Close
About
FAQ
Home
Collections
Login
USC Login
Register
0
Selected
Invert selection
Deselect all
Deselect all
Click here to refresh results
Click here to refresh results
USC
/
Digital Library
/
University of Southern California Dissertations and Theses
/
Essays on development economics
(USC Thesis Other)
Essays on development economics
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
Essays on Development Economics
by
Yiwei Qian
A Dissertation Presented to the
FACULTY OF THE GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(Economics)
August 2021
Copyright 2021 Yiwei Qian
Acknowledgements
First and foremost, I am extremely grateful to my advisors, Prof. John Strauss, Dr. Scott
Rozelle, and Prof. Vittorio Bassi, for their invaluable guidance, nonstop support, and
tremendous patience during my PhD study. Their deep understanding of the subject and
unbounded passion for research have encouraged me and will continue to encourage me
in all of my academic work. They also educated me on the professional codes and trained
me a both independent and collaborative researcher.
In addition to my advisors, I would like to thank faculty members in my qualifying
exam committee and my dissertation committee, Prof. Geert Ridder, Prof. Yongxiang
Wang, and Prof. Simone Schaner. I thank them for their insights and help to allow me to
reach milestones in my PhD study.
I would like thank to other faculty members at the Department of Economics at the
University of Southern California (USC) for their advice and enlightenment. I also would
like to express my gratitude for the help from staff at the Department of Economics at the
USC and the financial support of the Department.
I am indebted to my undergraduate advisor, Prof. Yaojiang Shi at Shaanxi Normal
University, Xi’an, China, for inspiring me to study Development Economics and showing
me the vastly uneven development in China.
I would like to thank all of my fellow PhD students and my friends for their compan-
ionship and constant encouragement to get me through the hardships during my journey.
I also thank my long list of coauthors for their collaborative efforts and nonstop inputs.
Finally, I would like to thank my wife Jiusi (Josie) Xiao, my parents, and all family
members for unconditional support and faith in me.
ii
Table of Contents
Acknowledgements ii
List of Tables vi
List of Figures viii
Abstract x
Chapter 1: Introduction 1
Chapter 2: Parental Investment, School Quality, and the Persistent Benefits of
Intervention in Early Childhood 4
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2 Contextual Background and Previous Literature of Preschool Education . . 7
2.3 Experimental Design and Data Collection . . . . . . . . . . . . . . . . . . . 10
2.3.1 Sampling and randomization . . . . . . . . . . . . . . . . . . . . . . 10
2.3.2 Parenting Program . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3.3 Data Collection and Measurement . . . . . . . . . . . . . . . . . . . . 11
2.3.3.1 Measuring Infant Skills . . . . . . . . . . . . . . . . . . . . . 11
2.3.3.2 Measuring Parental Investment . . . . . . . . . . . . . . . . 15
2.3.4 Summary Statistics, Balance, and Attrition . . . . . . . . . . . . . . . 16
2.3.5 Estimation of Program Effects . . . . . . . . . . . . . . . . . . . . . . 17
2.4 Medium-Term Impact of the Parenting Intervention at School Entry . . . . 18
2.4.1 Average Treatment Effects on Infant Skills . . . . . . . . . . . . . . . 18
2.4.2 Average Treatment Effects on Parental Investment . . . . . . . . . . 19
2.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
Chapter 3: The Role of Social Interactions in Early Childhood Development: Ev-
idence from Rural China 33
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2 Background of the Playground Intervention . . . . . . . . . . . . . . . . . . 38
3.2.1 Intervention Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.3 Experimental Design and Data Collection . . . . . . . . . . . . . . . . . . . 40
3.3.1 Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.3.2 Experimental Design . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
iii
3.3.3 Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.3.3.1 Measuring Child Development . . . . . . . . . . . . . . . . 41
3.3.3.2 Measuring Parental Time Investment . . . . . . . . . . . . . 42
3.3.3.3 Measuring Parenting Awareness . . . . . . . . . . . . . . . 42
3.3.3.4 Socioeconomic Characteristics . . . . . . . . . . . . . . . . . 42
3.3.4 Summary Statistics and Balance Check . . . . . . . . . . . . . . . . . 43
3.4 Impact of Having Access to the Playground . . . . . . . . . . . . . . . . . . 44
3.4.1 Impact on Child Development . . . . . . . . . . . . . . . . . . . . . . 44
3.4.2 Impact on Parental Time Investment . . . . . . . . . . . . . . . . . . 45
3.4.3 Impact on Parenting Awareness . . . . . . . . . . . . . . . . . . . . . 45
3.5 The Role of Social Interactions . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.5.1 The Impact of Social Interactions with Households that Have Similar-
age Children . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.5.2 The Impact of Social Interactions with Household from the Older
Cohort . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.5.3 Robustness Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.5.4 Social Learning from Knowledgeable Caregivers . . . . . . . . . . . 50
3.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
Chapter 4: Identification of Causal Effects in Cluster Randomized Experiments
with Spillovers and Noncompliance 60
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.2 The Setup and Parameters of Interest . . . . . . . . . . . . . . . . . . . . . . 62
4.2.1 The Potential Outcome Framework Setup . . . . . . . . . . . . . . . 62
4.2.2 Parameters of Interest . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.3 Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.4 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.5 Empirical Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.5.1 The Empirical Specification of the ITT effect for Compliers and
Never-takers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.5.2 Estimates of ITT for Compliers and Never-takers . . . . . . . . . . . 72
4.5.3 Tests of the Equal-trend Assumption . . . . . . . . . . . . . . . . . . 73
4.5.4 The Comparison with the LATE in Cr´ epon et al. (2015) . . . . . . . . 74
4.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
Chapter 5: Conclusion 78
References 80
Appendices 89
A Appendix A: Distribution of Latent Skills in WPPSI-IV . . . . . . . . . . . . 90
B Appendix B: Measurement System . . . . . . . . . . . . . . . . . . . . . . . 94
C Appendix C: Preschool Options and Enrollment Choice . . . . . . . . . . . 110
D Appendix D: Additional Figures and Tables for Chapter 3 . . . . . . . . . . 113
E Appendix E: Non-identification of ATT without exclusion restriction . . . . 121
iv
F Appendix F: Comparison to Usual DID Estimator . . . . . . . . . . . . . . . 122
G Appendix G: Additional Tables for Chapter 4 . . . . . . . . . . . . . . . . . 124
v
List of Tables
2.1 Descriptive Statistics and Balance . . . . . . . . . . . . . . . . . . . . . . . . 27
2.2 Program Treatment Impact on Infant Cognitive Skills at School-Entry . . . 28
2.3 Program Treatment Impact on Infant Non-Cognitive Skills at School-Entry 29
2.4 Program Treatment Impact on Parental Investment at School-Entry . . . . . 30
2.5a Preschool and Teacher Characteristics . . . . . . . . . . . . . . . . . . . . . . 31
2.5b Preschool Quality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.1 Balance Check . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.2 The Impact of Having Access to Playground on Child Development . . . . 54
3.3 The Impact of Having Access to Playground on Parenting Awareness . . . . 55
3.4 Suggestive Evidence of Social Learning from Similar-age Children Households 56
3.5 Impact of Having Access to the Playground on Social Interactions between
Children . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.6 Suggestive Evidence of Social Learning from Other Knowledgeable Caregivers 58
3.7 Evidence of Caregivers of the Younger Cohort Learning from Other Caregivers 59
4.1 ITT(co) & ITT(nt) estimates with Baseline Covariates & Pair Fixed Effect . . 75
4.2 Tests of Parallel Trends in the Control Sample . . . . . . . . . . . . . . . . . 76
4.3 Comparison between ITT effect for Compliers and LATE . . . . . . . . . . . 77
A1 Descriptive Statistics and Balance Adjusted for Attrition . . . . . . . . . . . 93
B1 Measurement System Non-Cognitive Skills . . . . . . . . . . . . . . . . . . . 104
B2 EFA Factor Selection Parental Investment . . . . . . . . . . . . . . . . . . . . 105
B3 Estimated factor loadings time and material parental investment . . . . . . 106
B4 EFA Factor Selection Preschool Quality . . . . . . . . . . . . . . . . . . . . . 107
B5 Estimated Factor Loadings Preschool Quality . . . . . . . . . . . . . . . . . 109
C1 Summary Statistics of Preschool Options and Enrollment Choices . . . . . 112
D1 The impact of having access to playground on parental investment . . . . . 115
D2 Suggestive Evidence on Whether Social Interactions Affect Parental Investment116
vi
D3 Robustness Check of Social Learning from Other Similar-age Children
Households . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
D4 Suggestive Evidence of Social Learning from Other Knowledgeable Caregivers118
D5 Suggestive Evidence of Social Learning from Other Similar-age Children
Households (-6 to 0 cohort) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
D6 Suggestive Evidence of Social Learning from Other Knowledgeable Care-
givers (-6 to 0 cohort) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
G1 Summary Statistics of the Non-attrited Sample . . . . . . . . . . . . . . . . 124
G2 Summary Statistics of the Non-attrited Sample with High Probability to
Borrow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
G3 ITT(co) & ITT(nt) estimates with baseline covariates but no fixed effect . . . 126
G4 ITT(co) & ITT(nt) estimates with pair fixed effect but no baseline covariates 127
vii
List of Figures
2.1 Program Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2 Distribution Cognitive Skill Factors by Treatment Assignment. . . . . . . . 24
2.3 Distribution Non-Cognitive Skill Factors by Treatment Assignment . . . . . 25
2.4 Preschool enrollment Decision . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.1 The Differences in Assigned Interventions between Cohorts . . . . . . . . . 39
3.2 Overview of the Playground ECD Intervention . . . . . . . . . . . . . . . . 41
A1 Distribution WPPSI-IV Verbal Comprehension Skills by Treatment Assign-
ment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
A2 Distribution WPPSI-IV Visual Spatial Skills by Treatment Assignment . . . 91
A3 Distribution WPPSI-IV Fluid Reasoning Skills by Treatment Assignment . 91
A4 Distribution WPPSI-IV Working Memory Skills by Treatment Assignment . 92
B1 Distribution of task item difficulty levels WPPSI-IV Verbal Comprehension 97
B2 Item Characteristic Curves (ICCs) WPPSI-IV Verbal Comprehension . . . . 97
B3 Distribution of task item difficulty levels WPPSI-IV Visual Spatial . . . . . 98
B4 Item Characteristic Curves (ICCs) WPPSI-IV Visual Spatial . . . . . . . . . 98
B5 Distribution of task item difficulty levels WPPSI-IV Fluid Reasoning . . . . 99
B6 Item Characteristic Curves (ICCs) WPPSI-IV Fluid Reasoning . . . . . . . . 100
B7 Distribution of task item difficulty levels WPPSI-IV Working Memory . . . 101
B8 Item Characteristic Curves (ICCs) WPPSI-IV Working Memory . . . . . . . 102
B9 Scree Plot of Eigenvalues of Parental Investment . . . . . . . . . . . . . . . . 105
B10 Scree Plot of Eigenvalues of Preschool Quality . . . . . . . . . . . . . . . . . 107
C1 Geographic locations of enrolled preschools and program villages . . . . . 111
D1 Photos of Parenting Center Playground in rural villages of China (Zhong
et al., 2020) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
D2 Number of Children across Treatment Assignments by Age of Children
Prior to the Intervention . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
viii
D3 Percentage of Children across Treatment Assignments by Age of Children
Prior to the Intervention . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
ix
Abstract
Randomized controlled trials (RCTs) have been widely adopted in the subject of De-
velopment Economics to study the causal impact of developmental interventions. In
this dissertation, I empirically study the impacts of two developmental interventions
on the skill formation during early childhood —a crucial period for the human capital
development —using RCTs and theoretically identify the causal impact of developmental
interventions under a particular but rather common circumstance in RCTs. The first
chapter evaluates the medium-term impact of a six-month early childhood home-visiting
program on child outcomes in rural China. Two and a half years after completion of
the program, we find persistent intervention effects on child working memory —a key
skill of executive functioning which plays a central role in children’s development of
cognitive and socio-emotional skills. We also find that the program had persistent effects
on both parental time investments and preschool enrollment, with children in the treat-
ment group enrolling earlier and in better quality preschools. Our finding of improved
parental preschool selection in treatment villages points to an important intervention-
induced persistent shift in parental investment behavior which might lead to long-term
benefits over the life-cycle. The second chapter examines the role of social interaction
in child development using a random experiment of ECD intervention in rural China.
The intervention promotes social interactions by providing both infants and caregivers
of infants free access to a playground in rural villages. After one year, the intervention
improved the language skills of infants, by 0.15 standard deviation (SD) and the parenting
awareness of their caregivers, by 0.21 SD. Evidence suggests infants and their parents
benefited from the intervention through social interactions: children benefited from the
intervention through their own interactions with other similar-age children as well as
their caregivers’ interactions with caregivers of other similar-age children; caregivers
improved parenting awareness by learning from other experienced caregivers. in the third
chapter, we study the causal identifications in cluster randomized controlled trials when
x
there are spillovers across individuals within clusters through social interactions and/or
general equilibrium of clusters. Under these circumstances, we show the traditional local
average treatment effect (LATE) can no longer identify the average treatment effect of the
subpopulation that is treated. Instead, we propose an analogous causal estimator of LATE
in clustered randomized experiments with spillovers across individuals within clusters.
Under a mild difference-in-differences type assumption, we point-identify the local causal
effect for the treated in clustered randomized experiments with spillovers and one-sided
noncompliance. Furthermore, we can identify the indirect effect of interventions for the
subpopulation that is not treated, which can be used to test whether the spillover effect
exists. We illustrate our method in our empirical analysis of a microcredit program in
rural Morocco.
xi
Chapter 1
Introduction
Early childhood —the period when the malleability of brain development is highest —is a
critical period for human capital development. Delayed development during this period
prevents children from reaching their full potential (Walker et al., 2007). Yet, over 250
million children under 5 years old are at risk of not reaching their full developmental
potential due to poverty and under-stimulation (Black et al., 2017). Such widespread
developmental delays carry potential long-term consequences, as delays in early childhood
have been shown to negatively impact educational attainment and income and perpetuate
an intergenerational cycle of poverty (Cunha and Heckman, 2008; Currie and Almond,
2011; Knudsen et al., 2006).
In response to these concerns, a growing strand of literature has documented the effec-
tiveness of early childhood development (ECD) interventions in addressing the adversity
that those children face, both in developed countries and developing countries(Baker-
Henningham and L´ opez B´ oo, 2010). Recent studies of ECD programs in the short term
have found positive effects on cognitive and non-cognitive development in a number of
developing settings, including Colombia, Pakistan, and China (Yousafzai et al., 2014;
Attanasio et al., 2014; Sylvia et al., 2020a). Further, long-term follow-up research shows
interventions launched decades ago have promising effects on a wide range of adult health,
labor market, and social outcomes: increased college attendance, employment, and earn-
ings and reductions in teen pregnancy and criminal activity (Heckman et al., 2010; Walker
et al., 2011; Gertler et al., 2014; Campbell et al., 2014)
However, little is know about the mechanism through which the successful intervention
at the early age was translated into the gains in later adulthood outcomes. Studies that
have followed children in the years following the conclusion of parenting programs have
in fact shown substantial variation in these medium-run effects. Although some studies
1
have found improved outcomes in the medium run, others have shown patterns of large
increases in skills at the conclusion of interventions, followed by rapid fade-out (Barnett,
2011; Bailey et al., 2017). An important question, therefore, is: what factors are driving
variation in the sustained benefits of these programs and can programs be designed to
better produce long-term gains?
In Chapter 2, I attempt to shed some light on this question by exploring the medium-
term effects of a parenting program in rural China two and a half years after completion
as children enter schooling.
1
Besides focusing on improving caregiver-child interactions at home, a growing body of
research has explored the impact of interventions that promote social interactions —in-
teractions refer both to those between children and those between caregivers of different
households —in the community both through group-based intervention and by allow-
ing for informal interaction between different families in a common play area (hereafter
group-based ECD interventions). Although various studies have claimed that informal
social interactions outside of formal training sessions could contribute to the success of
such interventions (Hamadani et al., 2019; Aboud et al., 2013; Grantham-McGregor et al.,
2020), to this date no study has empirically examined the existence and magnitude of
program effects due to informal social interactions. Thus, understanding the role of social
interactions in early childhood development would have important policy implications
regarding the cost-effectiveness of group-based ECD interventions.
In Chapter 3, I attempt to explore the role of social interactions in early childhood
development by studying the impact of an intervention in rural China that provides
households with free access to an indoor playground to promote social interactions.
Specifically, I measure the overall impact of the playground intervention (that promotes
informal social interactions) on developmental outcomes and parenting skills. Then, I
examine which channels of social interactions (including peer influence between children
and social learning between caregivers) were responsible for the impacts, if any.
Being aware of the importance of social interactions in decision making and the ex-
tensive adoptions of cluster randomized experiments, I study the identification issues
in causal effect potential raised by social interactions in clustered randomized experi-
ments. Given partial compliance of treatment assignment in randomized experiments,
1
In the short term of the program, Sylvia et al. (2020a) measured substantial increases in cognitive skills
and improvements in infant skill development were accompanied by increases in both parental investment
and parenting skills.
2
the previous literature adopts local average treatment effect (LATE) to understand the
average treatment effect of the subpopulation that is treated, which requires that treatment
assignment does not affect the outcome of interests other than being treated. In cluster
randomized controlled trials (cluster RCTs), this assumption fails to hold when there are
spillovers across individuals within clusters through social interactions and/or general
equilibrium of clusters.
In Chapter 4, I propose an analogous causal estimator of LATE in clustered randomized
experiments with spillovers across individuals within clusters. Under a mild difference-in-
differences type assumption, we point-identify the local causal effect for the subpopulation
that is being treated in cluster randomized experiments with spillovers and one-sided
noncompliance.
Taken together, Chapter 2 and Chapter 3 study the important developmental interven-
tions that address human capital accumulation challenges in developing context. Given
the extensive applications of cluster RCTs in the subject of Development Economics, Chap-
ter 4 provides a novel but straightforward approach to address causal identification issues
for studies in the subject.
3
Chapter 2
Parental Investment, School Quality, and the Persistent
Benefits of Intervention in Early Childhood
1
2.1 Introduction
A growing body of research establishes the effectiveness of ECD interventions in addressing
early cognitive deficiencies in disadvantaged populations in both lower- and higher-income
countries. Recent studies of at-scale parenting interventions find sizeable impacts on ECD
outcomes at program completion in Colombia (Attanasio et al., 2014), Pakistan (Yousafzai
et al., 2014) and China (Sylvia et al., 2020a; Heckman et al., 2020). While most long-term
follow-up research was published nearly a decade ago and study interventions that were
launched long before that, these studies analyze the long-term intervention effects of
smaller-scale ECD efficacy trials and find promising results on a wide range of adult
health, labor market, and social outcomes: increased college attendance, employment, and
earnings and reductions in teen pregnancy and criminal activity (Heckman et al., 2010;
Walker et al., 2011; Gertler et al., 2014; Campbell et al., 2014).
Despite this evidence that at-scale parenting programs can improve child outcomes
in the short-run at their conclusion and that more intensive interventions can lead to
substantial long-run impacts, an unresolved question is what mechanisms drive long-
run effects. Studies that have followed children in the years following the conclusion of
parenting programs have in fact shown substantial variation in these medium-run effects.
Although studies have found improved outcomes in the medium run, others have shown
patterns of large increases in skills at the conclusion of interventions, followed by rapid
fade-out (Barnett, 2011; Bailey et al., 2017). In low- and middle-income countries, two
1
Coauthored with Lei Wang, Nele Warrinnier, Sean Sylvia, Orazio Attanasio, and Scott Rozelle
4
recent studies for instance have found different results in medium-run follow-ups. A
follow-up study of a parenting intervention in Pakistan finds that initial improvements in
early skills of two-year-old infants persist two years after program completion (Yousafzai
et al., 2016). In contrast, a follow-up study of a similar integrated ECD intervention in
Colombia finds that initial gains in early skills faded out two years after that program
(Andrew et al., 2018). An important question, therefore, is: what factors are driving
variation in the sustained benefits of these programs and can programs be designed to
better produce long-term gains?
In this study, we attempt to shed some light on this question by exploring the effects
of a parenting program in rural China two and a half years after completion as children
enter schooling. Loosely modeled on the Jamaican Ready to Learn intervention (Grantham-
McGregor et al., 1991), the program promoted ECD in rural China through a home-based
parent training intervention implemented by officials associated with China’s Family
Planning Commission (FPC) (Sylvia et al., 2020a). At the end of the initial intervention
when the treatment and control children were 30 to 42 months old, Sylvia et al. (2020a)
measured substantial increases in cognitive skills in children assigned to receive weekly
home visits. The study also found that the improvements in infant skill development were
accompanied by increases in both parental investment and parenting skills, particularly
among children who were more disadvantaged at baseline.
Two and a half years after the initial intervention, we find large persistent effects on
child working memory - a key skill of executive functioning which plays a central role in
children’s cognitive functioning, behavior, emotional control, and social interaction. We
further find that the observed increases in parental investment behavior at the completion
of the ECD intervention persisted two and a half years later as parents in the treatment
villages continue to spend more time with their children. Beyond parental investments in
the home, we also find that children in treatment villages enroll earlier and in better quality
preschools. Our finding of persistent effects on caregivers’ preschool enrollment decision
points to an important intervention-induced persistent shift in investment behavior which
might lead to long-term benefits over the life-cycle.
Our results imply an important role for preschool quality in sustaining longer-term
effects of early childhood interventions, particularly where caregivers have low levels
of human capital. Previous research in diverse settings has shown that it matters when
and where parents decide to enroll their offspring in school. The age at which children
5
start school affects both short- and longer-term educational attainment as well as health
and labor market outcomes (Elder and Lubotsky, 2009; Carlsson et al., 2015). Several
studies show that the returns to preschool are substantial, especially for disadvantaged
children who typically have less stimulating counter-factual home environments (Havnes
and Mogstad, 2015; Felfe et al., 2015; Herbst, 2017; Cornelissen et al., 2018). Recent
evidence suggests that not just the quantity but, more importantly, the quality of preschool
is key to child development (Araujo et al., 2016; Andrew et al., 2019; Nores et al., 2019).
Araujo et al. (2016) find that preschool quality has positive effects on student learning
and executive functioning skills at school entry in Ecuador. Both Andrew et al. (2019)
and Nores et al. (2019) find that improvements in preschool quality led to significant
improvements in cognitive and language skills for preschool children in Colombia. Other
studies find that returns to preschool quality are persistent and lead to increased earnings,
higher college attendance, and lower teen pregnancy (Chetty et al., 2011, 2014).
Our findings contribute to the literature on early childhood programs, particularly
the ”puzzle” of intervention fade-out in the medium run but evidence of long-run effects
on labor market outcomes and health in adulthood (Barnett, 2011; Bailey et al., 2017).
Several hypotheses have been put forward in the literature to explain this apparent puzzle.
It is possible that the treatment intensity typical in smaller efficacy trials is difficult to
replicate in larger-scale integrated interventions (Andrew et al., 2018). It also is plausible
that short-term intervention effects might appear to fade in the medium run but then
re-appear later in the life-cycle because medium-run evaluations are not measuring the
relevant skills (Heckman et al., 2013) or they are measuring the relevant skills, but are
doing so with large measurement error (Cunha et al., 2010; Johnston et al., 2014; Laajaj
and Macours, 2017). Measurement error would make it more challenging to observe
medium-term intervention effects if initial skill differences are small but nevertheless
accumulate over time and end up resulting in long-term differences in health and labor
market outcomes.
Our analysis highlights another potential explanation for mixed findings on medium-
term effects: that ECD intervention programs differ in their ability to permanently shift
investment behavior. Most early intervention studies fail to follow up their study partici-
pants, but for those that have, there has often been no information collected on subsequent
investments from parents or schools. Two notable exceptions are Andrew et al. (2018)
and Yousafzai et al. (2016), who collect information on parental investment behavior two
6
years after the end of the parenting interventions. Andrew et al. (2018) find that initial
improvements in time and material investment observed at the end of the parenting inter-
vention had disappeared. Yousafzai et al. (2016), on the other hand, found persistently
improved parenting practices and a more stimulating home environment two years after
the intervention had ended.
In contrast to parental investments at home, however, how interventions affect sub-
sequent decisions regarding school enrollment may be more fundamental to sustaining
longer-run effects, particularly among children who would be more likely to have low
levels of skills absent intervention. Some previous evidence for this comes from Currie and
Thomas (2000) which finds that initial gains in test scores for children who attended Head
Start in the U.S. faded out more rapidly for black students compared to whites because
they were more likely to subsequently select into lower-quality schools. Distinguishing
between parental home investments and investments through schooling decisions may
help to explain the apparent persistence puzzle if the observed short-term improvements
in skill formation observed in the literature are mainly driven by the increased cognitive
stimulation during the intervention while the long-term health and labor market benefits
result from a shift in parental investment behavior after the intervention. As children
age, schooling is likely to become relatively more important to skill development than
parental investments of time and resources at home. And for parents who believe that
investments are important but perceive investment of their own time to have low returns,
the perceived net benefits of schooling may become larger at an earlier age and increase
with school quality.
The remainder of the paper is structured as follows. In section 2, we provide the
contextual background of preschool education in rural China and previous literature on
preschool education. In section 3, we describe the experimental design and data collection.
In section 4, we report the findings of the medium-run follow-up impact evaluation of the
parenting intervention. Section 5 concludes.
2.2 Contextual Background and Previous Literature of Preschool
Education
Preschool education in China consists of three years of kindergarten education for children
aged from 3-year-old to 6-year-old before the compulsory primary school. Preschool
7
enrollment is considered for the most part as a voluntary decision for parents of age-
appropriate children. Parents can enroll children once the children reach three-year-old,
but also have the option of enroll their children when they are older.
Enrollment rates in preschools are increasing rapidly in China. From 2009 to 2016, the
gross national preschool enrollment rate increased from 50.1% to 77.4% Wu (2017).
2
In
2029, the central government declared that by 85% of children are receiving three years
of preschool education. However, less is known know about preschool enrollment in the
rural parts of China, where 63.78% of China’s populations live (NBS, 2015).
Although increased enrollment rates in preschool are promising, they are not a guar-
antee for increased human capital outcomes. Research on the returns to preschool show
mixed results. Several studies find the returns to expansion of universal preschool to be
positive (Berlinski et al., 2008, 2009; Havnes and Mogstad, 2011; Felfe et al., 2015) whereas
other studies find mixed or no impacts (Magnuson et al., 2007; Gupta and Simonsen, 2010)
or even negative impacts on child development (Baker et al., 2008, 2019).
3
This mixed evidence might result from substantial heterogeneity in the quality of
alternative care environments. Disadvantaged children typically have less stimulating
home environments and might therefore benefit more from universal preschool than the
average child. Several studies indeed find that returns to universal preschool are higher
for disadvantaged children (Havnes and Mogstad, 2015; Felfe et al., 2015; Herbst, 2017;
Cornelissen et al., 2018).
Another important source of heterogeneity in returns to preschool is the preschool
quality (Ulferts et al., 2019). Araujo et al. (2016) find that preschool quality has positive
effects on student learning and executive functioning skills at school entry. Other studies
find that returns to preschool quality are persistent and lead to increased earnings, higher
college attendance and lower teen pregnancy (Chetty et al., 2011, 2014). Taken together,
this evidence shows that the effectiveness of preschool depends on the counterfactual care
arrangement the program is substituting for and the change in learning quality it presents
(Cascio, 2015).
