaining part of this chapter is structured as follows: Section 2 reviews the
previous literature on the topic. Section 3 describes the methodology of the study: data,
descriptive statistics, and identification strategy. Section 4 provides empirical estimation
results and robustness checks for alternative definitions of the outcome variable and
alternative model specifications. Section 5 discusses the major findings and draws
conclusions.
4.2 Literature Review
The increasing trend of mothers’ labor force participation in recent decades attracts
researchers’ attention to investigate the relationship between maternal employment and
children’s life outcomes. The relationship is ambiguous due to both the positive and
negative effects of mothers’ labor supply decisions on children’s health and development
(Heinrich, 2014). On the one hand, maternal employment can positively impact children, as
working mothers earn (additional) money that they can use to improve children’s nutrition,
healthcare, childcare, and learning opportunities. On the other hand, maternal employment
119
reduces the time mothers can spend with their children, particularly if they work long
hours.
There is ample literature that provides empirical evidence, mainly for the short-term
effect, on the effect of mothers’ labor market decisions on children’s life circumstances.
However, this
literatureprimarilyfocusedonchildrenatearlyagesandtheirchilddevelopmentasanoutcome,
perhaps due to the wider availability of objective measures such as academic performance
(Bernal, 2008; Blau & Grossberg, 1990; Desai et al., 1989; Kaestner & Corman, 1995;
Reynolds et al., 2017; Ruhm, 2004; Waldfogel et al., 2002).
The empirical findings on the effect of maternal work intensity on children’s health and
education show mixed evidence. Bernal and Keane’s (2005) review of empirical studies that
use data of children from birth to age 15 from the US National Longitudinal Survey of Youth
(NLSY) shows an inconclusive relationship between maternal employment and children’s
cognitive outcomes. Four of the 13 reviewed papers show a negative effect of maternal
employment on children’s cognitive ability, while two show a positive effect. Two of them
show that maternal employment has no significant effect on children’s cognitive ability. The
remaining papers show heterogeneous results depending on the households’ social group
and maternal employment timing.
ThemajorityofthepapersuseOLS(Bernal&Kean,2005)andhavelimitationsindrawinga
causal inference conclusion from their findings. However, few exceptions use causally
plausible identification (Reynolds et al., 2017). Blau and Grossberg (1990) employ an
instrumental variable (IV) approach and find that maternal employment has a negative
effect in the first year of children’s cognitive development and a positive and offsetting
impact in the second year. However, the net impact in the first overall three to four years is
insignificant. Similarly, Behrman et al. (2017) found that three years old children with
120
mothers who worked for higher fractions of their children’s lives in the previous two years
performed significantly better on all tests (cognitive, language, socio-emotional) compared
to children whose mothers had worked less. The study controls for a baseline test
performance. However, these main effects did not
remainsignificantwiththeinclusionofawiderangeofsocio-economicanddemographiccontrol
variables. Their IV and propensity score matching results reveal that the effect of maternal
work intensity on child development outcomes is insignificant.
In contrast, Hill et al., (2005) use propensity score matching and find small but
significantly negative effects of maternal employment on children’s cognitive outcomes.
They use different outcome variables including the Peabody Picture Vocabulary Test-
Revised (PPVT–R) (Dunn
&Dunn,1981),administeredatages3and4;thePeabodyIndividualAchievementTestinMath
(PIAT–M), administered at ages 5 or 6 and 7 or 8; and the Peabody Individual Achievement
TestinReading(PIAT–R),alsoadministeredatages5or6and7or8. Theycompareoutcomes
across four maternal employment patterns: no work in the first three years of post-birth,
work only after the first year, part-time work in the first year, and full-time work in the first
year. Their findings demonstrate small but significant negative effects of maternal
employment on children’s cognitive outcomes for full-time employment in the first-year
post birth as compared with employment postponed until after the first year.
Similarly, James-Burdumy’s (2005)’s individual and family fixed-effects estimation
results show that maternal hours and weeks worked in the first three years of the child’s
life have a mixed effect on cognitive outcomes (as measured by PPVT, PIAT Math, and PIAT
Reading scores).
Inparticular,theestimationresultsfromchildandmotherfixedeffectregressionsshow that the
PIAT Math and Reading scores were negatively affected by maternal employment in
121
thefirstyearofthechild. However, maternalemploymentinthesecondyeardoesnotaffectthe
test scores. Furthermore, maternal employment in the third year positively affects PIAT
Math scores.
Ruhm(2004)followsasimilarestimationprocedurewithamorecompletesetofcontrol
variablesandfindsamuchstrongernegativeimpactofmaternalemploymentonchildoutcomes.
Maternalemploymentduringthefirstthreeyearsofthechild’slifehasasmalldeleteriouseffect on
the estimated verbal ability of three- and four-year-olds and a larger negative impact on
readingandmathematicsachievementoffive-andsix-year-olds. Takingastructuralapproach,
Bernal(2008)hasperhapsbestmodeledthesimultaneityoftheworkingandchildcaredecisions
and finds having mothers working full-time and using childcare for an additional year
during the first five years of life is associated with a reduction in math, vocabulary, and
reading test scores for children ages three to six years.
Similar to the child development effect of maternal labor market choice, findings on the
effectsofmaternallabormarketchoiceonchildhealtharemixed. Baker, Gruber, andMulligan
(2008) estimate the effect of maternal labor supply on young children’s health by analyzing
the effect of a childcare subsidy program in Quebec. They employ a difference-in-
differences
identificationstrategyandfoundthatthepolicyresultedinanincreaseinmaternallaborsupply,
moreenrollmentinformalchildcare,andadecreaseinchildren’shealthoutcomes. However,the
studyonlyexaminestheimpactoneligiblechildren,makingitunabletodistinguishthespecific
effectsofchildcareversusmaternalemployment. BakerandMilligan(2008)alsousevariationin
maternityleavebenefitsinCanadatoanalyzetheshort-runeffectsofmaternalnon-employment
on infants’ health and development and find no significant effects. Similarly, Gwozdz et al.,’s
(2013) analysis based on European data provides little evidence for any association
between maternal employment and childhood obesity, diet, or physical activity. On the
122
other hand, Morrill (2011) using the restricted-access National Health Interview Survey
(1985–2004) data identifies the effects of overnight hospitalizations, asthma episodes, and
injuries/poisonings for childrenages7–17.
Maternalemploymentincreasestheprobabilityofeachadversehealthevent by nearly 200
percent. These effects are robust and do not reflect a non-representative local
effect.
