DOES MANDATED POLITICAL REPRESENTATION HELP THE POOR?
Theory and Evidence from India
BY ANIRBAN MITRA1,
November 2013, revised February 2014
A BSTRACT
Mandated political representation for minorities involves earmarking (“reserving”) certain electoral districts where only minority–group candidates are permitted to contest. This paper builds a
political–economy model to analyze the effect of such quotas on redistribution in equilibrium. Our
model predicts that, in situations where the minority–group is economically disadvantaged and
where voters favor candidates from their own group, such reservation actually reduces transfers to
poorer groups. Moreover, redistribution in reserved districts leads to a rise in within–(minority)
group inequality. These predictions are examined in the context of reservation for Scheduled Castes
(SCs) in the Indian parliament. We find evidence of lower implementation of poverty–alleviation
programs in districts reserved for SC politicians as well as an increase in SC–inequality.
JEL codes: D72, D78, O20
Keywords: Affirmative action, income distribution, political economy
1
ESOP, Department of Economics, University of Oslo. Email: [email protected]. I am deeply indebted to
Debraj Ray for his insightful comments, support and encouragement. I would like to thank Rajeev Dehejia, Oeindrila
Dube, Raquel Fernández, Shabana Mitra, Kalle Moene, Natalija Novta, Sarmistha Pal, Rohini Pande, Alessia Russo,
Rajiv Sethi, Adam Storeygard, Lore Vandewalle, Shing-Yi Wang and seminar participants at the Royal Economic
Society Annual conference 2011,the 19th BREAD Conference on Development Economics 2011 (pre-conference
session) at NYU; the APET workshop PED 2011 (Bangkok), the EconCon workshop 2011 (at NYU), the Graduate
Student Conference, Washington University, St. Louis, MO (2011), ESOP lunch seminar at the University of Oslo
(2012), the 9th Annual Conference on Economic Growth and Development (2013) at ISI Delhi for their comments
and suggestions. I would also like to express my gratitude to Rohini Somanathan for sharing relevant data from the
1971 Census of India. All existing errors are solely mine.
2
1. I NTRODUCTION
Serious concerns exist about the extent to which minority groups participate in policy–making.
These concerns are heightened when the minorities are socio–economically disadvantaged. For
example, blacks in the US are a minority who exhibit lower levels of educational attainments and
greater poverty, as compared to whites. Although in terms of population women may not be minority in many countries, their participation in various domains have been quite limited over several
centuries. Many countries have different quota requirements in various occupations and educational institutions (particularly in the public sector) to correct this anomaly. India has in place
wide–ranging affirmative action programs (often termed “reservation”) for minority groups called
the Scheduled Castes and Tribes, an important component of which has been mandated representation in the legislature. The mandate involves earmarking a fraction of electoral districts where
only these minority group candidates are permitted to contest. More recently, similar policies have
been put into action for women with the aim to enhance their presence in politics. A natural question that arises is the following: How does political reservation for a minority group affect the
conditions of the group–members living in these reserved districts?
There have been several empirical studies (discussed later) that confirm that such political reservation benefits the minority group in the aggregate. However, the question of who gains and
who loses within the minority group has not been studied before.2 So do these quotas benefit the
poor within the minority or the rich? Also, what are the possible channels through which this
effect takes place? This paper attempts to answer questions of this nature by putting forward a
political–economy model which is based on the standard probabilistic–voting framework. The
theory developed here generates two main predictions. They are: (i) In the context of an economically disadvantaged minority, political reservation reduces transfers to poorer (income) groups
when voters favor candidates from their own group. (ii) Such reservation leads to greater inequality within the minority group.
The mechanism underlying the theory does not rely upon either considerations of statistical discrimination, group–wise differences in reputation or perceived ability.3 Instead, our results stem
from some well–known features of models of political competition which have been quite popular
in the literature. The standard one–period probabilistic–voting setup is modified and extended to a
(simple) two–stage game to study the issue of affirmative action in the legislature. In the first stage,
parties choose candidates from different ethnic groups (majority or minority)4 while in the second
stage the fielded candidates propose redistribution policies. The presence of political quotas implies that the district in question may be reserved, i.e, contested only by members of the ethnic
minority; hence, all political parties are restricted in their first stage choice of candidates.
We build on the following intuition from a standard model of redistributive politics like Lindbeck and Weibull (1987). When parties maximize their expected votes by promising transfers
across different groups (or redistributing resources), the group with the least ideological/partisan
2
This question is especially relevant when the minority group — while economically disadvantaged — exhibits a fair
degree of heterogeneity in terms of income.
3
This is not to say that such considerations are not important. The point is that one could get the results we obtain even
without reliance on such features.
4
We use the term “ethnic” group here to refer to the two segments of the population: the majority and the minority.
However, the relevant marker need not be ethnicity; it could be language, race or even gender. However, for ease of
exposition we continue to use the term “ethnicity” while hoping for the reader’s indulgence in this matter.
3
bias (the “swing” group) is most favored by both parties in equilibrium. This idea is taken as a
stepping–stone on which our theory is developed. What role does political reservation play here?
Reservation of a district implies all contesting candidates must be from the ethnic minority. We
assume that ceteris paribus, every voter feels a positive bias for a candidate from his own ethnic
group.5 Reservation by influencing the ethnic identities of the party candidates potentially has
an effect on a voter’s ideological bias and thereby on the identity of the swing group. The other
component of the ideological bias stems from a party–wise affiliation, with poorer voters ex ante
preferring a certain party while their richer counterparts ex ante prefer the other one; call the former
a “pro–poor” party and the latter “pro–rich”.6
In this setup, a situation where both parties choose to field candidates from the same ethnicity —
be it majority or minority — is substantively different, in terms of redistribution policies chosen
in equilibrium, from a situation where one of the parties fields a minority group candidate while
its rival party fields a majority group candidate. In what follows, we shall call the latter situation a “mixed–ethnicity candidate” scenario. The intuition behind the main result is stated in the
following paragraphs.
Consider a reserved district. Here the ideological bias of any voter is driven by simply the party–
wise bias; the ethnicity bias loses relevance as the two candidates are from the same ethnic group
(by mandate, the minority). So the swing group is presumably some intermediate income group
which is neither too poor nor too rich. Next consider the following “mixed–ethnicity candidate”
situation. Suppose the “pro–poor” party fields a candidate from the minority while the rival party
picks one from the majority group. In this scenario, both biases — party–wise and ethnicity
(candidate–wise) — matter. Loosely–speaking, the swing group here is basically an income group
where the major ethnic group members (who form the bulk of the group) exhibit near–zero ex ante
bias. Note that it is now a poor, rather than some middle income, ethnic majority voter who will
exhibit low bias ex ante. Why? This voter prefers the “pro–poor” party but is averse to this party’s
candidate on ethnicity–based considerations; thus making him largely indifferent between the two
options. Therefore the swing group in this situation is relatively poor in contrast to the one in a
reserved district.
Political reservation — by eliminating such “mixed–ethnicity candidate” cases — results in penalizing lower–income groups. An implication is that the transfers in a reserved district end up
being concentrated at intermediate rather than lower income groups. This leads to a widening of
disparities within the economically disadvantaged minority group. Interestingly, this is also true
for the majority group.7 But if the minority group is indeed strictly economically disdvantaged
to begin with, the implications in terms of increased within–group inequality are more salient for
the minority group. Of course, the question whether one would observe candidates from different
ethnic groups pitted against each other is important. It may be that quotas are required to enhance
minority participation to the degree that they feel confident about running for elections in districts
which are not reserved. We take up this matter in detail later.
5
In fact, for India there exists abundant anecdotal evidence on voting behavior along caste lines. Banerjee and Pande
(2007) offer a formal model to study some implications (like the effect on legislator quality) of this phenomenon.
6
One could think of one party being more leftist in its ideological position and hence attracting the “toiling masses”
while the other party could be thought of as more “pro–business” and so more right–wing.
7
For our theoretical results, we only require that the ethnic minority be no richer than the majority; hence, identical
income distributions for the two ethnic groups is permitted.
4
Our model delivers that the nature of redistribution in equilibrium takes a particular form when the
parties decide — in the first stage of the game — to field candidates from the same ethnic group.
Moreover, this form is the same regardless of whether the candidates are all from the majority
ethnicity or the minority. This is perhaps expected when one considers that even in a reserved
district every citizen, regardless of ethnicity, is allowed to vote: when no candidate can count on
any additional ideological support on the basis of ethnic identity, they behave the same irrespective
of their own ethnicity.
Our simple model permits us to discuss the effect of an across–the–board reduction in the ethnicity–
based bias. Apart from an interesting theoretical exercise, it seems plausible (even if a trifle optimistic) that with the passage of time ethnic biases become less important. Perhaps somewhat
intriguingly, we find that the effect of (exogenously) lowering the ethnic bias is far from unambiguous. Although it is possible that reducing the bias makes the effect of reservation less pronounced
on the equilibrium redistribution policy, it also may make the “mixed–ethnicity candidate” scenario (first stage decision) more likely. Note, the “mixed–ethnicity candidate” case is essentially
where the introduction of the political quota has a marked impact. Thus, the two forces can work
in opposite directions resulting in ambiguity.
This paper also presents some empirical evidence consistent with the theoretical predictions. The
application of the theory is certainly not universal and given the constraints on data availability, we
chose to limit ourselves to examining one particular context in some detail — namely, the reservation for the Scheduled Castes (henceforth, SCs) in the Indian parliament. India has implemented
several employment–based poverty reduction programs (collectively known as Public Works). We
study the differences in the implementation of these programs across reserved versus unreserved
districts. These programs are the initiative of the central government and so the Members of Parliament (MPs) have a prominent role in shaping and implementing them.8 Using data from the
43rd round of the NSSO consumer expenditure survey (conducted in 1987-88), we show evidence
of lower implementation of the Public Works programs in constituencies reserved for SC MPs.9
Additionally, district–level regressions reveal evidence of greater within–SC inequality in reserved
districts. This effect persists for different measures of within–group disparities. This is also true
when looking at SC–inequality relative to overall inequality in the district. Our findings suggest
that mandated political representation may well be widening disparities within these disadvantaged
minorities. Not only is this worrying from the angle of poverty–reduction but it also serves to
alienate the bulk of the minority community by creating a sense of betrayal in them.
This paper is part of the broad literature which studies the impact of ethnicity/gender–based quotas on economic outcomes. There have been several studies on the impact of reservation on the
provision of publicly provided goods at local levels of government like village councils (see for
e.g., Besley et al (2004), Munshi and Rosenzweig (2008), Bardhan, Mookherjee and Parra Torrado
(2010)). Chattopadhyay and Duflo (2004a, 2004b) use political reservations for women in India to
study the impact of women’s leadership on policy decisions. Kotsadam and Nerman (2014) study
gender quotas in Latin American national elections and find no effects on policy or political participation. Pande (2003) finds that political reservation for SCs in the state legislature has resulted
8
The data corresponds to a period prior to the empowerment of local–level government bodies (“Panchayati Raj”).
Hence local bodies barely existed then (or even if they did, they had no say over local public finance).
9
An electoral district is referred to as a “constituency” in India.
5
in observable rise in targeted transfers towards these groups.10 Jensenius (2014) focuses on SC
reservation in the state assemblies (1971–2000) and finds no significant effect of such quotas on
several development indicators.
In terms of studying the redistributive outcomes of political reservation, there has been little work
— either theoretical or empirical — and this is precisely the gap this paper seeks to close. A
closely related (empirical) paper which studies the impact of political reservation on a sub–group
of the population is Chin and Prakash (2011). They estimate the impact of political reservation
in the state–level legislature on overall state–level poverty. They find that, at the state–level, ST
reservation reduces poverty while SC reservation has no such impact. In contrast, here we first
present a theory which links reservation to redistributive outcomes in general (and thereby, poverty)
and provides predictions regarding the effect of reservation on within–minority inequality while
emphasizing a particular (electoral) channel; then we present some evidence consistent with our
theory. In terms of the focus on diversity and redistribution, this paper is similar to Fernández and
Levy (2008). They study the relationship between redistribution and taste diversity using a model
with endogenous platforms involving redistribution and targeted public goods, and find a non–
monotonic relationship. However, there are important differences between their model and the
one presented in this paper.11 Austen-Smith and Wallerstein (2006) develop a theoretical model to
analyze the role of affirmative action (along racial lines) on labor–market outcomes. Interestingly,
they show how the less advantaged members of the majority can be harmed by affirmative action.
The remainder of the paper is organized as follows. Section 2 contains the theory while Section 2.4
considers a simple extension of the baseline model. Section 3 presents the empirical findings and
Section 4 concludes. Proofs of the results of extension of the baseline model (presented in Section
2.4), details on the data used, the background of political reservation in India and additional tables
can be found in the Appendix.
2. T HE M ODEL
Society is populated by a unit mass of individuals and every individual — indexed by i — is
characterized by two features. One is individual i’s income, denoted by wi , and the other is i’s
ethnicity, denoted by ei .12 Every individual belongs to either one of two groups: majority (h) or
minority (l).13 Hence for every i, ei ∈ {h, l}.
Let the distribution of incomes in society be represented by the cdf G with support on [0, w] where
w > 0 is “large”. Take any income level w. Let π(w) denote the proportion of w-earners who
are type l. Also, let π(w) be continuous in w. In society, the l–types form a minority. Hence,
Rw
π(w) dG(w) < 12 . Also, the l–types are economically disadvantaged; their numbers tend to
0
10
Pande (2003) finds that the effects of similar reservation for STs have been somewhat different (from SC reservation).
They have endogenous parties unlike the two-party framework in the current paper. The election process in their
paper is more in the spirit of the “citizen–candidate” model while the current paper is more in the Downsian tradition.
Also, they do not deal with affirmative action.
12
In principle ei could stand for i’s race, religion, gender or any such non–income marker. For ease of exposition, we
shall continue to refer to ei as ethnicity.
13
One can view h and l as “high” and “low”, respectively, in the sense of strength in numbers. Also, the SCs are often
referred to as “low castes” while non–SCs are called “high castes” in the literature.
11
6
dwindle as w increases. To capture this aspect, we assume that there is some threshold income
level, call it w, beyond which π(w) is weakly decreasing in w with π(w) < 1/2.
A balanced-budget redistribution — or simply redistribution for short — is a continuous function
z : [0, w] → R, where z(w) is the transfer to each individual earning w, and:
Rw
(i) z satisfies the budget constraint z(w) dG(w) = 0.
0
(ii) Every individual’s net consumption (≡ w + z(w)) is non–negative.
We will denote the set of all such functions by Z.
There are two political parties, denoted by R and P , where R is viewed as pro–rich, and P as
pro–poor, in a sense that we make precise below. Each party fields a candidate of either ethnicity
who proposes a particular redistribution. However, the constituency in question may be reserved
for l–type candidates. In this case, each party is constrained to field a l–type candidate. Note,
the game proceeds in two stages: in the first stage, the parties choose their respective candidates
concurrently while in the second stage, the candidates make simultaneous offers of redistribution.14
Let γ denote the candidate ethnicity configuration in the following manner:
γ = (e(P ), e(R))
where the first argument refers to P –candidate’s type and the second refers to party R–candidate’s
type. Therefore, γ ∈ {(h, h), (l, l), (l, h), (h, l)}. Note, γ is determined by the parties’ choices in
the first stage. For a reserved constituency, γ = (l, l) by definition.
An individual’s preferences over candidates (and their proposed policies) are described as follows.
First, individual i exhibits a bias αi , positive or negative, for party P . The corresponding payoff
from R is normalized to be zero; hence αi is really the net bias for party P . This bias has two
main components: one that stems from i’s emotive affiliation with party P (relative to R), and the
other which arises from i’s association with the ethnic identity of party P ’s candidate (relative to
R’s candidate).
Specifically, we assume that individual i feels an affiliation t(wi ) with party P (relative to R),
which is continuous and naturally decreasing in w.15 Thus, t : [0, w] → R. As for the “ethnic
bias”, assume that the voter feels some degree of association for a candidate of the same type as
the voter.16 This would depend upon the candidate configuration γ and hence we denote it by si (γ).