The importance of preschool quality might therefore be even more pressing in the
context of low and middle income countries as a large share of children grow up in
insufficiently stimulating home environments resulting in major risk of cognitive delays at
2
Enrollment rate is the ratio of the number of children in the kindergarten to the number of children
between 3-6 years old.
3
Elango et al. (2015) provides an extensive review of this literature.
8
school entry (Lu et al., 2016; Richter et al., 2017). A recent study in rural Shaanxi province
found that in most rural preschools, only the principals had participated in city-level
training, and most teachers had never participated in any training. Preschool teachers
in rural areas tend to be young, and the share of experienced teachers is small (Lai et al.,
2015).
Even within the rural areas in China, there is considerable heterogeneity in preschool
quality. Each township has one or two preschools, usually located in the township, that get
the majority of their funding through the government. Thanks to the support from public
finance, those preschools in the township are usually of higher-quality than preschools
located in the villages.
9
2.3 Experimental Design and Data Collection
2.3.1 Sampling and randomization
The study sample was selected from one prefecture located in a relatively poor province
located in Northwest China. The province ranks in the bottom half of provinces nationally
in terms of GDP per capita. The prefecture chosen for the study is located in a mountainous
and relatively poor region of the province. Nearly all of the people residing in this
prefecture are ethnically Han.
The research team used a systematic protocol to select the sample. As the first step
researchers selected townships from four nationally designated poverty counties in the
chosen prefecture. All townships in each county were included except the one township
in each county that housed the county seat. Within each township, administrative data
were used to compile a list of all villages reporting a population of at least 800 people.
Next, two villages were randomly selected from the list in each township. These exclusion
criteria were applied to ensure the sample villages had a sufficient number of children in
the target age range. All children in sample villages between 18 and 30 months of age
were enrolled in the study.
Before the start of the intervention, sample villages were randomly assigned to a
treatment (n=65) and control arm (n=66) and the randomization procedure was stratified
at the county level. Next, each parenting trainer was assigned a maximum of four families
chosen randomly from treatment villages to be enrolled in the program. In treatment
villages, this resulted in a sample of 212 children enrolled in the parenting intervention
and a remaining 79 that were not. In control villages, a total sample of 300 children was
enrolled.
2.3.2 Parenting Program
The parenting intervention, in essence, was a weekly home visiting program where par-
enting trainers trained caregivers to interact with their offspring through cognitively
stimulating and developmentally appropriate activities using a structured curriculum.
The teaching curriculum is based loosely on the Jamaican home visiting model (Grantham-
McGregor et al., 1991) and adapted by child development psychologists in China to the
local rural setting. During the weekly sessions, parenting trainers would introduce main
10
caregivers (typically, mother or grandmother) to the activity and assist caregivers to engage
in the activity with their child. At the end of each weekly session, the materials used for
that week’s activities (toys and books) were left in the household to be returned at the next
visit.
The 6-month long parenting intervention started in November 2014 and ended in
April 2015. The parenting trainers, selected by the FPC from among their cadres in each
township received an initial, one-week intensive training at the beginning of the program
which covered theories and principles of early childhood development, parenting skills,
and the curriculum. This initial training consisted of both classroom-based instruction
as well as field practice. Throughout the program, trainers received periodic training by
phone on curriculum activities which would vary according to the ages of children to
whom they were assigned. For more details about the intervention, we refer to Sylvia et al.
(2020a).
2.3.3 Data Collection and Measurement
A team of enumerators conducted a baseline survey in October 2014 and a follow-up survey
in May 2015 at the end of the parenting intervention.
4
In August 2017, the enumerator
team conducted a medium-term tracking survey to revisit all the households enrolled in
the initial randomized controlled trial (see Figure 2.1). Enumerators collected detailed
information on infant skills, parental investment behavior, and a wide range of household
and preschool characteristics.
2.3.3.1 Measuring Infant Skills
Infant skills are assessed using a battery of questions from the Weschler Preschool and
Primary Scale of Intelligence (WPPSI) and the Strengths and Difficulties Questionnaire
(SDQ). The Weschler Preschool and Primary Scale of Intelligence (Wechsler, 2012) is
designed to measure the cognitive development of pre-schoolers and young children. For
this study, we use the WPPSI-IV edition that has been translated into Chinese (Wechsler,
2014) and administered to children by trained enumerators. WPPSI-IV measures five
main skill domains:
i Verbal Comprehension Index: measures verbal comprehension and reasoning skills.
4
Details on the baseline and endline data collection procedure can be found in Sylvia et al. (2020a)
11
ii Visual Spatial Index: measures the ability to organise and understand visual parts
and information, assimilate visual and motor functions simultaneously, and see the
whole-part connection to objects.
iii Fluid Reasoning Index: measures the ability to utilize inductive reasoning (e.g. use
past observations to predict current situations)
iv Working Memory Index: measures the ability to balance focus and attention while
manipulating visual and auditory information in conscious awareness.
v Processing Speed Index: measures how fast a child can scan and differentiate visual
information.
Conceptually, WPPSI-IV is developed as an IQ test but in practice, several sub-domains
measure executive functioning (EF) skills as well. The Processing Speed Index of WPPSI is
related to the EF sub-domain of Information Processing whereas the Working Memory
Index of WPPSI-IV is related to the domain of Cognitive Flexibility of EF. We are not
aware of any study specifically linking WPPSI-IV to early executive functioning skills and
therefore assume that we are measuring a subset of executive functioning domains as well
as cognitive development more generally.
For index (i)-(iv) all administered test items increase in difficulty and the test is stopped
when the child can no longer provide a correct answer. Given this specific test structure, a
simple average score of all correctly answered items would provide a noisy measure of
underlying child ability. If all items are of identical difficulty, then the simple average is
the best estimate of a child’s underlying ability. However, to the extent that items differ
in their difficulty level, then a weighted average can provide a more precise estimate of
child ability by assigning higher weights to more difficult items. Hence, in the first step,
we estimate a two-parameter logistic IRT model which calculates the optimal weighted
average of all items taking into account response patterns. To fix ideas, we briefly discuss
how an IRT measurement system can help mitigate measurement error:
LetI
ij
define the performance measure for childi and itemj on test and let’s assume it
is determined as follows:
I
ij
=
j
+
j
i
+
ij
(2.1)
12
where
i
is childi’s latent skill for test and this is assumed to be independent from
the error term
ij
. In other words, we assume that a unidimensional skill is sufficient in
explaining a child’s response behavior on items in each sub-test We further assume that
a child’s response to an item is independent of his or her responses to other items after
conditioning on child latent skill.
The variableI
ij
is not observed to the enumerator or caregiver. Instead, we observe
I
ij
=1 ifI
ij
> 0 andI
ij
= 0 otherwise. We further assume that the measurement system is
invariant to treatment assignment. The estimated item-specific intercepts,
ˆ
j
, represent
the level of difficulty of itemj. The estimated parameter ˆ
j
represents the discrimination
ability of itemj. Hence, in the 2-parameter logistic IRT model, the probability of success
on an item j is a function of both the level of latent skill
i
and the difficulty level
and discrimination ability of itemj. We refer to Appendix B.1 for more details on the
two-parameter logistic IRT measurement model.
Each of the first four indexes (i)-(iv) of WPPSI-IV are administered by two separate
sub-tests which are described in more detail in Appendix B.1. We use the raw item data
for each sub-test to estimate a 2-parameter logistic IRT model as specified above.
5
We
estimate parameters
j
and
j
using maximum likelihood and integrate out the latent skill
i
. In a second step, we use empirical Bayes estimators of the latent skill
i
and take the
mean of the empirical posterior distribution of
i
, conditional on child’s item responses
I
j
, while imposing the estimated parameters ˆ
j
,
ˆ
j
. Appendix Figures A1-A4 in Appendix
A plot kernel density estimates of latent skill
i
for each sub-test in WPPSI-IV index
(i)-(iv). Next, we rank performance on each sub-test based on the estimated skill factor
scores
ˆ
i
. The final score for each of the four WPPSI-IV indexes (i)-(iv) is calculated as
the average rank-performance on the two sub-tests.
The last index of WPPSI-IV, (v) Processing Speed, is administered by two tests that
calculate the amount of time it takes a child to complete a task and therefore is not
dependent on increasing task difficulty as is the case for the four other WPPSI-IV indexes.
For this index, the score is calculated as the average time of test completion on both
sub-tests. All skill factors for index (i)-(v) are standardized non-parametrically for each
age-month group as infant skills mature rapidly over time.
6
Kernel density estimates of
5
Items with zero variance are excluded from the analysis and represent less than 1% of all item measures.
6
This standardization method is less sensitive to outliers and small sample size within age-category and
gives us normally distributed internally standardized scores with mean zero across each age-month group
(Attanasio et al., 2015).
13
the five infant skill distributions for control and treatment villages are plotted in Figure
2.2.
The Strengths and Difficulty Questionnaire (Goodman et al., 2000) is a 25-item carer-
reported instrument for the assessment of social, emotional, and behavioral functioning of
children and adolescents ages 2 to 17 years old. For this study, we use an SDQ questionnaire
that was translated and validated for the Chinese context (Du et al., 2008). SDQ measures
the following three non-cognitive sub-domains:
i Externalising behavior: measures behavioral problems that are manifested in chil-
dren’s outward behavior such as disruptiveness, hyperactivity, and aggressive behav-
ior.
ii Internalising behavior: measures behavioral problems affecting children’s internal
psychological environment such as withdrawn, anxious, and depressed behavior.
iii Pro-social behavior: measures positive behaviors, attitudes, and emotions directed
towards others.
All items for index (i)-(iii) are scaled on a 3-point likert scale (1 not true, 2 somewhat true, 3
certainly true). We estimate a dedicated measurement system relating all observed items to
a latent factor capturing the above three SDQ sub-domains:
I
ij
=
j
+
j
i
+
ij
(2.2)
withI
ij
the observed j
th
measure for child i;
j
the mean of the j
th
measure and
j
the
loading of the factor for measure j. The measurement error
ij
is the remaining proportion
of the variance in measure j that is not explained by the latent non-cognitive skill factor
i
and assumed to be independent and have a zero mean. We further assume again
that the measurement system is invariant to treatment assignment. The parameters of
the measurement system are estimated using maximum likelihood and can be found in
Appendix Table B1 in Appendix B.
We next use the estimated means and factor loadings from (3) to predict the three
latent non-cognitive skill factors,
i
, for each childi in the sample using the Bartlett
scoring method (Bartlett, 1937). The predicted non-cognitive skill factors are standardized
non-parametrically for each age-month group. Kernel density estimates of the three
non-cognitive skill distributions for control and treatment villages are plotted in Figure
2.3.
14
2.3.3.2 Measuring Parental Investment
Parental investment is measured on several dimensions. First, we measure the parent’s
decision to enroll his/her child into preschool, at what age parents decide to enroll their
child, and the quality of the preschool selection. enrollment rates and the enrollment age
distribution for control and treatment villages are plotted in Figure 2.4 in panel a and
panel c respectively. The preschool enrollment decision by rural Chinese parents might
present an important investment channel. We collect information on 165 preschools in
which children from control and treatment villages were enrolled at the time of the survey.
From the 429 children that enrolled in preschool, 221 are enrolled in village preschools
and 208 in township or county preschools.
7
Enumerators were sent out to survey both teachers and headteachers to collect informa-
tion on preschool characteristics, teacher characteristics as well as measures of structural
and process quality. Preschool characteristics are measured by the number of enrolled
pupils, the share of pupils receiving government need-based aid, and the tuition fees paid
by parents each semester.
8
Teacher characteristics are measured by age, gender, years of
experience on the job, salary, and educational attainment as well as whether teachers have
received any professional training in the past year.
Structural quality of preschools is measured by the pupil-teacher ratio, the number
of activity rooms, the size of the outdoor play area, and whether the preschool has a
designated playroom, exercise room, dormitories, and provides breakfast to pupils. To
assess process quality we collect information on whether teachers engage in a set of
teaching activities such as: reading books in class, organizing physical exercise activities,
art&music activities, and science activities. We also collect information on whether social,
and language skills are taught in class.
We use exploratory factor analysis (EFA) to guide us in the dimensionality reduction
of the preschool quality data and determine the optimal factor structure (Appendix B.4).
We derive a one-factor model which captures general preschool quality. The pattern
of estimated factor loadings (Appendix Table B5) shows that higher preschool quality
factor scores are associated with schools that are bigger in size, more likely to be located
in township or counties as compared to villages and have younger and more educated
7
We were not able to obtain the preschool information for 22 households who enrolled at preschools that
are out of enumeration area.
8
In 2016, the need-based aid in survey region covers up to 750 Yuan per year, equalling to about 110 USD
15
teachers that are more likely to have received a teacher training in the past year.
9
Higher
factor scores are also associated with larger indoor and outdoor space, the availability of
dormitories and breakfast, and several measures of process quality such as organizing
exercise and science activities and reading books in class. The distribution of preschool
quality for control and treatment villages is plotted in Figure 2.4.
Secondly, we measure parental investment at school-entry by asking parents how much
time and money they spend on their offspring. The main caregiver is asked how much time
he/she engaged in a set of child-rearing activities the previous day, including story-telling,
singing songs, interactive play activities, and/or how long the child spends watching
TV during the day. Caregivers are also asked to report how much money they spend on
children’s books, toys, clothes, and school expenditures in the last year.
10
. Exploratory
factor analysis indicates both time and material parental investment are best measured
using a one-factor measurement system (Appendix B.3) Estimated factor loadings of the
measurement system can be found in Appendix Table B3. Both investment factors are
standardized by the distribution of the control group.
2.3.4 Summary Statistics, Balance, and Attrition
Summary statistics and tests for balance across control and treatment groups during
baseline are shown in Table 2.1. Differences between study arms in individual child and
caregiver characteristics are insignificant. A joint significance test across all baseline
characteristics also confirms that the study arms are balanced.
11
Children in our sample are on average just over 24 months old at the start of the
program. Less than 5% of the sample children are born with low birth weight. A large part
of the children in our sample are firstborn in the family (60%). More than 80% of children
were ever breastfed and around 35% were breastfed for more than one year. More than
20% percent of sample children were anemic according to the WHO-defined threshold of
110 g/L. On average children were reported to be ill 4 days over the previous month.
12
At
9
A county in China usually consists of several townships and each township usually resides more than
10 villages. Township or county preschools typically offer higher process and structural quality hence the
decision to enroll offspring outside the village in a county or township preschools potentially captures an
important parental investment channel in rural China (Zhao and Hu, 2008; Hu et al., 2014).
10
School expenditures include both annual tuition fees and other school-related expenditures.
11
We test this by regressing treatment status on all baseline characteristics reported in Table 2.1 and test
that the coefficients on all characteristics were jointly zero. The p-value of this test is 0.529.
12
Caregivers were asked whether the child had suffered from fever, cough, diarrhea, indigestion, or
respiratory cold over the previous month.
16
baseline, around 40 percent of the sample is cognitively delayed with Bayley MDI scores
below 80 points, but few (10%) were delayed in their motor development. Around 30
percent of the children are at risk of social-emotional problems at baseline.
We also collected information on caregivers and families. Around 26 percent of
the sample receives social security support through the dibao, China’s minimum living
standard guarantee program, as reported in Panel B of Table 2.1. The biological mother is
the primary caregiver in only 60 percent of households, with grandmothers often taking
over child-rearing when mothers out-migrate to join the labor force in larger cities. We
find that slightly more than 70 percent of primary caregivers in the sample (mothers or
grandmothers) have at least 9 years of formal schooling. On average households report
being somewhat indifferent in their feelings toward the Family Planning Commission at
baseline.
Baseline statistics on parental inputs, shown in Panel C of Table 2.1, demonstrate that
caregivers engage in few stimulating activities with their children. Only 11% of caregivers
told a story to their child the previous day. Less than 5% read a book to their child (on
average households have only 1.6 books). Only around 1 in 3 caregivers report playing
with or singing to their child the previous day.
Overall attrition between November 2014 and May 2015 was less than 1 percent
and insignificantly correlated with treatment status. We define attrition as missing a
Bayley’s or Griffith outcome (depending on the age-cohort) measure at endline for children
with a Bayley baseline measure. Attrition at follow-up two and a half years after program
completion is 7.4% and balanced between treatment and control group. Appendix Table A1
in the Appendix A shows summary statistics of baseline characteristics of the non-attrited
sample and shows that our follow-up sample is also balanced on baseline characteristics.
2.3.5 Estimation of Program Effects
Given the random assignment of households into treatment and control groups, compar-
ison of outcome variable means across treatment arms provides unbiased estimates of
the effect of the parenting intervention on outcomes. However, to increase power (and
to account for our stratified randomization procedure) we condition our estimates on
randomization strata (Bruhn and McKenzie, 2009) and baseline values of the outcome
variable.
17
We use ordinary least-squares (OLS) to estimate the intention-to-treat (ITT) effects of
the parenting intervention with the following ANCOV A specification:
Y
ijt
=
1
+
1
T
jt
+
1
Y
ij(t1)
+
s
+
ij
(2.3)
whereY
ijt
is an outcome measure for childi in villagej at follow-up; T
jt
is a dummy
variable indicating the treatment assignment of villagej;Y
ij(t1)
is the outcome measure
for childi at baseline, and
s
is a set of strata fixed effects. We adjust standard errors for
clustering at the village level using the Liang-Zeger estimator.
2.4 Medium-Term Impact of the Parenting Intervention at School
Entry
2.4.1 Average Treatment Effects on Infant Skills
Medium-run average treatment effects of the parenting intervention measured on cognitive
and non-cognitive skills at school entry can be found in Table 2.2 and Table 2.3. We find
that the 6-months parenting intervention led to an 0.27 standard deviation increase
in working memory skills two and half years after program completion. For the other
subdomains of the WPPSI, we find no significant differences between the control and
treatment villages. Kernel density graphs of all latent skill measures by treatment status
are shown in Figure 2.2. The working memory skill distribution is clearly shifted to the
right for the treatment group and shows improvements have been achieved across the
entire ability distribution. A Kolmogorov-Smirnov (K-S) test confirms this visual evidence
as it rejects the equality of working memory skill distribution in the treatment and the
control group with a p-value of 0.004.
Table 2.3 reports medium-run average treatment effects of the parenting intervention
on non-cognitive skills. We find no significant differences in externalizing, internalizing,
and pro-social behavior between treatment and control groups. This is not entirely
surprising as we did not find any significant differences between the control and the
treatment group in socio-emotional skills at program completion either (Sylvia et al.,
2020a).
18
2.4.2 Average Treatment Effects on Parental Investment
At the time of the survey, 10% of children in our sample were not (yet) enrolled in preschool.
13
The preschool enrollment rates differ significantly by treatment status as shown in Table
2.4 (Panel A). We find that parents in treatment villages were significantly more likely to
enroll their offspring in preschool by the time of the survey. In treatment villages around
5% were not (yet) enrolled in preschool as opposed to 13% in control villages (Figure
2.4a). In addition, among enrolled children, children from the treatment villager are more
likely enrolled in the preschools located in townships (in comparison to located in villages)
(Figure 2.4b). We further find that parents in the treatment villages are also significantly
more likely to enroll their children at younger ages. Children in treatment villages are on
average enrolled 2 months earlier compared to children in control villages as shown in
Table 2.4.
14
Parents in treatment villages are also significantly more likely to enroll their
children in higher quality preschools. We find that children from treatment villages are
enrolled in preschools that score on average 0.30 SD higher on our preschool quality index.
Figure 2.4d plots the preschool quality distribution for control and treatment villages.
The distribution of the preschool quality factor for treatment villages is stochastically
dominant and a Kolmogorov-Smirnov (K-S) test rejects the equality of the two factors’
distributions with a p-value< 0.001.
15
Several mechanisms might drive this differential preschool selection decision between
parents from control and treatment villages. The parenting intervention might have
increased awareness of the importance of preschool quality for skill development and as
a result, increased the proportion of parents in treatment villages who decide to enroll
children more early and in better preschools. We find some evidence for this mechanism
13
We present descriptions of households’ preschool options and enrollment choices using the observed
preschool data in our sample in Appendix C.
14
As around 10% of children are not yet enrolled in school at the time of the survey, we have missing
information on their age of future enrollment. Therefore, we perform a bounding exercise to estimate
the program treatment effects on the age of preschool enrollment. We first assume that all children are
enrolled latest by the maximum age in the sample distribution which gives us an upper bound estimate of
the treatment effect as presented in Table 2.4 (point estimate: -2.085, std error: 1.018). Next, we assume
all remaining children enroll directly after the survey takes place, which allows us to estimate a lower
bound treatment effect (point estimate: -1.676, std error: 0.913). Both findings are statistically significant at
conventional levels.
15
Similarly, we bound the estimates of the program treatment effects on preschool quality. We first assume
that all remaining children are enrolled in worse quality preschool in the township where the households
reside which gives us an upper bound estimate of the treatment effect (point estimate:0.356, std error:
0.132). We assume all remaining children are enrolled in best quality preschool in the township where the
households reside which gives us an lower bound estimate of the treatment effect (point estimate:0.198, std
error: 0.126). The former is significant at the 1% level and the latter is marginally significant at 10% level.
19
as around 60% of children in the treatment villages are enrolled in township preschools
compared to 40% of children in the control villages (Figure 2.4b). Preschools located in
villages differ from preschools in townships in some important dimensions. As shown in
Table 2.5a-2.5b township preschools are significantly larger in terms of student enrollment
numbers but have a similar pupil-teacher ratio. Village preschools have around 7% more
students who receive government need-based aid and tuition fees are considerably cheaper
as parents pay on average 670 Yuan less each semester.
Teacher composition also varies somewhat between township and village preschools.
Teachers in village preschools are slightly older and have some more years of experience
on the job but are less educated. Around 25% of teachers in village preschools have a
bachelor’s degree compared to 53% in township preschools. Township preschool teachers
are also significantly more likely to have received teacher training in the past year. All
teachers in township preschools are female whereas around 14% of teachers in village
preschools are male. Preschool teachers earn on average 2200 Yuan per month, which is
equivalent to about 325 USD and salaries between the township and village teachers are
similar.
16
Township preschools not only have more educated teachers they also outperform
village preschools in physical capital or structural quality as shown in Table 2.5b. We
find that township preschools typically have more outdoor play area space and a larger
number of activity rooms in their facility. Township preschools are also more likely to
have a dedicated playroom, exercise room, and dormitories as well as provide breakfast.
Panel B of Table 2.5b compares measures of process quality between village and township
preschools. We find that township preschools also slightly outperform village level
preschools on this dimension of quality as teachers spend more time reading books in
class, organize exercise and science activities, art and music activities. Teachers in village
preschools spend a similar amount of time teaching social and language skills.
To the extent that rural Chinese parents are aware of these differences in human capital,
structural and process quality between village and township preschools the observed
enrollment differential might present an important medium-term change in parental
investment behavior resulting from the parenting intervention. An alternative explanation
for this observed differential is that parents in treatment villages were more likely to
out-migrate as a result of the parenting intervention. We check this but do not find that
16
The average USD to CNY exchange rate at the time of the survey is 6.76
20
households in treatment villages were more likely to have migrated between the end of
the intervention and the timing of the follow-up survey.
In Panel B we report the average treatment effects on parental time and material invest-
ment at school entry. At the end of the parenting intervention, we observed large increases
in parental time investment as a result of the intervention (0.69 SD) Sylvia et al. (2020a).
We find evidence in this follow-up study for the persistence of increased time investment
by parents as the overall parental time investment factor is 0.30 standard deviations higher
for the treatment group. We do not find any significant difference in parental material
investment. Taken together, these results indicate that the observed improvements in
parental investment behavior at the completion of the parenting intervention persisted
two and a half years later when children enter school.
2.5 Conclusion
This paper presents the medium-run impacts of a home-based parenting program deliv-
ered by cadres of China’s Family Planning Commission on child development and parental
investment at school-entry, two and a half years after program completion. We find large
persistent intervention effects on child working memory —a key skill of executive func-
tioning which plays a central role in children’s cognitive functioning, behavior, emotional
control, and social interaction. We further find that the observed change in parental
investment at the completion of the parenting intervention persisted at the time children
enrolled in preschool as parents in treatment villages continue to invest more time in
their children. Beyond parental investments in the home, we also find that children in
treatment villages enroll earlier and in better quality preschools. Our results imply an
important role for preschool quality in sustaining longer-term effects of early childhood
interventions, particularly where caregivers have low levels of human capital.
Our study has several limitations. First, the study took place in one disadvantaged
rural area in northwest China, and, therefore, the medium-run effectiveness of early
childhood interventions in other regions or for other populations might differ. The
fact that the population was all ethnically Han (which accounts for more than 90% of
China’s population), however, may mean that there is relevance for other areas of China.
Second, we measured a wide range of cognitive and non-cognitive skills but not all tests
administered might be developmentally relevant for the population under study and
21
may suffer from measurement error, especially a concern for caregiver-reported infant
skill assessments. Moreover, we might not measure all relevant skills and behaviors
and therefore not have a full picture of the benefits to child development and parental
investment of the intervention. Finally, we estimate effects two and a half years after the
conclusion of the intervention and a longer-run follow-up study of the children in the
study will be necessary to study whether benefits over the life-cycle can be sustained.
22
Figure 2.1: Program Overview
23
(a) Verbal Comprehension (b) Visual Spatial
(c) Fluid Reasoning (d) Processing Speed
(e) Working Memory
Figure 2.2: Distribution Cognitive Skill Factors by Treatment Assignment.
24
(a) Externalizing behavior (b) Internalizing behavior
(c) Prosocial behavior
Figure 2.3: Distribution Non-Cognitive Skill Factors by Treatment Assignment
25
(a) Preschool enrollment (b) Preschool enrollment Township
(c) Preschool enrollment Age (months) (d) Preschool Quality
Figure 2.4: Preschool enrollment Decision
26
Table 2.1: Descriptive Statistics and Balance
(1) (2) (3)
Control
(N=301)
Treatment
(N=212)
p-value
Panel A. Child Characteristics
(1) Age in months 24.46 24.45 0.747
(0.20) (0.22)
(2) Male 0.45 0.51 0.185
(0.03) (0.04)
(3) Low birth weight 0.04 0.04 0.774
(0.01) (0.01)
(4) First born 0.59 0.61 0.366
(0.03) (0.04)
(5) Ever breastfed 0.85 0.87 0.974
(0.03) (0.04)
(6) Breastfed 12 months 0.35 0.39 0.867
(0.05) (0.05)
(7) Anemia (Hb<110 g/L) 0.23 0.27 0.849
(0.03) (0.04)
(8) Days ill past month 4.32 4.55 0.620
(0.33) (0.37)
(9) Cognitive Delay (BSID MDI<80) 0.46 0.39 0.206
(0.04) (0.03)
(10) Motor Delay (BSID PDI<80) 0.12 0.10 0.476
(0.02) (0.02)
(11) Social-Emotional Problems (ASQ:SE>60) 0.25 0.28 0.401
(0.03) (0.03)
Panel B. Household Characteristics
(1) Social security support recipient 0.28 0.25 0.832
(0.03) (0.03)
(2) Mother at home 0.68 0.62 0.116
(0.04) (0.05)
(3) Caregiver education 9 years 0.72 0.74 0.487
(0.03) (0.04)
(4) Unfavourable perception of FPC 2.87 2.85 0.824
(0.06) (0.05)
Panel C. Parental Inputs
(1) Told story to child yesterday 0.11 0.11 0.960
(0.02) (0.02)
(2) Read book to child yesterday 0.05 0.04 0.872
(0.01) (0.01)
(3) Sang song to child yesterday 0.37 0.35 0.651
(0.03) (0.04)
(4) Played with child yesterday 0.34 0.34 0.996
(0.03) (0.03)
(5) Number of books in household 1.60 1.90 0.615
(0.24) (0.29)
P-values account for clustering within villages. Unfavourable perception of FPC is measured on a 5-point likert scale.
