Three essays on development economics

PDF

Download

Open document

Flip pages

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content THREE ESSAYS ON DEVELOPMENT ECONOMICS
by
Tao Chen
A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(ECONOMICS)
May 2025
Copyright 2025 Tao Chen

Acknowledgements
I sincerely thank my supervisor, Professor Paulina Oliva, for her invaluable guidance
and mentorship.
I would like to express my gratitude to my friends and cohort members, Zhan Gao and
Weizhao Huang, for their support and companionship.
I want to thank Kang Zhou and Li Fan for inspiring me with the idea that became the
foundation of my first chapter project.
I am also deeply grateful to my wife, Xinzhu, and to my parents for their unwavering
support throughout my life.
ii

Table of Contents
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xii
Chapter 1: The Ripple Effect of a Free Lunch: How School Nutrition Programs
Transform Children’s Wellbeing and Parental Labor Supply in Rural Areas . . . 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Nutrition Improvement Program (NIP) . . . . . . . . . . . . . . . . . . . 7
1.3 Theoretical Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.4 Data and Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . 14
1.4.1 Sample Selection . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.4.2 Main Outcome Variables and Summary Statistics . . . . . . . . . 15
1.5 Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.5.1 Direct Impact of the Program on Children and Adults . . . . . . . 20
1.5.2 Dynamic Effect . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.5.3 Spillover Effect on Parents Labour Supply . . . . . . . . . . . . . 22
iii

1.5.4 Heterogeneous Treatment Effect . . . . . . . . . . . . . . . . . . 23
1.5.5 Identification Assumption . . . . . . . . . . . . . . . . . . . . . 24
1.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.6.1 Direct Impact on Children . . . . . . . . . . . . . . . . . . . . . 26
1.6.2 Direct Impact on Adults . . . . . . . . . . . . . . . . . . . . . . 29
1.6.3 Estimation with CSDID Method . . . . . . . . . . . . . . . . . . 30
1.6.4 Spillover Effects of the NIP on Non-Eligible Parents . . . . . . . 32
1.6.5 Heterogeneous Treatment Effect . . . . . . . . . . . . . . . . . . 35
Heterogenous Impact on Children by Primary Caregiver and Income 35
Heterogeneous Impact on Adults by Primary Caregivers and Income Level . . . . . . . . . . . . . . . . . . . . . . . 36
In Which Occupational Sector Do Parents Experience the Greatest
Impact From the NIP? . . . . . . . . . . . . . . . . . . 42
At What Age Does the Treatment Have a Stronger Impact On
Parental Labor Outcomes When Children First Receive
It? . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
1.6.6 Robustness Check . . . . . . . . . . . . . . . . . . . . . . . . . 46
1.7 Discussion and Policy Recommendation . . . . . . . . . . . . . . . . . . 47
1.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
Chapter 2: The Consequence of China’s Supply-Side Structural Reform: Impact
of Coal Mine Shutdowns on Local Economic Development . . . . . . . . . . . 50
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
2.2 Theoretical Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
iv

2.3 Data and Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . 57
2.4 Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.4.1 Oaxaca-Blinder Regression (O-B) . . . . . . . . . . . . . . . . . 63
Placebo Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
2.4.2 OLS and Instrumental Variable . . . . . . . . . . . . . . . . . . . 69
2.4.3 Difference-in-difference (DD) . . . . . . . . . . . . . . . . . . . 76
Baseline Difference in Difference . . . . . . . . . . . . . . . . . 76
Event Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
How Does Variation in Transportation Infrastructure Affect the
Impact? . . . . . . . . . . . . . . . . . . . . . . . . . 81
Impact on the Listed Companies . . . . . . . . . . . . . . . . . . 82
2.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
2.5.1 Impact Of Coal Mine Closure On Counties, O-B Regression Results 84
2.5.2 Impact of a Coal Mine Closure on Cities, IV Regression Results . 86
2.5.3 Impact of a Coal Mine Closure on Cities, DD results . . . . . . . 87
Baseline Difference in Difference . . . . . . . . . . . . . . . . . 91
Exploiting the Magnitude of the Closure . . . . . . . . . . . . . . 95
Event Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
2.5.4 Discussion on the Results from the Three Empirical Strategies . . 98
2.5.5 The Mitigating Effect of Transportation Infrastructure on Negative
Shocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
2.5.6 Differential Performance of Private firms vs. SOEs . . . . . . . . 103
2.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
v

Chapter 3: White Elephant or Win-Win Projects: An Investigation into the Economic Impact of the Belt and Road Initiative . . . . . . . . . . . . . . . . . . . 106
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
3.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
3.3 Data and Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . 113
3.4 Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
3.4.1 Empirical Strategy I: Staggered Implementation . . . . . . . . . . 122
Effect of Project Implementation Initiation . . . . . . . . . . . . 122
Effect of Project Completion . . . . . . . . . . . . . . . . . . . . 123
Effect of Multiple Projects . . . . . . . . . . . . . . . . . . . . . 124
Effect of Electricity- and Metal & Mineral-Related Projects . . . 125
Identification Assumption of Difference-in-Difference . . . . . . 126
3.4.2 Empirical Strategy II: OLS and IV Regressions . . . . . . . . . . 127
OLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
IV Regressions . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
3.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
3.5.1 Difference in Difference with Staggered Implementation . . . . . 131
Effect of Project Implementation Initiation . . . . . . . . . . . . 131
Effect of Project Completion . . . . . . . . . . . . . . . . . . . . 132
Effect of Multiple Projects in a Province . . . . . . . . . . . . . . 135
Effect of Electricity- and Metal & Mineral-Related Projects . . . 138
3.5.2 OLS and IV Regressions . . . . . . . . . . . . . . . . . . . . . . 141
3.6 Robustness Check . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
vi

3.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
G.1 Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
G.2 Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
G.3 Modeling the NIP as an Improvement in AG . . . . . . . . . . . . 164
vii

List of Tables
1.1 Summary Statistics of Child and Adults Characteristics . . . . . . . . . . 18
1.2 Impact of NIP on Child Health, Cognitive Development, and Parental Engagement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
1.3 Impact of NIP on Adult Labor Outcomes . . . . . . . . . . . . . . . . . . 30
1.4 CSDID Aggregate Summary Results . . . . . . . . . . . . . . . . . . . . 33
1.5 Spillover Impact of NIP on the Non-Eligible Adults . . . . . . . . . . . . 34
1.6 Heterogeneous Impact of NIP on Children by Primary Caregiver . . . . . 37
1.7 Heterogeneous Impact of NIP on Children by Income Level . . . . . . . 38
1.8 Heterogeneous Impact of NIP on Adults by Primary Caregiver . . . . . . 40
1.9 Heterogenous Impact of NIP on Adults by Income Level . . . . . . . . . 41
1.10 Heterogenous Impact of NIP on Adults by Service Sector Jobs . . . . . . 43
2.1 Summary Statistics of County Level Data . . . . . . . . . . . . . . . . . 67
2.2 Oaxaca-Blinder Regression Placebo Test . . . . . . . . . . . . . . . . . . 70
2.3 Summary Statistics of City-Level Data . . . . . . . . . . . . . . . . . . . 75
2.4 Summary Statistics of Treated Cities and Neighboring Cities in 2014 . . . 79
2.5 O-B Regression: Impact Of Coal Mine Closure On Counties . . . . . . . 85
2.6 OLS: Effect of Cumulative Capacity Closure on the Local Economy . . . 88
viii

2.7 Reduced Form and 2SLS: Causal Effect of Cumulative Capacity Closure
on the Local Employment . . . . . . . . . . . . . . . . . . . . . . . . . . 89
2.8 Reduced Form and 2SLS: Causal Effect of Cumulative Capacity Closure
on the Local GDP and Prices . . . . . . . . . . . . . . . . . . . . . . . . 90
2.9 ATE by DD: Impact of Coal Mine Shutdown on Local Employment . . . 93
2.10 ATE by DD: Impact of Coal Mine Shutdown on Local GDP and Prices . . 94
2.11 ATE by DD: Impact of Coal Mine Shutdown on Local Economy - Exploiting Closure Magnitude . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
2.12 Empirical Results Summary . . . . . . . . . . . . . . . . . . . . . . . . 99
2.13 Heterogeneous ATE by DD: Impact of Coal Mine Shutdown by Variation
in Transportation Infrastructure . . . . . . . . . . . . . . . . . . . . . . . 101
2.14 Impact of Coal Mine Shutdown on Firms . . . . . . . . . . . . . . . . . 104
3.1 Number of BRI Projects in Each Province . . . . . . . . . . . . . . . . . 116
3.2 Summary Statistics: Mean of Each Variable by Province in 2014 . . . . . 121
3.3 ATE: the Effect of BRI Projects on GDP and GDP Components . . . . . 133
3.4 ATE: the Effect of BRI Projects on Other Social and Economic Outcomes 134
3.5 ATE of Multiple Projects: the Effect of BRI Projects on GDP and GDP
Components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
3.6 ATE of Multiple Projects: the Effect of BRI Projects on Other Social and
Economic Outcomes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
3.7 ATE across Industries: the Effect of BRI Projects on GDP and GDP Components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
ix

3.8 ATE across Industries: the Effect of BRI Projects on Other Social and
Economic Outcomes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
3.9 OLS: the Association between the Number of BRI Projects and Social and
Economic Outcomes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
3.10 2SLS: The causal Effect of BRI Projects on the Growth of GDP and other
Social and Economic Outcomes . . . . . . . . . . . . . . . . . . . . . . 143
3.11 Robustness Check: 2SLS Regressions with Different Lags of Chinese
Steel Production . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
A1 BJS Estimation Results: Impact of NIP on Children . . . . . . . . . . . . 161
A2 BJS Estimation Results: Impact of NIP on Adults . . . . . . . . . . . . . 162
A3 Heterogeneous Impact of NIP on Adults by Number of Children . . . . . 163
x

List of Figures
1.1 Geographic Distribution of NIP Rollout Counties in 2016 . . . . . . . . . 8
1.2 Event Study on Children Health Outcomes . . . . . . . . . . . . . . . . 29
1.3 Event Study on Adult Labor Market Outcomes . . . . . . . . . . . . . . 31
1.4 Event Study on Adult Outcomes by Child’s Age at the NIP Onset . . . . 45
2.1 Coal Mine Closure Trend in China (Unit: 10,000 ton) . . . . . . . . . . . 59
2.2 Provincial Ranking of Closed Coal Mine Capacity . . . . . . . . . . . . . 60
2.3 Distribution of Coal Mine Closures Across Counties . . . . . . . . . . . 61
2.4 Geographic Distribution of Treated Counties . . . . . . . . . . . . . . . . 65
2.5 Distribution of Oaxaca Weights Across Untreated Counties . . . . . . . . 68
2.6 Distribution of Coal Capacity Closure Across Cities . . . . . . . . . . . . 73
2.7 Distribution of Mining Tax Payments Across Cities in 1912 . . . . . . . . 74
2.8 Distribution of Treated and Neighboring Cities . . . . . . . . . . . . . . 77
2.9 Parallel Trends in Outcomes: Treated Cities vs. Neighboring Cities . . . . 81
2.10 Event Study of Coal Mine Closure Impact . . . . . . . . . . . . . . . . . 97
3.1 Number of Projects per Quarter . . . . . . . . . . . . . . . . . . . . . . . 115
3.2 Number of Projects by Industry . . . . . . . . . . . . . . . . . . . . . . . 117
3.3 Distribution of Projects by Industry Across Indonesia . . . . . . . . . . . 118
3.4 Indonesian GDP Distribution v.s. Projects Distribution . . . . . . . . . . 119
xi

A1 CSDID Event Study on Children Health Outcomes . . . . . . . . . . . . 157
A2 CSDID Event Study on Adults Labor Market Outcomes . . . . . . . . . 158
A3 CSDID Event Study on Adult Outcomes by Child’s Age at the NIP Onset 159
A4 Parallel Trend Plots for Instrumental Variable . . . . . . . . . . . . . . . 160
xii

Chapter 1
The Ripple Effect of a Free Lunch: How School Nutrition
Programs Transform Children’s Wellbeing and Parental
Labor Supply in Rural Areas
This paper investigates the impact of the Nutrition Improvement Program (NIP), a nationwide policy in rural China aimed at reducing child malnutrition, on children’s health,
educational outcomes, and parental labor market participation. Using a staggered rollout
difference-in-differences approach, I have found that the NIP increases children’s height
by 1.31 percent and increases the parental employment rate by 5.59 percent, weekly working hours by 4.1 hours, and incomes by 18%. The strongest gains occur in the service
sector. The program also produces positive spillover effects for untreated families, with
parents working 2.6 additional hours per week. Heterogeneous treatment analysis shows
these benefits are more pronounced for children primarily cared for by their parents, suggesting the program’s effects operate largely through an income effect rather than direct
nutritional improvements. These findings underscore the dual role of nutritional policies
1

in enhancing child welfare and household economic welfare, providing valuable insights
for designing welfare programs in low-income settings.
Keywords: Development Economics, Safety Net, School Lunch, Labor Supply
JEL Codes: I24, I38, J24, J08
2

1.1 Introduction
This paper investigates the Nutrition Improvement Program (NIP) in rural China, assessing its impact on both child development and parental labor market outcomes. Initiated by the Chinese government in 2011, the NIP aimed to address child malnutrition
in economically disadvantaged regions by providing free school lunches. While intended
primarily to improve children’s health and learning capacity, the program might also affect household dynamics by reducing parental caregiving burdens, thus enabling parents
to increase their workforce participation. From Kan and He (2024), it is documented that
the share of care work for women in rural China has increased much when they have children. They disproportionately bear the burden of household labor, including cooking and
preparing meals for their children, which significantly limits their time for economic or
personal activities. According to Mu and van de Walle (2011), rural women in China in
2006 spend 16.99 hours per week in family chores while men spend 4.61 hours only. Such
unequal time allocation places constraints on women’s participation in the labor market
and hinders their economic empowerment.
Programs that reduce caregiving burdens, such as providing school lunches, can significantly alleviate this time constraint. Studies in other contexts underscore how easing
these domestic demands can free women’s time for economically productive tasks. For
instance, Jagoe et al. (2020) conducted a field experiment in rural Kenya, demonstrating
that when cooking stoves were improved, participants reported reallocating the saved time
and energy to income-generating activities, enhancing their household economic stability.
Similarly, in rural China, where preparing lunch for children is a substantial time commitment, school-provided lunches could free up valuable hours for mothers to engage in work
3

or other activities. This reallocation of time could have far-reaching economic and social
implications, particularly in empowering women and supporting household incomes.
This project offers key insights into the effects of the NIP. First, the program leads to a
significant increase in children’s height in the medium term, showing that nutritional interventions can enhance physical and cognitive development, even when introduced beyond
early childhood. For instance, children who received free school lunch exhibit a 1.3%
greater increase in height compared to the control group. This finding challenges the conventional emphasis on early childhood as the sole critical period for effective interventions,
suggesting that support during later developmental stages remains important.
Second, the NIP generates notable effects on household economics by alleviating the
burden of caregiving and enabling parents to participate more actively in the labor market.
The results reveal that parents in NIP-benefiting households experience a 5.59% higher
employment rates, work 4.1 additional hours per week, and see an 18% increase in income,
with the strongest effects observed in flexible service-sector jobs. These findings highlight
the dual role of nutrition-based interventions: not only do they enhance child welfare, but
they also foster greater economic resilience for low-income families by freeing parental
time for more productive activities.
Third, heterogeneous treatment analysis shows that the program’s effects are particularly pronounced for children cared for primarily by their parents rather than grandparents
or other relatives. One possible explanation is that the program operates primarily through
an income effect rather than direct nutritional improvements from school meals. Parents
may reinvest additional time and income into better overall care and nutrition for their
children, amplifying the program’s impact. In addition, the program has little impact on
4

the labor supply and employment status for the parents with a high income level. These
findings highlight the importance of household economic improvements in helping disadvantaged children catch up in health and well-being. They also suggest that social safety
nets and family care can work complementarily to support child development.
To ensure the robustness of these findings, I have used advanced techniques for staggered treatment difference-in-difference methods, including the Callaway and Sant’Anna
(2021) and Borusyak et al. (2024) approaches to ensure the results are not biased due to
treatment effect heterogeneity. Furthermore, applying the method of Chaisemartin and
D’Haultfœuille (2020), I assessed the influence of negative weights in local treatment effects and found that only 4% of these weights were negative, with a negligible cumulative
sum. These methods mitigated biases associated with the “forbidden comparison” problem and enabled me to confirm the NIP’s impacts across diverse regions and treatment
timings.
This study contributes to a growing body of research on childhood interventions and
their effects on health, cognitive development, and household economic outcomes. Extensive literature has documented the impact of early-life interventions, particularly in nutrition and health, on later human capital development (Almond and Currie (2011), Heckman (2006), Almond, Currie, and Duque (2018), Currie and Vogl (2013), and Hoynes,
Schanzenbach, and Almond (2016)). The literature has shown that improved nutrition
during formative years enhances physical growth, cognitive abilities, and educational attainment, all of which are foundational for future development. This project focuses on a
program that targets children beyond early childhood and expands by exploring the continued importance of nutritional support in later developmental stages.
5

Research on school-based meal programs has shown effective results on children’s development. Studies in developing countries, such as India’s Midday Meal Scheme (Singh
et al. (2014)) and Ghana’s school feeding programs (Gelli et al. (2019)), have found positive effects on school attendance, academic performance, and child health. Lundborg et
al. (2022) evaluates the long-term effect of the school lunch in Sweden and shows that
mandatory free lunch increases the life time income, through an improvement on the educational attainment and health conditions. Kramer et al. (2021) uses a school-based nutri- ¨
tion intervention in India to improve the health condition of the students. Similarly, in the
U.S., programs like the National School Lunch Program (NSLP) have been linked with
improved nutrition and academic outcomes (Gundersen and Ziliak (2014)). Anderson et
al. (2018) shows healthy school meals help improve the students academic performance
in California. However, evidence on these programs’ broader economic effects, particularly on household labor supply and income, is limited, with this study filling the gap by
examining how school meal programs can influence parental employment and household
dynamics.
Another important area of related research is the impact of in-kind transfers on household labor supply and time allocation. In-kind transfers, such as food assistance or school
meals, are often introduced as part of broader safety net programs to support vulnerable
populations. While cash transfers have been more widely studied Baird et al. (2014) and
Banerjee et al. (2019), in-kind support has been shown to free up household resources, enabling greater flexibility in labor force participation. For instance, Hoynes and Schanzenbach (2012) found that the U.S. Supplemental Nutrition Assistance Program (SNAP) improved household food security and indirectly supported labor supply. Hoynes (1993)
6

also indicates that the labor supply should increase when the a disadvantaged family receives the cash or in-kind transfer. This research extends these findings by showing that
school-based nutritional support can similarly influence parental time allocation, particularly benefiting mothers’ participation in the labor market.
The paper proceeds as follows: Section 1.2 introduces the NIP. Section 1.3 presents
a simple theoretical framework to analyze how the NIP may impact the parental labor
supply. Section 1.4 presents data and summary statistics. Section 1.5 explains empirical
strategies. Section 1.6 reports results. Section 1.7 discusses the policy implication of NIP
and provides recommendations. Section 1.8 concludes.
1.2 Nutrition Improvement Program (NIP)
NIP was launched in the fall of 2011 with the primary goal of improving the nutritional
intake of students in rural areas. This program was part of a broader effort to address the
malnutrition and health disparities faced by children in China’s poorest regions. Initially,
the program was implemented in the eleven most impoverished regions, as designated by
the central government, along with specific areas in Tibet, the Zang areas of Sichuan and
Xinjiang provinces. In its first year, the NIP enrolled 20 million rural students across 1,296
counties. (Nutrition Improvement Program in Rurual China (n.d.))
Since its inception, the program has steadily expanded each year. By the mid-2021,
it encompassed 40 million rural students across 123,800 schools in 28 provinces in China
(Ten Years of the Nutrition Improvement Program and Future Recommendations (n.d.)).
This consistent expansion reflects the government’s commitment to improving rural students’ nutritional conditions and the growing recognition of the NIP importance.
7

Figure 1.1: Geographic Distribution of NIP Rollout Counties in 2016
Meal Standards and Government Support: When the program was first introduced,
the meal subsidy was set at three Chinese yuan per student per day (less than 50 U.S. cents
at the time). This subsidy aimed to provide children with basic nutritional needs that were
otherwise lacking in rural households. Over time, the subsidy increased to reflect both
inflation and the recognition of the need for greater nutritional support. In 2014, the meal
subsidy was raised to 4 yuan, and by 2020 it reached 5 yuan per student per day. This
increase in funding was accompanied by more substantial financial contributions from
8

both the central and local governments, signaling a greater national commitment to the
program’s goals.
The meals provided under NIP vary based on the school’s infrastructure. In schools
with dining halls, the program provides full lunches. However, in schools lacking these
facilities, the nutrition support is more flexible and may include supplemental food such as
milk and eggs, which are distributed between classes. This flexibility allows the program
to adapt to the varying conditions of rural schools while still ensuring that students receive
essential nutrients.
Target Population and Eligibility The NIP specifically targets rural students from
first to ninth grade, covering children during their compulsory education years in China.
Students living in urban areas, even within the same counties, are not eligible for the
program’s benefits. This distinction is based on the understanding that urban-based students typically have access to better resources and infrastructure compared to their rural
counterparts. By focusing on rural students, the program seeks to reduce inequalities in
educational and health outcomes between urban and rural populations.
Rollout and Regional Focus The rollout of the NIP is well-documented and publicly
available. In addition to publicly accessible information, I consulted government officials
to verify the implementation process, confirming that the rollout followed the timeline as
reported. The number of counties implementing the policy was 1,296 in 2012, 1,313 in
9

2014, 1,590 in 2016, 1,642 in 2018, and 1,762 in 20201
. Figure 1.1 illustrate the program’s rollout distribution in 2016, highlighting its focus on the southwestern regions of
China, which have lower socioeconomic status and higher rates of poverty. Notably, many
northern counties were also included in the early stages of the program.
The wealthier regions along China’s eastern coast, including the more affluent urban
counties, were not part of the rollout. This exclusion reflects the program’s targeted approach, designed to serve those regions most in need of nutritional support. By focusing on
economically disadvantaged areas, the NIP serves as a crucial safety net for rural students,
providing basic nutrition for their physical and cognitive development. As the program
continues to expand and evolve, it remains a cornerstone of China’s broader efforts to improve rural living standards and reduce inequalities between urban and rural populations.
1.3 Theoretical Framework
To demonstrate the impact of the NIP on family labor supply, I used a static time
allocation model based on the canonical time allocation model from Becker (1965). The
household maximizes the below utility function.
1The source includes Ten Years of the Nutrition Improvement Program and Future Recommendations
(n.d.), Significant Impact of the Rural Student Nutrition Improvement Program (n.d.), List of National and
Local Pilot Counties for the Rural Student Nutrition Improvement Program (n.d.), Report on the Implementation of the Rural Compulsory Education Student Nutrition Improvement Program (n.d.), Central Government
Allocates 2014 Special Funds for the Rural Student Nutrition Improvement Program (n.d.), and Supporting
Local Pilot Programs for the Rural Student Nutrition Improvement Program (n.d.)
10

max U (xa, xm, xl
, xg)
s.t. pmxm = pa (Q(Ld,AQ)−xa)−w

Ld −Lf

+I budget constraint
Lf +Lg +xl = T time constraint
xg = G(Lg,AG) = AGLg home produced good
where prices pa, pm, w and parameters T, AG, AQ and I are exogenous.
• Utility function U =U (xa, xm, xl
, xg): The household values agricultural goods, xa, market purchased goods, xm, leisure, xl
, home produced goods, xg.
• Non-marketed home produced goods, xg, including housekeeping, cooking, etc.
• Home-produced goods xg = G(Lg,AG) = AGLg has constant return to scale where AG
denotes the technology level and LG is the time spend on home produced goods. . In the
Appendix, I use a home-produced good with diminishing return to scale with labor to
show the impact of the increase in home-produced good technology.
• Agricultural production function: Q(Ld,AQ) where Ld is the total labor demand on a
farm and AQ denotes the agricultural production technology.
• The family supplies labor level Lf and the family labor and hired labor are perfect substitutes.
• Households are price takers for markets that exist.
1

We can substitute the time constraint and home-produced goods production function into
the budget constraint, and reach the maximization of interest:
maxxa,xm,xl
,xg,Ld U (xa, xm, xl
, xg)
s.t. pmxm + paxa +wxl +w

xg
AG

| {z }
= paQ(Ld,AQ)−wLd
| {z }
farm profit π
+ wT
|{z}
wage income
+I (1.1)
≜ Y
|{z}
Total income
We have created a static recursive model which is easy to solve and conduct comparative
static analysis.
We stick to the baseline model (1.1) and investigate the first order conditions. For labor
input Ld, the first order condition is:
pa
∂Q(Ld,AQ)
∂Ld
= w (1.2)
which contains only one endogenous variable Ld. We can solve (1.2) to determine L
∗
d
and
corresponding Q
∗ = Q

L
∗
d
,AQ

. Note that the once we know L
∗
d
and Q
∗
, the right-hand
side of the budget constraint of (1.1) is pinned down, i.e., we know π
∗
and Y
∗
.
(1.1) becomes
max
xa,xm,xl
,xg,Ld
U (xa, xm, xl
, xh)
s.t. pmxm + paxa +wxl +w
∗
xg = Y
∗
1

and first order condition for consumption variables, xa, xm, xl and xg:
∂U
∂ xi
= λ pi
where i = a,m,l,g and pl = w, pg = w
∗
.
Let U (xa, xm, xl
, xg) be Cobb-Douglas
The question becomes how to solve the following utility maximization problem of a
representative household:
max
xa,
,xm,xl
,xg
U(xa,
, xm, xl
, xg) =
∑
i∈{a,m,l,g}
ρix
ε
i
!1/ε
subject to the budget constraint
paxa + pmxm +wpl +
w
AG
xg = Y
Denote pl = w and pg =
w
AG
.
By FOCs, we have
xi =

pi
ρi

1
ε−1

ρg
pg

1
ε−1
xg
for i ∈ {a,m,l}. Plug in the budget constraint,
xg =
" ∑
i∈{a,m,l}

pi
ρi

1
ε−1
!
ρg
w

1
ε−1
A
1
ε−1
G +
w
AG
#−1
Y
and
13

Lg =
" ∑
i∈{a,m,l}

pi
ρi

1
ε−1
!
ρg
w
ε
ε−1
A
1
ε−1
G +w
#−1
Y
The function uniquely determines Xg. To assess the NIP’s impact on parental labor
supply, we can interpret the program as providing free Xg. By offering free meals to children, the NIP reduces the need for parents to allocate time for cooking at home, thereby
decreasing the home production labor (Lg). This reallocation frees up time for market
work or leisure. This theoretical framework guides the empirical analysis and predicts an
increase in parental labor supply in the NIP-treated family. An extended version of the
model is provided in Appendix G.3, demonstrating that under certain conditions, improving home production technology has a similar impact to providing free home goods. This
highlights a broader policy direction to enhancing labor market outcomes for disadvantaged families.
1.4 Data and Summary Statistics
The China Family Panel Studies (CFPS) dataset is particularly well-suited for investigating the impact of the NIP due to its comprehensive, longitudinal structure. The CFPS
surveyed baseline family members every two years from 2010 to 2020. This allows for
the analysis of both short- and medium-term effects on children’s health, educational outcomes, and parental labor market participation, with pre- and post-treatment data available.
CFPS offers a comprehensive set of variables critical to this study, including detailed data
on children’s physical and mental health, cognitive ability, parental involvement levels,
and household characteristics such as family relationship, income, employment information. This depth of information enables a nuanced assessment of the NIP’s impact on child
14

development and family dynamics. The dataset also clearly identifies whether individuals are located in rural or urban areas, facilitating a targeted analysis of rural populations
which are the primary beneficiaries of the program.
1.4.1 Sample Selection
For the purposes of this analysis, I restricted the dataset to include only families residing in rural areas. In each survey year, a specific variable indicates whether a family
lives in an urban or rural area. This classification is based on objective characteristics of
the community where the family resides, and it is officially designated by the government.
The government’s classification system considers various factors such as population density, infrastructure, administrative institutions, and the overall characteristics of the area to
determine whether a community is rural or urban. By limiting the dataset to rural households, I align analysis with the core objectives of the NIP, allowing us to focus on the
population most directly impacted by the program.
Additionally, for the parents’ data, I imposed an age restriction, limiting the analysis to individuals above 20 and under 60 years of age. This age range ensures that the
results capture only those individuals who are of working age and likely to be actively
participating in the labor market. Including individuals over 60 could introduce bias into
the analysis, as retirement decisions and pension eligibility could influence labor market
outcomes, independent of the NIP.
1.4.2 Main Outcome Variables and Summary Statistics
The data includes 7,085 children and 14,724 adults in total. The primary outcome
variables of interest for children were height, weight, mental health, cognitive test scores
15

(derived from a numerical sequence test and a word recall test), educational attainment
(derived from a math test and a word literacy test), and the frequency of parent-child
discussions about schoolwork each week. The main outcomes for adults of particular
interest included employment status, weekly work hours, and total income. Table 1.1
shows the summary statistics for the complete sample.
Height and weight, measured in centimeters and kilograms, served as a key indicator
of the program’s impact on children’s health. On average, the children are 134 centimeters
tall and weigh 65 kilograms. Sick is a dummy variable indicating whether the child had
been ill in the past month of the survey, with a mean of 0.2. Talk about Schoolwork
variable is the weekly count of schoolwork-related conversations, which measures how
often parents discussed schoolwork with their children. This variable offered an insight
into whether parents used time saved by the school lunch program to engage more actively
in their child’s education. On average, parents discussed schoolwork 2.9 times per week.
The Parent Take Care variable, a dummy indicator, identified whether children were
primarily cared for by their parents, as opposed to grandparents or other relatives. For
treated children, only 40.2% of children are directly cared for by their parents in 2010,
much lower than the control group which is 66.5%. This implies the program has targeted
the underdeveloped rural areas where parents might go out to work instead of taking care
of their children at home.
The depression evaluation consisted of six questions assessing emotional well-being
over the previous month:
How often have you felt down or depressed?
How often have you felt mentally tense or anxious?
16

