traditional institutions meet the modern world: caste, gender, and

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 theforces of globalization to shape the economic mobility and welfare of particulargroups of individuals in the new economy. We explore the role of one suchtraditional institution—the caste system—in shaping career choices by gender inBombay using new survey data on school enrollment and income over the past 20years. We find that male working-class—lower-caste—networks continue to channelboys into local language schools that lead to the traditional occupation, despite thefact that returns to nontraditional white-collar occupations rose substantially in the1990s, suggesting the possibility of a dynamic inefficiency. In contrast, lower-castegirls, who historically had low labor market participation rates and so did not benefitfrom the network, are taking full advantage of the opportunities that became availablein the new economy by switching rapidly to English schools. (JEL I21, J16, O15, Z13)

The collapse of the former Soviet Union,followed by the economic and financial liber-alization of the 1990s, has restructured and“globalized” many economies throughout theworld. One consequence of this restructuring,which has been widely observed, is that somegroups have taken advantage of the new ben-efits afforded by globalization, while othersappear to have been left behind. This paperaddresses the question of whether and howold institutions clash with the forces of glob-alization in shaping the response of particular

groups of individuals to the new economy.Traditional institutions, such as communitynetworks, are generally believed to play animportant role in low-income countries byfacilitating economic activity when marketsfunction imperfectly. Less well understood ishow traditional institutions affect the trans-formation of economies undergoing change,affecting in turn the distribution of benefitsfrom macroeconomic structural reform.

We explore the role of one such traditionalinstitution—the caste system—in shaping careerchoices by gender in a dynamic urban context,using new data on schooling choices and incomecovering the past 20 years in Bombay city, theindustrial and financial center of the Indian econ-omy. Bombay is a useful and important setting inwhich to study the role of institutional rigidities ina dynamic context, as the Bombay labor marketwas historically organized along caste lines, withindividual subcastes or jatis controlling particularoccupational niches over the course of many gen-erations.1 A particularly important feature of thesecaste networks is that they were most active in

* Munshi: Department of Economics, Brown University,Box B/64, Waterman St., Providence, RI 02912 (e-mail:[email protected]); Rosenzweig: Department of Economics,Yale University, 27 Hillhouse Ave., New Haven, CT 06520(e-mail: [email protected]). Leena Abraham andPadma Velaskar of the Tata Institute of Social Sciencescollaborated in the design of the survey instrument, carriedout the survey, and provided many valuable insights. We arealso very grateful to Suma Chitnis for her support andadvice at every stage of the project. We received helpfulcomments from Abhijit Banerjee, David Card, Jan Eeck-hout, Lakshmi Iyer, Duncan Thomas, three anonymous ref-erees, and seminar participants at Columbia University,Harvard University, the Massachusetts Institute of Technol-ogy, IZA, University of Pennsylvania, Princeton University,UCLA, the University of Southern California, WashingtonUniversity, and the World Bank. Research support from theMellon Foundation at the University of Pennsylvania andthe National Science Foundation (grant SES-0431827) isgratefully acknowledged. We are responsible for any errorsthat may remain.

1 Rajnarayan Chandavarkar (1994 pp. 122, 223), forinstance, describes how “[caste] clusters formed within par-ticular trades and occupations ... [this] occupational distri-bution reflected neither [traditional rural] caste vocation northe inheritance of special skills. It was produced partly byexclusionary practices by which social groups, once they

1225

working-class occupations dominated by lowercaste men. Women historically did not participatein Bombay’s labor market and so did not benefitfrom the caste networks, but both men and womenscrupulously adhered to the social rule of endog-amous marriage within the jati.

Although Bombay was a predominantly in-dustrial city for a hundred years beginning inthe last quarter of the nineteenth century, inthe early 1990s the liberalization of the Indianeconomy saw a shift in the city’s economytoward the corporate and financial sectors.We study how members of different jatis, bygender, responded to these changes in thereturns to different occupations, and we willshow that the historical pattern of networkingwithin the jati continues to shape gender-specific, individual responses to these newopportunities in ways that will importantlyaffect the future distributions of incomes, in-dependent of pre-schooling human capital ef-fects or liquidity constraints.2

Our strategy in this paper is to assess howschooling choice, measured by the languageof instruction, varied across jatis, across boysand girls within jatis, and over time. We focuson schooling choice because most adults werealready locked into their occupations whenthe unexpected economic changes occurred.Schooling choice is an important determinant offuture occupational outcomes in the Bombayeconomy and thus reflects the contemporaneousperceptions of expected occupational returns.University education in Bombay is entirely inEnglish, but children choose between Englishand Marathi (the local language) as the lan-guage of instruction at the time they enterschool. Schooling in Marathi channels the childinto working-class jobs, while more expensiveEnglish education significantly increases thelikelihood of obtaining a coveted white-collarjob. If the economic liberalization of the 1990seffectively increased white-collar incomes, and

by extension the returns to English education,then (future) occupational mobility can be iden-tified from changes in the choice of the lan-guage of instruction made by parents of school-age children. Examination of the changingpatterns of schooling choice by jati and genderthus permits an assessment of the interactionsbetween traditional institutions and the new re-alities of globalization.

Our empirical analysis is based on a surveyof 4,900 households belonging to the Maha-rashtrian community residing in Bombay’sDadar area and a survey of the schools in thatlocale that we conducted in 2001–2002. Sec-ondary schools in Bombay run from grade 1to grade 10. The household survey was basedon a stratified random sample of students whoentered 28 of the 29 schools in Dadar (in thefirst grade), over a 20-year period, 1982–2001.3 English is the language of instructionin ten schools in Dadar, while Marathi is thelanguage of instruction in the remaining 18schools.

The survey data suggest that the returns toEnglish education, for given years of school-ing, increased in the 1990s. Based on retro-spective information on the annual earningsof the parents of the sampled children, weestimated the returns to English and the re-turns to years of schooling at five points intime from 1980 through 2000 for workingadults between the age of 30 and 55.4 Figures 1and 2 provide the estimated returns to schoolingattainment and schooling language, for menand women, respectively, in each time period.As can be seen, the returns to years of school-ing increased only mildly over time for bothmen and women. In contrast, the English pre-mium increased sharply from the 1980s to the1990s for both sexes, rising from 15 percentin 1980 to 24 percent in 2000 for men andfrom approximately 0 percent in 1980 to 27percent in 2000 for women. The returns toEnglish for men increase from the mid-1980s,which is most likely due to the decline around

obtained a foothold in a particular occupation, would notadmit an outsider.”

2 A recent literature has shown that historical institutionshave long-run consequences for growth in low-incomecountries (Daron Acemoglu et al., 2001; Abhijit Banerjeeand Lakshmi Iyer, 2005). These empirical findings, how-ever, do not provide insight into the mechanisms underlyingsuch persistence.

3 One school refused to provide us with information on itsstudents and will be ignored in all the discussion that follows.

4 The details of the estimation procedure and the esti-mates of the returns to English and the returns to schooling(with standard errors) are provided in Munshi and Rosenz-weig (2003).

1226 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2006

that time in manufacturing jobs in Bombay(Darryl D’Monte, 2002), but continue to risethrough the 1990s.

The survey collected information on school-ing choice for 20 cohorts of students who en-tered the 28 neighborhood schools (in the firstgrade) over the 1982–2001 period. The times-series data on enrollments in English- and Mar-

athi-medium schools suggest that the changes inthe returns to English significantly affectedschooling choice for both boys and girls in thesample, across castes and over time. Figure 3and Figure 4 display the changing proportionsof students enrolled in English schools for the20 entering cohorts from 1982 (cohort ! 1) to2001 (cohort ! 20) for three caste group-

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

FIGURE 2. RETURNS TO ENGLISH AND SCHOOLING BY YEAR, 1980–2000: WOMEN AGE 30–55

1227VOL. 96 NO. 4 MUNSHI AND ROSENZWEIG: TRADITIONAL INSTITUTIONS MEET THE MODERN WORLD

ings—low, medium, and high—and by gen-der.5 The figures were constructed using theEpanechnikov kernel function to nonparametri-cally regress schooling choice (1 ! Englishmedium; 0 ! Marathi medium) on the cohortvariable for each caste group, taking into ac-count the strong intergenerational state-depen-dence with respect to the language of instructionwithin the family.6 Although jatis define the rel-

evant boundary for the labor-market networksand form the relevant social unit in our analy-sis,7 we aggregate the 59 subcastes in our datafor expositional convenience in these figures.

Figures 3 and 4 show that enrollment rates inEnglish-medium schools have grown substantiallyover time for both boys and girls and for allcastes.8 The trajectory is much steeper, how-ever, for the ten most recent cohorts, who wouldhave entered school in the post-reform 1990s.Thus, the increase in the returns to Englishobserved in Figure 1 and Figure 2 appears tohave shifted schooling choice toward Englisheducation. The figures also indicate substantialdifferences in English schooling between castesat the beginning of the sample period, reflecting inpart the circumstances of the colonial regime. Thehigh castes gained access to clerical and adminis-

5 Children enter first grade at the age of 6 and completetenth grade at the age of 15, so the current age of thestudents in our sample, with only a few exceptions, rangesfrom 6 to 25. Students in Bombay typically do not changethe language of instruction midstream or switch schoolsafter they enter first grade. High castes include all theBrahmin jatis, as well as a few other elite jatis (CKP andPathare Prabhus). Low castes include Scheduled Castes,Scheduled Tribes, and Other Backward Castes, as definedby the government of India. Medium castes are drawnmostly from the cultivator jatis, such as the Marathas andthe Kunbis, as well as other traditional vocations that werenot considered to be ritually impure.

6 If both parents have been schooled in English, it is veryunlikely that the child would be sent to a Marathi school,and all the regressions that we later report will also accountfor such state dependence at the level of the family. Detailsof the nonparametric estimation procedure and parametricestimates of the schooling regression (with standard errors)are provided in Munshi and Rosenzweig (2003).

7 As Morris David Morris (1965 p. 76) emphasizes in hishistorical account of the Bombay labor market, “for anyanalysis of labor recruitment [in Bombay] ... it is entirelyinappropriate to lump into larger groups because of simi-larity of name, function, social status, or region-of-originsubcastes that are not endogamous.”

