High Educ (2016) 71:611–649

DOI 10.1007/s10734-015-9927-1

Affirmative action in higher education in India:

targeting, catch up, and mismatch

Veronica Frisancho1 • Kala Krishna2

Published online: 24 July 2015

 Springer Science+Business Media Dordrecht 2015

Abstract Using detailed data on the 2008 graduating class from an elite engineering
institution in India, we evaluate the impact of affirmative action policies in higher edu-
cation focusing on three issues: targeting, catch up, and mismatch. We find that admission
preferences effectively target minority students who are poorer than average displaced
nonminority students. Moreover, we find that minority students, especially those in more
selective majors, fall behind their same-major peers in terms of grades as they progress
through college. We also identify evidence in favor of the mismatch hypothesis: Once we
control for selection into majors, minority students in more selective majors end up earning
less than they would have had if they had chosen a less selective major.

Keywords Affirmative action  Higher education  Targeting  Catch up  Mismatch

JEL Classification I23  I24  J15  J31  J71

We would like to thank Pradeep Kumar for answering many of our questions about the context. We are also
indebted to participants of the 7th Annual Conference on Economic Growth and Development at ISI Delhi,
the mini conference on the effects of racial preferences in higher education on student outcomes at The
Brookings Institution, and seminar participants from the Economics and Social Sciences area at the Indian
Institute of Management at Bangalore for suggestions and comments.

& Veronica Frisancho

vfrisancho@[Link]
Kala Krishna
kmk4@[Link]
1
Research Department, Inter-American Development Bank, 1300 New York Ave. NW, Washington,
DC 20577, USA
2
Department of Economics, The Pennsylvania State University, 523 Kern Graduate Building,
University Park, PA 16802, USA

123
612 High Educ (2016) 71:611–649

Introduction

Affirmative action (AA) policies in higher education are used in many countries to try
to socially advance historically disadvantaged groups. Although the underlying social
objectives of these policies are rarely criticized, there is intense debate over the actual
impact of minority preferences in higher education on educational performance and
labor outcomes. The debate has mainly focused on three issues: targeting, mismatch,
and catch up.
It is well known that family income is a strong predictor of performance. Thus, there is
great concern about the fairness of targeting based on race, ethnicity, or caste rather than
on income. If admission preferences only allow richer students within the minority group
to traverse the (lower) hurdles required for admission, then they may be displacing poor
students from the non-minority or general group. This is also called the ‘‘creamy layer
problem’’ in India.
The second issue is catch up. Students admitted to college under preferences often start
off far behind those admitted under regular admission criteria. But how does the gap
between these two groups change as both progress through college? Do they catch up or
fall further behind? If those admitted under preferences can catch up, even part of the way,
then the case for preferences is clearly stronger than if they fall further behind.
Opponents of AA also claim that the actual gains for the intended beneficiaries of the
policy may not exist. In the extreme case, minority students may even be worse off if they
are unprepared for the academic environment they obtain access to through the policy. This
argument is known as the mismatch hypothesis: Students who do not qualify for ordinary
admission would do better if they enrolled at schools and/or majors which are more in line
with their credentials. If there is severe mismatch, then preferences may even do more
harm than good.
Most of the studies to date are narrowly focused on the effects of AA on USA
minorities’ college performance and labor outcomes. The USA, we think, is a poor setting
in which to look for such evidence. In most USA higher education settings, selection
criteria are relatively nebulous. While institutions do want good students, they pay
attention to much more than grades or SAT scores in deciding whom to admit.1 SAT
scores, extracurricular activities, essays, alumni ties, interviews, the perceived likelihood
of the student coming,2 and donations, all matter. Moreover, AA policies in the USA are
themselves relatively nebulous: Even in their heyday, they basically consisted of adding
some ‘‘points’’ for race. There were rarely quotas or large and well-documented differences
in admission standards.3 Finally, American students have a huge amount of choice over
courses while in college. For example, if smart/serious students take harder courses in
which good grades are more difficult to obtain, while poor students take the ‘‘gut’’ courses
in which an A- is ensured with minimal effort, then grades may provide little information
on actual academic performance. That this factor is important is clear from the recent work
by Arcidiacono et al. (2011). They show that controlling for selection into majors virtually

1
This is partly explained by a large number of people having close to or perfect SAT scores. The best
schools could easily fill their seats with only such candidates. However, based on the USA experience, there
is reason to believe that this would result in a worse entering class (Blau et al. 2004; Bowen and Bok 1998).
2
Admissions officers are often rewarded on the basis of acceptance rates.
3
The Texas top 10 % law, which guaranteed admission to the top 10 % of graduates from all Texas high
schools to any state or public university may be one of the few exceptions. See the Texas Higher Education
Opportunity Project (THEOP) for more on this. The law was loosened in June 2010.

123
High Educ (2016) 71:611–649 613

eliminates any convergence in the black–white performance gap in college. For all these
reasons, the USA may not be the best place to evaluate the effects of AA.
We argue that other countries, with transparent selection criteria and rigid course
structure, provide a much more fertile ground to evaluate the effect of AA policies on
minority students. The evidence presented here is particularly important due to its focus on
India, which provides a better setting than the USA. In India, admission criteria are clear:
Performance in an open admission exam or in the school leaving exam is all that matters.
Moreover, admission preferences imposed by AA in India are far greater than the ones
given to African American or Hispanic applicants in the USA. India has very strict and
binding quotas in higher education in favor of scheduled castes (SC) and scheduled tribes
(ST). These groups include what were known as the ‘‘untouchable’’ castes, which used to
be relegated to the most menial occupations, as well as tribal populations who were
isolated from the mainstream and often treated as badly as the SC.4 The quotas result in
very large differences in admission standards so that the empirical results are not con-
founded by the program being a marginal one.5 Our focus on India also helps us overcome
selection problems present in USA college data. Most higher education institutions in India
have a very strict curriculum which minimizes the issue of self-selection into easier
courses. In this setting, grades are a very good indicator of college performance.
Using detailed data on the 2008 graduating class from an elite engineering institution
(EEI) in India, this article tries to cast some light on the effects of AA on Indian minorities.
In particular, we look at income and grade distributions of minority and non-minority
students at the EEI to provide some basic evidence on targeting. We find that SC/ST
students are in general poorer than other students at the EEI. Using a supplementary data
for all the applicants to a group of elite engineering colleges in India in 2009, we show that
AA seems to be effectively targeting minority students who are poorer than the average
displaced general (GE) student. By analyzing the college performance of minority and non-
minority students as they progress through college, we find no evidence of catch up: SC
and ST students actually seem to fall behind their same-major peers and this is more so
when they choose selective majors where they are further behind the rest of the class.
However, this does not mean that SC/ST students would learn less than the general stu-
dents when retaking a test at a given level of difficulty. Frisancho et al. (2014) in fact show
that underprivileged students improve more than those from better backgrounds when
retaking the same exam.
We also test the mismatch hypothesis using labor market outcomes and students’ self-
reports on emotional and social well-being while at the EEI. Without controlling for
selection into selective and nonselective majors, it looks like students in more selective
majors earn more. Propensity score matching methods that control for selection in
observables reduce the estimated effect of major selectivity on wages, though it remains
positive and significant for general students. However, if the wage and selection equations
are jointly estimated to take into account the role of unobservables, the positive effect
among general students goes away, suggesting that it was driven by selection. In other
words, general students earn more and choose more selective majors because they are
better in terms of unobservables. Even more interesting, minority students in more

4
Lower castes in India represent a greater share in total population than any minority in the USA. Even if
we only consider the most disadvantaged castes, SC and ST, their 22.5 % share surpasses the 13 % share of
African Americans in the USA.
5
In fact, the quotas are so much in favor of these disadvantaged groups that even with huge differences in
admissions cutoffs, some elite schools are not able to fill their quotas.

123
614 High Educ (2016) 71:611–649

selective majors end up earning significantly less than their same-race counterparts in less
selective majors, which supports the mismatch hypothesis. We also identify some evidence
in favor of social mismatch: even after controlling for selection, being enrolled in a more
selective major increases stress levels and feelings of not belonging among SC/ST stu-
dents, but the effect goes in the other direction among general students.
A note of caution is worth mentioning here. Our results on catch up and mismatch are of
course measuring the effect of affirmative action on minority students ex post, that is,
conditional on being accepted into the EEI. Unfortunately, the characteristics of the data
we have limit what we can say about the unconditional effects of affirmative action.
Our paper is also relevant for the USA, where the case of Fisher v. the University of
Texas has revived the debate around affirmative action. During a 5-year period of rulings
that started in the lower courts and continued up to the Supreme court, the pendulum has
swung in favor of and against race-based admission preferences. Our work suggests that
extreme policies of this kind may not be effective as minority students seem to fall behind
rather than catch up, and more so when the initial academic gap between them and non-
minority students is larger.
The rest of the paper proceeds as follows. Section 2 sketches out the major findings in
the literature so far. Much of this work is on the USA and is plagued by data problems.
Section 3 describes the EEI context and the reservation policies mandated by the Indian
government in higher education admissions. Section 4 describes the data. Section 5 pre-
sents the evidence we have put together on targeting, catch up, and mismatch. Section 6
concludes and describes the limitations of the study, as well as directions for future
research.

The evidence to date

AA policies are meant to help historically disadvantaged minorities. However, if only

richer students within the minority group benefit from AA preferences, displacing poor
students from the advantaged group, the fairness of the policy comes under scrutiny. In the
USA, for example, the use of race as a proxy for income when targeting the poor has been
strongly questioned over the last decade. The current debate focuses on a shift from race-
based to economic-based affirmative action policies as proposed by Kahlenberg (2004) and
Kahlenberg (1996) among others. Even though it is true that racial diversity has increased
in top colleges in the USA, income inequality may have increased as competition to get in
has, by all accounts, intensified. Carnevale and Rose (2003) find that 74 % of the students
at the top 146 colleges in the USA came from families in the richest economic quarter,
while only 3 % came from the least advantaged quarter. There is also some indication that
AA policies that favor African Americans have disproportionately benefited richer
minority students. At the 28 selective colleges studied by Bowen and Bok (1998), 86 % of
African Americans were middle- or upper-class students.
In India, however, work by Bertrand et al. (2010) suggests that affirmative action
successfully targets financially disadvantaged students applying to engineering colleges.
Even though caste-based targeting did not benefit the poorest SC/ST students, admission
was successfully reallocated from richer to poorer households. In particular, upper-caste
applicants displaced by AA were richer than the lower-caste applicants taking their place.
Brazil also provides a nice setting to evaluate the effects of affirmative action in college
admissions. In 2012, the country introduced the ’’Law of Quotas’’ in federal universities.

