Politician Identity and Religious Conflict in India

Politician Identity and Religious Conflict in India*
Sonia Bhalotra
University of Bristol
Irma Clots-Figueras
Universidad Carlos III de Madrid
Lakshmi Iyer
Harvard Business School
July 2012
Preliminary and Incomplete
Comments Welcome
*
Contact: s.bhalotra@bristol.ac.uk; irmaclots@gmail.com; liyer@hbs.edu. We are grateful toDebraj
Ray and Anirban Mitra for sharing their updating of the Varshney-Wilkinson data on Hindu-Muslim
riots. We thank Nina Kaysser, Maya Shivakumar, Peter Gerrish and Paradigm Data Services for
excellent research assistance, and Bradford City Council for sharing software used to decode religion
from name.
1
1.
Introduction
Civil violence, often representing ethnic, religious or racial conflict has been rising
through the past half-century (Gleditsch et al. 2002), but we still have only a limited
understanding of its causes. While there is considerable evidence that the outbreak of civil
conflict results from poverty (e.g. Miguel et al. 2004; Bohlken and Sergenti 2010; Do and Iyer,
2010), the evidence on other potential causes including the importance of social divisions and
political grievances is more controversial (Blattman and Miguel, 2010: p.45). This paper
examines Hindu-Muslim violence in India. Muslims constitute India’s largest religious minority,
and the observed patterns of Hindu-Muslim violence suggest that Muslims are more likely to
have been the victims of such violence (Mitra and Ray, 2010). Since Muslims are also underrepresented in elected office (constituting only 5% of members in the national legislature in
2009, down from nearly 9% in 1980), we investigate whether increasing Muslim political
representation lowers the incidence of religious conflict. We put together unique data on both the
religious identity of politicians and religious conflict for the period 1960-2007, merged at the
state and the district level. We account for the potential endogeneity of Muslim representation by
instrumenting the share of Muslim legislators with the share of Muslim legislators who win in
close elections against Hindus (a strategy similar to that implemented by Lee, 2001 and ClotsFigueras, 2011a, 2011b).
Our study is related to two important streams of the literature. The first is the importance
of political identity. Recent evidence suggests that the identity of political leaders, often
indicated by gender, race or ethnicity—has sizeable influences on policy choices, tending to shift
allocations in favour of the population group that shares the identity of the leader. For instance,
the presence of women in political office in India has been shown to result in more womenfriendly policies (Chattopadhyay and Duflo, 2004; Clots-Figueras, 2011a), better education and
health outcomes (Clots-Figueras, 2011b; Bhalotra and Clots-Figueras, 2011), improved
perceptions of women in leadership positions (Beaman et al, 2009), and greater voice for women
within the criminal justice system (Iyer et al, 2012). Similarly, there is some evidence of “ethnic
favouritism” in India and Kenya (e.g. Pande 2003, Burgess et al. 2011), although Kudamatsu
(2010) finds none in Guinea. There is very little evidence, however, of the relevance of the
religious identity of political leaders.
2
The second stream of research that this paper relates to is the growing literature on the
causes and consequences of civil wars. A major contribution of this paper is to link the literature
on political identity to the literature on conflict. Previous research on civil conflict has tended to
focus upon incidents that result in greater than a 1000 deaths (see Blattman and Miguel, 2010),
but smaller scale ethnic conflict is rife and may have other sorts of causes. We contribute to the
handful of studies that specifically examine ethnic conflicts (Sambanis, 2001), or crimes against
specific sections of society (see, among others, Iyer et al, 2012 and Miguel 2005). If we find that
legislator identity significantly reduces religious violence, this could provide a rationale for
mandated religious group representation in political office, as some political parties in India have
demanded.
Why should the religious identity of the politician matter for religious violence? Our
primary hypothesis is that the presence of a Muslim politician should reduce the incidence of
Hindu-Muslim riots either because the Muslim politician has a greater preference for preventing
such incidents (compared to a non-Muslim) or is electorally more accountable to the Muslim
community. In the latter case, we should note that if the Muslim share of the electorate is large
enough, even a non-Muslim politician will be incentivized to prevent riots. This suggests that the
relationship between politician identity and riot prevention might depend non-linearly on the
proportion of Muslims in the population. We plan to investigate such interaction effects in detail
in future work.
