Crossing Boundaries: How Social Hierarchy

Impedes Economic Mobility

Hanan G. Jacoby Ghazala Mansuri�

Abstract

How important are social hierarchies in explaining durable economic in-

equality? Using novel data from rural Pakistan on the caste composition of

village hamlets and the location of primary schools, we investigate the impact

of caste-based hierarchies on human capital accumulation, a key determinant

of economic mobility. We �nd that social stigma greatly discourages school

enrollment among low-caste children, with low-caste girls, the most educa-

tionally disadvantaged group, being the worst a¤ected. However, low-caste

households who can escape stigma invest at least as much in schooling as high

caste households, indicating similar returns to schooling across caste groups.

Keywords: Hierarchical identity, Social exclusion, Inequality of oppor-

tunity, Caste-based discrimination

�The World Bank, 1818 H St. NW, Washington DC 20433. e-mail: hjacoby@worldbank,gmansuri@worldbank.org. The data used in this study were collected by the Pakistan Instituteof Development Economics, funded by grant (no. 84150) from the World Bank Research SupportBudget, both of which we gratefully acknowledge. Thanks also to Salma Khalid for excellentresearch assistance, Naghma Imdad for help with the caste questions in PRHS-II, and to JishnuDas, Chico Ferreira, Sylvie Lambert, David McKenzie, Martin Ravallion, and especially to KarlaHo¤ for useful comments. The views expressed herein are those of the authors and should notnecessarily be attributed to the World Bank, its executive directors, or the countries they represent.

�Our children sit on the �oor in the school; benches are for [high

caste] children...We can pollute them, you understand?� (Low-caste

woman, Sindh).

1 Introduction

Hierarchical group identities, whether based on ethnicity, race, religion, or caste,

are a pervasive feature of human society. Even as the extractive institutions (e.g.,

slavery, patronage systems) from whence they sprang whither away or die out com-

pletely, the hierarchies themselves often persist.1 Marked by the high-status group

as possessing �spoiled�or disreputable identities (Go¤man, 1963), members of the

low-status group can be stigmatized and excluded from a range of institutions, both

public and private. When such social exclusion perpetuates cross-group inequality �

as was, arguably, the case with separate but (un)equal schooling of white and black

children in the American south (Margo, 1986) �a poverty trap arises. Public policy

blind to group identities may be ine¤ective in remediating such durable inequities.

Empirical work on the role of social exclusion in shaping economic outcomes is

scant,2 due largely to the di¢ culty of establishing a non-discriminatory counterfac-

tual in a context where discrimination is ubiquitous (Fryer, 2011, makes a similar

point). To �ll this void, we consider a setting, rural Pakistan, in which a vesti-

gial caste system operates� much like in neighboring India, but far less recognized.

Communities have a long history of strati�cation along caste lines, some hamlets

or settlements dominated by high-castes and others by low-castes. Since not every

settlement can have its own primary school, access to education invariably requires

children to cross social boundaries. As we will show, it is the interaction of school

placement and settlement caste-structure that provides us with the requisite non-

discriminatory counterfactual within a discriminatory environment; in particular, it

allows us to identify the impact of social exclusion on human capital accumulation,

an important source of economic inequality across groups.3 Indeed, we �nd that

1On American slavery and the subsequent Jim Crow laws as extractive institutions, see Ace-moglu and Robinson (2006).

2There is, of course, a vast literature on labor market discrimination (see Altonji and Blank,1999), but, as Loury (2000) exhorts, "when considering ethnic group inequality, economists shouldlook beyond what happens in markets." (p. 232).

3In this respect our work proposes a novel channel for community heterogeneity to a¤ect out-comes. By contrast, the collective action or "social capital" literature (Alesina and La Ferrara,2005) argues that, by inhibiting cooperation, social fragmentation reduces the provision of local

1

a hypothetical removal of discriminatory barriers to education would cut a 42 per-

centage point school enrollment rate gap between the most disadvantaged group,

low-caste girls, and the least disadvantaged, high-caste boys, to one of 7 percentage

points; a six-fold decline!

Our data set is unique in explicitly recognizing the geographic structure of vil-

lages and the social structure of their constituent settlements. Controlling for school

distance, we �nd that the propensity to ever enroll in school is substantially lower

for girls who would have to cross the physical boundaries of their settlement to at-

tend, an e¤ect not present for boys. This can be explained by the custom of purdah

or female seclusion, which is more rigidly enforced outside of the settlement than

within it.4 In other words, sending a daughter to a school located outside of her

settlement (even though still within her "village") entails greater psychic costs than

sending her to an equally distant school within her settlement. Caste status, how-

ever, strongly conditions this result: Unlike their high-caste counterparts, low caste

girls appear to be indi¤erent to the presence of a school in their own settlement, but

are much more likely to enroll when a school is available in a low caste dominant

settlement. Low-caste boys, too, are more likely to enroll when a caste-concordant

schooling option is available. These �nding appear to be the result of discrimination

or stigma operating against low-caste children wanting to access schools in high caste

dominant settlements.5 Indeed, we can rule out alternative explanations based on

school quality di¤erences or on di¤erential aspirations, motivations, or preferences

between low-caste children living in high and low-caste dominant settlements.

Finally, we demonstrate the quantitative importance of this stigma e¤ect by

showing that a policy of prioritizing school availability in settlements where low-

caste households dominate would actually increase overall enrollment by more than

a policy of placing schools in unserved settlements irrespective of their caste com-

position, and would do so at much lower cost.

public services such as schooling. Note that, while this mechanism might explain generally lowschool attainment in more diverse communities, it cannot easily rationalize inter-group di¤erenceswithin these communities, which is our primary concern.

4Purdah is not con�ned to Islamic societies, but is practiced throughout much of northern andcentral India as well. Fafchamps and Quisumbing (1999) discuss female seclusion in the context oflabor markets in Pakistan, arguing that this custom might explain low female returns to education.

5Note how this mechanism is distinct from that espoused by, e.g., Akerlof (1997), and othercommentators on the American underclass. Rather than facing ostracism by their out-group peersfor attending school�i.e., "the cost of education includes disutility from deviation from others inone�s social network" (Akerlof, p. 1017)�low-castes here face ostracism by the in-group (see alsoAkerlof and Kranton, 2000).

2

The empirical challenge that we face is well-known: Schools are not randomly

allocated across communities and, in particular, may be more likely to be built where

demand is high (or low, for that matter); in other words, the need to cross settle-

ment boundaries to attend a school may be endogenous. To uncover the causal

e¤ects of school location, we consider two identi�cation strategies that allow for

the possibility that unobserved education demand varies across settlements within

a village. The �rst strategy exploits the fact that even children living within the

same settlement may have had, on account of gender and, possibly, age variation,

di¤erent access to school at the point of enrollment. As a consequence, we can use

settlement �xed e¤ects to purge settlement level unobservables. The second strat-

egy involves an instrumental variable, the settlement�s share of village population,

which predicts whether a settlement within a village receives a school. To validate

the instrument, we show that migration into settlements over time has not been

driven by the establishment of new schools and that the instrument is uncorrelated

with settlement-level factors potentially related to the returns to, or the costs of,

schooling.

There is a growing economics literature featuring caste in South Asia, within

which the work of Anderson (2011) and Anderson et. al (2011) is closest in spirit

to ours. While the problems studied in their papers are di¤erent,6 we place com-

mon emphasis on the economic implications of power relations between elite and

subservient groups. In a separate vein, experiments conducted in India by Ho¤ and

Pandey (2006) suggest that caste-based stigma could play a powerful and early role

in human capital formation. Low-caste children perform poorly on cognitive tests,

but only if their caste status is made public; if caste is kept private, both groups do

equally well. Munshi and Rosenzweig (2006), meanwhile, argue that low-caste girls,

having looser ties to ancestral kin networks than their male counterparts, reap much

greater advantages from globalization through their schooling choices; i.e., in e¤ect,

their returns to education are higher. Our �ndings imply that low and high caste

parents, inasmuch as they would make similar schooling choices for their children

once social barriers are swept away, perceive similar returns to education.

This paper also links to a growing literature on barriers to school access in

developing countries. Pitt et al. (1993) and subsequently Du�o (2001) show that

better availability of village schools increased school attainment in Indonesia. More

6Anderson (2011), for example, considers how social hierarchy interferes with contract enforce-ment in private groundwater transactions.

3

recently, and in the context of South Asia, Andrabi et al. (2013) and Burde and

Linden (2013) �nd important school supply constraints, in the latter case with strong

e¤ects on girls�enrollment. None of this work, however, focuses on the exclusionary

role of caste.

The next section of the paper describes the data used in our empirical analysis.

Section 3 lays out the context, describing the structure of villages, the nature of fe-

male seclusion norms, and the salience of caste, following which, in sections 4, 5, and

6 we present, respectively, our empirical framework, results, and policy experiment.

We conclude in section 7.

2 Data

Our analysis combines representative household survey data with information from

a village and school census undertaken in the two most populous provinces of Pak-

istan, Punjab and Sindh. The household data come from the second round of the

Pakistan Rural Household Survey (PRHS-II) conducted in 2004-05, in which 3519

households were randomly drawn from 165 villages. The school census collects de-

tailed information on all schools inside each of these villages as well as schools lying

within a 15 minute walk (about 1 km) of the perimeter of each hamlet or settlement.7

GPS coordinates are available for households and schools, so the distance between

each household and each school can be calculated.8 Finally, the PRHS-II village

census provides land ownership data and the caste/clan (zaat/biradari) a¢ liation

of every household in each of the 165 villages.

The school census identi�es 1326 schools of which 1112 (84%) have classes at the

elementary level (up to grade 5). Of these, 63% are exclusively elementary schools,

while the rest had an elementary school attached to a middle or high school. Since

three-fourths of elementary schools are public (91% in Sindh province), government

school allocation is central in determining access to education (Appendix Table A.1,

Panel A). Among elementary schools, 51% are coeducational, 28% are exclusively

7Given the di¢ culty in identifying all schools to which children in a village may have access,the school census was conducted in two stages. In the �rst stage, all schools inside the village, aswell as all schools within a 15 minute walk of each settlement, were visually identi�ed and a schoollist was drawn up. This list was provided to all survey teams conducting the household survey.During the household survey, any non-listed schools identi�ed by sample households were addedto the initial list. Likewise, any additional schools identi�ed during the community focus groupexercise were incorporated into the �nal school census list.

