University of Southern California Dissertations and Theses

Three essays on human capital and family economics

(USC Thesis Other)

Three essays on human capital and family economics

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content

Three Essays on Human Capital and Family Economics

by
Jeonghwan Yun

A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(ECONOMICS)

August 2020

Copyright 2020 Jeonghwan Yun

ii
Acknowledgments

My six years of pursuit to become a researcher would not have been possible without the support
of many. I am grateful to my committee members for their guidance and patient support. John
Strauss was an attentive advisor who always demanded high academic integrity. He not only
trained me in research skills but also taught me critical thinking and that most essential attribute,
scholarly morale. His voice and words will ring in my head throughout my career. Paulina Oliva,
whose light helped me navigate around many dark corners, was a consistent source of advice and
sharp insight whenever I encountered challenges. Vittorio Bassi, who has witnessed my long
journey up close, served as a role model for me as a young researcher and taught me how to develop
and structure my unorganized raw ideas. Jinkook Lee, a patient supporter in all aspects, encouraged
and motivated me to broaden my perspective and ask the right questions. Emily Nix, with her
academic enthusiasm and expertise in the field, inspired me to pursue excellence tirelessly in every
smallest detail and to follow those towards a more profound understanding of the bigger questions.
I have learned immensely from Jeffery Nugent, Geert Ridder, Hyungsik Roger Moon, Jeffery
Weaver, Daniel Bennett, Hyeok Jeong, and other faculty members in the Department of
Economics. Being trained by my advisors has been an honor and privilege, and I cannot find
enough words to express my gratitude. I am grateful for the institutional support from Young
Miller, Alexander Karnazes, and many other members of the University of Southern California for
all the opportunities they have provided me.
My colleagues and friends have been my biggest assets over the past six years, as we navigated
our joys and struggles together. Specifically, I would like to express my most sincere gratitude to
my coauthors, Tushar Bharati, Yiwei Qian, Dawoon Jung, and Seungwoo Chin. They have been

iii
my biggest supporters and motivators, both academically and personally. In addition, I would like
to mention how much I appreciate my dearest friends: Mahrad Sharifvaghefi, Brian Finley,
Fabrizio Piasini, Bora Kim, Hae Yeun Park, Yu Cao, Jisu Cao, Jingbo Wang, Andreas Aristidou,
Eunjee Kwon, Bada Hahn, Donwook Kim, Grigory Franguridi, Ali Aboud, Daehyun Kim, Yoon
Jae Ro, and many more that I cannot list all of their names but I am indebted to. During the toughest
moments, I was able to get back up only because of their compassion and inspiration, and none of
my accomplishments would have been possible without this companionship.
Lastly, I would like to thank my family for their constant love and support. My parents have
given me the ability to move dauntlessly forward, and my little sister has been a constant reminder
of who I am and where I am headed.

iv
Table of Contents

1.1. Introduction ................................................................................................................................ 5
1.2. Data .......................................................................................................................................... 15
1.3. Theoretical Model: Unidimensional ......................................................................................... 19
1.4. Empirical Strategy and Estimation Method.............................................................................. 22
1.5. Unidimensional Analysis Results ............................................................................................. 30
1.6. Theoretical Model: Multidimensional ...................................................................................... 33
1.7. Empirical Strategy: Multidimensional ..................................................................................... 37
1.8. Results: Multidimensional ........................................................................................................ 39
1.9. Goodness of Fit ........................................................................................................................ 43
1.10. Policy .................................................................................................................................... 46
1.11. Conclusion and Discussion ................................................................................................... 50
1.12. Tables and Figures ................................................................................................................ 53
2.1. Introduction .............................................................................................................................. 77
2.2. Background .............................................................................................................................. 81
2.3. Data .......................................................................................................................................... 85
2.4. Estimation Strategy .................................................................................................................. 86
2.5. Result ........................................................................................................................................ 89
2.6. Conclusion ................................................................................................................................ 93
2.7. Tables and Figures .................................................................................................................... 95
2.8. Appendix ................................................................................................................................ 102

v
3.1. Introduction ............................................................................................................................ 105
3.2. Background ............................................................................................................................ 108
3.3. Data and Identification ........................................................................................................... 111
3.4. Empirical Specification .......................................................................................................... 113
3.5. Result ...................................................................................................................................... 118
3.6. Conclusion .............................................................................................................................. 132
3.7. Tables and Figures .................................................................................................................. 134

vi
List of Tables
Table 1.1: Descriptive Statistics_S1 ............................................................................................. 53
Table 1.2: Descriptive Statistics_S2 ............................................................................................. 54
Table 1.3: Moments ...................................................................................................................... 55
Table 1.4: Household income ....................................................................................................... 56
Table 1.5: Preference parameters .................................................................................................. 56
Table 1.6 : Wage parameter .......................................................................................................... 57
Table 1.7: Initial ability parameter ................................................................................................ 58
Table 1.8: Ability parameter ......................................................................................................... 59
Table 1.9: Additional Moments .................................................................................................... 60
Table 1.10: Household income ..................................................................................................... 61
Table 1.11: Preference parameter ................................................................................................. 62
Table 1.12: Wage parameter ......................................................................................................... 63
Table 1.13: Initial ability parameters ............................................................................................ 64
Table 1.14: Ability parameters ..................................................................................................... 65
Table 1.15: The correlation-coefficient—Simulated preference parameter ................................. 66
Table 1.16: Goodness of fit ........................................................................................................... 67
Table 1.17: Effects of policy intervention (%) ............................................................................. 68
Table 1.18: Mothers’ responses to policy intervention ................................................................. 69
Table 2.1: Number of Warning/Alert Issued Yearly .................................................................... 95
Table 2.2: Avoidance Behavior-All types of venues .................................................................... 96
Table 2.3: Avoidance Behavior-by DVSN Subgroups ................................................................. 97
Table 2.4: Avoidance Behavior-by SCTN Subgroups .................................................................. 98

vii
Table 3.1 : Summary Statistics (Data: IPC and SUPAS) ............................................................ 134
Table 3.2: Summary Statistics (Data: IFLS) ............................................................................... 135
Table 3.3 : Impact on household's LPG usage status (Data: IPC an SUPAS) ............................ 136
Table 3.4 : Impact on household’s LPG usage status (Data: IFLS) ............................................ 137
Table 3.5 : Impact on labor force participation status (Data: IPC and SUPAS) ......................... 138
Table 3.6 : Impact on labor force participation status (Data : IFLS) .......................................... 139
Table 3.7 : Impact on labor force participation in previous years (Data : IFLS) ........................ 140
Table 3.8 :Impact on measured health (Data: IFLS) ................................................................... 141
Table 3.9 : Impact on reported diagnosis of health conditions (Data: IFLS, for age above
40 only) ............................................................................................................... 142
Table 3.10 : Impact on expenditure (Data: IFLS) ....................................................................... 143
Table 3.11 : Impact on food items (Data : IFLS) ........................................................................ 144
Table 3.12 : Impact on expenditure on fuel and utilities (Data: IFLS) ....................................... 145
Table 3.13 : Subjective well-being (Data: IFLS) ........................................................................ 146
Table 3.14 : Correlates of fuel choice and decision-making power of women in 2000
(Data: IFLS) ........................................................................................................ 147
Table 3.15 : Impact on decision-making power of women (Data: IFLS) ................................... 148
Table 3.16 : Impact on household's amenities (Data: IPC and SUPAS) .................................... 149

viii
List of Figures
Figure 1.1: Female Labor Market Participation ............................................................................ 70
Figure 1.2: Fertility rate in South Korea(2000-2018) ................................................................... 70
Figure 1.3 : Earmarked Government Budget (2018) .................................................................... 71
Figure 1.4: Mothers’ Choice Variable (Data) ............................................................................... 72
Figure 1.5 : Distribution of mothers’ observed wage (Observed) ................................................ 73
Figure 1.6: Relative Motivation of Different Investments ............................................................ 74
Figure 1.7: Mothers’ choice variable (Simulated) ........................................................................ 75
Figure 1.8: Simulated demand for market care ............................................................................. 76
Figure 2.1 : Average PM10 by GoonGu(Municipality) ................................................................. 99
Figure 2.2 : Average PM10 by Year ............................................................................................ 100
Figure 2.3 : Average PM10 by Month ......................................................................................... 100
Figure 2.4: Golf Course- Domestic Visitors against PM10 Level ............................................... 100
Figure 2.5: Golf Course- Foreign Visitors against PM10 Level .................................................. 101
Figure 3.1 : Trends in GDP and education in Indonesia ............................................................. 150
Figure 3.2 : Labor force participation in Indonesian and worldwide ......................................... 150
Figure 3.3 : Staggered rollout of the LPG subsidy program across provinces ........................... 151
Figure 3.4 : Difference in LPG program roll-out across IFLS communities .............................. 152
Figure 3.5 : Primary cooking fuel (Survey: IPC and SUPAS) ................................................... 153
Figure 3.6 : Primary cooking fuel by program exposure status (Survey: IPC and SUPAS) ...... 154
Figure 3.7 : Primary cooking fuel by pre-program kerosene usage (Survey: IPC and
SUPAS) ............................................................................................................... 155

ix
Figure 3.8 : Change in LPG usage by pre-program kerosene usage (Survey: IPC and
SUPAS) ............................................................................................................... 156
Figure 3.9 : Primary cooking fuel (Survey: IFLS) ...................................................................... 157
Figure 3.10 : Change in LPG usage by pre-program kerosene usage (Survey: IFLS) ............... 158
Figure 3.11 : Labor force participation by program exposure status (Survey: IPC and
SUPAS) ............................................................................................................... 159

x
Abstract
This dissertation contains three essays on the topic of human capital formation and family
economics. Chapter 1, aims to discern how a mother’s time allocation and investment behavior
toward a child impact early childhood development, in both cognitive and non-cognitive abilities.
While the female labor market participation rate has been increasing in developing and developed
countries, it has not been fully explored whether market-provided childcare can perfectly substitute
mother’s care. To answer the question, to which extent market-provided childcare can replace
mother’s care, this paper takes a structural approach by solving mothers’ dynamic optimization
problem, and adopt SMD (simulated minimum distance) for the corresponding empirical strategy.
From two sets of analyses using data from PSKC (Panel Study of Korean Children), this study
extends existing literature’s exclusive focus on cognitive development to include multidimensional
development, and reveals that market-provided childcare may be superior in certain non-cognitive
dimensions, but not as much in cognitive development; neglecting this multidimensionality in
childhood abilities might lead to a biased estimation of market care’s contribution to the cognitive
dimension. This paper also conducts counterfactual policy simulations and compare four policies
commonly used to reduce mothers’ burdens: maternity leave, cash transfers, subsidized childcare
institutions, and quality improvement in childcare institutions. The result implies how current level
of the 50% wage replacement under paid maternity leave, may be too low to induce mothers to
take up the opportunity. Additionally, the result shows that while cash transfer benefits the most
in both mothers’ welfare and children’s ability, the quality improvement in childcare institutions
may benefit mothers and children, without sacrificing short-run economic vigor.
Chapter 2 investigates the effect of Particulate Matter 10 micrometers and smaller (PM10), and
a government policy that counteracts to this air pollution. Using difference-in-difference

xi
identification available due to the staggered schedule of monitoring system introduction in the
Seoul metropolitan area, we assess the impact on avoidance behavior. We present how, in days
with high PM10 level, the existence of a monitoring system significantly reduces the number of
visitors in tourist destinations. Additionally, such avoidance behavior is found the strongest for
outdoor venues, such as golf courses and ski resorts. However, foreigners, who are less likely to
benefit from the government policy, do not reveal this strong pattern in avoidance behavior. These
results provide valuable insights into how the provision of government policy may affect the public
differently due to strong heterogeneity in accessibility to the information provided to the public.
In Chapter 3, using the staggered roll out of the Indonesian “Conversion to Liquefied Petroleum
Gas (LPG) Program”, we show that a switch to the labor- and time-saving technology of cooking
with LPG increased the labor force participation of exposed women. The program caused an
increase in household expenditure on food and education and the subjective wellbeing of women.
We also show that the policy improves the decision-making power of women in the household,
especially in financial matters. A back-of-the-envelope calculation suggests that saving in
households expenditure on fuel far outweighed the cost of the conversion incurred by the
government. To the extent that intra-household externalities and gender differences in preferences
are a reason for the lack of adoption of such cost-effective technology, the results highlight that
temporary subsidies that empower women can encourage the adoption and sustained use of such
technology.

1
Introduction

The development and management of human capital have been undoubtedly one of the essential
factors for an individuals’ economic success and the economic prosperity of society. The
breakthrough research by Gary Becker [Becker (1967)] proposed a concrete concept. Numerous
studies have attempted to fully understand the dynamic process of human capital management at
various stages of life, its impact on labor market outcome, and its relation to aggregate economic
progress [Heckman (1976), Carneiro & Heckman (2003), Hanushek (2008), Doyle et al. (2009),
Heckman & Kautz (2012)]. However, until recently, the complexity of human capital formation
and its utilization have been the focus of research questions of labor, education, and family
economists.
Through three essays in this dissertation, I examine how human capital is developed,
maintained, and utilized over a lifetime, and how a family decides the investment pattern on it.
Furthermore, I assess the consequences of specific policies that aim to promote the investment
behavior or utilization of dormant human capital and discuss how this changes the decisions of the
families and individuals.
The first essay examines the development and dynamic evolution of human capital. Specifically,
multidimensional early childhood development is examined. Existing literature focuses on
unidimensional cognitive development in early childhood. However, the first essay expands the
scope to multidimensional development, which encompasses non-cognitive and cognitive
development. The importance of non-cognitive skills has drawn special attention since the 2000s
[Cunha & Heckman (2007), Cunha & Heckman (2010)], with the emphasis on the “windows of

2
opportunity” for development in early childhood and its significance in later life [Carneiro & Heckman
(2003), Doyle et al. (2009)].
The essay evaluates the efficacy of the different time investments on a child, including the more
traditional forms of childcare provided by a mother, and market-provided childcare, such as
kindergartens, daycare centers, or even babysitters. Previous studies have tried to pin down
whether maternal care is more efficient than the commercial forms of childcare. The study
demonstrates that this framework is overly restrictive and shows the necessity to generalize this
question for the appropriate evaluation of different types of investment.
Several government policies are available to support mothers raising young children. Maternity
leave or subsidization of childcare services in institutions are some of the typical policies available
to alleviate a mothers’ burden. In a unidimensional framework, the coexistence of such policies is
not strongly justified because one of these policies is likely to be more efficient than the other in
the development of human capital. However, I propose a theoretical background in which the
composition of these policies can work more efficiently. Therefore, the coexistence of seemingly
counterarguing policies may be justified as an optimal choice for policymakers.
The second chapter focuses on a specific type of human capital, health. While health and its
economic impact have consistently been of interest in the field of economics [Grossman (1972),
Thomas & Strauss (1997)], it has recently gained additional interest due to its relations to
environmental issues [Ponce et al. (2005), Neidell (2009), Zivin and Neidell (2009), Graff Zivin
and Neidell(2012), Arceo et al. (2016)]. Specifically, the challenges related to particulate matter
(size less than 10μg (PM10 or Particulate Matter 10)), has become a serious concern in North-East
Asia, one of the regions most industrialized areas over the last few decades [Kwon et al. (2002),
Jia and Ku (2015)].

3
The research question in this chapter is related to the policy on PM10. Therefore, whether there
is a substantial avoidance behavior exercised by the public once additional information becomes
available is examined here. While information access is the most commonly used policy to
counteract ambient air pollution, the benefit depends on the details of the policy. Unlike other
policies that have been reviewed in the literature, here the assessed government intervention 1) is
brought out at a geographically fine level, 2) provides a real-time measure of air pollutants, and 3)
information is spread out by new media primarily utilizing internet and mobile technology, as
opposed to traditional media such as newspapers.
Consequently, while the benefit of information might be larger because more accurate
information is provided compared to the traditional counterpart, the cost of acquiring the
information might become higher for those disadvantaged. Thus, Chapter 2 will pay further
attention to such heterogeneous patterns of avoidance behavior and discuss whether the
maintenance of human capital is efficient.
In the third chapter, I discuss the means to utilize the human capital that society is endowed with
fully. The effect of the introduction of new technology in a developing country and whether such
a policy can work as an engine to liberate women is analyzed [Greenwood et al. (2005)]. The third
chapter focuses on a developing economy, Indonesia, and its policy on conversion to liquefied
petroleum gas (LPG). This policy aims to promote the usage of cleaner cooking fuel, reducing
indoor air pollution caused by commonly used alternatives like kerosene, and other solid fuels.
While I assessed the effect of the introduction of cleaner cooking fuel in households in various
aspects of family behavior, three major dimensions that might have a direct impact due to the
policy are emphasized on female labor market participation and decision making within
households. Most likely, women are the main beneficiaries of cleaner cooking fuel in a

4
conservative society. By considering the labor market participation, an assessment was made of
whether this new technology indeed liberated women who were more likely to be caregivers of
families, and encouraged women to participate in the labor market. Additionally, the main channel
of such an impact is also discussed.
While Chapter 3 provides a conclusion stating that the benefit of a cleaner cooking fuel
surpasses its cost in multiple ways, the question remains that why the cleaner cooking fuel had not
been widely adopted before the government intervention. This question is answered by examining
the impact of introducing LPG on the decision-making process within a family. Conversion to
LPG does not only affect the short-run labor market decisions of the family but may even lead to
a perpetual structural change within a family, and work as an engine of liberation for women.

5
Is Market-Provided Childcare “Better” for Early Childhood
Development? A Structural Approach with a Korean Context

1.1. Introduction
This paper aims to answer the question regarding working mothers’ impact on children’s
development in their earliest stages of life.
Women’s participation in the labor force has steadily increased over the last few decades. Figure
1.1 illustrates this pattern among selected developing and developed countries in the most recent
three decades. Clearly, such a trend is not restricted to countries experiencing intensive economic
development, as this can also be observed in developed countries.
This has naturally led to an interest in how such labor market participation has affected
children’s development, and especially in their earliest stages of childhood. The relationship
between female labor force participation and economic growth and development has a longer
history that has been well-documented and explored in various contexts (Clark et al., 1991; Tsani
et al., 2015). However, literature has not fully examined the effects of this trend from a more
traditional perspective. This paper specifically focuses on the impact of mothers’ labor market
participation on children’s development regarding both cognitive and non-cognitive abilities.
As one of the earliest discussions on the topic, Preston (1984) notes that “it is not at all clear
that mothers’ work is a source of disadvantage for children, at least not as a direct determinant.
Recent reviews of studies of the effect of working mothers on child development find very few
and inconsistent effects...” However, studies’ responses to this issue vary only 30 years after
Preston’s (1984) work, depending on each study’s different contexts or methods used.

6
The answer to this topic ultimately depends on the socio-economic context. This paper examines
developed countries, in which working mothers are likely to have easy access to well-established
childcare services as alternatives in terms of childhood development. The natural question to ask
in such a context involves whether this market childcare can perfectly substitute for the mother’s
childcare, which is likely the only option for mothers in more conservative or underdeveloped
societies.
Some attempts have been made to answer the question, “To what extent can market childcare
replace the mother’s care?” (Bernal, 2008; Brilli, 2013). These noteworthy studies focus only on
one dimension of the childhood development: cognitive ability or sub-domains thereof. However,
different childhood abilities may develop through different types of childcare, or different types of
investments in general. Ignoring the multi-dimensionality in children’s abilities may lead academia
to false conclusions.
This paper’s primary contribution involves extending interests in literature to include non-
cognitive ability, which existing studies have neglected. As a preview of the main findings, this
paper shows that childcare provided by childcare institutions is superior in terms of developing
children’s sociability, although the mother’s care is still efficient in developing a child’s language
and communication skills. This finding contradicts this paper’s unidimensional analyses, which
conclude that market care more effectively promotes language development, as revealed in earlier
studies; this result reveals the need for a focus on multidimensional ability.
A comparison between the efficacy of mothers’ care compared to market care is also a policy
issue highly relevant to countries facing low fertility rates. Specifically, this paper utilizes data
from South Korea, in which low fertility rates have been a major concern for the past two decades.
Figure 1.2 displays the dramatic decrease in the nation’s fertility rates during this period. Since

7
2005, the government has invested 153 trillion South Korean won (KRW) on policies that might
relieve burdens on potential parents. Figure 1.3 illustrates how the government’s budget has been
distributed among its relevant policies. While the biggest proportion of the budget is assigned to
the enhancement of market care, such as customized childcare or educational reform, a fair amount
is also dedicated to policies that promote a balance between family and work . For example, the
career-family compatibility category includes funds budgeted for family leave. It is noteworthy
that different policies are likely to alleviate different child-rearing costs. A childcare support
system may relieve burdens for working mothers, and mothers on maternity leave are likely to
spend more quality time with their children. While these different types of policies intend to
alleviate the burdens of raising a child, their mechanisms and efficacy have not been fully assessed.
Based on estimation results, this paper aims to provide a comparison between these substantially
different policies.
Empirically, this paper’s interests can be represented by estimating some variation of a dynamic
evolution process for childhood ability, conceptually expressed as

ln 𝐴𝐴 𝑡𝑡 + 1
= 𝛿𝛿 1𝑡𝑡 ln 𝜏𝜏 𝑡𝑡 + 𝛿𝛿 2𝑡𝑡 ln 𝑖𝑖 𝑡𝑡 + 𝛿𝛿 3𝑡𝑡 ln 𝑚𝑚 𝑡𝑡 + 𝛿𝛿 4𝑡𝑡 ln 𝐴𝐴 𝑡𝑡 (1.1)

where 𝐴𝐴 𝑡𝑡 represents a child’s ability or multi-dimensional abilities. Such an ability will depend
on its lag variable, but is also developed through the mother’s different types of investment: 𝜏𝜏 𝑡𝑡 , 𝑖𝑖 𝑡𝑡 ,
and 𝑚𝑚 𝑡𝑡 represent the mother’s quality time with a child, the time spent in the childcare market,
and other material investments, respectively. A clear empirical challenge exists in verifying the
productivity coefficient 𝛿𝛿 , as each type of investment is the mother’s endogenous choice.
Additionally, such endogenous investments are chosen as the mother’s dynamically optimal

8
investments due to the dynamics of an ability’s evolution. The following subsection summarizes
some noteworthy earlier studies that attempt to examine this evolutionary process in early
childhood.

1.1.1. Childhood Development
Studies across multiple disciplines have agreed upon the importance of early childhood
development. Doyle et al. (2009) provide an overview on this topic in various fields. For example,
numerous developmental neuroscience studies have illustrated how a “greater plasticity of the
brain in early periods” defines the “windows of opportunities for certain developments to take
place”; these also mention how “inadequate stimulation in [an] early period can result in [a] large
and lasting negative effect on subsequent development.” Compared to this consensus from the
natural sciences, economics studies have focused more on early life interventions’ effects on later-
life outcomes. While some questions remain regarding longer-term dynamic complementarity, the
paper illustrates how investments’ rates of return are higher with earlier intervention, even
antenatal if possible. This impacts not only language skills, educational achievement, or labor
market earnings, but also social skills, behavioral problems, and physical and mental well-being
(Hack et al., 1992).
As previously mentioned, this study focuses on verifying whether mothers’ labor market
participation negatively affects children’s early development, with a concentration on the impacts
on the cognitive and non-cognitive dimensions. Brooks-Gunn et al. (2010) provide a correlational
study that documents the relationships between key variables, which are also this paper’s variables
of interest. Their study uses a sample of non-Hispanic white and African-American data from the
NICHD to discover a negative association between mothers working full-time and children’s

9
cognitive skills later in life. Regarding problematic behavior—a non-cognitive dimension—full-
time maternal employment during the child’s first year was negatively associated with
externalizing behavior. Additional to empirical findings based on ordinary least squares
specifications, the study’s structural equation modeling indicated no significant indirect effect
from center-based care, a common alternative means of childcare for working mothers.
Cunha et al. (2010) demonstrate how empirically analyzing early childhood development
without prudence may yield misleading conclusions. Although the key variables of interest differ
from those in this paper, the authors examine NLSY, the authors provide theoretical and empirical
evidence of not only how dynamics function in skill formation, but also why the endogeneity of
such investments should be considered. However, while Cunha et al (2010) provide valuable
insights under a more general setting, it is noteworthy that the current paper cannot be directly
compared with this work, as the latter focuses on the effects of specific types of investments by
modeling a problem pertaining to mothers’ dynamic utility maximization rather than specifying a
linear investment equation. However, as Cunha et al. (2010) advise, our model will seriously
consider 1) the time-varying nature of the skill-formation function, and 2) the multi-dimensionality
of children’s skill formation.
Prior studies illustrate the topic’s complicated nature. Clearly, the key challenges originate from
how the mother’s observable variables—such as her decision to work or the different types of
investments in her child—are all results of her endogenous choice, and these may be static as well
as dynamic. Literature reveals different sets of studies addressing such challenges regarding
endogeneity.
The first set of existing studies solves this problem by seeking proper instrumental variables.
These are derived either from policy interventions, such as maternity leave; market conditions,

10
such as the regional average wage rate; or field experiments. However, these studies do not reach
a consensus.
Other studies involve local labor market conditions, which are used as an instrument for the
mother’s choice variables. Baum II (2003) presents earlier research on the field, in which mothers’
employment status prior to the birth of a child is used as an instrumental variable for maternity
leave. The study uses US data from NLSY to reveal that mothers working in an infant’s first year
negatively affect their children’s cognitive development. This is measured by such tests as the
Peabody Picture Vocabulary Test (PPVT), performed on children three to five years old. Non-
maternal childcare was controlled, but not fully accounted for as a variable for endogenous choice.
The author addresses such concerns that the results might not provide total effects.
1
Another
noteworthy point from this study involves the positive association between family income and
cognitive development. Thus, the paper comments that the negative effect from working mothers
can partially be offset by additional income.
James-Burdumy (2005) also observes the US context using NLSY, but exploits the panel
structure with an IV-FE specification. The study instrument is mother’s employment, measured by
county-level employment rates. However, unlike Baum II (2003), James-Burdumy (2005) finds
mixed results, as the number of weeks worked in the first year negatively affect cognitive
development in the first year, while positive long-term impacts are observed in the third year. Such
a result reveals the importance of considering mothers’ dynamic decisions.

1
“However, results that control for nonmaternal childcare arrangements must be interpreted with caution: selected
childcare arrangements may be partially determined by maternal labor supply decisions. If this is true, then estimates
from these models will not give the total effect of maternal labor supply” Baum II (2003).

11
A few studies use policy variations—or typically, maternity leave—as an exogenous variation
of prior interest, and some papers take this approach to specifically examine long-term effects.
Dustmann and Schönberg (2012) and Rasmussen (2010) both analyze later-life labor market
outcomes for children with mothers who were exposed to maternity leave reform in Germany and
Denmark, respectively. Both studies found that expanding maternity leave significantly and
positively affected mothers’ return-to-job behaviors, income, and job opportunities. Regardless of
this significant first stage, both papers found no effects on either children’s educational
achievements or earned wages.
However, results from Carneiro et al. (2015) disagree with these previous studies. These authors
examined Norway’s 1977 maternity leave reform to discover decreased high school dropout rates
of 2% to 3%, and 4.5% to 7% higher earnings for these children around age 30 for those with
mothers exposed to mandatory maternity leave.

1.1.2. Multidimensional Childhood Ability
Aside from disagreements among studies that use administrative data, other studies have
analyzed the multidimensional nature of early childhood development using randomized control
trials (Attanasio et al., 2015; Meghir et al., 2015). Further, Attanasio et al. (2015) utilize data from
early childhood intervention in Colombia, in which intervention using randomized control trials
consists of both psychological stimulation—equivalent to the widely known Jamaican Study
(Grantham-McGregor et al., 1991; Walker et al., 2011; Gertler et al., 2014)—and nutrient
supplementation. This study is one of a few in which a fair amount of attention is given to the
formation of non-cognitive skills, and specifically socio-emotional skills. The authors analyzed
the behavioral changes of mothers due to intervention, and form a control function to control for

12
mothers’ endogenous investment decisions in their children’s skill formation function. Although
not directly comparable, they differ from Cunha et al. (2010) in the failure to find substantially
significant effects in the non-cognitive dimension.
The mother’s investment choice considered by Attanasio et al. (2015) include material
investments and the maternal time spent with her child. This is a natural choice considering such
intervention targeted underprivileged households in Colombia. In this context, one compelling
finding in their study is that policy interventions did not change the mother’s decision regarding
labor market participation, but did change the maternal time spent with children. Our paper
examines a different context, in which a clearer tradeoff might still exist between the mother’s
labor market participation and investment in children; market-based childcare may be an attractive
option for working mothers, in that this is a significant factor in childhood development.
Another noteworthy finding from work by Attanasio et al. (2015) is that while the empirical
estimations for the human capital production function are assumed as general CES function, its
estimate for the substitution of elasticity implies the production function is close to a Cobb-
Douglas production function, regardless of the choice of dimension. While it is possible to extend
the general CES function, our paper will take this result seriously to restrict our interests under a
Cobb-Douglas assumption.

1.1.3. Structural Approach in Developed Country Context
Another stream of literature exists in which mothers’ endogenous decisions were explicitly
considered from a dynamic decision model perspective; our study takes a similar route (Bernal,
2008; Brilli, 2013). While both Bernal (2008) and Brilli (2013) use the structural approach with
slight differences in their detailed assumptions, they fail to reach a consensus in their primary

13
conclusions; Bernal (2008) discovers that mothers’ labor market participation negatively affects
their children’s cognitive development, while Brilli (2013) finds no such effects.
The key difference in Brilli’s (2013) empirical strategy is the distinction between the quality
time spent with the child and the mother’s leisure, with elaborate time-use variables derived from
the Panel Study of Income Dynamics. Although the socio-economic context may not be directly
comparable, such a result is aligned with findings from Attanasio et al. (2015), in which mothers
spend more times with their children but no significant decrease occurs in labor market
participation. In such a case, using maternal care hours—proxied by non-working hours, as in
Bernal’s (2008) work—might yield false estimates for the childhood development function.
It is noteworthy here that both Bernal (2008) and Brilli (2013) only focus on cognitive ability,
and Bernal’s (2008) work uses NLSY data for its empirical analyses. As in many other previously
mentioned relevant studies that use NLSY, Bernal’s (2008) work concentrates on cognitive
development through such tests as the PPVT and the Peabody Individual Achievement Test’s
reading recognition and mathematics subtests. Similarly, the Brilli’s (2013) primary variable is the
letter-word score, which measures a child’s learning and reading skills. However, working
mothers’ effects on children’s non-cognitive dimension remains unaddressed in these studies.
While details will be provided later, we will use data from the Panel Study on Korean Children
(PSKC), which involves detailed time-use variables as well as measures of non-cognitive
temperament. Subsequently, this paper aims to identify the effects of mothers’ time-use decisions
on both cognitive and non-cognitive dimensions.
As the PSKC allows us to distinguish the time spent on maternal childcare and mothers’ leisure
time, we will consider Brilli’s (2013) model as our benchmark. One noteworthy aspect of Brilli’s
(2013) study is that it may not be directly comparable with our study, as the full structural model

14
depends upon the context, such as mothers’ wage offers. Thus, our motivation involves estimation
results from mechanically and separately applying Brilli’s (2013) specifications to our Korean
context regarding both cognitive and non-cognitive development. These results will motivate us
to present the modified specifications and results incorporating all non-cognitive dimensions.
While mothers’ comprehensive investments in their children’s ability and outcomes will be
jointly analyzed, this paper particularly takes Bernal’s (2008) and Brilli’s (2013) works into
account to ask: Is it sufficient to consider only cognitive ability? This will also be a question
relevant in policy-making, as we focus on the different types of maternity leave and childcare
subsidy programs widely provided in developed countries.
Therefore, this paper contributes to current literature as the first to fully trace the dynamic
evolution of early childhood development through both cognitive and non-cognitive dimensions,
and by considering mothers’ corresponding decisions. Such analyses will be performed under a
developed country context, in which institutional care—or more generally, the childcare market—
is an attractive external option for working mothers.

The following sections consist of the following topics: Section 1.2 outlines the PSKC data set
utilized in this paper. Section 1.3 presents the theoretical model and necessary empirical
specifications for our motivational exercise, which will not significantly differ from Brilli’s (2013)
work; and Section 1.4 provides the full empirical specifications for our benchmark model. Section
1.5 presents the results from our benchmark analyses. Sections 1.3, 1.4, and 1.5 are presented to
define our motivational objectives; as the structural model will also explore variables that may
depend on the socio-economic context, it would be worthwhile to verify its compatibility by
directly comparing this paper with existing literature. Subsequently, Section 1.6 will present the

15
main theoretical model, which extends our benchmark model, and provide additional empirical
specifications. Section 1.7 presents and discusses the primary analysis results. Section 1.8 will
explore not only the work’s goodness of fit from our simulation results, but also a relevant
discussion of potential future studies. Section 1.9 presents the results of our counter-factual
simulation followed by a discussion of their policy implications, and Section 1.10 concludes.

1.2. Data
The data set used for this study is derived from the Panel Study on Korean Children (PSKC),
which is a survey conducted annually by the Korean Institution of Child Care and Education since
2008. The sample was initially comprised of a random sample of 2,078 mothers who gave birth to
a child in 2008. Of these, 1,743 (84%) remained in Wave 5 (2012), which is the last period that
this paper utilizes for its analyses. The data was examined to determine the existence of systematic
attrition regarding 21 observable baseline characteristics, and no statistical evidence was found.
Due to this paper’s ability measures of interest, the data was restricted to those between Waves
3 to 5 for two reasons: First, the PSKC began collecting detailed information about the time
mothers spent with their children beginning with Wave 3. Further, the distinction between the
hours of maternal care and non-labor hours is the core characteristic differentiating Brilli’s (2013)
work from Bernal’s (2008). As mentioned in Section 1.1, Brilli’s (2013) results disagree with those
of Bernal (2008), and the former concludes that market care is more effective in children’s
development. This paper also maintains such a distinction.
Second, the analysis is restricted to Waves 3 to 5 to construct consistent measures for the three
dimensions of non-cognitive ability that this study focuses on: emotionality, activity, and
sociability. The test scores for these three dimensions were evaluated from Wave 1 except for

16
sociability, which is considered an inadequate measurement of children younger than two years,
and was only collected starting from the PSKC’s Wave 3. Paper-based EAS surveys (The Negative
Emotionality, Activity, and Sociability-Temperament Survey for Children-Parental Ratings) were
given to parents in Waves 3 to 5, and other personal trait-related surveys were distributed after
Wave 6. However, the dimensionality of personal traits and the validity of their measurement are
still in dispute in literature on both economics and psychology. Therefore, the analysis uses only
the three waves in which the same surveys were given to avoid any inconsistent measurements of
such personality traits.

The EAS survey consists of 10 questions that ask about sociability, and 5 each for emotionality
and activity. Answers are measured using a five-point scale ranging from one to five, with five as
“most likely.” The analysis uses the average score on the five-point scale for each trait. It is
noteworthy here that this paper does not aim to indicate that the EAS measurement, or
temperament in general, is the most important non-cognition factor that policy-makers should
consider. Rather, this paper observes whether considering other dimensions could provide a
broader perspective. Moreover, research on EAS measures—and on early childhood temperament
in general—as well as these indices’ impacts on outcomes later in life are still ongoing.
Regarding the corresponding measures of cognitive ability in these three waves of interest,
several dimensions were assessed. For example, children’s problem-solving skills were evaluated
in Wave 3, and their overall cognitive abilities were surveyed through childcare facilities in Wave
5. However, the only constant assessment conducted between Waves 3 to 5 involves language and
communication skills. Therefore, this paper’s analysis utilizes scores from the Korean Ages and Stage
Questionnaires in Wave 3 and the Receptive and Expressive Vocabulary Test in Wave 4; 11 questions

17
assessed various childcare institutions. The PSKC claims that these variables all measure the
language and communications domain in childhood development, but these are inconsistent in
terms of their scales and differ in their distribution. A lack of comparability between waves is
avoided by standardizing each wave to a mean of 5 and a standard deviation of 1 in the motivational
analyses, and to a mean of 10 and standard deviation of 1 for the main analyses; the relative ranking
among the sample of children is then exploited for identification purposes. One concern might
involve whether these different tests consistently measure the same domain.
Another example of insufficient data involves mothers’ wages, which were collected in Waves
4 and 5, but not in Wave 3. However, detailed information about mothers’ occupations—including
information on the industry, sector, and position—was collected in all three waves. Admittedly,
this method may be limited, and thus, the mother’s missing wages in Wave 3 was replaced with
the average wages for the same job type in Waves 4 and 5.
Two data sets were constructed for this paper’s analyses: the S1 sample, used for this work’s
motivational analyses; and the S2 sample, used for extensive primary analyses. Excluding any data
with missing observations in any of the three waves pares down the S1 sample to 763 observations.
Table 1.1 displays the summary statistics used for the analysis, as well as other important
characteristics that may be of interest. Another item to observe is the mean and standard deviation
of language scores, which have been standardized to a mean of five and standard deviation of one.
A small discrepancy occurs due to the standardization itself, which was performed before the
selection of sample regarding missing values. Therefore, the lack of a substantial difference in the
sample’s statistics reveals the sample has no systematic selection bias.
However, an additional variable was considered for this paper’s main analysis: monetary
expenditures on the child. A balanced panel subsample with full information including these

18
additional variables indicated three waves with 568 observations. Table 1.2 presents the
descriptive statistics of the second sub-sample (S2).
The time variables for Tables 1.1 and 1.2 are all presented in 24-hour averages, and the monetary
variables are noted in denominations of 10,000 KRW. Subsequently, it is safe to note that both
institutional and private care hours are defined as the total sum of different types of institutional
and private care as reported by the mothers. For example, institutional care hours may include
hours in kindergarten, preschool, or extra-curricular institutes. While these different types of
institutional care may be context-specific, Appendix 5 provides the exact variables used, as well
as their closest translation
The S2 sample is also used in this instance to the present dynamics of mothers’ investment
decisions as well as the mothers’ labor market as observed in Figure 1.4. The mother’s time
exhibits a significant cross-sectional variation, but does not display an obvious pattern over time.
The time spent in childcare institutions rapidly increases over time, with a tight 95% confidence
interval. Private surrogate care tends to mildly decrease over time, but the pattern is not as obvious
as with institutional care. Material expenditures tend to sharply increase in Wave 5, the period in
which the subject children reached ages 4 to 5. The mother’s labor supply is also noteworthy, as
mothers tended to work more, but mostly on an extensive rather than intensive margin.
As some further assumptions must be imposed on the structural model, it is nearly impossible
for simulation results to exactly replicate main data observations. However, it is still beneficial to
discuss how much of these are replicated; thus, Section 1.8 will discuss the counterpart to the
presented graph, as well as an analysis of the model’s goodness of fit.