Gennetian et al., (2010) examine whether maternal employment affects the health status
of low-income, elementary-school-aged children using instrumental variables estimation
and experimental data from a welfare-to-work program implemented in the early 1990s. IV
estimation results show a modest adverse effect of maternal employment on children’s
health.
In a secondary analysis using fixed effects techniques on longitudinal survey data collected
in 1998 and 2001, they find a comparable adverse effect of maternal employment on child
health that supports the external validity of the primary result found using an IV estimate.
Prior scholarly work on the effect of maternal employment on child circumstances
mainly focuses on the early years of childhood. This paper complements the literature by
focusing
attentiononexaminingtheeffectofmaternalworkintensityonchildhealthforthefullchildhood
period,frombirthto17yearsofage. Furthermore,Idisaggregatechildhoodintomultiplestages of
development to investigate the heterogenous effect of maternal work intensity across those
stages. Furthermore, none of the previous studies used a general self-reported measure of
the health status of the entire childhood. A self-reported measure of health is a good
predictor of
mortality,morbidity,oruseofmedicalcare(Appelsetal.,1996;Idler&Kasl,1995;McCallumet
al.,1994),stronglycorrelatedwithotherhealthoutcomes(Wagstaff&vanDoorslaer,1994)and
123
reliablyreportedbyrespondentsaboutitscurrentstatusandhowitchangesovertime(Vaillant &
Wolff, 2012). Hence, this study contributes to the existing literature through its empirical
evidence of the effect of maternal work intensity on childhood health outcomes reported by
childrenaftertheybecomeadults. Furthermore,sincefindingsofthepriorliteraturearemixed, it
calls for further study on the topic.
4.3 Empirical Methods
4.3.1 Data
This study also uses the PSID data to assess the impact of maternal work intensity on
children’shealth. BecausethePSIDdatasetcontainsintergenerationalinformationonparental
labor market outcomes and child health, it is ideal for studying the child health impact of
mothers’ labor market choice.
The PSID dataset was established in 1968 and for each year panel wave, it collects
parental labor market outcome information from a prior year and other household
information from the same year. In addition, it has retrospective maternal labor market
outcomes information and other household characteristics. The study includes individuals
who have information on
selfreportedchildhoodhealthandannualworkhoursofmothersthroughouttheirentirechildhoo
d. Theinclusioncriteriayield8,988childrenofwhom48.06%arefemale.
Whereasthemainmodel is based on the above sample size, the sample size varies slightly for
alternative specifications and robustness checks.
124
Table 4.1 presents definitions of all analytical variables used in this study. Most of the
variables are extracted from the PSID survey research center data set. In addition, statelevel
women’s labor force participation data used in the IV regression estimation is extracted
from the Bureau of Labor Statistics. The PSID dataset provides information about individual
characteristics: race,gender,andretrospectivelyreportedchildhoodhealthstatus. Inaddition,
the second group of variables provides information about parental characteristics:
maternal work intensity measured in hours worked per year, parents’ educational
attainment, Home Ownership status, number of adults living in the household and family
size of the household.
The main outcome variable is whether the individual had a healthy childhood or not. In
the PSID, childhood health is a retrospective self-evaluation using the standard 5-point
scale (excellent, very good, good, fair, or poor) of the general state of one’s health when one
was 17 years old or younger. Following the previous work of Smith (2009) and Finnas et al.
(2008), I create a dummy variable for a ‘healthy childhood’ that assumes the value of 1 if
the individual reported excellent, very good, or good childhood health and the value 0
otherwise. An alternative dichotomization is based on Smith’s (2009) classification of five-
category scale health status which assigns a value of 1 for excellent or very good health and
assigns a value 0 for less (good, fair, or poor) to do a robustness check.
Table 4.1: Definitions of Major Variables
Variable Name Definition
Healthy Childhood =1iftherespondenthasreportedanexcellent,verygood, or
good childhood health status, = 0 otherwise.
Mother’s HW Childhood Mother’s average annual hours worked during the
whole childhood period of a child (from birth - 17
years of age).
Mother’s HW Infancy Mother’saverageannualhoursworkedduringtheinfancy
childhood period of a child (from birth - 2years of age).
Mother’s HW Early Mother’s average annual hours worked during the
125
early childhood period of a child (3 to 6 years of age).
Mother’s HW Mid Mother’s average annual hours worked during the late
childhood period of a child (13 to 17 years of age).
Mother Age at Birth Age of mother when the child was born.
Mother HS Graduate = 1 if the mother completed high school, = 0 otherwise.
Father HS Graduate = 1 if the father completed high school, = 0 otherwise.
Mother College = 1 if the mother has some years of college, = 0
otherwise.
Father College = 1 if the father has some years of college, = 0
otherwise.
Mother School Dropout = 1 if the mother has not completed high school, = 0
otherwise.
Father School Dropout = 1 if the father has not completed high school, = 0
otherwise.
Male = 1 if the respondent is male, = 0 if the respondent is
female.
White = 1 if the respondent is White, = 0 otherwise.
Black = 1 if the respondent is Black, = 0 otherwise.
Other = 1 if the respondent is other than White or Black, = 0
otherwise.
Family Size Respondent’s family size at childhood home.
Single Mother = 1 if a child raised by a single mother, = 0 otherwise.
Home Ownership =1 if parents own their own home, = 0 otherwise.
Adults Number of adults in the household at respondent’s
childhood home.
State WLFP Rate The average rate of women’s labor force participation
at the state level during childhood years.
Year Year of birth of the respondent.
4.3.2 Descriptive Statistics of Major Variables
Table4.2presentsthemeanandstandarddeviationsofthemainvariablesrelatedtochildren,
parents and households’ characteristics.