Combining, we write
α(wi , ei ) = t(wi ) + si (γ).
Here, si (γ) takes on one of the three values: s, −s, or 0 where s > 0. In particular, when P 0 s
(R0 s) candidate is from the same ethnic group as voter i while R0 s (P 0 s) candidate is from the other
group, then the ethnic bias for voter i equals s (−s). Note, when both parties field candidates
14
The second stage game is also referred to as the “expected–plurality game” in the subsequent text.
In this sense, P is viewed as pro–poor. Also, in equilibrium, P ’s actions confirm this label.
16
There could be long–standing reasons for such bias; say, for e.g., asymmetric information (better knowledge of
members of own type). Such behavior has been dubbed as “homophily” in many social contexts. In India, there is
plenty of anecdotal evidence on voting behavior along caste lines. Also, there is evidence of bias against women
leaders. See Banerjee and Pande (2007) and Beaman et al. (2009), among others.
15
7
from the same ethnic group, this type–based association factor effectively cancels out. Hence,
si (h, h) = si (l, l) = 0.
This is not to say that the absolute ethnic bias any h–type voter feels towards either party is the
same whether γ = (h, h) or γ = (l, l). In other words, any h–type voter may strictly prefer (h, h)
over (l, l). However, when it comes to determining si (γ) this does not matter as si (γ) is the net
bias towards P 0 s candidate relative to R’s candidate. Analogous considerations may apply to any
l–type voter.17
Therefore, for any individual i:
si (γ) = s
if e(P ) 6= e(R) and ei = e(P )
si (γ) = −s if e(P ) 6= e(R) and ei = e(R)
si (γ) = 0
if e(P ) = e(R).
We can now write individual i’s bias as
αi = α(wi , ei ) + i ,
where the extra term i is just mean-zero noise. The individual sees the realization of i before
she votes; the politicians (the parties and their candidates) do not; more on this below. We assume
that i is independently and identically distributed across individuals, with a symmetric, unimodal
density f (and corresponding cdf F ) that has its support on R.
To this non-pecuniary bias αi we add the economic benefit from a proposed redistribution to arrive
at the overall payoff. Recall z(w) is the transfer to each individual earning w. We write the
economic benefit to an individual earning w from redistribution z ∈ Z as
m(z, w) ≡ u(w + z(w)),
where u denotes a utility function with u0 > 0, u00 < 0 and u0 (0) = ∞.
Say party P ’s candidate proposes x, and R’s candidate proposes y where x, y ∈ Z. An individual
i earning wi , with bias αi will vote for party P ’s candidate if
m(x, wi ) + αi > m(y, wi ),
will vote for R’s candidate if the opposite inequality holds, and will be indifferent in case of
equality.
From the perspective of the party, any individual’s vote is stochastic. The probability that citizen i
will vote for party P ’s candidate is given by
pi ≡ 1 − F (m(y, wi ) − m(x, wi ) − α(wi , ei )) .
R
The expected plurality for party P is proportional to pi ,18 and this is what party P ’s candidate
i
seeks to maximize — and party R’s candidate minimize — through the appropriate choice of
17
Here, every voter is assumed to cast his vote in favor of one party’s candidate or the other. Of course, one could
introduce the notion of “costly voting” where say any h–type voter may feel reluctant to vote for either party when he
observes γ = (l, l). This would no doubt lead to interesting predictions for turnout but we do not pursue this direction
here.
R
R
18
To be precise, the expected plurality is given by [pi − (1 − pi )] = [2pi − 1].
i
i
8
Candidates
propose
policies
Parties select
candidates
Each voter
draws bias αi
Voters cast
their votes
F IGURE 1. Timing of the game
policy in the second stage, conditional on candidate choice in the prior stage. Figure 1 depicts the
sequence of moves in the game.
Recall that in the first stage of the game, the two parties choose their respective candidates and
hence determine γ. Of course, this choice is trivial in a reserved constituency where γ = (l, l)
by mandate. We turn to an explicit discussion of party payoffs (which would drive the first stage
choices) in section 2.3.
The redistribution profile (x, y) proposed by the fielded candidates of P and R, respectively, in
equilibrium will — in principle — depend on γ.19 Therefore, for any given γ, an equilibrium of
the expected–plurality game is defined by a redistribution pair (x, y) which constitute mutual best
responses from the perspective of the two candidates. Note, that this expected–plurality game is,
by its nature, a zero–sum game.
2.1. Equilibrium. We use the standard notion of subgame perfection as the equilibrium concept
for this game. To be specific, an equilibrium of this game is given by a collection of candidate
choices and redistribution policies, {e(P ), e(R); x, y}, which satisfy the following:
(i) The redistribution policies x and y constitute mutual best–responses for the fielded candidates
given (e(P ), e(R)).
(ii) The candidate choices (e(P ), e(R)) constitute mutual best–responses for the two parties given
the redistribution policies x and y.
Existence: This can be guaranteed by simply assuming R— like in Lindbeck and Weibull (1987)
— that party P ’s candidate’s objective function, namely pi , is concave in x for any given y and
i
convex in y for any given x.
Characterization: We now proceed to describe the set of equlibria for this simple game. Given the
equilibrium notion adopted, we start by solving backwards.
It will be useful to define the term
σ(γ, w) ≡ π(w)f (α(w, l)) + (1 − π(w))f (α(w, h))
This term is important for subsequent analysis and hence requires some interpretation. First, fix a
candidate configuration γ; this effectively pins down the ethnic bias for every voter. Now consider
some level of income, say w, the values α(w, l) and α(w, h) are “small” in magnitude. This
basically means that the group characterized by income level w (group w henceforth, for brevity),
does not exhibit much partisan bias ex-ante. Recall that the density f is unimodal and symmetric
around 0. Hence, such lack of strong bias indicates that σ(γ, w) will exhibit a high value.
19
We will make this dependence explicit by indexing all relevant parameters by γ.
9
Alternatively, if the values α(w, l) and α(w, h) are “large” in magnitude — suggesting that the
members of group w feel strongly about one party vis–a–vis the other party — then σ(γ, w) will
exhibit a low value for this group w. In this sense, one can think of σ(γ, w) as representing the
“average swing propensity” of group w. It captures the extent to which the members of a group
may be willing to switch loyalties.
From the perspective of the parties, the groups with high “swing” propensity assume importance
as they are ex-ante more responsive to transfers. The following proposition makes this explicit.
P ROPOSITION 1. For any given candidate configuration γ, there is a unique equilibrium of this
expected–plurality game. In that equilibrium, there is a unique redistribution scheme xγ offered
by both parties, with the property that
σ(γ, w)[u0 (w + x(w))] = λγ
for every w in [0, w] and some λγ > 0. Moreover, w + x(w) varies positively with σ(γ, w).
Proof. The arguments here closely parallel those in Theorem 1 of Lindbeck and Weibull (1987).
Let V (x(w), y(w)) ≡ π(w)[1 − F (d(w) − α(w, l))] + (1 − π(w))[1
R − F (d(w) − α(w, h))]. Let
d(w) represent the utility differential m(y, w) − m(x, w). Rewrite pi as
i
Zw
[V (x(w), y(w))] dG(w)
0
P ’s candidate seeks to maximize the integral above by choosing redistribution x while R’s candidate seeks to minimize the same by choosing y. Suppose (x, y) is an equilibrium of this expected–
plurality game.
Pick any w in [0, w]. For this group w,
Vx(w) = [π(w)f (d(w) − α(w, l)) + (1 − π(w))f (d(w) − α(w, h))]u0 (w + x(w)).
Note, Vx(w) > 0 since both f , u0 > 0. Also, Vx(w) is continuous in w and this implies the value of
Vx(w) must be the same for every w in [0, w]. Suppose not. Let w1 and w2 be two distinct income
levels such that Vx(w1 ) > Vx(w2 ) . A marginal decrease in x(w) in a small enough interval around
w2 accompanied by a marginal increase in x(w) in a small interval around w1 while respecting the
budget constraint improves expected plurality for P , contradicting that (x, y) is an equilibrium.
Hence, we can write the following:
(1)
[π(w)f (d(w) − α(w, l)) + (1 − π(w))f (d(w) − α(w, h))]u0 (w + x(w)) = λγ
for every w ∈ [0, w] and some λγ > 0.
Now consider
(2)
∂V
.
∂y(w)
Analogous arguments apply in this case and hence we can claim
[π(w)f (d(w) − α(w, l)) + (1 − π(w))f (d(w) − α(w, h))]u0 (w + y(w)) = µγ
for every w ∈ [0, w] and some µγ > 0.
0
Comparing equations (1) and (2) for any group w yields uu0(w+x(w))
= µλγγ which is a constant. This
(w+y(w))
implies that in any equilibrium x = y given the strict concavity of u and that both x and y are
10
balanced–budget redistributions. Suppose not. Assume that for some w1 , w.l.o.g. x(w1 ) > y(w1 ).
0
By uu0(w+x(w))
= µλγγ , this implies x(w) > y(w) for every w in violation of the budget constraint.
(w+y(w))
Thus, d(w) = 0 for every group w. Imputing this in equation (1) and using the symmetry of f
around 0, we get for every group w:
(3)
σ(γ, w)[u0 (w + x(w))] = λγ .
This guarantees that w + x(w) varies positively with σ(γ, w) given the strict concavity of u. The
same equation can be utilized to show the uniqueness of equilibrium. Suppose that both (x, λ) and
(x0 , λ0 ) satisfy (3). If λ = λ0 then x = x0 by the strict concavity of u. Alternatively if λ < λ0 then
x > x0 for the same reason. However, this implies that both x and x0 cannot be balanced-budget
redistributions. Hence it must be that (x, λ) = (x0 , λ0 ).
Recall the “class bias” t(w) felt by an individual in group w. We will assume the following:
t(w) ≥ s and t(w) < 0.
By the continuity of t, there will be some w ∈ (w, w), call it w∗ , such that t(w∗ ) = 0.20 One can
think of this group w∗ as the “middle income” group which feels equally class–wise affiliated to
either party.
Proposition 1 provides a clue as to which type of w–groups will be most favored in terms of
final per–capita consumption by both parties in equilibrium. It is the groups with high values of
σ(γ, w). Specifically, the group(s) where σ(γ, w) is maximized — the “swing” group(s) — is (are)
the biggest gainer(s). We denote these groups as members of the set Wγ and a generic member
of this set as wγ . Clearly, depending upon γ, this set could be a singleton.21 Moreover, by the
continuity of σ(γ, w) in w we know that the post–transfer per–capita consumption is going to be
relatively high in income groups “close” to any wγ ∈ Wγ ; hence, the focus on Wγ .
The explicit dependence of the set Wγ on γ is the subject of the ensuing section (section 2.2).
2.2. Redistribution policies under different candidate configurations. We now examine, one
by one, what the equilibrium redistribution policies look like for the different possible choices of
candidate configuration γ in the first stage.
We start with the configuration (l, h), i.e., party P fields an l–group candidate while party R
fields a h–group candidate. Here, we show that any group which is swing, i.e. belongs to Wγ ,
must necessarily lie to the left of the income group w∗ . To put it in another way, the focus of
redistributive tranfers is on groups earning lesser than the middle income group. The following
proposition makes this point more formally.
P ROPOSITION 2. Take a constituency with γ = (l, h) and consider the swing group set Wγ . Here
the swing group(s) is (are) poorer than the group w∗ , i.e., wγ < w∗ for every wγ ∈ Wγ .
20
If t is weakly decreasing then there could be an income interval where the value of t is 0. The distinction between a
unique w∗ versus an interval makes no significant difference to the results that follow. So for ease of exposition, we
will proceed as if w∗ is unique.
21
We know that Wγ is non-empty for any given γ since the function σ(γ, w) is continuous in w and is defined over a
compact set [0, w].
11
Proof. Note, γ = (l, h) implies that every l–group voter associates positively with P 0 s candidate
while every h–group voter associates positively with R0 s candidate. Hence, α(w, l) = t(w) + s
and α(w, h) = t(w) − s.
The derivative of σ(γ, w) w.r.t. w when evaluated at w∗ is the following:
f 0 (s)t0 (w∗ )[2π(w∗ ) − 1].
This term is negative since π(w∗ ) < 1/2 and both f 0 (s) and t0 (w∗ ) are negative. So, wγ 6= w∗ .
Suppose wγ > w∗ . Hence, t(wγ ) ≤ 0 and
(4)
σ(γ, wγ ) = π(wγ )f (t(wγ ) + s) + (1 − π(wγ ))f (t(wγ ) − s).
It must be that t(wγ ) + s ≥ 0. Suppose not. Consider ŵ such that t(ŵ) + s = 0. Clearly, ŵ < wγ
since t is decreasing in w.
So,
σ(γ, ŵ) = π(ŵ)f (0) + (1 − π(ŵ))f (t(ŵ) − s).
Also
(5)
σ(γ, ŵ) ≥ π(wγ )f (0) + (1 − π(wγ ))f (t(ŵ) − s) > σ(γ, wγ ).
where the first inequality comes from π(ŵ) ≥ π(wγ ). The second inequality follows from the
unimodality and symmetry of f around 0 and by observing that
|t(ŵ) − s| = 2s < |t(wγ ) − s|.
Hence, it must be that s + t(wγ ) ≥ 0.
Now, corresponding to wγ , one can always find a group w̃ ∈ [w, w∗ ) such that t(w̃) = −t(wγ ) > 0.
This is possible since t(w) ≥ s by assumption. For this group w̃,
(6)
σ(γ, w̃) = π(w̃)f (t(w̃) + s) + (1 − π(w̃))f (s − t(w̃))
Now compare equations (4) and (6). Since t(w̃) = −t(wγ ) and 1 − π(w̃) > 1/2 > π(wγ ), it must
be that σ(γ, w̃) > σ(γ, wγ ). This leads to a contradiction which establishes the proposition.
The next proposition considers the case in which both parties field candidates from the same ethnic
group; hence, γ ∈ {(l, l), (h, h)}. Note, this subsumes the case of a reserved constituency. Recall
that for such configurations, we have si (γ) = 0 for every voter i. Thus, the ethnic–affiliation
component of every voter’s bias loses relevance.
P ROPOSITION 3. In a constituency in which both parties field candidates from the same ethnic
group, group w∗ is the “swing” group in equilibrium. Hence, Wγ = {w∗ } for γ ∈ {(l, l), (h, h)}.
Proof. In a constituency where e(P ) = e(R), for any group w:
σ(γ, w) = π(w)f (t(w)) + (1 − π(w))f (t(w)) = f (t(w)).
Clearly, the above is maximized at w = w∗ given that f is unimodal and symmetric around 0 and
that t(w∗ ) = 0.
12
A quick eye–balling of the two preceding propositions above provides some inkling about which
types of groups get preferential treatment under the different candidate configurations. Clearly,
γ = (l, h) is more geared towards benefitting lower income groups in relation to configurations
involving candidates from the same ethnic group, i.e., γ ∈ {(l, l), (h, h)}.
This leaves us with the case where γ = (h, l), i.e., e(P ) = h and e(R) = l. We will return to this
case later. Now we move to on the details of the first stage game.
2.3. Candidate choice by parties. So far the discussion has focused on equilibrium redistribution
policies and identity of “swing” groups given a candidate configuration. Recall that the parties are
free to choose their candidates in unreserved constituencies in the first stage of the game.22
Now we will describe the payoffs to the parties in detail. Both parties care about their respective
performance in the election, specifically, their (respective) expected pluralities; in other words,
they are office–seeking in some degree. Each party also cares about the effect fielding a candidate
of either ethnic group has on what we call the “cohesiveness” of the party.
The basic idea is the following. First, there is the issue of compliance of the fielded candidate
with regards to the party leadership should the former actually win the election. It makes sense
to assume that an l–group candidate is more compliant with the leadership’s decisions than an h–
group one. Secondly, there is the issue of within–party cooperation across the two ethnic groups;
alternatively, one could think of this as the degree of tolerance for l–group members by their h–
group counterparts. So, “cohesiveness” of a party depends upon both of these factors: compliance
and tolerance.