19
1.3. Theoretical Model: Unidimensional
This work’s motivational analyses examine mothers’ behavior and the corresponding
production function only by considering the unidimensional ability at a single point in time.
This section explains the theoretical model as a part of an additional empirical strategy used in
motivational exercises.
2
This will clarify that a unique solution does exist, while the following
subsection will derive a closed-form solution for further empirical works.
The mothers’ dynamic optimization problem takes the following form to maximize the infinite
horizon utility:
3

max
(ℎ
𝑡𝑡 , 𝑗𝑗 𝑡𝑡 , 𝜏𝜏 𝑡𝑡 )
𝑡𝑡 = 0
∞

� 𝛽𝛽 𝑡𝑡 𝑢𝑢 ( 𝑙𝑙 𝑡𝑡 , 𝑐𝑐 𝑡𝑡 , 𝐴𝐴 𝑡𝑡 )
∞
𝑡𝑡 = 0

s. t. 𝑐𝑐 𝑡𝑡 + 𝑝𝑝 𝑗𝑗 𝑡𝑡 = 𝑤𝑤 𝑡𝑡 ℎ
𝑡𝑡 + 𝐼𝐼 𝑡𝑡 (1.2)
TT = ℎ
𝑡𝑡 + 𝑙𝑙 𝑡𝑡 + 𝜏𝜏 𝑡𝑡
ln 𝐴𝐴 𝑡𝑡 + 1
= 𝛿𝛿 1 𝑡𝑡 ln 𝜏𝜏 𝑡𝑡 + 𝛿𝛿 2 𝑡𝑡 ln 𝑗𝑗 𝑡𝑡 + 𝛿𝛿 3 𝑡𝑡 ln 𝐼𝐼 𝑡𝑡 + 𝛿𝛿 4 𝑡𝑡 ln 𝐴𝐴 𝑡𝑡 𝑓𝑓𝑓𝑓 𝑓𝑓 ∀ 𝑡𝑡 < 𝑇𝑇 + 1
ln 𝐴𝐴 𝑡𝑡 + 1
= ln 𝐴𝐴 𝑡𝑡 𝑓𝑓𝑓𝑓 𝑓𝑓 ∀ 𝑡𝑡 ≥ 𝑇𝑇 + 1

A mother values leisure 𝑙𝑙 𝑡𝑡 , household consumption 𝑐𝑐 𝑡𝑡 , and her child’s ability 𝐴𝐴 𝑡𝑡 , in each period.
The first equation calculates the mother’s budget constraints; 𝑤𝑤 𝑡𝑡 denotes the wages offered to
mothers, which are considered exogenous in this model. Equation (1.6) displays its structure.
Further, 𝐼𝐼 𝑡𝑡 indicates the household income aside from the mother’s labor income, which this study
also considers exogenous. While the model includes no financial market, disposable income is

2
The theoretical model and corresponding empirical strategy in this motivational section is derived from Brilli’s
(2013) work.
3
Only the mother’s decision is considered because the data set does not fully observe fathers’ decisions. For
example, fathers’ time spent with children is only included in the questionnaire in Wave 5.

20
allocated between consumption and childcare services 𝑗𝑗 𝑡𝑡 , with a corresponding hourly price 𝑝𝑝 that
may potentially include not only the monetary costs to be paid, but also any other implicit costs,
such as transportation costs. The second equation reveals the mother’s time constraints. The total
hours TT is displayed in a 24-hour format to reveal the mother allocates her time between labor
ℎ
𝑡𝑡 , leisure 𝑙𝑙 𝑡𝑡 , and time spent for maternal care 𝜏𝜏 𝑡𝑡 .
The third constraint equation—which is of particular interest in this paper—corresponds to
Equation (1.1) in the Introduction. Conditional on its own lag term, children’s ability is developed
in a log-linear fashion by three types of investments: the mother’s care ( 𝜏𝜏 𝑡𝑡 ), market care ( 𝑗𝑗 𝑡𝑡 ), and
other material expenditures, which are assumed as proportional to household income ( 𝐼𝐼 𝑡𝑡 ).
Please note that this paper’s ultimate interest can be capitulated by the last two equations in the
optimization problem (2), which demonstrates the evolution of the child’s ability depending on the
mother’s investment decisions. Further, note that the 𝛿𝛿 𝑡𝑡 parameters of the ability production
function, which are assumed to take a Cobb-Douglas form (Attanasio et al., 2015), are time-
varying. This reflects the time-varying nature of the efficiency of investments depending on a
child’s age. For example, cognitive development may significantly depend on investments made
in earlier childhood, while non-cognitive abilities may be intensively developed in the child’s
teenage years. However, allowing for such a time-varying constraint hinders the problem to be
standard recursive dynamic model, without further manipulation.
The terminal period of childhood development T is set to assure the existence of a unique
solution to this problem. Children have a window of opportunity until period T, but mothers are
committed to this final level of ability for an infinite horizon, with a time depreciation of 𝛽𝛽 = 0.99.
Thus, our dynamic problem after T is a stationary problem from which the optimal investment can
be recursively solved.

21
For simplicity, our solution alleviates the computational burden of structural estimation, and the
functional form of instantaneous utility is imposed to follow a Cobb-Douglas form, as follows:

𝑢𝑢 ( 𝑙𝑙 𝑡𝑡 , 𝑐𝑐 𝑡𝑡 , 𝐴𝐴 𝑡𝑡 ) = 𝛼𝛼 1
𝑙𝑙𝑙𝑙 𝑙𝑙 𝑡𝑡 + 𝛼𝛼 2
𝑙𝑙 𝑙𝑙 𝑐𝑐 𝑡𝑡 + 𝛼𝛼 3
𝑙𝑙𝑙𝑙 𝐴𝐴 𝑡𝑡 (1.3)

With the factors mentioned thus far, the theoretical solution to illustrate the mother’s dynamic
choice takes a simple, closed-form solution that makes the structural estimation tractable by
avoiding any numerical optimization:

ℎ
𝑡𝑡 =
𝑇𝑇𝑇𝑇 (𝛼𝛼 2
+ 𝛽𝛽 𝛿𝛿 2 𝑡𝑡 𝐷𝐷 𝑡𝑡 + 1
)
𝛼𝛼 1
+ 𝛽𝛽 𝛿𝛿 1 𝑡𝑡 𝐷𝐷 𝑡𝑡 + 1
+ 𝛼𝛼 2
+ 𝛽𝛽 𝛿𝛿 2 𝑡𝑡 𝐷𝐷 𝑡𝑡 + 1
−
𝐼𝐼 𝑡𝑡 ( 𝛼𝛼 1
+ 𝛽𝛽 𝛿𝛿 1 𝑡𝑡 𝐷𝐷 𝑡𝑡 + 1
)
𝑤𝑤 𝑡𝑡 ( 𝛼𝛼 1
+ 𝛽𝛽 𝛿𝛿 1 𝑡𝑡 𝐷𝐷 𝑡𝑡 + 1
+ 𝛼𝛼 2
+ 𝛽𝛽 𝛿𝛿 2 𝑡𝑡 𝐷𝐷 𝑡𝑡 + 1
)
(1.4)
ℎ
∗
𝑡𝑡 = �
ℎ
𝑡𝑡 𝑖𝑖 𝑓𝑓 ℎ
𝑡𝑡 ≥ 0
0 𝑖𝑖 𝑓𝑓 ℎ
𝑡𝑡 ≤ 0
(1.5)

In addition to the mother’s investment decision, the solution is conditional on her labor market
decision:
𝑗𝑗 𝑐𝑐 𝑡𝑡 =
𝛽𝛽 𝛿𝛿 2 𝑡𝑡 𝐷𝐷 𝑡𝑡 + 1
𝑝𝑝 ( 𝛼𝛼 2
+ 𝛽𝛽 𝛿𝛿 2 𝑡𝑡 𝐷𝐷 𝑡𝑡 + 1
)
( 𝑤𝑤 𝑡𝑡 ℎ
∗
𝑡𝑡 + 𝐼𝐼 𝑡𝑡 ) (1.6)
𝜏𝜏 𝑐𝑐 𝑡𝑡 =
𝛽𝛽 𝛿𝛿 1 𝑡𝑡 𝐷𝐷 𝑡𝑡 + 1
𝛼𝛼 2
+ 𝛽𝛽 𝛿𝛿 1 𝑡𝑡 𝐷𝐷 𝑡𝑡 + 1
(𝑇𝑇𝑇𝑇 − ℎ
∗
𝑡𝑡 )
where 𝐷𝐷 𝑡𝑡 takes the following recursive form:
𝐷𝐷 𝑇𝑇 + 1
= 𝜌𝜌 𝛼𝛼 3
(1.7)
𝐷𝐷 𝑡𝑡 = 𝛼𝛼 3
+ 𝛽𝛽 𝛿𝛿 4 𝑡𝑡 𝐷𝐷 𝑡𝑡 + 1
for ∀t ∈ (1, T)

22
The intuition of this solution is clear; 𝐷𝐷 𝑡𝑡 captures the extent to which mothers’ value of the
continuation of childhood ability impacts the margin. With this value snowballed down to earlier
stage of childhood development, forward-looking mothers’ choice to care for their children is
proportional to their disposable income, and these mother-care hours are subject to their time
availability. This is again a result of the specific, assumed functional form. Appendix 7 presents
the details regarding the derivation of this solution.

1.4. Empirical Strategy and Estimation Method
1.4.1. Further specifications details
This subsection explains the estimations’ further specifications, some of which may seem
extreme. However, such assumptions are a result of a compromise to relive computation burden,
or data unavailability.

1.4.1.1. Household income
Household income 𝐼𝐼 𝑡𝑡 is assumed to follow a completely exogenous process:
𝐼𝐼 𝑡𝑡 ~ 𝑖𝑖 . 𝑖𝑖 . 𝑑𝑑 . 𝑁𝑁 ( 𝜇𝜇 𝑖𝑖 𝑖𝑖 𝑐𝑐 , 𝜎𝜎 𝑖𝑖 𝑖𝑖 𝑐𝑐 2
) (1.8)
Note that this process is independent of any other variables of interest; thus, the parameters 𝜇𝜇 𝑖𝑖 𝑖𝑖 𝑐𝑐
and 𝜎𝜎 𝑖𝑖 𝑖𝑖 𝑐𝑐 can be estimated separately in the first stage.
1.4.1.2. Wage offered
As previously mentioned, the mother’s wage offer is assumed to be exogenous; therefore, this
study controls for the mother’s characteristics and her unobservable ability 𝜇𝜇 0
:

𝑙𝑙𝑙𝑙 𝑤𝑤 𝑡𝑡 = 𝜇𝜇 0
+ 𝜇𝜇 1
𝑚𝑚 𝑓𝑓 𝑚𝑚 _ 𝑒𝑒 𝑑𝑑𝑢𝑢 + 𝜇𝜇 2
𝑎𝑎𝑎𝑎 𝑒𝑒 𝑡𝑡 + 𝜇𝜇 3
𝑎𝑎𝑎𝑎 𝑒𝑒 𝑡𝑡 2
+ 𝜇𝜇 4
𝑢𝑢 𝑓𝑓 𝑢𝑢 𝑎𝑎 𝑙𝑙 + 𝜀𝜀 𝑡𝑡 (1.9)

23
𝜀𝜀 𝑡𝑡 ~ 𝑖𝑖 . 𝑖𝑖 . 𝑑𝑑 . 𝑁𝑁 (0, 𝜎𝜎 𝜀𝜀 2
)

The mother’s education variable 𝑚𝑚 𝑓𝑓 𝑚𝑚 _ 𝑒𝑒 𝑑𝑑𝑢𝑢 is imposed as a continuous variable. Two variables
for the mother’s age— 𝑎𝑎𝑎𝑎 𝑒𝑒 𝑡𝑡 and 𝑎𝑎𝑎𝑎 𝑒𝑒 𝑡𝑡 2
—also continuously enter the equation. While mothers’ full
work history is unavailable in the data, age variables are used as a proxy for experience. A
residential status variable 𝑢𝑢 𝑓𝑓 𝑢𝑢 𝑎𝑎 𝑙𝑙 takes a value of one if the family resides in a “big” city and zero
if they are from either a “small” city or rural area. In addition to the variations in observable
characteristics, unobserved maternal characteristics are discretely allowed in the specifications;
specifically, the mother’s ability is noted as either high or low. The probability of the mother
having high ability 𝜋𝜋 𝑚𝑚 ℎ
is specified as follows:

𝜋𝜋 𝑚𝑚 ℎ
=
exp (𝑧𝑧 𝑚𝑚 )
1 + exp (𝑧𝑧 𝑚𝑚 )
(1.10)

High-type mothers will exhibit 𝜇𝜇 0
= 𝜇𝜇 0 ℎ 𝑖𝑖 𝑖𝑖 ℎ
results, while low-type mothers are noted as 𝜇𝜇 0
=
𝜇𝜇 0 𝑙𝑙 𝑙𝑙 𝑙𝑙 , where these analyses will also estimate 𝜇𝜇 0 ℎ 𝑖𝑖 𝑖𝑖 ℎ
and 𝜇𝜇 0 𝑙𝑙 𝑙𝑙 𝑙𝑙 as well as 𝑧𝑧 𝑚𝑚 .
Figure 1.5 displays the distribution of the mother’s observed wages using the primary S2
sample, which will reveal the validity of such specifications. Figure 1.5 and other explorations of
the data confirm that the mother’s education plays the biggest role in offered wages in our specific
context. After further controlling for education, age still seems to capture wage variations to a
certain extent. The wage distribution by residential area, or our 𝑢𝑢 𝑓𝑓 𝑢𝑢 𝑎𝑎 𝑙𝑙 variable, indicates a
compelling pattern. While some discrepancy exists in the intensive margin, the last panel in Figure
1.5 suggests that a difference might exist in the extensive margin, which implies the work decisions

24
of mothers in urban regions might be heterogeneous from their rural counterparts. This study
reflects such a possibility by allowing the mother’s utility parameter to vary by residential region.
The heterogeneity of these preferences will be discussed later in this subsection.
1.4.1.3. Preference
Heterogeneity may exist in mothers’ preferences depending on the mother’s observed or
unobserved characteristics. Instead of directly estimating preference parameters 𝛼𝛼 , which will
tightly restrict homogeneous mothers, we define a set of hyper-parameters that govern mothers’
preferences. First, in restating the Cobb-Douglas utility function, we have
𝑢𝑢 ( 𝑙𝑙 𝑡𝑡 , 𝑐𝑐 𝑡𝑡 , 𝐴𝐴 𝑡𝑡 ) = 𝛼𝛼 1
𝑙𝑙𝑙𝑙 𝑙𝑙 𝑡𝑡 + 𝛼𝛼 2
𝑙𝑙𝑙𝑙 𝑐𝑐 𝑡𝑡 + 𝛼𝛼 3
𝑙𝑙𝑙𝑙 𝐴𝐴 𝑡𝑡 (1.11)

We restrict preference parameters 𝛼𝛼 to add up to one, and allow for the individual heterogeneity
among mothers by assuming that our preference parameters take the following form:

𝛼𝛼 1
=
exp (𝛾𝛾 1
)
exp( 𝛾𝛾 1
) + exp( 𝑋𝑋′ 𝛤𝛤 2
) + exp (𝑋𝑋′ 𝛤𝛤 3
+ 𝐼𝐼 (𝜇𝜇 0
= 𝜇𝜇 0 ℎ 𝑖𝑖 𝑖𝑖 ℎ
))

𝛼𝛼 2
=
exp( 𝑋𝑋′ 𝛤𝛤 2
)
exp( 𝛾𝛾 1
) + exp( 𝑋𝑋′ 𝛤𝛤 2
) + exp (𝑋𝑋′ 𝛤𝛤 3
+ 𝐼𝐼 (𝜇𝜇 0
= 𝜇𝜇 0 ℎ 𝑖𝑖 𝑖𝑖 ℎ
))
(1.12)
𝛼𝛼 3
=
exp (𝑋𝑋′ 𝛤𝛤 3
+ 𝐼𝐼 (𝜇𝜇 0
= 𝜇𝜇 0 ℎ 𝑖𝑖 𝑖𝑖 ℎ
))
exp( 𝛾𝛾 1
) + exp( 𝑋𝑋′ 𝛤𝛤 2
) + exp (𝑋𝑋′ 𝛤𝛤 3
+ 𝐼𝐼 (𝜇𝜇 0
= 𝜇𝜇 0 ℎ 𝑖𝑖 𝑖𝑖 ℎ
))

where 𝐼𝐼 (𝜇𝜇 0
= 𝜇𝜇 0 ℎ 𝑖𝑖 𝑖𝑖 ℎ
) is an indicator function of mothers’ unobservable types, as discussed with
Equation (1.6). Observable characteristics 𝑋𝑋 involve mom_edu to indicate the mother’s education
and urban, a variable to flag the residential area. Therefore, 𝛤𝛤 2
and 𝛤𝛤 3
are vectors of ( 𝛾𝛾 2
, 𝛾𝛾 3
) and
( 𝛾𝛾 4
, 𝛾𝛾 5
), respectively. For identification purposes, 𝛾𝛾 1
is normalized to zero in this estimation.

25

1.4.1.4. Child’s ability
In restating the log-linear type ability-development process presented in Equation (1.2),
ln 𝐴𝐴 𝑡𝑡 + 1
= 𝛿𝛿 1 𝑡𝑡 ln 𝜏𝜏 𝑡𝑡 + 𝛿𝛿 2 𝑡𝑡 ln 𝑗𝑗 𝑡𝑡 + 𝛿𝛿 3 𝑡𝑡 ln 𝐼𝐼 𝑡𝑡 + 𝛿𝛿 4 𝑡𝑡 ln 𝐴𝐴 𝑡𝑡 𝑓𝑓𝑓𝑓 𝑓𝑓 ∀ 𝑡𝑡 < 𝑇𝑇 + 1 (1.13)

This study’s motivational analyses involve four separate regressions, assuming mothers only
care about one ability dimension in each iteration. Thus, 𝐴𝐴 𝑡𝑡 in Equation (1.2) as well as the
productivity parameter 𝛿𝛿 𝑖𝑖 𝑡𝑡 will take scalar forms in this benchmark exercise.
This paper further presents two assumptions in its specifications: 1) The time-varying
productivity of 𝛿𝛿 𝑖𝑖 𝑡𝑡 i ∈ {1, 2, 3, 4} over time t, and 2) the initial ability 𝐴𝐴 1
for the autoregressive
process.
First, we specifically assume the following time-varying productivity:

𝛿𝛿 𝑖𝑖 𝑡𝑡 = exp( 𝜉𝜉 𝑖𝑖 𝑡𝑡 ) (1.14)

Such a simple formation of ability-developing technology assumes each investment
monotonically increases or decreases the contribution to abilities’ evolution. As previously
discussed in Section 1.1.1, it is well-known that different dimensions of abilities are developed at
different times in life. For example, such cognitive abilities as language and communication
skills—which are skills of focus in this paper—may develop more extensively before age 10. In
contrast, social skills—another skill dimension that this paper scrutinizes—can be more
extensively developed in a child’s teenage years. On the one hand, the estimated 𝜉𝜉 𝑖𝑖 when 𝐴𝐴 𝑡𝑡
denotes language and communication skills will be negative with a significant magnitude in such

26
an example, implying a window of opportunity that closes early. On the other hand, 𝜉𝜉 𝑖𝑖 may be
estimated as a large, positive number if a window of opportunity opens in later life. Therefore,
while Equation (1.14) seems restrictive, it does calculate a minimum regarding the nature of
childhood development.
This study allows the initial levels of childhood abilities to vary depending on the parents’
educational level and the child’s unobserved type:
𝐴𝐴 1
= exp( ψ
𝑐𝑐 + 𝜂𝜂 1
𝑚𝑚 𝑓𝑓 𝑚𝑚 _ 𝑒𝑒 𝑑𝑑𝑢𝑢 + 𝜂𝜂 2
𝑝𝑝 𝑓𝑓 𝑝𝑝 _ 𝑒𝑒 𝑑𝑑𝑢𝑢 ) (1.15)
where mom_edu and pop_edu are variables to indicate the parents’ years of education. Similar
to the mother’s unobserved ability and the child’s unobserved type, ψ
𝑐𝑐 is also assumed to take one
of two possible values � ψ
𝑐𝑐 ℎ 𝑖𝑖 𝑖𝑖 ℎ
, ψ
𝑐𝑐 𝑙𝑙 𝑙𝑙 𝑙𝑙 � . For ψ
𝑐𝑐 , the probability of having either type is assumed
to follow a Bernoulli distribution governed by parameters 𝑧𝑧 𝑐𝑐 . Thus, 𝜋𝜋 𝑐𝑐 ℎ
is defined as the
probability of being a high type 𝜋𝜋 𝑐𝑐 ℎ
, which is formalized by the following Equation (1.16):
𝜋𝜋 𝑐𝑐 ℎ
=
exp (𝑧𝑧 𝑐𝑐 )
1 + exp (𝑧𝑧 𝑐𝑐 )
(1.16)
1.4.1.5. Ability Scores
Once a child’s ability is determined, observable scores are measured with specific measurement
errors. The following specification is assumed for language and communication skills:
𝑐𝑐𝑓𝑓 𝑎𝑎 _ 𝑠𝑠 𝑐𝑐𝑓𝑓 𝑓𝑓𝑒𝑒 𝑡𝑡 = 𝐴𝐴 𝑡𝑡 + 𝜈𝜈 𝑡𝑡 (1.17)
𝜈𝜈 𝑡𝑡 ~i. i. d. 𝑁𝑁 (0, 𝜎𝜎 𝜈𝜈 2
)

As discussed in Section 1.2, the cognitive skill of language 𝑐𝑐𝑓𝑓 𝑎𝑎 _𝑠𝑠 𝑐𝑐𝑓𝑓 𝑓𝑓𝑒𝑒 𝑡𝑡 is standardized to a
distribution of the mean of five and standard deviation of one due to inconsistent score measures.
However, note that unobserved true abilities are unit-free measures, and this paper does not intend

27
to interpret their magnitude other than any changes in percentage. Therefore, only a relative
ranking is used to identify our ultimate parameter of interest 𝛿𝛿 s. This identification is not limited
by the use of standardization, which does not respond to an affine transformation because we do
not interpret the change in magnitude for unobserved abilities.
Alternatively, the non-cognitive skill score is calculated as
𝑙𝑙𝑓𝑓 𝑙𝑙 _ 𝑐𝑐𝑓𝑓 𝑎𝑎 _𝑠𝑠 𝑐𝑐𝑓𝑓 𝑓𝑓 𝑒𝑒 𝑡𝑡 = 1 + 4 ×
exp (𝐴𝐴 𝑡𝑡 + 𝜈𝜈 𝑡𝑡 )
1 + exp (𝐴𝐴 𝑡𝑡 + 𝜈𝜈 𝑡𝑡 )
(1.18)
𝜈𝜈 𝑡𝑡 ~i. i. d. 𝑁𝑁 (0, 𝜎𝜎 𝜈𝜈 2
)
This is scaled between one and five to correspond with the original questionnaire.
In summary, this study’s empirical specifications assume the mother’s choice depends on a
predetermined variable that is either observed or unobserved to the researcher. Once the mother’s
choice of her optimal time use is determined, her child’s ability can consequently be determined.

1.4.2. Estimation Method
The estimation method used in this paper is the simulated minimum distance (SMD), which
minimizes the distance between the sample moments and simulated moments. Table 1.3
summarizes the moment conditions used to construct the distance.
The estimation is conducted in two steps. Due to the i.i.d. assumption of household income, it
is possible to estimate ( 𝜇𝜇 𝑖𝑖 𝑖𝑖 𝑐𝑐 , 𝜎𝜎 𝑖𝑖 𝑖𝑖 𝑐𝑐 2
) in the first step. Recursively, the rest of the parameters are
estimated in the second step using simulated data, which includes simulated household income
using the estimated � 𝜇𝜇 ̂ 𝑖𝑖 𝑖𝑖 𝑐𝑐 , 𝜎𝜎 �
𝑖𝑖 𝑖𝑖𝑐𝑐
2
� from the first step. The parameters are recapitulated by denoting
θ = � 𝛾𝛾 2
, 𝛾𝛾 3
, 𝛾𝛾 4
, 𝛾𝛾 5
, 𝜉𝜉 1
, 𝜉𝜉 2
, 𝜉𝜉 3
, 𝜉𝜉 4
, 𝜎𝜎 𝜈𝜈 , 𝜇𝜇 0 ℎ 𝑖𝑖 𝑖𝑖 ℎ
, 𝜇𝜇 0 𝑙𝑙 𝑙𝑙 𝑙𝑙 , 𝑧𝑧 𝑚𝑚 ℎ
, 𝜇𝜇 1
, 𝜇𝜇 2
, 𝜇𝜇 3
, 𝜇𝜇 4
, , ψ
𝑐𝑐 ℎ
, ψ
𝑐𝑐 𝑙𝑙 , 𝑧𝑧 𝑐𝑐 ℎ
,
𝜎𝜎 𝜀𝜀 , 𝜂𝜂 1
, 𝜂𝜂 2
, 𝑝𝑝 , 𝜇𝜇 𝑖𝑖 𝑖𝑖 𝑐𝑐 , 𝜎𝜎 𝑖𝑖 𝑖𝑖 𝑐𝑐 � as the full set of 23 parameters to be estimated. Additionally, the vector

28
of moments specified in Table 1.3 are denoted as M(θ), with M
�
(θ) specifically used for that
calculated from the simulated data under θ. The corresponding moment from the sample, which
presumably is derived from population moments M( θ), is denoted as 𝑚𝑚 � .
To outline the simulation process, the mothers’ and fathers’ observed predetermined
characteristics mom_edu, mom_age, and pop_edu are used, as defined in the previous section. The
sample size is denoted in each period N, and the number of simulations is set at R = 100; the N × R
mothers’ fixed type is randomly drawn from a binary distribution. I then set β = 0.99, which is
higher than in Brilli’s (2013) work, as the gap between each wave is only one year. After setting
the terminal child development period T to 20, as explained in Section 1.3, the preference
parameters 𝛼𝛼 𝑖𝑖 s are constructed according to Equation (1.9) and the 𝐷𝐷 𝑖𝑖 𝑡𝑡 s construct. By randomly
drawing the N × R × W realization of 𝜀𝜀 𝑡𝑡 s, where W hereafter is used to indicate the total number
of waves in our data, we can simulate the wage offer. Once wages are simulated, the mother’s
decision can be simulated using the closed form solutions from Equations (1.3), (1.4), and (1.5).
As all the input variables are generated in the simulated data set, simulating the N × R children’s
initial ability and N × R × W score measurement error can simulate the corresponding scores for
each child.
The simulated data enables the construction of simulated moments M
�
( θ), and the distance
between the original sample and simulated data is defined as
𝑎𝑎 � ( θ) = 𝑚𝑚 � − M
�
( θ) (1.19)
Thus, the SMD estimator θ
�
can be defined as
θ
�
= 𝑎𝑎 𝑓𝑓 𝑎𝑎 𝑚𝑚 𝑖𝑖𝑙𝑙 𝑎𝑎 � ( θ)
′
𝑊𝑊 𝑎𝑎 � ( θ) (1.20)

29
Note that this is an over-identified model with 23 parameters and 92 moment conditions;
theoretically, any positive, definite matrix would guarantee a consistent estimator. However, the
results are kept comparable with Brilli’s (2013) work by choosing a diagonal weighted matrix W
as presented in Equation (1.21) to construct the optimal minimum distance (OMD) estimator,
4
in
which each 𝑉𝑉 𝑎𝑎𝑓𝑓 (𝑚𝑚 �
𝚥𝚥 )
�
is calculated using a bootstrap variance from the S1 sample, with 500
repetitions:
𝑊𝑊 = �
𝑉𝑉 𝑎𝑎𝑓𝑓 (𝑚𝑚 �
1
)
�
⋯ 0
⋮ ⋱ ⋮
0 ⋯ 𝑉𝑉 𝑎𝑎𝑓𝑓 (𝑚𝑚 �
92
)
�
�
− 1
(1.21)

Another challenge in evaluating the estimates involves verifying the estimation’s standard
errors. Due to the structural model’s complexity, a closed-form estimate for standard errors is not
possible. Thus, we provide a bootstrap standard error, with 30 bootstrap samples.
Technically, Matlab’s fminsearch command for the minimization algorithm for Equation (1.16),
as this uses a Nelder-Mead algorithm. Although this algorithm is slower in its computation, it is
more robust in cases in which the objective function is not well-behaved. As we use a simulated
method in which the objective function depends on the specific realization of a randomly generated
number, using the Nelder-Mead algorithm was a better option compared to other such alternatives
as the Newton-Rapson method.

4
Cameron, C. A., & Trivedi, P. K. (2005). Microeconometrics: Methods and applications. Cambridge, MA:
Cambridge University Press, p. 205.

30
1.5. Unidimensional Analysis Results
This section presents the estimation results from the study’s motivational exercise. Please note
that each column’s results are derived from a separate regression, assuming that mothers only
appreciate one specific dimension of an ability at a time. Therefore, the motivational results
presented in this section will potentially experience misspecifications; subsequently, these
estimates must be interpreted with prudence.
While the estimates themselves have limitations, the estimates are still noteworthy for two
reasons. First, we use data from a context that has not been similarly analyzed, and thus, we must
confirm this study’s position in existing literature. Therefore, these estimates are juxtaposed with
those from Brilli’s (2013) work, as the latter provides results most comparable with those in the
current study. Second, this paper primarily aims to extend a unidimensional framework to one that
is multi-dimensional; consequently, this section’s results will serve as a benchmark for our primary
analyses as presented in Section 1.8.
Table 1.5 presents the set of estimates for preference-related parameters as well as average
mothers’ imputed 𝛼𝛼 . As anticipated, the estimates vary across columns, which implies each of the
specifications might be incomplete.
The first column in Table 1.5 concentrates only on the estimates for the language and
communication dimensions—or specifically, 𝐴𝐴 𝑡𝑡 = 𝐿𝐿 𝑡𝑡 , where 𝐿𝐿 𝑡𝑡 represents language. This
demonstrates that the average mother in our sample will face an optimization problem from the
following utility and budget constraints:
max
(ℎ
𝑡𝑡 , 𝑖𝑖 𝑡𝑡 , 𝜏𝜏 𝑡𝑡 )
𝑡𝑡 = 0
∞

� 𝛽𝛽 𝑡𝑡 (0.254 𝑙𝑙𝑙𝑙 𝑙𝑙 𝑡𝑡 + 0.486 𝑙𝑙𝑙𝑙 𝑐𝑐 𝑡𝑡 + 0.259 𝑙𝑙𝑙𝑙 𝐿𝐿 𝑡𝑡 )
∞
𝑡𝑡 = 0

s. t. 𝑐𝑐 𝑡𝑡 + 0.051 𝑗𝑗 𝑡𝑡 = 𝑤𝑤 𝑡𝑡 ℎ
𝑡𝑡 + 10.6811

31

This substantially differs from the average mother’s results in Brilli’s (2013) work:

max
(ℎ
𝑡𝑡 , 𝑖𝑖 𝑡𝑡 , 𝜏𝜏 𝑡𝑡 )
𝑡𝑡 = 0
∞

� 𝛽𝛽 𝑡𝑡 (0.413 𝑙𝑙𝑙𝑙 𝑙𝑙 𝑡𝑡 + 0.022 𝑙𝑙𝑙𝑙 𝑐𝑐 𝑡𝑡 + 0.564 𝑙𝑙𝑙𝑙 𝐿𝐿 𝑡𝑡 )
∞
𝑡𝑡 = 0

s. t. 𝑐𝑐 𝑡𝑡 + 5.15 𝑗𝑗 𝑡𝑡 = 𝑤𝑤 𝑡𝑡 ℎ
𝑡𝑡 + 𝐼𝐼 𝑡𝑡

Such a discrepancy may be due to differences in the socio-economic context, but it might also
be impacted by differences in the sample’s structure, as Brilli (2013) utilizes a broader range of
child ages, between 2 to 12 years. For example, the nature of market care might differ not only
because of the social context but also the stage of childhood development.
A certain discrepancy is also found in other estimates. For example, Table 1.5 displays the return
to school, which appears lower for Korean mothers in the early 2010s compared to US mothers in
late 1990s, although this result is not unexpected. For the reader’s convenience, this paper observes
the regression on language dimension to find our average mother experiences the following wage
offer process:

𝑙𝑙𝑙𝑙 𝑤𝑤 𝑡𝑡 = 3.651 + 0.037 𝑚𝑚 𝑓𝑓 𝑚𝑚 _ 𝑒𝑒 𝑑𝑑 𝑢𝑢 − 1.646 𝑎𝑎𝑎𝑎𝑒𝑒 𝑡𝑡 + 0.032 𝑎𝑎𝑎𝑎𝑒𝑒 𝑡𝑡 2
− 0.594 𝑢𝑢𝑓𝑓 𝑢𝑢 𝑎𝑎𝑙𝑙 + 𝜀𝜀 𝑡𝑡

while Brilli (2013) finds the average mother’s process as follows:

𝑙𝑙𝑙𝑙 𝑤𝑤 𝑡𝑡 = 1.8819 + 0.0788 𝑚𝑚 𝑓𝑓 𝑚𝑚 _ 𝑒𝑒 𝑑𝑑𝑢𝑢 + 0.0018 𝑎𝑎𝑎𝑎 𝑒𝑒 𝑡𝑡 − 0.0006 𝑎𝑎𝑎𝑎 𝑒𝑒 𝑡𝑡 2
− 0.0156 𝑢𝑢 𝑙𝑙𝑎𝑎𝑐𝑐 𝑏𝑏 + 𝜀𝜀 𝑡𝑡

32
Now, we turn to the ultimate interest of this paper: the formation of ability. Table 1.7 displays
the estimates for the initial ability parameters, which are included to allow for the heterogeneity in
initial values. However, nearly all of the estimates are not statistically significant or very intuitive,
and such a pattern is also found in Brilli’s (2013) work, in which effectively no unobservable
variations are observed, and the coefficients are insignificant.
Table 1.8 presents the dynamic evolution of children’s ability as well as the estimates of the
hyper-parameter
𝜉𝜉 𝑖𝑖 s and the imputed productivity parameter 𝛿𝛿 𝑖𝑖 𝑡𝑡 s. These reveal quite compelling patterns. First,
household income as a proxy for material investment seems to play the biggest role in development
for all dimensions, which agrees with Brilli’s (2013) results in a US context.
More importantly, a comparison of mothers’ care and market care demonstrates that market care
seems to play a more substantial role in all dimensions, which again agrees with Brilli’s (2013)
conclusion. This is especially the case when observing that the Table 1.7 estimates were performed
simultaneously with results presented in Tables 1.4 to 1.6. As these tables present noticeably larger
differences depending on the context, an agreement among deep parameters was unexpected.
Further, a comparison between this work’s non-cognition results and Brilli’s (2013) estimates
reveals some differences in the magnitude of productivity, while the estimates for language ability
present even stronger agreement. The current work’s estimates regarding the language ability
indicate the following imputed ability production between Waves 3 and 4:
ln 𝐴𝐴 𝑡𝑡 + 1
= 0.265 ln 𝜏𝜏 𝑡𝑡 + 0.523 ln 𝑖𝑖 𝑡𝑡 + 1.003 ln 𝐼𝐼 𝑡𝑡 + 0.664 ln 𝐴𝐴 𝑡𝑡 (1.22)
while Brilli (2013) notes the following imputed production for children of the same age:

ln 𝐴𝐴 𝑡𝑡 + 1
= 0.2681 ln 𝜏𝜏 𝑡𝑡 + 0.3485 ln 𝑖𝑖 𝑡𝑡 + 0.8438 ln 𝐼𝐼 𝑡𝑡 + 0.7351 ln 𝐴𝐴 𝑡𝑡

33
It would be injudicious to draw any strong conclusions based on this motivational exercise, as
the results could occur due to potentially incorrect specifications. However, the previously
discussed results assure that our result aligns with existing literature to some extent; this also
requires further analyses that fully incorporate a multi-dimensional ability.
Figure 1.6 may provide additional insights on our results thus far before proceeding to this
work’s multi-dimensional analyses. Noting that the log linear specification for ability production
is not homogenous for degrees of one, the relative motivations of mothers’ care and market care
were compared by imputed 𝛿𝛿 𝑖𝑖 𝑡𝑡 s normalized to one in each period; these were plotted up to age 15.
The L, E, S, and A in the label each represent the relative magnitude of 𝛿𝛿 𝑖𝑖 𝑡𝑡 from the regressions
based on language, emotionality, sociability, and activeness, respectively. Figure 1.6 implies that
while the motivation still exists for market care for language development, the dominating motivation for market
care might still be the development of sociability. Alternatively, the mother’s time may contribute to the
development of social skills, but a greater motivation may exist for the mothers to spend time with their children
to promote their language ability. However, this observation across separate regression results may not provide
a fair comparison, and thus, we will depart from the motivational analyses and extend our interest to a multi-
dimensional case.

