ESSAYS IN DEVELOPEMENT ECONOMICS

By

Sambojyoti Biswas

A DISSERTATION

Michigan State University

in partial fulﬁllment of the requirements

Submitted to

for the degree of

Economics – Doctor of Philosophy

2019

ABSTRACT

ESSAYS IN DEVELOPEMENT ECONOMICS

By

Sambojyoti Biswas

In chapter one I use a large nationally representative data set from India to analyze the causal eﬀect
of age at marriage of a woman on her post marriage health and fertility outcomes. To look at
this eﬀect I propose a new instrumental variable in the sparse causal literature. The instrumental
variable strategy stems from two social practices within the Indian society; minimum age targeting
at marriage and seasonality of marriage dates. I ﬁnd that delaying marriage causally decreases the
probability of a women being diabetic and having elevated blood pressure post marriage, increases
her age at ﬁrst birth and has a zero eﬀect on the number of children she has. I also ﬁnd that delayed
marriage causally increases the women’s educational attainment, improves her bargaining power
within the household and improves her spousal quality. The paper also contrasts its ﬁndings with
the existing causal literature by producing results using both the new and the existing instrumental
variable in the literature. I ﬁnd that both the IV’s produce similar estimates in terms of health,
education and spousal quality but the estimates diﬀer for fertility outcomes.

In the absence of universal social security, parents in India depend heavily on their oﬀ springs for
post-retirement consumption. The patriarchal nature of the Indian society combined with low labor
market returns for women skews this dependency towards sons. This skewedness in dependency
for old age support, towards the son, might be a potential reason for gender gap in human capital
investments of the children. In chapter two I use the 2004 New Pension Scheme reform in India
as a quasi-experiment in a Diﬀerence in Diﬀerence in Diﬀerence framework to identify the causal
eﬀect of post retirement security on gender gap in human capital investment.
I ﬁnd that with
the decrease in post-retirement security for the parents the gender gap against the female child
increases. Compared to the male child, a girl child is less likely to be enrolled in a private school or
a school where the medium of instruction is English, both these eﬀects are statistically signiﬁcant.

I also ﬁnd that the gender gap in test scores is larger post reform for the children of parents whose
retirement security decreases, however these results are not statistically signiﬁcant.

Chapter three looks at the relationship between risk correlation in borrower outcomes and
micro ﬁnance tools such as monitoring and audit. To look at these relationships I introduce risk
correlation in existing theoretical models of moral hazard and costly state veriﬁcation.
I ﬁnd
that as risk correlation in borrower outcomes increases auditing by the lender increases whereas
the relation with monitoring is conditional on the probability of success of the borrower and the
monitor. I use the Townsend Thai data base to check whether the empirical ﬁndings are consistent
with the theoretical predictions. I ﬁnd that the relationship between risk correlation and monitoring
depends on the measure of risk correlation I use and for auditing I ﬁnd that the empirical ﬁndings
are opposite to the theoretical prediction.

This thesis is dedicated to all the women in developing countries, majority of whom are still

strugling for equality and autonomy.

iv

ACKNOWLEDGEMENTS

I would like to thank my parents Mrs.Chandra Biswas and Mr.Anup Kumar Biswas for their
continual support. I also wish to thank my comitte members for their guidnace and support.Thank
you Drs. Todd Elder, Christian Ahlin and Leah Lakdawala. Finally I would like to thank Michigan
State University for giving me an opprtunity to be a part of an ever inspiring and innovating
community.

v

TABLE OF CONTENTS

LIST OF TABLES .
LIST OF FIGURES .

.
.

.
.

.
.

.
.

.
.

.
.

.
.

.
.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

xi

.
.

.

.
.

.
.

. .

. .

. .

.

.

. .

.

.

.

.

.

.

.

1.1
1.2

1.5 Results .

1.3 Data .
1.4

First Stage
Exogenity of the Instrument

1.4.2.1
1.4.2.2
1.4.2.3 Mechanisms

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . .

CHAPTER 1 CAUSAL EFFECT OF AGE AT MARRIAGE ON HEALTH AND FERTILITY 1
Introduction . .
.
1
Instrumental Variable .
7
1.2.1 Minimum age targeting at marriage
8
1.2.2 Auspicious Marriage Dates . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.2.3 Constructing the IV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
Identiﬁcation Strategy .
1.4.1
Instrumental Variable Approach . . . . . . . . . . . . . . . . . . . . . . . 16
1.4.2 Validity of the Instrument . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
. . . . . . . . . . . . . . . . . . . . 17
. . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
. .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.5.1 Health outcomes
Fertility Outcomes
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.5.2
1.5.3
Intermediary Outcomes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.5.4 Discussion and Mechanism . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.5.4.1 Health . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.5.4.2
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.6 Comparison with the Menarche IV . . . . . . . . . . . . . . . . . . . . . . . . . . 25
First Stage .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
Intermediary Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. 27
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. 28
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
.
1.7.1 Month of Birth .
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.7.2 Measurement Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
1.7.3 Non Monotonicity of the IV . . . . . . . . . . . . . . . . . . . . . . . . . 31
Seasonality versus Targeting: Non-Hindu Sample . . . . . . . . . . . . . . 32
1.7.4
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

1.6.1
1.6.2
1.6.3 Health .
1.6.4

Fertility .
1.7 Robustness Checks

1.8 Conclusion . .

Fertility .

. .
.
.
.
.

. .
.

.

.

.

.

.

.

.

. .

CHAPTER 2 RETIREMENT SECURITY AND GENDER GAP IN PARENTAL IN-

Introduction .
2.1
2.2 Background .

VESTMENT .
.
.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. 69
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
Pension System in India . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

. .
.
.

.
.
.

.
.
.

.
.

2.2.1

vi

.

.
.

.
.

.
.

. .

. .

. .

2.5 Empirical Framework .

2.3 Literature Review .
2.4 Model
.

The Reform . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.4.1 Maximization Problem for Households with Children of Both Gender
. .
2.4.2 Maximization Problem for Households with Children of one Gender . . .

2.2.1.1
. 75
2.2.1.2 Comparison between the NPS and CSP . . . . . . . . . . . . . . 76
2.2.2 Unequal Investment in Sons and Daughters in India . . . . . . . . . . . . . 77
2.2.3
Private versus Public Schools in India . . . . . . . . . . . . . . . . . . . . 78
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
.
. 81
. 82
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
2.5.1 Main Identiﬁcation Strategy . . . . . . . . . . . . . . . . . . . . . . . . . 83
2.5.2 Common Trend and Falsiﬁcation . . . . . . . . . . . . . . . . . . . . . .
. 85
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
. 88
. 89
Inter-Household Vs Intra-Household . . . . . . . . . . . . . . . . . . . . . 90
. . . . . . . . . . . . . . . . . . . . . . . . 92
. . . . . . . . . . . . . . . . . . . . . . . . . . 92
. 92
. . . . . . . . . . . . . . . . . . . 92
2.9 Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93

.
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

.
.
2.7.1 Current Students
2.7.2 Test Scores
. .
2.7.3

2.8.1 Trends in Gender Attitude
2.8.2
2.8.3

Falsiﬁcation Test
Selection into Private and Public Sector

. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.6 Data description .
2.7 Results . .
. .

2.8 Common Trends and Falsiﬁcation Test

. .

.
.

.
.

.
.

.

.

.
.

.
.

.
.

. .

. .

. .

. .

. .

. .
.

Introduction . .

3.1
3.2 Model

3.2.1.1 Case I .
3.2.1.2 Case II

3.2.1 Moral Hazard; BBG . . . . . . . . . . . . . . . . . . . . . . . . . . . .

CHAPTER 3 EFFECT OF RISK CORRELATION ON MONITORING AND AUDITING 103
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
. 105
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
3.2.2 Costly state Veriﬁcation; Maitreesh Ghatak , Timothy W. Guinnane 2009 . 111
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
3.3 Data .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
3.4
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
3.5 Results . .
3.6 Comparing the theoretical results with the empirical ﬁndings . . . . . . . . . . . . 117
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
3.7 Conclusion . .
.
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
CHAPTER I : APPENDIX . . . . . . . . . . . . . . . . . . . . . . 136
CHAPTER II : APPENDIX . . . . . . . . . . . . . . . . . . . . . . 147
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160

APPENDICES .
APPENDIX A
APPENDIX B
BIBLIOGRAPHY .

.
Identiﬁcation Strategy .
.

. .
.
.

. .

. .

. .

.
.
.

.
.
.

.
.

.
.

.
.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

vii

LIST OF TABLES

Table 1.1: Descriptive Statistics .

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

Table 1.2: First Stage . .

.

.

.

.

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

Table 1.3: Eﬀect of IV1 on Early Childhood Outcomes . . . . . . . . . . . . . . . . . . . . 58

Table 1.4: Eﬀect of Age at Marriage on Wealth . . . . . . . . . . . . . . . . . . . . . . . . 59

Table 1.5: Eﬀect of Age at Marriage on Fertility . . . . . . . . . . . . . . . . . . . . . . . 60

Table 1.6: Eﬀect of Age at Marriage on Spouse and Spousal Family Quality . . . . . . . . . 61

Table 1.7: Eﬀect of Age at Marriage on Education and Bargaining Power

. . . . . . . . . . 62

Table 1.8: Eﬀect of Age at Marriage on Health . . . . . . . . . . . . . . . . . . . . . . . . 63

Table 1.9: Eﬀect of Age at Marriage on Fertility . . . . . . . . . . . . . . . . . . . . . . . 64

Table 1.10: First Stage Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

Table 1.11: Comparison : Intermediary Variables

. . . . . . . . . . . . . . . . . . . . . . . 66

Table 1.12: Comparison : Health and Fertility . . . . . . . . . . . . . . . . . . . . . . . . . 67

Table 1.13: Comparison : Robustness

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

Table 2.1: Pension System in India

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

Table 2.2: Descriptive Statistics for Current Students . . . . . . . . . . . . . . . . . . . . . 95

Table 2.3: Descriptive Statistics for the Test Score Sample . . . . . . . . . . . . . . . . . . 96

Table 2.4: Distribution of Industry Across the Public and Private Sector . . . . . . . . . . . 97

Table 2.5: Main Results .

.

.

.

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

Table 2.6:

Inter versus Intra-Household . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99

Table 2.7: Trends & Falsiﬁcation .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

Table 2.8: Selection into Sector

.

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

viii

Table 3.1: Descriptive Statistics .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121

Table 3.2: Descriptive Statistics .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122

Table 3.3: Leader Monitorig 1 RC .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123

Table 3.4: Leader Monitorig 2 OH . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124

Table 3.5: Leader Monitorig 1 RC .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125

Table 3.6: Leader Monitorig 2 OH . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126

Table 3.7: Peer Monitoring RC .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. 127

Table 3.8: Peer Monitoring OH .

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128

Table 3.9: Monthly Meetings RC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. 129

Table 3.10: Monthly Meetings OH .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

Table 3.11: Borrower Income RC .

Table 3.12: Borrower Income OH .

Table 3.13: Oﬃcer Visit RC .

.

Table 3.14: Oﬃcer Visit OH .

.

.

.

.

.

.

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. 134

Table A.1: First Stage Summary Stat (Appendix Table 1) . . . . . . . . . . . . . . . . . . . 140

Table A.2: Eﬀect of Age at Marriage on Spousal Characteristics (Appendix Table 2)

. . . . 141

Table A.3: Eﬀect of Age at Marriage on Reproductive Health (Appendix Table 3) . . . . .

. 142

Table A.4: Wald IV Estimates (Appendix Table 4)

. . . . . . . . . . . . . . . . . . . . . . 143

Table A.5: First Stage Estimates (Appendix Table 5)

. . . . . . . . . . . . . . . . . . . . . 144

Table A.6: Muslim Population (Appendix Table 6)

. . . . . . . . . . . . . . . . . . . . . . 145

Table A.7: Eﬀect of Age at Marriage on Fertility (Appendix Table 7) . . . . . . . . . . . . . 146

Table B.1: A1 Diﬀerence-in-Diﬀerence Results for Boy child and Girl child . . . . . . . . . 149

Table B.2: A2 Triple Diﬀerence Estimate with sibling ﬁxed eﬀect

. . . . . . . . . . . . . . 150

ix

Table B.3: A3 Private School Enrollment

. . . . . . . . . . . . . . . . . . . . . . . . . . . 151

Table B.4: A4 Robustness Check for Age . . . . . . . . . . . . . . . . . . . . . . . . . . . 152

Table B.5: A5 Alternate deﬁnition of Public Sector . . . . . . . . . . . . . . . . . . . . .

. 153

Table B.6: A5 Including Mother’s Sector

. . . . . . . . . . . . . . . . . . . . . . . . . .

. 154

Table B.7: A6 Means of the Outcome Variables . . . . . . . . . . . . . . . . . . . . . . . . 155

Table B.8: A7 Diﬀerential Eﬀect of the reform on other forms of Investment . . . . . . . . . 156

Table B.9: A8 Household Characteristics by Household Composition . . . . . . . . . . . . 157

Table B.10: A9 Private School Enrollment by Household Composition . . . . . . . . . . .

. 158

Table B.11: A10 Results on Anthropocentric Measures

. . . . . . . . . . . . . . . . . . . . 159

x

LIST OF FIGURES

Figure 1.1: Distribution of Age at Menarche and Age at marriage . . . . . . . . . . . . . . 35

Figure 1.2: Distribution of fraction of marriages by months

. . . . . . . . . . . . . . . .

. 36

Figure 1.3: Distribution of Age at Marriage in months . . . . . . . . . . . . . . . . . . .

. 37

Figure 1.4: Distribution of the variable Target . . . . . . . . . . . . . . . . . . . . . . . . . 38

Figure 1.5: Fraction of Marriages by distance from a whole number age . . . . . . . . . . . 39

Figure 1.6: Back of the Envelope Distribution of the Variable Target . . . . . . . . . . . . . 40

Figure 1.7: Wealth by Distance from a whole number marriage age . . . . . . . . . . . . . 41

Figure 1.8: Education by Distance from a whole number marriage age . . . . . . . . . . . . 42

Figure 1.9: Rural by Distance from a whole number marriage age . . . . . . . . . . . . .

. 43

Figure 1.10: Distribution of Target By Wealth . . . . . . . . . . . . . . . . . . . . . . . . . 44

Figure 1.11: Distribution of Target By Education . . . . . . . . . . . . . . . . . . . . . . .

. 45

Figure 1.12: Distribution of marriages by month over time . . . . . . . . . . . . . . . . . .

. 46

Figure 1.13: Distribution of marriages by month across religion . . . . . . . . . . . . . . . . 47

Figure 1.14: Age at Menarche by distance from a whole number age . . . . . . . . . . . . . 48

Figure 1.15: Distribution of Education by month of marriage . . . . . . . . . . . . . . . . . 49

Figure 1.16: Distribution of Rrural by month of marriage . . . . . . . . . . . . . . . . . . . 50

Figure 1.17: Distribution of Wealth by month of marriage . . . . . . . . . . . . . . . . . . . 51

Figure 1.18: Distribution of month of marriage by month of marriage . . . . . . . . . . . . . 52

Figure 1.19: Distribution of fraction of births by months for respondents . . . . . . . . . . . 53

Figure 1.20: Distribution of fraction of births by months for children age 0-4 . . . . . . . .

. 54

Figure 1.21: Distribution of Target by Religion . . . . . . . . . . . . . . . . . . . . . . . .

. 55

xi

Figure 2.1: Diﬀerence in Means .

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

xii

CHAPTER 1

CAUSAL EFFECT OF AGE AT MARRIAGE ON HEALTH AND FERTILITY

1.1 Introduction

In developing countries the practice of women getting married at a young age is still prevalent.
In India the legal age of marriage for women is eighteen but we see that almost 41 percent of
women are getting married before turning the age of eighteen and 82 percent of women are getting
married by the age of 21 (DHS VII, 2016). The numbers are also troubling for other developing
countries such as Bangladesh and Niger where 51 percent and 74 percent of the women are married
before turning the age of 18 (UNFPA, 2007). Women getting married at a young age is negatively
correlated with the women’s educational, health and marital outcomes. We also see negative
correlation in terms of their child’s health and educational outcomes (Bruce 2003, Clark 2004,
Nour 2006, Santhya et al. 2010). Early marriage also has negative impact on national development
by stunting educational and vocational opportunities for a large sector of the population (Raj et al.
2009). In this paper I try to estimate the causal eﬀect of age at marriage of a women on her post
marriage health and fertility outcomes.

There has been a lot of correlational studies looking at the adverse eﬀects of early marriage,
but the causal literature on age at marriage is sparse. Field and Ambrus, (2008) was one of the ﬁrst
papers to establish a causal relationship. They found that delayed marriage is causally associated
with higher educational attainment and literacy. They also found that delayed marriage is causally
associated with an increase in use of preventative health services. Sekhri and Debnath, (2014) and
Chari et al., (2017) look at the causal eﬀect of age at marriage of the mother on her children’s
wellbeing. They ﬁnd that delayed marriage is causally associated with better health and educational
outcomes for the children. Chari et al, (2017) also found that the gain in child outcomes due to
delayed marriage is partially driven by a decrease in mother’s fertility. A recent study by Dhamuja
and Roychowdhury, (2018) also ﬁnd that delayed marriage is causally associated with a decline in

1

physical violence against the women.

Introduced by Field and Ambrus, (2008), all the aforementioned causal papers uses age at
menarche as an instrumental variable for age at marriage to establish causality. They argue that
a signiﬁcant portion in the variation of age at menarche is random, making it a good instrument
for age at marriage. There are two issues in using age at menarche as an instrumental variable
to establish a causal relationship on heath and fertility in this paper. First, if we look at graph 1
which plots the distribution of age at menarche and age at marriage, we see that 98.7 percent of
the women have gone through menarche by the age of sixteen whereas only 30.2 percent of the
women are married by the age of sixteen. Since an instrumental variable strategy estimates LATE,
the complier group from using age at menarche as an IV is small and might not be representative
of the current population of women. A second and more pressing issue is that age at menarche
might not be an exogenous variation when we look at women’s health and fertility outcomes, and
hence fail the exogenity condition of an instrument. In the medical literature we see that both early
and late onset of menarche is often associated with adverse pregnancy outcomes such as ectopic
pregnancies and spontaneous abortions (Liestol, 1980; Martin et al., 1983; Wyshak, 1983; Sandler
et al., 1984). Later onset of menarche is also associated with a slightly higher risk of sub fecundity
and infertility (Gulbrandsen et al., 2014). Hence my biggest concern is that age at menarche could
aﬀect health and fertility outcomes through channels other than age at marriage, such as underlying
reproductive health. For example if late onset of menarche increases the risk of sub fecundity
and infertility, then that might result in lower number of children. At the same time late onset
of menarche delays marriage age, thus the causal eﬀect of delayed age at marriage on number of
children obtained from using age at menarche as an instrument might be biased.

In this paper I propose a new instrumental variable to look at the causal eﬀect of age at marriage
of a women on her post marriage health and fertility outcomes. The exogenous variation for
the instrument comes from two social practices within Indian society; Seasonality of marriages
and parents preference of getting their daughter married oﬀ early and after they turn a minimum
marriage eligible age. First, the seasonality of marriages partly stems from the deep rooted belief

2

in astrology by the Hindu population in India. When it comes to choosing marriage dates certain
dates in the Hindu calendar year are considered more auspicious for getting married compared to
others. If we look at graph 2 which plots the distribution of marriages by month we see that certain
months of the year on average have more marriages compared to others. Second, parents in India
on average prefer to marry oﬀ their girl child at an early age. The legal age for marriage in India is
eighteen for women. This creates a weak lower bound on how early parents can get their daughters
married oﬀ. If we look at graph 3 which pots the distribution of marriage age in months, we see a
spike in the number of marriages at the exact age of eighteen (216 months). From the data we see
these spikes also at whole number ages from ﬁfteen through twenty one. I call this phenomenon
minimum age targeting at marriage. Where parents prefer to get their daughters married oﬀ early
and after they turn a minimum marriage eligible age. The minimum age after which parents can
get their daughters married oﬀ should be eighteen due legal reasons but we see such targeting at
the aforementioned ages of ﬁfteen to twenty one also. This could be driven by localized social and
cultural norms about minimum marriage age of a woman. If the minimum target age is binding
then the women who turns the minimum target age in the month which has low number of marriage
dates i.e. women who are born in the month with low number of marriage dates, has to wait longer
in expectation to get married. If being born in a month with low number of marriages versus being
born in a month with high number of marriages is random, then that produces an exogenous shock
to their age at marriage. I exploit this variation using an instrumental variable strategy.

I use a nationally representative data set from India to estimate the results. First I look at
the overall causal eﬀect of age at marriage of a women on her post marriage health and fertility
outcomes. The eﬀect of age at marriage on health and fertility are possibly driven by other
intermediary variables such as educational attainment, spousal quality, spousal family quality and
bargaining power of the women within the marriage. From the previous literature (Chari et al.,
2017; Field and Ambrus, 2008; Shekhri and Debnath, 2014; Dhamuja and Roychowdhury, 2018)
we know that the intermediary variables itself are also aﬀected by age at marriage. So in the
next step I look at the eﬀect of age at marriage on these intermediary variables. Finally, I try to

3

understand the mechanisms and potential channels through which age at marriage aﬀects health
and fertility. Also to bench mark my ﬁndings to the previous causal literature, I re-estimate and
compare the results obtained from using both the instrumental variables (age at menarche IV and
my IV) separately. The prior is that both the IV’s should produce similar results for the intermediary
variables and the estimates might diﬀer when we look at health and fertility outcomes.