We use a simple (reduced form) way of capturing these aspects. Let the payoff to party j, for
j = P, R, be given by
Wj (γ) + χcj (e(j))
where Wj (γ) is the expected plurality under configuration γ, cj (e(j)) is the cohesiveness for party
j from choosing candidate type e(j) and χ > 0 is the weight accorded to it. So, cj : {l, h} → R
for j = P, R.
Suppose that parties can only field their members as candidates and any citizen can potentially join
any political party. However, party R by its very nature attracts richer individuals, as compared to
party P .23 Hence, the relatively wealthier section of society will populate R.24 This means that the
proportion of l–group members in party R is strictly lower than in P (since l–type citizens are on
average poorer than the h–type ones). This suggests that tolerance is higher in party P than in R
when each of the parties consider fielding an l–type candidate.
The “cohesiveness” payoff from fielding a h–type candidate is normalized to 0 for both parties.
Why? Observe that there is a large pool of h–type members in both parties who are eligible for
candidacy. So it seems reasonable that this would affect cohesiveness in the same way in the two
22
In principle, there could be equilibria in which one or both political parties play mixed strategies, i.e., randomize
between fielding an l–group candidate and a h–group candidate. We restrict attention to only those equilibria in which
each political party fields a candidate rather than randomizing since this seems more natural in our setting.
23
Recall, R is pro–rich which can be interpreted as being “pro–business”, more “right–wing”, etc.
24
This joining of political parties by politicians is a coalition formation process which could have been written out in
a more sequenced manner. However, this reduced form approach pursued here is adequate for our purposes.
13
parties. Hence, cP (h) = cR (h) = 0. For ease of notation, we will denote cj (l) simply as cj for
j = P, R.
As noted earlier, it seems reasonable to assume that the compliance factor is positive for an l–type
candidate for either party as the l–group is a minority in either party and also from their minority
status in society. As per the tolerance part, we argued earlier than it is greater in party P .
Combining, we can write
cP > max{0, cR }.
This sets the ground for identifying the set of candidate configurations that can be observed in
equilibrium in any unreserved constituency. The following proposition is a step in that direction.
P ROPOSITION 4. In any unreserved constituency, the candidate configuration γ = (h, l) will not
be observed in equilibrium.
Proof. Suppose γ = (h, l) is the first–stage equilibrium choice in an unreserved constituency.
Since party R is fielding a low-caste candidate, it must be
WR (h, l) + χcR ≥ WR (h, h).
(7)
Now suppose that party P deviates to fielding an l–group candidate. Given that γ = (h, l) is part
of the equilibrium, such a deviation should not be profitable for party P . Hence,
WP (l, l) + χcP − WP (h, l) ≤ 0.
(8)
However,
(9) WP (l, l) + χcP − WP (h, l) = WP (l, l) + χcP + WR (h, l) = WP (h, h) + χcP + WR (h, l)
where the last equality follows from Proposition 3. However,
WP (h, h) + χcP + WR (h, l) ≥ WP (h, h) + WR (h, h) − χcR + χcP = χ(cP − cR ) > 0
where the first inequality follows from the relation in (7). Therefore,
WP (l, l) + χcP − WP (h, l) > 0.
This contradicts the relation in (8) and thus establishes the proposition.
Proposition 4 rules out the possibility of the configuration (h, l) in an unreserved constituency.
However one may ask — in the context of an unreserved constituency — if (h, h) is the only
possible candidate configuration in equilibrium. After all, if the candidate configuration (l, h) is
like–wise ruled out then political reservation would not affect the equilibrium redistribution policy
at all.
Next, we examine conditions under which configuration (l, h) can be part of the equilibrium of
this game.
Consider the configuration (h, h). In this configuration the ethnic bias component is irrelevant
for
R
every voter; hence, si (h, h) = 0 for every voter i. Here the payoff to P is WP (h, h) = pi since
i
14
the cohesion payoff from fielding a h–type candidate has been normalized to zero. Now,
Zw
Z
[π(w)[1 − F (−α(w, l))] + (1 − π(w))[1 − F (−α(w, h))]] dG(w)
pi =
i
0
Rw
which simplifies to [1 − F (−t(w))] dG(w) since α(w, h) = α(w, l) = t(w).
0
Note that under the configuration (l, h), we have α(w, l) = t(w) + s and α(w, h) = t(w) − s.
Hence,
Zw
WP (l, h) = [π(w)[1 − F (−t(w) − s)] + (1 − π(w))[1 − F (−t(w) + s)]] dG(w)
0
Now for P given that R is fielding h, fielding l will be (weakly) preferred to h if and only if
(10)
WP (l, h) + χcP ≥ WP (h, h).
Note that
1 − F (−t(w) − s) > 1 − F (−t(w)) > 1 − F (−t(w) + s)
for every w ∈ [0, w]. This introduces some ambiguity in ranking WP (l, h) and WP (h, h). Intuitively, in switching from (h, h) to (l, h), party P potentially gains the support of some l–type
voters while it potentially loses some h–type supporters.
To be specific, the extent of loss and gain of votes depends on the distribution of the biases among
the voters (F (.)), the proportion of l–type voters across income groups (π(.)), the function t(.),
the size of ethnic bias s as well as the income distribution (G(.)). Moreover, the term χcP (the
additional cohesiveness payoff to party P from fielding an l–type candidate as opposed to an h–
type candidate) is positive. Hence, the ambiguity in comparing the two payoffs for party P .25
However, one can make the following claim in this matter.
P ROPOSITION 5. Suppose in a unreserved constituency, the candidate configuration γ = (l, h)
is observed in equilibrium. Now take this constituency and increase the proportion of the l–types
(π(w)) in at least one income group w while making no other changes. Then γ = (l, h) continues
to be observed in equilibrium.
Proof. Increasing π(w) for at least one income group while making no other changes implies an
increase in WP (l, h) since F (−t(w) − s) < F (−t(w) + s). Note, WP (h, h) is unaffected. Clearly,
if (10) was satisfied earlier, it continues to be so after the change in the size of the l–group.
Thus, Proposition 5 suggests that the configuration γ = (l, h) is more likely to be observed in
constituencies where the l–group whilst a minority is relatively sizeable.
In sum, we note that in an unreserved constituency the only possible configurations in equilibrium
are (h, h), (l, l) and (l, h). The first two configurations yield identical redistributions in equilibrium
(see Proposition 3) while the last configuration produces a redistribution policy which favors poorer
25
This ambiguity is in sharp contrast with the case of γ = (h, l) which has been dealt with earlier.
15
groups (by Proposition 2). In the aggregate, reservation appears to bias policy against poorer
groups.
In the following section, the baseline model is amended so as to (additionally) justify labeling
parties as “pro–poor” and “pro–rich”. As discussed in detail below, the main intuition is robust to
such an extension.
2.4. An Extension of the Model. The baseline model presents two political parties P and R
towards whom voters feel affiliated on the basis of their incomes; in particular, poor voters tend
to favor P while rich voters tend to favor R. However, in equilibrium both parties propose the
same redistribution policy. This might raise the question as to why party affiliations take such a
form when both parties propose identical redistribution policies in equilibrium. However, there is
an important distinction in the actions of the two parties which justifies — at least, to some extent
— the “pro–poor” and “pro–rich” labels. Recall, P fields an l–type candidate (in some cases) in
response to R0 s fielding a h–type candidate and this improves the condition of poorer groups by
shifting the focus of targeting towards them. In this way, P does favor poorer groups through its
actions; specifically, through candidate choice.
However, this still leaves us open to the criticism that the fielded candidates in a constituency
behave identically regardless of their party affiliations. Here, we explicitly allow candidates to instrinsically care about their proposed redistribution policies alongside expected vote shares. In particular, one can model P ’s candidate to care more (less) about the consumption of poorer (richer)
groups relative to R’s candidate. In this manner, the parties can be more easily categorized as
“pro–poor” and “pro–rich”. What is interesting is that it is possible to introduce this aspect without altering the main findings in any significant way, although there is no longer policy convergence
in equilibrium.
In particular, now suppose party P 0 s candidate maximizes the following:
Zw
Z
pi +
i
Note,
R
ρ(w)d(w)dG(w).
0
pi is P 0 s expected vote share and d(w) is the utility differential u(w+y(w))−u(w+x(w)),
i
just as before. Here, ρ(w) is a real-valued, continuous function with ρ0 > 0 and ρ(w∗ ) = 0. Note,
w∗ represents the intermediate income group where t(w) is 0, as before.
Party R0 s candidate continues to maximize — as before — the following:
Z
1 − pi .
i
By construction, P ’s candidate now has an incentive to have d(w) (the utility shortfall from P 0 s
candidates’ perspective) as negative for groups with w < w∗ and d(w) > 0 for groups with
w > w∗ . So P ’s candidate has an incentive to “outbid” R’s candidate for groups with income
lower than w∗ while for richer groups, i.e., for w > w∗ , P ’s candidate has an incentive to “bid”
below R’s candidate.
16
2.4.1. Description of the second–stage equilibrium. Suppose (x, y) is an equilibrium of this (second–
stage) redistribution choice game, conditional on candidate configuration choice γ.
Like in Proposition 1 of the baseline model, the following relations must hold. The marginal gain
in payoff to party P from a marginal increase in transfer to any group w must be equalized across
all (income) groups. Therefore,
[ψ(γ, w) − ρ(w)]u0 (x(w) + w) = λ
The same consideration applies to party R. Hence,
ψ(γ, w)u0 (y(w) + w) = µ
where
ψ(γ, w) ≡ π(w)f (d(w) − α(w, l)) + (1 − π(w))f (d(w) − α(w, h)).
Note, ψ(γ, w) is positive for all (γ, w). However, ρ(w) is non-negative only for w ≥ w∗ . This
implies we need some restriction on ρ(.) so that marginal gain in payoff to party P from a marginal
increase in transfer to any group w is always positive. Intuitively, if ρ(.) is “sufficiently” bounded
then both λ and µ will be positive. It is easy enough to outline such a sufficient condition.
Note, |α(w, e)| ≤ s + max{t(0), −t(w)} for e = l, h. Also, d(w) is continuous in w and given
that both x and y are budget–balanced schemes ensures that there is some upper bound — call it d
— such that |d(w)| ≤ d for every w ∈ [0, w].
Hence, assuming
f (d + s + max{t(0), −t(w)}) > max{ρ(w), −ρ(0)}
is sufficient to guarantee that λ, µ > 0. In the ensuing discussion, it is assumed that this condition
is met.
Consider the following relation obtained from the above two equations:
[ψ(γ, w) − ρ(w)]u0 (x(w) + w)
λ
=
.
ψ(γ, w)u0 (y(w) + w)
µ
Now we examine each of the following possibilities. (i)
λ
µ
> 1, (ii)
λ
µ
< 1 and (iii)
λ
µ
= 1.
If we are in case (i), then it must be x(w∗ ) < y(w∗ ) since ρ(w∗ ) = 0 and u00 < 0. Hence,
d(w∗ ) > 0. Also, for every w > w∗ , it must be that d(w) > 0. To ensure that both x and y are
budget–balanced policies, it must be that d(w) < 0 for some interval of incomes for w < w∗ .
Hence, we have d(w) > 0 for “high” incomes and d(w) < 0 for some “low” income groups.
Case (ii) is analogous with d(w) < 0 for every w < w∗ and d(w) > 0 for some interval of incomes
for w > w∗ .
For case (iii), we have d(w) < 0 for every w < w∗ , d(w) > 0 for every w > w∗ and d(w∗ ) = 0.
In principle, any of the three possibilties can arise in equilibrium depending upon the details of
ρ(.). However, case (iii) is the most compelling candidate for the following reason. One can
always construct an equilibrium corresponding to case (iii); such is not true for cases (i) and (ii).
For certain functional forms for ρ(.), one could construct revealed–preference type arguments to
17
show that cases (ii) and (iii) are not possible in equilibrium. For this reason, in the following
analysis we only focus on those equilibria (x, y) which belong to case (iii).
Hence, the equilibrium policy profile (x, y) for any
lowing property:
x(w) > y(w) if
x(w) = y(w) if
x(w) < y(w) if
given candidate configuration γ has the folw < w∗
w = w∗
w > w∗
Thus, there is no longer policy convergence in equilibrium. Moreover, P 0 s policy is more tilted
towards poorer groups while R0 s policy is more tilted towards richer groups thus reflecting their
party identities. Next we study the equilibrium redistribution policy profile under different candidate configurations.
2.4.2. Redistribution under different candidate configurations. First consider the configuration
(l, h), i.e., party P fields an l–group candidate in response to party R fielding a h–group candidate.
Like in the baseline model, we can show that the focus of targeting — under (l, h) — is some group
poorer than the middle-income group w∗ . Recall, in this setup parties propose different policies
in equilibrium. Let party P 0 s most favored group — the group with the highest x(w) + w — be
denoted by wP ; let wR be defined similarly. In principle, now wP and wR could refer to distinct
income groups depending upon the configuration γ, and thus complicating matters relative to the
baseline model. Hence, to simplify the arguments we make an assumption concerning the share of
the l–group voters in group w∗ , i.e., π(w∗ ).
The assumption — call it C1 — is the following:
f (s) > π(w∗ )f (s − ) + (1 − π(w∗ ))f (s + )
for any > 0.
As long as the minority group is “sufficiently” poor in the sense that π(w∗ ) << 1/2, the above
condition will be easily met. The following proposition describes the relationship between wP , wR
and the middle income group w∗ .
P ROPOSITION 2*. Suppose condition C1 is satisfied. Consider an unreserved constituency with
γ = (l, h). Then, party P 0 s most favored group, namely wP , is poorer than the group w∗ , while
party R0 s most favored group wR is no richer than w∗ . Hence, wP < w∗ and wR ≤ w∗ .
Proof. (See Appendix.)
So under the configuration (l, h), the most favored group in equilibrium will either be wR ≤ w∗
(when party R wins) or some group wP < w∗ (when party P wins). Hence, in expected terms, the
most favored group is one which is poorer than the group w∗ .
Next, we move to candidate configurations where both parties field candidates from the same ethnic
group. Before proceeding, we make one more assumption concerning ρ(.). Let
[f (0) − f (|t(w)|) ≥ |ρ(w)|]
for all w ∈ [0, w].
This assumption — call it C2 — essentially imposes bounds on ρ(.).
18
Now, consider either of the two configurations: (h, h) or (l, l). Here, like in the baseline model,
we can show that the intermediate income group w∗ is most favored by both parties P and R. The
result is stated formally in the following proposition.
P ROPOSITION 3*. Under condition C2, in a constituency in which both parties field candidates
from the same caste, the group w∗ becomes the most favored group for both parties in equilibrium.
Therefore, wP = wR = w∗ in such a constituency.
Proof. (See Appendix.)
In this setup we allow that in every possible configuration γ, P 0 s candidate has an intrinsic bias
in favor of poorer groups relative to R0 s candidate. Also, this is something known by all players
— voters and party–members alike. However, for reasons of tractability, we continue to assume
that the first–stage considerations (specifically, the party payoff functions) are the same as those in
the baseline model. Hence for guaranteeing that the only possible equilibrium configurations in an
unreserved constituency are (h, h), (l, l) and (l, h), we can use the exact same arguments as in the
baseline model.
2.5. Some Remarks on the Ethnic Bias. There are two aspects relating to the ethnic bias component which we discuss here. The first relates to the question of allowing some degree of heterogeniety in the individual–level ethnic bias component. The second is a question of how redistribution
would respond — in equilibrium — to a shift in the size of ethnic bias, i.e., a change in s.
2.5.1. Heterogeneity of the bias. In the baseline model, the size of the ethnic bias for every individual was either 0 or a given s > 0. It possible to allow some extent of heterogeniety in this
respect without substantially affecting the results in any way. Suppose, instead of the baseline
model structure of ethnic bias, we have the following. For any individual i:
si (γ) = ψi
if e(P ) 6= e(R) and ei = e(P )
si (γ) = −ψi if e(P ) 6= e(R) and ei = e(R)
si (γ) = 0
if e(P ) = e(R).
where
ψi = s + ηi
and
−ψi = −s + ηi .