8 Details of the nonparametric estimation procedure usedto generate these figures are provided in Munshi and Rosen-zweig (2003).

FIGURE 3. ENGLISH SCHOOLING: NET PARENTAL EDUCATION EFFECT—BOYS


trative jobs under the British, while the lowercastes were confined for the most part to working-class jobs. Consistent with the view that Marathieducation channels students into working-classjobs, and that English education increases the like-lihood of obtaining a white-collar job, we see inFigure 3 and Figure 4 that high-caste boys andgirls currently 25 years old (the oldest cohort)were much more likely to have been schooled inEnglish, and that this caste difference in schoolingpersists over the next ten cohorts. But although thecaste gap narrows dramatically for the girls in the1990s, there is no convergence for the boys. Thus,it appears that caste continues to play a role inshaping schooling choices in the new economy ofthe 1990s, but only for boys. The key question iswhy the lower-caste boys seemingly fail to takeadvantage of the new economic opportunities.

The gender-specific explanation for the ob-served pattern in Figure 3 that we pursue inthis paper is based on network externalities.Numerous studies document higher levels ofnetworking in blue-collar occupations, possi-bly because the information and enforcementproblems that give rise to networks are more

acute in those jobs.9 These studies focus onmen, the primary occupants of blue-collarjobs. And among the household heads in Dadar,68 percent of the men in working-class jobsfound employment through a relative or a memberof the community, while the corresponding sta-tistic for white-collar workers was 44 percent.Once the (working-class) network is in place,there is a positive externality associated withparticipation in the network, and hence withthe traditional occupational choice in the jati.This externality could give rise to intergenera-tional occupational persistence at the level ofthe jati, with labor market networks channelingboys into particular (traditionally male) occupa-tions and hence toward particular schoolingchoices.

9 For example, Albert Rees (1966) found that informalsources accounted for 80 percent of all hires in eight blue-collar occupations versus 50 percent of all hires in four white-collar occupations in an early study set in Chicago. Similarly,68 percent of blue-collar workers and 38 percent of white-collar workers reported having received help finding a job inM. S. Gore’s (1970) study of migrants in Bombay.

FIGURE 4. ENGLISH SCHOOLING: NET PARENTAL EDUCATION EFFECT—GIRLS


Once the returns to the white-collar occupationgrow, however, schooling choice must ultimatelyconverge across castes. The explanation for theabsence of convergence in Figure 3 that we putforward in this paper is based on the idea that thecaste networks might place tacit restrictions on theoccupational mobility of their members to pre-serve the integrity of the network. We will showthat although these restrictions might have beenwelfare-enhancing and indeed equalizing whenthey were first put in place, such restrictions couldresult in dynamic inefficiencies when the structureof the economy changes.

The results in this paper provide empirical sup-port for the view that historical occupation pat-terns kept in place by caste-based networkscontinue to shape occupational choice, and henceschooling choice, for the boys in the new econ-omy. In contrast, the lower-caste girls who histor-ically kept away from the labor market, and sohave no network ties to constrain them, take fulladvantage of the opportunities that become avail-able in the new economy. The growing disparitiesin school choices between boys and girls withinthe traditional jatis not only suggest a new balanceof economic opportunities by gender, but alsocould threaten the long-run stability of the castesystem, which is based on endogamous marriageswithin the subcaste. Thus, a complete understand-ing of the development process must not only takeaccount of the initial conditions and the role ofexisting institutions in shaping the response tomodernization and globalization, but must alsoconsider how these traditional institutions areshaped in turn by the forces of change.

I. The Institutional Setting

A. Bombay’s Labor Market

Bombay’s industrial economy in the late nine-teenth century and through the first half of thetwentieth century was characterized by wide fluc-tuations in the demand for labor (Chandavarkar,1994). It is well known that such frequent jobturnover can give rise to labor market networks,particularly when the quality of a freshly hiredworker is difficult to assess and performance-contingent wage contracts cannot be imple-mented. The presence of such recruitmentnetworks has indeed been documented by numer-

ous historians studying Bombay’s economy priorto independence in 1947 (Chandavarkar, 1994;Morris, 1965; Alexander R. Burnett-Hurst, 1925).These networks appear to have been organizedaround the jobber, a foreman who was in chargeof a work gang in the mill, factory, dockyard, orconstruction site, and more importantly also incharge of labor recruitment.

Given the information and enforcement prob-lems associated with the recruitment of short-termlabor, it is not surprising that the “jobber had tolean on social connections outside his workplacesuch as his kinship and neighborhood connec-tions” (Chandavarkar, 1994, p. 107). Here theendogamous subcaste or jati served as a naturalsocial unit from which to recruit labor, becausemarriage ties strengthen information flows andimprove enforcement. This widespread use ofcaste-based networks thus led to a fragmentationof the Bombay labor market along social lines.10

Other studies also suggest that these patternstended to persist over many generations. For ex-ample, Hemalata C. Dandekar (1986) traces theevolution of a network of Jadhavs (a particularsubcaste) belonging to one village in interior Ma-harashtra. In 1942, 67 percent of the Jadhav mi-grants from that village were working in the textilemills and 4 percent in other factories. Thirty-fiveyears later, in 1977, 58 percent were still em-ployed in textile mills, while 10 percent were inother manufacturing industries.

A noticeable feature of historical descriptionsof caste-based networks in Bombay is that theywere restricted to working-class jobs. This is notsurprising, because the information and enforce-ment problems that give rise to such networkstend to be more acute in those occupations. Fur-ther, most studies of caste-based networks inBombay focus on male workers. Women wereconspicuously absent from Bombay’s labor

10 The presence of caste clusters has been historically doc-umented in the mills (R. G. Gokhale, 1957), among dockworkers (R. P. Cholia, 1941), construction workers, and in therailway workshops (Burnett-Hurst, 1925), in the leather anddyeing industries, and in the Bombay Municipal Corporationand the Bombay Electric Supply and Transportation Company(Chandavarkar, 1994). More recently, Kunj M. Patel (1963)surveyed 500 mill workers in the Parel area, close to the site ofour study, in 1961–1962 and found that 81 percent of theworkers had relatives or members of their jati in the textileindustry. Sixty-six percent of the workers got jobs in the millsthrough the influence of their relatives and friends.


force, particularly in the working-class jobs(Morris, 1965). These historical patterns oflabor force participation by gender will laterhelp explain the schooling choice dynamics,for boys and girls, that we saw in Figure 3 andFigure 4.

B. The Schools in Dadar

Our analysis highlights the medium of instruc-tion as the salient feature of schooling choice. It ispossible that the choice of the language of instruc-tion merely proxies for school quality. In parallelwith the household survey, we carried out a sur-vey of schools based on a questionnaire filled outby school principals. This questionnaire elicitedinformation on a variety of school characteristics,

as well as recent student performance on the stan-dardized school leaving examination (common toboth Marathi and English schools), which allowsus to compare the two types of schools as well asthe students across the schools. Table 1, panel A,describes school infrastructure and faculty quali-fications in the English and Marathi schools. Theaverage student-teacher ratio, class size, numberof students per desk, computers per student, andthe proportion of teachers with Bachelor of Edu-cation degrees and higher (postgraduate) degrees,are each very similar and statistically indistin-guishable for the two types of schools.11

11 A regression of the language of instruction on the setof school characteristics in Table 1, panel A, indicates that

TABLE 1—SCHOOL CHARACTERISTICS AND STUDENT PERFORMANCE

School type

English medium Marathi medium

(1) (2)

Panel A. School characteristicsStudent-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 expensesFees 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 (1997–2001)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 thehousehold survey. School characteristics are based on the secondary section (grades 5 through 10). School expenses aremeasured for 2000–2001 in thousands of 1980 Rupees. To convert to 2000 Rupees, multiply by 4.44. Other school expensesinclude transportation, coaching classes, textbooks, uniforms, and stationery. Scores above 35 percent are required to passSSC; 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.


Despite the increase in the demand for En-glish education in the last ten years, as seen inFigures 3 and 4, no new schools were added inthis period in Dadar.12 The English-languageschools accommodated this increased demandby adding divisions in each grade, increasingthe number of desks in each classroom, anddoubling students on each desk. Because thesupply of schools was effectively fixed, wewould expect the English schools to extractsome economic rents from this increased de-mand through higher fees and schooling costs ingeneral. In contrast, fees in the Marathi schoolsare subsidized by the state government. Ourhousehold survey collected information onschool fees and other expenses (transportation,coaching classes, textbooks, uniforms, and sta-tionary) in the last year. Table 1, panel B showsthat school fees (in 1980 Rupees) are currentlysignificantly higher in the English schools (480versus 200 Rupees), as are other expenses(1,100 versus 710 Rupees).

One other difference between the schools is inthe performance of the students on the SecondarySchool Certificate (SSC) school-leaving examina-tion. Table 1, panel C reports student performanceon this exam over a five-year period, 1997–2001.Students in the English schools perform muchbetter on this standardized test in terms of thepercentage that pass and receive a first class and adistinction.13 Although these substantial differ-ences in test performance can be explained bydifferences in school quality, they can also beexplained by differential selection by ability intoEnglish and Marathi schools, as implied by ournetwork model of school choice, and we willprovide evidence supporting this implication ofthe model below.

In addition, our survey provides direct evi-dence that the medium of instruction and itsimplications for children’s future role in soci-

ety, rather than differences in school quality,dominate the schooling choice of parents. Thesurvey elicited from parents the reasons for theirchoice of school for their child. The percentageof parents reporting that the “quality of educa-tion” was a factor in their choice was relativelylow and did not differ substantially across par-ents choosing English-medium schools andMarathi schools—43.7 percent versus 35.2 per-cent, respectively. In contrast, almost 87 percentof parents who chose English as the medium ofinstruction for their child reported that careeropportunities was a factor in choosing thatschool. And over 62 percent of parents whochose a Marathi-language school listed closercommunity ties as a reason.