For post marriage health outcomes, I ﬁnd that one month delay in marriage reduces the probabil-
ity of the woman being diabetic (2.8 percent), being pre-diabetic (3.5 percent) and having elevated
blood pressure (0.8 percent). The results are statistically and economically signiﬁcant. Diabetes
and blood pressure are strongly driven by genetics but they are also aﬀected by lifestyle choices
such as lack of exercise, unhealthy meal plans and obesity (Nahas et al., 2009). I also look at the
eﬀect of age at marriage on anemia (0.3 percent increase), I do not ﬁnd a statistically signiﬁcant
eﬀect. For fertility outcomes, I ﬁnd that, one month delay in marriage has no eﬀect on the total
number of children (0.02 percent decrease), decreases the ideal number of children reported by
mother (0.09 percent), increases the age at ﬁrst birth (0.3 percent) and decreases birth gap by (0.05
percent). The eﬀect on age at ﬁrst birth is highly statistically signiﬁcant whereas the eﬀect on birth
gap and number of children is not.

For intermediary variables, I ﬁnd that one month delay in marriage increases the number of
years of education of the woman by (0.5 percent) and having at least primary level of education
(0.04 percent). The eﬀect on years of education is statistically and economically signiﬁcant. One
month delay in marriage increases husband’s years of education (0.4 percent), decreases husbands
age1 (0.05 percent), increases the wealth of the in laws household (0.07 percent) and decreases
the probability that the in laws household belongs to a rural area (0.9 percent). The eﬀects are
statistically signiﬁcant. One month delay in marriage decreases the probability of the husband
limiting the women to contact her natal (3.6 percent), decreases the probability of the husband
insist of knowing her whereabouts (3.3 percent) and decreases the number of control questions
1A large age gap between men and women at time of marriage would generally indicate that the

younger partner has less power and less say in the relationship (Carmichael, 2011).

4

answered yes2 (0.2 percent). The eﬀects are statistically signiﬁcant. Overall we see that women
who get married later have higher education, better spousal and spousal family quality, and higher
bargaining power in the marriage.

To systematically unpack the channels for the eﬀect of age at marriage on health and fertility I
look at three diﬀerent speciﬁcations. I look at the results controlling for spousal and spousal family
characteristics, controlling for spousal, spousal family and bargaining power and including spousal
and spousal family controls for the zero education population. For the health outcomes I see that
adding spousal controls and bargaining power controls does not change the estimates for the health
outcomes, for the zero education I ﬁnd that the gains are even larger this rules out the possibility that
the gains in health are completely driven by gains in education and spousal quality. To conclude, I
see that delaying marriage causally results in better health outcomes for the women and the gains
are not completely driven by gains in education, spousal quality and bargaining power.

Next, I unpack the eﬀect of delaying age at marriage on fertility. Previously we saw that delaying
marriage age had no eﬀect (0.02 percent decrease) on the total number of children but the eﬀect
was not statistically signiﬁcant. One month delay in marriage increases the number of children
by 0.009 (0.3 percent) when I include spousal and spousal family controls, by 0.008 (0.3 percent)
when I include spousal controls and bargaining controls, and by 0.016 (0.5 percent) for the zero
education sample with spousal controls. The eﬀects are highly statistically signiﬁcant but small in
magnitude. For age at ﬁrst birth and birth gap I see that adding spousal and bargaining controls do
not change the estimate. Overall I conclude that delayed age at marriage has zero or a very small
positive eﬀects on the number of children and it increases the age at ﬁrst birth.

When I compare the results of the eﬀect of age at marriage on the intermediary variables across
the two IV’s (age at menarche and my IV), I ﬁnd that the estimates are equal in direction and
magnitude. I also compare these results with estimates from the previous literature and ﬁnd that
they are equal in direction and mostly similar in magnitude. When I compare the results on health
outcomes across the two IV’s, I ﬁnd that the estimates are similar in direction and magnitude. My
2Lower number of control questions answered yes to implies greater bargaining power for the

women.

5

prior was that, both the IV’s should produce similar estimates for the intermediary variables but the
estimates on health and fertility might diﬀer. A possible explanation for why I see similar results
across the two IVs for health outcomes could be that the health factors which are inﬂuenced by early
or late onset of menarche might not aﬀect health outcomes such as diabetes, blood pressure and
anemia. The main divergence in results is when I look at the eﬀect of age at marriage on number
of children. My IV estimates a zero or small positive eﬀect on the number of children. Whereas
age at menarche IV estimates a negative, statistically signiﬁcant eﬀect on the number of children.
The previous literature which uses age at menarche as an IV also found negative and signiﬁcant
eﬀect. There are two plausible explanations for this, the ﬁrst one which has already been stated,
age at menarche is associated with reproductive health using it as an IV to measure total number of
children might downward bias the results. Another explanation could be that the compliers from
both the IV’s are diﬀerent and since an IV estimates LATE, the results are diﬀerent.

There are three major contributions of this paper, ﬁrst this is the ﬁrst paper in the literature to
look at the causal eﬀect of age at marriage on diabetes, blood pressure and anemia. Second, the
paper introduces a new instrumental variable strategy in the literature to look at the causal eﬀects
of age at marriage, the IV is less demanding in terms of data compared to age at menarche IV and
can easily be applied to other countries which has minimum marriage age laws and seasonality of
marriage dates. Finally, the results from the paper on intermediary variables such as education,
spousal quality and bargaining power, concretizes the ﬁndings from the previous causal literature
and further justiﬁes their use of age at menarche as an IV.

The rest of the paper is structured as follows, section 2 formally introduces the instrument,
section 3 discusses the data, section 4 discusses the empirical strategy, section 5 presents the
results, section 6 compares my IV with the age at menarche IV, section 7 presents robustness
checks and ﬁnally section 8 is conclusion.

6

1.2 Instrumental Variable

The instrumental variable strategy stems from two social phenomenon within the Hindu pop-
ulation in India, preference to get their daughter married oﬀ early and after turning a particular
minimum age, and preference to get married on auspicious dates. Parents in India prefer to get their
daughter married oﬀ at an early age but how early they can get their daughter married oﬀ depends on
laws and social norms. I call the preference of getting your daughter married oﬀ early and after they
turn a minimum marriage eligible age "minimum age targeting at marriage". The girls who turn
these minimum marriage eligible age in a month of low auspicious marriage dates in expectation
has to wait longer to get married compared to girls who turn the minimum marriage eligible age
in a month of high auspicious marriage dates. For example imagine a girl born in January (an
unpopular month for getting married), 1980. If there were no preference for getting married on
auspicious marriage dates and no minimum age targeting (parents do not have a preference to get
their daughter married oﬀ early and after turning a particular minimum age), then she would get
married in any month/year. In presence of only minimum age targeting (say at the legal age of 18)
she would get married in January, 1998. If there is preference for getting married on auspicious
marriage month and minimum age targeting then she would likely get married in June (popular
month for marriage), 1998 at the age of 18 years and 6 months. Compare this to a girl born in
June, 1980. In absence of minimum age targeting and preference for getting married on auspicious
marriage dates she could get married whenever. In presence of minimum marriage age targeting
(again at the legal age of 18) she would get married in June, 1988. In presence of age targeting and
preference for getting married on auspicious dates she would still get married in June,1980 at the
age of 18. Hence the girl who was born in January in expectation has to wait longer to get married.
If being born in a popular versus unpopular marriage month is random then we get an exogenous
variation for age at marriage.

Before constructing the IV we look at the practice of minimum age targeting and preference to
get married on auspicious dates. This paper is the ﬁrst to introduce the phenomenon of minimum
age targeting in the literature. I present suggestive empirical evidence to establish it. The second

7

phenomenon comes from Hindu religious beliefs and astrology where people prefer to get married
on certain auspicious dates.

1.2.1 Minimum age targeting at marriage

Parents in Indian tend to marry oﬀ their daughters at an early age but how early the daughter
is married oﬀ depends on various social, cultural and economic reasons. One candidate for the
minimum age which parents can target their daughters to get married after is the minimum legal
marriage age of eighteen. Graph 3 plots the age at marriage in months. We see a spike at 216
months (Exactly 18 years) i.e. there is a jump in the number of marriages at the exact age of 18
compared to getting married at 17 years 11 months or 18 year 1 month. This is suggestive that
parents are getting their daughters married oﬀ as soon as they turn eighteen. In India 41 percent
of women get married before turning the age of eighteen. Thus for a majority of the population
the legal minimum age constrain is not binding. When we look at graph 3 we see spikes similar
to 216 months at 180, 192, 204, 228 and 240 months. I cite this as evidence for parents targeting
ages apart from the minimum legal age as a minimum age for their daughters to get married after.
I call this phenomenon minimum age targeting at marriage. This phenomenon is comprised of two
social preference i) Parents want their daughter to get married early ii) Parents want their daughter
to get married after turning a minimum marriage eligible3 age. In the existing literature there is
plenty of evidence of the ﬁrst preference. In this section I will present some evidence that both of
them holds together.

Let us deﬁne the variable Target = mod (age at marriage in months, 12). Graph 4 shows the
density of the variable target for our population. The maximum frequency for target is at the value
zero implying maximum number of girls are getting married in the same month as their birth month
i.e. as soon as they turn a particular minimum marriage eligible whole number age. Then we
see a monotonic decrease in density as the variable target increases, reaching a minimum at six4,
3The age itself could be driven by social and cultural norm, in certain communities it might be

frowned upon to get your daughter married of before turning sixteen

4This could be because of both targeting and the seasonality of birth and marriage months, in

8

implying least number of girls are waiting six months to get married after turning a particular whole
number age. We then see the variable monotonically rising towards zero. In Graph 5 I plot the
average number of marriages by the absolute distance in months from a whole number age. We
see that as the distance from the whole number ages decreases the average number of marriage
decreases.

In Appendix Section I, I do a back of the envelope calculation for the variable target. Given the
population distribution of marriages by month and the population distribution of births by month
I calculate the distribution of the variable target, under the assumption that month of birth is not
correlated to the month of marriage i.e. no age targeting. Graph 6 plots the actual (in presence of
targeting) and calculated (absence of targeting) distribution of the variable target. We see that at the
value zero there is gap between the two distributions implying that if women were not coordinating
their month of marriage with their month of birth then less number of women would get married
in their birth month. This gap is indicative of minimum age targeting at marriage.

In graphs 7-9 I plot the mean values of household characteristics by the distance from a whole
number age. We see that women who gets married closer to a whole number age are on average less
educated, gets married into families belonging to rural areas, and gets married into families with
lower wealth. Thus families which practice minimum age targeting usually belongs to the lower
socio-economic strata of the society. In graphs 10-11 I plot the distribution of the variable target
by wealth and education. We see that the women who gets married into the top wealth quintile
families do not practice minimum age targeting at marriage whereas women who gets married into
the bottom quintile practice age targeting at marriage. We see similar trends for education and
other women and family characteristics. This gives us an idea about the compliers for the IV. The
variables looked at in the above graphs are outcomes rather than covariates as they are determined
after marriage. In graph 14 in I plot the age at menarche5 by the distance from a whole number age
appendix Section I do a back of the envelope calculation for the variable target there also we see a
similar shape for the variable target at 6.

5For women who get married after menarche.

9

we see that the average age at menarche does not vary by the variable target6 , this is indicative of
the fact that targeting aﬀects women outcomes through marriage.7

1.2.2 Auspicious Marriage Dates

Hinduism is one of the most prominent religion in the world, it is the most dominant religion in
India. Indian people belonging to the Hindu religion have a deep rooted believe in astrology. They
depend on astrology to make many life decisions. So when it comes to setting a marriage date
Hindus normally consult a horoscope interpreted by the pundits (Narlikar 2013, Lal 2015). Now
when it comes to marriage dates certain months of the year are considered more auspicious than
others this results in the auspicious months getting more marriage dates which in turn translates into
more marriages occurring in certain months of the year than others. Graph 2 shows the distribution
of number of marriages by month. Graphs 12 and 13 show the distribution for other religions
and over years. We see that the distributions looks diﬀerent by religion but similar over the years.
Graph 18 plots the distribution of month of marriage by month of birth, we see that the distribution
looks similar across month of birth.

The distribution of marriage months might not solely be driven by auspicious dates it could
also be due to other reason such as overlap with the festive season which is usually post-harvest,
climate, availability of wedding services etc. Thus in equilibrium some we see some periods of
the year are more favorable for getting married compared to other hence we see a seasonality in
marriage.

To understand who gets married on these auspicious dates I plot the mean values of household
and woman characteristics by month of marriage, graph 15-17 presents these. We see that similar
to the population that practice age targeting people who gets married in the auspicious months are
on average less educated, gets married into families belonging to rural areas, and gets married into
6Menarche is an outcome realized by the women prior to their marriage and hence should not

vary by whether their families practice minimum age targeting at marriage.

7Ideally I would like to show more pre-marriage outcomes but unfortunately the data doesn’t

allow that.

10

families with lower wealth. Since the IV is relevant to the population that practice age targeting
and prefer getting married on auspicious dates this gives us an idea about the compliers for the IV.

1.2.3 Constructing the IV

To get an exogenous variation for age at marriage I make use of the practice of minimum age
targeting at marriage and seasonality of marriage dates. From the data we see that a lot of women
are getting married as soon as they turn a particular whole number age. Now if the women is
born in a period of low number of auspicious marriage dates then she turns the minimum marriage
eligible whole number age for getting married in that period of the year which has low number of
auspicious marriage dates. So in expectation she has to wait longer to ﬁnd a suitable auspicious
date to get married. Whereas if the individual is born in a period of high number of auspicious
marriage dates then the individual turns the minimum marriage age in that period of the year which
has high number of marriage dates. Then she has to wait less to ﬁnd a suitable auspicious marriage
date. Since the distribution of auspicious dates comes from astrology, I assume that is has no real
eﬀect on marriage outcomes. Thus if being born in a low marriage frequency period versus high
marriage frequency period is random8 then in presence of the two social phenomenon minimum
age targeting at marriage and the desire to get married at an auspicious date; we get an exogenous
shock to age of marriage. Where women born in low marriage frequency period waits longer in
expectation to get married.

I construct two IVs, IV1 and IV2. To capture the aforementioned eﬀect, the IVs should not
only take into account whether the woman was born in a month of low marriage frequency or not,
it should also take into account whether the period following her birth month has a high number
of marriages or not. For example the month of January and July have similar amount of marriages
but January is followed by months which have high marriage frequency and July is followed by
months with low marriage frequency. Hence someone born in January might have to wait less in

8I present evidence for this in section 4.2.

11

expectation9 To capture this I create a relative score for each birth months (IV2). The score should
capture two things, ﬁrstly months which have high number of marriage dates should have higher
relative score and secondly the months which are followed by months with high number of marriage
dates should also have higher relative score. I capture these two factors using the following equation

IV2=Relative Frequency of Birth Month + Relative Frequency of the Period after Birth Month

IV2it = Mit + (β1m1it + β2m2it + β3m3it + ....)

(1.1)

Let t be the whole number age at which the individual i gets married then Mit is the relative
frequency of marriages of the month she was born in, this term captures the ﬁrst factor that months
with high marriage number of marriage dates have higher relative score. Let m1it be the relative
frequency of the month after her birth month, m2it is the relative frequency of the month two months
after birth month and so on. Beta is a discounting factor with a value less than one. The power
of beta increases as the distance from the birth month increases implying lower wait on months
which are further away. The term (β1m1it + β2m2it + β3m3it + ....) captures the second factor that
months immediately followed by period of high number of marriages have a higher relative score.
Thus IV2 gives us a relative score for each birth month which captures how favorable the month is
for getting married. In Appendix Section II I give an example of IV2 for various values of beta.

I also construct a binary IV, IV1. IV1 is a binary variable which is equal to 1 if you are born
in a month which belongs to a period where there is low number of marriages and 0 otherwise.
IV1 takes the value 1 for the months (7,8,9,10) and 0 for the rest. These months were chosen as
they have the lowest relative frequency according to the ranking generated by the relative score
computed in equation 110.

9This rules out using some common instruments such as quarter of birth or quartile of birth by

marriage frequency.

10See Appendix section II for example

12

1.3 Data

The data used for the analysis comes from the 2015-16 National Family Health Survey (NFHS-
4). This is the fourth data set in the NFHS series11. The data set provides information on
population, health and nutrition for India and each state and union territory. The survey has four
diﬀerent questionnaire – woman’s, man’s, biomarker and household. All women age 15-49 and
men age 15-54 in the selected sample of households were eligible for individual interviews. The
household questionnaire covers basic information on all household members along with information
on socioeconomic, sanitation, water, health and deaths in past three years. For the women’s inter-
view two sets of questionnaire were used. The ﬁrst version (district module) collected information
on women’s basic characteristics, marriage, fertility, reproductive health, children characteristics.
This ﬁrst version was collected on all women in the household survey. The second version (state
module) was collected for sub sample of the households. It has additional four topics on sexual be-
havior, HIV/AIDS, husband’s background and woman’s work, and domestic violence. The man’s
questionnaire covers information on man’s characteristics, life style, sexual behavior, marriage,
fertility, nutrition, attitude towards gender roles, contraception and HIV/AIDS. The biomarker
questionnaire covered information on height, weight and hemoglobin levels for children; height,
weight, hemoglobin levels, blood pressure and random glucose for women and men.

For my analyses I restrict the sample to Hindu women who gets married before the age of 22,
this is 87 percent of the population. The minimum age targeting at marriage and auspicious dates
phenomenon is most relevant for this group, I call this the “full sample”12. I also analyze the results
on a sub-sample of women with zero years of education. I call this the “zero education sample”.
The purpose is mainly to understand that if we block the channel of education gains by delayed
marriage, do we still see an eﬀect13. The Age at Menarche question was asked to a sub sample
11I use this round as it is the latest round in the series with higher number of observation for my

12For women getting married after the age of 21 it is unlikely that her parents were waiting for

population

some minimum marriage eligible age.

13More on this in Section 3.3

13

of the women. To compare the results from my IV and Age at Menarche IV I use the sample of
women who were asked about their age at menarche. I call this the “Menarche Sample”. In all
the samples I drop the women for whom the information on month/year of birth or month/year of
marriage was imputed (18 percent of the population was dropped).

Table 1 provides the descriptive statics for all three samples. The uneducated sample as expected
performs worse than the full sample on dimensions of wealth, education, family size among others.
The menarche sample is much younger than the full sample, this is because the age at menarche
question was asked to women age less than 25.

The main outcomes includes measures of health and fertility. The intermediary outcomes
includes measures of educational attainment, spousal quality, spousal family quality and bargaining
power.

1. Health: It includes measures of diabetes, blood pressure and anemia. For diabetes I look at
two outcomes. a) A binary variable indicating blood glucose level greater than 200mg/dl, when
the blood sample is collected randomly. This indicates whether the person is diabetic. b) A binary
variable indicating blood glucose level greater than 100 mg/dl, when the blood sample is collected
at least after 8 hours of fasting. This indicates whether the person is pre diabetic. For high blood
pressure I look at two outcomes. c) A binary variable indicating systolic blood pressure level
greater than 120. This indicates high blood pressure. d) A binary variable indicating diastolic
blood pressure level greater than 80. This indicates high blood pressure. For anemia I look at
two outcomes. e) A binary variable indicating a hemoglobin level (g/dl) less than 6.9, indicating
severe anemia. b) A binary variable indicating a hemoglobin level (g/dl) less than 9.9, indicating
moderate or severe anemia.

2. Fertility:

It includes four measures a) Number of children. b) Age at ﬁrst birth.

c)

Desired/Ideal number of children reported by the ever married woman d) Birth Spacing14.

3. Women’s Educational attainment: It includes two measures of women’s education a) highest
grade attained by the woman. b) A binary indicator for whether the woman has at least primary

14Number of months between the birth of ﬁrst and second child.

14

level of education.

4. Husband and Husband’s Family Characteristics: 15 : It includes four measures a) Highest
grade attained by husband. b) Husband’s age in years. c) An index variable indicating wealth of
husband’s family. One being the poorest and ﬁve being the richest. d) A binary variable indicating
whether the family belongs to a rural or urban area.

5. Proxies for Bargaining Power: These are reported by an ever married women to whom the
domestic violence questionnaire was given. The questionnaire included questions to access control
issues within marriage. There are several control issue questions in the data, I report the results
from a few. a) A binary variable indicating whether the husband limits the woman from contacting
natal family b) A binary variable indicating whether the husband insists on know where she is. c)
Number of control issue questions answered yes.16

1.4 Identiﬁcation Strategy

I want to ﬁnd the causal eﬀect of age at marriage for a woman on her health and fertility. The
main empirical challenge is that age at marriage might be endogenous due to omitted variables.
Age at marriage might be correlated with unobserved family and woman characteristics which will
bias the results from simple correlational studies. To overcome this challenge I use an Instrumental
variable approach. I use functions of month of birth as an IV for age at marriage. The instrumental
variable approach is detailed in section 2. The exogenity assumption of this IV approach is that
women being born in low marriage frequency months are not systematically diﬀerent from women
who are born in high marriage frequency months.

15Ideally I should at other husband outcomes such as husband’s income and occupation. The
NFHS-4 data has a men’s questionnaire but it does not report income. It reports occupation but
the way it is reported makes it diﬃcult to create a relative ranking over the various occupations
reported, hence I do not include them

16The higher number of control issue questions answered yes indicates lower bargaining power

of the women in the marriage.

15

1.4.1

Instrumental Variable Approach

I consider the following regression model where Yi denotes outcome for women i.

Yi = βAgeo f Marriagei + γWomenControlsi + i

(1.2)

Women controls include age and month of marriage FE17. I report robust standard errors.
Now the major concern with estimating the above equation is that age of marriage might be
endogenous. To overcome this I use IV1 and IV2 as my instrument for age at marriage. The
motivation for the IVs are discussed in section 2.

1.4.2 Validity of the Instrument

1.4.2.1 First Stage

First I examine whether the IVs predict the age at marriage. Table 2 reports the ﬁrst stage estimates.
I control for women’s age and include month of marriage FE. We see that being born in a low
frequency month delays age at marriage by 3 months in the full sample and by 3.4 months in the
zero education sample. One unit increase in IV2 delays age at marriage by 0.15 months in the
full sample and 0.16 months in the zero education sample. All the estimates are statistically and
economically signiﬁcant.