8Note, however, that shape �les with GPS coordinates of settlement boundaries within villagesof Pakistan are as yet unavailable.

4

for boys and 21% exclusively for girls. Public elementary schools are far more likely

to be single-sex (62%) compared to private ones (13%).

We consider all children who were age 9-15 at the time of survey; practically all

of the children who ever enroll in school have already done so by the age of nine.

Non-enrolled children age 5-8, by contrast, may eventually enter school. For this

reason, we exclude this entire cohort from our analysis. Married girls, who no

longer co-reside with their natal family, but who constitute just one percent of 9-15

year-old females, are excluded as well (child fostering, incidentally, is negligible in

Pakistan). These criteria yield a sample of 4,662 children, of whom 2,167 (46%)

are girls and 2,495 (54%) are boys.

Table 1 provides some descriptive statistics for the sample, highlighting, in par-

ticular, the importance of the school entry margin in rural Pakistan. Also evident is

the substantial gender gap in schooling, one that varies tremendously by region. In

the more developed North Punjab, 93% of boys and 85% girls had enrolled in school

by age nine, whereas the poorer southern part of Punjab and the entire province of

Sindh have much lower enrollment rates (just above 50% for girls) and huge gender

gaps favoring boys (17 and 23 percentage points, respectively). Dropout rates from

primary school are much less di¤erentiated by gender and region: Conditional on

having enrolled in school and not having completed the primary grades, 88% of 9-15

year-old boys are still in school compared to 81% of girls, with gender gaps of 5, 5,

and 11 percentage points in North Punjab, South Punjab, and Sindh, respectively.9

3 Context

The term "village" in much of Pakistan, and indeed in large parts of south Asia, has

a dual connotation. The revenue village is the lowest tier of the government�s ad-

ministrative structure, at which public services, such as schools and health workers,

are allocated. Most revenue villages consist of several hamlets, or distinct clusters

of habitations, which we will refer to henceforth as "settlements" even though their

residents would typically refer to them, and not to the higher-level administrative

units, as "villages".

9Beyond primary school, strong gender-region di¤erences in dropout rates re-emerge. Condi-tional on having completed primary school, 81% of 9-15 year-old boys are still enrolled in schoolcompared to 78% of girls, with region-speci�c gender gaps of 1, 15, and 22 percentage points, re-spectively. Since these decisions are driven more by the location of middle and secondary schoolsthan of primary schools, we will have little more to say about them in this paper (see Sawada andLokshin, 2009, for related evidence)

5

Within the 165 PRHS-II villages, our census records 1031 named settlements

(see Table 2). These are typically very small, with a median population of 24

households. Villages of North Punjab are less geographically dispersed and tend

to have many fewer but more populous settlements than those of Southern Punjab

and Sindh.10 To help visualize our empirical strategy, Figure A.1 in the Appendix

shows part of a typical village of Sindh. Scattered amongst the agricultural �elds

can be seen clusters of habitations (settlements), the larger of which have a primary

school located in their midst.

3.1 Seclusion practices and mobility constraints

Seclusion of women (purdah) is practiced widely and in varying degrees throughout

rural Pakistan. As they enter puberty, girls usually experience increased enforce-

ment of purdah norms and attendant restrictions on mobility and social interaction.

Limits on mobility for young unmarried girls include prohibitions on travelling un-

accompanied or the need to obtain permission to do so from a male family member.

These mobility constraints stem from a desire to maintain family honor (izzat),

which, in a conservative society, is intimately bound up with female reputation and

hence behavior, whether real or perceived (see also Jacoby and Mansuri, 2010). Un-

chaperoned travel outside the immediate con�nes of the village and, especially, the

settlement can invite damage to a girl�s reputation.

Among households living in close proximity with long standing familial, caste,

or patronage ties, as is typically the case within a settlement, there tends to be less

harassment of each other�s womenfolk than of those from other settlements. In the

PRHS-II, which asks all married women age 15-40 about their perception of safety

within and outside their own settlement as well as their use of any type of veil

or purdah in public,11 80% report feeling �safe�alone inside their own settlement,

while only 27% report feeling safe alone outside of it. Similarly, only 14% of married

women report practicing full purdah (coverage of whole body including face) out

10Until the late 19th century, much of North-central Punjab (West Punjab in British times)was a vast desert with little or no cultivation. This changed dramatically with the introduction ofcanal irrigation under British rule and the transformation of six million acres into one of the richestagricultural regions of Asia. To settle these newly cultivable lands, �yeoman�farmers were broughtin from East Punjab. New villages were laid out for these settlers in neat �modern�squares withone settlement area by design. The much older villages of South Punjab and Sindh, by contrast,have more organic settlement patterns derived from a feudal social organization.11Veiling or purdah in this context is the use of a large shawl (chador) to cover the head and

the body, and sometimes also part of the face.

6

in public within their own settlement compared to 31% outside their settlement.

Thus, settlement boundaries appear to matter as far as female reputational risk is

concerned.

Other data also point to mobility restrictions as a key constraint on schooling for

girls. The nationally representative Pakistan Integrated Household Survey (PIHS)

of 2001-02 shows that enrollment decisions for girls in rural areas are more responsive

than those of boys to proximity to school (see World Bank, 2005), but this analysis

does not account for settlement boundary-crossing. More suggestive evidence comes

from the 2001-02 PRHS-I, in which parents of children never enrolled were asked

the main reason for the child�s non-enrollment. For boys, far and away the most

important reason is economic ("school too expensive" cited by 43% of parents), but

the picture is quite di¤erent for girls. While economic motives still dominate (32%

reported "school too expensive"), social constraints become an important consider-

ation. Respondents were much more likely to report that they did not "approve" of

their non-enrolled daughters going to school (30%) than to disapprove of schooling

for non-enrolled sons (7.5%).

Field interviews conducted in �ve villages randomly selected, by region, from

the PRHS-II sample also highlight the importance of mobility constraints for girls�

school enrollment, as the following quotes illustrate:12

�...I took my daughters out of school because it was too far to walk

and I feel that things are not safe there...�(Woman, Southern Punjab)

�When sons go to schools that are far away we don�t get worried,

but for our daughter we get worried.� (Women, Southern Punjab).

�The problem is, when children are so small to attend primary school,

we cannot expect them to walk all the way to a school in another settle-

ment or another village. But when they grow a bit more mature, 8-9

years, we cannot send girls alone to walk this distance. There are tall

sugar-cane �elds and all the way is deserted except young boys, who use

the same route for going to school. This is very dangerous!� (High-caste

women, Southern Punjab)

12Khan (1998) provides similar evidence from North Punjab.

7

3.2 Caste and Stigma

Caste in Pakistan has only recently come under scholarly scrutiny.13 Indeed, the

PRHS-II is the �rst large-scale survey, as far as we are aware, to collect systematic

data on caste a¢ liation in rural Pakistan. While caste groupings in Pakistan can-

not be precisely mapped into the much more extensively studied social hierarchy of

neighboring India, there are signi�cant similarities between the two due to their long

shared history. In particular, caste identity is embedded within occupational di¤er-

ences, which are associated with status and notions of purity and pollution. Various

exclusionary norms follow from these hierarchies and are exercised in relations of

mutual assistance, in social networks, and in the establishment and maintenance of

political power within the village economy and without. Unlike the Indian con-

text, there has been no o¢ cial acknowledgement of caste-based discrimination in

Pakistan and thus no a¢ rmative action programs to mitigate its impacts. Caste

mobility through land transactions is also extremely limited, if not impossible.

Working with a local anthropologist, we constructed a caste-status identi�er that

categorizes dozens of distinctly named caste/clan (zaat/biradari) groups into "high"

and "low" caste. High-caste includes all such groups that self-identify on the basis

of traditional access to land (zamindars). The low-caste group comprises zaats

that were historically considered either out-castes (similar to the dalits in India) or

were in clientalist relationships with zamindars as providers of services in the village

economy; i.e. barbers, metalworkers, clothes washers, etc. Based on this de�nition,

around 25% of the population from which we draw our sample consists of low-caste

households, with the highest proportion (35%) found in Sindh province.14

Village census data, summarized in Table 2, provide insight into residential pat-

terns by caste in rural Pakistan. Villages typically contain many more distinctly

named zaat/biradari groups than do their constituent settlements. Moreover, while

more than half of all settlements are comprised of just one caste-type (mostly high),

this is true of just 3% of villages. In short, settlements are far more segregated

along caste/clan lines than villages.

We de�ne a caste-type as dominant if it owns the majority of land in the settle-

ment, based on land shares calculated from village census data. Anderson (2011),

13Cheema et al. (2006), in an ethnographic study of two villages, �nd that caste hierarchiesremain an important impediment to social mobility, even in relatively developed central Punjab.Budhani et al. (2006) also talk about caste as a barrier to school access in their qualitative analysis.14The population shares for scheduled castes and scheduled tribes in India are similar, ranging

from 15 to 35 percent across regions.

8

citing Dumont�s (1970) emphasis on economic power rather than numerical pre-

ponderance, uses precisely this notion of caste dominance for her study of Indian

villages (see also Srinivas, 1955). However, we modify this de�nition by allowing

population to matter beyond a threshold level of either 60% or 80%. Thus, when

the low (high)-caste population exceeds the threshold, the settlement is deemed low

(high)-caste dominant even if most land is owned by high (low)-castes. This is

to avoid situations where, for example, high-caste households own all the land but

have only a small number of children in school. Of the 492 settlements from which

our sample children are drawn, only 81 (16%) are low-caste dominant using the 80%

threshold (85 (17%) using the 60% threshold), re�ecting both the smaller low-caste

population as well as their lower landownership.15 Most of these low-caste domi-

nant settlements are located in Southern Punjab or Sindh.16 Among the low-caste

children in our sample, 46% reside in high-caste dominant settlements. By contrast,

only 7% of high-caste children live in low-caste dominant settlements, a di¤erence

with important implications for our empirical analysis.

