ESSAYS IN THE ECONOMICS OF CASTE AND RELIGION
By
Muzna Fatima Alvi

A DISSERTATION
Submitted to
Michigan State University
in partial fulfilment of the requirements
for the degree of
Economics - Doctor of Philosophy
2017

ABSTRACT
ESSAYS IN THE ECONOMICS OF CASTE AND RELIGION
By
Muzna Fatima Alvi
Through my dissertation, I aim to uncover the processes that govern the relationship between social
or ethnic identity and economic outcomes such as employment and education. I focus in particular on
studying outcomes for Muslim women in India who have remained on the fringes of economic enquiry. A
majority of the literature on economic exclusion in India has focused on caste based discrimination, which
is warranted given the acute poverty and deprivation faced by lower caste communities in the country.
However, despite being comparable on measures of economic, social and educational deprivation, Indian
Muslims have remained absent from the development discourse in the country. I hope my research will
contribute to filling this lacuna in the literature.

The first chapter examines the evolution of labor force participation rate (LFPR) differences between
women from the Hindu, Schedule Caste/Schedule Tribe (SC/ST) and Muslim communities in India. I
find that affirmative action policies have led to convergence in the education levels of Hindus and SC/ST
which has consequently led to a narrowing of the LFPR gap of the two groups, however, the HinduMuslim gap in female labor participation remains persistently high. I find that part of the religious gap is
explained due to Muslim women being trapped in enclaves of low economic development due to a
combination of Muslim ghettoization and relatively localized marriage markets. Contrary to popular
rhetoric, I find that culture does not play a big role in explaining the absence of Muslim women from the
labor force.

In the second chapter my coauthor Dr Maitreyi Bordia Das and I, study the causes for large
differences in female labor force participation between West Bengal and Bangladesh. Despite historical
and cultural commonalities, both regions exhibit very different levels of socio-economic development, one

of which is the wide disparity in female labor force participation. We find that observable covariates do
not play a major role in explaining the wide gap in female LFPR between the two regions. This is driven
by the fact that in West Bengal, much like the rest of India, higher levels of education and improvements
in social status leads to withdrawal of women from the labor force. On the other hand, we find the opposite
to be true in Bangladesh. In comparing Muslims on either side of the border, our results remain
qualitatively similar to the aggregate results. We conclude that the differing political trajectories followed
by the two regions have created widely varying economic structures which have in turn led to divergent
demands for female labor in the two countries- the most obvious of which is the decline of agriculture in
West Bengal and the continued flourishing of both agriculture and manufacturing in Bangladesh.

The third chapter aims to study the impact of communal violence in India on the schooling and
education decisions of Muslims in India. The causes and consequences of violent conflict, particularly civil
war, mass violence and ethnic conflict, has attracted considerable scholarly attention across disciplines.
However, there is much less research on the how small isolated and geographically dispersed incidences
of violence can lead to long term deficit in human capital accumulation for entire communities. This effect
could be driven by the immediate loss of life and livelihood following conflict. However, repeated conflict
also ingrains a culture of fear in the minds of the victimized community that implicitly determines
economic behavior in the long-run. My paper deals with one such group, namely Indian Muslim. I use
data on religious violence from 1950-2006 combined with data on education from several consecutive
rounds of the National Sample Survey (NSS) on Employment and Unemployment (Schedule 10) and
Participation in Education (Schedule 25.2) covering the period between 1983 and 2014. My results
suggest that both contemporaneous and cumulative violence has no effect on the educational outcomes
and dropout decision of Muslim girls and boys.

To all the strong women in my life.
Thank you for blazing the trail on
which I wander.

iv

ACKNOWLEDGEMENTS
I would like to sincerely thank my advisor, Todd Elder, for his guidance and steadfast support. Thank
you for not allowing me to wallow in my setbacks and for elevating me in every minor triumph. This
dissertation has benefitted not only from your knowledge and insight but also by your encouragement,
positivity and good humor through every turn.
I would also like to thank my committee member, Leah Lakdawala, for her patience and support
throughout several iterations of this dissertation. Your detailed comments on all my papers and
presentations have compelled me to improve my writing and have been invaluable in my growth as a
researcher. I would like to extend my gratitude to my committee members, Christian Ahlin and Amita
Chudgar for their support and encouragement both in the writing of the dissertation and during the
process of my job search.
I would be remiss if I did not acknowledge the support of my dear friend Riju Joshi, without whom
this dissertation truly would not have been possible. Very few of us are fortunately to have a friend like
her, who is an ally both in academic endeavors and in personal adventures. You have enriched me as a
scholar and a friend and I am blessed to have you in my life. A big thank you to my friends Annie, PoChun, Doosoo, Patrick, Walter, Saman, Maggie, Meenakshi and Ashesh. You have been the shoulders on
which I have leaned through professional and personal challenges over these last five years.
Last and certainly not the least, I would like to thank my family, who have been my biggest
cheerleaders throughout. Thank you to Mummy, Aanie, Humera Qala and Nanima for leading by example
and for allowing me the luxury to make mistakes and the freedom to succeed. Finally, I would like to
thank my wonderful husband, Taha for supporting and loving me through every success and failure over
the last five years. In the process of writing this dissertation, we have also built our life together, and I
look forward to our adventures together as a family with Zidaan.

v

TABLE OF CONTENTS

LIST OF TABLES

viii

LIST OF FIGURES

x

CHAPTER 1 Caste, Religion and Economic Participation of Women in India

1

1 Introduction

1

2 System of Caste in India

3

3 Review of Existing Literature

5

3.1 Economic exclusion and caste in India

5

3.2 Indian Muslims

7

4 Methodology

9

4.1 DFL method

10

4.2 Oaxaca-Blinder method

11

5 Data and Summary Statistics

12

6 Role of All Covariates

15

7 Role of Individual Covariates

18

8 Extensions

21

8.1 How does culture influence participation gaps?

21

8.2 Do alternative mappings change the result?

22

8.3 How do results differ in rural and urban areas?

23

8.4 Has the predictive power of covariates declined?

24

9 Discussion and Conclusion

25

APPENDICES

27

APPENDIX A MAIN TABLES AND FIGURES

28

APPENDIX B SUPPLEMENTARY TABLES AND FIGURES

42

REFERENCES

60

CHAPTER 2 The Two Bengals: Female Labor Force Participation in West
Bengal and Bangladesh

64

1 Introduction

64

2 Review of Literature

65

vi

2.1 Female labor in Bangladesh

65

2.2 Women’s work in West Bengal

68

3 Methodology

71

3.1 DFL method

72

3.2 Oaxaca-Blinder method

73

4 Data and Descriptive Statistics

74

5 Main Results

78

6 Extensions

80

6.1 How do results vary for different sub-groups?

80

6.2 Do results change with different mappings?

81

6.3 How do the two regions compare with the rest of India?

81

7 Conclusion

83

APPENDICES

85

APPENDIX A MAIN TABLES AND FIGURES

86

APPENDIX B SUPPLEMENTARY TABLES

95

REFERENCES

105

CHAPTER 3 Communal Violence and Human Capital Accumulation in India

109

1 Introduction

109

2 Review of Literature

110

3 Data and Descriptive Statistics

112

3.1 Creating the retrospective panel

114

4 Methodology

116

5 Results

118

6 Extensions

119

6.1 Testing robustness of results with leads and lags

120

6.2 Testing results for unmarried respondents

120

6.3 Testing the results with a hazard model

121

7 Conclusion and Future Work

121

APPENDICES

125

APPENDIX A MAIN TABLES AND FIGURES

125

APPENDIX B SUPPLEMENTARY TABLES

136

REFERENCES

147

vii

LIST OF TABLES

Table 1.1: Decomposition of Hindu-SC/ST LFPR Gap (DFL Method)

29

Table 1.2: Decomposition of Hindu-Muslim LFPR Gap (DFL Method)

30

Table 1.3: Decomposition of Hindu-SC/ST LFPR Gap (Oaxaca-Blinder Method)

31

Table 1.4: Decomposition of Hindu-Muslim LFPR Gap (Oaxaca-Blinder Method)

32

Table 1.5: Decomposition of Hindu-SC/ST LFPR Gap (Oaxaca-Blinder Method
including Predicted Income)

33

Table 1.6: Decomposition of Hindu-Muslim LFPR Gap (Oaxaca-Blinder Method
including Predicted Income)

34

Table 1.7: Measuring Role of Culture-IHDS II (Oaxaca Blinder Method)

35

Table 1.8: Summary of Culture Variable- IHDS

36

Table 1.9: Descriptive Statistics- Hindus

43

Table 1.10: Descriptive Statistics- Muslims

45

Table 1.11: Descriptive Statistics- SC/ST

47

Table 1.12: Decomposition of Hindu-SC/ST LFPR Gap (Oaxaca-Blinder MethodSC/ST Mapping)

49

Table 1.13: Decomposition of Hindu-Muslim LFPR Gap (Oaxaca-Blinder MethodMuslim Mapping)

50

Table 1.14: Decomposition of Hindu-SC/ST LFPR Gap (Oaxaca-Blinder MethodPooled Mapping)

51

Table 1.15: Decomposition of Hindu-Muslim LFPR Gap (Oaxaca-Blinder MethodPooled Mapping)

52

Table 1.16: Decomposition of Rural Hindu-SC/ST LFPR Gap (Oaxaca-Blinder
Method)

53

Table 1.17: Decomposition of Rural Hindu-Muslim LFPR Gap (Oaxaca-Blinder
Method)

54

Table 1.18: Decomposition of Urban Hindu-SC/ST LFPR Gap (Oaxaca-Blinder
Method)

55

Table 1.19: Decomposition of Urban Hindu-Muslim LFPR Gap (Oaxaca-Blinder
Method)

56

viii

Table 1.20: Decomposition of Hindu-Muslim LFPR Gap (Oaxaca-Blinder MethodHindu 38th Round Mapping)

57

Table 1.21: Decomposition of Hindu-Muslim LFPR Gap (Oaxaca-Blinder MethodHindu 68th Round Mapping)

58

Table 2.1: OLS of LFP (x100) on Observables

87

Table 2.2: Decomposition of Female LFPR Gap

89

Table 2.3: Descriptive Statistics for Counterfactual Population

90

Table 2.4: Decomposition of Female LFPR Gap by Sub-Group

91

Table 2.5: Decomposition of Female LFPR Gap (West Bengal & India)

92

Table 2.6: Decomposition of Female LFPR Gap (Bangladesh & India)

93

Table 2.7: Summary Statistics

96

Table 2.8: Labor Force Participation Rates

97

Table 2.9: Distribution of Labor Force

99

Table 2.10: Distribution of Female Workforce by Top 10 Industry- West Bengal

100

Table 2.11: Distribution of Female Workforce by Top 10 Industry- All India

101

Table 2.12: Distribution of Female Workforce by Top 10 Industry- Bangladesh

102

Table 2.13: OLS of LFP (x100) on Observables

103

Table 3.1: Descriptive Statistics- Individual Sample

127

Table 3.2: Effect of Casualties on Enrolment 1986-2006

128

Table 3.3: Effect of Cumulative Violence (Casualties) on Individual Outcomes

130

Table 3.4: Effect of Incidents on Enrolment 1986-2006

137

Table 3.5: Effect of Cumulative Violence (Incidents) on Individual Outcomes

139

Table 3.6: Effect of Past and Future Casualties on Enrolment 1986-2006

140

Table 3.7: Effect of Casualties on Enrolment 1986-2006, Unmarried Sample

142

Table 3.8: Effect of Cumulative Violence (Casualties), Unmarried Sample

144

Table 3.9: Effect of Casualties on Enrolment 1986-2006, Hazard Model

145

ix

LIST OF FIGURES

Figure 1.1: Labor Force Participation Rates

37

Figure 1.2: Hindu-SC/ST Gap in LFPR

38

Figure 1.3: Hindu-Muslim Gap in LFPR

39

Figure 1.4: Income Distribution and LFPR

40

Figure 1.5: Muslim Concentration and Female LFPR

41

Figure 1.6: Comparing Counterfactual Populations

59

Figure 2.1: Actual and Predicted Gap using Different Reference Categories

94

Figure 3.1: Total Casualties by Year

131

Figure 3.2: Total Casualties by State (1950-2006)

132

Figure 3.3: Yearly Rate of Enrolment, Retrospective Data

133

Figure 3.4: Enrolment by Age, Retrospective Data

134

Figure 3.5: Dropout Rates for Children Age 5-16

135

x

Chapter 1
Caste, Religion and Economic Participation of Women in
India

1

Introduction
This paper aims to study gaps in labor force participation rates (LFPR) for women from the Hindu,

Schedule Caste/Schedule Tribe (SC/ST) and Muslim communities in India. The motivation behind such
a comparison is twofold: firstly, SC/ST and Muslims are the most deprived communities in India, and the
status of women from these communities merits in-depth analysis. Secondly, summary statistics show
that despite similarity on many indicators of economic development, the labor market participation
behavior of women from these communities is markedly different. My data spans over two decades
(1983/84 to 2011/12), and covers a period of remarkable growth in the Indian economy, boosted by
widespread economic reforms in the early 1990s. Have the benefits of this growth trickled down to these
marginalized communities? How have women from these communities responded to this period of rapid
growth1? Is lack of education or poverty holding back women from these communities from fully
participating in the economy? These are some of the questions this paper will attempt to answer.

Female labor force participation has received widespread attention in economic literature (Blau &
Kahn, 2013). While some studies have looked at how changes in fertility affect the decision of women to
enter or exit the labor force (Bloom et al. 2009; Hotz & Miller, 1988; Rosenzweig & Wolpin, 1980), others
have studied the effect of female labor supply on children’s cognitive development and health (Blau &
Grossberg, 1992; Brooks-Gunn et al. 2002; Ruhm, 2004). Of particular interest to us is the growing
literature that looks at the effect of culture on economic outcomes, particularly female labor supply

1 Munshi & Rosenzweig (2006) use survey data from Bombay spread over 20 years to show that lower caste girls
are exploiting opportunities presented by the new economy to enroll in English schools as a means of upward mobility,
in contrast to boys who are constrained by caste networks that restrict them to traditional occupation and local language
schools.

1

(Antecol, 2003; Fernandez, 2007). This paper attempts to not only study religion and caste based gaps in
female LFPR, but also examine how much of these gaps are attributable to differences in observable
covariates such as education, and how much can be attributed to unobservable components such as culture
and discrimination. I am also interested in knowing how this gap in female LFPR has evolved over time,
particularly when the Indian economy has gone through rapid structures changes and there have been
marked improvements in income levels and human capital development. This issue has come into renewed
focus with recent data suggesting that female LFPR in India has been declining over the last decade.
Klansen and Pieters (2015) study the phenomenon of stagnating female LFPR in urban India and conclude
that this trend is being driven by rising income and education of men coupled with low growth in sectors
that employ women2.

I use data from seven rounds of the National Sample Survey (NSS) on Employment and
Unemployment in India and find that there is considerable gap between the LFPR of Hindu women in
comparison to SC/ST and Muslim women. Despite being similar on various indicators of development,
SC/ST women have the highest LFPR and Muslim women have the lowest LFPR across all socioreligious categories in India. While the Hindu-SC/ST gap has been narrowing over time, the HinduMuslim gap remains high. Moreover, I find that observable characteristics such as education, poverty and
household size explain a large part of the Hindu-SC/ST gap, whereas very little of the Hindu-Muslim gap
is explained by these covariates. My findings are consistent with recent research on caste and religion
based disparities in India. Hnatkovska et al. (2012) find that between 1983/84 and 2004/05, there has
been significant convergence in wages and consumption levels between SC/STs and non-SC/STs, and
this is driven largely by convergence in levels of education. For both wage and consumption convergence
they find that SC/STs in lower percentiles have benefited more compared to the relatively better off.
Bhalotra et al. (2010) study gaps in infant mortality rates between Hindus, Muslims and SC/STs and find
that despite low indicators of socio-economic status, children born to Muslim mothers have, and continue

2

This finding is consistent with other cross-country studies that find economic growth and female LFPR follow a
U-shape pattern, where initial economic growth is accompanied by declining female participation. However as incomes
continue to grow, female LFPR also rises (see Mammen & Paxson, 2000)

2

to exhibit, higher survival rates. This paradoxical result is consistent with my finding that shows the
inability of background characteristics in explaining outcomes for Muslim women. I find that Muslim
women’s absence from the labor force is explained in large part due to their concentration in areas of low
economic activity. I use the second wave of India Human Development Survey to test the role of culture
and find that it plays only a small part in explaining Muslim women’s low LFP.

The paper is divided in the following way: The second section provides a brief introduction to the
caste system in India, section 3 reviews the literature on caste and religious inequality in India, and section
4 outlines the methodology while section 5 explains the data and summary statistics. Sections 6 and 7
explore the role of background characteristics in explaining the LFPR gaps, while section 8 provides
avenues for future research and concludes.

2

System of Caste in India
India has been and continues to be deeply stratified on the basis of caste and religion. Even today,

caste status plays an important role in determining opportunities and outcomes for the people of the
country. While there are thousands of castes in India, they are divided between five main religions,
namely, Hinduism, Islam, Christianity, Sikhism and Buddhism. Efforts to remove caste based inequality
led some monarchs and British colonials to recognize most of the lower caste and religious populations
as historically marginalized. Consequently, in independent India, the constitution went to great length in
formulating affirmative action for marginalized populations, namely, Dalits, Adivasis, and to a lesser
extent Shudras, and Muslims. In official lexicon Dalits are identified as Scheduled Castes (SCs), Adivasis
as Schedule Tribes (STs), Shudras as Other Backward Class (OBCs) and Muslims as Muslims. Together,
these four groups are identified to be placed lower in the caste and religious hierarchy while the Hindu
Upper Castes are understood to be placed higher.

The Hindu caste system can be divided into four hierarchical groups as follows: Brahmin (the priestly
caste), Kshatriyas (the warrior caste), Vaishyas (the trading caste), Shudras (the service caste such as

3

farmers, craftsmen and shepherds) and Ati-Shudras. The Ati-Shudras, now called Dalits, have been
designated the work of menial labor including manual scavenging and were segregated from the village
and main areas of the cities owing to their subordinate caste status (Omvedt, 2004; Thorat, 2002). While
most of them are still Hindus many converted to other religions such as Buddhism, Christianity, Islam
and Sikhism, to escape caste based oppression. According to the Indian constitution, Ati-Shudras who
converted to Islam or Christianity are no longer recognized as SC for the purpose of affirmative action
policies. According to the 2011 Census Ati-Shudras form 16.2% of the Indian population and in our
analysis they are referred to as Schedule Caste.

The next group is Adivasi, which translates to original inhabitants. Over a period of time some
Adivasis have been co-opted into the caste system alongside Dalits while others have continued to move
further away from the fertile plains and into forests (FĂźrer-Haimendorf, 1982). The social status of
Adivasis is found to be similar to that of the Dalits. Apart from their own tribal religions, which are
grouped under the larger category of Hinduism, they follow Christianity and to a much smaller extent,
Islam. Their current population as per the 2011 Census is 8.2%, which makes them smallest of the
population groups. In this paper they are referred to as Schedule Tribe.

Shudras or Other Backward Caste are the largest of the population group. Their main occupations
were farming, rearing cattle and other forms of skilled labor (Ambedkar, 1970), that were counted as
service to the castes above them in social hierarchy. Like Dalits, most of them are Hindu though many
converted to other religions for a chance at upward mobility. The Indian census does not ask persons
whether they identify as OBCs and there is significant heterogeneity across states in which caste is
officially defined as OBC. In fact, the National Sample Survey (NSS) did not include OBCs as a caste group
till 1999-00, and for this purpose OBCs are included with ‘Hindus’ in our analysis3.

3 Along with a smaller central OBC list, each state is allowed discretion in whom to include or exclude from the
state OBC list. For this reason, the definition of OBC varies from state to state, and it is not uncommon to find the
same caste defined as OBC in one state and upper caste in a neighboring state. Because of the inconsistent definition
of OBC across time and states, we decide to include OBCs under ‘Hindu’ category. Unlike SC/ST, OBC are eligible for
affirmative action policies only if they meet certain income criterion.

4

As indicated above, a large number of Muslims in modern India are former Shudras, Dalits and
Adivasis who converted to Islam. Apart from the caste baggage that the community carried over after
conversion, they face religion based marginalization. The dominant Hindu-nationalist rhetoric defines
Islam and Christianity as foreign religions (Rauf, 2011) and this contributes to their persecution. Their
right to freedom of religion is enshrined in the Indian constitution, which recognizes Muslims primarily
as ‘minorities’. In this paper they are referred to as Muslims.

3

Review of Existing Literature

3.1

Economic exclusion and caste in India

Much of the literature on economic discrimination in India has been centered on caste based
discrimination. Indeed, Scheduled Caste and Scheduled Tribes continue to be one of the most deprived
social groups in the country and still lag behind upper caste Hindus and other religions on indicators of
social and economic development. The relative as well as absolute deprivation of SC/ST is widely
documented (Hanna & Linden, 2012; Ito, 2009; Zacharias & Vakulabharanam, 2011). According to
Planning Commission statistics for 2009-10, in urban areas, ST and SC have a poverty rate of 30.4% and
34.1% respectively compared to 12.4% for Hindus and 21% for India as a whole. Similarly in rural areas,
compared to an overall poverty rate of 33.8%, ST and SC have a head count ratio (HCR) of 47.4% and
42.3% respectively (Government of India, 2012). Borooah (2005) decomposes the differences between
Hindu and SC/ST households on income levels, incidence of poverty and levels of poverty into a residual
effect which accounts for the difference between the ‘income generating’4 profile of the SC/ST and Hindu
households, and a discrimination effect. He finds that at the minimum, one-third of the difference in income
across households is attributable to discrimination or ‘unequal treatment’ of the SC/ST households.

Discrimination and exclusion in the Indian context comes from a multitude of institutions which
‘discriminate, isolate, shame, and deprive subordinate groups on the basis of identities like caste, religion

4

Income generating profile refers to the household’s ownership of land, regional location, non-land assets, number
of adult workers and household head’s education level

5

and gender’ (Thorat & Newman, 2010). Based on a field study which involved sending out three
applications in an upper caste Hindu, Dalit and Muslim name for advertised private sector jobs, Thorat
and Attwell (2007) find that Muslims elicited the least favorable response from potential employers,
followed by Dalits. On the other hand, in a similar experiment Banerjee et al.(2009) find no conclusive
evidence of caste or religion having an effect on the callback decision of software and call-center firms in
and around New Delhi, India. Deshpande (2011), in perhaps one of the most comprehensive analyses of
economic discrimination experienced by Dalits in recent times, shows that caste discrimination cuts across
all levels of employment, right from the market for unskilled labor to the market for graduates from elite
education institutions (ibid., p190). The book covers a gamut of issues from occupational distribution,
wages, education, asset ownership and landholding, and develops a CDI (Caste Development Index) using
various rounds of the National Family Health Survey, on the lines of the Human Development Index
developed by the UNDP. The analysis shows that in almost all the states and across all three rounds of
the NFHS, the CDI for SCs is consistently lower than that for non-SC/ST and the same is the case for
STs, though they fare marginally better than SCs. She finds considerable variation across states both
within and across rounds, and finds that the CDI for SCs is positively correlated with per capita real State
Domestic Product (NSDP), indicating their better performance in economically better performing states.
However, she finds that the level of inequality between groups is not correlated with state income. She
also finds no clear relationship between SC-CDI disparity and real NSDP growth rate. In a very telling
analysis of the relationship between caste and gender in labor markets, Deshpande (ibid.) finds that the
GCDI (Gender Caste Development Index), follows a pattern similar to the CDI. However, it does not
have any correlation with per capita real NSDP, indicating that the better economic performance of a
state does not necessarily translate to increased well-being for Dalit women.

Wage disparity between different groups is perhaps one of the most widely used measures of economic
discrimination. In India, the method of wage decomposition has been used to estimate the wage gap
between various caste groups, and more recently, religions as well (Bhaumik & Chakrabarty, 2006).
Banerjee and Knight (1985) find that the wage gap between SC and non-SC in their sample is almost 17%,

6

and of this they estimate, more than half is on account of discrimination. A more recent study by
Madheswaran and Attwell (2007), arrives at slightly different results. They estimate that a major part of
the difference in wages between SC/ST and non-SC/ST is due to differences in endowments, while 15%
is due to pure discrimination. They argue that the gap in wages is primarily due to differences in the kinds
of occupations that SC/ST and others hold, rather than pure wage discrimination. They however find
that returns to education for SC/ST are considerably lower than for other groups and that wage
discrimination exists in both public and private sector, and is much higher in the private sector.

3.2

Indian Muslims

The status of Indian Muslims has been a subject of historical (Hardy, 1972), sociological (Fazalbhoy,
1997; Robinson, 2004), political (Brass, 2005; Rauf, 2011) and economic (Khalidi, 2006) enquiry for many
decades. Until very recently, the discussion on Indian Muslims revolved around issues of terrorism,
communal violence and personal law, while the economic and social deprivation of the community was
rarely discussed. The setting up of the Prime Minister’s High Level Committee on Social, Economic and
Educational Status of the Muslim Community of India (more commonly known as the Sachar Committee)
in 2005, was perhaps the first official attempt at recognizing and studying the existence and extent of the
backward status of the Muslim community in India. The committee submitted its report in November
2006 and became the reference for studying the various dimensions that define Muslim existence in
modern India, from violence, exclusion and education to employment, health and poverty. In addition,
recently released official poverty estimates of the Planning Commission indicate that Muslims are by far
the most economically deprived religious community in India, across rural and urban areas. In rural
Assam, West Bengal and Uttar Pradesh, the headcount ratio (HCR) for Muslims stands at 53.6 (39.9),
34.4 (28.8) and 44.4 (39.6)5. Similarly in urban areas, Muslims have a high HCR in states such as Gujarat42.4 (17.7), Bihar- 56.5 (39.4) and Uttar Pradesh- 49.5 (31.7) (Government of India, 2012). That these
figures are high, even when using the controversially low poverty lines defined by the Planning
Commission, highlights not only the relative, but also the absolute deprivation faced by Indian Muslims.

5

Figures in parenthesis indicate overall state HCR

7

Bhaumik and Chakrabarty (2006) examine the difference in wages for Hindus and Muslims over the
period between 1987 and 2005 using NSS data. Similar to studies on Dalits, they find that differences in
education are the primary cause for differences in wages between the two groups. A large part of this is
due to the difference in the proportion of wage earners in the two groups with tertiary education. It will
be seen later in this research that this is not only due to the fact that a relatively lower proportion of
Muslims have tertiary education, but also because only a small proportion of them are engaged in higher
paying regular/salaried employment. This is corroborated in a study (Borooah et al. 2007) of the 55th
round of NSS which finds that compared to upper caste Hindus, Muslim workers have a significantly
lower probability of being in regular salaried employment. It will be shown in this paper that this is even
more the case for Muslim women than for Muslim men.

Muslim women have largely been absent from realm of economic enquiry. Hasan and Menon (2005),
best sum up the treatment of Muslim women in academic literature (p.3), ‘the literature on Indian women in
general is characterized by three broad tendencies: it ignores Muslim women, considers their status a product of
personal laws, and assumes sameness in the status and forms of oppression, cross-community’6. The assumption of
sameness of women across groups, religions and communities, has received considerable attention in the
works of the so-called ‘feminist economists’ who developed the Gender and Development (GAD) approach
to development as a response to the WID (Women in Development) approach championed by Esther
Boserup (1970). The GAD framework for the first time accepted the heterogeneity among women, and
recognized that patriarchy operates both within and across class and other divisions to oppress women.

The assumption of homogeneity of Indian women and the over-emphasis on personal law and issues
of religious identity in the context of Muslim women has meant that their socioeconomic, educational and
political status remained understudied in India. A survey, called the Muslim Women Survey (MWS), by
6 The Indian constitution allows religious minorities to follow their own religious laws concerning issues of marriage,
divorce, division of property and succession. Indian Muslims come under a version of Sharia law that is governed by
the All-India Muslim Personal Law board. While several successive governments have proposed the creation of a
Uniform Civil Law Code, it remains a matter of considerable political contention.

8

Hasan and Menon (2004), spread over 9541 households over 12 states was conducted to study the
situation of Muslim women in India, covering a gamut of issues such as: education, work, socio-economic
status, marriage, mobility, access to media, political participation, domestic violence and decision-making.
The findings of the MWS are largely consistent with the trend observed at the national level in the NSS
data, particularly in the area of education and employment. The study finds evidence of high drop-out
rates and very low participation in higher education. Moreover, it shows that southern and western states
perform better on indicators of education than northern and eastern states, and this is consistent with the
findings of this research. The study finds very low levels of work participation and a concentration of
women in self-employment and low-skill sectors. In an analysis of the 1993-94 round of the NSS, BordiaDas (2005) corroborates the findings of the MWS, by concluding that the major reasons for low work
participation of Muslim women is their limited engagement in agriculture in rural areas, and their
‘exclusion from professional, technical and clerical jobs’ in urban areas. The concentration of Muslims in
certain sectors, areas, and kinds of work is corroborated in another recent study (Bordia Das, 2008) which
found clear evidence of the presence of ethnic economic enclaves for Muslim men.

4

Methodology
My primary outcome of interest is the mean female labor force participation rate across the three

caste-religion groups in India and its evolution over the last 25 years. The aim of this paper is to explore
the contribution of observable background characteristics in explaining the LFPR gaps between the three
groups. My methodology is twofold- I first examine the combined role of all covariates in explaining the
LFPR gaps and this gives the total, predicted and unexplained gap. I then look at the role of individual
characteristics in explaining the total and the predicted gap. I use two methods for the overall
decomposition analysis- DFL method based on DiNardo, Fortin and Lemieux (1996) and the OB method
based on Oaxaca (1973) and Blinder (1973). For assessing the role of individual characteristics I use an
extension of the OB method. Observable covariates include age, marital status, level of education, rural
residence indicator, number of children below the age of 5, household size, household head characteristics,

9

and controls for local economic conditions. In addition, the model also includes a full set of controls for
state of residence to capture regional variations in the economic participation of women.

4.1

DFL method

Reweighting methods continue to be a popular tool in economics to study gender, ethnicity and race
based gaps in various health and labor market outcomes7. The intuition behind reweighting is simple- if
we want to study outcome differences between group L and M, we reweight group L so that its
distribution of observables closely matches that of M, while retaining its own mapping from observables
to outcome. Essentially, we give more weight to those Ls whose observables are similar to the Ms in our
sample and progressively less weight to the Ls whose background characteristics are different from the
Ms. In our sample, the L signifies Hindu women while M refers to SC/ST or Muslim women.
To begin, define the probability density for outcome y of caste/religion group c with background
characteristics X as:
(1)

𝐹(𝑦|𝑐) = ∫𝑥 𝐹(𝑦|𝑐, 𝑋)𝑑𝐹(𝑋|𝑐)

From here it is easy to construct a valid counterfactual density of the following form:
𝐹(𝑦|𝑐𝑦|𝑋 = 𝐿, 𝑐𝑋 = 𝑀) = ∫𝑋 𝐹(𝑦|𝑐 = 𝐿, 𝑋)𝑑𝐹(𝑋|𝑐 = 𝑀)

(2)

This counterfactual density is valid only when the changing the marginal distribution function from
𝑑𝐹(𝑋|𝑐 = 𝐿) to 𝑑𝐹(𝑋|𝑐 = 𝑀) leaves the conditional distribution 𝐹(𝑦|𝑐 = 𝐿, 𝑋) unchanged. Using
weights of the form 𝜑𝐿→𝑀 (𝑋) the counterfactual density in (2) can be written as
𝐹(𝑦|𝑐𝑦|𝑋 = 𝐿, 𝑐𝑋 = 𝑀) = ∫𝑋 𝐹(𝑦|𝑐 = 𝐿, 𝑋)𝜑𝐿→𝑀 (𝑋)𝑑𝐹(𝑋|𝑐 = 𝐿)

(3)

Where using Bayes’ rule, we can write 𝜑𝐿→𝑀 (𝑋) as
𝑑𝐹(𝑋|𝑐
𝑑𝐹(𝑋|𝑐

= 𝑀)
= 𝐿) ≡

Pr(𝑐=𝑀 |𝑋)
Pr(𝑐=𝐿|𝑋)

∗

Pr(𝑐=𝐿)
Pr(𝑐=𝑀)

(4)

I use logit (any binary model can be used) to calculate Pr(𝑐 = 𝑖 |𝑋), which is the probability of being
from caste/religion group i as a function of background characteristics X. The second part, that is

7 For a comprehensive review of decomposition methods in economics see Fortin et al. (2010). For an application
of DFL method to studying racial gaps in IMR see Elder et al. (2011)

10

Pr(𝑐 = 𝑖), is simply the unconditional proportion of population in group i. In constructing the
counterfactual weights, I use data for Hindu women pooled with the corresponding comparison group
(SC/ST or Muslims), and then obtain three set of means: mean Hindu LFPR (H), mean SC/ST or Muslim
LFPR (C) and mean counterfactual Hindu LFPR (H’). The gaps in labor force participation are calculated
as follows: Total Gap (T) = H - C, Explained Gap (E) = H - H’ and Unexplained Gap (U) = H’ - C.

4.2

Oaxaca-Blinder method

The Oaxaca-Blinder (OB) method for wage decomposition is the first and one of the most well-known
decomposition techniques in economics. The original model used differences in means to break down
male-female wage gaps into an explained component that can be attributed to differences in endowments
such as education and experience and a residual component that is attributed to discrimination (and other
unobservable characteristics).
In keeping with the notation developed in the previous section, let the groups be denoted by L and M.
I am interested in knowing how much of the total gap T can be explained by the covariates X and how
much remains unexplained.
𝑇 = 𝐸(𝑦𝐿 ) − 𝐸(𝑦𝑀 )
Where 𝑦𝑐 = 𝐹(𝑋′𝑐 𝛽𝑐 ) and F is a mapping of X to y. In our case, the mapping is linear, that is, 𝑦𝑐 =
(𝑋′𝑐 𝛽𝑐 ). According to the OB model, we can estimate T by using the sample differences in mean values
of 𝑦𝐿 and 𝑦𝑀 . More formally
̅̅̅̅̅̅̅̅̅̅
̅̅̅̅̅̅̅̅̅̅̅
̅̅̅̅̅̅̅̅̅̅̅ ̅̅̅̅̅̅̅̅̅̅̅
𝑇 = ̅̅̅
𝑦𝐿 − ̅̅̅̅
𝑦𝑀 = [(𝑋′
𝐿 𝛽𝐿 ) − (𝑋′𝑀 𝛽𝐿 )] + [(𝑋′𝑀 𝛽𝐿 ) − (𝑋′𝑀 𝛽𝑀) ] = 𝐸 + 𝑈

(5)

Here the first part is explained due to differences in characteristics, whereas the second part is
unexplained or due to differences in coefficients or returns to characteristics. An extension of the OB also
allows us to calculate the role of individual covariates (or group of covariates) in predicting the explained
𝑍
𝑍
̅̅̅̅̅̅̅̅̅̅
̅̅̅̅̅̅̅̅̅̅̅
gap. These estimates are calculated as [(𝑍′
𝐿 𝛽𝐿 ) − (𝑍′𝑀 𝛽𝐿 )] where 𝑍𝑖 is a subset of the variables from X

and 𝛽𝐿𝑍 is the corresponding coefficient from regressing y on X for group L.

11

5

Data and Summary Statistics
I use data from the 38th (1983-84), 43rd (1988), 50th (1993-94), 55th (1999-00), 61st (2004-05), 66th

(2009-10)8 and 68th (2011-12) rounds of the National Sample Survey on Employment and Unemployment
in India. These nationally representative surveys are usually conducted every five years to provide a
comprehensive assessment of the labor market situation in the country. Each round covers between
100,000-120000 households and 450,000-600,000 individuals.

My working sample is comprised of women between the ages of 15-59 belonging to 3 mutually
exclusive categories- Hindu, Schedule Caste/Schedule Tribe and Muslim. The Hindu category consists
of all those who report being Hindu but not SC/ST. The Muslim category contains all Muslims regardless
of caste. And finally, SC/ST includes all non-Muslim SC/ST persons. I drop all observations who are
non-SC/ST or non-Hindu/Muslim9. My measure of labor force participation rate is based on a one-year
reference period and includes work in a subsidiary status10. I drop observations for which information on
caste or religion is missing11.

To account for regional variation, I include a full set of state dummies. I restrict the sample to 15
major states of India that cover roughly 92-95% of India’s population. Three new states of Jharkhand,
Uttaranchal and Chhattisgarh were carved out of the states of Bihar, Uttar Pradesh and Madhya Pradesh
respectively, in 2000. To ensure comparability across rounds, I treat the new states as a part of the old
state resulting in a total of 15 states12. I also include controls for local economic conditions by controlling
8 Significant changes have been undertaken in designing the surveys beginning with the 66 th round (2009-10). One
of the most important changes was a marked reduction in the sample size as well as reclassification of rural and urban
areas. Moreover, 66th round was conducted in a year when many areas of the country experienced acute drought, which
may affect labor market outcomes. We thus need to proceed with caution when comparing results of the 66 th round
with previous rounds.
9 This group includes non-SC/ST members of religious minorities such as Sikhs, Christians, Jains, Buddhists and
Others. They represent roughly 4.9 percent of our sample.
10 The NSS defines this as usual status. This includes the work done as primary status in the last 365 days as well
as work done in a subsidiary status for at least 30 days in the last year. Women who reported the following activities
were counted as being in the labor force: worked in HH enterprise (self-employed) as own account worker or employer,
worked as helper in HH enterprise (unpaid family worker), worked as regular salaried/ wage employee, worked as casual
wage labor in public works or in other types of work, sought work, did not seek but was available for work
11 We drop 21/120909 households in 38 th round, none in 50th and 55th round (115409 and 120386 households
respectively), 96/124680 households in 61st Round and 87/100957 households in 66th round for missing data on caste
and/or religion.
12 The states included in our analysis are Uttar Pradesh, Bihar, Maharashtra, Madhya Pradesh, West Bengal,
Andhra Pradesh, Tamil Nadu, Karnataka, Gujarat, Rajasthan, Orissa, Kerala, Assam, Punjab and Jammu & Kashmir.

12

for adult male unemployment rate and the proportion of men with a secondary school education or higher
in the respondent’s region. I also drop all observations that are missing information on any of the
covariates such as monthly per capita income, age, years of education, household size13. The final sample
consists of 140,000 to 180,000 women per round.

I control for level of education by dividing it into five mutually exclusive categories based on highest
level of completed schooling-no education, incomplete primary school, complete primary school, middle
school and secondary and higher14. Controls for household head’s education are similarly constructed.
Controls for age of respondent and age of head are also added.

Concurrent with recent literature, household income and wealth is believed to be an important
component driving the decision of women to participate in the labor force. The NSS does not report
measures of income but does provide monthly household per capita consumption expenditure. However,
controlling for consumption directly would create endogeneity. I work around this problem using two
solutions. I first control for characteristics of the household head and male household members that I
believe proxy for some measure of permanent household income. In the second specification, I also add
predicted consumption expenditure rather than actual expenditure to more directly control for measures
of income.

A caveat of the NSS data is that each of the respondent is asked her/his relationship to the household
head only, hence, unless the adult female in the house is the head or spouse of the household head it is
impossible to determine exactly how many children in the household are her own. This is particularly
problematic in the case of joint families which are very prevalent in India, where children live in the same
household with the mother, grandmother and aunts. One solution to that would have been to include only

13

Less than 1 percent of our sample is dropped by excluding observations with missing data on any of our covariates.
This method does not account for technical education because technical education is reported by a very small
proportion of the population. Moreover, the technical education question has changed over the rounds making it
inconsistent across all our sample years. We do however try an alternative specification that converts level of education
(general and technical) into years of education and it only reinforces our main finding that covariates explain a sizeable
portion of the Hindu-SC/ST gap, but very little of the Hindu-Muslim Gap. Results available on request.
14

13

those women who report themselves to be heads or spouses of the head. They represent between 60-61%
of our working sample. However, this would have led to excluding a vast number of women who live in
joint families and whose labor market behavior may be quite different from women living in nuclear
families. To control for the effect of fertility decisions on labor market behavior, I include as a covariate,
the total number of children below 5 in the household. This is reasonable since child rearing
responsibilities in India are often shared among all women of the household when living in joint families.

The premise of the paper is that there are substantial differences in the LFPR of women from different
communities, and this is very evident from Figure 1. While on the one hand, SC/ST women have the
lowest indicators of human capital, they have the highest LFPR. On the other hand, Muslim women, who
fare similarly on indicators of education and income, have the lowest LFPR. As Figure 1 shows, this
disparity between groups is particular to women, since men have similar levels of LFPR across all the
three groups. Until the 61st round, the LFPR of women was steadily increasing, however it has gone
down substantially in the last two rounds. Some of this has been attributed to the fact that younger women
are staying in school/college longer, which does seem to be true. However, there has also been a
disproportionate increase in the number of women who are withdrawing from the labor force in favor of
domestic work, and this is a secular trend across all groups.

Historically Muslims have been involved in crafts and artisanal trade and this is mirrored in the fact
that a majority of working Muslim women are ‘self-employed’, most of them in home based micromanufacturing units. On the other hand, SC/ST women are disproportionately represented in casual
labor, most of which is in the rural agriculture sector. These clear trends in occupational distribution are
reflective of the fact that landholding in India has been concentrated in the hands of the upper caste, which
in the absence of substantive land reforms, has led to persistent ethnic employment enclaves. While
Muslim women have lowest LFPR, they also have the highest rate of unemployment, which incidentally
is the highest across all groups and genders. In our sample, Hindu and SC/ST women had an
unemployment rate of 4% in 2009-10, whereas the same for Muslim women was 7%.

14

The persistently low levels of education of women in India has been a major concern for policymakers.
However, our data shows that there is considerable heterogeneity across the three groups of women as
well. Descriptive statistics in Table 1.9-1.11 show that Hindu women continue to have the highest levels
of education. On the other hand, while SC/ST women have made remarkable improvements over the
years, they still have the lowest levels of education and high rates on illiteracy. Finally, Muslim women
who were previously performing significantly better than SC/ST women are now placed only marginally
better on indicators of education. According to the 66th Round, Hindu women had an average of 5.85
years, Muslim women had 4 years and SC/ST women had 3.5 years of education.

Household level statistics also show considerable heterogeneity across groups. Between 83-86 percent
of SC/ST households live in rural areas, whereas Muslims are the most urbanized community. Muslim
households have the most members and have the highest number of children under 5, which is indicative
of the community’s higher fertility rate and lower child and infant mortality (Bhalotra, 2009). Muslim
tend to live in areas where male unemployment rate is higher, and more than 80% Muslim and SC/ST
women live in households with no adult male engaged in regular salaried employment.

6

Role of All Covariates
For analysis based on the DFL method, I construct a counterfactual Hindu distribution that has the

characteristics of SC/ST and Muslims respectively, but retains its own mapping from background
characteristics to labor force participation behavior. Figures A1-A4 show that the reweighted Hindu
population looks quite similar to its corresponding reference group in terms of the distribution of
covariates. The counterfactual (CF) Hindu distribution is marginally younger than the SC/ST sample. In
both cases the CF population has slightly higher years of education, and smaller HH size compared to the
reference group. On all other indicators, our CF, on average, looks almost identical to the corresponding
comparison group. The gap between actual Hindus and the counterfactual Hindus is what I call predicted
or explained gap, whereas the gap between counterfactual Hindus and the comparison group is called the

15

unexplained gap. Point estimates based on DFL method are shown in Tables 1 and 2, and point estimates
based on OB method are shown in Tables 3 and 4. As is evident, both the DFL and OB method give
similar results for both groups. I report OB results with an alternative specification that includes predicted
income along with other proxy controls for household wealth. The results of this alternative specification
are in Tables 5 and 6. The alternative specification gives us results that are qualitatively similar to the
main specification.

Gap by Caste: Between 46-57 % percent of the gap in LFPR between SC/ST and Hindu women can be
explained by differences in covariates. What this means is that if Hindu women had the same distribution
of characteristics as SC/ST women, their LFPR would be between 5 to 8.7 percentage points higher than
it is. As an example, column 1 of Table 1 shows that the Hindu counterfactual LFPR is 51.5 percent,
which is the Hindu LFPR if they had the SC/ST characteristics on average. Compared to this, the actual
Hindu LFPR is 42.9%. This means that the gap in Hindu-SC/ST LFPR which can be predicted by
differences in observables is 8.6% compared to the overall gap of 18.1%. Table 2 shows that the results
using the OB method are also quite similar. Roughly 8.5% of the total gap between Hindu and SC/ST in
1983-84 is explained by differences in observables.

Lower caste women are overwhelmingly concentrated in rural areas, where opportunities to be
employed as unskilled marginal and casual labor are abundant. This is the sector that employs the highest
proportion of SC/ST women, and if they are increasingly getting pushed into low paying casual labor,
then the higher labor force participation of lower caste women could be a sign of economic distress rather
than economic empowerment. As Figure 2 shows, relative to total gap, the explained component of the
SC/ST gap has also been steadily, albeit slowly, increasing over the years, and this underscores the
importance of social and human capital indicators in predicting economic outcomes for lower caste
women.

16

The absolute gap in LFP between Hindu and SC/ST women has also been declining quite significantly
over the years. In 1983 the LFPR of SC/ST women was 18% higher than Hindu women, however by
2009-10 this had dropped to only 10%. When I include measures for predicted consumption expenditure,
the explained portion of the Hindu-SC/ST gap rises to almost 85% in 1999-00. Household income seems
to have a significant positive association with female labor force participation for SC/ST women, an
argument I will revisit in the next section.

Gaps by Religion: Muslims fare similarly to SC/ST on indicators of human development and poverty,
however the labor market response of Muslim women to these indicators is markedly different. If Muslim
women’s labor market behavior was similar to SC/ST women, we would expect Muslim LFPR to be
higher than Hindu LFPR, however the gap between Hindu and Muslim women is large and positive, the
opposite of what was seen between Hindu and SC/ST women. Also in contrast to lower caste women,
observables explain a very small part of the LFPR gap between Muslim and Hindu women, and this
explained component has been declining over the years from 18% in 1983 to 8% in 2009. Again, using
1983-84 as an example, column 1 of Table 2 shows that the Hindu counterfactual LFPR is 39.1% compared
to the actual Hindu LFPR of 42.9%. This implies that of the total gap of 16.8%, only 3.9% is explained by
differences in observables between Hindu and Muslim women. Table 4 reports the results using the OB
method and gives similar results, 3% of the Hindu-Muslim gap is explained.

The absolute gap between Hindu-Muslim LFPR has fluctuated significantly over the years, rising to
its highest level of 18 percentage points in 2004-05. If Hindus in our sample had the Muslim distribution
of characteristics, their LFPR would have been 0.5 to 4 percentage points lower than it is. As we will see
in the next section, almost the entire explained portion of the Hindu-Muslim gap can be attributed to
regional concentration of Muslims in low female LFPR states.

If low human capital indicators and economic deprivation predict high LFPR for SC/ST women, why
do these same characteristics have little to no power in explaining the low LFPR of Muslim women? One

17

reason for this paradox could perhaps be attributed to the presence of ethnic economic enclaves. Many
Muslims, particularly in rural areas, are involved in craft and artisanal trade, and traditionally the ‘skill’
of the craft is passed on to sons and other male family members rather than to daughters. Low levels of
education, combined with little to no specific ‘skills’, severely restrict the employment options available
to Muslim women. Moreover, there is some evidence to suggest that women working in home based
enterprises systematically under report their economic activity because they themselves don’t place
economic value on their work. Women working as a part of their husband’s or father’s small business may
consider such work to be part of their domestic duties, particularly because home based work does not
involve a fixed wage or salary. Finally, religious traditions and customs could also explain Muslim
women’s absence from the labor force. I explore each of these hypotheses in the next section.

7

Role of Individual Covariates
The results from the previous section show that despite similar levels of economic and human capital

development, Muslim and SC/ST women exhibit vastly different labor market behavior. Given these
paradoxical results that emerge when looking at the role of all covariates combined, it becomes important
to see what contribution individual covariates have in explaining LFPR gaps between the different groups.
I use an extension of the Oaxaca-Blinder method to examine the role of individual covariates. For an
easier and more intuitive interpretation of results, I group the covariates intro 3 categories: personal,
household and regional characteristics. The personal category includes age, level of education, and marital
status, household category includes household size, number of children under the age of five in the
household, indicator for adult household male characteristics and household head’s characteristics and the
regional category includes rural residence indicator, regional adult male outcomes as well as state fixed
effects.

Gaps by Caste: Together personal and household characteristics account for 80-90% of the explained
gap for Hindus and SC/ST, whereas regional variables play a smaller role. As an example, if Hindu women
had the personal characteristics of SC/ST women in 2004-05, their LFPR would have been higher by 3.2

18

percentage points. Overall personal and household characteristics, respectively, explained 48% of the total
gap in 2004-05, whereas this was 38% in 1983-94. This shows that personal and household characteristics
are increasingly playing a larger role in explaining the labor market behavior of SC/ST women. This
could be, at least in part, due to the rapid convergence in levels of education between SC/STs and Hindus.
While SC/ST levels of education are still the lowest among all groups, they have been the fastest to
increase over the period under study. This convergence has been aided by affirmative action policies that
guarantee 22.5% seats in higher education and public employment for SC/STs at the state and federal
level.

On the other hand, given that a vast majority of SC/ST women continued to be employed in the
agriculture sector, it become imperative to examine whether female LFPR among SC/ST and Hindus is
being driven by economic opportunities or economic distress. To this effect, I examine how female LFPR
for the three groups is associated with change in income. If the decision to enter the labor force is indeed
driven by economic need, then we would expect LFPR to fall as incomes increase. Moreover, the decline
will be sharper for those groups for which the relationship is stronger. To test this, I divide real income15
into 250 quantiles and aggregate female LFPR for the three groups within each quantile. Figure 4 shows
that both SC/ST and Hindu women’s LFPR has a strong negative correlation with income whereas for
Muslim women the relationship is relatively flat. The relationship has become more pronounced over the
years. Descriptive statistics showed that female LFPR in India has been declining in recent years. The
graph suggests two reasons why that may be happening- female LFPR is becoming more strongly
associated with increases in income, and real incomes themselves are rising- both of which are correlated
with lower rate of female employment.

Gaps by Religion: In contrast to SC/ST, the role of personal and household characteristics is almost
negligible in explaining the Hindu-Muslim gap in LFPR. Regional characteristics more than account for
15 We use household monthly per capita expenditure as a proxy for household income. Nominal values are converted
to real values using publically available poverty lines. We use 1983 rural Maharashtra as the base. The method of
calculating these poverty lines changed significantly in 2009-10, hence we restrict our analysis from the 38th to the 61st
round to maintain consistency.

19

the entire explained gap between the two groups. On closer examination of the sample, the reason for this
becomes clearer. More than half of the Muslim women in the sample come from the states of Bihar, Uttar
Pradesh and West Bengal, whereas only 30% of Hindu women live in these states. However, it is in these
three states that Hindu women have the lowest LFPR, hence when I give more weight to the Hindu
women from these states, I close some of the gap between Hindus and Muslims. Figure 5 offers suggestive
evidence for this hypothesis. I plot female LFPR for the two groups as a function of the proportion of
Muslims in the region16. If female LFPR is depressed in areas where concentration of Muslims is high,
we would expect LFPR to be negatively correlated with concentration of Muslims. Figure 5 confirms that
areas with higher concentration of Muslims exhibit lower levels of female participation for all groups, not
just Muslims. Descriptive statistics show that along with exhibiting higher rates of unemployment,
Muslim women are also more likely to live in areas with higher male unemployment rate compared to the
other two groups. This suggests that Muslim women are more likely to live in areas where opportunities
for employment are low for both women and men. According to the Indian census, a majority of women
in India move within the same district or state at the time of marriage. Such localized marriage markets
could mean that Muslim women remain trapped in enclaves of low economic activity.

Without this regional heterogeneity, the gap between these two groups would have remained the
same and would even have widened in some cases. For example, according to the 38th Round, if Hindus
had the personal characteristics of Muslims, their LFPR would have been 1.2 percentage points higher
than it was. This is similar to the result I saw in the Hindu-SC/ST comparison, where less educated and
poorer Hindu and SC/ST women are more likely to participate in the labor force. Thus covariates,
excluding spatial distribution, explain little to none of the reason behind the significantly lower LFPR of
Muslim women in India.

16 Each circle represents a sector-state-year combination and the size of each circle is proportional to the
corresponding population weights. There are 2 sectors- rural and urban, 15 states and 7 years of data available.

20

8

Extensions

8.1

How does culture influence participation gaps?

In this section, I examine the role of culture in explaining the persistence or absence of Muslim women
from the labor force. Das (2005) presents some anecdotal evidence from the Muslim Women’s Survey that
points to the practice of purdah or veiling among Muslim women as a factor in restricting their mobility
for purpose of employment. She does however conclude that there is considerable heterogeneity across
class divisions in the practice of purdah, and that lack of jobs (particularly agriculture based) outside the
house contribute to Muslim women’s concentration in home-based work. While most of the evidence on
the role of culture remains anecdotal and descriptive, I try to offer some empirical evidence on this issue.
The NSS does not collect any information on intra-household decision making, mobility or violence, so
for this section I use data from the India Human Development Survey. The IHDS is a two panel nationally
representative survey and I use the second wave (2011-12) of the survey. I am able to construct almost
all the controls used in the main analysis, along with some controls for “culture”. An advantage of this
dataset is that it asks for measures of more permanent household wealth (such as ownership of durables
and assets), along with flow variables such as income and consumption. I use a composite “asset” measure
to control for HH wealth. This variable takes values from 0 to 33 and is a simple sum of assets owned by
the household. I construct various measures for the culture variable using the “gender relations” part of
the women’s questionnaire. I divide gender relations this into 3 parts- mobility (0-4), attitude towards
violence (0-6) and decision making (0-6). Each of these is constructed using answers to questions about
these variables (higher is better)17. I also control for whether the woman observes covering of head/face
and whether the woman says work decision is made by self or husband/someone else. In total, I have 5
controls for culture. Decomposition results from Table 7 show that while culture does play a small role
in explaining the LFPR gap between Hindus and Muslims, the role of culture is not large enough to
explain the persistently low LFPR for Muslim women. For example, column 2 of Table 7 shows that of
the total Hindu-Muslim gap of 15.5 percentage points, culture explains only 1.8 percentage points. Much

17 As an example, the question on mobility asks- Can you go to health clinic by your self- YES (1) or NO (0). There
are 4 such questions. If a woman answers YES for all 4 she gets a mobility score of 4, if she answers YES for two NO
for two her mobility score is 2.

21

of the Hindu-Muslim gap is explained by regional variables, as was seen in the main analysis using NSS
data. For Muslim women, culture explains 28% of the explained and 11% of the overall Hindu-Muslim
gap in female labor force participation.

8.2

Do alternative mappings change the results?

In the main analysis, for both DFL and Oaxaca-Blinder, I have reweighted Hindus to look take on the
individual characteristics of Muslims or SC/STs. However, decomposition methods can be sensitive to
the choice of the reference category or group. To test the generalizability of my results, I can use
alternative reweighting methods and see how the results change with the alternative reference groups. In
this section, I present two alternative reference categories. In the first case, I use SC/ST (or Muslims) as
the reference category to compare with Hindus. In the second case, I use a pooled sample of the two
groups to create a composite reference category. The results using SC/STs and Muslims as the reference
category are reported in Tables A4 and A5 respectively, whereas the results using the pooled sample of
Hindu-SC/ST and Hindu-Muslim is reported in Tables A6 and A7.

The results show that for SC/STs, the predicted gaps are quite similar across the different reference
category choices, including when comparing with the predicted gaps obtained using DFL. On the other
hand, the results for Muslims are slightly sensitive to the choice of reference category. For Muslims, all
methodologies follow a similar pattern across the seven rounds of data, however they differ slightly in the
magnitude of the predicted gap based on which reference category is used. This is not unexpected since
we know that Muslim women’s participation rate is less sensitive to background characteristics, and is
also different in magnitude, compared to the other two groups. For example, figure 5 shows that while
Hindu and SC/ST women’s LFPR has a strong negative association with increases in income, Muslim
women’s participation rate is quite flat with respect to changes in real income.

The role of individual covariates is also largely similar for the different reference category choices.
Personal and household characteristics accounts for majority of the explained gap between Hindus and

22

SC/STs, whereas regional variation does not play a major role. On the other hand, for Muslims both
personal and household characteristics predict a negative gap, whereas regional variation account for
more than the entire predicted gap. As an example, whether we use the Hindus or Muslims as a reference
category, convergence in regional variation would reduce the Hindu-Muslim gap in 1983 by 4.9
percentage points. On the other hand, if we used the pooled reference category, convergence in regional
heterogeneity would reduce the Hindu-Muslim gap by 5.7 percentage points.

8.3

How do results differ in rural and urban areas?

There has been considerable recent debate and discussion about the causes and consequences of
stagnation of female participation in India’s urban areas (Klasen & Pieters, 2015), and sharp decline in
participation in rural areas (Chatterjee, Murgai, & Rama, 2015; Eswaran, Ramaswami, & Wadhwa, 2013).
Given these differing trends in rural and urban areas, accompanied by the fact that SC/STs are
concentrated in rural areas whereas Muslims are a more urban community relative to the average, it may
be insightful to look at how results vary when considering these geographic differences separately.

In rural areas, the gap in LFPR between Hindus and SC/STs has been steadily narrowing over timefrom 15% in 1983 to 8.9% in 2011. Unlike the aggregated sample, a smaller proportion of this gap is
explained by differences in observable characteristics. This is predominantly because, while personal and
household characteristics continue to explain much of the gap between the two groups, regional
characteristics predict a negative, but statistically insignificant, gap. On the other hand, the LFPR gap
between Hindus and Muslims in rural areas is higher than for the aggregate sample and continued to rise,
before declining in recent years. Like before, all the explained gap is due to regional variation, while
personal and household characteristics predict that the gap should be lower by 2.5-3 percentage points.

In urban areas as well, the gap between Hindus and SC/STs has been steadily narrowing from 12.4%
in 1983 to 3.3% in 2011. Depending on the round, between 35-75% of this gap can be explained by
differences in observable covariates, with household level variables accounting for much of this explained

23

gap. Notably, while the Hindu LFPR has remained stagnant around 21%, the urban SC/ST participation
rate has declined significantly over the years from 34% in 1983 to 24% in 2011. Like Hindus, the urban
Muslim LFPR has also been stagnant around 18-14%. The gap between the two groups has also remained
consistently between 4-6 percentage points. However, covariates predict a negative gap between the two
groups. That is, if Hindus had the characteristics of urban Muslim women, their LFPR would have been
2-5 percentage points higher than it is. This is because urban Muslim women have household level
characteristics that are strong predictors of female participation among Hindus, and this strong negative
isn’t offset by regional characteristics that only explain a small, and statistically insignificant, portion of
the gap.

8.4

Has the predictive power of covariates declined?

The main results show that not only has the gap in LFPR between Hindus and Muslims remained
consistently high, the explained portion of the gap has also declined over time. This could either be due
to convergence in the level of covariates between the two groups, such as equalizing levels of income and
education, or it could be due to a decline in the ability of the covariates to predict participation rates. To
see which of these could explain the declining explained gap between Hindus and Muslims, I run the
decomposition using a fixed Hindu mapping for all the rounds. In Table 1.20, the Hindu mapping from
1983-84 is used as the reference category and in Table 1.21 the 2009-10 Hindu mapping is used as the
reference category. By fixing the mapping to a specific group/year, we are holding constant the
explanatory power of the covariates (from the reference category), while allowing the level of the
covariates to change over time. The results for both the mappings are quite similar in magnitude to the
main results and show a similar declining trend in the explained Hindu-Muslim gap. This indicates that
the results are not driven by the declining predictive power of covariates, but by convergence in covariate
levels.

24

9

Discussion and Conclusion
I use data from 7 rounds of the NSS survey on employment in India to estimate the gaps in labor force

participation rates between women from 3 different caste/religion groups, namely Hindus, Schedule
Caste/Schedule Tribe and Muslims. I use two different decomposition methods to decompose the gap into
a part explainable by differences in background characteristics and an unexplained component.
Additionally, I also examined the role of individual covariates in explaining the predicted gap.

Several findings emerge from these results that can serve as input into formulating better targeted
redistribution policies, as well as input into future research on the economics of inequality in the Indian
context. My analysis showed that the gap in LFPR between Hindu and SC/ST women has been steadily
narrowing over the years, however the gap between Hindu and Muslim women has remained consistently
high. Moreover, while background characteristics explain over half of the Hindi-SC/ST gap, they explain
very little of the Hindu-Muslim gap. My results show that the higher LFPR of SC/ST and Hindu women
seems to be driven in large part due to poverty and economic distress and this is especially true for SC/ST
women. It is hence not surprising that rising incomes are associated with withdrawal of women from the
labor force in recent years across all groups.

Recognizing the need for looking at the role of culture and religion in explaining economic outcomes
for women, particularly from multi-ethnic and diverse countries such as India, I examine the role of
culture in explaining the persistently low LFPR of Muslim women. I find that culture plays a small but
significant role. It is important to recognize that while culture may operate in conjunction with
background characteristics to reinforce conventional economic relationships, it could also represent
behaviors and unobservables that lead to outcomes that are contrary to what economic models predict.

Based on personal and household characteristics we would expect Muslims to exhibit significantly
higher LFPR compared to their own current levels and compared to Hindus. However, I find that Muslim
women’s absence from the labor force cannot be explained by their low levels of education and income.

25

Moreover, contrary to popular rhetoric, culture does not seem to be an important predictor of the low
Muslim female LFPR either. Overall my results show that while personal and household characteristics,
such as low levels of education and income, play a critical role in explaining the high SC/ST female LFPR,
they provide little to no insight into the persistently low rates of labor force participation of Muslim
women. Finally, I find that geographic location plays an important role in explaining economic outcomes
for women. I find that Muslim women’s low LFPR is being driven largely by their concentration in areas
of low economic activity which is also reflected in their high rates of unemployment. If avenues for
mobility outside such areas are restricted by both lack of economic opportunities and cultural restrictions,
then women can become trapped in a cycle of deprivation. These findings highlight not only the markedly
different experiences of deprivation faced by women from minority communities in India, but also their
widely varying response to this deprivation, thus underscoring the need for specifically targeted policies
and intervention that address these needs.

26

APPENDICES

27

APPENDIX A

MAIN TABLES AND FIGURES

28

Table 1.1: Decomposition of Hindu-SC/ST LFPR Gap (DFL Method)

Hindu
LFPR

SC/ST
Hindu CF

Total
Gap

Predicted
Unexplained

38th
1983-84

43rd
1988-89

50th
1993-94

55th
1999-00

61st
2004-05

66th
2009-10

68th
2011-12

0.429
(0.0438)***
0.610
(0.0463)***
0.515
(0.0484)***

0.424
(0.0496)***
0.599
(0.0449)***
0.503***
(0.0532)***

0.412
(0.0503)***
0.586
(0.0480)***
0.495***
(0.0523)***

0.379
(0.0458)***
0.542
(0.0456)***
0.460***
(0.0523)***

0.432
(0.0426)***
0.560
(0.0430)***
0.505***
(0.0460)***

0.317
(0.0387)***
0.416
(0.0422)***
0.369***
(0.0447)***

0.302
(0.0364)***
0.399
(0.0363)***
0.351***
(0.0407)***

-0.181
(0.0309)
-0.0861
(0.0181)
-0.0950
(0.0232)

-0.175
(0.0282)
-0.0798
(0.0158)
-0.0955
(0.0219)

-0.174
(0.0293)
-0.0835
(0.0161)
-0.0910
(0.0224)

-0.162
(0.0223)
-0.0808
(0.0155)
-0.0815
(0.0187)

-0.128
(0.0247)
-0.0727
(0.0149)
-0.0551
(0.0180)

-0.0986
(0.0201)
-0.0518
(0.0126)
-0.0469
(0.0133)

-0.0965
(0.0167)
-0.0484
(0.0122)
-0.0481
(0.0132)

Note: Hindu CF is the Hindu population reweighted to look like SC/ST population, Standard errors are reported in parenthesis. *p-value≤0.10, ** pvalue≤0.05, *** p-value≤0.01

29

Table 1.2: Decomposition of Hindu-Muslim LFPR Gap (DFL Method)

Hindu
LFPR

Muslim
Hindu CF

Total
Gap

Predicted
Unexplained

38th
1983-84

43rd
1988-89

50th
1993-94

55th
1999-00

61st
2004-05

66th
2009-10

68th
2011-12

0.429
(0.0438)***
0.262
(0.0237)***
0.391
(0.0355)***

0.424
(0.0496)***
0.273
(0.0279)***
0.374
(0.0404)***

0.412
(0.0503)***
0.232
(0.0213)***
0.362
(0.0394)***

0.379
(0.0458)***
0.228
(0.0167)***
0.335
(0.0411)***

0.432
(0.0426)***
0.253
(0.0180)***
0.393
(0.0363)***

0.317
(0.0387)***
0.184
(0.0174)***
0.305
(0.0353)***

0.302
(0.0364)***
0.203
(0.0210)***
0.292
(0.0302)***

0.168
(0.0318)
0.0388
(0.0138)
0.129
(0.0212)

0.151
(0.0356)
0.0498
(0.0157)
0.101
(0.0218)

0.180
(0.0345)
0.0499
(0.0132)
0.130
(0.0238)

0.151
(0.0388)
0.0448
(0.0232)
0.106
(0.0363)

0.179
(0.0320)
0.0390
(0.0158)
0.140
(0.0256)

0.133
(0.0273)
0.0119
(0.0158)
0.121
(0.0244)

0.0996
(0.0272)
0.0101
(0.0131)
0.0894
(0.0240)

Note: Hindu CF is the Hindu population reweighted to look like Muslim population, Standard errors are reported in parenthesis. *p-value≤0.10, ** pvalue≤0.05, *** p-value≤0.01

30

Table 1.3: Decomposition of Hindu-SC/ST LFPR Gap (Oaxaca-Blinder Method)
38th
1983-84

43rd
1988-89

50th
1993-94

55th
1999-00

61st
2004-05

66th
2009-10

68th
2011-12

Hindu

0.4294

0.4236

0.4120

0.3794

0.4324

0.3174

0.3025

SC/ST

(0.0409)***
0.6104
(0.0433)***

(0.0462)***
0.5989
(0.0434)***

(0.0460)***
0.5864
(0.0452)***

(0.0421)***
0.5416
(0.0424)***

(0.0399)***
0.5601
(0.0405)***

(0.0358)***
0.4160
(0.0396)***

(0.0334)***
0.3989
(0.0339)***

-0.1811
(0.0305)***
-0.0867
(0.0182)***
-0.0944
(0.0270)***

-0.1753
(0.0282)***
-0.0803
(0.0192)***
-0.0950
(0.0226)***

-0.1745
(0.0278)***
-0.0870
(0.0173)***
-0.0875
(0.0228)***

-0.1622
(0.0231)***
-0.0799
(0.0149)***
-0.0823
(0.0197)***

-0.1278
(0.0248)***
-0.0758
(0.0149)***
-0.0520
(0.0195)***

-0.0986
(0.0204)***
-0.0475
(0.0136)***
-0.0511
(0.0195)***

-0.0965
(0.0173)***
-0.0481
(0.0119)***
-0.0484
(0.0157)***

-0.0301
(0.0057)***
-0.0408
(0.0061)***
-0.0159
(0.0134)

-0.0299
(0.0057)***
-0.0451
(0.0051)***
-0.0053
(0.0153)

-0.0429
(0.0062)***
-0.0408
(0.0043)***
-0.0033
(0.0143)

-0.0326
(0.0052)***
-0.0357
(0.0039)***
-0.0116
(0.0117)

-0.0331
(0.0049)***
-0.0308
(0.0049)***
-0.0118
(0.0108)

-0.0280
(0.0068)***
-0.0157
(0.0043)***
-0.0039
(0.0094)

-0.0238
(0.0054)***
-0.0166
(0.0040)***
-0.0077
(0.0094)

110,851

119,416

106,574

112,356

111,815

87,764

87,346

47.87

45.81

49.86

49.26

59.31

48.17

49.84

LFPR

Difference
Gap

Predicted
Unexplained

Personal
Predicted Gap

Household
Regional

N
Explained (% of total)

Note: Standard errors are reported in parenthesis. *p-value≤0.10, ** p-value≤0.05, *** p-value≤0.01

31

Table 1.4: Decomposition of Hindu-Muslim LFPR Gap (Oaxaca-Blinder Method)
38th
1983-84

61st
2004-05

66th
2009-10

68th
2011-12

0.4294

0.4236

0.4120

0.3794

0.4324

0.3174

0.3025

Muslim

(0.0462)***
0.2730
(0.0272)***

(0.0460)***
0.2323
(0.0219)***

(0.0421)***
0.2285
(0.0182)***

(0.0399)***
0.2529
(0.0189)***

(0.0358)***
0.1840
(0.0175)***

(0.0334)***
0.2029
(0.0223)***

0.1675
(0.0315)***
0.0206
(0.0305)
0.1470
(0.0214)***

0.1507
(0.0336)***
0.0281
(0.0305)
0.1225
(0.0201)***

0.1797
(0.0339)***
0.0295
(0.0286)
0.1502
(0.0205)***

0.1510
(0.0368)***
0.0269
(0.0287)
0.1240
(0.0272)***

0.1795
(0.0314)***
0.0122
(0.0259)
0.1673
(0.0194)***

0.1334
(0.0261)***
-0.0014
(0.0225)
0.1348
(0.0213)***

0.0996
(0.0268)***
-0.0019
(0.0206)
0.1015
(0.0199)***

-0.0127
(0.0044)***
-0.0159
(0.0059)***
0.0493
(0.0259)*

-0.0102
(0.0052)*
-0.0188
(0.0061)***
0.0572
(0.0253)**

-0.0158
(0.0065)**
-0.0196
(0.0055)***
0.0648
(0.0223)***

-0.0086
(0.0061)
-0.0175
(0.0061)***
0.0529
(0.0245)**

-0.0094
(0.0061)
-0.0192
(0.0061)***
0.0408
(0.0213)*

-0.0141
(0.0058)**
-0.0087
(0.0039)**
0.0214
(0.0198)

-0.0081
(0.0045)*
-0.0135
(0.0039)***
0.0197
(0.0189)

99,524

108,986

93,522

99,432

98,912

78,311

80,085

-1.05

-1.91

Predicted
Unexplained

Personal
Household
Regional

N

55th
1999-00

(0.0409)***
0.2618
(0.0241)***

Difference

Predicted Gap

50th
1993-94

Hindu
LFPR

Gap

43rd
1988-89

Explained (% of total)
12.30
18.65
16.42
17.81
6.80
Note: Standard errors are reported in parenthesis. *p-value≤0.10, ** p-value≤0.05, *** p-value≤0.01

32

Table 1.5: Decomposition of Hindu-SC/ST LFPR Gap (Oaxaca-Blinder Method including Predicted Income)

Hindu
LFPR
SC/ST

Difference
Gap

Predicted
Unexplained

Personal
Household
Predicted Gap
Regional
Predicted Inc.

N

38th
1983-84

43rd
1988-89

50th
1993-94

55th
1999-00

61st
2004-05

66th
2009-10

68th
2011-12

0.429
(0.0409)***
0.610
(0.0433)***

0.424
(0.0462)***
0.599
(0.0434)***

0.412
(0.0461)***
0.586
(0.0452)***

0.379
(0.0425)***
0.542
(0.0427)***

0.432
(0.0400)***
0.560
(0.0407)***

0.317
(0.0359)***
0.416
(0.0397)***

0.302
(0.0334)***
0.399
(0.0340)***

-0.1811
(0.0305)***
-0.0854
(0.0255)***
-0.0957
(0.0325)***

-0.1753
(0.0282)***
-0.0870
(0.0256)***
-0.0883
(0.0278)***

-0.1745
(0.0278)***
-0.1096
(0.0190)***
-0.0649
(0.0244)***

-0.1622
(0.0229)***
-0.1395
(0.0233)***
-0.0227
(0.0264)

-0.1278
(0.0248)***
-0.1144
(0.0166)***
-0.0134
(0.0217)

-0.0986
(0.0204)***
-0.0592
(0.0152)***
-0.0394
(0.0207)*

-0.0965
(0.0174)***
-0.0544
(0.0158)***
-0.0421
(0.0187)**

-0.0301
(0.0057)***
-0.0415
(0.0068)***
-0.0160
(0.0136)
0.0021
(0.0241)

-0.0298
(0.0057)***
-0.0416
(0.0089)***
-0.0040
(0.0157)
-0.0116
(0.0284)

-0.0427
(0.0062)***
-0.0257
(0.0079)***
0.0010
(0.0144)
-0.0422
(0.0200)**

-0.0323
(0.0049)***
0.0023
(0.0133)
0.0050
(0.0134)
-0.1145
(0.0377)***

-0.0327
(0.0048)***
-0.0064
(0.0074)
-0.0016
(0.0107)
-0.0738
(0.0188)***

-0.0280
(0.0067)***
-0.0075
(0.0098)
0.0009
(0.0113)
-0.0247
(0.0217)

-0.0237
(0.0054)***
-0.0127
(0.0092)
-0.0054
(0.0109)
-0.0125
(0.0261)

110,851

119,416

106,574

112,356

111,815

87,764

87,346

89.51

60.04

56.37

Explained (% of total)
47.16
49.63
62.81
86.00
Note: Standard errors are reported in parenthesis. *p-value≤0.10, ** p-value≤0.05, *** p-value≤0.01

33

Table 1.6: Decomposition of Hindu-Muslim LFPR Gap (Oaxaca-Blinder Method including Predicted Income)

Hindu
LFPR
Muslim

Difference
Gap

Predicted
Unexplained

Personal
Household
Predicted Gap
Regional
Predicted Inc.

N

38th
1983-84

43rd
1988-89

50th
1993-94

55th
1999-00

61st
2004-05

66th
2009-10

68th
2011-12

0.429
(0.0409)***
0.262
(0.0240)***

0.424
(0.0462)***
0.273
(0.0273)***

0.412
(0.0461)***
0.232
(0.0216)***

0.379
(0.0425)***
0.228
(0.0179)***

0.432
(0.0400)***
0.253
(0.0188)***

0.317
(0.0359)***
0.184
(0.0176)***

0.302
(0.0334)***
0.203
(0.0224)***

0.1675
(0.0315)***
0.0210
(0.0311)
0.1465
(0.0220)***

0.1507
(0.0336)***
0.0273
(0.0305)
0.1234
(0.0201)***

0.1797
(0.0339)***
0.0226
(0.0285)
0.1570
(0.0206)***

0.1510
(0.0371)***
-0.0005
(0.0303)
0.1515
(0.0292)***

0.1795
(0.0317)***
-0.0024
(0.0262)
0.1819
(0.0205)***

0.1334
(0.0263)***
-0.0059
(0.0228)
0.1392
(0.0218)***

0.0996
(0.0267)***
-0.0031
(0.0207)
0.1026
(0.0203)***

-0.0127
(0.0044)***
-0.0164
(0.0077)**
0.0493
(0.0259)*
0.0009
(0.0107)

-0.0102
(0.0052)*
-0.0158
(0.0091)*
0.0570
(0.0254)**
-0.0037
(0.0091)

-0.0156
(0.0065)**
-0.0058
(0.0077)
0.0637
(0.0216)***
-0.0197
(0.0097)**

-0.0084
(0.0060)
0.0241
(0.0137)*
0.0412
(0.0276)
-0.0575
(0.0236)**

-0.0092
(0.0060)
0.0079
(0.0071)
0.0321
(0.0215)
-0.0332
(0.0116)***

-0.0141
(0.0058)**
0.0006
(0.0101)
0.0208
(0.0202)
-0.0132
(0.0118)

-0.0080
(0.0046)*
-0.0085
(0.0108)
0.0194
(0.0191)
-0.0059
(0.0125)

99,524

108,986

93,522

99,432

98,912

78,311

80,085

-1.34

-4.42

-3.11

Explained (% of total)
12.54
18.12
12.58
-0.33
Note: Standard errors are reported in parenthesis. *p-value≤0.10, ** p-value≤0.05, *** p-value≤0.01

34

Table 1.7: Measuring Role of Culture-IHDS II (Oaxaca Blinder Method)

LFPR

Hindu-SC/ST

Hindu-Muslim

0.389
(.005)***
0.510
(.008)***
-0.122
(.009)***
-0.081
(.005)***
-0.040
(.009)***
-0.008
(.003)***
-0.016
(.003)***
-0.001
(.003)
-0.042
(.004)***
-0.014
(.002)**
26,208

0.389
(0.005)***
0.233
(0.009)***
0.155
(0.010)***
0.063
(0.007)***
0.092
(0.011)***
-0.001
(0.003)
-0.008
(0.003)***
0.071
(0.004)***
-0.017
(0.002)***
0.018
(0.004)***
20,955

Hindu
SC/ST or Muslim
Difference
Predicted
Unexplained

Explained

Personal
Household
Regional
Assets/Wealth
Culture

N

Note: Standard errors are reported in parenthesis. *p-value≤0.10, ** p-value≤0.05, *** p-value≤0.01

35

Table 1.8: Summary of Culture Variable- IHDS
Hindu

Muslim

SC/ST

Mobility (0-4)

2.65

2.56

2.70

Decision making (0-6)

1.46

1.39

1.51

Violence (0-6)

2.97

2.85

2.91

Woman decides whether to work (0/1)

0.35

0.30

0.45

Observe head/face covering (0/1)

0.56

0.87

0.57

36

37

38

39

40

41

APPENDIX B

SUPPLEMENTARY TABLES AND FIGURES

42

Table 1.9: Descriptive Statistics- Hindus
38

Age Group

Education

Head Age Group

Head Education

43

50

55

61

66

68

Mean

SE

Mean

SE

Mean

SE

Mean

SE

Mean

SE

Mean

SE

Mean

SE

15-24

35.29

0.201

34.66

0.195

33.44

0.208

31.79

0.227

31.12

0.241

28.93

0.338

28.72

0.328

25-39

37.07

0.203

37.98

0.2

38.84

0.217

40.47

0.243

39.98

0.255

40.76

0.37

40.81

0.356

40-49

16.35

0.154

16.27

0.15

16.57

0.163

16.62

0.182

17.64

0.195

18.6

0.29

18.56

0.276

50-59

11.29

0.137

11.09

0.127

11.15

0.139

11.13

0.149

11.26

0.159

11.71

0.245

11.92

0.229

No Education

63.39

0.199

59.56

0.197

52.28

0.221

46.07

0.245

40.17

0.256

31.93

0.368

29.77

0.353

Below Primary

7.86

0.108

8.94

0.114

9.36

0.128

8.81

0.134

8.72

0.145

8.35

0.222

8.6

0.208

Primary

11.7

0.13

11.87

0.127

11.23

0.141

10.87

0.156

12.2

0.17

12.12

0.252

10.92

0.226

Middle

9.02

0.112

9.29

0.112

11.97

0.138

14.58

0.171

16.23

0.189

16.83

0.27

16.61

0.258

Secondary & higher

8.04

0.112

10.34

0.112

15.15

0.151

19.67

0.196

22.68

0.21

30.77

0.324

34.1

0.326

0-30

15.85

0.159

15.18

0.151

13.44

0.157

11.51

0.166

9.85

0.166

8.63

0.22

8.79

0.22

30-45

35.46

0.199

36.46

0.2

37.21

0.214

38.13

0.24

38.2

0.255

37.95

0.364

38.61

0.357

45-60

35.82

0.202

35.99

0.195

37.27

0.213

36.68

0.235

37.62

0.247

39.4

0.369

38.67

0.347

60-99

12.86

0.141

12.38

0.13

12.08

0.143

13.69

0.159

14.34

0.172

14.02

0.245

13.93

0.23

No Education

41.72

0.207

38.88

0.205

34.67

0.214

32.59

0.231

29.99

0.242

25.8

0.35

25.03

0.338

Below Primary

16.33

0.167

17.33

0.154

15.98

0.162

14.35

0.168

12.59

0.166

10.67

0.237

11.64

0.24

Primary

16.4

0.15

16.1

0.146

14.36

0.154

13.11

0.176

14.43

0.182

14.2

0.266

12.17

0.231

Middle

11.69

0.131

11.22

0.123

13.58

0.151

14.77

0.17

15.95

0.186

16.37

0.271

16.54

0.261

Secondary & higher

13.87

0.146

16.46

0.144

21.41

0.175

25.18

0.211

27.03

0.23

32.95

0.338

34.61

0.329

Rural

74.39

0.174

75.18

0.161

72.62

0.18

71.58

0.221

71.61

0.236

68.82

0.31

68.38

0.302

Married

80.44

0.164

80.31

0.159

79.19

0.176

79.11

0.194

78.76

0.209

78.3

0.3

78.19

0.29

Salaried Men in HH

23.06

0.176

24.73

0.178

24.02

0.185

23.26

0.215

22.86

0.216

22.86

0.283

23.64

0.274

Female Headed

3.86

0.087

3.86

0.078

3.86

0.082

4.63

0.097

4.75

0.106

5.07

0.154

5.16

0.147

1

0.005

0.94

0.005

0.81

0.005

0.78

0.005

0.72

0.005

0.61

0.007

0.51

0.006

HH Size

6.54

0.013

6.33

0.012

5.95

0.011

6.06

0.014

5.8

0.014

5.43

0.018

5.3

0.017

Unemployment Rate in Region

2.1

0.007

2.62

0.007

2.03

0.006

2.04

0.008

1.81

0.007

1.34

0.007

1.56

0.008

14.99

0.045

17.59

0.044

23.14

0.053

27.11

0.065

28.86

0.072

36.65

0.108

38.21

0.101

Children Below 5

Secondary Education Rate in Region

43

Table 1.9 (cont’d)
Log MPCE

4.69
N

80,369

0.004

4.99
88,261

0.005

5.71
78,794

44

0.002

6.2
80,154

0.003

6.44
79,067

0.003

6.93
61,829

0.004

7.22
62,090

0.004

Table 1.10: Descriptive Statistics- Muslims
38

Age Group

Education

Head Age Group

Head Education

43

50

55

61

66

68

Mean

SE

Mean

SE

Mean

SE

Mean

SE

Mean

SE

Mean

SE

Mean

SE

15-24

37.26

0.447

37.31

0.455

36.23

0.481

37.4

0.582

37.67

0.52

36.87

0.731

35.72

0.674

25-39

37.45

0.47

37.82

0.462

39.59

0.49

39.38

0.558

37.29

0.512

38.39

0.744

38.53

0.692

40-49

15.54

0.33

14.77

0.326

14.58

0.347

14.6

0.413

15.74

0.391

16.08

0.546

16.44

0.517

50-59

9.75

0.269

10.1

0.283

9.59

0.292

8.62

0.278

9.31

0.309

8.66

0.437

9.31

0.396

No Education

72.03

0.441

69.29

0.408

61.84

0.475

54.92

0.563

48.81

0.535

41.92

0.768

37.88

0.708

Below Primary

9.81

0.341

10.96

0.275

11.9

0.317

11.03

0.311

11.85

0.33

11.13

0.473

11.43

0.478

Primary

9.48

0.264

9.76

0.254

10.21

0.291

11.73

0.319

13.33

0.345

14.91

0.545

14.48

0.499

Middle

5.28

0.213

5.94

0.189

9.38

0.275

12.19

0.352

13.54

0.331

15.1

0.496

15.6

0.473

Secondary & higher

3.39

0.156

4.05

0.155

6.68

0.215

10.13

0.273

12.46

0.364

16.95

0.518

20.62

0.507

0-30

17.63

0.364

18.3

0.39

17.07

0.391

14.74

0.511

11.87

0.365

11.57

0.527

10.38

0.444

30-45

37.53

0.467

36.01

0.454

38.6

0.487

39.82

0.57

38.21

0.517

38.3

0.733

37.21

0.694

45-60

33.64

0.435

34.08

0.443

33.8

0.464

34.45

0.531

37.74

0.522

37.01

0.724

39.85

0.685

60-99

11.2

0.274

11.62

0.275

10.52

0.32

11

0.295

12.18

0.312

13.13

0.548

12.56

0.436

No Education

53.43

0.472

51.19

0.47

48.63

0.501

44.79

0.591

44.05

0.536

40.04

0.755

39.25

0.718

Below Primary

16.87

0.393

17.61

0.344

17.51

0.374

17.26

0.436

15.5

0.365

13

0.506

14.29

0.49

Primary

13.81

0.305

14.7

0.309

13.04

0.333

12.65

0.326

15.57

0.364

16.3

0.568

15.74

0.495

Middle

8.22

0.24

8.11

0.236

9.79

0.282

11.35

0.31

11.32

0.309

13.83

0.496

14.13

0.468

Secondary & higher

7.67

0.251

8.39

0.228

11.03

0.289

13.95

0.372

13.57

0.381

16.82

0.549

16.6

0.438

Rural

65.31

0.438

68.55

0.401

65.18

0.441

65.02

0.484

66.48

0.506

65.7

0.641

63.05

0.632

Married

78.5

0.373

78.92

0.374

78.47

0.398

76.2

0.453

74.08

0.471

74.4

0.642

73.53

0.613

Salaried Men in HH

19.33

0.35

18.57

0.332

17.56

0.357

18.61

0.38

17.71

0.398

18.66

0.516

18.74

0.494

Female Headed

4.52

0.183

4.69

0.175

4.95

0.209

5.67

0.22

6.21

0.237

6.03

0.311

7.07

0.349

Children Below 5

1.23

0.012

1.18

0.01

1.06

0.011

1.05

0.013

0.92

0.012

0.75

0.017

0.73

0.014

HH Size

7

0.03

6.83

0.028

6.44

0.028

6.81

0.039

6.57

0.031

6

0.039

6.09

0.035

Unemployment Rate in Region

2.47

0.017

2.81

0.018

2.29

0.015

2.5

0.015

1.98

0.015

1.57

0.014

1.8

0.018

Secondary Education Rate in Region

17.26

0.11

19.41

0.1

24.9

0.126

28.66

0.158

29.45

0.147

36.16

0.242

38.32

0.216

45

Table 1.10 (cont’d)
Log MPCE

4.55
N

19,155

0.009

4.88
20,726

0.011

5.58
14,728

46

0.005

6.06
19,278

0.008

6.3
19,845

0.005

6.76
16,482

0.007

7.07
17,995

0.007

Table 1.11: Descriptive Statistics- SC/ST
38

Age Group

Education

Head Age Group

Head Education

43

50

55

61

66

68

Mean

SE

Mean

SE

Mean

SE

Mean

SE

Mean

SE

Mean

SE

Mean

SE

15-24

34.41

0.314

35.16

0.308

32.46

0.32

32.77

0.34

33.03

0.362

32.21

0.532

31.69

0.511

25-39

38.01

0.324

38.27

0.316

40.17

0.335

40.51

0.347

40.09

0.379

40.45

0.555

40.93

0.546

40-49

17.02

0.287

15.73

0.235

16.49

0.259

16.53

0.269

16.83

0.289

17.16

0.428

16.8

0.406

50-59

10.56

0.199

10.83

0.2

10.88

0.212

10.19

0.212

10.05

0.229

10.19

0.338

10.59

0.357

No Education

86.96

0.224

85.16

0.218

78.89

0.273

70.16

0.338

62.87

0.366

50.22

0.566

47.62

0.556

Below Primary

4.31

0.13

4.55

0.128

6.71

0.169

7.34

0.182

8.33

0.207

9.53

0.347

9.89

0.339

Primary

4.62

0.145

4.95

0.132

5.76

0.154

7.59

0.207

9.88

0.221

12.52

0.369

11.92

0.358

Middle

2.88

0.111

3.34

0.112

5.07

0.145

8.21

0.186

10.5

0.231

13.14

0.377

14.58

0.377

Secondary & higher

1.23

0.071

2

0.079

3.58

0.12

6.7

0.217

8.41

0.201

14.59

0.352

15.99

0.361

0-30

21.47

0.268

22.52

0.277

19.64

0.275

16.46

0.259

14.42

0.275

13.37

0.39

12.84

0.377

30-45

38.13

0.324

37.65

0.314

39.69

0.335

41.22

0.357

41.18

0.381

41.61

0.56

41.76

0.548

45-60

32.44

0.328

32.12

0.299

33.29

0.322

33.63

0.331

35.29

0.366

36.67

0.545

36.29

0.529

60-99

7.96

0.17

7.71

0.167

7.38

0.175

8.69

0.185

9.11

0.217

8.36

0.292

9.12

0.314

No Education

67.57

0.323

66.42

0.302

62.27

0.33

57.01

0.355

51.76

0.385

45.31

0.568

43.1

0.557

Below Primary

14.08

0.258

13.65

0.217

14.54

0.242

13.7

0.233

14.16

0.273

12.8

0.391

14.53

0.391

Primary

9.86

0.198

10.16

0.198

9.79

0.199

10.01

0.203

12.39

0.246

14.01

0.391

13.7

0.38

Middle

5.19

0.138

5.63

0.139

6.97

0.172

9.27

0.197

11.09

0.235

13.03

0.362

13.26

0.358

Secondary & higher

3.29

0.132

4.14

0.124

6.43

0.161

10.01

0.262

10.6

0.232

14.85

0.36

15.41

0.353

Rural

86.57

0.227

87.46

0.185

86.49

0.199

84.13

0.226

83.72

0.275

83.6

0.314

81.52

0.346

Married

83.95

0.235

83.78

0.233

83.08

0.252

80.88

0.291

79.67

0.304

78.35

0.461

78.48

0.445

Salaried Men in HH

16.34

0.242

16.74

0.231

14.26

0.226

15.11

0.293

15.96

0.274

15.52

0.342

16.83

0.369

2.7

0.1

3.06

0.11

3.21

0.119

3.44

0.119

4.66

0.155

4.95

0.239

5.48

0.244

1

0.008

1.01

0.007

0.88

0.007

0.83

0.007

0.8

0.008

0.65

0.01

0.58

0.01

HH Size

5.99

0.02

5.82

0.016

5.54

0.016

5.78

0.017

5.66

0.019

5.37

0.025

5.25

0.024

Unemployment Rate in Region

1.65

0.01

2.11

0.01

1.72

0.009

1.75

0.009

1.61

0.01

1.21

0.01

1.38

0.011

Secondary Education Rate in Region

11.45

0.062

13.78

0.054

18.37

0.065

22.68

0.075

24.45

0.088

31.36

0.128

33.11

0.134

Female Headed
Children Below 5

47

Table 1.11 (cont’d)
Log MPCE

4.39
N

30,482

0.007

4.65
31,156

0.009

5.42
27,780

48

0.003

5.94
32,202

0.003

6.13
32,748

0.003

6.62
25,935

0.005

6.9
25,256

0.005

Table 1.12: Decomposition of Hindu-SC/ST LFPR Gap (Oaxaca-Blinder Method- SC/ST Mapping)
38th
1983-84

43rd
1988-89

50th
1993-94

55th
1999-00

61st
2004-05

66th
2009-10

68th
2011-12

Hindu

0.4294

0.4236

0.4120

0.3794

0.4324

0.3174

0.3025

SC/ST

(0.0409)***
0.6104
(0.0433)***

(0.0462)***
0.5989
(0.0434)***

(0.0460)***
0.5864
(0.0452)***

(0.0421)***
0.5416
(0.0424)***

(0.0399)***
0.5601
(0.0405)***

(0.0358)***
0.4160
(0.0396)***

(0.0334)***
0.3989
(0.0339)***

-0.1811
(0.0305)***
-0.0894
(0.0223)***
-0.0917
(0.0236)***

-0.1753
(0.0282)***
-0.0914
(0.0234)***
-0.0839
(0.0197)***

-0.1745
(0.0278)***
-0.0938
(0.0223)***
-0.0806
(0.0177)***

-0.1622
(0.0231)***
-0.0915
(0.0177)***
-0.0707
(0.0157)***

-0.1278
(0.0248)***
-0.0707
(0.0187)***
-0.0571
(0.0159)***

-0.0986
(0.0204)***
-0.0448
(0.0162)***
-0.0538
(0.0130)***

-0.0965
(0.0173)***
-0.0600
(0.0135)***
-0.0365
(0.0140)***

-0.0267
(0.0056)***
-0.0385
(0.0055)***
-0.0242
(0.0182)

-0.0295
(0.0052)***
-0.0456
(0.0060)***
-0.0163
(0.0189)

-0.0327
(0.0053)***
-0.0453
(0.0065)***
-0.0158
(0.0179)

-0.0365
(0.0062)***
-0.0365
(0.0045)***
-0.0186
(0.0146)

-0.0325
(0.0061)***
-0.0246
(0.0046)***
-0.0136
(0.0152)

-0.0284
(0.0067)***
-0.0158
(0.0045)***
-0.0006
(0.0141)

-0.0265
(0.0061)***
-0.0195
(0.0035)***
-0.0140
(0.0104)

110,851

119,416

106,574

112,356

111,815

87,764

87,346

49.36

52.14

53.75

56.41

55.32

45.44

62.18

LFPR

Difference
Gap

Predicted
Unexplained

Personal
Predicted Gap

Household
Regional

N
Explained (% of total)

Note: Standard errors are reported in parenthesis. *p-value≤0.10, ** p-value≤0.05, *** p-value≤0.01

49

Table 1.13: Decomposition of Hindu-Muslim LFPR Gap (Oaxaca-Blinder Method-Muslim Mapping)
38th
1983-84

43rd
1988-89

50th
1993-94

55th
1999-00

61st
2004-05

66th
2009-10

68th
2011-12

Hindu

0.4294

0.4236

0.4120

0.3794

0.4324

0.3174

0.3025

Muslim

(0.0409)***
0.2618
(0.0241)***

(0.0462)***
0.2730
(0.0272)***

(0.0460)***
0.2323
(0.0219)***

(0.0421)***
0.2285
(0.0182)***

(0.0399)***
0.2529
(0.0189)***

(0.0358)***
0.1840
(0.0175)***

(0.0334)***
0.2029
(0.0223)***

0.1675
(0.0315)***
0.0362
(0.0220)*
0.1314
(0.0181)***

0.1507
(0.0336)***
0.0434
(0.0255)*
0.1073
(0.0185)***

0.1797
(0.0339)***
0.0353
(0.0192)*
0.1444
(0.0215)***

0.1510
(0.0368)***
0.0185
(0.0184)
0.1325
(0.0262)***

0.1795
(0.0314)***
0.0301
(0.0186)
0.1493
(0.0230)***

0.1334
(0.0261)***
0.0122
(0.0142)
0.1211
(0.0210)***

0.0996
(0.0268)***
0.0015
(0.0186)
0.0981
(0.0170)***

-0.0030
(0.0027)
-0.0097
(0.0045)**
0.0490
(0.0203)**

-0.0029
(0.0039)
-0.0062
(0.0045)
0.0525
(0.0231)**

-0.0034
(0.0035)
-0.0135
(0.0045)***
0.0521
(0.0171)***

-0.0051
(0.0053)
-0.0089
(0.0054)*
0.0325
(0.0147)**

-0.0011
(0.0038)
-0.0106
(0.0049)**
0.0418
(0.0155)***

-0.0005
(0.0030)
-0.0118
(0.0053)**
0.0246
(0.0119)**

-0.0138
(0.0056)**
-0.0007
(0.0037)
0.0159
(0.0167)

99,524

108,986

93,522

99,432

98,912

78,311

80,085

21.61

28.80

19.64

12.25

16.77

9.15

1.51

LFPR

Difference
Gap

Predicted
Unexplained

Personal
Predicted Gap

Household
Regional

N
Explained (% of total)

Note: Standard errors are reported in parenthesis. *p-value≤0.10, ** p-value≤0.05, *** p-value≤0.01

50

Table 1.14: Decomposition of Hindu-SC/ST LFPR Gap (Oaxaca-Blinder Method-Pooled Mapping)
38th
1983-84

43rd
1988-89

50th
1993-94

55th
1999-00

61st
2004-05

66th
2009-10

68th
2011-12

Hindu

0.4294

0.4236

0.4120

0.3794

0.4324

0.3174

0.3025

SC/ST

(0.0409)***
0.6104
(0.0433)***

(0.0462)***
0.5989
(0.0434)***

(0.0460)***
0.5864
(0.0452)***

(0.0421)***
0.5416
(0.0424)***

(0.0399)***
0.5601
(0.0405)***

(0.0358)***
0.4160
(0.0396)***

(0.0334)***
0.3989
(0.0339)***

-0.1811
(0.0305)***
-0.0854
(0.0191)***
-0.0957
(0.0252)***

-0.1753
(0.0282)***
-0.0812
(0.0200)***
-0.0941
(0.0210)***

-0.1745
(0.0278)***
-0.0874
(0.0188)***
-0.0871
(0.0203)***

-0.1622
(0.0231)***
-0.0828
(0.0159)***
-0.0795
(0.0176)***

-0.1278
(0.0248)***
-0.0727
(0.0161)***
-0.0551
(0.0176)***

-0.0986
(0.0204)***
-0.0478
(0.0140)***
-0.0508
(0.0156)***

-0.0965
(0.0173)***
-0.0511
(0.0121)***
-0.0454
(0.0144)***

-0.0275
(0.0057)***
-0.0403
(0.0058)***
-0.0176
(0.0149)

-0.0286
(0.0054)***
-0.0449
(0.0050)***
-0.0077
(0.0162)

-0.0384
(0.0058)***
-0.0420
(0.0045)***
-0.0070
(0.0153)

-0.0323
(0.0054)***
-0.0362
(0.0035)***
-0.0142
(0.0129)

-0.0320
(0.0049)***
-0.0291
(0.0044)***
-0.0116
(0.0124)

-0.0281
(0.0064)***
-0.0144
(0.0037)***
-0.0053
(0.0112)

-0.0243
(0.0053)***
-0.0168
(0.0033)***
-0.0099
(0.0094)

110,851

119,416

106,574

112,356

111,815

87,764

87,346

47.16

46.32

50.09

51.05

56.89

48.48

52.95

LFPR

Difference
Gap

Predicted
Unexplained

Personal
Predicted Gap

Household
Regional

N
Explained (% of total)

Note: Standard errors are reported in parenthesis. *p-value≤0.10, ** p-value≤0.05, *** p-value≤0.01

51

Table 1.15: Decomposition of Hindu-Muslim LFPR Gap (Oaxaca-Blinder Method-Pooled Mapping)
38th
1983-84

43rd
1988-89

50th
1993-94

55th
1999-00

61st
2004-05

66th
2009-10

68th
2011-12

Hindu

0.4294

0.4236

0.4120

0.3794

0.4324

0.3174

0.3025

Muslim

(0.0409)***
0.2618
(0.0241)***

(0.0462)***
0.2730
(0.0272)***

(0.0460)***
0.2323
(0.0219)***

(0.0421)***
0.2285
(0.0182)***

(0.0399)***
0.2529
(0.0189)***

(0.0358)***
0.1840
(0.0175)***

(0.0334)***
0.2029
(0.0223)***

0.1675
(0.0315)***
0.0304
(0.0289)
0.1371
(0.0195)***

0.1507
(0.0336)***
0.0327
(0.0302)
0.1180
(0.0185)***

0.1797
(0.0339)***
0.0403
(0.0290)
0.1394
(0.0206)***

0.1510
(0.0368)***
0.0325
(0.0276)
0.1185
(0.0253)***

0.1795
(0.0314)***
0.0275
(0.0257)
0.1520
(0.0184)***

0.1334
(0.0261)***
0.0099
(0.0206)
0.1235
(0.0187)***

0.0996
(0.0268)***
0.0047
(0.0204)
0.0949
(0.0170)***

-0.0119
(0.0041)***
-0.0151
(0.0055)***
0.0575
(0.0247)**

-0.0096
(0.0049)*
-0.0168
(0.0058)***
0.0591
(0.0255)**

-0.0144
(0.0060)**
-0.0187
(0.0053)***
0.0734
(0.0237)***

-0.0086
(0.0057)
-0.0166
(0.0058)***
0.0577
(0.0231)**

-0.0089
(0.0055)
-0.0177
(0.0056)***
0.0541
(0.0212)**

-0.0127
(0.0053)**
-0.0092
(0.0039)**
0.0317
(0.0173)*

-0.0099
(0.0045)**
-0.0112
(0.0032)***
0.0258
(0.0179)

99,524

108,986

93,522

99,432

98,912

78,311

80,085

18.15

21.70

22.43

21.52

15.32

7.42

4.72

LFPR

Difference
Gap

Predicted
Unexplained

Personal
Predicted Gap

Household
Regional

N
Explained (% of total)

Note: Standard errors are reported in parenthesis. *p-value≤0.10, ** p-value≤0.05, *** p-value≤0.01

52

Table 1.16: Decomposition of Rural Hindu-SC/ST LFPR Gap (Oaxaca-Blinder Method)
38th
1983-84

43rd
1988-89

50th
1993-94

55th
1999-00

61st
2004-05

66th
2009-10

68th
2011-12

Hindu

0.5024

0.4888

0.4827

0.4525

0.5051

0.3738

0.3457

SC/ST

(0.0538)***
0.6522
(0.0517)***

(0.0637)***
0.6342
(0.0509)***

(0.0663)***
0.6267
(0.0545)***

(0.0599)***
0.5899
(0.0519)***

(0.0544)***
0.6083
(0.0485)***

(0.0540)***
0.4468
(0.0506)***

(0.0530)***
0.4345
(0.0440)***

-0.1498
(0.0366)***
-0.0468
(0.0186)**
-0.1030
(0.0312)***

-0.1453
(0.0354)***
-0.0429
(0.0197)**
-0.1025
(0.0248)***

-0.1440
(0.0375)***
-0.0406
(0.0221)*
-0.1034
(0.0264)***

-0.1374
(0.0291)***
-0.0450
(0.0199)**
-0.0924
(0.0224)***

-0.1032
(0.0318)***
-0.0365
(0.0180)**
-0.0667
(0.0198)***

-0.0729
(0.0245)***
-0.0165
(0.0185)
-0.0564
(0.0189)***

-0.0889
(0.0238)***
-0.0266
(0.0197)
-0.0622
(0.0169)***

-0.0258
(0.0051)***
-0.0315
(0.0052)***
0.0106
(0.0178)

-0.0274
(0.0053)***
-0.0343
(0.0039)***
0.0188
(0.0197)

-0.0373
(0.0055)***
-0.0311
(0.0031)***
0.0278
(0.0211)

-0.0296
(0.0053)***
-0.0263
(0.0033)***
0.0109
(0.0179)

-0.0270
(0.0045)***
-0.0242
(0.0048)***
0.0147
(0.0157)

-0.0233
(0.0070)***
-0.0092
(0.0045)**
0.0160
(0.0150)

-0.0208
(0.0058)***
-0.0123
(0.0041)***
0.0065
(0.0188)

75,109

80,973

68,669

71,579

74,452

54,902

55,137

31.24

29.53

28.19

32.75

35.37

22.63

29.92

LFPR

Difference
Gap

Predicted
Unexplained

Personal
Predicted Gap

Household
Regional

N
Explained (% of total)

Note: Standard errors are reported in parenthesis. *p-value≤0.10, ** p-value≤0.05, *** p-value≤0.01

53

Table 1.17: Decomposition of Rural Hindu-Muslim LFPR Gap (Oaxaca-Blinder Method)
38th
1983-84

43rd
1988-89

50th
1993-94

55th
1999-00

61st
2004-05

66th
2009-10

68th
2011-12

Hindu

0.5024

0.4888

0.4827

0.4525

0.5051

0.3738

0.3457

Muslim

(0.0538)***
0.3087
(0.0324)***

(0.0637)***
0.3165
(0.0413)***

(0.0663)***
0.2576
(0.0349)***

(0.0599)***
0.2683
(0.0217)***

(0.0544)***
0.2889
(0.0280)***

(0.0540)***
0.2055
(0.0275)***

(0.0530)***
0.2317
(0.0328)***

0.1937
(0.0364)***
0.0408
(0.0324)
0.1529
(0.0268)***

0.1723
(0.0419)***
0.0545
(0.0364)
0.1178
(0.0257)***

0.2250
(0.0429)***
0.0785
(0.0355)**
0.1465
(0.0218)***

0.1843
(0.0483)***
0.0774
(0.0478)
0.1068
(0.0496)**

0.2162
(0.0404)***
0.0546
(0.0356)
0.1617
(0.0253)***

0.1683
(0.0343)***
0.0371
(0.0376)
0.1312
(0.0361)***

0.1140
(0.0350)***
0.0324
(0.0324)
0.0816
(0.0368)**

-0.0109
(0.0047)**
-0.0144
(0.0055)***
0.0661
(0.0308)**

-0.0093
(0.0066)
-0.0153
(0.0058)***
0.0790
(0.0330)**

-0.0144
(0.0082)*
-0.0160
(0.0055)***
0.1090
(0.0291)***

-0.0095
(0.0083)
-0.0152
(0.0057)***
0.1022
(0.0458)**

-0.0068
(0.0076)
-0.0175
(0.0061)***
0.0788
(0.0344)**

-0.0145
(0.0076)*
-0.0069
(0.0040)*
0.0585
(0.0352)*

-0.0070
(0.0057)
-0.0096
(0.0039)**
0.0491
(0.0332)

N

61,166

68,389

54,654

57,652

61,828

45,034

47,021

Explained (% of total)

21.06

31.63

34.89

42.00

25.25

22.04

28.42

LFPR

Difference
Gap

Predicted
Unexplained

Personal
Predicted Gap

Household
Regional

Note: Standard errors are reported in parenthesis. *p-value≤0.10, ** p-value≤0.05, *** p-value≤0.01

54

Table 1.18: Decomposition of Urban Hindu-SC/ST LFPR Gap (Oaxaca-Blinder Method)
38th
1983-84

43rd
1988-89

50th
1993-94

55th
1999-00

61st
2004-05

66th
2009-10

68th
2011-12

Hindu

0.2171

0.2261

0.2244

0.1953

0.2488

0.1927

0.2090

SC/ST

(0.0197)***
0.3411
(0.0194)***

(0.0249)***
0.3528
(0.0276)***

(0.0219)***
0.3286
(0.0240)***

(0.0208)***
0.2855
(0.0192)***

(0.0231)***
0.3125
(0.0230)***

(0.0217)***
0.2589
(0.0193)***

(0.0174)***
0.2419
(0.0161)***

-0.1239
(0.0136)***
-0.0534
(0.0095)***
-0.0706
(0.0140)***

-0.1267
(0.0149)***
-0.0556
(0.0131)***
-0.0711
(0.0191)***

-0.1041
(0.0106)***
-0.0535
(0.0097)***
-0.0506
(0.0127)***

-0.0903
(0.0143)***
-0.0367
(0.0115)***
-0.0536
(0.0142)***

-0.0636
(0.0151)***
-0.0468
(0.0111)***
-0.0168
(0.0155)

-0.0662
(0.0096)***
-0.0235
(0.0090)***
-0.0427
(0.0114)***

-0.0329
(0.0082)***
-0.0173
(0.0063)***
-0.0156
(0.0088)*

-0.0189
(0.0056)***
-0.0448
(0.0058)***
0.0103
(0.0050)**

-0.0112
(0.0054)**
-0.0562
(0.0074)***
0.0117
(0.0050)**

-0.0169
(0.0036)***
-0.0442
(0.0069)***
0.0076
(0.0034)**

-0.0077
(0.0052)
-0.0388
(0.0058)***
0.0098
(0.0060)

-0.0174
(0.0059)***
-0.0291
(0.0071)***
-0.0003
(0.0054)

-0.0057
(0.0040)
-0.0246
(0.0045)***
0.0068
(0.0060)

-0.0056
(0.0050)
-0.0171
(0.0045)***
0.0053
(0.0036)

35,742

38,443

37,905

40,777

37,363

32,862

32,209

43.10

43.88

51.39

40.64

73.58

35.50

52.58

LFPR

Difference
Gap

Predicted
Unexplained

Personal
Predicted Gap

Household
Regional

N
Explained (% of total)

Note: Standard errors are reported in parenthesis. *p-value≤0.10, ** p-value≤0.05, *** p-value≤0.01

55

Table 1.19: Decomposition of Urban Hindu-Muslim LFPR Gap (Oaxaca-Blinder Method)
38th
1983-84

43rd
1988-89

50th
1993-94

55th
1999-00

61st
2004-05

66th
2009-10

68th
2011-12

Hindu

0.2171

0.2261

0.2244

0.1953

0.2488

0.1927

0.2090

Muslim

(0.0197)***
0.1735
(0.0172)***

(0.0249)***
0.1780
(0.0218)***

(0.0219)***
0.1849
(0.0147)***

(0.0208)***
0.1545
(0.0119)***

(0.0231)***
0.1815
(0.0092)***

(0.0217)***
0.1428
(0.0105)***

(0.0174)***
0.1538
(0.0118)***

0.0436
(0.0125)***
-0.0468
(0.0113)***
0.0904
(0.0119)***

0.0481
(0.0172)***
-0.0456
(0.0170)***
0.0937
(0.0189)***

0.0395
(0.0213)*
-0.0406
(0.0128)***
0.0802
(0.0164)***

0.0408
(0.0220)*
-0.0312
(0.0125)**
0.0720
(0.0175)***

0.0674
(0.0281)**
-0.0379
(0.0148)**
0.1052
(0.0247)***

0.0499
(0.0273)*
-0.0201
(0.0124)
0.0700
(0.0246)***

0.0553
(0.0217)**
-0.0231
(0.0108)**
0.0783
(0.0223)***

-0.0176
(0.0060)***
-0.0438
(0.0093)***
0.0146
(0.0133)

-0.0081
(0.0055)
-0.0493
(0.0118)***
0.0119
(0.0106)

-0.0101
(0.0058)*
-0.0440
(0.0102)***
0.0135
(0.0104)

-0.0014
(0.0058)
-0.0335
(0.0097)***
0.0037
(0.0106)

-0.0131
(0.0076)*
-0.0297
(0.0108)***
0.0049
(0.0129)

-0.0045
(0.0047)
-0.0179
(0.0082)**
0.0023
(0.0118)

-0.0029
(0.0067)
-0.0260
(0.0077)***
0.0058
(0.0112)

38,358

40,597

38,868

41,780

37,084

33,277

33,064

-107.34

-94.80

-102.78

-76.47

-56.23

-40.28

-41.77

LFPR

Difference
Gap

Predicted
Unexplained

Personal
Predicted Gap

Household
Regional

N
Explained (% of total)

Note: Standard errors are reported in parenthesis. *p-value≤0.10, ** p-value≤0.05, *** p-value≤0.01

56

Table 1.20: Decomposition of Hindu-Muslim LFPR Gap (Oaxaca-Blinder Method-Hindu 38th Round Mapping)
38th
1983-84
0.4294
(0.0409)***
0.2618
(0.0241)***

43rd
1988-89
0.4236
(0.0462)***
0.2730
(0.0272)***

50th
1993-94
0.4120
(0.0460)***
0.2323
(0.0219)***

55th
1999-00
0.3794
(0.0421)***
0.2285
(0.0182)***

61st
2004-05
0.4324
(0.0399)***
0.2529
(0.0189)***

66th
2009-10
0.3174
(0.0358)***
0.1840
(0.0175)***

68th
2011-12
0.3025
(0.0334)***
0.2029
(0.0223)***

0.1675
(0.0315)***
0.0206
(0.0305)
0.1470
(0.0214)***

0.1507
(0.0336)***
0.0160
(0.0293)
0.1346
(0.0218)***

0.1797
(0.0339)***
0.0222
(0.0322)
0.1574
(0.0268)***

0.1510
(0.0368)***
0.0154
(0.0338)
0.1356
(0.0250)***

0.1795
(0.0314)***
0.0062
(0.0323)
0.1733
(0.0253)***

0.1334
(0.0261)***
-0.0058
(0.0297)
0.1391
(0.0248)***

0.0996
(0.0268)***
-0.0001
(0.0308)
0.0997
(0.0188)***

-0.0127
(0.0044)***
-0.0159
(0.0059)***
0.0493
(0.0259)*

-0.0118
(0.0052)**
-0.0188
(0.0059)***
0.0466
(0.0243)*

-0.0106
(0.0055)*
-0.0229
(0.0065)***
0.0558
(0.0262)**

-0.0083
(0.0058)
-0.0196
(0.0072)***
0.0432
(0.0270)

-0.0074
(0.0056)
-0.0231
(0.0073)***
0.0366
(0.0263)

-0.0065
(0.0051)
-0.0249
(0.0061)***
0.0257
(0.0260)

-0.0044
(0.0052)
-0.0248
(0.0067)***
0.0290
(0.0260)

N

99,524

108,986

93,522

99,432

98,912

78,311

80,085

Explained (% of total)

12.30

10.62

12.35

10.20

3.45

-4.35

-0.10

Hindu
LFPR
Muslim
Difference
Gap

Predicted
Unexplained
Personal

Predicted Gap

Household
Regional

Note: Standard errors are reported in parenthesis. *p-value≤0.10, ** p-value≤0.05, *** p-value≤0.01

57

Table 1.21: Decomposition of Hindu-Muslim LFPR Gap (Oaxaca-Blinder Method-Hindu 68th Round Mapping)

Hindu
LFPR
Muslim
Difference
Gap

Predicted
Unexplained
Personal

Predicted Gap

Household
Regional

N

38th
1983-84
0.4294
(0.0409)***
0.2618
(0.0241)***

43rd
1988-89
0.4236
(0.0462)***
0.2730
(0.0272)***

50th
1993-94
0.4120
(0.0460)***
0.2323
(0.0219)***

55th
1999-00
0.3794
(0.0421)***
0.2285
(0.0182)***

61st
2004-05
0.4324
(0.0399)***
0.2529
(0.0189)***

66th
2009-10
0.3174
(0.0358)***
0.1840
(0.0175)***

68th
2011-12
0.3025
(0.0334)***
0.2029
(0.0223)***

0.1675
(0.0315)***
0.0147
(0.0231)
0.1528
(0.0235)***

0.1507
(0.0336)***
0.0177
(0.0244)
0.1329
(0.0244)***

0.1797
(0.0339)***
0.0257
(0.0212)
0.1540
(0.0233)***

0.1510
(0.0368)***
0.0106
(0.0225)
0.1404
(0.0283)***

0.1795
(0.0314)***
0.0100
(0.0228)
0.1695
(0.0224)***

0.1334
(0.0261)***
0.0015
(0.0207)
0.1318
(0.0197)***

0.0996
(0.0268)***
-0.0019
(0.0206)
0.1015
(0.0199)***

-0.0101
(0.0042)**
-0.0050
(0.0029)*
0.0299
(0.0210)

-0.0102
(0.0047)**
-0.0078
(0.0031)**
0.0358
(0.0226)

-0.0116
(0.0050)**
-0.0093
(0.0035)***
0.0466
(0.0182)**

-0.0085
(0.0053)
-0.0086
(0.0037)**
0.0276
(0.0197)

-0.0069
(0.0054)
-0.0120
(0.0039)***
0.0290
(0.0207)

-0.0090
(0.0045)**
-0.0126
(0.0039)***
0.0231
(0.0183)

-0.0081
(0.0045)*
-0.0135
(0.0039)***
0.0197
(0.0189)

99,524

108,986

93,522

99,432

98,912

78,311

80,085

1.12

-1.91

Explained (% of total)
8.78
11.75
14.30
7.02
5.57
Note: Standard errors are reported in parenthesis. *p-value≤0.10, ** p-value≤0.05, *** p-value≤0.01

58

Figure 1.6: Comparing Counterfactual Populations

59

REFERENCES

60

REFERENCES

Ambedkar, B. R. (1970). Who Were the Shudras?: How They Came to be the Fourth Varna in the IndoAryan Society. Thackers.
Antecol, H. (2003). Why is there cross-country variation in female labor force participation rates? The role
of male attitudes toward family and sex roles. Working paper series/Claremont McKenna College. Retrieved
from http://www.econstor.eu/handle/10419/23378
Banerjee, A., Bertrand, M., Datta, S., & Mullainathan, S. (2009). Labor market discrimination in Delhi:
Evidence from a field experiment. Journal of Comparative Economics, 37(1), 14–27.
https://doi.org/10.1016/j.jce.2008.09.002
Banerjee, B., & Knight, J. B. (1985). Caste discrimination in the Indian urban labour market. Journal of
Development Economics, 17(3), 277–307. https://doi.org/10.1016/0304-3878(85)90094-X
Bhalotra, S., Valente, C., & van Soest, A. (2010). The puzzle of Muslim advantage in child survival in India.
Journal of Health Economics, 29(2), 191–204. https://doi.org/10.1016/j.jhealeco.2009.11.002
Bhaumik, S. K., & Chakrabarty, M. (2006). Earnings Inequality in India: Has the Rise of Caste and ReligionBased Politics in India Had an Impact? (SSRN Scholarly Paper No. ID 890287). Rochester, NY: Social Science
Research Network. Retrieved from http://papers.ssrn.com/abstract=890287
Blau, F. D., & Grossberg, A. J. (1992). Maternal Labor Supply and Children’s Cognitive Development. The
Review of Economics and Statistics, 74(3), 474–481. https://doi.org/10.2307/2109492
Blau, F. D., & Kahn, L. M. (2013). Female labor supply: Why is the US falling behind? National Bureau of
Economic Research. Retrieved from http://www.nber.org/papers/w18702
Blinder, A. S. (1973). Wage Discrimination: Reduced Form and Structural Estimates. The Journal of
Human Resources, 8(4), 436–455. https://doi.org/10.2307/144855
Bloom, D. E., Canning, D., Fink, G., & Finlay, J. E. (2009). Fertility, female labor force participation, and
the demographic dividend. Journal of Economic Growth, 14(2), 79–101. https://doi.org/10.1007/s10887-0099039-9
Bordia Das, M. (2008). Minority status and labor market outcomes: does India have minority enclaves?
Washington, DC: World Bank, South Asia Region, Sustainable Development Dep. Retrieved from
http://dx.doi.org/10.1596/1813-9450-4653
Borooah, V., Dubey, A., & Iyer, S. (2007). The Effectiveness of Jobs Reservation: Caste, Religion, and
Economic Status in India (MPRA Paper No. 19421). University Library of Munich, Germany. Retrieved from
https://ideas.repec.org/p/pra/mprapa/19421.html
Borooah, V. K. (2005). Caste, inequality, and poverty in India. Review of Development Economics, 9(3),
399–414.
Boserup, E. (1970). Woman’s role in economic development. New York: St. Martin’s Press.
Brass, P. (2005). The Production of Hindu-Muslim Violence in Contemporary India. Seattle u.a.:
University of Washington Press.
Brooks–Gunn, J., Han, W.-J., & Waldfogel, J. (2002). Maternal employment and child cognitive outcomes
in the first three years of life: The NICHD study of early child care. Child Development, 73(4), 1052–1072.
Chatterjee, U., Murgai, R., & Rama, M. (2015). Job Opportunities along the Rural-Urban Gradation and
Female
Labor
Force
Participation
in
India.
The
World
Bank.
Retrieved
from
http://elibrary.worldbank.org/doi/abs/10.1596/1813-9450-7412
Deshpande, A. (2011). The Grammar of Caste: Economic Discrimination in Contemporary India. New
Delhi: Oxford University Press.

61

DiNardo, J., Fortin, N. M., & Lemieux, T. (1996). Labor Market Institutions and the Distribution of
Wages, 1973-1992: A Semiparametric Approach. Econometrica Econometrica, 64(5), 1001–1044.
Elder, T. E., Goddeeris, J. H., & Haider, S. J. (2011). A deadly disparity: a unified assessment of the BlackWhite infant mortality gap. The BE Journal of Economic Analysis & Policy, 11(1). Retrieved from
http://www.degruyter.com/view/j/bejeap.2011.11.issue1/bejeap.2011.11.1.2821/bejeap.2011.11.1.2821.xml
Eswaran, M., Ramaswami, B., & Wadhwa, W. (2013). Status, Caste, and the Time Allocation of Women in
Rural India. Economic Development and Cultural Change, 61(2), 311 – 333.
Fernandez, R. (2007). Culture as Learning: The Evolution of Female Labor Force Participation over a
Century (Working Paper No. 13373). National Bureau of Economic Research. Retrieved from
http://www.nber.org/papers/w13373
Fortin, N. M., Lemieux, T., Firpo, S., & National Bureau of Economic Research. (2010). Decomposition
methods in economics. Cambridge, Mass.: National Bureau of Economic Research. Retrieved from
http://papers.nber.org/papers/w16045
FĂźrer-Haimendorf, C. von. (1982). Tribes of India: the struggle for survival. Berkeley: University of
California Press.
Government of India. (2012). Press Note on Poverty Estimates 2009-10. Planning Commission of India.
Retrieved from http://planningcommission.nic.in/news/press_pov1903.pdf
Hanna, R. N., & Linden, L. L. (2012). Discrimination in Grading. American Economic Journal: Economic
Policy, 4(4), 146–168. https://doi.org/10.1257/pol.4.4.146
Hardy, P. (1972). The Muslims of British India. London: Cambridge University Press.
Hasan, Z., & Menon, R. (2004). Unequal Citizens: A Study of Muslim Women in India. New Delhi: Oxford
University Press.
Hasan, Z., & Menon, R. (2005). The Diversity of Muslim Women’s Lives in India. Rutgers University
Press.
Hnatkovska, V., Lahiri, A., & Paul, S. (2012). Castes and Labor Mobility. American Economic Journal:
Applied Economics, 4(2), 274–307. https://doi.org/10.1257/app.4.2.274
Hotz, V. J., & Miller, R. A. (1988). An Empirical Analysis of Life Cycle Fertility and Female Labor Supply.
Econometrica, 56(1), 91–118. https://doi.org/10.2307/1911843
Ito, T. (2009). Caste discrimination and transaction costs in the labor market: Evidence from rural North
India. Journal of Development Economics, 88(2), 292–300. https://doi.org/10.1016/j.jdeveco.2008.06.002
Khalidi, O. (2006). Muslims In Indian Economy. Gurgaon, Haryana: Three Essays Collectives.
Klasen, S., & Pieters, J. (2015). What Explains the Stagnation of Female Labor Force Participation in
Urban India? The World Bank Economic Review, lhv003. https://doi.org/10.1093/wber/lhv003
Madheswaran, S., & Attewell, P. (2007). Caste Discrimination in the Indian Urban Labour Market:
Evidence from the National Sample Survey. Economic and Political Weekly, 42(41), 4146–4153.
Mammen, K., & Paxson, C. (2000). Women’s Work and Economic Development. Journal of Economic
Perspectives, 14(4), 141–164. https://doi.org/10.1257/jep.14.4.141
Munshi, K., & Rosenzweig, M. (2006). Traditional Institutions Meet the Modern World: Caste, Gender,
and Schooling Choice in a Globalizing Economy. American Economic Review, 96(4), 1225–1252.
https://doi.org/10.1257/aer.96.4.1225
Nasreen Fazalbhoy. (1997). Sociology of Muslims in India: A Review. Econpoliweek Economic and
Political Weekly, 32(26), 1547–1551.
Oaxaca, R. (1973). Male-Female Wage Differentials in Urban Labor Markets. International Economic
Review, 14(3), 693–709. https://doi.org/10.2307/2525981

62

Omvedt, G. (2004). Untouchables in the World of It. Contemporary Review. Retrieved from
http://www.highbeam.com/doc/1G1-117860278.html
Rauf, T. A. (2011). Violence Inflicted on Muslims: Direct, Cultural and Structural. Economic and Political
Weekly.
Retrieved
from
http://www.epw.in/notes/violence-inflicted-muslims-direct-cultural-andstructural.html
Robinson, R. (2004). Sociology of religion in India. New Delhi; Thousand Oaks: Sage Publications.
Retrieved from http://public.eblib.com/choice/publicfullrecord.aspx?p=475984
Rosenzweig, M. R., & Wolpin, K. I. (1980). Life-Cycle Labor Supply and Fertility: Causal Inferences from
Household Models. Journal of Political Economy, 88(2), 328–348.
Ruhm, C. J. (2004). Parental employment and child cognitive development. Journal of Human Resources,
39(1), 155–192.
Thorat, S. (2002). Oppression and Denial: Dalit Discrimination in the 1990s. Economic and Political
Weekly, 37(6), 572–578.
Thorat, S., & Attewell, P. (2007). The Legacy of Social Exclusion: A Correspondence Study of Job
Discrimination in India. Economic and Political Weekly, 42(41), 4141–4145.
Thorat, S., & Newman, K. S. (Eds.). (2010). Blocked by Caste: Economic Discrimination and Social
Exclusion in Modern India. New Delhi: Oxford University Press, USA.
Zacharias, A., & Vakulabharanam, V. (2011). Caste Stratification and Wealth Inequality in India. World
Development, 39(10), 1820–1833. https://doi.org/10.1016/j.worlddev.2011.04.026

63

Chapter 2
The Two Bengals: Female Labor Force Participation in
West Bengal and Bangladesh1

1

Introduction
Till the British granted independence to India and Pakistan in 1947, the area that now covers West

Bengal and Bangladesh was part of the larger Bengal presidency under colonial rule. The war of 1971 saw
East Pakistan gain independence from Pakistan and embark on a new political journey as the country now
known as Bangladesh. Both West Bengal and Bangladesh have undergone rapid changes, both economic
and political over the last 68 years but retain a shared culture, history and topography. Despite these
commonalities, both regions exhibit very different levels of socio-economic development, one of which is
the wide disparity in female labor force participation. Rapid development in garment manufacturing since
the 1990s and more recently, community and health service, has driven young women to join the labor
force in large numbers in Bangladesh. On the other hand, LFPR for women remains abysmally low in
West Bengal and is among the lowest across all major Indian states.

The aim of this paper is twofold- firstly, we want to study what proportion of the LFPR differences
between women from the two regions can be explained by differences in observable characteristics and
second, we want to know how much each category (individual, household, regional) of observables explain
this gap in LFPR2. We employ two widely used decomposition methods to disaggregate the gap into an
explained and unexplained part, and then use further decomposition to disaggregate the explained gap
into various components. The paper is divided as follows: the second section offers a brief review of

This paper is co-authored with Maitreyi Bordia Das.
Unless specified, throughout the paper we refer to ‘gap’ as the difference between the LFPR of women in Bangladesh and
the LFPR of women in West Bengal (or vice-versa).
1
2

64

literature related to economic and political participation of women in Bangladesh and West Bengal, the
third section details the methodology, the fourth section explains our data and summary statistics, the
fifth section explains the main results and the sixth section concludes.

2

Review of Literature
Female labor supply has been a subject of widespread economic enquiry (Altonji & Blank, 1999; Blau

& Kahn, 2013) for many decades. At the micro level, economists have studied the effect of mother’s labor
supply on outcomes such as cognitive development and schooling of children (Blau & Grossberg, 1992;
Brooks–Gunn, Han, & Waldfogel, 2002; Ruhm, 2004), whereas at the macro level various studies
document the relationship between female LFP and economic growth (Bhalotra & Umana-Aponte, 2010;
Klasen & Pieters, 2012). There is also an extensive literature on the complex relationship between female
labor supply and fertility decisions. Of particular interest to us is the wave of relatively recent literature
that studies the role of culture and attitudes in determining economic outcomes (Almond et. al, 2013),
particularly female labor supply (Antecol, 2003; Fernandez, 2007). Culture, though hard to quantify, could
play an important role in explaining difference in outcomes for females across social groups, religions and
even regions. This section outlines some of the major research on female labor force participation and
employment in Bangladesh and West Bengal to identify broad trends in the literature and areas for future
research.

2.1

Female labor in Bangladesh

Bangladeshi women’s participation in paid work has risen rapidly in the last two decades sparking a
renewed academic interest in studying their labor market behaviors and outcomes. This increase in female
LFP is driven largely by two major sectors- the NGO and services sector and the garment manufacturing
sector. The importance of the readymade garment (RMG) sector to the economy of Bangladesh cannot
be underscored. The RMG sector is the largest manufacturing industry in the country and the single
largest source of foreign exchange, making up almost 3/4th of total exports of Bangladesh. The industry

65

started small in the late 1970s, grew rapidly in the 1980s and flourished in the 1990s, aided largely by
tariff-free access to the US markets. International financing and buying houses entered the Bangladeshi
RMG sector in the early 1990’s opening the sector to a wider international market. The industry grew
rapidly with a large number of RMG factories concentrated heavily in downtown Dhaka, and Chittagong
serving as the main port. There were concerns that the removal of multi-fiber agreement quotas in 2005
would hurt Bangladeshi exporters however, contrary to all predictions the Bangladesh textiles and
garments industry has not only survived but has actually thrived in the years following 2005. Not only
have the exports been increasing in real and nominal terms, the contribution of the textiles and garments
industry to total exports also peaked at 74-75% in 2008-093. Owing to the fact that the sector is heavily
dependent on a large workforce consisting mainly of low paid young female labor, it is not surprising that
Bangladesh has managed to remain competitive in the international RMG market despite increased
competition from countries like Cambodia, Vietnam and the Philippines. However, women’s conditions of
work, wages and benefits at these factories has sparked national, and more recently international debate
following a slew of devastatingly fatal accidents involving garment factories (New York Times, 2013).
Islam and Zahid (2012) survey 110 respondents from 11 factories in Dhaka to study the socio-economic
status of female garment workers. Among other findings, they show that these women report a marked
increase in frequent and occasional illness after beginning work in RMG factories compared to before. An
earlier study (Paul-Majumder & Begum, 2000) uses survey data from 1990-1997 and finds evidence for
high gender-based gaps in wages and access to benefits as well as significantly higher job volatility for
women4. At the other of the spectrum, some authors have argued that the spread of the RMG sector in
Bangladesh has benefitted young women. Heath and Mobarak (2015) use a survey of close to 1400
households in 60 villages around Dhaka and find that girls who live in villages where a garment factory
is within commuting distance, are more likely to stay in school and to delay marriage and childbirth.
There are others who offer a more neutral stance on the RMG sector. Hossain (2012) concludes that
despite very obvious wage and benefits inequality among male and female workers in RMG

3
4

Author’s calculations from data available with Bangladesh Bank, Central Bank of Bangladesh
Also see War on Want (2011)

66

manufacturing, work in the sector has led to socio-economic empowerment of women and girls in the
form of more mobility, consciousness of rights and socio-political agency.

The trends seen in the RMG sector are mirrored in the general patterns of female employment in
Bangladesh. Aggregated data shows that despite agriculture being the single largest employer of women
in the country, the proportion of women in this sector is the lowest across all South Asian countries
(World Bank, 2008). Despite the very visible international presence that the RMG sector of Bangladesh
enjoys, the largest increase in female employment has been in the NGO and services sector. It is estimated
that between 1999-00 and 20020-03, female employment in this sector grew by an estimated 29% per
annum (ibid.). The rapid spread of microfinance in the country has also contributed to increasing women’s
employment in small microenterprises. Afrin et al (2009) use survey data from women belonging to two
microcredit institutions and find that borrowing small amounts of money from these institutions led to
an increase in female entrepreneurship. This effect, they argue, manifests itself through the development
of financial management skills and group identity. There is conflicting evidence on the growth (or decline)
of gender based gaps in employment and wages in Bangladesh over the last decade. Ahmed and Maitra
(2011) find that wage gaps between men and women exist at all wage levels and the gap is, in fact, larger
at lower levels of wages. In an updated paper using the 2009-10 LFS, Ahmed and McGillvray (2015) find
that between 1999-2010 wage gap between men and women fell in the last decade, driven by rapid
narrowing of the gap at low wage levels.

No discussion of female in labor in Bangladesh is complete without looking at the role of women in
agriculture, the sector that currently employs close to 60% of the female labor force. Researchers have
looked at various aspects of agriculture and female labor including whether female work on family farms
leads to increases in agricultural productive (Rahman, 2010) and an increase in autonomy (Anderson &
Eswaran, 2009). Rahman (2010) uses a survey of over 1800 households in 16 villages in Bangladesh to
study the role of female labor in improving productivity and finds that female labor contributed nearly

67

28% of total labor use on farms, with most of this coming from the family. He further finds that female
labor is relatively easily substitutable for male labor in terms of productivity, thus arguing in favor of
expanding markets for hired female agricultural labor, a conclusion that is strengthened by the findings
of Anderson and Eswaran’s (2009). They find that female labor outside the family farm contributed to
increasing female autonomy by providing women access to wage income5. On the other hand, while work
on own family farms generated non-wage income, it afforded no agency or autonomy to women whose
labor was seen as free and thus devalued in the household.

There is no consensus in academic research on whether Bangladeshi women’s increased economic
participation is something to be celebrated, particularly given reports of workplace harassment, unequal
wages6, lack of access to health and maternity benefits etc. Hossain and Tisdell (2005) use large scale data
to measure improvements in female education, wages and employment and find that not only have there
been remarkable improvements in education, but that higher education is positively correlated with
increased work participation. While this may seem like an obvious conclusion, the neighboring country
of India exhibits an entirely opposite trend, with LFP falling for women with progressively higher levels
of education.

2.2

Women’s work in West Bengal

The scholarship on female labor in West Bengal has been largely restricted to small scale qualitative
studies, with a few state-wide studies using survey data. One area that has attracted widespread attention
is political reservation for women in local governance, not only in West Bengal, but across all states of
India. Pardhan et al (2010) study the reservation of seats in local government for SC/ST and women and
its impact on, among other things, female headed households in West Bengal. Contrary to what theory
predicts, they find that constituencies that were reserved for women saw no improvement in benefits
targeted to female-headed households. On the other hand, female- headed households in constituencies
5
6

This is also corroborated by Kabeer et al (2011)
For a discussion of how the gender wage gap in Bangladesh has changed over time see Ahmed & McGillivray (2015)

68

reserved for SC/ST candidates experienced better targeted policies and the villages themselves received
more benefits. A contrary finding is reported by Chattopadhyay and Duflo (2004) who study female
participation in village meetings and public goods provisions in villages with and without a female villagehead. They find that in West Bengal, not only did women participate more in village meetings when their
head was a female, but they also demanded, and got, better access to drinking water, and more importantly
better roads. Roads are important for women in West Bengal because they are the main source of labor
on these roads (ibid).

While the spread of micro-finance in India is relatively nascent, the country has experimented
successfully with the concept of women cooperatives. This cooperative movement took root in western
India but slowly spread towards other parts of the country. However, Mayoux (1995) reports that the
success of such cooperatives may be limited. In a study of weaving and handloom cooperatives in West
Bengal, she finds that very few cooperatives operate as such, and many have either failed or serve as fronts
for male merchants who install female relatives as proxy heads. This is consistent with findings regarding
political reservations for women, where researchers have often found that women serve as lame-duck
candidates for their husbands who are incarcerated or otherwise ineligible to stand for elections.

It is argued that low female LFP in West Bengal is due to the fact that women are not involved in
agriculture as much as in other states. Despite that, agriculture and allied sectors employ close to 42% of
the rural female workforce in West Bengal. Sinha (2005) examines rural female work participation rates
(WPR) in 4 districts within West Bengal using data from the Indian Census and finds that districts with
better infrastructure and higher levels of education exhibit lower rural female WPR. While this may seem
counterintuitive, it points to a troubling trend of female LFP in India as whole. Entry into the labor force
does not appear to be a sign of empowerment, rather a sign of acute distress. In the previous chapter I
show how female LFPR is highest for the lowest income groups and begins to decline rapidly as incomes
increase, particularly for Hindus and lower caste groups.

69

Despite the commonality of language and culture across the state of West Bengal, there is also
considerable heterogeneity based on caste and religion. Muslims, who form roughly 14% of the Indian
population, make up more than 25% of the population of West Bengal. Muslim women have had, and
continue to have low rates of labor force participation. While there are some who argue that this is due to
the conservative nature of the community that restricts women’s mobility, the evidence for this is largely
anecdotal and not backed by data. One such anomaly to this norm are the districts of Murshidabad and
South 24 Parganas7 where female LFPR is comparatively quite high. Chakraborty & Chakraborty (2010)
use a primary survey data to study how observable covariates such as education, ownership of land and
age affect female LFPR in these districts, and to study how much of the male female wage gap is
attributable to differences in endowments or discrimination. They find that the high female LFPR in these
two districts is due to the concentration of women in home based self-employment. However, they find
that both Muslim men and women in these districts earn lower wages indicating lack of access to wellpaying jobs for members of this community8. These issues point to the complex nature of discrimination
faced by women in heterogeneous state such as West Bengal. This discrimination is mediated not only
through the lens of gender, but also through multiple layers of caste, religion and ethnicity. A recognition
of the multi-faceted nature of this exclusion is critical to understanding both the presence and absence of
women from the labor force in West Bengal.

Barring a recent World Bank report on Bangladesh (World Bank, 2008) we are not aware of any study
that has compared labor and employment outcomes between Bangladesh and West Bengal. Basu and
Amin (2000) study the remarkable decline in fertility in Bangladesh and compare it with the equally
impressive decline in fertility in West Bengal over the two decades following 1970. This, they argue, is
due to the exchange and diffusion of attitudes between both regions due to the common Bengali language
that facilitated exchange of progressive ideas and attitudes. The common language allowed the formation

7
8

Both these districts have a very high concentration of Muslims- Mushidabad (63.67%) and South 24 Parganas (33.3%)
For an analysis of caste and religion based ethnic employment enclaves see Bordia Das (2008)

70

of a common Bengali identity and nationalism that led to ‘a secular society that is somewhat at odds with the
general socioeconomic development of the regions’. Banerjee et al.(2002) study tenancy reforms in West Bengal
and compare it with Bangladesh where no such land reforms took place and find that the reforms led to a
significant improvement in agricultural productivity in West Bengal. Other than a handful of these
studies, no other research has exploited the homogenous culture and geography but contrasting political
organization that characterizes West Bengal and Bangladesh.

3

Methodology
Our primary outcome of interest is mean female labor force participation rate differences between

women from Bangladesh and West Bengal. The aim of this paper is to explore the contribution of
observable background characteristics in explaining the LFPR gaps between these two groups of women.
Our methodology is twofold- we first examine the combined role of all covariates in explaining the LFPR
gaps and this gives us the total, predicted and unexplained gap. We then look at the role of individual
characteristics in explaining the total and the predicted gap. We use two methods for our overall
decomposition analysis- DFL method based on DiNardo, Fortin and Lemieux (1996) and the OB method
based on Oaxaca (1973) and Blinder (1973). For assessing the role of individual characteristics, we use an
extension of the OB method. We use two decomposition methods as a robustness check and in a correctly
specified model, the two estimates should be quite similar to each other. Moreover, the OB method
provides a very simple and intuitive way of further decomposing the explained gap, something that the
DFL method does not. Our observable covariates include age categories, marital status, education
categories, rural residence indicator, number of children below the age of 5, household size, indicator for
female headed household, and age of the household head. We also include an indicator for whether the
district of the respondent is a border district between the two countries, in order to capture some regional
pattern in the labor market behavior of women9.

9 The assumption here is that women living in districts on either side of the border are likely to be similar in all respects
except their country of residence and hence the political regime under which their district operates. Including this indicator
allows us to potentially capture the effect of differences in political regime (and political history) on labor market participation.

71

3.1

DFL method

Reweighting methods continue to be a popular tool in economics to study gender, ethnicity and race
based gaps in various health and labor market outcomes10. The intuition behind reweighting is simple- if
we want to study outcome differences between group B and W, we reweight group B so that its
distribution of observables closely matches that of W, while retaining its own mapping from observables
to outcome. Essentially, we give more weight to those Bs whose observables are similar to the Ws in our
sample and progressively less weight to the Bs whose background characteristics are different from the
Ws. In our sample, the B signifies Bangladeshi women while W refers to women from West Bengal.

To begin, define the probability density for outcome y of country group c with background
characteristics X as:
(1)

𝐹(𝑦|𝑐) = ∫𝑥 𝐹(𝑦|𝑐, 𝑋)𝑑𝐹(𝑋|𝑐)

From here it is easy to construct a valid counterfactual density of the following form:
𝐹(𝑦|𝑐𝑦|𝑋 = 𝐵, 𝑐𝑋 = 𝑊) = ∫𝑋 𝐹(𝑦|𝑐 = 𝐵, 𝑋)𝑑𝐹(𝑋|𝑐 = 𝑊)

(2)

This counterfactual density is valid only when the changing the marginal distribution function from
𝑑𝐹(𝑋|𝑐 = 𝐵) to 𝑑𝐹(𝑋|𝑐 = 𝑊) leaves the conditional distribution 𝐹(𝑦|𝑐 = 𝐵, 𝑋) unchanged. Using
weights of the form 𝜑𝐵→𝑊 (𝑋) the counterfactual density in (2) can be written as
(3)

𝐹(𝑦|𝑐𝑦|𝑋 = 𝐵, 𝑐𝑋 = 𝑊) = ∫𝑋 𝐹(𝑦|𝑐 = 𝐵, 𝑋)𝜑𝐵→𝑊 (𝑋)𝑑𝐹(𝑋|𝑐 = 𝐵)
Where using Bayes’ rule, we can write 𝜑𝐵→𝑊 (𝑋) as
𝑑𝐹(𝑋|𝑐
𝑑𝐹(𝑋|𝑐

= 𝑊)
= 𝐵) ≡

Pr(𝑐=𝑊 |𝑋)
Pr(𝑐=𝐵|𝑋)

∗

Pr(𝑐=𝐵)
Pr(𝑐=𝑊)

(4)

We use logit (any binary model can be used) to calculate Pr(𝑐 = 𝑖 |𝑋), which is the probability of
being from country group i as a function of background characteristics X. The second part, that is

10 For a comprehensive review of decomposition methods in economics see Fortin et al (2010). For an application of DFL
method to studying racial gaps in infant mortality rates in the US see Elder et al. (2011)

72

Pr(𝑐 = 𝑖) is simply the unconditional proportion of population in group i. In constructing the
counterfactual weights, we use data for Bangladeshi women pooled with the corresponding comparison
group that is West Bengal women, and then obtain three set of means: mean Bangladeshi female LFPR
(B), mean West Bengal female LFPR (W) and mean counterfactual Bangladeshi female LFPR (B’). The
gaps in labor force participation are calculated as follows: Total Gap (T) = B – W, Explained Gap (E) =
B - B’ and Unexplained Gap (U) = B’ - W.

3.2

Oaxaca-Blinder method

The Oaxaca-Blinder (OB) method for wage decomposition is the first and one of the most well-known
decomposition techniques in economics. The original model used differences in means to break down
male-female wage gaps into an explained component that can be attributed to differences in endowments
such as education and experience and a residual component that is attributed to discrimination (and other
unobservable characteristics).

In keeping with the notation developed in the previous section, let the groups be denoted by B and
W. We are interested in knowing how much of the total gap T can be explained by the covariates X and
how much remains unexplained.
𝑇 = 𝐸(𝑦𝐵 ) − 𝐸(𝑦𝑊 )
Where 𝑦𝑐 = 𝐹(𝑋′𝑐 𝛽𝑐 ) and F is a mapping of X to y. In our case, the mapping is linear, that is, 𝑦𝑐 =
(𝑋′𝑐 𝛽𝑐 ). According to the OB model, we can estimate T by using the sample differences in mean values
of 𝑦𝐵 and 𝑦𝑊 . More formally
̅̅̅̅̅̅̅̅̅̅
̅̅̅̅̅̅̅̅̅̅̅
̅̅̅̅̅̅̅̅̅̅̅ ̅̅̅̅̅̅̅̅̅̅̅
𝑇 = ̅̅̅
𝑦𝐵 − ̅̅̅̅
𝑦𝑊 = [(𝑋′
𝐵 𝛽𝐵 ) − (𝑋′𝑊 𝛽𝐵 )] + [(𝑋′𝑊 𝛽𝐵 ) − (𝑋′𝑊 𝛽𝑊) ] = 𝐸 + 𝑈

(5)

Here the first part is explained due to differences in characteristics, whereas the second part is
unexplained or due to differences in coefficients or returns to characteristics. An extension of the OB also
allows us to calculate the role of individual covariates (or group of covariates) in predicting the explained

73

𝑍
𝑍
̅̅̅̅̅̅̅̅̅̅
̅̅̅̅̅̅̅̅̅̅̅
gap. These estimates are calculated as [(𝑍′
𝐵 𝛽𝐵 ) − (𝑍′𝑊 𝛽𝐵 )] where 𝑍𝑖 is a subset of the variables from X

and 𝛽𝐵𝑍 is the corresponding coefficient from regressing y on Z for group B.

4

Data and Descriptive Statistics
We use two main sources of data for our analysis- the 66th and 68th rounds of National Sample Survey

(NSS) of India and the 2010 Bangladesh Labour Force Survey (LFS). The NSS on Employment and
Unemployment in India is a nationally representative survey, usually conducted every five years to
provide a comprehensive assessment of the labor market situation in the country. Each round covers
between 100,000-120000 households and 450,000-600,000 individuals. We pool the data on West Bengal
for the 2009-10 (66th) and the 2011-12 (68th) rounds to provide us with adequate sample size for the state.
Breaking convention, the 68th round of the NSS was conducted 2 years after the previous round compared
to the norm of 5 years. Since the period between the two surveys is relatively short we can pool
observations from the two surveys without any loss of methodological rigor.

Like the NSS, the Bangladesh LFS is conducted at regular intervals to measure employment outcomes
in the country. It is a nationally representative survey that covers close to 200,000 individuals and we use
data from the last publicly available round which was conducted in 2010. Our working sample consists of
women between the ages of 15-59. We restrict our sample to women below the age of 60 as LFPR drops
quite dramatically for women over the age of 60 which incidentally, is also the official retirement age in
the countries. However, the retirement age legislation applies to only a small proportion of women
working in the organized and public sectors in the two countries, and LFPR begins to drop after the age
of 50 for women in both regions, particularly in Bangladesh. The NSS asks information on LFPR using 3
different measures- usual status (one year reference), weekly status (one week reference) and daily status.
The LFS however only asks information using the one-week reference period. We hence only use the
weekly status measurement for NSS to ensure comparability with the LFS. A person is considered as
being in the labor force if they report being employed or engaged in seeking employment. Unpaid own

74

domestic work is excluded from the definition of employment, however work in household based
enterprise and on own family farms is counted as employment.

We include various individual, household and regional level controls in our main analysis as well in
the extension. At the individual level, we control for marital status and control for level of education by
dividing it into five mutually exclusive categories based on highest level of completed schooling-no
education, complete primary school or lower, middle school, secondary and higher secondary and college
and above. Controls for household head’s education are similarly constructed. Controls for age of
respondent and age of head are also added. To control for the effect of fertility decisions and family size
on labor market behavior, we include as covariates, household size and the total number of children below
the age of 5 in the household. This is reasonable since child rearing responsibilities in India are often
shared among all women of the household when living in joint families.

Concurrent with recent literature, household income and wealth is believed to be an important
component driving the decision of women to participate in the labor force. The NSS does not report
measures of income but does provide monthly household per capita consumption expenditure. The LFS
provides no information on income or household consumption. To control household wealth and income
we control for characteristics of the household head and male household members that we believe proxy
for some measure of permanent household income. Unless specified otherwise, whenever the full sample
of women is used, we also control for minority status where minority is defined as being non-Muslim in
Bangladesh and non-Hindu in West Bengal and rest of India.

Table 2.7 shows a broad overview of our sample. Bangladeshi women in our sample are marginally
younger, with an average age of 31.7 years compared to 32.4 years for women from West Bengal. They
are also more likely to be married and in general have lower levels of education. The average Bangladeshi
household is more likely to be rural and has 5.15 members, whereas in West Bengal households are
slightly smaller with 4.8 members on average. Households in West Bengal however are slightly more

75

likely to be female headed. While the differences in individual and household characteristics between the
two groups is not very large, we find large gap in rates of labor force participation.

The stark difference in female LFPRs in the two countries is quite evident in the descriptive statistics.
Table 2.8 shows the LFPR for our sample of women in West Bengal is 20.9% compared to 38.2% for
women in Bangladesh. This disparity seems restricted to women, since men from the two regions show
similar LFPRs on average. Conventional wisdom dictates that LFPR should rise with education, since
accumulation of human capital affords more employment options to the individual and the LFPR of
women in Bangladesh confirms this particularly at the very highest level of education where LFPR rises
dramatically. Paradoxically, West Bengal exhibits the exact opposite trend with LFPR for women falling
with progressively higher levels of education before rising at the college and higher level. If we were to
argue that culture is an important driver of the decision to participate in the labor force, and that religion
is an important marker of culture, we would expect Muslim (or Hindu) women in both regions to have
similar rates of LFPR. However, we see that wide disparities in LFPR persist even when disaggregating
by religion. In fact, unlike the rest of India, Muslim women in West Bengal have a higher rate of female
LFP compared to Hindu women.

As Table 2.9 shows, the distribution of female labor force across types of employment also varies
dramatically between the two regions, whereas the male labor force is more similarly distributed. Close
to 80% of Bangladesh’s female labor force reports being self-employed, while only 50% of those in West
Bengal report working in self or family owned enterprises. The other staggering difference is the wide
gap in those who report being casual laborers- 25% for West Bengal versus 6% for Bangladesh. A cursory
look at these descriptive statistics may lead us to ask whether female employment in West Bengal could
be seen as a sign of distress rather than empowerment. A closer look at the distribution of workforce by
industry will help us better understand where women are being employed.

76

Tables A4 and A5 give details on the industry wise distribution of the female workforce in West
Bengal and Bangladesh. In contrast to the rest of India, where a large proportion of people continue to
depend on agriculture for employment, it is the manufacturing industry which is largest provider of
employment for women in West Bengal. Agriculture and allied sectors employ 28% of the female
workforce in the state followed by education at 9%. It is worth noting that industries that are traditionally
considered better paying such as financial services, communication, administrative services etc., employ a
very small proportion of the female work force in WB which is surprising considering close to 18% women
report being in regular or salaried employment. This means that even those women who are in regular
employment, are predominantly employed in traditionally low paying industries. One such industry is
paid domestic work, which is captured by the ‘activities of households as employers’ industry classification.
This sector employs around 9.4% of the West Bengal female workforce, a sizeable proportion.

Like West Bengal, the two most important industries employing women in Bangladesh are
Agriculture and Manufacturing. Despite the spectacular growth of manufacturing in the country over the
last two decades, agriculture and allied sectors continue to employ a vast majority of the country’s female
workforce. On the other hand, close to 12% of women in the country are employed in manufacturing, of
which the readymade garment industry is an important part. This is followed by domestic and other
service work at 9% and trade at 6%. The education sector, which employs close to 9% of West Bengal
women, plays a much smaller role in Bangladesh, however like West Bengal, the proportion of
Bangladeshi women in well-paying formal sector industries is limited.

We see that women from both the regions have very distinct observables and the workforce
distribution is also quite dissimilar. Before we move on to our main result we run a simple OLS regression
of labor force partition on all our observables. While this may not be a perfect specification, it helps us
understand the relation between LFP and observables for the two groups and allows us to examine which
variables matter in explaining LFP in the two regions.

77

As Table 2.1 shows, not only are the observables for the two groups very different, their LFP response
to the observables is also markedly different. Perhaps the most striking is the correlation between
education and LFP. While higher level of education is associated with a strong positive probability of
being in the labor force for Bangladeshi women, the opposite seems to be the case for women from West
Bengal. As an example, women with high school education and significantly less likely to be employed in
West Bengal compared to women with no education. In comparison, compared to women with no
education, Bangladeshi women’s probability of being in the labor force is higher for every level of
education. Similarly, while being married is a significant negative predictor of labor force participation in
West Bengal, it predicts the opposite in Bangladesh. West Bengal women from female headed households
are much more likely to be in the labor force, this could be driven by the fact that female headed households
tend to be more impoverished. Paradoxically, we see the opposite effect in Bangladesh. Having at least
one male salaried worker in the household is associated with a significantly lower probability of women
working in both regions, though the result is much more magnified for Bangladesh. Unlike the rest of
India, living in a rural area does not predict a higher rate of female LFP for women in West Bengal. In
fact, on aggregate the rural-urban female LFPR in West Bengal is quite similar. The magnitude and
direction of the coefficients on the other covariates are in the direction we would expect. These results
will help us better understand the findings from our main decompositions.

5

Main Results
Our first method for the decomposition analysis is the Oaxaca-Blinder method. We use this method

to decompose the wage gap into a component that is explained by differences in mean endowment levels
and the other that is unexplained or due to unobservables. The OB method also allows us to further
decompose the explained portion into three parts- individual, household and regional11.

11 Personal includes dummies for each of the age groups, dummies for each of the education categories, and an indicator for
marital status. Household category includes HH size, indicator for female headed household, indicator for being a religious
minority, number of children under the age of 5 in the household, and characteristics of male household members and the HH
head. Finally, the regional category includes an indicator for rural residence and indicator for living in a border district.

78

The OB method predicts a lower gap than the 17.3 percentage points that exists. That is, if both
groups of women had the same mean observable characteristics, we would expect the LFPR difference
between the two groups to be lower by 4 percentage points. More broadly, the mean characteristics of
women that predict a lower LFPR in West Bengal, predict a slightly higher LFPR in Bangladesh. More
disaggregated analysis shows that this effect is driven largely by personal and regional characteristics,
more specifically, by the education distribution of women in the two regions and the distribution of women
in districts on either side of the border12.

Our second decomposition method is the DFL method, which allows us to reweight the population so
that both groups have the same mean observable characteristics while retaining their own mapping from
observables to outcomes. In other words, we reweight the Bangladeshi population is such a way to
construct a counterfactual distribution of observables which is nearly identical to the distribution of
characteristics for the West Bengal group. To see whether we are in fact able to construct such a
distribution we can look at the summary statistics for the counterfactual Bangladeshi population and see
how well it compares to actual distribution of characteristics for the West Bengal group. As we can see in
Table 2.3, for all the indicators, the counterfactual distribution is very similar to the actual distribution.
In particular, the distribution across various levels of education, which was different between the original
two groups, now looks nearly identical.

Both the DFL and OB methods give us results that are similar in magnitude and direction. The DFL
method predicts that if women in Bangladesh had the average characteristics of women in West Bengal
then their LFPR would have been lower by 4.7 percentage points on average. This paradoxical result is
magnified even more if we consider the fact that, compared to the original group, the counterfactual group

12 Border districts in West Bengal include North 24, Nadia, Murshidabad, Maldah, Dakshin Dijnapur, Uttar Dijnapur,
Jalpaiguri & Koch Behar. Border districts in West Bengal include Chiadanga, Satkhira, Jessore, Jhenaidah, Meherpur, Kushtia,
Rajshahi, Nawabganj, Naogaon, Joypurhat, Dinajpur, Thakurgaon, Panchgarh, Nilphamari, Lalmonirghat & Kurigram.

79

has a smaller HH size, is slightly better educated and is less likely to be married on average, all of which
characteristics have been shown to increase participation rates for women.

6

Extensions
In this section, we conduct additional analysis to study the main results in more detail. In addition to

studying various sub-groups, we also introduce an additional comparison group to better understand the
processes that may explain the disparity in female LFPR between West Bengal and Bangladesh.

6.1

How do results vary for different sub-groups?

Table 2.4 shows the Oaxaca-Blinder decomposition for three different sub-groups in West Bengal and
Bangladesh. As mentioned previously, in our main specification we control for whether respondent lives
in a border-district between the two regions with the assumption that these two groups are likely to share
common cultural and agro-climatic conditions but differ in the political conditions under which the district
operates. In this sub-section, we restrict our sample to these border districts only. We find that LFPR is
higher for respondents who live in the border districts on either side, which is expected since this is the
region where a lot of agricultural activity is concentrated. However, like our main results, we find that
background characteristics are unable to explain the reason for the 20.3 percentage point gap in female
LFPR between the two groups. We conduct a similar analysis for the sub-groups of religious minorities
and Muslims. West Bengal has one of the highest proportion of Muslims across states in India and we
wanted to study whether religious identity matters in explaining some of the large gap in female LFPR
in the two regions. We find that the results for both the minority and Muslim sub-group mirror the main
results. In both cases, background characteristics predict a negative LFPR gap between the two regionswhile in case of minority religious group this effect is driven by regional variables, in case of Muslims it
is largely driven by household characteristics.

80

6.2

Do results change with different mappings?

Decomposition methods can give different results based on which group is used as the reference
category. In both our main and sub-group analysis of previous sections, we have used Bangladesh as the
reference category. The Oaxaca-Blinder methodology allows us to use as the reference category, either of
the two groups, as well as a pooled sample of both groups (Neumark, 1988; R. L. Oaxaca & Ransom,
1994)13. In this section, we show how the results vary based on the choice of reference category.

For each of the components in Figure 1, we graph three bars that show the predicted gap based on
using Bangladesh, pooled sample and West Bengal as the reference category respectively, in addition to
the total overall gap. While there are some differences in magnitude based on the different reference
categories, the direction of the gaps is largely similar overall. Disaggregating by individual covariates
also shows that the three reference groups give broadly similar results. As an example, regional
distribution explains -0.95% of the total gap if Bangladesh is used as the reference, in comparison to 0.99% and -0.80% if the pooled or the West Bengal sample is used respectively.

6.3

How do the two regions compare with the rest of India?

To further understand the disparate female LFPR in the two regions, we introduce an additional
comparison group, rest of India14. In Tables 5 & 6, we use the India group as our reference category and
compare it with the West Bengal and Bangladesh populations. Like the previous table, we also decompose
gaps between religious sub-groups. Female LFPR in the India sample is significantly higher than in West
Bengal, at 28.7% compared to 20.9%. However, decomposition predicts a negative LFPR gap between the
two groups. Detailed analysis15 indicates that this is driven predominantly by the Muslim and rural
variable. In West Bengal being Muslim is not a predictor of lower levels of LFPR, however in the India

For detailed implementation see Jann (2008)
The rest of India sample includes 17 major states of India, excluding WB. Together with West Bengal, these states cover
roughly 92-95% of India’s population. The states included in our analysis are Uttar Pradesh, Bihar, Maharashtra, Madhya
Pradesh, Andhra Pradesh, Tamil Nadu, Karnataka, Gujarat, Rajasthan, Orissa, Kerala, Assam, Punjab, Jammu & Kashmir,
Jharkhand, Uttaranchal and Chhattisgarh. In this subsection, we use India as the reference category.
15 Not shown
13
14

81

sample, being Muslim is associated strongly with low female participation16. Similarly, unlike the rest of
India, living in rural areas in West Bengal is not associated with higher levels of female employment. This
could be in part due to the falling importance of agriculture in the state, both in terms of its share in
employment and in state domestic product.

If we consider the minority sub-sample, we find that covariates more than explain the entire 1.09
percentage point gap between the two groups. As mentioned previously, West Bengal is one of the only
major states in India, where Muslim LFPR is at par or slightly higher than LFPR for Hindus. We see
that in the Muslim sub-sample, West Bengal female LFPR is significantly higher than the rest of India.
Of this 6.2 percentage point gap, 20% is explained by observable covariates, particularly household and
regional characteristics- unlike for Hindus, being in rural areas is associated with higher female LFPR
among Muslims in West Bengal.

We do a similar analysis comparing the rest of India with Bangladesh. For all the three sub-samples,
Bangladesh has significantly higher LFPR, than India. In the full sample, of the total 9.6 percentage point
gap, only 0.63% is explained by differences in covariates. This effect is largely driven by the household
characteristics, in particular minority status. In Bangladesh, being a religious minority is associated with
higher female LFPR, whereas in India it is the opposite, driven by the low LFPR of Muslims who are a
minority in India. Similarly, as per Table 2.13, being in rural areas in Bangladesh is associated with only
slightly higher LFPR, however for the rest of India, rural areas have significantly higher female
participation. This rural residence is what drives the explained portion of the minority and Muslim subsample gap between Bangladesh and Rest of India. On the other hand, personal characteristics predict a
positive LFPR gap- if the rest of India had lower levels of education like Bangladesh, they would have
exhibited between 1.3-1.5 percentage points higher rates of female participation. This is because unlike
Bangladesh, higher levels of education are associated with lower rates of female participation in India.

16

See Table 2.13

82

7

Conclusion
In this chapter, we attempt to study the factors driving the divergent female labor force participation

rates in West Bengal and Bangladesh, two historically, culturally, topographically and linguistically
similar regions who have had a disparate political and development paths over the last six decades. We
used decomposition methods to parse out the contribution of observable characteristics in explaining gaps
in LFPR for the entire population as well as for various regional and religious sub-groups. We found in
our analysis that background characteristics do not help explain the significantly lower rates of female
employment in West Bengal compared to Bangladesh and even compared to the rest of India. One
explanation given for the withdrawal of women from the labor force in West Bengal is the declining
importance of agriculture in the state. Table 2.10 shows that even though agriculture is the single largest
sector employing women in the state, it’s share has declined substantially in the last 25 years, from 45%
to 28%. In comparison, in the neighboring states of Bihar and Assam, more than 70% of women were
employed in agriculture in 2011-12. Agriculture’s share of employment is much lower even when
compared with the all-India average of 59%. While manufacturing has risen in relative importance, most
of the work in this sector is in the unorganized tobacco and textile industries, where wages are low and
erratic.

These results lead us to question what is driving the LFP behavior of women in the two regions.
Several sociological and anthropological studies have bemoaned the lack of mobility that women in the
South Asian region suffer from. These restrictions on mobility often lead to girls being pulled out of
school earlier than boys and ultimately to women and girls being disallowed from seeking employment
outside of their homes. Interestingly, this is even more the case for relatively affluent upper caste/class
households, where women are seen as bearers of household honor (Eswaran, Ramaswami, & Wadhwa,
2013). Perhaps the most glaring manifestation of this is the significantly higher LFPR of Schedule
Caste/Schedule Tribe women at the all-India level compared to upper caste Hindu women. If this is indeed

83

the case, then this could explain our results to some extent. By giving more weight to those Bangladeshi
women who are better educated, unmarried, urban and young, we may be capturing the effect of being
relatively ‘elite’. This effect of unobservables on LFP decisions is not entirely captured by observable
covariates and perhaps represents the unmeasured effect of culture. An additional factor that we cannot
explicitly control for is the widely varying political organization of the two countries. Until recently,
West Bengal had been under the leadership of the Communist Party of India for nearly four decades,
while Bangladesh oscillated between military rule and democracy with a stress on free-market policies.
These differing political trajectories have created widely varying economic structures which have in turn
led to divergent demands for female labor in the two countries.

84

APPENDICES

85

APPENDIX A

MAIN TABLES AND FIGURES

86

Table 2.1: OLS of LFP (x100) on Observables
West Bengal
Bangladesh
Education Level
No Schooling
Primary or Lower
Middle
Secondary & H Secondary
College & Higher
Age Group
15-24
25-39
40-49
50-59
Married^
HH Size
Children under 5
Female headed^
Adult Salaried Male in HH^
HH Head Age Group
Under 30
30-45
45-60
Above 60
HH Head Education
No Schooling
Primary or Lower
Middle

Excluded

Difference

Excluded

-4.414
(1.812)**
-10.018
(2.130)***
-11.353
(2.283)***
6.374
(2.945)**
Excluded

4.543
(0.625)***
6.572
(0.746)***
1.812
(0.749)**
38.143
(2.136)***

-8.957
(1.915)***
-16.591
(2.255)***
-13.166
(2.401)***
-31.769
(3.637)***

Excluded

12.799
(1.826)***
6.040
(2.189)***
-2.175
(2.320)
-12.121
(1.788)***
-0.292
(0.325)
-0.257
(1.124)
9.440
(2.645)***
-3.381
(1.372)**
Excluded

7.738
(0.649)***
9.125
(0.781)***
-28.521
(0.676)***
11.954
(0.619)***
0.962
(0.125)***
-1.644
(0.309)***
-8.573
(1.148)***
-21.594
(0.559)***

5.061
(1.936)***
-3.085
(2.323)
26.346
(2.415)***
-24.074
(1.891)***
-1.254
(0.348)***
1.387
(1.165)
18.013
(2.882)***
18.213
(1.48)***

Excluded

-6.010
(2.494)**
-4.077
(2.558)
-3.259
(2.892)
Excluded

-1.481
(0.749)**
-0.913
(0.770)
-3.204
(0.859)***

-4.529
(2.603)**
-3.164
(2.669)
-0.056
(3.015)

Excluded

-1.713
(1.685)
-3.784

87

1.744
(0.603)***
2.542

-3.456
(1.788)**
-6.326

Table 2.1 (cont’d)

Secondary & H Secondary
College & Higher
Minority^
Rural^
Border^
Constant

N
R2

(1.961)*
-3.726
(2.117)*
-3.203
(2.614)
-0.064
(1.492)
-1.146
(1.194)
4.180
(1.237)***
35.852
(3.193)***

(0.733)***
3.720
(0.704)***
-1.474
(1.129)
4.624
(0.655)***
0.644
(0.522)
5.206
(0.510)***
22.672
(1.039)***

15,396

53,857

0.05

(2.092)
-7.446
(2.229)***
-1.729
(2.846)
-4.688
(1.628)***
-1.790
(1.302)
-1.026
(1.338)
13.180
(3.355)***

0.11

Note: Sample includes persons aged 15-59, who are not heads of household. Activity based on selfreported status in week preceding survey. Robust standard errors in parenthesis ^Binary (1=yes, 0=no)
***Significant at 1% **Significant at 5% *Significant at 10%

88

LFPR

Gap

Explained Gap

Table 2.2: Decomposition of Female LFPR Gap
Oaxaca-Blinder method
Bangladesh
38.21
(0.221)***
West Bengal
20.92
(0.596)***
Total
17.30
(0.636)***
Explained
-4.064
(0.314)***
Unexplained
21.361
(0.690)***
Personal
-2.369
(0.239)***
Household
-0.740
(0.199)***
Regional
-0.955
(0.104)***
69,253

N

DFL Method
38.21
(0.221)***
20.92
(0.599)***
17.30
(0.638)***
-4.743
(0.419)***
22.04
(0.697)***

69,253

Note: Sample includes persons aged 15-59, who are not heads of household. Activity based on self-reported
status in week preceding survey. Robust standard errors in parenthesis ^Binary (1=yes, 0=no) ***Significant at 1%
**Significant at 5% *Significant at 10%

89

Table 2.3: Descriptive Statistics for Counterfactual Population
Bangladesh
West
Counterfactual
Bengal
Education Level
No Schooling
Primary or Lower
Middle
Secondary & H Secondary
College & Higher
Age Group
15-24
25-39
40-49
50-59
Married^
HH Size
Children under 5
Female headed^
Adult Salaried Male in HH^
HH Head Age Group
Under 30
30-45
45-60
Above 60
HH Head Education
No Schooling
Primary or Lower
Middle
Secondary & H Secondary
College & Higher
Minority^
Rural^
Border^
N

0.2790
0.3303
0.1781
0.1528
0.0598

0.2941
0.3265
0.1772
0.1462
0.0560

0.2805
0.4266
0.1825
0.1105
0.8150
4.7624
0.4058
0.0533
0.1865

0.2985
0.4180
0.1766
0.1070
0.8027
4.7615
0.4076
0.0557
0.1870

0.1067
0.3803
0.3741
0.1389

0.1081
0.3849
0.3669
0.1401

0.2737
0.3573
0.1444
0.1372
0.0874
0.2717
0.7317
0.3847

0.3048
0.3474
0.1409
0.1259
0.0809
0.2620
0.7339
0.4089

53,857

15,396

Note: Sample includes persons aged 15-59, who are not heads of household. Activity based on
self-reported status in week preceding survey. Robust standard errors in parenthesis ^Binary (1=yes,
0=no) ***Significant at 1% **Significant at 5% *Significant at 10%

90

LFPR

Gap

Explained Gap

N

Table 2.4: Decomposition of Female LFPR Gap by Sub-Group
Only Border
Minority
Districts
Bangladesh
44.1028
42.0292
(0.4756)***
(0.6511)***
West Bengal
23.8449
22.4814
(1.0129)***
(1.3237)***
Total
20.2579
19.5478
(1.1189)***
(1.4752)***
Explained
-1.2036
-5.9742
(0.7032)*
(0.8595)***
Unexplained
21.4615
25.5220
(1.2684)***
(1.6785)***
Personal
-0.9076
-1.6733
(0.4670)*
(0.5782)***
Household
-0.5058
-1.4484
(0.5301)
(0.4211)***
Regional
0.2098
-2.8525
(0.1048)**
(0.5858)***
17,409

10,116

Muslim
37.7405
(0.2350)***
22.3587
(1.3495)***
15.3818
(1.3698)***
-4.4035
(0.5094)***
19.7854
(1.4358)***
-0.7585
(0.4140)*
-2.0094
(0.2313)***
-1.6356
(0.1929)***
50,665

Note: Sample includes persons aged 15-59, who are not heads of household. Activity based on self-reported
status in week preceding survey. Robust standard errors in parenthesis ^Binary (1=yes, 0=no) ***Significant
at 1% **Significant at 5% *Significant at 10%

91

Table 2.5: Decomposition of Female LFPR Gap (West Bengal & India)
Full Sample
LFPR

Rest of India
West Bengal

Gap

Total
Explained
Unexplained

Explained Gap

Personal
Household
Regional

N

Minority

Muslim

28.6547
(0.1900)***
20.9161
(0.5960)***
7.7386
(0.6255)***
-1.1828
(0.1923)***
8.9214
(0.6353)***
-0.0647
(0.1354)
-0.9323
(0.1166)***
-0.1858
(0.0491)***

21.3907
(0.3613)***
22.4814
(1.3237)***
-1.0906
(1.3722)
-2.1073
(0.3419)***
1.0167
(1.4215)
1.1812
(0.2704)***
-1.4789
(0.2505)***
-1.8096
(0.1738)***

16.1950
(0.3980)***
22.3587
(1.3495)***
-6.1637
(1.4070)***
-1.2811
(0.3201)***
-4.8826
(1.4494)***
0.2604
(0.2111)
-0.5036
(0.2371)**
-1.0379
(0.1932)***

218,094

47,378

34,477

Note: Sample includes persons aged 15-59, who are not heads of household. Activity based on self-reported
status in week preceding survey. Robust standard errors in parenthesis ^Binary (1=yes, 0=no) ***Significant at
1% **Significant at 5% *Significant at 10%

92

Table 2.6: Decomposition of Female LFPR Gap (Bangladesh & India)
Full Sample
LFPR

Rest of India
Bangladesh

Gap

Total
Explained
Unexplained

Explained Gap

Personal
Household
Regional

N

28.6547
(0.1900)***
38.2130
(0.2211)***
-9.5582
(0.2915)***
-0.6311
(0.1260)***
-8.9272
(0.3155)***
1.3685
(0.0797)***
-1.6369
(0.1063)***
-0.3626
(0.0275)***
256,555

Minority

Muslim

21.3907
(0.3613)***
42.0292
(0.6511)***
-20.6385
(0.7446)***
-0.4349
(0.2281)*
-20.2036
(0.7801)***
1.5041
(0.1437)***
-0.7635
(0.1554)***
-1.1755
(0.1154)***

16.1950
(0.3980)***
37.7405
(0.2350)***
-21.5455
(0.4622)***
0.3862
(0.2676)
-21.9317
(0.5516)***
1.2875
(0.1360)***
-0.2819
(0.2111)
-0.6194
(0.1146)***

50,404

78,384

Note: Sample includes persons aged 15-59, who are not heads of household. Activity based on self-reported
status in week preceding survey. Robust standard errors in parenthesis ^Binary (1=yes, 0=no) ***Significant at
1% **Significant at 5% *Significant at 10%

93

94

APPENDIX B

SUPPLEMENTARY TABLES

95

Table 2.7: Summary Statistics
West Bengal

Education Level
No Schooling
Primary or Lower
Middle
Secondary & H Secondary
College & Higher
Age Group
15-24
25-39
40-49
50-59
Married^
HH Size
Children under 5
Female headed^
Adult Salaried Male in HH^
HH Head Age Group
Under 30
30-45
45-60
Above 60
HH Head Education
No Schooling
Primary or Lower
Middle
Secondary & H Secondary
College & Higher
Minority^
Muslim^
Rural^
Border^
N

Bangladesh

Rest of India

0.2941
0.3265
0.1772
0.1462
0.0560

0.4050
0.2149
0.1438
0.2242
0.0121

0.3701
0.2017
0.1557
0.2020
0.0705

0.2985
0.4180
0.1766
0.1070
0.8027
4.7615
0.4076
0.0557
0.1870

0.3424
0.3844
0.1679
0.1053
0.8180
5.1541
0.6812
0.0345
0.1893

0.3065
0.4025
0.1792
0.1118
0.7736
5.4888
0.5808
0.0540
0.2063

0.1081
0.3849
0.3669
0.1401

0.1381
0.3704
0.3334
0.1581

0.1004
0.3896
0.3859
0.1242

0.3048
0.3474
0.1409
0.1259
0.0809
0.2620
0.2525
0.7339
40.89

0.4627
0.2061
0.1150
0.1706
0.0455
0.1102
0.8898
0.7533
22.30

0.3230
0.2502
0.1536
0.1801
0.0931
0.1693
0.1203
0.7136

15,396

53,857

202,698

Note: For definition of Rest of India see Footnote 12 , Women age 15-59 ^Binary (1=yes, 0=no)

96

Table 2.8: Labor Force Participation Rates
West Bengal
Bangladesh
Mean
S.E
Mean
S.E

Difference
Mean

S.E

Total
Male
Female
By Religion
Minority
Non-Minority

73.208
20.916

0.869
0.599

63.898
38.213

0.333
0.221

9.310
-17.297

0.931
0.638

39.928
37.337

1.250
0.622

49.546
45.468

0.547
0.202

-9.619
-8.131

1.364
0.654

Muslim
Hindu
By Education
No schooling
Primary or lower
Middle
Secondary & Higher Secondary
College & Higher

32.103
41.915
34.875
33.141
59.485

1.281
1.075
1.252
1.112
1.392

43.210
54.171
51.910
37.102
76.164

0.324
0.403
0.480
0.354
1.232

-11.107
-12.256
-17.034
-3.961
-16.679

1.322
1.148
1.341
1.167
1.859

Male
By Religion
Minority
Non-Minority

76.065
72.202

1.796
0.987

66.341
63.581

0.924
0.356

9.724
8.621

2.020
1.050

Muslim
Hindu
By Education
No schooling
Primary or lower
Middle
Secondary & Higher Secondary
College & Higher

63.581
66.245

0.356
0.969

63.581
66.245

0.356
0.969

12.660
5.957

1.864
1.383

86.278
86.959
66.102
54.553
87.736

2.749
1.309
2.000
1.770
1.187

80.238
81.774
69.710
41.818
81.812

0.660
0.567
0.739
0.551
1.452

6.040
5.185
-3.608
12.735
5.925

2.827
1.427
2.132
1.854
1.875

Female
By Religion
Minority
Non-Minority

22.481
20.361

1.347
0.655

42.029
37.741

0.651
0.235

-19.548
-17.380

1.496
0.696

22.359
20.361

1.373
0.655

37.741
40.810

0.235
0.681

-15.382
-20.450

1.393
0.944

26.006
20.791

1.308
1.069

36.458
41.948

0.343
0.480

-10.451
-21.157

1.352
1.172

Muslim
Hindu
By Education
No schooling
Primary or lower

97

Table 2.8 (cont’d)
Middle
Secondary & Higher Secondary
College & Higher

15.151
13.413
32.747

1.207
1.082
1.937

42.154
33.624
68.785

0.592
0.458
2.061

-27.003
-20.212
-36.038

1.344
1.174
2.828

Note: Sample includes persons aged 15-59, who are not heads of household. Activity based on self-reported
status in week preceding survey.

98

Table 2.9: Distribution of Labor Force
West Bengal
Bangladesh
Mean
S.E
Mean
S.E

Difference
Mean

S.E

Male
Casual
Regular Wage/Salaried
Self Employed
Unemployed

0.332
0.167
0.410
0.091

0.012
0.008
0.011
0.006

0.266
0.220
0.459
0.055

0.004
0.004
0.004
0.002

0.066
-0.053
-0.049
0.036

0.012
0.009
0.012
0.007

Casual
Regular Wage/Salaried
Self Employed
Unemployed

0.222
0.161
0.545
0.072

0.015
0.010
0.016
0.008

0.051
0.094
0.820
0.034

0.002
0.002
0.003
0.001

0.171
0.067
-0.275
0.038

0.015
0.010
0.016
0.008

Casual
Regular Wage/Salaried
Self Employed
Unemployed
Non-Minority Casual
Regular Wage/Salaried
Self Employed
Unemployed

0.194
0.082
0.665
0.059
0.233
0.192
0.498
0.077

0.028
0.015
0.032
0.016
0.017
0.012
0.018
0.009

0.073
0.097
0.801
0.029
0.048
0.094
0.823
0.035

0.005
0.006
0.008
0.003
0.002
0.002
0.003
0.001

Female

By Religion
Female
Minority

0.122
0.028
-0.015
0.016
-0.136
0.033
0.029
0.016
0.185
0.017
0.098
0.013
-0.325
0.018
0.042
0.009
Note: Sample includes persons aged 15-59, who are not heads of household. Activity based on
self-reported status in week preceding survey.

99

Table 2.10: Distribution of Female Workforce by Top 10 Industry- West Bengal
International Standard Industrial
Classification – 2 digit level
1983
1999
2011
Agriculture
Manufacture of Tobacco
Other service activities
Manufacture of textiles
Education
Manufacture of food and beverage
Retail trade
Manufacture of apparel
Wood Work
Health and social work

100

44.82
4.71
13.61
3.74
4.73
6.38
4.71
2.54
2.22
1.92

43.12
14.24
8.18
5.7
4.24
4.1
5.2
1.15
3.63
1.53

28.31
20.44
8.23
9.65
9.61
1.26
3.88
3.65
0.71
2.44

Table 2.11: Distribution of Female Workforce by Top 10 Industry- All India
International Standard Industrial
Classification – 2 digit level
1983
1999
2011
Agriculture
Education
Other service activities
Retail trade
Construction
Manufacture of Tobacco
Manufacture of textiles
Manufacture of apparel
Manufacture of food and beverage
Public administration, Defence

101

74.49
2.11
3.44
3.09
2.09
2.41
1.96
1.61
1.22
1.01

73.84
2.86
3.04
3.78
1.89
2.48
1.99
0.57
1.22
1.11

58.6
5.66
4.12
4.08
5.8
3.48
3.07
3.45
1.24
0.91

Table 2.12: Distribution of Female Workforce by Top 10 Industry- Bangladesh
International Standard Industrial
Classification – 2 digit level
2010
Agriculture
Other service activities
Manufacture of apparel
Retail trade
Education
Manufacture of textiles
Land transport
Construction
Manufacture of food and beverage
Health and social work

102

64.77
9.16
7
6.13
2.04
1.74
1.38
1.34
1.13
1.02

Table 2.13: OLS of LFP (x100) on Observables
West Bengal

Bangladesh

Rest of India

Excluded

Excluded

Excluded

Education Level
No Schooling

Primary or Lower
Middle
Secondary & H Secondary
College & Higher
Age Group
15-24
25-39
40-49
50-59
Married^
HH Size
Children under 5
Female headed^
Adult Salaried Male in HH ^
HH Head Age Group
Under 30
30-45
45-60
Above 60
HH Head Education
No Schooling
Primary or Lower

-4.631
(1.816)**
-10.083
(2.134)***
-11.523
(2.289)***
6.490
(2.948)**
Excluded
12.848
(1.826)***
5.866
(2.188)***
-2.196
(2.309)
-12.070
(1.782)***
-0.344
(0.325)
-0.355
(1.129)
9.224
(2.647)***
-3.119
(1.369)**
Excluded
-6.070
(2.496)**
-3.965
(2.552)
-3.447
(2.888)
Excluded
-2.096

103

4.512
(0.626)***
6.643
(0.747)***
1.878
(0.749)**
38.154
(2.142)***
Excluded
7.755
(0.650)***
9.127
(0.783)***
-28.530
(0.677)***
12.165
(0.619)***
0.897
(0.125)***
-1.784
(0.309)***
-9.043
(1.148)***
-21.928
(0.558)***
Excluded
-1.476
(0.750)**
-0.955
(0.770)
-3.344
(0.859)***
Excluded
1.804

-1.873
(0.584)***
-7.177
(0.617)***
-9.800
(0.613)***
10.039
(0.839)***
Excluded
14.368
(0.576)***
15.269
(0.700)***
9.191
(0.781)***
-3.021
(0.570)***
-1.074
(0.087)***
-1.036
(0.261)***
4.940
(0.812)***
-1.798
(0.417)***
Excluded
-0.575
(0.736)
0.676
(0.760)
-0.266
(0.887)
Excluded
-1.659

Table 2.13 (cont’d)

Middle
Secondary & H Secondary
College & Higher
Minority^
Rural^
Constant
N
R2

(1.673)
-4.118
(1.957)**
-4.105
(2.110)*
-3.875
(2.608)
0.641
(1.486)
-1.022
(1.199)
37.901
(3.117)***
15,396
0.04

(0.603)***
2.523
(0.733)***
3.774
(0.704)***
-1.349
(1.131)
4.431
(0.656)***
1.229
(0.520)**
23.736
(1.032)***
53,857
0.11

(0.552)***
-5.224
(0.624)***
-9.676
(0.610)***
-14.504
(0.715)***
-7.165
(0.420)***
9.123
(0.353)***
29.826
(0.950)***
202,698
0.07

For definition of rest of India see Footnote 12, Robust standard errors in parenthesis ^Binary (1=yes, 0=no)
***Significant at 1% **Significant at 5% *Significant at 10%

104

REFERENCES

105

REFERENCES
Afrin, S., Islam, N., & Ahmed, S. U. (2009). A Multivariate Model of Micro Credit and Rural Women
Entrepreneurship Development in Bangladesh. International Journal of Business and Management, 3(8), p169.
https://doi.org/10.5539/ijbm.v3n8p169
Ahmed, S., & McGillivray, M. (2015). Human Capital, Discrimination, and the Gender Wage Gap in
Bangladesh. World Development, 67, 506–524. https://doi.org/10.1016/j.worlddev.2014.10.017
Almond, D., Edlund, L., & Milligan, K. (2013). Son Preference and the Persistence of Culture: Evidence
from South and East Asian Immigrants to Canada. Population and Development Review, 39(1), 75–95.
https://doi.org/10.1111/j.1728-4457.2013.00574.x
Altonji, J. G., & Blank, R. M. (1999). Race and gender in the labor market. Handbook of Labor Economics, 3,
3143–3259.
Anderson, S., & Eswaran, M. (2009). What determines female autonomy? Evidence from Bangladesh.
Journal of Development Economics, 90(2), 179–191. https://doi.org/10.1016/j.jdeveco.2008.10.004
Antecol, H. (2003). Why is there cross-country variation in female labor force participation rates? The role of male
attitudes toward family and sex roles. Working paper series/Claremont McKenna College. Retrieved from
http://www.econstor.eu/handle/10419/23378
Banerjee, A. V., Gertler, P. J., & Ghatak, M. (2002). Empowerment and Efficiency: Tenancy Reform in
West Bengal. Journal of Political Economy, 110(2), 239–280. https://doi.org/10.1086/338744
Basu, A. M., & Amin, S. (2000). Conditioning Factors for Fertility Decline in Bengal: History, Language
Identity, and Openness to Innovations. Population and Development Review, 26(4), 761–794.
https://doi.org/10.1111/j.1728-4457.2000.00761.x
Bhalotra, S. R., & Umana-Aponte, M. (2010). The Dynamics of Women’s Labour Supply in Developing Countries
(IZA Discussion Paper No. 4879). Institute for the Study of Labor (IZA). Retrieved from
https://ideas.repec.org/p/iza/izadps/dp4879.html
Blau, F. D., & Grossberg, A. J. (1992). Maternal Labor Supply and Children’s Cognitive Development. The
Review of Economics and Statistics, 74(3), 474–481. https://doi.org/10.2307/2109492
Blau, F. D., & Kahn, L. M. (2013). Female labor supply: Why is the US falling behind? National Bureau of
Economic Research. Retrieved from http://www.nber.org/papers/w18702
Blinder, A. S. (1973). Wage Discrimination: Reduced Form and Structural Estimates. The Journal of Human
Resources, 8(4), 436–455. https://doi.org/10.2307/144855
Bordia Das, M. (2008). Minority status and labor market outcomes: does India have minority enclaves?
Washington, DC: World Bank, South Asia Region, Sustainable Development Dep. Retrieved from
http://dx.doi.org/10.1596/1813-9450-4653
Brooks–Gunn, J., Han, W.-J., & Waldfogel, J. (2002). Maternal employment and child cognitive outcomes
in the first three years of life: The NICHD study of early child care. Child Development, 73(4), 1052–1072.
Chakraborty, I., & Chakraborty, A. (2010). Female Work Participation and Gender Differential in Earning in
West
Bengal
(Working
Paper
No.
id:2438).
eSocialSciences.
Retrieved
from
https://ideas.repec.org/p/ess/wpaper/id2438.html
Chattopadhyay, R., & Duflo, E. (2004). Impact of Reservation in Panchayati Raj: Evidence from a
Nationwide Randomised Experiment. Economic and Political Weekly, 39(9), 979–986.
DiNardo, J., Fortin, N. M., & Lemieux, T. (1996). Labor Market Institutions and the Distribution of
Wages, 1973-1992: A Semiparametric Approach. Econometrica Econometrica, 64(5), 1001–1044.
Elder, T. E., Goddeeris, J. H., & Haider, S. J. (2011). A deadly disparity: a unified assessment of the BlackWhite infant mortality gap. The BE Journal of Economic Analysis & Policy, 11(1). Retrieved from

106

http://www.degruyter.com/view/j/bejeap.2011.11.issue1/bejeap.2011.11.1.2821/bejeap.2011.11.1.2821.xml
Eswaran, M., Ramaswami, B., & Wadhwa, W. (2013). Status, Caste, and the Time Allocation of Women in
Rural India. Economic Development and Cultural Change, 61(2), 311 – 333.
Fernandez, R. (2007). Culture as Learning: The Evolution of Female Labor Force Participation over a Century
(Working Paper No. 13373). National Bureau of Economic Research. Retrieved from
http://www.nber.org/papers/w13373
Fortin, N. M., Lemieux, T., Firpo, S., & National Bureau of Economic Research. (2010). Decomposition
methods in economics. Cambridge, Mass.: National Bureau of Economic Research. Retrieved from
http://papers.nber.org/papers/w16045
Heath, R., & Mushfiq Mobarak, A. (2015). Manufacturing growth and the lives of Bangladeshi women.
Journal of Development Economics, 115, 1–15. https://doi.org/10.1016/j.jdeveco.2015.01.006
Hossain, M. A., & Tisdell, C. A. (2005). Closing the gender gap in Bangladesh: inequality in education,
employment and earnings. International Journal of Social Economics, 32(5), 439–453.
https://doi.org/10.1108/03068290510591281
Hossain, N. (2012). Exports, Equity, and Empowerment: The Effects Of Readymade Garments
Manufacturing Employment On Gender Equality In Bangladesh Bangladesh. Retrieved from
https://openknowledge.worldbank.org/handle/10986/9100
Jann, B. (2008). A Stata implementation of the Blinder-Oaxaca decomposition (ETH Zurich Sociology Working
Paper
No.
5).
ETH
Zurich,
Chair
of
Sociology.
Retrieved
from
http://econpapers.repec.org/paper/etswpaper/5.htm
Kabeer, N., Mahmud, S., & Tasneem, S. (2011, September). Does paid work provide a pathway to women’s
empowerment? Empirical findings from Bangladesh [Monograph]. Retrieved July 6, 2015, from
http://www.ids.ac.uk/about-us
K, B. P., Dilip, M., & Monica, P. T. (2010). Impact of Political Reservations in West Bengal Local
Governments on Anti-Poverty Targeting. Journal of Globalization and Development, 1(1), 1–38.
Klasen, S., & Pieters, J. (2012). Push or Pull? Drivers of Female Labor Force Participation during India’s
Economic Boom (IZA Discussion Paper No. 6395). Institute for the Study of Labor (IZA). Retrieved from
https://ideas.repec.org/p/iza/izadps/dp6395.html
Mayoux, L. (1995). Alternative vision or Utopian fantasy?: Cooperation, empowerment and women’s
cooperative development in India. Journal of International Development, 7(2), 211–228.
https://doi.org/10.1002/jid.3380070203
Neumark, D. (1988). Employers’ Discriminatory Behavior and the Estimation of Wage Discrimination.
The Journal of Human Resources, 23(3), 279–295. https://doi.org/10.2307/145830
New York Times. (2013, May 4). Worker Safety in Bangladesh and Beyond. The New York Times. Retrieved
from http://www.nytimes.com/2013/05/05/opinion/sunday/worker-safety-in-bangladesh-and-beyond.html
Oaxaca, R. (1973). Male-Female Wage Differentials in Urban Labor Markets. International Economic
Review, 14(3), 693–709. https://doi.org/10.2307/2525981
Oaxaca, R. L., & Ransom, M. R. (1994). On discrimination and the decomposition of wage differentials.
Journal of Econometrics, 61(1), 5–21. https://doi.org/10.1016/0304-4076(94)90074-4
Paul-Majumder, P., & Begum, A. (2000). The gender imbalances in the export oriented garment industry in
Bangladesh. Washington, D.C.: World Bank, Development Research Group/ Poverty Reduction and Economic
Management Network.
Rahman, S. (2010). Women’s Labour Contribution to Productivity and Efficiency in Agriculture: Empirical
Evidence From Bangladesh. Journal of Agricultural Economics, 61(2), 318–342. https://doi.org/10.1111/j.14779552.2010.00243.x

107

Ruhm, C. J. (2004). Parental employment and child cognitive development. Journal of Human Resources,
39(1), 155–192.
Sinha, S. (2005). Female work participation rates in rural West Bengal : a village level analysis.
War on Want. (2011). Stitched Up. Retrieved from http://www.waronwant.co.uk/overseaswork/sweatshops-and-plantations/hide/inform/17318-stitched-up
World Bank. (2008). Whispers to voices : gender and social transformation in Bangladesh (No. 43045) (pp. 1–
170).
The
World
Bank.
Retrieved
from
http://documents.worldbank.org/curated/en/2008/03/9163686/whispers-voices-gender-socialtransformation-bangladesh

108

Chapter 3
Communal Violence and Human Capital Accumulation in
India

1

Introduction
This paper aims to answer a straightforward albeit complex question- how does religious identity and

conflict determine long-term education outcomes for religious minorities? This pertinent question has
been asked and investigated by the scholars in different contexts over time. Religious identity has often
been used as a tool for both political and economic mobilization and exclusion. This exclusion manifests
itself in the form of wide inter-group inequality, ghettoization, ethnic enclaves and in extreme cases,
violent conflict. The causes and consequences of violent conflict, particularly civil war, mass violence and
ethnic conflict, has attracted considerable scholarly attention across disciplines. However, there is much
less research on the how small isolated and geographically dispersed incidences of violence can lead to
long term deficit in human capital accumulation for entire communities. This effect could be driven by the
immediate loss of life and livelihood following conflict. However repeated conflict also ingrains a culture
of fear in the minds of the victimized community that implicitly determines economic behavior in the
long-run. This paper deals with one such group, namely Indian Muslim.

I use data on education from the 64th (2007-08) and 71st (2014) round of the National Sample Survey
on Participation and Expenditure on Education in India to study the effect of both contemporaneous as
well as cumulative violence on both stock and flow education decisions of religious minorities. Data on
violence is derived from newspaper reports of Hindu-Muslim violence from 1950-2006, compiled by
various researchers using the Mumbai (Bombay) edition of the Times of India, a leading English language
daily in India (Iyer & Shrivastava, 2015; Mitra & Ray, 2014; Varshney & Wilkinson, 2006). Since the
education data I am able to use is relatively recent, in this paper I only consider reports of violence from

109

1986 to 2006. This period marked some of the deadliest incidents of communal violence in India, including
nationwide riots following the Babri Masjid Demolition in 1992 and the Gujarat riots of 2002, among
others. Some of the education outcomes I consider are, dropout rate, years of completed education, age at
dropout from school and probability of primary school completion. I find that communal violence has only
a small effect on educational enrolment and completion decisions of minorities and this effect is restricted
to girls.

2

Review of Literature
The persistent effect of violence on anthropometric outcomes has been widely studied in various

countries (Akresh, Caruso, & Thirumurthy, 2014; Akresh, Lucchetti, & Thirumurthy, 2012; Bundervoet,
Verwimp, & Akresh, 2009; Minoiu & Shemyakina, 2014). On the other hand, the scholarship on conflict
and schooling is relatively new and evolving (Baez, 2011; Chamarbagwala & MorĂĄn, 2011; LeĂłn, 2012;
Shemyakina, 2011). In the Indian context, most recent research has attempted to study the causes of
violence, ranging from ethnic and radical leftist to separationist and insurgent violence (Bohlken &
Sergenti, 2010; Gomes, 2015; Jha, 2014; Mitra & Ray, 2014; Sarsons, 2015; Sharma, 2015; Wilkinson,
2004), while there is very little evidence on the consequences of such violence, especially for children from
minority communities who suffer disproportionately due to such conflict. This paper attempts to bridge
this gap in literature by studying the effect of religious violence in India on schooling of Muslim children.

Most of the previous research on education and conflict has focused on large scale violence such as
civil war and ethnic conflict. Mass violence of this kind is marked by long periods of intense conflict
followed (and/or preceded) by periods of relative peace. This form of conflict allows researchers to more
accurately identify both cohorts affected, and not affected, by violence. This identification strategy has
been used by several researchers to identify the causal impact of violence on education. Chamarbagwala
& MorĂĄn (2011) study the impact of exposure to the Guatemalan civil war on educational attainment and
find large negative impact for Mayan men and women, who represent the poorest and most vulnerable

110

section of the Guatemalan population. Verwimp & Van Bavel (2013) follow a similar estimation strategy
and find that the negative effect of violence on schooling, measured as probability of primary school
completion, is exacerbated for boys relative to girls. Shemyakina (2011) on the other hand finds in her
study of violent conflict in Tajikistan, that girls suffer disproportionately compared to boys on measures
of educational attainment. One of the few such studies based on the Indian context (Singh & Shemyakina,
2016) looks at the effect of the Punjab insurgency on the education of girls and finds that girls exposed to
violence exhibit lower levels of educational attainment and this effect is likely driven by reduction in
education expenditure following conflict.

The nature of communal violence in India is distinct from such kinds of mass violence because it is
sporadic in nature, spread over time and is marked by short bursts of violence followed by extended
periods of peace. There is a growing body of economic research on this kind of sporadic localized violence
and its effect on human capital outcomes. Brown and Velasquez(2017) study the effect of gang-related
violence in Mexico on human capital outcomes of Mexican children using panel data and find that
children, particularly males, exposed to violence have lower levels of education and are more likely to be
engaged in employment. Duque (2013) similarly studies the effect of early life exposure to neighborhood
crime and violence in Colombia and finds that an increase in exposure to violence in-utero or in early life,
reduces school enrolment and increases the probability of being behind grade. Several such papers have
looked at the effect of exposure to gang violence and crime on education and health outcomes (Gershenson
& Tekin, 2017; Jarillo, Magaloni, Franco, & Robles, 2016; Torche & Villarreal, 2014).

This paper aims to add to this growing body of literature by looking at if and how small but consistent
instances of violence against a community can lead to permanent shortfalls in human capital accumulation
through withdrawal from education and lower student achievement. While the material and physical loss
following, religious violence is significant, it is also important to study the long-term impacts violence
has on a community. If violence leads to reduction in human capital accumulation by Muslim children,

111

who are often poor, it results in creating a generation of low skilled adults who are then trapped in a
vicious cycle of poverty. Violence itself, through systematic targeting of Muslim businesses and
institutions, could create conditions for long-term economic deprivation, which could negatively impact
the resources households can allocate toward education. Moreover, if these institutions and businesses are
the primary employers of Muslim youth, their decline and destruction could also disincentivize education.
I am also interested in studying how violence impacts Muslim boys and girls differently. While on the
one hand boys are at risk for being co-opted into the violence as participants, girls face the risk of being
victims of physical and sexual violence. Girls’ education is also often at the margins of the household
resource allocation decision, and any decline in household resources is likely to negatively impact their
education and enrolment. Through this paper, I try to shed light on the effect of short-term sporadic
violence on long term schooling outcomes, with the hope of informing policy decisions on controlling
violence and in directing rehabilitation efforts following periods of violence.

3

Data and Descriptive Statistics
I use data on religious violence from the Wilkinson-Varshney (2006) dataset on religious violence in

India from 1950-1995 and subsequently updated by Mitra and Ray (2014) till 2000 and by Iyer and
Srivastava (2015) till 20061. The original dataset, and the subsequent updates, were compiled using
newspaper reports from the Bombay (now Mumbai) edition of the Times of India. Each daily edition of
the newspaper was scanned for any report of violent conflict or confrontation involving Hindus and
Muslims2. I have compiled the violence dataset from three different sources to create a panel of district
level data covering five decades. I have also cleaned and coded the violence data to fill missing information
by matching the data with administrative and census records. To the best of my knowledge, my paper is

1 I use Mitra and Ray (2014) for data from 1996-2000 & Iyer and Srivastava (2015) for data from 2001-2006.
The latter have also collected data for the 1996-2000 period but I do not utilize their data for that period.
2 The Iyer and Srivastava (2015) dataset also includes instances of conflict between Hindus and other nonMuslim religious groups as well as between Muslims and non-Hindu religious groups. I do not include such
instances in my sample.

112

also the first to combine this violence dataset with large scale survey data to study the causal impacts of
communal violence in a large and democratic country such as India.

I use data on the 15 major states in India by population based on the 1991 census, these states are
Uttar Pradesh (including Uttarakhand), Bihar (including Jharkhand), Maharashtra, West Bengal, Madhya
Pradesh (including Chhattisgarh), Tamil Nadu, Rajasthan, Karnataka, Gujarat, Andhra Pradesh, Odisha,
Kerala, Assam, Haryana and Punjab. Together these states comprise close to 95% of India’s total
population. I exclude the state of Jammu and Kashmir from my analysis both because its low population
as well as the history and nature of conflict in the state which is markedly different compared with the
rest of the country. As the figure below shows, there is considerable heterogeneity across years in the
number of casualties, with some years with high number of casualties and some with very few. The spike
in violence following the Babri Masjid demolition in 1992 and the subsequent countrywide riots, as well
as the mass violence in Gujarat in 2002, is clearly reflected in the number of casualties in these two years.
There is also considerable heterogeneity across states with some states such as Gujarat and Maharashtra
reporting consistently high number of incidents and casualties, and others such as Punjab and Tamil Nadu
reporting very few.

Data on schooling outcomes comes from two consecutive rounds of the National Sample Survey (NSS)
Participation in Education (Schedule 25.2) covering the period between 1983 and 2014. The Participation
in Education survey includes detailed information on schooling choices and decisions. In particular, this
dataset contains self-reported information on age of entry and exit from the education system. I use this
information to create a novel panel dataset for each respondent using retrospective information on school
entry and exit and this allows me to exploit the time-series nature of schooling decisions to capture more
dynamic responses to incidences of violence.

113

Each round of the Participation in Education survey covers over 360,000 respondents and is
representative at the region level3. For preliminary analysis, I used four rounds of the survey conducted
in 1986-87 (42nd Round), 1995-96 (52nd Round), 2007-08 (64th Round) and 2014 (71st Round), however
the results reported in this chapter are restricted to the 64th and 71st round due to limited information on
background characteristics reported in the 42nd and 52nd round4. Combining the two round, I get a sample
of roughly 284,300 respondents of which roughly 15% are Muslim.

The average age of the respondents is 16.6 years and is equally divided between males and females.
Muslims are much more likely to report living in urban areas, which is consistent with the overall
distribution of Muslim population in India. Muslims also have slightly larger household size and are likely
to drop out of school at earlier ages than non-Muslims. The average age of starting school is about 5.5
years and more than 93% of the sample reports their school entry age as 6 years or less.

3.1

Creating the retrospective panel

I plan to conduct my analysis using both individual cross-section data as well as a constructed
retrospective panel. As mentioned previously, the NSS survey on education provides information on the
age at entry into school and if the respondent is currently not in school, then she is also asked about her
age at exit from school along with the grade and level of education completed before exit. This information
is collected from all individuals between the ages of 5 and 29. With this information I can create a
retrospective panel of enrolment for each individual from the time of entry in school to the time of exit5.

3 Over time, the number of districts in India has changed due to splitting of districts into two or more parts. A
majority of the districts were clean splits so I can trace the new districts back to the old parent district. For both the
violence and schooling data, I define the districts as they were in 1980 which is the time of the first school-choice
observation in my sample.
4 An alternative dataset that could have been used is the more frequently conducted NSS Employment and
Unemployment survey, which is larger in scope and covers close to 600,000 respondents. It has more detailed
information on post-education employment outcomes but also contains some information on education decisions at
terminal education levels such as primary, middle, secondary etc. However, since I want to capture the temporal
education decisions of minorities, I require information on the schooling decisions over time which this dataset does
not provide. Therefore, in this chapter, I report the results using only the Education survey of the NSS.
5 As an example, if in the 2014 survey a 15-year-old reports age at entry in school as 5 and age at exit as 11,
then she is considered to have entered school in 2004 and exited in 2010. In the years between 2004 and 2010 I
consider the child as being enrolled in school, and in the years following the child is considered dropped out.

114

This gives me multiple year-age combinations for each individual allowing me to create a panel of
schooling history from cross-section data. The data is right censored because I do not observe behavior
after the age of 29 or the age at survey, and I also do not observe the future behavior of individuals who
are interviewed while they are still in school or of school going age. For example, if the respondent is 10
years old when she is interviewed in 1995, then she is considered in the study only till 1995. I make no
ex-ante assumptions about future schooling choices. Since I am interested in schooling outcomes I
consider the outcomes for individuals till they attain the age of 18 or graduate from high-school,
whichever occurs sooner. This means that individuals enter the sample when they reach the age of 5 or 6,
depending on the school entering age for the state, and exit when they reach age 18 or exit school.
Expanding the data in this way gives me 2.1 million person-year observations, covering 402 districts in
15 states over a 21-year period6.

In an alternative specification, I model the dropout decision in a discrete time hazard model
framework. As in the main specification, I still use the retrospective panel but in this case a person enters
my sample when they turn 6 and exit when they drop out or turn 18, whichever is sooner. Unlike the
main specification, I have no observations for individuals beyond their year of dropout. As an example, if
an individual drops out at age 15, there are no entries for that individual at age 16-18, whereas in the
retrospective panel I would code the individual as “0” for being (un)enrolled at ages 16-19. Along with

Similarly, a 15-year-old child in the 2014 survey who never enrolled in school is counted as being unenrolled for the
entire duration.
6 This individual panel information can also be used for regional level analysis. I can collapse the data for each
individual-region-year combination to obtain the enrolment and/or dropout rate for each region by year. I can do
this for the entire sample combined and separated by religion and gender. Similarly, from the violence dataset I can
obtain region-year incidences of communal violence in terms of incidences of violence and number of casualties. It
would be useful to be able to do this kind of analysis at the intensive margin using information on duration of
violence, however since the violence information is extracted from newspaper reports, the data on duration of
violence is not very precise. In a previous draft of the paper I restrict my analysis to the extensive margin for region
level analysis. Dropout or enrolment rate in each region-year is used as the main dependent variable and lagged
incidence of violence or casualties as the main independent variable. In addition, I control for other region specific
characteristics based on the income, education, and employment and population profile of the region. Some income
and education data at the region level is available from the Census Bureau of India and from the Planning
Commission of India, the rest is interpolated from NSS data on Employment and Unemployment (Schedule 10). I
find no significant effect of violence experienced in the district on the aggregate enrolment or dropout rates for the
district, either overall or for Muslims separately. Results available on request.

115

district and time fixed effects, I also control for baseline hazard which is modeled as series of dummies to
indicate the number of years an individual was enrolled before dropping out. As an example, a 6-year-old
has a baseline 1, a 7-year-old is 2, an18-year-old is 13, etc. Since the maximum is age 18, this value is
between 1-13. This allows for a flexible baseline hazard and the vector of coefficients on these indicators,
indicates the hazard of dropping out in every period since being enrolled, where period is measured as
years. I also include LPM results for comparison. Failure is coded as going from “1” (enrolled) to “0”
(dropped out).

4

Methodology
The relationship between schooling and conflict can manifest itself both at the individual level, as well

as at the regional level. At the individual level the effect may be driven by concerns for safety (particularly
for girls), fear of recruitment into violence, destruction of local business which reduces incentives for
education, income loss and targeting of religious educational institutions. This could result in high
dropout rates, low completed years of schooling, schooling lags (due to gaps in education) and low test
scores. The effect of violence on schooling outcomes could also be seen at the aggregated district/regional
level. In particular, this is seen in the higher average dropout rates of Muslim versus Hindu children.
Dropout rates are an important outcome for policy makers in India because as the country approaches
100% enrolment for primary school-age children, the next challenge is in ensuring that children stay in
school to complete their education. For Muslim children, this challenge is exceptionally high, since for a
long time Muslim children, especially boys, have shown significantly higher rates of dropout across all
caste and religious groups. As the figure above shows, in 1983-84 both Muslim girls and boys started
with dropout rates that were comparable to or even better than that for SC/STs, the group recognized
by the government to be the most deprived in the country. Over time however, this gap narrowed and
now Muslim boys and girls have the highest dropout rate across all socio-religious groups in the country.
As of 2009-10, close to 22% of school age Muslim girls and 18% boys have either never enrolled in school

116

or have dropped out. That this is happening while the government is celebrating universal school
enrolment is indeed worrying.

I model the decision to stay in or drop out from school in each period as a binary decision that is based
on a host of household and individual covariates as well as experience of violence in the short and medium
term7. I model the probability that a respondent i, in district d is enrolled at time t as:
𝑃𝑟(Enrolledidt = 1 | ∙ ) = 𝑓(𝑐𝑜𝑛𝑓𝑙𝑖𝑐𝑡𝑑𝑡 , 𝑐𝑜𝑛𝑓𝑙𝑖𝑐𝑡𝑑𝑡−1 , 𝑐𝑜𝑛𝑓𝑙𝑖𝑐𝑡𝑑𝑡−2 , 𝑋𝑖 , 𝐷𝑑 , 𝛿𝑡 )
Where 𝑐𝑜𝑛𝑓𝑙𝑖𝑐𝑡𝑑𝑡−1 is a measure of incidents or casualties in district d at time t-1, 𝑋𝑖 is a vector of
time-invariant individual and household characteristics, and 𝐷𝑑 and 𝛿𝑡 are district and time fixed effects.

In addition to the analysis using panel and temporal information, I also conduct the analysis at the
individual level using the original cross-section dataset. However, unlike the retrospective panel which
measures the effect of contemporaneous violence, in case of the cross-section analysis I attempt to measure
the impact of cumulative experience of violence on aggregate schooling decisions. In this case my outcome
of interest is years of completed schooling, age at dropout and an indicator for whether the respondent
has completed primary school. The first two outcomes are tested on individuals who have completed the
school going age of 18 by 2006, and the results for primary school completion are tested on individuals
who have completed the age of 10 by 2006, which is the age of primary school completion in the country.
In the former case, I measure exposure to violence as the cumulative sum of incidents or casualties in the
respondent’s district of residence till the age of 18, and similarly in the latter, violence is measured as
cumulative experience of violence till age 10.

7 While I model a linear relationship between violence and schooling decision, we can also hypothesize that
dropout decisions follow an inverted U-shape that is, in districts that experience very little or a lot violence, the
effect on dropout rates is minimal. However, in the middle districts the effect on violence is likely to be larger because
violence comes as a large negative shock that is not accounted for by the household in their education demand
decision.

117

5

Results
I start with the results of the analysis using the retrospective panel. The sample is restricted to those

between ages of 6-18 in each year of the panel. I have used two measures of violence, the total incidents
of violence in the district in period t, t-1 and t-2 and similarly the total casualties in period t, t-1 and t-2
in the respondent’s district. I discuss here only the results using casualties as a measure of violence,
however the results using incidents of violence as a measure are included in the appendix tables8. The
outcome variable is binary9 - whether you are enrolled in a particular year or not. In addition, I control
for individual specific covariates that do not change over time such as gender and rural residence, type of
household, as well as birth cohort fixed effects in each period. All the specifications have district and year
fixed effects.

I estimate several specifications including interactions with indicators for being Muslim and female
to estimate heterogeneous effects of violence on the decision to be enrolled. The results of the linear
probability model are reported in Table 2. The probability of being enrolled is lower for older ages, rural
residents, women and Muslims. The results do not show any general impact of violence on the decision
to remain enrolled. While the coefficient of violence for Muslim women is negative, overall the effect of
violence on enrolment is positive, and close to zero, for all groups.

This results however, is quantitatively small, indicating that violence has to be significant enough for
it to have an impact on education decisions. This result is not entirely surprising given that anecdotal
evidence suggests dropout in India is a permanent and irreversible decision and in very rare cases is the
violence so extreme to merit a response of this kind. It is thus necessary to either use a finer measure of

8 I expected that casualties better capture violence since they are reported more accurately than incidents. As an
example, a news report that states “700 people were killed or injured in sporadic incidents of violence in Mumbai
over the last week” will be recorded as 1 incidence of violence in the dataset even though there were multiple
incidents across the district. However, the number of deaths or injuries is likely to be closer to the actual figure.
9 The binary enrolment variable is multiplied by 100 for easy interpretation. For example, the coefficient on
Muslim in Col 3. Table 1 says that Muslims are 9.72 percentage points less likely to be enrolled in each year
compared to non-Muslims.

118

education achievement that can capture the small effects of sporadic violence, or explore the possible
cumulative effect of violence on various schooling and education decisions. I discuss some of these
measures next section as well as in the final section.

Table 3 uses lifetime experience of violence as the variable of interest to measure its impact on
schooling decisions. As with the previous specification, I find that the coefficient on cumulative violence
for Muslim girls is negative for total years of completed education, and probability of having completed
primary school. Overall however, the effect of cumulative violence on all three indicators of education is
positive and close to zero for all groups. This indicates that it would take a particularly violent incidence
of rioting or for someone to have experienced significant violence during their childhood and youth for it
to have a significant effect on their schooling decision10. As descriptive statistics show, such communal
violence is rare- the average individual experiences less than 90 casualties from birth to age 18 in the form
of communal violence and less than two incidences of violence in the same period. Moreover, the results
indicate that for girls in general, experience of violence has a small positive effect on education outcomes,
which offsets any negative coefficient for Muslim females.

6

Extensions
I found in the previous sections that communal violence in India does not have any discernible effect

on the education outcomes of Muslim children, barring a small statistically significant but economically
insignificant effect for Muslim girls. In this section I test some other specifications as well as restrict my
analysis to other samples to see how this affects the results.

10 As an example, doubling the mean (overall) level of violence is associated with a 0.08 increase in years of
education for non-Muslim girls and a 0.04 increase in years of education for all other groups. Similarly, doubling
the mean total casualties is associated with a 0.15% increase in probability of completing primary school for all
groups except Non-Muslim men, for whom the probability is nearly zero.

119

6.1

Testing robustness of results with leads and lags

One the ways to test the robustness of the effect of lags of violence on contemporaneous education
decisions is to also test the effect of leads or future incidences of violence. If the results are indeed robust,
I would see no effect of violence in the future on education decisions of the present. Table 3.4 shows the
effect of both contemporaneous, past and future violence on enrolment decision. As the table shows, like
in the main analysis I see a small but significant negative coefficient on past and contemporaneous violence
on the decision to remain enrolled for Muslim girls, however I also see that violence in period t+1 and
t+2 exhibits a negative coefficient on the enrolment decision, which does not support the hypothesis that
violence lowers enrolment (or increases dropout rates). I see the same pattern when measuring violence
in terms of incidents of violence.

6.2

Testing results for unmarried respondents

Given the restrictions of the data, I assume that the place, more specifically the district, where the
respondent is enumerated is the district where the respondent engaged in and completed schooling. This
is not an entirely unreasonable assumption since inter-district migration in India is less common than
migration within the same district. However, it can be argued that women who migrate after marriage
are more likely to migrate out of district than other groups. If this is indeed the case, including married
women in the sample, who may have migrated from another district where their schooling decisions were
made in childhood, would incorrectly count them as having experienced violence in the district of
enumeration (or marital residence). To account for this I restrict my retrospective and individual sample
to unmarried persons, to see if and how the result changes. The results are reported in Table 3.7 and 3.8.
I find that restricting the sample does not significantly alter the magnitude of effect observed in either of
the two analysis. In case of the retrospective panel, I still observe a small but statistically significant
negative coefficient of violence on the probability of being enrolled for Muslim girls. As before, when I
test the same with adding leads of violence, I find that violence in the future also shows a negative
coefficient on the enrolment decision of the present. Similarly, for the individual cross-section sample of
unmarried persons, I find no difference between the full and the unmarried sample. As before, the

120

coefficient for Muslim girls is significant but small, the magnitude for the unmarried sample is almost the
same as that for the overall sample and overall violence shows a close to zero but positive correlation with
education outcomes for all groups, including Muslim girls.

6.3

Testing the results with a hazard model

For an alternative specification, I model the dropout decision in a discrete time hazard model. Along
with district and time fixed effects, I also control for baseline hazard which is measured as the duration
the individual was in the sample before exiting the sample, where exit is defined as reaching the age of 18
or dropping out, whichever happens sooner. I use the logistic functional form and for comparison, I include
results from a linear probability model. Table 3.9 reports the result for the main specification using both
lags as well as an extended specification that uses both leads and lags. Like the other results, I find that
the coefficient of violence on the enrolment decision of Muslim girls is small, however this result is not
robust to inclusion of lead measures of violence.

7

Conclusion and Future Work
In this chapter I attempt to measure the relationship between communal violence in India and the

schooling decision of Muslim children. While there has been considerable research on the causes of
communal, caste-based and Maoist violence in India, there is much less research on the consequences and
effects of this violence. I started the analysis with a straightforward ordinary least square regression. An
OLS regression specification gives us an unbiased estimate of the relationship between
dropout/enrolment rates and violence. If I were to relax the assumption of perfectly reported incidences
of riots, and assume that only a fraction of actual incidences of violence are reported, then our OLS
estimates may be biased downward. Even so, the OLS estimates give us, at worst, a lower bound for the
coefficient estimate. However, it should be noted that a simple OLS estimate gives us the correlation
between riots and dropout rates, and the implied relationship is not necessarily causal. A solution for this
is to use an instrument variable approach. Several methods have been discussed in previous literature to

121

instrument for the occurrence of violence. For future work, a method that is similar to that used by
Wilkinson (2004) can be used where riots are modelled as being affected by the presence of political
competition. It is widely speculated that violence is often used as a tool by political parties to polarize the
electorate and create plurality for winning in a first-past-the-post system. In closely contested elections,
it becomes even more imperative for the dominant parties to maintain their dominance by encouraging
inter-ethnic violence and polarizing the electorate for electoral dividends. Closeness to an upcoming
election and winning margin in the previous election can be used as an instrument for violence in a district
in any given year. It is hypothesized that the closer a constituency is to an upcoming election and the
smaller the margin of victory was in the previous election, the more likely it is for the district to experience
rioting. Data on state assembly elections results from the Election Commission of India can be used to
estimate the winning and margin and the dates for past and upcoming elections.

In this chapter I considered both stock and flow measures of education as outcomes of interest. In the
first case I modelled the decision to dropout (or remain enrolled), as a function of contemporaneous
experience of violence, where violence was measured as the total casualties or total incidents of HinduMuslim riots. I found that violence experienced in the period just before, had a significant but very small
negative relationship on the decision to remain enrolled in school for Muslim girls, however this was
offset entirely by the positive correlation between education and violence in general and for girls in
particular. This finding is unusual since girls are often at the margins of the household resource allocation
decision and any incidence of violence that affects household resources through disruption of economic
activity is likely to result in reduction of expenditure on the schooling of girls. In addition, concerns for
physical well-being and threats of sexual violence against women and girls is also more likely to result in
girls being forced to stop or delay their education attainment. While I did find a statistically significant
negative coefficient of violence on schooling for Muslim girls, this result was not robust, in particular, to
inclusion of future violence as an additional measure. This was true across several specifications and
samples. Moreover, I did not observe any correlation between school and violence for any other sub-group

122

as well. There could be several factors driving this result. Firstly, dropout from school in India is a drastic
and often permanent decision hence an event would have to be calamitous in nature to force parents to
withdraw their children from school permanently. Secondly, compared to previous studies that examine
the effect of civil war and large-scale violence on schooling, the nature of communal violence in India is
sporadic and dispersed over time. Most of the incidents of violence result in no casualties and are shortlived. Moreover, these incidents are concentrated in certain states, districts and cities, with some states
and urban areas within states much more likely to experience violence than other states and rural areas.
The markedly different nature of Hindu-Muslim violence in India may be driving the lack of any
significant relationship between enrolment decisions and experience of violence.

To consider if the dose-responses to violence matters more than immediate exposure, I also modelled
certain stock measures of education such as years of education, completion of primary school and age of
school dropout, as a function of cumulative lifetime experience of violence and found that long term
exposure to violence also had only a small effect on schooling decisions. As in the main specification, these
negative coefficients were significant only for Muslim girls. The magnitude of the effect was economically
insignificant considering the mean exposure to violence was very small and offset by the positive
coefficient for females in general. It is thus worth considering if a finer measure of educational deprivation
might better capture the effect of violence on schooling. One such measure could be school performance.
While sporadic and small-scale violence may not cause reduction in the quantity of schooling acquired, it
could certainly affect the quality of learning, which could manifest itself in lower achievement on exams
and test-scores. For future research on the relationship between communal violence and education
decisions, we could consider using test score data to get a more nuanced measure of education achievement
from various data sources. The only publicly available data on test scores in India at the moment is the
India Human Development Survey which is a two period panel first conducted in 2004-05 and followed
up in 2011-12. As a part of the survey, children aged 8-11 years were tested on reading, math and writing
and scored on a progressive scale. Each of the two rounds tested roughly 12000 children. An advantage

123

of the IHDS is that the children are tested at home and not in school, thus allowing the surveyors to
interview children who may have dropped out of school or are out of school at the time of the survey.
Following the vast literature that studies the impact of early life exposure to violence on later life
outcomes, we may also be able to examine the impact of exposure to violence in early life on test score
performance in later life.

124

APPENDICES

125

APPENDIX A

MAIN TABLES AND FIGURES

126

Age
Rural
Married
Female
HH Size
Muslim
Age at School Start
Age of Dropout
Casual Labor HH
Self Employed HH
N

Table 3.1: Descriptive Statistics- Individual Sample
All
Muslim
Non-Muslim
Male
Female
Male
Female
16.588
16.048
16.341
16.527
16.805
(0.012)
(0.044)
(0.046)
(0.018)
(0.020)
0.627
0.519
0.526
0.644
0.646
(0.001)
(0.003)
(0.004)
(0.001)
(0.001)
0.504
0.518
0.492
0.521
0.485
(0.001)
(0.003)
(0.004)
(0.001)
(0.001)
0.472
0.000
1.000
0.000
1.000
(0.001)
(0.000)
(0.000)
(0.000)
(0.000)
5.821
6.491
6.614
5.575
5.823
(0.005)
(0.019)
(0.021)
(0.007)
(0.007)
0.150
1.000
1.000
0.000
0.000
(0.001)
(0.000)
(0.000)
(0.000)
(0.000)
5.531
5.650
5.675
5.494
5.523
(0.001)
(0.005)
(0.005)
(0.002)
(0.002)
12.709
11.983
11.416
13.298
12.422
(0.009)
(0.031)
(0.032)
(0.014)
(0.015)
0.273
0.258
0.266
0.272
0.278
(0.001)
(0.003)
(0.003)
(0.001)
(0.001)
0.511
0.548
0.534
0.508
0.504
(0.001)
(0.003)
(0.004)
(0.001)
(0.001)
284,190
22,303
20,311
127,714
113,862

Source: Schedule 25.2 NSS on Education in India 64th and 71st Round
Standard Errors in parenthesis

127

Table 3.2: Effect of Casualties on Enrolment 1986-2006
(1)
Enrolled
0.000269
(0.000711)

(2)
Enrolled
-0.000214
(0.000827)

(3)
Enrolled
-0.00308**
(0.00101)

Casualties at t-1

0.000834
(0.000625)

0.000803
(0.000831)

0.000251
(0.00111)

Casualties at t-2

0.000811
(0.000732)

0.000564
(0.000866)

-0.00106
(0.00105)

Casual Labor HH

-23.65***
(0.543)

-23.09***
(0.524)

-23.07***
(0.524)

Self Employed HH

-6.438***
(0.390)

-5.854***
(0.376)

-5.874***
(0.380)

Rural

-6.925***
(0.435)

-8.041***
(0.434)

-7.980***
(0.432)

Muslim

-11.54***
(0.618)

-12.72***
(0.596)

Muslim* Casualties at t

0.00257*
(0.00105)

0.00473***
(0.00115)

Muslim* Casualties at t-1

0.000219
(0.00122)

0.00101
(0.00192)

Muslim* Casualties at t-2

0.00112
(0.00115)

0.00404**
(0.00154)

Casualties at t

Female

-10.02***
(0.341)

Female* Casualties at t

0.00622***
(0.000681)

Muslim*Female

2.572***
(0.597)

Muslim*Female* Casualties
at t

-0.00529***
(0.00127)

Female* Casualties at t-1

0.00110
(0.000888)

Muslim*Female* Casualties
at t-1

-0.00253
(0.00203)

Female* Casualties at t-2

0.00344***
(0.000785)

128

Table 3.2 (cont’d)
Muslim*Female* Casualties
at t-2
Constant
N
Adjusted R2
Mean casualties at t
Mean casualties at t-1
Mean casualties at t-2

-0.00734**
(0.00260)
84.22***
(1.176)
1941021
0.252

85.54***
(1.172)
1941021
0.259

3.640
(0.033)
3.545
(0.033)
3.942
(0.035)

Mean Enrolment All
Mean
Enrolment
Muslim
Mean
Enrolment
Muslim Female

90.09***
(1.151)
1941021
0.270
69.343
(.033)
60.688
(.092)
56.421
(.134)

Enrolment is measured as a whether the individual is enrolled a particular period or not, multiplied by 100.
All specifications include district, year and birth cohort fixed effects. Linear probability model results reported.
Robust standard errors clustered at the district level in parentheses * p < 0.05, ** p < 0.01, *** p < 0.001

129

Table 3.3: Effect of Cumulative Violence (Casualties) on Individual Outcomes
(1)
(2)
(3)
Years of
Age at Dropout
Completed
Education
Primary (0/1)
Total Casualties
0.000428
0.000604*
0.00000215
(0.000218)
(0.000246)
(0.0000145)
Muslim

-2.018***
(0.102)

-2.226***
(0.111)

-0.119***
(0.00743)

-0.00000929
(0.0000851)

0.0000116
(0.0000783)

0.0000393***
(0.00000916)

-1.672***
(0.0619)

-2.002***
(0.0669)

-0.104***
(0.00394)

0.000479***
(0.0000463)

0.000508***
(0.0000540)

0.0000289***
(0.00000789)

0.497***
(0.106)

0.553***
(0.120)

0.0229**
(0.00749)

-0.000426***
(0.000128)

-0.000414**
(0.000127)

-0.0000354**
(0.0000127)

Rural

-1.720***
(0.0779)

-1.764***
(0.0837)

-0.0683***
(0.00488)

Casual Labor HH

-4.063***
(0.0859)

-4.542***
(0.0963)

-0.204***
(0.00702)

Self Employed HH

-1.291***
(0.0809)

-1.543***
(0.0936)

-0.0349***
(0.00379)

Constant
N
Adjusted R2

11.07***
(0.172)
75460
0.211

18.17***
(0.190)
75396
0.209

0.888***
(0.0135)
176396
0.105

Mean Total Casualties till
age 18
Mean Total Casualties till
age 10

87.285
(1.351)
42.473
(0.524)

Muslim * Total Casualties
Female
Female * Total Casualties
Muslim * Female
Muslim * Female * Total
Casualties

Years of education is calculated using highest reported completed schooling level and status of current
attendance. Age at drop-out is self-reported. For those who completed high-school age at dropout is 18.
Columns 1 and 2 restricted to those who completed age 18 by 2006, the last year of available violence data
and sum of casualties is the sum of casualties from birth to age 18 in the district of residence. Column 3
restricted to those older than age 10 by 2006, and sum of casualties is sum of casualties from birth to age 10
in district of residence. All specifications include district and birth cohort fixed effects. Ordinary least square
results reported. Robust standard errors clustered at the district level in parentheses * p < 0.05, ** p < 0.01,
*** p < 0.001

130

131

132

133

134

135

APPENDIX B

SUPPLEMENTARY TABLES

136

Incidents at t

Table 3.4: Effect of Incidents on Enrolment 1986-2006
(1)
(2)
(3)
Enrolled
Enrolled
Enrolled
-0.0847
-0.0874
-0.332**
(0.0944)
(0.107)
(0.124)

Incidents at t-1

-0.0187
(0.0475)

-0.0217
(0.0536)

-0.0754
(0.0466)

Incidents at t-2

-0.123
(0.0859)

-0.106
(0.107)

-0.202
(0.140)

Casual Labor HH

-23.65***
(0.543)

-23.09***
(0.524)

-23.07***
(0.524)

Self Employed HH

-6.438***
(0.389)

-5.851***
(0.376)

-5.871***
(0.380)

Rural

-6.925***
(0.435)

-8.041***
(0.434)

-7.981***
(0.432)

Muslim

-11.48***
(0.622)

-12.70***
(0.602)

Muslim* Incidents at t

-0.0101
(0.224)

0.0942
(0.248)

Muslim* Incidents at t-1

-0.0109
(0.110)

0.0829
(0.162)

Muslim* Incidents at t-2

-0.222
(0.218)

0.0336
(0.259)

Female

-10.10***
(0.342)

Female* Incidents at t

0.541**
(0.172)

Muslim*Female

2.667***
(0.607)

Muslim*Female*
Incidents at t

-0.246
(0.242)

Female* Incidents at t-1

0.119
(0.0980)

Muslim*Female*
Incidents at t-1

-0.203
(0.181)

Female* Incidents at t-2

0.224
(0.167)

137

Table 3.4 (cont’d)
Muslim*Female*
Incidents at t-2
Constant
N
Adjusted R2
Mean Incidents at t
Mean Incidents at t-1
Mean Incidents at t-2

-0.599*
(0.280)
84.30***
(1.187)
1941021
0.252

85.61***
(1.183)
1941021
0.259

90.19***
(1.161)
1941021
0.270

0.145
(0.001)
0.148
(0.001)
0.157
(0.001)

Mean Enrolment
All
Mean Enrolment
Muslim
Mean Enrolment
Muslim Female

69.343
(.033)
60.688
(.092)
56.421
(.134)

Enrolment is measured as a whether the individual is enrolled a particular period or not, multiplied by 100.
All specifications include district, year and birth cohort fixed effects. Linear probability model results reported.
Robust standard errors clustered at the district level in parentheses * p < 0.05, ** p < 0.01, *** p < 0.001

138

Table 3.5: Effect of Cumulative Violence (Incidents) on Individual Outcomes
(1)
(2)
(3)
Years of
Age at Dropout
Completed
Education
Primary (0/1)
Total Incidents
0.0193
0.0302*
0.00242
(0.0141)
(0.0150)
(0.00138)
Muslim

-2.013***
(0.107)

-2.221***
(0.114)

-0.104***
(0.00937)

Muslim * Total Incidents

-0.00221
(0.00844)

-0.00121
(0.00855)

0.00132
(0.00107)

Female

-1.679***
(0.0627)

-2.011***
(0.0678)

-0.117***
(0.00524)

Female * Incidents Total

0.0169**
(0.00554)

0.0183**
(0.00569)

0.000754
(0.000748)

Muslim * Female

0.492***
(0.109)

0.554***
(0.123)

0.0205
(0.0119)

Muslim * Female * Incidents
Total

-0.0114
(0.0103)

-0.0126
(0.0107)

-0.00194
(0.00255)

Rural

-1.721***
(0.0780)

-1.765***
(0.0837)

-0.0749***
(0.00552)

Casual Labor HH

-4.062***
(0.0859)

-4.541***
(0.0964)

-0.226***
(0.00764)

Self Employed HH

-1.290***
(0.0811)

-1.542***
(0.0939)

-0.0313***
(0.00454)

Constant
N
Adjusted R2

11.05***
(0.173)
75460
0.210

18.14***
(0.191)
75396
0.208

0.890***
(0.0150)
76635
0.184

Mean Total Incidents till age
18
Mean Total Incidents till age
10

2.915
(0.029)
1.472
(0.010)

Years of education is calculated using highest reported completed schooling level and status
of current attendance. Age at drop-out is self-reported. For those who completed high-school age
at dropout is 18. Columns 1 and 2 restricted to those who completed age 18 by 2006, the last year
of available violence data and sum of incidents is the sum of incidents from birth to age 18 in the
district of residence. Column 3 restricted to those older than age 10 by 2006, and sum of incidents
is sum of incidents from birth to age 10 in district of residence. All specifications include district
and birth cohort fixed effects. Ordinary least square results reported. Robust standard errors
clustered at the district level in parentheses * p < 0.05, ** p < 0.01, *** p < 0.001

139

Table 3.6: Effect of Past and Future Casualties on Enrolment 1986-2006
(1)
(2)
Enrolled
Enrolled
Casualties t
0.00362***
-0.00176
(0.000952)
(0.00129)
Casualties t-1

0.00252**
(0.000857)

-0.000911
(0.00103)

Casualties t-2

0.00245*
(0.00116)

-0.000969
(0.00120)

Muslim

-12.58***
(0.615)

Muslim*Casualties t

0.00273*
(0.00125)

Female

-10.79***
(0.368)

Female*Casualties t

0.00411***
(0.000674)

Muslim*Female

2.633***
(0.649)

Muslim*Female*Casualties t

-0.00383***
(0.00109)

Muslim* Casualties t-1

0.00108
(0.00138)

Female*Casualties t-1

0.00205**
(0.000625)

Muslim*Female*Casualties t-1

-0.00251
(0.00185)

Muslim*Casualties t-2

0.00362*
(0.00148)

Female*Casualties t-2

0.00301***
(0.000837)

Muslim*Female*Casualties t-2

-0.00688**
(0.00233)

Casualties t+1

0.00319**
(0.00101)

-0.00131
(0.00128)

Casualties t+2

0.00390***
(0.00102)

-0.00349***
(0.000882)

140

Table 3.6 (cont’d)
Muslim*Casualties t+1

0.00276***
(0.000813)

Female*Casualties t+1

0.00403*
(0.00195)

Muslim*Female*Casualties t+1

-0.00499**
(0.00153)

Muslim*Casualties t+2

0.00397*
(0.00156)

Female*Casualties t+2

0.00734***
(0.000928)

Muslim*Female*Casualties t+2

-0.00706***
(0.00188)

Casual Labor HH

-23.52***
(0.544)

Self Employed HH

-5.781***
(0.402)

Rural

-8.438***
(0.457)

Constant

68.46***
(0.0176)
1616207
0.000

N
Adjusted R2

90.73***
(1.157)
1616207
0.266

Enrolment is measured as a whether the individual is enrolled a particular period
or not, multiplied by 100. All specifications include district, year and birth cohort
fixed effects. Linear probability model results reported. Robust standard errors
clustered at the district level in parentheses * p < 0.05, ** p < 0.01, *** p < 0.001

141

Table 3.7: Effect of Casualties on Enrolment 1986-2006, Unmarried Sample
(1)
(2)
(3)
(4)
Enrolled
Enrolled
Only lags
Leads and Lags
Casualties t
0.00615*** -0.000692
0.00709*
0.000768
(0.00118)
(0.00225)
(0.00299)
(0.00247)
Casualties t-1

0.00103
(0.00269)

0.00175
(0.00111)

0.00557
(0.00392)

0.000648
(0.00107)

Casualties t-2

0.00216
(0.00435)

-0.000190
(0.00166)

0.00498
(0.00613)

0.0000556
(0.00153)

Muslim

-11.43***
(0.773)

-11.66***
(0.815)

Muslim*Casualties t

0.00387
(0.00230)

0.00185
(0.00232)

Female

-11.58***
(0.410)

-12.49***
(0.447)

0.00490***
(0.00147)

0.00287
(0.00148)

2.613**
(0.844)

2.727**
(0.928)

-0.00970***
(0.00181)

-0.00666***
(0.00151)

Muslim* Casualties t-1

0.00239
(0.00229)

0.00303
(0.00162)

Female*Casualties t-1

-0.000523
(0.000846)

0.000655
(0.000870)

Muslim*Female*Casualties t-1

-0.00516
(0.00281)

-0.00551*
(0.00225)

Muslim*Casualties t-2

0.00430*
(0.00209)

0.00356
(0.00209)

Female*Casualties t-2

0.00446**
(0.00151)

0.00437**
(0.00134)

Muslim*Female*Casualties t-2

-0.0101***
(0.00232)

-0.00945***
(0.00229)

Female*Casualties t
Muslim*Female
Muslim*Female*Casualties t

Casualties t+1

0.00468
(0.00291)

0.00174
(0.00236)

Casualties t+2

0.00839***
(0.00164)

0.0000509
(0.00181)

142

Table 3.7 (cont’d)
Muslim*Casualties t+1

0.000745
(0.00230)

Female*Casualties t+1

0.00155
(0.00251)

Muslim*Female*Casualties t+1

-0.00572*
(0.00286)

Muslim*Casualties t+2

0.00324
(0.00199)

Female*Casualties t+2

0.00408*
(0.00179)

Muslim*Female*Casualties t+2

-0.0117***
(0.00229)

Casual Labor HH

-23.86***
(0.573)

-24.95***
(0.612)

Self Employed HH

-5.399***
(0.449)

-5.738***
(0.509)

Rural

-8.904***
(0.479)

-9.425***
(0.526)

Constant
N
Adjusted R2

63.24***
(0.0292)
868106
0.000

90.17***
(1.246)
868106
0.354

60.19***
(0.0634)
749919
0.000

91.11***
(1.268)
749919
0.337

Enrolment is measured as a whether the individual is enrolled a particular period or not,
multiplied by 100. All specifications include district, year and birth cohort fixed effects.
Linear probability model results reported. Robust standard errors clustered at the district
level in parentheses * p < 0.05, ** p < 0.01, *** p < 0.001

143

Table 3.8: Effect of Cumulative Violence (Casualties), Unmarried Sample
(1)
(2)
(3)
Years of Education Age at Dropout Completed Primary (0/1)
Sum of Casualties
0.00110
0.00146**
0.0000183
(0.000592)
(0.000536)
(0.0000234)
Muslim

-2.056***
(0.145)

-2.252***
(0.159)

-0.103***
(0.00926)

-0.0000202
(0.000126)

0.0000475
(0.000131)

0.0000386**
(0.0000122)

-1.939***
(0.0705)

-2.315***
(0.0782)

-0.117***
(0.00516)

0.000359***
(0.0000877)

0.000358**
(0.000118)

0.0000247
(0.0000132)

0.675***
(0.138)

0.778***
(0.160)

0.0205
(0.0112)

-0.000463***
(0.000123)

-0.000467**
(0.000149)

-0.0000726*
(0.0000305)

Rural

-1.750***
(0.0944)

-1.760***
(0.100)

-0.0750***
(0.00552)

Casual Labor HH

-4.104***
(0.105)

-4.499***
(0.118)

-0.226***
(0.00764)

Self Employed HH

-1.355***
(0.109)

-1.566***
(0.122)

-0.0313***
(0.00455)

Constant

10.86***
(0.205)
42507
0.212

17.89***
(0.220)
42487
0.210

0.894***
(0.0148)
76635
0.184

Muslim * Casualties
Female
Female * Casualties
Muslim * Female
Muslim * Female * Casualties

N
Adjusted R2

Years of education is calculated using highest reported completed schooling level and status of current
attendance. Age at drop-out is self-reported. For those who completed high-school age at dropout is 18.
Columns 1 and 2 restricted to those who completed age 18 by 2006, the last year of available violence data
and sum of casualties is the sum of casualties from birth to age 18 in the district of residence. Column 3
restricted to those older than age 10 by 2006, and sum of casualties is sum of casualties from birth to age
10 in district of residence. All specifications include district and birth cohort fixed effects. Ordinary least
square results reported. Robust standard errors clustered at the district level in parentheses * p < 0.05, ** p
< 0.01, *** p < 0.001

144

Table 3.9: Effect of Casualties on Enrolment 1986-2006, Hazard Model
(1)
(2)
(3)
(4)
Logit
OLS
Logit
OLS
Only Lags
Lags and Leads
Casualties t

-0.000211*
(0.0000892)

-0.00400***
(0.000798)

-0.0000898
(0.0000994)

-0.00219*
(0.000917)

-0.909***
(0.0463)

-10.74***
(0.662)

-0.895***
(0.0475)

-10.92***
(0.687)

0.000148
(0.000119)

0.00550***
(0.00109)

0.0000140
(0.000145)

0.00305*
(0.00127)

-0.894***
(0.0219)

-10.91***
(0.395)

-0.932***
(0.0229)

-11.83***
(0.426)

0.000392*
(0.000156)

0.00744***
(0.00132)

0.000225*
(0.0000892)

0.00526***
(0.000727)

0.264***
(0.0393)

1.237
(0.685)

0.265***
(0.0419)

1.406
(0.735)

Muslim * Female * Casualties t

-0.000276
(0.000283)

-0.00322
(0.00299)

-0.000133
(0.000188)

-0.00247
(0.00247)

Casualties t-1

0.0000439
(0.0000654)

0.0000153
(0.000870)

-0.0000256
(0.0000531)

-0.00150*
(0.000634)

Muslim * Casualties t-1

-0.0000957
(0.000164)

0.000887
(0.00209)

-0.0000804
(0.000108)

0.00145
(0.00138)

Female * Casualties t-1

-0.00000690
(0.0000584)

0.00203**
(0.000696)

0.0000797
(0.0000656)

0.00336***
(0.000597)

Muslim * Female * Casualties t-1

-0.0000189
(0.000111)

-0.00182
(0.00120)

-0.0000347
(0.0000929)

-0.00162
(0.00124)

Casualties t-2

-0.0000653
(0.000127)

-0.00179
(0.000912)

-0.0000600
(0.000118)

-0.00165
(0.000940)

Muslim * Casualties t-2

0.000155
(0.000135)

0.00510***
(0.00132)

0.000138
(0.000142)

0.00465**
(0.00144)

Female * Casualties t-2

0.000129
(0.000119)

0.00471***
(0.000868)

0.000109
(0.0000980)

0.00439***
(0.000733)

Muslim * Female * Casualties t-2

-0.000292*
(0.000117)

-0.00523***
(0.00129)

-0.000280*
(0.000114)

-0.00539***
(0.00131)

Casualties t+1

-0.0000808
(0.0000785)

-0.00132
(0.000858)

Muslim * Casualties t+1

0.0000616
(0.0000661)

0.00273***
(0.000752)

Muslim
Muslim * Casualties t
Female
Female * Casualties t
Muslim * Female

145

Table 3.9 (cont’d)
Female * Casualties t+1

0.000174**
(0.0000615)

0.00409***
(0.000869)

Muslim * Female * Casualties t+1

-0.000212
(0.000153)

-0.00351
(0.00209)

Casualties t+2

-0.000118
(0.0000855)

-0.00374***
(0.000752)

Muslim * Casualties t+2

0.000134
(0.000252)

0.00590**
(0.00199)

Female * Casualties t+2

0.000262**
(0.0000966)

0.00708***
(0.000980)

Muslim * Female * Casualties t+2

-0.000450**
(0.000141)

-0.00638***
(0.00186)

Self Employed HH

-0.520***
(0.0347)

-3.993***
(0.351)

-0.514***
(0.0359)

-4.035***
(0.376)

Casual Labor HH

-1.658***
(0.0366)

-20.85***
(0.631)

-1.668***
(0.0374)

-21.74***
(0.652)

Rural

-0.621***
(0.0356)

-7.537***
(0.455)

-0.643***
(0.0365)

-8.129***
(0.485)

Constant

2.697***
(0.0843)
1709006

91.57***
(1.195)
1709006
0.240

2.711***
(0.0847)
1428903

92.05***
(1.204)
1428903
0.247

N
Adjusted R2

Enrolment is measured as a whether the individual is enrolled in a period or not. For logit
specification, enrolled is a binary variable, for the OLS specification, enrolled is multiplied by 100.
All specifications include district and year fixed effects as well as controls for baseline hazard or
duration. Robust standard errors clustered at the district level in parentheses * p < 0.05, ** p <
0.01, *** p < 0.001

146

REFERENCES

147

REFERENCES
Akresh, R., Caruso, G. D., & Thirumurthy, H. (2014). Medium-Term Health Impacts of Shocks
Experienced In Utero and After Birth: Evidence from Detailed Geographic Information on War Exposure
(Working Paper No. 20763). National Bureau of Economic Research. Retrieved from
http://www.nber.org/papers/w20763
Akresh, R., Lucchetti, L., & Thirumurthy, H. (2012). Wars and child health: Evidence from the Eritrean–
Ethiopian
conflict.
Journal
of
Development
Economics,
99(2),
330–340.
https://doi.org/10.1016/j.jdeveco.2012.04.001
Baez, J. E. (2011). Civil wars beyond their borders: The human capital and health consequences of hosting
refugees. Journal of Development Economics, 96(2), 391–408. https://doi.org/10.1016/j.jdeveco.2010.08.011
Bohlken, A. T., & Sergenti, E. J. (2010). Economic growth and ethnic violence: An empirical investigation
of Hindu-Muslim riots in India. Journal of Peace Research. https://doi.org/10.1177/0022343310373032
Brown, R., & VelĂĄsquez, A. (2017). The effect of violent crime on the human capital accumulation of young
adults. Journal of Development Economics, 127, 1–12. https://doi.org/10.1016/j.jdeveco.2017.02.004
Bundervoet, T., Verwimp, P., & Akresh, R. (2009). Health and Civil War in Rural Burundi. Journal of
Human Resources, 44(2), 536–563. https://doi.org/10.3368/jhr.44.2.536
Chamarbagwala, R., & MorĂĄn, H. E. (2011). The human capital consequences of civil war: Evidence from
Guatemala. Journal of Development Economics, 94(1), 41–61. https://doi.org/10.1016/j.jdeveco.2010.01.005
Duque, V. (2013). The Hidden Costs and Lasting Legacies of Violence on Education: Evidence from
Colombia.
In
Session
111:
Crime,
Health
and
Human
Capital.
Retrieved
from
http://paa2013.princeton.edu/abstracts/130601
Gershenson, S., & Tekin, E. (2017). The Effect of Community Traumatic Events on Student Achievement:
Evidence from the Beltway Sniper Attacks. Education Finance and Policy, 1–54.
https://doi.org/10.1162/EDFP_a_00234
Gomes, J. F. (2015). The Political Economy of the Maoist Conflict in India: An Empirical Analysis. World
Development, 68, 96–123. https://doi.org/10.1016/j.worlddev.2014.11.021
Iyer, S., & Shrivastava, A. (2015). Religious Riots and Electoral Politics in India (SSRN Scholarly Paper
No. ID 2696360). Rochester, NY: Social Science Research Network. Retrieved from
http://papers.ssrn.com/abstract=2696360
Jarillo, B., Magaloni, B., Franco, E., & Robles, G. (2016). How the Mexican drug war affects kids and
schools? Evidence on effects and mechanisms. International Journal of Educational Development, 51, 135–146.
https://doi.org/10.1016/j.ijedudev.2016.05.008
Jha, S. (2014). “Unfinished business”: Historic complementarities, political competition and ethnic violence
in Gujarat. Journal of Economic Behavior & Organization, 104, 18–36.
LeĂłn, G. (2012). Civil Conflict and Human Capital Accumulation The Long-term Effects of Political
Violence in Perú. Journal of Human Resources, 47(4), 991–1022. https://doi.org/10.3368/jhr.47.4.991
Minoiu, C., & Shemyakina, O. N. (2014). Armed conflict, household victimization, and child health in CĂ´te
d’Ivoire. Journal of Development Economics, 108, 237–255. https://doi.org/10.1016/j.jdeveco.2014.03.003
Mitra, A., & Ray, D. (2014). Implications of an Economic Theory of Conflict: Hindu-Muslim Violence in
India. Journal of Political Economy, 122(4), 719–765. https://doi.org/10.1086/676316
Sarsons, H. (2015). Rainfall and conflict: A cautionary tale. Journal of Development Economics, 115, 62–
72. https://doi.org/10.1016/j.jdeveco.2014.12.007
Sharma, S. (2015). Caste-based crimes and economic status: Evidence from India. Journal of Comparative
Economics, 43(1), 204–226. https://doi.org/10.1016/j.jce.2014.10.005

148

Shemyakina, O. (2011). The effect of armed conflict on accumulation of schooling: Results from Tajikistan.
Journal of Development Economics, 95(2), 186–200. https://doi.org/10.1016/j.jdeveco.2010.05.002
Singh, P., & Shemyakina, O. N. (2016). Gender-differential effects of terrorism on education: The case of
the
1981–1993 Punjab insurgency.
Economics
of
Education
Review, 54,
185–210.
https://doi.org/10.1016/j.econedurev.2016.02.003
Torche, F., & Villarreal, A. (2014). Prenatal exposure to violence and birth weight in Mexico: Selectivity,
exposure,
and
behavioral
responses.
American
Sociological
Review,
79(5),
966–992.
https://doi.org/10.1177/0003122414544733
Varshney, A., & Wilkinson, S. (2006). Varshney-Wilkinson Dataset on Hindu-Muslim Violence in India,
1950-1995,
Version
2:
Version
1.
Retrieved
from
http://www.icpsr.umich.edu/ICPSR/studies/04342/version/1
Verwimp, P., & Van Bavel, J. (2013). Schooling, violent conflict, and gender in Burundi. The World Bank
Economic Review, lht010.
Wilkinson, S. (2004). Votes and violence electoral competition and ethnic riots in India. Cambridge; New
York: Cambridge University Press. Retrieved from http://site.ebrary.com/id/10131638

149