123
High Educ (2016) 71:611–649 615

Under this policy, universities were given 4 years to ensure that half of their admitted pool
of students came from public schools and that the racial composition of each incoming
class was representative of the racial mix in the local population. The law was unanimously
supported by the Senate since it was only validating the independent efforts of several
states that had been implementing racial and income quotas in public universities since
2002. Francis and Tannuri-Pianto (2012) estimate the effect of a voluntarily introduced 20
% quota for students who self-identify as black in the University of Brasilia in 2004. They
find that racial quotas raised the proportion of black and dark-skinned students at
University of Brasilia and that displacing applicants came from families with significantly
lower socioeconomic status than displaced applicants.
Besides targeting, two related issues plague the debate on the appropriateness of AA
policies: catch up and mismatch. In India, where quotas in some states reach 50 % of
college admissions, a large burden comes from the expected negative effect on the average
quality of students graduating from higher education institutions. If colleges are forced to
admit students from scheduled castes until the quota is met, large reductions in students’
average ability are expected. The magnitude of these reductions will be determined not
only by the level of the quota, but also by the initial differences in performance between
general and minority students (Kochar 2010).
Evidence from Brazil shows that AA preferences may even reduce students’ incentives
to exert effort in admission performance. Assuno and Ferman (2013) analyze the intro-
duction of large college admission quotas benefiting blacks and public school students in
three top universities in Brazil. Between 2002 and 2004, two schools in Rio de Janeiro and
one in Bahia were the first to implement state-level admission quotas which provides a
natural experiment setting. The authors find that whenever the quota had a larger relative
impact on the likelihood of begin accepted, i.e., black students in Rio de Janeiro, the
proficiency of the intended beneficiaries decreased by 6 % in relation to their performance
prior to the introduction of the AA admission policy. Other groups of favored students but
exposed to a less stringent policy due to higher existing levels of minority representation at
the baseline were not affected in terms of their incentives to accumulate human capital.
Nevertheless, Francis and Tannuri-Pianto (2012) show that the quotas introduced in Bra-
silia do not seem to reduce preuniversity effort among potential beneficiaries. Saeme
(2014) analyzes the introduction of racial admission quotas in a public research university
in Sao Paulo in 2011. The university implemented a 40 % quota for public school students
and mandated that 35 % of those seats had to be reserved for black students. By con-
structing a comparison group with states that were not affected by quotas, she finds that the
quotas generate a significant one-percent increase in admission test scores among bene-
ficiaries of the policy. However, the author is unable to rule out that other changes in the
educational setting between 2009 and 2011 may have changed the pool of applicants over
time.
However, if minority students catch up while in college, this particular cost of AA
policies can be greatly reduced. Alon and Tienda (2007) creatively use the 10 % rule in
Texas (where the top 10 % of the graduating class in public high schools was ensured
admission to UT Austin) to argue that the likelihood of graduation rose after the policy was
implemented, and did so significantly for blacks. Their evidence is consistent with previous
studies that find that those admitted under the top 10 % rule outperform those who were
admitted with a lower class rank but higher SAT scores, suggesting that SAT scores are not
a good indicator of college performance and that there may well be considerable catch up.
However, those admitted from worse schools may be less prepared for college and thus
more likely to choose less challenging majors. If this is the case, graduation rates per se

123
616 High Educ (2016) 71:611–649

may be less than fully informative. On the other hand, Sander (2004) finds that the average
performance gap between blacks and whites at selective law schools is large and, more
importantly, tends to get larger as both groups progress through college. He also finds that
boosting black applicants into more selective schools lowers their probability of graduation
mostly though reduced grades.
It is clear that minority students targeted by AA policies have initial academic cre-
dentials that are significantly weaker than those of their non-minority peers. If minority
students are not able to close the gap, AA policies that allow them into more selective
colleges and/or majors may end up hurting them. If minority students attending more
selective schools due to AA policies obtain lower grades than the ones they would have
obtained in less selective environments, their labor market outcomes could be worsened by
admission preferences.
Attempts to empirically evaluate the ‘‘mismatch hypothesis’’ in the USA provide mixed
evidence. Rothstein and Yoon (2009) and Sander (2004) find evidence of mismatch in law
school. Loury and Garman (1993) and Loury and Garman (1995) find that blacks in the
USA get considerable earning gains from attending more selective schools, but these gains
are offset for black students by lower performance both in terms of grades and probability
of graduation. Alon and Tienda (2005) assess the effect of college selectivity on graduation
probability. Using both propensity score matching methods and bivariate probit models,
they reject the mismatch hypothesis suggesting that blacks and Hispanics in the USA are
able to catch up. Arcidiacono (2005) estimates a structural model that incorporates
application decisions, admission decisions, attendance decisions and future earnings. He
argues that removing affirmative action reduces the presence of minority students, espe-
cially in top schools, but it does not affect income or college attendance by much.
Bertrand et al. (2010) use Indian data and find that the marginal effect of caste-based
admission preferences in Indian engineering colleges is positive for minority students: i.e.,
they do earn more as a result. However, they gain less than what the students they displace
lose. Though their data are better than ours as they have information on accepted and
rejected students, they have no information on grades, which account for a large part of the
differences in earnings. At the very least, their results are unable to distinguish between the
pure gains from graduating from more selective institutions and the loss arising from
poorer grades in these institutions.
In Brazil, Francis and Tannuri-Pianto (2012) rule out a mismatch effect of the AA
policy by showing that the racial gap in grades in selective programs does not widen.
Unfortunately, they lack data on wages to further confirm this hypothesis. Another limi-
tation of their work is that the policy is bound to have small effects because the magnitude
of the quota compared to the equilibrium share of black students already applying and
being admitted to the University of Brasilia is marginal. Alon and Malamud (2014) study
the introduction of class-based quotas in Israel in the four most selective universities
between 2001 and 2006. Eligibility for the program was determined by student’s neigh-
borhood of residence, high school of origin, parental education, family size, and individual
or family adverse circumstances, variables which were then weighed in an algorithm. This
allows them to use a regression discontinuity design to measure the causal effect of the
policy on student outcomes. The authors find that the preferences increased both the
probability of admission by 13 % and the probability of being admitted to a selective major
by 30 % (though the probability of applying to a more selective major was not affected).
Moreover, beneficiaries of the policy were also more likely to enroll in university. Within
the restricted sample of students who enrolled in university, the authors identify little
support for the mismatch hypothesis in this context. They find no effect of the policy on

123
High Educ (2016) 71:611–649 617

grades, neither in the first year nor at graduation, and show that the causal impact of the
preferences on graduation and graduation from selective majors is insignificant.

Admission process and reservation policies

Admissions to undergraduate programs and some graduate programs offered by the EEI are
conducted through a national examination that is also used by other engineering colleges.
The admission process is very competitive because of both the difficulty of the open
competitive exam and the large number of test takers who tend to be the very best students.
The undergraduate acceptance rate is less than 1 in 60: Over 300,000 annual test takers
compete for 5500 seats in undergraduate programs.
The entrance exam tests the candidate’s knowledge of three subjects: Chemistry,
Mathematics, and Physics. After the exam is administered, the average of the marks scored
by all candidates is computed for each of the three subjects. These averages give the
Minimum Qualifying Marks (MQM) for Ranking in each subject. All students above the
MQM in each subject are ranked in terms of their aggregate score to construct a common
merit list. This merit list contains as many students as the number of seats available in all
undergraduate and graduate programs offered by the colleges that use the exam for
admission. The aggregate score of the last candidate admitted from this list gives the
general cutoff score for admission.
Although minority students take the exam with the rest of the students, India’s law
entitles them to preferences. In traditional Hindu society, caste is hereditary and it used to
be occupation-specific. Thus, lower-caste individuals were trapped in less attractive
occupations, in terms of both prestige and wages.
According to the 2011 Census, SC and ST constitute 16.2 and 8.2 % of the Indian
population, respectively. Until this data collection effort, it was very difficult to get reliable
socioeconomic statistics by caste due to misreporting of the latter.6 Estimates from the
2000 National Sample Survey show that, in general, SC/ST are considerable poorer than
India as a whole, both in urban and in rural areas. While national urban poverty rates were
about 27 % in that year, 38 and 35 % of the SC and ST population lived below the poverty
line, respectively. In rural areas, where the national average was 27 %, the poverty rate of
SCs is also relatively higher at 36 % while STs fared even worse with 46 % below the
poverty line.
Although reservation policies in India were first applied in labor markets, they soon
appeared in higher education as a way to reduce the inequalities generated and perpetuated
by the caste system. According to the latest census, 10 % of the households in India do not
have a single literate person living with them, irrespective of household size. Relative to
the national average, SC and ST are clearly worse off, with 13 and 18 % of the households
in these groups only with illiterate members, respectively. The National Family Health
Survey from 1998—1999 reports that 21 and 30 % of SC and ST children aged 7 to 17 had
never attended school, respectively. This rate is below 8 % for other children in the
country. In 2012, the Ministry of Human Resource Development conducted the All India

6
In the past, there was an incentive to underreport belonging to SC/ST castes due to social status concerns.
Nowadays, there seems to be a tendency to overreport SC/ST status in order to capture government subsidies
and benefits targeted toward these historically disadvantaged populations.

123
618 High Educ (2016) 71:611–649

Survey on Higher Education.7 Though the country has registered a modest increase in the
gross enrollment ratio from 15 % in 2009–2010 to 19 % in 2010–2011, SC/STs populations
are still very much excluded from tertiary education. While other backward classes (OBC)
and general students comprise 86 % of student body in higher education institutions, only
10 and 4 % of the students in higher education institutions come from SC and ST,
respectively.
After independence from Britain, all central government higher education funded
institutions were mandated to comply with reservation requirements for traditionally dis-
advantaged castes. In particular, the central government requires that 15 % of the students
admitted to universities must be from SC while 7.5 % have to come from ST, reflecting
their share of the general population.8
To comply with the affirmative action policies imposed in higher education admissions,
the EEI has implemented caste-based reserved quotas since 1973. After the common merit list
is constructed, separate merit lists for SC and ST candidates are prepared. If the number of
candidates in each minority list is at least 1.4 times the number of seats available for that caste,
the merit list contains all these candidates. If the number of qualified minority candidates is
less than 1.4 times the number of available SC or ST seats, the general admission cutoff is
reduced by up to 50 percentage points to get the number of candidates as close as possible to
1.4 times the number of seats. Thus, if the general cutoff is 97 %, the SC/ST cutoff could be as
low as 47 %. However, even after extreme relaxation of the cutoff scores for minority
students, the aggregate quota of 22.5 % for SC and ST students is not always met.9
Each program at the EEI or at any other higher education institution that uses the
centralized exam to regulate admissions offers a fixed number of seats for general, OBC,
SC, and ST students. Once a student qualifies into the relevant merit list, he/she submits a
preference ranking over majors and colleges. Within each merit list, exam scores can be
thought of as bids for a particular program. Placement is offered until the reserved number
of seats for a caste group is filled or until all applicants in the corresponding merit list are
placed. Ex post, this system generates major and caste-specific cutoff scores. More pres-
tigious majors with higher salaries in the labor market tend to be more competitive and
hence have higher exam cutoff scores for both general and minority students. This allo-
cation process generates assortative matching within each major: Top students from the
general group are matched with top students in the minority group. However, as the quotas
are imposed on a major by major basis and the aggregate quota is not filled, SC/ST students
are more likely to be in selective majors so that the difference in performance between SC/
ST and general students tends to be greatest in selective majors. Note that once at the EEI all
students are evaluated under the same criteria in terms of grades and graduation require-
ments though SC/ST students are allowed to and usually do take longer to graduate.

7
The sample includes 448 universities, 8123 colleges, and 4076 stand-alone institutions who voluntarily
decided to answer the survey.
8
Reservations for OBC were recommended in 1978 and implemented in 1989 in private unaided institu-
tions as well as high-end government jobs for minority communities. The EEI we analyze did not make any
changes to its reservations policy until 2008. Since then, OBCs have also been provided with a 27 %
reservation, although their share in India’s population is about 50 %. However, it has been argued that the
OBC group is not ‘‘backward’’ and some privileged castes have made it on to this list.
9
Since 2008, a separate merit list is also constructed for OBC students. However, the relaxation in marks
applied to the admission cutoff for this group is at most 10 %.