27
Table 2.2: Program Treatment Impact on Infant Cognitive Skills at School-Entry
Treatment effect
Point estimate Std. error P-value FDR q-value
Wechsler Preschool Scale of Intelligence (N=465)
Verbal Comprehension 0.095 (0.089) f0.285g f0.384g
Visual Spatial 0.100 (0.093) f0.285g f0.384g
Fluid Reasoning 0.081 (0.086) f0.347g f0.384g
Working Memory 0.272
(0.095) f0.005g f0.027g
Processing Speed 0.086 (0.091) f0.344g f0.384g
Note: In all regressions we control for strata (county) fixed effects, child gender and baseline develop-
mental outcomes. All skill factors are non-parametrically standardized for each age-month group. To
control the potential bias caused by multiple hypothesis testing, we report the rate (q-value) of false
discovery rate (FDR)—the proportion of false positives among all positive results. All standard errors
are clustered at the village level. Significance levels are as follows:
p < 0:1,
p < 0:05,
p < 0:01.
28
Table 2.3: Program Treatment Impact on Infant Non-Cognitive Skills at School-Entry
Treatment effect
Point estimate Std. error P-value FDR q-value
Strengths and Difficulties Questionnaire (N=461)
Externalizing behavior 0.120 (0.080) f0.136g f0.617g
Internalizing behavior 0.067 (0.098) f0.496g f0.617g
Pro-Social behavior 0.088 (0.091) f0.332g f0.617g
Note: In all regressions we control for strata (county) fixed effects and baseline developmental out-
comes. All skill factors are non-parametrically standardized for each age-month group. To control
the potential bias caused by multiple hypothesis testing, we report the rate (q-value) of false discov-
ery rate (FDR)—the proportion of false positives among all positive results. All standard errors are
clustered at the village level. Significance levels are as follows:
p < 0:1,
p < 0:05,
p < 0:01.
29
Table 2.4: Program Treatment Impact on Parental Investment at School-Entry
Treatment effect
Point estimate Std. error P-value FDR q-value
Panel A: Preschool Enrolment (N= 474)
Preschool enrollment 0.074
(0.030) f0.016g f 0.022g
Township preschool enrollment 0.184
(0.066) f0.006g
Preschool enrollment age (in month) -2.085
(1.018) f0.043g
Preschool quality 0.275
(0.127) f0.033g
Panel B: Parental Investment (N=466)
Time investment factor 0.298
(0.108) f0.007g f 0.022g
Material investment factor 0.110 (0.101) f0.280g f 0.107g
Note: In all regressions we control for strata (county) fixed effects and baseline developmental out-
comes. Around 10% of children are not yet enrolled at the time of the survey hence we assume they
will enrol by the maximum age in the sample distribution. Program treatment effects on township
preschool enrollment, preschool enrollment age, and preschool quality selection are conditional on be-
ing currently enrolled (N=429). Preschool information is missing for 22 households. We impute the
preschool quality index for those missing observations with the township average preschool quality
by the locations (township or village) of of preschools that households enroll the child. To control
the potential bias caused by multiple hypothesis testing, we report the rate (q-value) of false discov-
ery rate (FDR)—the proportion of false positives among all positive results. All standard errors are
clustered at the village level. Significance levels are as follows:
p < 0:1,
p < 0:05,
p < 0:01.
30
Table 2.5a: Preschool and Teacher Characteristics
(1) (2) (3)
Village Township p-value
Number of pupils 69.585 251.087 0.000
(8.777) (19.146)
Share pupils receiving government need-based aid 0.268 0.208 0.043
(0.021) (0.018)
Tuition fee per semester (Yuan) 607.926 1275.704 0.001
(25.942) (226.249)
Teacher age 36.128 33.377 0.061
(1.041) (0.936)
Teacher male 0.138 0.000 0.001
(0.036) (0.000)
Teacher experience 6.452 4.819 0.119
(0.814) (0.498)
Teacher monthly salary (Yuan) 2183.436 2158.086 0.892
(136.098) (114.954)
Share of teachers with bachelor degree 0.251 0.528 0.000
(0.036) (0.042)
Teacher training in past year 0.691 1.000 0.000
(0.048) (0.000)
N 94 71
Standard errors in parentheses.
31
Table 2.5b: Preschool Quality
(1) (2) (3)
Village Township p-value
Panel A: Structural Quality
Pupil-teacher ratio 21.964 20.169 0.406
(1.620) (1.208)
Number of activity rooms 1.862 6.333 0.000
(0.216) (0.454)
Outdoor play-area 368.426 668.913 0.017
(77.270) (100.475)
Preschool has play room 0.160 0.406 0.000
(0.038) (0.060)
Preschool has exercise room 0.149 0.464 0.000
(0.037) (0.060)
Preschool has dormitories 0.213 0.609 0.000
(0.042) (0.059)
Preschool provides breakfast 0.500 0.696 0.012
(0.052) (0.056)
Panel B: Process Quality
Teacher reads books in class 0.926 1.000 0.020
(0.027) (0.000)
Teacher organizes exercise activities 0.702 0.971 0.000
(0.047) (0.020)
Teacher organizes art& music activities 0.936 0.971 0.312
(0.025) (0.020)
Teacher organizes science activities 0.809 0.942 0.014
(0.041) (0.028)
Teacher teaches social skills 0.957 0.971 0.652
(0.021) (0.020)
Teacher teaches language skills 0.989 0.971 0.392
(0.011) (0.020)
N 94 71
Standard errors in parentheses.
32
Chapter 3
The Role of Social Interactions in Early Childhood
Development: Evidence from Rural China
3.1 Introduction
A growing body of literature has established that the early childhood environment sub-
stantially affects later life outcomes. Because of the high plasticity of the brain during
infancy, the quality of the early childhood environment has long-term impacts on later
schooling and employment outcomes (Cunha and Heckman, 2008; Currie and Almond,
2011; Knudsen et al., 2006). Among all components of the early childhood environment,
stimulation from interactions – such as talk, play, and other forms of responsive attention
– have been shown to be one of the most significant (Black et al., 2017).
In response to the importance of early childhood development (ECD) and the lack of an
interactive environment in developing contexts (Yue et al., 2017; Luo et al., 2019), a strand
of literature has been working on showing the effectiveness of interventions to improve
interactive parenting in low-income households (Heckman et al., 2020; Sylvia et al., 2020b;
Attanasio et al., 2014; Yousafzai et al., 2014). Home-based interventions that involve one-
on-one psychosocial training sessions between a parent and a community health worker
have been shown to positively impact child developmental outcomes through raising the
quality of caregiver-child interactions (?Attanasio et al., 2020).
Besides caregiver-child interactions at home, the social environment outside the home
could also contribute to ECD outcomes. Social interactions outside the home refer both
to those between children and those between caregivers of different households. Social
interactions could contribute to ECD as both the development of children and the parent-
ing practices of caregivers are often influenced by their peers (Hanna and Meltzoff, 1993;
33
McCabe, 2008). A growing body of research has explored the impact of interventions that
promote social interactions in the community both through group-based parenting inter-
vention and by allowing for informal interaction between different families in a common
play area (hereafter group-based ECD interventions) (Aboud et al., 2013; Hamadani et al.,
2019; Nahar et al., 2012; Fernald et al., 2017).
While a growing body of research has shown that group-based ECD interventions
can significantly increase interactive parenting skills and improve child developmental
outcomes (Aboud et al., 2013; Hamadani et al., 2019; Nahar et al., 2012; Fernald et al.,
2017), researchers have yet to identify the precise mechanism of their impact. While
the effects might be attributed to the structured parenting training sessions, informal
social interactions before and after class could also play an important role. A child’s
interactions with similar-aged peers could directly stimulate their development (hereafter
peer influence between children) (Hanna and Meltzoff, 1993; Eckerman et al., 1975;
Garner and Bergen, 2006; Eckerman and Peterman, 2001; Brown et al., 2001). Likewise, a
caregiver’s interactions with other caregivers could improve their parenting knowledge
and skills (hereafter social learning between caregivers), thereby indirectly influencing
child development (Law et al., 2002; McCabe, 2008).
If informal social interactions contribute to the effect of group-based ECD interventions,
this could be significant for two reasons. First, the unstructured components of such
interventions that provide opportunities for informal social interactions (such as free
access to a common play area) are relatively inexpensive, easy to implement, and scalable
when compared to the structured training sessions, which often involve many costs such as
the design of a curriculum and the hiring of staff (Grantham-McGregor et al., 2020; Aboud
and Yousafzai, 2015). Second, it is possible that informal social interactions could lead to
spillover effects, in which an intervention can improve ECD outcomes of the community
as a whole, even in households that do not participate in the parenting sessions. However,
as of yet, the literature has neither isolated the effect of the unstructured components of
these interventions nor demonstrated the roles of different channels of social interactions
(i.e., peer influence between children and social learning between caregivers) on ECD
outcomes.
To address these gaps in the literature, this paper examines the effects of the unstruc-
tured component of a group-based ECD intervention. The specific objectives of the study
were as follows. The first was to measure the overall impact of the unstructured part of
34
the intervention (that promotes informal social interactions) on developmental outcomes
and parenting skills. The second was to examine which channels of social interactions
(including peer influence between children and social learning between caregivers) were
responsible for the impacts, if any.
In pursuit of these objectives, I draw on data from a clustered randomized experiment
evaluating a group-based ECD intervention in northwest China. From a random sample
of 100 rural villages, we randomly assigned 50 to be in the treatment group and granted
all parents with infants less than six months old one year of free access to an indoor
playground located in a parenting center in each village for one year. Parents in the
remaining 50 control villages had no such access. The treatment group also shared the
indoor space with a separate cohort of caregivers with older toddlers between six and
eighteen months old (hereafter, the older cohort) who received structured psycho-social
parenting training at specific times during the week, though the treatment group caregivers
themselves (those with children younger than six months, hereafter the younger cohort)
did not attend the training.
I leverage two features of the intervention to study the role of social interactions in the
intervention. The first is the free access to the playground, which was intended to promote
informal social interactions between children as well as those between caregivers. Using
the random assignment of granting access at the village level, I evaluate the intention-to-
treat (ITT) effect of having access to the playground on child development and parental
practices of the younger cohort households (i.e., those who did not have access to the
parenting training).
The second feature is the psycho-social parenting training received by the older cohort
households, who were also potential sources from whom the younger cohort could learn
through informal social interactions. I use the random assignment and ex-ante variations
in cohort sizes in each village, including the cohort size in the younger cohort and that
in the older cohort, as potential links of social interactions to estimate the treatment
effect heterogeneity by cohort size. This estimation provides suggestive evidence for
the magnitude of the treatment effect through social interactions and the specific social
interaction channels through which the intervention has an effect, as larger cohort sizes
increase the opportunities for social interactions. In addition, since only one of the two
cohorts received psycho-social parenting training in each village, I take advantage of this
partial-population intervention design to estimate a linear-in-means peer effect model on
35
parenting practices to validate the role of social interaction between caregivers suggested
by the treatment effect heterogeneity by cohort size (Moffitt et al., 2001).
The results indicate that providing free access to an indoor playground had a positive
impact on both child development and parenting practices. The ITT estimates show that
granting free access to the indoor playground, on average, improved the language skills
of children by 0.15 standard deviations (SD). They also show that caregivers’ parental
awareness, an index summarizing knowledge and confidence of parenting practices,
improved by 0.21 SD during the one-year intervention.
I also find suggestive evidence regarding the role of specific social interaction channels,
showing that both informal social interactions between younger cohort households as
well as those between younger and older cohort households played important roles in
the impact of the intervention. An additional younger cohort household in the treatment
village increased the language skills of the overall younger cohort by 0.04 SD, suggesting
children benefited from peer influence and interactions between caregivers. Apart from
that, an additional older cohort household (who received psycho-social training) increased
the parenting awareness index of the younger cohort by 0.02 SD, indicating that the newer
caregivers learned from the more knowledgeable and experienced caregivers. Strong peer
effects in the parenting awareness index from the linear-in-means estimate indicated that
social learning between caregivers took place during the intervention.
This paper builds upon two important strands of literature. The first is on the im-
portance of social interactions, which have been found to result in spillover effects of
certain interventions such as technology adoption among adults (Munshi, 2004; Duflo and
Saez, 2003; Kremer and Miguel, 2007; Foster and Rosenzweig, 1995; Conley and Udry,
2010) as well as peer effects on academic outcomes for school-aged youth and adolescents
(De Giorgi et al., 2010; Calv´ o-Armengol et al., 2009; Agostinelli, 2018; Sacerdote, 2001).
1
However, there is a lack of literature demonstrating whether social interactions contribute
to spillover effects of interventions in the first few years of life. List et al. (2019) shows
large spillover effects of an intervention aimed to improve preschool-aged children’s skill
formation in Chicago. In a very different social context, my paper shows the role of social
interactions in ECD with a sample at an even younger age. My paper also illustrates how
social interactions affect parental practices of the caregiver when the child is in infancy,
demonstrating that informal social interactions that occur as a result of group-based
1
See Mobius and Rosenblat (2014) and Foster and Rosenzweig (2010) for review of social learning in
technology adoption. See Epple and Romano (2011) for review of peer effects in education.
36
parenting intervention may lead to spillover effects on both overall child development and
parental awareness in a community.
Second, my paper contributes to the growing literature on the “critical period” of
ECD.
2
Although many studies have shown that early investment during infancy has a
high return to human capital(Cunha and Heckman, 2008), most studies draw on samples
of children who are more than 1.5 years old (J-PAL, 2020). Less is known about the
impact of investment during the very first year of life due to limited data availability. One
related exception is Ozier (2018), who identified a large positive externality of school-
based deworming treatment on the cognitive skills of infants less than one year old in
the community. My paper adds to the literature showing that interventions designed to
improve interactive parenting could be effective at this early stage as well.
The results of this paper also have important policy implications regarding the cost-
effectiveness of group-based ECD interventions. The paper sheds light on the mechanisms
of effective group-based ECD interventions in recent literature (Aboud et al., 2013; Nahar
et al., 2012; Fernald et al., 2017; Grantham-McGregor et al., 2020), which until now have
not been explored in detail. Although various studies have claimed that informal social
interactions outside of formal training sessions could contribute to the success of such
interventions (Hamadani et al., 2019; Aboud et al., 2013; Grantham-McGregor et al.,
2020), to this date no study has empirically examined the existence and magnitude of
program effects due to informal social interactions. The current study gives preliminary
evidence that social interactions indeed play a role in the effectiveness of group-based
ECD interventions. Moreover, since social interactions usually lead to spillover effects
that are often unaccounted for, the effectiveness of group-based interventions might be
larger than what current studies have documented, which could change the calculation of
cost-effectiveness of group-based ECD interventions.
The rest of this paper is organized as follows. The next section describes the background
of the intervention. Section III presents the methodology. Section IV reports the results.
Section V explores evidence of the role of social interactions in the intervention. Section
VI discusses the study’s implications and concludes.
2
See Currie and Almond (2011) for the review of the ”critical period” in early childhood development.
37
3.2 Background of the Playground Intervention
Among the different approaches for improving the stimulation received by infants and
toddlers, group-based ECD interventions have become a a promising model (Aboud and
Yousafzai, 2015). In addition to reducing the labor and travel costs associated with home
visit programs (Aboud and Yousafzai, 2015), group-based programs have been shown to be
just as effective in raising child developmental outcomes (McLeroy et al., 2003; Merzel and
D’Afflitti, 2003; Grantham-McGregor et al., 2020). Studies of group-based ECD programs
have reported positive impacts on the cognitive and non-cognitive skill development of
infants and toddlers in Bangladesh, Indonesia, Brazil, Mexico, and Mozambique (Aboud
et al., 2013; Eickmann et al., 2003; Fernald et al., 2017; Martinez et al., 2012).
In light of the successes of these interventions in other developing settings, China’s
National Health Commission (NHC) worked with the Rural Education Action Program
at Stanford University to conduct a large-scale, clustered randomized experiment in 100
villages located in an underdeveloped rural region of northwestern China. The aim was to
evaluate the effectiveness of the group-based ECD model in vulnerable communities.
3
3.2.1 Intervention Details
In each of the sample villages, two cohorts were identified for two different research
agenda. First, children in the age range of 6-24 months and their caregivers (the older
cohort) were enrolled in the study to understand the overall impact of the intervention,
including the structured psycho-social parenting training and the unstructured time
for informal social interaction in the indoor playground. Second, households that had
children who were due in less than 6 months and those with infants up to 6 months old
(the younger cohort) were enrolled in the study to test the impact of the informal social
interactions that have long been theorized as being the major channel for the effectiveness
of group-based ECD (Hamadani et al., 2019; Aboud et al., 2013; Grantham-McGregor
et al., 2020). Thus, the older cohort had access to both the structured and unstructured
components of the intervention (they could both attend the training sessions and access
the indoor playground), while the younger cohort only had access to the unstructured
component (i.e., they could enter the indoor playground, but they could not attend the
psychosocial training – see Figure 3.1). In this paper, I focus on the second aim of the
3
The AEA RCT Registry of the experiment can be found here.
38
Older cohort Younger cohort
Intervention component (6-24 month-old) (< 6 month-old)
Parenting center playground yes yes
One-to-one parenting training yes no
Figure 3.1: The Differences in Assigned Interventions between Cohorts
intervention – that is, to explore the effects of having access to the playground on the
younger cohort.
In each village, a parenting center was built by the research team at a site provided
by the local village committee. The research team also hired one local center manager for
each parenting center who was responsible for managing all of the activities at the center.
The intervention in the parenting center consisted of two components. First, all
parenting centers included a large playground, as well as toys, age-appropriate books, and
decorations provided by the research team. The parenting centers were designed to be
open five hours a day and six days a week; 90% of parenting centers were open for the
designated frequency and time. Caregivers were encouraged to bring their children to the
parenting centers during open hours, but they were not allowed to leave their children
alone in the parenting centers (see Appendix Figure D1).
Second, each parenting center was staffed by two trained parenting experts from
the local Family Planning Commission who conducted weekly one-to-one sessions for
caregivers on interactive parenting practices (commonly referred to as psycho-social
training). The curriculum is based loosely on the Jamaican home visiting model (Grantham-
McGregor et al., 1991) and adapted by child development psychologists in China to the
local rural setting.
Figure 3.1 displays the partial-population intervention design of the parenting training
sessions. It is important to note that, though the children and caregivers in the younger
cohort were excluded from the parenting training sessions, they could still interact with
those from the older cohort who had access to the other component of the intervention.
Both households in both cohorts received one year of intervention as described above.
39
3.3 Experimental Design and Data Collection
3.3.1 Sampling
The intervention took place in 20 nationally-designated poverty counties in the Qinling
mountain region of northwestern China. To select the sample for the study, the research
team followed a three-step sampling protocol.
First, all townships (the middle level of administration between county and village) in
the 20 counties were selected to participate in the study, with two exceptions: the study
excluded the township in each county that housed the county seat (as these tend to be
wealthier and more urbanized than the average rural township), as well as townships
that did not have any villages with a population of 800 or more. After applying the two
exclusion criteria, the sample consisted of 100 townships.
Second, one village per township was randomly selected for inclusion in the study to
avoid cross-village spillover. To ensure that all sample villages would have the potential
space to conduct the community-based parenting intervention, villages that could not
supply a 60-80m
2
space for the intervention site were excluded. If a village did not have
the available space, it was replaced with another randomly-selected village from within
the same township. In total, 100 villages were included in the sample.
Third, once villages were chosen, the research team enrolled the individual households,
including the primary caregivers and their children. In each village, children in the age
range of 6-24 months and their caregivers were enrolled in the study as the older cohort,
and the households that either were expecting newborns in less than 6 months or already
had newborns up to 6 months old were enrolled in the study as the younger cohort.
3.3.2 Experimental Design
After sampling, the research team randomly assigned 50 villages to the treatment group
and 50 villages to the control group, with stratification at the county level. One parenting
center was built in each treatment village, and the control villages did not receive any
intervention. Households from the older cohort were surveyed prior to the implementation
of the intervention, while households from the younger cohort were not. To meet the goal
of this study— examining impact of the social interactions in the unstructured components
40
Random Sample from Rural Northwest China (100 villages)
Randomization
Control group (50 villages) Treatment group (50 villages)
Having access to the parenting center playground No Intervention
Enrolled sample (50 villages, 581 households)
One-year follow-up survey One-year follow-up survey
Enrolled sample (50 villages, 469 households)
Figure 3.2: Overview of the Playground ECD Intervention
of the ECD intervention —I focused on the final sample of the younger cohort, consisting
of 1,067 households in 100 villages at the one-year follow-up survey (see Figure 3.2).
3.3.3 Data Collection
The data presented in this paper were collected in the one-year follow-up survey by teams
of trained enumerators. In this survey, enumerators collected detailed information on
infant language and cognitive skills, parental time investment, parenting awareness, and a
wide range of socio-economic household characteristics of both the younger and the older
cohorts.
3.3.3.1 Measuring Child Development
The survey team used the Bayley Scales of Infant Development-Third Edition (BSID-III) to
measure child development (Bayley, 2006). The BSID-III has been formally adapted to the
Chinese language and environment and used in multiple studies across rural China (Wang
et al., 2019). The BSID-III was administered in the home of each child in the sample using
a standardized set of toys and a detailed scoring sheet.
A child’s scores on the BSID-III are determined by the child’s performance on a series
of tasks, adjusted for age in months and (if applicable) premature birth. The caregiver
of each child was present but was not allowed to assist the child during the test. All
enumerators attended a one-week intensive training course on BSID-III administration,
including 2.5 days of experiential learning in the field, before the survey.
41
Child development was measured by two skill sets with sub-scales: cognitive skills and
language skills. Language development consists of the expressive language and receptive
language sub-scales. I compute age-adjusted internal z-scores from the raw scores for both
cognitive skills and language skills by subtracting the age-specific means and dividing by
the age-specific standard deviations estimated using non-parametric regression methods
Attanasio et al. (2020).
3.3.3.2 Measuring Parental Time Investment
I measured parental time investment in two ways. First, caregivers were asked whether
they were engaged in a set of child-rearing activities the previous day, including story-
telling, singing songs, and interactive play activities with toys. Second, caregivers were
asked how many minutes they spent with the child for each of those child-rearing activities
the previous day.
3.3.3.3 Measuring Parenting Awareness
Parenting awareness was assessed by asking the main caregiver a series of questions on
parenting knowledge and confidence. These include questions about the importance of
different developmentally-appropriate activities, such as reading and playing with their
offspring, and whether caregivers were confident in engaging in these activities. Caregivers
responded to these questions on a 5-point Likert scale. I constructed dummy measures
of these parenting awareness questions by categorizing the 4 points and 5 points on the
Likert scale as 1 (being confident/very confident) and 0 otherwise. I also constructed
an overall parenting awareness index using the first factor loading from the principal
component analysis with parenting awareness measures and then standardizing it by the
mean and the standard deviation of the control group.
4
3.3.3.4 Socioeconomic Characteristics
The survey team administered a detailed survey of child and household characteristics
in the one-year follow-up survey. Child characteristics include the child’s gender, age in
months, and whether the child was born preterm. The exact age of each child was obtained
from his or her birth certificate. Household characteristics include parental education
4
Exploratory factor analysis indicates that the dimensionality of the measures of parental time investment
equals one. I determined the number of relevant factors using Horn’s parallel analysis Horn (1965).
42
level, parental health status, parental age, number of household members, and residential
house value. I generate dummies for whether the father and mother of the child has less
than nine years of education (the length of compulsory education in China), as well as
whether the father and mother of the child self-report to be healthy.
3.3.4 Summary Statistics and Balance Check
Summary statistics and tests for balance across treatment and control groups are shown in
Table 3.1. Households on average had about 4.8 members under the same roof, indicating
parents were raising the child together with other relatives, i.e. grandparents. Most parents
were healthy as indicated by self-report (87% of mothers and 89% of fathers). More than
two-thirds of parents had less than nine years of education, while fathers were slightly
more educated than mothers (approximately 68% of fathers had less than nine years of
education versus about 73% of mothers). Fifty-two percent (52%) of the sample children
were male. About 5% of the children were born preterm. Children in the younger cohort
were approximately 13 months old when evaluated in the one-year follow-up survey.
To test for balance in the younger cohort across treatment and control groups, I chose a
collection of demographic characteristics as well as other characteristics that were time-
invariant or not likely to be affected by the one-year intervention, since the younger cohort
was not surveyed at baseline.
5
Though the number of households in the treatment group
was greater than that in the control group (581 versus 469), the differences between the
two groups in terms of summary statistics are mostly insignificant (Table 3.1), confirming
the random assignment resulted in balanced groups. Nevertheless, I control them as
covariates in the empirical analysis in the next section.
As shown in Appendix Figures D2 and D3, further examination of the difference in
sample size between the treatment and control groups shows that this is driven primarily
by relatively older children in the younger cohort: while the number of enrolled children
who were not yet born at the start of the intervention was the same between the treatment
and control groups, treatment villages had a greater number of children who were one
to six months old when the intervention began. However, since the summary statistics
are balanced between the treatment group and the control group, I believe this potential
difference in attrition does not bias the following results.
5
Due to the absence of baseline data of the younger cohort, I am not able to quantify the attrition of the
sample.
43
3.4 Impact of Having Access to the Playground
I use the following standard ANCOV A specification to estimate the ITT effect of having
access to the playground.
Y
iv
=
0
+
1
T
v
+X
iv
+
s
+
iv
(3.1)
whereY
iv
is outcome of interests of individual i in village v;T
v
is the village treatment
assignment;X
iv
consists of characteristics of individual i in village v (all variables shown
in the Table 3.1);
s
controls for strata (county) fixed effects. Standard errors are clustered
at the village level since treatment assignment is at the village level.
3.4.1 Impact on Child Development
The average treatment effect of providing free access to the playground on the child
development of the younger cohort can be found in Table 3.2. The intention-to-treat
estimates show that the impact is mainly concentrated on language development. The
average treatment effect on the receptive language skills without controlling for covariates
is 0.178 SD, significant at the 5% level. When controlling for covariates, the effect is 0.165
SD, significant at the 10% level. The impact on expressive language skills is 0.105 SD
without covariates and 0.099 SD with covariates, though neither are statistically significant.
The impact on overall language skills— a combined measure of receptive and expressive
language skills— is 0.147 SD when controlling for covariates, significant at the 10% level.
The intervention did not improve the cognitive skills of the younger cohort children.
The impact on cognitive skills is -0.046 SD when not controlling for covariates and -0.065
SD when controlling for covariates. Neither is statistically different from zero.
A possible explanation for why the intervention had an impact on language skills
but not on cognitive skills may be that stimulating cognitive development might require
more advanced parenting skills and knowledge. Given that previous research in the same
rural region of China found that the prevalence of passive parenting among caregivers
is high (Luo et al., 2019), it is intuitive that providing access to the playground—where
social interactions between children and those between caregivers can frequently happen
– would improve the language development of the children without targeted training.
