ESSAYS ON INDIA’S ECONOMIC DEVELOPMENT.
by
SAURABH GUPTA
A DISSERTATION
Presented to the Department of Economics
and the Division of Graduate Studies of the University of Oregon
in partial fulfillment of the requirements
for the degree of
Doctor of Philosophy
June 2022
DISSERTATION APPROVAL PAGE
Student: Saurabh Gupta
Title: Essays on India’s Economic Development.
This dissertation has been accepted and approved in partial fulfillment of the
requirements for the Doctor of Philosophy degree in the Department of Economics
by:
Dr. Shankha Chakraborty Chair
Dr. Anca D. Cristea Core Member
Dr. Alfredo Burlando Core Member
Dr. Nagesh Murthy Institutional Representative
and
Krista M. Chronister Vice Provost for Graduate Studies
Original approval signatures are on file with the University of Oregon Division of
Graduate Studies.
Degree awarded June 2022
ii
© 2022 Saurabh Gupta
All rights reserved.
iii
DISSERTATION ABSTRACT
Saurabh Gupta
Doctor of Philosophy
Department of Economics
June 2022
Title: Essays on India’s Economic Development.
This dissertation is on the economic development of India during the past
three decades with a focus on its changing industrial and household structure.
Chapter 1 provides a brief introduction of the Indian economy and motivates the
theme of the dissertation.
In Chapter 2, I study the effects of transportation infrastructure on regional
manufacturing activity. I exploit geographical and temporal variation in project
implementation to argue for causal effects on the regional industrial outcomes. I
investigate how highways can improve market competition between firms situated
in geographically distant locations, and as a result, create incentives to invest
in activities that improve productivity. The results show that highways had no
direct effect on India’s manufacturing output growth and led to a decline in average
manufacturing productivity. I argue that these results can be attributed to lack of
improvements in allocative efficiency within regions, and slow movement of skilled
labor into the manufacturing sector. These results are contrary to some recent work
on India but in line with evidence presented in the wider literature on low-income
countries
In Chapter 3, I investigate the relationship between highways and female
labor force participation (FLFP). Using census level data from India, I estimate
iv
how the construction of highways may have opened up market opportunities
for households and consequently affected FLFP. I find that the effects are
heterogeneous across districts with some districts experiencing an increase while
others experiencing a decline in FLFP. The decline was driven mostly by married
and educated women withdrawing from the manufacturing and services sectors.
I also find suggestive evidence that highways led to an increase in labor force
participation of low skilled women.
In Chapter 4, jointly with Dr. Shankha Chakraborty, we examine how
household decision making can explain India’s declining FLFP over the last
three decades. We propose a tractable analytical model in which married women
respond to opportunity costs of their labor hours when dividing their time between
household and market production. The model incorporates cultural costs attached
to female work and its negative effect on female labor supply. We highlight
competing mechanisms at play that suggest a U-shaped pattern of FLFP in
response to economic growth.
Finally, Chapter 5 summarizes the results of the dissertation and presents a
concluding remarks.
v
CURRICULUM VITAE
NAME OF AUTHOR: Saurabh Gupta
GRADUATE AND UNDERGRADUATE SCHOOLS ATTENDED:
University of Oregon, Eugene, Oregon, USA
Central University of Tamil Nadu, Chennai, Tamil Nadu, India
Panjab University, Chandigarh, Punjab, India
DEGREES AWARDED:
Master of Science, Economics, 2015, Central University of Tamil Nadu
Bachelor of Engineering, Mechanical Engineering, 2011, Panjab University
AREAS OF SPECIAL INTEREST:
Applied Econometrics
Development Economics
Industrial Economics
International Trade
PROFESSIONAL EXPERIENCE:
Graduate Employee, Department of Economics, University of Oregon, 2017-
2022
Research Associate, Indian Institute of Management, Bangalore, 2017
Research Assistant, Madras School of Economics, 2016
Data Scientist, Scienaptic Systems, 2015
GRANTS, AWARDS AND HONORS:
Graduate Teaching Fellowship, University of Oregon, 2017-2022
Homan Fellowship Award, University of Oregon, 2021
vi
PUBLICATIONS:
Gupta, S., & Bhaduri, S.N. (2019). Skin in the Game – Investor Behavior in
Asset Pricing, the Indian Context. Review of Behavioral Finance, Vol. 11
No. 4, pp. 373-392
vii
ACKNOWLEDGEMENTS
I thank the members of my committee for their guidance in conceiving
and executing my research agenda. In particular, I would like to thank Shankha
Chakraborty for his invaluable mentorship, encouragement, and patience. I thank
Anca Cristea, Alfredo Burlando, and Nagesh Murthy for many helpful comments
and advice. I extend my gratitude to Carol and family for providing support during
my stay in Eugene. A special thanks to Sarthak Bhatia for his efforts in helping me
access relevant data. Finally, I thank my friends and fellow graduate students for
always being there for me.
viii
TABLE OF CONTENTS
Chapter Page
I INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
II DO HIGHWAYS AFFECT MANUFACTURING
GROWTH? A CAUTIONARY TALE FROM INDIA . . . . . . . 5
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Relevant Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 The Golden Quadrilateral Project . . . . . . . . . . . . . . . . . . . . 10
2.4 Data Preparation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.5 Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.5.1 Basic Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.5.2 Main Specification . . . . . . . . . . . . . . . . . . . . . . . . 14
2.5.3 Identification Issues . . . . . . . . . . . . . . . . . . . . . . . . 15
2.6 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.6.1 Exploring Selection Bias . . . . . . . . . . . . . . . . . . . . . 18
2.6.2 Main Estimates: Manufacturing Output . . . . . . . . . . . . 23
2.6.3 Manufacturing Productivity . . . . . . . . . . . . . . . . . . . 26
2.7 Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.7.1 Local Average Treatment Effect . . . . . . . . . . . . . . . . . 33
2.7.2 Sampling Errors: Sub-sample Analysis . . . . . . . . . . . . . 36
2.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
III NATIONAL HIGHWAYS AND FEMALE LABOR
FORCE PARTICIPATION . . . . . . . . . . . . . . . . . . . . . . . . 40
ix
Chapter Page
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.2 Relevant Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.3 The Highway Project . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.4 Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.5 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.6.1 Baseline Comparison . . . . . . . . . . . . . . . . . . . . . . . 51
3.6.2 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
IV FEMALE LABOR FORCE PARTICIPATION IN
INDIA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.2.1 Additional Facts . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.2.2 Cultural values . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.2.3 Theoretical works . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.3 Household Decisions . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.3.1 Case I: I = 0 . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.3.2 Case II: I = 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.3.3 Extensive margin decision: to work or not . . . . . . . . . . . 83
4.4 Aggregate Female Labor Force Participation . . . . . . . . . . . . . . 89
4.5 Numerical Illustration . . . . . . . . . . . . . . . . . . . . . . . . . . 92
4.6 Conclusion and future plan . . . . . . . . . . . . . . . . . . . . . . . . 95
x
Chapter Page
V CONCLUSION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
APPENDICES
A CHAPTER II APPENDIX . . . . . . . . . . . . . . . . . . . . . . 100
A.1 Baseline Levels of Outcome . . . . . . . . . . . . . . . . . . . . . . . 100
A.2 Tables and Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
B CHAPTER III APPENDIX . . . . . . . . . . . . . . . . . . . . . . 106
B.1 Tables and Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
REFERENCES CITED . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
xi
LIST OF FIGURES
Figure Page
1 Share of India’s GDP by industry . . . . . . . . . . . . . . . . . . . . 2
2 Straight line and Least Cost Path Instrument Variables . . . . . . . . 16
3 Comparing manufacturing sector outcomes for different
district groups. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4 Comparing annual changes in outcomes after controlling
for district fixed effects. . . . . . . . . . . . . . . . . . . . . . . . . . 21
5 Women’s Employment Share by Industry Sectors . . . . . . . . . . . 55
6 Women’s Employment Shares by Industry and Skill level . . . . . . . 59
7 Simulation of Literacy Interaction Effects . . . . . . . . . . . . . . . 60
8 Female Labor Force Participation Rate, India . . . . . . . . . . . . . 69
9 Fertility rate 1990-2018 . . . . . . . . . . . . . . . . . . . . . . . . . 70
10 FLFP 1990-2019 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
11 FLFP and per capita income, Afridi, Bishnu and
Mahajan, 2019 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
12 Secondary enrollment 1990-2019 . . . . . . . . . . . . . . . . . . . . 74
13 LFPR by education (urban, married and not married,
age 20-45) Afridi et al. (2020) . . . . . . . . . . . . . . . . . . . . . . 74
14 Changes in female educational attainment in India . . . . . . . . . . 75
15 Changes in values if men should have more right to jobs . . . . . . . 77
16 Changes in values if university is important for a boy or
a girl . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
17 LFP decision at the household level . . . . . . . . . . . . . . . . . . . 87
xii
Figure Page
18 Aggregate FLFP, χ ∈ {χL, χH} . . . . . . . . . . . . . . . . . . . . . 90
19 Lf with continuum of types, effect of culture . . . . . . . . . . . . . . 91
20 Numerical Simulation for LFP decision at the household level . . . . 94
21 Numerical Simulation for Aggregate FLFP, χ ∈ {χL, χH} . . . . . . . 94
xiii
LIST OF TABLES
Table Page
1 Golden Quadrilateral - Progress (in kms.) . . . . . . . . . . . . . . . 11
2 GQ construction main specification results . . . . . . . . . . . . . . . 23
3 GQ treatment effect on productivity changes within districts . . . . . 30
4 GQ treatment effect on changes in labor composition
within districts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
5 Testing for Heterogeneous Effects . . . . . . . . . . . . . . . . . . . . 36
6 Main results with Census firms: Manufacturing Output . . . . . . . . 37
7 Baseline - Placebo Treatment Effects . . . . . . . . . . . . . . . . . . 52
8 Average Treatment Effects . . . . . . . . . . . . . . . . . . . . . . . . 54
9 Parameter Values for Simulation . . . . . . . . . . . . . . . . . . . . . 93
A.1 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
A.2 First Stage Results - Straight Line IV . . . . . . . . . . . . . . . . . . 103
A.3 Probability of Treatment . . . . . . . . . . . . . . . . . . . . . . . . . 104
A.4 GQ construction reduced-form results . . . . . . . . . . . . . . . . . . 105
B.1 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
B.2 Treatment Effects - FLFP with Literacy Interactions by Industry . . 108
B.3 Treatment Effects - Female Employment Levels by Industry . . . . . 109
B.4 Baseline - Placebo Treatment Effects with Literacy Interactions . . . 110
B.5 Average Treatment Effects - FLFP with Road
Connectivity Interactions by Industry . . . . . . . . . . . . . . . . . . 111
B.6 Average Treatment Effects - FLFP with Caste Diversity
Interactions by Industry . . . . . . . . . . . . . . . . . . . . . . . . . 112
xiv
Table Page
B.7 Treatment Effects - Education Wise Female Employment Levels . . . 113
B.8 Treatment Effects - Female Employment Levels by
Industry and Married Status . . . . . . . . . . . . . . . . . . . . . . . 114
xv
CHAPTER I
INTRODUCTION
India has experienced rapid growth in the past four decades starting
with a structural shift in the late 1980s. During this time, India liberalized its
economy and introduced gradual trade reforms that paved the way for growth in
international trade and investment. A dismal GDP growth of 3.4% during the
decade of 1970-1980 transformed into an average growth rate of 7.5% after the
late 1980s, peaking at 8.9% during 2000-2010. This economic transformation was
accompanied by secular improvements in the country’s development indicators and
large-scale structural shift away from traditional agriculture.
India’s growth and transformation has been unique in one respect. It has
largely been driven by the services sector instead of growth in the manufacturing
sector, a pattern at odds with the experiences of other developing countries such
as Bangladesh, China, or South Korea. Decomposing this growth further, Figure 1
plots the average sector-wise decomposition of value added to the GDP between
1950s and 2010s. These estimates show that between the decades of 1980s and
2010s the contribution of manufacturing sector towards the GDP increased from
14.5% to 17.7% whereas the contribution by the services sector increased by a much
larger margin from 46% to 61%. Not surprisingly, contribution by the agriculture
and allied activities decreased from 35% to 16.5% during the same time period.1
Several reasons have been proposed in the literature for this slow growth
in manufacturing. For example, the literature has identified a “missing middle”
in India’s manufacturing sector (Hsieh and Klenow, 2009): manufacturing firms
are unable to scale up due to which markets are dominated by a few very large
1These estimates have been complied from Reserve Bank of India’s annual publication on the
Indian economy, Handbook of Economics and Statistics
1
Figure 1. Share of India’s GDP by industry
firms and numerous small firms. These conditions are further exacerbated by a
large number of very small firms operating in the informal sector. The stunting
of manufacturing growth has also been attributed to India’s strict labor laws
making it difficult for firms to hire cheap labor, and bureaucratic hurdles such as
complicated tax laws that make it difficult to run businesses (see Mitra and Ural,
2008, Rao, 2005).
In addition to India’s unique sectoral growth, transformation in the labor
markets and demographics is also worth noting. The economic growth in the past
four decades has been associated with high educational attainment among younger
people and declining fertility rates for women. As per World Bank estimates, the
2
average literacy rate in India has grown from 48% in the year 1991 to 74% in the
year 2018. Simultaneously, the average fertility rate went down from 4.83 child
births per woman in 1980 to 2.22 child births per woman in 2018. We may expect
this fall in fertility and increase in education to encourage more women to join
the market work; however, the female labor force participation (FLFP) rate has
declined from 30% in 1990 to 19% in 2019.2
In this dissertation, my goal is to understand these patterns of slow
manufacturing growth and female labor force participation in conjunction with
investments in transportation infrastructure. The decades when India experienced
significant economic growth coincided with the construction of large-scale
transportation infrastructure projects. I hypothesize that construction of highways
may have helped transform regions closer to it in a manner different than what
we see through the national trends. For example, highways could have improved
market competition and efficiency, and hence, comparatively higher positive
gains for the manufacturing sector in regions closer to the highways. Similarly, if
highways improved access to schools, jobs, business opportunities for households,
they could have encouraged female labor force participation.
Exploration of these ideas situates this dissertation in the broader policy
framework that tries to understand the role played by transport infrastructure
in economic development. Significant efforts at the global and the country level
are devoted towards improving public infrastructure that can increase overall
productivity of the economy. Transport infrastructure is believed to be one of the
most significant factors that influences the economy in multiple ways. For example,
Woetzel et al. (2016) predict that global infrastructure demand will be $2.5 trillion
2These empirical facts have been compiled from World Bank’s DataBank that can be accessed
at this website. https://data.worldbank.org/country/IN
3
annually between 2016-2030. In 2021 alone, the World Bank approved transport
projects worth $3.6 billion. These facts emphasize the growing importance of
transport infrastructure in the global economy.
In Chapter 2 and 3, I investigate the role played by the construction of
highways on regional manufacturing industry and female labor force participation
in India. The results show that highways did not significantly impact the
manufacturing output in regions that were located closer to the highways; instead,
I notice a decline in average manufacturing productivity. In the labor markets, the
highways seem to have affected districts based on their initial market conditions.
Districts with lower literacy rate, higher road connectivity, and lower caste
fractionalization experienced an increase in FLFP.
Chapter 4, jointly with Dr. Chakraborty, investigates the economic reasons
behind the falling FLFP in India and proposes an analytical model that emphasizes
the role of gender norms and mother’s time devoted towards children’s human
capital production.
4
CHAPTER II
DO HIGHWAYS AFFECT MANUFACTURING GROWTH? A CAUTIONARY
TALE FROM INDIA
2.1 Introduction
Transport infrastructure is a crucial element of public investment.
Understanding the potential of transportation to facilitate economic gains is of
growing policy relevance. Intuitively, we can assume that railroads, highways,
and rivers provide easy access to markets across regions within a country, improve
market competition and contribute to welfare gains. They may also help alleviate
cost impediments to trade between firms located in different regions, thereby
providing an integrated market economy to firms as well as consumers. A lot
remains to be understood about this relationship between transport infrastructure
and the market economy.
The current literature offers mixed results on this relationship in the context
of low-income countries. For example, in the case of China, highways do not seem
to significantly stimulate any economic growth. Some studies find that locations
closer to the highways experienced negative GDP growth in manufacturing and
all other sectors combined (see Faber, 2014; Banerjee, Duflo and Qian, 2020).
In contrast, evidence from sub-Saharan Africa and India show robust positive
growth for locations situated near the highways (see Storeygard, 2016, Ghani,
Goswami and Kerr, 2016). The purpose of this paper is to revisit the case of India
to estimate the relationship between transport infrastructure and economic growth
using robust methodological tools. In addition, it also attempts to shed light on
barriers that may hold back economic gains from the highways.
5
I first examine the link between highway construction and subsequent
changes in the manufacturing sector. I use a large highway project, the Golden
Quadrilateral (GQ) Project, constructed in India during the period 2000-2009,
to examine the effects of national highways on local district-level manufacturing
output growth. To counter the problem of endogenous placement of the highway,
a combination of the difference-in-differences and instrument variable strategies is
used. Two hypothetical highway routes are constructed using detailed geographical
information to instrument the original GQ project, and bring quasi-random
variation in the placement of highways among districts. Finally, the estimates are
conditioned on a variety of district-level covariates including socio-economic and
geographical characteristics. The results suggest that contrary to previous work on
the effect of highways in India, districts closer to the GQ project did not experience
significant manufacturing growth compared to districts further away.
I then explore potential mechanisms that may have limited the effects of
the GQ project on manufacturing growth. Two notable roadblocks to India’s
manufacturing growth have been identified in the literature. First is the higher
degree of resource misallocation in India (Hsieh and Klenow, 2009). One
implication of resource misallocation is that market shares are not efficiently
allocated across firms; specifically, low productivity firms may command a higher
market share. I examine how the GQ project may have alleviated this problem.
The results suggest that for districts closer to the GQ project, market shares
adjusted towards less productive firms as opposed to more productive firms.
Additionally, these districts experienced a significant decline in their overall
productivity.
6
Another challenge to India’s manufacturing sector is labor market rigidity
(T. Besley and R. Burgess, 2004). Given India’s labor abundance, I hypothesize
that market integration through the GQ project would provide better access to
factors resources, incentivizing firms to adopt more labor to be more competitive.
Following the stylized results established in trade theory, this part of the chapter
investigates whether the removal of barriers between locations increases the
proportion of skilled labor in skill-abundant regions and non-skilled labor in skill-
scarce regions. We learn that labor composition did not change significantly due to
the construction of new highways, at least through the channels considered in the
chapter.
The chapter is divided into the following sections. Section 2 gives a brief
account of the relevant literature. Section 3 provides background information about
the GQ project I use as a quasi-natural experiment. Section 4 details the data
description and preparation. Section 4 lays out the empirical strategy. Section 6
presents the main results followed by some robustness checks in Section 7. Finally,
Section 8 presents the conclusion of the chapter.
2.2 Relevant Literature
Low-income countries provide a unique setup for analysis because of the
higher potential for returns on infrastructure investment. On the contrary, they
also suffer from market distortions and low quality of construction technology
that makes the efficacy of highway networks non-comparable to those in developed
countries. Therefore, returns on investment in infrastructure may vary along a wide
margin across countries.1
1The seminal paper by Chandra and Thompson (2000) leads the way for many studies in
recent decades, they study the interstate highway system in the US and their results show that
counties connected with the national highway experience significant growth in economic activity.
7
Two influential studies addressing the economic impact of highways in China
are Faber (2014) and Banerjee, Duflo and Qian (2020). Faber (2014) analyze a
more comprehensive highway network than the one I study in this chapter. He
finds that China’s National Trunk Highway System (NTHS), a highway project
connecting all provinces with an urban population of more than 500, 000, decreased
economic activity in non-targeted small peripheral counties. He argues that the
reason for this decreased activity is lower trade costs which shift economic output
towards larger metropolitan centers. To address the issue of endogenous placement
of highways he constructs a hypothetical highway network between counties based
on a minimum cost path network, and euclidean straight lines to instrument the
original highway.
Banerjee, Duflo and Qian (2020) focus on long term growth effects.
Confirming Faber’s results, they show that highways did not have any significant
impact on economic growth in the long run. While they did affect other economic
variables such as GDP per capita levels, number of firms, and firm profits, the
effects were too small to be considered of policy relevance. Their methodology also
involves constructing straight lines to proxy for highway and rail networks, while
controlling for a variety of other geographical controls.
In contrast, for the case of India we have have three studies who find robust
economic growth in locations close to the GQ project. The conclusion of this
research is quite different from these earlier studies. Ghani, Goswami and Kerr
(2016) use similar data and methods to find robust manufacturing growth in
districts closer to the GQ project whereas I find lack of evidence to support this
claim. I argue that difference arises due to different period of analysis and methods
On the contrary, the adjacent counties experience a decrease in economic activity. They establish
identification by assuming that non-metropolitan cities accidentally receive the highway network.
8
used to identify the causal parameters of interest. I restrict the study period to two
years before the GQ project construction started, allowing for better geographical
granularity of districts, and consistent definitions across years. In contrast, Ghani,
Goswami and Kerr use the year 1994, six years before construction, to validate
their results. Therefore, it becomes essential for them to aggregate a few districts
to keep the sample consistent across years. The second difference is that I do not
use the baseline levels of the outcome variable as an additional control for the
main specifications. I argue that this additional control leads to an alternate set
of identifying assumptions consequently changing the results with a wide margin.
A recent study by Chatterjee, Lebesmuehlbacher and Narayanan (2021) uses
firm-level manufacturing data similar to this chapter, to understand the returns on
public investment in the manufacturing sector. They proxy public investment with
the GQ project. A major deviation from this chapter is their identification strategy,
which uses variation in project completion times and geographical proximity to
identify the effects of the GQ project. Their results corroborate the evidence
provided by Ghani, Goswami and Kerr, and finds that firms located closer to the
GQ project had higher productivity. Khanna (2016) uses luminosity data to argue
for increased urbanization in more narrowly-defined locations near the GQ project.