Table 4.2: Summary Statistics of Major Variables
126
Variable Name Mean SD Min Max N
Healthy Childhood 0.61 0.49 0 1 8988
Mother’s HW Childhood 1,315.14 578.54 48 2838.75 8988
Mother’s HW Infancy 1,150.33 661.96 30 2877 8988
Mother’s HW Early 1,234.65 684.97 30 2967.5 8988
Mother’s HW Mid 1,325.57 666.50 40 2976 8988
Mother’s HW Late 1,425.08 656.14 40 3050 8988
Mother’s Age at Birth 27.71 7.00 14 58 8988
Mother HS Graduate 0.34 0.47 0 1 8988
Father HS Graduate 0.30 0.46 0 1 8988
Mother College 0.21 0.41 0 1 8988
Father College 0.22 0.41 0 1 8988
Mother School Dropout 0.15 0.36 0 1 8988
Father School Dropout 0.09 0.28 0 1 8988
Male 0.52 0.50 0 1 8988
White 0.65 0.48 0 1 8988
Black 0.33 0.47 0 1 8988
Other 0.02 0.14 0 1 8988
Family size 4.50 1.43 1 15 8988
Single Mother 0.23 0.42 0 1 8988
Home Ownership 0.58 0.49 0 1 8988
Adults 1.88 0.33 1 2 8988
State WLFP Rate 0.56 0.04 0.378 0.709 8988
Approximately 60 percent of the study sample of 8,988 observations reported that they
had excellent, very good, or good childhood health. Maternal labor market engagement
varied across stages of childhood development. The lowest average annual hours worked is
registered during the infancy stage of childhood development, with around 1,150 hours,
and the highest is registered during the late childhood stage of their children, with average
hours worked of 1,425. The average mothers’ labor market engagement during the entire
childhood period of their children is around 1,315 hours.
127
4.3.3 Empirical Models
The empirical method identifies the effect of individuals’ self-reported general
childhood health status over their mothers’ average annual hours worked and other
individual and household characteristics. I first model the effect of maternal work intensity
on childhood health using a probit model shown in equations (4.1 and 4.2).
Hi = α+ Hoursβi + Xθi + τt + ϵi
with
(4.1)
P(Hi = 1|Hoursi,Xi) = (Φ α+ Hoursβi + Xθi + τt) (4.2)
is the population probit model with Hours and covariates included in the vector X and (Φ.)
is thecumulativestandardnormaldistributionfunction. Hi isthechildhealthoutcome,Hoursis
themainindependentvariablethatgaugestheaverageannualhoursworkedbymothersduring
the individual childhood years and X is a vector of demographic characteristics of the child
and his/her family. The unit of observation, i, is the child. The childhood health status is
measured basedontheretrospectiveself-reportofchildrenaftertheybecomeadults.
Thechildhoodhealth status, Hi, is a binary variable constructed from the Likert scale
retrospective self-reported childhood health status reported by children after they become
adults. Variable Hi takes a value of 1 if the respondent reports excellent, very good, or good
childhood health, and 0 if they report it as poor or fair. The vector X contains household and
individual characteristics that include the mother’s age at birth, educational attainment of
parents, the sex of the respondent, race, family size, dummy variable that indicates the child
was raised by a single mother, home ownership status of the childhood household to proxy
the wealth status and the number of adults in the childhood household. Due to cohort
differences in self-rated health (Chen et al., 2007; Link et al., 2017), the model also includes
the year (cohort) fixed effects (τt).
In the model, coefficient βmeasures the effect of maternal work intensity on child health.
128
However, child health and maternal labor market choice have a simultaneity problem
(Bernal & Keane, 2011). Mothers of health-impaired children may decide not to work and
stay home to care for their children. On the other hand, mothers may choose to enter the
labor force to pay for these children’s health investments. Hence, the childhood health
status of children and mothers’ labor market choice constitutes an interdependent
decision-making process. In econometric terms, there is a feedback relationship between
both variables that causes the binary model estimates of the likelihood of a child having
healthy childhood to be biased. To deal with the endogeneity problem, the IV probit
regression approach is employed to estimate the causal effect of maternal hours worked on
child health.
One strategy for recovering a consistent estimate of βis to identify an instrumental
variable Z, i.e. a variable that partially determines annual maternal hours worked but is
uncorrelated with ϵi in equation (4.1).
Hoursi = δ+ Zγi + Xηi + νi(4.3)
Withsuchaninstrument, Z, atwo-stageregressionmodelcanbeestimated, withthefirststage
regression of Hours over the instrument and other covariates. The main model is estimated
by taking the predicted (fitted) value of Hours from the first stage and substituting it for Hi
in Equation (4.1). The coefficient estimate from the two-stage instrumental variable probit
estimation, βˆ
IV , measures the causal effect of the instrument on the outcome operates
solely through the endogenous variable, which is maternal hours worked in this case.
A suitable instrument must fulfill two conditions. First, it needs to be correlated with the
actual working time of the mother, which is assumed to be an endogenous variable. Second,
it may not affect childhood health except through the endogenous variable. Following
previous studies (Baum, 2003; James-Burdumy, 2005), I use local labor market
129
characteristics as instruments for maternal hours worked. In doing so, the instrumental
variable satisfies the following two conditions: (1) local labor market characteristics
strongly affect maternal labor market supply decisions, and (2) all the effects of local labor
market characteristics on child health work through maternal employment decisions only.
Reynoldetal.,(2017)uselocalwomen’slaborforceparticipationasaninstrumentalvariable
for maternal work intensity and found a strong correlation between maternal hours
worked and local women’s labor force participation rate. I expect the state-level labor force
participation rate to be strongly correlated to actual working hours, as both reflect women’s
attitudes and working preferences in the area. Furthermore, state-level labor participation
is exogenous to childhoodhealth. Hence, thestudyusesstate-
levellaborforceparticipationasaninstrumental variable.
Studies show that policy interventions targeted at reducing the prevalence of various
health problems among children are more effective for some groups of children than
others: girls are more responsive than boys (Kelder et al., 1995; Mu¨ller et al., 2005; Perry et
al., 1998; Plachta-Danielziketal., 2007). Hence,
toinvestigatetheheterogeneouseffectofmaternalwork intensity on childhood health for
females and males, I also estimate the IV Probit regression for the two subsamples
separately. Furthermore, to gauge the heterogeneous effect of maternal hours worked
across different age groups on childhood health, the average annual work hours
areseparatelycalculatedfordifferentstagesofchilddevelopment(infancy,earlychildhood,mid
childhood, and late childhood). Then, the dependent variables (childhood health)
regression overthefourvariablesandthevariablesincludedinthevectorX
toestimatethedifferenteffect of maternal work intensity at different stages of child
development on child health.
130
In addition, IV ordered probit estimation is also conducted using the PSID’s 5-point scale
(excellent, very good, good, fair, or poor) childhood health status as a dependent variable to
check the robustness of the binary probit estimation.