1.6. Theoretical Model: Multidimensional

Sections 1.6 and 1.7 will explain this work’s extension to multi-dimensional model and
corresponding empirical strategy. As most of the components from the motivational model will be
preserved, the following two sections will emphasize the changes made rather than repeating what
has been discussed in previous sections.
In the fully extended model, mothers now face the following dynamic optimization problem:

34
𝑚𝑚 𝑎𝑎 𝑚𝑚 (ℎ
𝑡𝑡 , 𝑖𝑖 𝑡𝑡 , 𝜏𝜏 𝑡𝑡 , 𝑖𝑖 𝑡𝑡 , 𝑚𝑚 𝑡𝑡 )
𝑡𝑡 = 0
∞

� 𝛽𝛽 𝑡𝑡 (𝛼𝛼 1
𝑙𝑙𝑙𝑙 𝑙𝑙 𝑡𝑡 + 𝛼𝛼 2
𝑙𝑙𝑙𝑙 𝑐𝑐 𝑡𝑡 + 𝛼𝛼 3
𝑙𝑙𝑙𝑙 𝐿𝐿 𝑡𝑡 + 𝛼𝛼 4
𝑙𝑙𝑙𝑙 𝐸𝐸 𝑡𝑡 + 𝛼𝛼 5
𝑙𝑙𝑙𝑙 𝑆𝑆 𝑡𝑡 + 𝛼𝛼 6
𝑙𝑙𝑙𝑙 𝐴𝐴 𝑡𝑡 )
∞
𝑡𝑡 = 0

𝑠𝑠 . 𝑡𝑡 . 𝑐𝑐 𝑡𝑡 + 𝑝𝑝 𝑖𝑖 𝑖𝑖 𝑡𝑡 + 𝑝𝑝 𝑖𝑖 𝑙𝑙 𝑡𝑡 + 𝑚𝑚 𝑡𝑡 = 𝑤𝑤 𝑡𝑡 ℎ
𝑡𝑡 + 𝐼𝐼 𝑡𝑡 (1.23)
𝑇𝑇𝑇𝑇 = ℎ
𝑡𝑡 + 𝑙𝑙 𝑡𝑡 + 𝜏𝜏 𝑡𝑡
�
𝑙𝑙𝑙𝑙 𝐿𝐿 𝑡𝑡 + 1
𝑙𝑙𝑙𝑙 𝐸𝐸 𝑡𝑡 + 1
𝑙𝑙𝑙𝑙 𝑆𝑆 𝑡𝑡 + 1
𝑙𝑙𝑙𝑙 𝐴𝐴 𝑡𝑡 + 1
� =
⎣
⎢
⎢
⎢
⎡
𝛿𝛿 1 𝑡𝑡 𝐿𝐿 𝛿𝛿 1 𝑡𝑡 𝐸𝐸 𝛿𝛿 1 𝑡𝑡 𝑆𝑆 𝛿𝛿 1 𝑡𝑡 𝐴𝐴 ⎦
⎥
⎥
⎥
⎤
𝑙𝑙𝑙𝑙 𝜏𝜏 𝑡𝑡 +
⎣
⎢
⎢
⎢
⎡
𝛿𝛿 2 𝑡𝑡 𝐿𝐿 𝛿𝛿 2 𝑡𝑡 𝐸𝐸 𝛿𝛿 2 𝑡𝑡 𝑆𝑆 𝛿𝛿 2 𝑡𝑡 𝐴𝐴 ⎦
⎥
⎥
⎥
⎤
𝑙𝑙𝑙𝑙 𝑖𝑖 𝑡𝑡 +
⎣
⎢
⎢
⎢
⎡
𝛿𝛿 3 𝑡𝑡 𝐿𝐿 𝛿𝛿 3 𝑡𝑡 𝐸𝐸 𝛿𝛿 3 𝑡𝑡 𝑆𝑆 𝛿𝛿 3 𝑡𝑡 𝐴𝐴 ⎦
⎥
⎥
⎥
⎤
𝑙𝑙𝑙𝑙 𝑙𝑙 𝑡𝑡 +
⎣
⎢
⎢
⎢
⎡
𝛿𝛿 4 𝑡𝑡 𝐿𝐿 𝛿𝛿 4 𝑡𝑡 𝐸𝐸 𝛿𝛿 4 𝑡𝑡 𝑆𝑆 𝛿𝛿 4 𝑡𝑡 𝐴𝐴 ⎦
⎥
⎥
⎥
⎤
𝑙𝑙𝑙𝑙 𝑚𝑚 𝑡𝑡 +
⎣
⎢
⎢
⎢
⎡
𝛿𝛿 5 𝑡𝑡 𝐿𝐿 𝛿𝛿 6 𝑡𝑡 𝐿𝐿 𝛿𝛿 7 𝑡𝑡 𝐿𝐿 𝛿𝛿 8 𝑡𝑡 𝐿𝐿 𝛿𝛿 5 𝑡𝑡 𝐸𝐸 𝛿𝛿 6 𝑡𝑡 𝐸𝐸 𝛿𝛿 7 𝑡𝑡 𝐸𝐸 𝛿𝛿 8 𝑡𝑡 𝐸𝐸 𝛿𝛿 5 𝑡𝑡 𝑆𝑆 𝛿𝛿 6 𝑡𝑡 𝑆𝑆 𝛿𝛿 7 𝑡𝑡 𝑆𝑆 𝛿𝛿 8 𝑡𝑡 𝑆𝑆 𝛿𝛿 5 𝑡𝑡 𝐴𝐴 𝛿𝛿 6 𝑡𝑡 𝐴𝐴 𝛿𝛿 7 𝑡𝑡 𝐴𝐴 𝛿𝛿 8 𝑡𝑡 𝐴𝐴 ⎦
⎥
⎥
⎥
⎤
�
𝑙𝑙𝑙𝑙 𝐿𝐿 𝑡𝑡 𝑙𝑙𝑙𝑙 𝐸𝐸 𝑡𝑡 𝑙𝑙𝑙𝑙 𝑆𝑆 𝑡𝑡 𝑙𝑙𝑙𝑙 𝐴𝐴 𝑡𝑡 � 𝑓𝑓𝑓𝑓 𝑓𝑓 ∀ 𝑡𝑡 < 𝑇𝑇 + 1
𝑙𝑙𝑙𝑙 �
𝑙𝑙𝑙𝑙 𝐿𝐿 𝑡𝑡 + 1
𝑙𝑙𝑙𝑙 𝐸𝐸 𝑡𝑡 + 1
𝑙𝑙𝑙𝑙 𝑆𝑆 𝑡𝑡 + 1
𝑙𝑙𝑙𝑙 𝐴𝐴 𝑡𝑡 + 1
� = 𝑙𝑙𝑙𝑙 �
𝑙𝑙𝑙𝑙 𝐿𝐿 𝑡𝑡 𝑙𝑙𝑙𝑙 𝐸𝐸 𝑡𝑡 𝑙𝑙𝑙𝑙 𝑆𝑆 𝑡𝑡 𝑙𝑙𝑙𝑙 𝐴𝐴 𝑡𝑡 � 𝑓𝑓𝑓𝑓 𝑓𝑓 ∀ 𝑡𝑡 ≥ 𝑇𝑇 + 1

The current models indicate that mothers appreciate all four dimension of children’s abilities;
𝐿𝐿 𝑡𝑡 , 𝐸𝐸 𝑡𝑡 , 𝑆𝑆 𝑡𝑡 , and 𝐴𝐴 𝑡𝑡 respectively represent language and communications, emotionality, sociability,
and activity. Correspondingly, the ability production function does not take a vector form, allowing
for a dynamic complementarity across dimensions. While these off-diagonal terms are not this
paper’s primary interest, it will still be beneficial for general possibilities.
Additional changes have been made. First, the market care in the motivational model 𝑗𝑗 𝑡𝑡 has been
designated as either institutional care 𝑖𝑖 𝑡𝑡 , which includes kindergarten; or surrogate private care 𝑙𝑙 𝑡𝑡 ,
which includes baby-sitters. These two may substantially differ in terms of both the quality and
motivation of use. As Section 1.2 illustrated, a majority of the market care 𝑗𝑗 𝑡𝑡 consists of

35
institutional care 𝑖𝑖 𝑡𝑡 , and not distinguishing these two qualitatively different types of market care
might lead to false conclusions.
Second, the data directly reveals monetary expenditures for a child 𝑚𝑚 𝑡𝑡 . As we have restrictive
budget constraints that do not allow for a credit market, the productivity of monetary investments
might also be restrictive. However, it would still be better to control for this variable to avoid any
uncontrolled endogeneity correlated with household income.
As the model’s primary structure has not changed, we can still narrow down the analytical
solution with additional algebra, as discussed in Appendix 7 and presented in the following form:
ℎ
𝑡𝑡 =
𝑇𝑇𝑇𝑇 (𝛼𝛼 2
+ 𝐾𝐾 𝑡𝑡 + 1
+ 𝐺𝐺 𝑡𝑡 + 1
+ 𝐹𝐹 𝑡𝑡 + 1
)
𝛼𝛼 1
+ 𝛼𝛼 2
+ 𝐻𝐻 𝑡𝑡 + 1
+ 𝐾𝐾 𝑡𝑡 + 1
+ 𝐺𝐺 𝑡𝑡 + 1
+ 𝐹𝐹 𝑡𝑡 + 1
−
𝐼𝐼 𝑡𝑡 ( 𝛼𝛼 1
+ 𝐻𝐻 𝑡𝑡 + 1
)
𝑤𝑤 𝑡𝑡 ( 𝛼𝛼 1
+ 𝛼𝛼 2
+ 𝐻𝐻 𝑡𝑡 + 1
+ 𝐾𝐾 𝑡𝑡 + 1
+ 𝐺𝐺 𝑡𝑡 + 1
+ 𝐹𝐹 𝑡𝑡 + 1
)
(1.24)
ℎ
∗
𝑡𝑡 = �
ℎ
𝑡𝑡 𝑖𝑖 𝑓𝑓 ℎ
𝑡𝑡 ≥ 0
0 𝑖𝑖 𝑓𝑓 ℎ
𝑡𝑡 ≤ 0

Conditional on the mother’s optimal working hours ℎ
∗
𝑡𝑡 , the following optimal investment
decisions are made:
𝜏𝜏 𝑐𝑐 𝑡𝑡 =
𝐻𝐻 𝑡𝑡 + 1
𝛼𝛼 2
+ 𝐴𝐴 𝑡𝑡 + 1
( 𝑇𝑇𝑇𝑇 − ℎ
∗
𝑡𝑡 )
𝑖𝑖 𝑐𝑐 𝑡𝑡 =
𝐾𝐾 𝑡𝑡 + 1
𝑝𝑝 𝑖𝑖 ( 𝛼𝛼 2
+ 𝐵𝐵 𝑡𝑡 + 1
)
( 𝑤𝑤 𝑡𝑡 ℎ
∗
𝑡𝑡 + 𝐼𝐼 𝑡𝑡 ) (1.25)
𝑙𝑙 𝑐𝑐 𝑡𝑡 =
𝐺𝐺 𝑡𝑡 + 1
𝑝𝑝 𝑖𝑖 ( 𝛼𝛼 2
+ 𝐶𝐶 𝑡𝑡 + 1
)
( 𝑤𝑤 𝑡𝑡 ℎ
∗
𝑡𝑡 + 𝐼𝐼 𝑡𝑡 )
𝑚𝑚 𝑐𝑐 𝑡𝑡 =
𝐹𝐹 𝑡𝑡 + 1
𝛼𝛼 2
+ 𝐹𝐹 𝑡𝑡 + 1
( 𝑤𝑤 𝑡𝑡 ℎ
∗
𝑡𝑡 + 𝐼𝐼 𝑡𝑡 )
where notations are recursively defined as follows:

36
𝐻𝐻 𝑡𝑡 + 1
= 𝛽𝛽 (𝛿𝛿 1 𝑡𝑡 𝐿𝐿 𝐷𝐷 𝑡𝑡 + 1
𝐿𝐿 + 𝛿𝛿 1 𝑡𝑡 𝐸𝐸 𝐷𝐷 𝑡𝑡 + 1
𝐸𝐸 + 𝛿𝛿 1 𝑡𝑡 𝑆𝑆 𝐷𝐷 𝑡𝑡 + 1
𝑆𝑆 + 𝛿𝛿 1 𝑡𝑡 𝐴𝐴 𝐷𝐷 𝑡𝑡 + 1
𝐴𝐴 )
𝐾𝐾 𝑡𝑡 + 1
= 𝛽𝛽 (𝛿𝛿 2 𝑡𝑡 𝐿𝐿 𝐷𝐷 𝑡𝑡 + 1
𝐿𝐿 + 𝛿𝛿 2 𝑡𝑡 𝐸𝐸 𝐷𝐷 𝑡𝑡 + 1
𝐸𝐸 + 𝛿𝛿 2 𝑡𝑡 𝑆𝑆 𝐷𝐷 𝑡𝑡 + 1
𝑆𝑆 + 𝛿𝛿 2 𝑡𝑡 𝐴𝐴 𝐷𝐷 𝑡𝑡 + 1
𝐴𝐴 ) (1.26)
𝐺𝐺 𝑡𝑡 + 1
= 𝛽𝛽 (𝛿𝛿 3 𝑡𝑡 𝐿𝐿 𝐷𝐷 𝑡𝑡 + 1
𝐿𝐿 + 𝛿𝛿 3 𝑡𝑡 𝐸𝐸 𝐷𝐷 𝑡𝑡 + 1
𝐸𝐸 + 𝛿𝛿 3 𝑡𝑡 𝑆𝑆 𝐷𝐷 𝑡𝑡 + 1
𝑆𝑆 + 𝛿𝛿 3 𝑡𝑡 𝐴𝐴 𝐷𝐷 𝑡𝑡 + 1
𝐴𝐴 )
𝐹𝐹 𝑡𝑡 + 1
= 𝛽𝛽 (𝛿𝛿 4 𝑡𝑡 𝐿𝐿 𝐷𝐷 𝑡𝑡 + 1
𝐿𝐿 + 𝛿𝛿 4 𝑡𝑡 𝐸𝐸 𝐷𝐷 𝑡𝑡 + 1
𝐸𝐸 + 𝛿𝛿 4 𝑡𝑡 𝑆𝑆 𝐷𝐷 𝑡𝑡 + 1
𝑆𝑆 + 𝛿𝛿 4 𝑡𝑡 𝐴𝐴 𝐷𝐷 𝑡𝑡 + 1
𝐴𝐴 )

𝐷𝐷 𝑇𝑇 + 1
𝐿𝐿 = 𝜌𝜌 𝛼𝛼 3

𝐷𝐷 𝑇𝑇 + 1
𝐸𝐸 = 𝜌𝜌 𝛼𝛼 4
(1.27)
𝐷𝐷 𝑇𝑇 + 1
𝑆𝑆 = 𝜌𝜌 𝛼𝛼 5

𝐷𝐷 𝑇𝑇 + 1
𝐴𝐴 = 𝜌𝜌 𝛼𝛼 6

𝐷𝐷 𝑡𝑡 𝐿𝐿 = 𝛼𝛼 3
+ 𝛽𝛽 (𝛿𝛿 5 𝑡𝑡 𝐿𝐿 𝐷𝐷 𝑡𝑡 + 1
𝐿𝐿 + 𝛿𝛿 5 𝑡𝑡 𝐸𝐸 𝐷𝐷 𝑡𝑡 + 1
𝐸𝐸 + 𝛿𝛿 5 𝑡𝑡 𝑆𝑆 𝐷𝐷 𝑡𝑡 + 1
𝑆𝑆 + 𝛿𝛿 5 𝑡𝑡 𝐴𝐴 𝐷𝐷 𝑡𝑡 + 1
𝐴𝐴 ) 𝑓𝑓𝑓𝑓 𝑓𝑓 ∀ 𝑡𝑡 ∈ (1, 𝑇𝑇 )
𝐷𝐷 𝑡𝑡 𝐸𝐸 = 𝛼𝛼 4
+ 𝛽𝛽 (𝛿𝛿 6 𝑡𝑡 𝐿𝐿 𝐷𝐷 𝑡𝑡 + 1
𝐿𝐿 + 𝛿𝛿 6 𝑡𝑡 𝐸𝐸 𝐷𝐷 𝑡𝑡 + 1
𝐸𝐸 + 𝛿𝛿 6 𝑡𝑡 𝑆𝑆 𝐷𝐷 𝑡𝑡 + 1
𝑆𝑆 + 𝛿𝛿 6 𝑡𝑡 𝐴𝐴 𝐷𝐷 𝑡𝑡 + 1
𝐴𝐴 ) for ∀t ∈ (1, T)
𝐷𝐷 𝑡𝑡 𝑆𝑆 = 𝛼𝛼 5
+ 𝛽𝛽 (𝛿𝛿 7 𝑡𝑡 𝐿𝐿 𝐷𝐷 𝑡𝑡 + 1
𝐿𝐿 + 𝛿𝛿 7 𝑡𝑡 𝐸𝐸 𝐷𝐷 𝑡𝑡 + 1
𝐸𝐸 + 𝛿𝛿 7 𝑡𝑡 𝑆𝑆 𝐷𝐷 𝑡𝑡 + 1
𝑆𝑆 + 𝛿𝛿 7 𝑡𝑡 𝐴𝐴 𝐷𝐷 𝑡𝑡 + 1
𝐴𝐴 ) for ∀t ∈ (1, T)
𝐷𝐷 𝑡𝑡 𝐴𝐴 = 𝛼𝛼 6
+ 𝛽𝛽 (𝛿𝛿 8 𝑡𝑡 𝐿𝐿 𝐷𝐷 𝑡𝑡 + 1
𝐿𝐿 + 𝛿𝛿 8 𝑡𝑡 𝐸𝐸 𝐷𝐷 𝑡𝑡 + 1
𝐸𝐸 + 𝛿𝛿 8 𝑡𝑡 𝑆𝑆 𝐷𝐷 𝑡𝑡 + 1
𝑆𝑆 + 𝛿𝛿 8 𝑡𝑡 𝐴𝐴 𝐷𝐷 𝑡𝑡 + 1
𝐴𝐴 ) for ∀t ∈ (1, T)
These notations might seem complicated, but the intuition behind it is quite simple. For example,
𝐷𝐷 𝑡𝑡 𝐿𝐿 in Equation (1.27) represents the continuation value of mother’s appreciation toward the child’s
language and communications; 𝐻𝐻 𝑡𝑡 + 1
in Equation (1.26) represents the marginal benefit from the
mother’s time considering the stream of marginal utility created by the ability that carries over to
the next period; 𝐾𝐾 𝑡𝑡 + 1
, 𝐺𝐺 𝑡𝑡 + 1
, and 𝐹𝐹 𝑡𝑡 + 1
all represent such marginal benefits corresponding to
investments in terms of institutional care, private care, and monetary expenditures for a child. The
analytical solutions in Equations (24) and (25) consider all four dimensions for the mother’s
optimal choice.

37
1.7. Empirical Strategy: Multidimensional
1.7.1. Additional Specifications
As mentioned in the last section, most of the specification assumptions remain unchanged for
our extended model, although an additional specification requires modification. Specifically,
mothers in the current model appreciate all four dimensions of children’s ability, and thus,
corresponding parameters must be specified differently. Consequently, we impose a specification
that depends additionally on the child’s sex and birth order. For brevity, this specification is
presented as follows:

𝛼𝛼 1
=
exp (𝛾𝛾 1
)
exp( 𝛾𝛾 1
) + exp( 𝛾𝛾 2
𝑚𝑚 𝑓𝑓 𝑚𝑚 _ 𝑒𝑒 𝑑𝑑𝑢𝑢 ) + exp (𝛾𝛾 4
𝑢𝑢 𝑓𝑓 𝑢𝑢 𝑎𝑎 𝑙𝑙 ) + exp (𝛾𝛾 5
𝑚𝑚 𝑓𝑓 𝑚𝑚 _𝑎𝑎𝑎𝑎 𝑒𝑒 _𝑎𝑎 𝑡𝑡 _𝑢𝑢 𝑖𝑖𝑓𝑓 𝑡𝑡 ℎ) +
exp (𝛾𝛾 6
𝑑𝑑 𝑎𝑎𝑢𝑢𝑎𝑎 ℎ𝑡𝑡 𝑒𝑒𝑓𝑓 + 𝛾𝛾 7
𝑢𝑢 𝑖𝑖𝑓𝑓 𝑡𝑡 ℎ_ 𝑓𝑓 𝑓𝑓𝑑𝑑𝑒𝑒𝑓𝑓 ) + exp (𝛾𝛾 8
𝐼𝐼 � 𝜇𝜇 0
= 𝜇𝜇 0 ℎ 𝑖𝑖 𝑖𝑖 ℎ
� )

𝛼𝛼 2
=
exp( 𝛾𝛾 2
𝑚𝑚 𝑓𝑓 𝑚𝑚 _ 𝑒𝑒 𝑑𝑑𝑢𝑢 )
exp( 𝛾𝛾 1
) + exp( 𝛾𝛾 2
𝑚𝑚 𝑓𝑓 𝑚𝑚 _ 𝑒𝑒 𝑑𝑑𝑢𝑢 ) + exp (𝛾𝛾 4
𝑢𝑢 𝑓𝑓 𝑢𝑢 𝑎𝑎 𝑙𝑙 ) + exp (𝛾𝛾 5
𝑚𝑚 𝑓𝑓 𝑚𝑚 _𝑎𝑎𝑎𝑎 𝑒𝑒 _𝑎𝑎 𝑡𝑡 _𝑢𝑢 𝑖𝑖𝑓𝑓 𝑡𝑡 ℎ) +
exp (𝛾𝛾 6
𝑑𝑑 𝑎𝑎𝑢𝑢𝑎𝑎 ℎ𝑡𝑡 𝑒𝑒𝑓𝑓 + 𝛾𝛾 7
𝑢𝑢 𝑖𝑖𝑓𝑓 𝑡𝑡 ℎ_ 𝑓𝑓 𝑓𝑓𝑑𝑑𝑒𝑒𝑓𝑓 ) + exp (𝛾𝛾 8
𝐼𝐼 � 𝜇𝜇 0
= 𝜇𝜇 0 ℎ 𝑖𝑖 𝑖𝑖 ℎ
� )

𝛼𝛼 3
=
exp (𝛾𝛾 4
𝑢𝑢 𝑓𝑓 𝑢𝑢 𝑎𝑎 𝑙𝑙 )
exp( 𝛾𝛾 1
) + exp( 𝛾𝛾 2
𝑚𝑚 𝑓𝑓 𝑚𝑚 _ 𝑒𝑒 𝑑𝑑𝑢𝑢 ) + exp (𝛾𝛾 4
𝑢𝑢 𝑓𝑓 𝑢𝑢 𝑎𝑎 𝑙𝑙 ) + exp (𝛾𝛾 5
𝑚𝑚 𝑓𝑓 𝑚𝑚 _𝑎𝑎𝑎𝑎 𝑒𝑒 _𝑎𝑎 𝑡𝑡 _ 𝑢𝑢 𝑖𝑖 𝑓𝑓𝑡𝑡 ℎ) +
exp (𝛾𝛾 6
𝑑𝑑 𝑎𝑎𝑢𝑢𝑎𝑎 ℎ𝑡𝑡 𝑒𝑒𝑓𝑓 + 𝛾𝛾 7
𝑢𝑢 𝑖𝑖𝑓𝑓 𝑡𝑡 ℎ_ 𝑓𝑓 𝑓𝑓𝑑𝑑𝑒𝑒𝑓𝑓 ) + exp (𝛾𝛾 8
𝐼𝐼 � 𝜇𝜇 0
= 𝜇𝜇 0 ℎ 𝑖𝑖 𝑖𝑖 ℎ
� )

𝛼𝛼 4
=
exp (𝛾𝛾 5
𝑚𝑚 𝑓𝑓 𝑚𝑚 _𝑎𝑎𝑎𝑎 𝑒𝑒 _𝑎𝑎 𝑡𝑡 _𝑢𝑢 𝑖𝑖𝑓𝑓 𝑡𝑡 ℎ)
exp( 𝛾𝛾 1
) + exp( 𝛾𝛾 2
𝑚𝑚 𝑓𝑓 𝑚𝑚 _ 𝑒𝑒 𝑑𝑑𝑢𝑢 ) + exp (𝛾𝛾 4
𝑢𝑢 𝑓𝑓 𝑢𝑢 𝑎𝑎 𝑙𝑙 ) + exp (𝛾𝛾 5
𝑚𝑚 𝑓𝑓 𝑚𝑚 _𝑎𝑎𝑎𝑎 𝑒𝑒 _𝑎𝑎 𝑡𝑡 _ 𝑢𝑢𝑖𝑖 𝑓𝑓 𝑡𝑡ℎ ) +
exp (𝛾𝛾 6
𝑑𝑑 𝑎𝑎𝑢𝑢𝑎𝑎 ℎ𝑡𝑡 𝑒𝑒𝑓𝑓 + 𝛾𝛾 7
𝑢𝑢 𝑖𝑖𝑓𝑓 𝑡𝑡 ℎ_ 𝑓𝑓 𝑓𝑓𝑑𝑑𝑒𝑒𝑓𝑓 ) + exp (𝛾𝛾 8
𝐼𝐼 � 𝜇𝜇 0
= 𝜇𝜇 0 ℎ 𝑖𝑖 𝑖𝑖 ℎ
� )

𝛼𝛼 5
=
exp (𝛾𝛾 6
𝑑𝑑 𝑎𝑎𝑢𝑢𝑎𝑎 ℎ𝑡𝑡 𝑒𝑒𝑓𝑓 + 𝛾𝛾 7
𝑢𝑢 𝑖𝑖𝑓𝑓 𝑡𝑡 ℎ_ 𝑓𝑓 𝑓𝑓𝑑𝑑𝑒𝑒𝑓𝑓 )
exp( 𝛾𝛾 1
) + exp( 𝛾𝛾 2
𝑚𝑚 𝑓𝑓 𝑚𝑚 _ 𝑒𝑒 𝑑𝑑𝑢𝑢 ) + exp (𝛾𝛾 4
𝑢𝑢 𝑓𝑓 𝑢𝑢 𝑎𝑎 𝑙𝑙 ) + exp (𝛾𝛾 5
𝑚𝑚 𝑓𝑓 𝑚𝑚 _𝑎𝑎𝑎𝑎 𝑒𝑒 _𝑎𝑎 𝑡𝑡 _𝑢𝑢 𝑖𝑖𝑓𝑓 𝑡𝑡 ℎ) +
exp (𝛾𝛾 6
𝑑𝑑 𝑎𝑎𝑢𝑢𝑎𝑎 ℎ𝑡𝑡 𝑒𝑒𝑓𝑓 + 𝛾𝛾 7
𝑢𝑢 𝑖𝑖𝑓𝑓 𝑡𝑡 ℎ_ 𝑓𝑓 𝑓𝑓𝑑𝑑𝑒𝑒𝑓𝑓 ) + exp (𝛾𝛾 8
𝐼𝐼 � 𝜇𝜇 0
= 𝜇𝜇 0 ℎ 𝑖𝑖 𝑖𝑖 ℎ
� )

𝛼𝛼 6
=
exp (𝛾𝛾 8
𝐼𝐼 � 𝜇𝜇 0
= 𝜇𝜇 0 ℎ 𝑖𝑖 𝑖𝑖 ℎ
� )
exp( 𝛾𝛾 1
) + exp( 𝛾𝛾 2
𝑚𝑚 𝑓𝑓 𝑚𝑚 _ 𝑒𝑒 𝑑𝑑𝑢𝑢 ) + exp (𝛾𝛾 4
𝑢𝑢 𝑓𝑓 𝑢𝑢 𝑎𝑎 𝑙𝑙 ) + exp (𝛾𝛾 5
𝑚𝑚 𝑓𝑓 𝑚𝑚 _𝑎𝑎𝑎𝑎 𝑒𝑒 _𝑎𝑎 𝑡𝑡 _ 𝑢𝑢𝑖𝑖 𝑓𝑓𝑡𝑡 ℎ) +
exp (𝛾𝛾 6
𝑑𝑑 𝑎𝑎𝑢𝑢𝑎𝑎 ℎ𝑡𝑡 𝑒𝑒𝑓𝑓 + 𝛾𝛾 7
𝑢𝑢 𝑖𝑖𝑓𝑓 𝑡𝑡 ℎ_ 𝑓𝑓 𝑓𝑓𝑑𝑑𝑒𝑒𝑓𝑓 ) + exp (𝛾𝛾 8
𝐼𝐼 � 𝜇𝜇 0
= 𝜇𝜇 0 ℎ 𝑖𝑖 𝑖𝑖 ℎ
� )

38
The new specification might be slightly more restrictive in terms of a functional form to
minimize the computational burden by the number of parameters to be estimated.

Additionally, it is fair to note that we still allow for cross-complementarity terms to vary across
time. Therefore, the following Equation (1.28) notes the off-diagonal complementarity terms as
well as standard productivity coefficients:
�
𝑙𝑙𝑙𝑙 𝐿𝐿 𝑡𝑡 + 1
𝑙𝑙𝑙𝑙 𝐸𝐸 𝑡𝑡 + 1
𝑙𝑙𝑙𝑙 𝑆𝑆 𝑡𝑡 + 1
𝑙𝑙𝑙𝑙 𝐴𝐴 𝑡𝑡 + 1
� =
⎣
⎢
⎢
⎢
⎡
𝛿𝛿 1 𝑡𝑡 𝐿𝐿 𝛿𝛿 1 𝑡𝑡 𝐸𝐸 𝛿𝛿 1 𝑡𝑡 𝑆𝑆 𝛿𝛿 1 𝑡𝑡 𝐴𝐴 ⎦
⎥
⎥
⎥
⎤
𝑙𝑙𝑙𝑙 𝜏𝜏 𝑡𝑡 +
⎣
⎢
⎢
⎢
⎡
𝛿𝛿 2 𝑡𝑡 𝐿𝐿 𝛿𝛿 2 𝑡𝑡 𝐸𝐸 𝛿𝛿 2 𝑡𝑡 𝑆𝑆 𝛿𝛿 2 𝑡𝑡 𝐴𝐴 ⎦
⎥
⎥
⎥
⎤
𝑙𝑙𝑙𝑙 𝑖𝑖 𝑡𝑡 +
⎣
⎢
⎢
⎢
⎡
𝛿𝛿 3 𝑡𝑡 𝐿𝐿 𝛿𝛿 3 𝑡𝑡 𝐸𝐸 𝛿𝛿 3 𝑡𝑡 𝑆𝑆 𝛿𝛿 3 𝑡𝑡 𝐴𝐴 ⎦
⎥
⎥
⎥
⎤
𝑙𝑙𝑙𝑙 𝑙𝑙 𝑡𝑡 +
⎣
⎢
⎢
⎢
⎡
𝛿𝛿 4 𝑡𝑡 𝐿𝐿 𝛿𝛿 4 𝑡𝑡 𝐸𝐸 𝛿𝛿 4 𝑡𝑡 𝑆𝑆 𝛿𝛿 4 𝑡𝑡 𝐴𝐴 ⎦
⎥
⎥
⎥
⎤
𝑙𝑙𝑙𝑙 𝑚𝑚 𝑡𝑡 +
⎣
⎢
⎢
⎢
⎡
𝛿𝛿 5 𝑡𝑡 𝐿𝐿 𝛿𝛿 6 𝑡𝑡 𝐿𝐿 𝛿𝛿 7 𝑡𝑡 𝐿𝐿 𝛿𝛿 8 𝑡𝑡 𝐿𝐿 𝛿𝛿 5 𝑡𝑡 𝐸𝐸 𝛿𝛿 6 𝑡𝑡 𝐸𝐸 𝛿𝛿 7 𝑡𝑡 𝐸𝐸 𝛿𝛿 8 𝑡𝑡 𝐸𝐸 𝛿𝛿 5 𝑡𝑡 𝑆𝑆 𝛿𝛿 6 𝑡𝑡 𝑆𝑆 𝛿𝛿 7 𝑡𝑡 𝑆𝑆 𝛿𝛿 8 𝑡𝑡 𝑆𝑆 𝛿𝛿 5 𝑡𝑡 𝐴𝐴 𝛿𝛿 6 𝑡𝑡 𝐴𝐴 𝛿𝛿 7 𝑡𝑡 𝐴𝐴 𝛿𝛿 8 𝑡𝑡 𝐴𝐴 ⎦
⎥
⎥
⎥
⎤
�
𝑙𝑙𝑙𝑙 𝐿𝐿 𝑡𝑡 𝑙𝑙𝑙𝑙 𝐸𝐸 𝑡𝑡 𝑙𝑙𝑙𝑙 𝑆𝑆 𝑡𝑡 𝑙𝑙𝑙𝑙 𝐴𝐴 𝑡𝑡 � (1.28)

which will also take the following form:
𝛿𝛿 𝑖𝑖 𝑡𝑡 𝑘𝑘 = exp(𝜉𝜉 𝑖𝑖 𝑘𝑘 𝑡𝑡 ) 𝑏𝑏 ∈ { 𝐿𝐿 , 𝐸𝐸 , 𝑆𝑆 , 𝐴𝐴 }, 𝑖𝑖 ∈ {1, 2, 3, 4, 5, 6, 7, 8} (1.29)

Finally, the initial ability parameters corresponding to Equation (1.15) will still be estimated,
and we assume a heterogeneous random effect will be independently assigned to each dimension.
With these additional specifications, the fully extended model includes 72 parameters to be
simultaneously estimated.

1.7.2. Changes to the Estimation Method
We also estimate additional parameters by including additional moments (Table 1.9) conditions
to identify the corresponding parameters. This drastically increases the number of moment
conditions to 849; most of these are added to identify the cross-complementarity terms, in which
we impose transitions from one quartile to another. Most of the additional moments are pairwise
transition moments.

39
One final change made for the primary analyses involves the use of an identity matrix as our
weighted matrix, instead of the OMD weighted matrix defined in Equation (1.21). While any
positive, definite weighted matrix will assure consistent estimators, a few cases occurred in which
the OMD matrix was not well-defined due to the smaller S2 sample and larger number of moment
conditions. Therefore, I decided to use an identity weighted matrix, which is also commonly
believed to function better in practice, rather than any further arbitrary manipulations.