123
High Educ (2016) 71:611–649 619

Data

The sample of students used in this study corresponds to the graduating class of 2008 from an
EEI, which includes 354 Bachelor students and 97 Dual Degree students.10 The data set
contains institutional records and some background information obtained from an exit survey
administered by the EEI to all graduating students.11 The institutional records contain data on
GPA and number of credits completed at the EEI on a semester by semester basis, as well as
some basic information on the students such as gender, caste, age, and major. Additionally,
the exit survey data provide information on previous schooling, family income, land, and
property ownership, parents’ and siblings’ educational levels and occupations, expenditures
in coaching to prepare for the admission exam, as well as some information about placement,
such as type of job and first salary after graduation. The survey also collects detailed infor-
mation on academic life experience, hostel life at the EEI12, and extracurricular activities. See
Table 3 in Appendix 1 for basic descriptive statistics.
A major limitation of the data is that it does not include students’ scores in the entrance
exam as a precollege performance measure. However, this problem can be partly cir-
cumvented by using the cumulative GPA (CGPA) of the student at the end of the first year
in the EEI. This is a good proxy for the exam score because the courses taken during the
first year of Bachelor’s and dual degree programs share a common structure across majors
and the material covered closely reflects the material evaluated in the exam.13 In what
follows, we define major selectivity based on students’ average performance in the major
in the first year.14
The second data limitation is the large number of students with missing values in one or
more of the variables obtained from the survey. Only 56 % of the sample has complete data
for all the 31 variables we use. Almost 36 % of the students have missing values in one or
two variables, and the remaining 8 % of the sample has between 3 and 14 variables with
missing values. However, most of the variables have missing values for a few individuals
(see Table 11 in Appendix 2). The variables with the greatest percentage of missing data
points are wages after graduation (10 %) and type of school attended (14.8 %), but even in
these cases the problem is not severe. To avoid dropping variables or observations, we rely
on multiple random imputation methods to generate an imputed complete data set.
Assuming the data are missing at random (MAR), a common assumption in most impu-
tations methods, the completed data set is obtained using sequential generalized regression
models. Tables 12 and 13 in Appendix 2 present the results of models that estimate the
probability of having missing data in one of the more problematic variables. Observables
explain very little, which confirms that the patterns of missing data are in line with the
MAR assumption. A complete description of the imputation procedures used is given in
Appendix 3.

10
Dual degree programs integrate undergraduate and postgraduate studies in selected areas of specializa-
tion. They are completed in 5 years, only one more year compared to conventional Bachelor’s degrees.
11
The survey was only administered in 2008. All individual records were anonymized prior to the analysis.
12
Throughout their stay in the EEI, students live in hostels located in campus.
13
Common courses usually include basic Electronics, Mechanics, Chemistry, and Physics.
14
Out of the nine bachelor majors and five dual degree majors offered at the EEI, only the Bachelor’s
programs in Computer Science and Electrical Engineering (Power) as well as dual degrees in Computer
Science and Electrical Engineering are defined as selective. See Table 4 in Appendix 1 for descriptive
statistics on the average performance in the first year and wages by major and caste.

123
620 High Educ (2016) 71:611–649

We also rely on a supplementary database on the scores of the universe of applicants

who took the centralized entrance exam used for admission to the EEI in 2009 (384,977
students). These records were obtained from the exam organizer’s Web site in September
2010. In addition to each applicant’s aggregate score and scores by subject, the records
contain information on the caste and place of residence as indicated by the All India Postal
Index Number (PIN). Using the PIN code identifiers, we merged the exam applicant data
with district-level poverty data in urban areas as well as data on the share of SC/ST
population and the share of rural population.15

Empirical evidence

Targeting

AA policies are usually implemented for two main reasons. First, governments may
directly target minority groups pursuing objectives related to diversity, social harmony,
and social advancement of historically excluded groups. Second, race (or caste in India)
can be a substitute for income to target social interventions. As mentioned above, minority
groups in India have been historically trapped in less prestigious occupations with lower
wages. Consequently, the government adopted caste-based targeting policies as a substitute
for income-based targeting policies with the hope that the former was a more efficient way
to identify disadvantaged groups as backward caste status is harder to falsify than
income.16 Verifying income is especially difficult in India since 92 % of the labor force
works as informal workers (NCEUS 2009).
Nevertheless, race- or ethnicity-based targeting strategies raise questions about the
fairness of the preferences. Although it is true that the proportion of people below the
poverty line among scheduled castes and tribes is about 50 % higher than that among the
general category (Chakravarty and Somanathan 2008), a common argument against AA is
that low-caste applicants may be far richer than the average low-caste household. Even
worse, advantaged minority students may take slots away from poorer students from the
non-minority group.
To address targeting issues, one needs to compare the background characteristics of
displaced general students to displacing minority students. Although the data from the
EEI’s 2008 graduating class provides us with detailed background information for ad-
mitted EEI students, we do not observe the background characteristics of displaced
general students who were not admitted in 2003 or 2004. Nevertheless, Figs. 1 and 2
provide some evidence in our data on the existence of a creamy layer effect at the EEI
resulting from minority quotas.
Figure 1 shows that minority students admitted to the EEI are on average poorer than
students from the general category (non-SC/ST). However, there are a few lower-caste
students who come from advantaged backgrounds. If we look at the mean CGPA on
entrance and at graduation, we notice that, on average, these ‘‘advantaged’’ students from
the ‘‘disadvantaged’’ group have much higher CGPAs both in the first year and in the
remaining periods. While there is at best a weak positive link between CGPA and family

15
Poverty data come from Table A2 in Chaudhuri and Gupta (2009), while data on minority and rural
population at the district level are obtained from the Indian Census 2001.
16
SC/ST status is documented by certificates issued by the Indian government. Given the widespread
income tax evasion in India, income is likely to be underestimated, especially in nonsalaried employment.

123
High Educ (2016) 71:611–649 621

(a) (b)

.25
.25

10
10

.2
.2

% Students
% Students

8
8

.15
.15

CGPA

CGPA
.1
.1

6
6

.05
.05

0
0

4
4
<1 lac 1−2 lac 2−3 lac 3−4 lac 4−6 lac 6−9 lac9−12 lac>12 lac <1 lac 1−2 lac 2−3 lac 3−4 lac 4−6 lac 6−9 lac9−12 lac>12 lac
Annual Family Income Mean Initial CGPA Annual Family Income Mean Initial CGPA
Mean Final CGPA Mean Final CGPA

Fig. 1 Is there a creamy layer effect? a General category, b SC/ST. Source: Survey data from the EEI’s
graduating class, 2008. Note: 1 lac = 100,000 rupees = US$ 1,572 at an exchange rate of 63.6 rupees per
dollar

(a) (b)
.8

.8
.6

.6
Density

Density
.4

.4
.2

.2
0

4 5 6 7 8 9 10 4 5 6 7 8 9 10
First Year CGPA First Year CGPA
Low Income (<3lac) Middle Income (3−6 lac) Low Income (<3lac) Middle Income (3−6 lac)
High Income (>6 lac) High Income (>6 lac)

Fig. 2 First Year CGPA distribution by caste and income level. a General category, b SC/ST. Source:
Survey data from the EEI’s graduating class, 2008. Note: 1 lac = 100,000 rupees = US$ 1572 at an exchange
rate of 63.6 rupees per dollar

income for general students, there is a clear jump in CGPA for higher-income minority
students. Rich SC/ST students perform much like their counterparts in the general cate-
gory, while poor ones look very different, suggesting that it is the interaction of poverty
and SC/ST status that is most harmful.
Figure 2 is consistent with this insight. Panel (a) shows that higher-income students
from the general group have slightly higher average grades. However, both first-order and
second-order stochastic dominance of the distribution of richer general students is rejected.
Panel (b) in Fig. 2 shows that high-income students among the SC/ST have much higher
grades than low income ones. In addition, while the general category has some rich
students with poor GPAs, the SC/ST does not. A Kolmogorov–Smirnov type test cannot
reject the null of second-order dominance of the grade distribution for rich SC/ST students
over that for poorer minority students, which is consistent with the higher mean and lower
dispersion for the rich observed by eyeballing the data.
Results from a regression of initial CGPA on income, caste, and controls confirm that
the interaction between high income (above 9 lacs or 900,000 rupees) and SC/ST is still

123
622 High Educ (2016) 71:611–649

positive and significant even after adding additional regressors and despite very few SC/ST
students in the highest income group (see Table 5 in Appendix 1). These results confirm the
evidence presented above: Poor minority students start off lagging behind general and rich
minority students. Grades are consistently lower among SC/ST students when compared to
general category students, but higher-income levels seem to ‘‘undo’’ the effect of caste.
The better socioeconomic conditions that these students face probably allow them to have
similar opportunities to those of general students, which makes an important difference in
terms of performance. In fact, rich minority students look more or less the same as general
students in terms of grades.
To summarize, better-off minority students who benefit from caste-based preferences
represent a very small proportion of all the SC/ST students admitted into the EEI (see
histogram in Fig. 1) and they look like students in the general category. These two facts
suggest that the extent of the creamy layer problem at the EEI might be small. However,
we must keep in mind that we are only comparing admitted students across income groups
and caste and we know nothing about general applicants who are not placed due to the
minority preferences.
Fortunately, we can learn more about targeting by using the 2009 entrance exam
applicant database.17 Given that the allocation of students through the centralized
admission system is implemented using caste-specific merit lists, we can evaluate the
targeting properties of the AA policy by comparing the economic background of minority
students admitted due to the preferences (‘‘displacing’’ students) to that of students who are
denied a seat but who would have been included in a global merit list in the absence of the
preferences (‘‘displaced students’’).
Though the 2009 applicant data lack of information on final college placement, we
define displaced and displacing groups using the information on the students in each caste-
specific merit list as well as the number of seats available for each caste group. In 2009, all
colleges using the centralized exam to regulate admission offered 8295 seats.18 Out of
these seats, 4784 were assigned to general students, while 1594, 1282, and 635 were
reserved for OBC, SC, and ST students, respectively. We construct the displaced GE group
with all general students who would have obtained a seat in a college (including the EEI) if
the 8295 seats were allocated without caste-specific preferences. These are general students
who made it into the common merit list but who were ranked worse than the 4784th
general student in this list and so ended up without a seat. The displacing SC/ST group
consists of two groups: (i) SC students who did not qualify into the common merit list but
who are better ranked than the 1282nd student within the SC merit list (967 applicants) and
(ii) ST students who did not qualify into the common merit list but who are better ranked
than the 635th student within the ST merit list (208 applicants). In 2009, 93 % (92 %) of
the SC (ST) applicants in the SC (ST) merit list are displacing students.19
Assuming that neither the number of applicants in each caste nor the number of seats
available would change in the absence of the AA policy, we compare district poverty rates
across displaced and displacing groups to evaluate the targeting properties of the program.
Figure 3 plots the cumulative distributions of the percentage of poor in urban areas in the

17
See Appendix 4 for more details on district-level patterns identified in these data.
18
Additionally, 251 seats were reserved for students with physical disabilities, but since we do not have
these applicants in the applicant data we exclude them from the analysis.
19
Although we might mislabel some applicants due to non-enrollment after placement, we expect this
potential bias to be small. Colleges who allocate seats using this centralized exam are top higher education
institutions in the country so students who are offered a seat in one of them are not very likely to reject it.

123
High Educ (2016) 71:611–649 623

Fig. 3 Does AA target students

1
from the poorest districts?
Source: centralized entrance

.8
exam, applicant data 2009

Cum. Density
.6
.4
.2
0 0 20 40 60 80 100
% Poor in district of residence (Urban)
Displaced GE Displacing SC/ST

applicants’ districts of residence for both groups. We find that displacing minority students
come from poorer districts compared to displaced general students, which suggests that the
use of minority admission preferences at the EEI is effectively redistributing educational
opportunities to students who live in poorer districts. However, lack of individual data on
income or consumption in the 2009 applicant database constrains us to compare living
standards of displacing SC/ST’s households to that of the average minority household.