Qualitative evidence gathered from �eld interviews documents how caste-based

stigma plays out in schools either to exclude low-caste children altogether or to make

them unwelcome. The following snippets are typical:

�As the school for girls is closed since about two years, the [high-

caste] girls are taught primary classes in the boys school. Our daughters

cannot attend classes together with the [high-caste] girls. So our girls

are not attending any school currently...� (Low-caste women, Sindh)

�The children of rich [high castes] are taught seriously but our chil-

dren are paid no attention to. Teachers treat children from poor house-

holds really badly. While our daughters have no access to the school at

all, our boys receive no attention from the teachers. All boys sit together

for classes but our boys sit on the �oor.� (Low-caste women, Sindh)

15From the �gures reported in Table 2 we can see that, based on the 80% (60%) populationthreshold, 22% (24%) of all settlements in the village census are low-caste dominant as comparedto only 5% (10%) of villages. This is further indication of the relative segregation of settlementsbased on caste in rural Pakistan.16As discussed in fn. 10, for historical reasons, villages in North Punjab tend to have far fewer

settlements than those in South Punjab and Sindh and, hence, less scope for caste-based residentialsegregation at the settlement level.

9

�The teachers make the daughters of [high castes] sit inside the rooms,

under the fans. Our poor children sit outside, under the sun and dust.�

(Low-caste women, Southern Punjab)

3.3 Enrollment decisions

To avoid excessive notation, we verbally sketch a model of school entry. The decision

is, �rst and foremost, a forward looking one, since completing each level or grade has

an option value consisting of the net return to going on to the next level. However,

there is also a cost to attending school, which includes: (1) foregone labor in the

home, (2) direct school fees and other monetary outlays, (3) disutility of walking

to school, and (4) psychic cost of social pressure, stigma, reputational harm, etc.

Following, e.g., Glewwe and Jacoby (1994), we can think of a �rst-stage problem

wherein parents decide on the best school available given considerations (2)-(4) and,

conditional on this best school, decide in a second stage whether to enroll at all.17

We have argued that purdah restrictions on girls come into force at the onset

of puberty, which occurs in the later primary school or early middle school years.

Yet, there are a couple of reasons to believe that mobility constraints a¤ect the

primary school enrollment decision. First, reputational risk to the family is still a

concern for younger girls (see, e.g., the third quote in section 3.1). Allowing them

unfettered mobility may signal a loose attitude about female honor in general, with

repercussions for other womenfolk. Second, because the decision to enroll a child is

driven in large part by the option value of future schooling, and because schooling

is costly, anything that raises the likelihood of dropping out of primary school in

the future will reduce the incentive to enroll at the outset.

With these considerations in mind, we state our two main hypotheses:

Settlement boundary e¤ect: Holding constant travel time to school, a girl whowould have to cross settlement boundaries to attend school will be less likely

to enroll than a girl who can attend a school in her own settlement. This

e¤ect should not be observed for boys.

Caste boundary e¤ect: Children will be more likely to be enrolled in school if17There are practical reasons why we focus only on the second stage problem in this paper. To

obtain a su¢ ciently large sample, we must consider 9-15 year-old children, a considerable fractionof whom have already left primary school (either dropped out or moved on to middle school) and,hence, for whom we have no information on the primary school actually attended.

10

their caste status coincides with that of the dominant caste of the settlement

containing their most convenient school (child-school caste concordance).

In the next section, we focus on the settlement boundary e¤ect, paying particular

attention to the endogeneity of school location but without considering caste-status,

whereas in section 5 we investigate the settlement and caste boundary e¤ects jointly.

4 Salience of Settlement Boundaries

4.1 Key variables

We use household and school GPS coordinates, and information on the year of school

establishment, to identify the location of each primary school that was available (i.e.,

not only present but functioning) in or near the child�s village by the time the child

turned 9. As new schools are established (and old ones closed down), the set of

available schools can di¤er between two children of di¤erent ages living in the same

settlement.18 Moreover, since single-sex schools are common in Pakistan, the set of

available schools will also vary within settlements by the gender of the child. From

this information, we construct:

�isv =

(1 if school for child i lies inside settlement s

0 otherwise(1)

where settlement s of village v is that in which child i resides, and

disv =

(distance to inside-settlement school if �isv = 1

distance to nearest outside-settlement school if �isv = 0: (2)

If there are multiple schools available within the settlement, then we chose the

nearest one to child i for the calculation of disv:19

The summary statistics in Table 1 show that around three-quarters of children

in our sample had an appropriate type of school available in their settlement at the

time they were nine years-old. While gender di¤erences in �isv are negligible, the

18Most existing public schools in Punjab and Sindh were already established by the early tomid-1990s. Since then, the main growth in the number of primary schools has been in the privatesector.19Note that, since we do not have GPS coordinates of settlement boundaries, we cannot calculate

distances between households or schools to settlement boundaries.

11

regional di¤erential is stark: 89% of children had access to inside-settlement schools

in North Punjab, a �gure which drops to 51% in Southern Punjab with its much

greater proliferation of settlements in a typical village.

Which primary schools do children actually attend? Considering the 40% of

our sample children who are currently enrolled in a primary school, we �nd that

22% have no appropriate school in their settlement. Of those children that do have

a school in their settlement, 73% choose to attend it. Among the children who

attend a primary school outside of their settlement (roughly 40% of the sample),

either because they choose to or because their settlement has no school, nearly half

attend the nearest one. In short, the vast majority (78%) of children who are

enrolled attend their nearest primary school.

A look at our school distance measure a¢ rms the limited geographical extent of

the typical settlement. The median distance between the child�s household and the

(closest) inside-settlement school is only 164 meters, with an interquartile range of

91 to 322 meters. Clearly, any potential travel time saving derived from locating

nearer to a school within one�s settlement is likely to be minuscule.

To separate the settlement boundary e¤ect from the e¤ect of distance, however,

it is important that the distribution of disv conditional on �isv = 1 share common

support with the distribution of disv conditional on �isv = 0: Figure 1 shows that

there is indeed considerable overlap; 85% of inside settlement-schools lie within half

a kilometer of the child�s residence compared to 36% of out-of-settlement schools.

Naturally, outside schools predominate at the tail of the distance distribution, and

so it is here where lack of common support may be problematic.

4.2 Empirical speci�cation

Let eisv be an indicator for whether a child has ever enrolled in school and consider

the regression

eisv = �G�isv �Gisv + �B�isv �Bisv + �G(disv) + �B(disv) (3)

+ G0Xisv �Gisv + B0Xisv �Bisv + �v + �sv + "isv

where Gisv and Bisv are girl and boy dummies, respectively, Xisv is a vector of child

(household) characteristics, including age to capture any cohort e¤ects, and �v;

�sv; and "isv are, respectively, village-speci�c, settlement-speci�c, and child-speci�c

12

error components.20 We assume throughout that "isv is orthogonal to all of the

regressors in equation (3). For each gender j = G;B; �j captures the settlement

boundary e¤ect and �j(�) the (possibly) non-linear mapping from school distance

to enrollment. Flexibly controlling for distance to school ensures that settlement

boundary e¤ects are not con�ated with distance e¤ects.

Non-random placement of schools is the principal threat to the validity of OLS

estimates of equation (3). If school allocation is based on village level unobservables,

then E[�isv�v] 6= 0: Village �xed e¤ects will, in this case, lead to consistent estimatesof the settlement boundary and distance e¤ects provided that E[�isv�sv] = 0. Of

course, there is ample reason to question this orthogonality condition. School

placement may re�ect unobserved settlement-level characteristics that also in�uence

enrollment rates. The next subsection details how we handle this possibility.

4.3 Identi�cation and testing strategies

It is important to avoid over�tting and the concomitant loss of statistical power when

testing for endogenous school placement. To this end, we impose the intuitively

appealing restriction that distance to school (or any function thereof) is conditionally

exogenous.21 Formally, letting �(disv) = �G(disv) + �B(disv) and Zisv = (�isv �Gisv; �isv �Bisv; Xisv �Gisv; Xisv �Bisv; �v); we assume that

�(disv)?�svjZisv (4)

which, if � is a linear function, is tantamount to assuming disv?�svjZisv. Thinkingof �sv as average settlement (unobserved) demand for education, condition (4) can

be justi�ed as follows: When �isv = 1; an available school lies inside the child�s own

settlement, which, of course, may indicate high settlement-level education demand,

but the distance between the child�s household and the within-settlement school

is essentially random (recall that settlements are quite small).22 When �isv = 0,

so that the child�s settlement does not have the right type of school, disv is the

20Of course, replacing the �rst two terms in equation (3) by �B�ivs + (�G ��B)�ivs �Givs is anequally valid parameterization. In either case, to allow for a gender-speci�c intercept, Xisv wouldinclude a constant term and the corresponding element of, say, B would be normalized to zero.21White and Lu (2010) discuss the role of conditional exogeneity in applied economics.22To anticipate a concern that might arise in section 5, there is no signi�cant relationship between

the distance to an inside-settlement school and a child�s caste status, whether we condition on thefull set of other covariates ultimately used in our regressions (p-value= 0.18) or not (p-value= 0.13).Thus, there is no evidence that schools are placed within settlements so as to be closer to highcaste neighborhoods.

13

distance between the child�s residence and an arbitrary location in a neighboring

settlement, the one containing the nearest available school. The distance between

the settlement where the child resides and the settlement where this school sits is

also, for all intents and purposes, a matter of pure chance, plausibly unrelated to

average education demand in the child�s settlement.