1.8. Results: Multidimensional
This section presents the study’s final results.
Tables 1.10 and 1.11 demonstrates that the average mothers in our sample make work and
investment decisions regarding the following utility and budget constraints:
� 0.99
𝑡𝑡 (𝟎𝟎 . 𝟖𝟖𝟖𝟖𝟖𝟖 𝑙𝑙𝑙𝑙 𝑙𝑙 𝑡𝑡 + 𝟎𝟎 . 𝟖𝟖𝟎𝟎 𝟏𝟏 𝑙𝑙𝑙𝑙 𝑐𝑐 𝑡𝑡 ∞
𝑡𝑡 = 0
+ 𝟎𝟎 . 𝟎𝟎𝟎𝟎 𝟏𝟏 𝑙𝑙𝑙𝑙 𝐿𝐿 𝑡𝑡 + 𝟎𝟎 . 𝟎𝟎𝟎𝟎𝟎𝟎𝟎𝟎𝟎𝟎𝟎𝟎 𝟖𝟖 𝑙𝑙𝑙𝑙 𝐸𝐸 𝑡𝑡 + 𝟎𝟎 . 𝟎𝟎𝟎𝟎 𝟏𝟏 𝑙𝑙𝑙𝑙 𝑆𝑆 𝑡𝑡 + 𝟎𝟎 . 𝟎𝟎 𝟖𝟖 𝟎𝟎 𝑙𝑙𝑙𝑙 𝐴𝐴 𝑡𝑡 )
s. t. 𝑐𝑐 𝑡𝑡 + 𝟐𝟐 . 𝟎𝟎𝟎𝟎 𝟏𝟏 𝑖𝑖 𝑡𝑡 + 𝟎𝟎 . 𝟎𝟎𝟎𝟎𝟎𝟎𝟎𝟎𝟎𝟎 𝑙𝑙 𝑡𝑡 + 𝑚𝑚 𝑡𝑡 = 𝑤𝑤 𝑡𝑡 ℎ
𝑡𝑡 + 𝟔𝟔 . 𝟗𝟗 𝟎𝟎𝟎𝟎

As the abilities themselves are again unit-free measures, we do not interpret the imputed
preference parameter’s magnitude. However, our hypothetical average mother seems to care about

40
development in all four dimensions. An extremely small coefficient for the emotionality parameter α
4

may indicate that mothers simply do not care certain dimensions, or may reflect strong correlations between this
dimension of ability and another ability, such that the marginal utility from emotionality itself is small.
Mothers’ offered wages tend to follow the following process, again on average:
𝑙𝑙𝑙𝑙 𝑤𝑤 𝑡𝑡 = − 𝟗𝟗 . 𝟔𝟔 𝟖𝟖𝟐𝟐 + 𝟎𝟎 . 𝟎𝟎 𝟖𝟖 𝟎𝟎 𝑚𝑚 𝑓𝑓 𝑚𝑚 _ 𝑒𝑒 𝑑𝑑𝑢𝑢 + 𝟎𝟎 . 𝟎𝟎 𝟖𝟖 𝟎𝟎 𝑎𝑎𝑎𝑎 𝑒𝑒 𝑡𝑡 − 𝟎𝟎 . 𝟎𝟎𝟎𝟎𝟎𝟎 𝟐𝟐 𝑎𝑎𝑎𝑎 𝑒𝑒 𝑡𝑡 2
+ 𝟎𝟎 . 𝟎𝟎𝟏𝟏𝟏𝟏 𝑢𝑢 𝑓𝑓 𝑢𝑢 𝑎𝑎 𝑙𝑙 + 𝜀𝜀 𝑡𝑡
𝜀𝜀 𝑡𝑡 ~ 𝑖𝑖 . 𝑖𝑖 . 𝑑𝑑 . 𝑁𝑁 ( 𝟎𝟎 , 𝟎𝟎 . 𝟖𝟖𝟐𝟐𝟖𝟖
𝟐𝟐 )
This result seems more intuitive compared to the motivational result. Specifically, the
coefficient for 𝑎𝑎𝑎𝑎 𝑒𝑒 𝑡𝑡 is now positive and significant but quite small in magnitude, and the
coefficient for 𝑎𝑎𝑎𝑎 𝑒𝑒 𝑡𝑡 2
is estimated as small and negative. These results are quite intuitive
considering the selective sample: females relatively early in their careers who decided to marry
and have children.
We now examine this paper’s topic of interest: the development of children’s ability. Table 1.12
presents the estimates for initial levels of heterogeneous childhood ability. This demonstrates that
no considerable heterogeneity exists that may confound the results, as it is either insignificant or
small in magnitude. One coefficient that is significant and may contradict intuition is the negative
relationship between mothers’ education and children’s language skill, which also appears in Table
1.6 as well as in the simple linear association between language scores and mothers’ education.
This may be due to the mother’s investment before Wave 3, either in their children’s first and
second years, or even as early as in utero. While this is beyond the scope of this paper, this requires
further exploration in future works.
Table 1.14 presents the dynamic evolution of childhood ability, and a result can be observed
that disagrees with that in our motivational result. The imputed productivity 𝛿𝛿 s between Waves 3
and 4 exhibits the following Equation (1.30):

41
�
𝑙𝑙 n 𝐿𝐿 𝑡𝑡 + 1
𝑙𝑙 n 𝐸𝐸 𝑡𝑡 + 1
𝑙𝑙 n 𝑆𝑆 𝑡𝑡 + 1
𝑙𝑙 n 𝐴𝐴 𝑡𝑡 + 1
� = �
𝟎𝟎 . 𝟗𝟗𝟗𝟗 𝟔𝟔 0.008
0.000
0.000
� ln 𝜏𝜏 𝑡𝑡 + �
0.266
0.314
𝟎𝟎 . 𝟖𝟖 𝟏𝟏 𝟎𝟎 0.006
� ln𝑖𝑖 𝑡𝑡 + �
0.000
0.289
0.000
0.000
� ln𝑙𝑙 𝑡𝑡 + �
0.000
0.058
0.611
0.000
� ln𝑚𝑚 𝑡𝑡 + �
0.097 0.000 0.051 0.005
0.014 1.002 0.999 0.999
0.103 0.000 0.960 0.917
0.362 0.000 0.006 0.731
� �
𝑙𝑙 n 𝐿𝐿 𝑡𝑡 𝑙𝑙 n 𝐸𝐸 𝑡𝑡 𝑙𝑙 n 𝑆𝑆 𝑡𝑡 𝑙𝑙 n 𝐴𝐴 𝑡𝑡 � (1.30)

Institutional and private care both seem to be superior options in terms of non-cognitive skills’
productivity. Emotionality 𝐸𝐸 𝑡𝑡 is most efficiently developed by institutional care, although the
efficiency of private care is somewhat comparable in magnitude. Institutional care is the superior
option to facilitate activity 𝐴𝐴 𝑡𝑡 compared to other investment options. One compelling pattern in
the activity dimension is that it does not develop but remains stable, while simultaneously and
quite significantly contributing to the other ability dimensions over time.
A most noteworthy observation in Equation (1.28) involves the comparison between the first
and the third rows, which correspond to language and sociability, respectively. As anticipated, the
latter is primarily and significantly developed through institutional care. This is an intuitive result
considering how institutional care collectively operates.
Alternatively, and contradicting the unidimensional result summarized in Equation (1.22), the
first row in Equation (1.28) demonstrates that the language and communication abilities are
primarily developed through mothers’ care. While institutional care seems to contribute to the
development of this specific cognitive skill, the magnitude of this contribution is significantly
smaller.
Although not directly comparable, this finding partially agrees with work by Fort et al. (2019),
in which the authors discovered a sharp decrease in cognitive ability (IQ) in those who experienced

42
daycare from birth to age two, which is the period directly preceding our own data. While these
authors did not control for mothers’ investments, they suggest a convincing mechanism, in that a
longer exposure to daycare negates the individual interactions with adults. Equation (1.28)
supports this mechanism by demonstrating that quality time with mothers better stimulates certain
cognitive skills, and therefore, the more frequent use of daycare may actually harm children’s
cognitive development.
Simultaneously, private care also has no crucial impact on language development; thus, neither
institutional nor private care functions as an alternative for language development. This contrasts
market care in the unidimensional approach (Section 1.5), which is a mere aggregation of
institutional and private care, and is ultimately superior on average to mothers’ care.
Multiple mechanisms may exist through which differences are created by ignoring multi-
dimensional abilities, such as dynamic complementarity. One mechanism may drive such
differences in the unidimensional approach.
Table 1.15 displays the correlation coefficient between the imputed preference parameters in
our main analysis presented in this section. Note that the parameters presented here are not those
from the motivational analyses in Section 1.5, but from the extended analysis in this section. The
most significant association involves the preferences for language ability and sociability, which
suggests that mothers concerned more with language are likely to appreciate developing
sociability. In this case, mothers who more highly value language ability are also more likely to
also send their children to childcare institutions to promote their sociability, controlling for any
other investment decision. If the correlation in preferences holds, a unidimensional analysis is
likely to detect the strong association between aggregated market care hours and improved
language development. This may lead to the false conclusion that market care also provides
superior language development.

43
In summary, this section’s results compared to those from the unidimensional model
demonstrate how omitting a multidimensional ability may to false conclusions due to the bias from
omitted variables. Specifically, language development, which previous studies primarily focused
on, seems to be enhanced by maternal care instead of any form of market care after incorporating
other ability dimensions. Therefore, I defer to the first question asked in this paper: “Is market-
provided childcare better?” The answer depends on how one defines “better,” as this may be
“better” for certain non-cognitive measures, such as those for EAS temperament, but not for
language development.

1.9. Goodness of Fit
This section discusses the goodness of fit from our simulated estimation; discuss some of its
virtues; and, more importantly, limitations that might contribute to future works in literature.
Table 1.16 presents some of the important averages of our choice variables of interest. All the
hourly variables are 24-hour averages of weekdays and weekends, surveyed separately. Monetary
units are denoted in 10,000 KRW, which may be approximated to 10 USD for the reader’s
convenience. Most of the choice variables align with the data, and are intuitive. Additionally, the
simulated consumption—which is the net of all other children’s expenditures—does not seem to
substantially deviate from intuition, considering our sample of relatively young households with
children.

Figure 1.7 illustrates the mothers’ dynamic choice variable, which is a simulation counterpart
of Figure 1.2. Most of the results replicate the sample quite well, although a mild decrease occurred

44
in hours of maternal and private care as well as an increase in institutional care. Mothers work
more on both intensive and extensive margins, which might vary slightly from Figure 1.4, in which
the labor supply increases in the extensive margin. This is a result of the monotonic change in
productivity terms as in Equations (1.13) and (1.29).
The material expenditures in the fourth panel do not replicate the increasing pattern we found
in our data. This might be due to our model’s restrictive budget set, in which only the dynamics in
disposable income enter the model through the mother’s age. This was an inevitable choice to
avoid any further computational burden, but will have to be seriously considered in any extension
of this paper.
One final topic that deserves further discussion involves the nature of private care. While Table
1.16 well-replicates the average use of private care, the supporting price is an estimated 0.00003,
which is beyond a reasonable market price, and this estimate was insignificant. Further, the
demand curve generated from the simulation, as presented in Figure 1.8, indicates that a small
increase beyond this price will drive the entire demand for private care to quickly diminish to
nearly 0.

While these observations support the decision to distinguish market care into two qualitatively
different types—or institutional and private care in our primary analysis—this calls for further
attention as to the characteristics of private care.
We now turn to the original data from the PSKC. In Wave3, which is the period in which private
surrogate care usage was the most prevalent, mothers were surveyed regarding their attitudes
toward childcare services as well as some characteristics of private care providers.

45
Regarding the question as to what mothers consider in their childcare options, a fair proportion
of mothers responded that providing an educational experience to their children is important
(37.38%), but a majority of mothers also prioritized a safe and healthy environment (56.81%), as
anticipated. For the mothers who actually used private care, they primarily chose such care either
because they felt their children were too young (46.43%), or private care was more reliable
(42.86%), but no one blamed the quality or accessibility of institutional care (both 0%). These
results imply that the demand for private care may be driven by motivations that differ from those
in using institutional care.
When we observe the characteristics of private surrogate caregivers, 87.29% were relatives, and
mostly the grandmothers of our subject children; 53.61% did not graduate high school, but 98.89%
had previous child-rearing experience. It was unexpected that 83.98% claimed monetary
compensation on a regular basis, but the average amount of this compensation was only 630,000
KRW. Wave 3 in our S2 sample revealed the hours for those who ever used private care were 6.6
hours on average, or an hourly wage of approximately 0.31 KRW. This is about 68% of the legal
minimum wage for the same year.
Clearly, the (semi-) market for private care operates differently in our context, as demand is
derived from some other motivation that is not captured by this model’s developmental motivation.
Childcare suppliers are not those who are trained as occupational caregivers, but rather, someone
who has self-trained through experience. It is beyond the scope of this study to verify the true
factors that make private care an attractive option, but it is important to note this shortcoming. In
terms of interpretation, if a significant motive exists that is not incorporated in this model, our
assessment of private care will tend to underestimate its productivity.

46
1.10. Policy
This section examines the impacts of childcare-relevant policies.
As mentioned in the Introduction, alleviating child-rearing costs was a major topic of interest
for the South Korean government, and the nation’s decreasing fertility rates have encouraged the
government to initiate various different means of support.
Four major policy interventions have been considered that resemble existing policies: maternity
leave, unconditional cash transfers, subsidized institutional care, and improved quality of
institutional care. This study provides details on the modified parameters before presenting their
results.

1.10.1. Maternity leave
We consider two different types of maternity leave.
The first is mandatory maternity leave. After mothers in Wave 3 made their labor supply
decisions, the government forced them to take maternity leave, and compensated them with 50%
of their income. This 50% compensation is a yearly average of Korea’s 2019 maternity leave
allowance, or 80% for the first three months and 40% for nine months. Mandatory leave may sound
unrealistic, and in some cases the government even enforced short-term mandatory parental leave
(Solaz et al., 2013).
However, our estimates indicate that a 50% compensation rate was too low to induce voluntary
changes in mothers’ labor supply. Therefore, I suggest a 75% income compensation rate for
mothers, with the option of taking optional maternity leave in Wave 3. While the endogenous
process of deciding whether to take this leave cannot be made without numerically omitting the
value function, this nearly impossible process is avoided due to the simple log linear form assumed

47
in our model. Appendix 10 provides the technical details; under this option, 44% of the mothers
do take this leave.

1.10.2. Unconditional Cash Transfer
Korea implemented an unconditional cash transfer program in 2019. While the amount varies—
from 100,000 to 200,000 KRW per month, depending on the region and child’s age—a fixed
amount of 13,600 KRW is provided per day, but only in Wave 3, to make the program comparable
with maternity leave. Further, this amount is equivalent to mandatory maternity leave in terms of
government expenditure. As the government cost is equalized between the two, a benefit
assessment will serve as a benchmark for a more sophisticated cost-benefit analysis.
1.10.3. Subsidized Institutional Care: Price Controls
Institutional care is highly subsidized by the government, both directly and indirectly. In 2018,
approximately 9.2% of existing daycare centers were public institutions operated by the
government, and 14.2% of children used public daycare. Privately operated institutions are also
directly subsidized for their operating costs, and more importantly, subsidies are provided to those
families who use private daycare, with amounts varying from 220,000 to 681,000 KRW per month.
Thus, a key government policy in our context involves controlling for effective prices.
Moreover, we set the size of three-year subsidy expenditures as equal to the government
expenditures used in Sections 1.10.2 and 1.10.3—using the demand schedule presented in Figure
1.8—to make it comparable with the two previously mentioned policies. As the government
expenditures are distributed over three years, the price decreases slightly, from 20,750 to 19,200
KRW per hour.

48
1.10.4. Quality Improvement
As mentioned in Section 10.3, institutional care in Korea is highly controlled by the government,
and the government has attempted to enhance the quality of childcare institutions through means
other than high subsidies. For example, one such effort involves regulating and incentivizing
privately operated institutions to hire certified caregivers. Specifically, the government initiated a
project in March 2012 to provide a universal curriculum for daycare centers, nursery schools, and
preschools.
The quality improvement process is too ambiguous to be included in a multidimensional model,
and especially since its implementation costs are unknown. Therefore, this study uniformly
simulates three levels of improvement in every dimension by adding 0.0033, 0.0162, and 0.0317
to the hyper-parameter 𝜉𝜉 2
𝑘𝑘 , which increases our institutional productivity parameters in Wave 3 by
approximately 1%, 5%, and 10%, respectively.
For example, a 1% quality improvement occurs by assigning the following new value:
𝜉𝜉 2
𝑘𝑘 ′ = 𝜉𝜉 2
𝑘𝑘 + 0.0033 𝑏𝑏 ∈ { 𝐿𝐿 , 𝐸𝐸 , 𝑆𝑆 , 𝐴𝐴 } (1.31)
which shifts the productivity parameter in every period to
𝛿𝛿 2 𝑡𝑡 𝑘𝑘 ′ = exp((𝜉𝜉 2
𝑘𝑘 + 0.0033)𝑡𝑡 ) 𝑏𝑏 ∈ { 𝐿𝐿 , 𝐸𝐸 , 𝑆𝑆 , 𝐴𝐴 } (1.32)
and enter the second column in our ability development function (28).
The number 0.0033 is chosen to satisfy the following in Wave 3:
𝛿𝛿 2 3
𝑘𝑘 ′ = 1.01 𝛿𝛿 2 3
𝑘𝑘 𝑏𝑏 ∈ { 𝐿𝐿 , 𝐸𝐸 , 𝑆𝑆 , 𝐴𝐴 } (1.33)

1.10.5. Simulation Results and Assessment
Table 1.17 summarizes the impacts of each policy intervention. The first column illustrates the
average increment of the mothers’ short-term utility. Clearly, each of these policies will contribute

49
to mothers’ long-term utility except for mandatory maternal leave, which will have negative long-
term effects; therefore, it will be more noteworthy to discuss the short-term effects. The latter four
columns in this table display the effect on childhood abilities in Wave 5, a separate point of interest
for policy-makers.
Optional leave seems to have a small impact on mothers’ welfare, possibility due to the low
implementation rate. While a substantial decrease can be observed in the emotionality and
sociability dimensions, the increment in the language dimension is marginal and limited, for two
reasons. First, the language ability exhibits a weak intertemporal complementarity to its own
dimension, as presented in Equation (1.30). Alternatively, activity seems to benefit from the higher
language ability in Wave 4. This quick dissipation of cognitive ability parallels Barnett’s (1995)
work. Second, mothers on maternity leave allocate their extra time more toward leisure and less
for maternal care, a pattern revealed in the last two columns in Table 1.17. While there is 0.76%
increase in leisure hours, maternal care hours exhibit a smaller increase, by 0.54%; this is intuitive,
as maternity leave is not solely implemented to promote childcare, but also involves other factors,
such as maternal health. It is noteworthy that such a result is aligned with well-documented
observations of maternity leave leading to maternal leisure, while paternity leave promotes the
father’s time spent in the home, such as childcare (Nepomnyaschy & Waldfogel, 2007; Yeung et
al., 2001).
Three policies equivalent to government expenditures exhibit results that agree with economic
principles without exception. The unconditional cash transfer is the most superior among the three
expenditure equivalent policies, and it assures uniform substantial increments across all four
dimensions. Further, subsidization in institutional care does not significantly improve the mother’s

50
welfare, for the same reason that mothers choose to bear short-term sacrifices for quality
improvement.
When quality improvement is smaller in magnitude—or 1% or 5% in this instance—mothers
choose to bear short-term sacrifices. Four corresponding columns in Table 1.17 indicate how these
mothers assume such burdens; specifically, mothers work more on the extensive margin in cases
involving both 1% and 10% quality improvement, while the increment on the intensive margin is
relatively smaller. Mothers accomplish this by reducing some time for maternal care, but not as
much as their leisure. However, they make even more substantial sacrifices in consumption to best
utilize this opportunity.
Quality improvement might have its own advantage from the policy-maker’s perspective. On
the one hand, other previously mentioned policies, such as maternity leave or unconditional cash
transfers, may consequently decelerate economic growth by discouraging female labor. On the
other hand, quality improvement does not sacrifice short-term economic vigor, and may improve
mothers’ long-term welfare. Additionally, if the improvement is substantial enough, or 10% in the
current case, the short-term sacrifice may even be partially remediated.
This policy evaluation reveals that each policy has their own advantages and disadvantages in
certain dimensions. As the policy-maker’s objectives are and should be multi-dimensional, each
policy deserves further exploration.

1.11. Conclusion and Discussion
This paper initially asked whether working mothers have any impact on early childhood
development. Specifically, the question in this study involves the extent to which market-provided
care can be an alternative for mothers’ care. Literature has yet to reach to a definite consensus

51
regarding this long-standing question. In a developed country context, in which mothers have
access to market-provided childcare, it is especially unclear as to how to identify this problem.
This paper attempted to answer this question through two steps: a structural approach utilizing
a unique data set involving details on mothers’ time; and childhood ability considering both the
cognitive and non-cognitive dimensions. The first half of this paper used a unidimensional
approach to determine that market care can be a perfect substitute for mothers’ care, as the former
seems to increase productivity in both cognitive and non-cognitive development. Therefore,
working mothers do not negatively impact childhood development from any perspective.
However, the multidimensional analysis presented in the second half of this paper refutes the
findings from the first half. The results from collectively considering cognitive and non-cognitive
abilities indicates how maternal care is still crucial in certain cognitive abilities, and a substantial
incentive exists to utilize market care for non-cognitive development. Specifically, we demonstrate
that collective institutional care might contribute largely to sociability in the early childhood.
This illustrates how overlooking the multidimensional nature of childhood abilities might
significantly bias the results. In addition to the dynamic complementarity across dimensions—
which is an obvious confounding factor—this study suggested selective investment due to
heterogeneous preferences might also create bias in a unidimensional model. While this paper’s
findings may not globally apply to every socio-economic context, further study in different
contexts is required to examine qualitatively different market care, or mothers’ characteristics.
However, this paper still contributes to an on-going interest in the mother’s role in early childhood
development by emphasizing the risk of focusing on only one specific ability.
This also shapes a particular policy question, as Section 10 revealed that no alternative is
universally better when considering every policy-maker’s interests. Instead, different policies

52
improve different objectives, and it is the policy-maker’s discretion as to which objectives to
select.
Therefore, the question “Is market care better?” should be rephrased to also consider “better in
what ways?”
This paper also illustrates some potential future research topics, the most obvious of which
involves verifying the other mechanisms that motivate certain types of investments. The main
challenge in this study was to explain the mechanism by which the private care market operates,
which we at least verified as significantly different from the well-defined institutional market.
While private care is quite prevalent in different parts of the world, studies at least exist to unveil
the mechanism.
Another noteworthy extension could incorporate this paper’s findings in a discussion of career
interruption and motherhood penalties. As the current study’s framework did not include wage
dynamics, we are ignoring the substantial costs of mothers’ care. However, the interaction between
the two sides of the mother’s choice are obvious, and both decisions include dynamic choices.
Some pioneering studies (e.g., Blundell et al., 2019) connect mothers’ dynamic career choices and
childcare motivations, which provide even further motivation to incorporate the multidimensional
investment motive and fully verify mothers’ life cycle choice.

53

1.12. Tables and Figures

Table 1.1: Descriptive Statistics_S1

Variable Obs. Mean
Std.
Dev.
Min. Max.
Mother’s
Decision
Mother’s work hours 2,289 2.11 3.06 0 12.29
Mother’s work hours (employed) 801 6.03 1.75 0.71 12.29
Market care hours 2,289 4.38 2.72 0 17.86
Maternal care hours 2,289 5.80 3.04 0 24
Mother’s leisure hours 2,289 16.09 3.53 0 24
Household
Status
Mother’s wages (hourly; in 10,000
KRW)
2,289 0.40 0.78 0 15.30
Mother’s wages (hourly; in 10,000
KRW | employed)
801 1.13 0.96 0 15.30
Urban 2,289 0.62 0.49 0 1
Household income (daily; in 10,000
KRW)
2,289 13.84 15.89 0 267.21
Mother’s years of education 763 14.34 1.91 0 21
Father’s years of education 763 14.70 2.01 12 21
Mother’s age 2,289 33.96 3.77 22 51
Child’s Ability
Language score—normalized to N(5, 1) 2,289 5.01 0.97 0.10 8.13
Emotionality (inversely coded) 2,289 3.17 0.61 1 5
Sociability 2,289 3.49 0.51 1.5 5
Activity 2,289 3.82 0.59 1.4 5
Characteristics
at Birth
Mother’s age at child’s birth 763 31.41 3.66 20 46
Birth order 763 1.70 0.71 1 4
Female 763 0.47 0.50 0 1

54
Table 1.2: Descriptive Statistics_S2

Variable Obs. Mean
Std.
Dev.
Min. Max.
Mother’s Decision
Mother’s work hours 1,704 2.03 3.00 0 12.29
Mother’s work hours (employed) 579 5.98 1.69 0.71 12.29
Institutional care hours 1,704 4.74 2.99 0 13.00
Private care hours 1,704 0.40 1.64 0 15.00
Material expenditures 1,704 0.88 0.80 0 10.69
Maternal care hours 1,704 6.17 3.03 0.004 24
Mother’s leisure hours 1,704 15.80 3.54 0 23.4
Household Status
Mother’s wages (hourly; in 10,000 KRW) 1,704 0.39 0.87 0 15.30
Mother’s wages (hourly; in 10,000 KRW |
employed)
579 1.14 1.17 0 15.30
Urban 1,704 0.63 0.48 0 1
Household income (daily; in 10,000
KRW)
1,704 8.15 12.69 0 194.10
Mother’s years of education (Wave 3) 1,704 14.16 1.86 6 21
Father’s years of education (Wave 3) 1,704 14.46 2.10 9 21
Mother’s age 1,704 33.98 3.63 22 51
Child’s Ability
Language score—normalized to N(10, 1) 1,704 9.98 0.99 5.10 13.13
Emotionality (inversely coded) 1,704 3.15 0.60 1 5
Sociability 1,704 4.09 0.51 2.1 5.6
Activity 1,704 3.81 0.59 1.4 5
Characteristics at
birth
Mother’s age at birth 568 31.43 3.53 20 46
Birth order 568 1.83 0.69 1 4
Female 568 0.46 0.50 0 1

55
Table 1.3: Moments

Moments Considered to Construct Distance

mean and standard deviation for the mother’s hours of work

mean and standard deviation for childcare hours

mean and standard deviation of the mother’s time with her child

the proportion of mothers working

average test score
1)

mean and standard deviation of mother’s wages

mean, standard deviation, and median of household income
2)

correlation coefficient between mother’s wages and mother’s hours of work

correlation coefficient between mother’s wages and household income

correlation coefficient between mother’s hours of work and household income

correlation coefficient between mother’s hours of work and time with her child

correlation coefficient between mother’s hours of work and childcare hours

correlation coefficient between household income and time with her child

correlation coefficient between household income and childcare hours

correlation coefficient between mother’s time with her child in Wave 3 and score in Wave 4

correlation coefficient between mother’s time with her child in Wave 4 and score in Wave 5

correlation coefficient between child care hours in t and score in (t + 1)

correlation coefficient between mother’s hours of work in t and score in (t + 1)

correlation coefficient between household income in t and score in (t + 1)

mean mother’s wages by mother’s education, age, and location
3)

mean mother’s hours of work by mother’s education, age, and location
3)

mean mother’s time with her child by mother’s education, age, and location
3)

mean childcare hours by mother’s education, age, and location
3)

average test score by parents’ education

proportion of children with scores in each quartile range in Wave 3 that transitioned to each quartile range in Wave 4

proportion of children with scores in each quartile range in Wave 4 that transitioned to each quartile range in Wave 5

1) This specific moment is effectively not in use for cognitive skill measure, as the score is standardized to a mean of five.
Therefore, there are a total of 91 moment conditions for the specific case.
2) These three moments are used for a separate estimation due to the i.i.d assumption of household income.
3) Each of the education and age variables were categorized in three groups. Education: High school education or lower,
community college, or bachelor’s degree or higher; Age: 20s, 30s, or 40s or older.

56

Table 1.4: Household income

𝐼𝐼 𝑡𝑡 ~ 𝑖𝑖 . 𝑖𝑖 . 𝑑𝑑 . 𝑁𝑁 ( 𝜇𝜇 𝑖𝑖 𝑖𝑖 𝑐𝑐 , 𝜎𝜎 𝑖𝑖 𝑖𝑖 𝑐𝑐 2
)
Parameter Estimate S.E. z-stat p-value
𝜇𝜇 𝑖𝑖 𝑖𝑖 𝑐𝑐 10.6811 0.182179 58.62967 0
𝜎𝜎 𝑖𝑖 𝑖𝑖 𝑐𝑐 19.58148 1.498496 13.06742 0

Table 1.5: Preference parameters

𝑢𝑢 ( 𝑙𝑙 𝑡𝑡 , 𝑐𝑐 𝑡𝑡 , 𝐴𝐴 𝑡𝑡 ) = 𝛼𝛼 1
𝑙𝑙𝑙𝑙 𝑙𝑙 𝑡𝑡 + 𝛼𝛼 2
𝑙𝑙𝑙𝑙 𝑐𝑐 𝑡𝑡 + 𝛼𝛼 3
𝑙𝑙𝑙𝑙 𝐴𝐴 𝑡𝑡
Parameter Description Language Emotion Sociability Active
Brilli
(2013)
𝛾𝛾 2

Mother’s
education
0.046 -0.013 0.154 0.121 -0.2094
(0.077) (1.186) (0.228) (0.551) (0.0878)
𝛾𝛾 3
Location
-0.023 -0.460 0.026 0.539 0.1369
(0.05) (1.453) (0.054) (0.912) (0.1064)
𝛾𝛾 4

Mother’s
education
-0.020 -0.015 -0.003 -0.036 0.0169
(0.033) (1.154) (0.007) (0.066) (0.0425)
𝛾𝛾 5
Location
0.001 -0.005 0.000 0.197 0.1104
(0.002) (0.941) (0.001) (0.102) (0.0514)
𝛼𝛼 1
Leisure 0.254 0.367 0.087 0.102 0.413
𝛼𝛼 2
Consumption 0.486 0.230 0.800 0.805 0.022
𝛼𝛼 3
Ability 0.259 0.403 0.113 0.093 0.564
𝑧𝑧 𝑚𝑚 ℎ

Mother’s
unobservable
type—
probability
-0.816 -0.217 1.852 6.902

(1.332) (2.143) (0.937) (13.39)

𝜋𝜋 𝑚𝑚 ℎ
0.307 0.446 0.864 0.999 0.3203
𝜋𝜋 𝑚𝑚 𝑙𝑙 0.693 0.554 0.136 0.001 0.6797
p
Price of market
care
0.051 0.149 0.042 0.063 5.1502
(0.041) (1.869) (0.028) (0.034) (0.6517)

57

Table 1.6 : Wage parameter

𝑙𝑙𝑙𝑙 𝑤𝑤 𝑡𝑡 = 𝜇𝜇 0
+ 𝜇𝜇 1
𝑚𝑚 𝑓𝑓 𝑚𝑚 _ 𝑒𝑒 𝑑𝑑 𝑢𝑢 + 𝜇𝜇 2
𝑎𝑎𝑎𝑎𝑒𝑒 𝑡𝑡 + 𝜇𝜇 3
𝑎𝑎𝑎𝑎𝑒𝑒 𝑡𝑡 2
+ 𝜇𝜇 4
𝑢𝑢𝑓𝑓 𝑢𝑢 𝑎𝑎𝑙𝑙 + 𝜀𝜀 𝑡𝑡

Parameter Description Language Emotion Sociability Active Brilli (2013)
𝜇𝜇 1
Mother’s education
0.037 0.466 0.019 0.745 0.0788
(0.064) (0.331) (0.025) (0.803) (0.0445)
𝜇𝜇 2
Mother’s age
-1.646 -1.296 -2.093 -2.901 0.0018
(1.356) (1.106) (3.711) (10.709) (0.0116)
𝜇𝜇 3

Mother’s age
squared
0.032 -0.033 0.007 0.016 -0.0006
(0.068) (0.541) (0.03) (0.072) (0.0008)
𝜇𝜇 4

Location
(“Race” for Brilli, 2013)
-0.594 4.657 3.138 3.932 -0.0156
(0.543) (0.47) (0.983) (7.186) (0.0038)
𝜎𝜎 𝜀𝜀 Wage offer error SD
0.000 0.001 0.000 0.000 0.3031
(0) (0.379) (0) (0) (0.0111)
𝜇𝜇 0ℎ 𝑖𝑖𝑖𝑖 ℎ
5

Constant term—
high-type mother
2.261 42.505 45.738 49.851 3.201
(2.717) (1.013) (114.241) (9.837)
𝜇𝜇 0 𝑙𝑙 𝑙𝑙 𝑙𝑙
Constant term—low-
type mother
4.267 42.311 0.057 -93.037 1.2603
(9.896) (0.944) (0.242) (197.222)
𝑧𝑧 𝑚𝑚 ℎ
Mother’s
unobservable type—
probability
-0.816 -0.217 1.852 6.902
(1.332) (2.143) (0.937) (13.39)
𝜋𝜋 𝑚𝑚 ℎ
0.307 0.446 0.864 0.999 0.3203
𝜋𝜋 𝑚𝑚 𝑙𝑙 0.693 0.554 0.136 0.001 0.6797

5
Note that the name for the mother’s type is simply a randomly given nomination. Thus, it is not peculiar that
“high-type” mothers actually have lower abilities. This similarly applies for children’s unobserved ability.

58
Table 1.7: Initial ability parameter

𝐴𝐴 1
= exp( ψ
𝑐𝑐 + 𝜂𝜂 1
𝑚𝑚 𝑓𝑓 𝑚𝑚 _ 𝑒𝑒 𝑑𝑑𝑢𝑢 + 𝜂𝜂 2
𝑝𝑝 𝑓𝑓 𝑝𝑝 _ 𝑒𝑒 𝑑𝑑𝑢𝑢 )

Parameter Description Language Emotion Sociability Active Brilli (2013)
𝜂𝜂 1

Mother’s
education
-12.521 -0.062 1.363 -107,250.293 2.0618
(20.591) (1.643) (4.354) (163,277.748) (1.2880)
𝜂𝜂 2

Father’s
education
0.054 -12.399 -6.507 -13.499 0.2255
(0.104) (2.408) (16.119) (18.132) (1.1359)
ψ
𝑐𝑐 ℎ

Constant
term—high-
type child
35.055 -1.257 -3.899 -0.126 -52.3704
(80.766) (0.852) (15.014) (0.329) (4.8011)
ψ
𝑐𝑐 𝑙𝑙
Constant
term—low-
type child
-10.584 27.019 -232.703 0.246 -108.0380
(22.78) (0.912) (1,951.941) (1.131) (13.7080)
𝑧𝑧 𝑐𝑐 ℎ

Child's
unobservable
type—
probability
0.051 0.588 0.000 -0.439
(0.116) (1.049) (0) (0.61)
𝜋𝜋 𝑐𝑐 ℎ
0.513 0.643 0.500 0.392 0.0001
𝜋𝜋 𝑐𝑐 𝑙𝑙 0.487 0.357 0.500 0.608 0.9999

59
Table 1.8: Ability parameter

ln 𝐴𝐴 𝑡𝑡 + 1
= 𝛿𝛿 1 𝑡𝑡 ln 𝜏𝜏 𝑡𝑡 + 𝛿𝛿 2 𝑡𝑡 ln 𝑗𝑗 𝑡𝑡 + 𝛿𝛿 3 𝑡𝑡 ln 𝐼𝐼 𝑡𝑡 + 𝛿𝛿 4 𝑡𝑡 ln 𝐴𝐴 𝑡𝑡

Parameter Description Language Emotion Sociability Active Brilli (2013)
𝜉𝜉 1
Mother’s care
-0.442 -0.906 -0.624 -0.756 -0.4388
(0.049) (0.774) (0.059) (0.23) (0.1299)
𝜉𝜉 2
Market care
-0.216 -0.205 -0.109 -0.119 -0.3514
(0.277) (0.718) (0.17) (0.229) (0.1169)
𝜉𝜉 3
Household income
0.001 -0.006 -0.105 0.249 -0.0566
(0.002) (0.695) (0.38) (0.636) (0.0510)
𝜉𝜉 4
Lag ability
-0.136 -0.018 -0.154 -0.045 -0.1026
(0.214) (0.674) (0.123) (0.074) (0.1215)
𝜎𝜎 𝜈𝜈
Measurement error
SD
4.469 0.116 0.000 0.000 15.2839
(10.397) (0.983) (0) (0) (1.2061)
𝛿𝛿 13
Mother’s care (W3) 0.265 0.066 0.154 0.104 0.2681
𝛿𝛿 23
Market care (W3) 0.523 0.540 0.720 0.700 0.3485
𝛿𝛿 33

Household income
(W3)
1.003 0.982 0.731 2.110 0.8438
𝛿𝛿 43
Lag ability (W3) 0.664 0.947 0.630 0.873 0.7351
𝛿𝛿 14
Mother’s care (W4) 0.171 0.027 0.083 0.049 0.1729
𝛿𝛿 24
Market care (W4) 0.422 0.440 0.646 0.621 0.2452
𝛿𝛿 34

Household income
(W4)
1.004 0.976 0.658 2.706 0.7974
𝛿𝛿 44
Lag ability (W4) 0.579 0.930 0.540 0.835 0.6634

60
Table 1.9: Additional Moments

mean and standard deviation of each type of market care hours by wave
mean of material investments by mother’s and child’s characteristics
mean and standard deviation of mother’s work hours, conditional on labor market participation by wave
proportion and standard deviation of mothers working by wave
mean and standard deviation of mothers’ leisure hours
mean and standard deviation of mothers’ leisure, conditional on mothers’ labor market participation
mean and standard deviation of mother’s wages, conditional on labor market participation
corr among mother’s work hours, each type of childcare hours, and material investments
corr among mother’s work hours, each type of childcare hours, and material investments
corr the maternal time with the child in Wave 3 and each score in Wave 4
corr maternal time with the child in Wave 4 and each score in Wave 5
corr childcare hours in t and score in (t + 1)
corr material investments in t and score in (t + 1)
proportion of mothers working by the mother’s education, age, location, age at birth, gender, and birth order 3)
mean mother’s wages by the mother’s education, age, urban, age at birth, gender, and birth order
3)

mean maternal time with the child by the mother’s education, age, urban, age at birth, gender, and birth order
3)

mean childcare hours by the mother’s education, age, urban, age at birth, gender, and birth order
3)

mean material investments by the mother’s education, age, urban, age at birth, gender, and birth order
3)

average of each test score by parents’ education
proportion of children with each score in each quartile range in Wave 3 that transitioned to each quartile range of each
test score in Wave 4
proportion of children with each score in each quartile range in Wave 4 that transitioned to each quartile range of each
test score in Wave 5

61
Table 1.10: Household income

𝐼𝐼 𝑡𝑡 ~ 𝑖𝑖 . 𝑖𝑖 . 𝑑𝑑 . 𝑁𝑁 ( 𝜇𝜇 𝑖𝑖 𝑖𝑖 𝑐𝑐 , 𝜎𝜎 𝑖𝑖 𝑖𝑖 𝑐𝑐 2
)

Parameter Estimate S.E. z-stat p-value
𝜇𝜇 𝑖𝑖 𝑖𝑖 𝑐𝑐 6.903 0.316 21.512 0
𝜎𝜎 𝑖𝑖 𝑖𝑖 𝑐𝑐 13.585 1.570 10.037 0

62
Table 1.11: Preference parameter

u(l
t
, c
t
, 𝐿𝐿 𝑡𝑡 , 𝐸𝐸 𝑡𝑡 , 𝑆𝑆 𝑡𝑡 𝐴𝐴 𝑡𝑡 ) = 𝛼𝛼 1
𝑙𝑙𝑙𝑙 𝑙𝑙 𝑡𝑡 + 𝛼𝛼 2
𝑙𝑙𝑙𝑙 𝑐𝑐 𝑡𝑡 + 𝛼𝛼 3
𝑙𝑙𝑙𝑙 𝐿𝐿 𝑡𝑡 + 𝛼𝛼 4
𝑙𝑙𝑙𝑙 𝐸𝐸 𝑡𝑡 + 𝛼𝛼 5
𝑙𝑙𝑙𝑙 𝑆𝑆 𝑡𝑡 + 𝛼𝛼 6
𝑙𝑙𝑙𝑙 𝐴𝐴 𝑡𝑡