The ﬁrst stage results using both the IV’s are highly signiﬁcant, which deals away with the weak
instrument issue. Stock and Yogo (2002) formalized the deﬁnition of “weak instruments” and it
suggested that a ﬁrst stage fstat of greater than 10 is suﬃcient to nullify concerns about weak IV
bias. Stock and Yogo (2005) go into more detail and provide useful rules of thumb regarding the
weakness of instruments based on a statistic due to Cragg and Donald (1993). To test for this I use
the Weak IV test in Stata which reports the Cragg-Donald Wald F statistic and reports the critical
values complied by Stock and Yogo (2005). Appendix 1 presents the results. We see a very high
17I include month of marriage FE to account for the possibility that women getting married in

diﬀerent months might vary systematically

16

fstat for IV1 and IV2 which rejects the null hypothesis that instruments are weak at the highest
critical value.

1.4.2.2 Exogenity of the Instrument

The instruments uses being born in a low marriage frequency month versus being born in a high
marriage frequency month as the primary source of exogenous variation for age at marriage. There
are two separate but related threats to the exogenity of the instruments. First if month of birth is
not randomly assigned between the low and high marriage frequency groups. This is problematic
if for example wealthy girls are born in the low frequency months and wealthy girls marry wealthy
boys (assortative matching). Then it might appear that girls born in low frequency marriage months
marry later and have better spousal quality, but this is not causal. To tackle this issue I look at
a population of unmarried girl child less than age 1018 and see whether the IV is predictive of
their background mother and family characteristics. For mother characteristics I look at mother’s
education, age at marriage, age at ﬁrst birth and number of children. For family characteristics I
look at a measure of wealth, whether the household belongs to a rural versus urban area, household
size and whether the household has electricity.

The second threat to the validity is if being born in low marriage frequency months versus
high marriage frequency months aﬀects outcomes for the girls through channels other than age
at marriage, e.g if girl child born in low marriage frequency months are less subjected to early
childhood diseases and are therefore healthier independent of age at marriage. Then the estimates
we get from this causal analysis would be biased. To refute this I show that the IV is not be
predictive of early childhood health outcomes of the girls.
I look at a population of girl child
between ages 0-4, for this population the NFHS-4 data set provide measures of natal health such
18The reason I chose this population is because in the data we have mother and family charac-
teristics only for girls whose mothers are sampled, i.e who live with their natal family which only
happens if they are unmarried and at older ages (e.g 14,15,16) the sample of unmarried girls living
with their natal families becomes more and more selected (because girls start getting married and
move out).

17

as size at birth; weight at birth and one year mortality. I see whether being born in low versus high
marriage frequency month is predictive of these early childhood health outcomes.

To show these I run an OLS regression model (Equation 3) without including any controls. The
purpose is to see whether the IV is correlated with early childhood health and family characteris-
tics. The main coeﬃcient of interest is Beta which captures the eﬀect of the IVs on background
characteristics and natal health.

Yi = βIVi + i

(1.3)

Table 3 panel C presents the results for the eﬀect on early childhood health outcomes. The
outcomes I look at are i) size at birth this is a categorical variable19 reported by the mother with 1
being the largest and 5 being the smallest ii) weight at birth in pounds iii) one year mortality. Being
born in a low frequency marriage months is associated with a 0.01 points (0.3 percent) increase in
log odds for being at a higher birth size, 0.001 pounds (0.03) percent decreases in child’s weight at
birth and 0.0003 percentage points (0.6 percent) increase in the probability of one year mortality.
None the eﬀects are statistically or economically signiﬁcant.

Table 3 panel A presents the results on mother characteristics. Being born in a low marriage
frequency month is associated with 0.008 (0.2 percent) increase in mother’s education, decreases
mother’s age at marriage by 0.097 years (0.5 percent), decreases mother’s age at ﬁrst birth by 0.002
years (0.01 percent) and increases total number of children by 0.02 (0.05 percent). None of the
results except mother’s age at marriage is statistically or economically signiﬁcant. From previous
literature (Chari et al, 2017); we see that a 0.097 years decrease in mother’s age at marriage is
causally associated with 0.01 years (0.19 percent) decrease in number of years of education for the
child, 0.25 percentage points (0.31 percent) decrease in probability of being enrolled and 0.3 percent
decrease in test scores, these eﬀects are small in magnitude and not economically signiﬁcant. The
mechanism is that women who are born in low marriage frequency months gets married later and

19Here I run an ordered logit regression without any controls

18

have better health and fertility outcomes. The fact that women who are born in low frequency
months have a lower mother age might slightly downward bias the beneﬁts we see from delayed
marriage.

Table 3 panel B presents the results for the eﬀect on household characteristics. Being born in
low frequency month increases household members by 0.008 (0.1 percent), increases the probability
of having electricity by 0.3 percentage points (0.4 percent), decreases the probability of being from
a rural area by 0.5 percentage points (0.8 percent) and increases wealth by 0.03(0.1 percent). None
of these eﬀects except the eﬀect on wealth is statistically or economically signiﬁcant. The eﬀect
on wealth is very small and the direction of the eﬀect is opposite to what I found for mother’s age
at marriage. Overall I conclude that being born in low frequency month does not systematically
aﬀect the woman’s early childhood health, mother and family characteristics.

1.4.2.3 Mechanisms

Conceptually the eﬀect of age at marriage on fertility and health should occur in steps.
In the
ﬁrst step age at marriage inﬂuences decisions regarding intermediary variables such as educational
attainment, spousal quality, spousal family quality. In the next step each of these decisions aﬀects
more tertiary outcomes such as bargaining power of the women in the household and ﬁnally all
these channel combined aﬀects her fertility decision and health. Note that these steps might not
always be sequential, for example bargaining power of the women can be directly aﬀected by the
age at which she enters the household at the same time the bargaining power of the women can be
inﬂuenced by her health and fertility outcomes, hence it’s hard to judge whether bargaining power is
an intermediary variable for ﬁnal outcomes such as health and fertility. Broadly categorizing, three
potential channels through which age at marriage might aﬀect health and fertility are educational
attainment, spouse and spousal quality and bargaining power. I ﬁrst establish the overall causal
eﬀect of age at marriage on health and fertility. Then I try to disentangle the eﬀects of the potential
channels. First I control for spouse and spousal quality then the eﬀect we see are thorough the
channels of education and bargaining power. Next I control for bargaining power along with spouse

19

and spousal family characteristics. Finally following Chari et al, (2017) and Field and Ambrus,
(2008) I restrict the sample to women with zero years of education and include the controls for
bargaining power and spousal quality. In this sample the eﬀect we see should be driven by factors
other than the above three channel. Since education itself is an outcome variable, restricting the
sample on it makes the results biased20. Also the zero education population itself might be very
diﬀerent form the rest of the sample and hence the estimates obtained on them is not reﬂective of
the true population eﬀect. The purpose of restricting the sample to zero education is to show that
age at marriage aﬀects health and fertility through channels other than gains in education.

1.5 Results

To understand the eﬀect of age at marriage of a women on her health and fertility outcomes I
start by analyzing the full sample. First I look at the eﬀect of age at marriage on her health outcomes
such as diabetes, blood pressure and anemia. Then I look at the eﬀect on her fertility outcomes
such as total number of children, age at ﬁrst birth and birth spacing. These results are estimated
using regression equation 2, I include minimal amount of controls for the above set of regressions
such as age at survey and month of marriage FE. Next I look at the eﬀect of age at marriage on more
intermediary variables such as education, spousal and spousal family characteristics and bargaining
power. Here I control for women21 and family characteristics.22 For the above set of outcomes I
present the results for both the instruments, IV1 and IV223. I also present the OLS results, though
the discussions would mainly focus on the estimates using IV1. In appendix table 4 I present the
Wald IV estimates for some of the outcomes Finally I present the results on fertility and health under
20Yes years of education is aﬀected by age at marriage but having zero versus some year of
education is not strongly aﬀected by age at marriage. The decision to give the female child some
education versus no education is taken at a much earlier age compared to the decision of getting them
married. When I do a simple OLS controlling for household char we see a very weak correlation
(0.03 percent) between age at marriage and having zero years of education versus some.

21Women characteristics includes age, and year of birth FE.
22Family characteristics includes number of household members, whether the family lives in
an urban versus rural area, age and sex of household head, whether the household has access to
electricity and wealth.

23With 0.7 as the value of Beta.

20

four speciﬁcations; speciﬁcation 1 presents the result for the full sample with minimal controls24,
speciﬁcation 2 presents the results including spousal25 and spousal family controls26, speciﬁcation
3 presents the results including spousal, spousal family and proxies for bargaining power controls
and speciﬁcation 4 presents the results including spousal and spousal family controls for the zero
education sample. This will help us understand the channels through which age at marriage for a
women aﬀects her health and fertility.

1.5.1 Health outcomes

Table 4 reports these results. One month delay in marriage decreases the probability of a woman
being diabetic by 0.07 percentage points (2.8 percent) and decreases the probability of being
pre-diabetic by 1.3 percentage points (3.5 percent). One month delay in marriage decreases the
probability of a woman having high blood pressure under the diastolic measure by 0.27 percentage
points (0.8 percent) and decreases the probability by 0.18 percentage points (0.6 percent) under the
systolic measure. These results are highly statistically and economically signiﬁcant. The magnitude
of some of these results are small but we need to keep in mind that these are the eﬀects of one month
delay in marriage and if we linearly27 transform the eﬀects into one year delay then the eﬀects are
pretty substantial. Next I look at anemia, one month delay in marriage increases the probability
of being severely anemic by 0.01 percentage points (0.1 percent) and increases the probability of
being moderately anemic by 0.05 percentage points (0.3 percent). The results for anemia are not
statistically or economically signiﬁcant. Overall we see that as age at marriage increases women
have better health outcomes.

24This is what I already report.
25Husband’s age and education.
26Family characteristics includes number of household members, whether the family lives in
an urban versus rural area, age and sex of household head, whether the household has access to
electricity and wealth.

27This might not be completely accurate but it helps us to understand the magnitude of these

eﬀects

21

1.5.2 Fertility Outcomes

Table 5 reports these results. One month delay in marriage decreases the total number of children
by 0.0005 (0.02 percent). The eﬀect is very small and statistically not signiﬁcant. When we look
at the ideal number of children the woman wants to have, we see that one month delay in marriage
decreases the ideal number of children by 0.002 (0.09 percent). Again the eﬀect is very small and
statistically not signiﬁcant. One month delay in marriage increases the age at ﬁrst birth by 0.06
months (0.3 percent) and decreases birth gap by 0.019 months (0.05 percent). The eﬀect on age at
ﬁrst birth is highly statistically signiﬁcant whereas the eﬀect on birth gap28 is not.

1.5.3

Intermediary Outcomes

Table 7 Panel A reports these results. Now we look at the eﬀect of age at marriage on some of the
intermediary outcomes such as women’s education, spouse and spousal family quality and proxies
for bargaining power within the marriage. One month delay in marriage increases the number of
years of education of the woman by 0.03 years (0.5 percent) and the eﬀect is statistically signiﬁcant.
One month delay in marriage increases the probability of the women having at least primary level
of education by 0.03 percentage points (0.04 percent) and the eﬀect is statistically not signiﬁcant.
Table 6 reports these results. One month delay in marriage increases husband’s years of
education by 0.03 years (0.4 percent), decreases husbands age by 0.024 years (0.05 percent),
increases the wealth of the in laws household by 0.02 (0.07 percent) and decreases the probability
that the in laws household belongs to a rural area by 0.7 percentage points (0.9 percent)29. Overall
we see that delayed marriage age improves spousal and spousal family quality. One thing to think
about here is the timing of the spousal match. Usually ﬁrst, the spousal match occurs and then the
search for the marriage date begins. Now being born in a low marriage frequency month delays
marriage age through the search for the marriage date process. In that scenario the spousal quality
28In appendix table X we see that one month delay in marriage delays the age at ﬁrst and second

birth by the same magnitude, hence we do not see an eﬀect on birth gap.

29Appendix table 2 presents some more results on in laws family characteristics

22

should not be aﬀected. One possibility is that the delayed marriage gives the husband more time to
increases his educational attainment which in turn increases the family’s wealth. This can explain
why we see a gain in spousal education and family characteristics but it doesn’t explain the decrease
in husband’s age. The decrease in husband’s age is small (0.05 percent) but still signiﬁcant. The
only other possible explanation is that, since women born in low marriage frequency months have
a longer wait time between spousal match and the marriage there might be a higher chance for the
match breaking before marriage and if bad matches (higher age diﬀerences) break more frequently,
then that might explain why we see an eﬀect on spousal age.

Table 7 Panel B reports these results. One month delay in marriage decreases the probability of
the husband limiting the women to contact her natal family by 0.6 percentage points (3.6 percent),
decreases the probability of the husband insist of knowing her whereabouts by 0.7 percentage points
(3.3 percent) and decreases the number of control questions answered yes by 0.01 (0.2 percent).
The eﬀects are statistically signiﬁcant.

1.5.4 Discussion and Mechanism

1.5.4.1 Health

I ﬁnd that delaying age at marriage decreases the probability of a woman having diabetes and high
blood pressure. Though partially genetic, high blood pressure and diabetes can also be driven by
lifestyle choices such as lack of exercise, unhealthy meals and obesity. Lifestyle choices can very
well be inﬂuenced by the intermediary variables we looked at such as education, spousal quality
and bargaining power in the marriage. Hence the health gains we saw could be partially driven
by the gains in the intermediary variables we looked at. To systematically unpack the channels I
look at three diﬀerent speciﬁcations, table 8 panel B presents the results on health controlling for
spousal and spousal family characteristics, panel C presents the results including spousal, spousal
family and bargaining controls and panel D presents the results including spousal and spousal
family controls for the zero education population.

23

We see that adding spousal controls (panel B) does not change the estimates for the health
outcomes, this rules out the theory that the gains in health is completely driven by the gains in
spousal quality. Next I add bargaining controls (panel C), when I control for bargaining power I see
that estimates are more negative implying higher gains in health outcomes. One thing to note here
is that the bargaining questions are at best, proxies to the actual bargaining dynamics. Also these
questions were asked to a sub-population of the sample, on further inspection I ﬁnd that the higher
gains are not driven by the controls itself rather it is driven by the chosen sub population. That is
if I run speciﬁcation 1 on the sub population to which the bargaining question were asked, I get a
bigger negative eﬀect. Hence I fail to conclusively state that whether the gains in health are driven
by a higher bargaining power within the marriage. Finally in speciﬁcation 4, which looks at the
zero education sample, I ﬁnd that the gains are even larger this rules out the theory that the gains
in health are completely driven by gains in education. To conclude we see that delaying marriage
causally results in better health outcomes for the women and the gains are not completely driven
by gains in education or spousal quality.

1.5.4.2 Fertility

I unpack the eﬀect of delaying age at marriage on fertility in table 9. Previously we saw that
(section 5.2) delaying marriage age had no eﬀect (0.02 percent decrease) on the total number of
children but the eﬀect was statistically not signiﬁcant. One month delay in marriage increases the
number of children by 0.009 (0.3 percent) in speciﬁcation 2, by 0.008 (0.3 percent) in speciﬁcation
3 and by 0.016 (0.5 percent) in speciﬁcation 4. The eﬀects are statistically signiﬁcant but small
in magnitude. Conceptually, delayed age at marriage can have opposing eﬀects on the number of
children the women has. Getting married later gives the woman less time to bear children whereas
getting married early and having children early possess health risks to the woman and child. When
I control for spousal quality in speciﬁcation 2 and block the education channel in speciﬁcation 4 I
see a positive eﬀect on fertility compared to the zero eﬀect in speciﬁcation 1. This might be driven
by the fact that higher education among women is correlated with lower number of children and

24

higher spousal education and wealth can lead to lower fertility through higher use of contraception
and lower son preference. These could also be the reason why we see a statistically signiﬁcant
positive eﬀect on ideal number of children under speciﬁcation 2 and 4 but a zero eﬀect under
speciﬁcation 1. For age at ﬁrst birth we see that adding spousal and bargaining controls do not
change the estimate.

1.6 Comparison with the Menarche IV

The previous causal literature which looks at the eﬀect of age at marriage uses age at menarche
as an instrumental variable (Field and Ambrus 2008, Sekhri and Debnath 2014, Chari et al 2017)
for age at marriage. In this section I presents the results comparing the two IVs using the menarche
sample. The menarche sample is a sub sample of the full sample. I obtain the results from both
IV on the menarche sample so that we can draw a direct comparison. The menarche sample is
quite diﬀerent from the full sample, table 1 provides the descriptive statistics for both the sample.
The question about age at menarche was asked to ever married women between the age of 15-
25 whereas the full sample consists of ever married women between the ages 15-49, hence the
menarche sample on average is much younger than the full sample. The menarche sample is also
on average more educated, has lower number of children, belongs to larger household and have
lower wealth compared to the full sample.

1.6.1 First Stage

Table 10 reports the ﬁrst stage estimates for both the instrumental variables.30 Both the instruments
are strong predictors of age at marriage. A one year delay in age at menarche delays age at marriage
by 2 months whereas for IV1 the ﬁrst stage estimate is 2.3 months, both the estimates are highly
signiﬁcant. For IV1 we saw a ﬁrst stage estimate of 3 months for the full sample, the diﬀerence
is attributed to the aforementioned diﬀerence in sample. For age at menarche, previous literature
30To draw direct comparison with the previous literature I control for age, height and include

caste and district FE, following Chari et al, (2017) and Field and Ambrus, (2008)

25

reports a ﬁrst stage of around 4.8 months (Chari et al, 2017), 4 months (Sekhri and Debnath, 2014)
and 8.8 months (Field and Ambrus, 2008). The diﬀerence from the previous literature could be
attributed to the fact that the menarche sample on average is much younger and consists of anewer
cohort. Both Chari et al, (2017) and Sekhri and Debnath (2014) uses the IHDS 2005 data set from
India and report an average age of 33, Field and Ambrus, 2008 uses the 1996 MHSS data set form
Bangladesh and report an average age 33. A more recent data set combined with lower average
age implies a more recent cohort. The lower magnitude of the ﬁrst stage implies that over time the
predictive power of age at menarche for age at marriage decreases.

1.6.2

Intermediary Variables

Before we compare the ﬁnal outcomes on health and fertility I look at the intermediary variables
such as education, spousal and spousal family quality. My prior was that both the IVs should
produce similar estimates for these outcomes. Table 11 reports the results, we see that one month
delay in marriage increases years of education by 0.14 months using IV1 and 0.13 months using age
at menarche, the eﬀects are statistically signiﬁcant. The full sample estimate is 0.03 months. Field
and Ambrus, 2008 found that delaying marriage by one year increases education by 0.22 years,
this is comparable to the baseline results converted to years (0.36 months). The menarche sample
estimates are higher when compared to the full sample and previous literature. Menarche sample
looks at a more recent cohort compared to the other samples and if the beneﬁts of delaying marriage
increases with time then that might explain why we see a larger eﬀect in the menarche sample. One
month delay in marriage increases the probability of having at least primary education by 0.5 (0.6
percent) percentage points using IV1 and 0.6 (0.6 percent) percentage points using age at menarche.
Again the eﬀect for the menarche sample is bigger than the full sample and statistically signiﬁcant.
Now we look at spousal and spousal family quality. One month delay in marriage increases
husbands education by 0.06 years using IV1 and 0.07 years using age at menarche, the results
are statistically signiﬁcant. Again the estimates are higher than the full sample (0.03). Field
and Ambrus, 2008 also looked at husband’s education and found that one year delay in marriage

26

increases husbands age by 0.04 years but the eﬀect is not statistically signiﬁcant. We see a pattern
similar to women’s education here, where the gain in husband’s education are higher for the more
recent cohorts. When I look at husband’s age I see that one month delay in marriage decreases
husbands age by 0.01 years using IV1 and 0.1 years using age at menarche. The full sample estimate
is a decrease of 0.02 years. The results using IV1 are not statistically signiﬁcant hence I cannot
draw a conclusion about why we see a bigger eﬀect for the menarche IV. For spousal household
quality we see a similar eﬀect from both the IVs and the full sample. Overall we see my prior that
both the IVs should produce similar results for the intermediary variable, holds.

1.6.3 Health

Since early or late onset of menarche is correlated with underlying health factors my prior was
that my IV and age at menarche IV will produce diﬀerent estimates. Table 12 reports the results,
I see that delaying age at marriage by one month decreases the probability of being diabetic by
0.01 percentage points (0.6 percent) using IV1 and 0.03 percentage points (1.8 percent) using age
at menarche IV. The full sample estimate showed a decrease of 2.8 percent. One month delay in
marriage decreases the probability of having high blood pressure by 0.2 percentage points (0.8
percent) using IV1 and 0.7 percentage points (2.8 percent) using age at menarche IV. The full
sample estimate is a decrease of 0.8 percent. For anemia we see that one month delay in marriage
has no eﬀect on the probability of being severely anemic using IV1 and 0.05 percentage points (5
percent) increase in probability using age at menarche IV. I didn’t ﬁnd any signiﬁcant eﬀect from
the full sample. Over all we see that the health estimates are not very diﬀerent across the two IVs.
The estimates from the menarche IV are slightly higher but overall has the same direction as IV1.
A possible explanation for why we see similar results across the IVs could be that the health factors
which are inﬂuenced by early or late onset of menarche might not aﬀect health outcomes such
as diabetes, blood pressure and anemia. These health outcomes are more driven by genetics and
lifestyle choices whereas late or early onset of menarche aﬀects the women’s reproductive health.