He even finds spillover effects of the GQ project over time. The luminosity data
captures overall economic activity which may include effects due to other sectors,
public amenities, or agglomeration. I attempt to provide a more complete picture of
the manufacturing sector which may not be captured by the luminosity data.
The chapter is also related to the literature on resource reallocation after
a reduction in trade barriers. Harrison, Martin and Nataraj (2013) explore a
related concept, the reallocation of market shares among Indian firms after trade
9
liberalization in 1991. My approach in understanding the misallocation of resources
is very similar to their paper. The main difference is that their treatment variable
is the trade liberalization episode whereas I consider the construction of highways
as an event that reduces trade barriers. Their results emphasize within-firm
productivity improvements as opposed to the reallocation of market shares as a
channel through which aggregate productivity has improved in India in the 1990s
and early 2000s.
None of the studies in the literature directly investigate the relationship
between labor composition and highways for low-income countries. Michaels (2008)
investigates this relationship in the US using an elaborate Heckscher-Olin model.
My objective is tangential to his approach and is limited to providing insights
on labor market rigidity and the subsequent impact on skill composition through
trade-related channels.
2.3 The Golden Quadrilateral Project
The Golden Quadrilateral (GQ) project was announced by the Government
of India in 1998. The objective of the project was to connect four major metro
cities in India, Delhi, Mumbai, Chennai, and Kolkata, situated at opposite
geographical ends of the country. Figure 2 presents the GQ project on the
Indian map. The GQ project was a significant part of a bigger initiative by the
government called the National Highway Development Project (NHDP). The
objective of this broader project was to introduce and upgrade the vast network of
national highways in India. The other major project of this scheme was the North
South East West Corridor (NS-EW), which was completed almost 9 years after the
GQ project but was started at a similar time.2
2Interestingly, the GQ project also overlapped with an ancient Indian road system called the
Grand Trunk Road (GTR) which supported majority trade and labor mobility in ancient and
10
The GQ project took almost 11 years to complete. A total stretch of 5846
kilometers was introduced/upgraded with 4/6 lane high-speed highways. The
planning for the project started in the year 1998. However, the actual construction
started only between late 2000 and early 2001. A substantial portion of project
execution was done during the years 2002-2005. Finally, it was around 95%
complete by 2007. A full account of the project progression is given in Table 1.3
Table 1. Golden Quadrilateral - Progress (in kms.)
2000-01 2001-02 2002-03 2003-04 2004-05 2005-06 2006-07
Completed 643 1063 1327 2612 4697 5278 5556
Under Constr. 2093 4594 4383 3234 1149 568 290
% Complete 10.9% 18.1% 22.7% 44.7% 80.3% 90.2% 95%
Total Length 5846 5846 5846 5846 5846 5846 5846
Note: The measures are calculated from the official annual reports of the National
Highway Authority of India (NHAI). The year 2000-01 represents the financial year
which runs from 1 April, 2000 to 31 March, 2001. All values are in kilometers
2.4 Data Preparation
I use firm-level survey data on formal Indian manufacturing published every
year by the Ministry of Statistics and Program Implementation (MOSPI), India.
The survey is called the Annual Survey of Industries (ASI) and has been used
in numerous influential firm-level studies related to India (for example, Harrison,
Martin and Nataraj, 2013, Martin, Nataraj and Harrison, 2017). It is a nationwide
survey of all formal manufacturing firms employing 10 or more workers. The
data classify firms into standardized industries defined as the National Industrial
medieval India. The British upgraded this road system by introducing metaled roads and at least
one-fourth of the GQ project was engaged in further upgrading this ancient transport system to
4/6 lane highways.
3I exclude the analysis of the NS-EW corridor in this chapter, to better understand the results
from the GQ project in isolation, and without the contamination of treatment effects due to the
NS-EW corridor.
11
Classification (NIC) which is regularly revised to meet the international industry
classification standards.
MOSPI collects the ASI data as a representative sample. The sampling
procedure varies and can be classified into two categories, “Census” or “Sample”.
A standard procedure is followed each year, a complete list of registered firms is
collected from the state-level authorities, and firms are broadly categorized into
“Census” or “Sample” sectors based on their size. The Sample firms constitute a
representative sample of the population of firms in a given state and NIC-4 digit
level industry. Only a certain percentage (20% for recent years) of total firms are
recorded in the data and are assigned appropriate survey weights to match the
population at the industry-state level. The Census firms include all firms with more
than 200 workers.4
In addition, the ASI data is disseminated in two formats, a panel set of
firms and a cross-section repeated over years. The panel data is protected behind
a paywall whereas the cross-sectional format is publicly available since 2019. For
this chapter, I use the publicly available cross-sectional data. A panel data analysis
would have been ideal since it gives information about the evolution of firm-level
parameters over time. It can also accurately predict the entry and exit rates of
firms in the market, and precisely estimate firm-level productivity.
The cross-sectional data set comes with an elaborate description of firm-
level characteristics. For example, product level input-output information, kinds
of labor employed, capital assets, and other important firm-level information. It
allows the construction of several firm-related parameters, I particularly make use
of total output, total input, fixed capital net of depreciation, total emoluments
4For a more detailed exposition of the ASI data please refer to Martin, Nataraj and Harrison
(2017).
12
paid to labor including bonus, other benefits, and total labor used. I clean the
data to eliminate any outliers and errors due to data entry. Observations having
zero or negative values for the variables of interest are dropped from the sample. I
only keep firm observations that are marked as open. Therefore, the results in this
chapter will apply only to the incumbent firms.
I gather other geographical data from different sources. The actual shapefiles
of the GQ and NSEW project are from Ghani, Goswami and Kerr (2016). The
shapefile data on the railway network and the Indian coastline is collected from
MIT GeoWeb where it has been published by the International Steering Committee
for Global Mapping. While this data was recorded in the year 2007, it is an
accurate description of railways in the 1990s since the rail network in India has
not changed significantly over the last three decades.
The descriptive statistics of the data are provided in the Table A.1. The
last three rows of the table show that the average distance from railways, coastline,
and the GQ project are positively correlated with each other. Therefore, adding
these geographical controls are necessary to properly identify the effects of the GQ
project. Second, the average output is the highest for the nodal districts, which
is expected because they represent the biggest metropolitan districts of India. In
addition, the districts closer to the GQ project have a substantially higher output
as compared to districts 40 km or 120 km away.
2.5 Empirical Strategy
2.5.1 Basic Setup. The basic strategy is to compare the changes
in outcome variables from the pre-construction period to the post-construction
period when the GQ project was completely built in difference-in-differences setup.
The unit of analysis is the sub-national regions in India called districts. I assume
13
that the districts closer to the highways would experience a higher impact of the
highway as compared to districts further away. This assumption motivates the
categorization of districts into four district groups based on their proximity to
the GQ project. The first group consists of nodal districts which were originally
intended to be connected with the GQ project. The second group is defined as the
districts between (0-40) km away, the third group has districts (40-120) km away,
and the fourth group has districts (>120) km away from the GQ project.
The main focus of the analysis is the (0-40) distance group which I consider
as the treated group because it is closest to the highway. I compare (0-40) distance
group with the (>120) distance group which acts as control. The second group
(40-120) km away acts as a buffer between the two groups to capture spillover
effects due to the highway, and also helps provide an overall picture of shifts in the
manufacturing activity. For example, one possible scenario is that the introduction
of the highway has heterogeneous effects across industries, for some industries
the effects may be captured within the (0-40) range whereas for others the effects
may extend beyond 40 km. To account for such differences I take into account an
alternate treatment group (40-120). In any case, I assume that the highway effects
do not extend beyond 120 km. Therefore, I consider (>120) as the control group.
2.5.2 Main Specification. The main specification follows the long
difference approach. I compare changes in the years 2007 and 2009 with the
baseline year 2000, when the GQ project construction had started. The baseline
specification is of the form:
( ) ( ) ∑
2009/2007
ln Yi − ln Y 2000i = α + βgGQdistgroupg + ηXi + i (2.1)
g∈G
14
here, Yi is the outcome of interest, subscript i represents the district, and g
represents the four distance groups. GQdistgroupg is a dummy which captures
which distance group does the district belongs to. Xi represents district-level
characteristics such as literacy rate, rural population, population density, and
SC/ST population. These district characteristics are recorded in the year 2001. It
also includes geographical controls such as distance from the nearest rail network
and distance from the Indian coastline. The coefficient of interest is βg which
captures the decade-long effects of the GQ project on the distance group g when
compared with (>120) distance group. I consider the average of the variables in the
years 2007 and 2009 as the post-construction outcomes since most of the highway
was complete by the year 2006 (see Table 1). The average of two post-construction
years helps overcome issues arising due to serial correlation.5 The year 2000 is
chosen as the baseline since the GQ project construction started at the end of that
year. The error term i could be correlated across years for districts. Therefore, I
cluster standard errors for all specifications at the district level.
2.5.3 Identification Issues. The most challenging issue while
investigating the relationship between highways and economic activity is the
problem of reverse causality. In most cases, highways are planned with the growth
potential of locations in mind. For the GQ project, the policymakers likely
selected districts based on certain economic characteristics. Annual reports and
geographical maps show that the highway passes through some important cities
in India. Descriptive statistics in the previous section show that (0-40) district
group had significantly higher manufacturing output than districts further removed.
This makes it hard to identify the true effects of the GQ project from other
5This is after assuming that the highways impacted all treated districts similarly. Another
reason is to compare the results with Ghani, Goswami and Kerr (2016)
15
potential mechanisms or confounders which may have existed before the highway
construction started.
Figure 2. Straight line and Least Cost Path Instrument Variables
(a) (b)
Note: Hypothetical networks used as instruments for the original highway are
presented in red and the original GQ project is presented in blue. Figure 2a presents the
straight line instrument and Figure 2b presents the least cost path instrument. The linear
cost function for the least cost path network is computed using factors such as elevation,
land cover and wetlands. The background map shows district boundaries in India. First
stage estimation results are reported in the appendix.
A related problem is selection bias, where the selection of districts into the
treatment group could be driven by potential confounders. I control for a variety
of factors under Xig to take care of socio-economic characteristics and particulars
related to geographical location. Additionally, the long differences help control for
any time-invariant district fundamentals confounding the estimates. However, there
may be other unobservable factors that may remain uncaptured. A simple way to
check if these unobservables are leading to any kind of bias is by examining the
growth trajectory of outcome variables before the construction of the GQ project
conditional on the observed covariates. The data restrictions do not allow for such
16
an extension before 1998, without discounting the quality of the data.6 However, I
do provide some empirical evidence in the next section that endogeneity is a crucial
concern.
To resolve the problem of endogeneity due to the non-random selection of
districts into districts groups, I use an instrument variable strategy commonly used
in the highways literature. The first instrument is a connection of four straight lines
between the nodal districts which were initially supposed to be connected with the
GQ project. An additional kink is added to the eastern side of the instrument in
order to keep it on land. Figure 2a presents this instrument on the Indian map
alongside the original GQ project. This instrument is similar to the Euclidean
network constructed by Faber (2014). The second instrument is also a hypothetical
construction based on the least cost path between the four nodal districts. The
least-cost path instrument (LCPIV) is constructed by assigning a cost structure
to a high-resolution location grids representing the map of the whole country.
For every grid on the map, the total cost is assumed to be a function of surface
elevation, land cover, and coverage due to wetlands.7 Using this cost structure a
path is constructed which minimizes the overall costs of travel between the nodal
districts. Figure 2b depicts this instrument on the map alongside the original GQ
network.
The identification assumption is that the two hypothetical instruments
affect the outcome variables of interest only through the original GQ network. To
6The district definitions change very frequently in the 1990s because of the creation of
three new states. The reshuffling of districts makes it hard to keep the district definitions
consistent over a long period. As a result, districts need to be aggregated to keep them consistent
throughout which leads to a loss in geographical granularity and sample size.
7More precisely, the cost function follows Faber (2014). For every grid n, I compute a cost
function such that cn = Elevation+ 25 ∗ (LandCover −Dummy) + 25 ∗ (Wetland−Dummy)
17
control for factors that may potentially violate the exclusion restriction, I include
the two sets of control variables discussed before, district-level characteristics and
geographical characteristics in all specifications estimated in the chapter. Table A.2
presents the first stage results with and without the additional controls. The first
stage is estimated using the two-stage least squares method. The first stage results
show that both instruments are highly correlated with the original highway with
F-statistic for the straight line IV of 38.8 and for the least cost path IV as 65.27.
In addition to the socio-economic and geographical characteristics, I also
estimate the main specification after including the baseline levels of district output
as an additional control. I do this only for the case of manufacturing output.
This additional control serves two purposes. First, it helps compare the results
in this chapter with Ghani, Goswami and Kerr (2016), who use similar data and
methodology. Second, it helps shed light on the sensitivity of results under two
cases, that is, with and without the baseline levels as an additional control. The
identification assumptions under these two cases are very different, discussed
more elaborately in the appendix. In a nutshell, including this additional control
amounts to conditioning the estimates on lagged dependent variables whereas
excluding it conditions on district fixed effects.
2.6 Empirical Results
2.6.1 Exploring Selection Bias. To provide a complete picture of
the long-term effects, I plot the dynamic trajectory of the raw outcome variables
(output, labor, capital and number of firms) ranging from the pre-construction
period in 1998 till the year 2009 when the GQ project was complete. Figure 3 plots
these graphs. A visual inspection of these plots shows that there is no significant
increase in any of these variables. The growth trends are similar for every distance
18
group. The level differences between distance groups confirm evidence from the
descriptive statistics that the distance groups are characteristically different from
each other. The raw plots show that the magnitude of effects from the GQ project,
if any, were not large enough to decisively change the growth trajectory of districts.
However, the (0-40) district group shows a marginally higher increase in total
output (Figure 3a) and fixed capital (Figure 3c) after the year 2000 when compared
with other distance groups. To put this into context, the (0-40) distance group
experienced grew at an average rate of 10% per annum between 2000 and 2009
highest when compared with other distance groups. This exploratory analysis
shows that the GQ project did not disproportionately affect the overall growth
trajectory of manufacturing activity in districts close to it. However, we do see
comparatively higher growth in output and capital in the (0-40) distance group.
In the next step, I investigate the more promising growth in manufacturing
output. It is important to understand which factors contributed to this growth
and how much of it is driven by the construction of the GQ project. I identify
the effects of the GQ project by controlling for as many district-level factors
as possible. The first set of specification assumes a dynamic panel framework
and estimates a fixed effects model with districts as the unit of analysis. I
include district and year fixed effects and re-estimate the growth trajectory of
manufacturing output. I estimate the following equation,
∑2009
ln(Yigt) = α + βgtGQdistgroupig × Y eart + λt + µi + igt (2.2)
t=1998
here, subscript i represents the district, subscript g represents the distance groups,
and subscript t represents the year. GQdistgroupg represents a dummy for the
distance category of either (0-40) km or (40-120) km away, two separate regressions
19
Figure 3. Comparing manufacturing sector outcomes for different district groups.
(a) (b)
(c) (d)
Note: This figure plots raw manufacturing sector variables for districts categorized
into different distance groups based on their proximity to the GQ project. Nodal
represents the group of four main districts connected by the GQ project, that is, Delhi,
Mumbai, Chennai, and Kolkata. It also includes the sub-urban districts located near the
main ones. All variables are log transformations.
are run for both distance groups with (>120) as the base category. Y eart is a
dummy for all years between 1998 and 2009 with the year 2000 as the base year.
Finally, µi represents district fixed effects and λt represents year fixed effects.
The district fixed effects help capture district fundamentals such as resource
endowments, demographic and cultural aspects whereas year fixed effects capture
overall changes to the Indian economy. The coefficients βgt capture the effect of the
20
GQ project on the distance groups for every year t. These coefficients are plotted
separately for (0-40) and (40-120) distance groups.
Figure 4. Comparing annual changes in outcomes after controlling for district fixed
effects.
(a) (b)
Note: This figure plots coefficients βgt from a fixed effects model defined by
Equation 2.2 for two separate distance groups. Figure 4a compares treated districts with
the control districts and Figure 4b compares intermediate districts with control districts.
The black dotted lines represent confidence intervals and the significance levels are marked
alongside the nodes following conventional norms (* p < 0.1, ** p < 0.05, *** p < 0.01).
The results show wide differences in growth between different distance
groups with districts (0-40) km away showing increased improvements in
manufacturing output. Figure 4a and Figure 4b show these contrasts. The point
estimates βgt show a growing trend for the (0-40) district groups whereas the trends
are almost flat throughout the sample period for (40-120) district groups. These
coefficient estimates are not significant except for a few at the end of the sample
period but have a positive magnitude suggesting a positive effect.
The major cause of concern in these estimates is the lack parallel trends
before the construction of the GQ project. Figure 4a shows that the growth
trend starts in the year 2000 as opposed to the years after 2000 when the GQ
construction started. This makes it hard to attribute the differences after 2000
21
to the GQ project specifically. The unobservable factors driving growth before the
construction of the GQ project could be driving the future trajectory of growth
in (0-40) district groups. More pre-construction periods could help give more
information on how these trajectories evolved, what underlying factors could be
driving them, and if we can specifically attribute this growth to the GQ project or
not.
I check the extent of bias due to the self-selection of districts into the
treatment group by estimating a logit model which predicts the construction
of the GQ project using economic characteristics recorded before the highway
was constructed. I use the dummy for the distance group (0-40) km away as the
dependent variable and regress it on a battery of district-level characteristics
recorded in the year 2001, and manufacturing industry characteristics recorded
in the year 1999. Ideally, the district characteristics must also be recorded before
2000, but due to lack of data, I assume that the demographic variation across
districts does not change significantly over these two years. The equation I estimate
is as follows, ( ) ( )
Pr GQdistgroupi(0−40) = 1 = Λ ψ0 + ψ
1999
1Xi1 + ψ2Xi2 (2.3)
Here, X1999i1 represents district characteristics such as literacy rate, rural population,
population density, distance to the rail network, and distance to the coastline. X2t
represents manufacturing industry characteristics recorded in the year 1999, these
include the outcome variables studied in the following sections, total output, share
weighted average productivity, skill intensity, and urban share.
Table A.3 presents results from the above model. The coefficient estimates
on total output are statistically significant even after controlling for other district
characteristics. This means that districts with higher output had a higher
22
Table 2. GQ construction main specification results
OLS SLIV LCPIV
(1) (2) (3) (4) (5) (6) (7) (8) (9)
No Controls Controls Controls No Controls Controls Controls No Controls Controls Controls
Nodal = 1 -0.414*** -0.083 0.282 -0.354** 0.110 0.604 -0.281 -0.078 0.308
(0.132) (0.208) (0.251) (0.168) (0.311) (0.381) (0.177) (0.256) (0.319)
Treated = 1 0.068 0.139 0.374*** 0.281 0.459 0.760 -0.237 -0.117 0.139
(0.123) (0.130) (0.138) (0.592) (0.677) (0.678) (0.339) (0.353) (0.377)
Intermediate = 1 0.177 0.213 0.241 0.199 0.174 0.560 1.033 1.022 1.219**
(0.181) (0.180) (0.161) (0.949) (1.057) (0.905) (0.635) (0.655) (0.602)
Baseline -0.249*** -0.266*** -0.219***
(0.043) (0.060) (0.053)
N 284 284 284 284 284 284 284 284 284
R sqr. 0.012 0.041 0.19 . . . . . .
Note: This table presents the main results for the effects of the GQ project on manufacturing output growth. The coefficient
estimates are produced by running Equation 2.1 with change in logged levels of output between 2000 and 2007/2009 as the
dependent variable. Columns (1)-(3) present the OLS estimates and (4)-(9) present the IV estimates. Standard errors are clustered
at the district level.
* p < 0.1, ** p < 0.05, *** p < 0.01, Standard errors in parentheses.
probability of getting treated with the GQ project. The negative sign in column
(3) on the variables representing the distance from the nearest rail network and
the coastline show that the GQ project was constructed close to these geographical
entities. The results show that manufacturing industry characteristics of districts
recorded in the year 1999 played a crucial role in determining the treatment of
districts with the GQ project. Therefore, the estimates from the long differences
presented in the next section will be biased. Therefore, the results presented by the
instrument variables will be crucial for causal interpretation of results.
2.6.2 Main Estimates: Manufacturing Output. The results from
the main specification defined in Equation 2.1 are presented in Table 2. The OLS
estimates presented in columns (1)-(2) show that the district groups (0-40) and
(40-120) experienced higher output growth than the (>120) district group. The
coefficient estimates are positive but statistically insignificant. Additionally, the
magnitude of estimates is higher for the (40-120) district group when compared
with the district group (0-40). This is contrary to the expectation that the treated
23
district group (0-40) will experience the highest growth due to the construction of
the GQ project.
The qualitative aspects of the OLS estimates are repeated in columns (4)-
(5) where the original highway is instrumented with the hypothetical highway
built along a straight line. The point estimates remain positive but the magnitude
increases, pointing at the fact that the GQ project targeted more prosperous
regions biasing the OLS estimates downward whereas the actual estimates may be
larger. In contrast, the least-cost path instrument produces different results. The
coefficient estimates on the (0-40) district group are almost zero with controls and
negative without controls whereas the (40-120) district group shows statistically
significant coefficients with higher magnitudes. These results predict that both
district groups (0-40) and (40-120) experienced positive output growth when
compared with (>120), but higher growth is seen in intermediate districts.
However, noise and uncertainty makes these point estimates non-significant.
The IV results for the straight line and the least-cost path instruments
are very different. This difference could be due to two possible reasons. Either
these two instruments are targeting fundamentally different types of districts and
these differences are not controlled for in the control variables, or, the implicit
assumption stating that the effects of the GQ project are homogeneous across
districts may not be true. The GQ project may have had heterogeneous effects
across districts along dimensions that may be difficult to identify. The two
instruments may pick up some of these heterogeneous effects separately. Later in
the chapter, I check how far we can extend these local average treatment effects
to the average treatment by showing differences in characteristics of complying
districts under these two instruments.