44
However, simply providing access to the playground might not be enough to result in any
impact on the cognitive development of the children.
3.4.2 Impact on Parental Time Investment
In Panel A of Appendix Table D1, I show the impact of the intervention on the parental
investment of caregivers of the younger cohort in terms of the likelihood that they per-
formed certain child-rearing behaviors. The intervention did not lead to an increased
probability of practicing any of the child-rearing behaviors yesterday, including using toys
to play with the child, telling stories to the child, using storybooks to tell stories to the
child, and singing songs to the child (Appendix Table D1, Panel A, Rows 1-4).
The intervention also did not increase the time that parents invested in the above
child-rearing behaviors, except for the time they spent playing with the child. The average
treatment effect is about 22 minutes without controlling for covariates and 20 minutes
with controlling for covariates, both significant at the 1% level (Panel B, Row 1). In sum,
there is some evidence, though limited, of the intervention leading to increases in the
parental time investment.
3.4.3 Impact on Parenting Awareness
In Table 3.3, I show the impact of the intervention on various parenting awareness mea-
sures. The intervention had a consistent and statistically significant impact on the parent-
ing awareness of the caregivers of the younger cohort.
I found that the intervention increased parenting knowledge and the confidence of
caregivers about playing with the child. The treatment increased the probability that
caregivers believe playing with their children is important by 11.6 percentage points,
significant at the 5% level (Table 3.3, Row 2, Column 2). The intervention also increased
whether parents know how to play with their child by 19.9 percentage points, significant at
the 1% level (Table 3.3, Row 3, Column 2). Similarly, the intervention increased parenting
knowledge and confidence of caregivers about reading to the child. The treatment effect
increases the probability that caregivers believe reading to their child is important by 18.9
percentage points, significant at the 1% level (Table 3.3, Row 4, Column 2). The inter-
vention also increases whether parents know how to read to the child by 16.3 percentage
points, significant at the 5% level (Table 3.3, Row 5, Column 2). The bottom row shows
45
that there is a strong treatment effect on parenting awareness overall – measured by the
parenting awareness index – which increased by 0.211 SD, significant at the 1% level.
Taken together, I found consistent evidence that access to the playground had a positive
impact on parenting awareness. This result is surprising as the caregivers were not directly
exposed to any parenting training. One possibility is that caregivers in the younger cohort
learned from the caregivers in the older cohort, who received parenting training in the
same treatment village and parenting center. In the next section, I will show the evidence
to test this hypothesis.
3.5 The Role of Social Interactions
In this section, I shed light on the mechanisms of the intervention’s impact on the younger
cohort. Since the intervention is conducted in a common area where frequent social
interactions could take place, the younger cohort could benefit from interactions with both
households from the younger and the older cohorts. In addition, (Zhong et al., 2020) shows
that the overall intervention, including both access to the playground and the psycho-social
parenting training, had a positive effect on the child development and parenting practices
of the older cohort. Thus, the younger cohort could potentially benefit from having access
to the playground through more social interactions in general as well as through social
interactions with a cohort of older toddlers and more experienced caregivers.
I explore several channels through which social interactions could occur. Intuitively,
two key factors that might affect the social learning in the intervention for the younger
cohort are (1) how many children or caregivers one can potentially interact with as well as
(2) whom one interacts with. In regard to the first factor, too many children or caregivers
accessing the parenting center might lead to crowding out others’ access to the parenting
center (e.g. limited space available, limited toys available., etc.). On the other hand,
children and their caregivers might not engage in many interactions if there are not
enough potential participants that they can interact with.
To test for these hypotheses, I use the following specification to explore the treatment
heterogeneity by cohort sizes:
Y
iv
=
0
+
1
T
v
+
2
T
v
N
O
v
+
3
T
v
N
NO
v
+
4
N
NO
v
+
5
N
O
v
+X
iv
+
s
iv
(3.2)
46
whereY
iv
is the outcome of interests of individual i in village v;T
v
is the village treatment
assignment;N
O
v
is the demeaned older cohort households size in the village v (number
of older cohort households in village v minus the average size of older cohort in sample
villages —17 households in the older cohort per village);N
NO
v
is the demeaned younger
cohort households size in the village v (number of younger cohort households in village
v minus the average size of younger cohort in sample villages —10 households in the
younger cohort per village); X
iv
are the covariates of individual i in village v that are
not likely to be changed by the intervention;
s
controls for strata (county) fixed effects.
Standard errors are clustered at the village level.
Specifically, I add two interaction terms – that of the treatment assignment with the
demeaned younger cohort size and that of the treatment assignment with the demeaned
older cohort size – to the regression for estimating the intention-to-treat effects, Equation
(1). Village covariates are also added to the regression as controls. The number of
households from the younger cohort and the number of households from the older cohort
in the village are controlled for as village covariates.
3.5.1 The Impact of Social Interactions with Households that Have Similar-
age Children
Suggestive results of social learning between similar-age children are presented in Table
3.4. The intervention’s impact on younger cohort children’s language skills is larger when
there are more children from the younger cohort in the village. This is the case for all
three language skill measurements (receptive skills, expressive skills, and overall language
skills). An additional child-parent dyad in the younger cohort improves child receptive
language skills by 0.04 SD, significant at the 1% level; it improves expressive language
skills by 0.027 SD, significant at the 10% level; and it improves overall language skills
by 0.04 SD, significant at the 5% level (Table 3.4, Row 2). In contrast, more children in
the older cohort do not appear to increase the impact of the intervention on the younger
cohort, and they also do not appear to have a crowding out effect either, as the coefficients
in Row 3 are not statistically significant. The treatment heterogeneity by the number of
child-parent dyads in the younger cohort suggests that children benefit from interaction
with other members of the younger cohort, though it is not completely clear whether
child-to-child interaction or parent-to-parent interaction is driving this heterogeneous
effect.
47
To support the theory that children benefited from increased social interactions with
peers, I show auxiliary regression results in Table 3.5. Results in Table 3.5 support this
interpretation, as having access to the playground increased the social interactions between
children in the village. The ITT estimates show that having access to the playground
significantly increased the number of children aged 6-18 months old who played together
and the total number of children of all ages who played together (significant at the 5%
level—Table 3.5, Column 1, and Column 4). Since the children in the younger cohort
at the follow-up survey are mostly between 6 and 18 months old, the increase in social
interactions overall was mainly driven by increased social interactions between similar-
age children in the younger cohort (Table 3.5, column 1). However, I cannot rule out
that the effects of social interactions between younger cohort households could also be
driven by increased social interactions between caregivers, which theoretically could also
lead to improvements in child development. Above all, this evidence helps to explain
the mechanism of the ITT effects on the language development of the younger cohort
presented in the previous subsection.
3.5.2 The Impact of Social Interactions with Household from the Older
Cohort
Treatment effect heterogeneity by cohort sizes on parenting awareness is shown in Table
3.6. Caregivers from the younger cohorts improve their parenting awareness when there
are more caregivers from the older cohort in the village. Each additional household from
the older cohort decreases the probability that caregivers felt nervous about parenting by
about 0.3 percentage points, significant at the 5% level (Column 1). It also increased the
probability of the following behaviors: knowing how to read to the child (0.7 percentage
points, significant at the 10% level, Column 3); knowing the importance of playing with
the child (1 percentage point, marginally significant at the 10% level, Column 4; and
knowing the importance of reading to the child (0.9 percentage points, significant at 10%
level, Column 5). Overall parenting awareness also increased (0.02 SD, significant at 5%
level, Column 6).
As parents from the older cohort had access to the psycho-social training, they were
more experienced and knowledgeable in child-rearing. This treatment heterogeneity
by the number of children in the older cohort therefore suggests that younger cohort
caregivers learned from the more experienced and knowledgeable older cohort caregivers.
48
I also present the results of treatment effect heterogeneity by cohort sizes on parental
investment outcomes in Appendix Table D2. I do not find consistent evidence that
social interactions improved the parental investment of the younger cohort, regardless
of whether the parent-to-parent interactions were with the older cohort or with other
younger cohort households. One of the reasons for this result could be that caregivers did
not perform the same parenting activities in the playground as they normally do inside
their households, thereby rendering it difficult for other caregivers to learn from or imitate
each other. Another possible explanation is that the improved parenting awareness from
social interactions did not shift their behavior (i.e., parenting investment).
3.5.3 Robustness Checks
I present two sets of robustness checks of the treatment heterogeneity by cohort size
results discussed in the previous sub-section. First, I address the potential issue that the
villages that have different cohort sizes might also differ in terms of other characteristics,
which could confound the results. To isolate the treatment heterogeneity by cohort sizes, I
included village characteristics and their interactions with the treatment assignment to
Equation (2). Results from this exercise can be found in Appendix Tables D3 (on child
development) and A4 (on parenting awareness). As shown in Appendix Table D3, an
additional household from the younger cohort in the treatment village improved receptive
language skills, expressive language skills, and overall language skills. The magnitude and
the statistical significance of the coefficients are similar to those in Table 3.4, suggesting
that the estimates in Table 3.4 already isolate the treatment heterogeneity by cohort sizes.
Similarly, in Appendix Table D4, I also find that including village characteristics and
their interactions with the treatment assignment does not change the results regarding the
suggestive evidence of social learning from more knowledgeable caregivers (Table 3.6).
Thus, results from Appendix Tables D3 and D4 further indicate that social interactions
played a major role in the impact of the playground intervention.
Second, I provide evidence to show that the potential attrition problem raised in section
III is unlikely to be the source of the treatment heterogeneity by cohort sizes shown in
Tables 3.4 and 3.6. Appendix Figures D2 and D3 show that there is a significant difference
between the treatment and control groups in the number of children in the younger cohort
who were at least 1 month old at baseline, while there is no such difference in the number
of households that are still expecting their newborns at baseline. In a similar exercise
49
as Equation (2), using the sample of children who were due less than 6 months up to
just born (the “-6-0” cohort), I examine whether the treatment effect varies by the “-6-0”
cohort size (children due in less than six months), the “1-6” cohort size (children one to
six months old), and the older cohort size (Appendix Table D5). I found results similar to
those in Table 3.4, which is that the effect on language skills depends on the number of
households that have similar-age children (the -6-0 cohort size – Appendix Table D5, Row
2). I also found results similar to those in Table 3.6, which is that the effect on parenting
awareness is affected by the older cohort size (Appendix Table D6, Row 4). Effect sizes
were similar to those in Tables 3.4 and 3.6, though they were less statistically significant.
The less significant coefficients might be due to the smaller sample size. In sum, I believe
that these results show that the potential attrition problem raised in section III is not likely
to be the source of the treatment heterogeneity by cohort sizes in Tables 3.4 and 3.6.
3.5.4 Social Learning from Knowledgeable Caregivers
To validate that caregivers indeed gained parenting awareness from interacting with the
older cohort, I investigated whether increases in the older cohort caregivers’ parenting
awareness lead to increases in that of the younger cohort—in other words, I identified
whether or not there were peer effects in parenting awareness. To overcome the iden-
tification problem in peer effects raised by Manski (1993), I use the partial-population
intervention design of parenting sessions in the experiment to estimate peer effects in
parenting awareness(Bobonis and Finan, 2009; Moffitt et al., 2001; Lalive and Cattaneo,
2009).
Consider the following linear-in-means framework: for the older cohort households:
Y
O
iv
= +X
O
iv
+
¯
X
v
+Z
v
+
¯
Y
v
+T
O
iv
+
iv
(3.3)
caregiver’s parenting awareness is affected byX
O
iv
individual characteristics,Z
v
village
characteristics (including mean of individual characteristics in the village),Y
v
is average
parenting awareness of the others in the village, and the direct treatment effect of psycho-
social parenting training from the intervention T
O
iv
; however for the younger cohort
households,
Y
NO
iv
= +X
NO
iv
+
¯
X
v
+Z
v
+
¯
Y
v
+
iv
(3.4)
50
Since the direct psycho-social parenting training is not available (shown in Figure
3.1), caregiver’s parenting awareness is affected byX
NO
iv
individual characteristics, Z
v
village characteristics (including mean of individual characteristics in the village), andY
v
is average parenting awareness of the others in the village. is the parameter of interests.
Combining Equations (3) & (4), average outcome of interests in the village can be
written as
¯
Y
v
=
1
+
+
1
¯
X
v
+
1
Z
v
+
1
m
O
v
T
v
(3.5)
wherem
O
v
is the fraction of older cohort households in the village.
Equations (4) & (5) consist of instrumental variable (IV) strategy to identify the param-
eter of interest, endogenous peer effect. Sincem
O
v
T
v
does not affect theY
NO
iv
—parenting
awareness of the younger cohort directly other than through village level parenting aware-
ness, it would be the choice of IV for
¯
Y
v
. Bobonis and Finan (2009) also suggests thatT
v
by
itself can also serve as the IV for
¯
Y
v
.
Under the strict assumption that there is no direct effect on parenting awareness from
having access to the playground for younger cohort caregivers alone (i.e., having access to
the storybooks, toys, and playground does not improve their parenting), Table 3.7 presents
evidence that caregivers learned from each other. In Columns (1) and (2), I present IV
results usingm
O
v
T
v
(the fraction of older cohort households in the village interacts with
the treatment assignment) as IV for the average parenting awareness of both the younger
cohort and the older cohort in the village. In Columns (3) and (4), I present IV results
usingT
v
(the treatment assignment) as IV for the average parenting awareness of both the
younger cohort and the older cohort in the village. Both IV strategy results show strong
peer effects in parenting awareness at the village level: every 1.0 SD increase in the village
parenting awareness index leads to more than a 0.9 SD increase in parenting awareness,
confirming the theory that caregivers of the younger cohort gained in parenting awareness
by interacting with caregivers of the older cohort.
3.6 Conclusion
This paper studies the role of social interactions in early childhood development. It
aimed to investigate two previously unexplored questions: first, whether promoting social
interactions among young children less than 1.5 years old and among their caregivers will
51
improve child development and parenting practices; second, through which mechanisms
social interactions affect child development and parenting practices.
To accomplish this, I used the data of a randomized controlled trial in rural China. In
the sample of 100 rural villages, households from half of the villages were provided free
access to an indoor playground that increased the social interactions among children and
those among their caregivers. Using random assignment, I identify the intention-to-treat
effects of the playground intervention on the child development and parenting practices of
caregivers after one year. To provide suggestive evidence about the potential mechanisms
of the effects, I examined the treatment heterogeneity using the ex-ante cohort sizes, a
representation of the number of potential social links and types of social links. Finally, I
use the partial population intervention design to estimate a linear-in-means peer effects
model to examine the existence of social learning between caregivers.
The results show that providing access to an indoor playground to children in the first
year of life improves child development and the parenting practices of their caregivers
in rural China. There is suggestive evidence that the impacts of providing access to the
playground on child development and parenting practices of caregivers are mediated by
social interactions.
The wider implications of these results is that social interactions might be one of the
mechanisms behind recent successes in group-based ECD interventions. Since social
interactions lead to spillover effects that are often unaccounted for, results from this paper
provide further evidence for the scalability of group-based ECD interventions.
52
Table 3.1: Balance Check
(1) (2) (3)
Variable name Treatment group Control group (1)-(2)
Number of household members 4.746 4.821 -0.1005
[1.002] [0.945] (0.0662)
Mother age (yrs) 28.126 28.361 -0.1766
[4.548] [5.335] (0.2911)
Father age (yrs) 30.653 31.026 -0.4456
[4.972] [5.291] (0.3369)
Mother self-report healthy 0.868 0.87 0.0019
[0.339] [0.336] (0.0225)
Father self-report healthy 0.894 0.888 0.0054
[0.308] [0.316] (0.0182)
Mother has less than 9yrs education 0.75 0.731 0.0134
[0.434] [0.444] (0.0314)
Father has less than 9yrs education 0.684 0.678 0.0115
[0.465] [0.468] (0.0322)
Male child 0.497 0.538 -0.0355
[0.5] [0.499] (0.0289)
Child preterm birth 0.055 0.049 0.0007
[0.228] [0.217] (0.0141)
Child age (month) 13.184 12.843 0.2918
[4.44] [4.432] (0.2533)
Housing value< 10 thousand (RMB) 0.018 0.024 -0.0010
[0.134] [0.152] (0.0081)
Housing value=10 - 50 thousand (RMB) 0.055 0.088 -0.0315
[0.228] [0.283] (0.0190)
Housing value=50 - 100 thousand (RMB) 0.137 0.152 -0.0048
[0.344] [0.359] (0.0208)
Housing value=100 - 300 thousand (RMB) 0.422 0.487 -0.0813**
[0.494] [0.5] (0.0349)
Housing value>300 thousand (RMB) 0.352 0.231 0.1224***
[0.478] [0.422] (0.0346)
1
Notes: Standard deviation in the brackets; Standard error in the parentheses are clustered at
village level;
2
(3) includes strata fixed effects;
3
p< 0:10,p< 0:05,p< 0:01.
53
Table 3.2: The Impact of Having Access to Playground on Child Development
Intention-to-treat estimates
without covariates with covariates
(1) (2) control
Dependent variables Coefficient S.E. Coefficient S.E. mean
Child Development (N=1043)
Cognitive Skills -0.0303 (0.0604) -0.0485 (0.0626) -0.0608
Receptive Language Skills 0.1779** (0.0896) 0.1653* (0.0912) -0.1510
Expressive Language Skills 0.1045 (0.0712) 0.0987 (0.0692) -0.1143
Overall Language Skills 0.1557* (0.0841) 0.1471* (0.0833) -0.1473
1
Standard errors in the parenthesis are clustered at the village level;
2
Individual covariates controlled are the child’s age, the child’s gender, an indicator
function of whether the child was a preterm birth, number of household members,
an indicator function of father of the child has less than 9 years of education, an
indicator function of mother of the child has less than 9 years of education,an
indicator function of father is self-reportedly healthy, an indicator function of mother
is self-reportedly healthy, father age, mother age, and four indicator functions of
housing value categories;
3
All scores of development skills have been internally standardized nonparametrically
for age and are expressed in standard deviation units;
4
p< 0:10,p< 0:05,p< 0:01.
54
Table 3.3: The Impact of Having Access to Playground on Parenting Awareness
Intention-to-treat estimates
without covariates with covariates
(1) (2) control
Dependent variables Coefficient S.E. Coefficient S.E. mean
Parental Awareness (N=1046)
Felt nervous around the child -0.0004 (0.0080) 0.037 (0.074) 0.0278
Playing with the child is important 0.1238** (0.0548) 0.1160** (0.0523) 0.5503
Know how to play with the child 0.1905*** (0.0636) 0.1987*** (0.0640) 0.2955
Reading to the child is important 0.1941*** (0.0544) 0.1890*** (0.0587) 0.3383
Know how to read to child 0.1608** (0.0731) 0.1632** (0.0703) 0.2313
Parenting awareness index 0.2014*** (0.0462) 0.2117*** (0.0724) 0
1
Standard errors in the parenthesis are clustered at the village level;
2
Individual covariates controlled are the child’s age, the child’s gender, an indicator function
of whether the child was a preterm birth, number of household members, an indicator
function of father of the child has less than 9 years of education, an indicator function of
mother of the child has less than 9 years of education,an indicator function of father is
self-reportedly healthy, an indicator function of mother is self-reportedly healthy, father age,
mother age, and four indicator functions of housing value categories;
3
Dependent variables in row one to five are dummy variables; Dependent variable in row
six, parenting awareness index, is expressed in standard deviation units standardized by the
mean and the standard deviation of the control group.
4
p< 0:10,p< 0:05,p< 0:01.
55
Table 3.4: Suggestive Evidence of Social Learning from Similar-age
Children Households
(1) (2) (3) (4)
Cognitive Receptive Expressive Overall
Skills Language Skills
Treatment -0.0827 0.0706 0.0195 0.0485
(0.0737) (0.0886) (0.0734) (0.0827)
Treatment 0.0039 0.0438*** 0.0273** 0.0395***
Demeaned younger cohort size (0.0116) (0.0143) (0.0137) (0.0147)
Treatment -0.0044 -0.0228 -0.0135 -0.0201
Demeaned the older cohort size (0.0110) (0.0191) (0.0159) (0.0187)
N 1013 1013 1013 1013
Individual covariates X X X X
Mean in control group -0.0583 -0.1521 -0.1118 -0.1467
1
Standard errors in the parenthesis are clustered at the village level;
2
Individual covariates controlled are the child’s age, the child’s gender, an indicator
function of whether the child was a preterm birth, number of household members,
an indicator function of father of the child has less than 9 years of education, an
indicator function of mother of the child has less than 9 years of education,an indi-
cator function of father is self-reportedly healthy, an indicator function of mother
is self-reportedly healthy, father age, mother age, and four indicator functions of
housing value categories;
3
All scores of development skills have been internally standardized nonparametri-
cally for age and are expressed in standard deviation units;
4
p< 0:10,p< 0:05,p< 0:01.
56
Table 3.5: Impact of Having Access to the Playground on Social Interactions between Children
(1) (2) (3) (4)
Number of children played together Total number of
aged 6-18 months old aged 18-30 months old aged 30-42 months old children played together
Treatment 0.3011** 0.0881 0.1168 0.5060**
(0.1300) (0.0915) (0.0833) (0.2195)
Individual covariates X X X X
N 1015 1015 1015 1015
Mean in Control 1.3326 0.9889 0.7206 3.0421
1
Standard errors in the parenthesis are clustered at the village level;
2
Individual covariates controlled are the child’s age, the child’s gender, an indicator function of whether the child was
a preterm birth, number of household members, an indicator function of father of the child has less than 9 years of
education, an indicator function of mother of the child has less than 9 years of education,an indicator function of father is
self-reportedly healthy, an indicator function of mother is self-reportedly healthy, father age, mother age, and four indicator
functions of housing value categories;
4
p< 0:10,p< 0:05,p< 0:01.
57
Table 3.6: Suggestive Evidence of Social Learning from Other Knowledgeable Caregivers
(1) (2) (3) (4) (5) (6)
Felt Nerves Know how to Know the importance of Parental
around play with read to playing with reading to awareness
the child the child the child the child the child index
Treatment 0.0086 0.0379 0.0519** 0.0738** 0.0686* 0.1758**
(0.0094) (0.0284) (0.0252) (0.0360) (0.0378) (0.0711)
Treatment 0.0008 0.0054 -0.0046 -0.0094 -0.0051 -0.0091
Demeaned younger cohort size (0.0014) (0.0047) (0.0040) (0.0069) (0.0067) (0.0119)
Treatment -0.0032** 0.0042 0.0071* 0.0095 0.0088* 0.0222**
Demeaned the older cohort size (0.0015) (0.0037) (0.0038) (0.0059) (0.0051) (0.0097)
N 1014 1014 1014 1014 1014 1014
Individual covariates X X X X X X
Mean in control 0.027 0.2905 0.2306 0.5499 0.3392 -0.0086
1
Standard errors in the parenthesis are clustered at the village level;
2
Individual covariates controlled are the child’s age, the child’s gender, an indicator function of whether the
child was a preterm birth, number of household members, an indicator function of father of the child has less
than 9 years of education, an indicator function of mother of the child has less than 9 years of education,an
indicator function of father is self-reportedly healthy, an indicator function of mother is self-reportedly
healthy, father age, mother age, and four indicator functions of housing value categories;
3
Dependent variables in column (1) to (5) are dummy variables; Dependent variable in (6), parenting awareness
index, is expressed in standard deviation units standardized by the mean and the standard deviation of the
control group.
4
p< 0:10,p< 0:05,p< 0:01.
58
Table 3.7: Evidence of Caregivers of the Younger Cohort Learning from Other Caregivers
IV-first stage IV (2SLS) IV-first stage IV (2SLS)
(1) (2) (3) (4)
Village mean of Parenting awareness (SD) Village mean of Parenting awareness (SD)
parenting awareness (SD) parenting awareness (SD)
Fraction of older cohort HHs 0.325
in the village Treatment (0.105)
Treatment 0.195
(0.057)
Village mean 0.934
0.975
of parenting awareness (0.049) (0.035)
Observations 1043 1043 1043 1043
Fraction of older cohort HHs 0.62
Mean of DV 0.074 0.075 0.069 0.069
First stage F-statistics 9.560 11.620
Individual covariates X X X X
Village covariates X X X X
1 Standard errors in the parenthesis are clustered at the village level;
2 Individual covariates are the child’s age, the child’s gender, and an indicator function of whether the child was a preterm birth;
3 Village covariates are percentage of children in the village taht are male, and percentage of children in the village that are preterm birth. The
village mean is calculated with characteristics from both the younger cohort and the older cohort.
4 Parenting awareness index is expressed in standard deviation units standardized by the mean and the standard deviation of the control group.
5p< 0:10,p< 0:05,p< 0:01.
59
Chapter 4
Identification of Causal Effects in Cluster Randomized
Experiments with Spillovers and Noncompliance
1
4.1 Introduction
Cluster randomized controlled trials (cluster RCTs) are widely applied in Development
Economics to study the impact of developmental interventions (Duflo et al., 2007; Bloom,
2005). In cluster RCT, the randomization took place at the cluster of interests, such as
communities, schools, or villages for theoretical or policy-relevant reasons. For the inter-
vention that seeks to have an impact on outcomes at the cluster level, the randomization
of intervention assignments at the cluster level fulfills its goal. (Chattopadhyay and Duflo,
2004). Alternatively, it is cost-efficient to have randomization at the cluster level when the
implementation of the intervention applies to large lump sum cost (Banerjee et al., 2011;
Miguel and Kremer, 2004).
In many cluster RCTs, the intended interventions either do not target or appeal to all in-
dividuals, leading to only a fraction of individuals take up the intended intervention—also
called partial compliance. For instance, schools are well-suited for the cost-effectiveness
of public health interventions, but only students present on the day of the intervention
administration would get the treatment (Miguel and Kremer, 2004; Del Rosso and Marek,
1996); on the other hand, encouragement designs only target individuals whose incentives
are compatible with the intervention (Duflo and Saez, 2003).
As it is often the case, given partial compliance, researchers estimate the intent-to-
treat (ITT) to identify the impact of intervention assignment. However, the effect of the
treatment take-up, rather than the assignment, is the parameter of interest in many cases.
1
Coauthored with Bora Kim
60
This is particularly true if the purpose of the research is to understand the effect of the
intervention that could be delivered by policy-makers with different experiment designs,
i.e., when one is interested in the impacts of counterfactual policy (Duflo et al., 2007).
To identify the effect of treatment take-up in the settings with partial compliance,
instrumental variable (IV) methods have been widely used. Specifically, Imbens and
Angrist (1994) showed that the IV estimator identifies the local average treatment effect
(LATE). However, the traditional IV estimator does not identify any causal parameter
when the nature of interventions in cluster-RCTs introduces spillovers within clusters,
i.e. through social interactions and/or the general equilibrium (GE) of clusters. This is
because if there are spillovers or general equilibrium effects, the exclusion restriction of
the treatment assignment is no longer satisfied.
In this paper, we systematically study how to define causal effects, and what effects
can be identified in the cluster RCTs with spillovers and noncompliance under simple
identifying assumptions. Our focus is to identify the treatment effect for the treated
as opposed to the general ITT effect. We show that in the general potential outcome
framework with spillovers within clusters, the usual LATE, taken as the treatment effect
of the compliers (or treatment effect of the treated under one-sided noncompliance), is
not well defined. Instead, we propose a simple difference-in-differences-style (DID-style)
estimator to estimate the ITT of compliers—the treatment effect of the treated in this
case—as long as we have baseline information and the case of one-sided noncompliance. In
addition, our estimator also identifies the ITT of never-takers, which captures the potential
spillover effects or general equilibrium effects introduced by the cluster-level intervention.