27

1.6.4 Fertility

Table 12 reports these results. The ﬁrst divergence in the results from the two IV is when I look at
total number of children. One month delay in marriage decreases the number of children by 0.0001
(0.005 percent) using IV1, the eﬀect is not statistically signiﬁcant. For the full sample we see a
0.02 percent decrease, not statistically signiﬁcant. When I include spousal controls I ﬁnd a small
(0.4 percent), positive and statistically signiﬁcant eﬀect. In summation IV1 estimates a zero or
small positive eﬀect on the number of children. Whereas age at menarche IV estimates a negative,
statistically signiﬁcant eﬀect on the number of children. Using age at menarche IV, one month
delay in marriage decreases the number of children by 0.017 (1 percent) decrease, and the eﬀect is
statistically signiﬁcant. Using age at menarche as an IV, Chari et al, (2017) ﬁnd that one year delay
in marriage decreases the number of children by 0.13 (4.64 percent). Hence both the menarche
sample and previous literature estimates a negative, statistically and economically signiﬁcant eﬀect
of age at marriage on number of children, whereas my IV1 predicts a zero or small positive eﬀect.
One reason for the diﬀerence in results could be that the set of compliers for both the IVs are
diﬀerent. In the menarche paper the compliers are women who delay marriage because of delay
in menarche whereas in my paper the compliers are women who delay marriage to get married at
an auspicious marriage month. The other possible reason is that age at menarche is not exogenous
when we look at the eﬀect of age at marriage on fertility. In the medical literature we see that
both early and late onset of menarche is often associated with adverse pregnancy outcomes such as
ectopic pregnancies and spontaneous abortions (Liestol, 1980; Martin et al., 1983; Wyshak, 1983;
Sandler et al., 1984). Later onset of menarche is also associated with a slightly higher risk of sub
fecundity and infertility (Gulbrandsen et al, 2014). Age of menarche has also been associated with
other health issues for women such as breast cancer and cardiovascular diseases (Petridou et al.,
1996; Garland et al., 1998). Since age at menarche is associated with reproductive health using
it as an IV to measure total number of children might bias the results. Women with late onset of
menarche might have lower reproductive health and delayed marriages. In this scenario the lower
number of children are driven by both lower reproductive health and delayed age at marriage, hence

28

the estimates could be biased.

In the data we have a few measure of reproductive health such as irregularities in period and
pregnancy termination. I use these variable to see whether age at menarche directly aﬀect these
variables even after controlling for age at marriage. I use an OLS regression model with age ﬁxed
eﬀects. I ran the model on the Menarche sample excluding woman who are currently pregnant.
I look at two variables, the ﬁrst variable is a binary variable indicating whether the woman had
periods in the last six weeks.
I view this as an indicator for period irregularities. The second
variable I look at is pregnancy termination. Appendix table 3 reports the results, I see that even
after controlling for age at marriage, women’s education, height, age, spousal and spousal family
characteristics age at menarche has a statistically signiﬁcant eﬀect on period irregularities. As age
at menarche increase irregularities decreases. This is again supportive evidence that the age at
menarche might directly aﬀect reproductive health and hence might not be a valid instrument for
this outcome.

When I look at other measures of fertility like ideal number of children and age at ﬁrst birth
I ﬁnd that both the IV’s produce similar estimates and the results are comparable to the baseline
estimates and previous literature.

1.7 Robustness Checks

1.7.1 Month of Birth

One non parametric way to capture the essence of the instrumental variable introduced in this
paper is to use month of birth dummies as instruments. Note that to use month of birth dummies
as an IV we need to make a stronger assumption for exogenity, compared to the current IV. For
the current IV the exogenity assumption is that women who are born in months with low number
of marriage dates are not systematically diﬀerent from women who are born in months with high
number of marriage dates, this I show in the paper. The exogenity assumption for using month of
birth dummies is, women born is diﬀerent months are not systematically diﬀerent from each other.
From the previous literature (Weber et al., 1998; Lokshin and Radyakin, 2009) we know that this

29

might not hold. In table 13 panel A, I present few of the results using month of birth dummies as
IV. I see that the estimates are similar to the ﬁndings of the paper.

1.7.2 Measurement Error

Graph 19 looks at the distribution of birth month. There is an unnatural jump at January. This
might be because of miss reporting. When we look at the distribution of birth months of children
between ages 0-4 in appendix graph 20 we do not see such jumps. Since their birth is more recent
there is less chance of misreporting. Since the IVs are created using month of births this might have
an eﬀect if there is non-random bias in reporting. Like if more people from low socioeconomic
background do not know their birth months and report as January. To account for this ﬁrst I drop
all the women for whom month of birth or month of marriage were imputed in the sample. This is
done for all the results presented in the paper. Next I present some of the results dropping women
who reports their month of birth as January. Table 13 panel B reports the results using IV1 for
the full sample dropping the women who reported their month of birth as January, I see that the
estimates are slightly bigger in magnitude, statistically signiﬁcant and have the same sign’s as the
main results.

Another form of misreporting could be if the women does not remember her month of marriage
and when asked, only remembers her age at marriage in years then she might report her month of
marriage as her month of birth. For example if the women remembers that she got married when
she turned eighteen and does not remember the month of marriage then she might just report the
month she turned eighteen in, i.e. her birth month. To see check for this I estimate some of the
results dropping women who reports the same month of marriage and birth, i.e I drop the women
for whom the variable target takes the value zero. I see that the estimates are statistically signiﬁcant
and similar to the main results.

30

1.7.3 Non Monotonicity of the IV

Monotonicity is a condition that the IV needs to satisfy for the estimates to be interpretable as
LATE (Hahn, Todd, and Van der Klaauw (2001)). Unlike the relevance and exogenity condition
this condition in rarely tested in the applied literature (Fiorini et al. 2016). There is no formal test
of the monotonicity condition and this is still an active area of research. Fiorini et al., (2016) in
their paper suggest that the interpretation of the estimates is not completely lost if the monotonicity
condition fails; they also suggest that the monotonicity condition should be tested using economic
intuitions and suggestive data patterns.

In my paper the monotonicity assumption might fail if being born in a period of low number
of marriages encourages some part of the population to get married earlier. The IV stems from
two social practices; seasonality of marriage dates and minimum age targeting at marriage.
I
deﬁned minimum age targeting as a preference where parents want their daughters to get married
oﬀ early and after turning a marriage eligible whole number age. Another form of targeting could
be if the parents want their daughters to get married before turning a particular age (type II). For
example if parents want their daughters to get married before they turn 21. Type II targeting in
itself would not aﬀect monotonicity. If the parents still want their daughter to get married early then
this kind of targeting would not cause bunching at any particular age. Type II targeting combined
with the parent’s preference to get their daughter married oﬀ as late as possible would challenge
the monotonicity condition. For example if the parents want their daughter to get married before
turning 21 and they want their daughter to get married as late as possible, now if the daughter
turns 21 in a period of low marriage dates then the parents might get her married early. Hence for
monotonicity to fail two things must hold i) Parents want their daughter to get married as late as
possible ii) Parents want their daughters to get married before turning a particular age (type II).
I argue that in context of India there is no evidence that parents have preference for getting their
daughters married as late as possible hence the monotonicity condition holds. In the next paragraph
I will presents suggestive evidence that the above conditions which invalidates monotonicity does
not hold.

31

Suppose both the conditions which invalidates monotonicity holds then when we look at the
distribution of target (Graph 4) we should see a spike at 11 similar to the one we see at 0. We do
not see any such spikes. When we plot the fraction of marriages by distance from a whole number
age (Graph 5) we see similar fraction of people getting married at 1 and -1. This implies that the
targeting is towards 0 and not towards -1 which would not have been case if the conditions for
non-monotonicity holds.

In this paragraph I provide another suggestive evidence that type II targeting does not exist.
When we look at the back of the envelope calculation for the distribution of the variable target
(Graph 6) we see that the actual distribution of the variable target diverges from the calculated one
at the value 0 and not at 11. Hence the fraction of people who would get married at 11 is the same
as if would have happened if month of birth and month of marriage is uncorrelated. This gives
further evidence to the fact that monotonicity conditions holds.

1.7.4 Seasonality versus Targeting: Non-Hindu Sample

When we look at the distribution of marriages by month over religion in graph 13 we see that there
is seasonality in marriages within some other religions also. If we look at the Muslim and Sikh
population we see that they have a seasonality of marriage similar to the Hindu population. As
I mentioned before that seasonality in marriage might be driven by reasons apart from religious
beliefs such as overlap with the festive season which is usually post-harvest, climate, availability
of wedding services etc. When we look at the distribution of the variable target by religion in
graph 21 we see that distribution of the variable target looks similar between the Hindu and Muslim
population whereas it is not present in the Sikh population. Hence for the Muslim population we
see both seasonality in marriage dates and minimum age targeting at marriage, whereas for the
Sikh population we see seasonality but not targeting.

In appendix table 5 I present the ﬁrst stage estimates of IV1 on the Muslim and Sikh population,
we see that for the Muslim population the ﬁrst stage eﬀect is pretty similar to that of the Hindu
population whereas for the Sikh population there is no ﬁrst stage eﬀect. This is suggestive evidence

32

that the IV strategy presented in this paper is only valid for populations which practice both
seasonality in marriage dates and minimum age targeting. Seasonality of marriage dates in absence
of minimum age targeting at marriage does not delay age at marriage for women who are born in
the low marriage frequency month.

In appendix table 6 I present some of the IV results for the Muslim population, I see that the
eﬀects are similar to the main estimates with the Hindu population in terms of direction but the
magnitudes of the eﬀects are higher. We also see that on average the Muslim women has higher
diabetes, higher blood pressure, more number of children and lower age at birth. This might be a
probable reason of why we see higher magnitudes compared to the Hindu population.

Hence this IV can easily be extended to any population which has seasonality in marriage dates

and practices age targeting at marriage.

1.8 Conclusion

In this this paper I use a nationally representative data set from India to look at the causal eﬀect
of age at marriage of a women on her health and fertility outcomes. To establish this causal relation
I propose a new instrument, the instrument uses two social practices. First parent’s preference to
get their daughter married oﬀ early after they turn a minimum marriage eligible age and second
seasonality of marriage dates. I ﬁnd that delayed age at marriage for a woman improves her health
post marriage in terms of being diabetic or having elevated blood pressure. Delayed marriage also
increases her age at ﬁrst birth and has a zero eﬀect on number of children. I also ﬁnd that delayed
marriage age also increases her educational attainment, increases her spousal and spousal family
quality and increases her bargaining power within the marriage. When I looked at the potential
channels through which delayed marriage improves the women’s health I ﬁnd that the beneﬁts are
not completely driven by gains in education and spousal quality.

I also compare my ﬁndings with the existing instrumental variable in the literature, age at
menarche. I ﬁnd that the results from both the IVs are similar when I looked at health, education
and, spousal and spousal family quality. The results diﬀer when I look at the eﬀect on total number

33

of children.

The instrumental variable strategy used in this paper can be easily extended to other countries
which has minimum age laws and seasonality in marriage dates. Since these features are not
uncommon in developing countries I plan to extend this study to those countries. Previous causal
literature used age at menarche as an IV but overtime as the average age of marriage increases
the IV would become less relevant. The IV introduced in this paper can be used to draw causal
inferences in those scenarios.

The paper is also ﬁrst in the literature to talk about minimum age targeting in marriage. It shows
how background characteristics vary for people who practice minimum age targeting at marriage.
I belief that there is a lot more to explore about this social phenomenon.

The ﬁndings from the paper supports the existing inference in the literature that increasing age
at marriage has many beneﬁts for women in developing countries. The paper provides support to
any policy which aims at delaying the age at marriage for girl child.

34

Figure 1.1: Distribution of Age at Menarche and Age at marriage

35

Figure 1.2: Distribution of fraction of marriages by months

Notes** 1=January, 2=February ....12=December

36

Figure 1.3: Distribution of Age at Marriage in months

37

Figure 1.4: Distribution of the variable Target

38

Figure 1.5: Fraction of Marriages by distance from a whole number age

39

Figure 1.6: Back of the Envelope Distribution of the Variable Target

40

Figure 1.7: Wealth by Distance from a whole number marriage age

41

Figure 1.8: Education by Distance from a whole number marriage age

42

Figure 1.9: Rural by Distance from a whole number marriage age

43

Figure 1.10: Distribution of Target By Wealth

44

Figure 1.11: Distribution of Target By Education

45

Figure 1.12: Distribution of marriages by month over time

Notes** 1=January, 2=February ....12=December

46

Figure 1.13: Distribution of marriages by month across religion

Notes** 1=January, 2=February ....12=December

47

Figure 1.14: Age at Menarche by distance from a whole number age

48

Figure 1.15: Distribution of Education by month of marriage

Notes** 1=January, 2=February ....12=December

49

Figure 1.16: Distribution of Rrural by month of marriage

Notes** 1=January, 2=February ....12=December

50

Figure 1.17: Distribution of Wealth by month of marriage

Notes** 1=January, 2=February ....12=December

51

Figure 1.18: Distribution of month of marriage by month of marriage

Notes** 1=January, 2=February ....12=December

52

Figure 1.19: Distribution of fraction of births by months for respondents

53

Figure 1.20: Distribution of fraction of births by months for children age 0-4

Notes** 1=January, 2=February ....12=December

54

Figure 1.21: Distribution of Target by Religion

55

Table 1.1: Descriptive Statistics

56

Table 1.2: First Stage

57

Table 1.3: Eﬀect of IV1 on Early Childhood Outcomes

58

Table 1.4: Eﬀect of Age at Marriage on Wealth

59

Table 1.5: Eﬀect of Age at Marriage on Fertility

60

Table 1.6: Eﬀect of Age at Marriage on Spouse and Spousal Family Quality

61

Table 1.7: Eﬀect of Age at Marriage on Education and Bargaining Power

62

Table 1.8: Eﬀect of Age at Marriage on Health

63

Table 1.9: Eﬀect of Age at Marriage on Fertility

64

Table 1.10: First Stage Comparison

65

Table 1.11: Comparison : Intermediary Variables

66

Table 1.12: Comparison : Health and Fertility

67

Table 1.13: Comparison : Robustness

68

RETIREMENT SECURITY AND GENDER GAP IN PARENTAL INVESTMENT

CHAPTER 2

2.1 Introduction

Gender gaps favoring the male child at the household level in various aspects such as health,
education and personal autonomy are systematically higher in developing countries (Jayachandran,
2014).
In the Indian context, there have been various studies trying to establish the reasons
behind these gaps. Some of the established reasons for these gaps include the patriarchal nature
of the Indian society, son preference, poverty, illiteracy, marriage norms and lack of employment
opportunities for women. These biases lead to diﬀerences in investment of resources between the
boy child and the girl child at the household level (Sen and Sengupta 1983, Das Gupta 1987 and
Behrman et al. 1988). In this paper, I will try and investigate the old age (post retirement) security
motive for the parents as a potential reason for this diﬀerential investment between a girl child and
a boy child.

With the absence of proper credit markets and retirement security in developing countries
parents are incentivized to invest in children as means of old age support (Nugent, 1987). It has
also been seen in the existing literature that transfer from children and well-being of the parents
depends on the education level of the children (Zimmer et al, 2007). These provide incentives
for the parents to invest in their children’s education, besides altruistic reasons.
In developing
countries like India, investment in the male child on average generates higher return to the parents
as compared to investment in their daughter, in terms of ﬁnancial support at an old age (Kishor,
1993). This could be attributed to labor market returns and marriage norms.1. Also parents in India
live with their sons in old age and it is the responsibility of the son to provide for parents whereas
the daughter is expected to take care of the husband’s family (Das Gupta, 1987). This skewedness
1When daughters get married they usually live with their husband and his family. She is also

expected to take care of her husband’s parents and her income goes to the husband’s family

69

in dependency towards the son for old age support can incentivise the parents to invest more in
sons compared to daughters. In this paper I will try and investigate this old age support (retirement
security) motive for unequal investment between a boy child and a girl child. Now, parents who
have other sources of post retirement income such as pension or savings will depend less on their
children for ﬁnancial support in their old age. Whereas parents who do not have any pension
income will primarily depend on their children and savings for post-retirement consumption. In
this paper I use a pension policy change as an exogenous shock to post retirement security to look
at the causal eﬀect of retirement security on gender gap in investment.

In 2004, the Government of India introduced the New Pension Scheme for its public sector
employees joining after January 1st, 2004. The public sector employees who joined prior to that
were enrolled in the Civil Servant Pension Scheme. This shift from the Civil Servant Pension
Scheme to the New Pension Scheme was seen as a decrease in the post retirement security by the
public sector employees. While the policy change decreased the post retirement security for the
public sector employees it had no eﬀect on the private sector employees. Due to the policy change
the parents who joined the public sector post 2004 (after the policy change) had lower retirement
security compared to the parents who joined prior to 2004 (before the policy change). This implies
that parents who joined the public sector post 2004 will be more dependent on their children for
old age support and since the dependency is skewed towards the son we might see a rise in gender
gap. In graph 1, I plot the diﬀerence in private school enrollment between boys and girls. We see
that there is a rise in gender gap in enrollment for children whose parents joined the public sector
post 2004 compared to those whose parents joined prior to 2004. We do not see such a rise for
children whose parents worked in the private sector or agricultural sector. This is indicative of the
fact that gender gap increased for children of public sector employees post policy reform.

To capture the causal eﬀect of retirement security on gender gap in child investment and
outcomes, I use the policy change in a triple diﬀerence framework. The ﬁrst diﬀerence between the
public sector employees who joined post 2004 and prior to 2004 will give us the eﬀect of decrease
in post-retirement security and time eﬀects on investment in children. A second diﬀerence between

70

parents who joined the private sector post 2004 and prior to 2004 will give us the time eﬀects
as there was no major pension plan change between these two time periods for the private sector
employees. Assuming that the time eﬀects are same for the two sectors, a diﬀerence in diﬀerence
between sectors (public versus private) and joining time (post versus pre) will give us the eﬀect of
decrease in retirement security on investment in children. Finally a third diﬀerence by the gender
of the child will capture the eﬀect of retirement security on gender gap in child investment. This
is captured by running a triple diﬀerence model across sector, time and gender. The prior is that,
as post retirement security decreases the girl child is worse oﬀ compared to the boy child in terms
of educational investment and outcomes. Alternatively stated, gender gap goes up for children of
public sector employees versus private sector employees, post reform.

I use a nationally representative panel data from India, the Indian Human Development Survey
2004 and 2011. The major outcomes that I look at are schooling choices2 and standardized test
scores for children.3. While schooling choice such as public versus private school is indicative of
the monetary investment made in the child, test scores reﬂect the sum total of various investments
made in the child. Private schools in India are more expensive than the public schools and are
associated with better outcomes for the children, so when a child is enrolled in a private school
compared to a public school it is indicative of higher investment in the child. Test scores on the
other hand are a reﬂection of the total investment made in the child.

I ﬁnd that a decrease in post-retirement security, increases the gender gap against the girl
child. We see that the gender gap in private school enrollment is 0.07 (19 percent)4 higher for
children of public versus private sector employees after the reform relative to before and the eﬀect
is statistically signiﬁcant. Also compared to a boy child, girl child is 0.0684 (16 percent) less likely
to be enrolled in a school where the medium of instruction is English5 and the eﬀect is statistically

2Private versus public.
3I do not look at enrollment as 97% of the kids in my sample report being in school (i.e. there

is no variation in school enrollment)

4This is driven by 15.3% increase in private school enrollment for boy child and 4% decrease

for girl child

5The magnitudes of change in enrollment for the boy child and girl child is similar to that of

71

signiﬁcant. The girl child also performs worse compared to the boy child on standardized test scores
in math(-0.07), writing(-0.04) and English ability(-0.06) however these results are not statistically
signiﬁcant. Overall we see that a decrease in post-retirement security for the parents are associated
with worse outcomes and lower investment for the girl child compared to the boy child.

The rest of the paper is organized in the following way.

In section II I discuss the relevant
institutional details to set up the empirical strategy, I discuss the pension system in India, the reform
and how the new pension scheme is worse for the public sector employees. I also brieﬂy discuss the
discrimination against girl child present in India and ﬁnally I discuss private versus public schools
in India. In Section III I provide a brief Literature Review where I talk about previous work on this
topic. In Section IV I develop a theoretical model to explain the linkage between old age security
and diﬀerential investment. In section V I discuss the empirical strategy, section VI I discuss the
data. Section VII presents the results followed by common trends and falsiﬁcation tests in section
VIII. Section IX presents the conclusion.

2.2 Background

2.2.1 Pension System in India

The Indian pension system suﬀers from a lot of drawbacks such as the absence of an universal
social security for the unorganized sector which constitutes the bulk of the work force, low coverage
in general, ever increasing ﬁscal stress on the government and long term structural instability for
the public sector pension system(Sadhak, 2009). In 2012 that the estimated population of elderly
people living in India was above 100 million, with an annual growth rate of 3.9 percent during the
last decade. Thus for a country with a high rate of growing elderly population it has a very low
pension coverage of around 12 percent of the working population. In the absence of a universal
pension system the elderly in India depend mostly on their own savings and transfers from relatives
for their old age consumption.
private school enrollment. This might be driven by the fact that private school and medium of
instruction being English is highly correlated (0.7)

72

Contrary to the present scenario India actually had a long history of pension beneﬁts, some
even dating back to the 3rd Century B.C, where the king had to pay half of the wages for people
who have completed 40 years of service (Gayithri, 2006). Pension system was also present during
the Colonial period of British India. Under the Royal Commission of Civil Establishments 1881,
pension beneﬁts were given to the public sector employees. Indian government further consolidated
these beneﬁts under the Acts of 1915 and 1935. Post-independence other schemes were created to
cover the private sector (Gowswami, 2001). The structure and beneﬁts from the diﬀerent pension
schemes currently available to diﬀerent sectors vary a lot. So it makes more sense to look at them
separately. In the next few paragraphs we will discuss the pension scenario in India prior to 2004,
for the three broad sectors.

First, I discuss the most generous pension scheme in India, the Civil Servant Pension scheme.
This scheme covered public sector employees under the Central and State government. This is a
pay-as-you-go deﬁned beneﬁt system, so the employees do not contribute anything towards their
retirement account. Thus the entire pension expenditure is paid from the government revenues.
By 2012 the scheme had 2.4 million subscribers (Sanyal, 2013). An Indian government employee
is entitled to a monthly pension post retirement or superannuation or invalidation. The pension
amount is ﬁxed and is calculated based on the number of years of service and the average salary
of the last ten months of service. Full superannuation beneﬁt is 50 percent of the average salary
of the last ten months of service. The beneﬁt upon retirement has two components:
the above
mentioned monthly pension and a lump sum payment called the retirement gratuity. Dearness
Pension beneﬁt in the form of family pension is given to the spouse after the death of the employee.
This compensation is approximately 30 percent of the monthly pension. The government sector
employees also have a commutation provision where they can forgo 40 percent of the pension and
take it as a lump sum restored after 15 years of retirement. The pension amount is indexed to
the consumer price index to provide real annuities to the retired government servants. To claim
the entire beneﬁts of the pension scheme the employee has to retire after the age of 60 or after
33 years of service. The penalty for early retirement varies by state and diﬀerent branches of the

73

government.