Suppose η is just mean–zero noise and each i draws ηi from some given distribution. As long as
η–distribution is independent of the income distribution (i.e., ηi and wi are independent) the game
is basically unchanged. We can rewrite αi as follows:
αi = t(wi ) + s + υi
where υi = ηi + i and let υi be distributed according to a symmetric, unimodal density f with
support on R (like i was in the baseline model). This makes it clear that the basic structure of the
game is the same as before; hence, all our results obtain. Of course, if the distribution of η and
income distribution were related in some way, then there could be important differences. We do
19
not pursue this avenue in this paper as it is not obvious as to what form such a correlation between
ethnic bias and income should take.26
2.5.2. Changes in the size of the bias. An interesting question one could ask is the following: what
happens to the nature of redistribution in equilibrium as society becomes more ethnically biased?
In the context of our baseline model, this tantamounts to a comparative statics exercise with respect
to s, the parameter which represents the size of the ethnic bias.
Now, it is straightforward to see that if we set s = 0, then the swing group would be the middle
income group w∗ regardless of the candidate configuration chosen in the first stage. In general for
s > 0, the set comprising the swing group(s), namely, Wγ need not be a singleton set for various
values of γ. One could, of course, choose specific forms for the different functions in the model to
guarantee a singleton Wγ for every feasible γ. One would require assumptions to guarantee that wγ
(which is defined by Wγ = {wγ }) is differentiable in s. Even at the sacrifice of some generality,
it is still interesting to understand how the swing group responds to shifts in s for the possible
candidate configurations.
For γ = (h, h), (l, l) the swing group is always w∗ for any s > 0. So what we need to check if the
relation between wγ and s for γ = (l, h). Basic intuition would suggest that increasing s would
make wγ (for γ = (l, h)) go down. Why? Roughly, the swing group in (l, h) is one where the
h–group has “neglible” bias whereas the l–group has a positive bias in favor of P . The “neglible
bias” for the h–group is because t(wγ ) and s effectively cancel out. Hence, increasing s suggests
that wγ must go down for γ = (l, h). The following example is in line with our basic inituition.
A N E XAMPLE . Suppose that the ideological bias follows a logistic distribution, i.e., F (x) =
1
and the downward–sloping function t(w) is simply w∗ − w. Note, this immediately implies
1+e−x
t(w∗ ) = 0. Also, let the share of the l–type members be the same across all income levels, so
π(w) = 1/2 − θ for every w ∈ [0, w] for some parameter θ ∈ (0, 1/2). It is easily checked that
these specific functional forms satisfy all the assumptions made in the baseline model.
Note, f (x) = f (−x) =
f (t(w) − s) =
e−x
.
(1+e−x )2
∗
ew−w +s
∗ +s 2
w−w
(1+e
)
∗
Hence for γ = (l, h), we have f (t(w) + s) =
ew−w −s
(1+ew−w∗ −s )2
and
.
Therefore, the function σ(γ, w) for γ = (l, h) is given by
∗
∗
1
h
i 1
h
i
ew−w −s
ew−w +s
−θ .
+
+θ .
.
2
(1 + ew−w∗ −s )2
2
(1 + ew−w∗ +s )2
The idea is to find the maxima of this function w.r.t w so as to identify the swing group. Differentiating this function w.r.t. w and setting ∂σ(γ, w)/∂w = 0 yields:
∗
∗
1
h
i h 1 − ew−w∗ −s i 1
h
i h ew−w∗ +s − 1 i
ew−w −s
ew−w +s
−θ .
.
=
+θ .
.
.
2
(1 + ew−w∗ −s )2 1 + ew−w∗ −s
2
(1 + ew−w∗ +s )2 1 + ew−w∗ +s
26
It is possible to explore a few options. For example, one could assume a negative correlation between ηi and wi
suggesting that poorer individuals have stronger ethnic biases. Alternatively, the relationship could be non–monotonic,
perhaps U–shaped with the middle income groups having the least degree of ethnic biases. We avoid pursuing these
possibilities as we see them as being somewhat tangential to the main issues in this paper.
20
Define
z ≡ w − w∗ .
Re-arranging terms and re-writing the above relation in terms of z yields
(11)
es−z (es−z − 1) (es+z + 1)3
.
.
=
es+z (es+z − 1) (es−z + 1)3
1
2
1
2
+θ
.
−θ
Noting that the RHS of equation (11) exceeds 1, we infer that z = 0 cannot be a solution. In fact,
applying Proposition 2 gives z < 0 for identifying the maxima. Also, s + z ≤ 0 does not satisfy
(11). Hence, all maxima must have z ∈ (−s, 0).
Define the function I(s, z) as follows:
I(s, z) ≡
es−z (es−z − 1) (es+z + 1)3
.
.
−
es+z (es+z − 1) (es−z + 1)3
1
2
1
2
+θ
= 0.
−θ
Using the Implicit function theorem, we get
∂z
Is
(cosh(z) − 2cosh(s))sinh(s)−1 sinh(z)
=− =
.
∂s
Iz
cosh(s) − 2cosh(z)
For the range s > 0 and z < 0 (the region relevant for the maxima), we have
turn, implies wγ (for γ = (l, h) must go down as s goes up.
∂z
∂s
< 0.27 This, in
However, this need not be true in general; non–monotonicity is a possibility. Moreover, even within
the context of this example, there is another (and perhaps more subtle) issue: while it is true that the
equilibrium redistribution policy in a constituency with configuration (l, h) will be more favorable
towards poorer groups in a more (ethnically) biased society, the probability of (l, h) being a first
stage equilibrium choice is affected by the size of s. In particular, the expected–plurality payoff
to P from fielding l against R’s h–group candidate (denoted by WP (l, h)) depends upon s. It may
well be the case that WP (l, h) falls in s and in that case can go below WP (h, h) − χcP for some
adequately high value of s.
In sum, the overall effect on greater ethnic bias is ambiguous even in the restricted environment
of our example.28 This arises from the fact that while the redistribution certainly changes in a
definite direction for the configuration (l, h), the possibility of this very configuration arising in
equilibrium may well go down.
2.6. Applying the model. The theory outlined above can be useful for understanding certain ramifications of various quotas in the political domain. In the context of India, mandated political
representation has been in place for various minority groups — like the Scheduled Castes (SCs)
and Tribes (STs) — since the 1950s. In fact, since the early 1990s this form of quotas has been
extended to women given their virtual absence in the political arena. Here we attempt to identify
the different scenarios under which our theory can be useful and also those where it is inapplicable.
One clear application is to the case of reservation for SCs in the Indian federal legislature, i.e., the
Parliament. The SCs are an economically–disadvantaged minority at the national level; even in
districts where they are more concentrated they happen to be never much above 50% of the district
27
The details of the calculation have been omitted in the interest of brevity.
We need to make assumptions on the income distribution G(.) to get a clear answer.
28
21
population. Moreover, since the early 1980s they have been politically active in part, due to the
impetus from pre–existing quotas. In fact, there are political parties dominated by certain caste
groups.
There is another relevant point: the way districts are chosen for SC reservation creates a bias
towards selecting those districts where SCs are relatively more numerous.29 Proposition 5 suggests
that it is precisely such constituencies where (l, h) is more likely to be observed prior to reservation.
All of this taken together make SC–reservation at the federal level an apt candidate for testing our
theory.30
What about the other closely related group — the Scheduled Tribes (STs)? Note, from a socio–
political angle STs are quite distinct from SCs on at least two counts: (i) STs are not as well
politically organized as SCs even today, so political reservation has had limited impact on their
political standing, unlike the SCs who are a political force to reckon with today. (ii) STs tend to
live in rural areas separate from other groups while SCs are spread more evenly in the geographical
sense. Hence, it would be unsatisfactory to combine SCs and STs and treat them as the l–group.
It makes little sense to treat the STs alone as the l–group as the notion of “caste identity” is much
more pervasive and salient than any notion of “tribe identity”. So the within–group bias s makes
logical sense for a caste–based division but not a tribe–based one.
The theory can be useful for studying gender quotas. If we recognize that women are a (population)
minority — perhaps only slightly — in many developing countries and that in many similar–
ranked occupations women typically are under–paid in comparison to their male colleagues, then
the model has the potential to address gender quotas at the political level. However, it is difficult to
develop — in an empirical exercise — a notion of within–women economic inequality especially
given that most households (in commonly used datasets) have both men and women co–habiting.
3. E MPIRICAL A NALYSIS
The theory above yields several testable predictions. For reasons outlined in section 2.6 above, we
restrict attention to the following ones.
[A] Reservation of seats for the Scheduled Castes disfavors the poorer sections of society as transfers to them are reduced.
[B] Given that the Scheduled Castes are poorer on average, political reservation hurts most SC
members and favors some relatively wealthier SCs. Thus, reservation widens income disparities
within the SCs.
In India, every Member of Parliament (MP) obtains funds from the central government for expenditure on local-area development projects in their respective constituencies.31 Ideally one would
29
We discuss this in detail in Section 3.1
It would be interesting to test the effect of changes in ethnic bias for any minority group (say, the SCs) if one could
somehow argue that these biases have reduced over time. Finding a measure of such bias in the data is a real challenge
and hence we avoid such an exercise.
31
Each MP has the choice to suggest to the District Collector (head of the district administration) for, works to the tune
of 20 million Indian Rupees (approximately equal to 0.45 million US $) per year to be taken up in his constituency.
30
22
look at the decomposition of the entire budgetary spending by every MP and note how the pattern
changes as one moves from a reserved constituency to one without reservation.32
However the data on MP spending which is publicly available, does not allow such a fine decomposition. The annual reports on MP spending inform us about the total number of “works” (i.e.
local-area small-scale developmental projects) the MP has commissioned during the past year;
there is no information on which types of projects were implemented. So it is not possible to
discern which groups are benefiting at the cost of others. Further, these reports decompose the
aggregate MP spending by sector only at the level of the states — no finer. So there is information
on how much is spent on each sector by state.33 Hence, this information is not particularly useful
in answering our questions.
Another possibility is to map the levels of different public goods (available from the census) to
constituencies and then check if reservation affects the pattern of public goods provision. In fact,
Banerjee and Somanathan (2007) perform an exercise of this nature although their main question
is different from the one examined here.34 However, this method of mapping public goods to
constituencies poses a problem. It is often not possible to unambiguously demarcate most public
goods as specifically being preferred by any particular income group. In some cases, the same
public good could be either “pro-poor” or “pro-rich” depending on its precise location.35 Therefore
looking at public goods does not cleanly identify losers and gainers across different income groups.
So what is required is hard information on some aspect of MP spending which the poor clearly
care about and then check if such kind of spending varies with SC reservation.
Given that India has implemented several employment based poverty reduction programs (collectively known as Public Works), differences in the implementation of these programs across reserved
versus unreserved districts provide useful information. One implication of the model could be the
lower implementation of these programs — something the poor clearly care about — in reserved
parliamentary constituencies. This is how prediction [A] is checked.
It is important to note that the empirical exercise (described below) has been conducted with data
from the time period before Panchayati Raj (local government bodies) came into de facto existence.
The role of the Panchayats in monitoring and disbursing central/state funds at the village level took
shape after 1993 (when the 73rd Amendement Act came into force). Hence, prior to 1993 the MPs
can be largely held accountable for the quality and implementation of such “works”.
For prediction [B], measures of inequality of per-capita expenditure of SC households are constructed at the district-level and matched with the reservation status of the district.
32
Data on the reservation status of all parliamentary constituencies of India is found from the reports of the Election
Commission of India.
33
The different sectors on which MPs typically allocate their funds (i.e. they commission “works” in these sectors) are
“Roadways, Paths and Bridges”, “Electricity facilities”, “Irrigation”, “Education”, “Health and Public Welfare” and
“Sanitation and Public Health”.
34
They look at the location of public goods between 1971 and 1991 in about 500 parliamentary constituencies in rural
India to assess the relative importance of social structures as regards the allocation of public resources over the period.
35
For e.g., if all the road-construction happens in such areas so that large wealthy farmers can access the market more
easily, then more roads do not necessarily benefit poor farmers. On the other hand, it could be that roads are developed
in a way that the transportation costs for all are reduced and perhaps the poor do gain more from this.
23
3.1. Empirical strategy. To identify the effect of reservation on redistributive outcomes, one has
to understand the process by which electoral districts are selected for reservation. The choice of
the constituencies (within a state) that are reserved for SCs is guided by the following considerations. The Delimitation Commission stipulates that constituencies for SCs are to be distributed in
different parts of the state and seats are to be reserved for SCs in those constituencies where the
percentage of their population to the total population is comparatively large.36 Also, there was a
freeze on rotation since 1977, fixing the identity of reserved and unreserved constituencies all the
way till 2009 (details contained in the Appendix).
One can check if the Delimitation Commission had indeed adhered to the afore-mentioned guidelines. Data from the 1971 census is used to verify if there are systematic differences — aside
from the differences in SC population proportion — between those districts subsequently selected
for reservation and those which were not. Table 1 contains some relevant summary statistics of
districts based on the 1971 cenus data. Panel A contains data on 296 districts out of which 68 contained at least one reserved constituency.37 Panel B contains data on 157 districts (out of this 296)
where each of these districts correspond to a single constituency and vice-versa. In both panels, SC
population proportion is significantly larger in reserved districts. Other differences — apart from
Electricity which measures any power facilities available in the district — do not seem to persist
when one looks across both panels. Thus, overall the evidence is re-assuring. Conditional on SC
population proportion, the selection process appears to have been fairly random.
A data–limitation issue which needs to be confronted is the following. There are 542 constituencies spread over 338 districts. Although the household survey used here (from the National Sample
Survey (NSSO)) is fairly extensive in terms of household characteristics, it does not permit identification of the household all the way down to the constituency; district is as far as one can go.
So, a dummy variable for reservation at the district level is used here. This variable takes the value
one when any constituency within the district is reserved; it takes the value of 0 otherwise. The
household-level analysis regarding the question of participation in poverty-reduction programs is
carried out using household data from all 338 districts. Out of these 338 districts, there are 182
cases where every district corresponds to a single constituency and vice-versa. The household
analysis is repeated with this subset as robustness check.
We first look at the question concerning the reduced implementation of poverty-reduction programs
in reserved constituencies. As briefly mentioned before, India has implemented a large number of
poverty alleviation programs which are employment-based. “Public Works” is about employing
poor households for building public infrastructure especially in the rural parts of the country.38
They were designed to address the issue of reduction of poverty via employment of poor household
members. At the same time, these programs aid in the creation of basic infrastructure like roads,
dams, bridges, canals, schools, etc. In fact, the wages for these employment programs are set at a
suitable level which are expected to attract only the poor.
36
The Delimitation Commission is a national-level commission whose orders have the force of law, and cannot be
questioned in court.
37
Districts are typically larger than constituencies. More on this later.
38
The National Rural Employment Programme, Rural Landless Employment Guarantee Programme (RLEGP), Minimum Needs Programme (MNP), and other schemes aim to provide employment in construction of roads, dams, bunds,
digging of ponds, etc. All of these come under this “Public Works” approach.
24
TABLE 1. D ISTRICT- LEVEL SUMMARY
served with non-reserved districts.