II. A Simple Model of Schooling Choice

Our first objective in this section is to showhow exogenous, historically determined occu-pational differences across otherwise identicaljatis can persist when network externalities arepresent. Because occupational choice translatesinto schooling choice, this explains the initialcaste gap that we observe for the boys inFigure 3 (the model that we lay out in this sectionapplies to the boys, as we will see later that labormarket networks are most active among the men).We will show, however, that jatis should start toconverge once the returns to English grow suffi-ciently large, which is inconsistent with what weobserved in that figure. Our second objective inthis section will consequently be to show hownetwork externalities could give rise to endoge-nous social restrictions on occupational mobility,and by extension schooling choice, preventingconvergence across social groups in a changingeconomic environment.

A. Population, Community Structure, andMarket Structure

Consider a population with a continuum ofindividuals. Each individual i is endowed with alevel of ability !i ! {0, 1⁄2, 1}. Note that abilityin this section, and throughout the paper, refersto pre-schooling human capital rather than ge-netic ability. He lives for three periods, studyingin the first period and working in the remainingperiods. Schooling choice is restricted to in-struction in English or Marathi, the local lan-

the joint set of characteristics is not significantly differentacross the school types.

12 The average establishment year for the 18 Marathischools is 1947 and the corresponding year for the 10English schools is 1959. All schools in the area have nowbeen operating for many decades.

13 Scores above 35 percent are required to pass, scoresabove 60 percent are required for a first class, and scoresabove 75 percent are required for a distinction. The sametest is administered to all schools, with the questions trans-lated into English and Marathi.


guage. Occupational choice is restricted towhite-collar and working-class jobs. Educationin English is required to obtain a white-collarjob, but is more expensive than Marathi educa-tion, which is assumed for simplicity to becostless. Occupational choice is based on thewage that the individual will receive in thewhite-collar and the working-class job, net ofthe pecuniary cost of schooling. Each individualthen makes his schooling decision based on thetype of job that he (correctly) anticipates he willoccupy in the subsequent period. If he prefers tohold a white-collar job, then he will study inEnglish, if not he will study in Marathi, which isless costly.

Each individual is born into a community orjati. There is a large number of communities inthis economy, and we normalize so that themeasure of individuals in each cohort or gener-ation of a jati is equal to one. To simplify andhighlight the role of network externalities inintergenerational occupational persistence, weassume that the distribution of pre-schoolinghuman capital does not vary over generations oracross jatis.14 Within each jati-generation thereis a measure PL of low types (with ability ! !0) and a measure PM of medium types (withability ! ! 1⁄2).

On the demand side of this labor market,firms operate competitively in both the work-ing-class and the white-collar sectors. We notedearlier that working-class jobs generally tend tobe more heavily networked. For the purpose ofthis simple model, we assume that the white-collar worker’s ability, and hence his produc-tivity, can be observed perfectly and so thewhite-collar wage (net of schooling costs) isspecified to be "!i. Here " represents the returnsto ability in the white-collar job, which in oursetup also reflects the returns to English educa-tion. In contrast, the nature of the productiontechnology prevents working-class firms fromdirectly observing their employees’ ability be-fore they commence work. We take it that thefirm is unable to specify a performance-contingentwage contract, and so will use referrals from itsincumbent workers to hire new employees, gen-erating a role for the network in the working-class

jobs alone. Munshi (2003) provides evidence thatexperienced workers contribute disproportionatelyto labor market networks in the United States, andwe would expect this pattern to hold up in othereconomies as well. In our model, the expectedworking-class wage over the individual’s workinglife is thus specified to be P, the proportion of theprevious generation (three-year-olds) in the jatithat will be employed in the working-class jobwhen he enters the labor force.

B. The Schooling Equilibrium

We now proceed to derive the different oc-cupational distributions, and hence schoolingequilibria, that can be sustained across jatis withthe same ability distribution in this setup. Eachindividual chooses the occupation, and hencethe language of instruction, that maximizes hisnet return. This return depends on his own abil-ity, as well as the proportion of his jati in theprevious generation employed in the working-class occupation, as described above.

Under conditions that we specify below, withthree levels of ability, three distinct schoolingequilibria can be sustained within jatis: (a) onlylow types choose Marathi education; (b) low andmedium types choose Marathi education; (c) ev-eryone in the jati chooses Marathi education.15

CONDITION 1: PL " "/2.

CONDITION 2: "/2 " PL # PM " ".

CONDITION 3: " " 1.

It is easy to verify that once a jati is exog-enously assigned a particular occupational distri-bution, this distribution will persist unchangedover many generations when the conditions aboveare satisfied.16 This intergenerational state depen-dence is a consequence of the network externalityassociated with the working-class occupation.It implies, in turn, that the probability that any

14 We will relax this assumption in the empirical work byallowing for heterogeneity in ability across jatis.

15 Munshi and Rosenzweig (2003) consider the generalcase with N types and N equilibria, without altering theresults that we present below.

16 It is merely necessary to show that no individualwishes to deviate from the occupation, and hence schoolingchoice, assigned to his type in his jati in the previousgeneration, for each of the schooling equilibria.


individual i drawn randomly from jati j will beschooled in English (Eij ! 1) is related to theproportion of men in the previous generation em-ployed in the working-class job, Pj:

(1) Pr$Eij # 1% # 1 $ Pj .

This expression will serve as the starting pointfor the empirical analysis described in SectionIV, where we will examine the relationship be-tween schooling choice in the current genera-tion and the occupational distribution in theprevious generation, to identify the presence ofan underlying network organized around thejati.

C. Schooling Choice as the Returns toEnglish Grow

The state dependence at the level of the jatiderived above is obtained under the assumptionthat the parameters of the model, PL, PM, ",remain stable over time. To explore the effect ofthe increase in the returns to English (") in the1990s, we now allow for multiple cohorts ofunit measure within each generation.

If " remains constant within a generation, theresults derived above follow through withoutmodification for all cohorts. If, however, " in-creases across successive cohorts, holding Pjconstant, then schooling choice within a jaticould change over the course of a single gener-ation. When " just crosses one, high-abilityboys belonging to jatis that were traditionally inequilibrium 3 switch to English. When " sub-sequently reaches 2(PL # PM), medium-abilityboys in jatis that were traditionally in equilib-rium 2 or equilibrium 3 switch to English, atwhich point schooling choice across all jatiswill converge.

Although the network externality describedabove can explain the persistence of traditionaloccupational patterns within the jati over manygenerations, and hence the initial caste gap ob-served in Figure 3, it cannot by itself explain theabsence of convergence over the 1990s as thereturns to English grew. To explain this absenceof convergence, we consider the possibility thatheavily networked (working-class) jatis mighthave put restrictions on occupational mobility,

and hence schooling choice, in place to preservethe viability of the community network.17

To understand why restrictions on mobilitymight emerge, define a social welfare functionthat places equal weight on all members of thejati. Now the welfare in a jati situated in equi-librium 3, in which everyone studies Marathi, issimply the unweighted average of all the pay-offs from the working-class occupation, W ! 1.When " just crosses one, in a given cohort, allhigh types in the jati can expect to earn more inthe white-collar sector than in the jati’s “tradi-tional” working-class occupation and will thusswitch to English schooling. Welfare from thatcohort onward is then W ! (PL # PM)2 # (1 &PL & PM). The new welfare level is a weightedaverage of PL # PM " 1 and 1, and so jati-levelwelfare must unambiguously decline whenschooling choice, and hence the occupationaldistribution, shifts. Historically there was in-tense competition for scarce working-class jobsin Bombay, as noted in Section I. Becauselarger numbers improve the jati’s competitive-ness, and increase the working class wage ingeneral, it is easy to see why social restrictionson occupational mobility could emerge endog-enously. Moreover, the fact that the lower-castegirls in our sample do not display a similarresistance to change can be attributed to thegender-specific nature of these job networks.

Social restrictions on occupational mobilitycan be welfare-enhancing for small and mediumchanges in ", as noted above. But they couldgive rise to substantial inefficiencies if theycontinue to persist when " grows large. Forexample, it is easy to verify that the socialrestrictions described above for equilibrium 3will be inefficient once " reaches 1 # (PL #PM), although a welfare calculation that identi-fies the presence of such a dynamic inefficiencyis beyond the scope of this paper.

17 Restrictions on mobility do not have to be associatedwith explicit punishment. Preferences for schooling or fu-ture career choices could be determined endogenously, forexample, by placing symbolic value on the traditional oc-cupation in the jati. Social interactions within the jati couldalso lead individuals to make similar schooling choices andcareer choices across generations. We do not attempt todistinguish between preferences that are complementary tothe network, and the network itself, in this paper.


While we conjecture that restrictions on oc-cupational mobility might be in place in theheavily networked jatis, no direct evidence oftheir presence in Bombay is available. We can,however, test one important implication that isconsistent with the presence of these restric-tions: the relationship between schooling choiceEij and the occupational distribution within thejati in the previous generation Pj must notweaken over successive cohorts in the currentgeneration, even as the returns to English grow.This stability in intergenerational state depen-dence would then explain the wedge betweenhigh-caste and lower-caste schooling choicesfor boys that was observed through the 1990s inFigure 3.

D. Selection into Schools

The model of schooling choice as laid out inthis section also has implications for selection,by ability, into English and Marathi schools.Within any jati, the average pre-schooling hu-man capital of the English students must begreater than that of the Marathi students. Takingthe average across all jatis, this implies thataverage ability must be greater among the En-glish students at any point in time. This obser-vation is consistent with the significantly highertest scores obtained by students in the Englishschools (Table 1), despite the fact that Englishand Marathi schools appear to be similar interms of the resources available per student andthe qualifications of the teachers. But how doesthe ability distribution within the English andMarathi schools change across successive co-horts in the current generation as the returns toEnglish grow? Without social restrictions, devi-ation to English education is ordered by ability,so as " grows there is a steadily worsening poolof Marathi students. Jatis that begin with agreater proportion of their members in working-class jobs have higher ability among the Mar-athi students, but their shift into English, andhence the decline in ability, must also be morerapid, because all jatis ultimately converge.With social restrictions, heavily networked jatiscontinue to begin with a superior ability distri-bution within Marathi schools, but now theremight be no convergence in ability among Mar-athi students across jatis.