24
The differences in magnitudes of the coefficient estimates for (0-40) and (40-
120) district groups warrant further discussion. One potential reason is diminishing
returns, due to which economic growth could be negatively correlated with the
baseline output levels. In other words, districts with higher baseline output
levels may experience lower economic growth than a district with lower baseline
output levels. Therefore, we may need to estimate the main specification while
conditioning on the baseline levels of district output. Columns (3), (6), and (9)
show results after controlling for the baseline levels of output for each district. The
results in column (3) are statistically significant, and the magnitude is higher for
the district group closer to the highways.
An important point to note is that adding baseline output levels changes
the conditional assumptions underlying the main specification. As discussed
in subsection 2.5.3, columns (2) and (3) are estimated under a different set of
assumptions. Column (2) conditions the regression equations on district-level
controls and time-invariant confounders. Column (3) on the other hand, conditions
based on district-level controls and the baseline level of output. As is evident in
columns (2) and (3) the point estimates are sensitive to these two alternative
specifications. It is difficult to know which of the controls correctly identify the
effects of the highways since both specifications are independent and not nested
into each other.
Ghani, Goswami and Kerr (2016) estimate long differences while controlling
for baseline outcome levels and find positive significant effects of the GQ project
on districts as shown in column (3). Based on that they conclude that districts
near highways gained significantly from the GQ project. This conclusion may be
25
misleading for the alternate specification for which, the intermediate district group
seems to gain more from the highways than the treated group.
2.6.3 Manufacturing Productivity. The previous section presents
suggestive evidence that the positive effects of the GQ project may not be as
large as we have seen in the literature. The effects are more pronounced in the
intermediate districts than the treated districts. In this section, I explore other
channels through which highways may have affected manufacturing output in the
districts. I explore two channels more relevant for the Indian economy. These
channels can be summarized under two basic questions I try to answer. First,
do highways shift economic activity towards more productive firms? Second, do
highways shift factors of production towards more favorable spatial locations?
These questions have been investigated in two different strands of the
literature. Hsieh and Klenow (2009) highlight the importance of misallocation of
factor resources in India’s manufacturing sector. They propose a theoretical model
which captures market distortions due to misallocation of resources among firms
and use it to compare the magnitude aggregate productivity between different
countries. They argue that misallocation helps explain the vast differences in
aggregate productivity between India and other developed countries. Resource
misallocation could emerge due to market distortions created by inefficient labor
markets, trade barriers, industrial policy, and other policy-related factors.8 As a
consequence, less productive firms can hold a vast proportion of labor and capital
resources. The introduction of the GQ project has the potential to reduce this
misallocation by optimally allocating factors of production across space. As a
8The literature has always struggled to find a dominant source of misallocation. The factors at
play are more context-dependent. A set of potential factors are explored in a survey by Restuccia
and Rogerson (2017)
26
result, we may see efficient firm selection into the market due to the increased
competitiveness of the manufacturing industry.
The second potential reason for negligible growth comes from frictions in
India’s large labor markets. Factor mobility in India has been low and to what
extent do highways alleviate this issue is a question of crucial policy importance.
Some policy reports provide evidence that labor migration across states is very
restricted, whereas a large portion of internal migration in India happens between
locations within states. In a different regulatory context, Banerjee, Duflo and
Qian (2020) point to a similar issue in the case of China, where factor mobility is
restricted due to peculiar government regulations which they argue could be one of
the critical factors why we see no effects of highways on economic growth.
In the backdrop of low factor mobility, I test the main results from the two-
factor Hecksher-Olin model which holds only under complete factor mobility. More
specifically, I test a version of Rybczynski theorem which states that regions with
higher skill endowment will specialize in skill-intensive goods, and vice versa. I test
whether the introduction of the GQ project changed the skilled labor composition
of districts along these lines.
District Productivity and Firm Selection. To capture shifts in the
firm selection, I construct two measures of district-level productivity. The first
measure is a simple average of productivity for all firms in a given district. The
second measure weights each firm’s productivity with its output share. The prime
focus of this section is the second measure, which helps capture the reallocation of
output shares among firms based on their productivity. If highways bring trade-
related competitive effects to regions, we will see more productive firms selecting
into the market and gaining higher market shares (Melitz, 2003). This measure
27
could shift due to two reasons. First, changes in average levels of firm productivity,
which represent within-firm enhancements in production efficiency. Second, due to
the reallocation of market shares between firms.
To construct this measure, I first compute firm-level productivity using a
method proposed by Levinsohn and Petrin (2003) (LP). The residuals of the log-
linearized Cobb-Douglas production function are used as a measure of total factor
productivity and can be estimated according to the following equation.
ln(GV Anjt) = αj + β1jln(Lnjt) + β2jln(Knjt) + φnjt (2.4)
Here, the subscript n represents the firm, j represents NIC 3-digit level industry,
and t represents the year.9 GV Anjt is the gross value added by the firm every year,
Lnjt is the total labor used, Knjt is the fixed capital assets of the firm, and φnjt is
the residual which captures the total factor productivity at the firm level. In the
above specification, the capital and labor coefficients estimated by ordinary least
squares may be biased. A common issue of endogeneity arises due to simultaneity
bias, where the level of output may determine firms’ level of fixed capital adoption.
The LP method estimates consistent labor β1j and capital β2j coefficients for each
industry while accounting for endogeneity in capital usage. They use control inputs
like materials, fuel, to inform about the productivity dynamics. More specifically,
firm productivity is assumed to follow a first-order Markov process and is proxied
by a polynomial expressed in capital and control inputs. This helps extract a
consistent labor coefficient which I then use to compute the capital coefficient
9NIC stands for national industrial classification, a standard for industry classification in
India which is comparable to the international industry classification standards. A 3-digit NIC
classification refers to the broader industry, for example, textile industry, paper industry, or, food
products.
28
assuming constant returns to scale.10 Due to the cross-sectional nature of the data
I cannot directly implement this approach. Therefore, I estimate one period lagged
values of relevant variables for each firm by taking the average of closest industry-
size-location matches from one period lagged years, a method similar to Sivadasan
(2009). After estimating labor and capital coefficient for each 3-digit level NIC
industry. I take the residuals as the measure of a firm’s total factor productivity.
Consequently, I estimate district-level share weighted average productivity as
follows,
∑
Φit = snitφnit (2.5)
n
here, snit is the market share of the firm n in a district i. There are two things
to note about this measure. First, it implicitly captures district-level changes
in productivity due to the entry/exit of firms and doesn’t parse these effects
separately. Second, it captures both within the industry and across industry market
share shifts. The literature argues that trade liberalization generally shifts market
shares between firms within well-defined industries (Melitz, 2003). Focusing on
within industry-district market share shifts would be more plausible. However, as
discussed before the data sample is not designed to represent the population at the
industry-district level unit. The industries become very sparse at the district level
and many of them are not recorded for the majority of the years, making it difficult
to track manufacturing activity within industry-district units. As an alternative,
I look at the overall changes within the district. Therefore, shifts in output shares
could come from within or across industry reallocations of market shares.
10Structural methods like these are regularly used in the industrial organization literature to
overcome endogeneity issues while estimating production functions.
29
Table 3. GQ treatment effect on productivity changes within districts
OLS SLIV LCPIV
(1) (2) (3) (4) (5) (6)
No Controls Controls No Controls Controls No Controls Controls
Panel A: ∆ Productivity (share-weighted mean)
Nodal 0.041 -0.374 -0.152 -0.728** 0.146 -0.309
(0.233) (0.230) (0.295) (0.350) (0.256) (0.259)
Treated -0.059 -0.203 -0.272 -0.676 -0.164 -0.276
(0.127) (0.145) (0.394) (0.517) (0.209) (0.252)
Intermediate -0.339** -0.371** -0.748 -0.575 0.058 0.070
(0.149) (0.157) (0.679) (0.605) (0.478) (0.429)
Panel B: ∆ Productivity (mean)
Nodal 0.096 -0.316* 0.006 -0.489** 0.271 -0.261
(0.177) (0.178) (0.222) (0.249) (0.207) (0.216)
Treated -0.243*** -0.316*** -0.192 -0.504 -0.326** -0.437**
(0.088) (0.099) (0.286) (0.361) (0.162) (0.188)
Intermediate -0.203** -0.195* -0.513 -0.402 0.385 0.331
(0.102) (0.110) (0.459) (0.411) (0.367) (0.319)
Note: This table follows a similar setup to Table 2 but with share weighted productivity as
the dependent variable. Standard errors are clustered at the district level.
* p < 0.1, ** p < 0.05, *** p < 0.01, Standard errors in parentheses.
The results are presented in Table 3. Panel (A) presents results from the
second measure representing changes in market share weighted district productivity.
The OLS specification results in columns (1) and (2) show negative and statistically
significant estimates for treated as well as intermediate districts, meaning that
market shares were distributed to less productive firms in these district groups.
Instrumenting the original highway with the straight line and least-cost path
switches the sign for intermediate districts but the treated districts retain their
negative sign and significance. These results show that the GQ project forced
the market shares to be reallocated towards less productive firms. One possible
explanation is that more productive firms are gaining ground in districts away from
the highways.
30
Panel (B) further solidifies these results, it shows the effects on average
district productivity without weighting on market shares. This measure captures
changes within-firm productivity and provides more insights into the overall
productivity of the districts. The results show that average productivity declined
decisively for the treated districts with negative point estimates significant
under OLS and IV specifications. Three possible scenarios explain these average
productivity results. First, incumbent firms in the treated group are leaving for
more favorable locations. Second, incumbent firms experience a decline in their
productivity which seems highly unlikely. Third, highways attract more economic
activity but due to market distortions, low productivity firms can enter the market.
Two plausible reasons could potentially bias these estimates. First, there are
strong assumptions on the way productivity is estimated. For example, firm-level
characteristics other than location, size, and industry could influence the firms’
adoption of resources, making the estimated lagged variables for firms inaccurate.
Second, the nature of the production function could be different. Following the
literature, I assume that the production function exhibits constant returns to scale,
this may or may not be appropriate in the Indian context.
District Labor Composition. In the previous exercise, I looked at the
firm selection and changes in district productivity. In this section, I explore India’s
low labor mobility and rigid labor laws which add to the challenge of efficient
allocation of resources across firms. The purpose of this section is not to directly
assess the labor mobility across India as that would require an elaborate analysis
beyond the scope of this chapter. Instead, I check for evidence of labor mobility
through trade-related channels more often associated with the highways. I use
results from the Hecksher-Olin (HO) model of trade to study the changing nature
31
of labor composition in districts. One of the important results from the HO model
is the Rybczynski theorem, which states that under an open economy with trade,
locations favor the production of goods for which they have a higher abundance of
relevant factor endowments. For example, districts with higher skill endowments
will start specializing in more skilled labor-intensive goods. These results only hold
under the assumption that there is non-frictional labor mobility across space and
the labor laws favor easy adoption of skilled labor.
I investigate whether districts close to the GQ project attract more skilled
labor into manufacturing.11 I construct a simple skill intensity measure by taking
the ratio of skilled labor to total labor engaged in formal manufacturing for a given
district year. Here, the treatment variable is the interaction between the dummy
for the distance groups and pre-construction endowment of skilled labor, measured
by districts’ literacy rate recorded in the year 2001. Specifically, I estimate the
following equation,
( ) ( ) ∑
2009/2007
ln S −ln S2000ig ig = α+ βgGQdistgroupig×Literacy2001ig +Xig+ig (2.6)
g∈G
here Literacy2001ig represents district literacy rate for districts in the year 2001.
The results are presented in Table 4. The results are too noisy to draw any useful
conclusions. The signs of point estimates change depending on which instrument is
used. The extreme volatility of results points at two possible explanations. First,
the effects of the GQ project on labor markets in the districts are heterogeneous
across regions. Second, the specification defined in Equation 2.6 may not be correct
11Michaels (2008) empirically tests the predictions of the Hecksher-Ohlin (HO) model and
concludes that, contrary to the predictions of the HO models, highways in the US didn’t affect
demand for high skilled workers relative to low skilled workers. However, he finds that it did
increase the wage bill of high-skilled workers in counties with relatively high skilled intensity
before the highway construction.
32
Table 4. GQ treatment effect on changes in labor composition within districts
OLS SLIV LCPIV
(1) (2) (3) (4) (5) (6)
NoControls Controls NoControls Controls NoControls Controls
Nodal 0.080 0.080 0.122* 0.222 0.098** 0.094
(0.049) (0.050) (0.067) (0.313) (0.049) (0.061)
Nodal X Literacy -0.115 -0.105 -0.186* -0.355 -0.153* -0.137
(0.078) (0.086) (0.113) (0.572) (0.079) (0.109)
Treated -0.016 -0.007 0.963 1.116 -0.002 0.006
(0.035) (0.033) (2.186) (2.377) (0.081) (0.087)
Treated X Literacy 0.058 0.051 -1.492 -1.721 0.023 0.025
(0.064) (0.060) (3.452) (3.782) (0.132) (0.147)
Intermediate 0.009 0.019 -0.273 -0.290 0.089 0.111
(0.030) (0.031) (0.933) (1.036) (0.058) (0.068)
Intermediate X Literacy 0.028 0.017 0.396 0.382 -0.138 -0.182
(0.054) (0.057) (1.367) (1.490) (0.114) (0.134)
Literacy 0.081*** 0.076*** 0.152* 0.116 0.119*** 0.101***
(0.028) (0.029) (0.087) (0.089) (0.035) (0.029)
N 284 284 284 284 284 284
R sqr. 0.12 0.19 . . . .
Note: This table estimates Equation 2.6 with the ratio of skilled labor to total labor as the
dependent variable. Standard errors are clustered at the district level.
* p < 0.1, ** p < 0.05, *** p < 0.01, Standard errors in parentheses.
and other functional forms may define the relationship between skill composition
and highway construction.
Overall, the results from this exercise confirm the findings in Michaels (2008)
who finds similar insignificant results in the case of US. These results also suggest
that the rigid labor markets in India may have restricted the incentives for firms to
adopt skilled labor as a resource.
2.7 Robustness
2.7.1 Local Average Treatment Effect. A potential concern is that
the IV estimates presented before only address the treatment effects for a subset
33
of “complier” districts.12 The complier districts are those which are randomly
assigned to the GQ project and whose nature of treatment was decided by their
allocation into the treated group based on the distance from straight line or least
cost path instruments. In a counterfactual scenario, they would not have received
the highway had they been assigned to the control group by these instruments.13
The characteristics of the compiler group define the local average treatment effects
(LATE) estimated in the previous sections. If this group is significantly different
from the overall population of districts then this is not a a valid instrument.
The LATE might not inform us about the average treatment effects (ATE). An
implicit assumption throughout the IV specifications has been that the effects are
homogeneous across all districts. However, there may exist heterogeneous treatment
effects with a potential to limit the results presented in the chapter. To investigate
this problem, I check whether the complier districts are different from the rest of
the population.
It is not possible to directly identify the complier districts in the data.
However, we can estimate the proportion of compliers in the population using the
following expression.
| Pr(Zi = 1)(E(Gi|Zi = 1)− E(Gi|Zi = 0))Pr(G1i > G0i Gi = 1) = (2.7)
Pr(Gi = 1)
Where Gi is a dummy for whether the original GQ project categorizes a district
into the treated group or not, G1i and G0i represent the potential for treatment
12Here the terminology “compliers” is similar to what is used in the program evaluation
literature, see Angrist and Pischke (2009).
13Contrast this with “never takers” who do not receive the GQ project irrespective of their
assignment to treatment or control groups, and “always takers” who receive the project under all
conditions. An implicit assumption is that there are no “defiers”, meaning a district does not have
the power to go against its initial assignment.
34
and only one of these observed in the real data. Pr and E are probability and
expectation operators, respectively. Zi is the dummy for whether the instrument
predicts treatment into the treated or control group. Using the above expression,
the proportion of compliers for the straight line instrument is 31.3% and for the
least cost path network is 36.2% of the districts.
The next crucial question is how different are these complier districts from
the rest of the population? Although it is not possible to identify the complier
distircts directly, Angrist and Pischke (2009) show that we can still know about
the characteristics of the compliers. To implement their strategy, I first create
dummies for district characteristics by categorizing districts into above or below
median value of a characteristic. Following this, we can measure the relative
likelihood of a complier district belonging to the above median category of a
district characteristic by taking the ratio of the first stage for the districts with
above median characteristics to the overall first stage. The results of this exercise
are reported in Table 5.
For the straight line instrument, column (2) and (7) show that the complier
districts had comparatively higher literacy rates and were generally closer to
the coast line. For the least cost path instrument, column (2), (3), (6), and (7)
show that districts close to this network had higher literacy rate, more rural
population, and were closer to both, existing rail network and the coastline. This
robustness check shows that there may be significant heterogeneity in the way
highways affected the districts. It also helps understand the main results where
both instruments show different estimated coefficient. One potential reason for
these differences could be characteristically different complier districts being picked
up by the two instruments.
35
Table 5. Testing for Heterogeneous Effects
(1) (2) (3) (4) (5) (6) (7)
FullSample literacy rural pop dens scst pop H rail H coast
Panel A: First Stage: Straight Line Instrument
GQ 40kms iv 0.626*** 0.720*** 0.807*** 0.662*** 0.631*** 0.658*** 0.823***
(0.068) (0.069) (0.044) (0.056) (0.085) (0.123) (0.039)
N 208 107 97 97 110 111 104
R sqr. 0.25 0.33 0.38 0.32 0.24 0.22 0.26
Panel B: First Stage: Least Cost Path Instrument
GQ 40kms iv2 0.702*** 0.784*** 0.838*** 0.681*** 0.738*** 0.832*** 0.849***
(0.056) (0.044) (0.042) (0.056) (0.062) (0.069) (0.037)
N 208 107 97 97 110 111 104
R sqr. 0.32 0.39 0.47 0.35 0.39 0.49 0.37
Note: The first-stage results are presented for the treated and control district groups only.
The first column presents full sample first-stage regression. The rest of the columns run
separate first-stage regressions after dividing the data into sub-samples based on district
with above median characteristics. Standard errors are clustered at the district level.
* p < 0.1, ** p < 0.05, *** p < 0.01, Standard errors in parentheses.
2.7.2 Sampling Errors: Sub-sample Analysis. The problem
with the manufacturing data I use is that due to unknown sampling errors it is
not representative of the population of firms at the district level. For the main
specifications, I assume that the sampling errors are small such that they do
not impact the main results presented in the chapter. However, if the sampling
errors are significant then the estimates could be biased in two potential ways.
First, if the proportion of firms sampled in the observed data is lower than the
real population of firms in districts closer to the highways, the results presented
could be downward biased. Second, if this proportion is higher the results could be
upward biased.
I make use of sampling details provided along with the ASI data to check
the extent of the bias generated by the sampling errors. As discussed before, the
ASI data is disseminated in two different categories. The Census category of firms
36
Table 6. Main results with Census firms: Manufacturing Output
OLS SLIV LCPIV
(1) (2) (3) (4) (5) (6)
∆lnOutput ∆lnOutput ∆lnOutput ∆lnOutput ∆lnOutput ∆lnOutput
Nodal -0.360** -0.152 -0.308* 0.174 -0.252 -0.121
(0.145) (0.227) (0.182) (0.335) (0.187) (0.269)
Treated 0.044 0.088 0.508 0.714 -0.177 -0.070
(0.117) (0.128) (0.633) (0.732) (0.302) (0.320)
Intermediate 0.211 0.250 -0.065 -0.076 0.877 0.890
(0.192) (0.194) (1.031) (1.170) (0.615) (0.635)
N 284 284 284 284 284 284
R sqr. 0.012 0.037 . . . .
* p < 0.1, ** p < 0.05, *** p < 0.01, Standard errors in parentheses.
Note: This table follows a setup similar to Table 2 with manufacturing output as the
dependent variable. The data considered is a sub-sample of firms with more than 200 workers.
These firms represent the whole population of bigger sized firms, therefore, sampling errors are
very minimal. Standard errors are clustered at the district level.
represents the population of manufacturing firms with more than 200 workers, no
sampling is done for this category and the whole population of firms is recorded.14
I drop the firms belonging to the Sample category which contributes the most to
the sampling errors, as only 10-20% of these firms are sampled every year. I then
re-estimate the main specifications from Equation 2.1 on manufacturing output
growth only using the Census firms. This helps confirm the robustness of results
presented in subsection 2.6.2, but only for firms with more than 200 workers.
If the results are driven by sampling errors, then this exercise should
output different results than those presented before. The results are presented in
Table 6. The magnitude and sign for the OLS and IV coefficients in all columns is
qualitatively similar to the main results presented in Table 2. This provides some
reassurance that the results presented before are not significantly affected by the
sampling errors.
14This may still have other measurement errors commonly encountered in any random sample of
a data.
37
2.8 Conclusion
This chapter finds no direct evidence related to the positive effects of
highways on manufacturing output growth. On the contrary, it finds that districts
located closer to the highways experienced a decline in their manufacturing
productivity and market shares shift towards less productive firms. Additionally,
no significant changes in labor composition are found in response to the highways.
The results in this chapter help consolidate a wider picture of growth in low-
income countries. Services and the informal sector has played an important role
in India’s GDP growth over the past years. The double-digit growth during the
decade of 2000s was propelled by large productivity gains in the services sector.
In addition, the informal sector employs a significantly large proportion (80% of
non-agricultural labor) of India’s labor markets.15 The effects of the highways may
not be as large in the manufacturing sector as has been found in other studies.
However, the GQ project may have led to developments in the service or informal
sector.
There are three potential ways to forward our understanding of the effect
of highway effects on economic activity in low-income countries. First, for the
manufacturing sector, other potential channels should be explored. For example,
labor market rigidity and resource misallocation are likely to be important margins
along which highways affect growth in low-income countries. Second, for an overall
picture, it is important to know how highways may affect the labor markets,
and the kind of opportunities they create for individual households by providing
access to markets. A tentative step in this direction is taken in Chapter 3. Third,
since highways bring large scale urbanization, it is worth exploring its impact on
15Rough estimates as per the International Labor Organization found here
38
the informal sector, a major component of employment in low-income countries.