There can be other possible relationships in addition to our primary one. Another
possibility is that the presence of a Muslim in such a prominent political position actually causes
backlash or resentment from the (majority) non-Muslim community, leading to an increase in the
incidence of religious violence. Finally, any riot prevention actions taken by legislators depend
both on their willingness to undertake such actions and their ability to do so. Riot-prevention
actions can include obtaining information on and preventing especially provocative incidents
(such as religious processions passing through localities dominated by the other religion),
resolving disputes between religious groups through peaceful means rather than violence, or
mobilizing the police and other state resources to prevent tensions from escalating into violence.
We therefore hypothesize that a legislator who wants to prevent riots will be more effective if
s/he belongs to the ruling party in the state, if s/he holds particularly influential ministerial
positions, and if lower level officials also have strong incentives to prevent riots. In future
3
research, we will investigate these mechanisms in more detail by collecting data on ministerial
positions and the religion of district-level administrative and police officials.
A further data contribution of our paper is the creation of a unique database on the
religious identity of every electoral candidate in India’s state elections over the period 19602008, coding religion from the candidates’ names. In addition to the effect on religious violence,
these data can be used to study the impact of religious identity on other public policy outcomes
as well. We have also updated the widely-used Varshney-Wilkinson database on Hindu-Muslim
riots (originally from 1950-1995) until the year 2010. The period after 1995 is an important
period to study religious violence: since 1995, India has witnessed significantly faster economic
growth, a secular decline in violent crimes such as murder, and a substantial increase in political
competition.
The rest of the paper is structured as follows: Section 2 describes the institutional setting
in India, the details of our newly collected data, and discusses some preliminary hypothesis about
the impact of politician identity on religious violence. Section 3 presents OLS results. Our
instrumental variables strategy and results are presented in Section 4, and Section 5 concludes
with some thoughts on future research.
2. Religion, Politics and Violence in India
India is a country of considerable religious diversity and the constitution enshrines
secularism. India is home to the world’s third largest Muslim population, with 138.2 million
Muslims recorded in the census of 2001. Muslims constituted 13.4% of the population in the
2001 census and form the single largest religious minority in India. Their share in the population
varies considerably across states, ranging from close to zero to more than 60% in the only
Muslim-majority state of Jammu & Kashmir. Their socioeconomic position is on average similar
to that of the low caste Hindu population, but the latter groups have access to a range of
affirmative action programs in the economic and political spheres, which Muslims do not have.2
A recent report to the Prime Minister’s Office cites survey evidence that Muslims feel
disenfranchised and somewhat marginalized in the allocation of public services and public sector
jobs (Besant and Shariff 2007).
2
The lowest castes (known as Scheduled Castes) and marginalized tribes have specific electoral constituencies set
aside for members of these communities; they also have mandated quotas in higher education and government jobs
and preferential access to secondary schooling.
4
2.1
Religious Identity of Elected State Legislators
We construct a unique data base on the religious identity of state legislators. India is a federal
country, with a parliamentary system of government at both the federal and state levels.
Elections are held every five years, on a first-past-the-post system in single-member
constituencies. Elections are very competitive in India, with more than 100 parties participating
in the 2009 national elections. There are no major “Muslim-only” parties, but some parties
appeal more to Muslims than others.
We obtained data on state legislative elections from the Election Commission of India
and they contain information on the name, sex, party affiliation and votes obtained by every
candidate in every election held in India since Independence. We used the legislator names to
infer religious identity. To minimize measurement error, we had two independent teams working
on the classification of legislator names. The first team used a software program called Nam
Pehchan, which was able to classify about 72% of the names, and manually classified the rest. A
second (India-based) team performed the whole classification manually using their judgment
gained from prior work with Election Commission files. The two teams agreed on more than
95% of the names; disagreements between the two teams’ classification were resolved by the
authors.3
In this draft, we focus on data from the period 1980-2007, for 17 major states of India
which account for over 95% of the total population.4 Over this period, electoral constituency
boundaries remained fixed, and therefore we do not have to worry about concerns such as
gerrymandering which might affect the proportion of Muslims elected to state legislatures.5 In
future research, we will also include the period from 1960-1980.