STATISTICS FROM
[A]
[B]
Non-R
R
Mean
Mean
13.974
(7.608)
Population density (log) 6.762
(0.827)
Population (log)
13.956
(0.596)
Gini (landholdings)
0.657
(0.114)
Electricity
287.638
(312.651)
Educational facilities
1,003.894
(580.541)
Medical facilities
112.465
(73.074)
Rainfall
1,197.446
(685.577)
Turnout 1971
0.581
(0.110)
23.012
(10.674)
6.761
(0.665)
14.202
(0.591)
0.674
(0.089)
450.908
(458.984)
1,144.085
(782.837)
118.268
(81.120)
1,084.219
(524.248)
0.599
(0.096)
Observations
68
Variables
SC population (%)
228
1971: Comparing re-
Non-R
R
Diff.(NR–R)
Mean
Mean
Diff.(NR–R)
-9.039
(1.161)***
0.001
(0.110)
-0.246
(0.082)***
-0.017
(0.015)
-163.270
(48.554)***
-140.191
(87.375)*
-5.804
(10.361)
113.227
(90.135)
-0.018
(0.015)
14.751
(7.828)
6.634
(0.752)
13.798
(0.533)
0.646
(0.117)
237.140
(269.523)
902.229
(536.748)
102.246
(70.226)
1,114.911
(599.377)
0.566
(0.109)
24.475
(8.180)
6.395
(0.671)
13.649
(0.637)
0.672
(0.088)
350.410
(523.703)
855.864
(663.022)
85.524
(71.706)
1,036.695
(618.923)
0.594
(0.108)
-9.724
(1.846)***
0.239
(0.174)
0.149
(0.128)
-0.026
(0.027)
-113.270
(73.644)*
46.365
(130.046)
16.723
(16.510)
78.216
(141.130)
-0.028
(0.026)
136
21
NOTES: Panel A pertains to data from 296 districts (out of the total 338) and panel B pertains to data from
157 districts (out of 182) where each district has just one constituency. All variables have been constructed
using data from the 1971 census. Standard errors in parentheses for the columns indicating differences;
standard deviation in parentheses for all other columns. T-test used for comparing differences. *significant
at 10% **significant at 5% ***significant at 1%
The 43rd round of NSS used here contains information regarding the participation of the households in these employment-based poverty-reduction programs. To be precise, the following question was put to every (head of) household:
“Did any member of the household work for at least 60 days on public works during the last 365
days?”
It is possible to identify the district a household belongs to, in the particular NSS round used here
(i.e. the 43rd round). Given that the reservation data are available from the Election Commission
of India Reports, the following question is posed: “Does the reservation status of a district to which
a household belongs, have any impact on the household’s participation in Public Works?”. In other
words, is a household more or less likely to participate in these programs if it comes from a district
which is reserved for SCs? Recall, the household survey is from 1987-88 while the freeze on
reservation has been in place since 1977. So the reservation status of the constituency has been
unchanged for at least a decade.
The paper then turns to the prediction regarding the impact of reservation on within-SC inequality.
Here, it is no longer possible to continue with all 338 districts (containing 542 constituencies within
25
them) as SC-inequality within a constituency inside a district may very well be different from SCinequality within the district overall. Hence, here the sample is restricted to those 182 cases where
constituencies and districts are synonymous.39
3.2. Empirical specification.
3.2.1. Household-level specification. First comes the issue of how the reservation status of a
household’s constituency may affect the household’s participation in Public Works projects. To
capture the relationship, a logit model of the usual form is employed.40
P (yh = 1| Xh ) = G(βo + β1 Reservedh + β2 xh2 + ... + βK xhK )
where G(z) ≡ exp(z)/[1 + exp(z)].
So yh = 1 if a household h has participated in Public Works for at least 60 days during the last 365
days and yh = 0 otherwise. Reservedh = 1 if h belongs to a district which has at least one reserved
constituency, and is 0 otherwise. Suitable survey weights have been employed for all the reported
regressions keeping the NSSO survey design in mind. Although the unit of analysis is a household,
the explanatory variable of interest, i.e. reservation status of the household’s constituency, varies
at the level of the constituency and not at the level of the household.
Several controls are available from the NSSO dataset and they have been suitably employed. Table 6 in the Appendix contains a brief summary of the relevant household-level variables. Any
household characteristic which influences the household’s choice of involvement in Public Works
and is possibly correlated with the district’s reservation status, needs to be controlled for. These
are roughly “demand side” considerations which come into play when examining the outcome of
a household’s participation. Some of the important controls are (the log of) household per-capita
monthly expenditure, household size, a dummy indicating whether the household is SC or not, a
dummy indicating whether the household is Hindu or not, a dummy for the household’s education
(whether any household member has completed primary school or not) and a dummy indicating
whether the household resides in an urban area or not.
Other controls which have been used can be broadly grouped into two types — (i) proxies for
the (dwelling-related) wealth of the household and (ii) variables for the primary occupation of the
household.41 In all regression tables the dwelling-related wealth proxies are collectively refered to
as “Home dummies” while the occupation-related dummies are called “Employ dummies”.
Potentially, there are some “supply side” considerations too — it might happen that in some districts poverty-reduction programs are far less implemented than in others. Hence, a typical household may be excluded from participation because there are very few Public Works in the district.
39
Table 8 in the Appendix compares both sets — ‘single-constituency’ districts (182 in number) and ‘non-single
constituency’ districts (156 in number) — in terms of relevant district characteristics. There are two main differences.
First, in the complete sample (338 districts with 542 constituencies), the proportion of reserved constituencies is
higher –79 reserved out of 542 – whereas in the restricted sample (182 constituencies) only 23 are reserved; see the
difference of means for Reserved dummy. Secondly, the mean district population is higher in the complete sample;
see the difference of means for Population. This makes sense because every constituency is supposed to represent a
similar mass of people and so more populous districts were split up into two or more constituencies.
40
The same analyses have been conducted using a probit model. The results are extremely similar.
41
There is clearly a fairly high degree of correlation between a household’s educational attainments, wealth and
primary occupation.
26
Clearly that is some property of the district per se which is not directly related to the participation
decision of the household. The key question is whether or not the reservation status of a district
has such a “supply side” impact — and if so, then in which direction.
3.2.2. District-level specification. In order to address the relationship between the reservation status of a district and the inequality among SCs therein, a standard linear model of the following
form is employed:
yd = α + βReservedd + ρXd + εd
where yd is some measure of inequality — in terms of monthly per capita expenditure — among
SC households in district d, Xd contains district-level controls and εd is the district level error term.
The chief explanatory variable is the “Reserved” dummy as before. Here, we restrict the sample to
only those districts which have only one constituency within them. So we drop all districts which
have more than one constituency within them reducing the sample size to 182.42 The Interquartile
range (normalized by the mean), the ratio of upper to lower quartiles, and the Gini coefficient have
been used to measure inequality.
Any factor which potentially affects the reservation status (reserved or not reserved) and has a
direct effect on the outcome of interest (SC inequality) must be controlled for. Given the procedure
of SC reservation, the choice of a constituency within a state for reservation is partly random and
partly determined by the SC population proportion in the constituency. The randomness comes
from the fact that one could pick several different sets of seats for SC reservation keeping in mind
the consideration about geographical dispersion of these seats.
The different controls used for the various regressions are the (log of) population % of SCs which is
the (log of) percentage of SC households in the district in 1987-88, the (log of) district population
which is the (log of) total households in the district over the total households in all districts in 198788 expressed as a percentage, the log of average per-capita monthly expenditure of all households
in the district to control for the general level of prosperity of the district in 1987-88. State dummies
are used in certain specifications in order to capture any possible state fixed effects.
Other important controls include certain district level characteristics from 1971 Census of India.
Recall, the Delimitation Commission which selected the set of reserved districts in 1972 did so
on the basis of the 1971 census. Additionally, the freeze on rotation in 1977 clearly implies that
the reservation status of districts in 1987-88 were determined solely by that Delimitation Commission’s orders.43
3.3. Regression Results.
3.3.1. Household-level regressions. Table 2 contains the basic results. The first four columns
report logit regressions, column 5 presents a probit regression while OLS is reported for column 6.
The dependent variable is a dummy, Public Works, which takes the value 1 if the household reports
participation in the program, and takes the value 0 otherwise. In all six columns, the coefficient
42
Inclusion of districts with multiple constituencies is problemmatic in this case since SC-inequality rising (falling)
within a constituency does not necessarily imply that the district level SC-inequality is rising (falling); and vice-versa.
43
Table 7 in the Appendix contains a brief summary of the single-constituency districts which are used for the districtlevel regressions.
27
on Reserved dummy is highly significant and negative suggesting that a household from a reserved
district is less likely to have participated in Public Works projects.
TABLE 2. H OUSEHOLD - LEVEL
ticipation in ‘Public Works’.
Reserved dum.
Pce
SC dum.
Urban dum.
Hindu dum.
HH size
Educ.
Home dum.
Employ dum.
Pseudo/Adj. R2
Observations
REGRESSIONS :
Reservation and household par-
[1]
[2]
[3]
[4]
[5]
[6]
Logit
Logit
Logit
Logit
Probit
OLS
-0.386*** -0.386*** -0.386*** -0.376*** -0.166*** -0.016***
(0.120)
(0.120)
(0.120)
(0.121)
(0.054)
(0.005)
-0.295*** -0.256*** -0.186**
-0.048
-0.020
-0.002
(0.078)
(0.080)
(0.080)
(0.074)
(0.033)
(0.003)
0.062
0.046
0.014
-0.130*
-0.056
-0.007*
(0.070)
(0.070)
(0.071)
(0.076)
(0.035)
(0.004)
-0.959*** -0.919*** -0.855*** -0.557*** -0.228*** -0.011**
(0.104)
(0.102)
(0.102)
(0.212)
(0.088)
(0.005)
0.361*** 0.367*** 0.373*** 0.392*** 0.176*** 0.016***
(0.091)
(0.091)
(0.089)
(0.089)
(0.039)
(0.003)
0.016*
0.019** 0.026*** 0.048*** 0.022*** 0.002***
(0.009)
(0.009)
(0.009)
(0.009)
(0.004)
(0.000)
-0.171*** -0.125**
-0.060
-0.025
-0.003
(0.057)
(0.059)
(0.058)
(0.026)
(0.002)
No
No
Yes
Yes
Yes
Yes
No
No
No
Yes
Yes
Yes
0.026
0.027
0.029
0.041
0.041
0.015
108,309
108,309
107,819
107,819
107,819
107,819
NOTES: Dependent variable is Public Works. Robust standard errors (clustered by district) in parentheses.
*significant at 10% **significant at 5% ***significant at 1%
Larger households are more likely to participate as is shown by the significant and positive coefficient on the household size variable (denoted by HH size). This is fairly intuitive as larger
households have more members who can participate. Also most large households typically include
children and people of advanced years — this puts more pressure on the earning members of the
household (“more mouths to feed”–type of argument).
The negative coefficient on per-capita expenditure (Pce) indicates that poorer households are more
likely to have worked in these programs. Also, the negative and significant coefficient on the
Urban dummy (indicating whether the household comes from an urban area or not) suggests that
these programs were primarily based in rural areas; more on this later. The negative coefficient
on Education (a dummy for whether any household member has completed primary schooling or
not) suggests that households with lower levels of education were participating in these programs.
Although, the district level poverty figures are similar for reserved and non-reserved districts (see
Table 6 in the Appendix), the analysis is repeated while explicitly controlling for district-level
poverty. Table 9 in the Appendix reports such regressions. We find that the coefficient on Reserved
dummy is negative and significant as in Table 2, while the coefficient on Poverty is not significant
even at the 10% level.
28
Table 3 reports similar regressions where the total sample of households is cut along different
lines. The sample is restricted, in turns, to SC households (column 1), rural households (column
2), urban households (column 3) and households from only single-constituency districts (columns
4–5). OLS regressions are reported in this table and all subsequent ones. This is mainly because of
two reasons: (i) the OLS results are substantively similar to logit or probit regressions and (ii) the
interpretation of coefficients (especially those on interaction terms) is much more straightforward
and intuitive with OLS.
TABLE 3. H OUSEHOLD - LEVEL REGRESSIONS (ROBUSTNESS ): Reservation and
household participation in ‘Public Works’ (different samples).
Reserved
Pce
SC dum.
Urban dum.
Hindu dum.
HH size
Educ.
Home dum.
Employ dum.
[1]
[2]
[3]
[4]
[5]
SC
Rural
Urban
Single
Single
-0.023*** -0.020***
-0.002
-0.006
-0.001*
(0.007)
(0.006)
(0.004)
(0.011)
(0.000)
0.011**
0.000
-0.001
0.005
0.005
(0.006)
(0.004)
(0.002)
(0.004)
(0.004)
-0.009*
0.000
-0.009
-0.008
(0.005)
(0.003)
(0.006)
(0.006)
-0.003
-0.021** -0.020**
(0.017)
(0.008)
(0.008)
0.024** 0.019*** 0.006*** 0.030*** 0.030***
(0.011)
(0.005)
(0.002)
(0.006)
(0.005)
0.003*** 0.003***
0.001* 0.003*** 0.003***
(0.001)
(0.001)
(0.000)
(0.001)
(0.001)
-0.007 -0.020***
0.002 -0.026*** -0.026***
(0.005)
(0.005)
(0.003)
(0.008)
(0.008)
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Adjusted R2
0.010
0.012
0.002
0.019
0.020
Observations
16,760
68,236
39,583
43,434
43,434
NOTES: OLS regressions where the dependent variable is Public Works. Reserved in columns 1–4 refers
to Reserved dummy and to Reserved Time in column 5. SC refers to regressions with only SC households;
likewise for Rural and Urban. Single means that the the sample is restricted to households from only those
districts which house a single constituency (there are 182 such districts). Robust standard errors (clustered
by district) in parentheses. *significant at 10% **significant at 5% ***significant at 1%
In Table 3, the dependent variable is still the dummy variable indicating the household’s participation in Public Works programs. For columns 1–4 the variable indicating reservation is Reserved
dummy as before. Here, once again the coefficient on Reserved dummy is negative in all specifications. However, it is no longer significant for columns 3–4. For urban households, this is
potentially due to the fact that Public Works projects were primarily implemented in rural areas. In
column 4, the loss of significance is most likely due to the large drop in the number of districts (182
from 338 in total; and 23 from 79 in terms of reserved constituencies) since the variation in Reserved dummy comes at the district level. In column 5, we introduce a related but distinct measure
of reservation, namely Reserved time. Reserved time is simply the sum of the total number of years
29
of SC reservation for all constituencies within a district divided by the number of constituencies
within a district. Therefore,
1X
Reserved time =
number of years c has been reserved
n c∈d
where c denotes a constituency in district d and n is the total number of constituencies within d.
Clearly, for “Single” in column 5, n = 1 and c ≡ d. This measure of reservation goes further than
incorporating whether a particular district was subjected to any SC reservation or not; the attempt
is to capture the extent of it in terms of duration. 44
The findings in Table 3 re-iterate what has been observed so far. More reservation means lower
chance of a houshehold’s participation in Public Works.
Although any citizen of India is allowed to work on Public Works projects, it was designed to attract
only the poor households.45 Viewed in this light, one could simply use a binary variable — whether
the household is poor or not — rather than the household’s average per-capita expenditure as a
relevant explanatory variable. A variable Poor HH dummy which captures whether the houshold
lies below the poverty line or not is introduced in Table 11 (see Appendix). Table 11 re-iterates the
previous findings.
In Table 4, the variable Reserved dummy is interacted with the (log of the) household’s per-capita
expenditure to yield Reserved dummy X Pce. Interestingly, the coefficient on Reserved dummy X
Pce is positive and significant. This suggests that higher income groups got favored — in terms of
benefits from Public Works — in reserved districts.
All these results strongly suggest that Public Works projects have been less implemented in those
districts which have witnessed SC reservation.46
3.3.2. District-level regressions. Table 5 contains the results for the regressions where the dependent variable is some measure of inequality among the SC households in a district.
The idea is to capture the dispersion in incomes within the Scheduled Caste group. In particular,
the focus is on measures which capture the divergence in incomes between the relatively well-off
SCs (the “creamy layer”) and the poorer ones. Specifically, in columns 1–3 the Interquartile range
(normalized by the mean SC exp.) is used, in columns 4–6 the ratio of the upper quartile to the
lower one (i.e. Q3 /Q1 ) is used while the Gini coefficient is used in columns 7–9.