Parameter Description Estimate
γ
2
Mother’s education
0.206
(0.012)
γ
3
Location
-0.059
(0.026)
γ
4
Mother’s age at birth
-0.439
(0.04)
γ
5
Child’s sex
0.013
(0.008)
γ
6
Birth order
-0.027
(0.017)
γ
7
Mother’s unobservable type
4.068
(0.129)
α
1
Leisure 0.841
α
2
Consumption 0.105
α
3
Language 0.005
α
4
Emotionality 0.00000001
α
5
Sociability 0.005
α
6
Activity 0.043
z
mh

Mother’s unobservable type—probability
0.000
(0.000001)
π
mh
0.500
π
ml
0.500
𝑝𝑝 𝑖𝑖𝑖𝑖𝑖𝑖 𝑡𝑡 Hourly price for institutional childcare
2.075
(0.096)
𝑝𝑝 𝑝𝑝 𝑝𝑝 𝑖𝑖 𝑝𝑝 Hourly price for private childcare
0.00003
(0.00019)

63

Table 1.12: Wage parameter

𝑙𝑙𝑙𝑙 𝑤𝑤 𝑡𝑡 = 𝜇𝜇 0
+ 𝜇𝜇 1
𝑚𝑚 𝑓𝑓 𝑚𝑚 _ 𝑒𝑒 𝑑𝑑 𝑢𝑢 + 𝜇𝜇 2
𝑎𝑎𝑎𝑎𝑒𝑒 𝑡𝑡 + 𝜇𝜇 3
𝑎𝑎𝑎𝑎𝑒𝑒 𝑡𝑡 2
+ 𝜇𝜇 4
𝑢𝑢 𝑓𝑓 𝑢𝑢 𝑎𝑎𝑙𝑙 + 𝜀𝜀 𝑡𝑡

Parameter Description Language
μ
1
Mother’s education
0.047
(0.007)
μ
2
Mother’s age
0.010
(0.002)
μ
3
Mother’s age squared
-0.0002
(0.0001)
μ
4
Location
0.355
(0.098)
σ
ε
Wage offer error SD
0.421
(0.05)
μ
0h i gh
Constant term—high-type mother
-18.420
(2.884)
μ
0l ow
Constant term—low-type mother
-0.805
(0.111)
z
mh

Mother’s unobservable type—probability
0.000001
(0.000001)
π
mh
0.500
π
ml
0.500

64

Table 1.13: Initial ability parameters

𝐴𝐴 1
= exp( ψ
𝑐𝑐 + 𝜂𝜂 1
𝑚𝑚𝑓𝑓𝑚𝑚 _𝑒𝑒 𝑑𝑑𝑢𝑢 + 𝜂𝜂 2
𝑝𝑝 𝑓𝑓 𝑝𝑝 _𝑒𝑒 𝑑𝑑𝑢𝑢 )

Parameter Description Language Emotion Sociability Activity
η
1
Mother’s education
-1.707 0.0003 -0.012 -0.0000005
(0.257) (0.0001) (0.004) (0.0000003)
η
2
Father’s education
0.000004 -0.040 0.018 -0.001
(0.000004) (0.019) (0.005) (0.001)
ψ
ch
Constant term—high-type child
-0.0000001 0.0001 -0.0001 -0.0000006
(0.0000004) (0.00018) (0.0001) (0.000025)
ψ
cl
Constant term—low-type child
0.004 0.0000294 -0.0000873 0.024
(0.002) (0.00002) (0.0001) (0.009)
z
ch

Child’s unobservable type—
probability
0.001 -0.0000001 -0.020 0.008
(0.0004) (0.0000007) (0.009) (0.005)
π
c h
0.500 0.500 0.495 0.502
π
cl
0.500 0.500 0.505 0.498

65
Table 1.14: Ability parameters

�
𝑙𝑙 n 𝐿𝐿 𝑡𝑡 + 1
𝑙𝑙 n 𝐸𝐸 𝑡𝑡 + 1
𝑙𝑙 n 𝑆𝑆 𝑡𝑡 + 1
𝑙𝑙 n 𝐴𝐴 𝑡𝑡 + 1
� =
⎣
⎢
⎢
⎢
⎡
𝛿𝛿 1 𝑡𝑡 𝐿𝐿 𝛿𝛿 1 𝑡𝑡 𝐸𝐸 𝛿𝛿 1 𝑡𝑡 𝑆𝑆 𝛿𝛿 1 𝑡𝑡 𝐴𝐴 ⎦
⎥
⎥
⎥
⎤
ln 𝜏𝜏 𝑡𝑡 +
⎣
⎢
⎢
⎢
⎡
𝛿𝛿 2 𝑡𝑡 𝐿𝐿 𝛿𝛿 2 𝑡𝑡 𝐸𝐸 𝛿𝛿 2 𝑡𝑡 𝑆𝑆 𝛿𝛿 2 𝑡𝑡 𝐴𝐴 ⎦
⎥
⎥
⎥
⎤
ln 𝑖𝑖 𝑡𝑡 +
⎣
⎢
⎢
⎢
⎡
𝛿𝛿 3 𝑡𝑡 𝐿𝐿 𝛿𝛿 3 𝑡𝑡 𝐸𝐸 𝛿𝛿 3 𝑡𝑡 𝑆𝑆 𝛿𝛿 3 𝑡𝑡 𝐴𝐴 ⎦
⎥
⎥
⎥
⎤
ln 𝑙𝑙 𝑡𝑡 +
⎣
⎢
⎢
⎢
⎡
𝛿𝛿 4 𝑡𝑡 𝐿𝐿 𝛿𝛿 4 𝑡𝑡 𝐸𝐸 𝛿𝛿 4 𝑡𝑡 𝑆𝑆 𝛿𝛿 4 𝑡𝑡 𝐴𝐴 ⎦
⎥
⎥
⎥
⎤
ln 𝑚𝑚 𝑡𝑡 +
⎣
⎢
⎢
⎢
⎡
𝛿𝛿 5 𝑡𝑡 𝐿𝐿 𝛿𝛿 6 𝑡𝑡 𝐿𝐿 𝛿𝛿 7 𝑡𝑡 𝐿𝐿 𝛿𝛿 8 𝑡𝑡 𝐿𝐿 𝛿𝛿 5 𝑡𝑡 𝐸𝐸 𝛿𝛿 6 𝑡𝑡 𝐸𝐸 𝛿𝛿 7 𝑡𝑡 𝐸𝐸 𝛿𝛿 8 𝑡𝑡 𝐸𝐸 𝛿𝛿 5 𝑡𝑡 𝑆𝑆 𝛿𝛿 6 𝑡𝑡 𝑆𝑆 𝛿𝛿 7 𝑡𝑡 𝑆𝑆 𝛿𝛿 8 𝑡𝑡 𝑆𝑆 𝛿𝛿 5 𝑡𝑡 𝐴𝐴 𝛿𝛿 6 𝑡𝑡 𝐴𝐴 𝛿𝛿 7 𝑡𝑡 𝐴𝐴 𝛿𝛿 8 𝑡𝑡 𝐴𝐴 ⎦
⎥
⎥
⎥
⎤
�
𝑙𝑙 n 𝐿𝐿 𝑡𝑡 𝑙𝑙 n 𝐸𝐸 𝑡𝑡 𝑙𝑙 n 𝑆𝑆 𝑡𝑡 𝑙𝑙 n 𝐴𝐴 𝑡𝑡 �

Parameter Description Language Emotion Sociability Activity
ξ
1
Mother’s care
-0.001 -0.839 -6.353 -10.926
(0.006) (0.319) (1.613) (3.796)
ξ
2

Institutional
care
-0.442 -0.386 0.382 -1.715
(0.162) (0.149) (0.016) (0.831)
ξ
3
Private care
-7.589 -0.413 -9.090 -10.318
(3.069) (0.089) (2.854) (2.894)
ξ
4
Material
-5.493 -0.950 -0.164 -3.743
(2.208) (0.432) (0.026) (1.244)
ξ
5
Language
-0.010 -1.429 -0.759 -0.339
(0.137) (0.469) (0.051) (0.032)
ξ
6
Emotion
-6.412 0.001 -2.861 -6.568
(1.438) (0.0003) (0.704) (1.887)
ξ
7
Sociability
-0.988 0.000 -0.014 -1.690
(0.523) (0.00001) (0.001) (0.563)
ξ
8
Activity
-1.735 0.000 -0.029 -0.104
(0.56) (0.0001) (0.03) (0.095)
σ
ν

Measurement
Error SD
10.225 3.610 0.00001 0.00001
(3.61) (0.941) (0.00001) (0.000004)

66
Table 1.15: The correlation-coefficient—Simulated preference parameter

Language
Emotional
ity
Sociabilit
y
Activity

𝜶𝜶 𝟎𝟎 𝜶𝜶 𝟖𝟖 𝜶𝜶 𝟏𝟏 𝜶𝜶 𝟔𝟔
Language 𝜶𝜶 𝟎𝟎 1

Emotionality 𝜶𝜶 𝟖𝟖 0.091 1

Sociability 𝜶𝜶 𝟏𝟏 0.9761 0.109 1

Activity 𝜶𝜶 𝟔𝟔 -0.9297 -0.0458 -0.9466 1

67

Table 1.16: Goodness of fit

Variables Real Data Simulated Data
Child Investment
Mother’s care (hours / day) 6.1716 6.1750
Institutional care (hours / day) 4.7397 4.5140
Private care (hours / day) 0.3953 0.3844
Material expenditures (10,000 KRW) 0.8777 0.9048
Resource Allocation
Household income (10,000 KRW) 8.1528 9.5479
Mothers’ leisure (hours / day) 15.7966 15.5970
Consumption (10,000 KRW)
Not
observed
2.5771
Labor Market Decision
Mothers’ LMP (proportion) 0.3398 0.3672
Employed mothers’ work (hours / day) 5.9794 6.0677
Employed mothers’ wages (10,000 KRW) 1.1373 1.4780

68

Table 1.17: Effects of policy intervention (%)

Mother’s
welfare
(W3–W5)
Language
(W5)
Emotionality
(W5)
Sociability
(W5)
Activity
(W5)
Optional maternity leave (75% compensation in Wave 3) 0.50 0.00 -0.72 -0.75 0.21
Mandatory maternity leave
(50% compensation in Wave 3)
1.76 0.50 -4.18 -3.95 0.62
Unconditional cash transfers
(13,600 KRW / day in Wave 3)
4.90 29.02 22.03 18.33 12.16
Subsidized institutional care (19,200 KRW / hour for three years) 0.11 3.81 33.19 80.89 0.62
Quality control (1% increase in all dimensions) -0.07 0.66 9.34 27.08 0.00
Quality control (5% increase in all dimensions) -0.12 3.483 56.70 235.67 0.21
Quality control (10% increase in all dimensions) 1.13 7.131 147.08 1,062.93 0.41

69

Table 1.18: Mothers’ responses to policy intervention

Variables
No
Intervention
1%
Quality Improvement
10%
Quality Improvement
75% Compensated
Optional Maternity
Leave
Level % Change Level % Change Level % Change
Child Investment
Mother’s care 6.1750 6.1712 -0.06 6.1365 -0.62 6.2050 0.49
Institutional care 4.5140 4.5413 0.60 4.7797 5.88 4.4958 -0.40
Private care 0.3844 0.3819 -0.63 0.3617 -5.89 0.3817 -0.69
Material expenditures 0.9048 0.8987 -0.68 0.8474 -6.35 0.8989 -0.66
Resource
Allocation
Mother’s leisure 15.5970 15.5783 -0.12 15.4079 -1.21 15.7158 0.76
Consumption 2.5771 2.5598 -0.67 2.3763 -7.79 2.5574 -0.76
Labor Market
Decision
Mother’s LMP (proportion) 0.3672 0.3728 1.52 0.3797 3.40 0.3210 -12.58
Employed mothers’ working
hours
6.0677 6.1053 0.62 6.4683 6.60 6.4776 6.76
Employed mothers’ wages 1.4780 1.4772 -0.05 1.4706 -0.51 1.4592 -1.27

70

Figure 1.1: Female Labor Market Participation

Figure 1.2: Fertility rate in South Korea(2000-2018)

71

0.04
0.98
1.38
1.40
1.45
3.56
4.23
11.13
0.00 2.00 4.00 6.00 8.00 10.00 12.00
Others
Childcare Support System
Youth Employment
Career-Family Compatibility
Medical Support
Residential Support
Educational Reform
Customized Childcare
Trillion KRW
Source: National Assembly Budget Office (2019)
Figure 1.3 : Earmarked Government Budget (2018)

72

Figure 1.4: Mothers’ Choice Variable (Data)

73

Figure 1.5 : Distribution of mothers’ observed wage (Observed)

74

0
0.02
0.04
0.06
0.08
0.1
0.12
3 4 5 6 7 8 9 10 11 12 13 14 15
Mothers' Care
L E S A
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
3 4 5 6 7 8 9 10 11 12 13 14 15
Market Care
L E S A
Figure 1.6: Relative Motivation of Different Investments

75

Figure 1.7: Mothers’ choice variable (Simulated)

76

Figure 1.8: Simulated demand for market care

77
Impacts of Introducing Air Quality Monitor-Alert System on
Health Behavior
6

2.1. Introduction
This study investigates individual health-related behavior in response to ambient air pollution
knowledge, specifically regarding particulate matter less than 10µm (PM10). The study examines
whether the government policy that provides objective, real-time information about air pollution
can intensify such behavior.
There have been some noteworthy studies on the impact of air pollution on individual health
behavior and health outcomes. The study most closely related here is by Neidell (2009). Using
data between 1989 and 1997 from Southern California, Neidell looks at the impact of the ozone
level and air quality information provided to the public in the form of a “Smog Alert.” As a
threshold needs to be reached in the ozone level for a Smog Alert, the study uses regression
discontinuity (RD) to look into two outcomes: self-restraint from outdoor activity as an avoidance
behavior; and asthma hospitalization as a health effect from the ozone.
For outdoor activities, Neidell (2009) finds a 12%-15% decrease in the number of Los Angeles
(LA) zoo visitors after a smog alert is issued, and a 3%-6% decrease in visitors to the Griffith
Observatory. This is smaller in magnitude since the Observatory is famous for its nightscape and
the ozone level drops after sunset. For asthma admissions, without controlling for Alerts and the
Pollutant Standard Index (PSI), Neidell finds that a 0.01ppm higher ozone level, results in a 1%-
2% increase in hospitalization in each age group. However, once Alerts and PSI are controlled,

6 Coauthored with Daniel Bennett and Jinkook Lee.

78
hospitalization increases by 2%-3%. This implies there is endogenous avoidance behavior caused
by the provision of information.
Another interesting study in the same context, but focusing on labor outcomes, is from Graff
Zivin and Neidell (2012) who look into the effect of the ozone level on the productivity of
individual farmworkers in California. They show that there is a significant negative impact from
the ozone level, demonstrating that air pollution not only has a negative impact on health outcomes,
such as mortality and morbidity, but also is likely to harm many aspects of life, each deserving
separate attention.
However, as mentioned in Neidell (2009), “Southern California has a very unique history of
smog and ozone, so the result may not necessarily generalize to other contexts.”
7
Thus, the impact
of information on avoidance behavior and health outcomes requires further scrutiny under different
contexts.
The context for this study is the region of Seoul, the metropolitan area in South Korea, between
2004 and 2016. The air pollution we focus on is PM10, which is a well-documented meteorological
phenomenon in South Korea due to westerly winds in the region. PM10 has been the subject of
substantial health concerns due to intensive industrialization in nearby China.
In an early study on this in the environmental literature, Kwon et al. (2002) investigate the
phenomenon of Asian Dust. Their study considers the level of Asian Dust as a proxy for overall
pollution. Although the results are not all significant, the results show that pollution levels

7
Neidell, M. (2009). Information, avoidance behavior, and health the effect of ozone on asthma hospitalizations.
Journal of Human resources, 44(2), Page 476

79
depending on different measures are all strongly correlated with Asian Dust. Additionally, it
suggests that Asian Dust is correlated with respiratory system related deaths.
One of the most recent studies that examines the consequences of PM10 specifically in South
Korea is by Jia and Ku (2015). The authors analyze the spillover effect of PM10 blown over to
South Korea by westerly winds. Using South Korea’s regional level data, their study consistently
shows a statistically significant positive spillover effect of Chinese air pollution on South Korea’s
respiratory and cardiovascular deaths, and approximates that PM10 causes an additional two to
three deaths, on average, per province each month.
Considering that PM10 has become a major concern for South Korea and even for North East
Asia in general, how air pollution impacts different aspects of individuals’ daily lives deserves
attention. As in Neidell (2009), this study extends the field of research that explores health impacts
due to PM10 by focusing on the individual response to such air pollution, such as avoidance
behavior, which is not addressed in studies looking solely into mortality.
Additionally, the Monitor–Alert system, a government initiative to aid in protecting citizens
from the effects of the air pollution, needs to be properly assessed. Although this study investigates
whether such government initiatives around air pollution induce and promote avoidance behavior
among the public, it is worth noting that the context of this analysis may differ from previous
comparable studies such as Neidell (2009).
First, the quality and accessibility of the information may matter. Neidell (2009) looks at data
between 1989 and 1997, and analyzes the impact of the Ozone Alert, which is based on a forecast
the day before, allowing enough time to disseminate the information through traditional media
such as the Los Angeles Times. However, this is not the case with PM10 intervention in Korea.

80
In this study, we will look at the period 2004 and 2016, when the internet and mobile phones
are now the common media used to disseminate information. The local governments we examine
utilized several channels, such as the internet and short message service (SMS), along with more
traditional media, such as local TV and radio, to provide real-time information to the public.
Hence, unlike in Neidell’s (2009) study, the Monitor–Alert system studied here, additional to
issuing forecasted alerts, provides actual, real-time measures of PM10 monitoring, and the alert is
based on that every hour.
However, whether this new technology and more reliable information enables information to
spread more efficiently and/or leads to more significant behavioral change is unconfirmed. While
Neidell (2009) studies the pollution forecasts and alerts printed in the Los Angeles Times as a
bundle, via the new technologies, individuals need to pay additional costs for access to such
information; such as visiting the websites or applying for SMS notification service. If such implicit
cost is not negligible, people may pay less attention to or be less aware of the pollution information.
Second, unlike the ozone level, which Neidell (2009) claims to have at least a correlation with
visible smog, PM10 and other pollutants might have higher correlation with other visible pollutants,
such as Asian Dust, which is not the major health concern itself if not for other pollutants that are
brought by Asian Dust. By analyzing the avoidance behavior of PM10, we can distinguish the size
of the impact depending on the level of additional objective information.
The contribution of this study is twofold. First, additional to mortality and morbidity analyses
in existing literature, we assess the individual avoidance behavior related to PM10, which is
relatively visible to the naked eye. More importantly, we assess the efficiency and efficacy of the
Monitor–Alert system, one of the most commonly used policies to alleviate the impact of air
pollution, heavily dependent on new media, the internet, and mobile technology usage. While these

81
new channels for alert dissemination may be efficient in terms of their cost to the government, we
hope to be able to measure their efficiency in terms of the benefit to the public.
The rest of this chapter is organized as follows. Section 2 discusses the historical background
of this government policy that allows us to develop the study strategy. In section 3, we discuss the
different sources of data and outline the data structure. Section 4 formally presents the estimation
method, followed by the results in Section 5. Section 6 concludes.

2.2. Background
2.2.1. Government policy introduction
Our study looks at the Seoul metropolitan area located in the northwest region of South Korea.
With a population of nearly 26 million, the area includes the cities of Seoul and Inchon in the
center of Gyung-gi province. While Seoul and Inchon, introduced city-wide PM10 alert systems in
February 2005 and March 2008, respectively, Gyung-gi initiated an alert system in 2007 in only
25 of the 31 cities in the province. It gradually expanded this system to all 31 cities by 2011. This
context can be exploited in our difference-in-difference response identification strategy discussed
in the next section.
The introduction of the PM10 monitor–alert system
8
was initially done not by the central
government, but by local governments as the first law for a national level PM10 monitor–alert
system passed congress in July 2013. After this legislation, in August 2013, the central government
enacted the policy in Seoul and Gyung-gi province as a pilot program since these provinces already
had relevant infrastructures established by their local governments. The rest of the country was

8
Please note that throughout this chapter, the terms “Monitor–Alert system,” Monitor system,” and “Alert system”
are used interchangeably.

82
introduced to the same program in December 2014. However, some regions in the country also
already had similar Monitor–Alert systems set by their local governments in place, such as Daegu
city, which introduced its alert system in 2011.
However, it is possible that the introduction of an alert system can be correlated with observed
or potentially unobserved characteristics. From the examples discussed, Seoul, Inchon, and
Gyung-gi area are located in the northwest of South Korea, an area significantly affected by Asian
Dust from China, while Daegu is located in the southeast of South Korea and less affected by
Asian Dust. The implication is that the introduction of the alert system may be a consequence of
differences in regional backgrounds such as geographical location, industrial composition, or
perceptions of citizens. Additionally, it is not well documented which of the 226 municipalities
introduced the alert system at which point in time, and at what intensity. Therefore, this study
limits it scope based on data availability to the Seoul metropolitan area, since this is an area
relatively homogeneous in terms of unobserved characteristics, and still allows us to exploit the
fine municipality level policy variations for our response behavior identification strategy.
Although this study focuses mainly on PM10, and not on the ozone, the expansion of the Ozone
Alert System is worth mentioning since it aligns with the rollout of the PM10 Alert System and
implies that the introduction of the PM10 Monitor–Alert system is, at least not solely, based on the
PM10 level itself. While PM10 has become a major concern to the government and to the public
relatively recently, the alert system for the ozone has a longer history in Korea, specifically in the
Seoul metropolitan area. The Ozone Alert system in Seoul city started in 1995 and in Inchon in
1996, followed by Gyung-gi province in 1997. However, while Seoul and Inchon introduced these
alerts across all the districts within their city boundaries, the province set up the system for only
seven cities in 1997. As in the case of the PM10 Monitor-Alert system, this was gradually expanded

83
to all 31 cities over the following 13 years. The schedule of expansion is presented in Appendix
Table 2.1. From 2007 onwards, the expansion schedules for the PM10 alert system and for the
Ozone Alert system aligned, and 25 cities with the Ozone Alert system in 2007 all initiated a PM10
Alert system that year. As a result, we take a conservative approach in our analysis since 1) the
region for the PM10 Monitor-Alert system is quite homogeneous; and 2) an endogenous
introduction schedule does not necessarily directly or solely correlate with historical PM10 levels.

2.2.2. PM10 Monitor-Alert System
In this subsection, we discuss the components of the PM10 Monitor–Alert system policy, which
has been heretofore referenced rather ambiguously without detailed explanation. Although the
system is referenced as the PM10 Monitor–Alert system, it consists of both a forecast system and
real-time information on the pollution level. However, the alert details might vary by provincial
government and change throughout time as well; for example, based on how strongly a mayor may
recommend that school principals restrain students from outdoor activities. However, the essential
intent of introducing the Monitor-Alert system package is understood to be that of providing
objective information to the public on the level of pollution, with only minimal direct intervention
from the government on this.
Based on the monitored air pollutants from a ground weather station, forecasts of daily average
PM10 are announced twice every 24 hours, once in a forecast of the following day and once at 9am
of the day. The forecasted level of pollution is announced to the public, along with color codes
that distinguish how bad the pollution may be; for example, in 2007, Gyung-gi province issued
color codes for six different levels: good (0-50μg/m
3
), normal (51-100μg/m
3
), bad for the sensitive
population (101-150 μg/m
3
), slightly bad (151-200 μg/m
3
), bad (201-300μg/m
3
), and very bad

84
(301μg/m
3
or over). Such information is spread to the public through a website, local media, and
outdoor electronic board displays, lower level government bodies and educational administrations,
and SMS to individuals. The forecasts for the following day are “correct” 57-65%, while same-
day forecasts have higher “correct” ratings of 73-75%, depending on the meteorological model
used.
Real-time hourly PM10 levels are also available through the same media, as well as actual
warnings or alerts based on these. There have been some changes in alert trigger levels of PM10
over time. For example, Gyung-gi province’s threshold for issuing a warning had been 200μg/m
3

for a consecutive two-hour average, and 300μg/m
3
for alert issuance until 2014. The threshold for
warning/alerts added another 24 hour criterion in 2015, and abolished it again in 2016, lowering
the 2 hours criteria to 150μg/m
3
for warnings and 300μg/m
3
for alerts. The full history of
warning/alert criteria is presented in Table 2.1.
The PM10 alerts have been issued based on entire regions, which means alerts have been issued
to all cities in the region as soon as at least one of the cities hits the threshold. While the pollution
level itself is available from the websites and outdoor displays, the public is notified of the issuance
of the alerts through channels similar to forecast announcements.
However, it should be noted that few alerts have been issued. Table 2.1, summarizes the actual
number of PM10 warnings or alerts issued yearly in Gyung-gi province. It is obvious, probably due
to the strict criteria for the warning/alert system, that the benefit to the public through the actual
warning/alerts may have been minimal. However, aside from the alert issuance, the introduction
of the Monitor–Alert system as a “package” might have an effect and benefit the public since
forecasts became available at a local level and, more importantly, allowed the public to access
information about real-time pollution levels.

85
In the study, the impact of introducing the Monitor–Alert system should not be mistaken for the
impact of the actual warning/alert. Rather, the study investigates the impact of the accessibility of
the information on avoidance behavior.
2.3. Data
In this subsection, the three sources of the main datasets used in this study are detailed. The
most crucial data required in the study are the air pollution levels measured by each ground
monitor. There are two different daily datasets and some hourly measuring PM10, which are
provided in a panel structure along with other air quality indices, such as O3, PM2.5, NO2, SO2, and
CO, which may be controlled in our analyses. This information comes from four different
institutions. The Korean Meteorological Administration (KMA)
9
provides nationwide data, which
cover the time between 2001 and 2016, while the Provincial Offices provide documentation for an
additional period: Seoul (SL) (1987-2018), Inchon (IC) (1997-2013), and Gyung-gi (GG) (2003-
2018). Most importantly, city level PM10 is provided by the major weather station. Therefore, from
KMA, we only use the meteorological information as a control variable, which is our first dataset.
Fortunately, each provincial government operates its own Environmental Research Institution,
and these three (SL, IC, and GG) are able to provide hourly, or at least, daily, data estimated from
each ground monitoring station, along with the geographical locations of all monitoring stations.
This is the second dataset which allows us to construct the main explanatory variables of interests.
The third dataset is the monthly number of visitors provided by the Korean Culture & Tourism
Institute (KCTI)
10
. While Neidell (2009) investigates only three tourist spots to assess avoidance
behavior, our dataset includes 421 venues within the Seoul Metropolitan Area, along with location,

9
https://data.kma.go.kr/data/ The data are publicly available from http://air.gg.go.kr/airgg/air_data/
10
http://www.kcti.re.kr/ The data are publicly available from https://tour.go.kr/

86
entrance fee, and type of venue. Additionally, the number of visitors is categorized into foreigners
and domestic Koreans, which allows us to analyze differences in the size of the effect depending
on possible accessibility to air quality information. This is analyzed using Greater Los Angeles
Zoo Association membership as a proxy in Neidell (2009). By analyzing type of venues, we can
revisit Neidell’s (2009) result that Griffith Park suffers less from avoidance behavior, since it is
more popular at night, thus less correlated with ozone. Although the number of visitors is available
for 1990-2017, the data before 2004, the year KCTI notes that data collection was renewed, seemed
highly unreliable, thus we did not use the data in this study.
The following Figures, 2.1 to 2.3, depict the PM10 levels in 2004-2017, the period of interest for
avoidance behavior. Figure 2.1 shows no huge gaps at the municipality (district) level (City of
Gyung-gi Province, Districts of Seoul and Incheon). The high PM10 level in the northern Gyung-
gi province is slightly notable due to low efficiency fossil fuel manufacturers in North Korea
11
.
Interestingly, there is no clear trend over time, as shown in Figure 2.2. However, Figure 2.3 shows
a clear seasonality in PM10 level, which should be considered in further analyses.

2.4. Estimation Strategy
The main response behavior identification is based on the differences in the points of time when
the Monitor–Alert system was introduced in the areas in the region.
We use the following specification to assess avoidance behavior.

11
https://imnews.imbc.com/replay/2018/nwdesk/article/4568746_30181.html

87
(𝑙𝑙𝑙𝑙 ) 𝑉𝑉 𝑖𝑖𝑠𝑠𝑖𝑖 𝑡𝑡𝑓𝑓 𝑓𝑓 𝑠𝑠 𝑘𝑘 𝑐𝑐 𝑝𝑝 𝑘𝑘 𝑚𝑚 = 𝛽𝛽 0
+ 𝛽𝛽 1
(𝑙𝑙𝑙𝑙 )𝑃𝑃𝑃𝑃 10
𝑘𝑘𝑘𝑘 𝑚𝑚 × 𝐴𝐴 𝑙𝑙𝑒𝑒𝑓𝑓𝑡𝑡 𝑆𝑆𝑆𝑆 𝑠𝑠 𝑡𝑡𝑒𝑒 𝑚𝑚 𝑐𝑐 𝑘𝑘𝑚𝑚
+ 𝛽𝛽 2
(𝑙𝑙𝑙𝑙 )𝑃𝑃𝑃𝑃 10
𝑘𝑘𝑘𝑘 𝑚𝑚 + 𝛽𝛽 3
𝐴𝐴 𝑙𝑙𝑒𝑒𝑓𝑓𝑡𝑡 𝑆𝑆𝑆𝑆 𝑠𝑠 𝑡𝑡𝑒𝑒 𝑚𝑚 𝑐𝑐 𝑘𝑘𝑚𝑚
+ 𝛽𝛽 4
𝐶𝐶 𝑓𝑓 𝑙𝑙 𝑡𝑡𝑓𝑓𝑓𝑓 𝑙𝑙𝑠𝑠 & 𝐹𝐹 𝐸𝐸𝑠𝑠 𝑘𝑘 𝑐𝑐 𝑝𝑝 𝑘𝑘 𝑚𝑚 + 𝜀𝜀 𝑘𝑘 𝑐𝑐 𝑝𝑝 𝑘𝑘 𝑚𝑚 (2.1)

where the subscripts k, c, 𝑣𝑣 , y, and m represent the entire region, the city (district), the venue, the
year, and the month, respectively. Note that (𝑙𝑙𝑙𝑙 )𝑃𝑃𝑃𝑃 10
𝑘𝑘𝑘𝑘 𝑚𝑚 is the log of the monthly PM10 level
averaged over the entire region instead of at city level. This is to mitigate any possible bias from
the endogenous process of the Monitor–Alert system rollout, which is not controlled.
𝑉𝑉 𝑖𝑖𝑠𝑠𝑖𝑖𝑡𝑡𝑓𝑓 𝑓𝑓𝑠𝑠 𝑘𝑘 𝑝𝑝 𝑘𝑘 𝑚𝑚 represents the monthly number of visitors to the venue; we also run separate
regressions for domestic and foreign visitors where data are available. Both of these are taken as
logs.
The 𝐴𝐴 𝑙𝑙𝑒𝑒𝑓𝑓𝑡𝑡 𝑆𝑆𝑆𝑆 𝑠𝑠 𝑡𝑡𝑒𝑒 𝑚𝑚 𝑐𝑐 𝑘𝑘𝑚𝑚
is a binary variable that takes the value of one if city 𝑐𝑐 has a Monitor–
Alert system and zero otherwise. For the main specifications, we control the MonthXVenue fixed
effect, where month is included to control for different seasonality for different types of venues,
and the year fixed effect for the time trend.
For the control, three sets of additional variables are considered. 1) Average PM10 for the whole
Gyung-in metropolitan area; 2) Monthly average of meteorological variables—temperature and its
square, humidity, atmospheric pressure, water vapor pressure, precipitation, wind speed, sunlight,
and ground surface temperature—in North Gyung-gi, South Gyung-gi, Seoul, and Incheon; 3)
Regional average of other pollutants: Ozone, SO2, and NO2. Using an average PM10 level for the
metropolitan area is important for our analysis, since there is the possibility of a substitution pattern

88
across the region. Since we are running regression analyses with venues that are heterogeneous,
standard errors are clustered at month and venue level.
The coefficient of interest is 𝛽𝛽 1
, which is expected to be negative if the public acquires better
information about air quality through the Monitor-Alert system provided by the government and
there is avoidance behavior on a bad day. 𝛽𝛽 2
represents the avoidance behavior that already exists
due solely to the visibility of PM10, even without additional information. If such avoidance
behavior does exist before the Alert system, we can expect 𝛽𝛽 2
to be negative. 𝛽𝛽 3
captures the
behavior of individuals on a good day, but with the Monitor–Alert system in place. Therefore, if
individuals are risk-averse agents, by providing them with the certainty of better air quality,
precautionary avoidance behavior may be limited, thereby encouraging the public to engage in
more activities. If such benefit does exist, we may find 𝛽𝛽 3
to be positive.
In addition to the main analysis above, the corresponding event study has also been conducted.
The main purpose of this exercise is to determine whether the government’s decision to introduce
the policy may have been endogenous. The following specification is used for this analysis.

(𝑙𝑙𝑙𝑙 ) 𝑉𝑉 𝑖𝑖𝑠𝑠𝑖𝑖𝑡𝑡𝑓𝑓 𝑓𝑓 𝑠𝑠 𝑘𝑘 𝑝𝑝 𝑘𝑘 𝑚𝑚 = 𝛼𝛼 0
+ � 𝛼𝛼 1
𝑙𝑙 [(𝑙𝑙𝑙𝑙 )𝑃𝑃𝑃𝑃 10
𝑘𝑘𝑘𝑘 𝑚𝑚 × 𝐷𝐷 𝑘𝑘𝑘𝑘 𝑚𝑚 𝑙𝑙 ]
1 2
𝑙𝑙 = − 1 2
+ 𝛼𝛼 2
(𝑙𝑙𝑙𝑙 )𝑃𝑃𝑃𝑃 10
𝑘𝑘𝑘𝑘 𝑚𝑚 + � 𝛼𝛼 3
𝑙𝑙 𝐷𝐷 𝑘𝑘𝑘𝑘 𝑚𝑚 𝑙𝑙 1 2
𝑙𝑙 = − 1 2
+ 𝛼𝛼 4
𝐶𝐶 𝑓𝑓 𝑙𝑙 𝑡𝑡𝑓𝑓𝑓𝑓 𝑙𝑙𝑠𝑠 & 𝐹𝐹 𝐸𝐸𝑠𝑠 𝑘𝑘 𝑝𝑝𝑘𝑘𝑚𝑚 + 𝜀𝜀 𝑘𝑘 𝑝𝑝 𝑘𝑘 𝑚𝑚 (2.2)

Subscripts 𝑏𝑏 , 𝑣𝑣 , 𝑆𝑆 , and 𝑚𝑚 are defined the same as in (2.1), and superscript 𝑙𝑙 represents the time
difference from the first introduction of the Monitor-Alert system. 𝐷𝐷 𝑘𝑘𝑘𝑘 𝑚𝑚 𝑙𝑙 is an indicator variable

89
that takes a value of 1 if and only if month 𝑚𝑚 of year 𝑆𝑆 is the 𝑙𝑙 th month after (or before) the
implementation of the Monitor-Alert system.
While we do not find evidence of endogenous decision of program introduction from this event
study, the result shows the impact of a Monitor-Alert system was not clearly noticeable. Therefore,
we do not discuss the event study in greater detail in the next subsection. The interested reader can
find our estimates of 𝛼𝛼 1
𝑙𝑙 and 𝛼𝛼 3
𝑙𝑙 in the Appendix Figure 2.2.

2.5. Result
Table 2.2 shows the main results of the regression analyses from equation (2.1), not
distinguishing types of venues. Each column shows the results controlling different sets of
variables. While the signs are robust regardless of the variables, there are some changes in
magnitude once metropolitan level average PM10 is controlled; however, there is scant change
from controlling the additional variables. Thus, column (2) is the focus of interpretation.
Starting with the estimate for 𝛽𝛽 3
, having a Monitor–Alert system seems to increase substantially
the number of visitors, by 80%, on average. As discussed in Section 2.4, this may be due to the
fact that providing additional certainty about the air quality alleviates precautionary avoidance
behavior that existed before the system. While such a benefit to the public is easily overlooked,
the magnitude of the estimate implies that this could be an indirect benefit that should not be
disregarded.
Looking at 𝛽𝛽 2,
which corresponds to the effects of a high PM10 level, we do not find concrete
evidence of clear avoidance behavior based on the visibility of PM10. The estimates fluctuate
significantly depending on the specification. Moreover, they are not statistically different from
zero under the most conservative criterion.