Exploring these additional channels can provide a more clear picture of the returns
to infrastructure investment in developing countries.
39
CHAPTER III
NATIONAL HIGHWAYS AND FEMALE LABOR FORCE PARTICIPATION
3.1 Introduction
Women may face asymmetrically higher socioeconomic costs to enter the
labor force compared to men due to unfavorable gender norms, and lack of suitable
job opportunities. These can negatively affect their economic prospects despite
significant economic growth. When it comes to female labor force participation
(FLFP), several mechanisms have been proposed regarding the effect of economic
growth. For example, a common argument is that growth has a U-shaped non-
monotonic effect on FLFP. Besides the culture-based explanation pursued in
Chapter 4, one explanation is that during the initial stages of economic growth,
women transition from necessary subsistence work to household production because
of an income effect. As the economy grows and real wages for women continue to
rise, the opportunity costs for household production increase and women substitute
household work for market work.
Due to regional disparities in opportunities for market work, the role of
transport infrastructure in this type of transition is critical. Roads can provide
access to schools, colleges, conducive work environment and better paying jobs.
Therefore, roads may not only help accelerate the movement along the U-shape
but also help with structural transformation in regions where growth is contingent
upon geographical connectivity. In this chapter, I provide suggestive evidence
that in rural areas of districts located closer to national highways, skilled women
working in manufacturing and services sector withdrew from market production. In
contrast, the agricultural sector shows signs of an increase in labor supply of low
skilled female workers. This evidence aligns with U-shaped relationship between
40
FLFP and economic growth, additionally suggesting that highways do open up
market opportunities for female workers.
India has seen a secular decline in FLFP during the past three decades. This
is surprising because at the same time fertility levels have fallen, and for married
women, that should have made more time available for market work. Moreover,
increases in labor productivity due to higher female educational attainment would
have opened up job opportunities. The rural regions have contributed more to
India’s FLFP decline whereas the urban centers have experienced stagnation
(Afridi, Dinkelman and Mahajan, 2018, Klasen and Pieters, 2015).
In this chapter, I explore how the construction of the Golden Quadrilateral
(GQ) Project may have led to FLFP decline in rural areas. The analysis shows
that the overall treatment effect of the GQ project is non-significant on FLFP in
districts located closer to it. Interestingly, pre-construction district characteristics
play an important role in determining the future impact of the project on FLFP
in non-agricultural sectors. In particular, I find that districts located closer to
the project that had higher literacy rates, lower road connectivity, and higher
caste fractionalization before the construction of the highway project experienced
a sharper decline in FLFP. These changes are primarily due to decreased female
employment rather than increase in working age population from migration towards
districts with higher literacy. The results are robust to several district level controls
and can be interpreted causally.
I establish these results using a difference-in-differences methodology. I
use the GQ project constructed in India between 2000-2009 as a quasi-natural
experiment and study its effects on districts located closer to it (treated group),
in comparison to districts located further away (control group). Consequently, the
41
universe of districts within India are divided into groups based on their proximity
to the GQ project. Treated and control groups are defined, and an incidental
group is considered to account for any spillover effects. FLFP is then compared
between these groups starting in the year 1999 (before construction) and 2009
(after construction). To estimate district level measures, I make use of data from a
wide variety of sources. I use the National Sample Survey (NSS) dataset to capture
employment information on households in districts, augmenting this with Census
data to add district level economic characteristics such as literacy rate, population
characteristics, and finally, extracting infrastructure related variables from the
publicly available SHRUG data set.
This chapter contributes to the literature in significant ways. First, I provide
evidence that large scale highway projects may have asymmetric effects across
genders. Second, I highlight the importance of prevailing market conditions in
determining how highways may impact a given location. Third, I add to the U-
shaped FLFP literature by providing evidence at a more dis-aggregated level and
show how the transformation may take place in the regional economies. Fourth,
I emphasize the role played by increased market access brought upon by the
construction of the highways.
3.2 Relevant Literature
Several literatures are relevant to this chapter. First, my work contributes
to the limited evidence on asymmetric effects of national highways across genders
in developing economies. Lei, Desai and Vanneman, 2019 address this issue on
the intensive margin, that is, hours worked. They use a household level panel
dataset, the India Human Development Survey (IHDS) collected between 2004-
2005 and 2011-2012, to predict the probability of individual working in the non-
42
farm as compared to traditional farm sector. They find that access provided
by paved or unpaved roads and frequent bus service increases the likelihood of
employment in the non-agriculture sector for both men and women. These effects
seem to be stronger for women. Interestingly, they also find that women’s non-
farm employment increases in communities with more egalitarian gender norms,
measured by the practice of purdah.1 The results in this chapter are estimated
at the district level and capture overall changes in local labor market rather
than looking at changes at the household-level. The results confirm the broader
argument of their paper. Similar to their conclusions, I show that even at the
district level rural areas with better road connectivity saw an increase in female
labor force participation due to the construction of the GQ project.
Melecky, Sharma and Subhash (2018) use a methodology similar to this
chapter to estimate the wider economic benefits of the GQ project. They also argue
that initial local market conditions play an important role in determining how
certain locations benefit from the highways.2 They emphasize that districts with
higher levels of secondary education attainment prior to the project construction
experienced a significant transition from farm to non-farm employment. In
addition, they look at a wider set of outcomes dealing with poverty, household
consumption, and environmental impacts. This study extends their analysis
by thoroughly examining female labor force participation in particular and
investigating the role played by the GQ project in its decline in India for the past
1According to Lei, Desai and Vanneman (2019), purdah is the practice of covering one’s face
with a piece of cloth like a scarf or shawl. This practice is followed in the presence of men and
while stepping out of the household.
2They also cover another national highway project built alongside the GQ project but that was
significantly delayed, the North South East West (NSEW) Corridor. I remove districts treated
with the NSEW corridor from my sample, and focus solely on the GQ project.
43
three decades. To improve upon their methodology, I incorporate a decade long
pre-construction time series to test for parallel trends between treated and non-
treated district groups.
In addition to gender specific effects, this chapter contributes to the
literature on the effects of transport infrastructure on rural economic development.
One of the few studies relevant in this context, specifically for India, is by Asher
and Novosad (2020). They argue that local rural roads facilitate structural
transformation by incentivizing workers to move out of agriculture. However,
they don’t find any major changes in agricultural outcomes, such as income, or
assets. Similarly, Shamdasani (2021) finds that road connectivity for remote villages
leads to crop portfolio diversification, and farmers start to invest in more labor
and capital intensive production methods. Overall, she augments the structural
transformation results presented by Asher and Novosad (2020) but adds that
agriculture sector grew as opposed to the non-farm sectors for some villages
despite road connectivity. I show that the gender specific effects are important,
highways may impact certain demographics through radically different channels.
I provide evidence that structural shifts from agriculture to non-agriculture
sectors may not be happening for women, even when the regional economies are
experiencing significant economic growth. I find that the effects of highways is
heterogeneous across districts with some districts experiencing a decline whereas
others experiencing an increase in FLFP. I propose potential channels that may
explain these shifts.
Finally, I provide evidence in support of the U-shaped theory of female
labor force participation proposed by Goldin (1990). More specifically, results in
this chapter support the predictions of the theory that during the initial period
44
of economic growth women’s labor supply might decline. In the context of India,
Afridi, Dinkelman and Mahajan (2018) demonstrate that increased education
levels among rural population of men and women leads to a fall in FLFP. They
argue that relatively higher returns to home production compared with market
production may be driving these results. The results presented in the study suggest
that these effects might be more localized and may as well be applicable to the
regional economies.
3.3 The Highway Project
To understand the relationship between highways and the local labor
markets, I use the introduction of a large infrastructure project in India, the
Golden Quadrilateral (GQ) project, as a quasi-experiment to parse out the effects
of highways on Indian districts. The GQ project was introduced with the objective
of connecting four major metro cities in India located at the four corners of the
country. I will use this project to study how it may have affected household
incentives through local market connectivity. Briefly, the project was proposed
in the year 1999 and the construction was started only in the second half of the
year 2000. The project was more than 90% complete by the year 2006. Therefore,
I consider the year 1999 as the pre-construction period and the year 2009 as the
post-construction period. A more detailed description of the project is given in
Chapter 1 of this dissertation.
3.4 Empirical Strategy
This section lays out the empirical strategy to analyse the impact of the GQ
project on the regional labor markets. I rely on a standard difference in difference
approach to estimate the causal effects of the GQ project. The publicly available
45
data makes it easier to construct a long time-series of outcomes variables before the
construction of the project that can be used to check for pre-construction trends.
I conduct the analysis at the district level. According to the census of
2001, India had 593 districts in total, a few of which were carved out of bigger
districts since the country’s independence. To keep a consistent panel of time
series at the district level, the districts are collapsed to their parent regions and
423 individual districts remain in the data set for the analysis. I allocate districts
into four distinct groups based on their distance from the GQ project. The nodal
districts lie at the four points that form the corners of the GQ project. The first
group contains districts that are 0-40 kms away from the highway, the second group
contains districts that are 40-120 kms away from the highway and finally, the last
group contains districts 120-500 kms away from the highway. The focus throughout
the analysis is on the first group that is assumed to benefit the most from the GQ
project. I also consider the second group to take into account any spillover effects
from the highway construction. The third group is the control group.3
I follow a simple identification strategy. I compare the labor market
outcomes in treated and control district groups before and after the construction
of highways. The hypothesis is that district groups closer to the GQ project (0-
40kms) will show significant effects in the outcome variables as compared to district
groups (120-500kms) further away from the highways. To check for these temporal
changes in labor market outcomes at the district level, I estimate the following
equation .
3This classification of districts is similar to Chapter 1, to keep a consistent and comparable
sample I keep the same classification for this Chapter as well.
46
∑3
∆Yi = βgDig +Xi + i (3.1)
g=1
Here Yi is the dependent variable such as district level FLFP. The subscript i
represents district, g represents distance group to which the district is allocated
based on its distance from the highway. For example, g = 1 represents a dummy for
districts falling on the nodal points of the highway, g = 2 represents a dummy for
districts 0-40 kms away from the GQ project. Xi represents district level controls
recorded before the highway construction was started.
As discussed later in this chapter, treatment heterogeneity will play a
crucial role in determining the causal effects of the GQ project. For example,
as per the U-shaped hypothesis, rural areas of the districts along the project
with higher presence of the manufacturing industry may experience a decline in
FLFP. I capture this treatment heterogeneity by interacting district characteristics
with the treatment dummies. Specifically, I interact the Dig dummy with a set
of district level characteristics recorded before the GQ construction had started.
Consequently, I will also estimate the following equation.
∑3 ∑3
∆Y = β1i gDig + β
2
gDig × Inti + Inti +Xi + i (3.2)
g=1 g=1
Here, Inti represents the interaction term for each district that captures the pre-
construction economic characteristics of the districts such as literacy rate, road
connectivity, and caste fractionalization.
Endogeneity Issues: A common issue while estimating the causal effects of
highways on economic activity is that highways are usually built where there is
already a potential for economic activity to grow. In the context of this project,
47
female labor participation may not directly influence the placement of highways but
it may influence the incidence of highways through other channels. For example, if
we assume that gender norms based on cultural factors such as societal outlook
towards women participating in market work is positively correlated with the
economic output of the region, then regions with higher economic output may
see more women participating in the labor force. It is likely that highways will be
placed in regions with high economic output or with gender norms favorable for
women’s participation in the labor force. Therefore, highway placement may not be
completely random and may be governed by the district characteristics correlated
with the outcome variables of interest.
The general approach to circumvent this identification problem in a
difference-in-differences framework is to test for parallel trends between treated
and control regions in the pre-treatment period. A precise time series analysis of
this kind is not possible because the survey data I use is conducted every 5 years.
However, I run placebo tests that mimic the GQ project construction before the
actual construction and check if any significant trends are prevalent before the GQ
construction started. I discuss this more in the next sections. In addition to this,
I take several other precautions to make sure that the estimated coefficients are
not biased. I consider regions which were “accidentally” treated with the highway
project. In other words, I exclude all nodal districts (the four metro cities) and
adjacent regions which were supposed to be connected with the highway project
from the treated group. Additionally, I add a set of control variables accounting
for district level characteristics that may potentially affect determine how districts
respond to the construction of the highway.
48
3.5 Data
The data I primarily use for this chapter is the household level survey
conducted after every 5 years in India, the national sample survey (NSS). The
survey is divided into four parts covering various aspects of the Indian household.
These parts are named schedules, and each schedule is roughly surveyed every 5
years. I make use of the schedule that covers the employment and unemployment
information of each member of the sampled household. The survey is conducted
separately for rural as well as urban areas of the country. The rural sample is
representative at the district level whereas the urban sample is representative at the
sub-state level but not at the district level.4 The survey helps construct labor force
participation rates and other labor market characteristics for the rural areas at the
district level for almost two decades between 1987-2009. I focus on the surveys that
cover periods during and around the time when the GQ project was constructed.
Surveys relevant for this study are National Sample Survey Office (NSSO)’s 43rd
round (July 1995-June 1996), 52nd round (July 1995-June 1996), 55th round (July
1999-June 2000), and 64th round (July 2007-June 2008), 66th round (July 2009-
June 2010), 68th round (July 2011-June 2012). In this version of the chapter, I
only make use of the 43rd, 55th, and 66th rounds to assess the impact of the GQ
project.
To construct the measures for labor force participation at the district
level, I make use of relevant survey responses from the household members. For
each member, the survey records “current activity status” which has more than
10 different categories that signify the status of an individual. For example,
“worked as a casual wage earner”, “sought work”, “did not seek work”, etc. I
4The sampling for the urban areas is done from the NSS regions, that are generally defined as
an aggregate of 4-5 districts within a state.
49
use a specific set of responses to categorize each sampled individual into whether
she participated in the labor force or not.5 The data also allows for capturing
the industry in which the worker is employed. For everyone employed, the data
provides industry classifications according to the National Industry Classification
(NIC). I use these NIC codes to identify the broader sectoral categories such as
agriculture, manufacturing, and services where an individual works. Additionally,
I use the education and marriage information of the household members to further
decompose the effects of the GQ project on different categories. In addition to the
NSS data, I also use census information from the 2001 survey to construct district
characteristics such as literacy rates, rural population, and caste fractionalization.
I extract the census information from Martin, Nataraj and Harrison (2017)
publicly provided along with their study. To check the baseline results before the
construction of the GQ project I also use census information from the 1991 survey
made available in the SHRUG data set.6 The SHRUG data set also helps provide
information about the number of colleges, schools, and other public amenities (local
roads, etc.) at the district live that I use to check heterogeneous impacts of the GQ
project. Summary statistics of selected variables are provided in the Table B.1
3.6 Results
This section presents the main results of the study. I first check the trends
of the outcome variables in treated versus control district groups before the
construction of the GQ project. Next, I present the main results. I start with
5For example, I consider an individual to be in the labor force is he/she belongs to either of
these categories defined by NSS 55, “self-employed”, “employer”, “unpaid family worker”, “wage
employee”, “casual wage laborer”, “other types of work”, and “did not work but was seeking”.
6SHRUG stands for Socioeconomic High-resolution Rural-Urban Geographic Platform for
India. It is initiative by economists to consolidate data on the Indian economy from a variety of
sources for easy access. More details about the data can be found at its website here.
50
average treatment effects that simply compare the treated districts with the control
districts. This is followed by an analysis of heterogeneous treatment effects where I
interact district economic characteristics and the treatment dummies.
3.6.1 Baseline Comparison. To causally interpret the results it
is critical to assess the pre-construction behavior of the districts. As mentioned
earlier, the difference-in-differences methodology stipulates that the treated and
control groups exhibit parallel trends in outcome variables before the treatment
takes effect. For this project, data restrictions make it impractical to perform such
an analysis. The primary data used for this study, specific schedules of the NSS
data, are surveyed only after every 5 years making it harder to visually inspect
the trajectory of outcome variables across time. Alternatively, I can run placebo
regressions. Placebo regressions repeat the main analysis of comparing the treated
and control districts but using a different time period, where the treatment variable
remain the same. Any significant changes from the main results will shed light on
pre-trends that can be a possible cause of worry for causal inference. I run these
placebo regressions for the period between 1987 and 1999 with the same treatment
dummies that are used in the main regressions. I use Equation 3.1 to estimate
these regressions.
The results of this exercise are presented in the Table 7. The regressions
are run separately for men and women. I also add a set of controls that account
for distance from the railways and the coastline.7 The coefficient estimates are
non-significant, except for women in the treated districts. This implies that the
districts closer to the GQ project were experiencing a decline in female labor
7I add more controls in the main equations that account for several other district
characteristics. Since data on district characteristics is not available for the year 1987, I have
not included those controls here.
51
Table 7. Baseline - Placebo Treatment Effects
Male Female
(1) (2) (3) (4)
∆FLFP ∆FLFP ∆MLFP ∆MLFP
Nodal 0.046∗∗ 0.020 0.015 0.007
(0.020) (0.029) (0.012) (0.012)
Treated -0.032 -0.054∗∗ -0.003 -0.010
(0.024) (0.024) (0.011) (0.009)
Intermediate -0.010 -0.016 0.013 0.012
(0.029) (0.028) (0.011) (0.008)
District Controls No Yes No Yes
N 294 294 294 294
R sqr. 0.0057 0.021 0.0036 0.013
Note: This table presents results from placebo regressions run
for the period 1987-1999 with the GQ project as treatment and
female labor force participation rate as the outcome variable.
The coefficient estimates are produced by running Equation 3.1
and standard errors are clustered at the district level.
* p < 0.1, ** p < 0.05, *** p < 0.01, Standard errors in
parentheses.
force participation. These trends are in line with the overall changes that were
happening in the female labor markets during that time (see Afridi, Dinkelman and
Mahajan, 2018). As I will show in the later sections, this falling trend on average
is reversed for the treated districts but hastened after the construction of the GQ
project for districts with higher literacy rates .
3.6.2 Main Results. I estimate the average treatment effects
by comparing treated and control districts. The results are estimated using
Equation 3.1 with labor force participation as the outcomes variable, estimated
separately for both men and women in the sample. The results are presented in
Table 8. The coefficient estimates of average treatment effects for both men and
women in treated districts is positive after controlling for the set of controls. The
magnitude of these positive effects is very low. For example, in the case of women,
the highway construction only led to 2 percentage points increase in the labor force
52
participation over the decade during which the GQ project was constructed. It is
hard to make any causal claims from these estimates since they are not statistically
significant, and consequently indistinguishable from zero. Taking into account the
placebo effects estimated earlier, the magnitude of effects in treated districts is now
opposite. This shows that there was a significant shift in the trends of labor supply
due to the construction of the GQ project. The secular decline for all districts
between the years 1987 and 1999 was replaced by a positive or zero net effect
between 1999 and 2009.
It is not surprising that the average treatment effects from the GQ
project are not significant. The high standard deviations of district economic
characteristics presented in Table B.1 suggests that they (treated, intermediate,
etc.) were different in many respects. These differences could have played an
important role in determining the future impact of the GQ project. For example,
the marginal benefits from the highway construction could have been much higher
for districts that had relatively low market opportunities available for women before
the construction. In other words, there may be diminishing returns to the highway
construction with low developed districts gaining the most. In addition, there could
be a variety of other channels where initial market conditions interact with the GQ
project construction to bring about specific changes that might not get captured
in the average treatment effects. I next explore these district-level heterogeneities
in terms of economic characteristics of districts recorded prior to the GQ project
construction and investigate how they may have determined the trajectory of
changes in the local labor markets.
The most important margins along which the highways affect districts are
highlighted when we look at the different production sectors separately. Much
53
Table 8. Average Treatment Effects
Women Men
(1) (2) (3) (4)
∆FLFP ∆FLFP ∆MLFP ∆MLFP
Nodal 0.062∗ 0.092∗∗ 0.003 0.026
(0.034) (0.040) (0.034) (0.040)
Treated -0.019 0.020 -0.012 0.008
(0.031) (0.033) (0.014) (0.013)
Intermediate -0.013 0.023 -0.026∗∗ -0.011
(0.028) (0.028) (0.012) (0.011)
District Controls No Yes No Yes
N 322 322 322 322
R sqr. 0.0038 0.079 0.0092 0.041
Note: This table presents the main results for the effects of
the GQ project on female labor force participation (FLFP).
The coefficient estimates are produced by running Equation 3.1
with change in FLFP between 1999 and 2009 as the dependent
variable. Standard errors are clustered at the district level.
* p < 0.1, ** p < 0.05, *** p < 0.01, Standard errors in
parentheses.
empirical research has emphasized the structural change in local economies due
to road construction (see Asher and Novosad, 2020, Shamdasani, 2021, Aggarwal,
2018). This change shifts surplus labor from agriculture to manufacturing or
services due to increases in agricultural productivity facilitated by economic
channels opened up by road infrastructure. The decade through which the GQ
project was built seems to have gone through similar economic shifts. This is shown
in Figure 5 where I plot the overall change in the share of women’s employment for
the three production sectors before and after the construction of the GQ project.
The shift in the labor share from agriculture to manufacturing and services is very
evident.
Taking this into account, I dissect the district-level employment data into
three primary production sectors and calculate female labor force participation
for each separately. To take into account the role played by initial economic
54
Figure 5. Women’s Employment Share by Industry Sectors
55
characteristics I then interact the variables representing these characteristics
with the district group dummies and finally estimate the regression coefficients
defined in Equation 3.2. I estimate these equations separately for the agriculture,
manufacturing, and services sector. I consider a variety of initial district level
characteristics as interaction terms but only present results for literacy rate, paved
roads, and caste diversity. I briefly talk about other parameters at the end but do
not emphasize those results because of the lack of statistical significance.
Literacy Rate: I construct the rural literacy rate of districts using the Indian
census data of 2001. I take the ratio of the total “literate” population in the rural
areas of a district to the total rural population. The definition of literacy in India
at an individual level that one must be able to read and write with understanding.