Average ability among the English studentsis greater than average ability among the Mar-athi students at any point in time, but among theMarathi students it is the group with the highestability that deviate as " grows. Thus, while thequality of the Marathi students unambiguouslydeclines over time as the returns to Englishincrease, the change in the quality of the pool ofEnglish students is ambiguous.18

III. The Household Data

A. The Survey

To examine empirically the role of caste net-works in shaping mobility during a period ofchange, we carried out a household surveybased on a random sample of students, stratifiedby caste, who entered the 28 secondary schoolsin Dadar (in the first grade) over a 20-yearperiod, 1982–2001. This design provides infor-mation for the periods before and after the ma-jor Indian economic reforms. We obtained acomplete list of all students enrolled in grades 1to 10 in 2001 (the year of the survey), as well asa list of students who were enrolled in grade 10from 1991 to 2000. Ignoring dropouts, thisleaves us with 20 cohorts of students who en-tered school over the 1982–2001 period. A totalof 101,567 students were enrolled in the schoolsin 2001 or studied in grade 10 over the previousten years. We drew the roll numbers of 20,596students randomly from these 20 cohorts, andrecovered their names and addresses from theschool records. Restricting attention to Maha-rashtrians residing in Dadar and the immedi-ately adjacent neighborhoods, we were left with8,092 eligible students to serve as the samplingframe for the survey. The student’s name is

18 For example, in the three-type case with no socialrestrictions, some jatis (in equilibrium 2) have only high-ability children in English schools, while other jatis (inequilibrium 1) have both medium- and high-ability childrenin English schools to begin with. This implies that thequality of the English pool must improve when " reachesone, because only the high types from jatis in equilibrium 3deviate at that point. But average ability drops below itsinitial level when " reaches 2(PL # PM), because mediumand high types in all jatis will have switched into Englishschools by that point. With social restrictions, the change inability within the English schools becomes even more dif-ficult to characterize.


typically a good indicator of the caste, and wewanted close to 1,000 upper castes in the sam-ple, so all 1,082 students from this populationwho appeared to be upper castes were selectedfor the survey. We drew randomly from theremaining students in the sampling frame untilthe target sample size was reached. The uppercastes account for 17.5 percent of the final sam-ple of 4,945 observations, which is slightlyhigher than the 13.4 percent that we began within the sampling frame.

The research team interviewed the parents ofthe selected students at their residences. Thesurvey instrument elicited detailed subcaste in-formation from the respondents and includedsections on grandparents’ education and occu-pation, parents’ education and occupational andincome histories (at five-year intervals from1980 to 2000), as well as the student’s andsiblings’ subsequent education, occupation, in-come, and marriage outcomes (where rele-vant).19 Information on transfers, assistance infinding jobs, and ties to the community was alsocollected.

Of the eligible households, 82.5 percentprovided completed schedules. This is a rel-atively high response rate, especially giventhat some of our addresses were 20 years old.But we might still have obtained a selectivesample of households, for a number of differ-ent reasons. First, households residing in Da-dar who sent their children to study outsidethe area would be missing from the sample.Second, households who moved out of thearea would be among the 17.5 percent of therespondents who did not complete the survey.And third, students from the first ten cohortswho did not reach the tenth grade, and currentstudents who have dropped out, would bemissing from the sample. In Section V, wewill discuss how our identification strategy is

unlikely to be undermined by these potentialsources of bias.

B. Descriptive Statistics: Caste, OccupationalNetworks and Schooling

The data provide empirical support for threefeatures of the model of schooling choice laidout in Section II. First, the occupational distri-bution, a product of historical circumstances,varies by caste, and persists across generations,particularly among the men. Second, working-class jobs are associated with a higher level ofreferrals (networking). And third, working-classjobs are associated with lower levels of Englishschooling.

The survey elicited information on parentaloccupations at five-year intervals from 1980 to2000. For the grandparents, we simply asked forthe main occupation over the individual’s work-ing life. The 90 occupations in the data weredivided by roughly increasing levels of humancapital into seven aggregate categories: un-skilled manual, skilled manual, organized blue-collar, petty trade, clerical, business, andprofessional. We further classified unskilledmanual, skilled manual, and organized blue-collar as working-class occupations. Clerical,business, and professional were classified aswhite-collar occupations. Petty trade is treatedas an intermediate unclassified occupation.

Table 2, panel A, describes the occupationaldistribution across broad caste categories (low,medium, high), separately for the employed fa-thers, based on information in 1995, and the pa-ternal grandfathers of the students in the sample.Columns 1 to 3 of the panel indicate that lower-caste fathers are much more likely to be employedin working-class occupations (54 percent and 43percent) as compared with high-caste fathers (18percent).20 The same cross-caste pattern is ob-tained for individual occupations within the work-ing-class and white-collar classifications, with theexception of clerical jobs. The comparison of thefathers in columns 1 to 3 with the grandfathers incolumns 4 to 6 also indicates that there has beenlittle change in the basic occupational distribution,as well as the percentage of working-class

19 The name is usually a good indicator of the individ-ual’s community and caste. For example, 98.7 percent of therespondents, whom we had selected on the basis of theirnames from the school records, said that Marathi was theirmother tongue, indicating that they were indeed Maharash-trian. The caste classification is potentially more problem-atic, however, because lower castes could in some caseschange their names or misreport their caste affiliation. Notethat such misreporting will not undermine the fixed effectsestimation strategy, described below, as long as it does notvary by the gender of the child, within the jati.

20 Note that we use only working-class and white-collaroccupations when computing this statistic.


TABLE 2—OCCUPATION, 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 grandfathersEmployment (%) 97.37 97.31 99.06 98.87 98.86 99.28Occupational distribution (%)

Unskilled manual 11.09 7.84 4.41 9.00 3.63 2.10Skilled manual 17.35 13.70 10.21 11.67 6.72 8.42Organized blue-collar 22.87 19.22 2.90 22.89 24.23 7.67Petty trade 4.00 4.51 2.52 3.11 3.20 3.34Clerical 28.09 36.64 20.81 22.22 23.79 28.84Business 7.95 8.79 15.51 6.11 4.72 13.00Professional 8.30 8.79 43.51 5.56 6.18 33.66Farming 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 grandmothersEmployment (%) 20.56 20.01 51.23 19.31 18.59 15.57Occupational distribution (%)

Unskilled manual 29.95 16.94 2.36 24.65 7.18 3.13Skilled manual 8.82 8.47 6.15 1.70 1.44 3.13Organized blue-collar 4.01 4.92 0.47 8.50 4.31 0.78Petty trade 3.74 3.83 1.18 1.13 0.57 0.00Clerical 31.55 40.71 46.34 4.25 2.30 19.53Business 4.55 2.46 3.78 2.27 1.44 3.91Professional 17.38 22.68 39.72 5.10 8.62 67.97Farming 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 andmothers is computed as of 1995. Statistics in columns 4–6 are reported for paternal grandfathers and maternal grandmothers. Workingclass ! 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 incomeis 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, assistantwatchman, 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 andTransportation (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.


jobs, across the generations within broad castecategories.21

Although most men are employed, we see thatlabor force participation (which includes part-timework) for the women in Table 2, panel B, isrelatively low but is growing. Only 15 percent ofhigh-caste grandmothers worked; whereas justover half of high-caste mothers entered the laborforce (based on their employment status in 1995).Among the lower castes, the percentage employedremains stable at 20 percent across the genera-tions, but notice that farming is listed as the pri-mary occupation for a large number of workinggrandmothers. This suggests that urban employ-ment must have increased sharply for the lower-caste women as well.

The occupational distribution across castes forthe mothers in panel B, columns 1 to 3, displays apattern similar to that for the fathers. Lower-castewomen are much more likely to be employed inworking-class occupations (44 percent and 32 per-cent) as compared with high-caste women (9 per-cent). There is an important difference, however,between men and women—although the large dif-ference within the working-class occupations forthe men was in access to blue-collar jobs for thelower castes, for women the major difference is inaccess to unskilled manual jobs; many of thelower-caste women work as sweepers and domes-tic servants.

Columns 1 to 3 and columns 4 to 6 in panelB suggest that there has been, in contrast to themen, significant intergenerational change in oc-cupational patterns for women within castes.The urban occupations that show the greatestincrease are skilled manual, clerical, and pro-fessional (with the exception of the highcastes).22 The decline in the percentage ofworking-class jobs among the lower-castewomen, across a single generation, is particu-larly dramatic. This contrasts with the stability

of the occupational distribution for the men, forall castes, that we noted earlier, consistent withthe view that labor networks are weak amongthe women.

Together with the occupational distribution,Table 2 reports the mean years of schooling andmonthly income separately by caste for men andwomen.23 As expected, high-caste mothers andfathers have significantly more years of school-ing and significantly higher incomes. Althoughthe model in Section II assumes that the distri-bution of pre-schooling human capital is thesame across castes (jatis), children in a wealthy,educated jati that has had access to white-collarjobs for many generations will be nurtured verydifferently from children in a jati that was his-torically confined to manual jobs. This suggeststhat pre-schooling human capital could vary inpractice across broad caste categories, andacross jatis, as well. When estimating the effectof the historical occupational distribution on thechild’s schooling choice, we will consequentlytake account of the possibility that the occupa-tional distribution could be correlated with theability distribution in the jati.

Table 3 indicates that, as assumed in themodel, working-class occupations are associ-ated with higher levels of networking (refer-rals).24 Column 1 shows that 68 percent of theworking-class men received help from a relativeor member of the community in finding theirfirst job (or starting their first business ifself-employed), which is significantly higher

21 The exception is farming, which is listed as the pri-mary occupation for a large proportion of lower-castegrandfathers. This implies, in turn, that roughly one-quarterof the lower-caste fathers are first-generation migrants. Mi-grants are by definition newcomers in the labor market, andso will be more susceptible to the information problems thatgenerate a need for the caste networks.

22 The decline in the proportion of high-caste women inprofessional jobs is most likely because only the highest-ability women of the older generation (grandmothers) en-tered the labor force.

23 Recall that income information was collected fromeach parent at five equal points in time from 1980 to 2000.We use the income (in 1980 Rupees) that coincides asclosely as possible with the year in which the child enteredschool. Thus, the income in 2000 is used for students age 6to 10, the income in 1995 for students 11 to 15, the incomein 1990 for students 16 to 20, and the income in 1985 forstudents 21 to 25. The same income statistic is used later inthe schooling regressions.