How often have you felt restless or unable to stay calm?
How often have you felt hopeless about the future?
How often have you found it difficult to do anything?
How often have you felt that life has no meaning?
Each question was to be answered on a five-point scale: 1. Almost every day 2. Two
to three times per week 3. Two to three times per month 4. Once per month 5. Never; The
total score ranged from 6 to 30, with higher scores indicating lower levels of depression. I
defined a score below 22 as indicating a tendency toward depression and created a dummy
variable, d depression, which equaled 1 if the respondent met this criterion.
The word recall and numeric sequence tests, administered in 2012 and 2016, serve as
measures of children’s cognitive ability. The word literacy and math tests, conducted in
2010 and 2014, capture educational attainment. As a result, the sample size was much
smaller for these outcome variables.
The word recall test involved showing respondents a set of common words. After five
minutes, the interviewer asked them to recall the ten words they had previously seen. The
number of correctly recalled words was recorded as the score. In the numeric sequence
test, respondents were given a series of numbers with missing values. They had to identify
the correct numbers based on the underlying pattern. A higher score was assigned to those
who correctly filled in more missing numbers.
In the word literacy test, participants were given with a set of Chinese characters,
and the number of words they could recognize was recorded. The math test evaluated
problem-solving skills and recorded the highest number of correctly answered questions.
To enhance testing efficiency, both literacy and math tests employed three starting points
17

tailored to the respondent’s education level. Children in elementary school or with lower
education started at the first point, those in middle school began at the second, and those in
high school began at the third. This adaptive design ensured that the tests were appropriate
for the educational background of each child.
Table 1.1: Summary Statistics of Child and Adults Characteristics
(1) (2) (3)
VARIABLES N mean sd
Children Data
Age 14,939 10.31 2.935
Height 14,939 134.0 22.02
Weight 14,939 65.36 24.21
Sick 14,939 0.207 0.405
Depression 3,149 0.137 0.343
Math Test 3,168 10.46 4.483
Word Test 3,168 20.70 7.455
Word Recall Test 2,449 5.290 2.032
Number Sequence Test 2,188 523.2 28.65
Talk about Schoolwork 14,939 2.978 1.185
Parent Take Care 14,939 0.489 0.500
Number of Children 7,085
Adults Data
Age 53,087 43.51 10.81
Employment 53,087 0.748 0.434
Working Hour 53,087 19.61 25.99
Income 53,087 9,405 22,702
Service Sector Job 53,087 0.120 0.325
Parent as Caregiver 53,087 0.336 0.472
More than 1 child 53,087 0.517 0.500
Number of Adults 14,724
Notes: The data source is China Family Panel Survey (CFPS). The sample is restricted to children aged 5 to 16 and individuals aged
20 to 60 who reside in rural areas. All of these observations are used in the regression analysis.
For adults, the sample consisted of 14,724 individuals, including 7,637 females and
7,087 males. Employment outcomes were measured using a dummy variable that indicated
whether the individual had worked in the previous month, with an average employment
18

rate of 74.8%. To capture the intensity of employment, I examined weekly work hours,
which averaged 19.61 hours per week across all adults. Additionally, I assessed total income to determine whether changes in work hours due to the NIP translated into increased
earnings, offering insights into the program’s broader economic effects on households.
The average income in the sample was 9,405 yuan.
Service Sector Job variable is a dummy variable, defined as working in the occupations excluding agriculture, mining, and manufacturing. On average, 12% of parents were
employed in the service sector. This sector was particularly relevant for understanding
the program’s impact, as service jobs typically offered more flexible schedules and lower
entry barriers. These characteristics make such jobs more accessible to parents, especially
mothers, who despite reduced the caregiving burden from the NIP, still need to balance
family and work responsibilities.
1.5 Empirical Strategy
To causally identify the impact of NIP, I utilized the canonical two-way fixed effects
(TWFE) model, a widely used econometric approach that allows for the control of both
time-invariant characteristics of individuals and common shocks across time periods. Additionally, to capture the dynamic effects of NIP over time, I employed an event study
framework, which provides a detailed view of how the program’s impact evolved before
and after its implementation. Given the staggered adoption of the program, I have employed more advanced techniques to verify the results. To further understand the spillover
effects of the program on parents, I modified the model to capture both the direct impacts
on parents whose children are directly benefited from the program and the broader effects
19

on other individuals within the community. Finally, I included the interaction between the
treatment and property dummies to examine heterogeneous treatment effects.
1.5.1 Direct Impact of the Program on Children and Adults
I estimated equation 1.3 to identify the treatment effect of NIP on children’s and adults’
outcomes. Since I could not directly observe whether each student consumed the provided
meals as intended, these estimates represent intent-to-treat (ITT) effects. This approach
captured the overall impact of the program, regardless of individual compliance, by evaluating the average effect on all eligible individuals in treated areas.
Yi jt = β0 +β1Treatmenti jt +ηt +ηi +εi jt (1.3)
Where:
Yi jt represents the outcome variable of child or adult i in county j at time t. For
children, these outcomes include height, weight, cognitive test scores, and the frequency
of schoolwork discussions with parents. For parents, the variable includes employment
status, weekly working hours, and log income. Treatmenti jt is an interaction result of
Treatedjt and Eligibleit. Treatedjt is a dummy variable indicating whether the nutrition
program was implemented in county j at time t, and Eligibleit denotes if there is a child
attending primary or middle school in the year t in the family. ηt and ηi represent time and
individual fixed effects, respectively, which control for time-invariant individual characteristics and common shocks affecting all observations in a given year. Additionally, age has
been included as a control variable in the regression. Standard errors are clustered at the
county level to account for potential within-county correlation. This design ensures that
20

the estimates of β1 reflect the causal impact of the NIP on children’s physical outcomes
and adults’ labor outcomes.
Taking into account treatment effect heterogeneity and forbidden comparison issues,
I conducted the CSDID proposed in Callaway and Sant’Anna (2021). The method only
uses the never treated group as the control group so to avoids the negative weight problem
which may have biased the result in TWFE. The aggregated treatment effect is a weighted
average treatment effect for each comparison.
I treated the NIP intervention as having an absorbing effect on both children and parents, meaning once introduced, its influence persisted over time. For parental labor outcomes, I assumed that these effects continued even after the child graduated from middle
school and was no longer eligible for the program. By this stage, children had grown up,
and the need for intensive caregiving had typically diminished compared to earlier years,
and thus parents could maintain their employment benefits achieved during the program
period. This approach captured the sustained influence of the NIP on household dynamics
and economic outcomes.
1.5.2 Dynamic Effect
I estimated equation 1.4 to examine the dynamic effects of the NIP.
Yi jt = β0 +
8
∑
α=−10[α̸=−2]
τα1[[t − Max{FirstGradei
,NIPj}] = α] +ηt +ηi +εi jt (1.4)
The indicator term, 1[[t − Max{FirstGradei
,NIPj}] = α] captures the event time,
where NIPj
is the year the NIP was adopted in county j, and FirstGradei
is the year
student i entered first grade. The event time, denoted by α, measures the number of years
21

since the child began receiving NIP treatment, with coefficients ranging from -10 to 8
years. Year -2 is used as the reference year.
The event-study framework enabled me to examine how the impact of the NIP unfolded
over time, capturing whether its benefits manifested immediately, accumulate gradually,
or required extended exposure to take full effect. By estimating the program’s impact
across various event years, I could differentiate between short- and medium-term effects.
Negative values of α indicate years before a student began receiving NIP benefits, serving
as a pre-treatment baseline. Positive values of α, on the other hand, tracked the program’s
effects over time, allowing me to observe whether improvements in children’s physical
outcomes, such as height or weight, strengthened with prolonged participation in the program. In addition to this OLS specification, I followed the methods proposed by Callaway
and Sant’Anna (2021) and Borusyak et al. (2024) to test the parallel trend assumption and
adjust for treatment effect heterogeneity in the Robustness Check section.
1.5.3 Spillover Effect on Parents Labour Supply
To examine the spillover effects of the NIP on families, particularly on parental availability for work in the county that the policy was implemented, I employed the following
model in Equation 1.5.
Yi jt = β0 +β1County Treatedjt +β2County Treatedjt ∗Eligibleit +ηt +ηi +εi jt (1.5)
County Treatedjt and Eligibleit are defined in Section 1.5.1, and their interaction
forms the treatment indicator Treatmenti jt in Equation 1.3. In Equation 1.5, β1 captures
22

the general equilibrium effects of the program on adults who do not have a child eligible
for the NIP (i.e., they do not have children or their children are not in primary or middle
school). This coefficient reflects how the program’s introduction might affect overall labor
market conditions within a treated county, potentially influencing even those households
that do not directly benefit from the free school meals.
1.5.4 Heterogeneous Treatment Effect
To analyze heterogeneous treatment effects, I extended Equation 1.3 to Equation 1.6
and Equation 1.7 by adding interaction terms. This helped explore the program’s full
impact and the mechanisms behind it. Specifically, I examined how effects varied based
on the child’s primary caregiver, household child amount, and parental occupation sector.
Yi jt = β0 +β1Treatmenti jt +β2Treatmenti jt ∗Parent Careit +ηt +ηi +εi jt (1.6)
Yi jt = β0 +β1Treatmenti jt +β2Treatmenti jt ∗Parent Careit
+β3Treatmenti jt ∗Parent Careit ∗Propertyit +ηt +ηi +εi jt
(1.7)
Parent Careit is a dummy variable equal to 1 if a parent is the primary caregiver
for the children. Propertyit is a dummy variable indicating whether the specified property condition is met. It includes the following two dummy conditions Multi Cit, and
ServiceSectorit. ServiceSectorit is 1 if the parent’s job is in the service sector, including
all industries except manufacturing, mining, and agriculture. Multi Cit equals 1 if the
family has more than one child. The coefficient β3 captures the differential impact of the
23

program between treated children or individuals who have the specified property and those
who are treated but do not have it.
1.5.5 Identification Assumption
To causally identify treatment effects with the models, two key assumptions should
be made: the no anticipation assumption and the parallel trend assumption (Callaway and
Sant’Anna (2021), Roth et al. (2023), Angrist and Pischke (2008), and Athey and Imbens
(2018)). In addition, the forbidden comparison problem and treatment effect heterogeneity
should be discussed.
Parallel Trend
In examining the impact of the NIP on children and adults, validating the parallel trends
assumption is crucial. This assumption posits that, in the absence of the NIP intervention,
outcome trends for both treatment and control groups would have evolved similarly over
time. In a staggered adoption setting, this assumption becomes especially challenging, as
each 2x2 comparison across different treatment timings requires parallel trends. Adults,
children, and even counties could differ significantly in healthcare, family economic condition, and infrastructure. Thus, it is difficult to simply assume parallel outcome trends
without the program.
I employed the event study framework and robust estimation methods, including CSDID proposed in Callaway and Sant’Anna 2021 and BJS imputation estimation proposed
in Borusyak et al. (2024) to conduct test for this assumption. CSDID event study figures
and BJS estimations are provided in the Appendix G.1 and G.2. The OLS event study
and the these heterogeneity robust estimation methods confirm that there is no significant
divergence in the pretend of treated and control groups.
24

No Anticipation
In addition, the no anticipation assumption posits that individuals or families did not
alter their behavior in anticipation of the NIP. There are several reasons why this assumption is likely credible in the context of the NIP. First, the NIP was implemented through
a sudden, top-down approach, with the government selecting target regions without prior
public notice. Families had little to no advance information, making it unlikely that they
could adjust their behavior in anticipation of the program. Furthermore, the benefits of
the NIP, such as improved nutrition and potential increases in parental free time, were
not sufficiently substantial to warrant major preemptive changes in family routines or employment. Finally, event study shows no significant changes in key outcomes before the
program’s implementation, confirming that families did not alter their behavior in anticipation of the NIP.
Forbidden Comparison and Treatment Effect Heterogeneity
When comparing outcomes using a traditional TWFE approach, it is possible for
already-treated units to be used as control groups for later-treated units, leading to what
is known as a “forbidden comparison”. This issue has been well-documented in recent
literature, with researchers highlighting the risks of biased estimates in staggered adoption
settings. (Goodman-Bacon (2021), Chaisemartin and D’Haultfœuille (2020), and Roth et
al. (2023))
To address this concern, in Section 1.6.6, I assessed the severity of the problem by
following the approach proposed by Chaisemartin and D’Haultfœuille (2020), which involves counting the number of negative weights in the estimations. Negative weights arise
25

when the TWFE model places improper weights on comparisons that can distort the overall results. By quantifying the extent of these negative weights, I have assured that the
forbidden comparison issue is minimal under my TWFE setting.
In addition, the TWFE model assumes homogeneous treatment effects across groups
and time, meaning it does not account for potential variations in how different groups
respond to treatment at different points in time. In settings with staggered policy adoption,
treatment effects are likely to vary across groups and locations. This creates a potential
source of bias. The advanced estimation approach of BJS estimator and CSDID have been
provided with the consideration of treatment effect heterogeneity in the Results section.
1.6 Results
1.6.1 Direct Impact on Children
Table 1.2 lists the TWFE estimates of NIP impact on children health condition, cognitive ability and parental engagement, derived from Equation 1.3. The analysis shows that
the program leads to a modest 0.01 percent improvements in children’s height but has no
statistically significant improvements in weight. Children in the treated group were able
to recognize, on average, 1.73 more words than the control group. However, in panel B,
limiting the sample to girls, the impact on height becomes insignificant.
Moreover, while the program alleviates parental burdens related to meal preparation,
the free-up time does not appear to have been reallocated to increasing school-related
interactions between parents and their children. On the contrary, the frequency of discussions about schoolwork decreased by 0.14-times per week, suggesting that instead parents
26

are using the additional time for other activities, such as employment or household responsibilities. This reallocation is consistent with subsequent findings that show increased
parental labor market participation. These results indicate that while the program indirectly supports children’s education through improved nutrition and parental time redistribution, it has reduced parental involvement in educational activities.
Figures 1.2a and 1.2b show the impact of NIP on children’s height and weight over
time, highlighting the program’s dynamic effect on physical development. The figures
demonstrate that the immediate effects of the program on height are minimal, with no
significant changes observed at the implementation year of the program. However, a clear
and delayed impact emerges as the duration of exposure to the program increases.
Starting from the second year, the effects on height become more pronounced, with
children in the treated group exhibiting height increases ranging from 0.01% to 0.05% following six years of intervention compared to their baseline. This progressive improvement
suggests that the NIP’s impact on physical development requires sustained participation to
achieve its full potential. Such delayed effects are consistent with the understanding that
nutritional interventions often have cumulative impacts, as consistent and improved dietary intake over time supports growth and development.
The NIP does not exhibit a noticeable impact on children’s weight, as revealed in the
event study in Figure 1.2b. This discrepancy likely stems from the program’s focus on improving nutritional quality, such as micronutrients essential for increases in height, rather
than caloric intake. Height reflects long-term nutritional improvements, while weight responds more to short-term caloric sufficiency.
27

Table 1.2: Impact of NIP on Child Health, Cognitive Development, and Parental Engagement
(1) (2) (3) (4) (5) (6) (7) (8) (9)
VARIABLES Log(Height) Log(Weight) Sick Depression Word Test Math Test Num Seq Test Word Recall Talk Schoolwork
Panel A: All Gender
Treatment 0.00969* -0.000368 -0.0266 -0.01000 1.731* -0.156 11.00 0.362 -0.141**
(0.00525) (0.0148) (0.0194) (0.0609) (0.975) (0.587) (7.300) (0.473) (0.0571)
Observations 14,939 14,939 14,939 3,153 3,168 3,168 2,188 2,449 14,939
R-squared 0.621 0.636 0.012 0.001 0.517 0.516 0.091 0.015 0.006
Number of pid 7,085 7,085 7,085 2,844 2,852 2,852 2,016 2,221 7,085
Mean 4.88 4.11 0.21 0.14 20.70 10.46 523.17 5.29 2.98
Individual FE YES YES YES YES YES YES YES YES YES
Year FE YES YES YES YES YES YES YES YES YES
Panel B: Girls Only
VARIABLES Log(Height) Log(Weight) Sick Depression Word Test Math Test Numb Seq Test Word Recall Talk School
Treatment 0.00902 -0.00282 -0.00500 0.0263 0.906 0.0218 34.87** 0.929 -0.162**
(0.00822) (0.0188) (0.0244) (0.0751) (1.172) (0.938) (14.41) (0.818) (0.0668)
Observations 6,998 6,998 6,998 1,518 1,527 1,527 1,015 1,152 6,998
R-squared 0.592 0.655 0.013 0.003 0.549 0.463 0.148 0.024 0.006
Number of pid 3,368 3,368 3,368 1,360 1,365 1,365 939 1,046 3,368
Mean 4.88 4.09 0.20 0.13 21.35 10.58 521.56 5.39 2.96
Individual FE YES YES YES YES YES YES YES YES YES
Year FE YES YES YES YES YES YES YES YES YES
Notes: Sick is a dummy variable indicating whether the child was ill in the past month. Depression is a dummy variable, representing a propensity for depression as defined
in the data section. The Number Sequence Test and Word Recall Test measure cognitive ability. The Word Test and Math Test assess educational attainment, with the Word
Test capturing chinese word recognition and the Math Test reflecting correct responses to math questions. Talk about Schoolwork measures the weekly frequency (0–5) of
parent-child discussions on schoolwork. Child’s age is included as a control variable. To account for potential correlations within counties, standard errors are clustered at the
county level. *** p
<0.01, ** p
<0.05, * p
<0.1.
28

(a) Event Study: Impact of NIP on Children
Log(Height)
(b) Event Study: Impact of NIP on Children
Log(Weight)
Figure 1.2: Event Study on Children Health Outcomes
1.6.2 Direct Impact on Adults
From Table 1.3, the static impact of the NIP on parents is evident through the immediate reduction in the caregiving burden, particularly that associated with meal preparation.
This freed-up time has been reallocated to employment, leading to a 5.36% increase in
employment rates among parents in the treated group compared to the control group. Additionally, average weekly working hours rose by 4.42 hours, and income increased by
34% relative to the control group. The average employment rate in the dataset is 75%,
and the mean weekly working hours are 19.61. The observed increase of 4.42 hours represents a 22% rise. The effects are especially strong for mothers, who traditionally bear
the majority of caregiving responsibilities. For women, employment rates increased by
5.95%, weekly working hours rose by 4.975 hours, and income levels grew by about 33
percent. These findings highlight the critical role of the NIP in enabling parents to re-enter,
or increase their participation in the labor force.
Dynamic analysis, as shown in Figure 1.3a, 1.3b, and 1.3c, reveals how these effects
become even more pronounced over time. As parents adjust to their newly available time,
29

Table 1.3: Impact of NIP on Adult Labor Outcomes
(1) (2) (3) (4) (5) (6)
VARIABLES Employment Working Hour Log Income Employment Working Hour Log Income
Treatment 0.0536** 4.419*** 0.297** 0.0595** 4.975*** 0.288**
(0.0249) (0.729) (0.123) (0.0252) (0.741) (0.116)
0.122
Observations 53,087 53,087 53,087 27,809 27,809 27,809
R-squared 0.082 0.152 0.107 0.077 0.161 0.115
Number of pid 14,724 14,724 14,724 7,637 7,637 7,637
Mean 0.75 19.61 3.39 0.70 15.89 2.52
Individual FE YES YES YES YES YES YES
Year FE YES YES YES YES YES YES
Gender All All All Female Female Female
Notes: The dependent variable, Employment, is a dummy variable that indicates whether the individual is currently employed.
Working Hours measures the number of hours the individual works per week. Log Income represents the total income the individual
earns, expressed in logarithmic form. Individual’s age is included as a control variable. Standard errors are clustered at the county level
to account for potential correlations within counties. The estimates are derived using the canonical Two-Way Fixed Effects (TWFE)
model. *** p<0.01, ** p<0.05, * p<0.1
employment rates continue to rise, and weekly working hours increase further, indicating
that the program fosters sustained labor market engagement.
The significant income gains for parents, especially mothers, are driven by longer
working hours and access to higher-paying jobs. The additional time available allows
parents to take on more consistent jobs. It also allows them to engage in training and
acquire professional skills, which can further enhance job prospects. These factors collectively generate a cumulative effect that drives wage growth and greater financial stability
over time. For mothers, the reallocation of time to employment activities improves their
economic empowerment, narrowing the gender gap in labor market participation.
1.6.3 Estimation with CSDID Method
Table 1.4 presents the aggregated treatment effects estimated using the CSDID method
proposed in Callaway and Sant’Anna (2021). This approach ensures that the estimates
30

(a) Event Study: Dynamic Impact of NIP on
Parent Employment
(b) Event Study: Dynamic Impact of NIP on
Parent Weekly Working Hour
(c) Event Study: Dynamic Impact of NIP on
Parent Log(Income)
Figure 1.3: Event Study on Adult Labor Market Outcomes
31

accurately capture the causal impact of the NIP by exclusively using the never-treated
group as a control and account for treatment effect heterogeneity. By excluding forbidden
comparison pairs, the method improves identification and yields unbiased results.
The findings are very consistent with those from the TWFE static analysis, indicating
that the NIP not only improves children’s physical health but also significantly enhances
parental labor force participation. Specifically, the program results in a 1.31 percent increase in children’s height relative to the never-treated group. Additionally, parents’ employment rates rose by 5.59 percent, average weekly working hours increased by 4.103
hours, and household income grew by 18 percent.
The observed gains in children’s height underscore the program’s effectiveness in supporting physical development. At the same time, by alleviating caregiving burdens, the
NIP enables parents to participate more actively in the labor market, translating time saving into tangible economic benefits. This reallocation of household labor highlights the
broader potential of targeted nutritional interventions to transform the supply of household labor and improve the economic condition of the family.
1.6.4 Spillover Effects of the NIP on Non-Eligible Parents
Although the NIP was primarily designed to support families with eligible children enrolled in primary and middle schools, Table 1.5 reveals notable spillover effects on other
parents in the community who do not have eligible children. One of the most significant
findings is that parents without eligible children increase their weekly working hours by
2.649 hours, with the effect being more pronounced among males. Although smaller than
the gains observed among directly supported households, this increase demonstrates that
32

Table 1.4: CSDID Aggregate Summary Results
(1) (2) (3) (4) (5)
Variable Log Height Log Weight Employment Work Hour Log(Income)
ATT 0.0131∗∗∗ -0.135 0.0559∗∗∗ 4.103∗∗∗ 0.169∗
2.35 (-1.21) (4.30) (7.54) (1.74)
Observations 9,500 9,500 49,576 49,576 49,576
Sample Children Children Adult Adult Adult
Gender All All All All All
Notes: The estimates are calculated by Stata package “csdid” proposed in Callaway and Sant’Anna
2021. Age serves as a control. Standard errors are clustered at the county level to account for
potential correlations within counties. t statistics in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗
p < 0.01.
such interventions can impact labor market dynamics beyond their primary targets. Notably, however, non-eligible female workers experienced a decline in income following the
program’s implementation.
Several mechanisms may underlie the observed increase in working hours among noneligible parents. First, the NIP likely improved the community’s overall economic environment. By alleviating caregiving responsibilities on families with eligible children,
the program may have indirectly freed up local resources and expanded labor market opportunities. For instance, increased employment and income among eligible households
may have increased local demand for goods and services, thereby generating further job
opportunities that benefited non-eligible parents.
Second, the NIP may have directly generated new labor demand through infrastructure
improvements in treated communities. The requirement to provide school lunches may
have prompted schools to expand facilities and infrastructure, leading to a local labor
demand shock. Parents without eligible children, particularly those involved in service
sector or part-time work, may have found new opportunities created by the construction or
33

renovation of school dining facilities and related infrastructure. This reflects the program’s
potential to stimulate broader economic engagement beyond its direct beneficiaries.
Table 1.5: Spillover Impact of NIP on the Non-Eligible Adults
(1) (2) (3) (4) (5) (6)
VARIABLES Employment Working Hour Log Income Employment Working Hour Log Income
County Level Treated -0.0172 2.649*** -0.0418 -0.0314 1.904* -0.231*
(0.0373) (0.967) (0.141) (0.0390) (1.002) (0.123)
County Treated*Eligible 0.0314*** 2.399*** 0.289*** 0.0401*** 2.825*** 0.348***
(0.0119) (0.631) (0.0901) (0.0138) (0.825) (0.111)
Observations 53,087 53,087 53,087 27,809 27,809 27,809
R-squared 0.082 0.153 0.107 0.076 0.161 0.115
Number of pid 14,724 14,724 14,724 7,637 7,637 7,637
Mean 0.75 19.61 3.39 0.70 15.89 2.52
Individual FE YES YES YES YES YES YES
Year FE YES YES YES YES YES YES
Gender All All All Female Female Female
Notes: The dependent variable, Employment, is a dummy variable that indicates whether the individual is currently employed.
Working Hours measures the number of hours the individual works per week. Log Income represents the total income the individual
earns, expressed in logarithmic form. Treatedjt is a dummy variable indicating whether the nutrition program was implemented in
county j at time t, and Eligibleit denotes if there is a child attending primary or middle school in the year t in the family. Individual’s
age is included as a control variable. Standard errors are clustered at the county level to account for potential correlations within
counties. The estimates are derived using the canonical Two-Way Fixed Effects (TWFE) model. *** p<0.01, ** p<0.05, * p<0.1
However, as an unintended consequence of the program, the decline in income among
non-eligible female workers points to potential general equilibrium effects. It is possible
that some higher-paying job opportunities were taken by newly available eligible parents,
leading to a crowding-out effect for non-eligible females. As a result, despite longer working hours, non-eligible women’s economic well-being may have declined. These findings
suggest that policymakers should carefully consider general equilibrium effects and be attentive to the broader distributional impacts of targeted interventions, particularly when
such programs influence local labor supply.
34

1.6.5 Heterogeneous Treatment Effect
This section examines the heterogeneous effects of the policy treatment, specifically
exploring how the impact varies depending on the primary caregiver, the parents’ occupation sector, and the age at which children first receive treatment.
Heterogenous Impact on Children by Primary Caregiver and Income
The NIP appears to influence children primarily through an income effect rather than
direct nutritional improvements. Table 1.6 supports this conclusion. Panel A shows that
children with parents as their primary caregivers experienced a 1.82% increase in height.
Furthermore, Panel B reveals a stronger effect when the sample is restricted to girls, with
height increasing by approximately 2.8%. However, girls’ weight has decreased by about
6%. Neither group shows an increase in schoolwork-related discussions with parents. This
suggests that parents used their freed-up time for employment or other responsibilities
rather than direct educational engagement.
If the program’s primary mechanism is improved nutrition intake, one would expect
to observe similar health benefits for all children, regardless of caregiving arrangements.
However, the results indicate otherwise. Children without parents as their main caregivers
exhibit no significant changes in height or weight before and after the program, suggesting
that the nutritional support provided by the program alone is not enough to deliver measurable improvements in their physical health. By contrast, children with parents as their
primary caregivers experience significant increases in height, which is a indicator of better
nutrition and overall care. This suggests that the program’s true impact lies in its ability to
35

free up parental time, allowing parents to engage more actively in the labor market, generate additional income, and reinvest these resources into their children’s well-being. The
section below on the heterogeneous effect on adults further substantiates this conclusion.
Table 1.7 presents the heterogeneous effects of the program on children between different household income levels. Children from families in the top 25% income bracket exhibit
no significant changes in physical health outcomes, likely due to their already sufficient
nutritional intake. In contrast, children from the bottom 25% income group experience
1.73 percent improvements in height. These findings suggest that the program effectively
functions as a safety net, targeting and benefiting disadvantaged students by helping them
catch up in terms of physical development.
Heterogeneous Impact on Adults by Primary Caregivers and Income Level
From Table 1.8, the impact on the labor market outcome is evident for parents who
are primary caregivers. The employment probability is significantly higher for 8.22 percent after the NIP program started for the parents with the main caregiving responsibility.
Average weekly working hours increased by 5 hours, compared to 3.1 hours for parents
without primary caregiving responsibilities. Consequently, the employment effect was
successfully translated into an income increase of about 32 percent for the main caregiving parents. Each effect is greater for the mothers.
36

Table 1.6: Heterogeneous Impact of NIP on Children by Primary Caregiver
(1) (2) (3) (4) (5) (6) (7) (8) (9)
VARIABLES Log(Height) Log(Weight) Sick Depression Word Test Math Test Num Seq Test Word Recall Talk School
Panel A: All gender
Treatment 0.000798 0.0119 -0.0173 -0.0444 2.765** 0.130 13.52* 0.502 -0.109
(0.00769) (0.0206) (0.0226) (0.0742) (1.151) (0.634) (8.098) (0.547) (0.0761)
Treatment*Parent Takecare 0.0182* -0.0250 -0.0190 0.0783 -2.334 -0.647 -15.15 -0.781 -0.0643
(0.00944) (0.0178) (0.0286) (0.0716) (1.547) (0.902) (14.36) (0.926) (0.102)
Observations 14,939 14,939 14,939 3,149 3,168 3,168 2,188 2,449 14,939
R-squared 0.622 0.636 0.012 0.005 0.523 0.517 0.095 0.017 0.006
Number of pid 7,085 7,085 7,085 2,840 2,852 2,852 2,016 2,221 7,085
Mean 4.88 4.11 0.21 0.14 20.70 10.46 523.17 5.29 2.98
Individual FE YES YES YES YES YES YES YES YES YES
Year FE YES YES YES YES YES YES YES YES YES
Panel B: Girls only
Treatment -0.00472 0.0269 0.00271 0.0401 1.703 0.194 38.51** 0.999 -0.201***
(0.0104) (0.0268) (0.0354) (0.0998) (1.256) (1.123) (15.55) (0.867) (0.0742)
Treatment*Parent Takecare 0.0280** -0.0605** -0.0157 -0.0270 -1.560 -0.336 -32.75** -0.625 0.0796
(0.0126) (0.0268) (0.0449) (0.0979) (1.452) (1.569) (15.15) (1.947) (0.123)
Observations 6,998 6,998 6,998 1,518 1,527 1,527 1,015 1,152 6,998
R-squared 0.593 0.656 0.013 0.004 0.551 0.463 0.158 0.024 0.006
Number of pid 3,368 3,368 3,368 1,360 1,365 1,365 939 1,046 3,368
Mean 4.88 4.11 0.21 0.14 20.70 10.46 523.17 5.29 2.98
Individual FE YES YES YES YES YES YES YES YES YES
Year FE YES YES YES YES YES YES YES YES YES
Notes: Sick is a dummy variable indicating whether the child was ill in the past month of the survey. Depression is a dummy variable, representing a propensity for depression
as defined in the data section. The Number Sequence Test and Word Recall Test measure cognitive ability. The Word Test and Math Test assess educational attainment, with the
Word Test capturing Chinese character recognition and the Math Test reflecting correct responses to math questions. Talk about Schoolwork measures the weekly frequency
(0–5) of parent-child discussions on schoolwork. Child’s age is included as a control variable. To account for potential correlations within counties, standard errors are
clustered at the county level. *** p
<0.01, ** p
<0.05, * p
<0.1.
37

Table 1.7: Heterogeneous Impact of NIP on Children by Income Level
(1) (2) (3) (4) (5) (6)
VARIABLES Log(Height) Log(Weight) Sick Log(Height) Log(Weight) Sick
Treatment 0.0173** 0.00617 -0.0387 0.0175 0.0143 -0.0240
(0.00766) (0.0204) (0.0353) (0.0156) (0.0275) (0.0506)
Observations 5,197 5,197 5,197 3,146 3,146 3,146
R-squared 0.628 0.635 0.009 0.650 0.629 0.018
Number of pid 2,866 2,866 2,866 2,049 2,049 2,049
Mean 4.89 4.13 0.20 4.87 4.07 0.21
Individual FE YES YES YES YES YES YES
Year FE YES YES YES YES YES YES
Parent Care YES YES YES YES YES YES
Income Level Low Low Low High High High
Notes: Sick is a dummy variable indicating whether the child was ill in the past month of the survey. Depression is a dummy variable, representing a propensity for depression
as defined in the data section. The Number Sequence Test and Word Recall Test measure cognitive ability. The Word Test and Math Test assess educational attainment, with the
Word Test capturing Chinese character recognition and the Math Test reflecting correct responses to math questions. Talk about Schoolwork measures the weekly frequency
(0–5) of parent-child discussions on schoolwork. Parent Careit is a dummy variable equal to 1 if a parent is the primary caregiver for the children. Child’s age is included as a
control variable. To account for potential correlations within counties, standard errors are clustered at the county level. *** p
<0.01, ** p
<0.05, * p
<0.1.
38

For parents who are the main caregivers, the NIP significantly reduces the time burden associated with meal preparation and caregiving. These parents may need to invest
a considerable amount of time each day preparing lunch for their children, especially in
households without additional support. With the school providing meals, this responsibility is partially alleviated, freeing up substantial time that these parents can redirect toward
employment. Since primary caregivers usually have less flexibility in time allocation due
to caregiving duties, the program’s support enables them to pursue work or extend their
working hours. This leads to a more notable increase in employment participation compared to parents who are not primary caregivers.
Table 1.9 indicates that the individuals in lower-income households who face tighter
budget constraints receive more benefit from the NIP. The NIP enables these families to
reallocate time from caregiving to employment activities, thus increasing income. In contrast, households in the higher-income group experience no significant change in employment rate or income, likely because they already possess sufficient resources and the NIP
does not relax their time constraint. These findings underscore the program’s effectiveness
in addressing the needs of financially disadvantaged and caregiving-constrained populations.
Appendix Table A3 shows that the NIP’s effects are more pronounced among parents
with multiple children. Compared to the control group, these parents experienced a 7.2%
increase in employment rate. Parents with more than one child typically face greater caregiving demands, and the program’s support helps relieve this burden, resulting in a greater
impact on their labor market participation.
39