4.4 Empirical Results
4.4.1 Maternal Work Intensity and Child Health
Table 4.3 reports the probit regression estimation results for the effect of maternal work
intensity on children’s health status. In addition to the main coefficient of interest, average
hours worked by the mother, the table also reports coefficients of the variables
representing individual characteristics, parental characteristics, the state where the child
grew up, and the year of birth of a child as a cohort indicator. The results in the first column
show that maternal work intensity has a positive and significant net effect on child health. It
indicates that the positive earning effect of maternal outside work engagement on child
health outweighs the negative effect of time deduction from taking care of the child. The
coefficient’s magnitude declines slightly but is still significant when I add cohort fixed
effects to the model in the second column. The table also provides the marginal effects of
maternal annual hours worked on child health. The coefficient is positive and significant at
the 1% level, which indicates that more maternal hours worked is associated with a higher
probability of having a healthy childhood.
Eachadditional1000hoursofmaternalhoursworkedisassociatedwitha0.02percentagepoints
increase in the likelihood of having a healthy childhood.
Table 4.3: Effect of Maternal Work Intensity on Child Health
131
Probit FE Probit Marginal Effect
Mother’s HW Childhood 0.000151*** 0.0000535** 0.0000200**
(0.0000241) (0.0000257) (0.00000962)
Mother’s Age at Birth 0.0229*** 0.0141*** 0.00527***
(0.00210) (0.00220) (0.000823)
Mother HS graduate 0.204*** 0.206*** 0.0773***
(0.0345) (0.0359) (0.0135)
Father HS graduate 0.247*** 0.231*** 0.0865***
(0.0353) (0.0369) (0.0138)
Mother College 0.321*** 0.293*** 0.110***
(0.0385) (0.0405) (0.0152)
Father College 0.158*** 0.164*** 0.0616***
(0.0378) (0.0393) (0.0147)
Male -0.0905*** -0.101*** -0.0378***
(0.0276) (0.0286) (0.0107)
White 0.667*** 0.691*** 0.259***
(0.0994) (0.0994) (0.0373)
Black 0.535*** 0.649*** 0.244***
(0.101) (0.101) (0.0380)
Family size -0.0381*** -0.0434*** -0.0163***
(0.0104) (0.0109) (0.00408)
Single Mother -0.404*** -0.429*** -0.161***
(0.0344) (0.0358) (0.0134)
Home Ownership 0.115*** 0.139*** 0.0521***
(0.0289) (0.0300) (0.0113)
Adults 0.0582 0.184*** 0.0692***
(0.0455) (0.0481) (0.0180)
Constant -1.257*** -1.582***
132
(0.147) (0.177)
Cohort Fixed Effect NO YES YES
The result also shows that the likelihood of an individual being healthy during childhood
is positively and significantly associated with parental educational attainment, if parents
own their residence, the number of adults in the family, and the mother’s age at birth. On
the other hand, the probability of having healthy childhood significantly decreases if a
single mother raises a child, the child grows up in a large-sized family, and if the child is
male.
4.4.2 Causal Relationship Between
Maternal Work Intensity and Child
Health
This section presents instrumental variable estimation results to examine the effect of
maternal work intensity on child health while the endogeneity issue is considered. The first
stage estimation results in Table 4.4 show that the instrumental variable, the state-level
women labor force participation, has a positive and significant correlation of 0.17 with the
actual working hours of mothers. The F-test for the first stage regression gives a test
statistic of 45.07; thus, we have no problem of weak instrument (Stock et al., 2002).
Furthermore, the Wald test of exogeneity has a chi-square statistic value of 5868 with a p-
133
value less than 1 percent. The latter provides evidence for the zero correlation between the
error term and the instrumental variable.
Table 4.4: First Stage Estimation of Maternal Work Intensity
Dependent: Maternal Work Intensity
IV: State Level Women LFP 1788.1***
(138.6)
Age at Birth 6.232***
(0.929)
HS graduate 23.48
(14.92)
Spouse HS graduate 120.4***
(15.17)
Some College 30.16*
(16.72)
Spouse some College 87.92***
(16.26)
Male 12.78
(12.03)
White -45.15
(47.81)
Black 143.5***
(48.36)
Family size -44.66***
(4.730)
Single Mother -3.013
(16.17)
Home Ownership 19.60
(12.72)
Adults 78.22***
(20.61)
Constant 84.75
(103.2)
134
>F
0.0000 The second stage IV probit estimation results presented in Table 4.5 show
that after addressing the issue of endogeneity for maternal work intensity, the sign of
the estimate is consistent with the previous conclusion. Hence, the net effect of
maternal work intensity on childhood health is positive and significant. The
coefficient and marginal effect coefficient of maternal hours worked are positive and
significant at the 1% level, which indicates that more maternal hours worked
increases the likelihood of having a healthy childhood. The magnitudes of the
estimated coefficients are larger than the coefficient estimates of the probit model in
the previous section. The results in the table show that each additional 1000 hours of
maternal hours worked is associated with a 0.6 percentage points increase in the
likelihood of having a healthy childhood.
Conversion of the dependent variable mothers’ annual hours worked into weekly hours
worked provides similar results but larger magnitudes that ease the interpretation. The
estimation result shows that an hour increase in maternal weekly hours worked increases
the likelihood of having a healthy childhood by 0.03 percentage points. The details of the
first stage, IV probit, and marginal effect estimations are in Table B1 and Table B2 in the
appendix.
Table 4.5: IV Estimation on Effect of Maternal Work Intensity on Childhood Health
Coefficients Marginal Effect
Mother’s HW Childhood 0.00149*** 0.000584***
(0.0000512) (0.0000217)
Mother’s Age at Birth -0.00182 -
0.000711
135
(0.00206) (0.000809)
Mother HS graduate 0.0660** 0.0258**
(0.0315) (0.0123)
Father HS graduate -0.0502 -0.0196
(0.0347) (0.0136)
Mother College 0.113*** 0.0443**
*
(0.0364) (0.0142)
Father College -0.0485 -0.0190
(0.0344) (0.0135)
Male -0.0773*** -0.0303***
(0.0240) (0.00939
)
White 0.428*** 0.168***
(0.0977) (0.0381)
Black 0.142 0.0554
(0.101) (0.0393)
Family size 0.0411*** 0.0161**
*
(0.0105) (0.00413
)
Single Mother -0.212*** -0.0829***
(0.0370) (0.0143)
Home Ownership 0.0520** 0.0204**
(0.0257) (0.0100)
Adults -0.0187 -
0.00732
(0.0411) (0.0161)
136
Constant -2.313***
(0.143)
Cohort Fixed Effect YES YES
Robust Standard errors in parentheses.