Note, Reserved dummy is significant in these regressions, except in column 9. Moreover, the
coefficient on Reserved dummy is always positive. The effect is robust to adding various controls
and it is persistent across different measures of inequality. Overall, this suggests that reservation
did in fact increase inequality among the SC group. In Table 12 (see Appendix) are regressions
44
See also Table 10 in the Appendix. The correlation between Reserved dummy and Reserved time is positive and
large (around 0.92). Although prior to 1977 there was rotation in reservation status (based on population figures from
the most prior census), this high correlation suggests that there was not a large amount of change in the identity of
reserved districts over the decades.
45
The wages were set at a level which would be attractive to poor households alone.
46
Unless one entertains the (rather implausible) notion that most of the Public Works in these reserved areas were
capital-intensive relative to those in unreserved areas.
30
TABLE 4. H OUSEHOLD - LEVEL REGRESSIONS : Reservation and household participation in ‘Public Works’ (Interaction effects).
[1]
[2]
[3]
[4]
Reserved dum.
-0.112*** -0.112*** -0.112*** -0.109***
(0.035)
(0.035)
(0.035)
(0.034)
Pce
-0.018*** -0.016*** -0.013***
-0.007*
(0.005)
(0.005)
(0.005)
(0.004)
Reserved dum. X Pce 0.019*** 0.019*** 0.019*** 0.018***
(0.006)
(0.006)
(0.006)
(0.006)
SC dum.
0.003
0.003
0.001
-0.007
(0.004)
(0.004)
(0.004)
(0.004)
Urban dum.
-0.031*** -0.029*** -0.027*** -0.012**
(0.003)
(0.003)
(0.003)
(0.005)
Hindu dum.
0.014*** 0.015*** 0.015*** 0.016***
(0.003)
(0.003)
(0.003)
(0.003)
HH size
0.001
0.001*
0.001** 0.002***
(0.000)
(0.000)
(0.000)
(0.000)
Educ.
-0.007*** -0.005**
-0.003
(0.002)
(0.003)
(0.002)
Home dum.
No
No
Yes
Yes
Employ dum.
No
No
No
Yes
Adjusted R2
0.009
0.009
0.010
0.016
Observations
108,309
108,309
107,819
107,819
NOTES: OLS regressions where the dependent variable is Public Works. Robust standard errors (clustered
by district) in parentheses. *significant at 10% **significant at 5% ***significant at 1%
where the dependent variable is a measure of SC inequality relative to the overall district inequality.
The results are substantively similar to those in Table 5.
3.4. Concerns. The empirical results presented above are of course consistent with our theory.
However, we would hesitate (given the limitations of our data) to view them as a direct test of
our theory. That said, we can rule out some alternative theoretical explanations for our empirical
findings.
A plausible alternative theory explaining the empirical results at the household level could be
one involving inexperience. One could argue that the low–caste leaders have had lesser training
and so are unable to implement policies efficiently. However, over time one would expect this
problem would go away. But the results obtained (see Table 10 in the Appendix) reveal a negative
relationship between Reserved Time and implementation of public works projects; this goes against
the spirit of the “inexperience” theory.
Alternatively, consider taste–based discrimination by individuals that the MP needs to interact with
(for e.g., bureaucrats) to get certain things done. If these people are unwilling to comply with the
low–caste politician, then one would observe less “Public Works” in reserved districts. There are
two things which go against such a story. First, in the Indian context, there is abundant anecdotal
evidence about how bureaucrats get transferred when they fail to do the politician’s bidding. So the
31
reins of power really lie in the MP’s hands. Secondly, such a story does not really make the case
for the benefits of “Public Works” to go to less–deserving wealthier households. This is what one
observes in Table 4; see, in particular, the positive and significant coefficient on the term Reserved
dum.X Pce.
The positive relationship between reservation and SC inequality in the district regressions raises
the following concern: perhaps the rise in SC inequality due to reservation is just a reflection of
the fact that the some of the poor SCs are actually benefitting from reservation and moving out of
poverty; it this rise in their expenditures that seems to exacerbate SC inequality. But this would
suggest a positive relationship between SC per–capita expenditures and reservation; something
which is not observed in Table 13 in the Appendix (see columns 1–3). Moreover, the regressions
in columns 4–9 in that table reveal that there is no significant association between SC poverty and
the reservation in the district.
There is some need to discuss the role of Scheduled Tribe (ST) reservation in this analysis. In the
results presented so far, constituencies reserved for STs have been treated as unreserved (since they
are not reserved for SCs). However, one can treat districts reserved for STs as different from both
SC reserved districts and unreserved districts. We do so and none of our main findings are altered.
Some of these results are collected in Table 14 and Table 15 (see Appendix).47
4. C ONCLUSION
This paper presents some novel results on certain economic impacts of political reservation for
minorities. The theory developed here exploits the “swing voter” idea from previous models of
electoral competition and provides some apparently counter–intuitive predictions. We show that
in the presence of ethnic biases, political quotas can potentially harm the interests of the very
minorities it was designed to benefit.
Our empirical exercise applied to SC reservation at the federal legislature confirms these theoretical
predictions. A rise in within–group inequality per se due to reservation may not be undesirable
from society’s view. However, if the rise in inequality comes at a cost to the poor then the welfare
effects of reservation may actually be negative.
On the whole, our model coupled with the empirical results suggests a grim possibility — political empowerment of minorities via mandated representation may be accompanied by creation of
greater inequities among them and may hurt the more economically vulnerable sections of society
when voting behavior is influenced by ethnic biases. Moreover, this possibility exists in a world
without statistical discrimination or perceived differences in group–wise ability or reputation.
47
Interestingly, the coefficient on the ST reservation dummy variable is positive and significant in our household–level
regressions suggesting greater Public Works implementation in such districts. This however does not refute our theory
in any way since the model is being tested for SC–reservation alone and not SC– and ST–reservation simultaneously
(see Section 2.6).
182
Observations
157
0.125
157
0.223
182
0.086
157
0.116
0.215**
(0.084)
0.249***
(0.081)
0.446**
(0.180)
-0.041
(0.074)
-0.570
(0.737)
0.278
(0.342)
-0.228***
(0.088)
No
Gini SC
[9]
157
0.202
182
0.099
157
0.120
157
0.184
0.217**
0.018*
0.017*
0.014
(0.103)
(0.010)
(0.010)
(0.018)
0.218*
0.018*
0.030
0.038
(0.127)
(0.011)
(0.020)
(0.027)
0.398** 0.095*** 0.089*** 0.090**
(0.188)
(0.020)
(0.032)
(0.037)
0.061
0.025*** 0.040***
(0.076)
(0.009)
(0.015)
-0.933
-0.128*
-0.132
(0.799)
(0.067)
(0.132)
0.296
0.043
0.048
(0.462)
(0.104)
(0.084)
-0.155
-0.023*
-0.017
(0.104)
(0.012)
(0.022)
Yes
No
No
Yes
Gini SC
[8]
NOTES: OLS regressions where the dependent variable is a measure of SC-inequality: Mean-normalized Inter-quartile range (IQR) in columns 1-3,
Q3 /Q1 ratio in columns 4-6 and Gini in columns 7-9. Robust standard errors in parentheses. Standard errors are clustered by state.*significant at
10% **significant at 5% ***significant at 1%
0.084
Adjusted R2
No
0.131*
(0.077)
-0.046
(0.058)
0.447***
(0.171)
Gini SC
[5]
SC IQR SC IQR Q3 /Q1 (SC) Q3 /Q1 (SC) Q3 /Q1 (SC)
[4]
[7]
[3]
[6]
[2]
0.082*** 0.093*** 0.087**
(0.028)
(0.031) (0.038)
Population
-0.003 0.128*** 0.108*
(0.025)
(0.039) (0.058)
Pce.
0.189*** 0.177** 0.139*
(0.066)
(0.081) (0.077)
SC%
0.024 0.075*
(0.055) (0.043)
SC% (1971)
-0.031
-0.049
(0.029) (0.037)
Gini 71 (landholdings)
0.161
0.103
(0.127) (0.171)
Population 71
-0.100**
-0.069
(0.039) (0.051)
State dummies
No
No
Yes
Reserved dum.
SC IQR
[1]
TABLE 5. D ISTRICT- LEVEL REGRESSIONS : Reservation and SC inequality.
32
33
R EFERENCES
A KERLOF, G. A., AND R. E. K RANTON (2000): “Economics and Identity,” The Quarterly Journal of Economics,
115(3), 715–753.
A RULAMPALAM , W., S. DASGUPTA , A. D HILLON , AND B. D UTTA (2009): “Electoral goals and center-state transfers: A theoretical model and empirical evidence from India,” Journal of Development Economics, 88(1), 103 –
119.
AUSTEN -S MITH , D., AND M. WALLERSTEIN (2006): “Redistribution and affirmative action,” Journal of Public
Economics, 90(10-11), 1789 – 1823.
BANERJEE , A., AND E. D UFLO (June 2008): “Mandated Empowerment: Handing Antipoverty Policy Back to the
Poor,” Annals of the New York Academy of Sciences, 1136, 333–341(9).
BANERJEE , A., AND R. PANDE (2007): “Parochial Politics: Ethnic Preferences and Politician Corruption,” Harvard
University, Kennedy School of Government, typescript.
BANERJEE , A., AND R. S OMANATHAN (2007): “The political economy of public goods: Some evidence from India,”
Journal of Development Economics, 82(2), 287 – 314.
BARDHAN , P., AND D. M OOKHERJEE (2010): “Determinants of Redistributive Politics: An Empirical Analysis of
Land Reforms in West Bengal, India,” American Economic Review, 100(4): 1572 – 1600.
BARDHAN , P., D. M OOKHERJEE , AND M. PARRA T ORRADO (2005): “Impact of Minority Reservations on Targeting in West Bengal Local Governments.,” Boston University Working Paper, December 2005.
(2010): “Impact of Political Reservations in West Bengal Local Governments on Anti-Poverty Targeting,”
Journal of Globalization and Development: Vol. 1 : Iss. 1, Article 5.
B EAMAN , L., R. C HATTOPADHYAY, E. D UFLO , R. PANDE , AND P. T OPALOVA (2009): “Powerful Women: Does
Exposure Reduce Bias?,” Quarterly Journal of Economics, 124(4), 1497 – 1540.
B ERTRAND , M., R. H ANNA , AND S. M ULLAINATHAN (2008): “Affirmative Action: Evidence from college admissions in India,” (April 2008). NBER Working Paper No. W13926.
B ESLEY, T., AND R. B URGESS (2000): “Land Reform, Poverty Reduction, and Growth: Evidence from India,” The
Quarterly Journal of Economics, 115(2), 389–430.
B ESLEY, T., R. PANDE , L. R AHMAN , AND V. R AO (2004): “The Politics of Public Good Provision: Evidence from
Indian Local Governments,” Journal of the European Economic Association, 2(2/3), 416–426.
B ESLEY, T., R. PANDE , AND V. R AO (2005): “Participatory Democracy in Action: Survey Evidence from South
India,” Journal of the European Economic Association, 3(2/3), 648–657.
C HATTOPADHYAY, R., AND E. D UFLO (2004a): “Impact of Reservation in Panchayati Raj: Evidence from a Nationwide Randomised Experiment,” Economic and Political Weekly, 39(9), 979–986.
(2004b): “Women as Policy Makers: Evidence from a Randomized Policy Experiment in India,” Econometrica, 72(5), 1409–1443.
C HAUDHURY, P. (2004): “The ’Creamy Layer’: Political Economy of Reservations,” Economic and Political Weekly,
39(20), pp. 1989–1991.
C HIN , A., AND N. P RAKASH (2011): “The redistributive effects of political reservation for minorities: Evidence
from India,” Journal of Development Economics, 96(2), 265 – 277.
DASGUPTA , I., AND R. K ANBUR (2007): “Community and class antagonism,” Journal of Public Economics, 91(9),
1816–1842.
D ESHPANDE , A. (2000): “Does Caste Still Define Disparity? A Look at Inequality in Kerala, India,” The American
Economic Review, 90(2), 322–325.
D IXIT, A., AND J. L ONDREGAN (1996): “The Determinants of Success of Special Interests in Redistributive Politics,” The Journal of Politics, 58(4), pp. 1132–1155.
(1998): “Fiscal federalism and redistributive politics,” Journal of Public Economics, 68(2), 153 – 180.
D UFLO , E. (2005): “Why Political Reservations?,” Journal of the European Economic Association, 3(2/3), 668–678.
F ERN ÁNDEZ , R., AND G. L EVY (2008): “Diversity and redistribution,” Journal of Public Economics, 92(5-6), 925
– 943.
F OSTER , A., AND M. ROSENZWEIG (2001): “Democratization, Decentralization and the Distribution of Local Public
Goods in a Poor Rural Economy.,” Working Paper, Department of Economics, Brown University.
G HOSH , S., AND S. PAL (2010): “Poverty, Heterogeneous Elite, and Allocation of Public Spending: Panel Evidence
from the Indian States,” Brunel Economics and Finance Working paper 10-20, London, UK.
G HURYE , G. S. (1969): “Caste and Race in India,” 5th edition, Popular Prakashan, India.
34
H NATKOVSKA , V., A. L AHIRI , AND S. B. PAUL (2011): “Breaking the Caste Barrier: Intergenerational Mobility in
India,” Journal of Human Resources, forthcoming.
J ENSENIUS , F. R. (2014): “Development from Representation? A Study of Quotas for Scheduled Castes in India,”
Working paper.
KOTSADAM , A., AND M. N ERMAN (2014): “Descriptive, Substantive, and Symbolic Effects of Gender Quotas in
Latin American National Elections,” Mimeo, University of Oslo.
L INDBECK , A., AND J. W. W EIBULL (1987): “Balanced-Budget Redistribution as the Outcome of Political Competition,” Public Choice, 52(3), 273–297.
L IZZERI , A., AND N. P ERSICO (2001): “The Provision of Public Goods under Alternative Electoral Incentives,” The
American Economic Review, 91(1), 225–239.
M UNSHI , K., AND M. ROSENZWEIG (2008): “The Efficacy of Parochial Politics: Caste, Commitment, and Competence in Indian Local Governments,” (September 2008). NBER Working Paper No. W14335.
M YERSON , R. B. (1993): “Incentives to Cultivate Favored Minorities Under Alternative Electoral Systems,” The
American Political Science Review, 87(4), 856–869.
N EUMARK , D., AND H. H OLZER (2000): “Assessing Affirmative Action,” Journal of Economic Literature, 38(3),
483–568.
PANDE , R. (2003): “Can Mandated Political Representation Increase Policy Influence for Disadvantaged Minorities?
Theory and Evidence from India,” The American Economic Review, 93(4), 1132–1151.
R AY, D., AND R. S ETHI (2010): “A Remark on Color-Blind Affirmative Action,” Journal of Public Economic Theory,
12, 399-406 (2010).
S EN , A. K. (1997): “On Economic Inequality,” (enlarged edition with a substantial new annex co-written with James.
E. Foster), Clarendon Press, 1997.
35
A PPENDIX
4.1. Proofs: Extension of the Model.
Proof. [P ROPOSITION 2*.]
Note, γ = (l, h) implies that α(w, l) = t(w) + s and α(w, h) = t(w) − s.
Recall P 0 s FOC for any w:
[ψ(γ, w) − ρ(w)]u0 (x(w) + w) = λ.
(12)
Given the strict concavity of u(.), wP is where ψ(γ, w) − ρ(w) is maximized. Analogously (from
R0 s FOC), wR is where ψ(γ, w) is maximized.
Note,
ψ(γ, w∗ ) = f (s).
For any w > w∗ ,
ψ(γ, w) = π(w)f (d(w) − t(w) − s) + (1 − π(w))f (d(w) − t(w) + s).
where d(w) − t(w) > 0. The RHS is lower than
π(w∗ )f (s − ) + (1 − π(w∗ ))f (s + )
for some ∈ (0, s] by the unimodality and symmetry of f around 0. Therefore, under C1,
ψ(γ, w) < f (s)
for any w > w∗ . This implies wj ≤ w∗ for j = P, R.