The conditional exogeneity restriction allows us to avoid the more data demand-

ing joint endogeneity test of �isv and �(disv). In particular, since (4) implies that

�(disv) = �0Zisv + �isv; where �isv is a random error term uncorrelated with �sv, we

may substitute into equation (3) to obtain

eisv = e�G�isv �Gisv+e�B�isv �Bisv+e G0Xisv �Gisv+e B0Xisv �Bisv+e�v+�sv+"isv+�isv:(5)

where e�j = �j + �j and �G (�B) is the element of � corresponding to �isv � Gisv(�isv � Bisv). To be sure, we are ultimately interested in �j, not in e�j, but, giventhe relatively high correlation between �isv and disv (cf., Figure 1), this formulation

allows a simple and relatively powerful test of the null E[�isv�sv] = 0. Once the

test is carried out, we can turn to the problem of disentangling �j and �(disv).

One strategy for obtaining consistent estimates of equation (5) under the alter-

native E[�isv�sv] 6= 0 is to purge �sv using settlement �xed e¤ects. Recall that theschools available at the time of potential entry may di¤er across children within the

same settlement according to the child�s age (cohort) and gender. Since �isv is not

a �xed settlement characteristic, e�G and e�B are, at least in principle, separatelyidenti�ed.23 In practice, however, within settlement variation in �isv is limited;

only 7% of settlements have at least two children with di¤erent values of �isv, al-

though nearly half of settlements exhibit variation in �isv �Gisv (or in �isv �Bisv): Amore fundamental issue with this identi�cation strategy is that (as in the analogous

scenario of Munshi and Rosenzweig, 2006), it assumes no gender-speci�c settlement-

level unobservable. In other words, we must rule out heterogeneity in preferences

for girls� education relative to boys at the settlement level, a preference which could

simultaneously increase girls�enrollments and the demand for a girl�s school in the

settlement. While a settlement-gender �xed e¤ect estimator would be robust to

this problem, the virtual absence of variation in �isv within settlement and child

gender renders this approach impractical.

This leads us to our second identi�cation strategy, which remains valid even in

23Even if �isv was constant within settlements, the di¤erence e�G � e�B would still be identi�ed(see Munshi and Rosenzweig, 2006, for an empirical strategy along these lines).

14

the presence of correlated settlement-gender-speci�c unobservables. We exploit the

fact that, within villages, public schools are largely allocated to high population

settlements. Indeed, private schools, most predominant in North Punjab, undoubt-

edly also locate within villages based on population considerations. Since, from the

village census, we know the number of households in each settlement as well as the

total number of settlements within each village, we can construct an instrumental

variable psv = proportion of households in village v residing in settlement s. The

school allocation mechanism implies that Cov(�isv; psvj�v) > 0; in other words, chil-dren residing in relatively high population settlements will be more likely to have a

school within their settlement. This relationship is con�rmed in Figure 2, which

plots a semiparametric version of our �rst-stage regression. Partialling out all of the

other covariates (see below) along with the village �xed e¤ects, we obtain a strongly

increasing and concave relationship between �isv and psv; across the interquartile

range of settlement population shares in our sample, for example, the proportion of

children with access to a within-settlement school more than doubles.24

4.4 Instrument validity

4.4.1 Migration across settlements

The validity of the exclusion restriction�that a settlement�s population share (psv)

a¤ects school enrollment only through school availability in the settlement (�isv)�

rests, in part, on the nature and extent of intra-village mobility. If those households

with a relatively strong demand for education (girls�education) tend to relocate to

settlements already having schools (girls�schools) or set up new settlements next

to existing schools, then a settlement�s population could re�ect the community�s

average strength of schooling preference, disqualifying psv as an instrument. Al-

though the establishment of new settlements is far-fetched in our context given

long-standing land ownership patterns, intra-village migration in response to school

availability cannot be dismissed a priori. To assess this possibility, we analyze settle-

ment level panel data covering 27 years constructed from retrospective information

on migration and school establishment, as described fully in the Appendix. To

summarize, we �nd no evidence that in-migration increases after the establishment

of a new primary school, whether girls�and boys�schools are treated separately or

24Figure A.2 in the Appendix, shows that this pattern is present in all three regions and, inparticular, is not driven by the expansion of private schools in North Punjab. Indeed, if anything,the pattern is even stronger in Sindh where there are mostly public schools.

15

jointly (see Table A.2). This result supports the exogeneity of settlement population

shares for our purposes.

4.4.2 Other settlement-level unobservables

We can also ask whether settlement population shares might be picking up some

other settlement-level factors, aside from schooling preferences, that a¤ect enroll-

ment behavior independently of the presence of a school. Agricultural productivity,

for example, might vary to a degree across settlements within a village; higher pro-

ductivity settlements attracting a larger share of the village population. Now, if it is

also true that the returns to education are greater in high productivity settlements,

then the exclusion restriction would be violated.

To assess the �rst part of this argument, we use data from the PRHS-II household

survey on land values at the plot level, taking means across all plots within the

same settlement. More productive land should translate into higher reported land

values. So, the question becomes: Do the more populous settlements within

a village have the more valuable land? The evidence, reported in the �rst two

columns of Appendix Table A.3, suggests not. There is no signi�cant relationship

between settlement population shares and the value of land (conditional on village

�xed e¤ects). And, this is not because land values are a poor proxy for productivity.

On the contrary, irrigation and soil type are highly signi�cant determinants of land

values at the settlement level (col. 3 of Table A.3).

Another line of argument is that relatively populous settlements are also those

with higher population density and population density in�uences parents�percep-

tions of child security and hence a¤ects school enrollment directly. To check for

this, we use village census data and regress settlement population density (number

of households per acre agricultural land, both cultivable and uncultivable) on psvand p2sv along with village �xed e¤ects. The results (cols. 4 and 5 of Table A.3)

again show no relationship between population shares and density.25

25Anticipating yet another concern that might arise later, the population share of a child�ssettlement bears no signi�cant relationship to a child�s caste status (p-value = 0:29), net of village�xed e¤ects. So, it is not the case, in particular, that low-caste children are concentrated inrelatively small settlements.

16

4.5 Results

4.5.1 School location endogeneity

The �rst four columns of Table 3 report alternative estimates of equation (5); i.e., all

speci�cations exclude distance to school. To facilitate IV estimation and comparison

across estimators, we mostly employ the linear probability model.26 Standard errors

are clustered at the settlement, the lowest-level sampling unit in the household

survey. Column (1) shows the OLS results with village �xed e¤ects (common for

boys and girls) and child age � gender controls. The presence of a school in the

settlement increases girls� enrollment rate by 18:9 (with a standard error of 3:4)

percentage points, which is very close to the gender gap in school enrollment (both

raw and conditional on village �xed e¤ects and cohort). In other words, girls with

a school in their settlement have practically the same enrollment rate as boys.

Next we control for household wealth using a dummy variable for whether the

household is landless along with the log of per-capita household expenditures mea-

sured at the time of the survey. Neither variable is a perfect proxy for wealth,

but together they should do a reasonable job of capturing such variation. Allow-

ing the coe¢ cients on these variables to di¤er by the gender of the child, we see

that, although girls�enrollment is more responsive to wealth than that of boys, the

coe¢ cients on �isv barely change.

Column (3) of Table 3 provides evidence on the exogeneity of �isv conditional on

the village �xed e¤ect, household wealth proxies, and other regressors. As explained

in the previous section, including settlement �xed e¤ects removes any correlation

between �isv and settlement-level unobservables, though at the cost of purging sub-

stantial variation from the data. Indeed, the standard errors on e�B and e�G morethan double between columns (2) and (3) as a consequence. Nevertheless, the co-

e¢ cient estimates hardly budge, with that of e�G remaining statistically signi�cant.Based on this result, we cannot reject the hypothesis that E[�isv�sv] = 0:

27

26Speci�cally, we want to compare estimates based on the settlement �xed e¤ects strategy tothose from the IV strategy. However, if we take the discrete choice route, we would have to use a�xed e¤ect logit in the �rst case (given the large number of settlements relative to children) and aprobit IV in the second case (noting that the former estimator does not produce marginal e¤ects).Moreover, as we will see later, probit (or logit) estimation is only possible on a reduced sample,which compromises e¢ ciency relative to a linear probability model. Nevertheless, for the �nalresults in the paper we present the corresponding probit estimates.27Formally, we would like to test whether E

hdeisvd�isv

i= e�GG+ e�BB di¤ers across the two estima-

tion methods. We do so by stacking equations and estimating the joint cluster-robust covariancematrix of the two models (village and settlement �xed e¤ects). A �2(1) distributed test statistic of

17

The 2SLS results, using psv and p2sv interacted with the child gender dummies

as excluded instruments, appear in column (4). But �rst, the instrument diagnos-

tics (see the footnote to Table 3) resoundingly reject underidenti�cation and weak

identi�cation and fail to reject overidenti�cation. Moreover, as might be expected

given the concavity exhibited in Figure 2, adding the squared population share to

the instrument set improves the e¢ ciency of the 2SLS estimate (i.e., instrument

redundancy is strongly rejected). Turning now to the second stage estimates, these

results also belie school placement endogeneity; once again we cannot reject the

null that E[�isv�sv] = 0 (p-value = 0:43 using the testing approach in fn. 27).

We also remain extremely con�dent that the enrollment impact of having an inside-

settlement school, inclusive of its e¤ect through distance to school, is greater for girls

than for boys. Finally, using the settlement rank instead of the share of population

as the instrument (e.g., dummies for �rst and second most populous settlements in

the village) makes virtually no di¤erence for the results.

4.5.2 Settlement boundary e¤ect

Having just failed to reject E[�isv�sv] = 0 and maintaining the assumption of or-

thogonality between �(disv) and �sv in equation (3), we next distinguish settlement

boundary and distance-to-school e¤ects. To do so, we return to the OLS estimator

with village dummies adding disv as a control. Whether distance enters linearly in

column (5) or as a cubic polynomial in column (6), the �isv coe¢ cients falls only

marginally relative to column (2); i.e., in terms of our earlier notation, the �j do

not di¤er much from the e�j. Figure 3 plots predicted enrollment for boys and girlsas a function of distance based on the estimates in column (6).28 Erecting a settle-

ment boundary between a girl�s residence and a next-door school would reduce her

enrollment by as much on average as moving that school two and a half kilometers

away within the same settlement.