The expectations to accomplish this are low, and, on average, anyone with primary
education is considered literate. I interact these literacy rates with treatment
dummies and run regressions as per Equation 3.2 for each production sector
separately. The results are presented in Table B.2. The coefficient estimates are
significant (at 1%) for the treated districts. Broadly, the results suggest that FLFP
declined more for districts that exhibit higher literacy rates before the construction
of the GQ project. This holds for manufacturing as well as the services sector. On
the contrary, the agricultural sector saw positive shifts but the coefficient estimates
are non-significant.
To interpret these results statistically and compute the overall effects, we
can take the coefficient on the interaction terms, multiply it with the literacy rate
in a district, and add the overall coefficient representing the effects on district
groups. For example, the services sector results outlined in column (6), there is a
3.12 percentage points decrease in FLFP due to the GQ project for districts with
56
a literacy rate of 70% (0.057 − (0.126) × 0.7). Similarly, one point increase in
the literacy rate of a district leads to a decline of 0.126 percentage points in the
FLFP. For a clearer exposition, I simulate these interaction effects in Figure 7
for all sectors. These simulations clearly show that FLFP in the manufacturing
and services sectors is declining for high-literacy treated districts . The threshold
literacy rate lies between 0.4 and 0.6. This implies that, on average, districts with
literacy rates higher than the threshold on average will experience a decline in
FLFP, whereas, districts with lower literacy rates on average will experience an
increase in FLFP.
Since the FLFP is the ratio of labor force and total population in a
district, changes in FLFP could be due to shifts in population across districts.
A potential concern could be that labor is migrating towards higher literacy
districts to access better education and other related amenities. Therefore, the
decline in FLFP presented in Table B.2 could be driven by population shifts rather
than changes in employment levels or contraction of labor force. To test what is
driving these results, I look at changes in the employment levels of districts. The
results presented in Table B.3 show that employment levels also decreased in the
manufacturing and services sector where rural literacy was higher.
To test whether these results are due to the GQ project instead of
unobservable factors already prevalent before the construction, I run the placebo
regressions with the treatment dummies interacted with the literacy rate of districts
in the year 1991. The outcome variable is similar to before, that is, the change in
FLFP between 1987 and 1999. For this analysis, I run the regression Equation 3.2.
The results are presented in Table B.4. Here, none of the coefficient estimates are
statistically significant. This provides confidence in the fact that there were no
57
alternate unobservable factors in the pre-treatment period that affected the treated
districts.
A possible explanation for these results is the U-shaped behavior of the
female labor force participation. The theory first proposed by Goldin (1990), posits
that in the initial stages of economic growth the female labor supply declines as
production moves from individual households to businesses in the market. This
happens because low skilled women generally have strong social stigma attached to
manual work that is available outside the household. Decline in labor participation
may also be seen for educated women as they may choose to focus more on
investing in children’s education, an argument that we pursue in Chapter 4.8 The
U-shaped hypothesis has been tested in the case of India by Afridi, Bishnu and
Mahajan (2019) where they use the same data as used in this chapter but perform
a country-wide study and argue that the U-shaped behavior is at play in India
as well. Their decomposition exercise shows that FLFP in rural India declined
between 1999 and 2011. They argue that this decline can be explained by the fact
that more number of young women are pursuing education and married women are
finding it more beneficial to invest their time in household production.
The literacy interaction results confirm this idea at the district level and
show that the infrastructure investments like the GQ project accelerate the
movement along the U-shaped curve. Significantly different results for agriculture
and non-agricultural sectors help magnify the movement along the U-shape curve.
Figure 6 shows that the average skill level of female workers in these sectors is
quite different. The services sector has the highest share of labor with graduate
education and the agriculture sector has the higher share of people with primary or
8It is assumed that women disproportionately bear the responsibility of childbearing in a
household and draw away from market production.
58
Figure 6. Women’s Employment Shares by Industry and Skill level
no education. In light of this distribution, we can assume that the declining female
labor supply in the services and the manufacturing sector is primarily driven by
more skilled women. In contrast, women with low skill levels seem to be engaging
in agricultural labor in larger numbers.
In summary, the GQ project led to a movement along the U-shaped curve
with FLFP in the manufacturing and services sector declining significantly, and the
agriculture sector drawing in more unskilled women into the markets.
Rural Infrastructure: Next, I check whether the existing road infrastructure
in the rural areas had a significant role to play in the way districts benefited
from the GQ project. I construct district aggregates that capture paved road
59
Figure 7. Simulation of Literacy Interaction Effects
connectivity to villages in 2001. I use the SHRUG data set to construct these
measures. The data provides binary response variables for every village describing
whether a village has a paved road connectivity or not. I aggregate these binary
responses to the district level and use it to proxy for rural road connectivity of the
district. In other words, the measure captures the number of villages that have
paved road connectivity in a district. I take log of this variable to interpret the
results in percentages.
The results are presented in the Table B.5. Similar to the literacy
interaction effects, the effects are more evident in the non-agricultural sectors.
However, two differences stand out. First, the coefficients on the interaction terms
60
are positive for all the sectors, suggesting that female labor for participation
increased in districts with better rural connectivity before the construction of
the GQ project. Second, the coefficients for the manufacturing and agricultural
sectors are non-significant, suggesting that the services sector was the most affected
through this channel. Examining the coefficient estimates in column (6) we can see
that the magnitude of positive change is high. For example, a one percent increase
in rural road connectivity increases the impact of the GQ project on services FLFP
by 0.5%. These results could mean that village road connectivity was augmented
by the GQ project, opening up wider range of market opportunities for women.
Since a major share of high-skilled workers is in the services sector, the results
suggest that the nature of these opportunities was specific to the skilled workers.
These results are opposite to the previous exercise with literacy interactions.
Ideally, economic growth is assumed to be positively correlated with the literacy
rate, rural road infrastructure, and development related characteristics of districts.
Therefore, the results in the above analysis should be similar to the literacy
interaction effects. In other words, if the FLFP is declining for districts with higher
literacy rates then it must decline for districts with better rural road infrastructure.
However, the best possible explanation for the contradictory results is that
two different economic forces are counter-acting each other. To understand this,
we must note that the cross-sectional variation in district-level literacy rates and
rural road infrastructure may not be positively correlated with each other. These
economic characteristics might be a product of district and state-level budget
allocations where there are trade-offs between education and road infrastructure
investments. A simple correlation exercise shows that these two characteristics
are negatively correlated in the case of India in the year 2001. Therefore, the GQ
61
project might be affecting FLFP through different channels that are actuated based
on the pre-construction district characteristics. I previously discussed the U-shaped
phenomena in the case of literacy interaction effects. The other channel, different
than the U-shaped phenomena discussed earlier, could be the opening up of market
opportunities. The GQ project could have augmented the already existing rural
road infrastructure and opened access to opportunities that may not be accessible
through local roads, as a consequence increasing FLFP. This argument is aligned
with the literature on intra-country trade where highways provide access to remote
markets.
Caste Fractionalization: The literature has actively explored the role
played by identity in economic decision-making. In the context of labor supply
decisions in India, Oh (2021) argues that these decisions are influenced by the
caste association of an individual. She finds that individuals favor jobs that are
associated with their caste/community and do not take up even a higher-paying
jobs if the job under consideration is traditionally performed by a lower caste.
Similarly, theoretical research has pointed out caste networks as a valuable resource
that opens up opportunities in urban areas (see Munshi, 2019). These social factors
could significantly influence the female labor force participation decisions in the
treated districts. It is interesting to see how these social factors interact with the
construction of the GQ project and affect the local labor markets. For example, at
the district level, a more homogeneous caste population will be able to benefit more
from caste networks and increase FLFP. Additionally, this would also mean that
individuals have a higher probability of finding jobs that are associated with their
own caste group, and hence have lower social costs to bear while taking up a job
offer.
62
To test this hypothesis, I construct a caste fractionalization index using the
Indian census data from the year 2001. The index represents caste diversity within
a district. Specifically, I use a Herfindahl concentration method to construct caste
concentration based on the population share of the scheduled castes, scheduled
tribes, and the general category. I subtract this concentration index from 1 and use
it as a measure for caste fractionalization. Higher levels of the index indicates more
caste diversity. More diversity restricts the set of caste specific job choices female
labor can choose from when they participate in the labor force. Therefore, having a
negative effect on average district-level FLFP. By the same logic, lower levels of the
index may have positive effects on FLFP.
The results presented in Table B.6 suggest that the caste fractionalization
did not play a significant role in districts located next to the GQ project.
Nevertheless, regressions for the services sector show some support for the
evidence underlining the mechanisms discussed by Oh (2021). The magnitude
of coefficient estimates is negative for all the sectors, meaning that higher caste
fractionalization (more diversity) leads to lower levels of FLFP. The interaction
terms are statistically significant only for the services sector, meaning that one
index point increase in the diversity of the district leads to 6.6% drop in female
labor force participation.
These results can be further interpreted in light of the evidence provided
by Oh (2021). She runs a field experiment with 640 male participants to study the
effect of identity on labor supply decisions. The study focuses on male subjects,
whereas the results presented in this chapter study similar effects for female labor
supply. Caste fractionalization is generally interpreted as the probability that
two randomly selected people come from different caste backgrounds. Higher
63
probabilities in this context would lead to more labor market conflicts while taking
up a job offer, negatively affecting the overall supply of female labor force in the
region.
3.7 Discussion
In the previous section, we saw that the GQ project on average did not have
any significant effects on the treated districts located 0-40 kms away. However,
we learn that significant heterogeneities across districts determined the future
gains from the GQ project. Higher literacy rates and caste fractionalization within
districts before the construction of the GQ project led to a decline in FLFP in the
treated districts. On the contrary, better road connectivity in the rural areas of the
districts led to positive changes in FLFP from the project. In this section, I discuss
some of these results in more detail and provide further support for the arguments
presented in the previous section.
I argued that the reason for decline in FLFP in districts with higher
literacy rates is that the GQ project could be accelerating movement along the
U-shaped curve. The main idea behind this theory is that relative returns to home
production increase (compared with market production) with increased economic
growth and activity due to the highway construction. To further strengthen this
argument, it may help to highlight changes in how women choose to allocate time
for different activities. The NSS data provides information on whether a household
member is pursuing education or engaged in domestic activities, among other
activities. However, this data becomes very thin at the district level and makes
the analysis infeasible. However, I can test whether this decline is primarily driven
by educated versus less educated and married versus unmarried women.
64
I first categorize district-level female employment according to women’s
education characteristics. I recognize three categories of education, primary
schooling and no education, secondary schooling, and graduate-level education. For
each of these education categories, I compute the share of their employment in year
1999 and 2009. I then run similar regressions as before, defined in Equation 3.2.
The results of this exercise are presented in Table B.7. As expected, there is a
significant decline in the share of women’s employment with graduate degrees in
districts that had higher literacy rates. Moreover, districts with higher literacy
rates also show an increased share of women employed with secondary education.
The coefficients on these interaction terms are statistically significant, providing
robust evidence in favor of the U-shaped hypothesis.
The U-shaped path of female labor force participation could explain the
significant effects we see on educated women. The theory underlines two competing
forces that govern women’s labor market decisions. First, women are motivated
to join market production since it increases household utility by bringing in more
income. Second, the opportunity cost of market production is time spent in the
household production raising children and performing other domestic duties. The
construction of the GQ project led to economic growth in the treated districts
(see Khanna, 2016). Households with lower income levels and human capital will
respond to this growth by joining the labor force. On the contrary, women in
households with higher income levels and human capital will withdraw from market
production as they devote more time in children’s education that becomes more
important with economic growth.
I also check whether this decline is concentrated among married women.
Since the choice of the market versus home production would more often arise in
65
the case of married women, married women would contribute more towards the
declining FLFP. Similar to the exercise above, I categorize women employed into
married vs non-married categories and compute the FLFP rates. For the married
category I also include women who are either divorced or widowed. Finally, I test
the change in employment levels of married vs unmarried women caused by the
GQ project. The results are presented in Table B.8. Here too, the married women
show a significantly larger decline in the female labor force participation in the
services and manufacturing sector, augmenting the previously presented results and
strengthening the U-shaped hypothesis.
3.8 Conclusion
This chapter presents evidence that the construction of a national highway
project led to heterogeneous effects on female labor force participation (FLFP)
in districts located closer to the highways. Districts with higher literacy rates,
lower levels of road connectivity, and higher caste fractionalization experienced
a significant decline in FLFP. This decline is predominantly concentrated in the
manufacturing and services sector.
The results are consolidated by proposing two broader theories of structural
transformation. First, suggestive evidence is presented in favor of acceleration
downward along a U-shaped transformation of FLFP. The highway-led economic
growth may have led to appreciation in the returns to children’s education, making
investment in home production more beneficial. Consequently, skilled women may
be withdrawing from the labor force to engage in child rearing duties, lowering
the district level FLFP. Second, the highway construction shows signs of opening
up market opportunities for households. Districts with lower literacy rates seem
to have gained the most from these new opportunities. In addition, the national
66
highways potentially augmented rural road connectivity, because districts with
better road infrastructure before highway construction also saw an increase in
FLFP.
In future work, I plan to extend the analysis in this chapter in three
important dimensions. First, the key implicit assumption that holds together the
explanation for the decline in FLFP is that highways create economic growth. I
plan to check this association by capturing economic growth through the luminosity
data from the SHRUG dataset and providing robust evidence on how the labor
market may have changed on the demand side. Along with this, I also plan to
strengthen the asymmetric effects on women by incorporating an analysis of change
in income levels and household consumption patterns data collected from the NSS
surveys. This will help shed light on the household level decision making may
have influenced changes at the extensive margin, at the district level. Second, I
wish to explore the kind of market opportunities that were created due to the
GQ project and how they might be different from the market access provided by
the local rural roads. Third, robustness checks are needed to provide confidence
in the causal claims made in the chapter. The first set of robustness checks will
incorporate more data from the NSS surveys that are not covered in this chapter.
This will provide an up-to-date complete trajectory of FLFP in time and help with
additional placebo regressions. Taking into account significant heterogeneity in
the treatment effects, I also plan to estimate the causal effects using alternative
techniques such as a combination of matching and difference-in-difference methods.
67
CHAPTER IV
FEMALE LABOR FORCE PARTICIPATION IN INDIA
This chapter is co-authored with Dr. Shankha Chakraborty. The empirical
analysis and numerical simulations were done by me under the supervision of
Dr. Chakraborty. The analytical model was developed by Dr. Chakraborty, and I
provided proof-checking assistance. The writing of sections 4.2.1, 4.2.2, and 4.5 was
done by me, and the rest of the chapter was written by Dr. Chakraborty.
4.1 Introduction
A central goal of economic development is to improve the living standards
and opportunities of a population. In practice, those improvements are unevenly
spread across socio-economic groups. Of interest is how they affect disadvantaged
groups, particularly women.
This chapter studies how economic development facilitates the participation
of women in the market economy and dissolves gender norms. It is reasonable to
think that economic development will raise wages for women and, through reduced
fertility, shift their time allocation from demanding childcare to compensated
market work. One measure of such work is the female labor force participation
(FLFP) rate, defined as the proportion of women 15 years and older who are
working. During 1880-1998, income (GDP) per person in the U.S. increased by
a factor of 11, full-time earnings of American women rose from 46% to 67% of
men’s earnings and fertility rates for white women fell from about 3.5 children
to 1-2 (Bailey and Hershbein, 2018; Goldin, 1990; Maddison, 2001). During that
same period, the FLFP rate steadily rose from 2% to more than 70% in the U.S.
(Fernández, 2013). This general increase in FLFP is observed in many other
advanced economies.
68
Contrast this to the case of India, an emerging economy. Between 1990 and
2016, India’s GDP per person grew by a factor of 3.5 and fertility fell from 4.5 to
2.33 per woman. Over the same period, India’s FLFP rate dropped from 30.38% to
23.18% (Figures 8 and 9). Why?
Recent work has been documenting how this drop in India’s FLFP is
occurring in several sectors, education groups and religious communities. Some
researchers also provide explanations relying on within-family resource allocation
and weak job opportunities for women (for example, Afridi, Bishnu and Mahajan,
2019, Klasen and Pieters, 2015, Neff, Sen and Kling, 2012). What is lacking,
though, is a comprehensive understanding of women’s labor supply decisions
in an economy that is undergoing rapid structural change. In particular, some
commentators have equated the falling FLFP with worsening gender equality
(Pande and Moore, 2015) and to treat the Indian experience as unique. Neither
is necessarily true.
Figure 8. Female Labor Force Participation Rate, India
To understand why not, compare the Indian case with contemporary and
historical ones. During 1990-2016, as China’s GDP per person grew by a factor of
11.8, its FLFP rate fell from 73.20% to 60.64%, even more than India’s. Morocco’s
69
Figure 9. Fertility rate 1990-2018
Figure 10. FLFP 1990-2019
70
FLFP rate, on the other hand, first rose from 23.18% (1990) to 26.56% (2004)
before falling to 22.27% (2016) during a time when its GDP per capita increased
by a factor of 1.86.1 The pattern across countries, as Figure 10 shows, is not
uniform and falling FLFP is hardly unique to India. In fact, falling FLFP is also
observed in historical data. Goldin (1990) finds, for example, that the labor force
participation rate of married – as opposed to all – women in the U.S. declined
from the late-nineteenth to early-twentieth centuries before increasing. A similar
decrease was observed in late-nineteenth century England when more and more
educated women were staying home to raise families (Clark, 2015). Viewed in this
light, decreasing FLFP may be a normal pattern of economic development as long
as it is temporary, that is, soon followed by increasing participation. To the extent
that one can extrapolate time series implications from cross-sectional data, Figure
11 shows that the association between FLFP and economic development has been
non-monotonic.
In this chapter, we propose a tractable analytical framework of household
decision-making that identifies factors behind a potentially U-shaped FLFP. The
model has several ingredients. Decisions are made in a unitary household on
consumption, fertility, child investment and labor supply by the wife. The latter
is subject to social costs: specifically, the very decision to work outside the house
involves a social stigma that the household internalizes.2 The wife’s decision to
work therefore depends on balancing the obvious benefits – higher household
income – with the costs – time away from childbearing and care giving, and the
1Contemporary FLFP data from the International Labor Organization, income and fertility
data from World Development Indicators (World Bank).
2Alternatively, the wife internalizes her husband’s resistance to the idea of her working. This
may not be as far fetched as it sounds; see Lowe and McKelway (2019) for experimental evidence
in a different context.
71
Figure 11. FLFP and per capita income, Afridi, Bishnu and Mahajan, 2019
psychic cost of breaking a social taboo. The tradeoff regarding time away from
children is particularly salient: working outside means not just less time in raising
young children, but also time away from monitoring them, looking after them and
other time investments that augment their human capital. Anecdotal evidence and
commentaries, for example, point to these as important margins for urban Indian
women (M. B. Das and Zumbyte, 2017).
The tractable household model outlined below shows three things. First,
unchanging culture in the sense of unchanging attitudes towards women’s work
can interact with economic forces in interesting ways to produce a non-monotonic
response of the female labor force participation rate to economic development.
Secondly, child quality investment within the household is an important margin
– particularly where fertility rates are falling – affecting women’s decision to work.
When households are primarily investing mother’s time in quality investment, labor
supply can be particularly costly unless wages were to be sufficiently high. Third,
FLFP is not necessarily non-monotonic. This depends on the importance of culture
72
and level of development: societies less resistant to the idea of working wives and
mothers and/or at a level of development where the gender wage gap is relatively
low would see FLFP rates rise with further development.
Section 4.2 below presents some empirical evidence on India’s FLFP and
related contributions to the literature. The household model is presented in section
3 which is used to construct the aggregate FLFP rate in section 4. Section 5
outlines the next steps we plan to take in the future and ideas for an extension.
4.2 Background
4.2.1 Additional Facts. Much has been written about the secular
increase in FLFP in the advanced economies since industrialization. Goldin’s
(1990) seminal work on the US labor market since the late nineteenth century has
informed modern commentaries about the challenges and nuances of labor force
participation of married women.
The quirkiness of India’s FLFP during last three decades of robust growth
has, not surprisingly, attracted a lot of attention. India’s overall FLFPR has
declined from an already low of 22% in 1987. If we also consider the status of
subsidiary activities then, Sarkar, Sahoo and Klasen (2019) show that the overall
FLFP has declined from 51.4% in 1983 to 38.7% in 2011 for the age group of 25-
55. These changes when decomposed at the rural and urban level, show contrasting
pictures.
In the rural areas, the decline has been more significant among the age
group of 25-65 years old. Afridi, Dinkelman and Mahajan, 2018 perform a
decomposition analysis of the National Sample Survey (NSS) data and conclude
that the higher decline in rural areas is because of an increased participation in
domestic work and higher levels of education attainment. They conjecture that
73
only having primary education is not sufficient, as it provides skills complementary
to household domestic work. Therefore, the opportunity cost of participating in the
labor force has increased, driving down rural FLFP.
Figure 12. Secondary enrollment 1990-2019
Figure 13. LFPR by education (urban, married and not married, age 20-45) Afridi
et al. (2020)
Another quirk of India’s FLFP has been the weak association between
education and work. It stands to reason that rapid economic growth creates both
demand for skilled labor and incentives to acquire education. As in several other
Asian economies, female education has been steadily rising in India during 1990-
2019 (Figure 12). Figure 14 plots changes in school level educational attainment
74
Figure 14. Changes in female educational attainment in India
for women in India above the age of 25. There has been a consistent increase in
educational attainment at all levels of schooling, primary, secondary and tertiary.
Also, a simultaneous decrease in the number of females who never went to school
has declined significantly between 1985 and 2010. Yet FLFP of educated women
has stagnated.
This is especially clear when we look at urban areas. The urban population
has more graduates and fewer illiterates, while the opposite is true in rural areas.