24 The parents of the selected students were asked howthey learned about their first job: through a childhood friend,through a college friend, through a relative, through a mem-ber of the community (jati), or by some other means (whichwas left open-ended in the questionnaire). This open-endedcategory included cases in which no help was received, or inwhich the job was found through newspaper advertise-ments, campus interviews, and other impersonal informa-tion channels. A binary referral variable was thenconstructed, taking the value of one if the parent learnedabout the first job from a relative or member of the com-munity, and zero otherwise.


than the 44 percent of men in the white-collarjobs who received a referral. The correspond-ing statistics for the women in column 3 re-veal essentially the same pattern, although thelevel of referrals for the women is generallylower than that for the men, perhaps because thenetworks for female jobs are less developed.

The model also assumes that Marathi school-ing channels the student into a working-classjob, while English schooling leads to the white-collar occupation. The survey elicited informa-tion on the language of instruction (Englishversus Marathi) for fathers and mothers, in sec-ondary school. Columns 2 and 4 of Table 3 showthat there is a clear distinction between working-class and white-collar jobs with respect to thelanguage of instruction in secondary school. Thepercentage of men in working-class jobs that at-tended secondary school in English is just over 1percent, compared with the 6 percent of men inwhite-collar jobs. In column 4, a similar pattern isobtained for the women. We have described therelationship between the broad occupational cate-gories (working-class versus white-collar), thelevel of referrals, and English schooling. But in-spection of Table 3 indicates that the level ofreferrals and English schooling vary systemati-cally within these categories as well. Later, wewill take advantage of this finer relationship be-

tween particular occupations, the level of referrals,and English schooling, to characterize the occu-pational distribution in the jati.

Table 2 suggests that lower-caste men andwomen are much more likely to hold work-ing-class jobs. Combining these cross-castepatterns with the results in Table 3, it is notsurprising that a much higher proportion oflower-caste men received referrals (60 per-cent versus 37 percent), and that these menare also much less likely to have beenschooled in English (2 percent versus 12 per-cent). In contrast, although lower-castewomen are also much less likely to beschooled in English, the level of referrals isstatistically indistinguishable across castes.25

The level of referrals is low in any case (13percent for the lower castes and 19 percentfor the high castes), especially when com-pared with the corresponding level for themen, and we will later establish that labor

25 Although we noted earlier that lower-caste womenwho work are more likely to hold working-class jobs, whichare associated with more referrals, we also saw that lower-caste women are less likely to enter the labor force. Thesetwo opposing effects appear to cancel each other, leavinglittle variation in the level of referrals across castes for thewomen.

TABLE 3—REFERRALS AND SCHOOLING BY OCCUPATION

Relationship to student Father Mother

Outcomes and choicesPercentage that

received referralsPercentage that

studied in EnglishPercentage that

received referralsPercentage that

studied in English

(1) (2) (3) (4)

OccupationUnskilled manual 65.95 0.80 61.29 0.00Skilled manual 60.13 2.24 45.56 5.56Organized 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 2.94Clerical 47.41 2.89 30.56 7.26Business 49.29 8.53 41.86 9.30Professional 32.77 11.38 29.25 14.47All 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 areferral if a relative or member of the community found him/her a job. A parent is said to have studied in English if he/shestudied in that language in secondary school. Occupational categories are defined in Table 2.


market networks are effectively available forthe men only.

IV. Empirical Analysis

A. Specification and Identification

The first implication of the model is that theoccupational distribution in the jati shouldpersist across generations when networks areactive. Because schooling choice maps intooccupational choice, equation (1) in SectionII expressed this implication in terms ofschooling choice in the current generation andthe occupational distribution in the previousgeneration:

Pr$Eij # 1% # 1 $ Pj .

Recall that Eij ! 1 if individual i belonging tojati j is schooled in English; Eij ! 0 if he isschooled in Marathi; and Pj is the proportion ofmen in the jati in the previous generation whoare employed in working-class jobs and so in aposition to provide referrals.

The particular relationship between Eij and Pjin the equation above is, of course, a conse-quence of the modelling assumption thatschooling choice maps perfectly into futureoccupational outcomes. More generally, wewould expect to see a negative coefficient, butnot necessarily with magnitude one, on Pj. Themodel laid out in Section II also does not allowfor intergenerational state dependence inschooling choice at the level of the household.Moreover, we noted above that pre-schoolinghuman capital and family incomes appeared tovary systematically across castes with differentoccupational backgrounds. The schooling re-gression that we estimate is consequently spec-ified as

(2) Pr$Eij # 1% # %Pj & Xij' & !j ,

where Xij includes the parents’ language ofschooling to reflect household-level state de-pendence in schooling choice, as well as acohort variable to capture the increase in thereturns to English over successive cohorts inthe current generation; !j measures unob-served or imperfectly observed pre-school hu-man capital and family income in the jati,

which could independently determine school-ing choice.

Pj measures the proportion of men in the pre-vious generation employed in working-class jobs.Although the model assumes that only two typesof jobs—working-class and white-collar—areavailable, as many as 90 occupations are listedin the data. A relatively strong relationship be-tween the level of referrals and the type ofoccupation was observed earlier in Table 3, andso one convenient statistic that accurately andparsimoniously describes the occupational dis-tribution in the jati would be the proportion offathers (the previous generation) who received ajob referral. Working-class jobs were also asso-ciated with lower levels of English schooling(Table 3). An alternative measure of the occu-pational distribution in the previous generationwould compute the proportion of fathers whoattended English secondary schools. Most of theregressions reported in this paper will use thereferrals statistic to measure Pj; English school-ing levels were generally low in the previousgeneration and so there is substantially morevariation in the referrals statistic across jatis.We will, however, verify that the results hold upwith the English-schooling statistic as well.

Following the discussion above, we expect tofind % " 0 when networks are active and theoccupational distribution persists across gener-ations. Recall that % must also remain stableacross cohorts in the current generation to ex-plain the absence of convergence in Figure 3.Although much of the analysis treats % as con-stant, we will later verify that % does indeed re-main stable across cohorts.

An identification problem arises when Pj and!j are correlated in equation (2). Although jatismight have been the same to begin with, wenoted in the previous section that their membersnow have very different characteristics (incomeand education), depending on the type of occu-pation that the jati has historically been engagedin. A traditionally working-class jati could thusbe associated with high Pj and low !j, in whichcase a family effect would be erroneously in-terpreted as a network effect because individu-als with lower family resources independentlyselect into Marathi schools.

Our solution to this identification problemexploits the fact, documented in Table 2, thatnetworks are concentrated in working-class jobs


dominated by men. We mentioned earlier thatthe levels of referrals for women were relativelylow, consistent with the significant change inthe occupational distribution across generationsfor women indicated in Table 2. Thus, althoughthe networks might affect schooling choice forthe boys, they should have had little or noimpact on the girls. The model in Section II thenapplies to boys only. Instead of using variationin the level of referrals across jatis to identifythe presence of networks, as in equation (2), weproceed instead to exploit this gender differencein the access to job networks by pooling bothsexes in the schooling regression to identify thepresence of the network within the jati:

(3) Pr$Eij # 1% # $% $ %̃%Pj ! Bij & Xij'̃

& Xij ! Bij$' $ '̃% & (Bij & fj ,

where %̃, '̃ represent the effect of the networkand parents’ language of schooling on the girls.Bij is a dummy variable that takes a value of onefor boys and zero for girls. The advantage ofpooling the boys and girls is that the schoolingregression can be estimated with jati fixed ef-fects, fj ' %̃Pj # !j. Although we can no longeridentify % directly, we can obtain a consistentestimate of % & %̃, the coefficient on the Pj ! Bijinteraction term. For the special case with ex-clusively male networks, %̃ ! 0 and the coeffi-cient on the interaction term identifies network-based occupational persistence for the boysdirectly. More generally, the coefficient on theinteraction term provides a conservative esti-mate of the effect of caste-based networks onschooling choices for the boys.

The identifying assumption in this estimationstrategy is that no variable )j ! Bij appears in theresidual of equation (3), where )j is correlatedwith Pj. A sufficient condition for this identifyingassumption to be satisfied is that no unobserveddeterminant of schooling choice should vary bygender or have a differential effect on schoolingchoice by gender, within the jati. Later in SectionV we will discuss alternative explanations for thenegative and significant % & %̃ coefficient weobtain in the schooling regression. These explana-tions either relax the assumptions of the model,made earlier in Section II, or build on the failure ofthe identifying assumption. We will argue that

none of these explanations fits the data quite aswell as the male labor market network explanationwe put forward in this paper.

B. Caste-Based Networks and SchoolingChoice

Table 4, column 1, reports the estimates ofthe schooling choice regression, equation (2),for the boys. As noted, the sample covers 20cohorts of students age 6 to 25, who enteredschool between 1982 (cohort ! 1) and 2001(cohort ! 20). The student’s cohort (1 to 20),the proportion of fathers in his jati who receiveda referral, and the father’s and the mother’slanguage of instruction in secondary school areincluded as regressors.

The cohort term is included in this regressionto account for the increase in the returns toEnglish over time. While the linear cohort effectwe specify in Table 4 is clearly restrictive, weverify below that the estimated referral coeffi-cient is unchanged when we allow for moreflexible cohort effects. The referral coefficientis also specified to be constant over time inTable 4, and we will subsequently relax thisrestriction as well. For now, we see that thereferral coefficient is negative and significant;children belonging to (historically) working-class and more heavily networked jatis are lesslikely to be schooled in English, consistent withthe first implication of the model. The cohorteffect is positive and significant, implying ashift into English over time, which is consistentwith the increase in the returns to English wesaw in Figure 1. Finally, the results imply that aboy is much more likely to be schooled inEnglish if his parents were educated in thatlanguage, indicating significant state depen-dence in schooling choice at the level of thehousehold.

Table 4, column 2, reports estimates from aspecification that includes variables that deter-mine the student’s pre-schooling human capitalas well as the household budget constraint,which could independently determine schoolingchoices. The parents’ years of education, condi-tional on their language of instruction in sec-ondary school and the level of referrals in thejati, are likely significant determinants of chil-dren’s pre-schooling human capital. The fami-ly’s access to own resources is measured by the


total income of the father and the mother at thetime when the child entered school.26 Inclusionof these variables results in a substantial declinein the referral coefficient, suggesting that thelevel of referrals was previously proxying tosome extent for unobserved, family-specific de-terminants of schooling choice, but it remainsnegative and significant. The coefficient on the

cohort variable is quite stable. And the coeffi-cients on the additional regressors all have sen-sible signs; the boy is more likely to be schooledin a more-expensive English-medium school ifhis father or mother are more educated, or if thefamily is wealthier.