In an empirical application, we apply our estimators to a cluster-RCT microcredit
program in Morocco (Cr´ epon et al., 2015). Given the possibility that the take up of the
microcredit by households could introduce spillover effects to non-takers and/or a general
equilibrium effect to the village, we estimate the ITT effect of the compliers and the never-
takers. We found the ITT effects for the never-takers are non-zero, confirming the incident
of spillovers. Besides, the average treatment effect of the treated measured by the ITT
effect for compliers is smaller than the LATE estimates proposed in Cr´ epon et al. (2015),
suggesting the differences between the estimates stem from the latter ignoring spillover
effects across households within the cluster.
Our research makes two contributions to the literature. First, we purpose an alternative
approach with minimal assumptions to estimate the treatment effect of the treated when
61
the exclusion restriction assumption of the treatment assignment fails. Previous literature
has been approaching this problem —sometimes characterized by complier average causal
effect —with partial identification(Flores and Flores-Lagunes, 2013; Mealli and Pacini,
2013) and Bayesian inference (Imbens and Rubin, 1997; Rubin and Zell, 2010). The
advantage of our approach is that our key identifying assumption is mild and testable.
Second, we contribute to the literature in estimating the spillover effect of intervention
programs when there is no information available to identify counter-factual never-takers.
With the aid of eligibility rule of intervention take-up from the programs, Lalive et al.
(2015) and Angelucci and De Giorgi (2009) were able to construct counterfactual never-
takers to identify the spillover effect of the unemployment insurance and conditional cash
transfer, respectively. Our approach can be applied to the program even without any
specific intervention eligibility rule.
The outline of this paper is as follows. In section II, we present our setup and param-
eters of interests; in section III, we show our identification assumptions and strategies;
in section IV, we explain our estimation strategy; finally, we apply our estimator to a
microcredit program and test the key identifying assumption empirically.
4.2 The Setup and Parameters of Interest
4.2.1 The Potential Outcome Framework Setup
We consider a model where there can be a possible treatment effect spillover within a
predetermined cluster structure such as classrooms or villages. Each cluster is indexed by
g2f1;;Gg. For each clusterg, there aren
g
individuals indexed byi = 1;;n
g
with the
total ofN =
P
G
g=1
n
g
. LetZ
g
2f0;1g indicate the binary policy intervention at the cluster
levelg. Specifically,Z
g
= 1 if the clusterg is assigned with the intervention andZ
g
= 0
otherwise. In a clustered RCT, everyone in a clusterg withZ
g
= 1 is eligible to receive
treatment. LetD
ig
2f0;1g denote the actual treatment received by individuali in a cluster
g so that we haveD
ig
= 1 for treated individuals andD
ig
= 0 for non-treated individuals.
Outcome of interest is denoted byY
ig
. Our aim is to assess causal effect of treatmentD
ig
on
the outcomeY
ig
while allowing for possible spillover effects across individuals in the same
cluster. To define causal effect of treatment, we utilize the potential outcome framework
popularized by Rubin (1974) as our basic framework.
62
LetZ
g
2f0;1g be the possible values thatZ
g
can take. Potential treatment is denoted
byD
ig
(z
g
). Specifically, (D
ig
(0);D
ig
(1)) denote the potential treatment when individual
is in the control group and treatment group, respectively. In a perfect compliance case,
we haveD
ig
(1) = 1 andD
ig
(0) = 0 for every observation. If there is a noncompliance,
as in many empirical cases where perfect enforcement of assignment is impossible, we
will haveD
ig
(0) = 1 and/orD
ig
(1) = 0. Observed treatmentD
ig
can be written asD
ig
=
Z
g
D
ig
(1) + (1Z
g
)D
ig
(0).
Letd
ig
2f0;1g be the possible values thatd
ig
can take. We model the potential outcome
forY
ig
byY
ig
(z
g
;d
ig
). Given that we have binary intervention and binary treatment, for
each individuali in clusterg, we have four potential outcomes:
fY
ig
(1;1);Y
ig
(1;0);Y
ig
(0;1);Y
ig
(0;0)g
Different from the traditional approach, we allow for the assignment variableZ
g
to
affectY
ig
directly. In other words, we do not require the exclusion restriction ofZ
g
onY
ig
.
Such exclusion restriction is unlikely to hold when there is a spillover effect or general
equilibrium effect of the policy intervention. To see this, note that
Y
ig
(1;d),Y
ig
(0;d); d = 0;1
For instance, if there is a program spillover effect to the non-treated (D
ig
= 0) in a treated
region (Z
g
= 1), then we expect to haveY
ig
(1;0),Y
ig
(0;0). Allowing for such spillover
effects is important especially when we have large-scale RCT. Even when the individual
is not treated, the intervention may affect her through the general equilibrium effect or
direct spillovers from social interactions.
2
4.2.2 Parameters of Interest
Throughout the paper, we maintain the following assumptions:
Assumption 1 (Independence) Program assignmentZ
g
is jointly independent to potential
outcomes. Formally,
Z
g
?
Y
ig
(z
g
;d
ig
);D
ig
(z
0
g
)
; 8d
ig
;z
g
;z
0
g
2
We note that we do not try to identify the exact mechanism of the spillover effect. In other words, we do
not attempt to distinguish whether the spillover effect comes from the general equilibrium effect or from
social interactions/peer effects, etc.
63
The Assumption 1 is trivially satisfied whenZ
g
is randomly assigned to clusters as in the
case of cluster RCTs.
Assumption 2 (One-sided noncompliance) There is only one-sided noncompliance where
D
ig
(0) = 0 andD
ig
(1) can be either 0 or 1.
The Assumption 2 dictates that individuals in an untreated region cannot take the treat-
ment. In many applications, this is satisfied by design.
We can classify each individual in terms of their potential treatment values. In the
terminology of Angrist et al. (1996), let us define the compliance type in the following
way:
T
ig
=co () (D
ig
(0);D
ig
(1)) = (0;1)
T
ig
=nt () (D
ig
(0);D
ig
(1)) = (0;0)
Under Assumption 2, the population consists only of compliers (co) and never-takers (nt).
While we cannot identify the type of each individual, we can identify the proportions
of types in the population. Let
co
=
def
Pr(T
ig
=co) and
nt
= Pr(T
ig
=nt). They are
identified by
nt
= Pr(D
ig
= 0jZ
g
= 1);
co
= 1
nt
Many empirical papers have largely focused on the effects ofZ
g
onY
ig
by focusing on
the intent-to-treat effect which is defined as follows:
ITT =
def
E[Y
ig
(1;D
ig
(1))Y
ig
(0;D
ig
(0))]:
With exogenousZ
g
, ITT is easily identified by E[Y
ig
jZ
g
= 1] E[Y
ig
jZ
g
= 0], difference in
mean outcome between treated clusters and control clusters.
While such ITT effect identifies the impact of the intervention on the outcome, it does
not identify the impact of the treatment on the outcome in experiments with noncompli-
ance. In contrast, in many applications, researchers are often interested in identifying the
64
effect of treatment take-up rather than the intervention itself. The following derivation
shows that ITT in fact can be decomposed into two distinct effects:
E[Y
ig
(1;D
ig
(1))Y
ig
(0;D
ig
(0))]
= E[Y
ig
(1;D
ig
(1))Y
ig
(0;0)]; * Assumption 2
= E[Y
ig
(1;1)Y
ig
(0;0)jD
ig
(1) = 1]
| {z }
=ITT(co)
Pr(D
ig
(1) = 1)
+E[Y
ig
(1;0)Y
ig
(0;0)jD
ig
(1) = 0]
| {z }
=ITT(nt)
Pr(D
ig
(1) = 0)
Hence, simple ITT identifies the weighted averages of two different effects. The first term
is the effect of the intervention for compliers. Therefore, we call it ITT(co).
ITT (co) =
def
E[Y
ig
(1;1)Y
ig
(0;0)jD
ig
(1) = 1]
= E[Y
ig
(1;1)Y
ig
(0;0)jco]
Note that with one-sided noncompliance, ITT(co) is equivalent to E[Y
ig
(1;1)Y
ig
(0;0)jD
ig
=
1]. Therefore, ITT(co) has an interpretation as the average treatment effect on the treated
(ATT). Note that
E[Y
ig
(1;1)Y
ig
(0;0)jco] = E[Y
ig
(1;1)Y
ig
(0;1)jco] + E[Y
ig
(0;1)Y
ig
(0;0)jco]
It measures the overall effect for compliers that consists of a direct effect of treatment and
an indirect effect of treatment assignment.
The second term is the effect of the intervention on never-takers or ITT(nt).
ITT (nt) =
def
E[Y
ig
(1;0)Y
ig
(0;0)jD
ig
(1) = 0]
ITT(nt) measures the spillover effect of the intervention for never-takers.
Our purpose is to separately identify ITT(co) and ITT(nt). Knowledge on ITT(co) and
ITT(nt) are important as it sheds light on the treatment effect heterogeneity: even when
the overall ITT effect is positive, ITT(co) may have a different sign depending on ITT(nt),
the spillover effect. In that case, using the overall ITT estimates to infer ATT can be
misleading.
65
In the following, we show that, unlike the overall ITT effect, local ITT effects such as
ITT(co) and ITT(nt) are not identified without further assumptions. Angrist et al. (1996)
shows that the Wald ratio identifies ITT(co) when the exclusion restriction is satisfied.
This is no longer true when the exclusion restriction is not satisfied.
3
In fact, ITT(co) and
ITT(nt) are both unidentified without strong assumptions. To see this point, note that
E[Y
ig
jZ
g
= 1;D
ig
= 0] = E[Y
ig
(1;0)jnt] (4.1)
E[Y
ig
jZ
g
= 1;D
ig
= 1] = E[Y
ig
(1;1)jco] (4.2)
E[Y
ig
jZ
g
= 0;D
ig
= 0] =
co
E[Y
ig
(0;0)jco] +
nt
E[Y
ig
(0;0)jnt] (4.3)
From Equation 4.1 and Equation 4.2, we see that E[Y
ig
(1;0)jnt] and E[Y
ig
(1;1)jco] are identi-
fied. However, Equation 4.3 shows that we cannot separate E[Y
ig
(0;0)jco] and E[Y
ig
(0;0)jnt]
implying thatITT (co) = E[Y
ig
(1;1)Y
ig
(0;0)jco] andITT (nt) = E[Y
ig
(1;0)Y
ig
(0;0)jnt] are
not point identified without additional assumptions.
One strong assumption is to assume E[Y
ig
(0;0)jco] = E[Y
ig
(0;0)jnt], i.e., in the absence
of intervention, mean outcomes are the same for never-takers and compliers. Under
such assumption, we can identify E[Y
ig
(0;0)jco] by E[Y
ig
jZ
g
= 0;D
ig
= 0]. However, this
assumption is very strong as it imposes the homogeneity of treatment effect across different
compliance types. In the next section, we show that when we have baseline information
on the outcome variable, point identification is possible under the mild assumption.
4.3 Identification
We incorporate baseline information to facilitate our identification. Let us assume that
there are two time periods for simplicity. DenoteT =t for pre-treatment period andT =
t + 1 for post-treatment period. Outcome variables are denoted by (Y
t
ig
;Y
t+1
ig
). Randomized
program assignment dummy is denoted by (Z
t
g
;Z
t+1
g
) whereZ
t
g
= 0 for all clustersg. On
the other hand, in the post-treatment period,Z
t+1
g
= 1 for treated clusterg whileZ
t+1
g
= 0
for control cluster g. Potential treatment status is denoted by (D
t
ig
(z
t
g
); D
t+1
ig
(z
t+1
g
)) =
(0; D
t+1
ig
(z
t+1
g
)).
3
See Appendix E to see what Wald ratio identifies when the exclusion restriction is not satisfied
66
We assume that baseline outcomesY
t
ig
=Y
t
ig
(0;0) are observed for all individuals. Our
parameter of interest becomes
ITT
t+1
(co) = E[Y
t+1
ig
(1;1)Y
t+1
ig
(0;0)jco]
where E[Y
t+1
ig
(0;0)jco] is again a counterfactual outcome as compliers do not exhibitY (0;0)
in the post-period. To identify this, we impose an equal-trend assumption:
Assumption 3 (Equal-Trend)
E[Y
t+1
ig
(0;0)Y
t
ig
(0;0)jnt] = E[Y
t+1
ig
(0;0)Y
t
ig
(0;0)jco] = E[Y
t+1
ig
(0;0)Y
t
ig
(0;0)]
Under the equal-trend Assumption 3, the common trend E[Y
t+1
ig
(0;0)Y
t
ig
(0;0)] is identified
by E[Y
t+1
ig
Y
t
ig
jZ
t+1
g
= 0]. That is, we take the changes in the outcomes of individuals in
the control group between two periods as the counterfactual trend for compliers and
never-takers in the treatment group.
Lemma 1 (Identification of E[Y
t+1
ig
(0;0)jco])
With equal-trend assumption, the counterfactual outcome E[Y
t+1
ig
(0;0)jco] is identified by
E[Y
t+1
ig
(0;0)jco] = E[Y
t+1
ig
(0;0)Y
t
ig
(0;0)jco] + E[Y
t
ig
(0;0)jco]
= E[Y
t+1
ig
(0;0)Y
t
ig
(0;0)] + E[Y
t
ig
(0;0)jco]
= E[Y
t+1
ig
Y
t
ig
jZ
t+1
g
= 0] + E[Y
t
ig
jD
t+1
g
= 1;Z
t+1
g
= 1]
Theorem 1 (Identification ofITT
t+1
(co))
ITT
t+1
(co) is identified by
ITT
t+1
(co) =
def
E[Y
t+1
ig
(1;1)Y
t+1
ig
(0;0)jco]
= E[Y
t+1
ig
jD
t+1
ig
=Z
t+1
g
= 1]
E[Y
t+1
ig
Y
t
ig
jZ
t+1
g
= 0] + E[Y
t
ig
jD
t+1
ig
= 1;z
t+1
g
= 1]
= E[Y
t+1
ig
Y
t
ig
jD
t+1
ig
=Z
t+1
g
= 1] E[Y
t+1
ig
Y
t
ig
jZ
t+1
g
= 0]
TheITT
t+1
(co) equals to the change in the outcome for compliers after netting out the
common trend.
67
Theorem 2 (Identification ofITT
t+1
(nt))
We can also identify the local ITT of never-takes in the post-period,
ITT
t+1
(nt) =
def
E[Y
t+1
ig
(1;0)Y
t+1
ig
(0;0)jnt]
This is the effect of the treatment assignment on the outcome for never-takers. Note that
when the exclusion restriction is satisfied, we should haveITT
t+1
(nt) = 0. Therefore, we
can test the plausibility of exclusion restriction with the estimates of theITT
t+1
(nt).
Our Difference-in-Differences (DID) style estimator is different from the traditional
DID estimator where exclusion restriction is assumed to be satisfied.
4
Specifically, in the
usual DID setting, the common trend assumption that requires the trends between the
treated sample and the not treated sample are the same is obtained from E[Y
it+1
Y
it
jD
it+1
=
0]. Whereas we assume the trend between the never-takers and compliers are the same.
The common trend is estimated with the sample with the control group, i.e. Z
t+1
= 0.
Nevertheless, our DID-style estimator has a similarity with the traditional DID. If we have
more than two periods of observations before the intervention as in many traditional DID
applications, we can evaluate the validity of our equal-trend assumption with observed
compliers and never-takers in the treatment group.
4.4 Estimation
In the last section, we have shown that we can identify
ITT
t+1
(co) = E[Y
t+1
ig
(1;1)Y
t+1
ig
(0;0)jco]
ITT
t+1
(nt) = E[Y
t+1
ig
(1;0)Y
t+1
ig
(0;0)jnt]
Using the sample analog, such effects can be estimated from data conditional on covariates
using nonparametric estimation method (Hirano et al., 2003). In this section, we instead
use a regression-based estimation approach given convenience and popularity in empirical
applications.
4
We show the proof in Appendix F
68
Let us denote the potential outcome byY
zd
igt
=Y
igt
(z;d). We specify4Y
00
ig;t+1
Y
00
ig;t+1
Y
00
igt
, the potential outcome changes without any intervention by the following linear
model:
4Y
00
ig;t+1
=X
0
ig
+4u
ig;t+1
CovariatesX
ig
are pretreatment characteristics of individuali in clusterg and may include
pretreatment cluster-specific characteristics of clusterg.
5
Under such specification, our
identifying DID-style assumption becomes
E[4u
ig;t+1
jX
ig
=x;nt] = E[4u
ig;t+1
jX
ig
=x;co]
That is, after controlling forX
ig
, idiosyncratic changes are the same across never-takes and
compliers without intervention.
Let us denote the treatment effects for compliers (
igt
) and the treatment effect for the
never-takers (
igt
) as
igt
= Y
11
igt
Y
00
igt
igt
= Y
10
igt
Y
00
igt
Then, the observed changes in outcome can be written as
4Y
ig;t+1
= Z
g;t+1
D
ig;t+1
(Y
11
ig;t+1
Y
00
ig;t
) +Z
g;t+1
(1D
ig;t+1
)(Y
10
ig;t+1
Y
00
ig;t+1
)
+(1Z
g;t+1
)(Y
00
ig;t+1
Y
00
ig;t
)
= 4Y
00
ig;t+1
+Z
g;t+1
(Y
10
ig;t+1
Y
00
ig;t+1
) +Z
g;t+1
D
ig;t+1
(Y
11
ig;t+1
Y
10
ig;t+1
)
= 4Y
00
ig;t+1
+
ig;t+1
Z
g;t+1
+
ig;t+1
Z
g;t+1
D
ig;t+1
= X
0
ig
+
ig;t+1
Z
g;t+1
+
ig;t+1
Z
g;t+1
D
ig;t+1
+4u
ig;t+1
Since there are only two periods, WLOG, we can write it
4Y
ig
=X
0
ig
+
ig
Z
g
+
ig
Z
g
D
ig
+4u
ig
5
CovaritesX
ig
can also be time-variant characteristics that are unlikely to be affected by the treatment—to
avoid post-treatment bias(Montgomery et al., 2018)
69
Similarly, our parameters of interest then become
ITT (co) = E[
ig
jco]
ITT (nt) = E[
ig
jnt]
Denoting
nt
E[
ig
jnt] and
co
E[
ig
jco]. We have
4Y
ig
= X
0
ig
+
nt
Z
g
+
co
Z
g
D
ig
+
ig
where
ig
4u
ig
+ (
ig
nt
)Z
g
+ (
ig
co
)Z
g
D
ig
so that
E[
ig
jX
ig
;Z
g
;D
ig
] = 0
That is because E[4u
ig
jX
ig
;Z
g
;D
ig
] = 0 by DID assumption; E[(
ig
nt
)Z
g
jX
ig
;Z
g
;D
ig
] = 0
when Z
g
= 0; E[
ig
nt
jZ
g
= 1;D
ig
= 0] = E[
ig
nt
jnt] = 0 when Z
g
= 1; E[(
ig
co
)Z
g
D
ig
] = 0 when eitherZ
g
= 0 orD
ig
= 0; E[
ig
co
jZ
g
= 1;D
ig
= 1] = 0 whenZ
g
= 1
andD
ig
= 1.
Therefore, the least square estimators of the regression model
ˆ
nt
and
ˆ
co
are consistent.
4Y
ig
= X
0
ig
+
nt
Z
g
+
co
Z
g
D
ig
+
ig
Since the randomization is at the cluster level in cluster RCTs, we cluster standard errors
at the cluster level. Given the number of clusters (g) and the cluster size (n
g
) are about
equal in most cluster RCTs, the inference of the regression model is standard (Cameron
and Miller, 2015; Duflo et al., 2007).
4.5 Empirical Application
In this section, we estimate the local intent-to-treat (ITT) effect for compliers of a microcre-
dit intervention in rural Morocco. Cr´ epon et al. (2015) evaluates a microcredit program,
operated by Al Amana, introduced in rural areas of Morocco in 2006. The program is
a pair-wise cluster randomized controlled trial where the intervention took place at the
70
village level. One village is randomly assigned to the treatment group and one village to
the control group in each pair. 81 villages were randomly selected as treated villages while
81 villages were selected as control villages. In the treatment villages of the program,
every household in treated villages was given access to a microcredit program where
households can borrow a loan at the low-interest rates. Thirteen percent of the households
in treatment villages took a loan, while households in the control villages were not able
to access the same intervention. This context is in line with assumptions in our analytic
framework that the intervention is randomly assigned to the clusters (villages in this
application) and the non-compliance in the intervention is only one-sided. We use the
online data of Cr´ epon et al. (2015) to conduct the following analysis.
Previous literature on the evaluation of microcredit programs has focused only on ITT
effect which identifies the effect of ”microcredit access” rather than the effect of actual
”microcredit take-up”. Due to the potential spillover of microcredit take-up, it is likely that
exclusion restriction does not hold. Therefore, the previous literature, rightly, avoids using
the standard LATE framework to identify the effect of take-up of microcredit. Though
Cr´ epon et al. (2015) estimated the externalities of borrowing and argued the externalities
of borrowing are not the main driving force for the program effect, they only claimed their
LATE estimates are suggestive. We use our DID assumption to point identify the ITT effect
for compliers. Though the ITT effect for compliers can be interpreted as the impact of
microcredit uptake on compliers, it is not the same as LATE as we have shown above.
We also make some adaptions to the data for our identification strategy. First, since we
do not consider the potential complications raised by attrition in this paper and Cr´ epon
et al. (2015) argued attrition was not a major concern in this program, we only use the
non-attrited sample for the following specification. Second, since our approach relies on
observations of individuals from two periods, we exclude the sample collected only from
the endline. We also provide the summary statistics of the sample used in our empirical
application: non-attrited sample in Appendix Table G1 and non-attrited sample with high
probability to borrow in Appendix Table G2, respectively.
4.5.1 The Empirical Specification of the ITT effect for Compliers and Never-
takers
We further adapt our key identification assumption to the pair-wise randomization feature
of the microcredit intervention. Thus, our identification in this empirical application
71
relies on the assumption that the parallel trends over time between the never-takers and
compliers within each pair of villages. Specifically,
E[4u
ig;t+1
jX
ig
=x;H
p
=;nt] = E[4u
ig;t+1
jX
ig
=x;H
p
=;co]
LetH
p
be an index for the village pair, i.e.,H
p
= 1;2;;81. Since the program is based
on pair-wise randomization, we are assuming that time-effects are the same only for two
villages in the same pair.
Following our assumption, the main specification for the estimation of the ITT (co) and
the ITT (nt) is as follows:
Y
i;g;p;t+1
=X
0
i;g;p;t+1
+
nt
Z
g;p;t+1
+
co
Z
g;p;t+1
D
i;g;p;t+1
+H
p
+
i;g;p;t+1
(4.4)
where i stands for individual, g for villages, p for village pairs; the treatment assignment
Z
g;p;t+1
was assigned at village level;X
0
i;g;p;t+1
time-variant covariates that are not likely
affected by the uptake of microcredit and the treatment assignment are controlled;H
p
pair-specific fixed effect controlled for the pair-specific time-variant changes.
4.5.2 Estimates of ITT for Compliers and Never-takers
The ITT estimates for compliers indicate that the take-up of the microcredit loan has a
substantial effect on households. The take-up of microcredit loan increased the asset value
by 4846, significant at 1 % level; the sales and home consumption combined by 25715,
significant at 1% level; expenses by 16886, significant at 5% level; and the profit of the
home business by 8138, significant at 10% level (Table 4.1, Panel A, Row 1, Column 1, 2, 3,
4). In addition, the take-up of microcredit loans increased self-employment by 5 hours per
week for compliers, significant at the 10% level (Table 4.1, Panel A, Column 8).
Besides, we also find the never-takers were also impacted by the program though
they didn’t borrow from the program. The microcredit intervention increased the sales
and home consumption combined and expenses by 4819 and 3294, respectively, both
significant at the 5% (Table 4.1, Panel A, Row 2, Column 2, 3). However, never-takers
decreased their monthly household consumption by 104, significant at the 5% level. The
statistically significant impacts on these dimensions of outcomes for never-takers even if
they did not borrow from the microcredit program indicate others’ take-ups of microcredit
72
loans might have a general equilibrium effect or spillover effects within the treatment
village. As we discussed in section 2, the significant ITT effects on never-takers suggest
that the exclusion condition is likely to fail in this context.
We also apply the same estimation procedure on the non-attrited sample that has a
high probability to borrow as Cr´ epon et al. (2015) suggested. We find the estimates are
slightly higher but still in similar magnitude. It is consistent with the logic that this group
of individuals is more likely to borrow, thus affected by the take-up of the loan to a greater
extent.
We estimate the ITT effect for compliers and never-taker with a few variations of speci-
fications, results from all specifications are robust. In Appendix Table G3, we show when
not controlling individual characteristics, the estimates are almost the same as estimates
in Table 4.1. Similarly, when not controlling for village pair fixed effect (Appendix Table
G4), we still have similar estimates as Table 4.1.
4.5.3 Tests of the Equal-trend Assumption
Since our estimator only identifies the treatment effect on the treated under the equal-
trend assumption (the Assumption 3), it is crucial for us to validate our assumption with
every possible effort. As we mentioned in section 3, if the cluster-RCTs had two periods
of surveys before the treatment assignment, we would be able to test the assumption by
indirectly testing whether the pre-trend between the compliers and never-takers are the
same using the treatment sample where types are later observed during the intervention
duration. Unfortunately, that is not the case for Cr´ epon et al. (2015)—the study does not
have pre-baseline data.
An alternative approach to test the equal-trend assumption uses the control sample.
We first use the treatment sample to estimate a prediction model of the borrowing decision.
Then, we predict the probability of borrowing for the control individuals with coefficients
from the estimated model. Finally, we classify the types of individuals in the control group
by predicted probability of borrowing—high predicted probability to borrow as compliers
and low predicted probability to borrow as nevertakers—and test whether types differ
in the trend of outcome variables. Since only about 13% individuals in the treatment
are compliers, we separately classify individuals in the top quintile, sextile, septile, and
octile of predicted probability to borrow as compliers and the rest as next-takers in four
separately regressions to test the assumption (See Table 4.2, row 1, 2, 3, and 4). No matter
73
how we choose our threshold for types, results consistently confirm our equal-trend
assumption is valid in this application.
4.5.4 The Comparison with the LATE in Cr´ epon et al. (2015)
Though the LATE in Cr´ epon et al. (2015) is potentially problematic, it is still relevant
to compare our ITT estimates for the treated to the LATE because both estimators are
intended to estimate the impact of the take-up. Since we only use the non-attrited sample
in this microcredit program to estimate the ITT for compliers, we restrict the sample to
the non-attrited individuals when estimating LATE as well (see Table 4.3). The LATE
estimates are much higher than ITT compliers in asset values, expenses (Table 4.3, Panel
A, Row 1, Column 1,3), but are comparable in consumption level and profit. The LATE
estimates also indicate the take-up of the microcredit reduces the probability of being
self-employed and working outside, in much larger magnitudes than the ITT effect for
compliers. The difference between our estimates of the ITT effect for compliers and LATE
estimates stem from the latter did not account for potential spillover within clusters, thus
violating the exclusion restriction.