Given the paucity of inside-settlement schools at the greater distances, noted

earlier, we also experimented with alternative distance cuto¤s to check whether our

estimates of �G and �B are sensitive to the exclusion of far away schools. The

results are, in fact, highly robust. For example, re-estimating the model in column

just 0:1 (p-value = 0:8) provides no evidence against the null.28To be clear, equation (3) is not a regression-discontinuity design (RDD) inasmuch as �isv

is restricted to shift the intercept but not the slope of the enrollment-distance relationship. Asan empirical matter, an RDD-type interaction between �isv and � is not supported by the data,providing virtually no additional explanatory power.

18

(6) using the 4,093 children whose most convenient school is within 1000 meters,

we obtain b�G = 0:150 (0:042) and b�B = �0:027 (0:031). Thus, lack of overlappingsupport at the tails of the school distance distribution does not seem to bias our

results.

Lastly, we investigate the impact of middle school location on primary school

enrollment decisions by including variables analogous to �isv and disv for middle

schools. The likelihood of having a middle school inside one�s settlement (20% for

boys and 29% for girls) is far lower than that of having a primary school (74% and

72%, respectively) and middle schools are considerably more distant as well (median

distance to outside-settlement middle schools is 2.5 km compared to only around

0.7 km for such primary schools). Indeed, since the degree of overlap between the

distribution of distance for inside and outside settlement middle schools is so much

lower than was the case for primary schools, distinguishing settlement boundary

and distance e¤ects for middle school is likely to be far more di¢ cult. Thus, we

�nd in column (7) that having a middle school inside the settlement signi�cantly

raises primary enrollment rates for girls but not for boys, paralleling the results

for primary school location. However, including a cubic polynomial in distance to

middle school in column (8) largely wipes out this settlement boundary e¤ect with

respect to middle school location (747 middle school distances are missing because

the child had no such school available). The important result, returning to our main

argument, is that the primary school settlement-boundary e¤ects are not driven by

correlations between primary and middle school placement; i.e., the estimates of �jare quite robust (compare columns (6) and (8)).

To summarize, in a sample of 9-15 year-olds, for whom the decision to ever

enroll in primary school has already been made, we have demonstrated the salience

of settlement boundaries conditional on distance to the nearest primary school. Note

that it is immaterial to our argument regarding female seclusion practices that many

of the sample girls are currently below the age of puberty, since, for one thing, safety

issues are still relevant for younger girls and, for another, expectations about the

ability to attend school in the future can in�uence the decision to enter school to

begin with.

19

5 Salience of Caste-based Stigma

5.1 Empirical strategy

The key variable for examining the caste-boundary e¤ect is an indicator of child-

school caste-concordance. Let

CCisv =

(child i�s caste = settlement s�s dominant caste if �isv = 1

child i�s caste = nearest school settlement�s dominant caste if �isv = 0:

(6)

If the nearest school outside the settlement is not contained within any one set-

tlement (5% of cases), we match it to the nearest settlement and assign it the

corresponding dominant caste.

Means of CCisv by gender, caste, and region are shown in Table 1. Re�ecting

the residential patterns already discussed, high-caste children are more than twice

as likely overall to have access to a caste-concordant school than low-caste chil-

dren. Regionally, southern Punjab fares the worst on these terms with only 11%

of low-caste children having their most convenient school in a low-caste dominant

settlement.

5.2 Results

5.2.1 Caste-boundary e¤ect

In the �rst column of Table 4, we introduce caste-status as a determinant of en-

rollment behavior in a stripped-down model with only gender-age interactions and

village �xed e¤ects; the equivalent of column (1) of Table 3. Whereas in the raw

data shown in Table 1 we found a 16 percentage point enrollment gap favoring

high-castes overall, once we look within villages, there is little di¤erence in school

enrollment between high and low-caste boys. By contrast, there is a 9:4 (3:9) per-

centage point gap in enrollment rates favoring high over low-caste girls. Controlling

for household wealth in column (2) using the landless dummy and the per-capita

expenditure variable as before lowers this marginal e¤ect of higher caste status to

7:3 (3:7) percentage points. So, a modest share of the caste di¤erence in girls�

enrollment can be explained by the greater wealth of high-caste households.

Next we investigate the settlement-boundary e¤ect by child caste, conditioning

20

as before on distance to school (column (3)). Remarkably, the positive enrollment

e¤ect of having a school in one�s settlement is concentrated among high-caste girls.

Low-caste girls appear to be indi¤erent to the presence of an inside-settlement school.

Two possible explanations spring to mind: First, the enforcement of female seclusion

norms may simply be weaker among low-caste households than among their high-

caste neighbors. Second, since low-caste children are far more likely to live in

a high-caste dominant settlement than vice-versa, the stigma of attending a local

school if one exists is particularly salient for this group.

The last columns of Table 4 support the second explanation.29 Column (4) uses

the 80% population threshold for landownership-based caste dominance, whereas

column (6) uses the 60% threshold. Since the two sets of results are practically iden-

tical, we discuss only the latter. Low-caste girls with access to a caste-concordant

school enroll at a rate 27:7 (6:2) percentage points higher than their counterparts

without such schools, whereas there is no signi�cant di¤erence for high-caste girls.

For low-caste boys, the child-school caste concordance e¤ect is also important but

only about half as large, at 13:7 (5:5) percentage points, as it is for low-caste girls,

and there is no e¤ect at all for high-caste boys. These �ndings are consistent with

the qualitative data suggesting that caste-based stigma cuts across gender lines.30

To completely rule out our �rst explanation for the caste-di¤erentiated settlement-

boundary e¤ect (i.e., weaker female seclusion enforcement among low-castes), we

would have to show that low-caste girls with access to a caste-concordant school are

not indi¤erent to settlement boundaries. To do so�i.e., to estimate the four-way in-

teraction between gender, caste, �isv, and CCisv�we would need signi�cant numbers

of low-caste girls with caste-concordant schools both inside and outside their own

settlements. Unfortunately, since low-caste children are concentrated in relatively

few low-caste dominant settlements, there are very few (around 20) low-caste girls

with caste-concordant schools outside their settlements. This fact renders credible

estimation of the four-way interaction practically impossible. Nevertheless, the ma-

29We had to drop 235 observations in this analysis because the most convenient school was notwithin the village and, therefore, information on the caste-dominance of the school settlement wasnot available.30An otherwise identical probit model yields caste-concordance e¤ects of comparable magnitude

and signi�cance: In terms of marginal probabilities, 0:344 (0:056) for low-caste girls; 0:132 (0:054)for low-caste boys; essentially zero for high-caste children. Relative to the linear probabilitymodel, the probit estimation drops 29 villages (632 observations) in which all sample children hadever enrolled in primary school. Nevertheless, the settlement-boundary e¤ects are also similaracross estimators, with the probit marginal probability for high-caste girls at 0:175 (0:042) andinsigni�cant for all other children.

21

jor �nding of this section stands (modulo robustness checks in the next subsection):

Caste-based stigma is an important facet of school enrollment decisions.

To get a sense of how big these stigma e¤ects are, we predict enrollment rates

assuming that all children face social barriers (i.e., �isv = 0 and CCisv = 0 8i)while �xing the covariates (age, household wealth, and village) at the overall sample

means.31 Focusing on girls, we obtain baseline enrollment percentages of 47:4

(6:4) for high-castes and 44:0 (6:5) for low-castes. Dropping settlement boundaries

alone (i.e., setting �isv = 1) brings high-caste girls�enrollment up to 66:3 percent

and low-caste girls�enrollment to merely 46:1 percent, thus opening a caste gap of

around 18 percentage points. Now removing caste boundaries as well (CCisv = 1)

yields a high-caste girls� enrollment rate of 71:5 percent as compared to a low-

caste girls� enrollment rate of 73:8 percent. Thus, absent social barriers, low-

caste girls would actually enroll at slightly higher rates than high-caste girls. A

similar pattern obtains for boys, with the enrollment gap switching from around

7 percentage points favoring high-castes to 5 percentage points favoring low-castes

after social barriers are removed. These results suggest that, if anything, the latent

demand for education is higher among low-caste households than among their high-

caste neighbors in the village. Low-status per se does not appear to dampen parents�

aspirations for their children.

5.2.2 Alternative mechanisms

Aspirations One concern about our results regarding access to caste-concordant

schools runs as follows: Perhaps there is something special about low-caste children

living in low-caste dominant settlements, or about low-caste dominant settlements

themselves, aside from better access to caste-concordant schools. To be sure, pref-

erences, aspirations, motivation, etc. may be shaped by one�s position in the social

hierarchy of one�s community. These in�uences may a¤ect demand for education

independently of the speci�c mechanism that we have emphasized in this paper,

namely caste-based stigma at the schoolhouse door.

We deal with this question in the �fth column (80% population threshold for

caste dominance) and seventh column (60% threshold) of Table 4, by introducing

two additional dummy variables indicating, respectively, low and high-caste children

living in low-caste dominant settlements. Along with the caste dummy already

31This is precisely the calculation used in the introduction in discussing the enrollment gapbetween high-caste boys and low-caste girls.

22

included, these variables completely saturate the 2 � 2 interaction between childcaste-status and settlement caste-dominance. And, the results provide no evidence

that school enrollment rates di¤er between low-caste children in low-caste dominant

settlements and each of the other groups once access to caste-concordant schools

is accounted for. We take this as further support for the view that di¤erences in

educational opportunities matter far more for schooling outcomes than di¤erences

in preferences or aspirations.

School quality Might di¤erences in schools across community types explain the

caste-boundary e¤ect? Imagine that schools catering to high-caste children are

posher or more academically rigorous, thus demanding a greater level of preparation

and/or involvement on the part of their students.32 Low-caste children, having less

of these prerequisites, might be discouraged from attending caste-discordant schools

for this reason alone and not because of stigma.