The estimates for girls are reported in col-umns 3 and 4 in Table 4. Column 3 reports theestimates based on equation (2); column 4 re-ports the estimates from the augmented speci-fication that adds the parents’ years of schoolingand family income as additional regressors. Theestimated cohort effects, and the coefficients on

26 We use the income in 2000 for students currently age 6to 10, the income in 1995 for students 11 to 15, the income in1990 for students 16 to 20, and the income in 1985 for students21 to 25. All incomes are computed in 1980 Rupees.

TABLE 4—CASTE-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)

Father’s years of education — 0.023 — 0.020 — 0.021(0.004) (0.003) (0.003)

Mother’s 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)

Father’s years of education - boy — — — — — 0.002(0.005)

Mother’s 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 0.299Number 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 inthe jati who received a referral. Boy ! 1 if the student is a boy, 0 if girl. Family income is measured in thousands of 1980 Rupeesin the year that is closest to the year in which the child entered school. Columns 1–2: 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.


parents’ language of instruction, parents’ edu-cation, and family income are similar to thosefor boys in columns 1 and 2. The referral coef-ficient, however, becomes negligible for thegirls in column 4 once the observed determi-nants of pre-schooling human capital and accessto own family resources are included. One ex-planation for this result is that girls receive helpfrom the women, not the men, in their jati. Butwe noted that the level of referrals for thewomen is very low, across all castes. Althoughnot reported, we also found no correlation be-tween referrals and schooling choice, for bothboys and girls, when we replaced the level ofreferrals for the fathers with the level of refer-rals for the mothers.

The results we have just described are con-sistent with the view that caste-based networks,net of individual and family characteristics, af-fect schooling decisions for the boys, but not forthe girls. But up to this point, we have con-trolled only for unobserved ability with a lim-ited number of family characteristics. A morerobust identification strategy estimates theschooling regression with jati fixed effects, as inequation (3). These estimates are reported incolumn 5 of Table 4. As noted, only the referral-boy interaction coefficient, and not the linearreferral coefficient, can now be identified. Thecoefficient on this term is negative and signifi-cant, and very similar to the referral coefficientfor the boys in column 2. Recall from equation(3) that the coefficient on the referral-boy inter-action term provides us with a direct estimate ofthe referral coefficient for the boys if the referralcoefficient for the girls is zero. The result weobtained earlier for the girls, in column 4, sug-gests that this might well be the case.

The regression specification with jati fixedeffects in Table 4, column 5, did not includefamily characteristics. Column 6 includes par-ents’ language of schooling, parents’ years ofschooling and family income, both interactedand uninteracted with the boy dummy, as addi-tional regressors. Including uninteracted familycharacteristics in the schooling regression hasno effect on the estimated referral-boy coeffi-cient by construction, once jati fixed effects areincluded. But we see that the inclusion of thefamily characteristics, interacted with the boydummy, has no effect on the estimated referral-boy coefficient as well. Indeed, this coefficient

is no smaller than the corresponding coefficientestimated earlier in column 5 without anyhousehold characteristics. This stability con-trasts with the decline in the referral coefficientin Table 4, columns 1 to 4, when family char-acteristics were included, providing some sup-port for the view that the jati fixed effectsabsorb much of the unobserved heterogeneity inthis environment.27

Notice also that parents’ years of schoolingand family income, which had a strong influ-ence on schooling choice for both boys and girlsin columns 1 to 4, do not differentially affectschooling choice by gender (column 6). It isonly the jati-level referral variable that has sucha differential effect on schooling choice, asmeasured by the negative and significant refer-ral-boy coefficient. This observation will beuseful later in Section V when we consideralternative explanations for the results pre-sented in this paper.

C. Schooling Choice over Time

The second implication of the model laid outin Section II is that the relationship betweenschooling choice and the occupational distribu-tion in the previous generation will weakenacross successive cohorts in the current gener-ation as the returns to English grow, unlessrestrictions on occupational mobility are inplace. To assess empirically the stability of thereferral coefficient, we create 4 cohort catego-ries that evenly divide the 20 cohorts, and thenestimate the referral coefficient separately foreach category.

We begin with a benchmark jati-fixed-effectsregression, which maintains a constant referralcoefficient but relaxes the restriction imposed

27 A previous version of the paper (Munshi and Rosen-zweig, 2003) reported a number of additional robustnesstests. First, we accounted for occupational persistence at thelevel of the family by including a full set of (90) dummiesfor the student’s father’s occupation. Second, we allowedfor the possibility that the scope of the network was deter-mined by caste and the region of origin (within Maharash-tra) by replacing the jati by the jati-region as the boundaryof the network. Third, we dropped very large jati-regions(more than 250 observations) and very small jati-regions(fewer than 10 observations). The estimated referral-boycoefficient with these alternative specifications was shownto be very similar to what we report in Table 4.


thus far that cohort effects are linear, by includ-ing the cohort categories in Table 5, column 1.The estimated negative referral-boy coefficientis unaffected by the inclusion of the flexiblecohort effect and remains very similar to theresults shown in Table 4. Inclusion of the familybackground variables, uninteracted and inter-acted with the boy dummy, as additional regres-sors again has no effect on the estimated referralcoefficient (column 2).

Table 5, column 3, allows for changes in thereferral coefficient across cohort categories. Allthe referral-boy-cohort coefficients are negativeand significant except for the coefficient on thefirst cohort category, which is slightly less pre-

cisely estimated. The referral coefficient is ac-tually increasing for the later cohorts, and wecan easily reject the convergence hypothesiswhich implies a decline in the referral effectover time. Once more, the estimated referralcoefficients are robust to the inclusion of thefamily background variables as regressors (col-umn 4). Although not reported here, the referralcoefficient remained stable when the schoolingregression was estimated with boys only, in-cluding parents’ years of education and familyincome as additional regressors. It is this jati-level effect that presumably sustains the gap inschooling choice between broad caste catego-ries observed for the boys in Figure 3.

TABLE 5—SCHOOLING CHOICE OVER TIME

Dependent variable English schooling

Additional regressors

Withoutfamily

characteristicsWith family

characteristics

Withoutfamily

characteristicsWith family

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)




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 0.301Number 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. Englishschooling ! 1 if the child is sent to an English school, 0 if the child is sent to a Marathi school. Referrals measures theproportion of fathers in the jati who received a referral. Boy ! 1 if the student is a boy, 0 if girl. Cohort 1: age 21–25; Cohort2: age 16–20; Cohort 3: age 11–15; Cohort 4: age 6–10. Column 2 and column 4 include family characteristics, separatelyand interacted with the boy dummy. Family characteristics include parents’ language of schooling and years of education, andtotal family income. A full set of jati dummies is included in all regressions. Sample includes boys and girls.


D. Robustness and Validation: AlternativeMeasures of the Occupational Distribution

and Schooling

The regression results reported thus far usedthe proportion of fathers who received a referralfor their first job to measure the occupationaldistribution. We now proceed to verify the ro-bustness of the results by repeating the school-ing regressions for boys, girls, and the pooledsample with the proportion of fathers schooledin English as the measure of the occupationaldistribution.

The coefficients on the cohort variable andthe household characteristics in Table 6, col-umns 1 to 2, are very similar to the estimatesreported in Table 4. The coefficient on the En-glish proportion, which can be interpreted asstate dependence in schooling choice at the jatilevel, is positive and significant as expected.

The coefficient on the English-boy interactionterm in Table 6, columns 3 to 4, which includejati fixed effects, is very similar to the Englishcoefficient for the boys in column 1, matchingthe results reported earlier with the referralsvariable. Although not reported, once again par-ents’ education and family income do not havea differential effect on schooling choice bygender.

The language of instruction measures thechild’s future occupation, and the referral sta-tistic measures the occupational distribution inthe previous generation, in most of the regres-sions that we report in this paper. The negativeand significant referral-boy coefficient in thefixed effects regressions then reflects the persis-tence in the occupational distribution acrossgenerations, differentially for boys and girlswithin the jati. To validate this interpretation ofthe results, we proceed to replace the language

TABLE 6—ALTERNATIVE MEASURES OF THE OCCUPATIONAL DISTRIBUTION AND SCHOOLING

Dependent variable English schooling Test scores

Occupational distribution measure Proportion of fathers schooled in EnglishProportion of fathers that

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)

Father’s 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)

Mother’s 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 0.354Number 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 incolumns 5–7 is restricted to children in cohorts 1–10, past the school-leaving age, who passed the SSC exam. Test scoresrange from 35 to 100. Boy ! 1 if the student is a boy, 0 if girl. Family income is measured in thousands of 1980 Rupees inthe year closest to the year in which the child entered school. Column 3 also includes cohort interacted with boy. Column 4and column 7 also include cohort, father/mother studied in English, father’s/mother’s 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.


of instruction with an alternative schooling out-come as the dependent variable.

We assume that test scores depend on schoolquality and the pre-schooling human capital ofthe student. The comparison of English andMarathi schools in Table 1, and the parents’ per-ception of these schools, indicates that schoolquality does not vary by the language of instruc-tion. Under the maintained assumption that pre-schooling human capital does not vary by genderwithin the jati, this implies that the referral-boycoefficient should be close to zero in the fixedeffects regression with test scores as the depen-dent variable. The fact that referrals have adifferential effect by gender on schooling choiceshould be irrelevant for test scores, if schoolquality does not vary by the language ofinstruction.