Therefore, we would expect an increase in labor supply and as a consequence higher
FLFP in the urban areas. FLFP for married women between the age group of
25-55 there fell slightly from 18.5 % in 1987 to 17.9% in 2011, even as it declined
significantly in rural regions (Klasen and Pieters, 2015).
Basu and Desai (2012) and M. B. Das and Zumbyte (2017) find that
children in one-child families to be advantaged: they are more likely to be sent
to private schools and English medium schools, and more likely to receive private
75
tutoring in addition to schooling, than children from larger families. Yet women
in one-child families are less likely to be employed than those in larger families.
Following on this theme, M. Das and Desai (2003) argue that standard economic
factors such as job opportunities may have more to do with low FLFP in India,
than cultural factors.
4.2.2 Cultural values. To get a sense of the role of culture, we look
at the changing values and perceptions of Indian households. The World Values
Survey (WVS) provides comprehensive data on multi-dimensional values and belief
systems people hold true in different countries. We look at two specific variables
from this dataset to gauge how values may have shifted in past two decades.
The WVS asks relevant questions to the participants and gives them choices
to select their answer from. Figure 15 shows responses to the question whether or
not men should have more right to a job than a woman. In 2010 more people have
shifted from disagreeing with the statement to being neutral, that is, more people
became undecided whether or not men have more right to a job than women.
In contrast, in 1990 a significant number of people disagreed to the statement.
While this change may be due to shifting cultural attitudes against FLFP, it is
also possible that it reflects more competition in the labor market and limited job
availability. In such a scenario, holding cultural norm constant, men may prefer to
be chosen for the job rather than an equally qualified woman.
Another variable of interest in the WVS is how much importance households
place on their children’s education and specifically, on their daughter’s. Figure 16
report survey data on the question whether university education is more important
for a boy than a girl. Responses do not conclusively point in one direction. We see
that between 1995-2005 there is a decline in population which strongly disagrees
76
Figure 15. Changes in values if men should have more right to jobs
with the fact that boys deserve more education than girls and simultaneously an
increase in population which agrees with the statement.
4.2.3 Theoretical works. Two sets of theoretical works bear upon
the analysis in this chapter. One set studies household decision-making under
alternative assumptions of the decision-making process. These range from collective
models of the household (a special case of this, the unitary model, is widely used
in the macro literature and adopted below), to cooperative models where the
bargaining strength of the spouses determine allocations (for example, Heath and
Tan (2020)), to non-cooperative models where the conflict arises from preference
mis-match and the sequence in which decisions are made.3
3A popular framework that has been increasingly adopted is Lundberg and Pollak’s (1993)
“separate spheres” model where each spouse’s outside option involves them spending on their
private good and the household public good they care more about.
77
Figure 16. Changes in values if university is important for a boy or a girl
Our work has less to do with this body of theoretical work on household
decision-making and more to do with analyses on gender and development itself.
This literature is relatively sparse but a few papers are closely related to ours.
Hiller, 2014 proposes an OLG model where inegalitarian cultural norms affect
parental investment in children’s human capital: “strong norm” parents invest
less in daughters. During the early stages of development, two opposing forces are
at work. There is a positive income effect that raises educational spending on all
children, while a negative cultural effect delays parental investment in daughters’
human capital. The author shows that the interaction between endogenous
inegalitarian norm and the gender gap in education can generate multiple
equilibria. Initial conditions determine convergence to either an inegalitarian norms
development trap, or to an egalitarian norms affluent society. More relevant for our
purposes, along convergence path, FLFP follows an U-pattern.
78
In Chichilniski and Frederiksen’s (2009) model of the gender wage
gap, women’s productivity in market work decreases with an increase in time
devoted towards child-bearing activities. The model produces multiple equilibria:
expectation of women’s lower quality work is self-fulfilling as the lower offered wage
prevents women from engaging in market work, thereby keeping their productivity
low. The evolution of the gender pay gap is the focus of Galor and Weil’s (1996)
model. They assume that women specialize in mental labor, men in physical labor.
As capital accumulation rises with economic development, returns to women’s labor
rise faster because mental labor complements capital in the production process.
Rising women’s relative wages, in turn, reduces fertility by raising the cost of
children more than household income. Lower fertility raises capital per worker even
more. These effects are complemented by increasing education among women that
augments their productivity in mental labor; see also Kimura and Yasui (2010).
4.3 Household Decisions
Consider a static framework where an economy is populated by a continuum
of unitary households. Households are heterogeneous along two dimensions, human
capital and culture.
Each household consists of a husband (m) and wife (f) who pool their
resources and jointly maximize the household’s welfare. This involves choosing
household consumption, fertility, child investment and wife’s labor force
participation. Culture enters the model through resistance to the idea of the wife
working outside the household. This is parameterized as χ, the subjective cost
or social stigma experienced should the wife decide to work. Besides paid market
work, women have the option staying home to raise and care for children and/or
engaging in unpaid home production, distinct from childcare. Home production of
79
the non-market good is valued differently (imperfect substitutes) from consumption
of the market good c. The wife’s opportunity cost of raising children depends on
her immediate alternative: home production or market production.
Let the human capital of the wife be represented by hf and of husband by
hm. A household is characterized by the vector (hf , hm, χ). Denote by wf and wm
the wage rates per unit of human capital for female and male workers, respectively;
we will refer to these as efficiency wage rates. The joint decision-problem for the
household is
max ln c+ blh + γ [θ lnn+ (1− θ) ln q]− χI
s.t.
c = wmhm + wfhf lf − φen
lh + lf = 1− (τn + τ q)n
lh ≥ 0, lf ≥ 0
q = max {q , (e+ aτ qh )α} , α ∈ [0, 1]
 0 f
I = 0, if l = 0 f1, if lf > 0
It is assumed that θ > α(1 − θ) and b > 1. lf and lh are the wife’s time allocations
to market and home production. Home production requires time away from child
rearing, besides withholding labor supply from the market.4 We assume that only
the wife contributes to child rearing and education time costs. It can be shown that
the household will optimally choose to do so as long as wm > wf .
4This serves an important objective: if lf = 0, a mother’s opportunity cost of time is no longer
the market wage rate and we need her time to be used for something else to get the opportunity
cost effect.
80
The production of child quality q takes as input mother’s education time
(attention to children’s welfare) τ q, and market-provided educational goods
and services (school quality for example) e. These two inputs are perfectly
substitutable. This is a simplification: as long as the two inputs are substitutable
to some degree, child quality will be time-intensive at low levels of income
(specifically wf ), resource-intensive at higher levels of income.
5
We analyze household decisions sequentially. First we derive their behavior
based on whether or not the wife chooses to work, then we look at the female labor
supply decision on the extensive margin.
4.3.1 Case I: I = 0. For lf = 0, we can simplify the decision problem
to be
max ln c+ blh + γ [θ lnn+ α(1− θ) ln (e+ aτ qhf )]
s.t.
c = wmhm − φen
lh = 1− (τn + τ q)n ≥ 0.
from which follows the optimality conditions
φe γθ
+ b(τn + τ q) = (4.1)
c n
− ahfbn+ αγ(1− θ) ≤ 0, τ q ≥ 0 (4.2)
e+ aτ qhf
−φn 1+ αγ(1− θ) ≤ 0, e ≥ 0 (4.3)
c e+ aτ qhf
for n, τ q and e respectively. This leads to one of two outcomes depending on wages
and human capital of the spouses.
5Introducing q0 in the quality production functions nests the possibility of no investment in
child quality. This assumption is often used in fertility models to produce a larger fertility gap
between parents who do not invest and parents who do. In what follows, we will assume this is
not binding. It may become necessary for computational work later.
81
For wmhm ≤ aφhf/b, the household invests only time in child quality
production (e = 0) and
γ[θ − α(1− θ)]
n =
bτn
α(1− θ)τn
τ q =
θ − α(1− θ)
(4.4)
c =w[mhm ]
aα(1− θ)τn αhf
q =
θ − α(1− θ)
For wmhm > aφhf/b on the other hand, only resources go into child quality
production:
γ[θ − α(1− θ)]
n =
bτn
α(1− θ) bτn
e = w h
1 + αγ(1− m mθ) φ[θ − α(1− θ)]
(4.5)
wmhm
c = [1 + αγ(1− θ) ]α
α(1− θ) bτn
q = w h
θ − α(1− θ) φ[θ − m mα(1− θ)]
Since the wife does not work outside, only the husband’s labor income matters for
whether or not the household can afford resource investment in children.
4.3.2 Case II: I = 1. Under lh = 0, the decision problem
max ln c+ γ [θ lnn+ α(1− θ) ln (e+ aτ qhf )]− χ
s.t.
c = wmhm + wfhf lf − φen
lf = 1− (τn + τ q)n ≥ 0
82
leads to the FONCs
φe+ (τ q + τn)wfhf γθ
= (4.6)
c n
−wfhfn ahf+ αγ(1− θ) ≤ 0, τ q ≥ 0 (4.7)
c e+ aτ qhf
−φn 1+ αγ(1− θ) ≤ 0, e ≥ 0 (4.8)
c e+ aτ qhf
for n, τ q and e respectively. Similar to above, when wf ≤ aφ, only time is invested
in child quality and we have ( )
γ[θ − α(1− θ)] wmhm
n = 1 +
(1 + γθ)τn wfhf
α(1− θ)τn
τ q =
θ − α(1− θ)
(4.9)
w h + w h
c = [ m m f f1 + γθ ]
aα(1− θ)τn αhf
q = .
θ − α(1− θ)
For w qf > aφ, τ = 0 and we have ( )
γ[θ − α(1− θ)] wmhm
n = 1 +
(1 + γθ)τn wfhf
α(1− θ) τnwfhf
e =
1 + αγ(1− θ) φ
(4.10)
wmhm + wfh
c = [ f1 + γθ ]α
α(1− θ)
q = τnw
− f
hf
θ α(1− θ)
Here, the husband’s income does not matter for whether or not the household
invests resources in child quality. Because time investment in children is undertaken
only by mothers in this setup, the relevant tradeoff is whether or not the female
wage rate is high enough: for a low wage rate, the opportunity cost of mother’s
time is low and working wives are also in charge of child quality production.
4.3.3 Extensive margin decision: to work or not. Armed with
these optimality conditions, the decision on whether or not the wife works outsides
83
the household boils down to comparing the indirect utility functions under I =
0 and I = 1. Observe, however, that this comparison will depend on which of
the parametric conditions for child quality production hold hold. We analyze two
scenarios below.
The model predicts gender equality in education, that is, same educational
outcomes for boys and girls in a household. We assume on top of that positive
assortative matching on the marriage market so that hf = hm = h. This reduces
the dimensionality of the household’s characteristics vector to (h, χ). Also assume
that µ ≡ wf/wm ∈ (0, 1) which we call the gender wage ratio or wage gap.
Scenario 1: Case I.A versus Case II.A. Recall that Case I.A
applies when wmhm ≤ aφhf/b and Case II.A when wf ≤ aφ. We are, therefore,
considering the scenario where both apply, that is,
wf < wm < aφ
assuming b > 1 an(d hm)= hf[. Th(e wif)e decides to w(ork )as]long as
c1 n1 q1
ln + γ θ ln + (1− θ) ln ≥ χ+ bl0
c0 n0 q0 h
where variables with superscript 0 refer to decisions (4.4) and those with
superscript 1 to decisions (4.9) under hm = hf = h. On the RHS of the condition
above are the social stigma faced and the utility from home production given up
when the wife chooses to work. The first term on the LHS is the utility gain from
consumption when the wife works: clearly working outside brings in additional
income, and that gain is higher for a higher gender wage ratio:
c1 1 wm + wf 1 + µ
= = .
c0 1 + γθ wm 1 + γθ
The second and third terms on the LHS represent the tradeoffs regarding quantity
and quality of children from the wife working. On the one hand, more time to
84
the labor market detracts from time available to raising children which lowers the
household’s utility. Here the loss is higher, higher is the wage ratio:
n1 b 1 + µ
= .
n0 1 + γθ µ
On the other hand, the household may want to compensate for the utility loss from
fewer children by investing more intensively in them when the wife works. As it
turns out, because in both situations it is the mother who is investing her time in
children’s human capital formation, the net effect is a wash:
q1
= 1.
q0
Simplifying these expressions, we arrive a[t the de]cision that lf > 0 as long as
≡ 1 + µG(µ) ln(1 + µ) + γθ ln > Γ (4.11)
µ
where Γ ≡ χ + b + ln(1 + γθ)− γθ. The associated female labor supply function for
the household [ ]
− θ 1lf = 1 1 + (4.12)
1 + θ µ
is increasing in the female wage rate, decreasing in the male wage rate; the latter is
due to the usual opportunity cost margin.
Note from (4.11) that χ and b enter additively: cultural stigma is
indistinguishable from the cultural value placed on the enjoyment of home
production. Both discourage women’s labor force participation similarly. Two
points are worth making about this. First, this implies cultural resistance to FLFP
can be broader than just social stigma against wives and mothers working outside.
Secondly, in some societies the cultural value of household production can be as
strong as, if not stronger than, stigma. Take this example from Dasgupta (1997):
. . . the Pathan tribeswomen of of a particular village in the Northwest Frontier
Province (Pakistan) have traditionally walked some 8 km each way to a small
85
aqueduct to wash clothes once a week. In order to ease the plight of these women,
a foreign aid project built a traditional washing facility nearby. The project was a
failure. While the women admired it, they refused to use it. A female anthropologist,
employed to discover why the facility, remained unused, discovered that the
tribeswomen are rarely permitted to leave their homes. They spend most of their
lives within the confines of the mud castles owned by each Pathan family. These
weekly trips were their only chance to talk to the other women in the community,
when they could laugh, gossip, and play, The event lasted the whole day, and was
much looked forward to. The women also reasoned that if they were to complete the
laundry in less time their husbands would merely find more things for them to do.
Rising male wages can allow wives to more intensively specialize in household
production of children and goods. Even if the female wage rate is rising at the same
time, this margin may be unaffected for households that strongly value home life.
Returning to (4.11), observe that it does not depend on the household’s
human capital h, an outcome of the positive-assortative-matching assumption. As
such the household’s potential income, more precisely whether or not the household
is (potentially) rich or poor in terms of human capital (hm, hf ), has no bearing on
the female labor supply decision. What does produce heterogeneity of outcomes is
heterogeneity in χ: all else equal, higher χ (Γ) households withhold female labor
supply.
Conversely, given a χ, higher wage ratio (lower gender wage gap) has a
non-monotonic effect on the labor market participation decision as illustrated in
Figure 17. If µ < µL or µ > µH , then I = 1. In this range, lf increases in wf
from (4.12). The tradeoff between staying at home and working for a wage depends
on opportunity cost and productivity in home production, and working mothers’
involvement in non-market production depends on whether or not they invest time
in children’s education. To see this clearly, first note that the non-monotonicity
86
Figure 17. LFP decision at the household level
of G depends on the increasing function ln(1 + µ) and on the decreasing function
ln(1+1/µ). The former comes from the household’s relative utility gain from higher
household consumption when the wife works (clearly higher when wage ratio is
higher), the latter is the relative utility loss from restricting family size were the
wife to work. Starting from low values of µ, the latter dominates: because µ is low,
the household’s consumption gain is not particularly high, while the need to restrict
family size owing to the mother’s time commitments on the labor market is very
costly.
Scenario 2: Case I.B versus Case II.A. Suppose now
wm > aφ/b and wf < aφ
which requires a lower wage ratio than the previous case. A more complicated
tradeoff arises here since the wife would not be investing time into child quality
87
production if she w(ere)not to[ wor(k. A)s before, parti(cipa)tio]n is positive as long as
c1 n1 − q
1
ln + γ θ ln + (1 θ) ln ≥ χ+ bl0
c0 n0 q0 h
or, substituting in optimal decisions, w(hen )
≡ 1 + µG(µ) ln(1 + µ) + γθ ln ≥ Γ + κ(wm) ≡ Ω(wm) (4.13)
µ
where
κ(wm) ≡ αγ(1− θ) [lnwm − ln(aφ)] + (1 + γθ) ln(1 + γθ)
− [1 + αγ(1− θ)] ln [1 + αγ(1− θ)] + αγ(1− θ) ln b,
and the associated female labor supply function for the household is the same as
(4.12). By participating in the labor market, women face wf as the time cost of
child bearing: as wf goes up, it lowers fertility and lowers utility. In the previous
case the lower fertility was not compensated by an increase in q, now it is. But
the increase is less than the decrease because θ > α(1 − θ). Hence the gain in
consumption has to be high enough to offset the loss of utility from children. That,
in turn, requires µ = wf/wm to be high enough.
This illustrates another channel that affects FLFP besides the conventional
income channel and culture: the ability to substitute child quality for quantity.
Moreover, here this ability depends on φ: a higher cost of education lowers q0,
which encourages labor force participation because in doing so, the household
switches to mother’s time-intensive quality production which can raise child quality
significantly if wf is not high enough. In (4.13), Ω is decreasing in φ: all else equal,
the constraint is therefore more likely to be satisfied for higher φ.
What is especially interesting in Scenario 2 is the Ω is now endogenous
through its dependence on the male wage rate wm. This is because when the wife
does not work, it is the husband’s labor income that is relevant for financing child
88
quality investment. In choosing to work, if the female wage rate is not high enough,
the wife has to balance her labor time with child quality time: higher is wm, the
more does the wife need to devote time to produce a commensurate level of child
quality and, therefore, less time she has to work. Given a wm, changes in wf affect
the decision to supply labor similar to Figure 17 above: labor supply is positive for
µ < µ′ and µ > µ′L H where (µ
′ ′
L, µH) are suitably defined. Were wm to change too, it
would, all else equal, disincentivize female labor supply at the margin.
We leave it to future work to analyze this particular scenario. In light of the
evidence discussed in section 2, it may be especially relevant to understanding how
urban, relatively educated women respond to changing labor market conditions.
4.4 Aggregate Female Labor Force Participation
Returning to the case above, we next construct the aggregate female labor
supply decision. Let us assume that χ ∼ U [χL, χH ] with a mean value χ̄ ≡ (χH −
χL)/2. Given a wage ratio µ, as we saw in Figure 17, those households for whom
G(µ) ≥ Γ.
To understand how this works, consider Figure 18, that shows only the low-
and high-types as described by χL and χH respectively, corresponding to ΓL and
ΓH . For a wage ratio µ < µL or a wage ratio µ > µH , both high-cost and low-
cost households are willing to supply female labor. For intermediate values of the
wage ratio, only the low-cost household is. As a result, the extensive margin at the
aggregate level looks like the lower panel of the Figure.6
6There is an intensive margin response: below µL and above µH , and increase in wf holding
constant wm increases the hours supplied by a household.
89
Figure 18. Aggregate FLFP, χ ∈ {χL, χH}
90
It follows that with a continuum of household types on [χL, χH ], the
aggregate labor force participation rate ∫
`f = Iidχi
i
will be U-shaped with respect to the gender wage ratio as seen in the lower panel of
Figure 19. Of course, the aggregate labor supply is a function of both extensive and
Figure 19. Lf with continuum of types, effect of culture
intensive margins, the latter being an increasing function of µ as long as I = 0. The
figure is drawn under the assumption the extensive margin dominates.
91
It is important to note that the theory does not predict the aggregate
FLFPR is always non-monotonic. For example, were ΓH is below G(µ) in Fig 17,
then all households choose I = 1 and therefore, via (4.12), `f = 1. Or, if an
economy were to start from relatively high gender wage ratio, say higher than µl,
then an increase in the relative wage can only increase `f .
Moreover, while cultural resistance by itself does not change in response to
µ, it can explain a low level of FLFP rate. An increase in χ̄, for example, shifts
the distribution of Γi up, making previously indifferent households opt for no
participation and lowering the aggregate FLFP rate.
Based on the discussion from the previous section, we summarize the key
features of `f : Comparative statics with respect to b, φ, α
– An increase in χ and/or b lowers the net benefit of FLFP, increases µL,
decreases µH and lowers `f ;
– An increase in φ affects FLFP only if the household were spending resources
in child quality investment in the absence of FLFP. In this case, an increase
in φ, can increase `f ;
– An increase in the return to education, α, has ambiguous effects on FLFP;
– An increase in the gender wage ratio µ from below µL (low) to above will
produce a non-monotonic effect on `f ; conversely, `f will increase if the
economy starts with relatively high gender wage ratio.
4.5 Numerical Illustration
This section numerically illustrates the household model presented in the
previous sections. The illustration simulates a U-shaped FLFP with respec to
92
Parameter Parameter Explanation Parameter Values
b Utility from Home Production 0.6
θ Elasticity of Substitution (Child Quantity and Quality) 0.6
γ Utility from Children 1.4
χh Negative Utility from Female Market Production (Max.) 0.905
χl Negative Utility from Female Market Production (Min.) 0.897
φ Costs for child-rearing 0.5
α Child Quality Production Function 0.5
a Child Quality Production Function 10000
τn Time spent child-bearing 2
Table 9. Parameter Values for Simulation
economic development, where the latter is proxied by the female/wage ratio. An
increase in the wage ratio is indicative of better economic development. This
exercise does not calibrate the model instead presents the mechanisms at work
using arbitrary parameter values. Calibration, together with dynamic simulation,
would be the next step of the project. Relevant parameters are considered as per
the constraints defined by the model. For example, because lf ≥ 0 requires µ ≥ θ
from (4.12) we restrict the gender wage ratio to the interval [θ, 1]. Parameter values
are presented in Table 9.
Household Level We first present simulations at the intensive margin. In
other words, how households make female labor force participation decisions as
cultural costs change. The simulation results are presented in Table 20. The figure
plots the threshold condition arrived at in equation 4.11. Here, the relationship
between G(µ) the relative wage of women is plotted. The variation in thresholds
for labor force participation at the household level are driven by cultural costs χL
and χH depicted by the horizontal curves. To interpret this, take an example of the
highest value of cultural cost χH , households facing these costs will not participate
in market production for the relative wage rate between [0.77, 0.98].
Aggregate FLFP Extending the intensive margin decision to the aggregate
level, we consider that the cultural costs are heterogeneous across the population.