Although we cannot observe academic standards directly, the school census pro-

vides a wealth of information on the characteristics of physical infrastructure and

teaching in all current village primary schools. Given the substantial correlations

among these variables, we construct two indices, one for school infrastructure and

one for teaching quality, using principal components (see notes to Appendix Table

A.4 for the underlying characteristics). Table A.3 reports tests for within-village

di¤erences in these two indices across high and low-caste dominant settlements, us-

ing both the 80% and 60% population thresholds. In no case are these di¤erences

statistically signi�cant, whether we consider all primary schools or only the subset

of primary schools catering to girls.

6 Policy Implications

Exclusion of low-caste children from educational opportunities has important im-

plications for the allocation of public school facilities. To illustrate, we simulate

the impacts of alternative school-placement polices on 9-12 year-olds in PRHS-II

villages. Consider, �rst, a

32Public schools in Pakistan, it should be noted, have standard requirements for uniforms, ex-pected hours of attendance, and curricula. However, in apartheid South Africa �a context relatedto ours in principle if not in degree �schools for Blacks were of lower quality than those for Whites(Case and Deaton, 1999).

23

Low-caste targeting policy: Provide a coed school to a low-caste dominant set-tlement in every village currently lacking one.

To avoid building schools with an infeasibly small student base, we restrict this

policy to villages with at least 30 low-caste households.33

Table 5 presents simulated enrollment impacts based on the statistically sig-

ni�cant coe¢ cients from column (4) of Table 4 and settlement population weights

derived from the PRHS-II village census.34 The targeted policy would greatly boost

enrollment rates for low-caste children of both genders (9.6 percentage points for

boys and 18.9 for girls),35 but would leave those of high-caste children unchanged.

Meanwhile, the overall enrollment rate, which weights both caste-types by their re-

spective settlement populations, rises by 3.5 percentage points. To achieve this

increase, 95 schools would have to be established, either built from scratch (93) or

converted to coed from existing single-sex facilities (2).36

By way of comparison, we next consider a

Caste-blind policy: Provide a coed school to every settlement currently lackingone.

As before, to limit �super�uous�schools, we restrict this policy to settlements

of at least 30 households. Our estimates imply that the caste-blind policy induces

three groups to increase their enrollment: high-caste girls, by relaxing the settle-

ment boundary constraint, and low-caste girls and boys, by (incidentally) relaxing

the caste boundary constraint. As shown in Table 5, the increase in high-caste girl

enrollment is modest, just 3.1 percentage points. Since higher population set-

tlements typically already have a school, the caste-blind policy disproportionately33Of course, these public schools would be open to students of any caste. Insofar as de facto

segregation by caste-type occurs, low-caste students might be deprived of positive peer e¤ectsbestowed by high-caste classmates. However, these potential bene�ts of mixing are decidedlysecond-order compared to those of enrolling in school in the �rst place.34The population proportion of girls among 9-12 year old children is, however, calculated from

our sample and, along with household size, is assumed to be constant across settlements andcaste-types.35How large of an unconditional cash transfer would be required to elicit an equivalent 18.9

percentage point enrollment gain among low-caste girls? Based on our estimate of 0:109 (0:017)for the coe¢ cient on G� ln(pce), the per-capita income of every low-caste household with a 9-12year-old girl would have to be boosted by a factor of 1:7!36We are assuming that these new schools can accommodate any number of additional students

induced to enroll. Even with our 30 household low-caste village population threshold, this isprobably not unreasonable given the size of these communities. Using a less �generous�thresholdof 60 households reduces the number of new schools to 72, but also slightly lowers the expectedenrollment gain to 3.2 percentage points.

24

serves smaller settlements (population threshold notwithstanding) and thus does

not greatly increase school access on a population-weighted basis. Among low-

caste children, the increase in enrollment is also much lower when caste is ignored

than when it is explicitly targeted; the targeted policy e¤ectively has a much lower

threshold for placing a caste-concordant school in a village, which is, after all, its

intention.

Overall, therefore, the caste-blind policy raises enrollment by just 2.2 percentage

points, only two-thirds as much as the caste-targeted policy, yet it requires twice as

many new schools to be established.37 Moreover, from a distributional perspective,

a school placement policy attentive to community-caste composition does far better

at reaching the educationally disadvantaged low-caste population, especially low-

caste girls.

7 Conclusion

Returning to the main theme of this paper, we have shown that social hierarchy, a

legacy of an ancient caste system, exerts a profound e¤ect on human capital accumu-

lation in rural Pakistan. Using intricate micro-data on caste and communities, we

have shone light on a speci�c mechanism, social exclusion, through which extractive

institutions continue to matter and to perpetuate economic inequality.

We have also seen that social constraints interact in surprising ways to limit

educational opportunities for girls. Entry into primary school is substantially dis-

couraged when girls have to cross settlement boundaries to attend, irrespective of

the distance they would have to travel. This e¤ect is concentrated among high-caste

girls, however; low-caste girls appear indi¤erent to settlement boundaries. But this is

because many low-caste children live in settlements dominated by high-caste house-

holds and thus face particularly high barriers to attending a local school, if one

exists.

A �nal lesson of this paper is that di¤erences in the returns to education (or in

household wealth, for that matter) do not fully account for school enrollment gaps

across gender and caste; there is a signi�cant role for social constraints. Indeed, our

results imply that if one could eliminate the in�uence of stigma, low-caste children

37Raising the settlement population threshold to 60 households greatly reduces the pro�igacy ofthe caste-blind policy (the number of new coed schools drops from 194 to 78), but also drasticallypares the expected enrollment gains (from 2..2 to 1.4 percentage points).

25

would enroll in primary school in greater proportions than their high-caste counter-

parts. Given the di¢ culty of designing policies that raise the returns to education

on a large scale, our �ndings thus provide hope that suitably targeted supply-side

interventions�in this case, speci�cally attentive to community caste structure�can

ameliorate educational inequities.

26

References

[1] Acemoglu, D. and J. Robinson, 2006. "De facto political power and institu-

tional persistence," American Economic Review Papers and Proceedings, 96(2):

325-30.

[2] Alesina, A. and E. La Ferrara, 2005. "Ethnic diversity and economic perfor-

mance," Journal of Economic Literature, 43(3):762-800.

[3] Akerlof, G. 1997. "Social distance and social decisions," Econometrica, 65(5):

1005-27.

[4] Akerlof, G. and R. Kranton. 2000. "Economics and identity,"Quarterly Journal

of Economics, 115 (3):715-753.

[5] Altonji, J. and R. Blank, 1999. "Race and gender in the labor market," chapt.

48 in O. Ashenfelter and D. Card, eds., Handbook of Labor Economics, Volume

3, Part C. Amsterdam: North Holland.

[6] Anderson, S., 2011. "Caste as an impediment to trade," American Economic

Journal: Applied Economics, 3(1):239-63.

[7] Anderson, S., P. Francois, and A. Kotwal, 2011. "One kind of democracy,"

unpublished manuscript, University of British Columbia.

[8] Andrabi, T., J. Das, and A. Khwaja. 2013. "Students today, teachers tomor-

row? Identifying constraints on the provision of education.�Journal of Public

Economics, 100:1�14.

[9] Budhani, A. A., H. Gazdar, H. B. Mallah and N. Masood, 2006. "Social mar-

ginalisation and education in Pakistan �Findings of a qualitative survey," Is-

lamabad: UNICEF.

[10] Burde, D. and L. Linden. 2013. "Bringing education to Afghan girls: A ran-

domized controlled trial of village-based schools," American Economic Journal:

Applied Economics, 5(3): 27-40.

[11] Case, A. and A. Deaton, 1999. "School inputs and educational outcomes in

South Africa," Quarterly Journal of Economics,114(3):1047-1084.

27

[12] Cheema, A., S. K. Mohmand, and S. A. A. Naqvi, 2006. "History, land settle-

ments and informal institutions in Sargodha," unpublished manuscript, Lahore

University of Management Sciences.

[13] Du�o, E. 2001. "Schooling and labor market consequences of school construc-

tion in Indonesia: Evidence from an unusual policy experiment.� American

Economic Review, 91(4): 795�813.

[14] Dumont, L., 1970. Homos hierarchicus: The caste system and its implications.

Chicago: University of Chicago Press.

[15] Fafchamps, M. and A. Quisumbing, 1999. "Human capital, productivity, and

labor allocation in rural Pakistan," Journal of Human Resources, 34(2): 369-

406.

[16] Fryer, R., 2011. "Racial inequality in the 21st century: the declining signi�cance

of discrimination," chapt. 10 in D. Card and O. Ashenfelter, eds., Handbook of

Labor Economics, Volume 4, Part B. Amsterdam: North Holland.

[17] Glewwe, P. and H. Jacoby, 1994. "Student Achievement and Schooling Choice in

Low Income Countries: Evidence from Ghana," Journal of Human Resources,

29(3):843-864, 1994.

[18] Go¤man, E. 1963. Stigma: Notes on the management of spoiled identity. New

York: Simon & Schuster.

[19] Ho¤, K. and P. Pandey, 2006. "Discrimination, social identity, and durable

inequalities," American Economic Review (Papers & Proceedings), 96(2): 206-

211.

[20] Jacoby, H. and G. Mansuri, 2010. "Watta Satta: Bride exchange and women�s

welfare in rural Pakistan," American Economic Review, 100:1804�1825.

[21] Khan, A., 1998. "Female mobility and social barriers to accessing health and

family planning services: A qualitative research study in three Punjabi vil-

lages," Islamabad: Ministry for Population Welfare, London School of Tropical

Hygiene and Medicine and Department for International Development, British

Government.

28

[22] Loury, G., 2000. "Social exclusion and ethnic groups: The challenge to eco-

nomics," Annual World Bank Conference on Development Economics 1999.

Washington DC: The World Bank.

[23] Margo, R., 1986. "Educational achievement in segregated school systems: The

e¤ects of �Separate-but-Equal�," American Economic Review, 76(4): 794-801.