Table 1.8: Heterogeneous Impact of NIP on Adults by Primary Caregiver
(1) (2) (3) (4) (5) (6)
VARIABLES Employment Working Hour Log Income Employment Working Hour Log Income
Treatment -0.00168 3.107** 0.117 -0.0215 3.202*** -0.00711
(0.0326) (1.200) (0.203) (0.0377) (1.119) (0.194)
Treatment*Parent Takecare 0.0822** 1.950* 0.268 0.115*** 2.514** 0.418*
(0.0330) (1.151) (0.206) (0.0388) (1.126) (0.221)
Observations 53,087 53,087 53,087 27,809 27,809 27,809
R-squared 0.083 0.153 0.107 0.078 0.161 0.115
Number of pid 14,724 14,724 14,724 7,637 7,637 7,637
Mean 0.75 19.61 3.39 0.70 15.89 2.52
Individual FE YES YES YES YES YES YES
Year FE YES YES YES YES YES YES
Gender All All All Female Female Female
Notes: The dependent variable, Employment, is a dummy variable that indicates whether the individual is currently employed. Working Hours measures the number of hours
the individual works per week. Log Income represents the total income the individual earns, expressed in logarithmic form. Parent Careit is a dummy variable equal to 1 if
a parent is the primary caregiver for the children. Individual’s age is included as a control variable. Standard errors are clustered at the county level to account for potential
correlations within counties. The estimates are derived using the canonical Two-Way Fixed Effects (TWFE) model. *** p
<0.01, ** p
<0.05, * p
<0.1
40

Table 1.9: Heterogenous Impact of NIP on Adults by Income Level
(1) (2) (3) (4) (5) (6)
VARIABLES Employment Working Hour Log Income Employment Working Hour Log Income
Treatment 0.0543 4.276*** 0.294*** 0.0300 3.748* -0.00708
(0.0362) (0.896) (0.0937) (0.0264) (2.083) (0.0528)
Observations 12,869 12,869 12,869 4,979 4,979 4,979
R-squared 0.115 0.171 0.036 0.015 0.031 0.265
Number of pid 3,938 3,938 3,938 2,430 2,430 2,430
Mean 0.72 12.51 0.95 0.93 39.52 10.23
Individual FE YES YES YES YES YES YES
Year FE YES YES YES YES YES YES
Parent Takecare YES YES YES YES YES YES
Income Level Low Low Low High High High
Notes: The dependent variable, Employment, is a dummy variable that indicates whether the parent is currently employed. Working Hours measures the number of hours
the parent works per week. Log Income represents the total income the parent earns in the past one year, expressed in logarithmic form. Parent’s age is included as a control
variable. Standard errors are clustered at the county level to account for potential correlations within counties. *** p
<0.01, ** p
<0.05, * p
<0.1
41

In Which Occupational Sector Do Parents Experience the Greatest Impact From the
NIP?
By analyzing the occupational categories of parents, I classified their employment into
service and non-service sectors to assess the differential impact of the NIP. The service
sector includes all industries except manufacturing, mining, and agriculture. This classification provides a more detailed understanding of how the program influences labor market
outcomes based on employment sector.
The results from Table 1.10 show that the program has a significantly stronger impact
on those employed in the service sector compared to those in the non-service sector. Parents working in the service sector increase their weekly working hours by 7.27 hours and
experience a notable increase in income. This effect is even more pronounced for women
in the service sector, who increase their weekly working hours by 11.85. Given women’s
average weekly working hours of 15.89, this proves that many mothers are re-entering the
labor market, mainly through service-sector jobs. These findings highlight the program’s
ability to release economic opportunities for parents, particularly mothers, with flexible
and accessible service-sector jobs.
One possible explanation for this greater impact in the service sector is the generally
lower skill requirements and greater accessibility of these jobs. Service sector positions,
such as those in retail, hospitality, and caregiving, often demand fewer specialized qualifications compared to jobs in non-service sectors such as manufacturing and technical
professions. For individuals who have been out of the labor force for an extended period,
re-entering the labor market is easier and more feasible in the service sector.
42

Table 1.10: Heterogenous Impact of NIP on Adults by Service Sector Jobs
(1) (2) (3) (4) (5) (6)
VARIABLES Employment Working Hour Log Income Employment Working Hour Log Income
Treatment -0.00160 3.127** 0.121 -0.0213 3.241*** 0.000768
(0.0326) (1.200) (0.203) (0.0377) (1.118) (0.194)
Treatment*Parent Takecare 0.0787** 1.097 0.0709 0.109*** 1.137 0.142
(0.0336) (1.185) (0.208) (0.0389) (1.106) (0.212)
Treatment*ParentCare*Service 0.0300 7.274*** 1.683*** 0.0499 11.85*** 2.375***
(0.0320) (2.123) (0.266) (0.0418) (2.585) (0.415)
Observations 53,087 53,087 53,087 27,809 27,809 27,809
R-squared 0.083 0.153 0.108 0.078 0.163 0.117
Number of pid 14,724 14,724 14,724 7,637 7,637 7,637
Mean 0.75 19.61 3.39 0.70 15.89 2.52
Individual FE YES YES YES YES YES YES
Year FE YES YES YES YES YES YES
Gender All All All Female Female Female
Notes: The dependent variable, Employment, is a dummy variable that indicates whether the individual is currently employed. Working Hours measures the number of
hours the individual works per week. Log Income represents the total income the individual earns, expressed in logarithmic form. Parent Careit is a dummy variable equal
to 1 if a parent is the primary caregiver for the children. ServiceSectorit is 1 if the parent’s job is in the service sector, including all industries except manufacturing, mining,
and agriculture. Individual’s age is included as a control variable. Standard errors are clustered at the county level to account for potential correlations within counties. The
estimates are derived using the canonical Two-Way Fixed Effects (TWFE) model. *** p
<0.01, ** p
<0.05, * p
<0.1
43

At What Age Does the Treatment Have a Stronger Impact On Parental Labor Outcomes
When Children First Receive It?
The question of whether the impact of the NIP treatment on parents varies depending
on the age at which a child receives it is an important consideration for informing future
policy design. To explore this, dynamic impact graphs were created in Figure 1.4a, 1.4b,
and 1.4c for three distinct age groups: 6-8, 9-12, and 13-15, based on the age at which the
child first enrolled in the NIP. The results suggest the treatment effect is most pronounced
for the parents of the children who receive the first treatment at ages 6-8 years and 9-11
years.
Children aged 6-8 years and 9–11 years typically require substantial levels of parental
care. They can perform basic tasks yet still depend on their parents for supervision and
support. The NIP, by providing free school lunches, relieves these parents from meal
preparation responsibilities and creates additional free time. Thus, this time-saving effect
is especially significant for them, and they are able to reallocate more time to employment
activities.
For parents of children aged 12–15 years, the level of caregiving required is much
lower. Adolescents are more independent enough to manage their schoolwork and daily
needs without much parental intervention. As a result, parents of children in this age group
are not subject to a caregiving burden and already have more flexibility to work. Thus, NIP
benefits may not have a significant enough impact on their time constraint to affect their
employment status.
44

(a) Dynamic Impact of NIP on Adult Employment for Children First Treated at Ages
6–8
(b) Dynamic Impact of NIP on Adult Employment for Children First Treated at Ages
9–11
(c) Dynamic Impact of NIP on Adult Employment for Children First Treated at Ages
12–15
Figure 1.4: Event Study on Adult Outcomes by Child’s Age at the NIP Onset
45

1.6.6 Robustness Check
To address the potential issue of forbidden comparison biasing the TWFE estimation
of the treatment effect, the methodology proposed by Chaisemartin and D’Haultfœuille
(2020) was applied to evaluate the proportion of local treatment effects with negative
weights, which could distort the overall estimates. Amongst the 4,237 local treatment
effects analyzed, only 143 exhibited negative weights, accounting for a cumulative total
of just 0.0073. This low proportion of negative weights suggests that their impact on the
aggregated treatment effect was minimal and unlikely to introduce significant bias. Regarding the children TWFE, 764 out of 2,948 local treatment effects have negative weight,
with cumulative total weight of 0.096. These results provide strong evidence that negative
weights are not a significant source of bias, supporting the validity and reliability of the
treatment effect estimates for both the child and adult samples.
To further validate the robustness of these findings, CSDID proposed in Callaway and
Sant’Anna (2021), and BJS estiamtor proposed in Borusyak et al. (2024), were implemented to construct event study estimators as a complement to the TWFE results. The
CSDID event study figures are attached in the Appendix Section G.1 and BJS estimates
are presented in the Appendix Section G.2 Table A1 and A2 .
The BJS estimation infers the never-treated potential outcome for each treated observation using the predicted value from a TWFE regression, allowing for an individual
treatment effect estimate for every treated unit. The individual treatment effects are then
aggregated to derive summary results. Only units and periods that have not been treated
are used in the analysis. The consistency between these estimates and those from the
46

canonical TWFE model confirms that the NIP induces durable and significant changes in
children development and parental labor supply, particularly for female parents.
1.7 Discussion and Policy Recommendation
NIP is a major safety net policy aimed at enhancing children’s nutrition intake and
health condition in rural China. By addressing nutritional deficits in children and reducing caregiving burdens on parents, the program has the potential to help disadvantaged
children and families to improve both their health and labor market outcomes. Without
considering the cost, these findings support the expansion and continuation of such programs, not only in China but also in other developing countries with the same objective.
Understanding the Mechanism: Income Effect vs. Direct Nutrition Gains
The mechanism behind the program’s effectiveness requires further discussion. The
findings suggest that nutrition alone is not the key driver of improved child health. Instead, the increase in parental income through higher labor participation seems to be the
main channel leading to better child health outcomes. By reducing the time parents spend
on meal preparation, the NIP allows for greater labor force participation, particularly for
mothers. The higher employment rates and longer working hours generate more household income, which translates into better children’s nutrition. These results highlight the
importance of improved household economic conditions in enhancing children’s health
outcomes.
Safety Net Policy Beyond Early Life
47

The empirical results demonstrate that safety net policies like the NIP effectively help
disadvantaged students catch up in physical development, which provides treatment beyond the childhood stage. While many programs focus on early childhood, health benefits
remain significant when the intervention continues throughout the growth years. Starting at age six, children who continue to receive nutritional support show greater increases
in height. These findings provide strong evidence that safety net policies should extend
beyond early childhood.
Work Contingency for Greater Effectiveness of Safety Net Policy
The results suggest that integrating work incentives into safety net policies like the NIP
may be more effective in achieving the objective of supporting disadvantaged families. Potential complementary interventions could include skill training and job referral programs
in collaboration with local state-owned enterprises and private firms. These programs can
equip parents with professional skills and reduce job search frictions. As a result, they
could help parents reenter the labor market more quickly or advance to better positions.
Policy Extensions to Enhance Household Productivity
In Appendix G.3, I extend the model introduced in Section 1.3 to illustrate that, under
certain conditions, enhancing home production technology/efficiency can generate similar
benefits with providing free school meals. When households receive support that boosts
their productivity in daily tasks, they can reallocate time to activities that enhance overall
welfare, such as joining the labor force or investing in children’s education.
For example, subsidizing household appliances like washing machines or stoves could
generate a similar effect with the NIP. In addition, affordable and high-quality childcare
48

can relax caregiving burdens and increase participation in the workforce, especially among
low-income women.
1.8 Conclusion
This study finds that the NIP improves children’s height growth and increases parental
labor force participation, especially among mothers. However, further research is needed.
Future studies should examine the long-term effects of the NIP by tracking children
into adulthood to understand its impact on education, employment, and health in the longer
term. In addition, a detailed field investigation is necessary to understand the format in
which the NIP delivered the nutrition supplement. Only by examining the implementation process and quality can the research identify why free lunch alone is not effectively
improving health, mental well-being, and cognitive ability. Lastly, research should also
explore community-level effects, including the program’s role in local economic growth
and public spending. A comprehensive evaluation of the program’s benefits and costs is
essential to assess its overall impact as a social safety net policy.
49

Chapter 2
The Consequence of China’s Supply-Side Structural
Reform: Impact of Coal Mine Shutdowns on Local
Economic Development
This project examines the impact of large-scale coal mine shutdowns across China during
the period from 2016 to 2019. Leveraging econometric methods of Oaxaca-Blinder regression, instrumental variable, and difference-in-difference, I conduct analysis using data at
county, city and firm levels. The negative labor demand shock from coal mine shutdowns
reduces total local employment, the average wage, and regional GDP. In addition, the general equilibrium effect also decreases local industries’ employment and the price of local
goods, represented by the average housing price. Improved transportation infrastructure
mitigates this negative impact by promoting labor mobility. Large state-owned enterprises
(SOEs) whose headquarters are located in a city with coal mine shutdowns lose a number
of employees, profitability, and total factor productivity (TFP).
Keywords: Coal Mine, Labor Demand, Public Policy
JEL Codes: J08, I38, O13, J0
50

2.1 Introduction
In November 2015, China president Xi Jinping raised the idea of supply-side structural reform for the first time in a national meeting planning for economic development.
China’s economic growth had been continuously slowing down since 2010 and the trend
was expected to continue in the years following 2015. At that time, the economy was characterized as overcapacity. Dozens of industries have to deal with the issue of overstock
and oversupply, including steel, cement, flat glass, coal, etc. The increased production
capacity of those industries has far more exceeded their consumption growth in the past
few years, resulting in a decrease in utilization rate and consequent business downturns
(Woo (2019)). To solve the overcapacity issue, the Chinese government started supplyside structural reform with five objectives: cutting excess capacity, lowering leverage in
high leverage corporations, de-strocking inventories, reducing running costs for firms and
shoring up weak industries (Boulter (2018)). One of the most resolute actions was to close
the dirty and small-capacity coal mines to cut the production level of coal.
This paper intends to examine how the shutdown of coal mines affects a local economy,
in terms of employment, wages, GDP growth, people’s well being, and firm performance.
In particular, as this regulation lowers the local labor demand in the mining industry, to
what extent does this exogenous shock affect local economic development and how large
is the spillover effect on other industries? How does the general equilibrium effect work
in the setting? Is there any mitigating practice to offset such a potential negative impact?
Following the theoretical framework introduced by Moretti (2010) and Moretti (2011),
this study adopts three empirical strategies to provide empirical evidence about the impact
51

of a negative labor demand shock brought on by coal mine shutdowns, including OaxacaBlinder regression, instrumental variables, and difference-in-difference. The negative labor demand shock from coal mine shutdowns reduces local total employment, average
wages, and regional GDP. In addition, the general equilibrium effect also decreases local
industries’ employment and the price of local goods, represented by the average housing
price. Improved transportation infrastructure mitigates this negative impact by promoting
labor mobility, which consequently increases the labor supply elasticity. Publicly traded
SOEs whose headquarters are located in the cities with coal mine shutdowns lose a number of employees, profit, and total factor productivity (TFP), while privately listed firms
thrive.
This paper is related to several strands of literature. First, it contributes to the literature about the empirical evidence of the local labor demand shock on the community. The
pioneering work, Black et al. (2005), reveals the positive effect of coal boom shocks on
the employment of locally traded industry. Greenstone et al. (2010) quantify the agglomeration spillover brought by new industrial plants. Aragon and Rud (2013) exploit the ´
opening of a giant gold mine in Pure and use a difference-in-difference method to show
its impact. They find households living near a gold mine have an improvement in income
and the price of local food and house rents increase due to the general equilibrium effect.
Kline and Moretti (2013) investigate the big push effect of Tennessee Valley Authority on
the local counties, where the direct effect comes from infrastructure improvements and indirect effect stems from agglomeration economies. Different to them, my setting focuses
on a negative shock which spans a large portion of China.
52

Second, this project contributes to the literature about the role of transportation infrastructure. Donaldson (2018) introduces a comprehensive overview of the welfare effect
brought by India’s railroad. Baum-Snow et al. (2017) documents how railroad decentralizes industrial activities in Chinese cities. Lin (2017) focuses solely on high-speed rail in
China and stresses its employment-boosting effect. I add to other publications in focusing on the impact of transportation infrastructure by introducing its role in hedging the
negative labor demand shock.
Third, this project is related to the scarce volume of literature documenting the impact
of China’s supply-side structure reform (SSSR). This paper provides empirical evidence
of the negative impact the policy brought to the local community during the process of
reducing coal mine production. Regarding the impact on the firms, Jiang, Shen, et al.
(2021) argue that SSSR helps firms adjust their debt ratio to the optimal level, while my
results show SOEs sacrifice profit and TFP to attain their political goal under SSSR.
This paper is organized as follows. Section 2.2 provides the analytical framework to
analyze the question. Section 2.3 introduces the data source and some summary statistics.
Section 2.4 explains the three empirical strategies used to yield credible results. Section
2.5 discusses the results. Section 2.6 concludes.
2.2 Theoretical Framework
The shutdown of coal mines creates a negative local labor demand shock. In this paper,
I use the spatial equilibrium model introduced by Moretti (2010) and Moretti (2011) as an
analytical framework, which is a realistic extension of the Rosen-Roback model. (Rosen
(1979), Roback (1982)) In this section, I briefly explain the context of the model and list
53

the predictions we can make from it regarding the impact on local economies. The model
can provide the guidance for me to proceed with the empirical analysis.
The spatial equilibrium model contains two types of goods/industries, traded goods and
nontraded goods. While the price of the traded goods is established by the outside larger
market, the local nontraded goods’ price is endogenous. The homogeneous labor serves
as the most important input, and some traded goods can be used as intermediate goods. In
particular, one typical nontraded good is housing, and other local industries include but are
not limit to restaurants, real estate, construction, retail, etc. On the other hand, the mining
industry and manufacturing industry can be considered as traded industries, as the price of
the associated goods is governed by the equilibrium of the national or even international
market. Both the local labor supply and the housing supply are assumed to be upward
sloping. The elasticity of the labor supply depends on workers’ preferences over the place
and their geographic mobility, while the elasticity of housing supply is determined by the
housing land regulation.
An exogenous shock on one traded industry may have a minor impact on the other
traded and non-traded industries. As an example, let us take a situation where we have a
negative labor demand shock on one traded industry, which directly reduces the number
of employees. As labor is perfectly mobile within the city, the wage of all workers should
decrease because of the shrinkage in total employment. Then, to analyze the impact on
the other traded goods industries, we need to consider three forces. First, the lowered
wage improves the competitiveness of the city, so the employment in other traded goods
industries may increase with firms being attracted to the area. Second, depending on the
magnitude of the local supply chain, the traded goods industry sector, i.e. intermediate
54

goods and services may be harmed because of the weaker demand. Third, a stronger local
agglomeration may exacerbate the negative impact of this industry on the other traded
goods industries. Overall, it is unclear of the direction of this negative labor demand
shock on the other traded industries.
It is clearer to see its effect on the other non-traded goods. As the wages decline, labor
budget constraints become more limited. Thus, the demand for other locally produced
goods drops. However, here we need to think deeper about an offsetting general equilibrium effect. The cheaper labor may induce an increase in the supply of local supply.
This increase in supply may offset partially the effect of demand decrease on the local
goods. In this case, the price of the local goods will still drop because of the negative labor
demand shock, with the magnitude depending on the elasticity of labor supply and local
goods supply. If the local labor supply is very elastic due to a better geographic mobility,
the fall in wage will be restricted and such offsetting general equilibrium effect may not
be noticeable.
In the setting of this paper, I specify the mining industry, manufacturing industry and
primary (agricultural) and secondary (industrial) sectors of the GDP as the traded goods.
I define the primary sector of the GDP to be traded goods because of the excellent transportation infrastructure in China, e.g., agricultural output can be easily shipped across
China. On the other hand, housing, construction, retail, services, and tertiary (the service)
sector of the GDP belongs to the non-traded/local goods/industries. The price of housing
is used to reflect the effect on the price of local goods. Each county or city in China is considered as a local economy with limited inter-regional mobility of the labor force. Based
55

on the model above, I would make the following predictions regarding the impact of coal
mines shutdown that can be tested:
1. Coal mine employment decreases as a direct effect of the shock.
2. The overall employment rate and the average wage both decrease in the locality.
3. The effect on other traded industries’ employment is unclear, including manufacturing, and the primary and secondary sectors of the GDP.
4. The effect on other local industries’ employment is negative, including construction,
retail, services, and the tertiary sector of the GDP
5. The housing price decreases.
6. The magnitude of the impact above depends on the elasticity of the local labor supply. The labor inter-regional mobility determines the labor supply elasticity so that
those places with better labor mobility may suffer from a lower impact from the
negative labor demand shock.
In addition, Notowidigdo (2020) underlines the asymmetric response of the local population regarding different labor demand shocks. The population is more responsive to a
positive labor demand shock than to a negative labor demand shock, while wages, housing
prices and rents do not exhibit such characteristics. Under the setting of negative demand
shock brought by coal mine shutdown, I want to test how the population flow change relative to the wage level and housing price. Moretti (2011) also explains the heterogeneous
effect on the skilled and unskilled labor force. Miners are more likely to be low-education
and low-skilled people. Given that the skilled and unskilled labor forces are not perfect
56

substitutes, the negative demand shock should affect unskilled worker’s income more.
Unfortunately, the dataset in this study does not provide sufficient information to test this
empirically.
The analysis above provides the rationale for my empirical analysis and the selected
outcome variables introduced in the next section. In addition, I also investigate the impact
on citizens’ well being and firms performance under the shock of coal mine shutdowns. I
introduce the detailed dataset in the next section.
2.3 Data and Summary Statistics
The data set is composed of four segments: 1). Magnitude of coal production closure measured per 10,000 tons in each county, 2). County-level outcomes, 3). City-level
outcomes, and 4). Listed firms’ data. I also elaborate on the source and some summary
statistics in this section.
The main dataset, coal production closures in each county, is produced by and was purchased from Fenwei Digital Information Technology Co. Ltd., an information consulting
firm in China, which specializes in data collection and consulting services for the mining
industry. The dataset contains information on coal mine closed in 839 counties during
the period from 2016 to 2019, which the firm claims as the most complete and exhaustive
dataset for China coal mine closures for the period. The reason for each shutdown is not
recorded and could have varied from forced closure due to government regulation from
SSSR, to purely business decisions made by the owners in order to exit the market. The
lack of data on closure reasoning does not weaken the conclusion that will made in this
57

paper, because despite the reasons for any coal mine closure, each coal mine shutdown
would have led to a negative labor demand shock.
Figure 2.1 shows the trend in closures from 2016 to 2019. The unit of the data is
10,000 tons coal production capacity per year. As expected, 2016 had the largest magnitude of coal production closures amounting to 330 million tons, as it was the first year
after the Chinese government officially implemented its SSSR. This amount is equal to
approximately 5.8% of total coal production capacity in China for 2015, which was about
5.7 billion tons (Hao et al. (2019)). In 2017, the reduction of coal production capacity
decreased to about 170 million tons. Based on the county-level data available, Figure 2.2
provides information on the ranking of total closure across provinces from 2016 to 2019.
Shanxi province, ranked no.1 in all the provinces, closed coal mines representing more
than 110 million tons coal production capacity. Guizhou province and Henan provinces
ranked no.2 and no.3 respectively, with a reduction in capacity of 78 million tons and 62
million tons. I have represented the distribution of the total capacity closure of each county
from 2016 to 2019 in Figure 2.3. We can tell that the counties with the highest capacity
closure are clustered in the southwest, central and northern parts of China.
Second, county-level outcome data is provided by the China County-Level Economic
Research Database, whose primary source is China County Statistical Yearbook. The
variables of interest include population, GDP, agricultural (primary) sector GDP, industrial
(secondary) sector GDP, service (tertiary) sector GDP, GDP of manufacturing industry,
income, saving, fiscal revenue, and fixed asset value. County-level data does not provide
detailed information on employment.
58

Figure 2.1: Coal Mine Closure Trend in China (Unit: 10,000 ton)
The most crucial drawback of the county-level data is that it contains a significant
amount of missing values for the variables except for GDP, population, and fiscal revenue.
Some variables have reported missing data of up to 30%. To solve this issue, I imputed
the missing values through the regressions using GDP, population and fiscal revenue as
dependent variables, instead of dropping them. While the dataset has become more complete, the imputation may have created data points that are less reliable and may result in
endogenous problems in the regression. For this reason, a more complete city-level data
was used to complement the analysis.
As mentioned above, data for city-level outcomes data has been utilized to collaborate
with the county-level results. City-level outcome data represents a more complete set with
most of the desired information we need to test the predictions, including average wage,
total employment, agricultural (primary) sector employment, industrial (secondary) sector
59

Figure 2.2: Provincial Ranking of Closed Coal Mine Capacity
employment, service (tertiary) sector employment, and employment in different industries, including mining, retail, construction, and manufacturing. I manually aggregated the
employment of financial services, restaurant, communication service, rental, education,
public service sectors into service employment. The data also provides information about
fiscal revenue, fiscal expenditure, number of uninsured unemployed people. These variables allowed me to investigate how the governments adjust their fiscal budgets and if they
have provided an extra safety net for their citizens in the event of a reduction in mining
output, which has decrease employment. In addition, I collected data on PM 2.5 in each
city for the period from 2013 to 2019 from the Ministry of Ecology and Environment. A
higher value than the PM 2.5 indicator signals a less desired air quality. Average housing
price per square meter was collected from the Macroeconomics and Real Estate Database
at the National Information Center. I also obtained data on the high-speed rail station of
60

Figure 2.3: Distribution of Coal Mine Closures Across Counties
each city from Xi Gao, who manually complied data using public information from the
Ministry of Railways.
The 1912 mining tax payment data, introduced as an instrumental variable in the empirical strategy section, is shared by the authors of He and Mou (2020). They manually
collected the data from historical books.
Lastly, company-level data was derived from the China Stock Market & Accounting
Research (CSMAR) database, which provides the most comprehensive information on
China’s public traded companies and their stocks. The variables of interest include total
asset, current debt ratio, percentage of debt from banks, return on asset (ROA), return on
equity (ROE), sales, number of employees, total payment to employees, fixed asset value,
profit, and total factor productivity. Current debt ratio is defined as total debt divided by total asset. The percentage of debt from banks can be interpreted as a measure of support by
61

the government to the firm. All the major banks in China are controlled by the government
and their credit expansions are usually directed to back certain companies. For example,
in the China 2009-2010 stimulus plan, the government showed favoritism towards SOEs
by giving disproportional credit to them (Cong et al. (2019)). Total factor productivity is
calculated using the Levinshon-Petrin method (Levinsohn and Petrin (2003)). To derive it
with STATA command “prodest”, I used cash payment to the employee as a free variable,
net fixed assets as a capital variable, operating cost minus depreciation and amortization
and cash payment to the employee as intermediate input.
I separately merged the coal mine closure data with county-level data, city-level data,
and company-level data. In terms of different empirical strategies, the data is reshaped to
a wide or long format. In particular, the difference-in-difference analysis uses the citylevel panel data from 2013 to 2018, while the Oaxaca-Blinder regression and instrumental
variable rely on wide format county- and city-level data..
2.4 Empirical Strategy
To identify the conclusive impact of the closed coal mines on the local economy, I
adopted three empirical strategies: Oaxaca-Blinder regression, OLS and instrumental variable regressions, a difference-in-difference (DD) and event study analysis. Each of the
strategies serves a different purpose and comes with certain drawbacks, which I have laid
out in detail in this section. Overall, I don’t believe the three strategies suffer from the
identical selection bias so I will draw conclusions that resonate with the results from all
three strategies.
62

The main challenge of evaluating the effect of a place-based policy is to identify or
construct a “similar” comparison group. The method of Oaxaca-Blinder regression is to
construct the counterfactual outcome with the weights assigned to each county in the untreated area, and compare the counterfactual outcome with the mean of the treated group.
This method has been proved to have a dual property as a propensity score reweighting estimator, by Kline (2011). Then, in order to circumvent data limitation and test the spillover
effect on the traded and non-treaded industries, I leveraged on another OLS and an instrumental variable to conduct analysis with a more comprehensive city-level dataset. Finally,
following Aragon and Rud (2013), I exploited the variation of the timing of large-scale ´
shut-down in each place and the reduction in the magnitude of the impact as the distance
to the coal mine increases to do a DD analysis. In addition, to capturing the dynamic effect of the policy, I conducted event analysis so that we could clearly see how the impact
evolved over time. Within the difference-in-difference framework, using high-speed rail
as a proxy for having better labor mobility, I tested if the locations which had better labor
mobility were able to resist the negative labor demand shock. I then attempted to identify
the shock’s impact on the listed firms. In this study, because the policy regulation had
only been executed a short while ago and has continued to be implemented, I could only
identify the contemporaneous effect, instead of any long-term effect.
2.4.1 Oaxaca-Blinder Regression (O-B)
In order to capture the county-level economic impact, I used the Oaxaca-Blinder regression to compare the counterfactual outcome derived in the following regression with
the outcome of the treated counties. This exercise was inspired by Kline and Moretti
63

(2013), which adopted this strategy to investigate the impact of the Tennessee Valley Authority on the local community.
∆Yc = α +βXc +εc (2.1)
∆Yc indicates the annual growth rate of the interested outcome variables from the period
between 2015 and 2018 in city c, calculated as (log(Y2018)−log(Y2015))/3. The outcomes
includes GDP and GDP sectors, population, people’s income and savings, local fiscal
revenue and fixed-asset value. For the placebo test, ∆Yi denotes the yearly average change
between the period from 2012 to 2014 before the treatment. The coefficients ˆβ is used
to predict the counterfactual mean outcome of the treated group. X contains a rich set
of 2014 characteristics, including population, population squared, GDP, GDP of primary
industry, GDP of secondary industry, GDP from territory industry, GDP of manufacturing
industry, fixed-asset value, fiscal revenue, fix asset investment, residential saving and total
agricultural output. These covariates control for the different characteristics before the
treatment began.
Considering the small number of coal mine closures may not create a noticeable labor
demand shock to the local community, I defined the county with a cumulative production
closure from 2016 to 2019 above 1,260,000 tons as a treated county, which is the median of
cumulative capacity closure over all the counties that have any capacity reduction. Figure
2.4 displays the distribution of the treated counties. In order to simplify the results, I did
not include the counties with coal mine capacity closure between 10,000 and 1,260,000
tons in the untreated sample. That is, the counties exposed to negative but mild labor
demand shock do not appear in the treated group nor untreated group. As the economic
64

activities are always spatially correlated, I have calculated two sets of standard errors to
corroborate my results: clustering at the city level and clustering at the provincial level.
Figure 2.4: Geographic Distribution of Treated Counties
Oaxaca-Blinder regression has an advantage in estimating the average treatment effect
when there is a heterogeneous treatment effect. The treatment heterogeneity in this case is
guaranteed as the treated counties span different regions across China and do not share the
same pretreatment economic characteristics. The regression assigns a weight to each of
the counties in the control group as 1560 counties do not have any coal mine closures. It
assigned a higher weight to the counties that showed more similar economic characteristics
to the treated group. Kline (2011) has proven the dual property of O-B regression with
a propensity score reweighting estimator. Given that the controls include a rich set of
economic characteristics before the beginning of the treatment, it is possible to believe
that a counterfactual outcome follows a pre-existing developing trend.
65

In order to further improve the credibility of the Oaxaca-Blinder regression, I eliminated the counties with low propensity to treatment from the sample. I used the logistic
regression to regress the treatment dummy on the controls and predict the propensity to
treatment for each untreated county. I then eliminated the counties with a propensity to
treatment below the 25 percentile. In total, 377 counties were dropped from the sample.
In Table 2.1, I list the mean of 2014 characteristics and the mean of the change of interested outcomes between 2012 and 2015 for the three groups of counties: treated counties,
full untreated counties and trimmed untreated counties. We can tell that the mean of the
trimmed untreated sample becomes much closer to the treated counties’.
The weight of the trimmed untreated counties has been drawn in Figure 2.5. The darker
color signals a stronger comparability between the untreated counties and the treated counties. Most of the counties with darker colors are found in the central and northern side of
China, which is indeed consistent with the economic trait of the treated counties. The
treated counties are not very wealthy so they are not akin to the developed counties on
the east coast of China. On the other hand, the counties in the western side of China are
even more undeveloped than the treated cities. Thus, they are dropped because of the
low propensity to treatment. Actually, these more similar counties surround the treated
counties, and this phenomenon implies the validity of the parallel trend assumption for my
later difference-in-difference analysis. The weight possess a positive relationship with the
county’s odds of being treated. It can be negative in some counties that are substantially
different than the treated group.
66