The results also reveal that the likelihood of an individual being healthy during
childhood ispositivelyandsignificantlyaffectedbymaternaleducationalattainment,
ifparentsowntheir residence, and the child grows up in a large-sized family. On the other
hand, the probability of having healthy childhood significantly decreases if a single mother
raises a child and if the child is male.
4.4.3 Heterogeneous Effects
Despite the positive effect of maternal work intensity for the overall childhood period,
the effect may vary across childhood years. Hence, this section investigates the
heterogeneous effect of maternal work intensity across stages of child development. To this
end, I divide the childhood period into four groups: Infancy (from birth - 2years of age),
early childhood (3 to 6 years of age), mid childhood (7 to 12 years of age), and late
childhood (13 to 17 years of age). I also conduct separate fixed effects regression analyses
for male and female subgroups and investigate whether there are heterogeneous effects of
maternal work intensity on child health. 4.4.3.1 Maternal Work Intensity at Different
Stages of Child Development
137
TheresultsinTable4.6showthatmaternalworkintensityhasaheterogeneouseffectacross
thefourstagesofchilddevelopment. Maternalworkintensityduringearlyperiodsofchildhood
has negative and significant effects on childhood health. The estimation results show that
the negative time effect outweighs the positive earnings effect of maternal employment
during infancy and early childhood stages of child development.
Table 4.6: Effect of Maternal Work Intensity at Different Stages
FE Probit Marginal Effect
Mother’s HW Infancy -0.000871*** -0.000347***
(0.000110) (0.0000439)
Mother’s HW Early -0.000807*** -0.000321***
(0.000136) (0.0000543)
Mother’s HW Mid 0.000252 0.000100
(0.000165) (0.0000657)
Mother’s HW Late 0.00195*** 0.000775***
(0.0000981) (0.0000389)
Mother Age at Birth 0.00774*** 0.00308***
(0.00193) (0.000768
)
Mother HS graduate -0.0379 -0.0151
(0.0284) (0.0113)
Father HS graduate -0.0473 -0.0188
(0.0319) (0.0127)
Mother College -0.0445 -0.0177
(0.0319) (0.0127)
Father College -0.0145 -0.00578
(0.0324) (0.0129)
138
Male -0.0531** -0.0211**
(0.0217) (0.00864
)
White 0.135 0.0538
(0.0823) (0.0327)
Black 0.167* 0.0663*
(0.0857) (0.0341)
Family size 0.00745 0.00296
(0.00875) (0.00348
)
Single Mother 0.0711*** 0.0283***
(0.0268) (0.0106)
Home Ownership -0.0820*** -
0.0326***
(0.0225) (0.00895
)
Adults -0.133*** -
0.0528***
(0.0315) (0.0125)
Constant -1.102***
(0.137)
Cohort Fixed Effect YES YES
Robust Standard errors in parentheses.
Specifically, the estimation results in Table 4.6 indicates that each additional 1000
average
annualhoursmothersworkdecreasesachild’slikelihoodofhavingahealthychildhoodbyaround
0.3percentagepoints. However,theeffectofmaternalworkintensityinoutsideworkduringthe
latechildhoodstageofchilddevelopmenthaspositiveandsignificanteffectongeneralchildhood
health(each1000hoursaverageannualhoursworkedincreasesthelikelihoodofhavingahealthy
139
childhood by around 0.7 percentage points). The magnitude of the coefficient estimates of
this variable is larger than any of the predecessor stages and offsets their negative effect.
Hence, in line with the previous section, the overall effect of maternal work intensity on
childhood health is positive and significant.
4.4.3.2 Gender Heterogeneity
The literature on gender differences in returns on investment in child health shows
ample evidence on performance differences between girls and boys (Gissler et. al., 1999;
Kelder et al., 1995;Mu¨lleretal.,2005;Perryetal.,1998;Plachta-Danielziketal.,2007).
Inlightofthisclaim, this section analyzes the effect of maternal work intensity on child health
for female and male
subgroupstogaugetheheterogeneouseffectsofmaternallabormarketchoice. Thesamplesizes
for the two subgroups are 4320 and 4668, respectively.
TheresultspresentedinthesecondandthirdcolumnsofTable4.7showevidencesupporting
the hypothesis that females’ health outcome response to the intervention is slightly
different from that of males. It shows that the maternal work intensity effect on child health
is negative at the early stages of child development for both. However, the magnitude is
slightly larger for females during the infancy stage of child development and for males
during the early childhood stagesofchilddevelopment.
Eachadditional1000averageannualhoursmothersworkdecreases a child’s likelihood of
having a healthy childhood by around 0.38 percentage points for females and 0.3
percentage points for males during the infancy stage of child development. The same
figures from the early childhood period estimation are 0.2 and 0.4 percentage points for
female and male groups, respectively. On the other hand, in the later stage of childhood
140
development, the effect of maternal outside work intensity has a positive and significant
effect on child health for both subgroups. But the magnitude is relatively larger for the
males’ subgroup (0.7 and 0.8, respectively).
The results show that the consideration of heterogeneous effects for females and males
is importantwhenweassesstheeffectofmaternallabormarketchoiceonchildhealth. Generally,
childhealthsignificantlydeteriorateswhenmaternalworkintensityishigherintheearlyyearsof
childhood for both female and male subgroups. In the later stage, the earnings effect
outweighs the time effect and brings a positive return on child health for both groups.