To rule out wP = w∗ , note the following. For any w < w∗ ,
ψ(γ, w) = π(w)f (d(w) − t(w) − s) + (1 − π(w))f (d(w) − t(w) + s).
where d(w) − t(w) < 0. Let δw ≡ t(w) − d(w) for any w ≤ w∗ . Clearly, δw ≥ 0 with δw∗ = 0. If
we can establish that there is some δw > 0 such that
π(w)f (s + δw ) + (1 − π(w))f (s − δw ) ≥ f (s) = ψ(γ, w∗ )
then we are done.
Define χ(w) ≡ (1 − π(w))[f (s − δw ) − f (s)] + π(w)[f (s + δw ) − f (s)]. Hence, showing that
χ(w) ≥ 0 for some w < w∗ will establish wP < w∗ .
Pick any w arbitrarily close to w∗ such that δw > 0 but infinitesimal. Hence,
χ(w) = δw {(1 − π(w))[
f (s − δw ) − f (s)
f (s + δw ) − f (s)
] + π(w)[
]}
δw
δw
Given that δw is arbitrarily close to 0,
χ(w) ≈ δw {(π(w) − 1)f 0 (s − δw ) + π(w)f 0 (s)} ≈ (2π(w) − 1)f 0 (s)δw > 0.
This implies wP < w∗ and establishes the proposition.
36
Proof. [P ROPOSITION 3*.] Under such same-caste configurations, the caste bias for every voter
loses relevance, i.e. α(w, h) = α(w, l) = t(w) for every income group w.
Consider party P’s FOC for any w:
[f (d(w) − t(w)) − ρ(w)]u0 (x(w) + w) = λ.
For w = w∗ , this takes the following form:
f (0)u0 (x(w∗ ) + w∗ ) = λ
Recall,
d(w), −t(w) < 0 if w < w∗
d(w) = t(w) = 0 if w = w∗
d(w), −t(w) > 0 if w > w∗
Now, we claim that the most favored group for party P , i.e., wP is actually w∗ . First, we show that
f (0) > f (d(w) − t(w)) − ρ(w)
∗
for any w < w .
Take any w < w∗ . Now,
f (0) − f (t(w)) < f (0) − f (d(w) − t(w))
since d(w), −t(w) < 0 and f is symmetric around 0. Hence,
f (0) − f (t(w)) + ρ(w) < f (0) − f (d(w) − t(w)) + ρ(w).
But the LHS is non-negative by C2. Hence,
f (0) − f (d(w) − t(w)) + ρ(w) > 0
which gives wP 6= w < w∗ .
For w > w∗ ,
f (0) > f (d(w) − t(w)) − ρ(w)
since ρ(w) > 0 for such w and by the unimodality and symmetry of f around 0. This establishes
wP = w ∗ .
Party R0 s FOC for any w is
f (d(w) − t(w))u0 (y(w) + w) = µ.
We have f (0) > f (d(w) − t(w)) for w 6= w∗ by the unimodality and symmetry of f around 0.
This establishes wR = w∗ .
4.2. Institutional Background. The economic backwardness of the Scheduled Castes is a direct
fallout of the caste system which has been an integral part of the Indian society for centuries.
Scheduled castes form roughly 16% of the Indian population. Caste membership is hereditary and
is largely reflective of the person’s traditional occupation choice. In particular, members of the
scheduled castes were traditionally assigned low-paying, labor-intensive tasks. In fact, they were
denied education and barred from high-paying occupations. Besides, there were restrictions on
their asset ownership.
In light of this history, the Constitution of India (which came into force in 1950) mandates political
reservation in favor of scheduled castes (SCs) and scheduled tribes (STs) in every state and national
37
election.48 Moreover, there are clear directives for the state governments to use public policy to
improve the economic conditions of these groups.
A brief word about the Indian political system is in order. The Indian Parliament is bicameral in
nature. However, the Lok Sabha is the popularly elected House and is de facto more powerful
than the other House (Rajya Sabha). The popularly elected Members of Parliament (MP) enjoy
a five-year term after which fresh Lok Sabha elections are held. There were 518 (Lok Sabha)
constituencies in 1971. This went up to 542 after a Delimitation order in 1976 and then to 543 in
1991.
Population is the basis of allocation of seats of the Lok Sabha. As far as possible, every State
gets representation in the Lok Sabha in proportion to its population as per census figures. Hence,
larger and more populous states have more seats in the Lok Sabha as compared to their smaller and
sparsely-populated counterparts. For example, Uttar Pradesh (a north Indian state) with a population of over 166 million has 80 Lok Sabha seats while the state of Nagaland with a population of
less than 2 million has only one Lok Sabha seat.
Allocation of seats for Scheduled Castes and Tribes in the Lok Sabha are made on the basis of
proportion of Scheduled Castes and Tribes in the State concerned to that of the total population
(provision contained in Article 330 of the Constitution of India). Thus, a state which has a relatively larger proportion of SCs will have a larger fraction of its total Lok Sabha seats reserved for
SC candidates. For example, Uttar Pradesh which has relatively large proportion of SCs has 17 out
80 seats reserved for SCs. On the other hand, the state of Gujarat with relatively lower proportion
of SCs has only 2 out of 26 seats reserved for SCs.
The choice of the constituencies (within a state) that will be reserved for SCs is guided by the
following considerations. The Delimitation Commission stipulates that constituencies for SCs
are to be distributed in different parts of the state and seats are to be reserved for SCs in those
constituencies where the percentage of their population to the total population is comparatively
large.
The rise in political participation among the Scheduled Castes is, to a large extent, due to reservation of constituencies for them. In the 1980s, the Scheduled Castes were able to establish a
successful caste-based party. This significantly enhanced their representation in national politics.
4.3. Data. For information on the reservation status of the Parliamentary constituencies in India,
the sources are the Statistical Reports on General Elections available from the Election Commission of India. As per the provisions of the Indian Constitution, readjustment and delimitation of
constituencies was to take place after every decennial census. Accordingly, Delimitation Commissions were constituted in 1952, 1963 and 1972.
48
Aside from political reservation, the Indian constitution mandates reservation of jobs (through job quotas) in all State
and Central departments and government institutions. In 1950, the constitution laid down 15% and 7.5% of vacancies
to government aided educational institutes and for jobs in the government/public sector, as reserved quota for the SC
and ST candidates respectively for a period of five years, after which the situation was to be reviewed. This period was
routinely extended by the following governments and the Indian Parliament, and no revisions were undertaken. Later,
reservations were introduced for other sections as well. The Supreme Court ruling that reservations cannot exceed
50% (which it judged would violate equal access guaranteed by the Constitution) has put a cap on reservations.
38
However, the Parliament by passing the 42nd Constitutional Amendment Act (1976) imposed a
moratorium on the number of seats allocated and the territorial boundaries of constituencies as
determined by Delimitation of Parliamentary and Assembly Constituencies Order 1976 until the
publication of the population figures of 2001 Census. Accordingly, the Fourth Delimitation Commission was set up under the Delimitation Act 2002 after a gap of as many as 30 years. This means
that there has been no change in the identity or the number of reserved SC seats after the 1977 Lok
Sabha elections. So whichever seats were reserved in 1977 remained so right through till 2001 and
no other seats were reserved between 1977 and 2001.
In India, large scale household surveys are conducted quinquennially as part of the National Sample Surveys (NSS). The survey rounds cover the entire nation and capture monthly expenditure
incurred by the sample household for the purpose of domestic consumption.49 The recall period
used is 30 days, i.e., the surveyed households are asked to provide information on consumption
expenditure incurred over the past 30 days. For this analysis, the 43rd round of NSS is extensively used for information on household monthly per capita expenditure. Data for this round was
collected during July 1987- June 1988. There is data on monthly per capita expenditure at the
household level for over 300 districts in India – these districts span 18 major states in India and
account for over 95% of the Indian population.
These large scale NSS rounds also have information regarding the surveyed households’ participation in the nationwide poverty-reduction programs. In fact, most of the poverty reduction programs
gathered momentum in the late 1970s/early 1980s50 This makes the 43rd NSSO round particularly
suitable for the task at hand.
Data on various district-level characteristics from the census of India conducted in 1971 has also
been used. This is primarily to facilitate comparison — along various dimensions — of districts
which underwent reservation in 1977 with those which did not.
49
Unfortunately, in the case of India there are no available sources of income data on a nationwide scale. Hence,
expenditure is the closest one can get.
50
See for e.g. Banerjee and Somanathan (2007) for details on “Garibi Hatao” (translates into “Eradicate Poverty”)
programs during this era.
39
TABLE 6. H OUSEHOLD - LEVEL
reserved districts
SUMMARY STATISTICS :
Reserved vs non-
[A]
[B]
[C]
All
Mean
Non-Reserved
Mean
Reserved
Mean
Difference
NR – R
0.05
(0.21)
Per-capita exp. (in Rs.)
196.41
(207.49)
SC dummy
0.18
(0.39)
Urban dummy
0.24
(0.43)
Hindu dummy
0.84
(0.37)
HH size
5.00
(2.60)
Primary Educ. dummy
0.36
(0.48)
Homeowner dummy
0.82
(0.38)
Dwelling type (1 or 2)
1.81
(0.39)
Dwelling structure (1, 2 or 3)
1.87
(0.82)
Self-Employ. Rural dummy
0.38
(0.49)
Laborer Rural
0.31
(0.46)
Self-Employ. Urban dummy
0.08
(0.27)
Wage/Salary earner dummy
0.11
(0.31)
Casual Laborer dummy
0.03
(0.17)
Poor HH dummy (Y=1, N=0)
0.33
(0.47)
Headcount Ratio (district)
0.43
(0 .21)
Observations
108,312
0.05
(0.22)
195.83
(191.82)
0.16
(0.37)
0.24
(0.43)
0.84
(0.37)
5.02
(2.57)
0.36
(0.48)
0.82
(0.38)
1.80
(0.40)
1.89
(0.82)
0.38
(0.49)
0.31
(0.46)
0.08
(0.27)
0.11
(0.31)
0.03
(0.18)
0.34
(0.47)
0.43
(0.22)
78,211
0.04
(0.19)
197.77
(240.55)
0.23
(0.42)
0.24
(0.43)
0.85
(0.36)
5.03
(2.63)
0.35
(0.48)
0.83
(0.37)
1.83
(0.38)
1.84
(0.83)
0.38
(0.49)
0.31
(0.46)
0.09
(0.28)
0.11
(0.31)
0.03
(0.17)
0.33
(0.47)
0.43
(0.18)
30,101
0.02
(0.00)***
-1.94
(1.59)
-0.07
(0.00)***
0.00
(0.00)
-0.01
(0.00)**
-0.00
(0.02)
0.00
(0.00)
-0.01
(0.00)***
-0.03
(0.00)***
0.05
(0.01)***
0.00
(0.00)
0.00
(0.00)
-0.01
(0.00)***
0.00
(0.00)
0.00
(0.00)***
0.00
(0.00)
-0.01
(0.03)
Variables (Household)
Public Works dummy
NOTES: Panel A pertains to households from all districts, while Panels B and C present household
characteristics from non-reserved (250) and reserved districts (77), respectively (a ‘reserved’ district
contains at least one reserved constituency). # Observations for Dwelling type and Dwelling structure are
107,909 (panel A), 77,918 (panel B) and 29,991(panel C). T-test used for differences. Standard errors in
parentheses for the columns indicating differences; standard deviation in parentheses for all others.
*significant at 10% **significant at 5% ***significant at 1%.
40
TABLE 7. D ISTRICT- LEVEL SUMMARY STATISTICS FROM 1987-88: Comparing
reserved with non-reserved districts (for single-constituency districts).
[1]
[2]
[3]
All
Non-R
R
Mean
Mean
Mean
Diff. (NR–R)
0.475
(0.180)
Q3/Q1 (SC)
1.747
(0.408)
Gini SC
0.239
(0.076)
IQR (overall)
0.520
(0.089)
Q3/Q1 (overall)
1.862
(0.223)
Gini (overall)
0.295
(0.052)
SC per-capita exp. (in Rs.)
158.749
(43.850)
Average per-capita exp. (in Rs.) 191.265
(50.985)
SC population (%)
18.565
(9.644)
Population
0.203
(0.086)
Headcount Ratio SC
0.420
(0.227)
Poverty Gap Ratio SC
0.102
(0.073)
Reserved dummy
0.126
(0.333)
0.464
(0.184)
1.729
(0.415)
0.237
(0.079)
0.517
(0.090)
1.859
(0.229)
0.296
(0.053)
158.876
(44.920)
190.934
(49.954)
17.424
(9.099)
0.205
(0.085)
0.422
(0.229)
0.102
(0.072)
0.547
(0.135)
1.866
(0.347)
0.253
(0.057)
0.543
(0.075)
1.883
(0.179)
0.288
(0.041)
157.868
(36.436)
193.547
(58.809)
26.452
(9.807)
0.191
(0.098)
0.403
(0.210)
0.104
(0.084)
-0.083
(0.040)**
-0.137
(0.091)*
-0.016
(0.017)
-0.027
(0.020)*
-0.024
(0.050)
0.008
(0.012)
1.008
(9.809)
-2.613
( 11.404)
-9.028
(2.050)***
0.014
(0.019)
0.019
(0.051)
-0.002
(0.016)
159
23
Variables (district)
IQR (SC)
Observations
182
[4]
NOTES: The sample of districts is restricted to those which correspond to a single constituency. All
variables except Reserved dummy have been constructed using data from the NSSO round 43 (1987-88).
Standard errors in parentheses for the columns indicating differences; standard deviation in parentheses for
all other columns. T-test used for comparing differences. *significant at 10% **significant at 5%
***significant at 1%
41
TABLE 8. D ISTRICT- LEVEL SUMMARY STATISTICS :
constituency’ with ‘Non-Single constituency’ districts.
Variables (district)
Comparing ‘Single-
[A]
[B]
Single
Non-Single
Difference
N
Mean
N
Mean
S–NonS
IQR (SC)
182
145
Q3/Q1 (SC)
182
Gini SC
182
IQR (overall)
182
Q3/Q1 (overall)
182
Gini (overall)
182
0.475
(0.180)
1.747
(0.408)
0.239
(0.076)
0.520
(0.089)
1.862
(0.223)
0.295
(0.052)
158.749
(43.850)
191.265
(50.985)
18.565
(9.644)
0.203
(0.086)
0.126
(0.333)
0.469
(0.134)
1.727
(0.309)
0.248
(0.068)
0.528
(0.099)
1.895
(0.267)
0.303
(0.057)
155.750
(39.527)
191.869
(54.656)
17.400
(8.599)
0.374
(0.179)
0.346
(0.477)
0.006
(0.018)
0.020
(0.041)
-0.009
(0.008)
-0.008
(0.010)
-0.033
(0.027)
-0.008
(0.006)
2.998
(4.573)
-0.605
(5.751)
1.165
(1.001)
-0.171
(0.015)***
-0.220
(0.044)***
SC per-capita exp. (in Rs.) 182
Per-capita exp. (in Rs.)
182
SC population (%)
182
Population
182
Reserved dummy
182
145
145
145
145
145
156
156
156
156
156
NOTES: Panel A corresponds to the 182 districts where each district has just one constituency while Panel
B has data from the remaining 156 districts. Standard errors in parentheses for the columns indicating
differences; standard deviation in parentheses for all others. T-test used for comparing differences.
*significant at 10% **significant at 5% ***significant at 1%
No
0.008
108,309
Employ dum.