Table 2.1: Summary Statistics of County Level Data
(1) (2) (3)
Treated Untreated-full Untreated-Trimmed
mean mean mean
Panel A: 2014 pretreatment characteristics
Log GDP 12.697 10.772 12.791
Log fiscal revenue 10.921 9.099 11.342
Log population 3.658 3.500 3.700
Log population sq 14.052 13.208 14.255
Log fixed asset 4.790 4.730 4.749
Log GDP manu 13.175 12.780 12.997
Log saving 4.419 4.224 4.491
Log GDP pri 11.964 12.009 12.010
Log GDP sec 13.494 13.210 13.352
Log GDP ter 12.781 12.137 12.897
Log agriculture 12.499 12.547 12.551
Panel B: Changes 2012 - 2015
GDP 0.052 0.078 0.076
Population -0.001 -0.000 0.001
Fiscal Rev 0.018 0.101 0.100
Per cap GDP 0.045 0.053 0.055
GDP pri 0.048 0.067 0.041
GDP Sec 0.049 0.141 0.070
GDP Ter 0.129 0.133 0.139
GDP Manu 0.032 0.080 0.047
Income 0.050 0.043 0.049
Saving 0.118 0.110 0.131
Fixed Asset 0.161 0.146 0.140
# of County 220 1508 1131
# of Province 21 28 24
Notes: The data in this summary statistics is used for O-B regression. Except for GDP, population, and fiscal Revenue, other variables
may suffer from a major missing value issue. I imputed the missing data using a regression with GDP, population, and fiscal revenue as
dependent variables. Treated counties include those with cumulative capacity closure above 1,260,000 tons. Untreated counties does
not include the counties with cumulative capacity closure between 1 and 1,260,000. Trimmed data have eliminated the counties with a
low propensity to treatment.
67

Figure 2.5: Distribution of Oaxaca Weights Across Untreated Counties
Placebo Test
In order to test the credibility of the controls in the O-B regression, I carried out a
placebo test to see if treated counties and untreated counties had differential trends in the
relevant outcomes between the year 2012 and 2015, before the government implemented
the supply-side structural reform to aggressively shut down the dirty coal mines. Ideally,
no selection bias means no different trends in the pretreatment period after conditioning
on the controls.
The results are shown in Table 2.2. We can see that while the some OLS estimates
have significant coefficients, none of the O-B coefficients are significant when the standard
error is clustered at the province level. This directly proves the effectiveness of the controls
chosen in capturing the selectivity biases. In particular, we see a significantly slower GDP
68

growth in the treated counties from the OLS results. However, after controlling for the
rich characteristics, the coefficient becomes smaller and non-significant.
There are certain caveats in using O-B regression. Even though it is better to allow
for the heterogeneous treatment effect than the standard OLS regression, I would still not
be able to control for any unobservable differences between the treated counties and untreated counties. In addition, one must remember that there are other concurring policies,
especially numerous economy-boosting programs implemented by the government. The
estimates are to be interpreted as the treatment effect of the closed coal mines on the local
economy, taking into account other equilibrium effects of the simultaneous policies. Finally, as mentioned in the Data section, the missing data issue is especially severe at the
county level. As a solution, I imputed all the missing data from the available characteristics. This practice may generate severe endogenous bias. Thus, I decided to use other
empirical strategies and leverage a more thorough and comprehensive city-level data set
to complement my analysis.
2.4.2 OLS and Instrumental Variable
I want to conduct further OLS regression and leverage on an instrumental variable to
do the analysis with a richer city-level dataset. Within this setting, I leveraged on the
magnitude of the shutdown capacity. I did not need to define the treated cities and thus I
kept all the cities in my sample, instead of dropping those cities exposed to mild shocks,
as I did in the O-B regression. By doing so, we can compare the economic performance
across the whole country. There are two benefits for the city-level dataset. First, city-level
data is more complete than county-level data. The level of missing data at the city-level
data is much lower so, when doing the regression, I discard those observations with the
69

Table 2.2: Oaxaca-Blinder Regression Placebo Test
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
GDP Population Fiscal Rev Per cap GDP GDP pri GDP Sec GDP Ter GDP Manu Income Saving Fixed Asset
Panel A: OLS with standard errors clustered in province
Treated -0.024** -0.001 -0.080** -0.011 0.007 -0.017 -0.009 -0.014 0.002 -0.017 0.012
(-2.45) (-0.06) (-2.12) (-1.18) (1.02) (-0.84) (-0.72) (-0.71) (0.25) (-1.67) (1.08)
Constant 0.076*** -0.000 0.099*** 0.055*** 0.041*** 0.068*** 0.136*** 0.048*** 0.048*** 0.129*** 0.141***
(12.44) (-0.06) (4.18) (4.19) (6.01) (4.85) (14.47) (2.94) (5.54) (31.42) (10.11)
Cluster prov prov prov prov prov prov prov prov prov prov prov
Observations 1219 1246 1251 1349 1350 1351 1299 1328 1351 1302 1351
Panel B: Oaxaca-Blinder regression with standard errors clustered in city
Treated -0.019 -0.002 -0.083*** -0.012 0.007 -0.029 -0.006 -0.025 0.004 -0.015 0.012
(-1.31) (-0.37) (-3.19) (-1.08) (1.23) (-1.54) (-0.47) (-1.14) (0.60) (-1.46) (1.03)
Cluster city city city city city city city city city city city
Observations 1199 1222 1235 1325 1326 1327 1278 1308 1327 1280 1327
Panel C: Oaxaca-Blinder regression with standard errors clustered in province
Treated -0.020 -0.001 -0.083 -0.010 0.006 -0.025 -0.006 -0.021 0.005 -0.013 0.014
(-0.79) (-0.10) (-1.35) (-0.49) (0.83) (-0.72) (-0.45) (-0.55) (0.36) (-1.13) (0.79)
Cluster prov prov prov prov prov prov prov prov prov prov prov
Observations 1219 1246 1251 1349 1350 1351 1299 1328 1351 1302 1351
t statistics in parentheses
* p < 0.10, ** p < 0.05, *** p < 0.01
Notes: This table is served as placebo test for Oaxaca-Blinder regression. The outcome variables are the growth rate from 2012 to 2014. No significant coefficient proves the
validity of the controls.
70

missing values, instead of imputing them. Without imputing the observations, the results
will be more realistic and will not create concerns relating to endogeneity. Second, citylevel data contains more variables, allowing me to capture the effect of the coal mine
closure on the trend of the traded and non-traded industries I defined in the Theoretical
Framework section. In order to see how the closure of coal mine affects the city where it
is located, I needed to run the following regression.
∆Yc = βL Capacity closurec +γL Popc2013 +εc (2.2)
∆Yc denotes the change of the variable of interest in city c from 2015 to 2018, e.g. log(yc2018)−
log(yc2015). L Capacity closurec is the log of the total production capacity closure in city
c over the years. L Pop2013, logged population in 2013, is the control for the pre-event
condition.
Obviously, this regression suffers from the omitted variable bias and reverse causality.
In addition to the population level, numerous elements associated with coal mine closure
can contribute to the local economic outcomes, such as the demographic composition,
geographic location, etc. On the other hand, those cities experiencing an economic boom,
such as receiving extra infrastructure investment, which is often favored by the Chinese
government, may create alternative, attractive business opportunities. Thus, the owners of
a coal mine may shut down the mine in order to pursue a financially better opportunity.
Thus, the economic boom may induce a closure of the coal mines, resulting in a reverse
causality.
In order to solve this endogeneity issue, I have adopted the instrumental variable from
He and Mou (2020). The mining tax payment of each city in 1912 was shared with me
71

from the authors to serve as an instrumental variable from a historical perspective. I ran
the first stage regression as:
L Capacityˆ closurec = θL 1912 Mine Taxc +λL Popc2013 +εc (2.3)
A valid instrumental variable should satisfy both the exclusion restriction and the relevance restriction. The 1912 mining tax payment has to affect the outcome growth only
through the coal mine closure. It has been documented that between 1912 and 1928, the
total revenue collected from the mining activities was virtually ignorable in relation to the
national GDP (He and Mou (2020)). That means that the initial tax payments did not lay
the foundation of economic prosperity or economic recession for any city today. Even
if one argues that valuable resources may give an economic advantage to city, this initial
advantage has been eliminated by taking the difference in ∆Yc.
On the other hand, the mining activities in 1912 largely depended on the spatial distribution of easily available mining resources, due to the technical limitations. Even though
this mining resources is not limited to the coal mines, it contains the places with an significantly valuable amount of coal mine resources. Only those places with high levels of coal
resources were likely subject to closure. I have represented the distribution of the coal
mine closures in Figure 2.6 and the distribution of the 1912 mining tax payment across
cities in Figure 2.7. It is evident that both the distributions have a similar pattern: a darker
color in the southwest and northeast part of China. It is undeniable that data for the closure of coal mines has a larger geographic coverage than that of the mining tax payment in
1912. However, their similarity is also obvious. More straightforward, the high F-statistics
72

in the results table, well above the rule of thumb of F-statistic of 10, confirms no weak IV
problem in the setting (Stock and Yogo (2002)).
Figure 2.6: Distribution of Coal Capacity Closure Across Cities
Table 2.3 provides the mean of interest outcomes in 2014. It is obvious that untreated
cities are more wealthy, shown by higher GDP and housing prices, while the treated cities
have greater levels of mining employment.
73

Figure 2.7: Distribution of Mining Tax Payments Across Cities in 1912
74

Table 2.3: Summary Statistics of City-Level Data
(1) (2)
Treated City Untreated City
mean mean
Log cumulative closure 5.3
Log 1912 mining payment 2.6
Log population 2013 5.8 5.9
Log GDP 16.3 16.7
GDP per Capita 43099.0 55831.6
Log GDP pri 18.7 18.8
Log GDP sec 20.2 20.5
Log GDP ter 19.9 20.3
House price 4217.1 5995.5
Log insured unemp 12.4 12.7
Log fiscal revene 12.8 13.5
Log fiscal exp 13.5 14.0
Average wage 46056.9 50570.9
Log employment 3.5 3.8
Log emp pri 3.3 2.9
Log emp sec 7.3 7.7
Log emp ter 7.5 7.6
Mining emp 29085.6 7729.3
Retail emp 19627.2 31935.8
Construction emp 65163.2 128192.8
Manufacture emp 103613.5 240732.8
Service emp 188771.8 246409.2
# of City 159 135
# of Province 23 25
Notes: The data in this summary statistics is used for 2SLS regression. It uses the full data set of coal capacity closure. The statistics
uses the data for 2014, except for cumulative closures, 1912 mininig payment and the 2013 population.
The OLS and instrumental variable regression has the advantage of telling us the precise impact of each shutdown. However, in the perspective of whole country, inter-regional
labor mobility should be quite high and this exercise may not be ideal when looking to
show the impact of a local labor demand shock. Thus, I want to introduce my third empirical strategy, difference-in-difference, relying on the limited inter-regional labor mobility
between a city and its neighboring cities.
75

2.4.3 Difference-in-difference (DD)
Finally, to capture the spatial equilibrium effect of a closed coal mine on a city where
it is located, I conducted a difference-in-difference analysis comparing the performance
of the treated cities and their neighboring cities. This method was inspired by Aragon´
and Rud (2013). Then, to show the dynamic effect, I presented graphs of event studies.
To study the effect of negative labor demand shock under different labor supply elasticity
(Prediction 6), I ran the regression about the heterogeneous treatment effect on the cities
with a high-speed rail station/connection and those without. Finally, I used the DD method
to check the impact on the listed firms whose headquarters were located in the treated city.
Baseline Difference in Difference
To identify the treated group, I first defined two dummies regarding the treated status.
The cities with capacity closure in any year between 2016 and 2019 above 900,000 tons
has a treated city dummy equal to one, which is the median of capacity closure over
the cities shut down which occurred across the years. The rationale is the same as the
treated dummy in the O-B setting: only a comparatively larger size of demand shock
can have a noticeable impact on the local community. Defining the first year the city
shutdown more than 900,000 tons capacity as its treatment-starting year. Then I created a
treated yearc,t dummy equal to one for any year after and at the starting year of treatment
for city c. Not all treated cities started to close large coal capacity in 2016. By multiplying
the Treated city and Treated year, I derived the Treated dummy for the difference-indifference analysis. After establishing the treated group, I used ArcGIS to select their
neighboring cities, which I defined as those cities which shared any part of a border with
76

the treated cities, and took them as the control group. Figure 2.8 has shown the distribution
of the treated cities and their neighboring cities.
Figure 2.8: Distribution of Treated and Neighboring Cities
I wanted to compare the cities which experienced a large number of coal mine closures with the cities around them. I leveraged two variances to conduct the difference-indifference analysis, the timing of mine capacity closure and the distance of each city to the
closed mines. Each city is considered as a local economy. The underlying assumption is
that the neighboring cities share similar developing trends with the treated cities, without
the effects of a coal mines closure. In addition, the labor forces in each of the treated
77

cities and their neighbors were not perfectly mobile. Otherwise, I would not observe any
difference in the trend. I ran the DiD regression as
Yc,t = αc +β1Treatedc,t +ηt +εc,t (2.4)
Yc,t
includes a number of interest economic outcomes in the city c at year t, the same with
the outcomes variables in the OLS and IV regression. Treatedc,t
is the treaded dummy
introduced above, equal to 1 if the city c is treated at year t. City and year fixed effect
are included as αc and ηt
. In order to detect if there was any selection problem, I also
ran a set of regressions with a control of logged 2013 population, L Pop2013, to inspect
how coefficients changed. To account for the characteristics of neighboring cities, I also
estimated the regression including XNeighborc,t
, a vector of four variables capturing the
neighboring cities’ GDP, average wage, PM2.5 concentration, and average housing price
for each city i at time t.
Table 2.4 provides the outcomes mean in 2014 for both the treated cities and neighboring cities. We can tell that even though the difference between these two groups is smaller
than the difference of the groups under the IV setting in Table 2.3, the treated cities still
had greater involvement in the mining industry.
In addition, to exploit the magnitude of the effect of the closure on a local economy, I
further conducted the regression:
Yc,t = αc +β1(Treatedc,t ∗ Mt,c) +ηt +εc,t (2.5)
78

Table 2.4: Summary Statistics of Treated Cities and Neighboring Cities in 2014
(1) (2)
Treated city Neighboring city
mean mean
GDP 16.4 16.4
GDP per cap 46318.3 44828.9
GDP Pri 18.6 18.9
GDP Sec 20.3 20.3
GDP Ter 20.0 20.0
House Price 4363.5 4561.3
Insured Unemp 341174.1 381040.5
Pub Exp 13.5 13.7
Pub Rev 12.9 13.0
PM 2.5 43.4 47.6
Population 5.8 5.9
Avg Wage 47069.4 46769.3
Emp 3.6 3.6
Emp Pri 3.1 3.1
Emp Sec 7.4 7.3
Emp Ter 7.5 7.5
Emp Mine 40862.0 9838.2
Emp Retail 19091.3 23968.0
Emp Consc 61677.1 91062.1
Emp Manufacturing 107394.3 128330.1
Emp Service 162556.9 182477.5
# of City 98 116
# of Province 23 24
Notes: The distribution of treated cities and neighboring cities is drawn in figure 2.8. Treated cities are the ones with at least one-year
capacity closure above 900,000 tons between 2016 and 2019. Some variables are transformed to log, including GDP, GDP primary, GDP
secondary, GDP tertiary, public expenditure, public revenue, population, employment, employment primary, employment secondary,
employment tertiary. Other employment data has the unit of person.
79

Mt,c is the measure of the mine closure activities. I used the cumulative production capacity closure until year t as a proxy to capture a cumulative effect of the closure and assumed
the impact takes time to have an influence on the local economy and to spread to other
industries.
To illustrate the validity of the parallel trend assumption, I represented the conditional
mean of the treated cities and that of their neighboring cities for four representative variables in Figure 2.9, including average housing price, service employment, average wage,
and mining employment. We can see that up to 2015, one year before the most common
treatment starting year of 2016, the treated and the control groups share a similar trend
and their difference has increased after that. However, there are possibly some unobserved
time-varying elements to affect the cities and their neighboring cities in differential ways.
In that case, the validity of my difference-in-difference strategy is challenged.
Event Study
In order to show the dynamic effect of the mine closures, I have conducted an yearly
event analysis, focusing on only the treated group, in the following regression.
Yc,t = αc +
3
∑
τ=−3
βτyearτ +ηt +εc,t (2.6)
yearτ denotes year dummies if year = yearτ
. βτ gives the impact of the leading or lagging year τ to the treatment year on the city. Similar to difference-in-difference analysis,
by including the fixed effects, the regression relies on the selection of characteristics that
don’t change over time. However, an event study does not take into account any systematic
change around the treatment time. The standard errors are clustered at province level.
80

Figure 2.9: Parallel Trends in Outcomes: Treated Cities vs. Neighboring Cities
How Does Variation in Transportation Infrastructure Affect the Impact?
Next, to test if a higher labor supply elasticity will mitigate the negative labor demand
shock, I intended to use cities which possessed at least one high-speed rail (HSR) station
as a proxy for higher labor mobility, and thus higher labor supply elasticity. In addition
to being beneficial to the knowledge diffusion and idea spillovers, the high speed rail
effectively increases market connectivity and labor agglomeration (Dong et al. (2020),
Chong et al. (2019), Jiang and Lin (2018)). The fact that a city has built an HSR station
81

signals the city’s status as a regional transportation center. In order to test its effect, I ran
the following regression:
Yc,t = αc +β1Treatedc,t +β2Treatedc,t ∗HighSpeedRailc,t +ηt +εc,t (2.7)
HighSpeedRailc,t refers to a dummy variable, equal to 1 if city c has at least one HSR
station at time t. The coefficient β2 gives us the effect of coal mine shutdown on cities
with higher labor elasticity relative to the ones with lower labor elasticity.
Impact on the Listed Companies
Finally, I wanted to look at how the government regulations about the coal mine closure
affects a listed firm located in a treated city. I wanted to see if such regulation affects
their operations, profit level, and total productivity factor. This impact can be the channel
through which the government can affect the local economy or the consequence of those
regulations. More importantly, I wanted to test the heterogeneous effect on the firms given
their shareholding structure. I ran the regression as:
Yi,c,t = αi +β1Treatedc,t +β1Treatedc,t ∗ SOEi
+β1Treatedc,t ∗ SOEi ∗ Miningi +ηt +εc,t
(2.8)
Yi,c,t contains 11 indicators/performance of firm i, whose headquarters are located in
city c in year t. Treatedc,t
is same defined in the baseline DD. SOEc equals to 1 if the the
firm is classified as an SOE. Miningi equals to 1 if the firm belongs to the mining industry,
however, not limited to coal mining. Firm and year fixed effects are included as αi and ηt
82

to take care the time invariant attributes. Standard errors are clustered at province level,
taking into consideration the spatial correlation.
We need to interpret the results cautiously and keep in mind three caveats. First, the
firms are limited to those already listed, which means they are more sizable than the average local city-based firms. Thus, the results are not representative of the local business
environment. Second, as the regulation is aimed at closing the smaller and dirty coal
mines, these listed firms usually own large and efficient mines and thus may not be directly affected. Third, the listed firms usually have lots of business outside the city where
their headquarters are located. For the mining industry, listed firms sometimes possess
mines across the country. In this case, the local labor demand shock does not necessarily
affect their performance.
2.5 Results
In this section, I show and discuss the empirical results in the sequence as follows:
1). Oaxaca-Blinder regression with county-level data, 2). OLS and 2SLS regression with
a richer city-level dataset, and 3). An difference-in-difference and event study analysis
using the sample of the treated cities and their neighboring cities. The idea is to start
from the most detailed geographic level, counties, to investigate the effect of the coal mine
closures on the overall economic environment, then escalate to a bigger geographic level,
cities, to leverage on their more available outcome variables and understand the impact on
employment, prices, average wage and other well-being indicators.
83

Then, I dive deeper into a DD analysis by 1). Exploiting the magnitude of production
capacity abatement, 2). Showing dynamic effect with event study, 3). Examining the role
of transportation infrastructure, and 4). Pointing out the impact on the public firms.
2.5.1 Impact Of Coal Mine Closure On Counties, O-B Regression
Results
In Table 2.5, I present the impact of coal mine closures on the treated counties by OLS
and Oaxaca-Blinder regression, with the OLS results having standard errors clustered at
province level in Panel A, OB results with standard errors clustered at city level in Panel
B, and OB results with standard errors clustered at province level in Panel C. From the
OLS results, only the coefficient on fixed assets demonstrates statistical significance. Even
though it makes sense that the closure of coal mines decreases the amount of fixed asset
in a county because of the liquidation of valuable equipment and properties, undoubtedly
this coefficient is endogenously biased and can not be trusted. We cannot establish any
other observable impact from the OLS results.
Panel B and Panel C of Table 2.5 show the O-B coefficients with different standard
error clusters. The comparison between the OLS and O-B coefficients demonstrates the
importance of the control of 2013 economic indicators. Their results are consistent and
indicate an underperformance in economic development and an outflow in population.
The significance in the coefficient to the outflow in population needs to be interpreted
cautiously, because it indicates a moderate mobility between the counties and other places.
As discussed in the Theoretical Framework section, greater labor mobility will mitigate the
local demand shock. The greater the labor mobility, the less effect on wages we can see.
84

Table 2.5: O-B Regression: Impact Of Coal Mine Closure On Counties
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
GDP Population Fiscal Rev Per cap GDP GDP pri GDP Sec GDP Ter GDP Manu Income Saving Fixed Asset
Panel A: OLS with standard errors clustered in province
Treated -0.002 -0.015 0.002 -0.010 0.014 -0.026 -0.024 -0.025 0.003 -0.003 -0.031**
(-0.35) (-1.58) (0.08) (-0.60) (1.28) (-1.37) (-1.37) (-0.93) (0.52) (-0.25) (-2.32)
Constant 0.075*** 0.013 0.148*** 0.059*** 0.027** 0.086*** 0.182*** 0.072** 0.052*** 0.177*** 0.147***
(13.39) (1.61) (3.10) (3.71) (2.48) (3.63) (4.52) (2.76) (6.62) (5.74) (7.31)
Cluster prov prov prov prov prov prov prov prov prov prov prov
Observations 1043 1009 1304 1345 1346 1346 1320 1292 1347 1339 1346
Panel B: Oaxaca-Blinder regression with standard errors clustered in city
Treated -0.033** -0.013*** -0.015 -0.006 0.012 -0.028 -0.049* -0.027 0.004 -0.026 -0.026
(-2.47) (-2.74) (-0.65) (-0.36) (0.94) (-1.26) (-1.81) (-1.02) (0.89) (-1.11) (-1.53)
Cluster city city city city city city city city city city city
Observations 1022 985 1286 1321 1322 1322 1298 1276 1323 1316 1322
Panel C: Oaxaca-Blinder regression with standard errors clustered in province
Treated -0.032* -0.015* -0.016 -0.007 0.011 -0.028 -0.051 -0.028 0.004 -0.027 -0.027
(-1.95) (-1.77) (-0.28) (-0.23) (0.50) (-0.82) (-0.97) (-0.67) (0.39) (-0.65) (-1.29)
Cluster prov prov prov prov prov prov prov prov prov prov prov
Observations 1043 1009 1304 1345 1346 1346 1320 1292 1347 1339 1346
t statistics in parentheses
* p
<
0.10, **
p
<
0.05, ***
p
<
0.01
Notes: The outcome variables are the growth rate from 2015 to 2018. Oaxaca-Blinder regression construct the counterfactual mean by assigning different weight to the
untreated cities. The distribution of weights is shown in figure 2.5. Untreated counties with low propensity to treatment have been dropped.
85

In addition, the coefficient on the change of service sector GDP is statistically negative,
which implies that the coal mine closure may have a negative spillover effect on the service
sector at the county level. It should also be noted that the coefficient on fixed asset is no
longer significant.
Even though the results document a strong negative impact on the population and GDP
of the counties with coal mine closures, the data limitation prevents me from analyzing the
effect on employment, wages and local prices. Thus, I aggregated the coal capacity closure
data to a city level and leverage on richer city-level data to conduct my analysis.
2.5.2 Impact of a Coal Mine Closure on Cities, IV Regression Results
In this section, I present the results of the OLS, reduced form, and 2SLS in Table 2.6,
2.7 and 2.8. In these regressions, I use the full coal mine closure dataset.
The OLS regression demonstrates a strong correlation between the variables of interest
and the closure of coal mines, controlling for the pretreatment 2013 population. Not only
has local employment dropped in multiple dimension, such as in the mining and services
sectors, but also the local housing price has dropped. Obviously, this OLS suffers from
omitted variable bias and reversal causality so I intend to use the instrument variable, the
1912 mining tax payment, to solve the endogeneity problem.
86

The IV regression results reinforce the negative impact of a coal mine closure on a local
economy, including the growth of population, employment, and GDP per capita. Consistent with the results from the OB regression, conditioning on the pretreatment 2013 population, the cities experiencing a large scale coal mine closure witness a population outflow.
Each percentage increase in the capacity closure of coal mines results in a 0.00375 percent
decrease in the population growth. Again, this result demonstrates that there is strong mobility across the cities and probably explains why a closure does not automatically affect
a local average wage.
When an aggregate employment level has dropped sharply, the magnitude is much
larger than that of any change in the population. In particular, this drop is mainly due
to a shrinking in employment in the service sector. Even if the employment level in the
industrial sector is unchanged, the negative spillover effect has chased some people in the
service industry out of the city. On the other hand, the GDP from the industrial sector has
shrunk and forced the GDP per capita to decrease. To summarize, the impact of the coal
mine closure is quite strong and the spillover effect leads to economic under-performance.
2.5.3 Impact of a Coal Mine Closure on Cities, DD results
Finally, I reveal the impact of a coal mine shutdown leveraging on a difference-indifference analysis. Assuming each city is a local economy and has an imperfect labor
mobility with its neighboring city, I compare the interest outcome of the treated cities with
their neighboring cities. The results in this section will be used to test the six predictions
from the spatial equilibrium we set up in the Theoretical Framework section.
87

Table 2.6: OLS: Effect of Cumulative Capacity Closure on the Local Economy
Panel A: Outcome variables regarding to employment
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Pop Avg Wage Emp Emp Pri Emp Sec Emp Ter Mine Retail Consc Manu Service
Cum Closure -0.00361*** -128.4 -0.00918*** 0.0103 -0.0178*** -0.00499* -1078.4*** -274.6 -925.5 1358.7 -1815.4*
(-3.03) (-1.28) (-2.65) (0.68) (-3.65) (-1.66) (-6.06) (-0.74) (-1.45) (1.29) (-1.90)
Pop 2013 0.0168** 297.7 0.0357* -0.138* 0.0582** 0.0251 -1733.4* -4585.7 -4511.5 -31319.1*** 17553.6***
(2.27) (0.60) (1.71) (-1.71) (2.42) (1.08) (-1.67) (-1.34) (-1.18) (-4.10) (2.66)
Constant -0.0802* 16280.9*** -0.266** 0.355 -0.525*** -0.114 8868.0 23481.5 25790.0 150032.3*** -86329.1**
(-1.77) (5.57) (-2.15) (0.74) (-3.67) (-0.80) (1.44) (1.16) (1.20) (3.46) (-2.33)
Observations 288 286 288 276 288 287 261 286 287 287 276
Panel B: Outcome variables regarding to GDP/Welfare
(1) (2) (3) (4) (5) (6) (7) (8) (9)
GDP GDP C GDP Pri GDP Sec GDP Ter House Price Ins Unemp Pub Rev Pub Exp
Cum Closure -0.00919*** -778.9*** -0.00453 -0.0155*** -0.00364 -150.2*** -0.00671** 0.0309** 0.0351***
(-3.18) (-4.30) (-1.36) (-3.64) (-1.30) (-4.93) (-2.42) (2.55) (2.80)
Pop 2013 0.0637*** 1676.5 -0.0160 0.0711*** 0.0585*** 449.8*** 0.0141 0.168*** 0.206***
(5.00) (1.63) (-1.05) (4.15) (4.12) (2.64) (1.02) (2.98) (3.49)
Constant -0.116 2030.5 0.193** -0.284*** 0.0923 -615.1 0.0295 -0.0803 0.0953
(-1.46) (0.32) (2.08) (-2.67) (1.06) (-0.59) (0.35) (-0.24) (0.27)
Observations 288 286 288 288 287 281 288 288 288
t statistics in parentheses
* p
<
0.10, **
p
<
0.05, ***
p
<
0.01
Notes: The dependent variable is log cumulative coal production capacity Closure in city c from 2016 to 2019.
88

Table 2.7: Reduced Form and 2SLS: Causal Effect of Cumulative Capacity Closure on the Local Employment
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Population Avg Wage Emp Emp Pri Emp Sec Emp Ter Mine Retail Consc Manu Service
Panel A: Reduced form regression
1912 Mining Tax -0.00129 -26.02 -0.00618** -0.00758 -0.00409 -0.00635** -90.18 -409.3 -376.2 815.7 -973.1
(-1.26) (-0.28) (-1.98) (-0.56) (-0.89) (-2.41) (-0.55) (-1.14) (-0.67) (0.96) (-1.10)
Population 2013 0.0185*** 328.6 0.0439*** -0.132* 0.0660*** 0.0334** -1661.9* -4169.0** -3988.6 -32447.2*** 19474.3***
(3.47) (0.68) (2.70) (-1.85) (2.76) (2.42) (-1.87) (-2.22) (-1.35) (-7.34) (3.95)
Constant -0.0975*** 15816.2*** -0.328*** 0.359 -0.608*** -0.165** 5621.1 20939.5* 20995.6 158808.1*** -100534.7***
(-3.12) (5.62) (-3.45) (0.86) (-4.34) (-2.05) (1.08) (1.90) (1.21) (6.13) (-3.47)
Observations 288 286 288 276 288 287 261 286 287 287 276
Panel B: 2SLS regression
Cum Mine Closure -0.00379* -75.73 -0.0182** -0.0226 -0.0120 -0.0186** -300.6 -1181.1 -1104.7 2395.0 -2924.0
(-1.76) (-0.30) (-2.09) (-0.62) (-0.96) (-2.45) (-0.57) (-1.01) (-0.69) (0.92) (-1.24)
Population 2013 0.0165** 291.0 0.0343* -0.144* 0.0596** 0.0240 -1799.2 -4755.2 -4560.0 -31208.3*** 17673.6***
(2.22) (0.59) (1.68) (-1.75) (2.42) (1.04) (-1.49) (-1.39) (-1.19) (-4.08) (2.67)
Constant -0.0783* 16183.1*** -0.236* 0.476 -0.547*** -0.0740 7101.4 26703.3 26513.4 146845.8*** -84214.4**
(-1.70) (5.54) (-1.90) (0.92) (-3.58) (-0.52) (0.94) (1.32) (1.17) (3.35) (-2.36)
Observations 288 286 288 276 288 287 261 286 287 287 276
fs 59.84 61.09 59.84 57.30 59.84 60.16 41.67 62.75 59.96 59.96 54.84
t statistics in parentheses
* p
<
0.10, **
p
<
0.05, ***
p
<
0.01
Notes: Some variables are in log. Please refer to table 2.3. The dependent variable is log cumulative coal production capacity Closure in city c from 2016 to 2019. The
instrumental variable is log 1912 mining tax payment in city c.
89

Table 2.8: Reduced Form and 2SLS: Causal Effect of Cumulative Capacity Closure on the Local GDP and Prices
(1) (2) (3) (4) (5) (6) (7) (8) (9)
GDP GDP per cap GDP Pri GDP Sec GDP Ter House Price Insured Unemp Pub Rev Pub Exp
Panel A: Reduced form regression
1912 Mining Tax -0.00190 -299.1* -0.00167 -0.00722* 0.00285 -33.73 0.00229 0.0319*** 0.0311***
(-0.68) (-1.80) (-0.56) (-1.76) (1.07) (-1.19) (0.84) (2.96) (2.76)
Population 2013 0.0672*** 2161.7** -0.0143 0.0814*** 0.0561*** 519.3*** 0.0117 0.129** 0.166***
(4.62) (2.48) (-0.93) (3.81) (4.03) (3.50) (0.83) (2.29) (2.83)
Constant -0.157* -2318.4 0.175* -0.373*** 0.0937 -1347.1 0.0237 0.183 0.369
(-1.84) (-0.45) (1.94) (-2.97) (1.15) (-1.55) (0.29) (0.56) (1.07)
Observations 288 286 288 288 287 281 288 288 288
Panel B: 2SLS regression
Cum Mine Closure -0.00560 -885.5** -0.00490 -0.0212** 0.00840 -99.48 0.00674 0.0938*** 0.0914**
(-0.76) (-2.10) (-0.63) (-1.96) (1.03) (-1.49) (0.98) (2.65) (2.49)
Population 2013 0.0642*** 1653.7 -0.0169 0.0702*** 0.0607*** 461.2*** 0.0153 0.178*** 0.215***
(4.95) (1.63) (-1.09) (4.14) (3.94) (2.71) (1.00) (2.84) (3.38)
Constant -0.128 2425.7 0.200** -0.265** 0.0499 -809.7 -0.0104 -0.293 -0.0944
(-1.53) (0.39) (2.06) (-2.36) (0.51) (-0.78) (-0.11) (-0.74) (-0.24)
Observations 288 286 288 288 287 281 288 288 288
fs 59.84 58.82 59.84 59.84 59.35 57.80 59.84 59.84 59.84
t statistics in parentheses
* p
<
0.10, **
p
<
0.05, ***
p
<
0.01
Notes: Some variables are in log. Please refer to table 2.3. The dependent variable is log cumulative capacity closure in city c from 2016 to 2019. The instrumental variable
is log 1912 mining tax payment in city c.
90