Catch up

The use of minority preferences on college admission in India is justified by the under-
representation of SC and ST students in higher education institutions. AA policies were
introduced as a way to alleviate the legacy of the caste system and allow SC and ST groups
to catch up to the educational and labor outcomes of upper castes. In turn, opponents of
affirmative action argue that reservation schemes undermine the quality of education since
the lack of preparation of minority students during childhood cannot be amended at such a
late stage. They argue that minority students who benefit from preferential admission start
off college far behind students accepted under regular criteria and that competition while in
college will then translate into poor performance among SC and ST students.
In fact, data on test results for the universe of students taking the exam in 2009 reveal a huge
difference in initial caste-based academic levels. Although we cannot observe final placement,
we can again proxy admission status by using information on the merit lists. We approximate
the group of entrants by: those general students who made it into the merit list and who were
ranked better than the 4784th general student and those SC, ST, or OBC students who made it
into their caste-specific merit lists. Figure 4 shows that the lowest performance decile among
potential entrants is entirely composed by SC/ST students. Besides an important share of 0.45 in
the second decile, SC/ST students are barely represented in higher deciles of the score distri-
bution. In turn, 90 % of the students in the top decile come from the general category while their
share is between 0.80 and 0.86 in the top 7 deciles. Participation of backward castes in the
common merit list is just as striking: out of the 8,295 students who made it into the common
merit list in 2009, only 78 come from SC while 15 come from ST.

123
624 High Educ (2016) 71:611–649

1
.8
Proportion of students
.6
.4
.2
0

1 (lowest) 2 3 4 5 6 7 8 9 10 (highest)
GE OBC SC/ST

Fig. 4 Proportion of general and SC/ST students by 2009 entrance exam score deciles. Source: centralized
entrance exam, applicant data 2009

But how is the educational gap between general and SC/ST students changing once at
the EEI? First of all, what do we mean by educational gap? One approach might be to look
at how these two groups fare at the end of their time at the EEI relative to the beginning. In
fact, if one compares the cumulative GPA of minority and non-minority students at the end
of the first year and by the end of their programs, net of their initial performance, it looks
like SC/ST students are catching up. For both SC/ST and the GE groups, the CGPA
distribution at the end seems to improve relative to that in the first year, and more so for
minority students. By this measure, the gap in average CGPA between general and SC/ST
students shrinks by 15 %. However, this pattern could also be explained by more lenient
grading in later years, especially in the majors chosen by SC/ST. In this case, GPA
improvements by SC/ST students may not translate into class rank improvements.20
Thus, if we are interested in catch up, we should not be looking at the absolute grades,
but at the relative grades within each program. In fact, Arcidiacono et al. (2011) show that
although the GPA gap between white and black students at Duke falls by half between the
first and the last year of college, this comes primarily from smaller variance in grading
during later years and a higher proportion of black students switching into easier majors.
To control for major and relative grades, we look at how students who were in a
particular percentile in terms of first year CGPA, relative to all those in their major, fared
in terms of their CGPA after the first year, relative to all those in their major. If students
who were in the 30th percentile of their major in terms of their first year CGPA are on
average in the 40th percentile in terms of their CGPA in the next 3–4 years of their careers,
catch up may be occurring. In turn, if those students average at the 20th percentile after the
first year, they are falling back instead of catching up. Since we are considering how
students fare relative to others in their major, we eliminate the effect of different grading
standards across majors. Moreover, by considering their ordinal standing rather than the
level of CGPA, we eliminate the effect of differences in grades in early versus late
semesters.

20
Arcidiacono et al. (2011) make a similar point for black students in Duke University.

123
High Educ (2016) 71:611–649 625

(a) (b)
Average CGPA percentile at graduation

100

1.5
net of 1st year, relative to major
80

1
60

.5
40
20

0
0

−.5
0 20 40 60 80 100
GE, GE, SC/ST, SC/ST,
First year CGPA percentile, relative to major Selective Non−Selective Selective Non−Selective
GE, Selective SC/ST, Selective Point Estimate
GE, Non−Selective SC/ST, Non−Selective

Fig. 5 Average Ranking at Graduation by Initial Ranking: Major-Specific Percentiles of CGPA. Source:
Survey data from the EEI’s graduating class, 2008. Notes: Bachelor programs in CS and EE Power as well as
dual degree programs in CS and EE are classified as Selective Majors. Average final rankings are locally
mean-smoothed using a kernel-weighted local polynomial smoother

Panel (a) in Fig. 5 plots average major ranks at graduation versus first year major ranks
both by caste and by major selectivity. Average percentiles at the end are calculated
through separate locally linear regressions within each group.21 Thus, for example, general
students in non-selective majors who were in the 20th percentile of their majors in terms of
their fist year CGPA are, on average, in the 28th percentile of their major at graduation,
measured in terms of their final CGPA net of the first year.
In general, the curves presented in Panel (a) show that the slope of the final average
ranking with respect to the initial ranking is less than unity, especially at the top and
bottom. This is to be expected as those initially at the bottom have no place to go but up
and those at the top have no place to go but down. In other words, reversion to the mean
should be present. Nevertheless, in the general category, irrespective of the selectivity of
the major, these curves are very close to the 45 degree line suggesting that, on average,
general students stay in the major-specific percentiles they start off in. However, this is not
the case for SC/ST students. Panel (a) shows that the slope for selective and non-selective
majors seems to be below one for minority students and it is even flatter for selective
majors, suggesting that SC/ST students are falling behind over time and more so in more
selective majors.
Panel (b) in Fig. 5 presents the results of separate regressions by caste and major
selectivity of the final major rank on the initial percentile relative to the major. The round
marker shows point estimates of the coefficient of the initial ranking, while the vertical line
represents the 95 % confidence interval of each coefficient. The estimates confirm the
pattern observed in panel (a): The slope of final ranking with respect to the initial ranking
is close to one for general students, and even so for SC/ST students in less selective majors.
In fact, we cannot reject the null of the coefficient being equal to unity for the latter. In

21
To avoid misleading patterns due to outliers, the five SC/ST students who ranked above the 50th
percentile in the first year are removed. Two of these dropped observations belong to the group of minority
students who have high family income.

123
626 High Educ (2016) 71:611–649

turn, the slope for SC/ST students in more selective majors is around 0.25 and unity is
clearly far outside the confidence interval.
One may worry about mean reversion as the main driver of the patterns identified in
Fig. 5. Using the approach proposed in Chay et al. (2005) and later used in Akyol and
Krishna (2014), we control for mean selection bias and find that the coefficients among
general students are quite robust to the correction. Those estimated for SC/ST are instead
almost entirely driven by mean reversion which leaves them even in a worse situation.
The evidence presented in Fig. 5 suggests that the catch up we seemed to find in the
aggregate was an illusion. When we look at how SC/ST students do relative to those in
their major over time, a different picture emerges. Although minority students in less
selective majors are able to at least keep up, SC/ST students in more selective majors seem
to be falling behind rather than catching up. This is not surprising considering that minority
students start college lagging far behind nonminority students. Consequently, the gap
between general and SC/ST students is likely to increase as both groups progress through
college, especially in more selective majors. Even by running as fast as they can, SC/ST
students can hope, at best, to stay in the same place they started but those in more
competitive majors cannot even do this and fall even further behind their general category
peers.

Mismatch

The evidence so far seems to suggest that minority students are not catching up in terms of
academic performance. On the contrary, SC/ST students in more selective majors tend to
fall behind in their major-specific rankings as they progress through college. This pattern
seems to be supportive of the mismatch hypothesis, which predicts lower success rates for
minority students who enroll in more selective colleges and/or majors relative to those in
colleges and/or majors where their academic credentials are better matched to the average.
We first explore the extent to which AA policies motivate minority students to aim for
more selective majors at the EEI. Figure 6 displays the fractions of general and minority
students in the EEI who are accepted into selective majors as functions of their first year
CGPA percentile.22 The fraction of students in selective majors at each percentile score is
computed from separate locally linear regressions within each group. Although the fraction
of students who attend selective majors is increasing in initial performance for both general
and SC/ST students, minority students are much more likely to attend selective majors for
all levels of first year CGPA. This evidence is consistent with Rothstein and Yoon (2009)
findings on school choice for white and black students in law school, and it suggests that
the reservation schemes in India are playing a significant role in the major choice of SC/ST
applicants.
The patterns in Fig. 6 are also in line with the evidence presented by Arcidiacono et al.
(2011) in terms of expected major: Compared to white students, black students are more
likely to exhibit an initial interest in natural science, engineering, and economics, majors
that have stricter grading standards and are perceived as more challenging. However, as
students progress through college, black students in the USA switch away from sciences
and into humanities and social studies at a greater pace than white students. In India,
however, mobility across majors is very limited and almost all students at the EEI graduate
from the major they initially enroll in. Two factors explain this lack of mobility. First, in
the Indian higher education system students are admitted to a major and not into college as

22
Again, the five minority students above the 50th percentile in terms of first year CGPA are dropped.

123
High Educ (2016) 71:611–649 627

Fig. 6 Fraction of Students

1
Attending Selective Majors by
Caste and Initial CGPA

Fraction in selective majors

Percentile. Source: Survey data

.75
from the EEI’s graduating class,
2008. Notes: Bachelor programs
in CS and EE Power as well as
dual degree programs in CS and

.5
EE are classified as Selective
Majors. Fractions are locally
mean-smoothed using a kernel-

.25
weighted local polynomial
smoother

0
0 25 50 75 100
Percentile CGPA Score in the First Year
GE SC/ST

in the USA case. Second, majors in India are rationed by test performance and this
generates an important prestige effect of being in a selective major. SC/ST students may
choose selective majors for their prestige value and find themselves unable to change them
ex post.
The rest of this subsection will evaluate the effect of major selectivity on labor market
outcomes for both general and SC/ST students. Since admission is driven by caste and
performance in the centralized entrance exam, the allocation of seats in selective and non-
selective majors is not exogenous with respect to future academic performance and wages.
Taking into account that allocation to selective and non-selective majors is not random, we
try to assess the causal link between major selectivity and first wage after graduation.
To evaluate the mismatch hypothesis, we compare the wages of students in selective
majors with their same-caste counterparts in less selective majors. This task requires taking
into account that graduating from a selective major depends on being admitted to that
major, which ultimately depends on ability (maybe partially unobserved) and other indi-
vidual characteristics that also determine wages.
Let Si be a latent continuous variable which is increasing in the selectivity of i’s major:
Si ¼ Zi c  li

where Zi denotes observed individual characteristics such as gender, first year CGPA
(proxy for entrance exam score), family income, father’s education, school type, a dual
degree dummy, and type of job. Define Si ¼ 1 if Si  0 and Si ¼ 0 if Si \0. Assuming that
li  Nð0; r2l Þ, the probability of being enrolled in a selective major is given by:

PrðS ¼ 1jZi Þ ¼ UðZi c; r2l Þ ð1Þ

Let wi denote individual i’s log wage at graduation (in logs of lacs):

wi ¼ a1 Si þ Xi b1  i

where Xi is contained in Zi . Our data give interval data for wages, so we define a discrete
variable, wi , which will be equal to category Wk if nk1  wi \nk . We use three wage
groups with known thresholds, nk , at n0 ¼ 1, n1 ¼ logð5Þ; n2 ¼ logð7Þ; and n3 ¼ 1.