The second major pension scheme which primarily covers the organized private sector em-
ployees is the Employee Provident Fund Scheme 1952 (EPF) and the Employee Pension Scheme
1995 (EPS). Any organization with more than 20 employees are mandatorily covered by EPF. The
provident fund is a deﬁned contribution, fully funded pension scheme. The employee and the
employer each contribute 12 percent of monthly earnings to the EPF account. This is returned to
the employee in lump sum with interest (Current annual interest rate is 11percent ). For employees
drawing a basic pay of up to Rs 6,500 ($10) per month it is mandatory to make the contribution
towards EPF and EPS. Employees drawing more than that have an option of getting PF deducted
from the salary. In normal circumstances the employee and the employer contribute 12 percent each
of the basic salary of the employee. The contribution of 12 percent from the employee is put into the
EPF account. Out of the 12 percent contribution from the employee 8.33percent is invested in the
EPS and the rest in EPF. The interest rates on these accounts are set by the government every year.
Though the contribution is monthly the interest rate is set annually. In case of death of the employee
the money in the EPF account is handed over to the nominee. The employer’s contribution is tax
free and the employee’s contribution is eligible for tax deduction. EPF withdrawals are permitted
under certain circumstances.

The government scheme along with employee provident fund schemes covers almost 93 percent
of the pension subscribers. There are other tertiary target pension programs which are usually
for the poor. One such pension program is the National Old Age Pension Scheme launched in
1995, targeted to people above 65 years old and below the poverty line. The contribution from the
government coupled with the state was as low as Rs.200 ($3) per month. Other such programs
include micro pension schemes run by MFI and NGOs. The purpose of these pension schemes is to
reach the disadvantaged people at an aﬀordable price. Since in our sample we only look at private
and public sector employees from white collar jobs these schemes are not relevant for our sample.

74

2.2.1.1 The Reform

The Civil Servant Pension Scheme was becoming unstable because of the ﬁscal stress it was creating
on the central and state government. With the demography aging and increased lifespan of the
elderly, the ﬁscal stress on the government increased three folds in the past decade. In 2004, 16
percent of the tax revenue from the state government and 12 percent from the central government
were spent on making pension payments (Sadhak, 2009). A reform in the government pension
scheme was long overdue. The Government of India introduced the New Pension System (NPS)
in 2004, new government employees joining after January 1st 2004 will be placed under the NPS
instead of the traditional CSP and other government pension schemes. The scheme was announced
to the public in December of 2003. The NPS is a deﬁned contribution pension scheme which
requires 10 percent contribution from the employee and ten percent contribution from the employer
(government). The NPS will have two tier, Tier – I is mandatory for all government employees
joining post 2004 and Tier-II is optional. In Tier-I account the employee will contribute 10 percent
of his basic pay, deducted directly from his salary and the government will make an equal matching
contribution. The government of India established the Pension Fund Regulatory and Development
Authority (PFDRA), as the prudential regulator for the NPS. Funds from Tier-I account are mostly
invested in government and corporate bonds. Tier-I is mostly non-withdraw able. Tier-II is optional
and at discretion of the government employee. Government will not make any contribution to the
Tier-II account. There is no limit to the number of withdrawals from the Tier-II account. There
is also a choice for the employee to which fund to invest in. Tier-II account can be looked at as a
medium term investment tool. Since there is no locking period for the funds there is no tax-beneﬁts
also. To make the NPS more eﬀective as a retirement saving tool there is provision for mandatory
annuitization. Individuals can normally exit at the age of 60 or at superannuation from Tier-I of
the pension system. At the time of exit the individual would be mandatorily required to invest
at least 40 percent of the pension wealth to purchase an annuity (from an Insurance Regulatory
and Development Authority (irda)-regulated life insurance company). Finally the existing deﬁned
beneﬁt CSP scheme would not be available to any government employee joining the government

75

sector post Jan 1st, 2004.

2.2.1.2 Comparison between the NPS and CSP

In this section I will discuss why the government sector employees who joined the government sector
post 2004 and fall under the NPS are worse oﬀ in terms of retirement security than government
sector employees who joined the sector prior to 2004 and fall under the CSP. The ﬁrst major
diﬀerence between the two schemes is the fact that the NPS is deﬁned contribution (DC) whereas
CSP is deﬁned beneﬁt (DB). A change from DB to DC schemes shifts the risk of investment
to the employee from the employer. The employee is more exposed to variability in retirement
income. (Broadbent, Palumbo and Woodman, 2006). This variability in retirement income in itself
is viewed as a dis-utility by the employee. DB beneﬁt pension schemes are back loaded and favors
the long tenured employees. Government sector jobs are one of the most secured jobs in India, so
people joining the government sector end up working in the sector all their lives. Thus with high
job security and long tenure the DB scheme was highly beneﬁcial for these employees. Further
a secured job with predictable growth and deﬁned income stream post retirement made working
in government sector very reliable. The change from DB to DC takes away this security from the
employees. Other issues which make the NPS worse than the CSP are the following: the income
under CSP was indexed to inﬂation whereas under NPS it is not and that the contribution rate went
from 0 percent to 10 percent for the government employees.

Now a move from DB to DC is not necessarily detrimental to the employee as they have more
labor force mobility among others. So what is more important is how this change was perceived by
the people. The introduction of the NPS was met by protests all over the country. The people who
joined the sector post 2004 viewed this as an act of stripping them of their pension rights. They
argued that since the return under NPS is market driven there is no guaranteed return.

Another major indication that the government themselves believed that the CSP is better than
the NPS is the fact that the Armed Forces were kept under CSP. Armed Forces are the most valued
functionaries of the nation, the fact that the government is keeping the armed forces under CSP was

76

further seen as a reason that the CSP is better than NPS.

Finally there are reports (K.V Ramesh, 2013) which calculate returns under the two pension
scheme, they look at returns under various market interest rate, ranging from 8-12 percent. Taking
into account that NPS is not inﬂation proof they show that at all plausible market return rates the
CSP outperforms the NPS.

2.2.2 Unequal Investment in Sons and Daughters in India

There is a large literature which analyses the issue of son preference in India. Pande and Astone
(2007) discuss the existence of son preference in rural India and analyze the role of individual
versus structural eﬀects that might generate this son preference. This taste based preference among
parents has been seen to result in pronounced eﬀects on the sex composition of children in families
(Clark, 2000; Mutharayappa, 1997; Arnold et al 1998) as well as in parental investments in child
care (Barcellos et al 2014), nutrition (Chen, Huq and D’Souza, 1981; Das Gupta, 1987; Behrman
1988), vaccinations (Borooah 2004) and healthcare (Basu, 1989). There is evidence that girls
receive less investment in their human capital as compared to boys in both developing as well as
developed nations (Behrman, Pollak and Taubman 1986). There is also evidence that discrimination
begins in the womb (Bharadwaj & Lakdawala 2013). They ﬁnd that mothers visit antenatal clinics
and receive tetanus shots more frequently when pregnant with a boy.

The preference based reasons cited for these gender gap in investment includes the patriarchal
nature of the Indian society, son preference, poverty, illiteracy and marriage laws. Unfortunately
there also exists economic reasons for gender gap in educational investment. Girls face poorer
economic incentives to invest in schooling than boys because they reap lower labor market returns
to education than boy (Kingdon, 1998). Women also suﬀer high levels of wage discrimination in
the Indian urban labor market, but education contributes little to this discrimination (Kingdon and
Unni, 2001). Overall in the literature we see that both economic and preference based reasons drive
gender gap against the girl child.

77

2.2.3 Private versus Public Schools in India

Public schools also knows as government schools have rapidly lost their enrollment to private
schools in India (Murlidharan and Kremer, 2006). Due to poor performance of over-crowded
government schools there has been a steady surge in private school enrollments (Kingdon, 2007;
Wadhwa, 2009). Parents have started to value education more and are sending their children to
private schools which are more expensive compared to government schools (Goyal and Pandey,
2009). The teachers in private school are more accountable as they can be ﬁred whereas as
government school teachers are mostly tenured and have low accountability (e.g. Probe Report,
1999). There have been a few a studies comparing the performance of private versus public schools
in India. These studies ﬁnd that the private schools out performs the public schools on grounds
of test scores, English ability and other child outcomes (Goyal and Pandey, 2009; Kingdon, 2007;
Tooley, Dixon, Shamsan, and Schagen, 2010).standardizing for home background and controlling
for sample selectivity greatly reduces the raw average achievement advantage of private school
students over public school students, but does not wipe it out (Kingdon 1996). Private schools
being more expensive are usually attended by children from better oﬀ and better informed families
(Goyal and Pandey, 2009). Thus enrolling your child in private schools could be viewed as a form
investment on the parent’s part as private schools are more expensive and associated with better
outcomes for the child. Appendix table A3 summarizes private school enrollment in the data, we
see that the average enrollment for boys are higher than girls.

2.3 Literature Review

The Old Age Security hypothesis views children as capital goods. They are used as investments
to transfer resources from present to future (Srinivasan, 1988). One implication of this statement,
which is the core of this paper is that if the male child produces higher return than the female
child then by the Old Age Security hypothesis the male child should get higher investment. There
is a huge literature which talks about old age security and son preference as potential reasons
for increased fertility (Jensen, 2003; De Vos, 1985; Hodinott, 1992; Nugent, 1985; Zhang and

78

Nishimura, 1993; Mutharrayappa, 1997), but there is a dearth of good empirical papers which
establishes a causal link between old age security of parents and diﬀerential investment in children
by gender. This dearth of papers could be attributed to the lack of good intergenerational data in
developing countries.

Ebenstein and Leung (2010) discuss a potential link between old age security and son preference.
They use a voluntary pension program introduced in rural China in 1990 to conﬁrm two hypotheses.
They establish that parents with sons are less likely to enroll in the program and less likely to save
for old age. They also establish that the implementation of rural old age pension programs mitigated
the increase in the sex ratio observed in counties that had at least partial adoption of the program.
Both the ﬁndings support the argument that sons provide higher old age security than daughters
and preference for sons could be driven by this old age security motive. What the paper fails to
capture is how signiﬁcant this old age security motive is in explaining the diﬀerential treatment of
a boy and girl child, which I intend to capture in my paper.

With the access to new data there has been some recent study which looks at the eﬀect of
Old Age Pension on inter generational resource allocation. Mu and Du (2015) link educational
investment in children and old age security for parents. They ﬁnd that as old age security increases
parents invest more in their children’s education. They talk about two reasons why increase in old
age security should aﬀect investment in children’s education. Firstly, the lifetime income of parents
is increasing; so by income eﬀect the investment should increase. Secondly, as old age security
increases, the parents are less reliant on their children for old age support so they should decrease
their investment. They conclude that the income eﬀect is greater than the second eﬀect so they see
a rise in the investment in children’s education. This paper again does not look at the diﬀerential
investment by gender which is the focus of my paper. Another discussion paper, Chen (2015) looks
at the eﬀect of old age security on inter generational living arrangements. In the paper the author
ﬁnds that as old age security increases for the parents, they are less likely to live with their children.

79

2.4 Model

I construct a very simple two period model to explain how dependency on children for post
retirement consumption brings in gender gap in investments.The model is constructed under the
assumption that there is no taste based gender discrimination by the parents when it comes to
investing in their children. The purpose of this assumption is to highlight the fact that even in
the absence of taste based gender discrimination, parents invest deferentially in their children by
gender.

I consider a simple two period model, were parents derive utility from their consumption in
period one and period two and they also derive utility from their children’s well being in period
two. It is a time separable concave utility function. In period one parents earn income Y1, they
choose how much to consume C1, how much to invest in their boy child SB, how much to invest in
their girl child SG and how much to invest in assets A. In period two parents consume C2, earn lump
sum pension income T , they get return from their son RBSB, they get return from their daughter
RGSG and the return from the asset RAA. For interior solution we assume that return from assets
is higher than return from children6.

I prove three propositions :
Proposition 1 : If the cumulative return from investing in a boy child is higher than the
cumulative return from investing in the girl child then parents invest more in the boy child
compared to the girl child.

Proposition 2 : As post retirement income increases the gender gap in investment decreases.
Proposition 3 : For household with children of one gender as pension income T decreases
investment in children increases if the cumulative return to investing in children is higher than
that of investing in the asset. The increase/decrease in investment is proportional to the diﬀerence
in return between the asset and the children

6If return to assets is lower than return to children then parents would not invest in assets as

investing in children produce greater returns and added altruistic utility

80

2.4.1 Maximization Problem for Households with Children of Both Gender

Agent maximizes

subject to:

{C1,C2,SB,SG,A} U(C1) + β{V(SB) + V(SG) + U(C2)}

max

C1 = Y1 − SB − SG − A

C2 = T + RBSB + RGSG + RAA

Combining the FOC w.r.t SG and SB we get:

V(cid:48)(SB) − V(cid:48)(SG)

U(cid:48)(C2)

= RG − RB

(2.1)

(2.2)

(2.3)

(2.4)

From the existing literature we know that RG < RB (Jayachandran 2014, Das Gupta et al. 2003,
Kishor 1993), i.e. the return and the dependency are higher for boys. This implies V(cid:48)
G(SG).
Since the utility functions are concave, we get SB > SG.

B(SB) < V(cid:48)

Proposition 1 : Proved
From the FOC w.r.t RA we get:

U(cid:48)(C1) = βRAU(cid:48)(C2)

Diﬀerentiating (5) w.r.t to Pension Income T we get:

βRAU(cid:48)(cid:48)(C2)
U(cid:48)(cid:48)(C1)

=

> 0

dC1
dT
dC2
dT

(2.5)

(2.6)

When pension income changes consumption in both period moves in the same direction. As

consumption in both period cannot decrease with an increase in income we get

dC2
dT > 0
Now diﬀerentiating (4) w.r.t T and using (6) we get:

V(cid:48)(cid:48)(SB) dSB
dT

− V(cid:48)(cid:48)(SG) dSG
dT

= U(cid:48)(cid:48)(C2) dC2
dT

(RG − RB) > 0

81

(2.7)

(2.8)

Under the Assumption that V() is concave we get:
V(cid:48)(cid:48)(SG)
V(cid:48)(cid:48)(SB)

dSB
dT <

dSG
dT

= k

dSG
dT

Where k is less than 1 as:

If

We get

V(cid:48)(cid:48)(SB) > V(cid:48)(cid:48)(SG)

dSG
dT

⊂ (−∞,

1
1 − k

]

dSB
dT <

dSG
dT

Proposition 2 : Proved under the assumption that dSG

dT ⊂ (−∞, 1
1−k]

2.4.2 Maximization Problem for Households with Children of one Gender

Agent maximizes

Where

subject to:

max

{C1,C2,SI,A} U(C1) + β{V(SI) + U(C2)}

I ⊂ (B, G)

C1 = Y1 − SI − A

C2 = T + RI SI + RAA

Combining the F.O.C w.r.t SI & A

βV(cid:48)(SI) = U(cid:48)(C2)(RA − RI)

Diﬀerentiating w.r.t T

dSI
dT
From the above equation we get

U(cid:48)(cid:48)(C2)
V(cid:48)(cid:48)(SI)

1
β

dC2
dT

=

(RA − RI)

82

(2.9)

(2.10)

(2.11)

(2.12)

(2.13)

(2.14)

(2.15)

(2.16)

(2.17)

We know that, U(cid:48)(cid:48)(C2) > 0 & V(cid:48)(cid:48)(SI) > 0 & dC2
If RA < RI

dT > 0

If RA > RI

And

dSI
dT < 0

dSI
dT > 0

dSI
dT

∝ (RA − RI)

(2.18)

(2.19)

(2.20)

This implies that the decrease or increase in investment in children is proportional to the
dT ∝ (RA − RI) the
diﬀerence between the return to asset and children. Since RB > RG, and dSI
decrease in investment would be higher for the girl child in an all girl household compared to a boy
in a all boys household.

Proposition 3 : Proved
Thus with a decrease in pension we should see an increase in the Intra-household gender gap
between all girls and all boys household if the return to the boy child is greater than that of the girl
child.

2.5 Empirical Framework

2.5.1 Main Identiﬁcation Strategy

The empirical strategy hinges on a diﬀerential decrease in retirement beneﬁts across the private and
the public sector employees. We use this variation in diﬀerence in diﬀerence in diﬀerence (DDD)
strategy. The ﬁrst diﬀerence is over time as public sector employees joining the sector post 2004 are
exposed to a diﬀerent retirement beneﬁt scheme compared to the people joining prior to 2004. The
second diﬀerence is over the work sector of the father of the child namely private versus public, this
diﬀerence will account for other time trends which aﬀects both the sector. The third diﬀerence is

83

among the child gender to capture the diﬀerential eﬀect of the retirement beneﬁt change by gender.
The net diﬀerential eﬀect by gender of the pension contraction for the parents in the public sector
on various child outcomes is estimated by the following regression equation.

Yi = β1Sectori + β2Posti + β3Genderi
+ β4Sectori × Posti + β5Sectori × Genderi + β6Posti × Genderi
+ β7Posti × Genderi × Sectori + γXi + i

(2.21)

Sector is a binary variable which takes value one if the parent works in the public sector and is
zero if the parent work in the private sector. Post is a binary variable which takes value one if the
parent joined the current sector post 2004 and is zero if the parent joined the current sector prior
to 2004. Gender is a binary variable for the sex of the child, zero for a boy and one for a girl. We
include a full set of interaction terms in the regression equation namely sector-post, sector-gender
and post-gender. The gender diﬀerential treatment eﬀect of the pension reform is captured by β7.
We hypothesize to see a negative a β7 which would imply that as the retirement security decreases
girls are worse oﬀ related to boys. The main identifying assumption is that the evolution of gender
gap in investment in child’s education, over time follows the same trend for the public and private
sector employees. In the above baseline speciﬁcation we are not explicitly using within household
variation i.e we are comparing across all household within children7. In section 7.3 we include
household ﬁxed eﬀects, this helps us compare inter versus intra-household eﬀect.

We also include a rich set of individual, father, mother and household speciﬁc control variables
in the above equation. Child characteristics includes age, grade and other variables depending on
the outcomes we are looking at. Father characteristics include age, salary8 and years of education9.
7Hence the sample consists of households with only boys, only girls and both boys and girls.
This is more representative of the population rather than just looking at household with both boys
and girls. Also this gives us a bigger sample size to work with

8One might argue that this is endogenous to the policy change, since I am interested in the
eﬀects of the policy that work through the dependence on children and not current income, I control
for current income. For the same reason I also control for mother’s work status.

9We are unable to control for father’s industry here as the measure for industry is not consistent

across the two panels at a micro level.

84

Mother characteristics include age, years of education and work status10. House hold characteristics
include caste, income, poverty level, highest education level for adult male and female, number of
children and district.

We analyze two diﬀerent set of outcomes, ﬁrst set contains outcome variables which are direct
indicators of investment in the child. These variables include the type of school the child is enrolled
in private or public school, medium of instruction at school. The second set of variables are more
derivative as they reﬂect the returns to the investment in a child. These variables include English
speaking ability of the child, mathematics ability, reading ability and writing ability.

2.5.2 Common Trend and Falsiﬁcation

The identifying assumption for the validity of our results is that the evolution of gender gap in
investment in child’s education, over time follows the same trend for the public and private sector
employees. This might not hold if gender attitudes evolve diﬀerently for the employees of the
two sector. Besides economic returns there are other social and cultural norms that make parents
discriminate against their female child if such taste based gender attitudes evolve diﬀerentially over
time for the two sectors then that will be a threat to the validity of the identifying assumption.
To show that this is not the case I will provide two arguments: ﬁrst the fathers selected in the
sample in private and public sector have a considerable amount of overlap when it comes to the
industry11 they are employed in and second we will use a diﬀerence in diﬀerence strategy to show
that gender attitudes does not evolve diﬀerentially across the two sector. We will use the data on
gender relation question asked to an ever married woman in the house hold, these gender relation
questions capture the gender attitudes of the household. I will use a diﬀerence in diﬀerence strategy
to see that if these gender attitudes have changed diﬀerentially for both the sectors over time. The
questions we speciﬁcally look at are the following: is the women frequently beaten, is she allowed
to visit her natal family, does she have her name on important rental documents , does she have
10I do not have data for years of work experience for mother and father but it is usually collinear

with age and years of education.

11Table 4 shows this.

85

to cover her face and does she have to take permission to go out of the house.These questions
capture the attitude of the household towards women and would capture the non-economic part
of discrimination against women. Another threat to the identiﬁcation strategy is if selection into
private versus public sector changes as a result of the policy change. To show that the people
entering the sectors post policy change are not diﬀerent I estimate equation 22 with background
individual and family characteristics as outcomes

Gi = β1Sectori + β2Posti + β3Sectori × Posti + γXi + i

(2.22)

Gi are variables that capture gender attitudes of household, these gender relation question were
asked to an ever married woman. We check that whether the time trend in gender attitudes vary
diﬀerentially across the employees of the two sector. Here sector represents her husbands work
sector. A β3 signiﬁcantly not diﬀerent from zero would imply that the time trends in gender attitudes
are similar for the two sector. We want to see if the taste based discrimination towards women have
changed diﬀerentially for the households of the current student.

I also run a falsiﬁcation test to make sure that what we are picking up is not some residual
time trend eﬀect. i set treat equal to one for private sector and treat equal to zero for agricultural
sector.
I run the baseline speciﬁcation (equation 21) between the private sector employees and
agricultural sector employees instead of public versus private. Since the reform was for the public
sector employees we should not see any signiﬁcant coeﬃcient when we change the sectors.