Baseline Difference in Difference
Table 2.9 and table 2.10 show the treatment effect on employment and GDP outcomes
respectively with baseline regression in Equation 2.4. Panel B takes the characteristics of
the neighboring city as controls, while panel A does not have any control. The coefficients
hardly change between the two settings. Now, I want to go through the testable predictions.
1. Mining employment: from Table 2.9 Column 7, we can tell that the treated cities
experience a drop of seven thousand employment positions in the mining industry,
confirming the first prediction of the negative direct effect on the mining industry.
2. Total employment and average wage: Even though the coefficient on total employment in Table 2.9 Column 3 is not significant, its negative direction is in alignment
with expectations. On the other hand, the coefficient of −1583 on Column 2 average wage exhibits consistency with the prediction. The negative labor demand may
reduce overall wage growth. This has a substantial large effect, equivalent to 3.3%
of the 2014 average wage in the treated cities, which is 47,069.4 yuan.
3. Traded industries: none of the variables of primary sector employment, secondary
sector employment, or employment in manufacturing from Table 2.9 Column 4, 5,
and 10 demonstrate differential trends between the control and treated cities. This
may be as a result of a complex impact on the traded industries; the three forces we
discussed are combined while none of them is strong enough to outweigh others.
However, the coefficient on the secondary sector GDP is significant, indicating the
negative labor demand negatively impacts lower secondary sector GDP growth rate
by 0.04 percent.
91

4. Local industries: The negative impact on other local industries is confirmed by the
negative coefficient sign on tertiary sector employment and employment in the industries of retail, construction and services, shown in Table 2.9 Column 6, 8, 9, and
11.
5. Housing price: The average housing price per square meter has been growing slower
by 233.4 RMB than that of their neighboring cities, in a statistically significant way.
(Table 2.10 Column 6)
6. Population flow: The coefficient on the logged population shows that a negative
labor demand does not bring any change in population flow in the treated cities and
their neighboring cities. This parameter shows population is less responsive to the
negative labor demand than the housing price and average wages. To some extent,
this result is compatible with Notowidigdo (2020)’s argument about asymmetric
response of populations: positive local labor demand shock increases the population
more than negative demand shocks, while there is no such effect on housing prices,
wages and rents.
Overall, the results are very consistent with the predictions from the spatial equilibrium
setting in Moretti (2010) and Moretti (2011). The negative labor demand shock creates a
profound spillover impact on the non-traded industries but its effect on traded industries is
not that clear.
There are some other interesting indicators evaluating people’s well being. First, the
impact on the overall GDP is negative, showing that neighboring cities are becoming more
prosperous than the treated cities. The economic performance of the treated cities has
92

Table 2.9: ATE by DD: Impact of Coal Mine Shutdown on Local Employment
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Population Avg Wage Emp Emp Pri Emp Sec Emp Ter Mine Retail Consc Manu Service
Panel A: Difference in difference regression without neighboring city characteristics as control
Treated -0.000703 -1583.1** -0.0376 0.0651 -0.0455 -0.0124 -7265.4*** -1815.5 -2630.1 1469.6 -6870.5*
(-0.09) (-2.45) (-1.48) (0.84) (-1.34) (-0.92) (-4.67) (-1.37) (-0.79) (0.42) (-1.71)
Year FE yes yes yes yes yes yes yes yes yes yes yes
City FE yes yes yes yes yes yes yes yes yes yes yes
Observations 1282 1278 1282 1265 1281 1280 1223 1280 1281 1281 1274
Panel B: Difference in difference regression with neighboring city characteristics as control
Treated -0.000406 -1513.0** -0.0368 0.0639 -0.0435 -0.00947 -7314.6*** -1737.7 -2142.5 1659.2 -7105.9
(-0.05) (-2.81) (-1.49) (0.84) (-1.33) (-0.76) (-4.60) (-1.22) (-0.66) (0.46) (-1.69)
Year FE yes yes yes yes yes yes yes yes yes yes yes
City FE yes yes yes yes yes yes yes yes yes yes yes
Observations 1256 1252 1256 1240 1255 1254 1197 1254 1255 1255 1248
t statistics in parentheses
* p
<
0.10, **
p
<
0.05, ***
p
<
0.01
Notes: A city is defined as the treated city if its capacity closure in any year between 2016 to 2019 is above 900,000 tons. Some variables are transformed to log, including
population, employment, employment primary, employment secondary, employment tertiary. Other employment data has the unit of person. The neighboring cities’ control
variables include average GDP, PM 2.5, average housing price, and average wage.
93

Table 2.10: ATE by DD: Impact of Coal Mine Shutdown on Local GDP and Prices
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
GDP GDP per cap GDP Pri GDP Sec GDP Ter House Price Insured Unemp Pub Exp Pub Rev PM 2.5
Panel A: Difference in difference regression without neighboring city characteristics as control
Treated -0.0458* -1013.5 0.000501 -0.0471* -0.0202 -230.5*** -14932.6 -0.000169 -0.00646 0.943
(-1.76) (-0.97) (0.02) (-1.89) (-1.26) (-3.08) (-1.32) (-0.00) (-0.20) (1.60)
Year FE yes yes yes yes yes yes yes yes yes yes
City FE yes yes yes yes yes yes yes yes yes yes
Observations 1282 1068 1068 1068 1069 1267 1283 1282 1282 1272
Panel B: Difference in difference regression with neighboring city characteristics as control
Treated -0.0475* -690.4 0.00578 -0.0318 -0.0148 -233.4*** -16191.8 0.00867 0.00190 1.058***
(-2.05) (-0.78) (0.30) (-1.46) (-1.16) (-3.14) (-1.38) (0.22) (0.06) (3.56)
Year FE yes yes yes yes yes yes yes yes yes yes
City FE yes yes yes yes yes yes yes yes yes yes
Observations 1256 1047 1047 1047 1048 1244 1257 1256 1256 1246
t statistics in parentheses
* p
<
0.10, **
p
<
0.05, ***
p
<
0.01
Notes: A city is defined as the treated city if its capacity closure in any year between 2016 to 2019 is above 900,000 tons. Some variables are transformed to log, including
GDP, GDP primary, GDP secondary, GDP tertiary, public expenditure, public revenue. The neighboring cities’ control variables include average GDP, PM 2.5, average housing
price, and average wage.
94

been dragged down by this labor demand shock. However, we don’t see any statistically
significant shift in public revenue and public expenditure. It is puzzling that the local fiscal revenue does not decrease with economic development slowing down and comparable
government expenditure indicates the government is not spending extra money to help unemployed people overcome the negative shock. This conjecture can be further confirmed
by the coefficient of the number of insured unemployed people – the government is not
helping more people to get unemployment insurance. Finally, the indicator of PM 2.5 does
not perform differently between the treated cities and their neighboring cities. One case
can be that both the condition of the air in the treated cities and their neighboring cities
both improve during these years and they don’t have a different trend.
Exploiting the Magnitude of the Closure
To exploit the magnitude of the production capacity being shut down, I ran equation
2.5 and reported the results in Table 2.11. Most of the results are consistent with what we
have in the baseline DD.
The two coefficients demonstrating statistical significance are mining employment and
average housing price. For each 10,000 tons of coal production capacity being shut down
in a city, ten mining jobs are lost and the average housing price decreases by 0.397 yuan.
95

Table 2.11: ATE by DD: Impact of Coal Mine Shutdown on Local Economy - Exploiting Closure Magnitude
Panel A: DiD analysis - Outcome variables regarding employment with neighboring city characteristics as control
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Population Avg Wage Emp Emp Pri Emp Sec Emp Ter Mine Retail Consc Manu Service
Cum Closure 0.0000106 -0.992 -0.0000201 0.0000403 -0.0000380 0.00000868 -10.37*** -2.315 -0.386 1.871 0.00942
(0.61) (-0.96) (-0.74) (0.39) (-0.87) (0.61) (-3.90) (-1.30) (-0.09) (0.30) (0.00)
Year FE yes yes yes yes yes yes yes yes yes yes yes
City FE yes yes yes yes yes yes yes yes yes yes yes
Observations 1256 1252 1256 1240 1255 1254 1197 1254 1255 1255 1248
Panel B: DiD analysis - Outcome variables regarding GDP with neighboring city characteristics as control
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
GDP GDP per cap GDP Pri GDP Sec GDP Ter House Price Insured Unemp Pub Exp Pub Rev PM 2.5
Cum Closure -0.0000234 -0.259 0.0000468 -0.0000303 0.00000970 -0.355*** -1.269 0.0000558 0.0000550 0.000760
(-0.62) (-0.16) (1.39) (-0.85) (0.55) (-4.68) (-0.06) (0.93) (0.81) (1.61)
Year FE yes yes yes yes yes yes yes yes yes yes
City FE yes yes yes yes yes yes yes yes yes yes
Observations 1256 1047 1047 1047 1048 1244 1257 1256 1256 1246
t statistics in parentheses
* p
<
0.10, **
p
<
0.05, ***
p
<
0.01
Notes: This regression table is for equation 2.5, exploiting the magnitude of the capacity closure. Dependent variable is Treated ∗Cum Closure. Cum Closurec,t refers the
cumulative capacity closure in city c until year t. Some variables are transformed to log, including GDP, GDP primary, GDP secondary, GDP tertiary, public expenditure,
public revenue, population, employment, employment primary, employment secondary, employment tertiary. Other employment data has the unit of person. The control
variables include neighboring cities’ average GDP, PM 2.5, average housing price, and average wage.
96

Event Study
Figure 2.10 displays the dynamic effect of the treatment with event study regressions
in Equation 2.6. The four outcomes displayed are mining employment, housing price, PM
2.5 and secondary (industrial) sector employment.
Figure 2.10: Event Study of Coal Mine Closure Impact
The sizable decrease in the level of employment in the mining industry happens with
the treatment. We can tell the downward trend has been continued. Similarly, the average
housing price shifts its trend and starts to drop sharply as the city suffers from a labor
demand shock. However, the downward trend has been curbed after two years of the
treatment. Then, regarding PM 2.5, we can tell that the closure of a coal mine does not
stop the air quality from deteriorating further and therefore the coal mines may not be
the main source of the air pollution in the local region. Finally, the decreasing trend
97

in industrial employment already existed in the treated cities prior to treatment but the
treatment seems to have a major heterogeneous effects to different cities as the variance
is getting larger. This is consistent with our model prediction about the traded industries:
the three general equilibrium effects behave differently in each treated city and result in an
ambiguous result.
2.5.4 Discussion on the Results from the Three Empirical Strategies
To conclude, I want to draw the common inference from the three empirical strategies.
Table 2.12 summarizes the impact of coal mine shutdowns on local economic development
derived from the three strategies.
Overall, all the three strategies demonstrate a negative impact of the shutdown on local
economies. Even though due to the data limitation, O-B regression can not drill down
the impact into the employment rates and price of different industries/goods, it indicates
a significant negative coefficient on the GDP growth. This convincing result holds in the
other two strategies. Then, regarding population, both O-B regression and IV regression
show a population outflow from the cities with coal mine shutdowns. There is no such
effect in DD regression.
To focus on the employment and price in different industries/goods, I find relatively
consistent results from IV regression and DD regression. The negative labor demand shock
reduces the non-traded sector employment and total employment. And both strategies
show an unclear impact on the level of traded-sector employment. These results match
with the model prediction. Finally, DD regression provides definite coefficients on mining
employment, average wage, and housing price, which are all negative.
98

Table 2.12: Empirical Results Summary
(1) (2) (3) (4) (5) (6) (7) (8)
GDP Population Mining Emp Total Emp Avg Wage Emp Traded Emp Non-traded Housing Price
O-B regressioncounty level (-) (-)
IV regressioncity level (-) (-) (?) (-) (?) (?) (-) (?)
DD regressioncity level (-) (?) (-) (-) (-) (?) (-) (-)
Notes: This table summarizing the empirical results from the three empirical strategies.
99

A contrast between the results from IV regression and DD regression is interesting,
especially on the variables of population, average wage, and housing price. In the IV
regression, the analysis covers the cities across China and thus the workforce has a high
mobility. As people are able to move away from the cities suffering negative labor demand
shock, the labor supply decreases and the average wage and housing price remain around
the original equilibrium prices. On the other hand, when I compare the treated cities and
their neighboring cities, labor is not that mobile between them, even though labor may
move to cities further away. In this case, compared with the neighboring cities, the treated
cities have to endogenously adjust their labor price and housing price. This may be the
economic logic behind why when the coefficient on a population is significant, there is no
clear change in average wage and housing price, and vice versa.
2.5.5 The Mitigating Effect of Transportation Infrastructure on Negative
Shocks
To test Prediction 6 of how the labor supply elasticity changes the impact of negative
labor demand shock, I have run regression 2.7 and reported the results in Table 2.13. Panel
A lists the coefficients on employment outcomes and Panel B on GDP outcomes.
Let us discuss its impact in the sequence of our predictions.
1. Mining employment: Better transportation exacerbates the decrease in mining employment, even though the coefficient is not significant. More miners are quitting
the industry maybe because the better transportation infrastructure gives them better
access to other low-skilled jobs.
100

Table 2.13: Heterogeneous ATE by DD: Impact of Coal Mine Shutdown by Variation in Transportation Infrastructure
Panel A: Heterogeneity treatment effect analysis - Outcome variables regarding employment
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Population Avg Wage Emp Emp Pri Emp Sec Emp Ter Mine Retail Consc Manu Service
Treated -0.0165* -2484.8** -0.0359 0.111 -0.0774* -0.00588 -5020.0** -659.7 -211.5 5723.7 -8163.4**
(-1.87) (-2.70) (-1.26) (1.21) (-1.79) (-0.29) (-2.44) (-0.46) (-0.06) (1.38) (-2.24)
Treated*High Speed 0.0277** 1579.6* -0.00289 -0.0801 0.0558 -0.0115 -3944.6 -2028.8 -4245.2 -7467.0** 2277.4
(2.15) (1.73) (-0.14) (-0.90) (1.44) (-0.56) (-1.37) (-1.19) (-1.32) (-2.44) (0.72)
Year FE yes yes yes yes yes yes yes yes yes yes yes
City FE yes yes yes yes yes yes yes yes yes yes yes
Observations 1282 1278 1282 1265 1281 1280 1223 1280 1281 1281 1274
Panel B: Heterogeneity treatment effect analysis - Outcome variables regarding GDP
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
GDP GDP per cap GDP Pri GDP Sec GDP Ter House Price Ins Unemp Pub Exp Pub Rev PM 2.5
Treated -0.118*** -3641.9 0.00447 -0.0979** -0.0456 -476.1*** -28551.8** 0.0387 0.0156 3.654***
(-3.07) (-1.61) (0.15) (-2.28) (-1.22) (-4.54) (-2.52) (0.73) (0.27) (3.68)
Treated*High Speed 0.126** 4659.6 -0.00703 0.0900* 0.0450 429.6*** 23902.5** -0.0682 -0.0387 -4.716***
(2.27) (1.71) (-0.31) (1.94) (1.01) (3.30) (2.71) (-1.10) (-0.57) (-3.37)
Year FE yes yes yes yes yes yes yes yes yes yes
City FE yes yes yes yes yes yes yes yes yes yes
Observations 1282 1068 1068 1068 1069 1267 1283 1282 1282 1272
t statistics in parentheses
* p
<
0.10, **
p
<
0.05, ***
p
<
0.01
Notes: This regression tests how the transportation infrastructure alter the impact of negative labor demand shock. High Speed refer to 1 if the city occupies at least one
high speed rail station. Some variables are transformed to log, including GDP, GDP primary, GDP secondary, GDP tertiary, public expenditure, public revenue, population,
employment, employment primary, employment secondary, employment tertiary. Other employment data has the unit of person.
101

2. Total employment and average wage: The insignificant coefficient on total employment is unchanged. However, the treated cities with HSR decrease much less where
average wages are concerned than other treated cities.
3. Traded industries: The employment level in secondary industries and GDP from
secondary sectors both increase, showing a positive impact of higher mobility on
local industrial business. However, employment in manufacturing has decreased.
This result is puzzling and needs extra exploration.
4. Local industries: HSR has a positive but not significant impact on service employment. The size is not big enough to counter the negative impact of a coal mine
shutdown.
5. Housing price: HSR demonstrates a strong positive impact on housing prices, completely hedging the negative result from the negative labor demand shock.
6. Population flow: Population in the treated cities with HSR experience population
inflow and cities without HSR experience population outflow.
Overall, better transportation infrastructure conditions equip cities with a better ability to
counter the negative labor demand shock from mine closures. The local industries, labor,
and land owners are somehow protected by the mitigating effect of better transportation
infrastructure on such negative shocks. The average wage and housing price in these cities
are much more resilient. Even though we don’t see any extra spending/transfer in fiscal
expenditure to help the unemployed people by local governments, the Chinese government is building extensive infrastructures which may help mitigate the negative impact. In
102

2019, China built 35,000 km of high-speed rail (Wong (2021)). Such infrastructure investment is strategic and visionary as the benefit from such infrastructure can be long-lasting.
However, it also brings unequal development to the regional economy.
2.5.6 Differential Performance of Private firms vs. SOEs
Finally, I want to examine the impact on the firms by comparing the performance of
the listed firms whose headquarters are located in the treated cities with that of the listed
firms whose headquarters are located in neighboring cities.
Table 2.14 lists the estimated coefficients from Equation 2.8. We can tell the firms
performed very differently by their shareholding structure: private firms gain in total sales,
employee number, and profit, while SOEs tumble in these indicates. The SOEs in the
mining industry undergo a large-scale layoff. Compared with the SOEs in other industries
and having headquarter in the treated cities, the SOEs in the mining industry have seen
employee number decreased by 0.116 percent and payment to employees reduced by 0.174
percent.
Furthermore, the negative and significant decrease in SOE’s TFP is noticeable. The
mechanism behind this divergent performance between private firms and SOEs may be
complicated. I provide one hypothesis. The SOEs are the policy instrument of the government to implement supply-side structural reform. They care more about achieving the
goal set the by the government, even at the cost of sacrificing profit. Given the goal of
the government is to reduce the mining production capacity, they have fired employee and
cut the production, resulting in a lower TFP. At the same time, private firms increase their
production and take more market share, resulting in higher sales and profit. However, this
paper is not able to test this hypothesis and it is left for future research.
103

Table 2.14: Impact of Coal Mine Shutdown on Firms
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Asset Current Debt Bank Debt ROA ROE Ln Sale Employee Payment Emp Fixed Asset Profit TFP
Treated 0.0228 0.00217 -0.0143 0.00463 -0.154 0.122** 0.0807* 0.0961* 0.102* 0.120* 0.0200
(0.68) (0.25) (-1.49) (0.24) (-1.60) (2.39) (1.72) (1.88) (1.93) (1.87) (0.81)
Treated*SOE -0.0236 0.00295 -0.00652 0.00101 0.0914 -0.154*** -0.0471** -0.0827*** -0.120*** -0.141*** -0.0614***
(-0.91) (0.28) (-1.10) (0.06) (0.17) (-6.05) (-2.12) (-4.17) (-3.84) (-3.23) (-3.36)
Treated*SOE*Mine 0.00819 0.00979 0.00200 -0.0163 0.0193 -0.0630 -0.116** -0.174*** 0.0989 -0.0979 0.0698
(0.25) (0.90) (0.14) (-1.28) (0.04) (-0.95) (-2.24) (-2.86) (1.26) (-0.52) (1.15)
Year FE yes yes yes yes yes yes yes yes yes yes yes
Firm FE yes yes yes yes yes yes yes yes yes yes yes
Observations 11040 10895 7964 11040 10961 10810 11035 11039 11036 9283 10806
t statistics in parentheses
* p < 0.10, ** p < 0.05, *** p < 0.01
Notes: This regression reveals the heterogeneous treatment effect on firms with different shareholding structure. SOEi equal to 1 if the firm i is classified as an SOE. Minei
equal to 1 if firm i is in mining industry. TFP is calculated using LP method. Levinsohn and Petrin (2003)
104

2.6 Conclusion
This paper examines the negative economic impact of coal mine shutdowns in China
since 2016, focusing on GDP growth, employment, average wages, and housing prices.
The findings show that despite the lack of direct government assistance for affected workers, the development of transportation infrastructure has played a crucial role in mitigating
economic losses. While not an immediate relief measure, this infrastructure investment
acts as a stabilizing force, easing economic transitions during reform.
Several improvements could strengthen this study. First, the analysis does not capture
micro-level effects on individuals living in affected cities. To address this, householdlevel data would be essential for examining how the impact varies by skill level. This
would help identify who benefits and who suffers under this policy shift. Additionally,
access to panel data tracking employment over time would allow for an investigation of
the long-term consequences of unemployment. Understanding these effects could provide
clearer policy implications for assisting displaced workers.
Further research is also needed on the transition costs for low-skilled labor in developing economies. Additionally, while this paper proposes a hypothesis on how coal mine
closures affect firms, it is not yet tested empirically. Future work will aim to validate this
hypothesis and explore the broader industrial consequences of mine shutdowns.
105

Chapter 3
White Elephant or Win-Win Projects: An Investigation
into the Economic Impact of the Belt and Road Initiative
Do the projects relating to the Belt and Road Initiative (BRI) harm or benefit the local
economy? Using the staggered implementation of BRI projects in Indonesia and a quarter
province panel dataset, I applied the econometric methods of difference-in-difference and
instrumental variable regressions to show empirical evidence of their impact on macroeconomic and social indicators. A BRI project does not affect the local economy when it
is first implemented in a province. However, after a BRI project is accomplished, it helps
the overall GDP to grows 0.02%, and the GDP of exports expand 0.312% more than the
areas where no BRI project has been completed. Regarding the heterogeneous effect of the
projects across industry, I found that metal and mineral projects are much more inclined
towards export growth and to the realization of foreign investment, while electricity-based
projects don’t have a significant impact. The BRI projects do not impose a heavy financial
burden on the community, measured by the level of commercial and rural bank loans.
Keywords: Foreign Lending, Infrastructure, Economic Growth
JEL Codes: F34, F43, O18, O53, R1
106

3.1 Introduction
Aiming to boost regional connectivity between China and countries in Central Asia,
Africa, Europe, the central government of China announced the Belt and Road Initiative
(BRI) in 2013 with plans to invest in and develop infrastructure projects such as transportation facilities and powerplants in the participating countries (henceforth, BRI countries).
The projects are financed through public financial institutions backed by the central government of China. BRI countries are allowed to borrow from these state-owned financial
vehicles at the prevailing commercial interest rate, while the infrastructure projects are
usually undertaken and fulfilled by Chinese state-owned enterprises (SOEs) under government contracts.
While the Chinese government and partner countries are strong proponents for the
benefits of BRI projects, the government’s intention for backing these projects has been
questioned. In particular, some BRI projects have been scrutinized as a white elephant,
a metaphor often used to describe a costly or extravagant venture with little use or value.
Furthermore, some of the projects have been identified as debt traps by the Western media,
possibly leading BRI countries into debt distress. John Hurley and Portelance (2019) has
warned that eight countries are at the highest risk of falling into this situation, which
include Djibouti, Kyrgyzstan, Laos, the Maldives, Mongolia, Montenegro, Pakistan, and
Tajikistan. These concerns are not unwarranted. As outlined by Abi-Habib (2018), in one
BRI case in Sir Lanka, due to the government’s inability to repay its loan, the Sir Lankan
government ceded operating rights of the Port of Hambantota to a Chinese SOE under a
99-year lease as a condition for debt forgiveness.
107

Furthermore, there is little literature studying the economic impact of BRI on the local
economy due to a lack of transparency and public data. Even though the Chinese Minister of Finance publishes the number and total dollar value of the BRI-related contracts
annually, the data is not available on a per-contract or per-region basis. For example, it
is reported that “the total investment related to BRI has exceeded $300 billion, bringing
300,000 more jobs to the BRI countries,” Minister of Finance (2019).
Whether a BRI project will benefit or harm the local economy is an open empirical question, especially during the project’s implementation. BRI projects are financed
through Chinese government-controlled financial institutions, which usually come with
conditions attached, such as a requirement to hire Chinese SOEs as contractors or use
materials imported from China. Thus, if the projects do not create local employment or
utilize the local supply chain, there may be no positive effect of these projects on the local
economy when it is being implemented.
On the other hand, there are several channels through which BRI projects can benefit the local economy. First, they can improve local well-being in regions where projects
are implemented by generating jobs and increasing earnings. Second, to the extent that a
project uses the local supply chain, it can produce positive demand shocks and give firms
an incentive to improve productivity. Lastly and most importantly, after the project is completed, the infrastructure level is enhanced, and market friction will be reduced substantially through better transportation connections and access to services such as electricity,
paving the way for economic growth. Thus, a BRI project may exert different effects,
depending on which stage it is at.
108

In this research project, I have taken the case of Indonesia as a sample to investigate the
impact of BRI projects on the social and economic outcomes of each province. Leveraging
hard-to-obtain project-level BRI data, I have exploited the staggered implementation of
BRI projects in Indonesia and a quarter-province panel dataset consolidated by myself, I
have used two econometric methods of difference-in-difference and instrumental variable
(IV) to show their causal impact on economic indicators.
This study provides ample empirical evidence of BRI projects. First, when a BRI
project is being implemented in an Indonesian province, it does not initially affect local
GDP growth while it only slightly decreases the amount of commercial and rural bank
loans. Second, after a BRI project had been completed, it helped overall GDP grow by
0.0199%, and the GDP of exports expanded 0.312% more than the areas where no BRI
project existed. Third, regarding the heterogeneous effect of the projects across industry,
I found metal and mineral projects were much more favorable to export growth and to the
realization of foreign investment while electricity-based projects didn’t have much of an
effect. To conclude, the results of this paper state that BRI projects bring a certain amount
of economic momentum to the local community without inducing high financial costs.
Both econometric methods provide consistent results, which substantiate this conclusion.
This paper is related to several strands in economic literature. Firstly, this research
provides empirical evidence of the effect of the infrastructure on the local economic growth
in developing economies. Donaldson and Hornbeck (2016) examines the considerable
impact of railroads on the U.S. economy through the increase in “market access” in 1890.
Additionally, Donaldson (2018) shows the positive impact of the railroad on overall social
welfare. In this paper, I complement this literature by focusing on a recent initiative.
109

Additionally, I attempt to provide a conclusion on how infrastructure projects improve the
poverty level.
A more directly related topic with my research project is the effect of foreign direct
investment (FDI) on recipient countries. Nair-Reichert and Weinhold (2001) shows that
the positive impact of FDI on a developing economy is heterogeneous across countries,
and it is higher in more open economies. By utilizing the detailed information of the
projects, I was able to analyze the heterogeneous effect of foreign direct investment across
different industries.
Yao and Wei (2007) clarifies the dual role of FDI in promoting economic growth for
a newly industrializing economy. The paper indicates that FDI improves business productivity and expands the production frontier. I intend to complement this literature by
uncovering the causal effect of FDI on other social and economic outcomes, such as the
development of the highway’s length and the expansion of the bank loan amount. However, my results indicate a very weak or insignificant relationship.
This project is closely connected to literature on foreign aid which is different to FDI
as foreign aid may not require repayment or investment returns. Qian (2014) reviews
the studies on foreign aid and found the objectives of donor countries determine the aid,
instead of the need of recipient countries. As this study finds, there is a positive effect of
BRI projects on the growth of household consumption and exports. While they do not have
any significantly harmful impact on the variables analyzed in this paper, I argue that BRI
projects work better than foreign aid in the perspective of fitting the need of the recipient
countries.
110

Dreher et al. (2021) examines the causal effect of Chinese aid on GDP growth across
138 countries between 2000 and 2014, using the dataset from AidData lab at William &
Mary. My study complements their results from three different angles. First, I examine
the effect of China-financed BRI projects after 2014. Second, I investigate the effect by
limiting the sample to one country. Third, I have collected six other social and economic
variables in addition to GDP, attempting to show an overall effect of the projects.
The paper is organized as follows. Section 3.2 explains the background and clarifies the
relationship between different stakeholders in BRI. Section 3.3 talks about the data source
and summary statistics. Section 3.4 gives empirical strategies. Section 3.5 discusses the
results. Section 3.6 provides robustness checks and Section 3.7 concludes.
3.2 Background
In September 2013, President Xi Jinping of China announced the launch of the Belt
and Road Initiative(BRI) at a conference in Pakistan. BRI intends to revive the ancient
trade route “Silk Road” connecting China and the rest of the world. “Belt” is called “the
Silk Road Economic Belt”, crossing Central Asia, West Asia, the Middle East, and Europe.
“Road” refers to the “21st Century Maritime Silk Road”, which travels through Southeast
Asia, Oceania, and Africa. In order to reinforce the dominant status in international trade,
China wants to reduce the trade cost along the routes. Hence, China decided to provide
the capital and labor support for those countries willing to become involved in building
fundamental infrastructure. The Chinese government has been actively negotiating with
the other governments to instigate the arrangement of BRI projects.
111

BRI Countries: The total number of countries that agreed to be involved in BRI
has continuously increased over the years. In 2014, Russia and Mongolia were the first
countries to join the initiative. Xinhua (2019) indicates that up to 2019, there were 137
countries involved with the BRI. In 2019, Italy joined as the first G7 country to cooperate
with China.
The Contracts: There are usually two kinds of contracts for BRI projects, an EPC
contract and a builder contract. EPC contracts mean that the contractor will be responsible
for Engineering, Procurement, and Construction, while the responsibility of a builder contract solely involves construction. The proportion of EPC contracts has been increasing.
Minister of Finance (2019) listed the following data to show the growth of the number of
EPC contracts. In 2019, Chinese enterprises newly signed 6,944 EPC contracts in 62 countries along the “Belt and Road,” worth 154.89 billion US dollars, accounting for 59.5% of
China’s total of newly signed foreign contract. The number of EPC contracts increased by
23.1%, compared with the previous year.
The SOEs: Most of the firms awarded BRI project contracts are Chinese state-owned
enterprises. I summarize three reasons behind it. First, the EPC project is usually largescale, such as building a highway or a power plant. Compared with private firms, SOEs
are more capable of handling these major tasks. Second, as each project’s bidding process
lacks transparency, we do not know how exactly the SOEs were awarded each project.
One projection is that the project may be assigned by the Chinese central government to
the SOE, instead of being arranged by fairly public bids. SOEs consider accomplishing
BRI projects as a political task. Third, the project does not guarantee to deliver profit, and
some of them may be risky, so private firms do not want to bear the risk.
112

The Financing Institutions: There are mainly three sources of financing: the Silk
Road Fund, the Asian Infrastructure Investment Bank, and the Export-Import Bank of
China. In 2014, the Chinese government pledged $40 billion for the creation of the Silk
Road Fund to support the BRI. Then, in 2015 the Asian Infrastructure Investment Bank
was formed under the cooperation between the Chinese government and 20 other Asian
countries, aiming to facilitate the financing of the infrastructure projects. Another primary financing institution for BRI projects is Export-Import Bank of China, a policy bank
supporting China’s international trade.
3.3 Data and Summary Statistics
In this study, I want to focus on Indonesia to reveal the effect of a BRI project on
local macroeconomic and social outcomes. To do so, I collected and merged the dataset
from two segments, one of projects information and the other of the economic and social
variables of each Indonesian province. I elaborate on the data source and the summary
statistics in this section.
The BRI project data was collected from the website www.bhi.com.cn. The website
releases BRI project information for potential downstream suppliers to contact the main
contractor so that they can bid on the project to become a supplier. The website belongs
to Beijing Huaxinjie Investment Consulting Company. The raw project information was
scraped from the website, and then the data was manually organized and cleaned using
STATA. There are 3,810 BRI-related projects across the world between 2014 and 2020.
As I am only focussing on Indonesia in this study, 198 projects were left in the data set.
113