The share of Muslim legislators in the country has remained between 7% and 8.5% over
the past three decades, which is considerably less than their population share of 13% (Figure 1).
To the best of our knowledge, this is the first estimate of the proportion of Muslims in India’s
state legislatures. On average, between 8 and 9% of all candidates in state elections are Muslims;
3
In doubtful cases, the default was to assign a “non-Muslim” classification; hence, our method is likely to
undercount the number of Muslim state legislators.
4
In 2001, the states of Bihar, Madhya Pradesh and Uttar Pradesh were split into two. We aggregate the data from
the split states to the original unsplit boundaries to maintain a balanced panel data set.
5
For state-level results, we use the period 1980-2008, since all redistricting was strictly within state and the total
number of constituencies did not change. We should also note that there are no reservations for Muslims in the
political sphere.
5
this has remained fairly steady over the past three decades. The share of Muslim legislators, on
the other hand, shows some variation over time, being notably lower around 1992 when the
country witnessed a major wave of religious violence.
Figure 1: Percentage of Muslim state legislators and Muslim state election candidates
1980-2008 (17 Major States)
Our data set also reveals that Muslims are systematically under-represented in state
legislatures, compared to their population share, in almost every state (Figure 2). The only
exception is the Muslim-majority state of Jammu & Kashmir, where the percentage of Muslim
legislators closely reflects the population proportion.
6
Figure 2: Percentage of Muslim population, Muslim state election candidates and Muslim
legislators across major Indian states, 1980-2008
There does not appear to be any specific pattern in the fraction of Muslim candidates
compared to Muslim legislators. In some states, the fraction of Muslim legislators is somewhat
smaller than the fraction of Muslim candidates, but the reverse holds true in other states. In our
district level data, the fraction of Muslim candidates is highly correlated with the fraction of
Muslim legislators (correlation=0.90).
Which parties are more likely to have Muslim legislators? Compared to non-Muslim
legislators, we find that Muslim legislators are significantly less likely to belong to the Bharatiya
Janata Party (BJP), and more likely to belong to the Communist parties and other leading state
parties (Figure 3). This is exactly what we would expect, given the BJP’s strong rhetoric of
“Hindutva” over much of this period and its links with explicitly Hindu organizations such as the
Vishwa Hindu Parishad (VHP). The share of legislators belonging to the Indian National
Congress (INC) is similar across Muslims and non-Muslims.
7
Figure 3: Party affiliations of Muslim and non-Muslim state legislators 1980-2008
2.2 Data on Religious Violence
We updated a data base on Hindu-Muslim violence originally put together by Ashutosh
Varshney and Steve Wilkinson (Varshney and Wilkinson, 1995). The original data set was based
on newspaper articles published in The Times of India (Mumbai edition), a national newspaper
over the period 1950-1995. This was the first systematic data set on religious violence in India
over time, and has been used in several previous academic studies (discussed in more detail
below). We extend this data base until 2010, using the same methodology as the original data
8
base (as documented in Varshney, 2002, Appendix 3), and building upon the work of other
researchers (notably Mitra and Ray, 2010, who extend the data base until 2000). In the empirical
analysis, we will use the data until 2007 to match with the time span of the electoral data (until
2008 for state level elections).
The original Varshney-Wilkinson data set has been widely used to examine the
determinants of religious violence in India. Previous work has identified several important
factors which contribute to the prevalence or prevention of religious violence. Varshney (2002)
highlights the importance of consociational links i.e. the strength of inter-religious civil society
organizations, based on shared business or economic interests. Jha (2008) also highlights the
importance of historically determined economic complementarities between Hindus and
Muslims. In particular, he shows that cities which used to be medieval ports have a greater
degree of such economic complementarity and a lower incidence of riots. Bolhken and Sargenti
(2010) find that a 1% increase in state-level GDP growth reduces the incidence of riots by 5%;
their estimation relies on rainfall shocks as an exogenous determinant of state-level GDP growth.
Mitra and Ray (2010) show that differential economic growth across Hindus and Muslims can
generate conflict, due to resentment over relative economic well-being; their analysis strongly
suggests that Hindus are the aggressors in such riots (Chua 2003 discusses the role of differential
economic growth in ethnic violence more generally).