2.6 Data description

The data I use for this paper is panel data from India Human Development Survey Data (IHDS),
Waves I and II, which are two nationally representative surveys conducted in 2005 and 2011. The
data set is very rich as it contains data on household and individual level characteristics. The
richness of the data regarding gender speciﬁc questions makes it ideal for our study. We will have
two samples of interest for whom we looks various outcomes. The purpose of this is to better
identify the causality between old age security and child investments at various stages for various
outcomes. The ﬁrst sample consists of all current students who are enrolled in school or some

86

sort of degree program and are less than 18 years old. Table II presents the descriptive statistics
for this sample. The outcome variables we look at for this sample is type of school public or
private, medium of instruction in school English or regional language and the English ability12 of
the child. I present the descriptive statistics decomposed by parent’s sector of work and whether
they joined the sector post or prior to 2004. We see that the public sector households does better
than the private sector households, they are more educated, wealthier and belongs to higher caste.
This trend is common across all three samples. Compared to the private sector household we see
that the children in the public sector households are more likely to be enrolled in private schools
or schools where the medium of instruction is English. When it comes to comparing between the
people who joined post and prior to 2004 we see that the prior people are in general a little older
but other outcomes are comparable across the two sectors.

The second sample consists of children age 8-11, IHDS administers reading, writing and math
tests to this group of children. Table II presents the descriptive statistics for this sample. The
outcome variable we look at for this sample is test score for mathematics, reading and writing.
The scores are standardized13 at a district level. We also construct a variable Score which is the
summation of the three scores. Again I present the descriptive statistics decomposed by parent’s
sector of work and whether they joined the sector post or prior to 2004.

For both the sample the fathers work in either the public or the private sector. I further restrict
the sample to fathers who work in white collar jobs.
I identiﬁed jobs as white collar jobs 14
and then segregated them into private and public sector. This was done to ensure that the time
trends are comparable across the two sectors. Table IV presents the distribution of percentage of
people employed in various industries across the two sectors, the list is not exhaustive. We see a
considerable overlap of industry across the two sectors, private and public.

For analyzing common trends in the two sector I use the data on eligible woman from the IHDS
12English ability is a categorical variable taking the value 0 if the child has no knowledge of the

language, 1 if the child has some knowledge and 2 if the child is ﬂuent in English.

13I compute the z score at a district level by child age
14see appendix for the full list

87

survey. In the data gender relation questions were asked to a married woman in the household.
These question captures the gender attitudes of the household toward females. Questions such as
is the women frequently beaten, is she allowed to visit her natal family, does she have her name
on important rental documents and does she have to take permission to go out of the house. The
answer to these questions reﬂect the patriarchal nature of the household.

Appendix table 6 presents the means of the outcome variables by the three diﬀerence groups
public versus private, post reform versus pre reform and boy versus girl. These could be read as
the raw DDD results.

2.7 Results

2.7.1 Current Students

Table V15 illustrates the overall gender diﬀerential eﬀect of contraction of post-retirement secu-
rity. First we look at the sample of current students. We are looking at three outcomes for this
sample, whether the child is enrolled in a public schools or a private school, whether the medium
of instruction is English or a regional language and the English ability of the child. Speciﬁcation
one contains full set of controls including father characteristics, mother characteristics, household
characteristics and survey year dummy. Speciﬁcation two does not contain household characteris-
tics, the purpose of the two speciﬁcations is to show that the estimates does not vary a lot under
diﬀerent speciﬁcations. Under speciﬁcation one which contains the full set of controls we see
that compared to a boy child a girl child is 0.070116 points less likely to be enrolled in a private
15Table V presents the DDD estimates from equation 21. The top panel presents the results from
the current student sample and the bottom panel presents the results from the test scores. Each
row presents the result from equation 21 with a particular outcome variable. For example row 1
presents the DDD estimate with enrollment in private school as the outcome variable. Column 1
presents the results from speciﬁcation 1 which includes the full set of controls: child, parent and
household characteristics and column 2 presents the result from speciﬁcation 2 which does not
control for household characteristics.

16The pre-reform means for private school enrollment are 0.366 and 0.276 for boys and girl
respectively. If we look at the DD separately for boys and girls we see that the 7pp increase in
gender gap is driven by 1.62 pp decrease in private school enrollment for the girl and 5.61 percentage
point increase in private school enrollment for the boys. I think these are reasonable compared to

88

school post reform in the public sector, the eﬀect is signiﬁcant. This means that as post retirement
security is decreasing the gender gap in enrollments in private schools increases compared to public
schools. We see that the estimates do not vary much under the second speciﬁcation which does not
include household controls. The next outcome is whether the child is enrolled in a school where
the medium of instruction is English or a regional language. Under speciﬁcation one we see that
compared to a boy child a girl child is 0.0684 points less likely to be enrolled in a school where the
medium of instruction is English, post reform in the public sector. The result is signiﬁcant and is in
line with our hypothesis. The estimate does not vary much under the second speciﬁcation. Finally
we look at the eﬀect on the English ability17 of the child we see that compared to a boy child a girl
child is 0.0684 points less ﬂuent in conversing in English. The eﬀect is not signiﬁcant but in line
with our hypothesis.

We report the Diﬀerence-in-Diﬀerence for the boy child and girl child separately in Appendix
table A1. We see that the decrease in retirement security for the parents increases the investment in
the boy child whereas it decreases the investment in the girl child. The eﬀects are not signiﬁcant,
but the overall eﬀect on the gender gap is signiﬁcant.

2.7.2 Test Scores

For the second sample we look at standardized test scores administered by IHDS. The results are
presented in Table V. Under speciﬁcation one, we see that compared to a boy child a girl child
scores 0.07 points less in math post reform in the public sector , the estimate is not signiﬁcant. This
implies that compared to boy child a girl child performs worse post reform, thus as old age security
decreases the performance of the girl child falls compared to the boy child which in indicative of
increasing gender gap in human capital investment. We see similar results for writing and SCORE
which is the summation of all the scores. None of the results are signiﬁcant but the signs are
consistent with our hypothesis.
the pre-reform means.

17English Ability is reported by the parent or grandparent so there might be some non-standard

measurement error, the sibling ﬁxed eﬀect model will account for this

89

Overall we see that as post retirement security decrease girl child are getting worse of compared
to the boy child. They are less likely to be enrolled in good private school compared to boys or in
school where the medium of instruction is English. They also perform worse in exams compared
to boys and have lower scores in mathematics and writing, they are also less likely to be speaking
in English.

2.7.3

Inter-Household Vs Intra-Household

Appendix table A2 reports the results from our DDD estimation (equation 21) including household
ﬁxed eﬀect. We see that the signiﬁcant eﬀect we were getting in our triple diﬀerence estimates
vanishes after accounting for sibling ﬁxed eﬀect. To analyze this we divide our sample into families
with children of only one gender and household with children of both gender. Table VI reports
the results from running the triple diﬀerence separately for the two samples. We see that the rise
in gender gap is primarily driven by families with children of one gender. We further divide the
sample of one gender household into families with one child and families with more than one child.
We see that the eﬀect is primarily driven by families with more than one child and child of only
one gender.

To further deepen our understanding of why the results are being driven by them we look at
the characteristics of the two kinds of households. Appendix Table A8 presents the characteristics
of household with children of one gender versus household with children of multiple gender. We
see that household with children of only one gender looks similar to household with children of
multiple gender in terms of age, education and wealth.

There could be number of reasons why we see this, in the theoretical model in proposition 3 we
saw that for families with children of only one gender, as retirement security decreases there will
be an increase in investment in children if the return from investing in a child is greater than that of
investing in an asset. If the return from the asset lies between the return from a boy child and the
return from a girl child (RB > RA > RG) then that explains the ﬁnding. Proposition 2 states that
for household with both girl child and boy child gender gap increases as pension income decreases

90

but proposition 3 holds under less stricter assumption than proposition 2. Proposition 2 holds if
dSG/dT ⊂ (−∞, 1
1−k whereas proportion 3 holds for all values of SG. This might explain why we
see a inter-household eﬀect rather and not a intra-household eﬀect.

Other possible reason could be that parents of families with only girl children are more worried
about their retirement security and they shift resources from their girl children into other assets.
In appendix table A7 I present the result from running a DD with sector(public versus private)
and reform (post 2004 versus pre) to look at the eﬀect of the policy change on investment in other
assets18 we see that families with only girl child invest more in retirement saving and other long
term savings instrument compared to families with only boys but the eﬀect is not signiﬁcant.

If we look at table A6 we see that for both pre and post reform and public and private sector,
the girl child on an average is less likely to be send to a private school. In table A8 we look at
private school enrollment for a girl child and boy child by family composition. For the boy child we
see that the average private school enrollment is similar for household with one child, household
with children of all gender (all boys household) and household with children of all gender (mixed).
Whereas for the girl child we see that the private school enrollment varies by family composition.
Girls from household with only one child on average has the highest private school enrollment.
Followed by girls from household with children of both gender (mixed). Finally the girls from
household with children of one gender (all girls household) has the lowest average enrollment.
Hence in the entire population, irrespective of the policy reform and working sector of the parents,
we see that girls from all girls household has the lowest private school enrollment whereas for boys
the sibling composition does not matter. This supports the ﬁnding that the impact from the policy
reform were mainly driven by household who has only children of one gender. The girls from the
all-girls household in general receive less investment compared to girls in mixed household, the
policy reform might further magnify this discrimination.

18A binary variable indicating whether the household has invested in some other pension plans

or LIC policy

91

2.8 Common Trends and Falsiﬁcation Test

2.8.1 Trends in Gender Attitude

In table VI I present the results from equation twenty two. Since the people joining the sectors
post 2004 are relatively younger than people joining prior to 2004 the gender relation might evolve
diﬀerentially over time for the two sector. The ﬁrst outcome I look at is whether the woman is made
to practice “ghunghat” or “purdah”, this is an old patriarchal Indian tradition where the married
woman is made to cover her face at all times. We see that the coeﬃcient on the interaction term
is 0.0011, and it is not signiﬁcant. Since the coeﬃcient is not signiﬁcantly diﬀerent from zero it
implies that the trend does not vary diﬀerentially for the two groups. The next outcome we look
at is if the women is beaten frequently by her husband, the coeﬃcient on the interaction term is
-0.0164 and is not signiﬁcant. We ﬁnd very similar results for the other indicators of gender attitude
all of them have small coeﬃcients which are not signiﬁcantly diﬀerent from zero. Overall we can
summarize that the gender attitudes does not evolve diﬀerentially over time for the two sectors.

2.8.2 Falsiﬁcation Test

The results are presented in the lower panel of table VI. To further establish that the eﬀects that we
are picking up in our main estimation are not some spurious eﬀect driven by the private sector, we
look at the outcomes for current students across the private19 and agricultural sector. We see that
the coeﬃcients are smaller, of the opposite sign and insigniﬁcant.

2.8.3 Selection into Private and Public Sector

Another threat to the identiﬁcation strategy is if selection into public versus private sector changes
as a result of the policy change. To see if people entering the public sector versus private sector post
policy change are diﬀerent from the people who joined prior to the policy change I use a diﬀerence
in diﬀerence model across Sector ( Public versus Private) and Post (joined post or prior to policy

19The sector dummy takes the value 1 for private and 0 for agriculture

92

change). The outcome variables I look at are individual characteristics and family characteristics
such as education, father’s education, caste, religion, urban or rural, district and below poverty line.
Table VII reports the results from the diﬀerence in diﬀerence. I see that these characteristics do not
vary between the people who entered the two sector over time. Variables such as fertility, individual
income, household income, assets and mother’s income might be aﬀected by the policy change and
they in turn might aﬀect gender gap. To account for this I control for them in speciﬁcation1 for
equation 21.

2.9 Discussion and Conclusion

We see that the reform of 2004 for the public sector employees is associated with worse outcomes
for the girl child compared to the boy child. We established that the 2004 reform was viewed as
a reduction in post-retirement security by the public sector employees, which implies that when
the perceived retirement security decreases, gender gap in various outcomes increases. As we
have seen this change in gender gap is not driven by change in gender attitudes as gender attitudes
remain the same for the two sets of household. This would further imply that the eﬀect we see is
solely driven by economic incentives rather than taste based change in gender attitudes. What we
are capturing is the eﬀect of the contraction of post-retirement beneﬁt and not some interaction
eﬀect between taste and incentive. This result has good implication for the future, if people react
to economic incentives to change their behavior towards gender then with the right incentives the
gender gap between male and female could be reduced in India. The New Pension System has also
extended its coverage to the private sector and unorganized sector in the recent years. The take up
is slow but it is expected to expand the safety net. This will give us another chance to look into this
topic in the future.

93

Table 2.1: Pension System in India

94

Table 2.2: Descriptive Statistics for Current Students

95

Table 2.3: Descriptive Statistics for the Test Score Sample

96

Table 2.4: Distribution of Industry Across the Public and Private Sector

97

Table 2.5: Main Results

*Table V contains estimates of β7 from equation (8) for various outcome variables.
*The top panel presents the results from the current student sample and the bottom panel presents
the results from the test scores.
*Each row presents the DDD estimate from equation 21 with a particular outcome variable.
*Standard errors are reported in brackets and are clustered at the district level.
*Speciﬁcation (1) includes full set of controls : child characteristics, father characteristics, mother
characteristics and household characteristics.
*Speciﬁcation (2) does not include household characteristics.

98

Table 2.6: Inter versus Intra-Household

*Table V contains estimates of β7 from equation (8) for various outcome variables.
*Each row presents the DDD estimate from equation 21 with a particular outcome variable.
*Standard errors are reported in brackets and are clustered at the district level.
*Column 1 of the top panel shows the DDD estimate from equation 21 for the entire sample.
*Column 2 of the top panel shows the DDD estimate from equation 21 for household which has
children of both gender.
*Column 3 of the top panel shows the DDD estimate from equation 21 for the household which
has only girl child or boy child.
*Column 1 of the bottom panel shows the DDD estimate from equation 21 for household which
has only one girl child or boy child.
*Column 2 of the bottom panel shows the DDD estimate from equation 21 for household which
has only more than one girl child or boy child.

99

Table 2.7: Trends & Falsiﬁcation

*Table VII top panel presents the estimates of β3 from equation (22) for various outcome variables.
*Each row presents the DD estimate from equation 22 with a particular outcome variable.
*Standard errors are reported in brackets and are clustered at the district level.
*Speciﬁcation one includes all women who were interviewed and is married to someone in the
private sector or public sector. Speciﬁcation one includes all women who were interviewed and is
married to someone in the private sector or public sector and has at least one child.
*Table VII bottom panel presents the estimates of β7 from equation (21) for various outcome
variables, where sector is equal to 1 for private sector employees and 0 for Agricultural sector
employees.

100

Table 2.8: Selection into Sector

*Table VIII presents the estimates of β3 from equation (22) for various outcome variables.
*Each row presents the DD estimate from equation 22 with a particular outcome variable.
*Standard errors are reported in brackets and are clustered at the district level.

101

Figure 2.1: Diﬀerence in Means

*Graph 1 plots the raw diﬀerence in means for the boy child versus girl child from the current
student sample.
*Fathers characteristics is used to assign sector and joining date.

()

102

CHAPTER 3

EFFECT OF RISK CORRELATION ON MONITORING AND AUDITING

3.1 Introduction

Micro ﬁnance has been looked upon by policy makers as a viable means of alleviating poverty
and providing credit to the poor in developing countries. Bangladesh, Brazil, India, Bolivia,
Uganda, Morocco and several other countries have developed such institutions of their own to
cater to the ﬁnancial needs of the poor. For the past forty years, micro ﬁnance institutions (also
known as MFIs) have developed mechanisms to maintain ﬁnancial sustainability so as to serve the
poor customers, who were deemed unproﬁtable by the commercial banks, by providing small loans
without collateral, selling insurance, all in an eﬀort to bring about poverty reduction.

Group loans have been the dominant model for lending in such institutions. There are numerous
papers in the literature that investigates the reasons for why group loans outperform individual loans
in terms of repayment. In this paper I am interested in looking at the eﬀect of risk correlation in
outcomes among group members, on various aspects of group lending. Risk correlation among
group members is an important factor for repayment as most forms of group lending rely on joint
liability for repayment. The intuition is that if one of the group members is unable to repay then
the other members pay for that member, ensuring future credit. Higher risk correlation increases
the probability of multiple group members failing to repay thus breaking the credit line.

Ahlin and Townsend (EJ 2007) is one of the ﬁrst paper to look at the eﬀect of risk correlation
on repayment in group lending. They introduced risk correlation in some of the most cited models
in the literature; Stiglitz (1990) and Banerjee et al (1994) both of which focuses on moral hazard
problem which joint liability lending can mitigate; Besley and Coate (1995) which focuses on
limited liability contracts which village sanctions can mitigate; Ghatak (1999) which focuses on
how adverse selection can be dealt with using joint liability contracts. When they introduced risk
correlation in the limited enforcement model Stiglitz (1990), they ﬁnd that higher risk correlation

103

lowers repayment from the borrower. Whereas in the Beasley and Coates (1995) and Ghatak (1999)
models they ﬁnd higher risk correlation increases repayment by the borrower. They conclude that
risk correlation might have counter intuitive positive eﬀect on repayment. Empirically they ﬁnd a
weak positive relation between risk correlation and repayment.

Since the overall relation between risk correlation and repayment is ambiguous, I look at the
eﬀect of risk correlation on more intermediary variables which inﬂuences repayment. For example
in the moral hazard model by Banerjee et al (1994) the adverse eﬀect of moral hazards are mitigated
through monitoring within the group members. So rather than directly looking at the eﬀect of risk
correlation on repayment, I look at the relationship between risk correlation and monitoring by
group members. In the costly state veriﬁcation model in Ghatak and Guinnane (1999), the high
auditing cost on the part of the lenders are mitigated by joint liability contracts.
In this model
I look at the relationship between risk correlation and probability/cost of auditing by the lender.
Looking at the relation between risk correlation and these intermediary variables will give us a
better understanding on how risk correlation aﬀects the mechanisms of group loans, as we are
unable to understand its eﬀects on more observable outcomes such as repayment.

The structure of the paper is very similar to Ahlin and Townsend (EJ 2007). I ﬁrst introduce
risk correlation to some of the most cited theoretical models in the existing literature. This gives us
a theoretical understanding of the eﬀects of risk correlation. Then I test these theories empirically
in a non-structural way using the Townsend Thai data base.

To look at the relationship between risk correlation and monitoring I introduce risk correlation
in the moral hazard model by Banerjee et al (1994).
I ﬁnd that the relationship between risk
correlation and monitoring depends on the probability of success of the borrower and cosigner1.
If the probability of success for the borrower is lower than that of the cosigner then monitoring
decreases as risk correlation increases.
If the probability of success for the borrower is higher
than that of the cosigner then monitoring increases as risk correlation increases. Hence from the
theoretical model the eﬀect is ambiguous. On the empirical side we see that the results are also
1The cosigner is the group member who is liable to repay the borrower’s loan if the borrower is

unable to repay

104

ambiguous. The empirical relation between risk correlation and monitoring also depends on the
proxies for risk correlation I am using. When I proxy for risk correlation using a village level
measure for risk correlation2 I ﬁnd that as risk correlation increases monitoring decreases whereas
when I use occupational homogeneity as a measure of risk correlation I ﬁnd that as risk correlation
increases monitoring decreases. These results are not causal rather correlations.

To look at the relation between risk correlation and audit3 I introduce risk correlation in the
costly state veriﬁcation model in Ghatak & Guinnane (1999). I ﬁnd that as risk correlation increases
the probability/cost of audit increases. Since in the model the lender only audits when all group
members fail to pay; a higher risk correlation increases the probability of that state of the world
occurring and hence increases the cost/probability of audit. Empirically I ﬁnd a negative correlation
between audit and risk correlation but the results are not statistically signiﬁcant.

The paper is structured as follows. Section 2 introduces the theoretical model. Section 3
introduces the data. Section 4 talks about the empirical strategy. Section 5 discusses the results.
Section 6 compares the theoretical and empirical results. Section 7 presents the conclusion.

3.2 Model

3.2.1 Moral Hazard; BBG

To look at the eﬀect of risk correlation on Monitoring I introduce risk correlation in the Moral
hazard model by Banerjee et al (1994). Moral hazard in this context is the temptation by borrowers
to gamble with risky projects when they have a limited liability. I use a modiﬁed version of the
original model as used in Ahlin and Townsend (EJ 2007).

In this model the groups are asymmetric; it consists of one borrower and one cosigner (monitor).
The borrower receives one unit of capital from the lender and chooses to invest in a project Pb.
Where Pb ∈ (Pb,1) and Pb >0. The project generates a return of Y(Pb) with probability Pb and 0
with probability (1-Pb). The lender collects a return r from the borrower if the project is successful.

2deﬁned in section 3
3When the lender has to verify the borrowers income, which is costly to do.

105

In the original model the lender collects q from the cosigner if the borrower fails to repay. This is
the point I slightly diverge from the original model. In the original model there was no uncertainty
in the cosigner’s ability to pay the liability q but in my model the cosigner pays the liability only if
he is also successful in his own project4. Hence in my model the lender collects q from the cosigner
if the cosigner’s project is successful and gets 0 otherwise. The cosigner is successful in his project
with Pm and fails with probability (1-Pm).

The risk neutral borrowers pay oﬀ from the project without any penalties5 is E(Pb)-Pbr; where
E(Pb)=Y(Pb)Pb. By assumption E(Pb) is increasing in Pb. It is also assumed that the interest rate
and loan size is such that given limilted liability the borrower would always prefer risky projects to
safer ones. So in the absence of a cosigner/monitor he will gamble with the riskiest project Pb=Pb.
I assume that;

0<E(cid:48)(Pb)<r and E(cid:48)(cid:48)(Pb)< 0

The assumption implies that the borrowers expected payoﬀ is increasing as riskiness of the project
increases but the expected output is decreasing in risk.

The cosigner/monitor can penalize the borrower based on his project choice. To impose a
penalty c on the borrower the monitor has to incur a cost M(c) where the function M() is increasing
and convex in c.