The information about each project includes a brief introduction of the project content,
the industry it belongs to, the main contractor (usually a Chinese SOE), and an approximate starting date. Figure 3.1 shows the number of projects starting at each quarter, and
Figure 3.2 demonstrates the number of projects by industry, including electricity, transportation, construction, and metal and mineral production. Electricity-related projects are
the dominant type of BRI projects in Indonesia, accounting for 98 out of the 198 projects.
They relate to projects where the contractor builds a power plant, including but not limited to hydroelectric and thermal power plants. Transportation projects include building
highways and railways. Construction projects do not account for a large percentage of the
projects, but usually refer to projects such as building an industrial park or a hotel. Metal
and mineral projects relate to energy resources, mineral extraction, and metal production.
Figure 3.3 depicts the distribution of projects by industry across the provinces. We can
see that they don’t share the same distribution. Kalimantan has more electricity-based and
metal-related projects, while Java has most of the construction projects. Transportation
projects span the whole country.
I manually extracted the information of the province where each project was located
from its brief introduction so that I could match them with the provincial outcome data. I
also tried to extract each project’s duration so that I knew when the project ended. However, 80 percent of projects do not disclose how long the project will last or lasted, and
this created difficulty in identifying the effects of the accomplished projects, which will
be discussed later. Similarly, I attempted to gather the financial value of each project to
measure the return and cost of the project more precisely, but most of the projects did not
114

0 10 20 30 40 Number of Projects
2014Q4 2015Q1 2015Q2 2015Q4 2016Q4 2017Q1 2017Q2 2017Q3 2017Q4 2018Q1 2018Q2 2018Q3 2018Q4 2019Q1 2019Q2 2019Q3 2019Q4 2020Q1 2020Q2
Figure 3.1: Number of Projects per Quarter
come with that information. Table 3.1 lists the total number of projects for each Indonesian province. Central Sulawesi has the highest number of projects at 27, while six out of
34 provinces have no BRI-related projects.
Regarding the regional macroeconomic and social data, I collected the quarterly GPD
data of each province from Indonesian Central Bureau of Statistics for 2010Q1 to 2020Q1.
The GDP data can be broken down by expenditure or by industry, so I extract two GDP
components from each break down. From the method of expenditure, I collected the GDP
of household expenditure and the GDP of exports. By looking at the effect of BRI projects
on the export GDP, I can test if the projects captures their intention to promote trade.
Household consumption indicates how family living condition can or will benefit through
employment directly or indirectly created by the BRI projects. From the classification by
industry, I gathered the data for the GDP of electricity and gas and the GDP of construction, the two industries most likely to benefit from BRI projects.
115

Table 3.1: Number of BRI Projects in Each Province
Province Total Projects Percent% Cumulative Perc.%
Aceh 10 5.05 5.05
Bali 2 1.01 6.06
Banten 9 4.55 10.61
Bengkulu 5 2.53 13.13
Central Java 8 4.04 17.17
Central Kalimantan 2 1.01 18.18
Central Sulawesi 27 13.64 31.82
DKI Jakarta 14 7.07 38.89
East Java 7 3.54 42.42
East Kalimantan 12 6.06 48.48
East Nusatenggara 1 0.51 48.99
Jambi 1 0.51 49.49
Lampung 2 1.01 50.51
Maluku 4 2.02 52.53
North Kalimantan 6 3.03 55.56
North Maluku 15 7.58 63.13
North Sulawesi 3 1.52 64.65
North Sumatera 3 1.52 66.16
Riau 6 3.03 69.19
Riau Islands 8 4.04 73.23
South Eastsulawesi 8 4.04 77.27
South Kalimantan 3 1.52 78.79
South Sulawesi 6 3.03 81.82
South Sumatera 16 8.08 89.9
West Java 9 4.55 94.44
West Kalimantan 9 4.55 98.99
West Papua 1 0.51 99.49
West Sumatera 1 0.51 100
Bangka Belitung 0 0 100
DI Yogyakarta 0 0 100
Gorontalo 0 0 100
Papua 0 0 100
West Nusatenggara 0 0 100
West Sulawesi 0 0 100
Total 198
Notes: There are totally 34 provinces in Indonesia. 6 provinces don’t have any BRI projects.
116

0 20 40 60 80 100 Number of Projects
Construction Electricity Metallurgical Industry Transportation
Figure 3.2: Number of Projects by Industry
I produced Figure 3.4 to show the correlation between 2010 and 2019 GDP distribution
and projects distribution. We can see that East Kalimantan and South Sumatera, which are
located in the highest quartile of the GDP distribution, are also the states that have the
highest number of BRI projects. However, the projects do not seem to target the welldeveloped areas. For example, quite a large number of projects were carried out in North
Kalimantan and Central Sulawesi, whose GDP levels are below average in both 2010 and
2019.
In addition, I have collected other social and economic variables to complement my
analysis, including foreign investment realization, length of road, percentage of poor people, permanent worker index of construction sector, commercial and rural banks loans, and
actual state government expenditure. Besides, I took the population size of each province
117

Figure 3.3: Distribution of Projects by Industry Across Indonesia
as a control. Foreign investment realization is defined as “data about investment that were
realized by foreign companies in form of goods or services and the company had received
permanent licenses from the government”, according to the Central Bureau of Statistics.
As infrastructure projects usually involve high-value investment, I expect BRI projects to
directly affect it.
The variable length of road indicates the overall length of highways within a province.
As one of the most important instruments to connect the goods and services from its production location to the target market, an increase in the length of highway signals the
118

Figure 3.4: Indonesian GDP Distribution v.s. Projects Distribution
mitigation for market friction. The percentage of poor people measures the impact the
BRI projects have on alleviating the poverty level. Even though it is not the main intention of BRI projects, it is worth understanding its social welfare impact. Government
expenditure serves as an indicator of policy orientation. Finally, I also collected data on
regional, commercial, and rural banks loans to test the criticism to BRI projects, which
claims that they may cause debt distress to the recipient country. By examining the effect
of BRI projects on these social and economic outcomes, I could evaluate their impact more
comprehensively.
119

Table 3.2, summary statistics, shows the 2014 mean value of each variable by province,
ranking by the percentage of poor people. The social and economic factor distributions
behave consistently with the GDP distribution we have seen in the graph, with Jakarta,
Bali, and Kalimantan having the smallest percentage of poor people in the country. The
investment realization amount varied considerably among the provinces, ranging from 0
to USD 1,640 million. Regarding the credit expansion, Jakarta has a greater volume of
bank loans compared to other regions. Overall, the economic conditions in Indonesia are
not distributed equally across the country.
In addition, to construct the instrumental variable, I downloaded the Chinese steel
production per quarter from the website of the National Bureau of Statistics of China. I
plotted the trend of logged Chinese steel production in the left upper corner of Appendix
Figure A4. The production has continued to grow, with a temporary drop only in the first
quarter of 2018.
3.4 Empirical Strategy
I wanted to test whether any change of local economic and social factors was affected
by BRI projects. Exploiting the panel data structure and the staggered implementation of
projects across provinces in Indonesia, I first used a difference in difference as my main
econometric method. Under this method, I assumed the implementation of projects as a
treatment and evaluate the average treatment effect. Because the non-testable unconfoundness assumption may not have held in this setting, I further constructed an instrumental
variable to estimate the causal relationship to substantiate the results. In this section, I
explain the empirical strategies and the assumption related to the staggered difference in
120

Table 3.2: Summary Statistics: Mean of Each Variable by Province in 2014
Province Percentage of
Poor (%)
Permanent Worker
Index of Construction
GDP (IDR mn) GDP of Household
Consumption (IDR
mn)
GDP of Export
(IDR mn)
GDP of Construction (IDR mn) GDP of Electric- ity and Gas (IDR
mn)
Investment
Realization
(USD mn)
Commercial and
Rural Bank Loan
(IDR bn)
Length of
Road (km)
Population
(thousand)
State Government
Expenditure (IDR
bn)
DKI Jakarta 4.09 117.1 343,300,000 202,000,000 57,510,989 47,073,677 956,593 1,127 1,134,670 6,951 10,075 51,418
Bali 4.76 115.2 30,446,894 15,973,259 10,784,950 2,860,338 68,466 107 71,226 861 4,105 5,617
South Kalimantan 4.81 111.3 26,694,849 12,337,771 21,433,829 1,918,886 24,810 126 46,109 852 3,923 5,657
Bangka Belitung 4.97 103.6 11,039,860 5,599,168 6,800,383 888,095 8,906 26 12,635 899 1,344 1,922
Banten 5.51 103.8 87,337,807 50,429,161 27,233,235 7,909,117 1,099,792 509 194,252 853 11,705 8,138
Central Kalimantan 6.07 111.6 18,431,131 7,601,738 3,645,996 1,563,170 11,850 238 31,900 1,100 2,440 3,516
East Kalimantan 6.31 126.7 111,500,000 15,810,704 74,840,100 7,746,878 39,529 536 100,765 1,640 3,970 12,336
Riau Islands 6.4 113.7 36,581,309 13,333,353 26,595,615 6,488,707 331,555 98 41,242 895 1,917 3,458
West Sumatera 6.89 108.7 33,335,209 17,511,291 4,980,428 2,880,895 35,003 28 41,449 1,231 5,132 3,876
North Maluku 7.41 112.0 4,802,190 2,877,587 77,759 301,827 3,897 25 5,197 1,867 1,139 1,533
Riau 7.99 118.6 112,000,000 35,119,848 52,196,416 8,093,700 54,657 342 71,157 3,033 6,188 9,584
West Kalimantan 8.07 112.9 26,778,741 14,343,544 2,172,315 2,930,465 23,444 242 44,762 1,562 4,716 3,848
North Sulawesi 8.26 105.3 16,590,189 7,914,711 3,099,187 2,100,726 17,953 25 27,523 940 2,387 2,570
Jambi 8.39 111.5 29,997,861 13,230,999 8,560,251 2,139,605 14,958 13 32,600 1,505 3,344 3,679
West Java 9.18 116.0 287,300,000 180,800,000 67,458,422 23,150,873 1,593,322 1,640 458,110 2,191 46,030 25,897
South Sulawesi 9.54 118.5 58,497,013 31,917,331 4,517,817 6,916,650 58,419 70 85,028 1,148 8,432 5,843
North Sumatera 9.85 112.8 104,900,000 53,930,036 23,509,477 12,852,840 145,179 138 152,900 3,049 13,767 7,823
West Sulawesi 12.05 109.4 6,048,914 3,162,942 633,196 462,473 3,633 5,503 722 1,258 1,370
East Java 12.28 107.2 315,700,000 195,300,000 48,658,961 29,124,557 1,136,280 451 371,842 1,761 38,610 22,619
Southeast Sulawesi 12.77 111.0 17,072,946 8,305,035 1,016,566 2,094,240 9,099 40 16,913 906 2,448 2,521
Central Java 13.58 120.1 191,200,000 116,300,000 17,130,800 19,170,469 216,622 116 228,130 2,566 33,523 16,847
Central Sulawesi 13.61 107.5 17,919,383 10,023,919 978,886 2,200,137 8,926 374 22,802 1,619 2,831 2,588
South Sumatera 13.62 118.7 60,824,443 39,107,681 10,834,933 7,093,682 56,024 264 81,998 1,466 7,942 6,248
Lampung 14.21 109.5 47,448,565 28,072,462 9,545,312 4,255,972 52,718 39 52,359 1,703 8,026 4,568
DI Yogyakarta 14.55 114.3 19,884,020 11,980,973 1,069,562 1,877,136 31,240 16 27,664 690 3,637 3,529
Aceh 16.98 117.3 28,372,590 15,581,566 1,235,422 2,569,442 37,093 8 28,133 1,702 4,907 13,006
West Nusatenggara 17.05 111.0 18,343,241 12,952,309 1,012,192 1,804,827 16,671 138 24,839 1,772 4,774 2,803
Bengkulu 17.09 108.7 9,051,787 5,698,790 722,259 404,290 7,896 5 14,947 1,563 1,845 2,196
Gorontalo 17.41 117.7 5,193,951 3,147,586 44,448 617,530 3,822 1 8,537 433 1,116 1,301
Maluku 18.44 109.5 5,891,934 3,859,655 550,404 405,589 6,411 3 8,629 1,297 1,657 1,913
East Nusatenggara 19.6 113.0 13,526,994 10,527,339 252,458 1,433,348 8,945 4 16,832 1,737 5,037 3,023
West Papua 26.26 122.3 12,564,977 3,174,056 7,677,542 1,365,164 4,826 38 8,569 1,425 850 7,097
Papua 27.8 110.2 30,347,808 12,626,690 4,272,793 3,200,029 10,361 315 19,816 1,499 3,091 11,543
North Kalimantan N/A N/A 11,924,089 2,057,534 3,915,555 1,376,843 5,649 36 N/A N/A N/A 1,827
Notes: This study investigates the impact of BRI projects on the 11 GDP and economic outcomes. Population serves as a control in the regressions.
121

difference method, and then talk about how the instrument is constructed and test its assumptions.
3.4.1 Empirical Strategy I: Staggered Implementation
Effect of Project Implementation Initiation
As it is not possible to be sure whether any contractor hires local employees or buy
materials locally, the effect of the BRI projects when they are being implemented cannot
be clearly predicted. In order to investigate this effect, I have defined a treatment dummy
Dp,q equal to 1 if at least one project has been started at province p at or before quarter q,
equal to 0 otherwise. The baseline regression is in equation (3.1).
Yp,q = αq +βDp,q +ηp +εp,q (3.1)
where Yp,q denotes economic outcome of interest in province p at quarter q, including GDP,
GDP of household consumption, GDP of exports, GDP of construction, GDP of electricity and gas, permanent worker index (PWI), percentage of poor, investment realization,
length of road, commercial and rural bank loans, and state government expenditure, all
log-transformed except for PWI and percentage of poor. αq and ηp denote a quarter fixed
effect and a province fixed effect, respectively, used to control over time and provincial
invariant factors. β provides us with an average treatment effect (ATE) of the introduction
of a BRI project on economic factors. The standard errors of the estimators are at province
level.
However, this analysis ignores two most important considerations about how BRI
projects may affect the local economy. First, we need to separately clarify the effects
122

of a project being built with the effect after a project has been completed. When a project
is in progress, the project itself may create a local demand shock for the supply chain,
resulting in GDP growth. After the project is finished, it ameliorates the infrastructure
condition and reduces market friction, and is supposed to boost the economy in another
channel. Second, the heterogeneous effect of BRI projects, which depend on their value
and scale, should be analyzed further. In order to fix these two problems, I introduce the
subsections below.
Effect of Project Completion
In order to investigate the effect of the accomplished BRI projects on the local economy, it is ideal to have the finishing date of each project and use it as a treatment dummy.
Unfortunately, as I mentioned in the data section, this information is not fully available.
Through my readings of project introductions, most of the projects lasted for one or two
years. To overcome this difficulty and see how the project’s effect evolves over time, I
defined two dummy variables to proxy the treatment of accomplished BRI projects. First,
I defined a treatment dummy D lagged 1p,q equal to 1 if at least one project had been
started a year ago at province p at quarter q, and equal to 0 otherwise. Second, I defined
another treatment dummy D lagged 2p,q equal to 1 if at least one project had been started
two year ago at province p at quarter q, equal to 0 otherwise.
By replacing Dp,q in equation 3.1 with the dummies, D lagged 1p,q and D lagged 2p,q,
I ran regressions with Equation 3.2 and Equation 3.3 to see the effect of completed projects.
Compared with the results from Equation 3.1, these results also tell us whether the impact
123

increased or decreased over time. If the projects fulfilled their intended role, we should
see an improvement in the economic factors.
Yp,q = αq +βD lagged 1p,q +ηp +εp,q (3.2)
Yp,q = αq +βD lagged 2p,q +ηp +εp,q (3.3)
Effect of Multiple Projects
The starting of a BRI project may not be a good indicator for BRI projects treatment,
as we are neglecting the different financial value of each project, the most important and
fundamental measure of a project’s ability to influence economic outcomes. In order to
fix this problem, I created a dummy variable Multi Pro jp,q to indicate if there have been
relatively more BRI projects implemented in province p by quarter q. I want to use this
dummy variable to proxy the treatment of a big-value project, which I believe is legitimate
as several small projects should be able to have a similar effect on the economy as one
big-value project does. I used the median number of total projects across all the provinces
as a cutoff, which is 6. Precisely, MultiPro jq,p is set to be 1 if the total number of projects
in the province p by quarter q is greater than or equal to 6, and equal to 0 otherwise. I ran
Equation 3.4 to see the impact of multiple projects in a province.
Yp,q = αq +β1Dp,q +β2Dp,q × Multi Pro jp,q +ηp +εp,q (3.4)
β2 gives us the effect of having six or more projects on a province relative to the ones
which have one to five projects. Similar to the above subsection, I also replaced Dp,q with
D lagged 1p,q and D lagged 2p,q to check the effect of the completed projects.
124

Effect of Electricity- and Metal & Mineral-Related Projects
As we have seen in Figure 3.2, 98 projects are in the electricity-related industry, and
60 are in the metal- and mineral-related industries, out of 198 projects. As one of the
most significant entry barriers for firms in developing economies, access to electricity
can be particularly important for local economic growth. On the other hand, metal- and
mineral-related projects improve the ability to produce raw materials and exploit natural
resources, providing basic material for manufacturing production. Thus, I want to see the
economic consequence if a province has implemented more electricity-related projects or
more metal- and mineral-related projects than other regions. I define dummy variables,
Multi Elec Pro jp,q, equal to 1 if the province p at quarter p has implemented more than
three electricity-related projects, the median number of the total electricity-related projects
across the provinces that have any electricity-related projects, and Multi Metal Pro jp,q,
equal to 1 if the province p at quarter p has implemented more than two metal-related
projects, the median number of the total metal-related projects across the provinces that
have any such projects. The regression is
Yp,q = αq +β1Dp,q +β2Dp,q × Multi Elec Pro jp,q
+β3Dp,q × Multi Metal Pro jp,q +ηp +εp,q
(3.5)
β2 and β3 tells us the economic impact of having more electricity-related or metal-related
projects in a province, compared with the provinces which have fewer of these projects.
Again, I replaced Dp,q with D lagged 1p,q and D lagged 2p,q to check how the effects
change with project completion.
125

Identification Assumption of Difference-in-Difference
The key assumption for the difference-in-difference is the parallel pre-treatment trends
between control and treatment groups. Only if this assumption holds, β is an unbiased
estimator for the average treatment effect. To test this assumption, I drew a graph of
the outcome trend before and after the starting of any BRI project in the provinces, to
see if the pre-treatment trends are similar. The graph shows that the economic outcomes
grow at a stable rate across the provinces. Furthermore, I also listed the regression results
with the lagged one-year population size as a control to furtively show that the exclusion
assumption holds in my setting. If the estimators do not vary much with and without
including population control, the assumption is likely to hold, according to Angrist and
Pischke (2008).
Athey and Imbens (2018) discusses the assumptions for Difference-in-Difference settings in a staggered adoption, and they show that different treatment effects and interpretations depend on which assumption holds. I want to briefly talk about how my setting
adopts the three main assumptions discussed in their paper.
• Random Adoption Date: This assumption implies the marginal distribution of adoption
date does not change, and it is non-testable. Given the data availability, I cannot find
many pre-treatment variables that are not affected by the treatment to test this assumption, so I assume this assumption holds.
• No anticipation: This assumption stresses that the anticipation for a future adoption
should not affect the outcome today. This assumption may not hold in my setting. For
example, if the Indonesia local government anticipates there will be a future BRI project,
mainly financed by foreign investment, it may reduce the level of domestic investment
126

and invest in other regions. Nevertheless, BRI projects can work in the opposite way by
increasing the domestic investment to strengthen the economic advantage of the area.
No matter in which scenario, the anticipation of a BRI project may have an effect on the
local economy.
• Invariance to history: This assumption assumes the outcomes do not matter how long
the observation has been exposed to the adopted policy. This assumption does not hold
in my setting. When a BRI project is completed, even though it immediately reduces
the market friction, enterprises need time to accommodate and adjust their new strategy.
Thus, the momentum for economic growth from a project may be released gradually,
and the treatment effect is not invariant to history.
To conclude, my setting only holds the assumption of the random adoption date. In
terms of Athey and Imbens (2018), under the random adoption date assumption, the variance should be a conservative variance estimator. Thus, I use the variance estimator clustered at province level for inference in this study.
3.4.2 Empirical Strategy II: OLS and IV Regressions
Suppose there are circumstances that change around the same time cutoff with the
starting or completing of BRI projects in all the regions. In that case, my identification
assumption of using staggered implementation to test the causal effect of BRI projects
on Indonesian economic and social outcomes will be challenged. As the data for these
outcomes are collected at the province-level, subject to being influenced by state orders or
the geographic environment, they may be affected by some unforeseen factors, and it is
hard to rule out every possible coincidence.
127

For example, a simultaneous policy change with the occurrence of BRI projects can
confound my results. If the government has been developing policies to attract investment
and these policies were introduced at the same time a BRI project began, we cannot tell if
the economic impact is due to the BRI project or other investments induced by the policy.
To circumvent this alternative mechanism problem and complement my analysis on the
causal effect of the accomplished BRI projects on the growth of Indonesian economic
and social outcomes, an OLS and instrumental variable regressions are introduced in this
subsection to serve as my second empirical strategy.
OLS
To first uncover the association between the economic and social factors and the accomplished BRI projects, I ran the following regression:
Growthp,q = αq +β1BRI Pro jectsp,q−8 +β2Log(Population)p,q−4 +ηp +εp,q (3.6)
where Growthp,q denotes the economic factor’s growth of province p at quarter q. BRI Pro jectsp,q−8
is the measurement of BRI projects with a two-year lag. Log(Population)p,q−4 serves as
a control, standing for the one-year before logged population of province p. Similar to the
setting in staggered adoption, αq and ηp are province and quarter fixed effects. Standard
errors are clustered at province level.
I used two measurements for BRI projects. The first is the number of projects being
implemented at province q at quarter q − 8 (2-year lagged). This result should show an
association in the most straightforward way. The second is the cumulative number of
projects that have been implemented in the province up until two years ago. This result
128

attempts to uncover the long-term effect of the projects. Ideally, it will be preferable and
more insightful if I can use the projects’ financial value as the measurement. However, as
mentioned in the Data section, most of the projects did not disclose their financial aspects.
IV Regressions
It is evident that the estimators in the above regression will be biased, as it is most
likely suffer from omitted variable bias and reverse causality. There are many variables
correlated to both the local social and economic factors and the implementation of BRI
projects, inducing omitted variable bias. If the Chinese government wants to minimize
the risk of default projects, it can organize more projects for more wealthy areas. On
the other hand, if the Chinese government intends to help under-developed areas to build
infrastructure, a project may become prone to cluster in those regions with slower growth.
To mitigate this endogeneity problem, I followed Nunn and Qian (2014) and Dreher et al.
(2020) to construct an instrumental variable and ran the first-stage IV regressions.
BRI Pro jectsp,q = αq+γ1C Steelq−τ ∗Propp+γ2Log(Popultion)p,q−4+ηp+νp,q (3.7)
and second-stage regression:
Growthp,q = αq +β1BRI Pro jects ˆ
p,q−8 +β2Log(Population)p,q−4 +ηp +εp,q (3.8)
I constructed the instrumental variable for the BRI projects of province p at quarter q
by interacting the Chinese steel production at quarter q − τ (lagged τ quarters) with the
proportion of quarters in which province p has at least one BRI project among the total 32
quarters under this study (from 2012Q1 to 2019Q4), calculated as Propp =
1
32 ∑
32
q=1
1p,q,
129

where 1q,q stands for the indicator variable indicating at least one BRI project implemented
at province p at quarter q. I used τ = 2 to report the main results in the Results section and
conduct a robustness check by changing its value from 1 to 5 in Section 3.6. Province and
quarter fixed effects and population size are taken as controls.
In essence, this IV approach can be considered as a mimic to the difference-in-difference
approach, where we compare the differential effect of Chinese steel production on the
number of BRI projects in the provinces with a high probability of implementing projects
with the provinces with a low probability of implementing projects. The most fundamental assumption in this setting is that Chinese steel production does not differentially affect
the provinces with a different probability, other than through the channel of BRI projects.
In the main results, I used τ = 2 to have a half-year lagged and logged Chinese steel production, C Steelq−2, to predict the change of BRI projects. One may have concerns that
the exclusion restriction may be violated because the province with a high growth rate
may have a high probability of getting BRI projects. However, this factor has been controlled by the province fixed effect, as each province’s probability is constant. On the other
hand, as I also controlled the quarter fixed effect, the Chinese steel production should not
be correlated with the growth rate of Indonesian economic and social factors and can be
considered exogenous.
There are several noticeable features in my setting. Firstly, I only included the logged
and lagged population size as a control. Second, I used quarter-level data to demonstrate
a more precise change of the social and economic outcomes. Third, similar to OLS, I
reported two measurements of BRI projects: the number of projects implemented in the
quarter and the cumulative number of projects that have been implemented up until that
130

quarter. Fourth, to test the robustness of the effectiveness of using Chinese steel production
to predict the amount of BRI projects, I tested a different lag τ.
Akin to the difference-in-difference approach, this method requires the parallel trend
assumption to hold (Christian and Barrett (2017)). To test it, in Appendix Figure A4, I
follow Dreher et al. (2020) to draw the trends of several outcome variables for three groups
of provinces: the group with lowest 33% probability to have BRI projects commence
within the 32 quarters, the middle 33%, and the highest 33%. The figure shows that the
provinces with low and high BRI project probabilities exhibit similar trends in GDP, export
GDP, and household consumption GDP. These findings suggest that the parallel trends
assumption is likely to hold in this context, supporting the validity of the interaction-based
instrumental variable approach.
3.5 Results
3.5.1 Difference in Difference with Staggered Implementation
Effect of Project Implementation Initiation
Panel A of Table 3.3 and Table 3.4 give the regression results of Equation 3.1, with
and without the population control. We can see that the coefficient of beta barely changes
between those with and without population control. This reassures the unconfoundness
assumption in the empirical strategy. The implementation initiation of a BRI project does
not have an effect on the local GDP and its components. This result is not unexpected.
One single project should not substantially affect the local economy, especially without
analyzing the project’s financial value. This tells us the starting of a project does not
131

create a demand shock to the local supply chain, a major channel influencing the GDP and
its components.
Regarding other social and economic variables, the implementation of a BRI project
can be seen to decrease the amount of commercial and rural bank loans by 0.05%. This
decrease in local bank loans cannot be used as evidence to show that BRI projects do
not bring the local community into debt distress. As we don’t quantify the effect of BRI
projects on government debt, it does tell us if the project brings any benefit to the region,
but it does not place any great burden on local financial institutions.
Effect of Project Completion
As I discussed in the Empirical Strategy section, the effect of a project when it is in
progress and the effect of the project after it has been complated are completely different
and should be tested separately. To do so, I created the dummy variables using dates of one
year and two years after the starting date of the first project in a province and took them as
proxies for when a project was completed. The regression results are given in Panel B and
Panel C of Table 3.3 and Table 3.4.
After two years from the implementation of the first BRI project in a region, the GDP
grew 0.02% more than the areas where no BRI project had been implemented in the prior
two years. If we focus on the development of the coefficient with timing, the coefficient has
been reverting from a negative non-significant value in Panel A to a strong and positive
number in Panel C, demonstrating that the positive effect of a BRI project on GDP is
delivered slowly over time. I do not have sufficient observations to test the effect after
three years since the start of a project, but I conjecture the impact on the export GDP will
be further increased, following the trajectory. Furthermore, the effect on export GDP also
132

Table 3.3: ATE: the Effect of BRI Projects on GDP and GDP Components
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
GDP HHConsump Export Construction Elec&Gas GDP HHConsump Export Construction Elec&Gas
Panel A: Treatment equal to 1 after first BRI project being implemented
Treatment -0.00139 -0.00298 -0.270 -0.00972 -0.0327 -0.00137 -0.000998 -0.205 -0.00768 -0.0299
(-0.11) (-0.29) (-0.95) (-0.66) (-0.94) (-0.13) (-0.11) (-0.80) (-0.63) (-0.97)
Log(Poplation)q−4 0.263 0.460∗∗∗ 1.745 0.829∗∗∗ -0.591
(0.56) (3.26) (0.96) (3.15) (-0.88)
Quarter FE yes yes yes yes yes yes yes yes yes yes
Province FE yes yes yes yes yes yes yes yes yes yes
Observations 1224 1224 1179 1224 1224 1072 1072 1027 1072 1072
Panel B: 1 Year After 1st Proj. equal to 1 after 1 year when first BRI project implemented
1 Year After 1st Proj. 0.00819 -0.00212 -0.0862 -0.000128 -0.0319 0.00823 -0.000970 -0.0257 0.00248 -0.0284
(0.61) (-0.21) (-0.30) (-0.01) (-0.87) (0.74) (-0.11) (-0.10) (0.20) (-0.87)
Log(Poplation)q−4 0.272 0.460∗∗∗ 1.991 0.839∗∗∗ -0.586
(0.57) (3.26) (1.04) (3.17) (-0.87)
Quarter FE yes yes yes yes yes yes yes yes yes yes
Province FE yes yes yes yes yes yes yes yes yes yes
Observations 1224 1224 1179 1224 1224 1072 1072 1027 1072 1072
Panel C: 2 Year After 1st Proj. equal to 1 after 2 years when first BRI project implemented
2 Year After 1st Proj. 0.0202 -0.00396 0.254 0.00823 -0.0291 0.0199
∗ -0.00287 0.312 0.0130 -0.0282
(1.41) (-0.33) (1.02) (0.44) (-0.64) (1.76) (-0.31) (1.31) (0.96) (-0.65)
Log(Poplation)q−4 0.282 0.459∗∗∗ 2.189 0.848∗∗∗ -0.585
(0.60) (3.23) (1.15) (3.29) (-0.86)
Quarter FE yes yes yes yes yes yes yes yes yes yes
Province FE yes yes yes yes yes yes yes yes yes yes
Observations 1224 1224 1179 1224 1224 1072 1072 1027 1072 1072
t statistics in parentheses
∗ p
<
0.10, ∗∗
p
<
0.05, ∗∗∗
p
<
0.01
Notes: The dependent variable is a dummy variable, equal to 1 if at least one project has been started at province p at or before the cutoff stated by the different panel, equal
to 0 otherwise. Regressions from column 6 to 10 take population size as a control. All GDP and GDP component variables are in log. Standard errors are clustered at province
level.
133