Remarkably, there has been relatively little work on documenting the effects of
politicians on Hindu-Muslim violence, though leading politicians have sometimes been
implicated in such incidents.6 The major focus on politics has been the work of Wilkinson
(2004), who shows that greater political competition results in a lower incidence of riots against
Muslims.7 In our analysis, we will control for some of the variables identified by these previous
researchers, most importantly the “effective number of parties” used as a proxy for electoral
competition in the state.8
6
For instance, Chief Minister Narendra Modi has been accused of gross negligence and failure to prevent violence
against Muslims during the Gujarat riots of 2002.
7
Wilkinson finds that the proportion of Muslims in the state cabinet has no significant relationship with the
incidence of Hindu-Muslim riots, but does not examine the role of overall Muslim representation in the legislature,
or the presence of Muslim legislators in specific districts.
8
This is the primary variable used by Wilkinson (2004). See Chibber and Nooruddin (2004) for a definition of this
variable, which they show to be a significant determinant of public service provision by state governments.
9
The updated V-W data set shows some interesting trends in the post-1995 period.9 The
incidence of Hindu-Muslim riots is lower in the post-1995 period compared to the period 19801995, except for the upsurge in violence in 2002, which was concentrated in the state of Gujarat
(Figure 4). A similar trend is visible for the number of people killed in the riots.10 This overall
decline in the incidence of religious violence is in line with the overall decline in other violent
crimes in India (such as murders) in the period after 1990.
Figure 4: Number of Hindu-Muslim riots and riot deaths in India, 1980-2010
This decline in the incidence of Hindu-Muslim violence in the post-1995 period is
observed in almost all the states (Figure 5). However, there remains a strong correlation between
the incidence of riots in the two periods i.e. states which witnessed a high level of riots before
1995 also witnessed a high level of riots after 1995. As in the original V-W data set, the most
riot-prone states are Gujarat, Maharashtra and Uttar Pradesh.
We have also updated the data for the years 1994 and 1995 from the Times of India Mumbai edition for
consistency over time; these entries were based on the Delhi edition in the original data set.
10
The data for the number of people killed is an underestimate, especially for the Gujarat riots of 2002, where a
large number of missing people (several hundred) were declared dead only after seven years.
9
10
Figure 5: Number of Hindu-Muslim riots across Indian states 1980-2010
3. Legislator Identity and Religious Violence: OLS Estimates
We will examine the effect of legislator identity on religious violence using regression
analysis at both the state and district levels. At the state level, our main empirical specification is
as follows:
(1)
Log (0.1 + NRiotsit) = ai + bt + dMuslimit + fXit + eit
where NRiotsit is the number of Hindu-Muslim riots in state i and year t; ai is a state
fixed effect to control for all time-invariant state characteristics, bt is a time fixed effect to
control for nationwide changes in year t, Muslimit is the proportion of Muslim legislators in the
state in year t, Xit is a vector of other time-varying state characteristics and eit is an error term.
Since almost half of all state-year observations in our data have zero riots, we use the log
transformation above to avoid dropping these observations. Another way to deal with this highly
skewed count data is to run a negative binomial specification with the number of riots (NRiotsit)
as the dependent variable. Finally, we will also use a linear probability model with the dependent
variable as a dummy for whether any riots occurred in state i in year t (RiotDummyit).
11
We exclude the state of Jammu & Kashmir from our analysis, because of several factors.
It is the only Muslim-majority state in the country, which means that the interpretation of HinduMuslim riots as an attack by the majority community on the minority community may not be
valid for this state. Also this state is the scene of a long-running territorial dispute between India
and Pakistan and such an international dimension to religious relations makes this state very
different from the rest of the country. Finally, the state was under direct federal administrative
control (“President’s Rule”) for several years, which means that state legislators had much less
of a role to play.
In all state-level regressions, we cluster standard errors at the level of the state and
electoral cycle.11 We include controls for time-varying demographic and economic
characteristics of the states (proportion Muslim, proportion urban, proportion female, proportions
belonging to Scheduled Castes and Scheduled Tribes, proportion engaged in farming, per capita
state domestic product), as well as the effective number of parties as a proxy for electoral
competition.