4.6 Conclusion
In this paper, we study the causal identification of average treatment effect on the treated
in the cluster randomized controlled trials with spillover and noncompliance. We first
present the case that the average treatment effect of the treated can not be identified with
the local average treatment effect due to spillovers through social interactions within
clusters/general equilibrium effect of clusters. Then, we propose a simple DID-style
estimator for the average treatment on the treated—ITT effect for compliers —in this
setting. We also point identify ITT effect for never-takers which can be used to validate the
existence of the spillovers. In the empirical application of a microcredit program (Cr´ epon
et al., 2015), we show the ITT effect for compliers and LATE estimates are different and the
differences stem from the latter does not account for potential spillover within clusters.
74
Table 4.1: ITT(co) & ITT(nt) estimates with Baseline Covariates & Pair Fixed Effect
Panel A: Whole Non-attrited Sample
Asset Sales+home Has a self- Income from day Weekly hours worked Monthly HH
(stock) consumption Expense Profit employment activity labor/ salaried Self-employment Outside Consumption
ITT, complier (
1
) 4846.328
25715.447
16886.286
8138.491
-0.007 -1386.387 4.993
-2.101 329.545
(2047.501) (8155.426) (6948.572) (4918.364) (0.027) (1505.514) (2.992) (3.168) (292.178)
ITT, never taker 751.717 4819.461
3294.213
1575.185 0.015 -518.162 2.468
-0.673 -104.341
(703.962) (2028.907) (1571.594) (1504.569) (0.012) (634.364) (1.395) (1.245) (46.663)
Baseline Covariates X X X X X X X X X
Pair fixed effect X X X X X X X X X
Observations 4049 4049 4049 4049 4049 4049 4037 4037 4039
control mean 1798.850 4721.573 4266.519 1152.695 0.030 3692.911 21.208 0.313 614.806
Panel B: Non-attrited Sample with High Probability to Borrow
ITT, complier (
1
) 4807.066
26789.667
16614.551
9728.067
-0.009 -1853.078 3.567 -2.758 370.639
(2227.440) (8846.859) (7125.683) (5841.032) (0.028) (1592.119) (3.040) (3.376) (314.811)
ITT, never taker 1189.898 7038.431
5324.645
2072.388 0.018 -699.157 1.709 -0.772 -85.426
(753.183) (2387.208) (1847.035) (1620.218) (0.013) (710.586) (1.384) (1.260) (53.021)
Baseline Covariates X X X X X X X X X
Pair fixed effect X X X X X X X X X
Observations 3479 3479 3479 3479 3479 3479 3467 3467 3470
control mean 1811.076 4421.010 4150.940 772.065 0.033 4057.138 22.350 0.381 604.385
p< 0:10,p< 0:05,p< 0:01. Covariates are changes of number of adults in the households and number of children in the households.
75
Table 4.2: Tests of Parallel Trends in the Control Sample
Panel A: Propensity to Borrow Estimated by Logit Regression
Asset Sales+home Has a self- Income from day Weekly hours worked Monthly HH
(stock) consumption Expense Profit employment activity labor/ salaried Self-employment Outside Consumption
High Propensity 2048.809 24230.696 14289.524 11773.418 -0.028 -2336.005 0.360 -9.868 157.274
to Borrow (top quintile) (3624.546) (18169.483) (15987.412) (8953.303) (0.053) (2491.958) (5.174) (7.590) (274.078)
High Propensity 4278.175 6801.342 4701.645 5720.039 0.018 881.950 8.421 5.312 89.592
to Borrow (top sextile) (3278.394) (16157.358) (11296.354) (10171.640) (0.058) (2803.242) (5.599) (6.133) (238.546)
High Propensity 3414.918 -2744.137 -1148.447 658.314 0.054 -1001.388 4.711 -2.640 -70.289
to Borrow (top septile) (2812.556) (16094.317) (10094.848) (10938.956) (0.051) (2151.421) (5.596) (6.376) (229.447)
High Propensity 2566.991 -494.170 2269.574 -197.027 0.065 -1480.037 4.659 -0.797 41.617
to Borrow (top octile) (2669.526) (15153.216) (9821.940) (9940.516) (0.051) (1972.569) (5.251) (5.341) (232.643)
Baseline Controls X X X X X X X X X
Pair Fixed Effect X X X X X X X X X
Observations 2045 2045 2045 2045 2045 2045 2039 2039 2039
p< 0:10,p< 0:05,p< 0:01. Baseline Controls are number of household members, number of adults, head age, does animal husbandry, does other non-agricultural
activity, had an outstanding loan over the past 12 months, HH spouse responded to the survey, and other HH member responded to the survey.
76
Table 4.3: Comparison between ITT effect for Compliers and LATE
Panel A: Whole Non-attrited Sample
Asset Sales+home Has a self- Income from day Weekly hours worked Monthly HH
(stock) consumption Expense Profit employment activity labor/ salaried Self-employment Outside Consumption
ITT, complier (
1
) 4846.328
25715.447
16886.286
8138.491
-0.007 -1386.387 4.993
-2.101 329.545
(2047.501) (8155.426) (6948.572) (4918.364) (0.027) (1505.514) (2.992) (3.168) (292.178)
LATE (
2
) 10106.961
59442.802
39939.713
19161.886 0.095 -5012.568 22.319
-6.825 -404.695
(5075.838) (16495.541) (13034.667) (11961.646) (0.094) (4926.633) (10.748) (8.951) (379.467)
1
=
2
(P-value) 0.280 0.018 0.036 0.291 0.223 0.409 0.074 0.585 0.021
Baseline Covariates X X X X X X X X X
Pair fixed effect X X X X X X X X X
Observations 4105 4105 4105 4105 4105 4105 4093 4093 4095
control mean 1879.392 4958.848 4451.599 1190.183 0.034 3756.474 21.127 0.607 617.163
Panel B: Non-attrited Sample with High Probability to Borrow
ITT, complier (
1
) 4807.066
26789.667
16614.551
9728.067
-0.009 -1853.078 3.567 -2.758 370.639
(2227.440) (8846.859) (7125.683) (5841.032) (0.028) (1592.119) (3.040) (3.376) (314.811)
LATE (
2
) 12610.750
72949.654
51535.051
23319.366
0.109 -6438.343 14.813 -7.841 -192.674
(5232.957) (18286.628) (14338.229) (12286.062) (0.097) (5168.270) (10.199) (8.519) (420.003)
1
=
2
(P-value) 0.109 0.003 0.004 0.195 0.168 0.320 0.212 0.535 0.095
Baseline Covariates X X X X X X X X X
Pair fixed effect X X X X X X X X X
Observations 3525 3525 3525 3525 3525 3525 3513 3513 3516
control mean 1924.312 4704.979 4247.166 951.485 0.038 3982.046 22.325 0.526 608.610
p< 0:10,p< 0:05,p< 0:01. Baseline Controls are number of household members, number of adults, head age, does animal husbandry, does other non-agricultural
activity, had an outstanding loan over the past 12 months, HH spouse responded to the survey, and other HH member responded to the survey.
77
Chapter 5
Conclusion
The chapters of this dissertation provide empirical evidence of the importance of early
childhood development interventions. Chapter 2 shows that the impact of an early
childhood development intervention could extend to a medium-term impact on skill
formations of children as well as changing parental time investments and preschool
enrollment. These impacts potentially could contribute to long-term benefits on the labor
market founded by the small efficacy trials in the literature(Heckman et al., 2010; Walker
et al., 2011; Gertler et al., 2014; Campbell et al., 2014). Chapter 3 documents the role
of social interactions in early childhood development. The results of Chapter 3 indicate
that providing free access to an indoor playground had a positive impact on both child
development and parenting practices and show suggestive evidence that the impacts are
mediating through social interactions within the village. These also have important policy
implications regarding the cost-effectiveness of group-based ECD interventions. Since
social interactions usually lead to spillover effects that are often unaccounted for, the
effectiveness of group-based interventions might be larger than what current studies have
documented, which could change the calculation of cost-effectiveness of group-based ECD
interventions.
The chapters also present the case that policy interventions could affect the develop-
ment outcome of interests through social interactions, which requires research to pay more
attention to social interactions as the mechanism of the impacts and careful identification
strategies. Although limited by the experiment design and data availability, Chapter 3
still shows suggestive evidence that social interactions played a role in an early childhood
development interaction. When the stable unit treatment variable assumption breaks
down because of spillovers caused by social interactions or general equilibrium, Chapter
4 provides a solution to causally identify the overall impact of the intervention on the
78
treated in the cluster RCTs that have spillovers in the clusters, such as the intervention in
Chapter 2.
1
.
1
However, I cannot apply the estimator in Chapter 4 to Chapter 3 intervention as the playground
intervention does not have baseline data
79
Bibliography
Abbring, J. H. and Heckman, J. J. (2007). Econometric evaluation of social programs, part
iii: Distributional treatment effects, dynamic treatment effects, dynamic discrete choice,
and general equilibrium policy evaluation. Handbook of econometrics, 6:5145–5303.
Aboud, F. E., Singla, D. R., Nahil, M. I., and Borisova, I. (2013). Effectiveness of a parenting
program in bangladesh to address early childhood health, growth and development.
Social Science & Medicine, 97:250–258.
Aboud, F. E. and Yousafzai, A. K. (2015). Global health and development in early childhood.
Annual review of psychology, 66:433–457.
Agostinelli, F. (2018). Investing in children’s skills: An equilibrium analysis of social
interactions and parental investments. Unpublished Manuscript, Arizona State University.
Andrew, A., Attanasio, O., Bernal, R., Sosa, L. C., Krutikova, S., and Rubio-Codina, M.
(2019). Preschool quality and child development. Technical report, National Bureau of
Economic Research.
Andrew, A., Attanasio, O., Fitzsimons, E., Grantham-McGregor, S., Meghir, C., and Rubio-
Codina, M. (2018). Impacts 2 years after a scalable early childhood development
intervention to increase psychosocial stimulation in the home: A follow-up of a cluster
randomised controlled trial in colombia. PLoS medicine, 15(4):e1002556.
Angelucci, M. and De Giorgi, G. (2009). Indirect effects of an aid program: how do cash
transfers affect ineligibles’ consumption? American Economic Review, 99(1):486–508.
Angrist, J. D., Imbens, G. W., and Rubin, D. B. (1996). Identification of causal effects using
instrumental variables. Journal of the American statistical Association, 91(434):444–455.
Araujo, M. C., Carneiro, P ., Cruz-Aguayo, Y., and Schady, N. (2016). Teacher quality and
learning outcomes in kindergarten. The Quarterly Journal of Economics, 131(3):1415–
1453.
Attanasio, O., Cattan, S., Fitzsimons, E., Meghir, C., and Rubio-Codina, M. (2020). Estimat-
ing the production function for human capital: results from a randomized controlled
trial in colombia. American Economic Review, 110(1):48–85.
Attanasio, O., Meghir, C., and Nix, E. (2015). Human capital development and parental
investment in india. Discussion paper, National Bureau of Economic Research.
80
Attanasio, O. P ., Fern´ andez, C., Fitzsimons, E. O., Grantham-McGregor, S. M., Meghir, C.,
and Rubio-Codina, M. (2014). Using the infrastructure of a conditional cash transfer
program to deliver a scalable integrated early child development program in colombia:
cluster randomized controlled trial. BMJ, 349:g5785.
Bailey, D., Duncan, G. J., Odgers, C. L., and Yu, W. (2017). Persistence and fadeout in
the impacts of child and adolescent interventions. Journal of research on educational
effectiveness, 10(1):7–39.
Baker, M., Gruber, J., and Milligan, K. (2008). Universal child care, maternal labor supply,
and family well-being. Journal of political Economy, 116(4):709–745.
Baker, M., Gruber, J., and Milligan, K. (2019). The long-run impacts of a universal child
care program. American Economic Journal: Economic Policy, 11(3):1–26.
Baker-Henningham, H. and L´ opez B´ oo, F. (2010). Early childhood stimulation interventions
in developing countries: a comprehensive literature review.
Banerjee, A., Duflo, E., and Glennerster, R. (2011). Is decentralized iron fortification a
feasible option to fight anemia among the poorest? In Explorations in the Economics of
Aging, pages 317–344. University of Chicago Press.
Barnett, W. S. (2011). Effectiveness of early educational intervention. Science,
333(6045):975–978.
Bartlett, M. S. (1937). The statistical conception of mental factors. British journal of
Psychology, 28(1):97–104.
Bayley, N. (2006). Bayley scales of infant and toddler development: administration manual.
Harcourt assessment.
Berlinski, S., Galiani, S., and Gertler, P . (2009). The effect of pre-primary education on
primary school performance. Journal of public Economics, 93(1-2):219–234.
Berlinski, S., Galiani, S., and Manacorda, M. (2008). Giving children a better start:
Preschool attendance and school-age profiles. Journal of public Economics, 92(5-6):1416–
1440.
Black, M. M., Walker, S. P ., Fernald, L. C., Andersen, C. T., DiGirolamo, A. M., Lu, C.,
McCoy, D. C., Fink, G., Shawar, Y. R., Shiffman, J., et al. (2017). Early childhood
development coming of age: science through the life course. The Lancet, 389(10064):77–
90.
Bloom, H. S. (2005). Randomizing groups to evaluate place-based programs. Learning
more from social experiments: Evolving analytic approaches, pages 115–172.
Bobonis, G. J. and Finan, F. (2009). Neighborhood peer effects in secondary school enroll-
ment decisions. The Review of Economics and Statistics, 91(4):695–716.
81
Brown, W. H., Odom, S. L., and Conroy, M. A. (2001). An intervention hierarchy for
promoting young children’s peer interactions in natural environments. Topics in early
childhood special education, 21(3):162–175.
Bruhn, M. and McKenzie, D. (2009). In pursuit of balance: Randomization in practice in
development field experiments. American economic journal: applied economics, 1(4):200–
232.
Calv´ o-Armengol, A., Patacchini, E., and Zenou, Y. (2009). Peer effects and social networks
in education. The Review of Economic Studies, 76(4):1239–1267.
Cameron, A. C. and Miller, D. L. (2015). A practitioner’s guide to cluster-robust inference.
Journal of human resources, 50(2):317–372.
Campbell, F., Conti, G., Heckman, J. J., Moon, S. H., Pinto, R., Pungello, E., and Pan,
Y. (2014). Early childhood investments substantially boost adult health. Science,
343(6178):1478–1485.
Carlsson, M., Dahl, G. B.,
¨
Ockert, B., and Rooth, D.-O. (2015). The effect of schooling on
cognitive skills. Review of Economics and Statistics, 97(3):533–547.
Cascio, E. U. (2015). The promises and pitfalls of universal early education. IZA World of
Labor.
Cattell, R. B. (1966). The scree test for the number of factors. Multivariate behavioral
research, 1(2):245–276.
Chattopadhyay, R. and Duflo, E. (2004). Women as policy makers: Evidence from a
randomized policy experiment in india. Econometrica, 72(5):1409–1443.
Chetty, R., Friedman, J. N., Hilger, N., Saez, E., Schanzenbach, D. W., and Yagan, D. (2011).
How does your kindergarten classroom affect your earnings? evidence from project star.
The Quarterly Journal of Economics, 126(4):1593–1660.
Chetty, R., Friedman, J. N., and Rockoff, J. E. (2014). Measuring the impacts of teachers
ii: Teacher value-added and student outcomes in adulthood. American economic review,
104(9):2633–79.
Conley, T. G. and Udry, C. R. (2010). Learning about a new technology: Pineapple in
ghana. American economic review, 100(1):35–69.
Cornelissen, T., Dustmann, C., Raute, A., and Sch¨ onberg, U. (2018). Who benefits from
universal child care? estimating marginal returns to early child care attendance. Journal
of Political Economy, 126(6):2356–2409.
Cr´ epon, B., Devoto, F., Duflo, E., and Parient´ e, W. (2015). Estimating the impact of
microcredit on those who take it up: Evidence from a randomized experiment in
morocco. American Economic Journal: Applied Economics, 7(1):123–50.
82
Cunha, F. and Heckman, J. J. (2008). Formulating, identifying and estimating the technol-
ogy of cognitive and noncognitive skill formation. Journal of human resources, 43(4):738–
782.
Cunha, F., Heckman, J. J., and Schennach, S. M. (2010). Estimating the technology of
cognitive and noncognitive skill formation. Econometrica, 78(3):883–931.
Currie, J. and Almond, D. (2011). Human capital development before age five. In Handbook
of labor economics, volume 4, pages 1315–1486. Elsevier.
Currie, J. and Thomas, D. (2000). School quality and the longer-term effects of head start.
Journal of Human Resources, 35(4):755–774.
De Giorgi, G., Pellizzari, M., and Redaelli, S. (2010). Identification of social interac-
tions through partially overlapping peer groups. American Economic Journal: Applied
Economics, 2(2):241–75.
Del Rosso, J. M. and Marek, T. (1996). Class action: Improving school performance in the
developing world through better Health, Nutrition and Population. The World Bank.
Dickey, W. C. and Blumberg, S. J. (2004). Revisiting the factor structure of the strengths
and difficulties questionnaire: United states, 2001. Journal of the American Academy of
Child & Adolescent Psychiatry, 43(9):1159–1167.
Du, Y., Kou, J., and Coghill, D. (2008). The validity, reliability and normative scores of the
parent, teacher and self report versions of the strengths and difficulties questionnaire in
china. Child and adolescent psychiatry and mental health, 2(1):8.
Duflo, E., Glennerster, R., and Kremer, M. (2007). Using randomization in development
economics research: A toolkit. Handbook of development economics, 4:3895–3962.
Duflo, E. and Saez, E. (2003). The role of information and social interactions in retirement
plan decisions: Evidence from a randomized experiment. The Quarterly journal of
economics, 118(3):815–842.
Eckerman, C. O. and Peterman, K. (2001). Peers and infant social/communicative develop-
ment.
Eckerman, C. O., Whatley, J. L., and Kutz, S. L. (1975). Growth of social play with peers
during the second year of life. Developmental Psychology, 11(1):42.
Eickmann, S. H., Guerra, M. Q., Lima, M. C., Huttly, S. R., and Worth, A. A. (2003).
Improved cognitive and motor development in a community-based intervention of
psychosocial stimulation in northeast brazil. Developmental Medicine & Child Neurology,
45(8):536–541.
Elango, S., Garc´ ıa, J. L., Heckman, J. J., and Hojman, A. (2015). Early childhood education.
In Economics of Means-Tested Transfer Programs in the United States, Volume 2, pages
235–297. University of Chicago Press.
83
Elder, T. E. and Lubotsky, D. H. (2009). Kindergarten entrance age and children’s achieve-
ment impacts of state policies, family background, and peers. Journal of human Resources,
44(3):641–683.
Epple, D. and Romano, R. E. (2011). Peer effects in education: A survey of the theory and
evidence. In Handbook of social economics, volume 1, pages 1053–1163. Elsevier.
Felfe, C., Nollenberger, N., and Rodr´ ıguez-Planas, N. (2015). Can’t buy mommy’s love?
universal childcare and children’s long-term cognitive development. Journal of population
economics, 28(2):393–422.
Fernald, L. C., Kagawa, R., Knauer, H. A., Schnaas, L., Guerra, A. G., and Neufeld, L. M.
(2017). Promoting child development through group-based parent support within a
cash transfer program: Experimental effects on children’s outcomes. Developmental
psychology, 53(2):222.
Flores, C. A. and Flores-Lagunes, A. (2013). Partial identification of local average treatment
effects with an invalid instrument. Journal of Business & Economic Statistics, 31(4):534–
545.
Foster, A. D. and Rosenzweig, M. R. (1995). Learning by doing and learning from oth-
ers: Human 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.
Annu. Rev. Econ., 2(1):395–424.
Garner, B. P . and Bergen, D. (2006). Play development from birth to age four. Play from
birth to twelve: Contexts, perspectives, and meanings, pages 3–12.
Gertler, P ., Heckman, J., Pinto, R., Zanolini, A., Vermeersch, C., Walker, S., Chang, S. M.,
and Grantham-McGregor, S. (2014). Labor market returns to an early childhood stimu-
lation intervention in jamaica. Science, 344(6187):998–1001.
Goodman, A., Lamping, D. L., and Ploubidis, G. B. (2010). When to use broader inter-
nalising and externalising subscales instead of the hypothesised five subscales on the
strengths and difficulties questionnaire (sdq): data from british parents, teachers and
children. Journal of abnormal child psychology, 38(8):1179–1191.
Goodman, R. (1997). The strengths and difficulties questionnaire: a research note. Journal
of child psychology and psychiatry, 38(5):581–586.
Goodman, R., Ford, T., Simmons, H., Gatward, R., and Meltzer, H. (2000). Using the
strengths and difficulties questionnaire (sdq) to screen for child psychiatric disorders in
a community sample. The British journal of psychiatry, 177(6):534–539.
Gorsuch, R. L. (2003). Factor analysis. weiner, i. Handbook of Psychology: Research Methods
in Psychology, pages 143–164.
84
Grantham-McGregor, S., Adya, A., Attanasio, O., Augsburg, B., Behrman, J., Caeyers, B.,
Day, M., Jervis, P ., Kochar, R., Makkar, P ., et al. (2020). Group sessions or home visits for
early childhood development in india: a cluster rct. Pediatrics.
Grantham-McGregor, S. M., Powell, C. A., Walker, S. P ., and Himes, J. H. (1991). Nutri-
tional supplementation, psychosocial stimulation, and mental development of stunted
children: the jamaican study. The Lancet, 338(8758):1–5.
Gupta, N. D. and Simonsen, M. (2010). Non-cognitive child outcomes and universal high
quality child care. Journal of Public Economics, 94(1-2):30–43.
Hamadani, J. D., Mehrin, S. F., Tofail, F., Hasan, M. I., Huda, S. N., Baker-Henningham,
H., Ridout, D., and Grantham-McGregor, S. (2019). Integrating an early childhood
development programme into bangladeshi primary health-care services: an open-label,
cluster-randomised controlled trial. The Lancet Global Health, 7(3):e366–e375.
Hanna, E. and Meltzoff, A. N. (1993). Peer imitation by toddlers in laboratory, home,
and day-care contexts: Implications for social learning and memory. Developmental
psychology, 29(4):701.
Havnes, T. and Mogstad, M. (2011). No child left behind: Subsidized child care and
children’s long-run outcomes. American Economic Journal: Economic Policy, 3(2):97–129.
Havnes, T. and Mogstad, M. (2015). Is universal child care leveling the playing field?
Journal of public economics, 127:100–114.
Heckman, J., Pinto, R., and Savelyev, P . (2013). Understanding the mechanisms through
which an influential early childhood program boosted adult outcomes. American Eco-
nomic Review, 103(6):2052–86.
Heckman, J. J., Liu, B., Lu, M., Zhou, J., et al. (2020). Treatment effects and the measure-
ment of skills in a prototypical home visiting program. Technical report.
Heckman, J. J., Lochner, L., and Taber, C. (1998a). Explaining rising wage inequality: Explo-
rations with a dynamic general equilibrium model of labor earnings with heterogeneous
agents. Review of economic dynamics, 1(1):1–58.
Heckman, J. J., Lochner, L., and Taber, C. (1998b). General equilibrium treatment effects:
A study of tuition policy. Technical report, National Bureau of Economic Research.
Heckman, J. J., Lochner, L., and Taber, C. (1998c). Tax policy and human capital formation.
Technical report, National Bureau of Economic Research.
Heckman, J. J., Moon, S. H., Pinto, R., Savelyev, P . A., and Yavitz, A. (2010). The rate of
return to the highscope perry preschool program. Journal of public Economics, 94(1):114–
128.
Herbst, C. M. (2017). Universal child care, maternal employment, and children’s long-
run outcomes: Evidence from the us lanham act of 1940. Journal of Labor Economics,
35(2):519–564.
85
Hirano, K., Imbens, G. W., and Ridder, G. (2003). Efficient estimation of average treatment
effects using the estimated propensity score. Econometrica, 71(4):1161–1189.
Horn, J. L. (1965). A rationale and test for the number of factors in factor analysis.
Psychometrika, 30(2):179–185.
Hu, B. Y., Zhou, Y., Li, K., and Killingsworth Roberts, S. (2014). Examining program
quality disparities between urban and rural kindergartens in china: Evidence from
zhejiang. Journal of Research in Childhood Education, 28(4):461–483.
Imbens, G. W. and Angrist, J. D. (1994). Identification and estimation of local average
treatment effects. Econometrica, 62(2):467–475.
Imbens, G. W. and Rubin, D. B. (1997). Bayesian inference for causal effects in randomized
experiments with noncompliance. The annals of statistics, pages 305–327.
J-PAL (2020). Encouraging early childhood stimulation from parents and caregivers to
improve child development.
Johnston, D., Propper, C., Pudney, S., and Shields, M. (2014). Child mental health and
educational attainment: multiple observers and the measurement error problem. Journal
of Applied Econometrics, 29(6):880–900.
Knudsen, E. I., Heckman, J. J., Cameron, J. L., and Shonkoff, J. P . (2006). Economic,
neurobiological, and behavioral perspectives on building america’s future workforce.
Proceedings of the National Academy of Sciences, 103(27):10155–10162.
Kremer, M. and Miguel, E. (2007). The illusion of sustainability. The Quarterly journal of
economics, 122(3):1007–1065.
Laajaj, R. and Macours, K. (2017). Measuring skills in developing countries. The World Bank.
Lai, Y., Xue, X., and Yang, R. (2015). On the allocation problems of teacher resources of
pre-school education in rural areas and its optimized path. 36(03):103–111.
Lalive, R. and Cattaneo, M. A. (2009). Social interactions and schooling decisions. The
Review of Economics and Statistics, 91(3):457–477.
Lalive, R., Landais, C., and Zweim¨ uller, J. (2015). Market externalities of large unemploy-
ment insurance extension programs. American Economic Review, 105(12):3564–96.
Law, M., King, S., Stewart, D., and King, G. (2002). The perceived effects of parent-led
support groups for parents of children with disabilities. Physical & Occupational Therapy
in Pediatrics, 21(2-3):29–48.
List, J. A., Momeni, F., and Zenou, Y. (2019). Are estimates of early education programs
too pessimistic? evidence from a large-scale field experiment that causally measures
neighbor effects.
86
Lu, C., Black, M. M., and Richter, L. M. (2016). Risk of poor development in young children
in low-income and middle-income countries: an estimation and analysis at the global,
regional, and country level. The Lancet Global Health, 4(12):e916–e922.
Luo, R., Jia, F., Yue, A., Zhang, L., Lyu, Q., Shi, Y., Yang, M., Medina, A., Kotb, S., and
Rozelle, S. (2019). Passive parenting and its association with early child development.
Early Child Development and Care, 189(10):1709–1723.
Magnuson, K. A., Ruhm, C., and Waldfogel, J. (2007). Does prekindergarten improve
school preparation and performance? Economics of Education review, 26(1):33–51.
Manski, C. F. (1993). Identification of endogenous social effects: The reflection problem.
The review of economic studies, 60(3):531–542.
Martinez, S., Naudeau, S., and Pereira, V. (2012). The promise of preschool in africa: A
randomized impact evaluation of early childhood development in rural mozambique.
McCabe, H. (2008). The importance of parent-to-parent support among families of
children with autism in the people’s republic of china. International Journal of Disability,
Development and Education, 55(4):303–314.