Table 3.4: ATE: the Effect of BRI Projects on Other Social and Economic Outcomes
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)
PWI PercPoor Invest RoadLength BankLoans GovExp PWI PercPoor Invest RoadLength BankLoans GovExp
Panel A: Treatment equal to 1 after first BRI project being implemented
Treatment 1.128 0.124 0.107 -0.0483 -0.0573
∗ -0.00962 0.737 0.136 0.0683 -0.0264 -0.0494
∗ -0.0105
(1.51) (0.59) (0.47) (-0.83) (-2.02) (-0.26) (1.04) (0.75) (0.32) (-0.50) (-2.03) (-0.34)
Log(Poplation)q−4 -12.18 -1.747 3.790 2.820∗∗∗ 1.350
∗ 0.661
(-0.43) (-0.44) (1.13) (3.23) (1.89) (0.33)
Quarter FE yes yes yes yes yes yes yes yes yes yes yes yes
Province FE yes yes yes yes yes yes yes yes yes yes yes yes
Observations 1122 1188 1180 1072 1203 1216 990 1056 1036 936 1071 1072
Panel B: 1 Year After 1st Proj. equal to 1 after 1 year when first BRI project implemented
1 Year After 1st Proj. 1.231 -0.00158 0.166 -0.0533 -0.0290 0.0283 0.851 0.0174 0.113 -0.0225 -0.0230 0.0238
(1.49) (-0.01) (0.67) (-1.08) (-0.78) (0.78) (1.12) (0.09) (0.50) (-0.51) (-0.70) (0.79)
Log(Poplation)q−4 -12.28 -1.881 3.821 2.831∗∗∗ 1.381
∗ 0.694
(-0.43) (-0.48) (1.16) (3.27) (1.90) (0.34)
Quarter FE yes yes yes yes yes yes yes yes yes yes yes yes
Province FE yes yes yes yes yes yes yes yes yes yes yes yes
Observations 1122 1188 1180 1072 1203 1216 990 1056 1036 936 1071 1072
Panel C: 2 Year After 1st Proj. equal to 1 after 2 years when first BRI project implemented
2 Year After 1st Proj. 1.661 -0.0911 0.483 -0.0248 -0.000226 0.0807∗∗ 1.343 -0.0702 0.433 -0.00804 0.00498 0.0689∗∗
(1.51) (-0.32) (1.32) (-0.53) (-0.00) (2.13) (1.33) (-0.29) (1.31) (-0.18) (0.12) (2.39)
Log(Poplation)q−4 -12.67 -1.956 4.092 2.853∗∗∗ 1.407
∗ 0.733
(-0.44) (-0.51) (1.30) (3.26) (1.95) (0.37)
Quarter FE yes yes yes yes yes yes yes yes yes yes yes yes
Province FE yes yes yes yes yes yes yes yes yes yes yes yes
Observations 1122 1188 1180 1072 1203 1216 990 1056 1036 936 1071 1072
t statistics in parentheses
∗ p
<
0.10, ∗∗
p
<
0.05, ∗∗∗
p
<
0.01
Notes: The dependent variable is a dummy variable, equal to 1 if at least one project has been started at province p at or before the cutoff stated by the different panel, equal
to 0 otherwise. Regressions from column 6 to 10 take population size as a control. All dependent variables, except for PWI and PercPoor, are in log. PWI stands for permanent
worker index in construction sector, signaling the job growth in construction industry, with each province’s 2010 value as its benchmark 100 to calculate the index of each
quarter. PercPoor indicates the percentage of poor people. Invest denotes the foreign investment realization. RoadLength gives the total highway length. BankLoans stands
for the amount of local commercial and rural bank loan. GovExp contains the local government actual expenditure. Standard errors are clustered at province level.
134

demonstrates a positive result, even though not as statistically significant as the coefficient
of the overall GDP. The projects increase the local GDP of exports by 0.3%. On the other
hand, we see that completed BRI projects result in government expenditure increasing
0.07% in these areas. Government expenditure may serve as a media variable for the BRI
projects to affect GDP, which I cannot check in this subsection so it will be taken care of
using my second empirical strategy instrumental variable.
Using a dummy variable indicating one year after the first project commences does not
produce any coefficient with statistical significance. The reason for this may be that the
projects are in the transition period from development to fully functioning so their effects
are not fully released.
Effect of Multiple Projects in a Province
To see the effect of big-value projects on the economy without having the data of
project financial value, I used a dummy variable indicating that the province has implemented 6 or more projects as a proxy. Panel A, B, and C of Table 3.5 and Table 3.6 include
Dp,q, D lagged 1p,q and D lagged 2p,q and the indicator of having 6 or more projects as
the dependent variables respectively to isolate the effect of multiple projects. Surprisingly,
the big-value projects seem not have a strong economic impact on the local community. If
we focus on the coefficients on the GDP components, none of them are statistically significant, besides the one on GDP of electricity and gas in Panel A. It is unclear why the
big-value projects decrease the GDP of electricity and gas when under development.
135

Table 3.5: ATE of Multiple Projects: the Effect of BRI Projects on GDP and GDP Components
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
GDP HHConsump Export Construction Elec&Gas GDP HHConsump Export Construction Elec&Gas
Panel A: Treatment equal to 1 after first BRI project being implemented
Treatment -0.00297 -0.00263 -0.290 -0.00831 -0.0275 -0.00272 -0.000551 -0.225 -0.00590 -0.0258
(-0.24) (-0.26) (-1.07) (-0.58) (-0.82) (-0.27) (-0.06) (-0.93) (-0.51) (-0.87)
Multiple Proj. 0.0248 -0.00545 0.327 -0.0223 -0.0818
∗ 0.0218 -0.00724 0.331 -0.0289 -0.0674
(0.94) (-0.38) (0.73) (-0.71) (-1.70) (0.95) (-0.57) (0.81) (-1.21) (-1.54)
Log(Poplation)q−4 0.255 0.463∗∗∗ 1.667 0.840∗∗∗ -0.565
(0.54) (3.25) (0.94) (3.34) (-0.85)
Observations 1224 1224 1179 1224 1224 1072 1072 1027 1072 1072
Panel B: Dummy variable, 1 Year After 1st Proj., equal to 1 after 1 year when first BRI project implemented
1 Year After 1st Proj. 0.00404 -0.00111 -0.135 0.00466 -0.0172 0.00459 0.000455 -0.0746 0.00888 -0.0162
(0.31) (-0.11) (-0.53) (0.30) (-0.49) (0.45) (0.05) (-0.32) (0.74) (-0.53)
Multiple Proj. 0.0224 -0.00544 0.320 -0.0258 -0.0791 0.0193 -0.00754 0.313 -0.0338 -0.0647
(0.82) (-0.35) (0.74) (-0.79) (-1.68) (0.83) (-0.57) (0.80) (-1.35) (-1.53)
Log(Poplation)q−4 0.263 0.464∗∗∗ 1.890 0.856∗∗∗ -0.553
(0.56) (3.27) (1.01) (3.42) (-0.83)
Observations 1224 1224 1179 1224 1224 1072 1072 1027 1072 1072
Panel C: Dummy variable, 2 Year After 1st Proj., equal to 1 after 2 years when first BRI project implemented
2 Year After 1st Proj. 0.0109 -0.000595 0.132 0.00443 -0.00245 0.0114 0.000550 0.200 0.0122 -0.00451
(0.66) (-0.04) (0.66) (0.21) (-0.05) (0.97) (0.05) (1.07) (0.91) (-0.10)
Multiple Proj. 0.0310 -0.0112 0.570 0.0127 -0.0887 0.0279 -0.0112 0.513 0.00274 -0.0774
(1.10) (-0.82) (1.14) (0.38) (-1.13) (1.24) (-1.10) (1.13) (0.12) (-1.11)
Log(Poplation)q−4 0.278 0.461∗∗∗ 2.042 0.848∗∗∗ -0.572
(0.60) (3.23) (1.09) (3.29) (-0.85)
Observations 1224 1224 1179 1224 1224 1072 1072 1027 1072 1072
t statistics in parentheses
∗
p
<
0.10, ∗∗
p
<
0.05, ∗∗∗
p
<
0.01
Notes: The first dependent variable is a dummy variable, equal to 1 if at least one project has been started at province p at or before the cutoff stated by the different panel,
equal to 0 otherwise. Multi.Pro j. is a interaction, equal to the first dependent variable times a dummy variable equal to 1 if the total number of project in the province p by
quarter q is greater than or equal to 6, and equal to 0 otherwise. This variable serves as a proxy for a big value project. All GDP and GDP component variables are in log. All
regressions include province and quarter fixed effects. Standard errors are clustered at province level.
136

Table 3.6: ATE of Multiple Projects: the Effect of BRI Projects on Other Social and Economic Outcomes
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)
PWI PercPoor Invest RoadLength BankLoans GovExp PWI PercPoor Invest RoadLength BankLoans GovExp
Panel A: Treatment equal to 1 after first BRI project being implemented
Treatment 1.114 0.112 0.0983 -0.0479 -0.0553
∗ -0.0102 0.717 0.124 0.0607 -0.0264 -0.0470
∗ -0.0109
(1.56) (0.55) (0.44) (-0.82) (-1.99) (-0.27) (1.06) (0.71) (0.29) (-0.50) (-2.03) (-0.35)
Multiple Proj. 0.306 0.169 0.150 -0.0206 -0.0318 0.00838 0.409 0.186 0.138 -0.00122 -0.0379 0.00682
(0.14) (0.60) (0.61) (-0.49) (-0.92) (0.19) (0.22) (0.73) (0.63) (-0.03) (-1.18) (0.17)
Log(Poplation)q−4 -12.32 -1.822 3.733 2.819∗∗∗ 1.365
∗ 0.658
(-0.44) (-0.46) (1.12) (3.23) (1.95) (0.33)
Observations 1122 1188 1180 1072 1203 1216 990 1056 1036 936 1071 1072
Panel B: 1 Year After 1st Proj. equal to 1 after 1 year when first BRI project implemented
1 Year After 1st Proj. 1.244 -0.0404 0.147 -0.0534 -0.0232 0.0295 0.833 -0.0242 0.0934 -0.0231 -0.0156 0.0250
(1.51) (-0.19) (0.56) (-1.06) (-0.62) (0.74) (1.13) (-0.14) (0.40) (-0.51) (-0.50) (0.78)
Multiple Proj. -0.0945 0.208 0.103 0.00168 -0.0316 -0.00658 0.127 0.219 0.108 0.00733 -0.0390 -0.00605
(-0.04) (0.77) (0.38) (0.04) (-0.91) (-0.13) (0.07) (0.91) (0.46) (0.17) (-1.27) (-0.14)
Log(Poplation)q−4 -12.33 -1.992 3.766 2.831∗∗∗ 1.401
∗ 0.697
(-0.44) (-0.51) (1.13) (3.26) (1.98) (0.35)
Observations 1122 1188 1180 1072 1203 1216 990 1056 1036 936 1071 1072
Panel C: 2 Year After 1st Proj. equal to 1 after 2 years when first BRI project implemented
2 Year After 1st Proj. 2.007
∗ -0.153 0.541 -0.0231 -0.00257 0.0816
∗ 1.613 -0.135 0.484 -0.00618 0.00495 0.0726∗∗
(1.81) (-0.54) (1.21) (-0.49) (-0.05) (1.83) (1.57) (-0.58) (1.23) (-0.13) (0.13) (2.34)
Multiple Proj. -2.151 0.207 -0.194 -0.0419∗∗ 0.00781 -0.00297 -1.630 0.213 -0.170 -0.0446 0.0000993 -0.0121
(-0.69) (0.79) (-0.42) (-2.11) (0.19) (-0.07) (-0.58) (0.87) (-0.44) (-1.46) (0.00) (-0.34)
Log(Poplation)q−4 -12.50 -1.991 4.119 2.854∗∗∗ 1.407
∗ 0.734
(-0.43) (-0.52) (1.31) (3.25) (1.95) (0.37)
Observations 1122 1188 1180 1072 1203 1216 990 1056 1036 936 1071 1072
t statistics in parentheses
∗ p
< 0.10, ∗∗
p
< 0.05, ∗∗∗
p
< 0.01
Notes: The first dependent variable is a dummy variable, equal to 1 if at least one project has been started at province p at or before the cutoff stated by the different
panel, equal to 0 otherwise. Multi.Pro j. is a dummy variable, equal to 1 if the total number of project in the province p by quarter q is greater than or equal to 6, and equal
to 0 otherwise, proxy for a big value project. All dependent variables, except for PWI and PercPoor, are in log. PWI stands for permanent worker index in construction
sector, indicating the job growth in construction industry, with each province’s 2010 value as its benchmark 100 to calculate the index of each quarter. PercPoor indicates the
percentage of poor people in each province. Invest denotes the foreign investment realization. RoadLength is the total highway length. BankLoans stands for the amount of
local commercial and rural bank loan. GovExp contains the local government actual expenditure. All regressions include province and quarter fixed effects. Standard errors
are clustered at province level.
137

Regarding to the coefficients on the other social and economic factors, the provinces
with fewer projects experience higher growth on the permanent worker index (PWI), indicating the better expanding construction sector. Recapping the result from Table 3.3 and
Table 3.4, the government expenditure has grown in these provinces while we don’t see
this situation in the provinces with more projects.
All of these results seems puzzling. One conjecture for the confusing results in this
subsection is due to the timing issue: Until 2018, most of the provinces had fewer than
six BRI projects underway or completed. The limited observations of after-treatment outcomes jeopardize the results.
Effect of Electricity- and Metal & Mineral-Related Projects
To see the heterogeneous effect of having multiple projects across the industries, I ran
the regression for Equation 3.5 and show the results in Table 3.7 and Table 3.8. In general,
metal- and mineral-related projects serve as a better driver for economic growth. Specifically, compared with other provinces that have at least one BRI project, the provinces with
more than 2 metal- and mineral-related projects experience a 0.04% increase in overall
GDP and 0.9% increase in GDP of exports. In addition, these metal- and mineral-related
projects result in an increase of 1.163% in investment realization in the local economy.
One possible explanation for this is that in this industry the project value is usually substantial and takes a large proportion of the total investment of that period. This coefficient
on investment realization of metal- and mineral-related is robust to different treatment
dummies.
138

Table 3.7: ATE across Industries: the Effect of BRI Projects on GDP and GDP Components
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
GDP HHConsump Export Construction Elec&Gas GDP HHConsump Export Construction Elec&Gas
Panel A: Treatment equal to 1 after first BRI project being implemented
Treatment -0.00323 -0.00259 -0.332 -0.0101 -0.0319 -0.00276 -0.000383 -0.266 -0.00710 -0.0287
(-0.28) (-0.26) (-1.27) (-0.69) (-0.86) (-0.31) (-0.04) (-1.14) (-0.61) (-0.89)
Multi. Elec. Proj. 0.00450 -0.0167 0.0798 -0.0266 0.0271 0.00824 -0.0124 0.102 -0.0196 0.0225
(0.20) (-1.31) (0.33) (-1.24) (0.70) (0.41) (-1.01) (0.44) (-0.94) (0.62)
Multi Metal proj. 0.0140 0.00842 0.478 0.0224 -0.0277 0.00788 0.00210 0.462 0.00704 -0.0265
(0.46) (0.66) (1.64) (1.01) (-0.68) (0.30) (0.22) (1.57) (0.37) (-0.70)
Log(Poplation)q−4 0.257 0.449∗∗∗ 1.236 0.806∗∗∗ -0.539 (0.55) (3.09) (0.67) (3.05) (-0.77)
Observations 1224 1224 1179 1224 1224 1072 1072 1027 1072 1072
Panel B: 1 Year After 1st Proj. equal to 1 after 1 year when first BRI project implemented
1 Year After 1st Proj. 0.00138 -0.000683 -0.263 -0.000959 -0.0346 0.00228 0.000665 -0.197 0.00350 -0.0292
(0.11) (-0.07) (-1.05) (-0.06) (-0.75) (0.26) (0.08) (-0.84) (0.27) (-0.72)
Multi. Elec. Proj. 0.00546 -0.0148 0.120 -0.0300 0.0331 0.00917 -0.0103 0.125 -0.0222 0.0271
(0.24) (-1.13) (0.45) (-1.25) (0.77) (0.46) (-0.83) (0.49) (-0.99) (0.68)
Multi Metal proj. 0.0238 0.00539 0.636∗∗ 0.0264 -0.0141 0.0175 0.000912 0.613∗∗ 0.0122 -0.0166
(0.84) (0.40) (2.07) (1.11) (-0.29) (0.76) (0.09) (2.05) (0.68) (-0.36)
Log(Poplation)q−4 0.261 0.454∗∗∗ 1.678 0.814∗∗∗ -0.554 (0.55) (3.16) (0.91) (3.02) (-0.80)
Observations 1224 1224 1179 1224 1224 1072 1072 1027 1072 1072
Panel C: 2 Year After 1st Proj. equal to 1 after 2 years when first BRI project implemented
2 Year After 1st Proj. 0.00435 -0.00117 -0.103 -0.00224 -0.0449 0.00475 -0.00114 -0.0244 0.00275 -0.0376
(0.31) (-0.10) (-0.51) (-0.11) (-0.69) (0.47) (-0.12) (-0.13) (0.20) (-0.63)
Multi. Elec. Proj. -0.0000340 -0.0188 -0.0851 -0.0216 0.0433 0.00565 -0.0105 -0.140 -0.00778 0.0346
(-0.00) (-1.21) (-0.30) (-0.87) (0.83) (0.31) (-0.80) (-0.51) (-0.40) (0.73)
Multi Metal proj. 0.0428
∗ 0.00921 0.938∗∗∗ 0.0475 0.00429 0.0359
∗ 0.00460 0.931∗∗∗ 0.0346
∗ -0.00503
(1.73) (0.56) (3.38) (1.67) (0.06) (1.97) (0.33) (3.43) (1.85) (-0.08)
Log(Poplation)q−4 0.278 0.452∗∗∗ 1.898 0.837∗∗∗ -0.565 (0.60) (3.16) (1.01) (3.27) (-0.83)
Observations 1224 1224 1179 1224 1224 1072 1072 1027 1072 1072
t statistics in parentheses
∗ p
< 0.10, ∗∗
p
< 0.05, ∗∗∗
p
< 0.01
Notes: The first dependent variable is a dummy variable, equal to 1 if at least one project has been started at province p at or before the cutoff stated by the different panel,
equal to 0 otherwise. Multi.Pro j. is a dummy variable, equal to 1 if the total number of project in the province p by quarter q is greater than or equal to 6, and equal to 0
otherwise, proxy for a big value project. Multi.Elec.Pro j. and Multi.Metal Pro j.” are set equal to 1 if the province has implemented 3 eclectic projects or 2 metal and mineral
projects, respectively. All GDP and GDP component variables are in log. All regressions include province and quarter fixed effects. Standard errors are clustered at province
level.
139

Table 3.8: ATE across Industries: the Effect of BRI Projects on Other Social and Economic Outcomes
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)
PWI PercPoor Invest RoadLength BankLoans GovExp PWI PercPoor Invest RoadLength BankLoans GovExp
Panel A: Treatment equal to 1 after first BRI project being implemented
Treatment 1.186 0.151 0.0250 -0.0531 -0.0567
∗ -0.00632 0.794 0.151 0.000290 -0.0287 -0.0475∗∗ -0.00688
(1.69) (0.76) (0.12) (-0.90) (-2.03) (-0.18) (1.22) (0.85) (0.00) (-0.55) (-2.14) (-0.25)
Multi. Elec. Proj. -1.992 -0.274 0.194 -0.0128 -0.0274 0.0560 -1.960 -0.295 0.179 0.00802 -0.0135 0.0513
(-1.23) (-0.74) (0.56) (-0.32) (-0.63) (1.27) (-1.39) (-0.89) (0.51) (0.28) (-0.37) (1.24)
Multi. Metal proj. 1.025 -0.0473 0.682∗∗ 0.0654 0.0140 -0.0711 0.904 0.0635 0.582∗∗ 0.0220 -0.00863 -0.0668
(0.60) (-0.13) (2.51) (1.08) (0.28) (-0.89) (0.62) (0.18) (2.13) (0.58) (-0.17) (-0.91)
Log(Poplation)q−4 -13.71 -2.010 3.052 2.807∗∗∗ 1.354
∗ 0.790
(-0.52) (-0.50) (0.91) (3.21) (1.98) (0.40)
Observations 1122 1188 1180 1072 1203 1216 990 1056 1036 936 1071 1072
Panel B: 1 Year After 1st Proj. equal to 1 after 1 year when first BRI project implemented
1 Year After 1st Proj. 1.690
∗ 0.0940 -0.0788 -0.0572 -0.0304 0.0303 1.308
∗ 0.0857 -0.108 -0.0291 -0.0228 0.0258
(1.92) (0.48) (-0.34) (-1.08) (-0.81) (0.77) (1.70) (0.48) (-0.52) (-0.64) (-0.75) (0.89)
Multi. Elec. Proj. -1.529 -0.325 0.353 0.0145 -0.0433 0.0400 -1.473 -0.346 0.348 0.0239 -0.0272 0.0376
(-0.98) (-0.88) (1.05) (0.47) (-0.97) (0.95) (-1.09) (-1.01) (1.03) (0.89) (-0.69) (0.96)
Multi. Metal proj. -0.877 -0.147 0.748∗∗∗ 0.00940 0.0385 -0.0385 -0.924 -0.0178 0.655∗∗ 0.0167 0.0193 -0.0357
(-0.58) (-0.40) (2.84) (0.24) (0.82) (-0.55) (-0.79) (-0.05) (2.44) (0.41) (0.44) (-0.64)
Log(Poplation)q−4 -12.06 -2.056 3.389 2.836∗∗∗ 1.347
∗ 0.750
(-0.43) (-0.53) (1.09) (3.26) (1.87) (0.38)
Observations 1122 1188 1180 1072 1203 1216 990 1056 1036 936 1071 1072
Panel C: 2 Year After 1st Proj. equal to 1 after 2 years when first BRI project implemented
2 Year After 1st Proj. 2.813∗∗ 0.158 0.102 -0.0215 -0.00940 0.0694 2.549∗∗ 0.127 0.0905 -0.0202 -0.00703 0.0614
∗
(2.46) (0.68) (0.27) (-0.43) (-0.20) (1.53) (2.38) (0.60) (0.28) (-0.42) (-0.18) (1.77)
Multi. Elec. Proj. 0.614 -0.283 -0.301 -0.00123 -0.108
∗ -0.0202 0.125 -0.338 -0.266 0.0449 -0.0821 -0.00197
(0.35) (-0.67) (-0.80) (-0.09) (-1.69) (-0.58) (0.08) (-0.83) (-0.75) (1.26) (-1.33) (-0.06)
Multi. Metal proj. -3.443
∗ -0.423 1.300∗∗∗ -0.00889 0.121 0.0486 -3.267
∗ -0.236 1.163∗∗∗ 0.00775 0.105 0.0221
(-2.00) (-1.07) (5.70) (-0.40) (1.66) (1.08) (-2.03) (-0.60) (4.69) (0.21) (1.54) (0.78)
Log(Poplation)q−4 -13.01 -2.089 3.710 2.870∗∗∗ 1.340
∗ 0.727
(-0.45) (-0.56) (1.25) (3.25) (1.86) (0.36)
Observations 1122 1188 1180 1072 1203 1216 990 1056 1036 936 1071 1072
Notes: The first dependent variable is a dummy variable, equal to 1 if at least one project has been started at province p at or before the cutoff stated by the different panel,
equal to 0 otherwise. Multi.Pro j. is a dummy variable, equal to 1 if the total number of project in the province p by quarter q is greater than or equal to 6, and equal to
0 otherwise, proxy for a big value project. Multi.Elec.Pro j. and Multi.Metal Pro j.” are set equal to 1 if the province has implemented 3 eclectic projects or 2 metal and
mineral projects, respectively. All dependent variables, except for PWI and PercPoor, are in log. PWI stands for permanent worker index in construction sector, indicating the
job growth in construction industry, with each province’s 2010 value as its benchmark 100 to calculate the index of each quarter. PercPoor indicates the percentage of poor
people in each province. BankLoans stands for the amount of local commercial and rural bank loan. GovExp contains the local government actual expenditure. All regressions
include province and quarter fixed effects. Standard errors are clustered at province level. t statistics in parentheses * p < 0.10, ** p < 0.05, *** p < 0.01
140

On the other hand, the result of the electricity-related projects is confusing as they
do not exert a significant effect on the local economy. Neither the coefficients on overall
GDP and GDP components nor the ones on other social and economic factors demonstrate
statistical significance. One explanation is that even though these projects can improve
access to electricity, they don’t mitigate other market friction and this friction constitutes
more critical barriers for local economic growth.
3.5.2 OLS and IV Regressions
To complement and substantiate my results and overcome the potential endogeneity
and reverse causality problem, I employed the instrumental variable as my second empirical strategy to demonstrate the causal effect of BRI projects on the growth rate of the
GDP components and other variables. In Table 3.9, I first list the results of OLS as stated
in Equation 3.6 and then that of reduced-form regression. As mentioned in the Empirical
Strategy section, I used two measurements of BRI projects -the number of projects implemented in the quarter and cumulative project number up until that quarter. Few coefficients
are significant in the table, showing a very week association between the dependent and
independent variables. One exception is that one additional cumulative project is associated with a 0.372 increase in the growth rate of the permanent worker index. Regarding the
reduced form regression, the association between the instrumental variable and dependent
variables is weak and small. Undoubtedly, these regressions suffer from endogeneity and
reverse causality, resulting in biased and insignificant estimators.
I show the 2SLS results of Equation 3.8 in Table 3.10. All the regressions include
lagged and logged population size, quarter fixed effect, and province fixed effect as controls. The Cragg-Donald F-statistics are reported. Even though the 2SLS regressions using
141

Table 3.9: OLS: the Association between the Number of BRI Projects and Social and Economic Outcomes
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
GDP HHConsump Export Construction Elec&Gas PWI PercPoor Invest RoadLength BankLoans GovExp
Panel A: Using project number implemented in the quarter with a 2-year lag as a measurement of BRI projects
Project Number 0.00476 -0.000923 0.0350 0.00851 0.00596 -0.161 0.000420 -0.0596 0.0271 0.00429 0.00349
(1.35) (-0.39) (0.34) (0.87) (0.53) (-0.34) (0.01) (-0.41) (1.55) (0.37) (0.42)
Log(Poplation)q−4 -0.124∗∗∗ 0.0353 0.460 -0.160∗∗ 0.0598 16.24 -0.219 2.086 0.522 0.127 -0.629
∗
(-3.50) (0.72) (0.38) (-2.15) (0.24) (1.50) (-0.21) (1.12) (0.82) (0.47) (-1.90)
Quarter FE yes yes yes yes yes yes yes yes yes yes yes
Province FE yes yes yes yes yes yes yes yes yes yes yes
Observations 1072 1072 1027 1072 1072 990 1056 1032 936 1067 1072
Panel B: Using cumulative project number until the quarter with a 2-year lag as a measurement of BRI projects
Cumu. Proj. Num. 0.00345 0.0000489 0.0654 0.000593 0.00542 0.372
∗ -0.0415 0.0808 0.0224 -0.00864 -0.00698
(1.13) (0.04) (0.54) (0.12) (0.57) (1.75) (-1.05) (0.88) (1.48) (-1.42) (-0.71)
Log(Poplation)q−4 -0.118∗∗∗ 0.0356 0.542 -0.161∗∗ 0.0696 16.85 -0.304 2.267 0.534 0.108 -0.644
∗
(-3.62) (0.73) (0.44) (-2.12) (0.28) (1.56) (-0.29) (1.27) (0.86) (0.38) (-1.98)
Quarter FE yes yes yes yes yes yes yes yes yes yes yes
Province FE yes yes yes yes yes yes yes yes yes yes yes
Observations 1072 1072 1027 1072 1072 990 1056 1032 936 1067 1072
Panel C: Reduced form regression
C Steelq−τ ×Propp 0.0000120 0.00000540∗∗ 0.000830∗∗ -0.0000348 -0.00000375 0.000451 0.0000886 0.000186 0.0000398 -0.00000173 0.0000298
(0.97) (2.05) (2.53) (-1.29) (-0.14) (0.66) (0.51) (0.67) (0.61) (-0.12) (1.13)
Log(Poplation)q−4 -0.123∗∗ -0.0819∗∗ -0.764 -0.179
∗ 0.430 18.83 -0.592 1.410 0.347 0.121 -0.871
(-2.14) (-2.14) (-0.39) (-2.03) (0.75) (1.09) (-0.34) (0.55) (0.34) (0.77) (-1.20)
Quarter FE yes yes yes yes yes yes yes yes yes yes yes
Province FE yes yes yes yes yes yes yes yes yes yes yes
Observations 742 742 697 742 742 660 726 719 606 737 742
t statistics in parentheses
∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
Notes: OLS results listed above are suffered from endogeneity and reverse causality. Standard errors are clustered at province level.
142

Table 3.10: 2SLS: The causal Effect of BRI Projects on the Growth of GDP and other Social and Economic Outcomes
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
GDP HHConsump Export Construction Elec&Gas PWI PercPoor Invest RoadLength BankLoans GovExp
Panel A: Using project number implemented in the quarter with a 2-year lag as a measurement of BRI projects
Project Number 0.0918 0.0413
∗ 6.531∗∗ -0.266 -0.0287 4.708 0.676 1.410 0.382 -0.0132 0.228
(1.01) (1.83) (2.29) (-1.19) (-0.14) (0.65) (0.53) (0.72) (0.64) (-0.12) (1.13)
Log(Poplation)q−4 -0.0985 -0.0710
∗ 0.542 -0.249 0.422 20.44 -0.436 1.767 0.369 0.117 -0.811
(-1.32) (-1.92) (0.26) (-1.60) (0.79) (1.25) (-0.25) (0.81) (0.41) (0.75) (-1.13)
Quarter FE yes yes yes yes yes yes yes yes yes yes yes
Province FE yes yes yes yes yes yes yes yes yes yes yes
Observations 742 742 697 742 742 660 726 719 606 737 742
Cragg-Donald F 7.791 7.791 8.730 7.791 7.791 5.539 7.668 7.642 5.207 7.802 7.791
Panel B: Using cumulative project number until the quarter with a 2-year lag as a measurement of BRI projects
Cumu. Proj. Num. 0.0237 0.0107
∗ 2.079∗∗ -0.0689 -0.00743 1.302 0.175 0.366 0.157 -0.00342 0.0590
(0.98) (1.83) (2.27) (-1.26) (-0.14) (0.67) (0.54) (0.72) (0.62) (-0.13) (1.09)
Log(Population)q−4 -0.0709 -0.0585
∗ 2.062 -0.330 0.413 20.59 -0.232 2.194 0.425 0.113 -0.742
(-0.71) (-1.70) (0.67) (-1.62) (0.78) (1.32) (-0.12) (1.02) (0.51) (0.67) (-0.95)
Quarter FE yes yes yes yes yes yes yes yes yes yes yes
Province FE yes yes yes yes yes yes yes yes yes yes yes
Observations 742 742 697 742 742 660 726 719 606 737 742
Cragg-Donald F 13.91 13.91 12.20 13.91 13.91 5.718 13.84 13.41 3.827 13.88 13.91
t statistics in parentheses
∗ p
<
0.10, ∗∗
p
<
0.05, ∗∗∗
p
<
0.01
Notes: Dependent variables Pro j.Num. and Cumu.Pro j.Num are the number of project implemented at the quarter 2 years ago and the cumulative amount of projects until
the quarter 2 years ago, respectively. 2SLS results listed above use as the instrumental variable the intersection of lagged 2-quarter Chinese steel production and the province’s
probability of having BRI projects. Standard errors are clustered at province level.
143

the number of projects may suffer from weak instrument problems, marked by the F statistics less than 10, the positive and significant estimators of GDP of household consumption
and GDP of exports could suggest a beneficial role of BRI projects. For each BRI project
implemented 2 years ago, it causes the current period GDP of household consumption
growth rate to increase 0.0413 percentage point. Surprisingly, the causal effect of BRI
projects on GDP of exports is massive: a 2-year-old BRI project accelerates the growth
rate of GDP of exports by 6.531 percent points. Even though the coefficient on overall
GDP is insignificant, the positive estimator with t-statistics of 1.01 is not meaningless.
It infers a constructive effect of a BRI project on overall GDP growth. Specifically, one
project results in the local GDP growth rate improving by 0.0918 percentage points. This
number is comparable with the result from Dreher et al. (2021), indicating the effect of
between 0.41 and 1.49 percentage points of Chinese aid projects on the recipient country’s
GDP growth. On the other hand, the coefficient of BRI projects on commercial and rural
bank loans is not significant, a piece of evidence showing the projects do not affect the
original growth path of banking credit expansion. All these statistics demonstrate that BRI
projects benefit the local economy while not creating a considerable financial burden to it.
The results from using a two-years lagged cumulative project number as the measurement are reassuring, given that the first-stage regressions do not suffer from weak
instrument problems, as in the case above. By Panel B of Table 3.10, one more 2-year
lagged cumulative project causes the growth rate of the GDP of exports and the growth
rate of GDP of household consumption to improve 2.079 percentage points and 0.0107
percentage point, respectively. The magnitude of the estimators are smaller when using
144

cumulative projects as the measurement. One possible reason for this could be the diminishing return of the BRI projects: one additional project in the area which already
has several projects does not produce as great a benefit as in the area where few projects
have been carried out. To conclude, the positive causal effect of BRI projects on GDP of
exports uncovered using the instrumental variable approach is consistent with the results
from using difference-in-difference with staggered implementation.
3.6 Robustness Check
To test the credibility of the results by using the interaction of Chinese steel production and the probability of receiving BRI projects as the instrumental variable, I ran several robustness check regressions by changing the lag period of Chinese steel production.
As mentioned in the Empirical Strategy section, the 2SLS results reported in the Results
section use the half-year (two quarters) lagged Chinese steel production to construct the
instrumental variable. The intuition behind this is that I believe the previous half-year Chinese steel production is able to anticipate the volume of overseas Chinese BRI projects.
In Table 3.11, I list the regressions applying Chinese steel production lagged from one
quarter to five quarters to demonstrate the robustness of the results.
145