93
Figure 20. Numerical Simulation for LFP decision at the household level
Figure 21. Numerical Simulation for Aggregate FLFP, χ ∈ {χL, χH}
We assume that the population is distributed normally over the space of low and
high cultural costs. More specifically, we assume that the population is distributed
normally with the meanof χ̄ ≡ (χH − χL)/2 and standard deviation of χsd ≡
94
(χ̄ − χH)/3 with parameters of χH and χL presented in Table 9. The simulation
results are presented in Figure 21. As per the simulations, a relative wage increase
from 0.77 to 0.81 decreases aggregate female labor force participation by 75%.
4.6 Conclusion and future plan
The household decision problem outlined so far identifies three margins
that contribute to FLFP: consumption gain from both spouses working, women’s
ability to balance child care responsibilities with working outside the home,
and the cultural cost society imposes on working wives and mothers. If people
are heterogeneous in how they perceive this cost, then they will differ in their
willingness to supply labor given the same set of economic conditions. This simple
idea produces a naturally U-shaped response of the aggregate FLFP rate to female
wages (more precisely, to the gender wage ratio).
Future iterations of this work will pursue two related strategies. The first
strategy would be to introduce a role for human capital for the FLFP decision. It
is unclear, at this stage, how that can be done without breaking the assumption of
positive assortative matching. One possible option is to alter the nature of human
capital production. We are also interested in making the model dynamic. That
would require explicit assumptions about why the gender wage ratio is less than
one and why it increases with economic development. Galor and Weil (1997),
discussed in section 2, provide a roadmap for this using an aggregate production
function that relies on physical and mental labor. A more parsimonious approach is
to assume exogenous productivity growth rates of male and female labor that differ
in the short- and medium run but converge in the long run. This “reduced-form”
specification would be sufficient to understand the dynamics of FLFP, though not
the deeper causes of gender inequality.
95
A second strategy for future research is to study a more realistic labor
market choice. In deciding to work outside home in developing countries, women
face a choice between traditional market activities (agrarian, low-skill occupations)
and non-traditional ones (manufacturing, formal sector service). The former is
organized along family and community lines in which home and market activities
are unified (Goldin, 1990; Galor and Weil, 1997). This implies two things. First,
through tradition and upbringing, women are well-trained in these activities.
Secondly, they have an easier time balancing childcare with market production.
In comparison, jobs in non-traditional occupations require information about
participating in them, at least initially, and a clear separation of time devoted to
home versus market production. The former implies women start out in the new
economy with an informational disadvantage, the latter makes it harder for them to
give up time from childcare.
When an economy is experiencing rapid structural change, resources
gradually shift from the agricultural sector and labor-intensive cottage industries
to capital-intensive manufacturing and, then, skill-intensive services. These new
jobs offer a higher wage but the time-allocation and information problems can
dissuade many women from pursuing these jobs. Recent works, on US FLFP since
the nineteenth century, have identified information barriers as deciding factors.
Models such as Fernández (2013) and Fogli and Veldkamp (2011) offer feasible ways
these can be incorporated in a model of structural change.
96
CHAPTER V
CONCLUSION
This dissertation presented empirical evidence on the role played by
highway investments on India’s manufacturing industry and female labor force
participation (FLFP), and explored the broader economic reasons behind India’s
FLFP decline using an analytical model. My research shows that highways did
not have significant effects on the manufacturing sector, suggesting that there
are other structural reasons behind India’s slow manufacturing growth. Similarly,
highways did not significantly affect average district-level FLFP, but had significant
heterogeneous effects across districts conditional on initial market conditions. This
led to FLFP decreasing in some districts while increasing in others. Finally, an
analytical model of household decision-making shows that India’s FLFP decline
can be accounted for by gender specific norms that discourage FLFP, and mothers
choosing to invest in their children’s human capital over market production.
In Chapter 2, no significant evidence supports the idea that highways
lead to growth in the manufacturing output. On the contrary, the results show
that districts near the highways experienced a decline in their manufacturing
productivity. I speculate that India’s strict labor laws and complicated tax regime
may have restricted the positive gains from the highway project. These results are
different from what other studies have found for India but in line with the evidence
collected from China. Therefore, this chapter recommends that exploring specific
channels that connect transport infrastructure with manufacturing sector may help
consolidate contradictory results in the literature.
The third chapter explored the relationship of highways with FLFP in
India. The results are heterogeneous across districts, where districts with higher
97
literacy rates, lower levels of road connectivity and higher caste fractionalization
experienced a relative decrease in FLFP. These effects are predominantly
concentrated in the manufacturing and services sector. I study the districts that
experienced a decline and present suggestive evidence that this decline is driven
by married and skilled women. Future work will closely examine the trajectory of
FLFP growth in districts with low and high literacy rate districts separately. This
will help understand the heterogeneity across districts and propose mechanisms
explaining the variation.
The fourth chapter deconstructs the decline in FLFP by examining the
household decision-making process through the lens of an analytical model. A U-
shaped behavior of aggregate FLFP with respect to gender wage ratio is simulated.
Three margins are identified that contribute to FLFP: consumption gain from
both spouses working, women’s ability to balance child care responsibilities with
working outside the home, and the cultural cost society imposes on working wives
and mothers. We assume that people are heterogeneous in how they perceive these
costs, therefore, they differ in their willingness to supply labor given similar set of
economic conditions. This simple idea produces a naturally U-shaped response of
the aggregate FLFP rate to female wages (more precisely, to the ratio of female-to-
male wages).
The analyses in Chapters 3 and 4 suggest that when the economy is going
through the declining phase of the U-shaped relationship, simply increasing
opportunities for market production, for example due to better highway
infrastructure, may not be enough to increase FLFP. This is particularly true for
regions where returns to investing in children’s human capital are high. This may
also account for the evidence presented in Chapter 3, that high literacy regions
98
and more educated mothers experienced a bigger drop in FLFP due to the GQ
project.
99
APPENDIX A
CHAPTER II APPENDIX
A.1 Baseline Levels of Outcome
The chapter discusses the sensitivity of results with and without the
baseline level of manufacturing output as an additional control in Equation 2.1.
To demonstrate this, consider a simple setup without the instrument variables
and district-level characteristics as controls. This simple setup can be extended
to two versions, one with district fixed effects and the other with baseline level of
manufacturing output as a control variable. In a regression framework these two
scenarios can be presented as follows.
Yit = β1GQi + αi + λt + it (A.1)
Yit = α + β2GQi + γYt−h + λt + it (A.2)
here, Yit is the outcome variable with subscript i for district and t for the year.
GQi is a dummy for the treated districts, intermediate districts are not considered
for simplicity. The district fixed effects are represented by αi and the baseline
outcome variable is represented by Yt−h. λt represents year fixed effects, and it
is the residual term. Since we only have two periods, the baseline and the post
construction period, the the fixed-effects specification can also be represented as the
long differences. Therefore, Equation A.1 can also be written as,
′
Yit − Yit−h = α + β1GQi + it (A.3)
′
here, it = it − it−h. The above equation represents the main specification used
to estimate the results in this paper. However, adding a baseline outcome levels
100
to the above equation makes it algebraically equivalent to Equation A.2. This
can be seen by first adding the baseline outcome levels as an additional control
to Equation A.3,
Yit −
′ ′ ′ ′
Yit−h = α + β1GQi + γ Yit−h + it (A.4)
′
in the above specification β1 is algebraically equivalent to β2 in Equation A.2.
In conclusion, adding a baseline outcome level to the long difference equation
eliminates fixed effects and conditions the estimates on baseline levels instead.
Angrist and Pischke (2009) argue that these two alternate scenarios are not nested
in each other and are independent. Therefore, for empirical estimation we have to
choose between the two options, and there is no clear way to figure out which of
the specification is correct.
101
A.2 Tables and Figures
Table A.1. Summary Statistics
Nodal 0 - 40 Kms 40-120 kms > 120 kms
Total Output (Billion Rupees) 135.62 79.31 17.61 22.83
Average Productivity 3.74 3.35 3.47 3.55
Literacy Rate 0.65 0.56 0.50 0.54
(%) Rural Population 0.38 0.73 0.79 0.77
(%) SC/ST Population 0.17 0.24 0.27 0.25
Population Density (100 per sq.km) 68.96 8.64 4.75 5.54
Skill Intensity 0.11 0.10 0.10 0.08
Urban Share (Output) 0.77 0.29 0.30 0.29
Distance from Highway 19.27 18.21 80.13 249.53
Distance from Railways 6.67 10.35 16.45 20.71
Distance from Coast 338.88 314.25 413.86 487.87
Note: The summary statistics are calculated primarily from the ASI data. There
are a total of 441 districts with an unbalanced panel of 4885 district-years in total.
Districts are the sub-national administrative units after the states, geographically
bigger than villages. All districts are divided into separate distance bands based
on their proximity to the GQ project. Nodal districts represent Delhi, Mumbai,
Chennai, Kolkata, and other smaller districts areas closer to them. Geographical
distances from railways and coastline are computed from the data provided by MIT
Geodata and socio-economic characteristics are compiled using the 2001 Census
data.
102
Table A.2. First Stage Results - Straight Line IV
SLIV LCPIV
(1) (2) (3) (4) (5) (6)
(0− 40)km (0− 40)km (0− 40)km (0− 40)km (0− 40)km (0− 40)km
SLIV 0.621∗∗∗ 0.562∗∗∗ 0.484∗∗∗
(0.068) (0.071) (0.078)
LCPIV 0.694∗∗∗ 0.647∗∗∗ 0.582∗∗∗
(0.057) (0.064) (0.072)
Literacy Rate 0.003 -0.337 0.004 -0.215
(0.229) (0.256) (0.226) (0.249)
Rural Pop. -0.057 -0.031 -0.177 -0.118
(0.204) (0.203) (0.215) (0.216)
Pop. Density 0.144∗∗∗ 0.109∗∗ 0.117∗∗∗ 0.088∗
(0.042) (0.045) (0.044) (0.046)
SC/ST Population 0.281 0.238 0.046 0.031
(0.246) (0.229) (0.223) (0.221)
Distance to Rail Network -0.058∗∗ -0.059∗∗∗
(0.024) (0.023)
Distance to Coast Line -0.068∗∗ -0.048∗∗
(0.026) (0.024)
N 208 208 208 208 208 208
R sqr. .24 .29 .33 .31 .36 .38
Robust F-stat 84.67 62.54 38.8 148.5 101.15 65.27
Note: First stage results are presented. The dependent variables is a dummy which takes the value 1 if a
district is within 0-40 kms away from the highway and 0 otherwise. The districts falling only on the NSEW
corridor have been dropped and the districts 40-120 kms have also been dropped. Standard errors are
clustered at the district level and are in parentheses below point estimates.
* p < 0.1, ** p < 0.05, *** p < 0.01, Standard errors in parentheses.
103
Table A.3. Probability of Treatment
(1) (2) (3)
(0− 40)km (0− 40)km (0− 40)km
Output 0.192∗∗ 0.231∗∗ 0.194∗∗
(0.085) (0.099) (0.098)
Average Productivity (share-weighted) -0.262 -0.323 -0.307
(0.191) (0.213) (0.209)
Average Productivity -0.117 -0.217 -0.211
(0.214) (0.242) (0.261)
Skill Intensity 5.233∗∗ 5.847∗∗ 6.016∗∗
(2.282) (2.365) (2.413)
Capital to Labor Ratio -0.000 -0.000∗∗ -0.000∗
(0.000) (0.000) (0.000)
Urban Share 0.376 0.503 0.442
(0.492) (0.610) (0.632)
Literacy Rate 0.803 -1.344
(1.449) (1.647)
Rural Pop. 2.575∗∗ 2.278
(1.298) (1.426)
Pop. Density 0.741∗∗∗ 0.530∗∗
(0.212) (0.236)
SC/ST Population -0.171 -0.722
(1.362) (1.298)
Distance to Rail Network -0.328∗∗∗
(0.113)
Distance to Coast Line -0.332∗∗
(0.140)
N 264 264 264
Note: This table presents the coefficient estimates after predicting the
probability of highway treatment using district levvel manufacturing sector
characteristics, socio-economic characteristics and geographical factors. The
estimates are produced by running a logit regression as defined in Equation 2.3.
Column (1) presents results with manufacturing characteristics as explanatory
variables, Column (2) with socio-economic characteristics added, and finally
Column (3) with geographical factors added to the specification.
* p < 0.1, ** p < 0.05, *** p < 0.01, Standard errors in parentheses.
104
Table A.4. GQ construction reduced-form results
OLS SLIV LCPIV
No Controls Controls Controls No Controls Controls Controls No Controls Controls Controls
Nodal -0.414*** -0.083 0.282 -0.407*** -0.059 0.296 -0.433*** -0.141 0.200
(0.132) (0.208) (0.251) (0.128) (0.206) (0.252) (0.135) (0.202) (0.253)
Treated 0.068 0.139 0.374*** 0.174 0.234 0.425** -0.100 -0.052 0.144
(0.123) (0.130) (0.138) (0.167) (0.177) (0.172) (0.179) (0.197) (0.198)
Intermediate 0.177 0.213 0.241 0.136 0.176 0.357** 0.214 0.231* 0.401***
(0.181) (0.180) (0.161) (0.161) (0.175) (0.172) (0.138) (0.136) (0.129)
Baseline -0.249*** -0.252*** -0.247***
(0.043) (0.045) (0.044)
N 284 284 284 284 284 284 284 284 284
R sqr. 0.012 0.041 0.19 0.012 0.042 0.19 0.019 0.045 0.19
Note: This table presents the main results for the effects of the GQ project on economics growth. The coefficient estimates
are produced by running Equation 2.1 with change in logged levels of output between 2000 and 2007/2009 as the dependent
variable. Columns (1)-(3) present the OLS estimates and (5)-(6) present the IV estimates. The IV estimates do not include
the nodal districts. Standard errors are clustered at the district level.
* p < 0.1, ** p < 0.05, *** p < 0.01, Standard errors in parentheses.
105
APPENDIX B
CHAPTER III APPENDIX
B.1 Tables and Figures
106
Table B.1. Summary Statistics
Mean S.D. N
Population Statistics
Total Population 1045.87 755.37 841
Male Population 527.43 380.59 841
Female Population 518.44 378.75 841
Population (Age 15-30) 435.34 318.96 841
Population (Age 30-45) 334.84 248.88 841
Population (Age 45-65) 275.69 207.71 841
Employment Statistics
Total Employment 622.22 464.91 841
Male Employment 448.43 326.97 841
Female Employment 173.79 187.72 841
Agricultural Employment 430.23 352.85 839
Manufacturing Employment 92.20 94.18 829
Servies Employment 89.92 84.18 838
Employment (Age 15-30) 214.72 168.45 841
Employment (Age 30-45) 231.77 176.52 841
Employment (Age 45-65) 175.74 139.82 841
Labor Force Participation Rates
Male LFP 0.84 0.08 841
FeMale LFP 0.34 0.23 841
FLFP Agriculture 0.28 0.22 821
FLFP Manufacturing 0.04 0.05 689
FLFP Services 0.03 0.03 758
Distance Measures
Distance to the GQ project 241.06 216.01 841
Distance to the nearest Railway 34.47 91.64 841
Distance to the nearest Coastline 481.91 340.67 841
Note: The summary statistics are calculated primarily
from the NSS data. There are a total of 423 districts
with an unbalanced panel of 1236 district-years in total.
Districts are the sub-national administrative units
after the states, geographically bigger than villages.
Geographical distances from railways and coastline are
computed from the data provided by MIT Geodata, socio-
economic characteristics are compiled using the 2001
Census and the SHRUG data set.
107
Table B.2. Treatment Effects - FLFP with Literacy Interactions by Industry
Agriculture Manufacturing Services
(1) (2) (3) (4) (5) (6)
∆FLFP ∆FLFP ∆FLFP ∆FLFP ∆FLFP ∆FLFP
Nodal 0.309 0.039 -0.094∗∗∗ -0.109∗∗ 0.086 0.066
(0.241) (0.257) (0.027) (0.050) (0.071) (0.077)
Treated -0.152 -0.179 0.117∗∗∗ 0.116∗∗∗ 0.057∗∗∗ 0.057∗∗∗
(0.166) (0.168) (0.044) (0.043) (0.021) (0.021)
Intermediate -0.059 -0.030 0.040 0.029 0.020 0.026
(0.136) (0.136) (0.043) (0.047) (0.019) (0.020)
Nodal×Rural Literacy(2001) -0.473 0.050 0.187∗∗∗ 0.237∗∗ -0.137 -0.102
(0.417) (0.456) (0.047) (0.097) (0.135) (0.149)
Treated×Rural Literacy(2001) 0.252 0.359 -0.211∗∗∗ -0.201∗∗ -0.124∗∗∗ -0.126∗∗∗
(0.302) (0.310) (0.079) (0.081) (0.040) (0.040)
Intermediate×Rural Literacy(2001) 0.144 0.109 -0.069 -0.044 -0.045 -0.061
(0.279) (0.276) (0.076) (0.087) (0.039) (0.040)
District Controls No Yes No Yes No Yes
N 311 311 218 218 265 265
R sqr. 0.031 0.056 0.027 0.044 0.100 0.13
Note: This table presents the main results for the effects of the GQ project on female labor force
participation (FLFP) for each industrial sector separately. The treatment effects are interacted with
the literacy rate of districts in the year 2001. The coefficient estimates are produced by running
Equation 3.2 with change in FLFP between 1999 and 2009 as the dependent variable. Standard errors
are clustered at the district level.
* p < 0.1, ** p < 0.05, *** p < 0.01, Standard errors in parentheses.
108
Table B.3. Treatment Effects - Female Employment Levels by Industry
Agriculture Manufacturing Services
(1) (2) (3) (4) (5) (6)
∆FLFP ∆FLFP ∆FLFP ∆FLFP ∆FLFP ∆FLFP
Nodal 1.072 0.537 -3.799∗∗∗ -3.688∗∗∗ -1.117 -1.000
(1.101) (1.512) (0.862) (1.005) (6.363) (6.595)
Treated 0.012 -0.117 2.087∗∗∗ 2.093∗∗∗ 1.922∗∗ 1.997∗∗
(1.053) (1.050) (0.776) (0.784) (0.975) (0.966)
Intermediate -0.025 0.074 -0.501 -0.631 0.582 0.704
(0.667) (0.661) (1.213) (1.209) (0.726) (0.706)
Nodal×Rural Literacy(2001) -1.753 -0.740 7.972∗∗∗ 7.964∗∗∗ 4.645 4.115
(1.896) (2.668) (1.590) (1.915) (12.526) (13.017)
Treated×Rural Literacy(2001) -0.777 -0.289 -3.448∗∗ -3.418∗∗ -3.875∗ -4.279∗∗
(2.002) (2.012) (1.487) (1.553) (1.994) (1.981)
Intermediate×Rural Literacy(2001) -0.014 -0.020 0.908 1.181 -1.037 -1.523
(1.336) (1.300) (2.109) (2.157) (1.492) (1.479)
District Controls No Yes No Yes No Yes
N 311 311 218 218 265 265
R sqr. 0.034 0.053 0.044 0.049 0.095 0.12
Note: This table presents the main results for the effects of the GQ project on levels of female
employment for each industrial sector separately. The treatment effects are interacted with the literacy
of districts measured in the year 2001. The coefficient estimates are produced by running Equation 3.2
with change in female employment levels between 1999 and 2009 as the dependent variable. Standard
errors are clustered at the district level.
* p < 0.1, ** p < 0.05, *** p < 0.01, Standard errors in parentheses.
109
Table B.4. Baseline - Placebo Treatment Effects with Literacy Interactions
Agriculture Manufacturing Services
(1) (2) (3) (4) (5) (6)
∆FLFP ∆FLFP ∆FLFP ∆FLFP ∆FLFP ∆FLFP
Nodal 0.020 0.275∗ 0.024 -0.022 0.045 0.039
(0.057) (0.164) (0.024) (0.032) (0.059) (0.054)
Treated -0.043 -0.053 -0.029 -0.035 -0.011 -0.011
(0.091) (0.084) (0.031) (0.032) (0.010) (0.010)
Intermediate -0.058 -0.057 0.010 0.020 0.007 0.006
(0.084) (0.078) (0.026) (0.027) (0.012) (0.012)
Nodal×Total Literacy(1991) 0.033 -0.731∗ -0.037 0.097 -0.122 -0.101
(0.118) (0.399) (0.057) (0.095) (0.123) (0.113)
Treated×Total Literacy(1991) 0.069 -0.024 0.052 0.081 0.037 0.040
(0.219) (0.198) (0.076) (0.080) (0.024) (0.024)
Intermediate×Total Literacy(1991) 0.179 0.114 -0.032 -0.049 -0.030 -0.025
(0.225) (0.217) (0.059) (0.062) (0.030) (0.030)
District Controls No Yes No Yes No Yes
N 287 287 208 208 239 239
R sqr. 0.0043 0.058 0.0082 0.036 0.053 0.064
Note: This table presents results from placebo regressions run for the period 1987-1999 with the GQ
project as treatment, literacy rate in 1991 as the interaction term, and female labor force participation
rate as the outcome variable. The coefficient estimates are produced as per Equation 3.2 and standard
errors are clustered at the district level.
* p < 0.1, ** p < 0.05, *** p < 0.01, Standard errors in parentheses.