Columns 5 to 7 of Table 6 replace the lan-guage of instruction with performance on theschool-leaving SSC examination as the depen-dent variable. Referrals are once more used tomeasure the occupational distribution in the jati,to be consistent with the specifications usedelsewhere in the paper. We restrict attention tothe first ten cohorts (age 16–25), which havealready attained school-leaving age, in theseregressions. Only 17 percent of the students age16–25 in the sample never passed the SSCexamination, so we focus on the test score con-ditional on having passed the exam in theseregressions.28

Table 6, column 5, restricts attention to boys,and includes the cohort, family characteristics,and the level of referrals in the jati as regres-sors. The cohort effect is negative and signifi-cant, suggesting a decline in the quality ofstudents over time. The referral coefficient isalso negative and precisely estimated, whichwould be the case if students from high-referraljatis have lower levels of pre-schooling humancapital. Consistent with this interpretation, fam-ily characteristics, particularly parents’ years ofeducation, have a very large positive effect ontest performance. Subsequently we repeat theexercise just described for the girls (Table 6,

column 6). The cohort effect is now absent, butthe coefficient on referrals remains negative andstatistically significant—both boys and girlsfrom high-referral jatis do less well on exams.The fixed-effects estimates, reported in Table 6,column 7, of the cohort effect, the cohort-boyinteraction, and the boy dummy are not statis-tically significantly different from zero. Moreimportantly, the coefficient on the referral-boyinteraction term is small and statistically insig-nificant, in contrast to the specifications withlanguage of instruction as the dependent vari-able. Caste networks affect the language but notthe quality of instruction of their members.

E. Selection into Marathi Schools over Time

The framework laid out in Section II also hasimplications for the compositional change in thestudents who attend Marathi schools over timeby jati: first, the pre-schooling human capital ofboys entering Marathi schools should declineon average as the returns to English grow. Sec-ond, when there are no restrictions on mobilityput in place to exploit network externalities, thedistribution of pre-schooling ability among theboys entering Marathi schools will convergeacross all jatis over time. It is possible that suchconvergence across jatis will be absent whenrestrictions are in place. Note that the model hasno prediction for selection by ability into En-glish schools.

We do not have a direct measure of pre-schooling human capital. The results in Table 6suggest, however, that, net of income, parentalschooling has a positive and significant effecton school performance. In particular, father’sschooling has a significant positive effect on testscores for boys and girls, and the effects do notdiffer significantly by the gender of the child.We thus use the father’s schooling level as aproxy for pre-schooling ability.29 The questionwe address is whether boys with more educatedfathers increasingly exit Marathi schools andwhether, and how, the rate of decline in thepre-schooling ability of boys entering Marathischools varies by jati.

28 Munshi and Rosenzweig (2003) studied the effect ofreferrals on the probability of success in the SSC exam andobtained results that are qualitatively the same as what wereport below with the test score, conditional on success, asthe dependent variable.

29 The results reported below are essentially the same ifwe replace father’s schooling by mother’s schooling.


To test the implications for school selectivitydescribed above, we estimate regressions on thesubsample of boys entering Marathi schools ofthe form

(4) E$Sij!Eij # 0% # * & +Rj & ,Cij

& -Rj ! Cij & .!j ,

where Sij is boy i in jati j’s father’s years ofschooling; Cij is the boy’s cohort; Rj mea-sures the level of referrals in the jati; and !jmeasures pre-schooling ability in the entirejati. These terms reflect the fact that pre-schooling ability conditional on selection intoMarathi is, in general, a function of ability inthe jati and the level of referrals. The cohortterms reflect the change in this selection pro-cess over time. The model, which ignores thevariation in ability !j across jatis, predictsthat + ( 0, , " 0, - " 0 without restrictions,and (possibly) - ( 0 with restrictions.

Because the major shift into English school-ing occurred in the 1990s, we first estimateequation (4) for the boys in cohorts 11–20. The

estimates are reported in Table 7, column 1. Thecohort coefficient is negative and significant aspredicted, which implies that the pre-schoolinghuman capital of the boys who entered Marathischools was declining substantially in the 1990s.The coefficient on the referral-cohort interac-tion term is positive, consistent with the resultsin Table 5 showing that restrictions on mobilityin the high-referral networks were still in placeduring this period—the more-able boys in high-referral jatis were shifting to English-mediumschools at lower rates.

The referral coefficient is negative and sig-nificant, consistent with the results in Table 6,which indicate that jatis with higher referrals Rjhave lower ability !j; the negative Rj & !jcorrelation appears to dominate the positive se-lection + ( 0 effect in this case. But this tells usthat the positive referral-cohort coefficient thatwe reported above might also be spurious. Toassess the robustness of the results in column 1,we add jati fixed effects, which subsume * #+Rj # .!j (Table 7, column 2). The referralcoefficient + is no longer identified, but theestimated cohort and referral-cohort coefficientsare very similar to the results in column 1.

TABLE 7—SELECTION INTO MARATHI SCHOOLS OVER TIME

Dependent variable Father’s years of education

Cohort 11–20 1–10

Sample Boys Girls

Boysandgirls Boys Girls

Boysandgirls

(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)

R2 0.106 0.205 0.138 0.205 0.215 0.136 0.254 0.184 0.278 0.285Number of

observations839 839 815 815 1,654 866 866 851 851 1,717

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


The within-jati estimates allow ability to varyacross jatis but assume that ability is constantover time (both within and across generations).The level of parental schooling could, however,also depend on the access to education, whichmight have changed over time. If there wasconvergence in the access to education acrossjatis in the parent generation, that could explainthe positive referral-cohort coefficient in col-umns 1 to 2 without requiring networks to beactive. One test to rule out this alternative in-terpretation of our result would be to estimatethe school selectivity regression for girls ratherthan boys; we have already seen that the net-work has no effect on schooling choice for thegirls, and so both the cohort and the referral-cohort effect should be absent. In contrast, if thereferral-cohort term is picking up convergencein (fathers’) schooling levels across jatis, thenthis coefficient should be positive and signifi-cant for the girls as well.

Table 7, column 3, reports the basic selectiv-ity regression for the girls attending Marathischools with cohort, referral-cohort, and refer-rals included as determinants of father’s school-ing, while Table 7, column 4, repeats thisregression with jati fixed effects. The referralcoefficient in column 3 is again negative (butinsignificant), consistent with the lower levelsof ability in high-referral jatis. The cohort co-efficient is positive but insignificant. More im-portantly, the referral-cohort coefficient is smallin magnitude and statistically insignificant—thepoint estimate is actually negative—and consis-tent with the results obtained earlier that girls infamilies belonging to high-referral jatis are notrestricted in their mobility.

An alternative strategy to control for the con-founding effect of changes in access to school-ing among the fathers across jatis and over timepools boys and girls in the selectivity regression,which can then be estimated with a full set of jatidummies interacted with the cohort variable

(5) E$Sij!Eij # 0% # $, $ ,̃%Cij ! Bij

# $- $ -̃%Rj ! Cij ! Bij

# fj & gj ! Bij & hj ! Cij ,

where ,̃, -̃ are the coefficients on the cohortvariable and the referral-cohort interaction for

the girls, and Bij is a boy dummy as before. Thefixed effects, fj, which subsume *̃ # +̃Rj # .!j,allow for the possibility that ability variesacross jatis. The fixed effects interacted with theboy dummy gj ! Bij , which subsume (* & *̃)Bij #(+ & +̃)Rj ! Bij , also allow ability to vary bygender across jatis. Finally, the fixed effectsinteracted with the cohort variable hj ! Cij , sub-sume ,̃Cij # -̃Rj ! Cij and control for changesin access to schooling for the fathers both acrossjatis and over time.

For the special case with ,̃ ! 0, -̃ ! 0, as isconsistent with the model, the estimated coeffi-cients in equation (5) should match the cohortcoefficient and the referral-cohort coefficientwhen equation (4) is estimated with jati fixedeffects for boys only. Table 7, column 5, sug-gests that this is indeed the case: the cohort-boycoefficient is negative and significant, the refer-ral-cohort-boy coefficient is positive and signif-icant, and the point estimates are very similar tothe corresponding coefficients in columns 1 and2. These results confirm that in the most heavilynetworked jatis, high-ability girls were exitingto English-medium schools at significantlyfaster rates than were boys.30 The 0.1 quantile–0.9 quantile of the referrals distribution rangesfrom 0.2 to 0.7. The point estimates in column5 thus suggest that over the period of the 1990sthe gap in father’s schooling between boys andgirls schooled in Marathi grew by 2.3 years inthe highest-referral jatis (at the 0.9-quantilelevel). In contrast, the ability-differential mea-sured by the difference in the father’s schoolingbetween boys and girls, declined by as much asfive years in the low-referral jatis (at the 0.1-quantile level) over the same period. This in-creasing mismatch in ability levels between thesexes within jatis and school types could haveimportant implications for the future stability ofthe caste system, which relies on endogamousmarriage, as discussed below.

Columns 6 to 10 of Table 7 report the esti-mates of the selectivity equations for the firstten cohorts of students, who entered school in

30 The negative cohort-boy coefficient implies that theboy-girl pre-schooling human capital differential is declin-ing over time, independent of the influence of the male jobnetwork. This may be due to differences in labor forceparticipation or changes in the returns to English by gender(as in Figures 1 and 2).


the 1980s. Schooling choices were stable overthis period and thus we do not expect to findchanging selectivity effects for the boys or thegirls. As before, the referral coefficient, in col-umn 6 and column 8, is negative and significant,reflecting the persistent differences in abilityacross jatis. As expected, however, and in con-trast to the cohorts making schooling choices inthe post-1990s new economy, the cohort effectand the referral-cohort effect, both uninteractedand interacted with the boy dummy, are insig-nificant in the pre-reform period.

F. Alternative Interpretations of the EmpiricalResults

The discussion that follows considers alter-native explanations for the results we have pre-sented. The identifying assumption in the fixedeffects schooling choice regression is that un-observed determinants of schooling choice thatare correlated with the occupational distributionshould not vary by gender, or have a differentialeffect by gender on schooling choice, within thejati. Some of the alternative explanations wepursue are associated with the failure of thisidentifying assumption. Other explanations aregenerated by relaxing the assumptions of themodel. We will argue that none of these alter-native explanations matches all the results aswell as our preferred explanation, based on un-derlying male labor market networks.