141
Table 4.7: Heterogeneous Effect of Maternal Work Intensity on Child Health Across Age and
Gender
Female ME Male ME
Mother’s HW Infancy -0.000964*** -0.000383*** -0.000756*** -
0.000301***
(0.000145) (0.0000575) (0.000164) (0.0000652)
Mother’s HW Early -0.000541*** -0.000215*** -0.00108*** -
0.000430***
(0.000179) (0.0000713) (0.000195) (0.0000774)
Mother’s HW Mid 0.000291 0.000116 0.000198 0.0000789
(0.000227) (0.0000901) (0.000235) (0.0000935)
Mother’s HW Late 0.00188*** 0.000746*** 0.00200*** 0.000794***
(0.000143) (0.0000567) (0.000132) (0.0000525)
Mother Age at Birth 0.00551* 0.00219* 0.00945*** 0.00376***
(0.00283) (0.00113) (0.00258) (0.00103)
Mother HS graduate 0.0184 0.00731 -0.0873** -0.0347**
(0.0415) (0.0165) (0.0396) (0.0158)
Father HS graduate -0.122*** -0.0484*** 0.0312 0.0124
(0.0449) (0.0178) (0.0452) (0.0180)
Mother College 0.102** 0.0405** -0.165*** -0.0658***
(0.0457) (0.0182) (0.0446) (0.0178)
Father College -0.129*** -0.0515*** 0.0908** 0.0361**
(0.0452) (0.0180) (0.0459) (0.0182)
White 0.227* 0.0903* 0.0964 0.0383
(0.120) (0.0477) (0.113) (0.0451)
Black 0.233* 0.0927* 0.153 0.0607
(0.123) (0.0488) (0.121) (0.0481)
Family size 0.0250** 0.00995** -0.00728 -0.00290
(0.0124) (0.00492) (0.0122) (0.00487)
Single Mother 0.119*** 0.0474*** 0.0363 0.0145
142
(0.0404) (0.0161) (0.0365) (0.0145)
Home Ownership -0.0892*** -0.0355*** -0.0722** -0.0287**
(0.0327) (0.0130) (0.0314) (0.0125)
Adults -0.105** -0.0418** -0.162*** -0.0643***
(0.0450) (0.0179) (0.0455) (0.0181)
Constant -1.416*** -0.862***
(0.197) (0.190)
Cohort Fixed Effect YES YES YES YES
Furthermore, Figure 4.1 and Table 4.8 show that the marginal effect of maternal work
intensity on childhood health has a slight difference for the two groups when we take the
entire childhood years as a single period. Each additional 1000 average annual hours
worked by a mother decreases a child’s likelihood of having a healthy childhood by around
0.6 percentage points for both groups. However, the predicted probability of having a
healthy childhood at a given level of hours worked is larger for the female group as Figure
4.1 shows.
143
Figure 4.1: Predicted probability of childhood health by gender
Table 4.8: Heterogeneous Effect of Maternal Work Intensity on Child Health across Gender
Female ME Male ME
Mother’s HW Childhood 0.00147*** 0.000571*** 0.00152*** 0.000599***
(0.0000765) (0.0000328) (0.0000678) (0.0000285
)
Mother Age at Birth -0.00229 -0.000892 -0.00164 -0.000645
(0.00307) (0.00120) (0.00278) (0.00110)
Mother HS graduate 0.113** 0.0438** 0.0201 0.00789
(0.0472) (0.0183) (0.0427) (0.0168)
Father HS graduate -0.0658 -0.0256 -0.0387 -0.0152
(0.0490) (0.0191) (0.0490) (0.0193)
Mother College 0.164*** 0.0640*** 0.0638 0.0251
144
(0.0526) (0.0204) (0.0503) (0.0198)
Father College -0.0838* -0.0327* -0.0134 -0.00526
(0.0504) (0.0197) (0.0475) (0.0187)
White 0.556*** 0.216*** 0.305** 0.120**
(0.146) (0.0569) (0.132) (0.0516)
Black 0.261* 0.102* 0.0280 0.0110
(0.151) (0.0587) (0.135) (0.0530)
Family size 0.0434*** 0.0169*** 0.0371*** 0.0146***
(0.0154) (0.00603) (0.0143) (0.00564)
Single Mother -0.182*** -0.0708*** -0.243*** -0.0957***
(0.0571) (0.0220) (0.0484) (0.0189)
Home Ownership 0.0554 0.0216 0.0514 0.0202
(0.0369) (0.0144) (0.0357) (0.0140)
Adults -0.0322 -0.0126 -0.00915 -0.00360
(0.0569) (0.0222) (0.0600) (0.0236)
Constant -2.350*** -2.325***
(0.211) (0.199)
Cohort Fixed Effect YES YES YES YES
4.4.4 Robustness Check
4.4.4.1 Redefinition of Childhood Health Status
For many survey data sets, including the PSID, self-reported health status is a Likert
scale variable ranging from very poor to excellent health status. Finnas et al., (2008)
demonstrated thattheself-reportedhealthstatusvariableonthefive-
pointLikertscalecouldbedichotomized. However, certain methodological concerns need to
145
be addressed. Empirical findings of Finnas et al. (2008) and Bourne (2009) show that
exclusion or inclusion of moderate self-reported health in creating the dichotomous
variable may yield different estimates. Hence, this section checks the robustness of the
previous results when the cutoff of the healthy childhood dummy variable changes. For this
section, the dummy variable healthy childhood has a value =1 when the self-assessed health
status during childhood is reported as very good or excellent and =0 otherwise.
TheestimatespresentedinTable4.9showthattherecategorizationoftheoutcomevariable
does not affect the results. For the main model that aggregates the analysis for the entire
childhood as a single period, the estimation results show that the increase in maternal work
intensityincreasesthelikelihoodofhavingexcellentorverygoodchildhoodhealth. Thisresultis
consistent with the previous results based on dichotomization of the healthy childhood
dummy variable which gets value 1 when the self-reported childhood health status is
excellent, very good, or good and 0 otherwise. The magnitude of the coefficient estimate is
also equivalent. The results in the table show that each additional 1000 hours of maternal
hours worked is
associatedwitha0.6percentagepointsincreaseinthelikelihoodofhavingahealthychildhood.
Table 4.9: IV Probit Estimation for Alternative Definition of Childhood Health
Coefficient Marginal Effect
Mother’s HW Childhood 0.00149*** 0.000585***
(0.0000514) (0.0000218)
Mother’s Age at Birth -0.00171 -
0.000669
(0.00207) (0.000812)
Mother HS graduate 0.0636** 0.0249**
(0.0314) (0.0123)
Father HS graduate -0.0514 -0.0202
146
(0.0347) (0.0136)
Mother College 0.110*** 0.0429**
*
(0.0362) (0.0141)
Father College -0.0490 -0.0192
(0.0344) (0.0135)
Male -0.0763*** -0.0299***
(0.0240) (0.00940
)
White 0.429*** 0.168***
(0.0979) (0.0382)
Black 0.137 0.0537
(0.101) (0.0394)
Family size 0.0410*** 0.0161**
*
(0.0105) (0.00414
)
Single Mother -0.208*** -0.0813***
(0.0368) (0.0143)
Home Ownership 0.0541** 0.0212**
(0.0257) (0.0101)
Adults -0.0153 -
0.00599
(0.0412) (0.0161)
Constant -2.323***
(0.144)
Cohort Fixed Effect YES YES
Robust Standard errors in parentheses.
147
Table 4.10 presents the effect of maternal work intensity at different stages of child
development.