123
628 High Educ (2016) 71:611–649

Assuming that i  Nð0; r2 Þ; and representing the normal cumulative distribution with
mean zero and variance r2 by Uðx; r2 Þ; the probability of wi ¼ Wk is given by:

Prðwi ¼ Wk jXi Þ ¼ Prðnk1  wi \nk Þ

ð2Þ
¼ Uða1 Si þ Xi b1  nk1 ; r2 Þ  Uða1 Si þ Xi b1  nk ; r2 Þ

Our model can be summarized by the system (1, 2). In particular, the parameter of interest
is a1 : if a1  0, we cannot reject the mismatch hypothesis for SC/ST students in selective
majors. However, we need to take into account the endogenous selection process that
determines allocation into majors, summarized by (1). Measuring the wage premium of
selective majors as the difference in wages of students in selective and non-selective
majors gives a biased picture: Upward bias is likely since observable and unobservable
personal traits that make a student more likely to get into a selective major also make a
student more likely to get higher wages. For example, students with higher exam scores,
measured by first year CGPA, are more likely to be in selective majors and earn higher
wages. If we ignore the role of exam score in selection into majors, the beneficial effect of
major selectivity on wages is overestimated as selective majors have students with rela-
tively higher scores. If we assume that both i and li are uncorrelated random shocks, so
that selection into more selective majors is only driven by observables, propensity score
matching techniques yield an unbiased effect of Si on wages. However, if i and li contain
unobserved traits that increase the probability of being in a more selective major and that
of getting higher wages, propensity score matching methods yield biased estimates of a1 :
In this case, the bias caused by the correlation between i and li can be prevented by jointly
estimating (1, 2).
Assuming selection is exclusively driven by observables, we obtain propensity scores
(or the probability of a student being enrolled in a selective major) separately for general
and minority students based on their observable characteristics in Xi as well as additional
controls such as coaching expenditures for the entrance exam, number of household
members, and age of the student. We then include the estimated propensity score as an
additional regressor in the wage equation, assuming that enrollment in selective majors is
random with respect to wages once we control for Xi and the propensity score.23
If we also allow selection into selective majors to be related to students’ unobservable
characteristics (i.e., i and li are correlated through unobservables), the system in (1, 2)
must be jointly estimated. Despite the availability of extensive survey data at the student
level, we could not find a good exclusion restriction. Finding an instrument that would only
affect the major selectivity equation but not the wage equation was impossible given that
the determinants of both equations overlap a great deal. Therefore, identification comes
from the normality assumptions on the error terms.24
Table 1 summarizes our results (see Tables 6, 7 in Appendix 1 for more details).25
Column 1 provides a baseline estimate of a1 when the endogeneity of the major choice is

23
See Rosenbaum and Rubin (1983). Unfortunately, we have interval data on wages. If the outcome
variable were binary or continuous, we could also implement the traditional matching approach which
provides a nonparametric estimator of a1 without making assumptions on the relationship between wages
and Xi (or Zi ).
24
We also experimented with instruments. For the GE group, the results were very similar to those relying
on functional form assumptions. The sample size was too small in the SC/ST group for us to consider using
instruments there.
25
Tables 8, 9 in Appendix 1 report the same regression results when the final CGPA is excluded as an
independent variable.

123
High Educ (2016) 71:611–649 629

Table 1 Effect of attending a selective major at the EEI on wages (a^1 )

Interval regressiona PSb Joint estimationc

GE 0.185*** 0.176*** 0.115

(0.036) (0.037) (0.149)
q ¼ 0:196
SC/ST 0.055 -0.001 -0.380*
(0.085) (0.093) (0.199)
q ¼ 0:913

Standard errors in parentheses

Source: Survey data from the EEI’s graduating class, 2008
*** p \ 0.01, ** p \ 0.05, * p \ 0.1
a
Coefficient on observed major choice in an interval regression for wages.
b
Propensity score included as control in interval regression for wages.
c
Joint estimation of selective major choice and wages equations, no instruments

ignored and the wage Eq. (2) is estimated including Si and Xi as regressors. Columns 2
provides estimates based on propensity score matching methods, where the propensity
score has been added to equation (2) as an additional control.
Among general students, being enrolled in a more selective major seems to have a
positive and significant effect even after controlling for selection on observables. Of
course, a^1 is lower once we control for selection on observables, which is expected if
students who enroll in more selective majors are also more likely to earn higher wages.
Propensity score matching methods show that general students in highly competitive
majors earn 17 % more than their same-caste peers in less selective majors. Among
minority students, this wage premium is not significantly different from zero, suggesting
that admission preferences that facilitate admission of SC/ST into selective majors do not
increase future wages in this group.
However, if students with higher unobserved ability tend to choose more selective
majors, then controlling for selection as we have done so far will overestimate the effect of
being in a selective major. The results in the last column of Table 1 show that the estimated
correlation in the errors across equations is positive within each group (general and SC/ST)
and, as a result, a^1 falls, becoming insignificant for the GE group and significantly negative
for the SC/ST group. Being in a selective major has no effect on the earnings of GE
students: The wage premium previously identified for them was only due to unobserved
ability differences. This is consistent with the work of Altonji et al. (2005) in the USA on
the effect of attending Catholic schools on the probability of graduation. For the SC/ST
group, it suggests mismatch. Minority students who enroll in selective majors as a con-
sequence of AA policies obtain lower wages than they would have had if they had chosen a
less selective major.26,27
Estimates from the caste-specific wage equations run to obtain the results reported in
Table 1 (see Tables 6, 7 in Appendix 1) show that the constant terms vary very little,

26
A weakness in our estimates is that the estimated qs, though positive, are not significantly different from
zero as we can see from Table above.
27
Our results are robust to the exclusion of potentially intermediate variables such as final CGPA and type
of job. See Tables 8, 9 in Appendix 1.

123
630 High Educ (2016) 71:611–649

implying that caste-specific mean wages net of observables are very close. Indeed, even
after controlling for selection into majors, grades, and type of job, we cannot reject the null
that the constant in the wage equation is equal across groups. Of course, the same result is
obtained if we run the usual wage regression for all students controlling for observables:
The coefficient on caste is not significant.28
In addition to its effect on wages, being enrolled in a more selective major can also
affect minority students’ well-being while at the EEI. Since they see themselves falling
behind general students in the same major, they could also be facing higher instantaneous
costs of going to college by being more stressed and frustrated than their same-caste peers
in less selective majors.29
We rely on survey data on academic life experience and hostel life to analyze students’
well-being while in college. In particular, we focus on two survey questions. The first one
asks the students whether they had ever felt stressed, depressed, lonely, or discriminated at
the EEI where the possible answers are Never, Occasionally, Often, and Regularly. We
also rely on a question that asks the student whether he/she felt that the hostel was like a
home away from home.
Let variable yi be a latent variable that denotes an aspect of student’s well-being such as
stress, depression, loneliness, discrimination, or feeling comfortable at campus hostels:
yi ¼ a2 Si þ Xi b2  ei

In this case, if a2  0 we cannot reject social mismatch among SC/ST students in selective
majors. Define yi ¼ 1 if yi  0 and yi ¼ 0 if yi \0: Assuming that ei  Nð0; r2e Þ; the
probability of feeling stressed (depressed, lonely, discriminated, or comfortable at the
hostels) is then:

Prðyi ¼ 1jXi Þ ¼ Uða2 Si þ Xi b2 ; r2e Þ ð3Þ

Table 2 reports our estimates of a2 :30 Again, we control for selection into more selective
majors using propensity score matching methods and jointly estimating the system (1–3).
Once selection into observables and unobservables is taken into account, general students
in more selective majors tend to be slightly happier during their college experience than
their same-race counterparts in less selective majors. The estimated qs for general students
suggest that more able or motivated students who are more likely to choose tougher majors
are also more likely to face higher levels of emotional discomfort during their college

28
Although there are no earnings differentials by caste within occupations, minority students seem to be
placed in worse occupations than general students, which may be indicative of discrimination in terms of
occupations. These results do not necessarily imply that minority students are discriminated in their first job
placement. In fact, a higher probability of being placed in a lower paying job may be explained by individual
choices of the students or by labor market reservations that are in place in government jobs, which pay less
but are more stable than jobs in finance firms and in the private sector. Results can be found in Frisancho and
Krishna (2012).
29
Table 10 in Appendix 1 shows that indeed, students whose first year CGPA is below the average
performance in their major seem to be the most emotionally affected as the coefficient on the demeaned first
year CGPA shows.
30
We drop individuals with missing observations in each of the well-being variables. We did not include
these variables in the imputation process because the missing at random assumption is less defendable in
these cases. The problem is small though; missing values in the stress, depression, loneliness, discrimina-
tion, or hostel is like home variables only represent between 6 and 10 % of the individuals in each caste.

123
High Educ (2016) 71:611–649 631

Table 2 Effect of attending a

Probita PSb Joint estimationc
selective major at the EEI on
emotional and social well-being
(a^2 ) Stress
GE 0.112* 0.125* -0.082***
(0.065) (0.068) (0.010)
q ¼ 0:85
SC/ST 0.351*** 0.320** 0.115**
(0.120) (0.137) (0.046)
q ¼ 0:34
Depression
GE 0.015 0.019 -0.110***
(0.052) (0.053) (0.014)
q ¼ 0:86***
SC/ST -0.249 -0.212 -0.013
(0.152) (0.162) (0.353)
q ¼ 0:38
Loneliness
GE 0.035 0.042 -0.093*
(0.052) (0.054) (0.049)
q ¼ 0:96
SC/ST -0.089 -0.068 -0.061
Standard errors in parentheses
(0.141) (0.161) (0.109)
Source: Survey data from the
q ¼ 0:32
EEI’s graduating class, 2008
Discrimination
Note: Stress, depression,
loneliness, and discrimination GE -0.026 -0.026 -0.096**
dummies are constructed by (0.043) (0.045) (0.049)
coding Often and Regularly as q ¼ 0:68**
ones. Hostel feels like home
dummy is obtained by coding SC/ST -0.018 -0.018 -0.082
with ones all students who agree (0.172) (0.212) (0.124)
with the statement q ¼ 0:56
*** p \ 0.01, ** p \ 0.05, * p \ Hostel does not feel like home
0.1
a GE -0.016 -0.008 -0.003
Coefficient on observed major
choice in a probit model. (0.057) (0.058) (0.085)
b
Propensity score included as q ¼ 0:01
control in probit model. SC/ST 0.155 0.088 0.121***
c
Joint estimation of selective (0.122) (0.125) (0.036)
major choice and well-being q ¼ 0:86*
equations, no instruments

experience.31 Once this correlation is taken into account, general students in more selective
majors are more likely to fit in socially than nonminority students in less selective majors.
However, the story is different for SC/ST students. Compared to minority students in less
selective majors, SC/ST students in tougher majors are significantly more stressed and feel

31
The exception is the last variable, hostel does not feel like home, where the correlation is very close to
zero.

123
632 High Educ (2016) 71:611–649

less comfortable at the hostel facilities. Mismatch is not only generating lower wages, but it
is also imposing a higher cost of going to college on them.32
In sum, the results suggest that general students do not benefit from being in selective
majors; they earn more because they are better. However, they do tend to face relatively
lower emotional and social costs of studying at the EEI compared to general students in
less selective majors. Taking unobservables into account shows that SC/ST students in
selective majors earn less than minority students in less selective majors, supporting the
mismatch hypothesis. These students also seem to experience what we call social mis-
match as they also feel more stressed and less comfortable at the campus hostels when
compared to SC/ST in less selective majors. Though we do our best to evaluate the
mismatch hypothesis, our results on this are only suggestive given the data limitations we
face and the lack of an adequate exclusion restriction.

Conclusion

AA policies always generate divided opinions and provoke intense debate. In the USA, for
example, evidence showing that AA policies have only benefited richer black students has
promoted proposals to shift from race-based to economic-based affirmative action. This
paper offers evidence that can contribute to the debate on reservation policies in higher
education admissions in India by analyzing the real impact of such policies on their
intended beneficiaries.
Using detailed data on the 2008 graduating class from an elite engineering institution
(EEI) in India and applicant data on students taking the joint entrance exam for engineering
colleges in 2009, this article tries to cast some light on the effects of AA on Indian
minorities. Our paper is particularly relevant because it overcomes the limitations of USA
studies. First, admission criteria are clear and rigid so that performance in the entrance
exam is all that matters to get into the EEI in contrast to the much more nebulous
admissions process in the USA. Moreover, the use of AA policies in India results in very
large differences in admission standards, which implies that our empirical results are
unlikely to be confounded by the program being a marginal one. Finally, the strict curricula
in higher education Indian institutions minimize the issue of self-selection into easier
courses within a major so that grades are a better indicator of performance.
We provide some basic evidence on three issues central to the debate over AA: tar-
geting, catch up, and mismatch. We find that minority admission preferences seem to be
doing a reasonable job targeting poorer populations, though there seems to be little evi-
dence of catch up. In fact, minority students in more selective majors seem to be falling
behind, suggesting that mismatch effects might be present. Finally, we find no effect of
major selectivity on wages for minority students when only observables are assumed to
drive selection into selective majors. However, when the effect of unobservables is taken
into account, the results suggest that minority students do not benefit from being in
selective majors to which they are attracted by the preferences. In fact, SC/ST students in
selective majors earn less than minority students in less selective majors, supporting the
mismatch hypothesis. Although there are no earnings differentials by caste within occu-
pations, minority students seem to be placed in worse occupations than general students.
As Rothstein and Yoon (2008) point out, the pervasive nature of affirmative action
policies eliminates the possibility of finding a control group for the minorities benefited by
32
Social mismatch does not seem to affect wages or grades directly.