Adjusted R2
Observations
[3]
[4]
[5]
[6]
[7]
[8]
108,309
0.009
No
108,309
0.008
No
108,309
0.009
No
107,819
0.009
No
107,819
0.010
No
107,819
0.014
Yes
107,819
0.016
Yes
0.016
0.015
0.016
0.013
0.014
0.015
0.016
(0.013)
(0.013)
(0.013)
(0.013)
(0.013)
(0.013)
(0.013)
-0.017***
-0.017***
-0.016***
-0.016***
(0.005)
(0.005)
(0.005)
(0.005)
-0.011*** -0.009** -0.009**
-0.006*
-0.006*
-0.001
-0.000
(0.004)
(0.004)
(0.004)
(0.004)
(0.004)
(0.003)
(0.003)
0.003
0.001
0.002
-0.001
0.001 -0.008**
-0.007
(0.004)
(0.004)
(0.004)
(0.004)
(0.004)
(0.004)
(0.004)
-0.031*** -0.029*** -0.029*** -0.027*** -0.027*** -0.012** -0.011**
(0.003)
(0.003)
(0.003)
(0.003)
(0.003)
(0.005)
(0.005)
0.014*** 0.015*** 0.015*** 0.015*** 0.015*** 0.016*** 0.016***
(0.003)
(0.003)
(0.003)
(0.003)
(0.003)
(0.003)
(0.003)
0.001*
0.001*
0.001**
0.001**
0.001** 0.002*** 0.002***
(0.000)
(0.000)
(0.000)
(0.000)
(0.000)
(0.000)
(0.000)
-0.007*** -0.008*** -0.006** -0.005**
-0.003
-0.003
(0.003)
(0.002)
(0.003)
(0.003)
(0.002)
(0.002)
No
No
No
Yes
Yes
Yes
Yes
[2]
NOTES: OLS regressions where the dependent variable is Public Works. The Headcount Ratio (HCR) is used as the measure of (district-level)
poverty. Robust standard errors (clustered by district) in parentheses. *significant at 10% **significant at 5% ***significant at 1%
No
-0.011***
(0.004)
0.002
(0.004)
-0.031***
(0.003)
0.014***
(0.003)
0.001
(0.000)
0.015
(0.013)
Home dum.
Educ.
HH size
Hindu dum.
Urban dum.
SC dum.
Pce
Reserved dum.
Poverty (HCR)
[1]
TABLE 9. H OUSEHOLD - LEVEL REGRESSIONS : Reservation, district-level poverty and household participation in ‘Public Works’.
42
All
All
SC
[3]
SC
[4]
Rural
[5]
Rural
[6]
Single
[7]
Single
[8]
107,819
Observations
107,819
0.015
Yes
16,760
0.007
No
16,760
0.009
Yes
68,236
0.006
No
68,236
0.011
Yes
43,434
0.011
No
43,434
0.020
Yes
NOTES: OLS regressions where the dependent variable is Public Works. All refers to regressions with all households; likewise for SC and Rural.
Single means that the the sample is restricted to households from only those districts which house a single constituency (there are 182 such districts).
Robust standard errors (clustered by district) in parentheses. *significant at 10% **significant at 5% ***significant at 1%
0.009
No
Adjusted R2
Employ dum.
Reserved Time -0.001*** -0.001*** -0.001** -0.001** -0.001*** -0.001***
-0.001*
-0.001*
(0.000)
(0.000)
(0.001)
(0.001)
(0.000)
(0.000)
(0.000)
(0.000)
Pce
-0.007**
-0.002
0.008 0.011**
-0.007
0.001
-0.003
0.005
(0.004)
(0.003)
(0.005)
(0.005)
(0.005)
(0.004)
(0.005)
(0.004)
SC dum.
0.000
-0.007*
-0.001 -0.010**
0.003
-0.008
(0.004)
(0.004)
(0.004)
(0.005)
(0.006)
(0.006)
Urban dum.
-0.028*** -0.011** -0.042***
-0.004
-0.029*** -0.020**
(0.003)
(0.005)
(0.007)
(0.017)
(0.006)
(0.008)
Hindu dum.
0.015*** 0.016***
0.021** 0.021** 0.017*** 0.018*** 0.029*** 0.030***
(0.003)
(0.003)
(0.011)
(0.011)
(0.004)
(0.004)
(0.005)
(0.005)
HH size
0.001** 0.002*** 0.003*** 0.003*** 0.002*** 0.003***
0.002** 0.003***
(0.000)
(0.000)
(0.001)
(0.001)
(0.001)
(0.001)
(0.001)
(0.001)
Educ.
-0.005**
-0.003
-0.010*
-0.007 -0.025*** -0.020*** -0.032*** -0.026***
(0.003)
(0.002)
(0.005)
(0.005)
(0.005)
(0.005)
(0.008)
(0.008)
Home dum.
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
[2]
[1]
TABLE 10. H OUSEHOLD - LEVEL REGRESSIONS (ROBUSTNESS ): Duration of reservation and household participation
in ‘Public Works’.
43
All
All
All
[3]
All
[4]
Poor
[5]
Poor
[6]
Poor
[7]
Poor
[8]
0.009
108,309
Adjusted R2
Observations
108,309
0.009
No
107,819
0.010
No
107,819
0.015
Yes
32,618
0.012
No
32,618
0.012
No
32,463
0.012
No
32,463
0.018
Yes
NOTES: OLS regressions where the dependent variable is Public Works. All refers to regressions with all households and Poor means that the
sample is restricted to households which are below the official poverty line. Robust standard errors (clustered by district) in parentheses. *significant
at 10% **significant at 5% ***significant at 1%
No
Employ dum.
Reserved dum. -0.016*** -0.016*** -0.016*** -0.016*** -0.028*** -0.028*** -0.028*** -0.027***
(0.005)
(0.005)
(0.005)
(0.005)
(0.008)
(0.008)
(0.008)
(0.008)
Poor HH dum. 0.014*** 0.013***
0.010**
0.005
(0.004)
(0.004)
(0.004)
(0.004)
Pce
-0.009
-0.008
-0.007
-0.001
(0.012)
(0.012)
(0.012)
(0.012)
SC dum.
0.003
0.002
0.000
-0.007*
-0.006
-0.006
-0.008 -0.015**
(0.003)
(0.004)
(0.004)
(0.004)
(0.006)
(0.006)
(0.006)
(0.007)
Urban dum.
-0.036*** -0.033*** -0.030*** -0.012** -0.046*** -0.045*** -0.041***
-0.015
(0.004)
(0.003)
(0.003)
(0.005)
(0.006)
(0.006)
(0.005)
(0.009)
Hindu dum.
0.015*** 0.015*** 0.016*** 0.016*** 0.017*** 0.017*** 0.018*** 0.019***
(0.004)
(0.004)
(0.003)
(0.003)
(0.006)
(0.006)
(0.006)
(0.006)
HH size
0.001*
0.001** 0.001*** 0.002***
0.001**
0.001**
0.002** 0.002***
(0.000)
(0.000)
(0.000)
(0.000)
(0.001)
(0.001)
(0.001)
(0.001)
Educ.
-0.008*** -0.005**
-0.002
-0.006
-0.004
-0.001
(0.003)
(0.003)
(0.002)
(0.004)
(0.004)
(0.004)
Home dum.
No
No
Yes
Yes
No
No
Yes
Yes
[2]
[1]
TABLE 11. H OUSEHOLD - LEVEL REGRESSIONS : Reservation and below-poverty line household participation in ‘Public Works’.
44
IQR
IQR
[4]
IQR Q3 /Q1
[3]
[6]
Q3 /Q1 Q3 /Q1
[5]
Gini
[7]
Gini
[8]
Gini
[9]
182
Observations
157
0.057
157
0.167
182
0.029
157
0.037
157
0.137
182
0.010
157
0.086
157
0.065
IQR SC
IQR All
in
Gini SC
columns 1-3,
in columns 4-6 and Gini
All in columns 5-8. IQR stands for the Inter-quartile range (normalized by the mean). Robust
standard errors in parentheses. Standard errors are clustered by state.*significant at 10% **significant at 5% ***significant at 1%
Q3 /Q1 SC
Q3 /Q1 All
NOTES: OLS regressions where the dependent variable is a measure of SC-inequality relative to overall district-level inequality :
0.030
Adjusted R2
0.113** 0.127** 0.132* 0.055 0.083** 0.085* 0.081***
0.041
0.037
(0.053) (0.054) (0.069) (0.042) (0.042) (0.047)
(0.031)
(0.035) (0.054)
Population
-0.036 0.133**
0.074 -0.057
0.033 0.004
0.013
-0.059
-0.005
(0.053) (0.065) (0.107) (0.035) (0.043) (0.066)
(0.039)
(0.058) (0.083)
Pce.
0.190
0.203
0.145 0.082
0.093 0.057
0.084*
0.095
0.047
(0.123) (0.150) (0.130) (0.083) (0.101) (0.088)
(0.050)
(0.091) (0.125)
SC%
0.079 0.157**
0.014 0.053
0.086*** 0.123**
(0.107) (0.076)
(0.068) (0.043)
(0.028) (0.050)
SC% (1971)
-0.066
-0.103
-0.035 -0.052
0.021
0.043
(0.057) (0.066)
(0.039) (0.041)
(0.169) (0.393)
Gini 71 (landholdings)
0.169
0.379
0.077 0.245
0.078
0.119
(0.222) (0.346)
(0.147) (0.227)
(0.235) (0.243)
Population 71
-0.123*
-0.069
-0.053 -0.029
0.034
0.017
(0.070) (0.086)
(0.046) (0.052)
(0.034) (0.067)
State dummies
No
No
Yes
No
No
Yes
No
No
Yes
Reserved dum.
[2]
[1]
TABLE 12. D ISTRICT- LEVEL REGRESSIONS : Reservation and SC inequality relative to overall district inequality.
45
SC pce
SC pce
SC pce
[3]
HCR SC
[4]
HCR SC
[5]
HCR SC
[6]
PGR SC
[7]
PGR SC
[8]
PGR SC
[9]
182
Observations
157
0.694
157
0.715
182
0.497
157
0.504
157
0.614
182
0.391
157
0.455
157
0.579
NOTES: OLS regressions where the dependent variable is log SC per-capita exp. (in the district) for columns 1–3 and a measure of SC-poverty in
the remaining columns (Headcount Ratio in columns 4–6 and Poverty Gap Ratio in columns 7–9). Robust standard errors in parentheses. Standard
errors are clustered by state.*significant at 10% **significant at 5% ***significant at 1%
0.697
Adjusted R2
-0.006
0.020
0.032
-0.015
-0.010
-0.038
0.002
0.006
-0.003
(0.023)
(0.023)
(0.030)
(0.019)
(0.027)
(0.024)
(0.012)
(0.015)
(0.012)
Population
-0.042
-0.058
-0.034
0.027
0.055
0.119**
-0.002
0.011
0.033*
(0.028)
(0.039)
(0.065)
(0.026)
(0.050)
(0.060)
(0.011)
(0.018)
(0.018)
Pce.
0.830*** 0.866*** 0.936*** -0.607*** -0.618*** -0.785*** -0.177*** -0.183*** -0.225***
(0.047)
(0.065)
(0.077)
(0.050)
(0.052)
(0.074)
(0.017)
(0.023)
(0.024)
SC%
-0.007
-0.007
-0.020
0.013
0.003
0.012
(0.019)
(0.032)
(0.038)
(0.036)
(0.010)
(0.009)
SC% (1971)
-0.094
-0.219
0.064
0.266
-0.051
-0.009
(0.208)
(0.271)
(0.259)
(0.259)
(0.083)
(0.090)
Gini 71 (landholdings)
-0.142
-0.034
0.117
0.148
0.081
0.086
(0.126)
(0.140)
(0.176)
(0.160)
(0.060)
(0.056)
Population 71
0.018
-0.000
-0.013
-0.072
-0.012 -0.036**
(0.032)
(0.049)
(0.039)
(0.048)
(0.015)
(0.015)
State dummies
No
No
Yes
No
No
Yes
No
No
Yes
Reserved dum.
[2]
[1]
TABLE 13. D ISTRICT- LEVEL REGRESSIONS (ROBUSTNESS ): Reservation and SC pce, SC poverty.
46
47
TABLE 14. H OUSEHOLD - LEVEL REGRESSIONS : Household participation in
‘Public Works’ while accounting for ST reservation.
Reserved dum.
[1]
[2]
[3]
[4]
[5]
[6]
Logit
Probit
OLS
Logit
Probit
OLS
-0.292**
(0.125)
-0.129**
(0.055)
ST resv. dum.
Pce
SC dum.
Urban dum.
Hindu dum.
HH size
Educ.
Home dum.
Employ dum.
Pseudo/Adj. R2
Observations
0.045
0.021
(0.063)
(0.028)
-0.064
-0.026
(0.070)
(0.032)
-0.563*** -0.230***
(0.212)
(0.088)
0.338*** 0.151***
(0.092)
(0.040)
0.051***
0.024**
(0.010)
(0.005)
-0.075
-0.031
(0.063)
(0.028)
Yes
Yes
-0.112** -0.329*** -0.143*** -0.014***
(0.005)
(0.120)
(0.053)
(0.005)
0.482**
0.229**
0.030**
(0.194)
(0.094)
(0.015)
0.002
-0.031
-0.012
-0.001
(0.002)
(0.069)
(0.031)
(0.003)
-0.003
-0.089
-0.039
-0.005
(0.003)
(0.071)
(0.032)
(0.004)
-0.012** -0.556*** -0.228*** -0.011**
(0.005)
(0.212)
(0.088)
(0.005)
0.013*** 0.357*** 0.161*** 0.014***
(0.003)
(0.086)
(0.038)
(0.003)
0.002*** 0.049*** 0.023*** 0.002***
(0.000)
(0.009)
(0.004)
(0.000)
-0.003
-0.057
-0.025
-0.003
(0.003)
(0.059)
(0.027)
(0.002)
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
0.035
0.035
0.012
0.044
0.044
0.017
98,123
98,123
98,123
107,819
107,819
107,819
NOTES: Dependent variable is Public Works. In columns (1) — (3), all districts which have at least
constituency reserved for STs have been dropped. In columns (4) — (6), a dummy for ST reservation (=1 if
any constituency within the district is reserved for STs, and 0 otherwise) is included as an added control.
Robust standard errors (clustered by district) in parentheses. *significant at 10% **significant at 5%
***significant at 1%
48
TABLE 15. D ISTRICT- LEVEL REGRESSIONS : Reservation and SC inequality
while accounting for ST reservation.
[1]
SC IQR
[2]
[5]
[6]
SC IQR Q3 /Q1 (SC) Q3 /Q1 (SC) Gini SC
Gini SC
Reserved dum.
[3]
0.215**
(0.088)
[4]
0.116*** 0.118***
(0.032)
(0.033)
ST resv. dum.
0.077**
(0.025)
Population
0.102** 0.113**
(0.045)
(0.044)
Pce.
0.200*** 0.210**
(0.084)
(0.078)
SC%
0.010
0.013
(0.27)
(0.032)
SC% (1971)
-0.289
-0.315
(0.243)
(0.290)
Gini 71 (landholdings)
0.115
0.060
(0.151)
(0.152)
Population 71
-0.076* -0.086**
(0.040)
(0.039)
0.221**
(0.087)
0.424**
(0.197)
-0.048
(0.066)
-0.296
(0.586)
0.281
(0.345)
-0.202**
(0.093)
0.218** 0.019*
0.018
(0.089) (0.010)
(0.010)
0.225***
0.036***
(0.066)
(0.012)
0.227**
0.029
0.027
(0.092) (0.022)
(0.022)
0.482** 0.080** 0.095**
(0.185) (0.030)
(0.034)
-0.036 0.026** 0.026**
(0.077) (0.010)
(0.009)
-0.369
-0.103
-0.096
(0.702) (0.062)
(0.070)
0.212
0.019
0.033
(0.352) (0.102)
(0.104)
-0.209**
-0.022
-0.020
(0.092) (0.013)
(0.013)
Adjusted R2
0.112
0.131
0.110
0.132
0.109
0.130
Observations
143
157
143
157
143
157
NOTES: OLS regressions where the dependent variable is a measure of SC–inequality: Mean-normalized
Inter-quartile range (IQR) in columns (1) and (2), Q3 /Q1 ratio in columns (3) and (4) and Gini in columns
(5) and (6). In columns (1), (3) and (5) all constituencies reserved for STs are dropped while in columns
(2), (4) and (6) a dummy for ST reservation (=1 if the constituency is reserved for STs, and 0 otherwise) is
included as an added control. Robust standard errors in parentheses. Standard errors are clustered by
state.*significant at 10% **significant at 5% ***significant at 1%