McLeroy, K. R., Norton, B. L., Kegler, M. C., Burdine, J. N., and Sumaya, C. V. (2003).
Community-based interventions.
Mealli, F. and Pacini, B. (2013). Using secondary outcomes to sharpen inference in ran-
domized experiments with noncompliance. Journal of the American Statistical Association,
108(503):1120–1131.
Merzel, C. and D’Afflitti, J. (2003). Reconsidering community-based health promotion:
promise, performance, and potential. American journal of public health, 93(4):557–574.
Miguel, E. and Kremer, M. (2004). Worms: identifying impacts on education and health in
the presence of treatment externalities. Econometrica, 72(1):159–217.
Mobius, M. and Rosenblat, T. (2014). Social learning in economics. Annu. Rev. Econ.,
6(1):827–847.
Moffitt, R. A. et al. (2001). Policy interventions, low-level equilibria, and social interactions.
Social dynamics, 4(45-82):6–17.
Montgomery, J. M., Nyhan, B., and Torres, M. (2018). How conditioning on posttreatment
variables can ruin your experiment and what to do about it. American Journal of Political
Science, 62(3):760–775.
Munshi, K. (2004). Social learning in a heterogeneous population: technology diffusion in
the indian green revolution. Journal of development Economics, 73(1):185–213.
87
Nahar, B., Hossain, M., Hamadani, J., Ahmed, T., Huda, S., Grantham-McGregor, S., and
Persson, L. (2012). Effects of a community-based approach of food and psychosocial stim-
ulation on growth and development of severely malnourished children in bangladesh: a
randomised trial. European Journal of Clinical Nutrition, 66(6):701–709.
NBS (2015). China statistical yearbook 2015.
Nores, M., Bernal, R., and Barnett, W. S. (2019). Center-based care for infants and toddlers:
The aeiotu randomized trial. Economics of Education Review, 72:30–43.
Ozier, O. (2018). Exploiting externalities to estimate the long-term effects of early child-
hood deworming. American Economic Journal: Applied Economics, 10(3):235–62.
Richter, L. M., Daelmans, B., Lombardi, J., Heymann, J., Boo, F. L., Behrman, J. R., Lu, C.,
Lucas, J. E., Perez-Escamilla, R., Dua, T., et al. (2017). Investing in the foundation of
sustainable development: pathways to scale up for early childhood development. The
Lancet, 389(10064):103–118.
Rubin, D. B. (1974). Estimating causal effects of treatments in randomized and nonran-
domized studies. Journal of educational Psychology, 66(5):688.
Rubin, D. B. and Zell, E. R. (2010). Dealing with noncompliance and missing outcomes in
a randomized trial using bayesian technology: Prevention of perinatal sepsis clinical
trial, soweto, south africa. Statistical Methodology, 7(3):338–350.
Sacerdote, B. (2001). Peer effects with random assignment: Results for dartmouth room-
mates. The Quarterly journal of economics, 116(2):681–704.
Skrondal, A. and Rabe-Hesketh, S. (2009). Prediction in multilevel generalized linear
models. Journal of the Royal Statistical Society: Series A (Statistics in Society), 172(3):659–
687.
Sylvia, S., Warrinnier, N., Luo, R., Yue, A., Attanasio, O., Medina, A., and Rozelle, S.
(2020a). From quantity to quality: Delivering a home-based parenting intervention
through china’s family planning cadres. The Economic Journal.
Sylvia, S., Warrinnier, N., Luo, R., Yue, A., Attanasio, O., Medina, A., and Rozelle, S.
(2020b). From Quantity to Quality: Delivering a Home-based Parenting Intervention
through China’s Family Planning Cadres. The Economic Journal. ueaa114.
Thompson, B. (2004). Exploratory and confirmatory factor analysis. American Psychological
Association.
Ulferts, H., Wolf, K. M., and Anders, Y. (2019). Impact of process quality in early child-
hood education and care on academic outcomes: Longitudinal meta-analysis. Child
development, 90(5):1474–1489.
Walker, S. P ., Chang, S. M., Vera-Hern´ andez, M., and Grantham-McGregor, S. (2011).
Early childhood stimulation benefits adult competence and reduces violent behavior.
Pediatrics, 127(5):849–857.
88
Walker, S. P ., Wachs, T. D., Meeks Gardner, J., Lozoff, B., Wasserman, G. A., Pollitt, E., and
Carter, J. A. (2007). Child development: risk factors for adverse outcomes in developing
countries. The Lancet, 369(9556):145–157.
Wang, L., Liang, W., Zhang, S., Jonsson, L., Li, M., Yu, C., Sun, Y., Ma, Q., Bai, Y., Abbey,
C., et al. (2019). Are infant/toddler developmental delays a problem across rural china?
Journal of Comparative Economics, 47(2):458–469.
Wechsler, D. (2012). Wechsler preschool and primary scale of intelligence—fourth edition. The
Psychological Corporation San Antonio, TX.
Wechsler, D. (2014). Wechsler preschool and primary scale of intelligence—fourth cn
edition (wppsi-iv cn).(y. li & z. j, trans. y. li & z. j eds.). King-May Company China.
Wu, Z. H. (2017). China rural education development report 2017.
Yousafzai, A. K., Obradovi´ c, J., Rasheed, M. A., Rizvi, A., Portilla, X. A., Tirado-Strayer,
N., Siyal, S., and Memon, U. (2016). Effects of responsive stimulation and nutrition
interventions on children’s development and growth at age 4 years in a disadvantaged
population in pakistan: a longitudinal follow-up of a cluster-randomised factorial
effectiveness trial. The Lancet Global Health, 4(8):e548–e558.
Yousafzai, A. K., Rasheed, M. A., Rizvi, A., Armstrong, R., and Bhutta, Z. A. (2014). Effect
of integrated responsive stimulation and nutrition interventions in the lady health
worker programme in pakistan on child development, growth, and health outcomes: a
cluster-randomised factorial effectiveness trial. The Lancet, 384(9950):1282–1293.
Yue, A., Shi, Y., Luo, R., Medina, A., Kotb, S., and Rozelle, S. (2017). Parenting at arm’s
length: Early development and the absence of interactive parenting in rural china.
Parenting.
Zhao, L. and Hu, X. (2008). The development of early childhood education in rural areas
in china. Early Years, 28(2):197–209.
Zhong, J., Luo, R., Sylvia, S., Dill, S.-E., Medina, A., and Rozelle, S. (2020). Passive versus
active service delivery: Comparing the effects of two parenting interventions on early
cognitive development in rural china.
89
Appendices
A Appendix A: Distribution of Latent Skills in WPPSI-IV
(a) Information Test (b) Similarities Test
Figure A1: Distribution WPPSI-IV Verbal Comprehension Skills by Treatment Assignment
90
(a) Block Design Test (b) Object Assembly Test
Figure A2: Distribution WPPSI-IV Visual Spatial Skills by Treatment Assignment
(a) Matrix Reasoning Test (b) Picture Concepts Test
Figure A3: Distribution WPPSI-IV Fluid Reasoning Skills by Treatment Assignment
91
(a) Zoo Locations Test (b) Picture Memory Test
Figure A4: Distribution WPPSI-IV Working Memory Skills by Treatment Assignment
92
Table A1: Descriptive Statistics and Balance Adjusted for Attrition
(1) (2) (3)
Control
(N=276)
Treatment
(N=199)
P-value
Panel A. Child Characteristics
(1) Age in months 25.019 25.075 0.9514
[3.284] [3.367]
(2) Male 0.454 0.525 0.1241
[0.499] [0.501]
(3) Low birth weight 0.038 0.041 0.9411
[0.191] [0.199]
(4) First born 0.568 0.601 0.3566
[0.496] [0.491]
(5) Anemia (Hb<110 g/L) 0.241 0.283 0.7963
[0.428] [0.452]
(6) Days ill past month 4.459 4.589 0.8192
[5.150] [5.463]
(7) Cognitive Delay (BSID MDI<80) 0.457 0.389 0.3045
[0.499] [0.489]
(8) Motor Delay (BSID PDI<80) 0.113 0.102 0.7960
[0.318] [0.303]
(9) Social-Emotional Problems(ASQ:SE>60) 0.255 0.283 0.5454
[0.437] [0.452]
Panel B. Household Characteristics
(1) Social security support recipient 0.351 0.333 0.8354
[0.478] [0.473]
(20 Mother at home 0.684 0.636 0.2079
[0.466] [0.482]
(3) Caregiver education 9 years 0.652 0.657 0.7494
[0.477] [0.476]
(4) Unfavorable perception of FPC 2.832 2.832 0.9460
[0.600] [0.630]
Panel C. Parental Inputs
(1) Told story to child yesterday 0.115 0.117 0.9338
[0.319] [0.322]
(2) Read book to child yesterday 0.045 0.041 0.7929
[0.208] [0.198]
(3) Sang song to child yesterday 0.373 0.354 0.6359
[0.485] [0.479]
(4) Played with child yesterday 0.324 0.354 0.5966
[0.469] [0.479]
(5) Number of books in household 1.512 1.924 0.4193
[3.475] [4.644]
Standard deviation in the bracket; P-values account for clustering within villages.
93
B Appendix B: Measurement System
This appendix provides further detail about the dimensionality reduction techniques and
measurement systems of infant skills, parental investment and school quality.
B.1. Infant Cognitive Skills
Infant cognitive skills are measured using the Wechsler Preschool and Primary Scale of
Intelligence (WPPSI-IV)(Wechsler, 2012). The WPPSI-IV consists of several individually
administered subtests, each of which measures a specific area of cognitive ability. In each
sub-test in index (i)-(iv) administered test-items increase in difficulty level and the test
is stopped when a child can no longer provide a correct answer. Given this specific test
structure, we first estimate a two-parameter logistic IRT measurement system for each
of the eight sub-tests which calculates the optimal weighted average of all items taking
into account response patters. Conceptually, IRT models can be viewed as an extension of
confirmatory factor analysis (CFA) to binary or categorical outcomes.
2
LetI
ij
define the performance measure for childi and itemj on test and let’s assume
it is determined as follows:
I
ij
=
j
+
j
i
+
ij
(B.1)
where
i
is childi’s latent skill for test and this is assumed to be independent from
the error term
ij
. In other words, we assume that a unidimensional skill is sufficient in
explaining a child’s response behavior on items in each sub-test. We further assume that
a child’s response to an item is independent of his or her responses to other items after
conditioning on child latent skill.
2
See Skrondal and Rabe-Hesketh (2009) for detailed overview of IRT estimation methods and Zhao and
Hu (2008) for practical examples.
94
The variableI
ij
is not observed to the enumerator or caregiver. Instead, we observe
I
ij
=1 ifI
ij
> 0 andI
ij
= 0 otherwise. The model is identified by assuming
i
N(0;1). We
further assume that the measurement system is invariant to treatment assignment. Hence,
the probability of observingI
ij
= 1 given the child’s latent skill
i
is denoted as follows:
Pr(I
ij
= 1j
i
) = 1Pr(
ij
j
j
i
j
i
) = 1F
(
j
j
i
) (B.2)
whereF
is the distribution of
ij
. The estimated item specific intercepts,
j
, represent
the level of difficulty of itemj. The higher the value of
j
, the lower the success rate of
itemj is for a given latent skill level and hence the more difficult itemj is. The parameter
j
represents the discrimination ability of itemj as the rate at which the probability of
answering correctly changes with a child’s latent skill. Items with large discrimination
value have a high correlation between latent skill and the probability of success and can
distinguish better between low and high levels of latent skill. Hence, in the 2-parameter
logistic IRT model, the probability of success on an itemj is a function of both the level of
latent skill
i
and the difficulty level,
j
, and discrimination ability,
j
, of itemj. Below
we describe the individual tests in more detail and provide test diagnostics from the
two-parameter logistic IRT model.
(i) Verbal Comprehension
The verbal comprehension index measures a child’s verbal reasoning and comprehension
abilities and is assessed using the WPPSI-IV Similarities test and the Information test.
During the administration of the Similarities test the child is read incomplete sentences
containing two concepts that share a common characteristic. The child is asked to complete
the sentence by providing a response that reflects the shared characteristic. During the
Information test, the child is asked to respond to questions by choosing pictures from four
95
response options and answer questions addressing a broad range of general knowledge
topics.
Figure B1 plot the distribution of estimated item-difficulty parameters ,
ˆ
j
, for each of
the 2 administered tests from the two-parameter logistic IRT measurement system. Items
with negative estimated difficulty parameters are considered relatively easy, and items
with positive difficulty parameters are relatively hard. The distribution of the estimated
difficulty parameters provides information about whether the test is well-designed for the
population under study. In an ideal test, the difficulty parameters smoothly transition
from easy to more difficult and cover the whole skill distribution. By this metric, the
Similarities test is not well designed as the values of the difficulty level are relatively flat
after item 11. This means that the test is not able to distinguish well between children
with medium and high latent skill levels.
Figure B2 plots the Item Characteristic Curves (ICCs) for both tests. The ICC plots the
probability that a person is successful on a given itemj as a function of child’si latent
skill
i
for test. The ICCs for the tests of the WPPSI-IV Verbal Comprehension index
confirm that the Similarities test fails to differentiate between medium and higher levels
of latent skill. Moreover, the estimated discrimination parameters, ˆ
j
, for more easy items
are relatively low indicating that the test is also not very good in differentiating between
low and medium latent skill.
96
(a) Information Test (b) Similarities Test
Figure B1: Distribution of task item difficulty levels WPPSI-IV Verbal Comprehension
(a) Information Test (b) Similarities Test
Figure B2: Item Characteristic Curves (ICCs) WPPSI-IV Verbal Comprehension
(ii) Visual Spatial
The visual spatial index measures the ability to organise and understand visual parts and
information, assimilate visual and motor functions simultaneously, and see the whole-part
connection to objects. Visual spatial ability is assessed using the WPPSI-IV Block Design
and Object Assembly test. During the Block design test, a child is asked to use one- or
two-color blocks to recreate the design of a picture in a stimulus book within a specific
97
time limit. In the Object Assembly test, the child is presented with pieces of a puzzle
which needs to be fit together within the time span of 90 seconds. Figure B3- B4. show
that the Block Design test is good in distinguishing between low and high ability, but
less accurate in measuring medium ability. The Object Assembly test, on the other hand,
is only precise in measuring very low latent skill as almost all estimated item difficulty
parameters are negative.
(a) Block Design Test (b) Object Assembly Test
Figure B3: Distribution of task item difficulty levels WPPSI-IV Visual Spatial
(a) Block Design Test (b) Object Assembly Test
Figure B4: Item Characteristic Curves (ICCs) WPPSI-IV Visual Spatial
98
(iii) Fluid Reasoning
The fluid reasoning index measures a child’s ability to utilize inductive reasoning which is
the ability to use past observations to predict current situations. Fluid reasoning ability is
administered by the WPPSI-IV Matrix Reasoning and Picture Concept test. In the Matrix
Reasoning test, a child is presented with an incomplete matrix and asked to select missing
parts from 4 or 5 response options. During the Picture Concepts test, the child is shown
two or three rows of pictures and needs to choose one picture from each row to form a
group with a common characteristic. Figure B5-B6. show that both tests are relatively
better at measuring and distinguishing between higher latent skill levels. Figure B5-B6
show that both tests are relatively better at measuring and distinguishing between higher
latent skill levels.
(a) Matrix Reasoning Test (b) Picture Concepts Test
Figure B5: Distribution of task item difficulty levels WPPSI-IV Fluid Reasoning
99
(a) Matrix Reasoning Test (b) Picture Concepts Test
Figure B6: Item Characteristic Curves (ICCs) WPPSI-IV Fluid Reasoning
100
(iv) Working Memory
The working memory index measures the ability to balance focus and attention while
manipulating visual and auditory information in conscious awareness and is administered
by the WPPSI-IV Zoo Location and Picture Memory tests. In the Zoo Locations test, a
child is shown one or more animal card placeds on a zoo layout and than asked to place
the animal cards in the previously displayed locations. During the Picture Memory test, a
child is shown one or more pictures for a specific duration of time and than asked to select
the same picture from options on a response page.
Figures B7- B8 show that the Zoo Locations test is relatively good in distinguishing
between different latent skill levels but has a small amount of items which reduces overall
accuracy. The Picture Memory test is well designed to distinguish between low, medium
and high latent skills and has high discrimination ability across all items.
(a) Zoo Locations Test (b) Picture Memory Test
Figure B7: Distribution of task item difficulty levels WPPSI-IV Working Memory
101
(a) Zoo Locations Test (b) Picture Memory Test
Figure B8: Item Characteristic Curves (ICCs) WPPSI-IV Working Memory
(v) Processing Speed
The processing speed index analyses how quickly a child can scan and differentiate visual
information and is administered by the WPPSI-IV Bug Search and Cancellation tests. Both
tests measure the amount of time a child requires to finish the task. During the Bug Search
test, a child uses an ink dauber to mark the image of a bug that matches the target bug
in a collection of different bugs. For the Cancellation test, a child is asked to scan two
arrangements of objects on a page and mark all the target objects. Performance on both
test is not dependent on increasing task difficulty as is the case for the other WPPSI-IV
indexes and hence there is no need to estimate an IRT model.
B.2. Infant Non-Cognitive Skills
Infant non-cognitive skills are measured using the Strengths and Difficulty Questionnaire
(SDQ), a widely used behavioral screening tool translated to Chinese and validated on
a Chinese sample (Goodman et al., 2000; Du et al., 2008). The SDQ comprises of 25
items assessing social, emotional and behavioral functioning of children reported by the
main caregiver on a 3-point likert scale ( 1 not true, 2 somewhat true, 3 certainly true).
102
Items are both positively and negatively phrased to avoid the effect of acquiescence bias.
The original proposed factor structure of the SDQ includes five scales of five items each
corresponding with five sub-domains: (i) conduct problems; (ii) hyperactivity/inattention;
(iii) emotional symptoms; (iv) peer problems and (v) prosocial behavior (Goodman, 1997).
Alternatively sub-domains (i) and (ii) can be combined to measure externalizing behavior
which assesses behavioral problems that are manifested in children’s outward behavior
such as disruptiveness, hyperactivity, and aggressive behavior. The sub-domains (iii)-(iv)
can be combined to measure internalizing behavior which assesses behavioral problems
affecting children’s internal psychological environment such as withdrawn, anxious, and
depressed behavior. More recent studies using exploratory and confirmatory factor analysis
indicate that this three-factor structure might be more appropriate (Dickey and Blumberg,
2004; Goodman et al., 2010).
We follow the literature and estimate a three-factor dedicated measurement system in
which each item is associated with at most one factor (Gorsuch, 2003; Thompson, 2004).
Parameters of the dedicated measurement system are estimated using maximum likelihood
and can be found in Table B1. The first column in Table B1 reports factor loadings of each
of the 25 items for the three non-cognitive skill factors. We normalize factor loadings of
the first measure for each skill factor to one. In the second column of Table B1 we report
the signal-to-noise ratio which indicates how much of the variance in each of the 25 items
is driven by signal relative to noise. The signal-to-noise ratios for the j
th
item is calculated
as:
S
j
=
2
j
Var(
)
2
j
Var(
) +Var(
j
)
Several items of the SDQ have poor signal-to-noise ratios, confirming previous findings in
the literature that document measurement error in early childhood skills (Cunha et al.,
2010), especially for caregiver assessments as correlations between questions are partially
103
Table B1: Measurement System Non-Cognitive Skills
Latent skill factor Measurement item Factor loading % signal
Externalizing Behavior Often loses temper 1 0.317
Generally well behaved, usually does what adults request -0.220 0.014
Often fights with other children or bullies them 0.485 0.151
Often lies or cheats 0.461 0.145
Steals from home, school or elsewhere 0.121 0.060
Restless, overactive, cannot stay still for long 0.728 0.139
Constantly fidgeting or squirming 1.014 0.247
Easily distracted, concentration wanders 0.686 0.177
Think things out before acting -0.291 0.030
Good attention span, sees work through to the end -0.503 0.081
Internalizing Behavior Often complains of headaches, stomach-aches or sickness 1 0.174
Many worries or often seems worried 1.000 0.215
Often unhappy, depressed or tearful 1.031 0.209
Nervous or clingy in new situations, easily loses confidence 1.084 0.150
Many fears, easily scared 1.344 0.271
Rather solitary, prefers to play alone 0.747 0.090
Has at least one good friend -0.135 0.005
Generally liked by other children -0.244 0.013
Picked on or bullied by other children 0.947 0.181
Gets along better with adults than with other children 0.247 0.008
Pro-Social Behavior Considerate of other people’s feelings 1 0.124
Shares readily with other children, for example toys, treats, 0.998 0.145
Helpful if someone is hurt, upset or feeling ill 1.722 0.334
Kind to younger children 1.108 0.208
Often offers to help others (parents, teachers, or other children) 1.697 0.411
driven by answering patterns of respondents (Johnston et al., 2014; Laajaj and Macours,
2017).
104
B.3. Parental Investment
In a first step we use exploratory factor analysis (EFA) to determine the number of factors
we need to extract from our list of time and material parental investment measures. We
use Horn’s parallel analysis (Horn, 1965) and Cattell’s scree plot (Cattell, 1966) to guide
us in the factor selection process (Table B2).
Table B2: EFA Factor Selection Parental Investment
Cattell’s scree plot Horn’s parallel analysis
Time investment 1 1
Material investment 1 1
(a) Time investment Test (b) Material investment
Figure B9: Scree Plot of Eigenvalues of Parental Investment
Both factor selection methods indicate we should use a one-factor measurement model
for time and material investment. We next proceed with estimating factor loadings which
are reported in Table B3. We find that the first three investment measures load positively
105
on the latent time investment factor. The last investment measure, number of hours per
day toddler spends watching tv, loads negatively on the latent time investment factor. The
signs of the estimated factor loadings give us confidence that we are indeed measuring
positive parenting time investment practises. The cost measures of books, toys, clothes
and school all load positively on the latent material investment factor. All measures
load relatively strongly on the latent factor loadings hence we retain all measures and
estimate means and factor loadings using maximum likelihood. We next predict factor
scores for both parental investment dimensions and further standardize the factors by the
distribution of the control group.
Table B3: Estimated factor loadings time and material parental investment
Factor Loading
Time investment
Number of times family uses toys to play with toddler 0.577
Number of times per week family reads to toddler 0.549
Number of times per week family sings to toddler 0.564
Number of hours per day toddler spends watching TV -0.216
Material investment
Cost of children’s books last year 0.527
Cost of children’s toys last year 0.572
Cost of children’s clothes last year 0.572
Cost of children’s school last year 0.262
B.4. Preschool Quality
We collect data on preschool and teacher characteristics, process and structural quality.
It is a priori unclear which measures best predict (perceived) preschool quality in rural
China hence, in a first step, we use EFA to explore along which dimensions preschools can
106
be best classified. Horns parallel analysis (Horn, 1965) indicates there are 5 main latent
dimensions in the preschool data we collected. However, Cattell’s scree plot Cattell (1966)
shows that a large part of the variation in preschools can be summarized by one latent
factor. Estimated factor loadings for the 5-factor measurement model can be found in
Table B4.
Table B4: EFA Factor Selection Preschool Quality
Cattell’s scree plot Horn’s parallel analysis
Preschool quality 1 5
Figure B10: Scree Plot of Eigenvalues of Preschool Quality
The pattern of estimated factor loadings in Table B5 suggest that the first latent factor is
indeed a good measure of general preschool quality as it captures preschool and teacher
characteristics as well as structural and process quality. Higher factor scores are associated
with schools that are bigger in size, more likely to be located in township or counties as
compared to villages and have younger and more educated teachers that are more likely to
107
have received a teacher training in the past year. Higher factor scores are also associated
with larger indoor and outdoor space, the availability of dormitories and breakfast and
several measures of process quality such as organising exercise and science activities and
reading books in class.
Variation in the second latent factor is driven by preschools with smaller pupil-teacher
ratios and older teachers that receive higher salaries but score low on process quality
measures. Hence this second latent factor might capture more informal village nurseries
with small numbers of enrolled children. Variation in the subsequent latent factors is
driven by a small number of items, none of which present a clear pattern. We hence
proceed with a one-factor model and use estimated means and factor loadings to predict a
latent preschool quality score for each preschool in the sample. Factor scores are further
standardized by the distribution of the control group.
108
Table B5: Estimated Factor Loadings Preschool Quality
Factor 1 Factor 2 Factor 3 Factor 4 Factor 5
Preschool & Teacher Characteristics
Number of pupils 0.774 0.250 -0.103 -0.075 0.057
Share pupils receiving government need-based aid -0.252 -0.162 0.138 0.104 0.230
Tuition fee per semester (Yuan) 0.272 0.076 -0.262 0.042 -0.166
Township preschool 0.627 0.224 -0.026 -0.060 0.138
Teacher age -0.408 0.501 -0.136 0.300 0.012
Teacher male -0.346 0.543 0.188 -0.088 -0.191
Teacher experience -0.273 0.326 -0.277 0.452 0.051
Teacher monthly salary (Yuan) -0.184 0.650 0.414 -0.119 0.028
Share of teachers with bachelor degree 0.335 0.082 0.321 -0.206 0.220
Teacher training in past year 0.519 0.130 0.273 0.013 -0.083
Structural Quality
Pupil-teacher ratio 0.037 -0.314 -0.046 0.029 0.079
Number of activity rooms 0.715 0.281 0.058 -0.079 0.144
Outdoor play-area 0.347 0.149 -0.097 -0.043 0.061
Preschool has play room 0.245 0.118 -0.135 0.135 0.359
Preschool has exercise room 0.243 0.207 -0.434 0.185 0.239
Preschool has dormitories 0.525 0.067 -0.248 0.069 -0.287
Preschool provides breakfast 0.443 -0.064 -0.185 -0.050 -0.345
Process Quality
Teacher reads books in class 0.331 -0.233 0.025 0.126 0.097
Teacher organizes exercise activities 0.541 -0.132 0.205 -0.027 -0.053
Teacher organizes art& music activities 0.276 -0.316 0.212 0.328 0.057
Teacher organizes science activities 0.313 0.131 0.132 0.291 -0.241
Teacher teaches social skills 0.235 -0.096 0.354 0.556 -0.058
Teacher teaches language skills 0.019 0.032 0.300 0.291 -0.022
109
C Appendix C: Preschool Options and Enrollment Choice
In our study, 90.5% of households enrolled their children in preschools. Among those
enrolled children, 55.9% of them enrolled at local village preschools, while the rest 44.1%
enrolled at preschools in townships or county seats. Figure C1 shows the geographic
locations of enrolled preschools and program villages. Although there is no administra-
tive obstacle to enroll children in preschools (except for enrollment across the county
line), households still encounter a few challenges. First, there is no immediate available
preschool. In Table C1, we show that the average minimum distance to any preschool is
about 5 kilometers (km) or by 10 minutes of driving. Not to mention that the popular
modes of transportation in the rural villages usually take longer than driving. Second, the
high-quality preschools in townships or county seats are even less available due to the
distance. The average minimum distance from the village where households reside to a
preschool to township or county seat is about 12 km or 23 minutes by driving (Table C1,
Panel A, Row 3 and 4). Lastly, the non-negligible pecuniary cost of attending preschools for
rural households. For every half a year, the total cost of enrolling a child in the preschool
is 1210 Yuan (Table C1, Panel A, Row 7).