Mc() > 0; Mcc() > 0

If the monitor wants to enforce a project Pb then he should penalise the borrower in such a way
that the borrower’s gains from deviating to the riskiest project is exactly equal to the gains from
choosing the project Pb.

c(Pb,r) = E(Pb) − Pbr − (E(Pb) − Pbr)
4This is done to introduce risk correlation among group members
5Punishment from the monitor

106

The monitor then chooses Pb to maximize his payoﬀ; which includes paying the joint liability
fee q when the borrower is unsuccessful and the monitor is successful in his project. It also includes
the cost of imposing a penalty c, which the monitor has to pay in all states of the world.

((1 − Pb)(Pm) − )(−q) − M(c)

 is the correlation in project outcomes between the borrower and cosigner. I introduce risk
correlation in exactly the same way as Ahlin and Townsend (JoE 2007). They introduce risk
correlation in a unique way to preserve the individual probabilities of success for the borrower and
monitor.

Monitor Succeeds

Borrower Succeeds

Borrower Fails

PbPm+
(1−Pb)Pm−

Monitor Fails
Pb(1−Pm)−
(1−Pb)(1−Pm)+

 = 0 is the case where there is no correlation in the projects outcomes of the borrower and
monitor,  > 0 is the positive correlation case which I focus on in this paper. In general  may
depend on Pb and Pm. Since no general structure is available Ahlin and Townsend (JoE 2007)
assume two speciﬁcations.

Speciﬁcation I : The constant correlation case.

(Pb, Pm) = 

Speciﬁcation II :

(Pb, Pm) = µ × min{Pb(1 − Pm), Pm(1 − Pb)}

Under speciﬁcation II if the groups are homogeneous then the correlation coeﬃcient over
project returns is equal to µ. For non-homogeneous groups, that is for whom Pb (cid:44) Pm, µ
is not the correlation coeﬃcient, but something closely related:
it is the correlation, expressed
as a fraction of the maximum correlation possible given individual probabilities of Pb and Pm.
When Pb = Pm, the maximum possible correlation coeﬃcient is one, so µ equals the correlation

107

coeﬃcient. This formulation is the unique way of aﬀecting each potential group’s (appropriately
normalized) correlation coeﬃcient symmetrically. I will operate under speciﬁcation II in this paper.

The monitor chooses the project Pb to maximize,

Where,

((1 − Pb)(Pm) − )(−q) − M(c)

c(Pb,r) = E(Pb) − Pbr − (E(Pb) − Pbr)

Now I am interested in ﬁnding the eﬀect of risk correlation on monitoring dM/dµ; the risk

correlation structure under speciﬁcation II takes two forms given the value of Pb and Pm.

Case I,

Case II,

3.2.1.1 Case I

Pb < Pm =⇒ (Pb, Pm) = µ × Pb(1 − Pm)

Pb > Pm =⇒ (Pb, Pm) = µ × Pm(1 − Pb)

First I look at the eﬀect of risk correlation on monitoring under case I.

The ﬁrst order condition (F.O.C) w.r.t Pb;

q(Pm + µ(1 − Pm)) = MC(c)(r − E(cid:48)(Pb) (2.1.1)

The lender’s break even condition is given by;

Pbr + ((1 − Pb)Pm − µ × Pb(1 − Pm))q = ρ (2.1.2)

Where ρ is the outside option for the lender. To analyze the eﬀect of risk correlation (µ) on
monitoring (M(Pb,r)) we will use the ﬁrst order condition. Here we are assuming that the lender

108

is not changing the liability q or the interest rate r to account for the change in risk correlation. 6
Since M() is a function of PB I ﬁrst ﬁnd the sign of dPb
dµ . To look at the eﬀect of risk correlation on
Pb I diﬀerentiate the F.O.C wrt µ.

The F.O.C when diﬀerentiated w.r.t µ gives;

dPb
dµ

(McE(cid:48)(cid:48)(Pb) − Mcc(r − E(cid:48)(Pb))2) = −q(1 − Pm) (2.1.3)

Rearranging equation 2.1.3 we get;

= −q(1−Pm)

A

dPb
dµ

> 0

(McE(cid:48)(cid:48)(Pb) − Mcc(r − E(cid:48)(Pb))2) = A

A < 0 as M() is convex

−q(1 − Pm) < 0,

Where;

Hence,

dPb
dµ > 0

Now we look at the eﬀect of risk correlation on monitoring;

= (r − E(cid:48)(Pb)) dPb
dµ > 0
dc

= Mc

dµ > 0

dc
dµ

dM
dµ

Thus if Pb < Pm then monitoring increases as risk correlation increases.
6Since we are working with cross sectional data where liability and interest doesn’t change, this

assumption ﬁts.

109

3.2.1.2 Case II

Pb > Pm =⇒ (Pb, Pm) = µ × Pm(1 − Pb)

The ﬁrst order condition (F.O.C) w.r.t Pb;

qPm(1 − µ) = MC(c)(r − E(cid:48)(Pb)) (2.2.1)

The lender’s break even condition is given by;

Pbr + (1 − Pb)Pm(1 − µ)q = ρ (2.2.2)

Again we will use the ﬁrst order condition. Since M() is a function of PB I ﬁrst ﬁnd the sign of

dPb
dµ . To look at the eﬀect of risk correlation on Pb I diﬀerentiate the F.O.C wrt µ.

The F.O.C when diﬀerentiated w.r.t µ gives;

dPb
dµ

A

Pm(1−µ) = q(1−µ) (2.2.3)

From equation 2.2.3,

Hence,

This implies,

and,

dPb
dµ

= qPm
A

dPb
dµ < 0

= (r − E(cid:48)(Pb)) dPb

dµ < 0

dc
dµ

dM
dµ

= Mc

dµ < 0
dc

110

If Pb > Pm then monitoring decreases as risk correlation increases.
We see that the eﬀect of risk correlation on monitoring depends on the values of Pb and Pm.
If Pb < Pm then monitoring increases as risk correlation increases. Whereas if Pb > Pm then
monitoring decreases as risk correlation increases. The result is driven by the structure of risk
correlation used in the model. Hence form the theoretical model we ﬁnd that the eﬀect of risk
correlation on monitoring is ambiguous.

3.2.2 Costly state Veriﬁcation; Maitreesh Ghatak , Timothy W. Guinnane 2009

In this section I want to look at eﬀect of risk correlation on the probability and hence the cost
of audit. I will introduce risk correlation in the model of State Veriﬁcation used in (Maitreesh
Ghatak et al, 1999). Formal lenders sometimes cannot lend to poor borrowers because such lenders
cannot cost eﬀectively verify the borrowers state. In such a scenario joint liability contracts reduce
expected audit cost. The intuition is that suppose group members have a lower cost of verifying
each other and they are liable for each other’s loan then the bank has to verify only when the
entire group defaults. In their paper they introduce a small theoretical model to show how joint
liability contracts improve eﬃciency by lowering audit costs. I will use the same model to show
that introducing risk correlation into the model decreases the gains from moving to joint liability
contract ie it increases the audit cost.

There are two borrowers and one lender. There are two outcomes from the borrowers {YH,0}
with probability (p,(1− p))and(p(cid:48),(1− p(cid:48))). The lender has to incur a cost γ to verify the individual
borrowers output, whereas the borrowers can do it costlessly. Each borrower has an exogenous
outside option U. Assume that the borrowers can write side-contracts with each other costlessly and
that there is no cost for a borrower to observer her partner’s project returns. There are no problems
of moral hazard, adverse selection or enforcement of contracts. The ﬁnancial contract speciﬁes the
interest rate r and the probability of an audit λ when both the members default.

The Optimal contract is given by;

max pp’(YH − r) + p(1 − p(cid:48))(YH − 2r) − U

111

The truth telling constrain is given by7 ;

Y H − 2r ≥ max{0,(1 − λ)Y H}

The break even constrain is given by8;

PP(cid:48)r + P(1 − P(cid:48))r + (1 − P)P(cid:48)r − λ(1 − P)(1 − P(cid:48))γ = ρ

Since p,YH and U are exogenously given the optimal contract is given by r and λ which satisﬁes

both the constrain. The equilibrium value of λ and r is given by;

r = λY H
2

λ =

Y H

2 (PP(cid:48)+P(1−P(cid:48))+(1−P)P(cid:48)−(1−P)(1−P(cid:48))γ

ρ

Now, I add risk correlation among borrower outcomes in the model. I will show the proof under

speciﬁcation I where (P, P(cid:48)) =  but the proof can be easily extended for speciﬁcation II.

Let the new equilibrium probability of audit and interest rate be λ(cid:48) and r(cid:48).
The truth telling constrain remains the same after adding risk correlation and is given by;

Y H − 2r(cid:48) ≥ max0,(1 − λ(cid:48))Y H

The break even constrain changes, the new break even condition is given by,
(PP(cid:48) + )r + (P(1 − P(cid:48)) − )r + ((1 − P)P(cid:48) − )r − λ(cid:48)((1 − P)(1 − P(cid:48)) + )γ = ρ

The new equilibrium value of r and λ is given by,
r(cid:48) = λ(cid:48)Y H
2

λ(cid:48) =

Y H

2 (PP(cid:48)+P(1−P(cid:48))+(1−P)P(cid:48)−(1−P)(1−P(cid:48))γ−( Y H

2 +γ))

ρ

Now we compare the equilibrium value of r and λ,
7This constraint ensures that it is more proﬁtable for the lender to reveal his true state.
8This ensures that the lender is operating at zero proﬁt

112

λ(cid:48) − λ = (Y H/2+λ)
> 0
r(cid:48) − r = Y H/2(λ(cid:48) − λ) > 0

Z

As long as  > 0, λ(cid:48) > λ and r(cid:48) > r. Also λ(cid:48) − λ increases as  increases. Thus we see
introducing risk correlation increases the probability/cost of monitoring. The increase in cost
increases in . This result can readily be extended to speciﬁcation II of , as long as  > 0 the
above results hold.

3.3 Data

The data used in this project comes from the Townsend Thai data base. To be more precise the
data comes from a large cross section survey conducted in May of 1997 covering 192 villages in two
regions of Thailand. The regions covered in the survey includes the central and northeast region.
Each of these regions in Thailand consists of provinces and the provinces are further divided into
sub-counties. To conduct the survey ﬁrst sub-counties were selected in a province and then the
survey was conducted on a cluster of four villages within the sub-counties. Borrowing groups
in the villages belonging to the Bank of Agriculture and Agricultural Cooperatives (BAAC) were
interviewed. The BAAC is a government operated development bank in Thailand9. Within the
selected villages they interviewed as many groups as possible up to two. In the data there are 262
groups, 62 of which are the only groups in their village. Each group had a designated leader who
responded to the questions on behalf of the group. The analyses is at the group level.

The dependent variable of interest are measures of risk correlation. Following Ahlin and
Townsend (EJ 2007) I proxy for risk correlation by two measures of correlation; Occupational
homogeneity within a group and a village-level measure taken from the household survey. Oc-
cupational Homogeneity is taken from the BAAC survey and equals to the probability of two
randomly chosen group members to have the same occupation. For the village level measure I use
9The BAAC requires some kind of collateral for all loans but it allows smaller loans to be backed
with social collateral in the form of joint liability. The nature of the BAAC loans makes it ideal for
me to test my theoretical prediction on.

113

the household survey where villagers were asked which of the previous ﬁve years were the best and
worst for income, respectively. The variable is constructed as the probability that two randomly
selected respondents from the same village reported the same year as worst. Table 1 summarizes
these variables.

The independent variable of interest includes measures of monitoring and auditing. For mea-

sures of monitoring within the group I use four questions asked in the BAAC instrument

i) Leader Monitoring 1 : In a typical week how many group members does the leader talks to
more than ﬁve times; ii) Leader Monitoring 2 : Number of people in the group the leader talks
to less than once a week; iii) Peer Monitoring : Are the group members aware of the work of the
other members; iv) Meetings : How many meetings did the group had in the past month. These
variables are reported in Table 2.

I use two measures of auditing i) Verify Income : Does anyone veriﬁes the members income ii)
Oﬃcer visit: How many times does the BAAC oﬃcer visits. The summary of these variables are
reported in Table 2.

I also include group and village level control. I control for group size and group age. I control
for village level risk, percent in village claiming production credit group membership and percent
in village claiming to be clients of commercial banks.

3.4 Identiﬁcation Strategy

To empirically see whether risk correlation aﬀects monitoring within a group and audit cost I

use a simple OLS framework (Equation1).

Yi = α + βXi + γGroupCharacteristicsi + δVillageCharacteristicsi + i (1)

The coeﬃcient of interests is β, which is the coeﬃcient on the two measures of risk correlation
described in section 3 and the dependent variables include various measures monitoring and audit
costs. The results are not causal rather correlations. I control for group and village characteristics.

114

I partial out the eﬀect of group size and group age by controlling for them as they are usually strong
predictors of group behavior. I also run the OLS including region ﬁxed eﬀects.

3.5 Results

In this section I present the results from equation 1, the regressions results are estimated for both
proxies of risk correlation; occupational homogeneity and village level measure of risk correlation.
I also presents the results from including region ﬁxed eﬀects.

First, I look at the eﬀect of risk correlation on monitoring. Table 3, panel A presents the
eﬀect of occupational homogeneity on how many group members the leader talk to more than
ﬁve times in a week (Leader Monitoring 1). The more number of group members the leader
talks to more than ﬁve times a week implies greater monitoring by the leader on the group
members. We see that as occupational homogeneity increases leader monitoring 1 increases and
the eﬀect is statistically signiﬁcant. This implies that even after controlling for group and village
characteristics, occupational homogeneity among the group members is positively correlated with
leader monitoring. When I include region ﬁxed eﬀect the estimate become smaller and is still
statistically signiﬁcant at a lower level of conﬁdence. Table 4, panel A presents the eﬀect of village
level measure of risk correlation on how many group members the leader talk to more than ﬁve
times in a week (Leader Monitoring 1). We see that as village level measure of risk correlation
increases leader monitoring 1 decreases and the eﬀect is not statistically signiﬁcant.

Table 3, panel B presents the eﬀect of occupational homogeneity on number of people the
leader talks to less than once a week (Leader Monitoring 2). The more number of group members
the leader talks less than once a week implies lower monitoring by the leader. We see that as
occupational homogeneity increases leader monitoring 2 decreases and the eﬀect is statistically
signiﬁcant. This implies that after controlling for group and village characteristics occupational
homogeneity among the group members is positively correlated with leader monitoring. When I
include region ﬁxed eﬀect the estimate become smaller and is still statistically signiﬁcant. Table
4, panel B presents the eﬀect of village level measure of risk correlation on number of people the

115

leader talks to less than once a week (Leader Monitoring 2). We see that as village level measure
of risk correlation increases leader monitoring 1 decreases and the eﬀect is statistically signiﬁcant.
Table 3, panel C presents the eﬀect of occupational homogeneity on group awareness of each
other’s work (Peer Monitoring). The higher awareness of the group members about each other’s
work implies higher peer monitoring. We see that as occupational homogeneity increases peer
monitoring decreases and the eﬀect is not statistically signiﬁcant. When I include region ﬁxed
eﬀect the estimate become smaller and is still not statistically signiﬁcant. Table 4, panel C presents
the eﬀect of village level measure of risk correlation on peer monitoring. We see that as village level
measure of risk correlation increases peer monitoring decreases and the eﬀect is not statistically
signiﬁcant. When we include region FE the sign changes but the eﬀect is still not statistically
signiﬁcant.

Table 3, panel D presents the eﬀect of occupational homogeneity on how many meetings the
group had in the past month. We see that as occupational homogeneity increases number of
meetings decreases and the eﬀect is statistically signiﬁcant. This implies that after controlling for
group and village characteristics occupational homogeneity among the group members is negatively
correlated with number of meetings. When I include region ﬁxed eﬀect the estimate become more
negative and is still statistically signiﬁcant. Table 4, panel D presents the eﬀect of village level
measure of risk correlation on number of meetings. We see that as village level measure of risk
correlation increases number of meetings increases and the eﬀect is statistically signiﬁcant.

Overall we saw that as occupational homogeneity increases monitoring by both the leader and
among group members increases whereas as village level measure of risk correlation increases
monitoring decreases. In section 6 we compare the results from both measures in further details.
Second, I look at the eﬀect of risk correlation on cost/probability of audit. Table 5, panel
A presents the eﬀect of occupational homogeneity on the probability of someone verifying the
borrower’s income. The higher number of borrowers incomes veriﬁed implies higher auditing cost.
We see that as occupational homogeneity increases the probability of borrower’s income being
veriﬁed decreases and the eﬀect is statistically signiﬁcant. This implies that after controlling for

116

group and village characteristics occupational homogeneity among the group members is negatively
correlated with audit cost. When I include region ﬁxed eﬀect the estimate is similar and statistically
signiﬁcant. Table 6, panel A presents the eﬀect of village level measure of risk correlation on the
probability of someone verifying the borrower’s income. We see that as village level measure of
risk correlation increases the probability of someone verifying the borrower’s income decrease and
the eﬀect is not statistically signiﬁcant.

Table 5, panel B presents the eﬀect of occupational homogeneity on the probability on number
of oﬃcer visits. The higher number of oﬃcer visits implies higher auditing cost. We see that
as occupational homogeneity increases the number of oﬃcer visit decreases and the eﬀect is
statistically not signiﬁcant. When I include region ﬁxed eﬀect the estimate is more negative but
still statistically not signiﬁcant. Table 6, panel B presents the eﬀect of village level measure of risk
correlation on the number of oﬃcer visits. We see that as village level measure of risk correlation
increases the number of oﬃcer visits decreases and the eﬀect is not statistically signiﬁcant.

3.6 Comparing the theoretical results with the empirical ﬁndings

When I introduce risk correlation in the BBG model I ﬁnd that the relation between risk
correlation and monitoring depends on the probability of success of the borrower and cosigner. If
the probability of success for the borrower is less than that of the cosigner then as risk correlation
increases monitoring decreases. Whereas if the probability of success for the borrower is greater
than that of the monitor then as risk correlation increases monitoring decreases. It would have been
ideal if we had some individual level data which measures the relative success of a borrower and
cosigner then we could have tested the theoretical predictions more meaningfully but unfortunately
we do not have such individual level measures. When we look at the empirical result of risk
correlation on monitoring we see that the two measures of risk correlation consistently delivers
opposite results. The occupational homogeneity measure predicts a positive correlation between
monitoring and risk correlation whereas the village level measure of risk correlation predicts a
negative correlation.

117

Firstly I acknowledge that both the measures of risk correlation are at best proxies of the
true risk correlation. Occupational homogeneity captures the risk co-relatedness that comes from
having group members in the same occupation whereas the other measure captures a more macro
risk correlation at the village level.
In the context of the BBG model I think the occupational
measure does a better job of capturing correlation than the village level measure. If the cosigner
and borrower are in the same occupation they are more likely to have correlated outcomes. The
issue here could be that the increase in monitoring we see from the data might be coming from
channels which the model doesn’t capture such as information and marginal cost of monitoring.
For example if the borrower are cosigner are in the same profession then they might have similar
professional circles in the village and the cosigner gets information from the circles and has a
reduced cost of monitoring or higher capability to punish. Being in the same profession might also
aﬀect the marginal cost of monitoring. For example it might be easier for a peach farmer to verify
the output of another peach farmer compared to a shop keeper. The eﬀect from these channels
might over/under estimate the eﬀect of risk correlation on monitoring. On the other hand the village
level measure of risk correlation might not be representative of the risk correlation at the group
level as the BAAC borrowers constitute a very small portion of the total population. Hence the
eﬀect we capture from this measure of risk correlation might just be driven by other confounding
macro level village variables.

To get the relation between risk correlation and the cost/probability of audit I introduced risk
correlation in the Ghatak and Guinnane (2009) model. I found that as risk correlation increases the
cost/probability of audit increases. The intuition being that the lender only has to audit when both
borrowers fail to repay and as risk correlation increases the probability of that happening increases.
From the empirical side we see that risk correlation is negatively correlated with audit using both
measures of risk correlation. Again the measures of risk correlation we are using are proxies at best
and the divergence in the theoretical and empirical results could be driven by that. To understand the
contribution of this paper and importance of the correlational results, we should ﬁrst try and think
about how the ideal causal inference would look like. The ideal causal inference would require an

118

exogenous variation in the risk correlation of project outcomes among the group members. Since
groups are form endogenously and the lending mechanism heavily depends on this endogenous
matching process, such a variation might not be feasible. If we think of an experimental setting
then this will require us to randomly assign borrowers to groups in order to create some groups with
high risk correlation and some groups with low risk correlation. But again such randomization
would take away the very essence of group lending which heavily depends on the endogenous
matching by the group members which reﬂects network strength. For example one of the reasons
why group lending succeeds is because the borrowers are liable for each other’s loan and they are
able to eﬀectively punish each other in case someone defaults. If the groups are formed randomly
then the borrower’s ability to eﬀectively punish or monitor each other would substantially decrease
and hence the results we see from such an experiment would capture eﬀects apart from variation
in risk correlation. To sum up, an experiment to identify the causal relation would require some
sort of exogenous intervention in group formation but the endogenous group formation itself is an
important driver of group lending so the intervention would aﬀect a lot of other variables apart
from risk correlation. This is why a causal inference might not be feasible in this context.

3.7 Conclusion

In this paper I look at the eﬀect of risk correlation on monitoring and auditing cost in the context
of group lending. To identify this I take an approach very similar to Ahlin and Townsend (EJ 2007).
First I introduce risk correlation to some of the well established models in the literature which look
at monitoring and auditing. These models gives us a theoretical understanding of the relationship
between risk correlation and monitoring and auditing costs. Next I use the Thai Townsend data
base to get the empirical results. The empirical results are correlations and not causal.