Table 3.11: Robustness Check: 2SLS Regressions with Different Lags of Chinese Steel Production
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
HHConsump HHConsump HHConsump HHConsump HHConsump Export Export Export Export Export
Panel A: Using project number implemented in the quarter with a 2-year lag as a measurement of BRI projects
Project Number 0.0282 0.0413
∗ 0.0391 0.0424∗∗ 0.0135 9.070 6.531∗∗ 10.32 3.832 2.324
(1.28) (1.83) (1.33) (2.05) (1.17) (1.59) (2.29) (1.01) (1.45) (0.83)
Log(Population)q−4 -0.0667
∗ -0.0710
∗ -0.0843∗∗ -0.0865∗∗ -0.0957∗∗ 1.898 0.542 1.077 -1.021 -1.704
(-1.83) (-1.92) (-2.15) (-2.30) (-2.44) (0.58) (0.26) (0.27) (-0.57) (-0.80)
Instrument Lag 1 2 3 4 5 1 2 3 4 5
Quarter FE yes yes yes yes yes yes yes yes yes yes
Province FE yes yes yes yes yes yes yes yes yes yes
Observations 775 742 709 676 643 730 697 664 631 598
Cragg–Donald F 14.74 7.791 4.942 8.194 10.78 5.652 8.730 1.802 5.154 4.433
Panel B: Using cumulative project number until the quarter with a 2-year lag as a measurement of BRI projects
Cumu. Proj. Num. 0.00925 0.0107
∗ 0.0112 0.00988
∗ 0.00486 2.452∗∗ 2.079∗∗ 2.035
∗ 0.991 0.726
(1.31) (1.83) (1.47) (1.89) (1.17) (1.98) (2.27) (1.81) (1.48) (0.97)
Log(Population)q−4 -0.0537 -0.0585
∗ -0.0693
∗ -0.0756∗∗ -0.0868∗∗ 3.266 2.062 1.652 -0.379 -1.033
(-1.49) (-1.70) (-1.81) (-2.19) (-2.22) (0.89) (0.67) (0.53) (-0.18) (-0.42)
Instrument Lag 1 2 3 4 5 1 2 3 4 5
Quarter FE yes yes yes yes yes yes yes yes yes yes
Province FE yes yes yes yes yes yes yes yes yes yes
Observations 775 742 709 676 643 730 697 664 631 598
Cragg–Donald F 13.08 13.91 8.411 16.97 10.05 9.231 12.20 6.587 12.46 7.286
t statistics in parentheses
∗ p
<
0.10, ∗∗
p
<
0.05, ∗∗∗
p
<
0.01
Notes: The main results reported in Table 3.10 use the lagged 2-quarter Chinese steel production as a component of instrumental variable.
146

To mainly show the robustness of the statistically significant results from 2SLS regressions, I focused on the GDP of household consumption and GDP of exports. As we can tell
from column 1 to column 5, the coefficient estimators of household consumption do not
differ significantly as the Chinese steel production changes from lagging by one quarter to
lagging by five quarters, no matter whether we use the project number or the cumulative
project number as the measurement of BRI projects, with the variation range (maximum
estimator minus minimum estimator) of 0.0289 percentage points and 0.0063 percentage
points, respectively. On the other hand, even though the coefficient estimators of the GDP
of exports vary in value, they still share a positive sign. Furthermore, the estimators with
one-quarter lag, two-quarter lag, and three-quarter lag from column 6 to column 8 are very
similar. These three significant estimators testify to the robustness of the main results.
To conclude, I believe the critical estimators in the 2SLS regression pass the robustness
check as they all point to a similar conclusion of the beneficial role of BRI projects in
promoting GDP of household consumption and GDP of exports.
3.7 Conclusion
This study provides empirical evidence on the economic impact of BRI projects in the
provinces of Indonesia. Regarding the results, my tentative conclusion is that the projects
have a win-win property for both Indonesia and China. While the projects being implemented do not have any effect on local economic factors, the completed projects benefit
the growth of GDP of exports. This conclusion has been backed up by both of my empirical strategies - difference-in-difference with staggered implementation, and instrumental
variable regressions. Moreover, even though this paper does not identify any critical effect
147

from a big-value project (using the multiple projects dummy as a proxy), it uncovers the
importance of metal- and mineral-related projects. From the cost perspective, the projects
do not impose any extra financial burden to local commercial and rural banks.
There are several areas where I can improve this study. First, even though I disclose the
causal relationship between BRI projects and local macroeconomic and social indicators, I
am still not able to clarify the channel and drivers through which the BRI projects achieve
the effect, without looking at micro-level data, such as a change of companies’ characteristics and a variation in the labor force. It would be insightful to see the channels so that
we can provide concrete policy implications, like how to target the specific shortcomings
in the economic development of a region. Second, this study does not consider the debt
cost of the projects for the government. In other words, even though the project benefits
local economic growth, we are unsure whether the benefit is worth the cost. To perform a
comprehensive cost-benefit analysis, future studies should include a detailed examination
of the financing conditions of the projects
148

References
Abi-Habib, Maria (2018). How China Got Sri Lanka to Cough Up a Port. URL: https:
//www.nytimes.com/2018/06/25/world/asia/china-sri-lanka-port.html
(visited on 08/12/2020).
Almond, Douglas and Janet Currie (2011). “Killing me softly: The fetal origins hypothesis”. In: Journal of Economic Perspectives 25.3, pp. 153–72.
Almond, Douglas, Janet Currie, and Valentina Duque (2018). “Childhood Circumstances
and Adult Outcomes: Act II”. In: Journal of Economic Literature 56.4, pp. 1360–1446.
DOI: 10.1257/jel.20171164.
Anderson, Michael L. et al. (2018). “School meal quality and academic performance”. In:
Journal of Public Economics 168, pp. 81–93. DOI: https://doi.org/10.1016/j.
jpubeco.2018.09.013.
Angrist, Joshua D. and Jorn-Steffen Pischke (2008). ¨ Mostly Harmless Econometrics: An
Empiricist’s Companion. Princeton University Press.
Aragon, Fernando M. and Juan Pablo Rud (2013). “Natural Resources and Local Com- ´
munities: Evidence from a Peruvian Gold Mine”. In: American Economic Journal:
Economic Policy 5.2, pp. 1–25. DOI: 10.1257/pol.5.2.1.
Athey, Susan and Guido Imbens (2018). “Design-based Analysis in Difference-In-Differences
Settings with Staggered Adoption”. In.
Baird, Sarah et al. (2014). “Conditional, unconditional and everything in between: a systematic review of the effects of cash transfer programmes on schooling outcomes”. In:
Journal of Development Effectiveness 6.1, pp. 1–43.
149

Banerjee, Abhijit et al. (2019). “The impact of a multifaceted poverty program on human capital and economic self-sufficiency in Indonesia: Evidence from the TNP2K
program”. In: American Economic Journal: Economic Policy 11.2, pp. 261–297.
Baum-Snow, Nathaniel et al. (2017). “Roads, Railroads, and Decentralization of Chinese
Cities”. In: The Review of Economics and Statistics 99.3, pp. 435–448. DOI: 10.1162/
REST_a_00660.
Becker, Gary S. (1965). “A Theory of the Allocation of Time”. In: The Economic Journal
75.299, pp. 493–517. DOI: 10.2307/2228949.
Black, Dan et al. (2005). “The Economic Impact Of The Coal Boom And Bust*”. In: The
Economic Journal 115.503, pp. 449–476. DOI: https : / / doi . org / 10 . 1111 / j .
1468-0297.2005.00996.x.
Borusyak, Kirill et al. (2024). “Revisiting Event-Study Designs: Robust and Efficient Estimation”. In: The Review of Economic Studies 91.6, pp. 3253–3285. DOI: 10.1093/
restud/rdae007.
Boulter, John (2018). “China’s Supply-side Structural Reform”. In: RBA Bulletin, December, viewed 13.
Callaway, Brantly and Pedro HC Sant’Anna (2021). “Difference-in-differences with multiple time periods”. In: Journal of Econometrics 225.2, pp. 200–230.
Central Government Allocates 2014 Special Funds for the Rural Student Nutrition Improvement Program (n.d.). http://www.gov.cn/xinwen/2014-08/11/content_
2733513.htm. Accessed: 2025-03-05.
Chaisemartin, Clement de and Xavier D’Haultfœuille (2020). “Two-Way Fixed Effects ´
Estimators with Heterogeneous Treatment Effects”. In: American Economic Review
110.9, pp. 2964–96. DOI: 10.1257/aer.20181169.
Chong, Zhaohui et al. (2019). “Estimating the economic benefits of high-speed rail in
China: A new perspective from the connectivity improvement”. In: Journal of Transport and Land Use 12.1. DOI: 10.5198/jtlu.2019.1264.
150

Christian, Paul and Christopher B Barrett (2017). “Revisiting the effect of food aid on
conflict: A methodological caution”. In: World Bank Policy Research Working Paper
8171.
Cong, Lin William et al. (2019). “Credit Allocation Under Economic Stimulus: Evidence
from China”. In: The Review of Financial Studies 32.9, pp. 3412–3460. DOI: 10.1093/
rfs/hhz008.
Currie, Janet and Tom Vogl (2013). “Early-Life Health and Adult Circumstance in Developing Countries”. In: Annual Review of Economics 5.1, pp. 1–36. DOI: 10.1146/
annurev-economics-081412-103704.
Donaldson, Dave (2018). “Railroads of the Raj: Estimating the Impact of Transportation
Infrastructure”. In: American Economic Review 108.4-5, pp. 899–934. DOI: 10.1257/
aer.20101199.
Donaldson, Dave and Richard Hornbeck (2016). “ Railroads and American Economic
Growth: A “Market Access” Approach *”. In: The Quarterly Journal of Economics
131.2, pp. 799–858. DOI: 10.1093/qje/qjw002.
Dong, Xiaofang et al. (2020). “The role of transportation speed in facilitating high skilled
teamwork across cities”. In: Journal of Urban Economics 115. Cities in China, p. 103212.
DOI: https://doi.org/10.1016/j.jue.2019.103212.
Dreher, Axel et al. (2020). “Aid, China, and Growth: Evidence from a New Global Development Finance Dataset”. In: American Economic Journal: Economic Policy Forthcoming.
— (2021). “Aid, China, and Growth: Evidence from a New Global Development Finance
Dataset”. In: American Economic Journal: Economic Policy 13.2, pp. 135–74. DOI:
10.1257/pol.20180631.
Gelli, Aulo et al. (2019). “A school meals program implemented at scale in Ghana increases height-for-age during mid-childhood in girls and in children from poor households: A cluster randomized trial”. In: The Journal of Nutrition 149.8, pp. 1434–1442.
151

Goodman-Bacon, Andrew (2021). “Difference-in-differences with variation in treatment
timing”. In: Journal of Econometrics 225.2. Themed Issue: Treatment Effect 1, pp. 254–
277. DOI: https://doi.org/10.1016/j.jeconom.2021.03.014.
Greenstone, Michael et al. (2010). “Identifying Agglomeration Spillovers: Evidence from
Winners and Losers of Large Plant Openings”. In: Journal of Political Economy 118.3,
pp. 536–598. DOI: 10.1086/653714.
Gundersen, Craig and James P Ziliak (2014). “Childhood Food Insecurity in the US:
Trends, Causes, and Policy Options”. In: The Future of Children, pp. 1–19.
Hao, Xuguang et al. (2019). “De-capacity policy effect on China’s coal industry”. In: Energies 12.12, p. 2331.
He, Xiaoping and Dunguo Mou (2020). “Impacts of mineral resources: Evidence from
county economies in China”. In: Energy Policy 136, p. 111088. DOI: https://doi.
org/10.1016/j.enpol.2019.111088.
Heckman, James J. (2006). “Skill Formation and the Economics of Investing in Disadvantaged Children”. In: Science 312.5782, pp. 1900–1902. DOI: 10.1126/science.
1128898.
Hoynes, Hilary, Diane Whitmore Schanzenbach, and Douglas Almond (2016). “LongRun Impacts of Childhood Access to the Safety Net”. In: American Economic Review
106.4, pp. 903–34. DOI: 10.1257/aer.20130375.
Hoynes, Hilary W (1993). Welfare transfers in two-parent families: labor supply and welfare participation under AFDC-UP.
Hoynes, Hilary Williamson and Diane Whitmore Schanzenbach (2012). “Work incentives
and the Food Stamp Program”. In: Journal of Public Economics 96.1, pp. 151–162.
Jagoe, Kirstie et al. (2020). “Sharing the burden: Shifts in family time use, agency and
gender dynamics after introduction of new cookstoves in rural Kenya”. In: Energy
Research & Social Science 64, p. 101424.
152

Jiang, Mingyi and Xiaoyan Lin (2018). “The Impact of High-Speed Railway on Labor
Agglomeration”. In: 2018 5th International Conference on Industrial Economics System and Industrial Security Engineering (IEIS), pp. 1–5. DOI: 10.1109/IEIS.2018.
8598118.
Jiang, Xiaochen, Jim Huangnan Shen, et al. (2021). “Supply-side structural reform and
dynamic capital structure adjustment: Evidence from Chinese-listed firms”. In: PacificBasin Finance Journal 65.C. DOI: 10.1016/j.pacfin.2020.101.
John Hurley, Scott Morris and Gailyn Portelance (2019). “Examining the debt implications
of the Belt and Road Initiative from a policy perspective”. In: Journal of Infrastructure,
Policy and Development 3, p. 139. DOI: 10.24294/jipd.v3i1.1123.
Kan, Man-Yee and Guangye He (2024). “Resource bargaining and gender display in housework and care work in modern China”. In: Fertility and Childcare in East Asia. Routledge, pp. 162–194.
Kline, Patrick (2011). “Oaxaca-Blinder as a Reweighting Estimator”. In: American Economic Review 101.3, pp. 532–37. DOI: 10.1257/aer.101.3.532.
Kline, Patrick and Enrico Moretti (2013). “ Local Economic Development, Agglomeration Economies, and the Big Push: 100 Years of Evidence from the Tennessee Valley Authority *”. In: The Quarterly Journal of Economics 129.1, pp. 275–331. DOI:
10.1093/qje/qjt034.
Kramer, Marion et al. (2021). “Improving Child Health and Cognition: Evidence from ¨
a School-Based Nutrition Intervention in India”. In: The Review of Economics and
Statistics 103.5, pp. 818–834. DOI: 10.1162/rest_a_00950.
Levinsohn, James and Amil Petrin (2003). “Estimating Production Functions Using Inputs
to Control for Unobservables”. In: The Review of Economic Studies 70.2, pp. 317–341.
Lin, Yatang (2017). “Travel costs and urban specialization patterns: Evidence from China’s
high speed railway system”. In: Journal of Urban Economics 98. Urbanization in Developing Countries: Past and Present, pp. 98–123. DOI: https://doi.org/10.1016/
j.jue.2016.11.002.
153

List of National and Local Pilot Counties for the Rural Student Nutrition Improvement
Program (n.d.). http://www.moe.gov.cn/jyb_xwfb/xw_zt/moe_357/s6211/
s6329/s6371/201904/t20190419_378881.html. Accessed: 2025-03-05.
Lundborg, Petter et al. (2022). “Long-term effects of childhood nutrition: evidence from a
school lunch reform”. In: The Review of Economic Studies 89.2, pp. 876–908.
Minister of Finance (2019). Statistics of Outward Investment and Economic Cooperation.
URL: http://hzs.mofcom.gov.cn/article/date/ (visited on 04/30/2020).
Moretti, Enrico (2011). “Local Labor Markets”. In: ed. by O. Ashenfelter and D. Card.
1st ed. Vol. 4B. Elsevier. Chap. 14, pp. 1237–1313.
— (2010). “Local Multipliers”. In: American Economic Review 100.2, pp. 373–77. DOI:
10.1257/aer.100.2.373.
Mu, Ren and Dominique van de Walle (2011). “Left behind to farm? Women’s labor reallocation in rural China”. In: Labour Economics 18. Labour markets in developing
countries, S83–S97. DOI: https://doi.org/10.1016/j.labeco.2011.01.009.
Nair-Reichert, Usha and Diana Weinhold (2001). “Causality Tests for Cross-Country Panels: a New Look at FDI and Economic Growth in Developing Countries”. In: Oxford Bulletin of Economics and Statistics 63.2, pp. 153–171. DOI: 10.1111/1468-
0084.00214.
Notowidigdo, Matthew J. (2020). “The Incidence of Local Labor Demand Shocks”. In:
Journal of Labor Economics 38.3, pp. 687–725. DOI: 10.1086/706048.
Nunn, Nathan and Nancy Qian (2014). “US Food Aid and Civil Conflict”. In: American
Economic Review 104.6, pp. 1630–66. DOI: 10.1257/aer.104.6.1630.
Nutrition Improvement Program in Rurual China (n.d.). http://www.moe.gov.cn/jyb_
xwfb/xw_zt/moe_357/s6211/s6329/. Accessed: 2024-11-01.
Qian, Nancy (2014). Making Progress on Foreign Aid. Working Paper 20412. National
Bureau of Economic Research. DOI: 10.3386/w20412.
154

Report on the Implementation of the Rural Compulsory Education Student Nutrition Improvement Program (n.d.). http : / / www . moe . gov . cn / jyb _ xwfb / gzdt _ gzdt /
s5987/201703/t20170302_297934.html. Accessed: 2025-03-05.
Roback, Jennifer (1982). “Wages, Rents, and the Quality of Life”. In: Journal of Political
Economy 90.6, pp. 1257–1278.
Rosen, Sherwin (1979). “Wage-based indexes of urban quality of life”. In: Current issues
in urban economics, pp. 74–104.
Roth, Jonathan et al. (2023). “What’s trending in difference-in-differences? A synthesis of
the recent econometrics literature”. In: Journal of Econometrics 235.2, pp. 2218–2244.
DOI: https://doi.org/10.1016/j.jeconom.2023.03.008.
Significant Impact of the Rural Student Nutrition Improvement Program (n.d.). http :
//www.jyb.cn/rmtzcg/xwy/wzxw/202005/t20200521_328660.html. Accessed:
2025-03-05.
Singh, Abhijeet et al. (2014). “School Meals as a Safety Net: An Evaluation of the Midday Meal Scheme in India”. In: Economic Development and Cultural Change 62.2,
pp. 275–306.
Stock, James and Motohiro Yogo (2002). Testing for Weak Instruments in Linear IV Regression. NBER Technical Working Papers 0284. National Bureau of Economic Research, Inc.
Supporting Local Pilot Programs for the Rural Student Nutrition Improvement Program
(n.d.). http://www.moe.gov.cn/jyb_xwfb/xw_zt/moe_357/s6211/s6329/
s6371/201203/t20120331_133445.html. Accessed: 2025-03-05.
Ten Years of the Nutrition Improvement Program and Future Recommendations (n.d.).
http://www.jyb.cn/rmtzcg/xwy/wzxw/202209/t20220913_2110945681.html.
Accessed: 2025-03-05.
Wong, Samantha (2021). China: length of high speed rail operation network 2019.
155

Woo, Wing Thye (2019). “China’s soft budget constraint on the demand-side undermines
its supply-side structural reforms”. In: China Economic Review 57, p. 101111. DOI:
https://doi.org/10.1016/j.chieco.2017.09.010.
Xinhua (2019). A Brief Introduction to Belt and Road Initiative. URL: http : / / www .
xinhuanet.com/politics/2019-04/26/c_1124418156.htm (visited on 04/30/2020).
Yao, Shujie and Kailei Wei (2007). “Economic growth in the presence of FDI: The perspective of newly industrialising economies”. In: Journal of Comparative Economics
35.1, pp. 211–234. DOI: https://doi.org/10.1016/j.jce.2006.10.007.
156

Appendices
Appendix
G.1 Figures
(a) CSDID Event Study: Impact of NIP on
Children Log(Height)
(b) CSDID Event Study: Impact of NIP on
Children Log(Weight)
Figure A1: CSDID Event Study on Children Health Outcomes
157

(a) CSDID Event Study: Impact of NIP on
Parent Employment
(b) CSDID Event Study: Impact of NIP on
Parent Weekly Working Hour
(c) CSDID Event Study: Impact of NIP on
Parent Log(Income)
Figure A2: CSDID Event Study on Adults Labor Market Outcomes
158

(a) Impact of NIP on Adult Employment for
Children First Treated at Ages 6–8
(b) Impact of NIP on Adult Employment for
Children First Treated at Ages 9–11
(c) Impact of NIP on Adult Employment for
Children First Treated at Ages 12–15
Figure A3: CSDID Event Study on Adult Outcomes by Child’s Age at the NIP Onset
159

Figure A4: Parallel Trend Plots for Instrumental Variable
Notes: Plot the treads of growth for 3 groups: the provinces with lowest 33% probability to have BRI projects, the provinces with the
middle 33% probability, and the group with the highest 33% probability. The parallel trend plots serves as a validity check for the
instrumental variable, which is constructed as the interaction between steel production and the probability of BRI project
implementation.
160

G.2 Tables
Table A1: BJS Estimation Results: Impact of NIP on Children
(1) (2) (3)
Log(Height) Log(Weight) Sick
T-4 -0.00116 -0.0263 -0.150
(-0.03) (-0.32) (-0.99)
T-2 0.00342 0.0103 -0.172
(0.07) (0.13) (-1.19)
T0 0.0109∗
-0.00757 -0.0206
(2.31) (-0.59) (-1.17)
T2 0.0243∗∗ 0.00834 -0.0132
(2.95) (0.50) (-0.53)
T4 0.0420∗∗∗ -0.00836 -0.0821
(3.29) (-0.42) (-1.89)
T6 0.0287 -0.0376 -0.147∗
(1.20) (-1.26) (-2.28)
Observations 11988 11988 11988
Notes: Individual and year fixed effect is included. The estimates are calculated by Stata package
“did imputation” proposed in Borusyak et al. 2024. Age serves as a control. Standard errors are
clustered at the county level to account for potential correlations within counties. t statistics in
parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.
161

Table A2: BJS Estimation Results: Impact of NIP on Adults
(1) (2) (3)
Employment Working Hour Log Income
T-4 0.0195 1.928 0.363
(0.55) (1.75) (1.94)
T-2 0.0411 1.190 -0.174
(0.84) (0.71) (-0.80)
T0 0.0167 4.340∗∗∗ 0.295∗
(0.63) (5.23) (2.00)
T2 0.0947∗ 4.596∗∗∗ 0.149
(2.51) (3.64) (1.01)
T4 0.0816∗ 4.407∗∗∗ 0.356∗
(2.00) (4.86) (2.30)
T6 0.119∗ 7.237∗∗∗ 0.837∗∗∗
(2.45) (5.32) (3.44)
T8 0.0984 8.097∗∗∗ 0.357
(1.65) (3.79) (1.08)
Observations 53023 53023 53023
Notes: Individual and year fixed effect is included. The estimates are calculated by Stata package
“did imputation” proposed in Borusyak et al. 2024. Age serves as a control. Standard errors are
clustered at the county level to account for potential correlations within counties. t statistics in
parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.
162

Table A3: Heterogeneous Impact of NIP on Adults by Number of Children
(1) (2) (3) (4) (5) (6)
VARIABLES Employment Working Hour Log Income Employment Working Hour Log Income
Treatment -0.00172 3.108** 0.117 -0.0216 3.204*** -0.00719
(0.0326) (1.200) (0.203) (0.0377) (1.119) (0.194)
Treatment*Parent Takecare 0.0216 3.465* 0.176 0.0394 4.538** 0.359
(0.0386) (1.761) (0.306) (0.0500) (2.039) (0.341)
Treatment*Parent Take*Multi C 0.0721*** -1.803 0.110 0.0881*** -2.366 0.0699
(0.0246) (1.538) (0.204) (0.0330) (1.932) (0.299)
Observations 53,087 53,087 53,087 27,809 27,809 27,809
R-squared 0.083 0.153 0.107 0.079 0.161 0.115
Number of pid 14,724 14,724 14,724 7,637 7,637 7,637
Mean 0.75 19.61 3.39 0.70 15.89 2.52
Individual FE YES YES YES YES YES YES
Year FE YES YES YES YES YES YES
Gender All All All Female Female Female
Notes: The dependent variable, Employment, is a dummy variable that indicates whether the individual is currently employed. Working Hours measures the number of hours
the individual works per week. Log Income represents the total income the individual earns, expressed in logarithmic form. Parent Careit is a dummy variable equal to 1 if a
parent is the primary caregiver for the children. Multi
Cit equals 1 if the family has more than one child. Individual’s age is included as a control variable. Standard errors are
clustered at the county level to account for potential correlations within counties. The estimates are derived using the canonical Two-Way Fixed Effects (TWFE) model. ***
p<0.01, ** p
<0.05, * p
<0.1
163

G.3 Modeling the NIP as an Improvement in AG
With the household production function to be xg = AGL
γ
g, where γ ∈ (0,1).
We have simplified the following utility maximization problem of a representative
household:
max
xa,
,xm,xl
,xg
U(xa,
, xm, xl
, xg)
subject to the budget constraint
paxa + pmxm +wpl +w

xg
AG

1
γ
= Y
where the parameter which characterizes the return to scale in the household production
γ ∈ (0,1).
In the problem, xa, xm, xl
, xg represent the consumption in agricultural goods, manufactural goods, leisure and household production good, respectively, which are the decision
variables. The prices pa, pm, w are exogenously given. AG is the household production
technology parameter.
Assume the utility function follows a CES form, i.e.
U(xa, xm, xl
, xg) =
∑
i∈{a,m,l,g}
ρix
ε
i
!1/ε
where the parameter ε < 0.
We first explore the first order conditions, and represent xi
, i ∈ {a,m,l} in terms of xg,
and plug those into the budget constraint to get an equation in terms of xg.
164

From the first-order conditions, we can derive:
xa =

αa
αg
·
pg(xg,AG)
pa

1
1−ε
xg
xm =

αm
αg
·
pg(xg,AG)
pm

1
1−ε
xg
xl =

αl
αg
·
pg(xg,AG)
w

1
1−ε
xg
where the effective price of household production is:
pg(xg,AG) = w
γ
·
1
A
1/γ
G
· x
1
γ −1
g
Substituting these expressions into the budget constraint:
pa

αa
αg
·
pg(xg,AG)
pa

1
1−ε
xg+ pm

αm
αg
·
pg(xg,AG)
pm

1
1−ε
xg+w

αl
αg
·
pg(xg,AG)
w

1
1−ε
xg+w

xg
AG
1/γ
=Y
Let’s simplify this to:
pg(xg,AG)· xg · ∑
i∈{a,m,l}

αi
αg

1
1−ε

pg(xg,AG)
pi
ε
1−ε
+w

xg
AG
1/γ
= Y
Where pi represents the respective price (pa, pm, or w for xl
).
Let’s use the implicit function theorem to directly determine the sign of dxg
dAG
. This
approach will give us a more direct derivation of the condition.
165

The household’s optimization problem gives us a system of equations that implicitly
defines xg as a function of AG. Using the budget constraint after substituting the first-order
conditions, we have:
F(xg,AG) = paxa(xg,AG) + pmxm(xg,AG) +wxl(xg,AG) +w

xg
AG
1/γ
−Y = 0
By the implicit function theorem:
dxg
dAG
= −
∂F/∂AG
∂F/∂ xg
Step 1: Calculate ∂F
∂AG
Let’s first recall the expressions for each consumption good:
xi =

αi
αg
·
pg(xg,AG)
pi

1
1−ε
xg for i ∈ a,m,l
Where pg(xg,AG) = w
γ
·
1
A
1/γ
G
· x
1
γ −1
g
The effect of AG on the budget constraint comes through two channels:
1. Its effect on the optimal consumption of xa, xm, and xl
through pg 2. Its direct effect
on the household production cost w

xg
AG
1/γ
∂F
∂AG
= ∑
i∈a,m,l
pi
∂ xi
∂AG
+w·
∂
∂AG
"
xg
AG
1/γ
#
For each good i:
∂ xi
∂AG
=
∂ xi
∂ pg
·
∂ pg
∂AG
166

Given ∂ pg
∂AG
= −
1
γ
·
pg
AG
, this becomes:
∂ xi
∂AG
= −
1
γ
·
pg
AG
·
∂ xi
∂ pg
For the direct effect on household production cost:
∂
∂AG
"
xg
AG
1/γ
#
= −
1
γ
·
xg
AG
·

xg
AG
1/γ−1
Combining these terms:
∂F
∂AG
= −
1
γ
·
pg
AG
· ∑
i∈a,m,l
pi
∂ xi
∂ pg
−
w
γ
·
xg
AG
·

xg
AG
1/γ−1
Step 2: Calculate ∂F
∂ xg
The derivative with respect to xg has both direct and indirect effects:
∂F
∂ xg
= ∑
i∈a,m,l
pi
∂ xi
∂ xg
+w·
∂
∂ xg
"
xg
AG
1/γ
#
For each good i, the effect works through both the direct proportionality and the indirect price effect:
∂ xi
∂ xg
=
xi
xg
+
∂ xi
∂ pg
·
∂ pg
∂ xg
With ∂ pg
∂ xg
=
w
γ
·
1
A
1/γ
G
·

1
γ −1

· x
1
γ −2
g
167

For the direct effect on household production cost:
∂
∂ xg
"
xg
AG
1/γ
#
=
1
γ
·
1
AG
·

xg
AG
1/γ−1
Step 3: Evaluating the Sign of dxg
dAG
After substituting and simplifying, we get:
dxg
dAG
= −
−
1
γ
·
pg
AG
·∑i∈a,m,l pi
∂ xi
∂ pg
−
w
γ
·
xg
AG
·

xg
AG
1/γ−1
∑i∈a,m,l pi
∂ xi
∂ xg
+
w
γ
·
1
AG
·

xg
AG
1/γ−1
The denominator is typically positive due to the second-order conditions for utility
maximization. Therefore, the sign of dxg
dAG
depends on the sign of the numerator.
For CES utility, we can show that:
∑
i∈a,m,l
pi
∂ xi
∂ pg
= −
σ
pg
· ∑
i∈a,m,l
pixi + (σ −1)·
pgxg
p
2
g
· ∑
i∈a,m,l
pixi
Where σ =
1
1−ε
is the elasticity of substitution.
This simplifies to:
∑
i∈a,m,l
pi
∂ xi
∂ pg
= −
1
1−ε
·
∑i∈a,m,l pixi
pg
+
ε
1−ε
·
xg ·∑i∈a,m,l pixi
pg
Substituting into our expression for dxg
dAG
, and after algebraic manipulation, we get the
condition for dxg
dAG
< 0:
168

αg <
∑i̸=gαi
1+
γ
1−γ
·
1−ε
−ε
This is precisely the condition we derived earlier, now obtained directly through the
implicit function theorem. This confirms our result and provides a clearer understanding of
when technology improvements lead to decreased household production, enabling parents
to reallocate time to work or leisure.
The condition shows that dxg
dAG
< 0 when:
1. The preference weight for household production (αg) is sufficiently small
2. Goods are complements (ε < 0)
3. The returns to scale parameter (γ) is in an intermediate range
This happens because the income effect (redirecting resources to more valued consumption) outweighs the substitution effect (household production becoming relatively
cheaper).
169

Asset Metadata

Creator Chen, Tao (author)

Core Title Three essays on development economics

Contributor Electronically uploaded by the author (provenance)

School Dana and David Dornsife College of Letters, Arts and Sciences

Degree Doctor of Philosophy

Degree Program Economics

Degree Conferral Date 2025-05

Publication Date 05/12/2025

Defense Date 03/10/2025

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

Tag development economics,OAI-PMH Harvest

Format theses (aat)

Language English

Advisor Oliva, Paulina (committee chair), Quach, Simon (committee member), Wong, TJ (committee member)

Creator Email taoc@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-oUC11399KHCT

Unique identifier UC11399KHCT

Legacy Identifier etd-Chen-36957-47733

Document Type Dissertation

Format theses (aat)

Rights Chen, Tao

Internet Media Type application/pdf

Type texts

Source 202505w03-usctheses (batch), University of Southern California Dissertations and Theses (collection), University of Southern California (contributing entity)

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 author, as the original true and official version of the work, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright.

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

Repository Email uscdl@usc.edu

Abstract (if available)

Abstract This dissertation examines the economic impact of three major policies in China: the Nutrition Improvement Program (NIP), coal mine shutdowns, and the Belt and Road Initiative (BRI). Each chapter explores a distinct policy’s effects on labor markets, household welfare, and regional economic development.