Compare households’ actual enrollment choices to the statistics about the landscapes
of rural preschools in the region, we found that households are not limited by geographic
challenges. Many households did not enroll their children in the first available preschool
—the average distance from the village where the household resides to the enrolled
preschools is 11 km (Table C1, Panel B, Row 1), further away than the average minimum
distance to any preschool. This fact further confirms that many households decided to
enroll their children at higher-quality preschools in the township or even county seat.
110
Figure C1: Geographic locations of enrolled preschools and program villages
111
Table C1: Summary Statistics of Preschool Options and Enrollment Choices
Panel A: Rural Preschool Landscapes
Minimum distance to any preschool (kms) 4.785
[7.095]
Minimum duration to a any preschool (mins) 10.018
[13.674]
Minimum distance to a preschool 11.873
in township or county seat (km) [9.490]
Minimum travel time to a preschool 22.616
in township or county seat (mins) [17.197]
Average Tuition (Yuan) 885.821
[850.391]
Average Miscellaneous Fees (Yuan) 324.945
[502.057]
Average Total Costs (Yuan) 1210.766
[1053.522]
Panel B: Enrollment Choices
Distance to enrolled preschools (km) 11.147
[15.934]
Travel time to enrolled preschools (mins) 19.271
[24.409]
Note: Standard deviations are in the brackets.
112
D Appendix D: Additional Figures and Tables for Chapter 3
Appendix D Figures
Figure D1: Photos of Parenting Center Playground in rural villages of China (Zhong et al.,
2020)
Clockwise from left: public village building used for parenting center in a sample village;
established parenting center playground in a sample village; caregiver and child reading
in a parenting center playground; children and caregivers playing in parenting center
playground.
113
0 20 40 60 80 100
Num of Infants
-6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6
Source: Author's survey
Breakdown by age of childeren prior to the intervention (age in months)
Number of Childern from the Younger Cohort by the Treatment Assignment
Treatment Group Control Group
Figure D2: Number of Children across Treatment Assignments by Age of Children Prior
to the Intervention
0 .2 .4 .6 .8
Percentage
-6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6
Source: Author's survey
Breakdown by age of infants prior to the intervention (age in months)
Percent of Childern from the Younger Cohort by the Treatment Assignment
Treatment Group Control Group
Figure D3: Percentage of Children across Treatment Assignments by Age of Children Prior
to the Intervention
114
Appendix D Tables
Table D1: The impact of having access to playground on parental investment
Intention-to-treat estimates
without covariates with covariates
(1) (2) control
Dependent variables Coefficient S.E. Coefficient S.E. mean
Parental Investment (N=1046)
Panel A: Whether Caregivers Engaged in Parenting Activities Yesterday
Played with toy yesterday (0/1) 0.0006 (0.0279) 0.0113 (0.0270) 0.5439
Story telling yesterday (0/1) 0.0036 (0.0176) -0.0027 (0.0182) 0.0835
Story telling with storybook yesterday (0/1) 0.0064 (0.0100) 0.0024 (0.0103) 0.0300
Sang song yesterday (0/1) 0.0426 (0.0326) 0.0446 (0.0319) 0.2655
Panel B: The Time Caregivers Engaged in Parenting Activities Yesterday
Time playing with the child (minutes) 23.7629*** (7.2324) 21.3548*** (7.3116) 53.0815
Story telling time (minutes) -5.7584 (4.3071) -6.8254 (4.3403) 30.0236
Story telling with storybook time (minutes) 0.4659 (0.4871) 0.4032 (0.5587) 1.9036
Singing time (minutes) 0.5754 (1.1198) 0.6155 (1.2273) 5.8116
Total time spent (min) 18.9544* (9.5996) 15.4788 (9.7718) 90.9013
1
Standard errors in the parenthesis are clustered at the village level;
2
Individual covariates controlled are the child’s age, the child’s gender, an indicator function of whether
the child was a preterm birth, number of household members, an indicator function of father of the child
has less than 9 years of education, an indicator function of mother of the child has less than 9 years of
education,an indicator function of father is self-reportedly healthy, an indicator function of mother is
self-reportedly healthy, father age, mother age, and four indicator functions of housing value categories;
3
p< 0:10,p< 0:05,p< 0:01. 115
Table D2: Suggestive Evidence on Whether Social Interactions Affect
Parental Investment
Panel A: Whether Caregivers Engaged in Parenting Activities Yesterday
(1) (2) (3) (4)
Play Story-telling Storybook Sing
yesterday
Treatment 0.0335 -0.0069 0.0050 0.0513
(0.0307) (0.0220) (0.0128) (0.0345)
Treatment -0.0055 -0.0023 0.0012 0.0042
Demeaned younger cohort size (0.0039) (0.0032) (0.0016) (0.0052)
Treatment -0.0042 0.0052* -0.0008 -0.0075
Demeaned older cohort size (0.0037) (0.0027) (0.0014) (0.0060)
N 1014 1012 1014 1014
Individual covariates X X X X
Mean in control group 0.5521 0.0865 0.0310 0.2683
Panel B: The Time Caregivers Engaged in Parenting Activities Yesterday
(1) (2) (3) (4)
Play Story-telling Storybook Sing
yesterday (minutes)
Treatment 18.1523** -4.4813 0.8613 0.1399
(8.3582) (4.3656) (0.7355) (1.2357)
Treatment 0.6470 -0.2267 -0.1161* 0.3039*
Demeaned younger cohort size (1.0390) (0.5529) (0.0634) (0.1585)
Treatment 0.4037 -0.7509 0.0030 -0.0525
Demeaned older cohort size (1.2389) (0.5701) (0.0691) (0.1397)
N 1011 1013 1014 1014
Individual covariates X X X X
Mean in control group 54.4400 30.4213 1.9712 5.9734
1
Standard errors in the parenthesis are clustered at the village level;
2
Individual covariates controlled are the child’s age, the child’s gender, an indicator
function of whether the child was a preterm birth, number of household members,
an indicator function of father of the child has less than 9 years of education, an
indicator function of mother of the child has less than 9 years of education,an indi-
cator function of father is self-reportedly healthy, an indicator function of mother
is self-reportedly healthy, father age, mother age, and four indicator functions of
housing value categories;
3
Dependent variables in Panle A are dummy variables;
4
p< 0:10,p< 0:05,p< 0:01. 116
Table D3: Robustness Check of Social Learning from Other Similar-age
Children Households
(1) (2) (3) (4)
Cognitive Receptive Expressive Overall
Skills Language skills
Treatment -0.4685 0.8065 -0.1549 0.2543
(0.7718) (0.8104) (0.7211) (0.7545)
Treatment 0.0042 0.0397*** 0.0262* 0.0366**
Demeaned younger cohort size (0.0117) (0.0147) (0.0145) (0.0152)
Treatment 0.0004 -0.0222 -0.0111 -0.0183
Demeaned older cohort size (0.0104) (0.0178) (0.0155) (0.0179)
N 1013 1013 1013 1013
Individual covariates X X X X
Treatment* village covariates X X X X
Village covariates X X X X
Mean in control group -0.0583 -0.1521 -0.1118 -0.1467
1
Standard errors in the parenthesis are clustered at the village level;
2
Individual covariates controlled are the child’s age, the child’s gender, an indicator
function of whether the child was a preterm birth, number of household members,
an indicator function of father of the child has less than 9 years of education, an
indicator function of mother of the child has less than 9 years of education,an
indicator function of father is self-reportedly healthy, an indicator function of
mother is self-reportedly healthy, father age, mother age, and four indicator
functions of housing value categories;
3
Village covariates are percentage of mothers of children in the village who have
less than 9 years of education, percentage of fathers of children in the village who
have less than 9 years of education, percentage of children in the village who
are male, and percentage of children in the village who are preterm births. The
village mean is calculated with characteristics from both the younger cohort and
the older cohort.
4
All scores of development skills have been internally standardized nonparametri-
cally for age and are expressed in standard deviation units;
5
p< 0:10,p< 0:05,p< 0:01.
117
Table D4: Suggestive Evidence of Social Learning from Other Knowledgeable Caregivers
(1) (2) (3) (4) (5) (6)
Felt Nerves Know how to Know the importance of Parental
around play with read to playing with reading to awareness
the child the child the child the child the child index
Treatment -0.1250 0.2499 -0.2745 0.1964 -0.0728 0.0970
(0.0807) (0.2326) (0.2298) (0.4230) (0.2992) (0.6037)
Treatment 0.0007 0.0061 -0.0037 -0.0084 -0.0060 -0.0074
Demeaned younger cohort size (0.0013) (0.0047) (0.0048) (0.0073) (0.0064) (0.0122)
Treatment -0.0030** 0.0056 0.0093** 0.0098 0.0121** 0.0283***
Demeaned older cohort size (0.0015) (0.0039) (0.0039) (0.0063) (0.0053) (0.0107)
N 1014 1014 1014 1014 1014 1014
Individual covariates X X X X X X
Treatment* village covariates X X X X X X
Village covariates X X X X X X
Mean in control 0.027 0.2905 0.2306 0.5499 0.3392 -0.0086
1
Standard errors in the parenthesis are clustered at the village level;
2
Individual covariates controlled are the child’s age, the child’s gender, an indicator function of whether the
child was a preterm birth, number of household members, an indicator function of father of the child has less
than 9 years of education, an indicator function of mother of the child has less than 9 years of education,an
indicator function of father is self-reportedly healthy, an indicator function of mother is self-reportedly healthy,
father age, mother age, and four indicator functions of housing value categories;
3
Village covariates are percentage of mothers of children in the village have less than 9 years of education,
percentage of fathers of children in the village have less than 9 years of education, percentage of children in the
village is male, and percentage of children in the village is preterm birth. The village mean is calculated with
characteristics from both the younger cohort and the older cohort.
4
Dependent variables in column (1) to (5) are dummy variables; Dependent variable in (6), parenting awareness
index, is expressed in standard deviation units standardized by the mean and the standard deviation of the
control group.
5
p< 0:10,p< 0:05,p< 0:01.
118
Table D5: Suggestive Evidence of Social Learning from Other Similar-
age Children Households (-6 to 0 cohort)
(1) (2) (3) (4)
Cognitive Receptive Expressive Overall
Skills Language skills
Treatment -0.0551 0.1135 0.0637 0.0963
(0.1089) (0.1238) (0.0856) (0.0981)
Treatment -0.0097 0.0416 0.0322 0.0431**
Demeaned (-6-0) cohort size (0.0255) (0.0271) (0.0203) (0.0200)
Treatment 0.0063 -0.0003 0.0060 0.0037
Demeaned (1-6) cohort size (0.0155) (0.0167) (0.0147) (0.0154)
Treatment 0.0098 0.0029 0.0071 0.0059
Demeaned older cohort size (0.0156) (0.0166) (0.0146) (0.0154)
N 509 509 509 509
Covariates controls X X X X
Mean in control -0.1217 -0.2458 -0.1610 -0.2249
1
Standard errors in the parenthesis are clustered at the village level;
2
Individual covariates controlled are the child’s age, the child’s gender, an indi-
cator function of whether the child was a preterm birth, number of household
members, an indicator function of father of the child has less than 9 years of
education, an indicator function of mother of the child has less than 9 years of
education,an indicator function of father is self-reportedly healthy, an indicator
function of mother is self-reportedly healthy, father age, mother age, and four
indicator functions of housing value categories;
3
All scores of development skills have been internally standardized nonpara-
metrically for age and are expressed in standard deviation units;
4
p< 0:10,p< 0:05,p< 0:01.
119
Table D6: Suggestive Evidence of Social Learning from Other Knowledgeable Caregivers (-6 to
0 cohort)
(1) (2) (3) (4) (5) (6)
Felt Nerves Know how to Know the importance of Parental
around play with read to playing with reading to awareness
the child the child the child the child the child index
Treatment -0.0062 0.0512 0.0121 0.0434 0.0474 0.1240
(0.0140) (0.0485) (0.0380) (0.0428) (0.0396) (0.0951)
Treatment 0.0002 -0.0397*** 0.0003 -0.0170 -0.0076 -0.0544*
Demeaned (-6-0) cohort size (0.0062) (0.0100) (0.0129) (0.0160) (0.0138) (0.0322)
Treatment 0.0062 0.0223* 0.0083 0.0011 -0.0145 0.0215
Demeaned (1-6) cohort size (0.0021) (0.0061) (0.0050) (0.0060) (0.0062) (0.0133)
Treatment -0.0019 0.0104* 0.0039 0.0147** 0.0023 0.0256*
Demeaned older cohort size (0.0021) (0.0061) (0.0050) (0.0060) (0.0062) (0.0133)
N 509 509 509 509 509 509
Invididual covariates X X X X X X
Mean in control 0.0402 0.5783 0.3293 0.3855 0.2450 0.0941
1
Standard errors in the parenthesis are clustered at the village level;
2
Individual covariates controlled are the child’s age, the child’s gender, an indicator function of whether the
child was a preterm birth, number of household members, an indicator function of father of the child has less
than 9 years of education, an indicator function of mother of the child has less than 9 years of education,an
indicator function of father is self-reportedly healthy, an indicator function of mother is self-reportedly
healthy, father age, mother age, and four indicator functions of housing value categories;
3
Dependent variables in column (1) to (5) are dummy variables; Dependent variable in (6), parenting awareness
index, is expressed in standard deviation units standardized by the mean and the standard deviation of the
control group.
4
p< 0:10,p< 0:05,p< 0:01.
120
E Appendix E: Non-identification of ATT without exclusion re-
striction
First note that the observed outcome is
Y
ig
= Y
ig
(Z
g
;D
ig
(Z
g
))
= Z
g
D
ig
Y
ig
(1;1) +Z
g
(1D
ig
)Y
ig
(1;0) + (1Z
g
)Y
ig
(0;0)
= Z
g
(Y
ig
(1;0)Y
ig
(0;0)) +Z
g
D
ig
(Y
ig
(1;1)Y
ig
(1;0)) +Y
ig
(0;0)
so that
E[Y
ig
jZ
g
= 1] = E[Y
ig
(1;0)Y
ig
(0;0)] + E[D
ig
(Y
ig
(1;1)Y
ig
(1;0))] + E[Y
ig
(0;0)]
= E[Y
ig
(1;0)] + E[Y
ig
(1;1)Y
ig
(1;0)jD
ig
(1) = 1]Pr(D
ig
(1) = 1)
E[Y
ig
jZ
g
= 0] = E[Y
ig
(0;0)]
so that
E[Y
ig
jZ
g
= 1]E[Y
ig
jZ
g
= 0] = E[Y
ig
(1;0)Y
ig
(0;0)]+E[Y
ig
(1;1)Y
ig
(1;0)jD
ig
(1) = 1]Pr(D
ig
(1) = 1)
while
E[D
ig
jz
g
= 1] E[D
ig
jz
g
= 0] = Pr(D
ig
(1) = 1)
so that
E[Y
ig
jZ
g
= 1] E[Y
ig
jZ
g
= 0]
E[D
ig
jz
g
= 1] E[D
ig
jz
g
= 0]
=
E[Y
ig
(1;0)Y
ig
(0;0)]
Pr(D
ig
(1) = 1)
+ E[Y
ig
(1;1)Y
ig
(1;0)jD
ig
(1) = 1]
121
F Appendix F: Comparison to Usual DID Estimator
Traditional Difference-in-Difference estimator
ATET = E[Y
t+1
ig
Y
t
ig
jD
ig
= 1] E[Y
t+1
ig
Y
t
ig
jD
ig
= 0]
= E[Y
t+1
ig
Y
t
ig
jD
ig
= 1;Z
g
= 1]
E[Y
t+1
ig
Y
t
ig
jD
ig
= 0;Z
g
= 0]Pr(Z
g
= 0jD
ig
= 0)
| {z }
p
+E[Y
t+1
ig
Y
t
ig
jD
ig
= 0;Z
g
= 1]Pr(Z
g
= 1jD
ig
= 0)
| {z }
1p
= E[Y
t+1
ig
(1;1)Y
t
ig
(0;0)jD
ig
= 1;z
g
= 1]
E[Y
t+1
ig
(0;0)Y
t
ig
(0;0)jD
ig
= 0;z
g
= 0]p
+E[Y
t+1
ig
(1;0)Y
t
ig
(0;0)jD
ig
= 0;Z
g
= 1](1p)
= E[Y
t+1
ig
(1;1)Y
t
ig
(0;0)jD
ig
= 1]
E[Y
t+1
ig
(0;0)Y
t
ig
(0;0)jD
ig
= 0]p
+E[Y
t+1
ig
(1;0)Y
t
ig
(0;0)jD
ig
= 0](1p)
= E[Y
t+1
ig
(1;1)Y
t
ig
(0;0)jD
ig
= 1]
+pE[Y
t+1
ig
(1;0)Y
t+1
ig
(0;0)jD
ig
= 0] E[Y
t+1
ig
(1;0)Y
t
ig
(0;0)jD
ig
= 0]
= E[Y
t+1
ig
(1;1)Y
t+1
ig
(0;0)jD
ig
= 1]
+pE[Y
t+1
ig
(1;0)Y
t+1
ig
(0;0)jD
ig
= 0] E[Y
t+1
ig
(1;0)Y
t+1
ig
(0;0)jD
ig
= 0]
= E[Y
t+1
ig
(1;1)Y
t+1
ig
(0;0)jD
ig
= 1] (1p)E[Y
t+1
ig
(1;0)Y
t+1
ig
(0;0)jD
ig
= 0]
the first equality is derived fromfD = 1g =fD = 1;Z = 1g, the second equality uses the
independence of the randomization assignment, the fifth equality from the equal-trend
assumption: E[Y
t+1
(0;0)Y
t
(0;0)jD = 1] = E[Y
t+1
(0;0)Y
t
(0;0)jD = 0]. Therefore, DID
122
estimator identifiesITT (co) (1p)ITT (nt), and not their effects separately. Note that the
last term disappears only when eitherp = 1 or exclusion restriction holds. Here,
Pr(Z
g
= 0jD
ig
= 0) =
Pr(D
ig
= 0jZ
g
= 0)Pr(Z
g
= 0)
Pr(D
ig
= 0)
=
Pr(D
ig
= 0jZ
g
= 0)Pr(Z
g
= 0)
Pr(D
ig
= 0jZ
g
= 0)Pr(Z
g
= 0) + Pr(D
ig
= 0jZ
g
= 1)Pr(Z
g
= 1)
=
Pr(D
ig
= 0jZ
g
= 0)Pr(Z
g
= 0)
Pr(D
ig
= 0jZ
g
= 0)Pr(Z
g
= 0) + Pr(D
ig
= 0jZ
g
= 1)Pr(Z
g
= 1)
=
Pr(Z
g
= 0)
Pr(Z
g
= 0) +
nt
Pr(Z
g
= 1)
so thatp = 1 happens only when
nt
Pr(Z
g
= 1) = 1 when
nt
Pr(Z
g
= 1) = 1, p=0 because
Pr(z
g
= 0) = 0, p=1 only whenPr(z
g
= 0) = 1.
If the researcher ignores the spillover effect when there is such effect, the above equation
shows that DID estimator does not identify ATET
3
.
3
Regarding this point, see Abbring and Heckman (2007), p.5277: ”The analysis of Heckman et al.
(1998a), Heckman et al. (1998b), Heckman et al. (1998c) has important implications for the widely-used
difference-in-differences estimator. If the tuition subsidy changes the aggregate skill prices, the decisions
of nonparticipants will be affected. The ”no treatment” benchmark group is affected by the policy and the
difference-in-differences estimator does not identify the effect of the policy for anyone compared to a no
treatment state.”
123
G Appendix G: Additional Tables for Chapter 4
Table G1: Summary Statistics of the Non-attrited Sample
Control Group Treatment-Control
Obs Obs Mean SD coefficient p-value
Number members 4105 2101 5.168 2.68 0.061 0.453
Number adults (> 16) 4105 2101 3.467 1.99 0.041 0.472
Household age 4105 2101 47.892 15.912 1.115 0.011
Animal Husbandry 4105 2101 0.537 0.499 0.041 0.031
Run a non-farm business 4105 2101 0.218 0.413 -0.036 0.011
Has an outstanding loan over the past 12 months 4105 2101 0.261 0.439 0.057 0.006
HH spouse responded to the survey 4105 2101 0.061 0.24 0.02 0.007
other HH member responded to the survey 4105 2101 0.04 0.195 0.006 0.243
Missing value in HH spouse responded to the survey 4105 2101 0.155 0.362 0.001 0.91
Missing value in other HH member responded to the survey 4105 2101 0.155 0.362 0.001 0.91
124
Table G2: Summary Statistics of the Non-attrited Sample with High Probability to Borrow
Control Group Treatment-Control
Obs Obs Mean SD coefficient p-value
Number members 3525 1793 5.202 2.675 0.11 0.207
Number adults (> 16) 3525 1793 3.473 1.998 0.078 0.203
Household age 3525 1793 47.423 15.884 1.75 0
Animal Husbandry 3525 1793 0.52 0.5 0.053 0.007
Run a non-farm business 3525 1793 0.228 0.42 -0.039 0.012
Has an outstanding loan over the past 12 months 3525 1793 0.255 0.436 0.067 0.002
HH spouse responded to the survey 3525 1793 0.063 0.243 0.021 0.009
other HH member responded to the survey 3525 1793 0.041 0.198 0.006 0.205
Missing value in HH spouse responded to the survey 3525 1793 0.149 0.356 0.004 0.679
Missing value in other HH member responded to the survey 3525 1793 0.149 0.356 0.004 0.679
125
Table G3: ITT(co) & ITT(nt) estimates with baseline covariates but no fixed effect
Panel A: Whole Non-attrited Sample
Asset Sales+home Has a self- Income from day Weekly hours worked Monthly HH
(stock) consumption Expense Profit employment activity labor/ salaried Self-employment Outside Consumption
ITT, complier 6197.366
33232.931
21591.955
10768.910
0.017 -1813.735 7.914 -1.573 325.791
(2653.878) (10424.903) (9612.722) (5546.228) (0.044) (1651.587) (6.548) (4.069) (330.997)
ITT, never taker 560.941 3865.499 2574.157 1395.822 0.008 -544.950 1.274 -0.979 -115.155
(1057.104) (3249.111) (2119.834) (2332.300) (0.032) (1017.368) (4.150) (2.077) (83.182)
Covariates X X X X X X X X X
Observations 4049 4049 4049 4049 4049 4049 4037 4037 4039
control mean 1798.850 4721.573 4266.519 1152.695 0.030 3692.911 21.208 0.313 614.806
Panel B: Non-attrited Sample with High Probability to Borrow
ITT, complier 6217.462
35708.779
22073.686
13009.093
0.017 -2208.563 6.425 -2.299 362.481
(2790.298) (11481.811) (10265.416) (6524.136) (0.044) (1690.263) (6.626) (4.383) (356.603)
ITT, never taker 939.905 5775.956 4363.076
1835.398 0.008 -736.753 0.039 -1.270 -103.082
(1099.590) (3564.183) (2405.483) (2419.284) (0.034) (1050.461) (4.103) (2.095) (88.776)
Covariates X X X X X X X X X
Observations 3479 3479 3479 3479 3479 3479 3467 3467 3470
control mean 1811.076 4421.010 4150.940 772.065 0.033 4057.138 22.350 0.381 604.385
p< 0:10,p< 0:05,p< 0:01. Covariates are changes of number of adults in the households and number of children in the households.
126
Table G4: ITT(co) & ITT(nt) estimates with pair fixed effect but no baseline covariates
Panel A: Whole Non-attrited Sample
Asset Sales+home Has a self- Income from day Weekly hours worked Monthly HH
(stock) consumption Expense Profit employment activity labor/ salaried Self-employment Outside Consumption
ITT, complier 4853.085
28454.159
18445.625
9659.636
-0.009 -1161.752 6.027
-2.026 352.029
(2053.512) (8313.713) (7207.594) (4863.209) (0.031) (1506.700) (3.012) (3.211) (295.871)
ITT, never taker 643.233 4805.897
3273.928
1629.004 0.013 -539.652 2.458
-0.724 -103.047
(691.388) (2005.170) (1571.865) (1493.556) (0.014) (622.442) (1.436) (1.226) (48.378)
Pair fixed effect X X X X X X X X X
Observations 4105 4105 4105 4105 4105 4105 4093 4093 4095
control mean 1879.392 4958.848 4451.599 1190.183 0.034 3756.474 21.127 0.607 617.163
Panel B: Non-attrited Sample with High Probability to Borrow
ITT, complier 4881.309
30061.156
18734.238
11133.811
-0.012 -1573.765 4.886 -2.683 395.330
(2240.657) (8999.712) (7437.518) (5782.038) (0.033) (1586.692) (3.121) (3.447) (320.694)
ITT, never taker 989.909 6810.969
5268.462
1958.207 0.013 -667.858 1.731 -0.826 -91.098
(729.018) (2333.616) (1824.041) (1624.414) (0.015) (697.904) (1.434) (1.271) (55.295)
Pair fixed effect X X X X X X X X X
Observations 3525 3525 3525 3525 3525 3525 3513 3513 3516
control mean 1924.312 4704.979 4247.166 951.485 0.038 3982.046 22.325 0.526 608.610
p< 0:10,p< 0:05,p< 0:01.
127
Abstract (if available)
Linked assets
University of Southern California Dissertations and Theses
Conceptually similar
PDF
Essays on development and health economics: social media and education policy
PDF
Essays on econometrics
PDF
Essays on treatment effect and policy learning
PDF
Essays on causal inference
PDF
Three essays on econometrics
PDF
Essays on health and aging with focus on the spillover of human capital
PDF
The EITC, labor supply, and child development
PDF
Essays on health insurance programs and policies
PDF
Essays on health economics
PDF
Essays on price determinants in the Los Angeles housing market
PDF
Essays in health economics: evidence from Medicare
PDF
Three essays on human capital and family economics
PDF
Essays on development and health economics
PDF
Essays on economics of education
PDF
Three essays on the evaluation of long-term care insurance policies
PDF
Essays on urban and real estate economics
PDF
Two essays in economic policy: The influence of perceived comparative need on financial subsidy requests; &, Unintended consequences of the FOSTA-SESTA legislation
PDF
Three essays on economics of early life health in developing countries
PDF
Essays on beliefs, networks and spatial modeling
PDF
Childhood adversity across generations and its impact on externalizing behavior
Asset Metadata
Creator
Qian, Yiwei
(author)
Core Title
Essays on development economics
School
College of Letters, Arts and Sciences
Degree
Doctor of Philosophy
Degree Program
Economics
Degree Conferral Date
2021-08
Publication Date
06/28/2021
Defense Date
04/23/2021
Publisher
University of Southern California
(original),
University of Southern California. Libraries
(digital)
Tag
development economics,early childhood development,OAI-PMH Harvest,randomized controlled trials,social interactions,spillover
Language
English
Contributor
Electronically uploaded by the author
(provenance)
Advisor
Strauss, John (
committee chair
), Ridder, Geert (
committee member
), Rozelle, Scott (
committee member
), Wang, Yongxiang (
committee member
)
Creator Email
yiweiqia@usc.edu
Permanent Link (DOI)
http://doi.org/10.25549/usctheses-c89-470715
Unique identifier
UC13012630
Identifier
etd-QianYiwei-9661.pdf (filename), usctheses-c89-470715 (legacy record id)
Legacy Identifier
etd-QianYiwei-9661
Dmrecord
470715
Document Type
Dissertation
Rights
Qian, Yiwei
Internet Media Type
application/pdf
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
Tags
development economics
early childhood development
randomized controlled trials
social interactions
spillover