Table 4.10: Heterogeneity over Stages of Child Development after Redefinition of Childhood
Health Dummy
Coefficient Marginal Effect
Mother’s HW Infancy -0.000878*** -0.000349***
(0.000111) (0.0000440)
Mother’s HW Early -0.000797*** -0.000317***
(0.000137) (0.0000546)
Mother’s HW Mid 0.000247 0.0000981
(0.000166) (0.0000660)
Mother’s HW Late 0.00195*** 0.000776***
(0.0000984) (0.0000391)
Mother Age at Birth 0.00783*** 0.00311***
(0.00194) (0.000771)
Mother HS graduate -0.0387 -0.0154
(0.0284) (0.0113)
Father HS graduate -0.0481 -0.0191
(0.0319) (0.0127)
Mother College -0.0459 -0.0183
(0.0319) (0.0127)
Father College -0.0147 -0.00583
(0.0324) (0.0129)
Male -0.0526** -0.0209**
(0.0217) (0.00864)
White 0.136* 0.0540*
(0.0824) (0.0328)
Black 0.164* 0.0654*
(0.0858) (0.0341)
148
Family size 0.00750 0.00298
(0.00876) (0.00348)
Single Mother 0.0730*** 0.0290***
(0.0268) (0.0106)
Home Ownership -0.0809*** -0.0322***
(0.0225) (0.00895)
Adults -0.131*** -0.0522***
(0.0316) (0.0126)
Constant -1.109***
(0.138)
Cohort Fixed Effect YES YES
Robust Standard errors in parentheses.
For the model that assumes heterogeneity of the effect of maternal work intensity over
the four stages of child development, the estimation results in Table 4.10 show that
maternal work intensity during the early stages of child development decreases the
likelihood of having excellent or very good childhood health. These findings are also
consistent with the previous results which dichotomize childhood health with a value =1
when the self-reported childhood health status is excellent, very good, or good and =0
otherwise.
4.4.4.2 Ordered Probit Estimation Results
To examine the robustness of maternal work intensity on child health, another model
which takes the child health variable as an ordinal 5-point scale category (excellent, very
good, good, fair, or poor) is estimated using the ordered probit regression approach. The
extended ordered probit (eoprobit), which accounts for the endogeneity problem,
149
estimation technique using stata is employed to estimate the coefficient and marginal
effects. The result reinforces the IV
probitestimationresultsofthepositiveeffectofmaternalhoursworkedonchildhoodhealth, as
presented in Table 4.11 and Table 4.12. The estimation results in Table 4.11 show that
mother hours worked positively affect the child’s likelihood of having excellent childhood
health.
Table 4.11: Ordered Probit Estimation of Childhood Health
Coefficients (Healthychildhood=Excellent)
Mother’s HW Childhood 0.00137***
(0.0000671
)
Mother’s Age at Birth -0.000770
(0.00200)
Mother HS graduate 0.0267
(0.0287)
Father HS graduate -0.0799**
(0.0319)
Mother College 0.0720**
(0.0328)
Father College -0.0511
(0.0327)
Male -0.0554**
(0.0226)
White 0.407***
(0.0965)
Black 0.159
(0.100)
Family size 0.0435***
(0.0101)
Single Mother -0.211***
(0.0349)
Home Ownership 0.0739***
(0.0249)
Adults -0.0149
150
(0.0402)
Constant 84.75
(103.1)
Cohort Fixed Effect YES
Robust Standard errors in parentheses.
The marginal effect estimation results also show that maternal hours worked decreases
the child’s probability of having poor, fair, or good health and increases the likelihood of
having verygoodandexcellentchildhoodhealth.
Eachadditional1000averageannualhoursamother works decreases a child’s likelihood of
having poor, fair, or good childhood health by around 0.5, 0.003, and 0.006 percentage
points, respectively. In contrast, each additional 1000 average annual hours the mother
works increases a child’s likelihood of having very good or excellent childhood health by
around 0.03 and 0.5 percentage points, respectively.
Conversion of the dependent variable mothers’ annual hours worked into weekly hours
worked provides similar results but larger magnitudes that ease the interpretation. The
estimation shows that an hour increase in maternal weekly hours worked decreases a
child’s likelihood of having poor, fair, or good childhood health by around 0.02, 0.0001, and
0.0003 percentage points, respectively. In contrast, each additional 1000 average annual
hours the mother works increases a child’s likelihood of having very good or excellent
childhood health by around 0.0015 and 0.027 percentage points, respectively. The details of
these marginal effect estimations are in Table B3 in the appendix.
151
Table4.12:MarginalEffectEstimationsofOrderedProbitModelofChildhoodHealth
PoorFairGoodVeryGoodExcellent
Mother’sHWChildhood-0.0005368***-0.000003***-0.0000062***0.0000296***0.0005161***
(0.0000276)(0.0000004)(0.0000011)(0.000003)(0.000029)
Mother’sAgeatBirth0.00030220.00000150.0000035-0.0000167-0.00029
(0.00078)(0.0000039)(0.0000088)(0.000042)(0.00075)
MotherHSgraduate-0.0104894-0.0000534-0.00012070.00057850.010085
(0.0112609)(0.0000599)(0.0001353)(0.0006495)(0.0108097)
FatherHSgraduate0.0313615**0.0001597***0.0003609***-0.0017297***-0.0301524**
(0.0125563)(0.0000596)(0.0001388)(0.0006315)(0.0121499)
MotherCollege-0.0282473**-0.0001439*-0.0003251*0.0015579*0.0271584**
(0.0128544)(0.0000762)(0.0001748)(0.0008132)(0.012311)
FatherCollege0.02005730.00010220.0002308-0.0011062*-0.0192841
(0.0128393)(0.0000626)(0.0001428)(0.000671)(0.0123845)
Male0.0217359**0.0001107**0.0002501**-0.0011988**-0.0208979**
(0.0088474)(0.0000503)(0.0001165)(0.0005318)(0.0085018)
White-0.1596195***-0.000813***-0.0018368***0.0088034***0.1534659***
(0.0378034)(0.0002598)(0.0006123)(0.0027347)(0.0361386)
Black-0.0623327-0.0003175-0.00071730.00343780.0599297
(0.0393656)(0.0002209)(0.0005017)(0.002392)(0.0377252)
Familysize-0.0170538***-0.0000869***-0.0001962***0.0009406***0.0163963***
(0.0039941)(0.0000209)(0.0000516)(0.0002076)(0.003886)
SingleMother0.0826392***0.0004209***0.000951***-0.0045578***-0.0794533***
(0.0135943)(0.0001215)(0.000291)(0.0012606)(0.0128353)
HomeOwnership-0.0289948***-0.0001477**-0.0003337**0.0015991**0.027877***
(0.0097309)(0.0000618)(0.0001443)(0.0006538)(0.0093087)
Adults0.00585070.00002980.0000673-0.0003227-0.0056251
(0.0157949)(0.0000795)(0.0001801)0.000859)(0.0151963)
N8988
DeltaMethodStandarderrorinparentheses. ∗p <
,
0.10 ∗
∗p <
,
0.05 ∗∗
∗p < 0.01.