[24] Munshi, K. and M. Rosenzweig, 2006. "Traditional Institutions Meet the Mod-

ern World: Caste, Gender,and Schooling Choice in a Globalizing Economy,"

American Economic Review, 96(4):1225-1252.

[25] Pitt, M., M. Rosenzweig, and D. Gibbons. 1993. "The determinants and conse-

quences of the placement of government programs in Indonesia," World Bank

Economic Review, 7(3):319-348.

[26] Sawada, Y. and M. Lokshin, 2009. "Obstacles to school progression in rural

Pakistan: An analysis of gender and sibling rivalry using �eld survey data,"

Journal of Development Economics, 88(2): 335-347.

[27] Srinivas, 1955. "The social structure of a Mysore village," in Village India,

McKim Marriott ed. Chicago: University of Chicago Press.

[28] White, H. and X. Lu, 2010. "Robustness checks and robustness tests in ap-

plied economics," unpublished manuscript. Dept. of economics, University of

California, San Diego.

[29] World Bank, 2005. Bridging the gender gap: Opportunities and challenges,

Pakistan Country Gender Assessment, Washington, D.C.

29

AppendixSettlement In-migration and School Availability

Migration information, while not provided in the PRHS-II village census, is

available from a di¤erent village census conducted in 2006. This census includes

158 villages (420 settlements) in �ve districts, two in Punjab (one of which overlaps

with a district in PRHS-II), two in Sindh, and one in NWFP. The census was

supported by the World Bank and the Pakistan Poverty Alleviation Fund (PPAF).

Selected villages comprise the sample for the evaluation of the third phase of the

PPAF. The census questionnaire asks about the number of years of residence in the

settlement. We also have data on all of the schools within the boundaries of each

settlement, including their year of establishment.

We construct a settlement-level panel from 1980-2006 containing the number of

households migrating into a settlement in each year and indicators for the presence

of a boys�and a girls�primary school by settlement-year. Excluding settlements in

which the �rst primary school was established prior to 1980 as well as those that

had no primary school at all up to 2006 leads to a sample of 115 settlements in 67

villages. The median total number of in-migrants over the whole period is only

6 households, but the median settlement population (as of 2006) is also just 42

households. There are 34 settlements that did not receive a single new household

since 1980. Overall, zero migrants are recorded in 79 percent of the settlement-years.

Appendix Table A.2 reports regressions of the number of in-migrants on the

presence of a boy�s or girl�s primary school that include year and settlement �xed

e¤ects. Thus, we are asking whether migration into a particular settlement (we

do not have data on out-migration) increases after a school is established, or at

least after a one-year lag. OLS estimates in column (1)-(4) provide no evidence

for this, either for boys and girls schools separately or together for any primary

school. A panel-estimator suitable for count data, the �xed e¤ect Poisson, gives

similar results in columns (5)-(8), automatically dropping all observations in the 34

settlements that received no migrants over the entire period.38

38An alternative negative binomial panel estimator (not reported), which accounts for overdis-persion, also gives similar results; i.e., in no case does the presence of a school have a signi�cantimpact on migration.

30

0.2

.4.6

.81

Pr <

 dis

tanc

e

500 1000 2000 3000 4000Meters to (nearest) school

Schl. outside settlement Schl. inside settlement

Figure 1: CDF of distance to (nearest) school for inside and outside schools.

31

0.4

.81.

2

0 .2 .4 .6 .8 1Settlement proportion of village population

fitted value of theta 95% confidence band

Figure 2: Non-parametric component of partially-linear �rst-stage regression withbootstrapped con�dence band clustered on settlement.

32

­.50

.51

prop

ortio

n ev

er e

nrol

led

0 1000 2000 3000 4000Meters to (nearest) school

Girls w/inside school Girls w/outside schoolBoys w/inside school Boys w/outside school

Figure 3: Predicted school enrollment by school distance and location.

33

34

Figure A.1: Part of a typical village in Sindh province showing household residences (purple dots) and primary school locations (green buildings).

35

Figure A.2: Non-parametric component of first-stage regression by region.

0.2

.4.6

.81

0 .2 .4 .6 .8 1Settlement proportion of village population

North Punjab South Punjab Sindh

36

Table 1 Descriptive Statistics for Sample

Boys 9-15 Girls 9-15 All Children 9-15

Region Low-caste

High-caste Total

Low-caste

High-caste Total

Low-caste

High-caste Total

North Punjab Ever enrolled 0.80 0.94 0.93 0.79 0.86 0.85 0.80 0.90 0.89

Θprimary school 0.86 0.92 0.91 0.87 0.87 0.87 0.86 0.89 0.89

CC 0.22 0.94 0.85 0.25 0.94 0.85 0.23 0.94 0.85

N 111 824 935 101 734 835 212 1,558 1,770

South Punjab

Ever enrolled 0.60 0.70 0.68 0.34 0.55 0.51 0.48 0.63 0.60

Θprimary school 0.30 0.57 0.52 0.18 0.57 0.49 0.24 0.57 0.51

CC 0.12 0.92 0.78 0.09 0.92 0.76 0.11 0.92 0.77

N 96 385 481 90 341 431 186 726 912

Sindh

Ever enrolled 0.69 0.79 0.76 0.45 0.56 0.53 0.59 0.69 0.66

Θprimary school 0.74 0.67 0.69 0.75 0.67 0.69 0.74 0.67 0.69

CC 0.56 0.88 0.78 0.64 0.89 0.82 0.59 0.88 0.80

N 334 745 1,079 257 644 901 591 1,389 1,980

All Pakistan

Ever enrolled 0.70 0.84 0.81 0.50 0.68 0.65 0.61 0.77 0.73

Θprimary school 0.69 0.75 0.74 0.66 0.74 0.72 0.67 0.75 0.73

CC 0.42 0.91 0.81 0.45 0.92 0.82 0.43 0.91 0.81

N 541 1,954 2,495 448 1,719 2,167 989 3,673 4,662

Notes: All statistics are means of indicator variables. All Pakistan refers to Punjab and Sindh provinces. Θprimary school

and CC are dummies for, respectively, whether a school is available within the child’s settlement and whether the most convenient school is in a settlement where the child’s own caste is dominant (see text for details).

37

Table 2 Descriptive Statistics for Village Census

Region

All Sindh South Punjab North Punjab

Settlement Village Settlement Village Settlement Village Settlement Village

Settlements (median no.) --- 10 --- 12 --- 11 --- 4

No. of Households (median) 24 356 19 272 25 370 101 437

Zaat/biradari groups (median no.) 2 9 2 10 3 8 4 8

Low-caste households (%) 30 29 36 34 23 23 20 22

Single Caste-type (%) 52 3 60 2 50 0 28 6

Of which, high-caste (%) 75 100 68 100 87 --- 98 100

High-caste dominant (%)

80% population threshold 78 95 70 95 81 97 96 94

60% population threshold 76 90 67 84 81 94 95 94

N 1031 165 581 61 268 32 182 72 Note: A zaat/biradari group represents a distinctly named caste/clan affiliation, whereas a caste-type is a categorization of the latter into high and low status as described in the text. Caste dominance is based on which caste-type owns the majority of the land in the settlement/village unless the other caste-type has a preponderance of population (either 60% or 80%).

38

Table 3 Settlement Boundary Effects on Primary School Enrollment

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

(5) (6) (7) (8)

Girl × θprimary school 0.189*** 0.183*** 0.185** 0.206*** 0.146*** 0.144*** 0.111*** 0.127***

(0.034) (0.034) (0.076) (0.062) (0.037) (0.038) (0.040) (0.044)

Boy × θprimary school 0.023 0.017 -0.012 -0.082 -0.026 -0.035 -0.026 -0.043

(0.027) (0.027) (0.078) (0.051) (0.029) (0.030) (0.030) (0.032)

Girl × landless

-0.089*** -0.073*** -0.090*** -0.085*** -0.086*** -0.086*** -0.101***

(0.022) (0.023) (0.022) (0.022) (0.022) (0.022) (0.023)

Boy × landless

-0.067*** -0.062*** -0.070*** -0.065*** -0.067*** -0.069*** -0.078***

(0.018) (0.021) (0.018) (0.018) (0.018) (0.018) (0.019)

Girl × ln(pce)

0.105*** 0.108*** 0.104*** 0.106*** 0.105*** 0.102*** 0.091***

(0.018) (0.019) (0.017) (0.017) (0.017) (0.017) (0.019)

Boy × ln(pce)

0.049*** 0.056*** 0.049*** 0.048*** 0.049*** 0.053*** 0.043**

(0.018) (0.021) (0.018) (0.018) (0.018) (0.018) (0.018)

Girl × θmiddle school

0.119*** 0.057

(0.038) (0.044)

Boy × θmiddle school

0.005 0.024

(0.040) (0.045)

Gender × dprimary school No No No No Linear Cubic Cubic Cubic

Gender × dmiddle school No No No No No No No Cubic

Fixed effects village village settlement village village village village Village

N 4,662 4,662 4,662 4,662 4,662 4,662 4,662 3,915

Adj. R2 0.247 0.265 0.327 0.045 0.270 0.272 0.275 0.266 Notes: Standard errors in parentheses clustered on settlement (*** p <0.01, ** p <0.05, * p <0.1). There are 492 settlements in 163 villages. Dependent variable is an indicator for whether the child was ever enrolled in school. All regressions include a girl dummy and the age of child interacted with girl and boy dummy variables. Θj

is a dummy for whether a school of type j=primary,middle is available within the child’s settlement; dj is distance to (nearest) school of type j; and ln(pce) is the log of household per-capita expenditures. a2SLS estimate using village population share of settlement and its square, both interacted with the Girl and Boy dummies, as instruments. Instrument diagnostics are as follows: Kleibergen-Paap underidentification test Wald stat = 64.8 (3); Kleibergen-Paap weak identification Wald rk F-stat = 38.1; Hansen overidentification J-test = 3.90 (2); LM test of redundancy of instruments involving squared population shares = 44.6 (4), where numbers in parentheses are degrees of freedom for chi-square statistics.