123
High Educ (2016) 71:611–649 633

the policy. Rothstein and Yoon (2009) argue that much of the existing evidence on catch
up and mismatch is flawed since differences in the control groups chosen have been an
important source of variation across studies. One of our planned extensions is to develop a
structural model that can help us overcome a difficulty faced by most empirical studies on
the topic: finding a control group.

Appendix 1

See Tables 3, 4, 5, 6, 7, 8, 9, and 10.

Table 3 Summary statistics

Variable All students (N = 451) SC/ST (N = 68) General category (N = 383)

Mean SD Mean SD Mean SD

Male 0.9 0.3 0.93 0.26 0.90 0.31

Age 18.45 1.03 18.99 1.34 18.36 0.93
General category 0.85 0.36
Master program 0.22 0.41 0.15 0.36 0.23 0.42
Major
CE 0.12 0.33 0.13 0.34 0.12 0.32
CH 0.14 0.34 0.13 0.34 0.14 0.34
CS 0.13 0.34 0.13 0.34 0.13 0.33
EE 0.17 0.38 0.24 0.43 0.16 0.37
PH 0.06 0.25 0.01 0.12 0.07 0.26
MT 0.04 0.2 0.03 0.17 0.04 0.21
BB 0.05 0.22 0.03 0.17 0.06 0.23
ME 0.2 0.4 0.29 0.46 0.18 0.39
TT 0.09 0.28 0 0 0.1 0.3
First year CGPA 7.2 1.17 5.85 0.83 7.44 1.05
CGPA at graduation 7.4 1.02 6.23 0.70 7.61 0.92
Enter the EEI in first attempt 0.49 0.5 0.32 0.47 0.52 0.5
Log(Coaching expenditures) 8.4 4.45 7.21 5.05 8.60 4.31
School type
State/Government 0.17 0.37 0.41 0.5 0.13 0.33
KV/Minority 0.08 0.27 0.1 0.31 0.08 0.27
Private 0.75 0.43 0.49 0.5 0.8 0.4
First wage after graduation (lac)
No job 0.09 0.29 0.19 0.4 0.07 0.26
3–4 0.06 0.25 0.12 0.32 0.05 0.23
4–5 0.11 0.31 0.12 0.32 0.1 0.31
5–7 0.35 0.48 0.38 0.49 0.35 0.48
7–10 0.18 0.38 0.13 0.34 0.18 0.39
10–15 0.11 0.31 0.03 0.17 0.13 0.33
15–25 0.03 0.17 0.01 0.12 0.03 0.18
[25 0.07 0.25 0.01 0.12 0.08 0.26

123
634 High Educ (2016) 71:611–649

Table 3 continued
Variable All students (N = 451) SC/ST (N = 68) General category (N = 383)

Mean SD Mean SD Mean SD

Family income (lac)

\1 0.1 0.29 0.19 0.4 0.08 0.27
1–2 0.17 0.38 0.24 0.43 0.16 0.37
2–3 0.19 0.39 0.21 0.41 0.19 0.39
3–4 0.17 0.37 0.21 0.41 0.16 0.37
4–6 0.16 0.37 0.06 0.24 0.18 0.38
6–9 0.08 0.27 0.04 0.21 0.09 0.28
9–12 0.05 0.22 0.03 0.17 0.05 0.22
[12 0.08 0.28 0.03 0.17 0.09 0.29
Mother’s education
High school or less 0.30 0.46 0.66 0.48 0.24 0.43
College 0.35 0.48 0.19 0.4 0.38 0.49
Postgraduate 0.35 0.48 0.15 0.36 0.38 0.49
Father’s education
High school or less 0.18 0.39 0.41 0.5 0.14 0.35
College 0.5 0.5 0.4 0.49 0.51 0.5
Postgraduate 0.32 0.47 0.19 0.4 0.34 0.48
Mother’s occupation
Non-main earner–retired 0.72 0.45 0.82 0.38 0.70 0.46
Government or public sector 0.17 0.37 0.1 0.31 0.18 0.38
Private sector 0.11 0.32 0.07 0.26 0.12 0.33
Father’s occupation
Non-main earner–retired 0.08 0.26 0.07 0.26 0.08 0.26
Government or public sector 0.52 0.5 0.72 0.45 0.49 0.5
Private sector 0.4 0.49 0.21 0.41 0.44 0.5
Number of household members 2.25 0.82 2.46 0.92 2.22 0.8

Table 4 Average first year CGPA by major and caste

Average first year GPA Share with wage [7 lacs

All GE SC/ST All GE SC/ST

Bachelor’s
Computer Science 8.2 8.5 6.5 0.79 0.84 0.57
Mechanical Engineering, Production 7.3 7.7 6.0 0.26 0.31 0.13
Mechanical Engineering 7.2 7.5 5.5 0.36 0.33 0.50
Electrical Engineering, Power 7.6 8.0 6.4 0.56 0.66 0.22
Electrical Engineering 6.7 7.1 5.3 0.30 0.41 0.00
Chemical Engineering 7.4 7.7 5.6 0.38 0.42 0.17
Engineering Physics 6.9 6.9 6.4 0.34 0.36 0.00
Textile Technology 6.9 6.9 – 0.21 0.21 –
Civil Engineering 6.7 7.0 5.1 0.22 0.24 0.11

123
High Educ (2016) 71:611–649 635

Table 4 continued
Average first year GPA Share with wage [7 lacs

All GE SC/ST All GE SC/ST

Dual Degrees
Computer Science 7.9 8.1 6.8 0.84 0.94 0.00
Electrical Engineering 7.3 7.4 6.5 0.38 0.42 0.00
Chemical Engineering 6.5 6.6 5.9 0.23 0.26 0.00
Biotechnology 7.1 7.3 5.3 0.25 0.27 0.00
Mathematics and Computing (MA) 7.0 7.1 5.5 0.47 0.47 0.50

Table 5 First year CGPA as a

First year CGPA
function of income and caste
Constant 7.517***
(0.205)
Selective major 0.832***
(0.104)
Male -0.402***
(0.148)
Dual degree/master -0.314***
(0.109)
SC/ST -1.545***
(0.168)
Household income (base: \3 lac)
½39 lac 0.167
(0.107)
[9 lac 0.194
(0.149)
Household income 9 caste
½3  9 lac 9 SC/ST -0.349
(0.271)
[9 lac 9 SC/ST 1.367***
(0.516)
Father’s education (base: High School or less)
College or technical training 0.053
(0.130)
Grad education -0.052
(0.143)
School type (base: Public)
KV or minority -0.141
(0.194)
Standard errors in parentheses Private 0.070
Source: Survey data from the (0.129)
EEI’s graduating class, 2008 Observations 451
*** p \ 0.01, ** p \ 0.05, R-squared 0.383
* p \ 0.1

123
636 High Educ (2016) 71:611–649

Table 6 Wage regressions for general students

Interval regressiona PSb Joint estimationc

Constant 1.060*** 1.205*** 1.012***

(0.133) (0.196) (0.166)
Selective major 0.185*** 0.176*** 0.115
(0.036) (0.037) (0.149)
Male -0.032 -0.048 -0.027
(0.043) (0.046) (0.045)
Dual degree/master -0.013 -0.041 -0.002
(0.032) (0.042) (0.040)
CGPA first year 0.043* 0.020 0.056
(0.025) (0.035) (0.036)
CGPA at graduation 0.033 0.033 0.028
(0.028) (0.028) (0.030)
Household income (base: \3 lac)
½3  6 lac 0.055* 0.043 0.060*
(0.029) (0.031) (0.031)
[6 lac 0.129*** 0.112** 0.137***
(0.043) (0.046) (0.046)
Father’s education (base: High School or less)
College or technical training 0.045 0.078 0.032
(0.040) (0.052) (0.049)
Grad education 0.009 0.034 -0.001
(0.043) (0.050) (0.048)
School type (base: Public)
KV/Minority 0.160*** 0.164*** 0.160***
(0.057) (0.058) (0.058)
Private 0.144*** 0.136*** 0.148***
(0.039) (0.040) (0.040)
Type of job (base: nontechnical)
Core 0.069* 0.071* 0.068*
(0.041) (0.041) (0.041)
Management/consulting 0.021 0.024 0.021
(0.045) (0.045) (0.045)
Analytical 0.058 0.057 0.058
(0.046) (0.046) (0.046)
Financial 0.106** 0.106** 0.107**
(0.046) (0.046) (0.046)
Propensity score 0.181
(0.182)
q 0.196
(0.396)
Observations 383 383 383

123
High Educ (2016) 71:611–649 637

Table 6 continued
Interval regressiona PSb Joint estimationc

r 0.203 0.203 0.204

Standard errors in parentheses

Source: Survey data from the EEI’s graduating class, 2008
*** p \ 0.01, ** p \ 0.05, * p \ 0.1
a
Coefficient on observed major choice in an interval regression for wages
b
Propensity score included as control in interval regression for wages
c
Joint estimation of selective major choice and wages equations, no instruments

Table 7 Wage regressions for SC/ST students

Interval regressiona PSb Joint estimationc

Constant 1.247*** 1.459*** 0.830*

(0.370) (0.392) (0.497)
Selective major 0.055 -0.001 -0.380*
(0.085) (0.093) (0.199)
Male -0.174 -0.092 -0.331*
(0.136) (0.148) (0.182)
Dual degree/master -0.155 -0.144 -0.159
(0.112) (0.109) (0.125)
CGPA first year 0.046 -0.010 0.172
(0.070) (0.080) (0.106)
CGPA at graduation 0.056 0.049 0.053
(0.077) (0.075) (0.092)
Household income (base: \3 lac)
[3–6] lac -0.102 -0.115 -0.069
(0.078) (0.077) (0.095)
[6 lac -0.195 -0.223 -0.136
(0.158) (0.157) (0.199)
Father’s Education (base: High School or less)
College or technical training 0.001 -0.005 0.004
(0.074) (0.073) (0.093)
Grad education -0.001 -0.024 0.037
(0.092) (0.092) (0.116)
School type (base: Public)
KV/Minority or private 0.098 0.132* 0.028
(0.072) (0.075) (0.091)
Type of job (base: Nontechnical)
Core 0.021 0.029
(0.140) (0.138)

123
638 High Educ (2016) 71:611–649

Table 7 continued
Interval regressiona PSb Joint estimationc

Management/consulting 0.014 0.010

(0.178) (0.175)
Analytical 0.079 0.097
(0.189) (0.187)
Financial -0.037 -0.026
(0.213) (0.208)
Management/consulting, analytical, or financial 0.010
(0.094)
Propensity score 0.297
(0.214)
q 0.913
(0.182)
Observations 68 68 68
r 0.219 0.215 0.289

Standard errors in parentheses

Source: Survey data from the EEI’s graduating class, 2008
*** p \ 0.01, ** p \ 0.05, * p \ 0.1
a
Coefficient on observed major choice in an interval regression for wages
b
Propensity score included as control in interval regression for wages
c
Joint estimation of selective major choice and wages equations, no instruments