152
4.4.4.3 Comparison of Part-time and Full-time Working Mothers
Some previous works (for example, Ruhm, 2004) focus on the distinction between
fulltime versus part-time employment rather than on the number of work hours. This
section analysis allows for nonlinearities by separating part-time and full-time jobs, using
1700 hours per year as the threshold between the two. Table 4.13 shows that average
annual work hours were replaced by dummy variables indicating whether the mother was
employed full-time or part-time during the specified period. The estimates imply that full-
time employment yields a better (38 percentage points higher probability) childhood
health return and indicate that the positive income effect of maternal employment on child
health outweighs the negative impact of employment due to time reduction from childcare.
Comparable results are obtained using
1800 and 1900 hours per annum as a cutoff.
153
Table 4.13: IV Probit Estimation for Full-time and Part-time Working Mothers
Main Model Marginal Effect
Mothers Fulltime 1.233*** 0.3845***
(0.0557) (0. 0162)
Mother’s Age at Birth 0.00602*** 0.0019***
(0.00204) (0.0006)
Mother HS graduate 0.183*** 0.0574***
(0.0330) (0.0103)
Father HS graduate 0.0778** 0.0244**
(0.0356) (0.0112)
Mother College 0.212*** 0.0664***
(0.0382) (0.0119)
Father College 0.0838** 0.0263**
(0.0367) (0.0115)
Male -0.102*** -0.0319***
(0.0264) (0.0083)
White 0.629*** 0.1973***
(0.0935) (0.0292)
Black 0.437*** 0.1372***
(0.0960) (0.0301)
Family size -0.00566 -0.0018
(0.0105) (0.0033)
Single Mother -0.351*** -0.1102***
(0.0339) (0.0106)
Home Ownership 0.110*** 0.0345***
(0.0278) (0.0087)
Adults 0.0883** 0.0277**
(0.0444) (0.0139)
154
Constant -1.358***
(0.160)
Cohort Fixed Effect YES YES
Robust Standard errors in parentheses.
4.4.4.4 What about Fathers?
The preceding analyses investigate the importance of maternal work intensity on
children’s health status. In those analyses, the presence of fathers in the household is
controlled by the dummy variable “Single Mother,” which indicates whether a single mother
was raising a child. But what about fathers? While mothers may provide unique inputs,
there seems likely to be at least some substitutability between parents. However, since men
are typically paid more than women, more considerable income benefits could accrue to
paternal employment. These
issuesareaddressedinTable4.14,whichpresentsIVprobitestimationresultsthatincludeboth
parents’ average annual work hours as part of the covariates.
The maternal average annual work hours coefficient is virtually unchanged, indicating
that this omission does not affect the findings of previous sections. Each additional 1000
average annual hours a mother works decreases a child’s likelihood of having a healthy
childhood by around 0.6 percentage points. However, the labor supply of fathers appears to
have either no impact or a beneficial consequence.
Table 4.14: IV Probit Estimation with Fathers’ Hours Worked Covariate
Main Model Marginal Effect
Mother’s HW Childhood 0.00151*** 0.000591***
(0.0000484) (0.0000206)
Father’s HW Childhood 0.0000102 0.00000401
155
(0.0000301) (0.0000118)
Mother’s Age at Birth -0.00407** -0.00159**
(0.00196) (0.000769)
Mother HS graduate 0.0699** 0.0274**
(0.0315) (0.0123)
Father HS graduate -0.0491 -0.0192
(0.0350) (0.0137)
Mother College 0.126*** 0.0493***
(0.0365) (0.0142)
Father College -0.0411 -0.0161
(0.0343) (0.0135)
Male -0.0799*** -0.0313***
(0.0239) (0.00936)
White 0.423*** 0.166***
(0.0970) (0.0379)
Black 0.0963 0.0377
(0.0987) (0.0387)
Family size 0.0386*** 0.0151***
(0.0104) (0.00409)
Home Ownership 0.0636** 0.0249**
(0.0256) (0.0100)
Adults -0.00909 -0.00356
(0.0408) (0.0160)
Constant -2.347***
(0.153)
Cohort Fixed Effect YES
Robust Standard errors in parentheses.
156
Generally, the effect of maternal hours worked on childhood health is robust for
alternative forms of specification for alternative definitions of the dependent variable.
4.5 Discussion and Conclusion
Mothers’ labor force participation dramatically increased over the last few decades and
this calls for researchers’ attention to investigate the effect on child development and
health. Prior literature has produced mixed results on the topic (Bernal & Keane, 2005). On
the one hand, some studies demonstrate a negative and significant impact of mothers’
employment on child development(James-Burdumy,2005;Bernal,2008;Hilletal.,2005).
Ontheotherhand,certain studies find no significant relationship between maternal
employment and child development (Blau & Grossberg, 1990) and between maternal
employment and child health (Gwozdz et al., 2013; Baker & Milligan, 2008). Others show a
negative and significant effect of maternal employment on child health (Gennetian et al.,
2010; Morrill, 2011).
Most of the studies focus on mothers’ labor force participation effect on child
development because of the wider availability of objective measures of child development
indicators like academic performance. None of the previous studies investigates the effect
of maternal work intensity on child health using a self-reported general health outcome
variable. Furthermore, no study investigates the heterogeneity of maternal work intensity
on child health across stages of child development. These research gaps left by previous
studies and the mixed findings discussed earlier are the motivation for conducting this
study. In addition, most of the studies on the area use OLS (Bernal & Keane, 2005) and have
limitations in drawing causal interaction between maternal labor market choice and child
health due to the endogeneity problem. This study aims to gauge the causal association
between mothers’ work intensity and childhood health using the data from the PSID. Due to
157
the endogeneity problem of the main independent variable (mothers’ annual hours
worked), the study employs IV Probit regression estimation techniques.