To look at the eﬀect of risk correlation in monitoring I introduce risk correlation in the moral
hazard model by Banerjee et al (1994).
I ﬁnd that the eﬀect of risk correlation on monitoring
depends on the probability of success of the borrower and cosigner. If the probability of success
for the borrower is lower than that of the cosigner then monitoring decreases as risk correlation

119

increases. If the the probability of success for the borrower is higher than that of the cosigner then
monitoring increases as risk correlation increases. Hence from the theoretical model the eﬀect
is ambiguous. On the empirical side we see that the results are also ambiguous. When we use
occupational homogenity as a measure of risk correlation we ﬁnd that monitoring increases as risk
correlation increases whereas when we use the village level measure of risk correlation we ﬁnd that
monitoring decreases.

To look at the eﬀect of risk correlation on auditing cost I introduce risk correlation in the model
of State Veriﬁcation used in (Maitreesh Ghatak et al, 1999). I ﬁnd that as risk correlation increases
the probability/cost of audit increases. When we look at the empirical results we see that as risk
correlation increases audit cost decreases.

The empirical ﬁndings of the paper are not causal but correlation. It suﬀers from the usual
demerits of correlation results such as omitted variable bias and reverse causality. The measures
of risk correlation are also not perfect as they are at best proxies for the true risk correlations. In
the future I plan to dig further into the empirical side and try and establish more robust causal
results. The theory model in the paper speciﬁcally the BBG model used to look at the eﬀect of risk
correlation on monitoring has its drawbacks as the results from the theoretical model is partially
driven by the structure of risk correlation. Overall the study should not be treated as a ﬁnal answer
to question of how risk correlation among group members aﬀect group lending. Rather it should
be treated as a staring point to motivate research in the sparse literature.

120

Table 3.1: Descriptive Statistics

121

Table 3.2: Descriptive Statistics

122

Table 3.3: Leader Monitorig 1 RC

123

Table 3.4: Leader Monitorig 2 OH

124

Table 3.5: Leader Monitorig 1 RC

125

Table 3.6: Leader Monitorig 2 OH

126

Table 3.7: Peer Monitoring RC

127

Table 3.8: Peer Monitoring OH

128

Table 3.9: Monthly Meetings RC

129

Table 3.10: Monthly Meetings OH

130

Table 3.11: Borrower Income RC

131

Table 3.12: Borrower Income OH

132

Table 3.13: Oﬃcer Visit RC

133

Table 3.14: Oﬃcer Visit OH

134

APPENDICES

135

APPENDIX A

CHAPTER I : APPENDIX

A.1 Back of the Envelope Calculation

In this section I will calculate the distribution of the variable target given the sample distribution
of births by month and the sample distribution of marriages by month. Let m1, m2, m3, m4. . . m12
be the number of marriages by month. Where m1 is the number of marriages in January, m2 is the
number of marriages in February and so on. Similarly let b1,b2,b3...b12 be the number of births
by month. Where b1 is the number of births in January, b2 is the number of births in February and
so on.

Similar to section 2 we deﬁne the variable target=mod (age at marriage in months,12). Hence
the variable target takes the value zero when month of birth is the same as month of marriage.
That is if the individual is born in January and gets married in January or the individual is born in
February and gets married in February and so on. Takes the value 1 if the month of marriage is
one month later than the month of birth and so on.

Prob(target=0)=Prob(b1=m1) + Prob(b2=m2) . . . . . . . . . . . . . . . . . . + Prob(b12=m12)
Prob(target=1)=Prob(b1=m2) + Prob(b2=m3). . . . . . . . . . . . . . . . . . .. + Prob(b12=m1)

Prob(t=11)=Prob(b1=m12) + Prob(b2=m1) + . . . . . . + Prob(b12=m11)

...........

Now if there was no age targeting at marriage then; that is people had no preference of matching
their month of birth to their month of marriage; that is month of marriage is independent of month
of birth then;

Prob(bi=mi) = Prob(bi)Prob(mi)

Where Prob(bi)=bi/B and Prob(mi)=mi/M

M = mi

136

B = bi

We obtain the distribution of birth by month and the distribution of marriages by month from
our sample. Given the sample distribution we calculate the distribution of the variable target.
Graph 6 plots the actual and calculated distribution of the variable target. The diﬀerence between
the actual and calculated distribution at zero is what I deﬁne as age targeting at marriage, where
parents target certain whole number age for their daughters to get married after.

A.2 Example for IV2

IV2 captures the one year future discounted number of marriage dates from the month of birth
of the woman. IV2 to not only captures the fact that certain months have more marriages than
other but it also captures the fact that certain months are followed by a period of high number of
marriage dates whereas certain months are followed by a period of low number of marriage dates.
For example January and July has similar number of marriage dates but January is followed by a
period of high number of marriage dates and July is followed by a period of low number of marriage
dates. IV2 would be able to capture this essence whereas IV1 does not. IV2 could be thought of
as a relative ranking of a month compared to other months. I provide an example with beta equal
to 0 and 0.5.

Example for IV2

137

A.3 Compliers for the IVs

To better understand why we get diﬀerent results on fertility outcomes across the two IV’s, I
take a closer look at the compliers for both the IVs in this section. The complier for the age at
menarche IV are women who delays marriage due to delay in age at menarche. Since 94 percent of
the women reach menarche by the age of 15, the compliers for the menarche IV primarily consists of
young girls who get married at an early age of sixteen or lower. For my IV the compliers are women
belonging to families who practice minimum age targetting at marriage and exhibit a preference for
getting married on auspicious marriage months. The preference for getting married in an auspicious
marriage months is relevant for Hindu women getting married at any age but the age targetting at
marriage is relevant for women getting married before the age of 22. Hence the compliers for my
IV belongs to the group of women who gets married before the age of 22. When we look at the
socio-economic background of the compliers across the two IV. Both minimum age targeting at
marriage and preference for getting married on auspicious dates are exhibited by women who are
less educated and belonging to families with lower wealth and education 1. Families who get their
daughters married at an early age of less than sixteen (population for the menarche complier) also

1Graphs 7-9 and graphs 15-17.

138

belongs to families with lower wealth and education. Hence the diﬀerential results are most likely
not driven by the fact that the compliers belong to diﬀerent socio-economic backgrounds.

One potential diﬀerence between the compliers for the two IVs is that the compliers for the age
at menarche IV are younger than the compliers for my IV. Thus one possible explanation for why
we see diﬀerent results for fertility across the two IVs is that how delayed age at marriage aﬀects
fertility diﬀers by age at which the women gets married. A three month delay in marriage at the
young age of 12-16 might have a diﬀerent eﬀect on fertility compared to a three month delay at a
higher age of 17-21. To check that whether the diﬀerence in results across the two IVs is driven
by non linearity in the response function by age, I estimate the results from my IV on a set of age
windows. Appendix table 7 presents the results. We see that for women who gets married before
the age of 16 one month delay in marriage increases the number of children by 0.01 (0.3 percent)
and the eﬀect is statistically signiﬁcant. For women who got married after the age of 16 one month
delay in age at marriage increases the number of children by 0.006 (0.27 percent) and the eﬀect is
statistically signiﬁcant. We also ﬁnd similar results for delay in age at ﬁrst birth and ideal number
of children. Thus I conclude that the diﬀerential fertility results we get for both IV’s is not driven
by the fact that the menarche IV compliers are younger than the compliers for my IV.

139

Table A.1: First Stage Summary Stat (Appendix Table 1)

140

Table A.2: Eﬀect of Age at Marriage on Spousal Characteristics (Appendix Table 2)

141

Table A.3: Eﬀect of Age at Marriage on Reproductive Health (Appendix Table 3)

142

Table A.4: Wald IV Estimates (Appendix Table 4)

143

Table A.5: First Stage Estimates (Appendix Table 5)

144

Table A.6: Muslim Population (Appendix Table 6)

145

Table A.7: Eﬀect of Age at Marriage on Fertility (Appendix Table 7)

146

APPENDIX B

CHAPTER II : APPENDIX

B.1 Accounting for age

One of the major diﬀerence between people who joined the sectors prior and post 2004 is that
the people who joined prior are older. As a robustness check i include interaction terms with age
such as age*sector, age*gender and age*sector*gender. Appendix table A4 reports the results of
the triple diﬀerence estimates including the interaction terms. We see that the estimates do not vary
from the original estimates.

B.2 Deﬁnition of White Collar Jobs

There are two variable which helps in identifying my desired population, ﬁrst variable directly
asks whether the person works in public sector or private sector or other. The second variable
asks the speciﬁc occupation of the person. They jobs are grouped into 99 occupations. I separate
the occupations into white collar and non-white collar jobs based on my knowledge of occupation
in India, like teachers are white collar while bus drivers are not. Now in my analysis I use the
population which is the intersection of white collar and public sector or private sector. To see if
my analysis is robust to the deﬁnition is white collar I use the following speciﬁcation. I just use
the variable which asks whether the person works in a public or private sector. Hence I include
both white collar and non-white collar in my analysis. Table A5 reports the estimate from the triple
diﬀerence with the alternate speciﬁcation. We see that the estimates do not vary from the original
estimates. Hence the estimates are not sensitive to the deﬁnition of white collar.

B.2.1

Including Mothers Sector

Sector is deﬁned by fathers occupation as the percentage of female in the public sector is 14% in
my data and usually fathers have a higher say in matters of ﬁnancial decision in the household. I

147

present the results including mother’s sector, now sector is deﬁned as public if either mother or
father is in public sector and private if both mother and father are not in public and one of them is
in private. Table A6 reports the estimate from the triple diﬀerence with the alternate speciﬁcation.
We see that the estimates do not vary from the original estimates. Hence the estimates are not
sensitive to the deﬁnition of white collar.

B.3 Anthropometric Measures of Early Childhood Investments

In this section I look at some anthropometric measures of early child childhood investment
such as height for age, weight for age, weight for height and BMI for age. To compute these
scores I use the zscore06 function in STATA. It calculates anthropometric z-scores using the WHO
child growth standards in developing countries. Height-for-age, weight-for-height, BMI-for-age
and weight-for-age Z-scores are calculated for children 0 to 5 years of age. Appendix table I
presents the results. The sample consists of children age 0-5. Overall I do not ﬁnd any statistically
signiﬁcant or economically meaningful results for this sample. The most likely reason for this is
that the data is plagued with measurement errors and missing values. Also since the birth months
of a lot of children were imputed I had to use age in years to calculate these scores which might
also drive the ﬁndings.

148

Table B.1: A1 Diﬀerence-in-Diﬀerence Results for Boy child and Girl child

149

Table B.2: A2 Triple Diﬀerence Estimate with sibling ﬁxed eﬀect

150

Table B.3: A3 Private School Enrollment

151

Table B.4: A4 Robustness Check for Age

152

Table B.5: A5 Alternate deﬁnition of Public Sector

153

Table B.6: A5 Including Mother’s Sector

154

Table B.7: A6 Means of the Outcome Variables

155

Table B.8: A7 Diﬀerential Eﬀect of the reform on other forms of Investment

156

Table B.9: A8 Household Characteristics by Household Composition

157

Table B.10: A9 Private School Enrollment by Household Composition

158

Table B.11: A10 Results on Anthropocentric Measures

*Table A10 contains estimates of β7 from equation (8) for various outcome variables.
*The sample consists of children age 0-5.
*Each row presents the DDD estimate from equation 21 with a particular outcome variable.
*Standard errors are reported in brackets and are clustered at the district level.
*Speciﬁcation (1) includes full set of controls : child characteristics, father characteristics, mother
characteristics and household characteristics.
*Speciﬁcation (2) does not include household characteristics.

159

BIBLIOGRAPHY

160

BIBLIOGRAPHY

Bruce, Judith. "Married adolescent girls: human rights, health, and developmental needs of a
neglected majority." Economic and Political Weekly (2003): 4378-4380.

Clark, Shelley. "Early marriage and HIV risks in sub-Saharan Africa." Studies in family plan-

ning 35.3 (2004): 149-160.

Nour, Nawal M. "Health consequences of child marriage in Africa." Emerging infectious dis-

eases 12.11 (2006): 1644.

Santhya, Kidangamparampil G., et al.

"Associations between early marriage and young
women’s marital and reproductive health outcomes: evidence from India." International perspec-
tives on sexual and reproductive health (2010): 132-139.

Raj, Anita, et al. "Prevalence of child marriage and its eﬀect on fertility and fertility-control
outcomes of young women in India: a cross-sectional, observational study." The Lancet 373.9678
(2009): 1883-1889.

Field, Erica, and Attila Ambrus. "Early marriage, age of menarche, and female schooling

attainment in Bangladesh." Journal of political Economy 116.5 (2008): 881-930.

Sekhri, Sheetal, and Sisir Debnath. "Intergenerational consequences of early age marriages
of girls: Eﬀect on children’s human capital." The Journal of Development Studies 50.12 (2014):
1670-1686.

Chari, A. V., et al. "The causal eﬀect of maternal age at marriage on child wellbeing: Evidence

from India." Journal of Development Economics 127 (2017): 42-55.

Dhamija, Gaurav, and Punarjit Roychowdhury. "The Causal Impact of Women’s Age at Mar-

riage on Domestic Violence in India." (2018).

Liestøl, Knut. "Menarcheal age and spontaneous abortion: a causal connection?." American

Journal of Epidemiology 111.6 (1980): 753-758.

MARTIN, ELIZABETH J., LOUISE A. BRINTON, and ROBERT HOOVER. "Menarcheal

age and miscarriage." American journal of epidemiology 117.5 (1983): 634-636.

Wyshak, Grace. "Age at menarche and unsuccessful pregnancy outcome." Annals of human

biology 10.1 (1983): 69-73.

162

SANDLER, DALE P., ALLEN J. WILCOX, and LOUISE F. HORNEY. "Age at menarche and

subsequent reproductive events." American Journal of Epidemiology 119.5 (1984): 765-774.

Guldbrandsen, Karen, et al. "Age of menarche and time to pregnancy." Human Reproduction

29.9 (2014): 2058-2064.

Hossain, Parvez, Bisher Kawar, and Meguid El Nahas. "Obesity and diabetes in the developing

world—a growing challenge." (2009).

Carmichael, Sarah. "Marriage and power: Age at ﬁrst marriage and spousal age gap in lesser

developed countries." The History of the Family 16.4 (2011): 416-436.

Stock, James H., and Motohiro Yogo. "Testing for weak instruments in linear IV regression."

(2002).

Stock, James H., Jonathan H. Wright, and Motohiro Yogo. "A survey of weak instruments and
weak identiﬁcation in generalized method of moments." Journal of Business & Economic Statistics
20.4 (2002): 518-529.

Hausman, Jerry, James H. Stock, and Motohiro Yogo. "Asymptotic properties of the Hahn–Hausman

test for weak-instruments." Economics Letters 89.3 (2005): 333-342.

Cragg, John G., and Stephen G. Donald. "Testing identiﬁability and speciﬁcation in instrumen-

tal variable models." Econometric Theory 9.2 (1993): 222-240.

Petridou, Eleni, et al. "Determinants of age at menarche as early life predictors of breast cancer

risk." International journal of cancer 68.2 (1996): 193-198.

Garland, Miriam, et al. "Menstrual cycle characteristics and history of ovulatory infertility in
relation to breast cancer risk in a large cohort of US women." American journal of epidemiology
147.7 (1998): 636-643.

Weber, Gerhard W., Hermann Prossinger, and Horst Seidler. "Height depends on month of

birth." Nature 391.6669 (1998): 754.

Lokshin, Michael, and Sergiy Radyakin. Month of birth and children’s health in India. The

World Bank, 2009.

Hahn, Jinyong, Petra Todd, and Wilbert Van der Klaauw. "Identiﬁcation and estimation of

treatment eﬀects with a regression-discontinuity design." Econometrica 69.1 (2001): 201-209.

Edwards, Ben, et al. "Is Monotonicity in an IV and RD design testable? No, but you can still

check it." (2013).

163

Jayachandran, Seema. "The roots of gender inequality in developing countries." economics 7.1

(2015): 63-88.

Sen, Amartya. "More Than 100 Million Women Are Missing." New York Review of Books

(2010).

Behrman, Jere R., and Anil B. Deolalikar. "Health and nutrition." Handbook of development

economics 1 (1988): 631-711.

Zimmer, Zachary, et al. "Education of adult children and mortality of their elderly parents in

Taiwan." Demography 44.2 (2007): 289-305.

Sadhak, H. "Does Not india Need a Default Option in the New pension system?." Economic

and Political Weekly (2009): 59-68.

Gayathri, S. R. I. N. I. V. A. S. A. N. "The elderly in an Indian retirement community: the

Inﬂuence of the caste system." Bold 17 (2006): 2-10.

Goswami, Ranadev. "Indian Pension System: Problems and Prognosis." Indian Institute of

Management, Bangalore, January 262001 (2001).

Sanyal, Ayanendu. Civil Service Pension Reforms in India. Diss. INSTITUTE FOR SOCIAL

AND ECONOMIC CHANGE BANGALORE, 2013.

Broadbent, John, Michael Palumbo, and Elizabeth Woodman. "The shift from deﬁned beneﬁt
to deﬁned contribution pension plans—Implications for asset allocation and risk management.
Paper prepared for a working group on institutional investors, global savings, and asset allocation
established by the Committee on the Global Financial System." (2006).

Pande, Rohini P., and Nan Marie Astone. "Explaining son preference in rural India: the inde-
pendent role of structural versus individual factors." Population Research and Policy Review 26.1
(2007): 1-29.

Mutharayappa, Rangamuthia. "Son preference and its eﬀect on fertility in India." (1997).

Arnold, Fred, Minja Kim Choe, and Tarun K. Roy. "Son preference, the family-building process

and child mortality in India." Population studies 52.3 (1998): 301-315.

Barcellos, Silvia Helena, Leandro S. Carvalho, and Adriana Lleras-Muney. "Child gender and
parental investments in India: are boys and girls treated diﬀerently?." American Economic Journal:
Applied Economics 6.1 (2014): 157-189.

Bharadwaj, Prashant, and Leah K. Lakdawala. "Discrimination begins in the womb: Evidence

of sex-selective prenatal investments." Journal of Human Resources 48.1 (2013): 71-113.

164

Chen, Lincoln C., Emdadul Huq, and Stan d’Souza. "Sex bias in the family allocation of food

and health care in rural Bangladesh." Population and development review (1981): 55-70.

Gupta, Monica Das. "Selective discrimination against female children in rural Punjab, India."

Population and development review (1987): 77-100.

Behrman, Jere R., and Anil B. Deolalikar. "Health and nutrition." Handbook of development

economics 1 (1988): 631-711.

Borooah, Vani K. "Gender bias among children in India in their diet and immunisation against

disease." Social science & medicine 58.9 (2004): 1719-1731.

Behrman, Jere R., Robert A. Pollak, and Paul Taubman. "Do parents favor boys?." International

Economic Review (1986): 33-54.

Chaudhury, Nazmul, et al. "Missing in action: teacher and health worker absence in developing

countries." The Journal of Economic Perspectives 20.1 (2006): 91-116.

Wadhwa, Wilima. "Are private schools really performing better than government schools."

Annual Status of Education Report (Rural) New Delhi (2009).

Kingdon, Geeta Gandhi. "The progress of school education in India." Oxford Review of Eco-

nomic Policy 23.2 (2007): 168-195.

Kingdon, Geeta Gandhi. "Does the labour market explain lower female schooling in India?."

The Journal of Development Studies 35.1 (1998): 39-65.

Kingdon, Geeta Gandhi, and Jeemol Unni. "Education and women’s labour market outcomes

in India." Education Economics 9.2 (2001): 173-195

Kingdon, Geeta. "The quality and eﬃciency of private and public education: a case-study of

urban India." Oxford Bulletin of Economics and Statistics 58.1 (1996): 57-82.

Tooley, James, et al. "The relative quality and cost-eﬀectiveness of private and public schools
for low-income families: a case study in a developing country." School Eﬀectiveness and School
Improvement 21.2 (2010): 117-144.

Pandey, Priyanka, Sangeeta Goyal, and Venkatesh Sundararaman. "Community participation
in public schools: impact of information campaigns in three Indian states." Education Economics
17.3 (2009): 355-375.

Srinivasan, T. N. "Population growth and economic development." Journal of Policy Modeling

10.1 (1988): 7-28.

165

Jensen, Robert. 2003. “Equal Treatment, Unequal Outcomes? Generating Sex Inequality

Through Fertility Behavior” Mimeo, Harvard University

De Vos, Susan. "An old-age security incentive for children in the Philippines and Taiwan."

Economic Development and Cultural Change 33.4 (1985): 793-814.

Zhang, Junsen, and Kazuo Nishimura. "The old-age security hypothesis revisited." Journal of

Development Economics 41.1 (1993): 191-202.

Ebenstein, Avraham, and Steven Leung. "Son preference and access to social insurance: ev-
idence from China’s rural pension program." Population and Development Review 36.1 (2010):
47-70.

Mu, Ren, and Yang Du. "Pension coverage for parents and educational investment in children:

Evidence from urban China." The World Bank Economic Review (2015): lhv060.

Samanta, Tannistha, Feinian Chen, and Reeve Vanneman. "Living arrangements and health of
older adults in India." The Journals of Gerontology Series B: Psychological Sciences and Social
Sciences (2014): gbu164.

Kishor, Sunita. "" May God Give Sons to All": Gender and Child Mortality in India." American

Sociological Review (1993): 247-265.

Ahlin, Christian, and Robert M. Townsend. "Using repayment data to test across models of

joint liability lending." The Economic Journal 117.517 (2007): F11-F51.

Stiglitz, Joseph E. "Peer monitoring and credit markets." The world bank economic review 4.3

(1990): 351-366.

Banerjee, Abhijit V., Timothy Besley, and Timothy W. Guinnane. "Thy neighbor’s keeper: The
design of a credit cooperative with theory and a test." The Quarterly Journal of Economics 109.2
(1994): 491-515.

Besley, Timothy, and Stephen Coate. "Group lending, repayment incentives and social collat-

eral." Journal of development economics 46.1 (1995): 1-18.

Ghatak, M. "Group Lending, Local Information and Peer Selection ‘, Journal of Development

Economics, vol. 60." (1999).

Ghatak, Maitreesh, and Timothy W. Guinnane. "The economics of lending with joint liability:

theory and practice1." Journal of development economics 60.1 (1999): 195-228.

166