Table 8 Wage regressions for general students, excluding final GPA and type of job as regressors
Interval regressiona PSb Joint estimationc

Constant 1.141*** 1.291*** 1.098***

(0.120) (0.188) (0.160)
Selective major 0.192*** 0.183*** 0.139
(0.036) (0.037) (0.133)
Male -0.044 -0.061 -0.039
(0.043) (0.046) (0.045)
Dual degree/master -0.006 -0.035 0.002
(0.031) (0.042) (0.037)
CGPA first year 0.074*** 0.049* 0.081***
(0.014) (0.027) (0.022)
Household income (base: \3 lac)
½3  6 lac 0.053* 0.040 0.057*
(0.028) (0.031) (0.030)
[6 lac 0.131*** 0.113** 0.136***
(0.042) (0.045) (0.044)
Father’s education (base: High School or less)
College or technical training 0.049 0.084 0.039
(0.040) (0.052) (0.047)

123
High Educ (2016) 71:611–649 639

Table 8 continued
Interval regressiona PSb Joint estimationc

Grad education 0.015 0.041 0.007

(0.043) (0.050) (0.047)
School type (base: Public)
KV/Minority 0.176*** 0.180*** 0.175***
(0.057) (0.058) (0.058)
Private 0.150*** 0.141*** 0.153***
(0.039) (0.040) (0.039)
Propensity score 0.188
(0.183)
q 0.145
(0.352)
Observations 383 383 383
r 0.206 0.206 0.207
Standard errors in parentheses
Source: Survey data from the EEI’s graduating class, 2008
*** p \ 0.01, ** p \ 0.05, * p \ 0.1
a
Coefficient on observed major choice in an interval regression for wages
b
Propensity score included as control in interval regression for wages
c
Joint estimation of selective major choice and wages equations, no instruments

Table 9 Wage regressions for SC/ST students, excluding final GPA and type of job as regressors
Interval regressiona PSb Joint estimationc

Constant 1.399*** 1.600*** 0.987**

(0.303) (0.332) (0.432)
Selective major 0.062 0.009 -0.361**
(0.081) (0.088) (0.183)
Male -0.166 -0.079 -0.320*
(0.134) (0.148) (0.174)
Dual degree/master -0.128 -0.116 -0.149
(0.094) (0.093) (0.116)
CGPA first year 0.081 0.020 0.198**
(0.050) (0.066) (0.083)
Household income (base: \3 lac)
[3–6] lac -0.093 -0.106 -0.066
(0.076) (0.076) (0.094)
[6 lac -0.182 -0.210 -0.134
(0.158) (0.158) (0.197)
Father’s education (base: High School or less)
College or technical training -0.010 -0.017 -0.004
(0.073) (0.072) (0.090)

123
640 High Educ (2016) 71:611–649

Table 9 continued
Interval regressiona PSb Joint estimationc

Grad education 0.002 -0.021 0.040

(0.091) (0.091) (0.113)
School type (base: Public)
KV/Minority or private 0.100 0.135* 0.034
(0.071) (0.075) (0.090)
Propensity score 0.295
(0.215)
q 0.883
(0.188)
Observations 68 68 68
r 0.221 0.217 0.284

Standard errors in parentheses

Source: Survey data from the EEI’s graduating class, 2008
*** p \ 0.01, ** p \ 0.05, * p \ 0.1
a
Coefficient on observed major choice in an interval regression for wages
b
Propensity score included as control in interval regression for wages
c
Joint estimation of selective major choice and wages equations, no instruments

Table 10 Well-being as a function of relative academic disadvantage and controls

Stressed Depressed Lonely Discriminated Hostel feels like
home

Selective major 0.005 -0.114 -0.161** -0.106* 0.003

(0.089) (0.075) (0.071) (0.060) (0.077)
Male -0.072 -0.024 0.012 0.003 -0.115
(0.081) (0.070) (0.067) (0.054) (0.076)
Dual degree/master 0.047 0.064 0.079* 0.038 0.025
(0.060) (0.049) (0.047) (0.037) (0.054)
CGPA first year (CGPA1) 0.120 0.078 0.158** 0.035 -0.061
(0.082) (0.068) (0.066) (0.053) (0.072)
GE 0.069 -0.076 0.193 0.020 -0.046
(0.136) (0.105) (0.136) (0.093) (0.126)
(CGPA1—Major’s mean -0.194* -0.103 -0.296*** -0.095 0.099
CGPA1) (0.113) (0.090) (0.097) (0.073) (0.101)
GE x (CGPA1–Major’s mean -0.012 -0.032 0.052 0.021 -0.052
CGPA1) (0.091) (0.070) (0.083) (0.059) (0.080)
Household income (base: \3 lac)
[3–6] lac -0.016 -0.040 0.026 0.025 -0.020
(0.053) (0.043) (0.042) (0.034) (0.047)
[6 lac 0.046 -0.016 -0.052 0.039 -0.115*
(0.076) (0.063) (0.065) (0.049) (0.063)
Father’s education (base: High School or less)

123
High Educ (2016) 71:611–649 641

Table 10 continued
Stressed Depressed Lonely Discriminated Hostel feels like
home

College or technical training 0.080 0.005 -0.024 0.039 0.003

(0.069) (0.055) (0.055) (0.047) (0.060)
Grad education -0.005 -0.033 0.050 0.043 -0.001
(0.076) (0.061) (0.059) (0.051) (0.066)
School type (base: Public)
KV/Minority -0.096 -0.018 -0.174* 0.027 -0.025
(0.100) (0.079) (0.094) (0.053) (0.082)
Private 0.029 0.026 -0.031 -0.040 -0.007
(0.066) (0.053) (0.052) (0.039) (0.056)
Observations 422 413 420 412 425
Log likelihood -270.5 -194.5 -189.4 -128.8 -226.0
Standard errors in parentheses
Source: Survey data from the EEI’s graduating class, 2008
*** p \ 0.01, ** p \ 0.05, * p \ 0.1

Appendix 2: Pattern of missing observations

See Tables 11, 12 and 13.

Table 11 Summary statistics

Variable % Missing (N = 453) Original data set Imputed data set

Mean SD Mean SD

Male 0.0 0.90 0.90

General Category 0.0 0.85 0.85
First year CGPA 0.0 7.20 1.17 7.20 1.17
CGPA at graduation 0.0 7.40 1.03 7.40 1.03
Bachelor’s degree (Dual/master’s = 0) 0.0 0.79 0.79
Major
CS 0.0 0.13 0.13
ME 0.0 0.20 0.20
EE 0.0 0.17 0.17
CH 0.0 0.13 0.13
BB 0.0 0.05 0.05
MT 0.0 0.04 0.04
PH 0.0 0.06 0.06
TT 0.0 0.09 0.09
CE 0.0 0.12 0.12
Family income (lac) 1.8
\1 0.09 0.10

123
642 High Educ (2016) 71:611–649

Table 11 continued
Variable % Missing (N = 453) Original data set Imputed data set

Mean SD Mean SD

1–2 0.17 0.17

2–3 0.19 0.19
3–4 0.17 0.17
4–6 0.16 0.16
6–9 0.08 0.08
9–12 0.05 0.05
[12 0.08 0.08
Number of household members 2.4 2.22 0.75 2.25 0.82
Enter the EEI in first attempt 3.1 0.49 0.49
Father’s occupation 3.5
Non-main earner—retired 0.05 0.08
Government or public sector 0.54 0.52
Private sector 0.42 0.40
Mother’s occupation 3.5
Non-main earner—house wife 0.71 0.72
Government or public sector 0.17 0.17
Private sector 0.12 0.11
Father’s education 5.3
Middle school or less 0.04 0.09
High school 0.10 0.09
College 0.53 0.50
Postgraduate 0.34 0.32
Mother’s education 5.3
Middle school or less 0.11 0.14
High school 0.16 0.16
College 0.36 0.35
Postgraduate 0.36 0.35
First wage after graduation (lac) 10.0
No job 0.03 0.09
3–4 0.07 0.06
4–5 0.12 0.11
5–7 0.39 0.35
7–10 0.19 0.17
10–15 0.12 0.11
15-25 0.03 0.03
[25 0.05 0.07
Type of school 14.8
State/Government 0.17 0.17
KV/Minority 0.05 0.08

123
High Educ (2016) 71:611–649 643

Table 11 continued
Variable % Missing (N = 453) Original data set Imputed data set

Mean SD Mean SD

Private 0.78 0.75

Source: Survey data from the EEI’s graduating class, 2008

Table 12 Checking missing at random assumption (I)

(1) (2) (3) (4) (5) (6)
School Medium Coaching Wage Father’s Mother’s
type instruction exp. education education

Male -0.272 -0.430 -0.007 -0.358 -0.050 -0.249

(0.247) (0.284) (0.244) (0.294) (0.383) (0.344)
General category 0.342 -0.056 -0.238 -0.321 -0.031 -0.119
(0.262) (0.314) (0.220) (0.261) (0.306) (0.320)
Cumulative GPA -0.126 -0.111 -0.109 -0.273** -0.124 -0.111
(0.096) (0.124) (0.089) (0.117) (0.140) (0.144)
Selective major -0.002 -0.115 -0.066 -0.178 0.446* 0.272
(0.193) (0.246) (0.178) (0.237) (0.247) (0.252)
Master 0.198 -0.032 0.414** -0.485** 0.120 -0.304
(0.200) (0.239) (0.194) (0.206) (0.274) (0.243)
Middle annual income 3–6 lacs) -0.069 0.031 0.005 -0.450** -0.406 -0.516**
(0.169) (0.215) (0.157) (0.221) (0.249) (0.246)
High annual income ([6 lacs) -0.002 0.061 0.063 0.477* -0.088 -0.650
(0.238) (0.304) (0.223) (0.249) (0.317) (0.423)
Number of hh members -0.023 -0.021 -0.165 0.055 -0.024 0.106
(0.103) (0.133) (0.102) (0.121) (0.133) (0.136)
Repeated attempts to enter EEI 0.224 0.288 0.178 0.165 -0.098 0.034
(0.168) (0.210) (0.154) (0.198) (0.238) (0.244)
Number of observations 434 434 434 434 434 434

Standard errors clustered at the school level

* Significant at 10 %; ** significant at 5 %; *** significant at 1 %

123
644 High Educ (2016) 71:611–649

Table 13 Checking missing at random assumption (II)

(1) (2) (3) (4) (5)
Father’s Mother’s English Type of Number of
occupation occupation deficiency job missing values

Male -0.105 -0.105 0.118 -0.181 -0.200

(0.463) (0.463) (0.226) (0.268) (0.190)
General category 0.664 0.664 0.045 -0.059 -0.047
(0.510) (0.510) (0.227) (0.242) (0.182)
Cumulative GPA -0.122 -0.122 0.085 -0.212** -0.152**
(0.159) (0.159) (0.084) (0.099) (0.069)
Selective major -0.083 -0.083 -0.125 -0.150 -0.050
(0.364) (0.364) (0.166) (0.206) (0.138)
Master 0.067 0.067 -0.115 -0.240 -0.032
(0.351) (0.351) (0.163) (0.190) (0.139)
Middle annual income (3–6 lacs) 0.175 0.175 0.262* -0.182 -0.085
(0.283) (0.283) (0.147) (0.185) (0.124)
High annual income ([6 lacs) -0.137 -0.137 -0.009 0.546** 0.146
(0.475) (0.475) (0.216) (0.227) (0.176)
Number of hh members 0.324** 0.324** 0.013 0.038 0.043
(0.128) (0.128) (0.090) (0.105) (0.075)
Repeated attempts to enter EEI -0.263 -0.263 0.026 0.003 0.131
(0.286) (0.286) (0.147) (0.173) (0.122)
Number of observations 434 434 434 434 434