90
The parameter of our ultimate interest, 𝛽𝛽 1
, shows there is clear avoidance behavior responding
to a higher PM10 once the Monitor-Alert system is in place. For a 1% increase in the average
monthly PM10 level, the number of average monthly visitors to the tourist destination falls by
0.14%.
While the main result in Table 2.2 aligns with the intuition previously discussed, the
implications of these results deserve further exploration since they only show an average impact
of PM10 and the Monitor-Alert system without considering the characteristics of the tourist
destinations.
Fortunately, KCTI categorized the venues(RES) as follows : Category (CTGRY), Division
(DVSN), and Section (SCTN). Although this categorization does not exactly match the study
characteristics of interest, some insights about the nature of the venues are provided. For example,
the Mountains and Flatlands category would include more outdoor attractions and their visitors
are more likely to be engaged in physical activities. In contrast, the number of such venues under
the Accommodation/Food Category, which includes hotels and restaurants, may not be as
susceptible to air pollution.
To consider such heterogeneity, Table 2.3 presents results from the analyses of the different
subgroups at the DVSN level. Out of 19 DVSNs, we chose three most likely to be outdoor facilities
and three most likely to be indoor ones. Columns (1) to (3) are likely to be outdoors, while columns
(4) to (6) are likely to be indoors.
The obvious implication from Table 2.3 is that we should not rule out the possibility of
avoidance behavior patterns that differ depending on the characteristics of the tourist attractions.
Outdoor venues are likely to see fewer visitors when there is information of bad air condition.
Sports Facilities are likely to lose 0.27% of their visitors for every 1% increase in the PM10 level.

91
Waterside and Marine related tourist spots, such as piers or ports, also see a decrease of 0.73% in
the number of visitors, and the result is statistically significant at 5%, despite the small sample
size. This contrasts with the results for the indoor venues. These experience a non-significant
increase in the number of visitors when there is information about high PM10 levels, despite the
relatively larger sample size.
Furthermore, the results agree with the possibility of cross-sectional substitution. That is, with
bad news about air quality, there is a notable drop in the number of visitors to outdoor venues.
This seems to drive the negative coefficient for the interaction term in Table 2.2. In contrast, there
are relatively few indoor venues which experience anything other than an increase in the number
of visitors. Adding to the dynamic substitution in avoidance behavior proposed in Zivin and
Neidell (2009), our results also support the possibility of substitution.
Table 2.3 calls for further research into finer category definitions. In Table 2.4, we present
results from the subsamples of the finest subcategory, SCTN. We pick two specific subgroups of
venues from the Sports Facilities DVSN, Golf Courses and Ski Resorts. These tourist destinations
are clearly outdoor venues and involve physical activity for visitors. In addition, these two reflect
the most obvious avoidance behavior from the PM10 information under the Monitor–Alert system.
The avoidance behavior is strong for these two specific types of venues. Columns (1) and (4)
show the changes in total number of visitors to each venue, regardless of visitor nationality. Given
an Alert revealing a 1% increase in the PM10 level, the number of total visitors to the Golf Courses
drops by around 0.24%, and the Ski Resorts show a 0.99% decrease for the same percentage
change in PM10. Considering the seasonality of PM10 presented in Figure 2.3, it is not surprising
that Ski Resorts see a more dramatic impact than Golf Courses.

92
However, notably, in other SCTNs that fall under the same Sports Facilities DVSN, no such
obvious avoidance pattern is found, as most are statistically insignificant despite having larger
sample sizes than that for Ski Resorts; these include Swimming Pools, Gyms, and Snow Sled
Resorts.
One possible explanation for this is the heterogeneity in the characteristics of potential visitors.
As mentioned in the introduction in discussing the distinction between this study and that of
Neidell’s (2009), the Monitor–Alert system of interest here disseminates the information mainly
through new media, such as the internet and mobile sources. Since visitors to Golf Courses and
Ski Resorts are among those coming from relatively better socio-economic backgrounds, they are
likely to have easier access to such information for many reasons. They may have better skills to
access information on the internet because of their educational background or have more
experience. They may have heard of this small-scale city level alert through word of mouth because
of their social networks. They may be early adopters of smartphones and are able to have full
access to information on the mobile internet.
The above represents one possible explanation, but there may be others. Table 2.4 presents a
result that is consistent with this potential mechanism driving these differences. Columns (2) and
(3) show the impact on the number of domestic Korean visitors and foreign visitors, respectively,
in contrast to the total number of visitors to Golf Courses presented in column (1). Similarly,
columns (5) and (6) show domestic visitors and foreigners, respectively, to Ski Resorts.
Comparing the impact of the PM10 level given in the Monitor–Alert system on domestic visitors
versus foreign ones, the comparison shows that the group that exercises intensive avoidance
behavior is domestic visitors rather than foreign ones.

93
Figure 2.4 visualizes this pattern for domestic Korean visitors to golf courses. Without the
Monitor–Alert system, the drop in the number of visitors on bad months, defined as monthly
average PM10 over the median level (60μg/m
3
), is not notable, while the drop is notable once the
alert system is introduced. However, Figure 2.5 shows that the drop in the bad months is not
signified by the introduction of Monitor–Alert system. The main beneficiary of the Monitor–Alert
system are domestic Koreans who are likely to have full access to the information provided by
local governments.
There may be alternative explanations based on the benefit in contrast to the explanation based
on access to the information. As proposed by Neidell (2009), one possibility may be that tourists
are likely to visit these tourist attractions in spite of bad air quality, given their limited time in the
area. Similarly, potential customers of Golf Courses and Ski Resorts might value their health at a
higher level than those from lower socio-economic backgrounds, thus exercising avoidance more
intensively.
Compared with Neidell (2009), the larger heterogeneity in avoidance behavior found in this
study may be a consequence of differences in both costs and benefits. Thus, this study adds another
layer to the existing argument based solely on the benefit of the Alerts. Therefore, the mechanism
that drives the difference in avoidance behavior is still inconclusive and requires further
exploration in future studies.

2.6. Conclusion
This study shows that providing the public with accurate real-time information about ambient
air pollution, PM10, effectively induces and promotes avoidance behavior. However, the fact that
the number of visitors to indoor venues is not affected, implies such avoidance behavior is

94
selectively decided, and the Alert policy might affect different individuals differently. In addition,
we found that there were more visitors in general once the Monitor–Alert system was introduced,
regardless of air pollution. Therefore, our study also demonstrates how the government policy to
help counteract possible health consequences from air pollution might have unexpected benefits.
Our more interesting finding is that the intensity of such avoidance behavior may differ not only
by the characteristics of the venues, but also by the potential users of these venues; Golf Courses
and Ski Resorts, which potentially have customers from better socio-economics backgrounds were
found to lead our empirical support for avoidance behavior. Considering how the main channel for
information is through the internet and other mobile technology, this may be evidence that the
monitoring system in our context has not been effective for those unable to receive this information
based on a higher access cost. This explanation aligns with the finding that visits from foreigners
did not drop.
However, there still is a competing explanation. It is possible that such discrepancy in avoidance
behavior is not a result of difference in information accessibility, but of the benefit of the
information; that is, individuals with higher socio-economic status may value their health at a
higher level.
It is still ambiguous which of the two aforementioned hypotheses explain the heterogenous
response of the public, and it is most likely both aspects affect individuals’ decisions
simultaneously. While further studies have to be conducted to disentangle these suggested
mechanisms, our study indicates the possibility of an implicit cost to the public where the
infrastructure does not properly support the government intention, which, in turn, limits the benefit
of even the most commonly applied policy.

95
2.7. Tables and Figures

Table 2.1: Number of Warning/Alert Issued Yearly
year warning[alerts] Threshold condition for Warning [Alert] (μg/m3) issuance
2007 0[0]
200 for 2hrs [300 for 2hrs]
2008 5[0]
2009 3[0]
2010 4[0]
2011 0[0]
2012 0[0]
2013 1[0]
2014 1[0]
2015 48[0] (MA)120 in 24hr or 200 for 2hr [(MA)250 in 24hr or 400 for 2hr]
2016 25[2]
150 for 2hrs [300 for 2hrs]
2017 40[4]

96
Table 2.2: Avoidance Behavior-All types of venues

(1) (2) (3) (4)
VARIABLES (log)Visitors (log)Visitors (log)Visitors (log)Visitors

Alert system 1.318*** 0.832*** 1.090*** 1.064***

(0.305) (0.287) (0.289) (0.284)
PM10
-0.140* 0.141 0.236 0.331**

(0.079) (0.158) (0.157) (0.165)
Alert system X PM10
-0.273*** -0.139** -0.197*** -0.189***

(0.073) (0.069) (0.069) (0.068)
Metro_PM10

0.002 0.002 0.001

(0.002) (0.002) (0.003)

Observations 35,221 35,221 35,221 35,221
R-squared 0.440 0.470 0.471 0.471
Mean of dep var 32279.04 32279.04 32279.04 32279.04
Mean PM10 level 56.18 56.18 56.18 56.18
Proportion of alertsys 0.86 0.86 0.86 0.86
Controls None 1 1 2 1 2 3
FE Year SCTNxMonth Year SCTNxMonth Year SCTNxMonth Year SCTNxMonth
SubGroup All All All All
Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

97
Table 2.3: Avoidance Behavior-by DVSN Subgroups
(1) (2) (3) (4) (5) (6)
Outdoor Indoor
VARIABLES
(log)
Visitors
(log)
Visitors
(log)
Visitors
(log)
Visitors
(log)
Visitors
(log)
Visitors

Alert system 0.798 3.205** 1.213*** -4.332 -0.553 -1.673*

(1.445) (1.521) (0.232) (2.827) (0.532) (0.905)
PM10 0.607 1.709 0.399*** -2.933* -0.013 -0.431

(0.882) (1.113) (0.117) (1.527) (0.178) (0.429)
Alert system X
PM10 -0.165 -0.733** -0.276*** 0.989 0.094 0.395*

(0.353) (0.345) (0.053) (0.649) (0.126) (0.220)
Metro_PM10 -0.007 -0.018 0.000 0.038** 0.001 0.001

(0.012) (0.017) (0.002) (0.017) (0.003) (0.007)

Observations 407 387 9,602 270 8,604 2,018
R-squared 0.976 0.893 0.842 0.933 0.925 0.871
Mean of dep var 36154.95 48818.23 10327.10 108770.22 22035.07 8724.49
Mean PM10 level 53.39 55.61 56.80 59.05 54.97 57.76
Proportion of
Alert system 0.86 0.82 0.83 0.81 0.89 0.74

FE
Year
RESxMonth
Year
RESxMonth
Year
RESxMonth
Year
RESxMonth
Year
RESxMonth
Year
RESxMonth

SubGroup(DVSN)
Mountain
and Flatland
Waterside
and Marine
Sports
Facilities
Shopping Exhibition
Accommoda
tion / Food
Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

98
Table 2.4: Avoidance Behavior-by SCTN Subgroups
(1) (2) (3) (4) (5) (6)
Golf Course Ski Resorts
VARIABLES
(log)
Visitors
(log)
Domestic Visitors
(log)
Foreign Visitors
(log)
Visitors
(log)
Domestic Visitors
(log)
Foreign Visitors

Alert system 1.074*** 1.083*** 0.009 4.242** 4.411** -1.070

(0.224) (0.226) (0.786) (2.022) (2.012) (3.445)
PM10 0.298** 0.299** 0.020 1.750** 1.566** -4.772**

(0.119) (0.119) (0.366) (0.718) (0.703) (1.813)
Alert system X
PM10
-0.243*** -0.244*** -0.060 -0.987** -1.033** 0.232

(0.051) (0.051) (0.192) (0.450) (0.451) (0.733)
Metro_PM10 0.002 0.002 -0.001 -0.021 -0.019 0.079***

(0.002) (0.002) (0.006) (0.013) (0.014) (0.024)

Observations 8,695 8,650 4,151 279 267 139
R-squared 0.822 0.820 0.691 0.822 0.819 0.919
Mean of
Dependent Var
8622.45 8594.31 34.89 19380.07 17998.16 3053.71
Mean PM10
level
56.95 56.99 57.15 58.33 58.75 61.68
Proportion of
Alert system
0.83 0.83 0.77 0.77 0.76 0.73

FE
Year
RESxMonth
Year
RESxMonth
Year
RESxMonth
Year
RESxMonth
Year
RESxMonth
Year
RESxMonth

Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

99

Figure 2.1 : Average PM
10
by GoonGu(Municipality)

100

Figure 2.2 : Average PM
10
by Year
Figure 2.3 : Average PM
10
by Month

101
Figure 2.4: Golf Course- Domestic Visitors against PM10 Level

Figure 2.5: Golf Course- Foreign Visitors against PM
10
Level

102
2.8. Appendix
Appendix Table 2.1: Expansion of Ozone Alert in Gyung-gi Province
Year
Number of
Cities and
Clusters
List of Cities with Alert System Establish
Newly
Introduced
1997
Schedule of O3
Alert system
Introduction
7 cities Suwon, Sungnam, Uijeongbu, Anyang, Buchun, Gwangmyung, Ansan

1 99 8 9 cities
Suwon, Sungnam, Uijeongbu, Anyang, Buchun, Gwangmyung, Ansan, Gwacheon,
Guri
Gwacheon, Guri
1999 9 cities
Suwon, Sungnam, Uijeongbu, Anyang, Buchun, Gwangmyung, Ansan, Gwacheon,
Guri

2000 12 cities
Suwon, Sungnam, Uijeongbu, Anyang, Buchun, Gwangmyung, Ansan, Gwacheon, Guri,
G oya ng, G oo npo, S hi h e ung
Goyang,
Goonpo,
Shiheung
2001 14 cities
Suwon, Sungnam, Uijeongbu, Anyang, Buchun, Gwangmyung, Ansan, Gwacheon, Guri,
G oya ng, G oo npo, S hi h e ung , P y u n g taek , U iw a n g
Pyungtaek,
Uiwang
2002 15 cities
Suwon, Sungnam, Uijeongbu, Anyang, Buchun, Gwangmyung, Ansan, Gwacheon, Guri,
G oya ng, G oo npo, S hi h e ung , P yun gt a e k, U i w a n g, N a m y a ngj u
Namyangju
2003 19 cities
Suwon, Sungnam, Uijeongbu, Anyang, Buchun, Gwangmyung, Ansan, Gwacheon, Guri,
G oya ng, G oo npo, S hi h e ung , P yun gt a e k, U i w a n g, N a m y a ngj u , Y o ngi n, G i m po , H a na m , O s a n
Yongin, Gimpo,
Hanam, Osan
2004 19 cities
Suwon, Sungnam, Uijeongbu, Anyang, Buchun, Gwangmyung, Ansan, Gwacheon, Guri,
G oya ng, G oo npo, S h i h e ung, P yun gt a e k, U i w a ng, N a m ya ngj u, Y ongi n, G i m po, H a na m , O s a n

2005 22 cities
Suwon, Sungnam, Uijeongbu, Anyang, Buchun, Gwangmyung, Ansan, Gwacheon, Guri,
G oya ng, G oo npo, S hi h e ung , P yun gt a e k, U i w a ng, N a m ya ngj u, Y ongi n, G i m po, H a na m , O s a n,
H w a ns ung, I c he on, P a j oo
Hwansung,
Icheon, Pajoo
2006 23 cities, 8 clusters
Suwon, Sungnam, Uijeongbu, Anyang, Buchun, Gwangmyung, Ansan, Gwacheon, Guri,
G oya ng, G oo npo, S hi h e ung , P yun gt a e k, U i w a ng, N a m ya ngj u, Y ongi n, G i m po, H a na m , O s a n,
H w a ns ung, I c he on, P a j oo, P ocheon
Pocheon
2007
Scheduel of
PM10 and O3
Alert system
Introduction
25 cities
4 clusters
Suwon, Sungnam, Uijeongbu, Anyang, Buchun, Gwangmyung, Ansan, Gwacheon, Guri,
G oya ng, G oo npo, S hi h e ung , P yun gt a e k, U i w a ng, N a m ya ngj u, Y ongi n, G i m po, H a na m , O s a n,
Hwan s ung, I c he on, P a j oo, P oc he o n, Y a n gj u, D ongdu c he o n
Yangju,
Dongducheon
2008
26 cities,
4 clusters
Suwon, Sungnam, Uijeongbu, Anyang, Buchun, Gwangmyung, Ansan, Gwacheon, Guri,
G oya ng, G oo npo, S hi h e ung , P yun gt a e k, U i w a ng, N a m ya ngj u, Y ongi n, G i m po, H a na m , Osan,
H w a ns ung, I c he on, P a j oo, P oc he o n, Y a ngj u, D ongdu c he o n, G w a ng j u
Gwangju
2009
27 cities,
4 clusters
Suwon, Sungnam, Uijeongbu, Anyang, Buchun, Gwangmyung, Ansan, Gwacheon, Guri,
G oya ng, G oo npo, S hi h e ung , P yun gt a e k, U i w a ng, N a m ya ngj u, Y ongi n, G i m po, H a n a m , O s a n ,
H w a ns ung, I c he on, P a j oo, P oc he o n, Y a ngj u, D ongdu c he o n, G w a ng j u, A ns ung
Ansung
2010
27 cities,
4 clusters
Suwon, Sungnam, Uijeongbu, Anyang, Buchun, Gwangmyung, Ansan, Gwacheon, Guri,
G oya ng, G oo npo, S hi h e ung , P yun gt a e k, U i w a ng, N a m ya ngj u, Y ong i n, G i m po, H a n a m , O s a n,
H w a ns ung, I c he on, P a j oo, P oc he o n, Y a ngj u, D ongdu c he o n, G w a ng j u, A ns ung

2011
31 cities,
4 clusters
Suwon, Sungnam, Uijeongbu, Anyang, Buchun, Gwangmyung, Ansan, Gwacheon, Guri,
G oya ng, G oo npo, S hi h e ung , P yun gt a e k, U i w a ng, N a m ya ngj u, Y ongi n , G i m po, H a na m , O s a n,
Hwansung, Icheon, Pajoo, Pocheon, Yangju, Dongducheon, Gwangju, Ansung, Yeoju,
Y e o nc he o n, G a py e o ng , Y a ngpye o ng
Yeoju,
Yeoncheon,
Gapyeong,
Yangpyeong

103

2007.6
2008.1
2008.3
2009.1
2011.1
2005.2
Appendix Figure 2.1: Partition of Seoul Metropolitan Area

104

Appendix Figure 2.2: Impact of Monitor-Alert system Introduction Event

105
Fueling the engines of liberation with cleaner cooking fuel:
Evidence from Indonesia
12

3.1. Introduction
Women held back from participating in productive market activities is human capital wasted. It
is now well-established that the difference in rates of female labor force participation (FLFP) is an
important explanation behind the persistent differences in GDP per capita across countries (Bloom
et al. (2009)). Despite this, females form a little more than a third of the formal labor force of the
world with their participation rates ranging from as low as 6% in Yemen to as high as 84% in
Rwanda and Madagascar (World Bank Indicators, 2018). What explains these large differences in
FLFP across countries?
Previous research has suggested several factors, including the desirability of the jobs avail-able,
medical and production technology, discrimination, availability of childcare, and cultural attitudes,
affect FLFP.
13
While it is likely that a combination of factors are driving these differences, one
potential explanation that not received enough attention in the context of developing countries is
that of “engines of liberation” Greenwood et al. (2005). The emergence of cheap, time-saving
household technology has often been credited with liberating women from the burden of household
responsibilities and facilitating their integration into the labor force (Cutler et al. (2003); Goldin
(2006); Aguiar and Hurst (2007); de V. Cavalcanti and Tavares (2008)). But there is only limited
evidence on the liberating effect of such technology in developing countries and household

12
Coauthored with Tushar Bharati and Yiwei Qian.
13
See, among others, Goldin et al. (1992), Galor and Weil (1996), Costa (2000), Goldin and Katz (2002), Attanasio
et al. (2008), Albanesi and Olivetti (2009), and Fernandez´ (2013).

106
responsibilities are still one of the biggest impediments to female labor force participation (Schaner
and Das (2016)).
Against this backdrop, we study the potential role of household cooking technology in
determining female labor force participation in Indonesia. Indonesia, like many other low-and
middle-income countries, has grown steadily over the last few decades. While the welfare gains
from this phase of rapid growth in Indonesia have been shared equally between males and females
in domains like education (Figure 3.1), the female labor force participation in Indonesia has
remained below the world average.
14
An opportunity to examine the role of household cooking
technology in determining FLFP presented itself when, in 2007, Indonesia implemented the
national “Conversion to Liquefied Petroleum Gas (LPG) Program”.
The Conversion to LPG program, also known as the “No-Kero” or “Zero-Kero” program,
subsidized the use of LPG. Studies from Indonesia have found that LPG is a labor- and time-saving
cooking technology (ASTAE (2015); Thoday et al. (2018)). Using the exogenous staggered roll-
out of the program, we show that a switch to LPG increased the labor force participation of exposed
women. We also find that the policy was associated with an increase in household expenditure on
food and education and the subjective wellbeing of women. We explore two possible mechanisms
through which the switch to LPG might have affected the labor force participation of women -
better health and time savings. Consistent with previous research on the topic, we do not find major
effects on the health of the exposed women (Smith-Sivertsen et al. (2009); Duflo et al. (2012);
Thoday et al. (2018)). While we do not have information on the time use of the exposed women,

14
In comparison, the labor force participation of Indonesian men has stayed well above the world average and rela-
tively stable in the last three decades. See figure 2.

107
building on information from related studies and some suggestive evidence, we postulate that time
saved due to the technology is an important pathway through which the switch to LPG affected
labor force participation of women.
A back-of-the-envelope calculation suggests that saving in households expenditure on fuel far
outweighed the cost of the conversion incurred by the government. We conjecture that households
fail to switch to LPG despite the unambiguous net gains because of intra-household externalities
and gender differences in preferences - the benefits from switching to a cleaner fuel are greatest
for the woman in the household but the monetary price is most-often paid by the earning male
(Miller and Mobarak (2013); Pitt et al. (2006)). We also show that the policy improves the
decision-making power of women in the household, especially in financial matters. Given the role
of intra-household externalities and gender differences in preferences in the setting, this has
important implications for the sustained use of LPG even after the subsidy is withdrawn.
Our paper makes three main contributions. It is the first paper to evaluate the impact of the
“Conversion to Liquefied Petroleum Gas (LPG) Program” on the labor force participation of those
exposed. The results show that the benefits of the policy went far beyond the saved subsidy
expenditure, the main motivation behind the program. Second, the findings suggest that switching
to faster cooking methods, like cooking with LPG, can liberate women to join the labor force in
developing countries. This is especially important for countries like Indonesia that does not fair
too well on gender equality indices, where the working status of women is an important correlate
of women’s decision-making power within the household and attitudes towards domestic violence
(Schaner and Das (2016)). Third, our findings are also related to the strand of literature that
investigates the seemingly low rates of adoption of simple, relatively inexpensive, highly effective
technologies in developing countries that hold promises of improving the quality of life through

108
their impacts on health and productivity
15
To the extent that intra-household externalities and
gender differences in preferences drive the lack of adoption (Miller and Mobarak (2013)), the
impact of the policy on the decision-making power of the women provides insights into how
temporary subsidies that mitigate such externalities and empower women can encourage adoption
and sustained use of such technology.
We have organized the rest of the paper as follows. Section 3.2 describes the program. Section
3.3 talks about the data and identification strategy. Section 3.4 describes the empirical specification
used. Section 3.5 presents the results and section 3.6 concludes.

3.2. Background
At the turn of this millennium, kerosene was the main fuel used by Indonesian households for
their cooking requirements. In 2004, 48 out of the 52 million Indonesian households depended on
kerosene, mostly for their daily cooking requirement and as lighting fuel (Budya and Arofat
(2011)). The government had provided large subsidies on kerosene for decades and the subsidy
payouts were turning out to be a huge burden on the state, sometimes as high as 18 percent of the
state’s total expenditures.
16
In its attempt to reduce the subsidy burden, in 2007, the Indonesian

15
See, for example, Foster and Rosenzweig (1995), Miguel and Kremer (2004), Bandiera and Rasul (2006), Duflo
et al. (2008), Ashraf et al. (2010), Cohen and Dupas (2010), Conley and Udry (2010) and Foster and Rosenzweig
(2010).

16
The situation was worsened by the reduction of subsidies for industrial fuels (diesel, industrial diesel oil, and
marine fuel oil) in the early 2005, pricing them at international prices. The price disparity between the fuel prices for
industries and households led to a substitution of kerosene for industrial fuels wherever possible and, as a result, an
arbitrage opportunity. This subsequent smuggling caused large leakages in the subsidy increasing the cost even further.

109
government launched the “Conversion to LPG Program” to promote the use of Liquefied
Petroleum Gas (LPG) in Indonesian households.
LPG was the replacement choice for a variety of reasons. First, it was estimated that LPG would
greatly reduce the subsidy cost per unit of end-use calorific value of energy delivered for cooking
and subsidy per unit of fuel. Based on calculations by a team from the University of Trinity in
Jakarta and the State Ministry for Women’s Empowerment that included laboratory experiments
under various cooking conditions in Indonesia, it was found that one liter of kerosene was
equivalent to 0.39 kg LPG in terms of its end-use energy value (Budya and Arofat
(2011)).
17
According to Budya and Arofat (2011), based on the 2006 calculations alone, this would
have saved the state 2.17 billion USD. Second, LPG was a cleaner substitute with lower indoor
pollution, which directly affected the health of the users, and lower levels of greenhouse-related
pollutants compared to solid fuels.
18
Third, the infrastructure required to implement the transition
to a cleaner fuel was more developed for LPG than for other alternatives like electricity. Successful
implementation of subsidized LPG programs in neighboring countries of Malaysia and Thailand
provided additional motivation.
Depending on the readiness of the the LPG procurement, storage, and distributional
infrastructure in the region, the program was rolled out at different times in different regions.
Urban regions often got the program earlier (Budya and Arofat (2011)). By 2008, entire of Jakarta,
Bali, Yogyakarta, Banten, and parts of West, Central, and East Java had been covered. By 2009,
the entire of Java and Bali, parts of Lampung, South Sulawesi, East and West Kalimatan, South

17
This does not take into account the possible misuse of kerosene for industrial purposes, which would further tilt
the scale in favor of LPG. See Budya and Arofat (2011) for a detailed calculation, accounting for such leakages.
18
See Lam et al. (2012) and WHO (2014) for a review.

110
and North Sumatra, and Riau had received the program. By 2011, the program covered the entire
of Aceh, North Sumatra, Riau, Jambi, Bengkulu, Lampung, entire of Kalimatan except central
Kalimatan, and entire of Sulawesi except central and Southeast Sulawesi. By 2013, West Sumatra,
West Nusu Tenggara, Bangka Belitung, and the remaining regions of Kalimatan and Sulawesi
were covered. Some regions, like East Nusu Tenggara, Malaku, North Malaku, and Irani Jaya were
not covered by the program. As is clear, there was a substantial level of variation in the roll-out
date across provinces. Figure 3.3 depicts the variation in roll-out of the program.
Under the program, all eligible citizens were to receive a free ‘initial pack’ comprising a 3-kg
LPG cylinder with the gas, a one-burner stove, a hose, and a regulator. A few trials runs were
conducted before the launch of the program to gauge the society’s perception and acceptance of
LPG as a cooking fuel. The first test was carried out in Cempaka Baru Village, Kemayoran District,
Central Jakarta, on August 1, 2006. 500 families were given the ‘initial pack’ and their responses
and behaviors of the users were noted through surveys and observational methods. A second test
was carried out with 18,800 households in Kemayoran District, Central Jakarta, and 6700 families
in Karawaci District, Tangerang, Banten in December 2006. This test was not accompanied by a
survey, and evaluations were based on observations of people’s reaction. The general picture from
these market tests was that households were willing to switch to LPG under the subsidy (See Budya
and Arofat (2011) for details). A third test was carried out in February 2007 when the Ministry of
State-Owned Enterprises, under the State-Owned Enterprises Care program to help flood victims
in Jakarta, distributed 10,000 LPG cylinders in Kampung Makassar, East Jakarta. Here too the
results were in favor of scaling up the program.
The program had a significant impact on the use of LPG as cooking fuel in Indonesia (Andadari
et al. (2014)). The share of LPG in household consumption expenditure increased from 1.9 percent

111
in 2005 to 13.5 percent in 2013, while the share of kerosene dropped considerably from 18 percent
in 2005 to 1.8 percent in 2013. (Toft et al. (2016)). Besides the savings in subsidy cost for the
government, switching from Kerosene to LPG might have had implication on community-level
pollution and depletion of natural resources like forests, on food habits, budget allocations,
resources distribution and bargaining within the household, and on health, education, time use, and
labor force participation of individuals from the exposed household. A cost-benefit analysis in
terms of subsidy cost-savings alone is likely to understate the net benefits of the program.
However, there have hardly been any systematic evaluations of the impact on the program,
especially on factors affecting the health and economic wellbeing of those covered by the
program.
19

3.3. Data and Identification
For our main analysis, we use the information from a geographically stratified systematic 10%
sample of the 2010 Indonesian Population Census. The census interviews the entire population of
Indonesia, Indonesian and foreign, residing in the territorial area of Indonesia, regardless of
residence status and includes homeless, refugees, ship crews, and people in inaccessible areas.
Diplomats and their families residing in Indonesia are excluded. The census collected information
on a wide range of variables including the district and province of current residence and the
primary fuel used by responding households, the educational attainment, employment status, age

19
Andadari et al. (2014) look at the impact of the program on energy poverty. They find that the programs led
to increased stacking of fuels, increasing consumption of both electricity and traditional biomass. It failed to
reduce the overall number of energy-poor people although it was somewhat effective at reducing extreme energy
poverty. Permadi et al. (2017) find that the program led to significant reductions in emissions of greenhouse gases
and air pollutants

112
and gender of the individual respondents. Wherever required, we use earlier waves of the
Population Censuses and Intercensal Population Survey of Indonesia to examine time trends in the
independent and dependent variables of interest.
Using information from these censuses, we first examine the impact of the program on the
household’s primary fuel of choice and the educational attainment and the employment status of
individual respondents. While the large sample size of these censuses allow us to estimate the
impact of the program on these variables with great precision, they lack additional details about
the households and the individuals respondents preventing further analysis of the program. To get
around this problem, we then use the information the third, fourth and fifth wave of the Indonesian
Family Life Survey (IFLS). IFLS is an on-going longitudinal household survey representative of
about 83 % of Indonesian population living in 13 of the 27 provinces in the country (Strauss et al.
(2016)). The first wave was administered in 1993 to over 22,000 individuals living in 7,224
households. The follow-up waves 1997, 2000, 2007, and 2014, sought to follow the original
respondents and their off-springs in the same or split-off house-hold. In IFLS 5, 50,148 individuals
living in 16,204 households were interviewed. The survey is remarkable for its low levels of
attrition, with the recontact rate of original IFLS 1 dynasties (any part of the original IFLS 1
household) in IFLS 5 as high as 92%. We make use of waves 4 and 5 of the survey for our analysis.
The survey contains information on a wide variety of topics at the individual, the household and
the community level. At the individual-level, we make use of information on health, education,
employment, migration, etc., of respondents. At the household level, we utilize the information on
the main cooking fuel of the household and whether the household’s kitchen is inside the house.
Here, we first show that the impact of the program on LPG usage, education, and employment are

113
robust across the two data sets. Then, we examine the impact of the program on a wide range of
outcomes, including health and decision-making within the household.
The information on the variation in program roll-out across regions is obtained from Budya and
Arofat (2011) and Thoday et al. (2018). As described above, in certain cases only a part of a
province was covered in a given year. The rest of the province was covered in the following years.
Unfortunately, we do not have precise data on variation in roll-out at a finer level
(district/village/communities). Instead, we define a province to have received the program only if
the entire province was covered. This induces some degree measurement error that will bias the
estimates downwards. The variation in roll out of the program across the communities in the IFLS
dataset is presented in figure 3.4 and tables 3.1 and 3.2 reports the summary statistics for the two
data sets we use.

3.4. Empirical Specification
By the time of the 2010 census, some provinces in Indonesia had received the LPG program
while others had not. If the program had been randomly assigned to the provinces, we could have
have attributed the differences in the outcome variables of interest across the provinces that had
received the program (hereon, exposed provinces) and the provinces that had not (hereon, control
provinces) as the causal impact of the program. But as we point out in Section 3.2, the rollout of
the program was not random. The regions that had ready-infrastructure for LPG procurement,
storage, and distribution had received the program. It is likely that the exposed provinces were
different from the control provinces along a number of dimensions including our outcome
variables of interest or the factors that drive these outcomes. To account for this, we use a
difference-difference strategy. We compare the changes in our outcome variables of interest

114
between 2005 and 2010 for provinces that had received the program by 2010 with provinces that
had not received the program by 2010. Accounting for pre-existing differences across the
provinces, we expect that the household in provinces that had received the program by 2010 must
have increased their LPG usage more than those in control provinces.
The identifying assumption here is that in the absence of the program, the change in these
outcome variables of interest should have been the same in the exposed and control provinces.
Said differently, the trend in a variable of interest over time in the exposed provinces in the absence
of the program is assumed to have been the same as the trend in the variable in the control provinces
(hereon, the parallel trends assumption). We first provide support in favor of the parallel trend
assumption by showing that the variables of interest trended parallel in exposed and control
provinces before 2005. Then, we estimate the following equation:
𝑌𝑌 𝑖𝑖 𝑖𝑖 𝑝𝑝 𝑡𝑡 = 𝛼𝛼 + 𝛽𝛽 × 𝑃𝑃 𝑓𝑓 𝑠𝑠𝑡𝑡
𝑡𝑡 × 𝑇𝑇 𝑓𝑓𝑒𝑒𝑎𝑎 𝑡𝑡 𝑖𝑖𝑝𝑝
+ 𝜏𝜏 𝑡𝑡 + 𝛿𝛿 𝑖𝑖𝑝𝑝
+ 𝜀𝜀 𝑖𝑖 𝑖𝑖 𝑝𝑝 𝑡𝑡 (3.1)
where 𝑌𝑌 𝑖𝑖𝑖𝑖 𝑝𝑝𝑡𝑡 is the outcome variable of interest for household or individual i living in district
(kabupaten in Indonesia) d of province p in year t. At the household level, the outcomes of interest
are whether or not the household used LPG as the primary cooking fuel. At the individual level,
we are most interested in the impact of the program on the labor force participation of those
exposed to the program, especially that of females. 𝑃𝑃 𝑓𝑓 𝑠𝑠𝑡𝑡
𝑡𝑡 denotes the pre-and post-rollout period.
It takes value ‘0’ for year 2005 and ‘1’ for 2010. 𝑇𝑇 𝑓𝑓𝑒𝑒𝑎𝑎 𝑡𝑡 𝑖𝑖𝑝𝑝
is an in-dicator variable that takes value
‘1’ for all districts in all the provinces that had received the program by 2010, ‘0’ otherwise. 𝜏𝜏 𝑡𝑡
controls for time-varying factors that were common to exposed and control province and could
have affected the outcome of interest. 𝛿𝛿 𝑖𝑖𝑝𝑝
controls for time-invariant differences across districts

115
that could have affected the outcome.
20
To maintain consistency with the specifications that
follow, we cluster the standard errors at the level of the district. Clustering them at the level of the
province does not affect the statistical significance of the results.
However, provinces in Indonesia are considerably different. Not only in their population
(ranging from a few hundred thousands to well over 40 millions) and their geographical area (from
a little over 250 square miles to over 120000 square miles) but also in their distance from the
government’s seat in Jakarta or other bigger urban commercial centers in the country. As a result,
it is possible that even though the time trends in variables of interest for the exposed and control
provinces are parallel on an average, there are time-varying unobservable differences across
provinces that might bias our results. For example, consider a scenario where some provincial
administrations in-charge of the LPG program bundled the LPG program with other programs that
affected the outcomes of interest while other did not. If so, if we estimate the model in (3.1), we
will attribute any affect of these other programs on the outcome to the LPG program.
To get around this problem, we use a modified version of the shift-share instrument - we interact
𝑃𝑃𝑓𝑓 𝑠𝑠 𝑡𝑡 𝑡𝑡 × 𝑇𝑇 𝑓𝑓𝑒𝑒𝑎𝑎 𝑡𝑡 𝑖𝑖𝑝𝑝
with the proportion of household in district 𝑑𝑑 of province 𝑝𝑝 that used kerosene
as their primary cooking fuel in 2005.
21
The proportion of households in different districts within
the provinces in Indonesia that used kerosene as their primary cooking fuel was vastly different.
For the 258 districts included in the IPC and SUPAS, it ranges from as low was 0.03 % to as high
as 94% in 2005. In the IFLS survey, out of the 311 communities, none of the households in nine

20
Replacing district fixed effects with province fixed effects does not change our results.
21
The shift-share instrument, often referred to as the Bartik instrument (Bartik (1991)), is used extensively in the
migration literature. Some early applications of the instrument include Altonji and Card (1989), Card (2001), and Card
(2009). It leverages the observation that a national policy will have differential impact across different regions of the
country depending on the size of the population in each region affected by the policy.