Liquidity Constraints.—The model assumesthat schooling choices are based entirely on theindividual’s ability and the historical occupa-tional distribution in his jati, which determinesthe labor market network that he inherits. Whencredit markets function imperfectly, liquidityconstraints could, in addition, prevent individualsbelonging to working class jatis from choosingmore expensive English schooling. Liquidity doesnot vary by the gender of the child within the jati,and so the jati fixed-effects regression wouldappear to rule out this alternative explanation.Boys and girls have different labor market op-portunities, however, and it is thus conceivablethat liquidity could have a differential effect onschooling choice by gender. The schooling re-gression accounts for liquidity constraints byincluding family income at the time the childentered school. Schooling expenses for the chil-

dren in our sample are relatively low (6.3 per-cent of family income in English schools and6.0 percent in Marathi schools), and, not sur-prisingly, family income has a relatively weakeffect on schooling choice for both boys andgirls. Moreover, the effect of family income onschool choice does not differ by gender at con-ventional levels of significance (Table 4).

Differences in Ability.—The jati-level fixedeffects absorb all variation in the jati that is notgender specific. But in an economy where menand women historically performed very differ-ent roles, the parental and societal inputs thatboys and girls received in childhood might havebeen very different. The results reported earlier,however, provide no evidence of gender distinc-tions in pre-schooling human capital withinhouseholds or jatis. The estimates reported inTable 4 do not reject the hypothesis that theeffects of parental human capital characteristicson school choice are equal for boys and girls.The results reported in Table 6 with test perfor-mance as the dependent variable are also con-sistent with the assumption in the fixed effectsschooling regression that pre-schooling humancapital does not vary by gender within the jati.

Discrimination.—Unless there is a gender-based component to caste discrimination, it willbe subsumed entirely by the jati fixed effects.But it is possible that firms or schools discrim-inate against boys from working-class back-grounds, perhaps because they are difficult todiscipline, while treating girls from differentbackgrounds more equally. The referral-boy co-efficient would proxy for underlying discrimi-nation in that case.

Recall that household characteristics, such asparental education and family income, had thesame effect on schooling choice for boys andgirls within the jati. It was only the jati-levelreferrals statistic that had a gender-specific ef-fect on schooling. If these results are attributedto discrimination, then it implies that firms orschools do not discriminate by family back-ground within the jati, but by jati affiliationalone. It is not obvious why we would expect tosee gender discrimination purely along castelines. Family characteristics, such as parentaleducation and income, were seen to be corre-lated with pre-schooling human capital and are


at least as easy to observe as caste identity.Historically there does not appear to have beena policy of caste discrimination by employers inany industry in Bombay in any case (Morris,1965).

Restrictions on High-Caste Women.—We fo-cused in Figure 3 and Figure 4 on the absence ofconvergence for the boys, which was attributedto restrictions on occupational mobility amongthe lower castes. An alternative interpretation ofthese figures is that the ability distribution var-ies across the population such that it remainsoptimal for individuals to sort by caste intodifferent careers, even as the returns to Englishgrow. The convergence among the girls withthis alternative interpretation is attributed to re-strictions on the high-caste girls.

There is no evidence that such restrictions arein place, or have been in place historically.High-caste women in Bombay have always hadhigher labor-force participation rates and moreEnglish schooling than lower-caste women, asobserved in Table 2. Within the high castes,boys are substantially more likely to beschooled in English than girls, and so the girlscould easily switch into English schools withoutcreating a mismatch on the marriage market.

Moreover, although the pooled schoolingchoice regression with fixed effects cannot dis-tinguish between the alternative explanation,based on female restrictions, and our view thatmale networks shape schooling choice for theboys alone, recall that we also ran regressionsseparately for boys and girls. The jati-level sta-tistic, measured either by the proportion of re-ferrals or the proportion of fathers with Englishschooling, affects schooling choice for the boysbut not for the girls in these regressions. Wethus appear to be picking up restrictions onmobility that are specific to the boys.

Sampling Bias.—We noted three potentialsources of sampling bias in Section IIIA. First,particular households might school their chil-dren outside the Dadar area. Second, particularhouseholds might have moved from Dadar overthe past 20 years. Third, children from particu-lar households might have dropped out ofschool.

The first two sources of sampling bias areeasily accommodated in the fixed-effects re-

gression framework. Although school locationsand out-migration might vary by jati, there is noreason to expect these decisions to vary by thegender of the child within the jati. The thirdsource of sampling bias is potentially moreproblematic, because drop-outs could vary bygender within the jati. The decision to drop outwould depend on the child’s pre-schooling hu-man capital and future employment opportuni-ties, both of which determine schooling choice.Selective dropouts, by gender across jatis, couldconsequently violate the identifying assumptionunderlying the fixed-effects estimation procedure.

However, the sex-ratio of students in themost recent eight cohorts (grades one througheight) in which there would be relatively fewdropouts is statistically indistinguishable fromthe sex-ratio in the older 12 cohorts. Regres-sions not reported here also reveal that the sex-ratio is uncorrelated with the level of referrals inthe jati, both in the first 12 cohorts and in the 8most recent cohorts.31

V. Conclusion

As modernization proceeds around the world,there is a perception that indigenous existing in-stitutions may importantly shape the course of thedevelopment process across different countries.Yet little is known about how such institutionsactually affect the transformation of economiesundergoing change, or their impact on the eco-nomic mobility of particular groups of individuals.This paper examines the role of one long-standingtraditional institution—the Indian caste system—in shaping career choices by gender in a rapidlyglobalizing economy.

We have found that male working-classnetworks, organized at the level of the sub-caste or jati, continue to channel boys intotraditional occupations, despite the fact thatreturns to nontraditional (white-collar) occu-pations have risen substantially during thepost-1990s reform period. In contrast, girls,who have had historically low labor-marketparticipation rates and few network ties to

31 As an additional test, we also verified that the sex-ratioin the most recent cohort that entered school in 2001 isuncorrelated with the level of referrals in the jati. Thus,there does not appear to be selective enrollment by genderand jati either.


constrain them, appear to be taking full ad-vantage of the opportunities that have becomeavailable in the new economy. It is generallybelieved that the benefits of globalizationhave accrued disproportionately to the elitesin developing countries. In this setting wefind, instead, that a previously disadvantagedgroup (girls) might surpass boys in educationalattainment and employment outcomes in thefuture in the most heavily networked jatis.

Although we have focused on how traditionalinstitutions shape the responses of particulargroups of individuals to the new opportunitiesthat accompany globalization, our findings sug-gest that these institutions are likely to be af-fected in turn by the forces of change. In ourframework, an individual schooled in Englishno longer needs the traditional caste network;indeed, it has been remarked that “the Englisheducated form a caste by themselves” (M. P.Desai, quoted in Julian Dakin et al., 1968 p. 24).Simple statistics on marriage and migration thatwe computed for the elder siblings of the stu-dents in our sample would appear to supportthe view that English education will ultimatelyundermine the caste network. Among the 825married siblings in our sample, 11.9 percentmarried outside their jati. This contrasts withthe parent generation, in which only 3.7 percentof the partners were not members of the samejati. Schooling in English appears to be contrib-uting to this increase in inter-caste marriage, as31.6 percent of the English-educated siblingsmarried outside their jati, versus only 9.7 per-cent of the Marathi-educated siblings. Andamong the 1,073 siblings who are currentlyemployed, 13.9 percent of the English-educatedwork outside Maharashtra, versus only 2.1 per-cent of the Marathi-educated (these differencesbetween the Marathi-educated and the English-educated are statistically significant at the 5-per-cent level). Both marriage outside the jati andout-migration weaken caste ties and the castenetwork. Increasing exposure to the moderneconomy through English education, and themismatch in educational choices and future oc-cupational outcomes between boys and girls inthe same jati that we have documented, suggestthat the forces of modernization could ulti-mately lead to the disintegration of a system thathas remained firmly in place for thousands ofyears.

REFERENCES

Acemoglu, Daron; Johnson, Simon and Robinson,James A. “The Colonial Origins of Compar-ative Development: An Empirical Investiga-tion.” American Economic Review, 2001,91(5), pp. 1369–1401.

Banerjee, Abhijit and Iyer, Lakshmi. “History,Institutions, and Economic Performance: TheLegacy of Colonial Land Tenure Systems inIndia.” American Economic Review, 2005,95(4), pp. 1190–213.

Burnett-Hurst, Alexander R. Labour and hous-ing in Bombay: A study in the economicconditions of the wage-earning classes ofBombay. London: P. S. King and Son,1925.

Chandavarkar, Rajnarayan. The origins of in-dustrial capitalism in India: Business strate-gies and the working classes in Bombay,1900–1940. Cambridge: Cambridge Univer-sity Press, 1994.

Cholia, R. P. Dock labourers in Bombay. Bom-bay: Longmans, Green and Co. Ltd., 1941.

Dakin, Julian; Tiffen, Brian and Widdowson,Henry G. Language in education: The prob-lem in Commonwealth Africa and the Indo-Pakistan sub-continent. Oxford: OxfordUniversity Press, 1968.

Dandekar, Hemalata C. Men to Bombay, womenat home: Urban influence on Sugao Vil-lage, Decan Maharashtra, India, 1942–1982. Ann Arbor: Center for South andSoutheast Asian Studies, University ofMichigan Press, 1986.

D’Monte, Darryl. Ripping the fabric: The de-cline of Mumbai and its mills. New Delhi:Oxford University Press, 2002.

Gokhale, R. G. The Bombay cotton mill worker.Bombay: Millowners’ Association, 1957.

Gore, M. S. Immigrants and neighborhoods:Two aspects of life in a metropolitan city.Bombay: Tata Institute of Social Sciences,1970.

Morris, Morris David. The emergence of an in-dustrial labor force in India: A study of theBombay cotton mills, 1854–1947. Berkeley:University of California Press, 1965.

Munshi, Kaivan. “Networks in the ModernEconomy: Mexican Migrants in the U.S. La-bor Market.” Quarterly Journal of Econom-ics, 2003, 118(2), pp. 549–99.


Munshi, Kaivan and Rosenzweig, Mark. “Tradi-tional Institutions Meet the Modern World:Caste, Gender and Schooling Choice in aGlobalizing Economy.” Bureau for Researchand Economic Analysis of Development,Working Paper: No. 038, 2003.

Patel, Kunj M. Rural labour in industrial Bom-bay. Bombay: Popular Prakashan, 1963.

Rees, Albert. “Information Networks in LaborMarkets.” American Economic Review, 1966(Papers and Proceedings), 56(2), pp. 559–66.


traditional institutions meet the modern world: caste, gender, and

Documents