39

Table 4 Settlement and Caste Boundary Effects on Primary School Enrollment

(1) (2) (3) (4) (5) (6) (7)

Low-caste boys --- --- --- --- --- --- ---

High-caste boys 0.033 0.022 0.024 0.050 0.038 0.040 0.030

(0.030) (0.029) (0.047) (0.068) (0.080) (0.067) (0.078)

Low-caste girls -0.086 -0.575** -0.696*** -0.889*** -0.888*** -0.902*** -0.901***

(0.080) (0.229) (0.237) (0.243) (0.243) (0.243) (0.243)

High caste girls 0.009 -0.502** -0.709*** -0.877*** -0.890*** -0.895*** -0.905***

(0.082) (0.230) (0.237) (0.243) (0.248) (0.243) (0.247)

Low-caste boys × θprimary sch. -0.027 -0.069 -0.071 -0.067 -0.070

(0.053) (0.059) (0.059) (0.058) (0.058)

High-caste boys × θprimary sch. -0.035 -0.016 -0.017 -0.016 -0.016

(0.033) (0.035) (0.035) (0.035) (0.035)

Low-caste girls × θprimary sch. 0.057 0.021 0.018 0.021 0.017

(0.074) (0.079) (0.081) (0.078) (0.080)

High caste girls × θprimary sch. 0.171*** 0.188*** 0.187*** 0.189*** 0.188***

(0.042) (0.042) (0.042) (0.042) (0.042)

Low-caste boys × CC

0.142** 0.149** 0.137** 0.143**

(0.055) (0.068) (0.055) (0.065)

High-caste boys × CC

-0.051 -0.041 -0.037 -0.029

(0.051) (0.060) (0.051) (0.060)

Low-caste girls × CC

0.278*** 0.284*** 0.277*** 0.283***

(0.063) (0.076) (0.062) (0.075)

High-caste girls × CC

0.040 0.050 0.052 0.060

(0.072) (0.075) (0.069) (0.071)

Low-caste × LCD

-0.009 -0.008

(0.063) (0.059)

High-caste × HCD

0.015 0.012

(0.055) (0.055)

Caste dominance population threshold

--- --- --- 80% 80% 60% 60%

Gender × wealth No Yes Yes Yes Yes Yes Yes

Gender × dprimary sch No No cubic cubic cubic cubic cubic

Fixed effects village village village village village village village

N 4,662 4,662 4,662 4,424 4,424 4,424 4,424

Adj. R2 0.238 0.254 0.273 0.286 0.285 0.286 0.286

Notes: Standard errors in parentheses clustered on settlement (*** p <0.01, ** p <0.05, * p <0.1). There are 492 settlements in 163 villages. Dependent variable is an indicator for whether the child was ever enrolled in school. All regressions include the age of child interacted with girl and boy dummy variables. CC is a dummy for whether the most convenient school is in a settlement where the child’s own caste is dominant (see notes to Table 3 for definitions of other variables). LCD and HCD are dummies for whether low-castes or high-castes, respectively, are dominant in the child’s own settlement. Caste dominance defined by landownership up to a population threshold of 60-80%.

40

Table 5 Enrollment Impacts of Alternative School Placement Policies

School-placement policy

Low-caste targeting

Caste-blind Baseline %

ever enrolled

Low caste children

Boys 9.6 3.0 71.4

Girls 18.9 6.0 50.7

Total 14.0 4.5 61.6

High caste children

Boys 0.0 0.0 85.5

Girls 0.0 3.1 69.5

Total 0.0 1.5 77.9

All children

Boys 2.4 0.8 81.9

Girls 4.7 3.8 64.8

Total 3.5 2.2 73.8

No. of coed schools required under policy

New 93 168

Single-sex conversions 2 26

Total 95 194

Notes: Figures in the top panel, except for baseline enrollment rates, are changes in the percentage of 9-12 year-olds in a given group who ever enroll in school. Baseline enrollment rates are subgroup means from the PRHS-II sample, weighted using census data for PRHS-II villages to achieve representativeness. Likewise, simulated policy impacts are weighted by settlement population, for each caste-type as appropriate, taken from the village census. Figures in the bottom panel are numbers of schools required under either policy. See text for a description of the policies.

41

Table A.1 Descriptive Statistics for Schools

Total (N)

School Typea Caste Dominance of School Settlement

Public (%) Private (%) High (%) Low (%) No info (%)b

Panel A: Elementary-level Schools in the Census

N Punjab All Boys 179 93 6 59 3 38 All Girls 114 94 6 61 4 34 Co-Ed 228 29 68 66 1 33 Total 521 65 33 63 2 35 S Punjab All Boys 82 93 4 41 4 55

All Girls 65 95 2 51 5 45 Co-Ed 88 23 69 45 5 50 Total 235 67 28 46 4 50

Sindh All Boys 48 100 0 56 15 29

All Girls 48 92 2 63 8 29 Co-Ed 260 89 9 48 13 39 Total 356 91 7 51 12 37

All Schools 1,112 74 24 55 6 39

Panel B: Elementary-level Schools in the Sample

N Punjab All Boys 74 99 1 95 5 0 All Girls 66 97 2 88 8 4 Co-Ed 135 31 65 83 1 16 Total 275 65 33 87 4 9 S Punjab All Boys 32 94 3 71 9 20

All Girls 36 97 0 75 6 19 Co-Ed 44 36 52 70 7 23 Total 114 73 21 72 7 21

Sindh All Boys 29 100 0 73 17 10

All Girls 29 93 3 73 10 17 Co-Ed 158 95 1 68 16 16 Total 216 95 1 69 16 15

All Schools 605 77 19 78 9 13

Notes: Census data in Panel A reflect an exhaustive list of schools to which children in the 165 sample villages had access to and hence include many peripheral schools. Schools in Panel B are those nearest to the sample children or within the same settlement. aOther types of schools include community, NGO, and religious.

bSchool located outside of village boundaries, hence no caste information from village census.

42

Table A.2

School Establishment Effect on Settlement In-migration

OLS Poisson-Count

(1) (2) (3) (4) (5) (6) (7) (8)

Girls’ primary schl. established yr. t

0.065

0.195

(0.186)

(0.183)

Girls’ primary schl. established yr. t-1

0.017

0.065

(0.244)

(0.184)

Boys’ primary schl. established yr. t

0.052

-0.105

(0.198)

(0.177)

Boys’ primary schl. established yr. t-1

0.050

0.145

(0.245)

(0.184)

Any primary schl. established yr. t

0.085

-0.070

(0.130)

(0.184)

Any primary schl. established yr. t-1

0.056

0.115

(0.139)

(0.201)

Year fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Settlement fixed effects Yes Yes Yes Yes Yes Yes Yes Yes

N 3105 2990 3105 2990 2187 2106 2187 2106

Settlements 115 115 115 115 81 81 81 81

Notes: Standard errors in parentheses clustered on settlement (*** p <0.01, ** p <0.05, * p <0.1). Dependent variable is the annual number of in-migrating households in the settlement for 1980-2006.

43

Table A.3 Instrument Excludability Tests

Log land value per acre Population per acre

(1) (2) (3) (4) (5)

Proportion of village population -0.202 -0.103

-0.047 -0.117

(0.173) (0.450)

(0.040) (0.165)

Proportion of village population squared

-0.130

0.098

(0.635)

(0.189)

Proportion of plots in settlement with tubewell irrigation

0.465*

(0.248) Proportion of plots in settlement with

irrigation at head of canal

0.776***

(0.270)

Proportion of plots in settlement with irrigation at middle of canal

0.618**

(0.251) Proportion of plots in settlement with

irrigation at tail of canal

0.389

(0.252)

Proportion of plots in settlement with clay soil

0.259*

(0.135) Proportion of plots in settlement

with sandy soil

0.266*

(0.136)

Proportion of plots in settlement with chikni soil

0.158

(0.134)

Village fixed effects Yes Yes Yes Yes Yes

No. of settlements 441 441 441 1511 1511 R2 0.904 0.904 0.917 0.201 0.201 F-test (p-value) 0.245 0.490 0.007 0.232 0.412

Notes: Standard errors in parentheses clustered on village (*** p <0.01, ** p <0.05, * p <0.1). Dependent variable is average log land value per acre in settlement in cols. 1-3, number of resident households per acre of agricultural land in cols. 4-5. Regressions in cols. 1-3 are weighted by number of plots sampled in each settlement and use only PRHS-II villages.

44

Table A.4 School Quality Differences by School Settlement

School Settlement is Low Caste Dominant

80% Population Threshold

60% Population Threshold

All Schools Coed or Girls’ Schools

All Schools Coed or Girls’ Schools

School Infrastructure

0.216

0.292

0.167

0.288

(0.249)

(0.286)

(0.242)

(0.266)

Village fixed effects

Yes

Yes

Yes

Yes

Number of schools

661

516

661

516

Number of villages

160

158

160

158

Teaching Quality

-0.083

-0.082

-0.172

-0.184

(0.112)

(0.129)

(0.123)

(0.133)

Village fixed effects

Yes

Yes

Yes

Yes

Number of schools

681

531

681

531

Number of villages 161 159 161 159

Notes: Standard errors in parentheses clustered on village (*** p <0.01, ** p <0.05, * p <0.1). The sample includes all 681 primary schools that are located inside sample villages. Dependant variables are principal component indices of school infrastructure and teaching quality. The infrastructure index includes indicator variables which take the value 1 if the school has electricity, potable water, toilets, a playground, a library, a boundary wall and furniture (desks, chairs and a blackboard) in classrooms. It also includes the number of classrooms in the school building. When the sample is restricted to coed or all girls’ schools, a separate toilet for girls is added. The index of teaching quality includes the qualifications of teachers (proportion of teachers with at least a BA degree and proportion with some teacher training), the student teacher ratio, the presence of at least one female teacher, the proportion of teachers present on the day of the school visit, the number of multi-grade classrooms and the number of classes held outside.