116
communities and all of the households in 3 communities used kerosene in 2000. The LPG program
was a national-level policy intervention and, therefore, should be exogenous to the variation in
kerosene usage within the province.
22
Therefore, while the timing and nature of the program could
have differed across provinces (shift), it is unlikely that it was associated with the differences
across districts within a province and the districts with a higher proportion of kerosene users before
the program within a province would have benefited more from the program (share).
23

There are two reasons for why the districts with a higher incidence of kerosene usage stood to
benefit more from the program. One, the LPG subsidy was rolled out to replace the kerosene
subsidy. As a result, there was a high correlation between the phase in of the LPG subsidy and the
phase out of the kerosene subsidy. This meant that while the cost of LPG decreased for all
household in the regions that received the LPG subsidy, the relative price of kerosene went up
even more for household that used kerosene before. Second, before the LPG program, kerosene
was a highly subsidized fuel. Households that chose not to use kerosene even with the high subsidy
must have had a relatively inelastic demand for the fuel they used instead.
24
It is likely that a
reduction in LPG prices might have been equally unsuccessful in getting these households to
switch from their fuel of choice. Therefore, one can think of the variation in pre-program kerosene
usage across districts as a variation in the magnitude of the subsidy or the extent of its coverage.
We estimate the following specification:

22
“National specification of targeted localities for conversion would be done centrally under control of the
conversion team established by Pertamina.” - (Budya and Arofat (2011))
23
Our strategy is similar to Bleakley (2007) who combines the introduction of the hookworm eradication campaign
in the American South in the 1910s with the variations in the hookworm infection rates prior to the campaign across
regions to identify the impacts of hookworm eradication on later-life outcomes. The author points out that different
areas of the US had distinct incidences of the hookworm disease and, therefore, stood to gain differentially from the
campaign. The innovations in treatment of hookworm were not related to or in anticipation of the future growth
prospects of the affected areas.
24
Firewood was the second most important primary fuel of choice before the program.

117

𝑌𝑌 𝑖𝑖 𝑖𝑖 𝑝𝑝 𝑡𝑡 = 𝛼𝛼 + 𝛽𝛽 1
× 𝑃𝑃 𝑓𝑓 𝑠𝑠𝑡𝑡
𝑡𝑡 × 𝑇𝑇 𝑓𝑓𝑒𝑒𝑎𝑎 𝑡𝑡 𝑖𝑖𝑝𝑝
× 𝐾𝐾 𝑒𝑒𝑓𝑓𝑓𝑓
𝑖𝑖𝑝𝑝 , 2 0 0 5
+ 𝛽𝛽 2
× 𝑃𝑃 𝑓𝑓 𝑠𝑠𝑡𝑡
𝑡𝑡 × 𝑇𝑇 𝑓𝑓𝑒𝑒𝑎𝑎 𝑡𝑡 𝑖𝑖𝑝𝑝
+ 𝜏𝜏 𝑡𝑡 × 𝐾𝐾 𝑒𝑒𝑓𝑓𝑓𝑓
𝑖𝑖𝑝𝑝 , 2 0 0 5
+ 𝛾𝛾 𝑡𝑡 𝑝𝑝 + 𝛿𝛿 𝑖𝑖𝑝𝑝
+ 𝜀𝜀 𝑖𝑖 𝑖𝑖 𝑝𝑝 𝑡𝑡 (3.2)
where the terms common with (3.1) are defined as before. 𝐾𝐾 𝑒𝑒𝑓𝑓𝑓𝑓
𝑖𝑖𝑝𝑝 , 2 0 0 5
is the percentage of
households in district 𝑑𝑑 of province 𝑝𝑝 who used kerosene as their primary cooking fuel in 2005.
𝛽𝛽 2
captures the impact of the program in districts where no one used kerosene as the primary
cooking fuel in 2005. 𝛽𝛽 1
measures the increase in the impact of the program with increase in the
pre-program usage rate of kerosene. Following Acemoglu et al. (2004), Hoynes and Schanzenbach
(2009) and Hoynes et al. (2016), we also include interactions of the year fixed effects with the pre-
program proportion of kerosene users in the districts to control for possible differences in trends
across districts with different levels of kerosene users. In addition, we include province-year fixed
effects 𝛾𝛾 𝑡𝑡 𝑝𝑝 to account for time-varying difference across provinces and 𝛿𝛿 𝑖𝑖𝑝𝑝
to account for time-
invariant differences across districts. Even if some provinces rolled out the program in
combination with other programs, the province-year fixed effects will control for such differences.
Since there is no variation in 𝑇𝑇 𝑓𝑓𝑒𝑒𝑎𝑎 𝑡𝑡 𝑖𝑖𝑝𝑝
, 𝐾𝐾 𝑒𝑒𝑓𝑓𝑓𝑓
𝑖𝑖𝑝𝑝 , 2 0 0 5
, and 𝑇𝑇 𝑓𝑓𝑒𝑒𝑎𝑎 𝑡𝑡 𝑖𝑖𝑝𝑝
× 𝐾𝐾 𝑒𝑒𝑓𝑓𝑓𝑓
𝑖𝑖𝑝𝑝 , 2 0 0 5
within a district,
their effects are absorbed in the district fixed effect 𝛿𝛿 𝑖𝑖𝑝𝑝
. The effects of 𝑃𝑃𝑓𝑓 𝑠𝑠 𝑡𝑡 𝑡𝑡 and
𝑃𝑃 𝑓𝑓 𝑠𝑠𝑡𝑡
𝑡𝑡 × 𝐾𝐾 𝑒𝑒𝑓𝑓𝑓𝑓
𝑖𝑖𝑝𝑝 , 2 0 0 5
are absorbed in the 𝜏𝜏 𝑡𝑡 × 𝐾𝐾 𝑒𝑒𝑓𝑓𝑓𝑓
𝑖𝑖𝑝𝑝 , 2 0 0 5
and 𝛾𝛾 𝑡𝑡 𝑝𝑝 .
Once we establish the impact of the program using data from the censuses and the intercensal
surveys, we move to the IFLS to examine other outcomes and mechanism variables of interest.
None of provinces had received the program by 2000 when the third wave of IFLS was fielded.
By the time of the IFLS wave 4 in 2007 while the program had started, it was still in its initial
stages and none of the provinces had been covered completely. By the time of the fifth wave of

118
IFLS, all the provinces included in the IFLS surveys had been covered. As a result, in contrast to
data from the IPC and SUPAS, we do not have distinct exposed and control provinces in IFLS
and, therefore, cannot use 𝑃𝑃 𝑓𝑓 𝑠𝑠𝑡𝑡
𝑡𝑡 × 𝑇𝑇 𝑓𝑓𝑒𝑒𝑎𝑎 𝑡𝑡 𝑖𝑖𝑝𝑝
identification strategy laid out in (3.1). However,
IFLS, besides the in-depth information on individuals and households, has one more advantage
that helps the identification of the program impacts. IFLS provides geographical identifiers for
communities that are smaller geographical units than districts. This allows us to use variations in
pre-program kerosene usage at a finer level to identify the impact of the program. We begin by
estimating the following specification:
𝑌𝑌 𝑖𝑖 𝑐𝑐 𝑖𝑖 𝑝𝑝 𝑡𝑡 = 𝛼𝛼 + 𝛽𝛽 1
× 𝑃𝑃 𝑓𝑓 𝑠𝑠𝑡𝑡
𝑡𝑡 × 𝐾𝐾 𝑒𝑒𝑓𝑓𝑓𝑓
𝑐𝑐 𝑖𝑖𝑝𝑝 , 2 0 0 0
+ 𝜏𝜏 𝑡𝑡 × 𝐾𝐾 𝑒𝑒𝑓𝑓𝑓𝑓
𝑐𝑐 𝑖𝑖𝑝𝑝 , 2 0 0 0
+ 𝛾𝛾 𝑡𝑡 𝑝𝑝 + 𝛿𝛿 𝑖𝑖𝑝𝑝
+ 𝜀𝜀 𝑖𝑖 𝑐𝑐 𝑖𝑖 𝑝𝑝 𝑡𝑡 (3.3)
where c denotes the community recorded in the IFLS survey. 𝐾𝐾 𝑒𝑒𝑓𝑓𝑓𝑓
𝑐𝑐 𝑖𝑖𝑝𝑝 , 2 0 0 0
is the proportion of
households in community 𝑐𝑐 of sub-district (kecamatan) 𝑑𝑑 of province 𝑝𝑝 who used kerosene as the
primary cooking fuel in 2000. Similar to (3.2), we include interaction of the time fixed effects with
the pre-program rate of kerosene usage, sub-district-year fixed effects, and community fixed
effects. We cluster the standard errors at the level of the community.

3.5. Result
3.5.1. Fuel of Choice
Figure 3.5 reports the change in proportion of respondent households cooking with different
kinds of fuel. The proportion of households using LPG increased substantially from below 10 %
in 2005 to almost 50 % in 2010. We also observe a corresponding decline in the use of kerosene.
Consistent with earlier findings, we find that there were no sharp trend breaks in the proportion of

119
households using solid fuels between 2005 and 2010 but the number of solid-fuel users was
declining throughout the 1995-2010 period (Thoday et al. (2018)). The LPG conversion program
started in 2007-08. Therefore, it seems likely that the increase in LPG usage rate was a result of
the program. To probe this further, in Figure 3.6, we break down the LPG usage rate by whether
or not the district was exposed to the program by the time of the survey. There was an increase in
the LPG usage rate in all districts between 2005 and 2010.
25
However, the increase in LPG usage
in districts that had received the program was visibly greater than that in districts that had not
received the program.
We verify these findings using a regression framework that controls for district-level differences
and province-level changes over time. Table 3.3 presents the results. In column (1), we compare
the differences the probability of a household using LPG across time in exposed and control
provinces. We find that households in regions that received the LPG program were more likely to
use LPG by almost 40%. In columns (2) - (4), we show that this finding is not sensitive to the level
of geography that we include fixed-effects and cluster the standard errors at. In column (5), using
the strongest and our most-preferred specification from equation (2) that allows us to exploit finer
geographical variation, we show that the impact of the program was much higher in districts with
higher pre-program kerosene usage rate. As expected, the program had a bigger impact on the fuel
of choice in districts with a high rate of pre-program kerosene usage. The difference between the
change in LPG usage rate across two exposed districts, one where no one used kerosene before the
program and the other where everyone used kerosene before the program, was almost 40

25
Remember, according to our definition of exposure, districts in a province are not considered exposed until the
entire province has been covered by the program. This means that we might categorize some districts that have already
received the program as control districts. As explained in section 3, this will bias our coefficients downwards. This
may also explain some of the increase in the LPG usage rate in control districts in Figure 6

120
percentage points. The findings from table 3.3 are consistent with the broad trends presented in
Figures 3.6 and 3.7 - the program had a causal effect on the LPG usage rate.
Next, we verify these findings using information from IFLS using community-level variations.
We present the results in Table 3.4. According to column (1), controlling for differences across
time and time-invariant differences across communities, communities where everyone used
kerosene in 2000 were 40% more likely to be using LPG after the program in 2014 compared to
communities where no one used kerosene in 2000. Controlling for household fixed effects and
kecamatan-year fixed effects do not change the results. The impact magnitudes estimated using
information from IFLS are strikingly close to those from IPC and SUPAS, suggesting that
estimated impacts are robust across data-sets.
3.5.2. Labor supply
As discussed before, adoption of modern household technology can significantly impact on the
labor force participation of household members. Figure 3.11 presents the unconditional trend in
the labor force participation of men and women in the exposed and control provinces. The labor
force participation appears to have followed a roughly parallel trend in the two groups until 2005.
However, the labor force participation of both men and women in 2010 was significantly more in
province exposed to the program. Table 3.5 presents the difference in the labor force participation
status controlling for pre-program difference across regions. According to column (1), the labor
force participation increased significantly in regions exposed to the program. In column (2), we
find that though the labor force participation of status of both men and women increased over the
period, the increase in labor force participation of women was far 26 percentage points higher than
that for the males. In column (3), we examine the increase in labor force participation by pre-
program kerosene usage rate. As expected, we find that individuals in regions where the program

121
had a bigger impact on LPG usage see a higher increase in labor force participation. Finally, in
column (4), we break down the impact on males and females in high and low pre-program usage
rate. We find that the program had a negative effect on the labor force participation rate of males
in districts with low rates of pre-program kerosene usage but this effect was more than offset by
an increase in the female labor force participation in these districts. The effect was not significantly
different for males in districts with high rates of pre-program kerosene usage. However, the
increase in labor force participation of women in these regions was much higher. In summary, the
findings suggest a change in intra-household allocation. Controlling for province-level time
variation observable factors, we find that men might have decreased their labor force by a small
amount and women increased their labor force participation in all districts, more so in districts
more affected by the program. This is consistent with the findings from the OECD countries that
modern household technology has led to an increase in female labor force participation
(Greenwood et al. (2005); Goldin (2006); Aguiar and Hurst (2007); de V. Cavalcanti and Tavares
(2008)).
Data from the IFLS allows us to examine the impact of the program on the type of work that
men and women do. Table 3.6 presents the results. Women exposed to the program in regions that
had a high pre-program usage rate of kerosene were much more likely to report working for pay
as their primary activity in week prior to the survey. There is a corresponding decline in women
reporting housekeeping as their primary activity in the previous week. In terms of all activities
performed in the previous week, exposed women report having worked with or without pay and
searched for jobs more often and to have done housekeeping less often. Interestingly enough,
housekeeping activities for the men also seem to have gone down due to the program. This suggest
that the change due to the program was not a mere reassignment of household and other

122
responsibilities. The increase in labor force participation of the exposed women is also visible in
the increase in their probability of having ever held a job in the years preceding the survey (Table
3.7).
There are two important differences between the estimated labor market impacts of the program
in tables 3.5 and 3.6. First, the impacts are smaller for women when we use information from IFLS.
This could be a result of the fact that IFLS is representative of only 83% of the Indonesian
population living in 13 provinces on the main islands and misses out on the remoter areas of the
country (Strauss et al. (2016)). It is conceivable that the program had a bigger impact on the labor
force participation of women in these remoter areas. Comparing the labor force participation of
women in across the summary statistic tables 3.1 and 3.2, it is clear that those areas not included
in IFLS but included in IPC and SUPAS have a lower rate of female labor force participation.
This, in turn, could have been a result of the differences in household cooking technology used
across these regions. The IFLS regions had a higher rate of LPG us-age than the IPC and SUPAS
regions before the program. The program, therefore, might have liberated more women from the
burden of household responsibilities in these remote regions. Second, there appears to be no
negative impact of the program on male labor force participation of men when we use information
from IFLS. This too could be due to the difference is the representativeness of the IFLS from that
of IPC and SUPAS. Removal of kerosene subsidy negatively affected some cottage industries in
the coastal areas. For example, the Batik textile production, a textile production technique
indigenous to Indonesia, in coastal regions suffered when the kerosene subsidy was withdrawn as
LPG could not be used in place of kerosene to melt the Batik wax.
26

26
We thank Mari Pangestu, erstwhile Minister of Trade, and Tourism and Creative Economy, for pointing this out.

123
3.5.3. Time use
As is clean from table 3.6, while women exposed to the program were less likely to report having
performed housekeeping activities in the week prior to the survey, there was no discernible
increase in the housekeeping activities performed by men from the exposed households. This
suggests that women must have found the time to do both - perform housekeeping activities and
work for pay. Since it is unlikely that the program changed the list of housekeeping activities to
be performed, women must have been able to perform their housekeeping activities in a smaller
amount of time.
This is not unlikely. An advantage of cooking with LPG is the smaller amount of time required
for cooking compared to cooking with kerosene or other solid-fuels. Igniting a solid-fuel or a
kerosene stoves to full capacity is substantially more work than switching on the LPG stove by
turning a knob. Unlike some other fuels, it also does require the women to spend time collecting
the fuel and preparing it for usage. Since the cooking activities in most developing countries are
predominantly carried out women, the benefits of a switch to LPG, especially in terms of time
saved, are likely to be higher for women (Pitt et al. (2006); Miller and Mobarak (2013)).
Unfortunately, we do not have time use data for exposed women to be able to examine this
mechanism explicitly. However, earlier research on related topics provide suggestive evidence.

In their 2016 study of the Indonesian domestic biogas program of 2009, Gurung and Sety-owati
(2016) found that women save well over one hour per day when they switch to domestic bio-gas
for their cooking needs. This time saving, they report, is net of activities like cleaning the stable,
collecting dung, putting the dung into bio-digester, putting bio-slurry into the pit, etc., required to
fuel a bio-gas plan that requires close to forty minutes. LPG stoves do not require these elaborate

124
processes to keep it running. Therefore, the time saved from switching to LPG might have been
higher. Gurung and Setyowati (2016) also find that most of the saved time is spent in productive
activities. Similarly, an in-depth survey of cooking fuel consump-tion and cooking habits in peri-
urban households outside Yogyakarta City, in central Java by the World Bank found that cooking
with LPG was significantly faster than other methods (AS-TAE (2015)). When examining
preference for fuels and cooking stoves, the survey finds that households preferred technologies
that saved time.
It is likely that the LPG program, since it was similar to the bio-gas program but only faster, had
similar effects on the time use of the women in the household and on their labor force participation.
Is the time-saving enough to generate impacts on labor force participation? Building on the
findings from Gurung and Setyowati (2016), even if we use a conservative estimate of one hour
saved everyday, it amounts to seven hours in a week. Aggregating time saved over a week is
especially important in this case since some of the activities it replaces, like collection of firewood
and chopping it into usable blocks, is done on a weekly basis and often performed collectively by
female members of the households. With such activities no longer required, it is plausible that
women might have had enough time to work for pay for at least one day during the week.
Unfortunately, it is difficult to make claims about time use as a mechanism with certainty without
data on time use and future research should aim to test with hypothesis explicitly.

3.5.4. Health outcomes
Time savings from switching to LPG might not be the only pathway through which the program
might have affected labor supply. Cleaner cooking fuel generates less indoor air pollution. This
could have improved the respiratory health of the household members. In fact, much of the

125
motivation behind the large subsidies on cleaner cooking stoves and fuels comes from their
potential positive impact on health, and in particular, the respiratory health of women and young
children through reduction in indoor air pollution. And while better health is a desirable result in
itself, it might also affect the labor supply of the household members.
However, despite the perceived potential benefits, there is a dearth of empirical evidence on the
respiratory health benefits of using cleaner cooking fuels or technologies. Duflo et al. (2012)
examine the impact of a randomized distribution of cleaner cooking stoves in rural Orissa in India
on respiratory health of those who received the cook stove. They find reduction in the amount of
smoke inhaled in the first year but no improvements in lung capacity or other measures of health.
RESPIRE study, an experiment involving randomized distribution of concrete stoves in
Guatemala, finds similar results - reduction in CO and pm2.5 exposure but no improvement in
lung function and other respiratory symptoms like chronic cough, wheezing, tightness of chest,
etc. (Smith-Sivertsen et al. (2009)).
Using information from IFLS waves 2, 3, and 4, Silwal and McKay (2015) find that individuals
living in households that cook with firewood have 11.2 per cent lower lung capacity than others.
But their instrument of choice for household’s fuel choice, the availability of an all-whether road
in the community, might have affected health via other channels like access to health care facilities.
Gajate-Garrido (2013) uses a two-wave panel survey of Peruvian children and a household fixed
effects specification to show that young boys in households cooking with firewood are more likely
to report respiratory illnesses. The household fixed effects model does not account for household-
level time varying factors that might affect the choice of cooking fuel and child health. Besides, it
is not clear why the effect might be differential effects on girls, for whom she finds no impact, and
boys.

126
Since IPC and SUPAS do not contain health measures for the respondents, we turn to the IFLS
to examine the impact of the program on health. As a part of the IFLS survey, a professionally
trained nurse collects an extensive array of biomarker measurements. In table 3.8, we examine the
impact of the program on some of these measures. The program had no effect on the maximum
lung capacity of those exposed to the program. Among other measured health biomarkers, we do
not find any significant impact of the program on the probability of being underweight, grip
strength, systolic or diastolic blood pressure of any one in the household. However, exposure to
the program is associated with a significant increase in the proportion of overweight males and
females. We also see a significant increase in the pulse rate of males. IFLS also collects self-
reported information on doctor-diagnosed chronic conditions. Table 3.9 reports the impact of the
program on the probability of having been diagnosed with certain chronic conditions. Consistent
with our earlier findings on lung capacity in table 3.8, we find no effect of the program on
respiratory conditions like asthma and other lung conditions.
Exposure to the program is associated with a small decrease in the incidence of hypertension in
females. But taken together, the findings suggest that there were no major impact of the program
on the health of those exposed to the program. Our findings, that are consistent with Smith-
Sivertsen et al. (2009) and Duflo et al. (2012), appear to be driven by two factors. First, most of
the households that changed their primary cooking fuel switched from kerosene to LPG. Studies
find that kerosene is almost as clean as LPG in household cooking settings (Mehta and Shahpar
(2004)). Second, there is a significant positive association between those who cook with solid fuels
and those who have the kitchen outside their main housing building. This is consistent with the
findings of Pitt et al. (2006) who find that households in Bangladesh understand the harmful effects
of indoor air pollution generated due to cooking and invest in mitigation mechanisms. Similarly,

127
Kan et al. (2011) find that households in Anhui, China tend to use griddle stoves with smoke
removed by a hood or a chimney and cook in a separate room or building to mitigate the harmful
effects of cooking with solid-fuels. If the Indonesian households choose the location of the kitchen
strategically to mitigate the negative impact of indoor air pollution due to cooking, it seems
plausible that these household also invest in other methods of mitigation, including better
ventilation in the kitchen. The lack of any major significant effects on the respiratory health of
those who received the program are, therefore, not surprising.
The programs impact on lifestyle diseases, chances of being overweight and suffering from
hypertension are unlikely to be a result of reduction in indoor air pollution. While a reduction of
labor market activities could have been a possible explanation for increasing weight-related issues
in men, we do not find a reduction in the labor market or household activities for males in the IFLS
dataset that we use to evaluate the health effects of the program. In addition, change in labor market
activities cannot explain the results for women who were working more often. A more plausible
pathway is the income effect. An increase in labor force participation of women is likely to increase
the household income. This additional income may have changed the composition of household’s
food consumption that lead to these effects. But these effects are too small to explain the magnitude
of the effect on female labor force participation.

3.5.5. Other benefits
The increase in participation of women in work for pay activities, even though small, should
imply an increase in household income and expenditures. We examine this by looking at the impact
on different types of expenditure for the households. We report the results in table 3.10. For pre-
program kerosene-user households exposed to the program, weekly expenditure on food items

128
increased significantly after the program. While an increase of USD 3.61 might not look to high,
it is important to compare it with the average food expenditure per week. In percentage terms,
there was a 14% increase in food expenditure for the households affected by the program in the
week prior to the survey. In table 3.11, we examine the impact of the program on food composition.
We find that the program led to an increase in consumption of fruits, especially by women, but did
not have significant effects on the consumption of protein-rich food items. But we must point out
that the results in table 3.11 capture the impact of the program on the extensive margin of the food
items reported and fail to capture any changes in the quantity and quality of the food items at the
intensive margin. A key takeaway from table is that the food consumption benefits accrue to both
males and females in the family. That is, the increased food expenditure due to the program
benefited both males and females in the household.
The change in non-food and education expenditures, in comparison, have a lot of variation to
infer a clear impact of the program (columns (2)-(4) of table 3.10). The impact of the program on
non-food expenditure is, a priori, theoretically ambiguous. While increased female labor force
participation might have led to increase in household non-food expenditure, a reduction in price
of fuel due to the program, a non-food commodity with a relatively inelastic demand, might have
meant a reduction in non-food expenditure. In table 3.12, we separate out the impact of the program
on fuel and other expenditure on other utilities. We find that fuel and expenditures on other utilities
form a significantly smaller share of household non-food expenditure for the households affected
by the program. Therefore, it is difficult to rule out the positive impact on non-food consumption
even though we do not see a significant effect on the house-hold’s non-food expenditure. It is
entirely possible that the money saved in fuel expenses was used to increase the consumption of
other non-food items.

129
But though everyone in the household benefits from the program’s impact on household
expenditures, women were working more often. It is not clear by itself that the women pre-ferred
the arrangement where an increase in consumption expenditure came at the cost of them working
more. It is possible that women would have preferred to enjoy their time savings as leisure but
were pressurized by household members to work for pay instead. While there is no way to verify
that with the data we have, we might expect such a situation to have negative effect on the
subjective wellbeing of women. Table 3.13 reports the impact of the program on the subjective
wellbeing of members of the exposed household. While there is no change in the subjective
wellbeing of men except for increased optimism about the future, women are significantly more
optimistic about the future, less concerned about the situation of their standard of living and food
consumption, and happier. this makes it unlikely that women were pressured into work against
their wishes. In the next section, we provide further evidence on increased decision-making power
of women that rules out the possibility of women being pressured into work further.

3.5.6. Cost-benefit analysis and female decision-making power
In 2007, the cost of LPG/kg (US$ 0.89) was marginally higher than the cost of a liter of kerosene
(US$ 0.61). However, 1 liter of kerosene was equivalent to 0.39 kgs of LPG in terms of end use
energy generated (Budya and Arofat (2011)). Even if we assume that the two fuels generated the
same amount of energy per kg, and the average LPG requirement for one household to be between
4 to 5 kgs per household per week (Thoday et al. (2018)), the benefits of switching to LPG on
household food expenditure alone outweighed the costs. The question that then arises is why did
the household not switch to LPG themselves?

130
The lack of adoption cannot be explained as a supply side constraint. In 2007, the aver-age rate
of LPG usage across different IFLS communities was close to 20%. Out of the 312 communities,
237 had at least one household using LPG. But even among communities with at least one LPG
user, the LPG usage rate was around 26%. Later, the single-most important rea-son for choosing
LPG as the replacement fuel was that “… elements of the supply chain were already in place and
it was the easiest fuel to distribute to rural and remote populations across a vast territory” (Thoday
et al. (2018)). This suggests that even in 2007, LPG was readily available. Since the difference
between the expenditure on fuels would have been around five percent of the average household
weekly food expenditure, it is unlike that credit constraints prevented around 80 percent of the
Indonesian households from using LPG. Another often-cited reason is that the LPG cylinders
before the program had a capacity of 12 kgs while the those distributed during the program were
3-kg cylinders and the 12 kg-cylinders were difficult to transport and store. We cannot rule this
out as a possible explanation. But a 12-kg cylinder would have meant a single trip to the retailer
in a month in comparison to multiple trips for those using kerosene. Storage at home is also
unlikely to be a factor since the two types of cylinders were significantly different only in their
height.
A more likely reason seems to be the one suggested by Miller and Mobarak (2013) and alluded
to by Pitt et al. (2006) - intrahousehold externalities and gender differences in preferences. In
Indonesia, mostly women are in charge of cooking activities. As a result, they bear the maximum
brunt of the negative impact of the conventional cooking methods. How- ever, expenditure
decisions are often taken by the males in the family who might sometimes be somewhat reluctant
to spend money on commodities that do not benefit them directly. That is, there might be intra-

131
household externalities of the decision to switch fuels and there might be a difference in
preferences across different genders within the household.
It is possible that if women had more say in financial decisions, there might have been a higher
rate of adoption of cleaner cooking fuel. To examine this further, we examine the as-sociation
between the woman’s choice of cooking fuel and her decision-making power within the household.
We use two measures of a woman’s decision-making power within the house-hold. IFLS surveys
ask a respondent 18 questions about who among their household members makes decisions
pertaining to different household matters. For example, one of the questions asked that pertains to
financial decision-making is “In your household, who makes decisions about money for monthly
savings?” The respondent can choose more than one person as the decision-maker. For our first
measure, we count the respondent as having complete say in the matter if the respondent reports
that he or she takes decisions in the matter alone. For the second measure, we count the individuals
as having some say in the matter, if the respondent reports more than one person, including himself
or herself, as the decision-makers. We use a count measure of the number of domains in which an
individual has complete or some say in the matters. In addition to the general measure that
aggregates our decision-making variable over all 18 questions, we also define similar measures of
financial decision-making using eight questions related to financial matters. As reported in table
3.14, we find that the probability of a woman cooking with LPG (or solid fuels) before the program
was significantly and positively (negatively) associated with the decision-making power of
women.
Among other correlated, working status of a woman was also associated with a higher likelihood
of cooking with LPG. Since the subsidy program increased female labor force participation., we
might expect that the program to have increased the decision-making power of women in the

132
exposed households. We examine the possibility in table 3.15. Women affected by the program
report an increase in their decision-making power, especially in financial matters. This change in
decision-making power is, quite possibly, a result of increased work-force participation of women.
If the unwillingness of the husbands to pay for LPG was, in fact, a reason that explained low
adoption of the fuel, the increase in labor force participation and decision-making power of
women, especially in financial matters, might ensure that they buy the beneficial technology on
their own even in the absence of the subsidy.

3.6. Conclusion
In an attempt to reduce the subsidy burden of kerosene, the Indonesian government sought to
replace it with subsidized LPG. Cooking with LPG is less time consuming than cooking with
kerosene or solid fuels. Previous research has found that modern time-saving household
technologies have implication on female labor force participation. Consistent with this, we find
large impacts on the female labor force participation of women exposed to the LPG subsidy
program. The results reinforce the effectiveness of relatively inexpensive policy incentives for the
adoption of modern household technology in ensuring greater integration of women in the labor
force.
We explore two possible pathways through which a switch to LPG for cooking might have
affected labor force participation of women - better health and time saving. We rule out the health
mechanism but do not have adequate data to verify the time-saving mechanism. Based on previous
research on the topic, we posit that the time-saving mechanism might have been operation. We
leave a more rigorous examination of this mechanism to future research. We show that the program
had benefits for the entire households, and not just for women. Household expenditure on food

133
items increased significantly. Women were more optimistic, less worried, happier, and had more
decision-making power within the household, especially in financial matters.

The results have important implications on the cost-benefit analysis of the programs of the kind.
Focusing on the health alone might underestimate the benefits of such programs. The recent
developments in consumer technologies have been impressive not only in their pace but also in
the increasing number of feature they incorporate. A comprehensive analysis of the benefits of any
such technology should examine the effects on a number of dimensions of well-being. Another
important take away pertains to private incentives to adopt modern technology. Even in situations
where the private benefits of adoption might surpass the cost for a house-hold, intra-household
externalities and differences in preferences within the household might hinder adoption. We must,
therefore, revisit the question of low adoption of welfare-enhancing technology and evaluate the
extent to which difference in preferences of the potential beneficiaries can explain the puzzle.
Temporary subsidies that mitigate externalities might go a long way in solving the low-adoption
problem in such contexts.
Our analysis leaves a lot to be desired. An direct examination of the causal analysis of the impact
of the decision-making power with women on the adoption of modern technology is essential in
the identification of possible virtuous cycle of greater adoption and welfare. Similarly, an
understanding of the pathways through which technologies such as cooking with LPG affects labor
force participation of women is of crucial importance for designing policies aimed at improving
female labor force participation. Due to data limitation, we leave this to future research.

134

3.7. Tables and Figures

Table 3.1 : Summary Statistics (Data: IPC and SUPAS)
1995 2000 2005 2010

Observations 718,837 20,112,539 1,090,892 20,337,271
Number of households 166,033 5,124,971 266,732 5,364,132
Number of districts 200 267 258 206
Number of provinces 17 26 25 18
Mean (S.D. in brackets)
Kerosene usage rate 0.35 NA 0.42 0.09
[0.48] [0.49] [0.29]
LPG usage rate in 0.06 NA 0.09 0.51
[0.24] [0.28] [0.50]
Labor force participation rate of men 0.53 0.55 0.53 0.72
[0.50] [0.50] [0.50] [0.45]
Labor force participation rate of women 0.30 0.38 0.30 0.65
[0.46] [0.49] [0.46] [0.48]
Notes: Information on cooking fuel was not collected during the IPC of 2000. The SUPAS did not interview the
province of Aceh due to the 2004 Indian Ocean earthquake and tsunami that affected the province.

135
Table 3.2: Summary Statistics (Data: IFLS)
2000 2007 2014

Observations 20,729 21,487 23,226
Number of households 7,360 8,224 8,816
Number of communities 311 310 311
Number of kecamatan 29 30 30
Number of provinces 15 15 15
Mean (S.D. in brackets)
Kerosene usage rate 0.49 0.40 0.05
[0.50] [0.49] [0.22]
LPG usage rate in 0.12 0.16 0.69
[0.33] [0.36] [0.46]
Labor force participation rate of men 0.74 0.76 0.77
[0.43] [0.42] [0.42]
Labor force participation rate of women 0.46 0.43 0.41
[0.50] [0.50] [0.49]

136
Table 3.3 : Impact on household's LPG usage status (Data: IPC an SUPAS)

(1) (2) (3) (4) (5)
Primary cooking fuel is LPG

Post × Treat 0.37*** 0.37*** 0.37*** 0.37*** 0.20***
Post × Treat × Pre-program kerosene usage rate
(0.06) (0.03) (0.05) (0.03) (0.05)
0.39***
(0.11)
District FE NO NO YES YES YES
Province FE YES YES NO NO NO
Year FE YES YES YES YES NO
Province-year FE NO NO NO NO YES
Year FE by pre-program LPG usage NO NO NO NO YES
SE Clusters Province District Province District District
Mean of DV 0.47 0.47 0.47 0.47 0.48
District-level mean of the pre-program kerosene
usage rate 0.43
Observations 21,971,944 21,971,944 21,971,944 21,971,944 21,393,142
Note: *p < 0:10, **p < 0:05, ***p < 0:01. Robust standard errors in parentheses are clustered at the level of
the district.

137
Table 3.4 : Impact on household’s LPG usage status (Data: IFLS)
(1) (2) (3)
Primary cooking fuel is LPG

Post × Pre-program kerosene usage rate 0.40*** 0.46*** 0.44***
(0.042) (0.052) (0.055)
Household FE NO YES YES
Community FE YES NO NO
Year FE YES YES NO
Province-year FE NO NO YES
Mean of DV 0.32 0.32 0.32
Community-level mean of the pre-program LPG usage rate 0.48 0.48 0.48
Observations 24,564 24,564 24,564
Note: *p < 0:10, **p < 0:05, ***p < 0:01. Robust standard errors in parentheses are clustered at the level of the
community

138
Table 3.5 : Impact on labor force participation status (Data: IPC and SUPAS)
(1) (2) (3) (4)
VARIABLES Labor force participation indicator

Post × Treat 0.33*** 0.20*** -0.03 -0.10***
Post × Treat × Female
(0.01) (0.01) (0.02) (0.02)
0.26*** 0.15***
Post × Treat × Pre-program kerosene usage rate
(0.01) (0.03)
0.10*** -0.02
Post × Treat × Female × Pre-program kerosene usage rate
(0.04) (0.04)
0.24***
(0.05)
District FE YES YES YES YES
Year FE YES YES YES YES
Province-year FE YES YES YES YES
Mean of DV 0.57 0.57 0.57 0.57
Pre-program kerosene usage rate 0.43 0.43
Observations 42,247,030 42,247,030 41,424,338 41,424,338
Note: *p < 0:10, **p < 0:05, ***p < 0:01. Robust standard errors in parentheses are clustered at the level of the
district.

139
Table 3.6 : Impact on labor force participation status (Data : IFLS)
(1) (2) (3) (4) (5) (6)
Primary activity Activities past week

work for pay housekeeping work for pay
work with
w/o pay
housekeeping job search

Post × Pre-program kerosene rate 0 .03 0.00 0.03 0.03 -0.06** 0.01
Post × Pre-program kerosene rate ×
Female
(0.02) (0.01) (0.02) (0.02) (0.03) (0.01)
0.07* -0.07** 0.02 0.04 0.02 0.00
(0.04) (0.03) (0.03) (0.03) (0.03) (0.01)
Estimated Effect for females 0.10*** -0.07*** 0.05* 0.06** -0.04* 0.01**
(0.03) (0.03) (0.03) (0.03) (0.02) (0.01)
p-value for females 0.00 0.04 0.06 0.01 0.08 0.04
Community Fixed-effect YES YES YES YES YES YES
Province-Year Fixed-effect YES YES YES YES YES YES
Mean of DV 0.59 0.24 0.64 0.65 0.47 0.04
Pre-program kerosene usage rate 0.49 0.49 0.49 0.49 0.49 0.49
Observations 63,633 63,633 63,838 65,341 63,841 63,837
Note: * p < 0.10, **p < 0.05, ***p < 0.01. Robust standard errors in parentheses are clustered at the level of the
community.

140
Table 3.7 : Impact on labor force participation in previous years (Data : IFLS)
(1) (2) (3) (4) (5) (6)
Ever held a job in the previous
year two years three years four years five years six years

Post × Pre-program kerosene usage rate

0.02 0.02 0.01 0.01 0.01 0.01
(0.02) (0.02) (0.02) (0.02) (0.02) (0.02)
Post × Pre-program kerosene usage rate × Female
0.06** 0.06** 0.07*** 0.07*** 0.07*** 0.07***
(0.03) (0.03) (0.03) (0.03) (0.03) (0.03)
Estimated Effect for females

0.07*** 0.08*** 0.08*** 0.08*** 0.08*** 0.08***
(0.02) (0.02) (0.02) (0.02) (0.02) (0.02)
p-value for females

0.00 0.00 0.00 0.00 0.00 0.00
Community Fixed-effect

YES YES YES YES YES YES
Province-Year Fixed-effect

YES YES YES YES YES YES
Mean of DV

0.72 0.73 0.74 0.75 0.75 0.76
Pre-program kerosene usage rate

0.49 0.49 0.49 0.49 0.49 0.49
Observations

65,341 65,341 65,341 65,341 65,341 65,341
Note: *p < 0:10, **p < 0:05, ***p < 0:01. Robust standard errors in parentheses are clustered at the level of the
community.