Standard errors clustered at the school level

* Significant at 10 %; ** significant at 5 %; *** significant at 1 %

Appendix 3: Multiple random imputation

Allison (2001) proposes multiple imputation methods as an alternative to maximum

likelihood function methods. Like maximum likelihood, multiple imputation estimates are
consistent and asymptotically normal and close to being asymptotically efficient. In
addition, multiple imputation has two big advantages over maximum likelihood: (i) It can
be applied to any kind of data or model and (ii) the imputation procedure can be imple-
mented using conventional software. Since imputed values are random draws, the major
disadvantage of multiple imputation is that it produces different imputed databases every
time they are used.
The most widely used method for multiple imputation is the Markov Chain Monte Carlo
algorithm based on linear regression. However, in the case of the EEI data, important
complications arise from the fact that some of the missing variables are categorical. The
Monte Carlo method presumes that every variable with missing data is normally dis-
tributed and that is clearly not the case for categorical variables. An alternative approach is
known as ‘‘sequential generalized regression’’ or ‘‘multiple imputation for chained equa-
tions’’ (MICE). Instead of assuming a single multivariate model for all the data, this
method specifies a separate regression model for each variable which is used to impute

123
High Educ (2016) 71:611–649 645

missing values. This method is thus flexible and allows us to incorporate logistic, binomial,
or multinomial models for categorical variables.33
Each model is estimated sequentially using available data, starting with the variable that
has the fewest missing data and proceeding to the variable with the most missing data.
After each model is estimated, the parameter estimates are used to generate imputed
values. Once imputed values have been generated for all the missing data, the sequential
imputation process is repeated, except now the imputed values of the previous round are
used as predictors for imputing other variables. The main drawback of sequential gener-
alized regressions is that no theory guarantees convergence to the correct distribution for
the missing values. However, simulation-based evidence in Van Buuren et al. (2006)
suggests that the method works well.
The MICE method is implemented in the following way:
1. For each dependent variable to be imputed, choose a model that reflects the type of
data.
2. First round: Imputation starts with dependent variable with the fewest missing data and
proceeds to dependent variable with the most missing data.
• Order-dependent variables according to amount of missing data from Y1 to Yk :
Denote variables with complete data values as X.
• Regress Y1 on X and obtain b b ðb
b and V bÞ: Generate imputed values using observed
covariates and coefficients drawn from N( b b ðb
b; V bÞÞ.
• Regress Y2 on X and Y1 (including imputed values of Y1 ) and obtain imputed
values.
• Continue until all regression models have been estimated.
3. Second and subsequent rounds repeat the process, but each variable is regressed on all
other variables, using imputed values from previous rounds.
4. Process ends when stable imputed values are reached or after a specified number of
rounds.

Appendix 4: 2009 Entrance exam applicant data

In general, Fig. 7 shows that there are greater concentrations of exam applicants in north
eastern areas, especially in richer districts. Moreover, central regions of the country with
higher urban poverty rates are contributing less to the applicant pool. Although this evi-
dence is suggestive, we need to take into account the population ‘‘at risk’’ of taking the
exam in each district to check whether certain district characteristics are related to a
relatively higher proportion of high school graduates taking the entrance exam. We
approximate the probability of taking the entrance exam in a given district as the number of
2009 applicants from a given PIN code divided by the number of high school students
enrolled in grades 9th through 12th in 2006.34 Even though the 2009 applicant data do not

33
Allison (2001).
34
The latter comes from the State Profile 2005–2006 prepared by the Indian Ministry of Human Resource
and Development. We acknowledge that the total population ‘‘at risk’’ of taking the exam will be overes-
timated, especially due to higher dropout in the last years of high school, and more so in poorer districts.
Unfortunately, data for 12th grade enrollment in 2008 were not publicly available at the district level, only at
the state level.

123
646 High Educ (2016) 71:611–649

Fig. 7 District-level poverty rate and number of exam applicants. Source: Centralized entrance exam,
applicant data 2009. Poverty rates from and Chakravarty and Somanathan (2008)

contain information about students’ placement, we code all students who made it into their
respective merit list and are ranked above the maximum number of seats available for their
group as admitted students. We can then proxy the district’s probability of getting into
college as the number of admitted students divided by the total number of applicants from
the corresponding PIN code.

123
High Educ (2016) 71:611–649 647

(a) .02 (b)

.05
Probability of taking the JEE

Probability of Admission
.04
.015

.03
.01

.02
.005

.01
0

0
0 20 40 60 80 100 0 20 40 60 80 100
% %
By poverty rate By rural population By poverty rate By rural population
By share of SC/ST By share of SC/ST

Fig. 8 Probability of taking the entrance exam and probability of admission. a Probability of taking Exam,
b Probability of admission. Source: Centralized entrance exam, applicant data 2009. Note: Probabilities are
locally mean-smoothed using a kernel-weighted local polynomial smoother
Average Marks in the District (100 scale)

Average Marks in the District (100 scale)

(a) (b)
10

10
8

8
6

6
4

4
2

2
0

0 10 20 30 40 0 20 40 60 80 100
Share of SC/ST Applicants in the District % of Urban Poor in the District
Total Marks Math Total Marks Math
Chemistry Physics Chemistry Physics

Fig. 9 District Average Total Marks and Marks by Subject in the 2009 Entrance Exam. a SC/ST share,
b Poverty. Source: Centralized Entrance Exam, Applicant data 2009. Note: Average Marks in panel (a) are
locally mean-smoothed using a kernel-weighted local polynomial smoother

Using the results from non-parametric locally linear regressions in the sample of dis-
tricts with exam applicants in 2009, Fig. 8 plots districts’ probability of taking the exam
and the probability of getting into college as functions of urban poverty rates, share of rural
population, and share SC/ST population. The blue lines in panels (a) and (b) show that both
the probability of taking the college entrance exam and the probability of getting in do not
seem to be affected by the share of minority population in the district. This pattern is
particularly interesting if one takes into account that areas with higher concentrations of
SC/ST population tend to have lower average performance in the entrance exam (both in
the aggregate and by subject) as shown in panel (a), Fig. 9. In the absence of AA policies,
the lower prospects of success in the exam in areas with higher concentrations of minority
population would lead to a lower probability of taking the exam as well as lower chances of
being admitted. The fact that we do not identify a relationship between the district’s share

123
648 High Educ (2016) 71:611–649

of SC/ST population and the probabilities of writing the exam or getting in suggests that
admission preferences particularly motivate minority students to take the exam and
facilitate their admission into college.
When we order districts by their poverty rate, panel (a) in Fig. 8 shows a modest decline
in the probability of taking the exam as the percentage of urban poor rises. However, the
black line in panel (b) suggests that there is no relationship between the probability of
getting in and poverty. This could be due to preferences for SC/ST who tend to be poorer
than the GE group. The small differences in the proportion of people getting in may follow
from the small differences in average performance among poor and rich districts exhibited
in panel (b) in Fig. 9. Finally, the gray lines in panels (a) and (b) in Fig. 8 show that the
probability of taking the exam and the probability of getting into college are both
decreasing in the share of rural population, which reflects AA’s lack of focus on rural
students.

References
Akyol, P., & Krishna, K. (2014). Preferences, selection, and value added: A structural approach. NBER
working paper no. 20013.
Allison, P. (2001). Missing data. Thousand Oaks, CA: Sage.
Alon, S., & Tienda, M. (2005). Assessing the ‘‘Mismatch’’ hypothesis: Differences in college graduation
rates by institutional selectivity. Sociology of Education, 78, 294–315.
Alon, S., & Tienda, M. (2007). Diversity, opportunity, and the shifting meritocracy in higher education.
American Sociological Review, 72(4), 487–511.
Alon, S., & Malamud, O. (2014). The impact of Israel’s class-based affirmative action policy on admission
and academic outcomes. Economics of Education Review, 40, 123–139.
Altonji, J. G., Elder, T. E., & Taber, C. R. (2005). Selection on observed and unobserved variables:
Assessing the effectiveness of catholic schools. Journal of Political Economy, 113(1), 151–184.
Arcidiacono, P. (2005). Affirmative action in higher education: How do admission and financial aid rules
affect future earnings? Econometrica, 73(5), 1477–1524.
Arcidiacono, P., Ausejo, E., & Spenner, K. (2011). What happens after enrollment? An analysis of the time
path of racial differences in GPA and major choice. Working paper.
Assuno, J., & Ferman, B. (2013). Does affirmative action enhance or undercut investment incentives?
Evidence from quotas in Brazilian public universities. Working paper
Bertrand, M., Hanna, R., & Mullainathan, S. (2010). Affirmative action in education: Evidence from
engineering college admissions in India. Journal of Public Economics, 94(1–2), 16–29.
Blau, J., Moller, S., & Jones, L. (2004). Why test? Talent loss and enrolment loss. Social Science Research,
33, 409–434.
Bowen, W., & Bok, D. (1998). The Shape of the River. Princeton: Princeton University Press.
Carnevale, A. P., & Rose, S. J. (2003). Socioeconomic status, race/ethnicity and selective college admis-
sions. New York: The Century Foundation Press.
Chakravarty, S., & Somanathan, E. (2008). Discrimination in an elite labour market? Job placements at IIM-
Ahmedabad. Economic and Political Weekly, 43(44), 45–50.
Chaudhuri, S., & Gupta, N. (2009). Levels of living and poverty patterns: A district-wise analysis for India.
Economic and Political Weekly, 44(9), 94–110.
Chay, K., McEwan, P. J., & Urquiola, M. (2005). The central role of noise in evaluating interventions that
use test scores to rank schools. American Economic Review, 95, 1237–1258.
Francis, A., & Tannuri-Pianto, M. (2012). Using Brazil’s racial continuum to examine the short-term effects
of affirmative action in higher education. The Journal of Human Resources, 47, 754–784.
Frisancho, V., & Krishna, K. (2012). Affirmative action in higher education in India: Targeting, catch up,
and mismatch. NBER working paper no. 17727.
Frisancho, V., Krishna, K., Lychagin, S., & Yavas, C. (2014). Better luck next time: Learning through
retaking. Working paper.
Kahlenberg, R. D. (Ed.). (1996). The Remedy: Class, race, and affirmative action. New York: Basic Books.
Kahlenberg, R. D. (Ed.). (2004). America’s untapped resource: Low-income students in higher education.
New York: The Century Foundation Press.

123
High Educ (2016) 71:611–649 649

Kochar, A. (2010). Affirmative action through quotas: The effect on learning In India. Stanford Center for
International Development, Working paper no. 430.
Loury, L. D., & Garman, D. (1993). Affirmative action in higher education. American Economic Review,
83(2), 99–103.
Loury, L. D., & Garman, D. (1995). College selectivity and earnings. Journal of Labor Economics, 13(2),
289–308.
NCEUS. (2009). The Challenge of Employment in India An Informal Economy Perspective. National
Commission for Enterprises in the Unorganised Sector: Technical Report.
Rosenbaum, P. R., & Rubin, D. B. (1983). The Central Role of the Propensity Score in Observational
Studies for Causal Effects. Biometrika, 70, 41–55.
Rothstein, J., & Yoon, A. (2008). Affirmative action in law school admissions: What do racial preferences
do? University of Chicago Law Review, 75(2), 649–714.
Rothstein, J., & Yoon, A. (2009). Mismatch in Law School. Working paper.
Saeme, D. (2014). does the implementation of affirmative action in a competitive setting incentivize
underrepresented public school applicants performance? Evidence from So Paulo. The Yale Journal of
Economics, 2, 93–114.
Sander, R. (2004). A systemic analysis of affirmative action in American law schools. Stanford Law Review,
57(2), 367–483.
Van Buuren, S., Brand, J. P. L., & Groothuis-Oudshoorn, C. G. M. (2006). Fully conditional specification in
multivariate imputation. Journal of Statistical Computation and Simulation, 76, 1046–1064.

123