110
Table B.5. Average Treatment Effects - FLFP with Road Connectivity Interactions by Industry
Agriculture Manufacturing Services
(1) (2) (3) (4) (5) (6)
∆FLFP ∆FLFP ∆FLFP ∆FLFP ∆FLFP ∆FLFP
Nodal 0.001 -0.144 -0.067∗ -0.040 -0.039 -0.042
(0.216) (0.232) (0.039) (0.078) (0.063) (0.064)
Treated -0.143 -0.058 -0.018 -0.007 -0.048∗∗∗ -0.039∗∗
(0.101) (0.109) (0.036) (0.039) (0.015) (0.017)
Intermediate 0.173 0.128 0.109∗ 0.124∗∗ -0.010 -0.025
(0.141) (0.138) (0.058) (0.058) (0.021) (0.023)
Nodal×ln(Rural Paved Roads(2001)) 0.009 0.033 0.012∗∗ 0.011 0.008 0.009
(0.029) (0.034) (0.006) (0.012) (0.010) (0.010)
Treated×ln(Rural Paved Roads(2001)) 0.020 0.010 0.004 0.004 0.007∗∗∗ 0.005∗∗
(0.015) (0.017) (0.005) (0.006) (0.002) (0.002)
Intermediate×ln(Rural Paved Roads(2001)) -0.025 -0.016 -0.015∗ -0.017∗∗ 0.001 0.003
(0.021) (0.021) (0.009) (0.009) (0.003) (0.003)
District Controls No Yes No Yes No Yes
N 306 306 215 215 260 260
R sqr. 0.012 0.057 0.027 0.068 0.050 0.11
Note: This table presents the main results for the effects of the GQ project on female labor force
participation (FLFP) for each industrial sector separately. The treatment effects are interacted with a measure
that captures road connectivity in rural areas of districts in the year 2001. The coefficient estimates are
produced by running Equation 3.2 with change in FLFP between 1999 and 2009 as the dependent variable.
Standard errors are clustered at the district level.
* p < 0.1, ** p < 0.05, *** p < 0.01, Standard errors in parentheses.
111
Table B.6. Average Treatment Effects - FLFP with Caste Diversity Interactions by Industry
Agriculture Manufacturing Services
(1) (2) (3) (4) (5) (6)
∆FLFP ∆FLFP ∆FLFP ∆FLFP ∆FLFP ∆FLFP
Nodal -0.276∗ -0.164 0.073∗∗∗ 0.122∗∗∗ 0.076 0.088∗
(0.151) (0.203) (0.022) (0.044) (0.049) (0.050)
Treated 0.121 0.226 0.028 0.055 0.010 0.034
(0.226) (0.225) (0.048) (0.051) (0.024) (0.024)
Intermediate -0.145 -0.034 0.053 0.059 -0.018 0.006
(0.123) (0.124) (0.034) (0.037) (0.015) (0.016)
Nodal×Caste Index(2001) 0.543∗∗ 0.370 -0.103∗∗∗ -0.161∗∗ -0.104 -0.123
(0.212) (0.284) (0.037) (0.063) (0.083) (0.082)
Treated×Caste Index(2001) -0.225 -0.360 -0.036 -0.069 -0.029 -0.066∗
(0.365) (0.361) (0.078) (0.082) (0.038) (0.039)
Intermediate×Caste Index(2001) 0.246 0.093 -0.083 -0.087 0.022 -0.016
(0.201) (0.199) (0.056) (0.060) (0.025) (0.026)
District Controls No Yes No Yes No Yes
N 311 311 218 218 265 265
R sqr. 0.015 0.059 0.011 0.033 0.033 0.11
Note: This table presents the main results for the effects of the GQ project on female labor force
participation (FLFP) for each industrial sector separately. The treatment effects are interacted
with the caste fractionalization index constructed for the year 2001. The coefficient estimates are
produced by running Equation 3.2 with change in FLFP between 1999 and 2009 as the dependent
variable. Standard errors are clustered at the district level.
* p < 0.1, ** p < 0.05, *** p < 0.01, Standard errors in parentheses.
112
Table B.7. Treatment Effects - Education Wise Female Employment Levels
Primary Secondary Graduate
(1) (2) (3) (4) (5) (6)
∆FLFP ∆FLFP ∆FLFP ∆FLFP ∆FLFP ∆FLFP
Nodal 0.831∗ 0.801∗ -1.024∗∗∗ -0.964∗∗∗ 0.205 0.182
(0.426) (0.408) (0.202) (0.234) (0.329) (0.328)
Treated -0.035 -0.022 -0.344∗∗ -0.358∗∗ 0.504∗∗ 0.508∗∗
(0.201) (0.205) (0.163) (0.167) (0.213) (0.218)
Intermediate -0.268∗ -0.246 0.185 0.147 0.078 0.101
(0.161) (0.160) (0.152) (0.147) (0.160) (0.156)
Nodal×Rural Literacy(2001) -1.381∗ -1.407∗∗ 1.921∗∗∗ 1.891∗∗∗ -0.561 -0.534
(0.705) (0.688) (0.370) (0.433) (0.544) (0.559)
Treated×Rural Literacy(2001) 0.125 0.067 0.644∗∗ 0.704∗∗ -0.984∗∗ -1.003∗∗
(0.379) (0.390) (0.325) (0.337) (0.394) (0.411)
Intermediate×Rural Literacy(2001) 0.594∗ 0.520 -0.420 -0.312 -0.158 -0.212
(0.325) (0.327) (0.311) (0.304) (0.317) (0.306)
N 317 317 316 316 240 240
R sqr. 0.017 0.034 0.036 0.060 0.068 0.076
Note: This table presents the results for the effects of the GQ project on the employment share
of women belonging to different education categories. The treatment effects are interacted with the
literacy rate in districts measured for the year 2001. The coefficient estimates are produced by running
Equation 3.2 with change in employment share between 1999 and 2009 as the dependent variable.
Standard errors are clustered at the district level.
* p < 0.1, ** p < 0.05, *** p < 0.01, Standard errors in parentheses.
113
Table B.8. Treatment Effects - Female Employment Levels by Industry and Married Status
Agriculture Manufacturing Services
(1) (2) (3) (4) (5) (6)
∆FLFP ∆FLFP ∆FLFP ∆FLFP ∆FLFP ∆FLFP
Panel A: Married Women
Treated 14.967 5.813 63.565∗∗∗ 64.498∗∗∗ 30.568∗∗ 31.531∗∗
(93.339) (93.010) (22.965) (22.922) (13.170) (13.264)
Intermediate -5.001 1.609 -11.132 -15.157 11.120 13.182
(54.291) (54.112) (36.804) (36.502) (8.939) (8.929)
Treated×Rural Literacy(2001) -91.134 -54.714 -108.740∗∗ -109.991∗∗ -63.872∗∗ -68.458∗∗∗
(176.010) (176.787) (45.220) (46.224) (25.580) (25.854)
Intermediate×Rural Literacy(2001) 9.044 11.372 17.320 25.433 -22.912 -29.501
(108.242) (106.102) (64.609) (65.537) (18.272) (18.871)
Panel B: Un-Married Women
Treated -32.743 -35.088 -6.976 -6.667 -8.365 -9.442
(22.399) (22.873) (7.292) (7.693) (17.334) (16.932)
Intermediate -1.632 -1.720 6.884 1.378 -23.697 -27.609
(16.886) (16.372) (12.040) (13.454) (21.863) (21.614)
Treated×Rural Literacy(2001) 53.835 62.069 13.566 12.668 25.317 24.004
(46.292) (47.016) (14.859) (15.607) (35.104) (34.513)
Intermediate×Rural Literacy(2001) -4.744 -1.369 -9.409 1.078 45.602 50.872
(35.758) (33.906) (20.157) (24.020) (43.082) (42.518)
District Controls No Yes No Yes No Yes
N 201 201 76 76 92 92
R sqr. 0.046 0.063 0.18 0.21 0.20 0.24
Note: This table presents the results for the effects of the GQ project on female labor force participation
(FLFP) for married and unmarried women in different industrial sectors. The treatment effects are
interacted with the literacy rate in districts measured for the year 2001. The coefficient estimates are
produced by running Equation 3.2 with change in FLFP between 1999 and 2009 as the dependent variable.
Standard errors are clustered at the district level.
* p < 0.1, ** p < 0.05, *** p < 0.01, Standard errors in parentheses.
114
REFERENCES CITED
Ackerberg, Daniel A., Kevin Caves and Garth Frazer (2015). “Identification
Properties of Recent Production Function Estimators”. In: Econometrica 83.6,
pp. 2411–2451.
Afridi, Farzana, Monisankar Bishnu and Kanika Mahajan (2019). What Determines
Women’s Labor Supply? The Role of Home Productivity and Social Norms.
SSRN Scholarly Paper 3468613. Rochester, NY: Social Science Research
Network.
Afridi, Farzana, Taryn Dinkelman and Kanika Mahajan (2018). “Why Are Fewer
Married Women Joining the Work Force in Rural India? A Decomposition
Analysis over Two Decades”. In: Journal of Population Economics 31.3,
pp. 783–818. issn: 1432-1475.
Aggarwal, Shilpa (2018). “Do Rural Roads Create Pathways out of Poverty?
Evidence from India”. In: Journal of Development Economics 133,
pp. 375–395. issn: 03043878.
Alder, Simon (2016). “Chinese Roads in India: The Effect of Transport
Infrastructure on Economic Development”. In: SSRN Electronic Journal. issn:
1556-5068.
Amiti, Mary and Jozef Konings (2007). “Trade Liberalization, Intermediate Inputs,
and Productivity: Evidence from Indonesia”. In: The American Economic
Review 97.5, pp. 1611–1638. issn: 0002-8282.
115
Angrist, Joshua D. and Jörn-Steffen Pischke (2009). Mostly Harmless Econometrics:
An Empiricist’s Companion. Princeton University Press.
Asher, Sam, Tobias Lunt et al. (2021). “Development Research at High Geographic
Resolution: An Analysis of Night Lights, Firms, and Poverty in India Using
the SHRUG Open Data Platform”.
Asher, Sam and Paul Novosad (2020). “Rural Roads and Local Economic
Development”. In: American Economic Review 110.3, pp. 797–823. issn:
00028282.
Asturias, Jose, Manuel Garćıa-Santana and Roberto Ramos (2019). “Competition
and the Welfare Gains from Transportation Infrastructure: Evidence from
the Golden Quadrilateral of India”. In: Journal of the European Economic
Association 17.6, pp. 1881–1940. issn: 15424766.
Bailey, Martha J. and Brad J. Hershbein (2018). “U.S. Fertility Rates and
Childbearing, 1800 to 2010”. In: Cain, Louis P., Price V. Fishback and
Paul W. Rhode. Oxford Handbook of American Economic History. Vol. 1.
Oxford University Press. isbn: 978-0-19-093706-5.
Banerjee, Abhijit, Esther Duflo and Nancy Qian (2020). “On the Road: Access to
Transportation Infrastructure and Economic Growth in China”. In: Journal of
Development Economics 145, p. 102442. issn: 0304-3878.
Basu, Alaka M and Sonalde Desai (2012). Middle Class Dreams: India’s One-Child
Families, p. 40.
116
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.
issn: 0034-6535, 1530-9142.
Besley, T. and R. Burgess (2004). “Can Labor Regulation Hinder Economic
Performance? Evidence from India”. In: The Quarterly Journal of Economics
119.1, pp. 91–134. issn: 0033-5533, 1531-4650.
Besley, Timothy and Robin Burgess (2000). “Land Reform, Poverty Reduction, and
Growth: Evidence from India”. In: The Quarterly Journal of Economics 115.2,
pp. 389–430. issn: 0033-5533.
Chandra, Amitabh and Eric Thompson (2000). “Does Public Infrastructure Affect
Economic Activity?: Evidence from the Rural Interstate Highway System”. In:
Regional Science and Urban Economics 30.4, pp. 457–490. issn: 0166-0462.
Chatterjee, Santanu, Thomas Lebesmuehlbacher and Abhinav Narayanan (2021).
“How Productive Is Public Investment? Evidence from Formal and Informal
Production in India”. In: Journal of Development Economics 151, p. 102625.
issn: 0304-3878.
Chichilnisky, Graciela and Elisabeth Hermann Frederiksen (2008). “An Equilibrium
Analysis of the Gender Wage Gap”. In: International Labour Review 147.4,
pp. 297–320. issn: 1564-913X.
Das, Maitreyi and Sonalde Desai (2003). Why Are Educated Women Less Likely to
Be Employed in India? Testing Competing Hypotheses. 0313. World Bank.
117
Das, Maitreyi Bordia and Ieva Zumbyte (2017). The Motherhood Penalty and
Female Employment in Urban India. World Bank, Washington, DC.
Datta, Saugato (2012). “The Impact of Improved Highways on Indian Firms”. In:
Journal of Development Economics 99.1, pp. 46–57. issn: 0304-3878.
Dhar, Diva, Tarun Jain and Seema Jayachandran (2019). “Intergenerational
Transmission of Gender Attitudes: Evidence from India”. In: The Journal of
Development Studies 55.12, pp. 2572–2592. issn: 0022-0388.
Donaldson, Dave (2015). “The Gains from Market Integration”. In: Annual Review
of Economics 7.1, pp. 619–647. issn: 1941-1383, 1941-1391.
— (2018). “Railroads of the Raj: Estimating the Impact of Transportation
Infrastructure”. In: American Economic Review 108.4-5, pp. 899–934. issn:
0002-8282.
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. issn: 0033-5533, 1531-4650.
Duflo, Esther (2001). “Schooling and Labor Market Consequences of School
Construction in Indonesia: Evidence from an Unusual Policy Experiment”.
In: The American Economic Review 91.4, pp. 795–813. issn: 0002-8282.
Duranton, Gilles, Peter M. Morrow and Matthew A. Turner (2014). “Roads and
Trade: Evidence from the US”. In: The Review of Economic Studies 81.2,
pp. 681–724. issn: 0034-6527.
118
Faber, Benjamin (2014). “Trade Integration, Market Size, and Industrialization:
Evidence from China’s National Trunk Highway System”. In: The Review of
Economic Studies 81.3, pp. 1046–1070. issn: 0034-6527.
Fernández, Raquel (2013). “Cultural Change as Learning: The Evolution of Female
Labor Force Participation over a Century”. In: American Economic Review
103.1, pp. 472–500. issn: 0002-8282.
Fogli, Alessandra and Laura Veldkamp (2011). “Nature or Nurture? Learning and
the Geography of Female Labor Force Participation”. In: Econometrica 79.4,
pp. 1103–1138. issn: 1468-0262.
Galor, Oded and David N. Weil (1996). “The Gender Gap, Fertility, and Growth”.
In: The American Economic Review 86.3, pp. 374–387. issn: 0002-8282.
Ghani, Ejaz, Arti Grover Goswami and William R. Kerr (2016). “Highway to
Success: The Impact of the Golden Quadrilateral Project for the Location and
Performance of Indian Manufacturing”. In: The Economic Journal 126.591,
pp. 317–357. issn: 1468-0297.
Gibbons, Stephen et al. (2019). “New Road Infrastructure: The Effects on Firms”.
In: Journal of Urban Economics 110, pp. 35–50. issn: 0094-1190.
Goldin, Claudia (1990). Understanding the Gender Gap: An Economic History of
American Women. New York: Oxford University Press. isbn: 9780195050776.
119
Harrison, Ann E., Leslie A. Martin and Shanthi Nataraj (2013). “Learning
versus Stealing: How Important Are Market-Share Reallocations to India’s
Productivity Growth?” In: The World Bank Economic Review 27.2,
pp. 202–228. issn: 1564-698X, 0258-6770.
Heath, Rachel and Xu Tan (2020). “Intrahousehold Bargaining, Female Autonomy,
and Labor Supply: Theory and Evidence from India”. In: Journal of the
European Economic Association 18.4, pp. 1928–1968. issn: 1542-4766, 1542-
4774.
Hiller, Victor (2014). “Gender Inequality, Endogenous Cultural Norms, and
Economic Development”. In: The Scandinavian Journal of Economics 116.2,
pp. 455–481. issn: 0347-0520.
Holl, Adelheid (2016). “Highways and Productivity in Manufacturing Firms”. In:
Journal of Urban Economics 93, pp. 131–151. issn: 0094-1190.
Hsieh, Chang-Tai and Peter J Klenow (2009). “Misallocation and Manufacturing
TFP in China and India”. In: The Quarterly Journal of Economics 124.4,
pp. 1403–1448.
Jedwab, Remi and Alexander Moradi (2016). “The Permanent Effects of
Transportation Revolutions in Poor Countries: Evidence from Africa”. In:
Review of Economics and Statistics 98.2, pp. 268–284. issn: 0034-6535, 1530-
9142.
Khanna, Gaurav (2016). “Road Oft Taken: The Route to Spatial Development”.
120
Kimura, Masako and Daishin Yasui (2010). “The Galor–Weil Gender-Gap Model
Revisited: From Home to Market”. In: Journal of Economic Growth 15.4,
pp. 323–351. issn: 1381-4338, 1573-7020.
Klasen, Stephan and Janneke Pieters (2015). “What Explains the Stagnation
of Female Labor Force Participation in Urban India?” In: The World Bank
Economic Review 29.3, pp. 449–478. issn: 0258-6770.
Lei, Lei, Sonalde Desai and Reeve Vanneman (2019). “The Impact of
Transportation Infrastructure on Women’s Employment in India”. In: Feminist
Economics 25.4, pp. 94–125. issn: 1354-5701, 1466-4372.
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. issn: 0034-6527.
Lowe, Matt and Madeline McKelway (2019). “Bargaining Breakdown: Intra-
Household Decision-Making and Female Labor Supply”.
Lundberg, Shelly and Robert A. Pollak (1993). “Separate Spheres Bargaining and
the Marriage Market”. In: Journal of Political Economy 101.6, pp. 988–1010.
issn: 0022-3808.
Maddison, Angus (2001). Development Centre Studies The World Economy A
Millennial Perspective: A Millennial Perspective. OECD Publishing. 385 pp.
isbn: 978-92-64-18998-0.
121
Martin, Leslie A., Shanthi Nataraj and Ann E. Harrison (2017). “In with the Big,
Out with the Small: Removing Small-Scale Reservations in India”. In: The
American Economic Review 107.2, pp. 354–386. issn: 0002-8282.
Melecky, Martin, Siddharth Sharma and Hari Subhash (2018). Wider Economic
Benefits of Investments in Transport Corridors and the Role of Complementary
Policies. World Bank, Washington, DC.
Melitz, Marc J. (2003). “The Impact of Trade on Intra-Industry Reallocations and
Aggregate Industry Productivity”. In: Econometrica 71.6, pp. 1695–1725. issn:
1468-0262.
Michaels, Guy (2008). “The Effect of Trade on the Demand for Skill: Evidence from
the Interstate Highway System”. In: Review of Economics and Statistics 90.4,
pp. 683–701. issn: 0034-6535, 1530-9142.
Mitra, Devashish and Beyza P. Ural (2008). “Indian Manufacturing: A Slow Sector
in a Rapidly Growing Economy”. In: The Journal of International Trade &
Economic Development 17.4, pp. 525–559. issn: 0963-8199.
Munshi, Kaivan (2019). “Caste and the Indian Economy”. In: Journal of Economic
Literature 57.4, pp. 781–834. issn: 0022-0515.
Nataraj, Shanthi (2011). “The Impact of Trade Liberalization on Productivity:
Evidence from India’s Formal and Informal Manufacturing Sectors”. In:
Journal of International Economics 85.2, pp. 292–301. issn: 0022-1996.
122
Neff, Daniel, Kunal Sen and Veronika Kling (2012). The Puzzling Decline in Rural
Women’s Labor Force Participation in India: A Reexamination. GIGA Working
Paper 196. German Institute of Global and Area Studies.
Oh, Suanna (2021). Does Identity Affect Labor Supply? SSRN Scholarly Paper
3998025. Rochester, NY: Social Science Research Network.
Olley, G. Steven and Ariel Pakes (1996). “The Dynamics of Productivity in
the Telecommunications Equipment Industry”. In: Econometrica 64.6,
pp. 1263–1297. issn: 0012-9682.
Pande, Rohini and Charity Troyer Moore (2015). “Opinion — Why Aren’t India’s
Women Working?” In: The New York Times. Opinion. issn: 0362-4331.
Pavcnik, Nina (2002). “Trade Liberalization, Exit, and Productivity Improvements:
Evidence from Chilean Plants”. In: The Review of Economic Studies 69.1,
pp. 245–276. issn: 0034-6527.
Rao, M. Govinda (2005). “Tax System Reform in India: Achievements and
Challenges Ahead”. In: Journal of Asian Economics 16.6, pp. 993–1011. issn:
1049-0078.
Redding, Stephen J. and Matthew A. Turner (2015). “Chapter 20 - Transportation
Costs and the Spatial Organization of Economic Activity”. In: Handbook of
Regional and Urban Economics. Ed. by Gilles Duranton, J. Vernon Henderson
and William C. Strange. Vol. 5. Handbook of Regional and Urban Economics.
Elsevier, pp. 1339–1398.
123
Restuccia, Diego and Richard Rogerson (2017). “The Causes and Costs of
Misallocation”. In: Journal of Economic Perspectives 31.3, pp. 151–174. issn:
0895-3309.
Sarkar, Sudipa, Soham Sahoo and Stephan Klasen (2019). “Employment
Transitions of Women in India: A Panel Analysis”. In: World Development
115, pp. 291–309. issn: 0305-750X.
Shamdasani, Yogita (2021). “Rural Road Infrastructure & Agricultural Production:
Evidence from India”. In: Journal of Development Economics 152, p. 102686.
issn: 0304-3878.
Sivadasan, Jagadeesh (2009). “Barriers to Competition and Productivity: Evidence
from India”. In: The B.E. Journal of Economic Analysis & Policy 9.1. issn:
1935-1682.
Storeygard, Adam (2016). “Farther on down the Road: Transport Costs, Trade and
Urban Growth in Sub-Saharan Africa”. In: The Review of Economic Studies
83.3, pp. 1263–1295. issn: 0034-6527.
Syverson, Chad (2011). “What Determines Productivity?” In: Journal of Economic
Literature 49.2, pp. 326–365. issn: 0022-0515.
Topalova, Petia (2010). “Factor Immobility and Regional Impacts of Trade
Liberalization: Evidence on Poverty from India”. In: American Economic
Journal: Applied Economics 2.4, pp. 1–41. issn: 1945-7782.
Woetzel, Jonathan et al. (2016). “Bridging Global Infrastructure Gaps”. In:
McKinsey Global Institute 14.
124