141

Table 3.8 :Impact on measured health (Data: IFLS)
(1) (2) (3) (4) (5) (6) (7)

Max.lung
capacity
BMI< 18 BMI≥25
Grip
strength
Pulse
Systolic
BP
Diastolic
BP
Post × Pre-program kerosene usage rate -5.15 0.00 0.02*** -0.34 2.09*** 1.21 0.32
Post × Pre-program kerosene usage rate ×
Female
(6.24) (0.01) (0.01) (1.25) (0.59) (0.74) (0.51)
2.27 0.00 0.01 0.48 -1.34* -1.79* -0.59
(5.61) (0.02) (0.01) (0.61) (0.71) (0.99) (0.66)
Estimated Effect for females
-2.88 0.00 0.03** 0.13 0.74 -0.58 -0.27
(4.34) (0.01) (0.01) (1.22) (0.62) (0.90) (0.60)
p-value for female 0.51 0.86 0.01 0.91 0.23 0.52 0.66
Community Fixed-effect YES YES YES YES YES YES YES
Province-Year Fixed-effect YES YES YES YES YES YES YES
Mean of DV 341.70 0.14 0.06 23.82 78.15 128.87 79.92
Pre-program kerosene usage rate 0.49 0.49 0.49 0.49 0.49 0.49 0.49
Observations 65,502 54,326 54,326 41,296 62,324 62,254 62,254
Note: *p < 0:10, **p < 0:05, ***p < 0:01. Robust standard errors in parentheses are clustered at the level of the community.

142

Table 3.9 : Impact on reported diagnosis of health conditions (Data: IFLS, for age above 40 only)
(1) (2) (3) (4) (5) (6) (7) (8) (9)

Hypertension Diabetes TB Asthma
Other lung
conditions
Heart
conditions
Liver
problems Stroke Arthritis

Post × Pre-program kerosene
usage rate
-0.004 -0.014 0.003 -0.013 0.006 0.011 -0.002 0.006 -0.028
(0.023) (0.013) (0.008) (0.012) (0.011) (0.010) (0.007) (0.008) (0.018)
Post × Pre-program kerosene
usage rate × Female
-0.045 0.029* 0.004 0.003 0.001 -0.001 0.006 -0.005 0.052**
(0.032) (0.016) (0.010) (0.015) (0.013) (0.012) (0.008) (0.011) (0.026)
Estimated Effect for females
-0.049* 0.014 0.007 -0.010 0.007 0.010 0.004 0.001 0.024
(0.026) (0.011) (0.005) (0.010) (0.010) (0.009) (0.005) (0.008) (0.022)
p-value for females 0.059 0.181 0.160 0.323 0.470 0.272 0.355 0.946 0.272
Community Fixed-effect YES YES YES YES YES YES YES YES YES
Province-Year Fixed-effect YES YES YES YES YES YES YES YES YES
Mean of DV 0.21 0.04 0.01 0.03 0.02 0.03 0.01 0.02 0.11
Pre-program kerosene usage
rate 0.49 0.49 0.49 0.49 0.49 0.49 0.49 0.49 0.49
Observations 19,252 19,249 19,256 19,256 19,253 19,253 19,256 19,257 19,252
Note: *p < 0:10, **p < 0:05, ***p < 0:01. Robust standard errors in parentheses are clustered at the level of the community.

143

Table 3.10 : Impact on expenditure (Data: IFLS)
(1) (2) (3) (4)
Expenditure on

Food last week
†
Non-food last month
†
Non-food last year
†
Education last year
†
Post × Pre-program kerosene usage
rate 3.612*** -63.879 -319.554 7.756
(1.354) (65.243) (508.615) (56.029)
Community FE YES YES YES YES
Province-year FE YES YES YES YES
Mean of DV 24.66 123.95 610.49 236.23
Pre-program kerosene usage rate 0.48 0.48 0.48 0.48
Observations 24,564 24,564 24,564 24,564
Note: *p < 0:10, **p < 0:05, ***p < 0:01. Robust standard errors in parentheses are clustered at the level of the community.
†
Expenditure converted to USD according to the exchange rate at the time of each survey.

144

Table 3.11 : Impact on food items (Data : IFLS)
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)
Consumed [...] in the previous week

Sweet
Potatoes
Eggs Fish Meat Dairy
Green
Leafy
vegetables
Banana Papaya Carrot Mango Any fruit
Any
protein

Post × Pre-program kerosene usage rate 0.009 0.018 0.020 0.024 0.005 -0.029 0.105*** 0.065** 0.010 0.016 0.035 0.004
(0.039) (0.025) (0.022) (0.038) (0.031) (0.019) (0.031) (0.031) (0.041) (0.093) (0.034) (0.023)
Post × Pre-program kerosene usage rate
× Female 0.028 -0.013 0.010 0.010 0.005 0.019 0.008 0.028 0.010 0.007 0.025 0.003
(0.029) (0.021) (0.022) (0.026) (0.026) (0.019) (0.026) (0.028) (0.025) (0.026) (0.020) (0.022)

Estimated Effect for females 0.037 0.005 0.030 0.035 0.011 -0.010 0.113*** 0.093*** 0.020 0.022 0.060* 0.006
(0.037) (0.023) (0.024) (0.037) (0.027) (0.017) (0.031) (0.032) (0.044) (0.094) (0.031) (0.022)
p-value for females 0.312 0.832 0.222 0.353 0.690 0.549 0.000 0.004 0.648 0.811 0.056 0.771
Community FE YES YES YES YES YES YES YES YES YES YES YES YES
Province-year FE YES YES YES YES YES YES YES YES YES YES YES YES
Mean of DV 0.44 0.82 0.83 0.61 0.31 0.91 0.61 0.30 0.48 0.36 0.86 0.81
Pre-program kerosene usage rate 0.49 0.49 0.49 0.49 0.49 0.49 0.49 0.49 0.49 0.49 0.49 0.49
Observations 41,099 41,099 41,098 41,098 41,098 41,098 41,098 41,098 41,098 41,098 41,098 39,967
Note: p < 0:10, p < 0:05, p < 0:01. Robust standard errors in parentheses are clustered at the level of the community.

145
Table 3.12 : Impact on expenditure on fuel and utilities (Data: IFLS)
(1) (2) (3) (4) (5)

Fuel expenditure
last month
†

Ratio of fuel to
non-food
expenditure
Utility
expenditure last
month
†

Ratio of utility to
non-food
expenditure
Community-level
kerosene price per
liter
†

Post × Pre-program
kerosene usage rate
-3.088 -0.032** -5.926 -0.143*** -0.092***
(10.628) (0.014) (12.517) (0.046) (0.023)
Community FE YES YES YES YES YES
Province-year FE YES YES YES YES YES
Mean of DV 7.613 0.120 20.092 0.443 0.494
Pre-program kerosene
usage rate 0.479 0.479 0.480 0.480 0.528
Observations 17,066 17,018 24,439 24,342 928
Note: *p < 0:10, **p < 0:05, ***p < 0:01. Robust standard errors in parentheses are clustered at the level of the community.
†
Expenditure converted to USD according to the exchange rate at the time of each survey.

146

Table 3.13 : Subjective well-being (Data: IFLS)
(1) (2) (3) (4) (5) (6) (7)
On which economic step Concerned about Happiness
Today Five year ago
Five year
later
Standard of
living
Food
consumption
Health
status Scale (0-5)

Post × Pre-program kerosene
usage rate -0.001 -0.036 0.125* -0.020 -0.009 -0.029 0.009
(0.049) (0.053) (0.068) (0.020) (0.017) (0.022) (0.024)
Post × Pre-program kerosene
usage × Female -0.014 0.009 0.043 -0.020 -0.030 0.003 0.037
(0.046) (0.054) (0.061) (0.020) (0.019) (0.022) (0.028)
Estimated Effect for females -0.015 -0.027 0.168 -0.039 -0.039 -0.026 0.046
0.053 0.055 0.069 0.020 0.017 0.018 0.024
p-value for females 0.775 0.625 0.016 0.048 0.022 0.158 0.060
Community FE YES YES YES YES YES YES YES
Province-year FE YES YES YES YES YES YES YES
Mean of DV 2.897 2.713 3.576 0.187 0.126 0.172 2.988
Pre-program kerosene usage
rate 0.490 0.491 0.489 0.490 0.490 0.490 0.486
Observations 61,539 61,319 58,849 61,781 61,781 61,781 41,257
Note: *p < 0:10, **p < 0:05, ***p < 0:01. Robust standard errors in parentheses are clustered at the level of the
community.

147
Table 3.14 : Correlates of fuel choice and decision-making power of women in 2000 (Data: IFLS)
(1) (2) (3) (4) (5)
VARIABLES Cooking with Some say in

LPG kerosene solid fuel
all decisions
(Score out of 18)
Financial decisions
(Score out of 8)

Some say in all decisions
(Score out of 18) -0.004* -0.003 0.007***
(0.002) (0.003) (0.003)
Some say in financial decisions
(Score out of 8) 0.010** 0.010 -0.021***
(0.005) (0.008) (0.006)
Primary activity is work for pay 0.026*** -0.030*** 0.004 0.506*** 0.182***
(0.007) (0.009) (0.008) (0.091) (0.043)
Years of education 0.017*** -0.002 -0.015*** 0.087*** 0.054***
(0.001) (0.002) (0.001) (0.012) (0.005)
Head of the household 0.040*** 0.025 -0.075*** -2.785*** -1.113***
(0.014) (0.021) (0.017) (0.246) (0.103)
Wife of the head of the household 0.013 0.027* -0.044*** 10.596*** 4.342***
(0.011) (0.015) (0.012) (0.141) (0.062)
Household head is female -0.053*** 0.065** -0.011 1.106*** 0.503***
(0.020) (0.026) (0.018) (0.246) (0.105)
Community FE YES YES YES YES YES
Province-year FE YES YES YES YES YES
Mean of dependent variable 0.14 0.53 .33 7.92 3.22
Observations 8,766 8,766 8,766 8,766 8,766
Note: *p < 0:10, **p < 0:05, ***p < 0:01. Robust standard errors in parentheses are clustered at the level of the community.

148
Table 3.15 : Impact on decision-making power of women (Data: IFLS)
Complete say in Some say in

all decisions financial decisions all decisions financial decisions
(Score out of 18) (Score out of 8) (Score out of 18) (Score out of 8)

Post × Pre-program kerosene rate 0.03 -0.04 -0.49* -0.47***
(0.21) (0.11) (0.27) (0.14)
Post × Pre-program kerosene × Female
0.43 0.37** 0.80** 0.62***
(0.28) (0.15) (0.35) (0.18)
Community FE YES YES YES YES
Province-year FE YES YES YES YES
Mean of dependent variable 3.52 1.3 10.84 4.58
Pre-program kerosene usage rate 0.48 0.48 0.48 0.48
Observations 44,456 44,456 44,456 44,456
Note: *p < 0:10, **p < 0:05, ***p < 0:01. Robust standard errors in parentheses are clustered at the level of the
community. As an example, one of the questions asked to elicit financial decision-making power is “In your household,
who makes decisions about money for monthly savings?” Response options are respondent, spouse, son, daughter,
mother, father, etc.

149
Table 3.16 : Impact on household's amenities (Data: IPC and SUPAS)
(1) (2) (3) (4) (5)
Does your household have the access to the following?
Sewage System Electricity Piped Water Flush Toilet Finished Floor

Post × Treat × Pre-program kerosene usage
rate -0.00 0.06 -0.02 0.06 -0.09**
(0.06) (0.05) (0.07) (0.06) (0.03)
District FE YES YES YES YES YES
Province-Year FE YES YES YES YES YES
Mean of DV 0.54 0.94 0.16 0.60 0.88
Observations 21,548,424 21,550,574 21,551,010 21,547,060 21,527,414
Note: *p < 0:10, **p < 0:05, ***p < 0:01. Robust standard errors in parentheses are clustered at the level of the district.

Finished floor takes value ‘1’ if the house some kind of concrete, wood, or stone flooring. It is ‘0’ for earth floors.

150

Notes: Based on the World Bank national accounts data, and OECD National Accounts data files.
Figure 3.2 : Labor force participation in Indonesian and worldwide
Notes: Based on the World Bank national accounts data, and OECD National Accounts data files. GDP per capita in constant US$ terms.
Figure 3.1 : Trends in GDP and education in Indonesia

151

Notes: In some cases, the program was rolled out in different areas within a province in two consecutive years. However, we do not have
information on roll-out at a finer level. For this reason, we define a province to have received the program only once all areas within the
province were covered.
Figure 3.3 : Staggered rollout of the LPG subsidy program across provinces

152

Note: In some cases, the program was rolled out in different areas within a province in two consecutive years. However, we do
not have information on roll-out at a finer level. For this reason, we define all communities within a province to have received the
program only once all areas within the province were covered.
Figure 3.4 : Difference in LPG program roll-out across IFLS communities

153

Notes: We use information from the Indonesian Population Census (IPC) of 2010 and Intercensal Population Survey of Indonesia
(SUPAS) waves 1995 and 2005 for the figure. IPC 2000 does not contain information about household’s primary cooking fuel.
Figure 3.5 : Primary cooking fuel (Survey: IPC and SUPAS)

154

Notes: We use information from the Indonesian Population Census (IPC) of 2010 and Intercensal Population Survey of Indonesia
(SUPAS) waves 1995 and 2005 for the figure. IPC 2000 does not contain information about household’s primary cooking fuel.
Figure 3.6 : Primary cooking fuel by program exposure status (Survey: IPC and SUPAS)

155

Notes: We use information from the Indonesian Population Census (IPC) of 2010 and Intercensal Population Survey of Indonesia
(SUPAS) waves 1995 and 2005 for the figure. IPC 2000 does not contain information about household’s primary cooking fuel.
Figure 3.7 : Primary cooking fuel by pre-program kerosene usage (Survey: IPC and SUPAS)

156

Notes: We use information from the Indonesian Population Census (IPC) of 2010 and Intercensal Population Survey of Indonesia
(SUPAS) wave 2005 for the figure.
Figure 3.8 : Change in LPG usage by pre-program kerosene usage (Survey: IPC and SUPAS)

157

Notes: We use information from the third (2000), fourth (2007), and fifth (2014) waves of Indonesian Family Life Survey for the figure.
Figure 3.9 : Primary cooking fuel (Survey: IFLS)

158

Notes: We use information from the third (2000), fourth (2007), and fifth (2014) waves of Indonesian Family Life Survey for the figure.
Figure 3.10 : Change in LPG usage by pre-program kerosene usage (Survey: IFLS)

159

Notes: We use information from the Indonesian Population Census (IPC) of 2000 and 2010 and Intercensal Population Survey of Indonesia
(SUPAS) waves 1995 and 2005 for the figure.
Figure 3.11 : Labor force participation by program exposure status (Survey: IPC and SUPAS)

160
Conclusion

In this dissertation, I have presented how human capital is developed, maintained, and utilized,
and how the relevant policies closely interact with this process.
The first chapter explores the formation of human capital at an early stage in life. The
multidimensional analysis shows that different types of abilities are developed through different
types of investment. Through this study, I identified how non-cognitive skills are enhanced by
market provided institutional care, such as daycare centers or kindergartens. In contrast, the
cognitive skills are primarily improved by one-to-one interaction between child and mother. While
such a finding is subject to a specific social context, the result of the multidimensional approach,
compared with that of the unidimensional approach, demonstrates how ignoring other dimensions
of childhood development leads to false conclusions.
Furthermore, the findings in the first essay indicate that we may have been asking the wrong
question in the literature by assuming there is one type of investment that universally promotes
every dimension of childhood ability. The departure from such an assumption makes it reasonable
for the coexistence of different government policies, while it is still unclear why the government
uses different policies simultaneously.
The second essay discusses the Monitor-Alert system policy that counteracts the PM10 by
providing information about air pollution levels to the public. Compared to similar policies that
have been analyzed in previous studies, the key feature of this policy is its accuracy. However, it
is less accessible due to its main channel of information provision. The results from this study
indicated a significant drop in the number of visitors in tourist destinations, which shows that
individuals engage in practice avoidance behavior when informed about the dangers. Additionally,

161
there was an increase in the number of visitors on good days. By setting up the Monitor-alert
system resulted in risk-averse individuals discontinuing precautionary avoidance behavior when it
was unnecessary.
A noteworthy finding in the second chapter was that such avoidance behavior was more
significant for outdoor facilities, and those involving physical activities. This shows that people
fully utilize available information at their discretion. Furthermore, a distinct impact of the policy
was identified. Tourist destinations that attracted customers from better socio-economic
backgrounds suffered more from a lower number of visitors. The group that practiced such
avoidance behavior intensively was domestic Korean visitors and not foreign visitors.
The third chapter analyzes the consequences of introducing cleaner cooking fuel to Indonesian
families and examines whether this could further serve as an engine for the liberation of women
in developing countries. The policy was introduced with the appropriate infrastructure and
contributed to the adoption of new technology. This also led to a higher rate of women participation
in the labor market. Further looking into the mechanism that encourages women to become more
involved in the labor market, I found little evidence of significant improvement in women’s health.
However, the use of LPG is time-saving and efficient, increasing labor market participation.
Furthermore, I found that the benefit from increased food expenditure due to LPG adoption far
exceeded the cost of converting to LPG without any subsidy. This does not clarify the low rate of
adoption of LPG before the intervention by the government. There may have been an intra-
household externality in decision making, indicating that men make important financial decisions
while women are the primary beneficiaries of the cleaner cooking fuel. If this externality has been
the main impediment to adopting a new technology, it is essential to check whether this adoption
also alleviates such a discrepancy of interest. Chapter 3 supports this possibility. The introduction

162
of LPG has helped these women to become more vocal in important financial decisions. This
shows that our liberation engine not only has a short-run effect but may continue to help
households make better decisions in the long run.
In summary, the three essays presented in this dissertation demonstrate how human capital
interacts with family decisions in different stages. They also exemplify the importance of an
understanding of the nature of human capital. The relevant policies will be subject to the
complexity of human capital, including, but not limited to, its multidimensionality and
heterogeneous perception, and might result in unintended consequences if such complexity is
overlooked by policymakers.

163
Bibliography
Acemoglu, D., Autor, D. H., and Lyle, D. (2004). Women, war, and wages: The effect of female
labor supply on the wage structure at mid-century. Journal of Political Economy, 112(3):497–551.

Aguiar, M. and Hurst, E. (2007). Measuring trends in leisure: The allocation of time over five
decades. The Quarterly Journal of Economics, 122(3):969–1006.

Ailshire, J. A., & Clarke, P. (2014). Fine particulate matter air pollution and cognitive function
among US older adults. Journals of Gerontology Series B: Psychological Sciences and Social
Sciences, 70(2), 322-328.

Albanesi, S. and Olivetti, C. (2009). Home production, market production and the gender wage
gap: Incentives and expectations. Review of Economic Dynamics, 12(1):80–107.

Altonji, J. G. and Card, D. (1989). The effects of immigration on the labor market outcomes of
natives. Technical report, National Bureau of Economic Research.

Andadari, R. K., Mulder, P., and Rietveld, P. (2014). Energy poverty reduction by fuel switch-
ing: Impact evaluation of the LPG conversion program in Indonesia. Energy Policy, 66:436– 449.

164
Angst, F., Drerup, S., Werle, S., Herren, D. B., Simmen, B. R., and Goldhahn, J. (2010).
Prediction of grip and key pinch strength in 978 healthy subjects. BMC Musculoskeletal Disorders,
11(1):94.

Ashraf, N., Berry, J., and Shapiro, J. M. (2010). Can higher prices stimulate product use?
Evidence from a field experiment in Zambia. American Economic Review, 100(5):2383–2413.

ASTAE (2015). Clean biomass cookstoves in central Java, Indonesia: A quantitative market
analysis. Technical report, Asia Sustainable and Alternative Energy Program, The World Bank
Group.

Arceo, E., Hanna, R., & Oliva, P. (2016). Does the effect of pollution on infant mortality differ
between developing and developed countries? Evidence from Mexico City. The Economic Journal,
126(591), 257-280.

Attanasio, O., Low, H., and Sanchez´-Marcos, V. (2008). Explaining changes in female labor
supply in a life-cycle model. American Economic Review, 98(4):1517–52.

Bandiera, O. and Rasul, I. (2006). Social networks and technology adoption in northern
Mozambique. Economic Journal, 116(514):869–902.

Barnett, W. S. (1995). Long-term effects of early childhood programs on cognitive and school
outcomes. The Future of Children, 5(3):25–50.

165

Bartik, T. J. (1991). Who benefits from state and local economic development policies?

Baum II, C. L. (2003). Does early maternal employment harm child development? An analysis
of the potential benefits of leave taking. Journal of Labor Economics, 21(2):409–448.

Becker, G. S. (1967). Human capital and the personal distribution of income: An analytical
approach (No.1). Institute of Public Administration.

Bernal, R. (2008). The effect of maternal employment and child care on children’s cognitive
development. International Economic Review, 49(4):1173–1209.

Bianchi, S. M. (2000). Maternal employment and time with children: Dramatic change or
surprising continuity? Demography, 37(4):401–414.

Bleakley, H. (2007). Disease and development: Evidence from hookworm eradication in the
American South. Quarterly Journal of Economics, 122(1):73–117.

Bloom, D. E., Canning, D., Fink, G., and Finlay, J. E. (2009). Fertility, female labor force
participation, and the demographic dividend. Journal of Economic Growth, 14(2):79–101.

Blundell, R., Dias, M. C., Goll, D. A., & Meghir, C. (2019). Wages, experience and training of
women over the lifecycle (No. w25776). National Bureau of Economic Research.

166

Brilli, Y. (2013). Mother or market care? A structural estimation of child care impacts on child
development. Unpublished manuscript, Collegio Carlo Alberto, CHILD, and Catholic University
of Milan.

Brunekreef, B., Janssen, N. A., de Hartog, J., Harssema, H., Knape, M., and van Vliet, P. (1997).
Air pollution from truck traffic and lung function in children living near motorways. Epidemiology,
pages 298–303.

Budya, H. and Arofat, M. Y. (2011). Providing cleaner energy access in Indonesia through the
megaproject of kerosene conversion to LPG. Energy Policy, 39(12):7575–7586.

Buss, A. H., & Plomin, R. (1984). Theory and measurement of EAS. Temperament: Early
Developing Personality Traits, 84, 104.

Card, D. (2001). Immigrant inflows, native outflows, and the local labor market impacts of
higher immigration. Journal of Labor Economics, 19(1):22–64.

Card, D. (2009). Immigration and inequality. American Economic Review, 99(2):1–21.

Carneiro, P. M., & Heckman, J. J. (2003). Human capital policy.

167
Carneiro, P., Løken, K. V., & Salvanes, K. G. (2015). A flying start? Maternity leave benefits
and long-run outcomes of children. Journal of Political Economy, 123(2):365–412.

Caselli, F. (2005). Accounting for cross-country income differences. Handbook of Economic
Growth, 1:679–741.

Clark, R. D., Ramsbey, T. W., & Adler, E. S. (1991). Culture, gender, and labor force
participation. Faculty Publications, 61.

Cohen, J. and Dupas, P. (2010). Free distribution or cost-sharing? Evidence from a randomized
malaria prevention experiment. The Quarterly Journal of Economics, pages 1–45.

Conley, T. G. and Udry, C. R. (2010). Learning about a new technology: Pineapple in Ghana.
The American Economic Review, 100(1):35–69.

Costa, D. L. (2000). From mill town to board room: The rise of women’s paid labor. Journal of
Economic Perspectives, 14(4):101–122.

Cubas, G. (2016). Distortions, infrastructure, and female labor supply in developing countries.
European Economic Review, 87:194–215.

Cunha, F. & Heckman, J. (2007). The technology of skill formation. The American Economic
Review, 97(2):31.

168

Cunha, F., Heckman, J. J., & Schennach, S. M. (2010). Estimating the technology of cognitive
and noncognitive skill formation. Econometrica, 78(3):883–931.

Cutler, D. M., Glaeser, E. L., and Shapiro, J. M. (2003). Why have Americans become more
obese? Journal of Economic Perspectives, 17(3):93–118.

de V. Cavalcanti, T. V. and Tavares, J. (2008). Assessing the “engines of liberation”: Home
appliances and female labor force participation. The Review of Economics and Statistics, 90(1):81–
88.

Del Boca, D., Flinn, C., & Wiswall, M. (2013). Household choices and child development.
Review of Economic Studies, 81(1):137–185.

Doepke, M. (2004). Accounting for fertility decline during the transition to growth. Journal of
Economic Growth, 9(3):347–383.

Doyle, O., Harmon, C. P., Heckman, J. J., & Tremblay, R. E. (2009). Investing in early human
development: Timing and economic efficiency. Economics & Human Biology, 7(1):1–6.

Duflo, E., Kremer, M., and Robinson, J. (2008). How high are rates of return to fertilizer?
Evidence from field experiments in Kenya. The American economic review, 98(2):482– 488.

169
Dustmann, C., & Scho ¨nberg, U. (2012). Expansions in maternity leave coverage and children’s
long-term outcomes. American Economic Journal: Applied Economics, 4(3):190–224.

Fernandez, R. (2013). Cultural change as learning: The evolution of female labor force
participation over a century. American Economic Review, 103(1):472–500.

Fort, M., Ichino, A., & Zanella, G. (2019). Cognitive and non-cognitive costs of daycare 0–2
for children in advantaged families. Journal of Political Economy (forthcoming).

Foster, A. D. and Rosenzweig, M. R. (1995). Learning by doing and learning from others: Hu-
man capital and technical change in agriculture. Journal of Political Economy, 103(6):1176– 1209.

Foster, A. D. and Rosenzweig, M. R. (2010). Microeconomics of technology adoption. Annual
Review of Economics, 2(1):395–424.

Gagliardone, L. (2015). Women’s Allocation of Time in India, Indonesia, and China. The
Closing Circle, 21-29.

Gajate-Garrido, G. (2013). The impact of indoor air pollution on the incidence of life threatening
respiratory illnesses: Evidence from young children in Peru. The Journal of Development Studies,
49(4):500–515.

170
Galor, O. and Weil, D. N. (1996). The gender gap, fertility, and growth. American Economic
Review, 86(3):374–387.

Gharaibeh, N. S. (1996). Effects of indoor air pollution on lung function of primary school
children in Jordan. Annals of Tropical Pediatrics, 16(2):97–102.

Goldin, C. (2006). The quiet revolution that transformed women’s employment, education, and
family. American Economic Review, 96(2):1–21.

Goldin, C. et al. (1992). Understanding the gender gap: An economic history of American
women. OUP Catalogue.

Goldin, C. and Katz, L. F. (2002). The power of the pill: Oral contraceptives and women’s career
and marriage decisions. Journal of Political Economy, 110(4):730–770.

Graff Zivin, J., & Neidell, M. (2012). The impact of pollution on worker productivity. American
Economic Review, 102(7), 3652-73.

Grossman, M. (1972). On the concept of health capital and the demand for health. Journal of
Political economy, 80(2), 223-255.

Greenwood, J., Seshadri, A., and Yorukoglu, M. (2005). Engines of liberation. Review of
Economic Studies, 72(1):109–133.

171

Gurung, B. and Setyowati, A. (2016). Measuring time savings generated by the Indonesia
domestic biogas programme (IDBP).

Hanna, R., Duflo, E., & Greenstone, M. (2016). Up in smoke: the influence of household
behavior on the long-run impact of improved cooking stoves. American Economic Journal:
Economic Policy, 8(1), 80-114.

Hanushek, E. A., & Woessmann, L. (2008). The role of cognitive skills in economic
development. Journal of economic literature, 46(3), 607-68.

Heckman, J. J. (1976). Estimates of a human capital production function embedded in a life-
cycle model of labor supply. In Household Production and Consumption (pp. 225-264). NBER.

Heckman, J. J., & Kautz, T. (2012). Hard evidence on soft skills. Labour economics, 19(4), 451-
464.

Hoynes, H., Schanzenbach, D. W., and Almond, D. (2016). Long-run impacts of childhood
access to the safety net. American Economic Review, 106(4):903–34.

Hoynes, H. W. and Schanzenbach, D. W. (2009). Consumption responses to in-kind transfers:
Evidence from the introduction of the food stamp program. American Economic Journal: Applied
Economics, 1(4):109–39.

172

Hsin, A., & Felfe, C. (2014). When does time matter? Maternal employment, children’s time
with parents, and child development. Demography, 51(5):1867–1894.

Huh, K., Squires, J., Lee, S., & Lee, J. (2006). K-ASQ (Korean age & stages questionnaires)
parent recorded infant monitoring system. Technical report, Seoul: Seoul Community
Rehabilitation Center. (in Korean)

James-Burdumy, S. (2005). The effect of maternal labor force participation on child
development. Journal of Labor Economics, 23(1):177–211.

James Gauderman, W., McConnell, R., Gilliland, F., London, S., Thomas, D., Avol, E., Vora,
H., Berhane, K., Rappaport, E. B., Lurmann, F., et al. (2000). Association between air pollution
and lung function growth in southern California children. American Journal of Respiratory and
Critical Care Medicine, 162(4):1383–1390.

Jia, R., & Ku, H. (2019). Is China's Pollution the Culprit for the Choking of South Korea?
Evidence from the Asian Dust. The Economic Journal, 129(624), 3154-3188.

Joseph, O., Pailhé, A., Recotillet, I., & Solaz, A. (2013). The economic impact of taking short
parental leave: Evaluation of a French reform. Labour Economics, 25, 63–75.

173
Kan, X., Chiang, C.-Y., Enarson, D. A., Chen, W., Yang, J., and Chen, G. (2011). Indoor solid
fuel use and tuberculosis in China: A matched case-control study. BMC Public Health, 11(1):498.

Kwon, H. J., Cho, S. H., Chun, Y., Lagarde, F., & Pershagen, G. (2002). Effects of the Asian
dust events on daily mortality in Seoul, Korea. Environmental research, 90(1), 1-5.

Lam, N. L., Smith, K. R., Gauthier, A., and Bates, M. N. (2012). Kerosene: A review of house-
hold uses and their hazards in low-and middle-income countries. Journal of Toxicology and
Environmental Health, Part B, 15(6):396–432.

Lin, H.-H., Ezzati, M., and Murray, M. (2007). Tobacco smoke, indoor air pollution and
tuberculosis: A systematic review and meta-analysis. PLoS Medicine, 4(1):e20.

Martin, W. J., Glass, R. I., Balbus, J. M., and Collins, F. S. (2011). A major environmental cause
of death. Science, 334(6053):180–181.

Mehta, S. and Shahpar, C. (2004). The health benefits of interventions to reduce indoor air
pollution from solid fuel use: A cost-effectiveness analysis. Energy for Sustainable Development,
8(3):53–59.

Miguel, E. and Kremer, M. (2004). Worms: Identifying impacts on education and health in the
presence of treatment externalities. Econometrica, 72(1):159–217.

174
Miller, G. and Mobarak, A. M. (2013). Gender differences in preferences, intra-household
externalities, and low demand for improved cookstoves. Technical report, National Bureau of
Economic Research.

Nepomnyaschy, L., & Waldfogel, J. (2007). Paternity leave and fathers’ involvement with their
young children: Evidence from the American ECLS–B community. Work and Family, 10(4):427–
453.

Neidell, M. (2009). Information, avoidance behavior, and health the effect of ozone on asthma
hospitalizations. Journal of Human resources, 44(2), 450-478.

Park, E., (1995). Types of infant-mother affection and related factors. Unpublished Doctoral
Dissertation, Ehwa Women’s University, Seoul. (in Korean)

Permadi, D. A., Sofyan, A., and Oanh, N. T. K. (2017). Assessment of emissions of green-house
gases and air pollutants in indonesia and impacts of national policy for elimination of kerosene use
in cooking. Atmospheric Environment, 154:82–94.

Pitt, M. M., Rosenzweig, M. R., and Hassan, M. N. (2006). Sharing the burden of disease:
Gender, the household division of labor and the health effects of indoor air pollution in Bangladesh
and India. In Stanford Institute for Theoretical Economics Summer Workshop, volume 202.

175
Ponce, N. A., Hoggatt, K. J., Wilhelm, M., & Ritz, B. (2005). Preterm birth: the interaction of
traffic-related air pollution with economic hardship in Los Angeles neighborhoods. American
journal of epidemiology, 162(2), 140-148.

Preston, S. H. (1984). Children and the elderly: Divergent paths for America’s dependents.
Demography, 21(4):435–457.

Rasmussen, A. W. (2010). Increasing the length of parents’ birth-related leave: The effect on
children’s long-term educational outcomes. Labour Economics, 17(1):91–100.

Schaner, S. and Das, S. (2016). Female labor force participation in Asia: Indonesia country
study. Asian Development Bank Economics Working Paper Series, (474).

Silwal, A. R. and McKay, A. (2015). The impact of cooking with firewood on respiratory health:
Evidence from Indonesia. The Journal of Development Studies, 51(12):1619–1633.

Slama, K., Chiang, C. Y., Hinderaker, S., Bruce, N., Vedal, S., and Enarson, D. (2010). Indoor
solid fuel combustion and tuberculosis: Is there an association? The International Journal of
Tuberculosis and Lung Disease, 14(1):6–14.

Smith, K. R. (2010). What’s cooking? A brief update. Energy for Sustainable Development,
14(4):251–252.

176
Smith-Sivertsen, T., Diaz, E., Pope, D., Lie, R. T., Diaz, A., McCracken, J., Bakke, P., Arana,
B., Smith, K. R., and Bruce, N. (2009). Effect of reducing indoor air pollution on women’s
respiratory symptoms and lung function: The RESPIRE randomized trial, Guatemala. American
Journal of Epidemiology, 170(2):211–220.

Strauss, J., Witoelar, F., & Sikoki, B. (2016). The fifth wave of the Indonesia family life survey: overview
and field report (p. 5). Santa Monica California USA: RAND.

Sumpter, C. and Chandramohan, D. (2013). Systematic review and meta-analysis of the
associations between indoor air pollution and tuberculosis. Tropical Medicine & International
Health, 18(1):101–108.

Thoday, K., Benjamin, P., Gan, M., and Puzzolo, E. (2018). The mega conversion program from
kerosene to lpg in Indonesia: Lessons learned and recommendations for future clean cooking
energy expansion. Energy for Sustainable Development, 46:71–81.

Thomas, D., & Strauss, J. (1997). Health and wages: Evidence on men and women in urban
Brazil. Journal of econometrics, 77(1), 159-185.

Toft, L., Beaton, C., & Lontoh, L. (2016). International experiences with LPG subsidy reform.
International Institute for Sustainable Development.

177
Tsani, S., Paroussos, L., Fragiadakis, C., Charalambidis, I., & Capros, P. (2013). Female labour
force participation and economic growth in the south Mediterranean countries. Economics Letters,
120(2):323–328.

Viard, V. B., & Fu, S. (2015). The effect of Beijing's driving restrictions on pollution and
economic activity. Journal of Public Economics, 125, 98-115.

Waldfogel, J., Han, W. J., and Brooks-Gunn, J. (2002). The effects of early maternal
employment on child cognitive development. Demography, 39(2):369–392.

World Health Organization. (2014). WHO guidelines for indoor air quality: household fuel
combustion. World Health Organization.

Yeung, W. J., Sandberg, J. F., Davis-Kean, P. E., & Hofferth, S. L. (2001). Children’s time with
fathers in intact families. Journal of Marriage and Family, 63(1):136–154.

Zivin, J. G., & Neidell, M. (2009). Days of haze: Environmental information disclosure and
intertemporal avoidance behavior. Journal of Environmental Economics and Management, 58(2),
119-128.

Abstract (if available)

Abstract This dissertation contains three essays on the topic of human capital formation and family economics. Chapter 1, aims to discern how a mother’s time allocation and investment behavior toward a child impact early childhood development, in both cognitive and non-cognitive abilities. While the female labor market participation rate has been increasing in developing and developed countries, it has not been fully explored whether market-provided childcare can perfectly substitute mother’s care. To answer the question, to which extent market-provided childcare can replace mother’s care, this paper takes a structural approach by solving mothers’ dynamic optimization problem, and adopt SMD (simulated minimum distance) for the corresponding empirical strategy. From two sets of analyses using data from PSKC (Panel Study of Korean Children), this study extends existing literature’s exclusive focus on cognitive development to include multidimensional development, and reveals that market-provided childcare may be superior in certain non-cognitive dimensions, but not as much in cognitive development

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Essays on health and aging with focus on the spillover of human capital

PDF

Essays on development economics

PDF

Essays on family and labor economics

PDF

The EITC, labor supply, and child development

PDF

Essays in financial economics

PDF

Three essays on health & aging

PDF

Essays in development and experimental economics

PDF

Essays on development and health economics: social media and education policy

PDF

Intergenerational transfers & human capital investments in children in the era of aging

PDF

Selection and impacts of early life events on later life outcomes

PDF

Long-term impacts of childhood adversity on health and human capital

PDF

Essays in applied microeconomics

PDF

Essays on monetary policy and international spillovers

PDF

Three essays in education and urban economics

PDF

Essays in development economics

PDF

Essays in information economics and marketing

PDF

More knowledge, better plans: a study of heterogeneity in dementia prevalence and mid-life cognitive changes

PDF

Formal volunteering and human capital for youth in Senegal

PDF

Essays on innovation, human capital, and COVID-19 related policies

PDF

Three essays on cooperation, social interactions, and religion

Asset Metadata

Creator Yun, Jeonghwan (author)

Core Title Three essays on human capital and family economics

School College of Letters, Arts and Sciences

Degree Doctor of Philosophy

Degree Program Economics

Publication Date 07/16/2020

Defense Date 04/27/2020

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag Air pollution,avoidance behavior,childcare,childhood development,cognition,Decision making,female labor,female labor force,household technology,human capital,non-cognition,OAI-PMH Harvest,particulate matter,time saving

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Strauss, John (committee chair)

Creator Email 3rdclasseconopprentice@gmail.com,jeonghwy@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c89-331112

Unique identifier UC11664173

Identifier etd-YunJeonghw-8685.pdf (filename),usctheses-c89-331112 (legacy record id)

Legacy Identifier etd-YunJeonghw-8685.pdf

Dmrecord 331112

Document Type Dissertation

Rights Yun, Jeonghwan

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA