Traditional Institutions Meet the Modern World: Caste, Gender,

and Schooling Choice in a Globalizing Economy

By KAIVAN MUNSHI AND MARK ROSENZWEIG*

This paper addresses the question of how traditional institutions interact with the
forces of globalization to shape the economic mobility and welfare of particular
groups of individuals in the new economy. We explore the role of one such
traditional institutionthe caste systemin shaping career choices by gender in
Bombay using new survey data on school enrollment and income over the past 20
years. We find that male working-classlower-castenetworks continue to channel
boys into local language schools that lead to the traditional occupation, despite the
fact that returns to nontraditional white-collar occupations rose substantially in the
1990s, suggesting the possibility of a dynamic inefficiency. In contrast, lower-caste
girls, who historically had low labor market participation rates and so did not benefit
from the network, are taking full advantage of the opportunities that became available
in the new economy by switching rapidly to English schools. (JEL I21, J16, O15, Z13)

The collapse of the former Soviet Union, groups of individuals to the new economy.
followed by the economic and financial liber- Traditional institutions, such as community
alization of the 1990s, has restructured and networks, are generally believed to play an
globalized many economies throughout the important role in low-income countries by
world. One consequence of this restructuring, facilitating economic activity when markets
which has been widely observed, is that some function imperfectly. Less well understood is
groups have taken advantage of the new ben- how traditional institutions affect the trans-
efits afforded by globalization, while others formation of economies undergoing change,
appear to have been left behind. This paper affecting in turn the distribution of benefits
addresses the question of whether and how from macroeconomic structural reform.
old institutions clash with the forces of glob- We explore the role of one such traditional
alization in shaping the response of particular institutionthe caste systemin shaping career
choices by gender in a dynamic urban context,
using new data on schooling choices and income
* Munshi: Department of Economics, Brown University,
Box B/64, Waterman St., Providence, RI 02912 (e-mail: covering the past 20 years in Bombay city, the
[email protected]); Rosenzweig: Department of Economics, industrial and financial center of the Indian econ-
Yale University, 27 Hillhouse Ave., New Haven, CT 06520 omy. Bombay is a useful and important setting in
(e-mail: [email protected]). Leena Abraham and which to study the role of institutional rigidities in
Padma Velaskar of the Tata Institute of Social Sciences
collaborated in the design of the survey instrument, carried
a dynamic context, as the Bombay labor market
out the survey, and provided many valuable insights. We are was historically organized along caste lines, with
also very grateful to Suma Chitnis for her support and individual subcastes or jatis controlling particular
advice at every stage of the project. We received helpful occupational niches over the course of many gen-
comments from Abhijit Banerjee, David Card, Jan Eeck- erations.1 A particularly important feature of these
hout, Lakshmi Iyer, Duncan Thomas, three anonymous ref-
erees, and seminar participants at Columbia University, caste networks is that they were most active in
Harvard University, the Massachusetts Institute of Technol-
ogy, IZA, University of Pennsylvania, Princeton University,
1
UCLA, the University of Southern California, Washington Rajnarayan Chandavarkar (1994 pp. 122, 223), for
University, and the World Bank. Research support from the instance, describes how [caste] clusters formed within par-
Mellon Foundation at the University of Pennsylvania and ticular trades and occupations ... [this] occupational distri-
the National Science Foundation (grant SES-0431827) is bution reflected neither [traditional rural] caste vocation nor
gratefully acknowledged. We are responsible for any errors the inheritance of special skills. It was produced partly by
that may remain. exclusionary practices by which social groups, once they
1225
1226 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2006

working-class occupations dominated by lower by extension the returns to English education,

caste men. Women historically did not participate then (future) occupational mobility can be iden-
in Bombays labor market and so did not benefit tified from changes in the choice of the lan-
from the caste networks, but both men and women guage of instruction made by parents of school-
scrupulously adhered to the social rule of endog- age children. Examination of the changing
amous marriage within the jati. patterns of schooling choice by jati and gender
Although Bombay was a predominantly in- thus permits an assessment of the interactions
dustrial city for a hundred years beginning in between traditional institutions and the new re-
the last quarter of the nineteenth century, in alities of globalization.
the early 1990s the liberalization of the Indian Our empirical analysis is based on a survey
economy saw a shift in the citys economy of 4,900 households belonging to the Maha-
toward the corporate and financial sectors. rashtrian community residing in Bombays
We study how members of different jatis, by Dadar area and a survey of the schools in that
gender, responded to these changes in the locale that we conducted in 20012002. Sec-
returns to different occupations, and we will ondary schools in Bombay run from grade 1
show that the historical pattern of networking to grade 10. The household survey was based
within the jati continues to shape gender- on a stratified random sample of students who
specific, individual responses to these new entered 28 of the 29 schools in Dadar (in the
opportunities in ways that will importantly first grade), over a 20-year period, 1982
affect the future distributions of incomes, in- 2001.3 English is the language of instruction
dependent of pre-schooling human capital ef- in ten schools in Dadar, while Marathi is the
fects or liquidity constraints.2 language of instruction in the remaining 18
Our strategy in this paper is to assess how schools.
schooling choice, measured by the language The survey data suggest that the returns to
of instruction, varied across jatis, across boys English education, for given years of school-
and girls within jatis, and over time. We focus ing, increased in the 1990s. Based on retro-
on schooling choice because most adults were spective information on the annual earnings
already locked into their occupations when of the parents of the sampled children, we
the unexpected economic changes occurred. estimated the returns to English and the re-
Schooling choice is an important determinant of turns to years of schooling at five points in
future occupational outcomes in the Bombay time from 1980 through 2000 for working
economy and thus reflects the contemporaneous adults between the age of 30 and 55.4 Figures 1
perceptions of expected occupational returns. and 2 provide the estimated returns to schooling
University education in Bombay is entirely in attainment and schooling language, for men
English, but children choose between English and women, respectively, in each time period.
and Marathi (the local language) as the lan- As can be seen, the returns to years of school-
guage of instruction at the time they enter ing increased only mildly over time for both
school. Schooling in Marathi channels the child men and women. In contrast, the English pre-
into working-class jobs, while more expensive mium increased sharply from the 1980s to the
English education significantly increases the 1990s for both sexes, rising from 15 percent
likelihood of obtaining a coveted white-collar in 1980 to 24 percent in 2000 for men and
job. If the economic liberalization of the 1990s from approximately 0 percent in 1980 to 27
effectively increased white-collar incomes, and percent in 2000 for women. The returns to
English for men increase from the mid-1980s,
which is most likely due to the decline around
obtained a foothold in a particular occupation, would not
admit an outsider.
2 3
A recent literature has shown that historical institutions One school refused to provide us with information on its
have long-run consequences for growth in low-income students and will be ignored in all the discussion that follows.
4
countries (Daron Acemoglu et al., 2001; Abhijit Banerjee The details of the estimation procedure and the esti-
and Lakshmi Iyer, 2005). These empirical findings, how- mates of the returns to English and the returns to schooling
ever, do not provide insight into the mechanisms underlying (with standard errors) are provided in Munshi and Rosenz-
such persistence. weig (2003).
VOL. 96 NO. 4 MUNSHI AND ROSENZWEIG: TRADITIONAL INSTITUTIONS MEET THE MODERN WORLD 1227

FIGURE 1. RETURNS TO ENGLISH AND SCHOOLING BY YEAR, 1980 2000: MEN AGE 30 55

that time in manufacturing jobs in Bombay athi-medium schools suggest that the changes in
(Darryl DMonte, 2002), but continue to rise the returns to English significantly affected
through the 1990s. schooling choice for both boys and girls in the
The survey collected information on school- sample, across castes and over time. Figure 3
ing choice for 20 cohorts of students who en- and Figure 4 display the changing proportions
tered the 28 neighborhood schools (in the first of students enrolled in English schools for the
grade) over the 19822001 period. The times- 20 entering cohorts from 1982 (cohort ! 1) to
series data on enrollments in English- and Mar- 2001 (cohort ! 20) for three caste group-

FIGURE 2. RETURNS TO ENGLISH AND SCHOOLING BY YEAR, 1980 2000: WOMEN AGE 30 55
1228 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2006

FIGURE 3. ENGLISH SCHOOLING: NET PARENTAL EDUCATION EFFECTBOYS

ingslow, medium, and highand by gen- evant boundary for the labor-market networks
der.5 The figures were constructed using the and form the relevant social unit in our analy-
Epanechnikov kernel function to nonparametri- sis,7 we aggregate the 59 subcastes in our data
cally regress schooling choice (1 ! English for expositional convenience in these figures.
medium; 0 ! Marathi medium) on the cohort Figures 3 and 4 show that enrollment rates in
variable for each caste group, taking into ac- English-medium schools have grown substantially
count the strong intergenerational state-depen- over time for both boys and girls and for all
dence with respect to the language of instruction castes.8 The trajectory is much steeper, how-
within the family.6 Although jatis define the rel- ever, for the ten most recent cohorts, who would
have entered school in the post-reform 1990s.
Thus, the increase in the returns to English
5 observed in Figure 1 and Figure 2 appears to
Children enter first grade at the age of 6 and complete
tenth grade at the age of 15, so the current age of the have shifted schooling choice toward English
students in our sample, with only a few exceptions, ranges education. The figures also indicate substantial
from 6 to 25. Students in Bombay typically do not change differences in English schooling between castes
the language of instruction midstream or switch schools at the beginning of the sample period, reflecting in
after they enter first grade. High castes include all the
Brahmin jatis, as well as a few other elite jatis (CKP and part the circumstances of the colonial regime. The
Pathare Prabhus). Low castes include Scheduled Castes, high castes gained access to clerical and adminis-
Scheduled Tribes, and Other Backward Castes, as defined
by the government of India. Medium castes are drawn
mostly from the cultivator jatis, such as the Marathas and
7
the Kunbis, as well as other traditional vocations that were As Morris David Morris (1965 p. 76) emphasizes in his
not considered to be ritually impure. historical account of the Bombay labor market, for any
6
If both parents have been schooled in English, it is very analysis of labor recruitment [in Bombay] ... it is entirely
unlikely that the child would be sent to a Marathi school, inappropriate to lump into larger groups because of simi-
and all the regressions that we later report will also account larity of name, function, social status, or region-of-origin
for such state dependence at the level of the family. Details subcastes that are not endogamous.
8
of the nonparametric estimation procedure and parametric Details of the nonparametric estimation procedure used
estimates of the schooling regression (with standard errors) to generate these figures are provided in Munshi and Rosen-
are provided in Munshi and Rosenzweig (2003). zweig (2003).
VOL. 96 NO. 4 MUNSHI AND ROSENZWEIG: TRADITIONAL INSTITUTIONS MEET THE MODERN WORLD 1229

FIGURE 4. ENGLISH SCHOOLING: NET PARENTAL EDUCATION EFFECTGIRLS

trative jobs under the British, while the lower acute in those jobs.9 These studies focus on
castes were confined for the most part to working- men, the primary occupants of blue-collar
class jobs. Consistent with the view that Marathi jobs. And among the household heads in Dadar,
education channels students into working-class 68 percent of the men in working-class jobs
jobs, and that English education increases the like- found employment through a relative or a member
lihood of obtaining a white-collar job, we see in of the community, while the corresponding sta-
Figure 3 and Figure 4 that high-caste boys and tistic for white-collar workers was 44 percent.
girls currently 25 years old (the oldest cohort) Once the (working-class) network is in place,
were much more likely to have been schooled in there is a positive externality associated with
English, and that this caste difference in schooling participation in the network, and hence with
persists over the next ten cohorts. But although the the traditional occupational choice in the jati.
caste gap narrows dramatically for the girls in the This externality could give rise to intergenera-
1990s, there is no convergence for the boys. Thus, tional occupational persistence at the level of
it appears that caste continues to play a role in the jati, with labor market networks channeling
shaping schooling choices in the new economy of boys into particular (traditionally male) occupa-
the 1990s, but only for boys. The key question is tions and hence toward particular schooling
why the lower-caste boys seemingly fail to take choices.
advantage of the new economic opportunities.
The gender-specific explanation for the ob-
served pattern in Figure 3 that we pursue in 9
For example, Albert Rees (1966) found that informal
this paper is based on network externalities. sources accounted for 80 percent of all hires in eight blue-
Numerous studies document higher levels of collar occupations versus 50 percent of all hires in four white-
collar occupations in an early study set in Chicago. Similarly,
networking in blue-collar occupations, possi- 68 percent of blue-collar workers and 38 percent of white-
bly because the information and enforcement collar workers reported having received help finding a job in
problems that give rise to networks are more M. S. Gores (1970) study of migrants in Bombay.
1230 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2006

Once the returns to the white-collar occupation ous historians studying Bombays economy prior
grow, however, schooling choice must ultimately to independence in 1947 (Chandavarkar, 1994;
converge across castes. The explanation for the Morris, 1965; Alexander R. Burnett-Hurst, 1925).
absence of convergence in Figure 3 that we put These networks appear to have been organized
forward in this paper is based on the idea that the around the jobber, a foreman who was in charge
caste networks might place tacit restrictions on the of a work gang in the mill, factory, dockyard, or
occupational mobility of their members to pre- construction site, and more importantly also in
serve the integrity of the network. We will show charge of labor recruitment.
that although these restrictions might have been Given the information and enforcement prob-
welfare-enhancing and indeed equalizing when lems associated with the recruitment of short-term
they were first put in place, such restrictions could labor, it is not surprising that the jobber had to
result in dynamic inefficiencies when the structure lean on social connections outside his workplace
of the economy changes. such as his kinship and neighborhood connec-
The results in this paper provide empirical sup- tions (Chandavarkar, 1994, p. 107). Here the
port for the view that historical occupation pat- endogamous subcaste or jati served as a natural
terns kept in place by caste-based networks social unit from which to recruit labor, because
continue to shape occupational choice, and hence marriage ties strengthen information flows and
schooling choice, for the boys in the new econ- improve enforcement. This widespread use of
omy. In contrast, the lower-caste girls who histor- caste-based networks thus led to a fragmentation
ically kept away from the labor market, and so of the Bombay labor market along social lines.10
have no network ties to constrain them, take full Other studies also suggest that these patterns
advantage of the opportunities that become avail- tended to persist over many generations. For ex-
able in the new economy. The growing disparities ample, Hemalata C. Dandekar (1986) traces the
in school choices between boys and girls within evolution of a network of Jadhavs (a particular
the traditional jatis not only suggest a new balance subcaste) belonging to one village in interior Ma-
of economic opportunities by gender, but also harashtra. In 1942, 67 percent of the Jadhav mi-
could threaten the long-run stability of the caste grants from that village were working in the textile
system, which is based on endogamous marriages mills and 4 percent in other factories. Thirty-five
within the subcaste. Thus, a complete understand- years later, in 1977, 58 percent were still em-
ing of the development process must not only take ployed in textile mills, while 10 percent were in
account of the initial conditions and the role of other manufacturing industries.
existing institutions in shaping the response to A noticeable feature of historical descriptions
modernization and globalization, but must also of caste-based networks in Bombay is that they
consider how these traditional institutions are were restricted to working-class jobs. This is not
shaped in turn by the forces of change. surprising, because the information and enforce-
ment problems that give rise to such networks
tend to be more acute in those occupations. Fur-
I. The Institutional Setting ther, most studies of caste-based networks in
Bombay focus on male workers. Women were
A. Bombays Labor Market conspicuously absent from Bombays labor

Bombays industrial economy in the late nine-

teenth century and through the first half of the 10
The presence of caste clusters has been historically doc-
twentieth century was characterized by wide fluc- umented in the mills (R. G. Gokhale, 1957), among dock
workers (R. P. Cholia, 1941), construction workers, and in the
tuations in the demand for labor (Chandavarkar, railway workshops (Burnett-Hurst, 1925), in the leather and
1994). It is well known that such frequent job dyeing industries, and in the Bombay Municipal Corporation
turnover can give rise to labor market networks, and the Bombay Electric Supply and Transportation Company
particularly when the quality of a freshly hired (Chandavarkar, 1994). More recently, Kunj M. Patel (1963)
worker is difficult to assess and performance- surveyed 500 mill workers in the Parel area, close to the site of
our study, in 19611962 and found that 81 percent of the
contingent wage contracts cannot be imple- workers had relatives or members of their jati in the textile
mented. The presence of such recruitment industry. Sixty-six percent of the workers got jobs in the mills
networks has indeed been documented by numer- through the influence of their relatives and friends.
VOL. 96 NO. 4 MUNSHI AND ROSENZWEIG: TRADITIONAL INSTITUTIONS MEET THE MODERN WORLD 1231

TABLE 1SCHOOL CHARACTERISTICS AND STUDENT PERFORMANCE

English medium Marathi medium

School type (1) (2)
Panel A. School characteristics
Student-teacher ratio 36.71 35.76
(2.40) (2.17)
Class size 61.90 62.28
(3.69) (3.15)
Students per desk 2.40 2.36
(0.10) (0.11)
Proportion of teachers with B.Ed. 0.72 0.70
(0.07) (0.05)
Proportion of teachers with higher degree 0.08 0.10
(0.03) (0.03)
Computers per student 0.02 0.02
(0.004) (0.005)
Student enrollment in secondary section 1528.40 1059.00
(360.64) (175.73)
Panel B. School expenses
Fees 0.48* 0.20*
(0.01) (0.01)
Other expenses 1.10* 0.71*
(0.04) (0.01)
Panel C. SSC school-leaving exam results (19972001)
Percentage passed 92.59* 51.62*
(2.04) (5.95)
Percentage first class among passed 36.2* 24.23*
(1.69) (3.35)
Percentage distinction among passed 23.94* 6.90*
(3.92) (1.87)
Number of schools 10 18

Notes: Standard errors are in parentheses. Panels A and C use data from the school survey; panel B uses data from the
household survey. School characteristics are based on the secondary section (grades 5 through 10). School expenses are
measured for 2000 2001 in thousands of 1980 Rupees. To convert to 2000 Rupees, multiply by 4.44. Other school expenses
include transportation, coaching classes, textbooks, uniforms, and stationery. Scores above 35 percent are required to pass
SSC; scores above 60 percent are required for first class, and above 75 percent for distinction.
* Denotes rejection of the equality of means for the two types of schools with greater than 95-percent confidence.

force, particularly in the working-class jobs as well as recent student performance on the stan-
(Morris, 1965). These historical patterns of dardized school leaving examination (common to
labor force participation by gender will later both Marathi and English schools), which allows
help explain the schooling choice dynamics, us to compare the two types of schools as well as
for boys and girls, that we saw in Figure 3 and the students across the schools. Table 1, panel A,
Figure 4. describes school infrastructure and faculty quali-
fications in the English and Marathi schools. The
B. The Schools in Dadar average student-teacher ratio, class size, number
of students per desk, computers per student, and
Our analysis highlights the medium of instruc- the proportion of teachers with Bachelor of Edu-
tion as the salient feature of schooling choice. It is cation degrees and higher (postgraduate) degrees,
possible that the choice of the language of instruc- are each very similar and statistically indistin-
tion merely proxies for school quality. In parallel guishable for the two types of schools.11
with the household survey, we carried out a sur-
vey of schools based on a questionnaire filled out
by school principals. This questionnaire elicited 11
A regression of the language of instruction on the set
information on a variety of school characteristics, of school characteristics in Table 1, panel A, indicates that
1232 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2006

Despite the increase in the demand for En- ety, rather than differences in school quality,
glish education in the last ten years, as seen in dominate the schooling choice of parents. The
Figures 3 and 4, no new schools were added in survey elicited from parents the reasons for their
this period in Dadar.12 The English-language choice of school for their child. The percentage
schools accommodated this increased demand of parents reporting that the quality of educa-
by adding divisions in each grade, increasing tion was a factor in their choice was relatively
the number of desks in each classroom, and low and did not differ substantially across par-
doubling students on each desk. Because the ents choosing English-medium schools and
supply of schools was effectively fixed, we Marathi schools 43.7 percent versus 35.2 per-
would expect the English schools to extract cent, respectively. In contrast, almost 87 percent
some economic rents from this increased de- of parents who chose English as the medium of
mand through higher fees and schooling costs in instruction for their child reported that career
general. In contrast, fees in the Marathi schools opportunities was a factor in choosing that
are subsidized by the state government. Our school. And over 62 percent of parents who
household survey collected information on chose a Marathi-language school listed closer
school fees and other expenses (transportation, community ties as a reason.
coaching classes, textbooks, uniforms, and sta-
tionary) in the last year. Table 1, panel B shows II. A Simple Model of Schooling Choice
that school fees (in 1980 Rupees) are currently
significantly higher in the English schools (480 Our first objective in this section is to show
versus 200 Rupees), as are other expenses how exogenous, historically determined occu-
(1,100 versus 710 Rupees). pational differences across otherwise identical
One other difference between the schools is in jatis can persist when network externalities are
the performance of the students on the Secondary present. Because occupational choice translates
School Certificate (SSC) school-leaving examina- into schooling choice, this explains the initial
tion. Table 1, panel C reports student performance caste gap that we observe for the boys in
on this exam over a five-year period, 19972001. Figure 3 (the model that we lay out in this section
Students in the English schools perform much applies to the boys, as we will see later that labor
better on this standardized test in terms of the market networks are most active among the men).
percentage that pass and receive a first class and a We will show, however, that jatis should start to
distinction.13 Although these substantial differ- converge once the returns to English grow suffi-
ences in test performance can be explained by ciently large, which is inconsistent with what we
differences in school quality, they can also be observed in that figure. Our second objective in
explained by differential selection by ability into this section will consequently be to show how
English and Marathi schools, as implied by our network externalities could give rise to endoge-
network model of school choice, and we will nous social restrictions on occupational mobility,
provide evidence supporting this implication of and by extension schooling choice, preventing
the model below. convergence across social groups in a changing
In addition, our survey provides direct evi- economic environment.
dence that the medium of instruction and its
implications for childrens future role in soci- A. Population, Community Structure, and
Market Structure

the joint set of characteristics is not significantly different Consider a population with a continuum of
across the school types.
12 individuals. Each individual i is endowed with a
The average establishment year for the 18 Marathi
schools is 1947 and the corresponding year for the 10 level of ability !i ! {0, 12, 1}. Note that ability
English schools is 1959. All schools in the area have now in this section, and throughout the paper, refers
been operating for many decades.
13
to pre-schooling human capital rather than ge-
Scores above 35 percent are required to pass, scores netic ability. He lives for three periods, studying
above 60 percent are required for a first class, and scores
above 75 percent are required for a distinction. The same in the first period and working in the remaining
test is administered to all schools, with the questions trans- periods. Schooling choice is restricted to in-
lated into English and Marathi. struction in English or Marathi, the local lan-
VOL. 96 NO. 4 MUNSHI AND ROSENZWEIG: TRADITIONAL INSTITUTIONS MEET THE MODERN WORLD 1233

guage. Occupational choice is restricted to jobs alone. Munshi (2003) provides evidence that
white-collar and working-class jobs. Education experienced workers contribute disproportionately
in English is required to obtain a white-collar to labor market networks in the United States, and
job, but is more expensive than Marathi educa- we would expect this pattern to hold up in other
tion, which is assumed for simplicity to be economies as well. In our model, the expected
costless. Occupational choice is based on the working-class wage over the individuals working
wage that the individual will receive in the life is thus specified to be P, the proportion of the
white-collar and the working-class job, net of previous generation (three-year-olds) in the jati
the pecuniary cost of schooling. Each individual that will be employed in the working-class job
then makes his schooling decision based on the when he enters the labor force.
type of job that he (correctly) anticipates he will
occupy in the subsequent period. If he prefers to B. The Schooling Equilibrium
hold a white-collar job, then he will study in
English, if not he will study in Marathi, which is We now proceed to derive the different oc-
less costly. cupational distributions, and hence schooling
Each individual is born into a community or equilibria, that can be sustained across jatis with
jati. There is a large number of communities in the same ability distribution in this setup. Each
this economy, and we normalize so that the individual chooses the occupation, and hence
measure of individuals in each cohort or gener- the language of instruction, that maximizes his
ation of a jati is equal to one. To simplify and net return. This return depends on his own abil-
highlight the role of network externalities in ity, as well as the proportion of his jati in the
intergenerational occupational persistence, we previous generation employed in the working-
assume that the distribution of pre-schooling class occupation, as described above.
human capital does not vary over generations or Under conditions that we specify below, with
across jatis.14 Within each jati-generation there three levels of ability, three distinct schooling
is a measure PL of low types (with ability ! ! equilibria can be sustained within jatis: (a) only
0) and a measure PM of medium types (with low types choose Marathi education; (b) low and
ability ! ! 12). medium types choose Marathi education; (c) ev-
On the demand side of this labor market, eryone in the jati chooses Marathi education.15
firms operate competitively in both the work-
ing-class and the white-collar sectors. We noted CONDITION 1: PL " "/2.
earlier that working-class jobs generally tend to
be more heavily networked. For the purpose of CONDITION 2: "/2 " PL # PM " ".
this simple model, we assume that the white-
collar workers ability, and hence his produc- CONDITION 3: " " 1.
tivity, can be observed perfectly and so the
white-collar wage (net of schooling costs) is It is easy to verify that once a jati is exog-
specified to be "!i. Here " represents the returns enously assigned a particular occupational distri-
to ability in the white-collar job, which in our bution, this distribution will persist unchanged
setup also reflects the returns to English educa- over many generations when the conditions above
tion. In contrast, the nature of the production are satisfied.16 This intergenerational state depen-
technology prevents working-class firms from dence is a consequence of the network externality
directly observing their employees ability be- associated with the working-class occupation.
fore they commence work. We take it that the It implies, in turn, that the probability that any
firm is unable to specify a performance-contingent
wage contract, and so will use referrals from its
incumbent workers to hire new employees, gen- 15
Munshi and Rosenzweig (2003) consider the general
erating a role for the network in the working-class case with N types and N equilibria, without altering the
results that we present below.
16
It is merely necessary to show that no individual
wishes to deviate from the occupation, and hence schooling
14
We will relax this assumption in the empirical work by choice, assigned to his type in his jati in the previous
allowing for heterogeneity in ability across jatis. generation, for each of the schooling equilibria.
1234 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2006

individual i drawn randomly from jati j will be and hence schooling choice, in place to preserve
schooled in English (Eij ! 1) is related to the the viability of the community network.17
proportion of men in the previous generation em- To understand why restrictions on mobility
ployed in the working-class job, Pj: might emerge, define a social welfare function
that places equal weight on all members of the
jati. Now the welfare in a jati situated in equi-
(1) Pr$Eij # 1% # 1 $ Pj .
librium 3, in which everyone studies Marathi, is
simply the unweighted average of all the pay-
This expression will serve as the starting point offs from the working-class occupation, W ! 1.
for the empirical analysis described in Section When " just crosses one, in a given cohort, all
IV, where we will examine the relationship be- high types in the jati can expect to earn more in
tween schooling choice in the current genera- the white-collar sector than in the jatis tradi-
tion and the occupational distribution in the tional working-class occupation and will thus
previous generation, to identify the presence of switch to English schooling. Welfare from that
an underlying network organized around the cohort onward is then W ! (PL # PM)2 # (1 &
jati. PL & PM). The new welfare level is a weighted
average of PL # PM " 1 and 1, and so jati-level
welfare must unambiguously decline when
C. Schooling Choice as the Returns to schooling choice, and hence the occupational
English Grow distribution, shifts. Historically there was in-
tense competition for scarce working-class jobs
The state dependence at the level of the jati in Bombay, as noted in Section I. Because
derived above is obtained under the assumption larger numbers improve the jatis competitive-
that the parameters of the model, PL, PM, ", ness, and increase the working class wage in
remain stable over time. To explore the effect of general, it is easy to see why social restrictions
the increase in the returns to English (") in the on occupational mobility could emerge endog-
1990s, we now allow for multiple cohorts of enously. Moreover, the fact that the lower-caste
unit measure within each generation. girls in our sample do not display a similar
If " remains constant within a generation, the resistance to change can be attributed to the
results derived above follow through without gender-specific nature of these job networks.
modification for all cohorts. If, however, " in- Social restrictions on occupational mobility
creases across successive cohorts, holding Pj can be welfare-enhancing for small and medium
constant, then schooling choice within a jati changes in ", as noted above. But they could
could change over the course of a single gener- give rise to substantial inefficiencies if they
ation. When " just crosses one, high-ability continue to persist when " grows large. For
boys belonging to jatis that were traditionally in example, it is easy to verify that the social
equilibrium 3 switch to English. When " sub- restrictions described above for equilibrium 3
sequently reaches 2(PL # PM), medium-ability will be inefficient once " reaches 1 # (PL #
boys in jatis that were traditionally in equilib- PM), although a welfare calculation that identi-
rium 2 or equilibrium 3 switch to English, at fies the presence of such a dynamic inefficiency
which point schooling choice across all jatis is beyond the scope of this paper.
will converge.
Although the network externality described
above can explain the persistence of traditional
occupational patterns within the jati over many 17
Restrictions on mobility do not have to be associated
generations, and hence the initial caste gap ob- with explicit punishment. Preferences for schooling or fu-
served in Figure 3, it cannot by itself explain the ture career choices could be determined endogenously, for
absence of convergence over the 1990s as the example, by placing symbolic value on the traditional oc-
returns to English grew. To explain this absence cupation in the jati. Social interactions within the jati could
also lead individuals to make similar schooling choices and
of convergence, we consider the possibility that career choices across generations. We do not attempt to
heavily networked (working-class) jatis might distinguish between preferences that are complementary to
have put restrictions on occupational mobility, the network, and the network itself, in this paper.
VOL. 96 NO. 4 MUNSHI AND ROSENZWEIG: TRADITIONAL INSTITUTIONS MEET THE MODERN WORLD 1235

While we conjecture that restrictions on oc- Average ability among the English students
cupational mobility might be in place in the is greater than average ability among the Mar-
heavily networked jatis, no direct evidence of athi students at any point in time, but among the
their presence in Bombay is available. We can, Marathi students it is the group with the highest
however, test one important implication that is ability that deviate as " grows. Thus, while the
consistent with the presence of these restric- quality of the Marathi students unambiguously
tions: the relationship between schooling choice declines over time as the returns to English
Eij and the occupational distribution within the increase, the change in the quality of the pool of
jati in the previous generation Pj must not English students is ambiguous.18
weaken over successive cohorts in the current
generation, even as the returns to English grow. III. The Household Data
This stability in intergenerational state depen-
dence would then explain the wedge between A. The Survey
high-caste and lower-caste schooling choices
for boys that was observed through the 1990s in To examine empirically the role of caste net-
Figure 3. works in shaping mobility during a period of
change, we carried out a household survey
based on a random sample of students, stratified
D. Selection into Schools by caste, who entered the 28 secondary schools
in Dadar (in the first grade) over a 20-year
The model of schooling choice as laid out in period, 19822001. This design provides infor-
this section also has implications for selection, mation for the periods before and after the ma-
by ability, into English and Marathi schools. jor Indian economic reforms. We obtained a
Within any jati, the average pre-schooling hu- complete list of all students enrolled in grades 1
man capital of the English students must be to 10 in 2001 (the year of the survey), as well as
greater than that of the Marathi students. Taking a list of students who were enrolled in grade 10
the average across all jatis, this implies that from 1991 to 2000. Ignoring dropouts, this
average ability must be greater among the En- leaves us with 20 cohorts of students who en-
glish students at any point in time. This obser- tered school over the 19822001 period. A total
vation is consistent with the significantly higher of 101,567 students were enrolled in the schools
test scores obtained by students in the English in 2001 or studied in grade 10 over the previous
schools (Table 1), despite the fact that English ten years. We drew the roll numbers of 20,596
and Marathi schools appear to be similar in students randomly from these 20 cohorts, and
terms of the resources available per student and recovered their names and addresses from the
the qualifications of the teachers. But how does school records. Restricting attention to Maha-
the ability distribution within the English and rashtrians residing in Dadar and the immedi-
Marathi schools change across successive co- ately adjacent neighborhoods, we were left with
horts in the current generation as the returns to 8,092 eligible students to serve as the sampling
English grow? Without social restrictions, devi- frame for the survey. The students name is
ation to English education is ordered by ability,
so as " grows there is a steadily worsening pool
of Marathi students. Jatis that begin with a 18
For example, in the three-type case with no social
greater proportion of their members in working- restrictions, some jatis (in equilibrium 2) have only high-
class jobs have higher ability among the Mar- ability children in English schools, while other jatis (in
equilibrium 1) have both medium- and high-ability children
athi students, but their shift into English, and in English schools to begin with. This implies that the
hence the decline in ability, must also be more quality of the English pool must improve when " reaches
rapid, because all jatis ultimately converge. one, because only the high types from jatis in equilibrium 3
With social restrictions, heavily networked jatis deviate at that point. But average ability drops below its
continue to begin with a superior ability distri- initial level when " reaches 2(PL # PM), because medium
and high types in all jatis will have switched into English
bution within Marathi schools, but now there schools by that point. With social restrictions, the change in
might be no convergence in ability among Mar- ability within the English schools becomes even more dif-
athi students across jatis. ficult to characterize.
1236 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2006

typically a good indicator of the caste, and we unlikely to be undermined by these potential
wanted close to 1,000 upper castes in the sam- sources of bias.
ple, so all 1,082 students from this population
who appeared to be upper castes were selected B. Descriptive Statistics: Caste, Occupational
for the survey. We drew randomly from the Networks and Schooling
remaining students in the sampling frame until
the target sample size was reached. The upper The data provide empirical support for three
castes account for 17.5 percent of the final sam- features of the model of schooling choice laid
ple of 4,945 observations, which is slightly out in Section II. First, the occupational distri-
higher than the 13.4 percent that we began with bution, a product of historical circumstances,
in the sampling frame. varies by caste, and persists across generations,
The research team interviewed the parents of particularly among the men. Second, working-
the selected students at their residences. The class jobs are associated with a higher level of
survey instrument elicited detailed subcaste in- referrals (networking). And third, working-class
formation from the respondents and included jobs are associated with lower levels of English
sections on grandparents education and occu- schooling.
pation, parents education and occupational and The survey elicited information on parental
income histories (at five-year intervals from occupations at five-year intervals from 1980 to
1980 to 2000), as well as the students and 2000. For the grandparents, we simply asked for
siblings subsequent education, occupation, in- the main occupation over the individuals work-
come, and marriage outcomes (where rele- ing life. The 90 occupations in the data were
vant).19 Information on transfers, assistance in divided by roughly increasing levels of human
finding jobs, and ties to the community was also capital into seven aggregate categories: un-
collected. skilled manual, skilled manual, organized blue-
Of the eligible households, 82.5 percent collar, petty trade, clerical, business, and
provided completed schedules. This is a rel- professional. We further classified unskilled
atively high response rate, especially given manual, skilled manual, and organized blue-
that some of our addresses were 20 years old. collar as working-class occupations. Clerical,
But we might still have obtained a selective business, and professional were classified as
sample of households, for a number of differ- white-collar occupations. Petty trade is treated
ent reasons. First, households residing in Da- as an intermediate unclassified occupation.
dar who sent their children to study outside Table 2, panel A, describes the occupational
the area would be missing from the sample. distribution across broad caste categories (low,
Second, households who moved out of the medium, high), separately for the employed fa-
area would be among the 17.5 percent of the thers, based on information in 1995, and the pa-
respondents who did not complete the survey. ternal grandfathers of the students in the sample.
And third, students from the first ten cohorts Columns 1 to 3 of the panel indicate that lower-
who did not reach the tenth grade, and current caste fathers are much more likely to be employed
students who have dropped out, would be in working-class occupations (54 percent and 43
missing from the sample. In Section V, we percent) as compared with high-caste fathers (18
will discuss how our identification strategy is percent).20 The same cross-caste pattern is ob-
tained for individual occupations within the work-
ing-class and white-collar classifications, with the
19
The name is usually a good indicator of the individ- exception of clerical jobs. The comparison of the
uals community and caste. For example, 98.7 percent of the
respondents, whom we had selected on the basis of their
fathers in columns 1 to 3 with the grandfathers in
names from the school records, said that Marathi was their columns 4 to 6 also indicates that there has been
mother tongue, indicating that they were indeed Maharash- little change in the basic occupational distribution,
trian. The caste classification is potentially more problem- as well as the percentage of working-class
atic, however, because lower castes could in some cases
change their names or misreport their caste affiliation. Note
that such misreporting will not undermine the fixed effects
20
estimation strategy, described below, as long as it does not Note that we use only working-class and white-collar
vary by the gender of the child, within the jati. occupations when computing this statistic.
VOL. 96 NO. 4 MUNSHI AND ROSENZWEIG: TRADITIONAL INSTITUTIONS MEET THE MODERN WORLD 1237

TABLE 2OCCUPATION, EDUCATION, AND INCOME BY CASTE ACROSS GENERATIONS

Relationship to student Parent Grandparent

Caste Low Medium High Low Medium High
(1) (2) (3) (4) (5) (6)
Panel A. Fathers and grandfathers
Employment (%) 97.37 97.31 99.06 98.87 98.86 99.28
Occupational distribution (%)
Unskilled manual 11.09 7.84 4.41 9.00 3.63 2.10
Skilled manual 17.35 13.70 10.21 11.67 6.72 8.42
Organized blue-collar 22.87 19.22 2.90 22.89 24.23 7.67
Petty trade 4.00 4.51 2.52 3.11 3.20 3.34
Clerical 28.09 36.64 20.81 22.22 23.79 28.84
Business 7.95 8.79 15.51 6.11 4.72 13.00
Professional 8.30 8.79 43.51 5.56 6.18 33.66
Farming 0.35 0.51 0.13 19.44 27.53 2.97
Percent working class 53.64 42.91 18.01 56.24 49.92 19.42
(1.23) (1.21) (1.38) (1.33) (1.40) (1.44)
Years of schooling 9.63 10.22 13.82
(0.07) (0.07) (0.10)
Monthly income 1.92 1.99 4.61
(0.04) (0.04) (0.25)
Total number of observations 1,860 1,774 793 1,866 1,934 839

Panel B. Mothers and grandmothers

Employment (%) 20.56 20.01 51.23 19.31 18.59 15.57
Occupational distribution (%)
Unskilled manual 29.95 16.94 2.36 24.65 7.18 3.13
Skilled manual 8.82 8.47 6.15 1.70 1.44 3.13
Organized blue-collar 4.01 4.92 0.47 8.50 4.31 0.78
Petty trade 3.74 3.83 1.18 1.13 0.57 0.00
Clerical 31.55 40.71 46.34 4.25 2.30 19.53
Business 4.55 2.46 3.78 2.27 1.44 3.91
Professional 17.38 22.68 39.72 5.10 8.62 67.97
Farming 0.00 0.00 0.00 52.41 74.14 1.56
Percent working class 44.44 31.53 9.09 75.00 51.14 7.14
(2.62) (2.48) (1.41) (3.39) (5.36) (2.30)
Years of schooling 8.03 8.73 13.49
(0.09) (0.09) (0.10)
Monthly income 0.23 0.30 1.37
(0.02) (0.02) (0.07)
Total number of observations 1,887 1,954 857 1,885 1,953 854

Notes: Occupational distribution within each caste group is computed using employed individuals only. Employment for fathers and
mothers is computed as of 1995. Statistics in columns 46 are reported for paternal grandfathers and maternal grandmothers. Working
class ! 1 if unskilled manual, skilled manual, organized blue-collar; 0 if clerical, business, professional. Standard errors in parentheses.
Schooling and income statistics are computed using all parents in the sample, regardless of whether they are employed. Monthly income
is measured in thousands of 1980 Rupees in the year closest to the year in which the child entered school.
Occupational categories
Unskilled manual: daily wage labor, deliveryman, servant, hotel worker, helper, cleaner/sweeper, porter, assistant
watchman, fisherman, gardener, barber, cobbler (chambhar), unskilled laborer, seaman.
Skilled manual: machine operator, plumber, welder, technician, electrician, mechanic, carpenter, fitter/turner, tailor,
painter, film developer, goldsmith, artist, priest, lab assistant, skilled worker, traditional healer (vaidhya), computer operator.
Organized blue collar: mill worker, factory worker, peon, Bombay Port Trust (BPT) worker, Bombay Electric Supply and
Transportation (BEST) worker, Bombay Municipal Corporation (BMC) worker.
Petty trade: hawker, storeman (storekeeper), salesman, agent, shopkeeper.
Clerical: supervisor, driver, police, clerk, conductor, stenographer, postmaster, receptionist, foreman/draftsman, secretary.
Business: self business, medical representative, transporter, marketing, consultant, employer, contractor, politician (social
worker/leader), merchant.
Professional: tutor, teacher, programmer, engineer, officer, manager, doctor, lawyer, nurse, lecturer, vice-chancellor,
librarian, superintendant, director, principal, architect, salaried employee (service), chartered accountant, big businessman.
Farming: farmer, agricultural laborer.
1238 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2006

jobs, across the generations within broad caste of the occupational distribution for the men, for
categories.21 all castes, that we noted earlier, consistent with
Although most men are employed, we see that the view that labor networks are weak among
labor force participation (which includes part-time the women.
work) for the women in Table 2, panel B, is Together with the occupational distribution,
relatively low but is growing. Only 15 percent of Table 2 reports the mean years of schooling and
high-caste grandmothers worked; whereas just monthly income separately by caste for men and
over half of high-caste mothers entered the labor women.23 As expected, high-caste mothers and
force (based on their employment status in 1995). fathers have significantly more years of school-
Among the lower castes, the percentage employed ing and significantly higher incomes. Although
remains stable at 20 percent across the genera- the model in Section II assumes that the distri-
tions, but notice that farming is listed as the pri- bution of pre-schooling human capital is the
mary occupation for a large number of working same across castes (jatis), children in a wealthy,
grandmothers. This suggests that urban employ- educated jati that has had access to white-collar
ment must have increased sharply for the lower- jobs for many generations will be nurtured very
caste women as well. differently from children in a jati that was his-
The occupational distribution across castes for torically confined to manual jobs. This suggests
the mothers in panel B, columns 1 to 3, displays a that pre-schooling human capital could vary in
pattern similar to that for the fathers. Lower-caste practice across broad caste categories, and
women are much more likely to be employed in across jatis, as well. When estimating the effect
working-class occupations (44 percent and 32 per- of the historical occupational distribution on the
cent) as compared with high-caste women (9 per- childs schooling choice, we will consequently
cent). There is an important difference, however, take account of the possibility that the occupa-
between men and womenalthough the large dif- tional distribution could be correlated with the
ference within the working-class occupations for ability distribution in the jati.
the men was in access to blue-collar jobs for the Table 3 indicates that, as assumed in the
lower castes, for women the major difference is in model, working-class occupations are associ-
access to unskilled manual jobs; many of the ated with higher levels of networking (refer-
lower-caste women work as sweepers and domes- rals).24 Column 1 shows that 68 percent of the
tic servants. working-class men received help from a relative
Columns 1 to 3 and columns 4 to 6 in panel or member of the community in finding their
B suggest that there has been, in contrast to the first job (or starting their first business if
men, significant intergenerational change in oc- self-employed), which is significantly higher
cupational patterns for women within castes.
The urban occupations that show the greatest 23
increase are skilled manual, clerical, and pro- Recall that income information was collected from
each parent at five equal points in time from 1980 to 2000.
fessional (with the exception of the high We use the income (in 1980 Rupees) that coincides as
castes).22 The decline in the percentage of closely as possible with the year in which the child entered
working-class jobs among the lower-caste school. Thus, the income in 2000 is used for students age 6
women, across a single generation, is particu- to 10, the income in 1995 for students 11 to 15, the income
in 1990 for students 16 to 20, and the income in 1985 for
larly dramatic. This contrasts with the stability students 21 to 25. The same income statistic is used later in
the schooling regressions.
24
The parents of the selected students were asked how
21
The exception is farming, which is listed as the pri- they learned about their first job: through a childhood friend,
mary occupation for a large proportion of lower-caste through a college friend, through a relative, through a mem-
grandfathers. This implies, in turn, that roughly one-quarter ber of the community (jati), or by some other means (which
of the lower-caste fathers are first-generation migrants. Mi- was left open-ended in the questionnaire). This open-ended
grants are by definition newcomers in the labor market, and category included cases in which no help was received, or in
so will be more susceptible to the information problems that which the job was found through newspaper advertise-
generate a need for the caste networks. ments, campus interviews, and other impersonal informa-
22
The decline in the proportion of high-caste women in tion channels. A binary referral variable was then
professional jobs is most likely because only the highest- constructed, taking the value of one if the parent learned
ability women of the older generation (grandmothers) en- about the first job from a relative or member of the com-
tered the labor force. munity, and zero otherwise.
VOL. 96 NO. 4 MUNSHI AND ROSENZWEIG: TRADITIONAL INSTITUTIONS MEET THE MODERN WORLD 1239

TABLE 3REFERRALS AND SCHOOLING BY OCCUPATION

Relationship to student Father Mother

Percentage that Percentage that Percentage that Percentage that
Outcomes and choices received referrals studied in English received referrals studied in English
(1) (2) (3) (4)
Occupation
Unskilled manual 65.95 0.80 61.29 0.00
Skilled manual 60.13 2.24 45.56 5.56
Organized blue - collar 76.43 0.91 69.44 5.56
All working class 68.44 1.36 57.69 2.24
(standard error) (1.11) (0.28) (2.80) (0.84)
Petty trade 57.89 1.75 61.76 2.94
Clerical 47.41 2.89 30.56 7.26
Business 49.29 8.53 41.86 9.30
Professional 32.77 11.38 29.25 14.47
All white - collar 43.76 6.20 30.64 10.13
(standard error) (1.02) (0.49) (1.60) (1.05)
Number of observations 4,515 4,513 1,215 1,215

Notes: Statistics are computed using employed individuals only. Farmers are excluded. A parent is said to have received a
referral if a relative or member of the community found him/her a job. A parent is said to have studied in English if he/she
studied in that language in secondary school. Occupational categories are defined in Table 2.

than the 44 percent of men in the white-collar tween particular occupations, the level of referrals,
jobs who received a referral. The correspond- and English schooling, to characterize the occu-
ing statistics for the women in column 3 re- pational distribution in the jati.
veal essentially the same pattern, although the Table 2 suggests that lower-caste men and
level of referrals for the women is generally women are much more likely to hold work-
lower than that for the men, perhaps because the ing-class jobs. Combining these cross-caste
networks for female jobs are less developed. patterns with the results in Table 3, it is not
The model also assumes that Marathi school- surprising that a much higher proportion of
ing channels the student into a working-class lower-caste men received referrals (60 per-
job, while English schooling leads to the white- cent versus 37 percent), and that these men
collar occupation. The survey elicited informa- are also much less likely to have been
tion on the language of instruction (English schooled in English (2 percent versus 12 per-
versus Marathi) for fathers and mothers, in sec- cent). In contrast, although lower-caste
ondary school. Columns 2 and 4 of Table 3 show women are also much less likely to be
that there is a clear distinction between working- schooled in English, the level of referrals is
class and white-collar jobs with respect to the statistically indistinguishable across castes.25
language of instruction in secondary school. The The level of referrals is low in any case (13
percentage of men in working-class jobs that at- percent for the lower castes and 19 percent
tended secondary school in English is just over 1 for the high castes), especially when com-
percent, compared with the 6 percent of men in pared with the corresponding level for the
white-collar jobs. In column 4, a similar pattern is men, and we will later establish that labor
obtained for the women. We have described the
relationship between the broad occupational cate-
gories (working-class versus white-collar), the 25
Although we noted earlier that lower-caste women
level of referrals, and English schooling. But in- who work are more likely to hold working-class jobs, which
spection of Table 3 indicates that the level of are associated with more referrals, we also saw that lower-
caste women are less likely to enter the labor force. These
referrals and English schooling vary systemati- two opposing effects appear to cancel each other, leaving
cally within these categories as well. Later, we little variation in the level of referrals across castes for the
will take advantage of this finer relationship be- women.
1240 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2006

market networks are effectively available for which could independently determine school-
the men only. ing choice.
Pj measures the proportion of men in the pre-
IV. Empirical Analysis vious generation employed in working-class jobs.
Although the model assumes that only two types
A. Specification and Identification of jobsworking-class and white-collarare
available, as many as 90 occupations are listed
The first implication of the model is that the in the data. A relatively strong relationship be-
occupational distribution in the jati should tween the level of referrals and the type of
persist across generations when networks are occupation was observed earlier in Table 3, and
active. Because schooling choice maps into so one convenient statistic that accurately and
occupational choice, equation (1) in Section parsimoniously describes the occupational dis-
II expressed this implication in terms of tribution in the jati would be the proportion of
schooling choice in the current generation and fathers (the previous generation) who received a
the occupational distribution in the previous job referral. Working-class jobs were also asso-
generation: ciated with lower levels of English schooling
(Table 3). An alternative measure of the occu-
Pr$Eij # 1% # 1 $ Pj . pational distribution in the previous generation
would compute the proportion of fathers who
Recall that Eij ! 1 if individual i belonging to attended English secondary schools. Most of the
jati j is schooled in English; Eij ! 0 if he is regressions reported in this paper will use the
schooled in Marathi; and Pj is the proportion of referrals statistic to measure Pj; English school-
men in the jati in the previous generation who ing levels were generally low in the previous
are employed in working-class jobs and so in a generation and so there is substantially more
position to provide referrals. variation in the referrals statistic across jatis.
The particular relationship between Eij and Pj We will, however, verify that the results hold up
in the equation above is, of course, a conse- with the English-schooling statistic as well.
quence of the modelling assumption that Following the discussion above, we expect to
schooling choice maps perfectly into future find % " 0 when networks are active and the
occupational outcomes. More generally, we occupational distribution persists across gener-
would expect to see a negative coefficient, but ations. Recall that % must also remain stable
not necessarily with magnitude one, on Pj. The across cohorts in the current generation to ex-
model laid out in Section II also does not allow plain the absence of convergence in Figure 3.
for intergenerational state dependence in Although much of the analysis treats % as con-
schooling choice at the level of the household. stant, we will later verify that % does indeed re-
Moreover, we noted above that pre-schooling main stable across cohorts.
human capital and family incomes appeared to An identification problem arises when Pj and
vary systematically across castes with different !j are correlated in equation (2). Although jatis
occupational backgrounds. The schooling re- might have been the same to begin with, we
gression that we estimate is consequently spec- noted in the previous section that their members
ified as now have very different characteristics (income
and education), depending on the type of occu-
(2) Pr$Eij # 1% # %Pj & Xij' & !j , pation that the jati has historically been engaged
in. A traditionally working-class jati could thus
where X ij includes the parents language of be associated with high Pj and low !j , in which
schooling to reflect household-level state de- case a family effect would be erroneously in-
pendence in schooling choice, as well as a terpreted as a network effect because individu-
cohort variable to capture the increase in the als with lower family resources independently
returns to English over successive cohorts in select into Marathi schools.
the current generation; ! j measures unob- Our solution to this identification problem
served or imperfectly observed pre-school hu- exploits the fact, documented in Table 2, that
man capital and family income in the jati, networks are concentrated in working-class jobs
VOL. 96 NO. 4 MUNSHI AND ROSENZWEIG: TRADITIONAL INSTITUTIONS MEET THE MODERN WORLD 1241

dominated by men. We mentioned earlier that none of these explanations fits the data quite as
the levels of referrals for women were relatively well as the male labor market network explanation
low, consistent with the significant change in we put forward in this paper.
the occupational distribution across generations
for women indicated in Table 2. Thus, although B. Caste-Based Networks and Schooling
the networks might affect schooling choice for Choice
the boys, they should have had little or no
impact on the girls. The model in Section II then Table 4, column 1, reports the estimates of
applies to boys only. Instead of using variation the schooling choice regression, equation (2),
in the level of referrals across jatis to identify for the boys. As noted, the sample covers 20
the presence of networks, as in equation (2), we cohorts of students age 6 to 25, who entered
proceed instead to exploit this gender difference school between 1982 (cohort ! 1) and 2001
in the access to job networks by pooling both (cohort ! 20). The students cohort (1 to 20),
sexes in the schooling regression to identify the the proportion of fathers in his jati who received
presence of the network within the jati: a referral, and the fathers and the mothers
language of instruction in secondary school are
included as regressors.
(3) Pr$Eij # 1% # $% $ % %Pj ! Bij & Xij' The cohort term is included in this regression
to account for the increase in the returns to
& Xij ! Bij $' $ ' % & (Bij & fj , English over time. While the linear cohort effect
we specify in Table 4 is clearly restrictive, we
where % , ' represent the effect of the network verify below that the estimated referral coeffi-
and parents language of schooling on the girls. cient is unchanged when we allow for more
Bij is a dummy variable that takes a value of one flexible cohort effects. The referral coefficient
for boys and zero for girls. The advantage of is also specified to be constant over time in
pooling the boys and girls is that the schooling Table 4, and we will subsequently relax this
regression can be estimated with jati fixed ef- restriction as well. For now, we see that the
fects, fj ' % Pj # !j. Although we can no longer referral coefficient is negative and significant;
identify % directly, we can obtain a consistent children belonging to (historically) working-
estimate of % & % , the coefficient on the Pj ! Bij class and more heavily networked jatis are less
interaction term. For the special case with ex- likely to be schooled in English, consistent with
clusively male networks, % ! 0 and the coeffi- the first implication of the model. The cohort
cient on the interaction term identifies network- effect is positive and significant, implying a
based occupational persistence for the boys shift into English over time, which is consistent
directly. More generally, the coefficient on the with the increase in the returns to English we
interaction term provides a conservative esti- saw in Figure 1. Finally, the results imply that a
mate of the effect of caste-based networks on boy is much more likely to be schooled in
schooling choices for the boys. English if his parents were educated in that
The identifying assumption in this estimation language, indicating significant state depen-
strategy is that no variable )j ! Bij appears in the dence in schooling choice at the level of the
residual of equation (3), where )j is correlated household.
with Pj. A sufficient condition for this identifying Table 4, column 2, reports estimates from a
assumption to be satisfied is that no unobserved specification that includes variables that deter-
determinant of schooling choice should vary by mine the students pre-schooling human capital
gender or have a differential effect on schooling as well as the household budget constraint,
choice by gender, within the jati. Later in Section which could independently determine schooling
V we will discuss alternative explanations for the choices. The parents years of education, condi-
negative and significant % & % coefficient we tional on their language of instruction in sec-
obtain in the schooling regression. These explana- ondary school and the level of referrals in the
tions either relax the assumptions of the model, jati, are likely significant determinants of chil-
made earlier in Section II, or build on the failure of drens pre-schooling human capital. The fami-
the identifying assumption. We will argue that lys access to own resources is measured by the
1242 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2006

TABLE 4CASTE-BASED NETWORKS AND SCHOOLING CHOICE

Dependent variable English schooling

Sample Boys only Girls only Boys and girls
(1) (2) (3) (4) (5) (6)
Referrals &1.060 &0.377 &0.646 0.124
(0.164) (0.148) (0.160) (0.167)
Referral - boy &0.398 &0.464
(0.091) (0.105)
Cohort 0.013 0.009 0.013 0.009 0.017 0.010
(0.002) (0.002) (0.002) (0.002) (0.002) (0.002)
Father studied in English 0.320 0.236 0.388 0.309 0.301
(0.037) (0.033) (0.037) (0.026) (0.026)
Mother studied in English 0.351 0.220 0.441 0.269 0.259
(0.041) (0.028) (0.071) (0.045) (0.043)
Fathers years of education 0.023 0.020 0.021
(0.004) (0.003) (0.003)
Mothers years of education 0.023 0.026 0.024
(0.003) (0.003) (0.003)
Family income 0.005 0.009 0.007
(0.005) (0.003) (0.003)
Boy 0.270 0.297
(0.049) (0.077)
Cohort - boy &0.002 &0.001
(0.002) (0.002)
Father studied in English - boy &0.091
(0.044)
Mother studied in English - boy &0.044
(0.042)
Fathers years of education - boy 0.002
(0.005)
Mothers years of education - boy &0.001
(0.004)
Family income - boy &0.003
(0.005)
R2 0.173 0.274 0.146 0.272 0.163 0.299
Number of observations 2,405 2,286 2,228 2,093 4,635 4,379

Notes: Standard errors in parentheses are robust to heteroskedasticity and clustered residuals within each jati. English schooling !
1 if the child is sent to an English school, 0 if the child is sent to a Marathi school. Referrals measures the proportion of fathers in
the jati who received a referral. Boy ! 1 if the student is a boy, 0 if girl. Family income is measured in thousands of 1980 Rupees
in the year that is closest to the year in which the child entered school. Columns 12: schooling choice for boys. Columns 3 4:
schooling choice for girls. Columns 5 6: schooling choice for both boys and girls, including a full set of jati dummies.

total income of the father and the mother at the cohort variable is quite stable. And the coeffi-
time when the child entered school.26 Inclusion cients on the additional regressors all have sen-
of these variables results in a substantial decline sible signs; the boy is more likely to be schooled
in the referral coefficient, suggesting that the in a more-expensive English-medium school if
level of referrals was previously proxying to his father or mother are more educated, or if the
some extent for unobserved, family-specific de- family is wealthier.
terminants of schooling choice, but it remains The estimates for girls are reported in col-
negative and significant. The coefficient on the umns 3 and 4 in Table 4. Column 3 reports the
estimates based on equation (2); column 4 re-
26
ports the estimates from the augmented speci-
We use the income in 2000 for students currently age 6
to 10, the income in 1995 for students 11 to 15, the income in fication that adds the parents years of schooling
1990 for students 16 to 20, and the income in 1985 for students and family income as additional regressors. The
21 to 25. All incomes are computed in 1980 Rupees. estimated cohort effects, and the coefficients on
VOL. 96 NO. 4 MUNSHI AND ROSENZWEIG: TRADITIONAL INSTITUTIONS MEET THE MODERN WORLD 1243

parents language of instruction, parents edu- is no smaller than the corresponding coefficient
cation, and family income are similar to those estimated earlier in column 5 without any
for boys in columns 1 and 2. The referral coef- household characteristics. This stability con-
ficient, however, becomes negligible for the trasts with the decline in the referral coefficient
girls in column 4 once the observed determi- in Table 4, columns 1 to 4, when family char-
nants of pre-schooling human capital and access acteristics were included, providing some sup-
to own family resources are included. One ex- port for the view that the jati fixed effects
planation for this result is that girls receive help absorb much of the unobserved heterogeneity in
from the women, not the men, in their jati. But this environment.27
we noted that the level of referrals for the Notice also that parents years of schooling
women is very low, across all castes. Although and family income, which had a strong influ-
not reported, we also found no correlation be- ence on schooling choice for both boys and girls
tween referrals and schooling choice, for both in columns 1 to 4, do not differentially affect
boys and girls, when we replaced the level of schooling choice by gender (column 6). It is
referrals for the fathers with the level of refer- only the jati-level referral variable that has such
rals for the mothers. a differential effect on schooling choice, as
The results we have just described are con- measured by the negative and significant refer-
sistent with the view that caste-based networks, ral-boy coefficient. This observation will be
net of individual and family characteristics, af- useful later in Section V when we consider
fect schooling decisions for the boys, but not for alternative explanations for the results pre-
the girls. But up to this point, we have con- sented in this paper.
trolled only for unobserved ability with a lim-
ited number of family characteristics. A more C. Schooling Choice over Time
robust identification strategy estimates the
schooling regression with jati fixed effects, as in The second implication of the model laid out
equation (3). These estimates are reported in in Section II is that the relationship between
column 5 of Table 4. As noted, only the referral- schooling choice and the occupational distribu-
boy interaction coefficient, and not the linear tion in the previous generation will weaken
referral coefficient, can now be identified. The across successive cohorts in the current gener-
coefficient on this term is negative and signifi- ation as the returns to English grow, unless
cant, and very similar to the referral coefficient restrictions on occupational mobility are in
for the boys in column 2. Recall from equation place. To assess empirically the stability of the
(3) that the coefficient on the referral-boy inter- referral coefficient, we create 4 cohort catego-
action term provides us with a direct estimate of ries that evenly divide the 20 cohorts, and then
the referral coefficient for the boys if the referral estimate the referral coefficient separately for
coefficient for the girls is zero. The result we each category.
obtained earlier for the girls, in column 4, sug- We begin with a benchmark jati-fixed-effects
gests that this might well be the case. regression, which maintains a constant referral
The regression specification with jati fixed coefficient but relaxes the restriction imposed
effects in Table 4, column 5, did not include
family characteristics. Column 6 includes par-
ents language of schooling, parents years of 27
A previous version of the paper (Munshi and Rosen-
schooling and family income, both interacted zweig, 2003) reported a number of additional robustness
and uninteracted with the boy dummy, as addi- tests. First, we accounted for occupational persistence at the
level of the family by including a full set of (90) dummies
tional regressors. Including uninteracted family for the students fathers occupation. Second, we allowed
characteristics in the schooling regression has for the possibility that the scope of the network was deter-
no effect on the estimated referral-boy coeffi- mined by caste and the region of origin (within Maharash-
cient by construction, once jati fixed effects are tra) by replacing the jati by the jati-region as the boundary
included. But we see that the inclusion of the of the network. Third, we dropped very large jati-regions
(more than 250 observations) and very small jati-regions
family characteristics, interacted with the boy (fewer than 10 observations). The estimated referral-boy
dummy, has no effect on the estimated referral- coefficient with these alternative specifications was shown
boy coefficient as well. Indeed, this coefficient to be very similar to what we report in Table 4.
1244 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2006

TABLE 5SCHOOLING CHOICE OVER TIME

Dependent variable English schooling

Without Without
family With family family With family
Additional regressors characteristics characteristics characteristics characteristics
(1) (2) (3) (4)
Referral - boy &0.426 &0.478
(0.090) (0.106)
Referral - boy - cohort1 &0.269 &0.416
(0.168) (0.167)
Referral - boy - cohort2 &0.352 &0.333
(0.100) (0.112)
Referral - boy - cohort3 &0.523 &0.540
(0.145) (0.143)
Referral - boy - cohort4 &0.607 &0.665
(0.256) (0.238)
Cohort 1 &0.261 &0.161 &0.261 &0.161
(0.031) (0.032) (0.030) (0.032)
Cohort 2 &0.231 &0.146 &0.231 &0.146
(0.031) (0.028) (0.031) (0.028)
Cohort 3 &0.161 &0.121 &0.161 &0.121
(0.030) (0.023) (0.030) (0.023)
Boy 0.236 0.261 0.338 0.364
(0.065) (0.091) (0.156) (0.149)
Cohort 1 - boy 0.033 0.031 &0.152 &0.106
(0.038) (0.037) (0.209) (0.169)
Cohort 2 - boy 0.052 0.031 &0.090 &0.153
(0.042) (0.035) (0.174) (0.151)
Cohort 3 - boy 0.041 0.041 &0.007 &0.030
(0.032) (0.024) (0.117) (0.114)
R2 0.164 0.301 0.164 0.301
Number of observations 4,635 4,379 4,635 4,379

Notes: Standard errors in parentheses are robust to heteroskedasticity and clustered residuals within each jati. English
schooling ! 1 if the child is sent to an English school, 0 if the child is sent to a Marathi school. Referrals measures the
proportion of fathers in the jati who received a referral. Boy ! 1 if the student is a boy, 0 if girl. Cohort 1: age 2125; Cohort
2: age 16 20; Cohort 3: age 1115; Cohort 4: age 6 10. Column 2 and column 4 include family characteristics, separately
and interacted with the boy dummy. Family characteristics include parents language of schooling and years of education, and
total family income. A full set of jati dummies is included in all regressions. Sample includes boys and girls.

thus far that cohort effects are linear, by includ- cisely estimated. The referral coefficient is ac-
ing the cohort categories in Table 5, column 1. tually increasing for the later cohorts, and we
The estimated negative referral-boy coefficient can easily reject the convergence hypothesis
is unaffected by the inclusion of the flexible which implies a decline in the referral effect
cohort effect and remains very similar to the over time. Once more, the estimated referral
results shown in Table 4. Inclusion of the family coefficients are robust to the inclusion of the
background variables, uninteracted and inter- family background variables as regressors (col-
acted with the boy dummy, as additional regres- umn 4). Although not reported here, the referral
sors again has no effect on the estimated referral coefficient remained stable when the schooling
coefficient (column 2). regression was estimated with boys only, in-
Table 5, column 3, allows for changes in the cluding parents years of education and family
referral coefficient across cohort categories. All income as additional regressors. It is this jati-
the referral-boy-cohort coefficients are negative level effect that presumably sustains the gap in
and significant except for the coefficient on the schooling choice between broad caste catego-
first cohort category, which is slightly less pre- ries observed for the boys in Figure 3.
VOL. 96 NO. 4 MUNSHI AND ROSENZWEIG: TRADITIONAL INSTITUTIONS MEET THE MODERN WORLD 1245

TABLE 6ALTERNATIVE MEASURES OF THE OCCUPATIONAL DISTRIBUTION AND SCHOOLING

Dependent variable English schooling Test scores

Proportion of fathers that
Occupational distribution measure Proportion of fathers schooled in English received a referral
Sample Boys only Girls only Boys and girls Boys only Girls only Boys and girls
(1) (2) (3) (4) (5) (6) (7)
Occupational distribution 0.847 0.083 &23.151 &23.650
(0.262) (0.427) (5.045) (4.080)
Occupational distribution - boy 0.701 0.869 &0.734
(0.224) (0.221) (5.761)
Cohort 0.008 0.010 0.017 0.010 &0.505 &0.180 &0.190
(0.002) (0.002) (0.002) (0.002) (0.134) (0.204) (0.223)
Boy 0.026 &0.006 3.794
(0.028) (0.031) (4.357)
Father studied in English 0.217 0.301 0.313 4.901 2.323 1.847
(0.032) (0.027) (0.026) (1.397) (3.711) (4.028)
Mother studied in English 0.220 0.266 0.259 3.312 &2.596 &2.772
(0.030) (0.046) (0.042) (2.200) (1.905) (1.632)
Fathers years of education 0.024 0.019 0.020 0.929 0.765 0.812
(0.004) (0.002) (0.003) (0.195) (0.225) (0.238)
Mothers years of education 0.023 0.025 0.023 0.617 1.074 0.984
(0.003) (0.003) (0.003) (0.221) (0.199) (0.200)
Family income 0.005 0.008 0.007 0.260 0.122 0.107
(0.004) (0.003) (0.003) (0.118) (0.076) (0.076)
R2 0.275 0.272 0.162 0.298 0.322 0.334 0.354
Number of observations 2,286 2,093 4,635 4,379 849 775 1,624

Notes: Standard errors in parentheses are robust to heteroskedasticity and clustered residuals within each jati. The sample in
columns 57 is restricted to children in cohorts 110, past the school-leaving age, who passed the SSC exam. Test scores
range from 35 to 100. Boy ! 1 if the student is a boy, 0 if girl. Family income is measured in thousands of 1980 Rupees in
the year closest to the year in which the child entered school. Column 3 also includes cohort interacted with boy. Column 4
and column 7 also include cohort, father/mother studied in English, fathers/mothers years of education, and family income,
interacted with boy. Regressions pooling boys and girls (columns 3 4 and 7) include a full set of jati dummies.

D. Robustness and Validation: Alternative The coefficient on the English-boy interaction

Measures of the Occupational Distribution term in Table 6, columns 3 to 4, which include
and Schooling jati fixed effects, is very similar to the English
coefficient for the boys in column 1, matching
The regression results reported thus far used the results reported earlier with the referrals
the proportion of fathers who received a referral variable. Although not reported, once again par-
for their first job to measure the occupational ents education and family income do not have
distribution. We now proceed to verify the ro- a differential effect on schooling choice by
bustness of the results by repeating the school- gender.
ing regressions for boys, girls, and the pooled The language of instruction measures the
sample with the proportion of fathers schooled childs future occupation, and the referral sta-
in English as the measure of the occupational tistic measures the occupational distribution in
distribution. the previous generation, in most of the regres-
The coefficients on the cohort variable and sions that we report in this paper. The negative
the household characteristics in Table 6, col- and significant referral-boy coefficient in the
umns 1 to 2, are very similar to the estimates fixed effects regressions then reflects the persis-
reported in Table 4. The coefficient on the En- tence in the occupational distribution across
glish proportion, which can be interpreted as generations, differentially for boys and girls
state dependence in schooling choice at the jati within the jati. To validate this interpretation of
level, is positive and significant as expected. the results, we proceed to replace the language
1246 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2006

of instruction with an alternative schooling out- column 6). The cohort effect is now absent, but
come as the dependent variable. the coefficient on referrals remains negative and
We assume that test scores depend on school statistically significant both boys and girls
quality and the pre-schooling human capital of from high-referral jatis do less well on exams.
the student. The comparison of English and The fixed-effects estimates, reported in Table 6,
Marathi schools in Table 1, and the parents per- column 7, of the cohort effect, the cohort-boy
ception of these schools, indicates that school interaction, and the boy dummy are not statis-
quality does not vary by the language of instruc- tically significantly different from zero. More
tion. Under the maintained assumption that pre- importantly, the coefficient on the referral-boy
schooling human capital does not vary by gender interaction term is small and statistically insig-
within the jati, this implies that the referral-boy nificant, in contrast to the specifications with
coefficient should be close to zero in the fixed language of instruction as the dependent vari-
effects regression with test scores as the depen- able. Caste networks affect the language but not
dent variable. The fact that referrals have a the quality of instruction of their members.
differential effect by gender on schooling choice
should be irrelevant for test scores, if school
quality does not vary by the language of E. Selection into Marathi Schools over Time
instruction.
Columns 5 to 7 of Table 6 replace the lan- The framework laid out in Section II also has
guage of instruction with performance on the implications for the compositional change in the
school-leaving SSC examination as the depen- students who attend Marathi schools over time
dent variable. Referrals are once more used to by jati: first, the pre-schooling human capital of
measure the occupational distribution in the jati, boys entering Marathi schools should decline
to be consistent with the specifications used on average as the returns to English grow. Sec-
elsewhere in the paper. We restrict attention to ond, when there are no restrictions on mobility
the first ten cohorts (age 16 25), which have put in place to exploit network externalities, the
already attained school-leaving age, in these distribution of pre-schooling ability among the
regressions. Only 17 percent of the students age boys entering Marathi schools will converge
16 25 in the sample never passed the SSC across all jatis over time. It is possible that such
examination, so we focus on the test score con- convergence across jatis will be absent when
ditional on having passed the exam in these restrictions are in place. Note that the model has
regressions.28 no prediction for selection by ability into En-
Table 6, column 5, restricts attention to boys, glish schools.
and includes the cohort, family characteristics, We do not have a direct measure of pre-
and the level of referrals in the jati as regres- schooling human capital. The results in Table 6
sors. The cohort effect is negative and signifi- suggest, however, that, net of income, parental
cant, suggesting a decline in the quality of schooling has a positive and significant effect
students over time. The referral coefficient is on school performance. In particular, fathers
also negative and precisely estimated, which schooling has a significant positive effect on test
would be the case if students from high-referral scores for boys and girls, and the effects do not
jatis have lower levels of pre-schooling human differ significantly by the gender of the child.
capital. Consistent with this interpretation, fam- We thus use the fathers schooling level as a
ily characteristics, particularly parents years of proxy for pre-schooling ability.29 The question
education, have a very large positive effect on we address is whether boys with more educated
test performance. Subsequently we repeat the fathers increasingly exit Marathi schools and
exercise just described for the girls (Table 6, whether, and how, the rate of decline in the
pre-schooling ability of boys entering Marathi
28
schools varies by jati.
Munshi and Rosenzweig (2003) studied the effect of
referrals on the probability of success in the SSC exam and
obtained results that are qualitatively the same as what we
29
report below with the test score, conditional on success, as The results reported below are essentially the same if
the dependent variable. we replace fathers schooling by mothers schooling.
VOL. 96 NO. 4 MUNSHI AND ROSENZWEIG: TRADITIONAL INSTITUTIONS MEET THE MODERN WORLD 1247

TABLE 7SELECTION INTO MARATHI SCHOOLS OVER TIME

Dependent variable Fathers years of education

Cohort 1120 110
Boys Boys
and and
Sample Boys Girls girls Boys Girls girls
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Cohort &0.697 &0.577 0.423 0.219 0.153 0.240 &0.543 &0.087
(0.230) (0.155) (0.230) (0.189) (0.191) (0.146) (0.326) (0.156)
Cohort - boy &0.792 0.117
(0.252) (0.232)
Referral - cohort 1.430 1.204 &0.596 &0.280 &0.147 &0.258 1.054 0.310
(0.400) (0.260) (0.376) (0.311) (0.323) (0.250) (0.554) (0.270)
Referral - cohort - boy 1.469 &0.231
(0.414) (0.415)
Referrals &30.991 &3.651 &10.256 &18.624
(6.894) (6.866) (2.035) (2.793)
2
R 0.106 0.205 0.138 0.205 0.215 0.136 0.254 0.184 0.278 0.285
Number of 839 839 815 815 1,654 866 866 851 851 1,717
observations

Notes: Standard errors in parentheses are robust to heteroskedasticity and clustered residuals within each jati. Referrals
measures the proportion of fathers in the jati who received a referral. Column 2, column 4, column 7, and column 9 include
a full set of jati dummies. Column 5 and column 10 include a full set of jati dummies, jati-boy dummies, and jati-cohort
dummies. All regressions are restricted to students in Marathi schools.

To test the implications for school selectivity estimates are reported in Table 7, column 1. The
described above, we estimate regressions on the cohort coefficient is negative and significant as
subsample of boys entering Marathi schools of predicted, which implies that the pre-schooling
the form human capital of the boys who entered Marathi
schools was declining substantially in the 1990s.
The coefficient on the referral-cohort interac-
(4) E$S ij !E ij # 0% # * & + R j & , C ij tion term is positive, consistent with the results
in Table 5 showing that restrictions on mobility
& - R j ! C ij & .! j , in the high-referral networks were still in place
during this periodthe more-able boys in high-
where S ij is boy i in jati js fathers years of referral jatis were shifting to English-medium
schooling; C ij is the boys cohort; R j mea- schools at lower rates.
sures the level of referrals in the jati; and ! j The referral coefficient is negative and sig-
measures pre-schooling ability in the entire nificant, consistent with the results in Table 6,
jati. These terms reflect the fact that pre- which indicate that jatis with higher referrals Rj
schooling ability conditional on selection into have lower ability !j; the negative Rj & !j
Marathi is, in general, a function of ability in correlation appears to dominate the positive se-
the jati and the level of referrals. The cohort lection + ( 0 effect in this case. But this tells us
terms reflect the change in this selection pro- that the positive referral-cohort coefficient that
cess over time. The model, which ignores the we reported above might also be spurious. To
variation in ability ! j across jatis, predicts assess the robustness of the results in column 1,
that + ( 0, , " 0, - " 0 without restrictions, we add jati fixed effects, which subsume * #
and (possibly) - ( 0 with restrictions. +Rj # .!j (Table 7, column 2). The referral
Because the major shift into English school- coefficient + is no longer identified, but the
ing occurred in the 1990s, we first estimate estimated cohort and referral-cohort coefficients
equation (4) for the boys in cohorts 1120. The are very similar to the results in column 1.
1248 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2006

The within-jati estimates allow ability to vary the girls, and Bij is a boy dummy as before. The
across jatis but assume that ability is constant fixed effects, fj , which subsume * # + Rj # .!j ,
over time (both within and across generations). allow for the possibility that ability varies
The level of parental schooling could, however, across jatis. The fixed effects interacted with the
also depend on the access to education, which boy dummy gj ! Bij , which subsume (* & *) Bij #
might have changed over time. If there was (+ & + ) Rj ! Bij , also allow ability to vary by
convergence in the access to education across gender across jatis. Finally, the fixed effects
jatis in the parent generation, that could explain interacted with the cohort variable hj ! Cij , sub-
the positive referral-cohort coefficient in col- sume , Cij # - Rj ! Cij and control for changes
umns 1 to 2 without requiring networks to be in access to schooling for the fathers both across
active. One test to rule out this alternative in- jatis and over time.
terpretation of our result would be to estimate For the special case with , ! 0, - ! 0, as is
the school selectivity regression for girls rather consistent with the model, the estimated coeffi-
than boys; we have already seen that the net- cients in equation (5) should match the cohort
work has no effect on schooling choice for the coefficient and the referral-cohort coefficient
girls, and so both the cohort and the referral- when equation (4) is estimated with jati fixed
cohort effect should be absent. In contrast, if the effects for boys only. Table 7, column 5, sug-
referral-cohort term is picking up convergence gests that this is indeed the case: the cohort-boy
in (fathers) schooling levels across jatis, then coefficient is negative and significant, the refer-
this coefficient should be positive and signifi- ral-cohort-boy coefficient is positive and signif-
cant for the girls as well. icant, and the point estimates are very similar to
Table 7, column 3, reports the basic selectiv- the corresponding coefficients in columns 1 and
ity regression for the girls attending Marathi 2. These results confirm that in the most heavily
schools with cohort, referral-cohort, and refer- networked jatis, high-ability girls were exiting
rals included as determinants of fathers school- to English-medium schools at significantly
ing, while Table 7, column 4, repeats this faster rates than were boys.30 The 0.1 quantile
regression with jati fixed effects. The referral 0.9 quantile of the referrals distribution ranges
coefficient in column 3 is again negative (but from 0.2 to 0.7. The point estimates in column
insignificant), consistent with the lower levels 5 thus suggest that over the period of the 1990s
of ability in high-referral jatis. The cohort co- the gap in fathers schooling between boys and
efficient is positive but insignificant. More im- girls schooled in Marathi grew by 2.3 years in
portantly, the referral-cohort coefficient is small the highest-referral jatis (at the 0.9-quantile
in magnitude and statistically insignificantthe level). In contrast, the ability-differential mea-
point estimate is actually negativeand consis- sured by the difference in the fathers schooling
tent with the results obtained earlier that girls in between boys and girls, declined by as much as
families belonging to high-referral jatis are not five years in the low-referral jatis (at the 0.1-
restricted in their mobility. quantile level) over the same period. This in-
An alternative strategy to control for the con- creasing mismatch in ability levels between the
founding effect of changes in access to school- sexes within jatis and school types could have
ing among the fathers across jatis and over time important implications for the future stability of
pools boys and girls in the selectivity regression, the caste system, which relies on endogamous
which can then be estimated with a full set of jati marriage, as discussed below.
dummies interacted with the cohort variable Columns 6 to 10 of Table 7 report the esti-
mates of the selectivity equations for the first
(5) E$S ij !E ij # 0% # $ , $ , %C ij ! B ij ten cohorts of students, who entered school in

# $- $ - %Rj ! Cij ! Bij

30
The negative cohort-boy coefficient implies that the
# fj & gj ! Bij & hj ! Cij , boy-girl pre-schooling human capital differential is declin-
ing over time, independent of the influence of the male job
network. This may be due to differences in labor force
where , , - are the coefficients on the cohort participation or changes in the returns to English by gender
variable and the referral-cohort interaction for (as in Figures 1 and 2).
VOL. 96 NO. 4 MUNSHI AND ROSENZWEIG: TRADITIONAL INSTITUTIONS MEET THE MODERN WORLD 1249

the 1980s. Schooling choices were stable over dren in our sample are relatively low (6.3 per-
this period and thus we do not expect to find cent of family income in English schools and
changing selectivity effects for the boys or the 6.0 percent in Marathi schools), and, not sur-
girls. As before, the referral coefficient, in col- prisingly, family income has a relatively weak
umn 6 and column 8, is negative and significant, effect on schooling choice for both boys and
reflecting the persistent differences in ability girls. Moreover, the effect of family income on
across jatis. As expected, however, and in con- school choice does not differ by gender at con-
trast to the cohorts making schooling choices in ventional levels of significance (Table 4).
the post-1990s new economy, the cohort effect
and the referral-cohort effect, both uninteracted Differences in Ability.The jati-level fixed
and interacted with the boy dummy, are insig- effects absorb all variation in the jati that is not
nificant in the pre-reform period. gender specific. But in an economy where men
and women historically performed very differ-
F. Alternative Interpretations of the Empirical ent roles, the parental and societal inputs that
Results boys and girls received in childhood might have
been very different. The results reported earlier,
The discussion that follows considers alter- however, provide no evidence of gender distinc-
native explanations for the results we have pre- tions in pre-schooling human capital within
sented. The identifying assumption in the fixed households or jatis. The estimates reported in
effects schooling choice regression is that un- Table 4 do not reject the hypothesis that the
observed determinants of schooling choice that effects of parental human capital characteristics
are correlated with the occupational distribution on school choice are equal for boys and girls.
should not vary by gender, or have a differential The results reported in Table 6 with test perfor-
effect by gender on schooling choice, within the mance as the dependent variable are also con-
jati. Some of the alternative explanations we sistent with the assumption in the fixed effects
pursue are associated with the failure of this schooling regression that pre-schooling human
identifying assumption. Other explanations are capital does not vary by gender within the jati.
generated by relaxing the assumptions of the
model. We will argue that none of these alter- Discrimination.Unless there is a gender-
native explanations matches all the results as based component to caste discrimination, it will
well as our preferred explanation, based on un- be subsumed entirely by the jati fixed effects.
derlying male labor market networks. But it is possible that firms or schools discrim-
inate against boys from working-class back-
Liquidity Constraints.The model assumes grounds, perhaps because they are difficult to
that schooling choices are based entirely on the discipline, while treating girls from different
individuals ability and the historical occupa- backgrounds more equally. The referral-boy co-
tional distribution in his jati, which determines efficient would proxy for underlying discrimi-
the labor market network that he inherits. When nation in that case.
credit markets function imperfectly, liquidity Recall that household characteristics, such as
constraints could, in addition, prevent individuals parental education and family income, had the
belonging to working class jatis from choosing same effect on schooling choice for boys and
more expensive English schooling. Liquidity does girls within the jati. It was only the jati-level
not vary by the gender of the child within the jati, referrals statistic that had a gender-specific ef-
and so the jati fixed-effects regression would fect on schooling. If these results are attributed
appear to rule out this alternative explanation. to discrimination, then it implies that firms or
Boys and girls have different labor market op- schools do not discriminate by family back-
portunities, however, and it is thus conceivable ground within the jati, but by jati affiliation
that liquidity could have a differential effect on alone. It is not obvious why we would expect to
schooling choice by gender. The schooling re- see gender discrimination purely along caste
gression accounts for liquidity constraints by lines. Family characteristics, such as parental
including family income at the time the child education and income, were seen to be corre-
entered school. Schooling expenses for the chil- lated with pre-schooling human capital and are
1250 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2006

at least as easy to observe as caste identity. gression framework. Although school locations
Historically there does not appear to have been and out-migration might vary by jati, there is no
a policy of caste discrimination by employers in reason to expect these decisions to vary by the
any industry in Bombay in any case (Morris, gender of the child within the jati. The third
1965). source of sampling bias is potentially more
problematic, because drop-outs could vary by
Restrictions on High-Caste Women.We fo- gender within the jati. The decision to drop out
cused in Figure 3 and Figure 4 on the absence of would depend on the childs pre-schooling hu-
convergence for the boys, which was attributed man capital and future employment opportuni-
to restrictions on occupational mobility among ties, both of which determine schooling choice.
the lower castes. An alternative interpretation of Selective dropouts, by gender across jatis, could
these figures is that the ability distribution var- consequently violate the identifying assumption
ies across the population such that it remains underlying the fixed-effects estimation procedure.
optimal for individuals to sort by caste into However, the sex-ratio of students in the
different careers, even as the returns to English most recent eight cohorts (grades one through
grow. The convergence among the girls with eight) in which there would be relatively few
this alternative interpretation is attributed to re- dropouts is statistically indistinguishable from
strictions on the high-caste girls. the sex-ratio in the older 12 cohorts. Regres-
There is no evidence that such restrictions are sions not reported here also reveal that the sex-
in place, or have been in place historically. ratio is uncorrelated with the level of referrals in
High-caste women in Bombay have always had the jati, both in the first 12 cohorts and in the 8
higher labor-force participation rates and more most recent cohorts.31
English schooling than lower-caste women, as
observed in Table 2. Within the high castes, V. Conclusion
boys are substantially more likely to be
schooled in English than girls, and so the girls As modernization proceeds around the world,
could easily switch into English schools without there is a perception that indigenous existing in-
creating a mismatch on the marriage market. stitutions may importantly shape the course of the
Moreover, although the pooled schooling development process across different countries.
choice regression with fixed effects cannot dis- Yet little is known about how such institutions
tinguish between the alternative explanation, actually affect the transformation of economies
based on female restrictions, and our view that undergoing change, or their impact on the eco-
male networks shape schooling choice for the nomic mobility of particular groups of individuals.
boys alone, recall that we also ran regressions This paper examines the role of one long-standing
separately for boys and girls. The jati-level sta- traditional institutionthe Indian caste system
tistic, measured either by the proportion of re- in shaping career choices by gender in a rapidly
ferrals or the proportion of fathers with English globalizing economy.
schooling, affects schooling choice for the boys We have found that male working-class
but not for the girls in these regressions. We networks, organized at the level of the sub-
thus appear to be picking up restrictions on caste or jati, continue to channel boys into
mobility that are specific to the boys. traditional occupations, despite the fact that
returns to nontraditional (white-collar) occu-
Sampling Bias.We noted three potential pations have risen substantially during the
sources of sampling bias in Section IIIA. First, post-1990s reform period. In contrast, girls,
particular households might school their chil- who have had historically low labor-market
dren outside the Dadar area. Second, particular participation rates and few network ties to
households might have moved from Dadar over
the past 20 years. Third, children from particu- 31
lar households might have dropped out of As an additional test, we also verified that the sex-ratio
in the most recent cohort that entered school in 2001 is
school. uncorrelated with the level of referrals in the jati. Thus,
The first two sources of sampling bias are there does not appear to be selective enrollment by gender
easily accommodated in the fixed-effects re- and jati either.
VOL. 96 NO. 4 MUNSHI AND ROSENZWEIG: TRADITIONAL INSTITUTIONS MEET THE MODERN WORLD 1251

constrain them, appear to be taking full ad- REFERENCES

vantage of the opportunities that have become
available in the new economy. It is generally Acemoglu, Daron; Johnson, Simon and Robinson,
believed that the benefits of globalization James A. The Colonial Origins of Compar-
have accrued disproportionately to the elites ative Development: An Empirical Investiga-
in developing countries. In this setting we tion. American Economic Review, 2001,
find, instead, that a previously disadvantaged 91(5), pp. 1369 1401.
group (girls) might surpass boys in educational Banerjee, Abhijit and Iyer, Lakshmi. History,
attainment and employment outcomes in the Institutions, and Economic Performance: The
future in the most heavily networked jatis. Legacy of Colonial Land Tenure Systems in
Although we have focused on how traditional India. American Economic Review, 2005,
institutions shape the responses of particular 95(4), pp. 1190 213.
groups of individuals to the new opportunities Burnett-Hurst, Alexander R. Labour and hous-
that accompany globalization, our findings sug- ing in Bombay: A study in the economic
gest that these institutions are likely to be af- conditions of the wage-earning classes of
fected in turn by the forces of change. In our Bombay. London: P. S. King and Son,
framework, an individual schooled in English 1925.
no longer needs the traditional caste network; Chandavarkar, Rajnarayan. The origins of in-
indeed, it has been remarked that the English dustrial capitalism in India: Business strate-
educated form a caste by themselves (M. P. gies and the working classes in Bombay,
Desai, quoted in Julian Dakin et al., 1968 p. 24). 1900 1940. Cambridge: Cambridge Univer-
Simple statistics on marriage and migration that sity Press, 1994.
we computed for the elder siblings of the stu- Cholia, R. P. Dock labourers in Bombay. Bom-
dents in our sample would appear to support bay: Longmans, Green and Co. Ltd., 1941.
the view that English education will ultimately Dakin, Julian; Tiffen, Brian and Widdowson,
undermine the caste network. Among the 825 Henry G. Language in education: The prob-
married siblings in our sample, 11.9 percent lem in Commonwealth Africa and the Indo-
married outside their jati. This contrasts with Pakistan sub-continent. Oxford: Oxford
the parent generation, in which only 3.7 percent University Press, 1968.
of the partners were not members of the same Dandekar, Hemalata C. Men to Bombay, women
jati. Schooling in English appears to be contrib- at home: Urban influence on Sugao Vil-
uting to this increase in inter-caste marriage, as lage, Decan Maharashtra, India, 1942
31.6 percent of the English-educated siblings 1982. Ann Arbor: Center for South and
married outside their jati, versus only 9.7 per- Southeast Asian Studies, University of
cent of the Marathi-educated siblings. And Michigan Press, 1986.
among the 1,073 siblings who are currently DMonte, Darryl. Ripping the fabric: The de-
employed, 13.9 percent of the English-educated cline of Mumbai and its mills. New Delhi:
work outside Maharashtra, versus only 2.1 per- Oxford University Press, 2002.
cent of the Marathi-educated (these differences Gokhale, R. G. The Bombay cotton mill worker.
between the Marathi-educated and the English- Bombay: Millowners Association, 1957.
educated are statistically significant at the 5-per- Gore, M. S. Immigrants and neighborhoods:
cent level). Both marriage outside the jati and Two aspects of life in a metropolitan city.
out-migration weaken caste ties and the caste Bombay: Tata Institute of Social Sciences,
network. Increasing exposure to the modern 1970.
economy through English education, and the Morris, Morris David. The emergence of an in-
mismatch in educational choices and future oc- dustrial labor force in India: A study of the
cupational outcomes between boys and girls in Bombay cotton mills, 1854 1947. Berkeley:
the same jati that we have documented, suggest University of California Press, 1965.
that the forces of modernization could ulti- Munshi, Kaivan. Networks in the Modern
mately lead to the disintegration of a system that Economy: Mexican Migrants in the U.S. La-
has remained firmly in place for thousands of bor Market. Quarterly Journal of Econom-
years. ics, 2003, 118(2), pp. 549 99.
1252 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2006

Munshi, Kaivan and Rosenzweig, Mark. Tradi- Patel, Kunj M. Rural labour in industrial Bom-
tional Institutions Meet the Modern World: bay. Bombay: Popular Prakashan, 1963.
Caste, Gender and Schooling Choice in a Rees, Albert. Information Networks in Labor
Globalizing Economy. Bureau for Research Markets. American Economic Review, 1966
and Economic Analysis of Development, (Papers and Proceedings), 56(2), pp. 559
Working Paper: No. 038, 2003. 66.