We also estimate district level regressions using a similar specification as (1) above:12
(2)
Log (0.1 + NRiotsidt) = Ait + dMuslimidt + fXidt + eidt
where NRiotsidt is the number of riots occurring in district d of state i in year t, Ait is a state-year
fixed effect which proxies for all state level happenings in state i and year t, Muslimidt is the
fraction of Muslim legislators elected to the state assembly from district d, and Xidt are other
controls. Each administrative district contains 5-10 electoral constituencies on average. Standard
errors are clustered at the level of district. We also run a robustness check where we include
district fixed effects instead of state-year fixed effects (note that this specification assumes that
the occurrence of riots in a given district is independent of their occurrence in any other district).
In our state-level OLS regressions, we do not find any statistically significant relationship
between the fraction of Muslim legislators and the occurrence of Hindu-Muslim riots (Table 2,
11
We also run a robustness exercise where we cluster standard errors at the level of the state, even though there are
only 16 major states.
12
In most cases, it is not possible to identify the exact electoral constituency where the riot occurred. We therefore
aggregate our data on legislator identity to the district level for analysis. In future work, we will also conduct the
analysis at the level of the town or city (a key finding in Varshney, 2002, is that the vast majority of religious
violence happens in urban areas).
12
columns 1-3). This holds true regardless of the inclusion of state and year fixed effects (Table 2,
column 2) or additional controls (Table 2, column 3).
Our district-level analysis also yields inconclusive results. The specification using stateyear fixed effects shows a significant positive relationship between the presence of Muslim
legislators and Hindu-Muslim riots (Table 3, column 1). But this appears to be driven by
unobservable district characteristics: once we control for district fixed effects, the relationship is
negative, albeit statistically insignificant (Table 3, column 2). Another way to control for omitted
characteristics which determine whether Muslims get elected to state legislatures is to restrict our
sample to places which had “close” elections i.e. places where the vote margin between a
Muslim and a non-Muslim is less than 5% (or 3%). This restricted sample also does not show
any strong positive relationship (Table 3, columns 3 and 4).
4. Legislator Identity and Religious Violence: Instrumental Variable Estimates
In the analysis described above, the fraction of Muslim legislators is potentially endogenous.
There might be omitted factors which determine both the presence of Muslim legislators and the
occurrence of riots (e.g. the relative economic progress of the two communities, changing norms
about minority engagement in politics, changing relations between religious groups in the local
area); in addition, the occurrence of religious violence itself might change the incentives for
Muslims to participate in politics. Our OLS estimates therefore can be biased in unknown
directions.
4.1
Exploiting the Presence of “Close Elections”
We therefore implement an instrumental variable strategy, where we instrument the
fraction of Muslim legislators with the fraction of Muslim legislators who are elected in close
elections against a non-Muslim (MuslimCloseit). The identification assumption here is that the
outcomes of close elections are decided on an essentially random basis.13 Of course, places in
which a Muslim competes in a close election against a non-Muslim may be different in
unobservable ways from places in which Muslim candidates are not competitive. To account for
13
A similar instrument was used by Clots-Figueras (2011a, 2011b) and Bhalotra and Clots-Figueras (2011) to
estimate the effect of female legislators.
13
these differences we control for the fraction of seats in which close elections between a Muslim
and a non-Muslim are observed (Closeit). Our two-stage specification is therefore as follows:
(3)
(4)
Muslimit = αi + βt + λMuslimCloseit + θCloseit + μXit + eit
Log (NRiotsit) = ai + bt + dMuslim*it + kCloseit + fXit + eit
where Muslim*it represents the predicted values from the first stage regression in equation (3).
To demonstrate the validity of our instrumental variable strategy, we show that the
proportion of Muslims winning in close elections is a strong and significant predictor of the
overall proportion of Muslim legislators. Table 4 reports the first stage regressions for our
instrumental variables strategy, based on specification (3) for district level. We obtain a positive
and statistically significant estimate of λ across all specifications: with state-year fixed effects
(column 1), district and year fixed effects (column 2), when restricting to areas which had close
elections (column 3), and when changing the margin for defining close elections from 5% to 3%
(columns 4-6). Furthermore, the magnitude of the λ coefficient remains extremely stable across
all these specifications.
We now examine whether places which feature close elections between Muslims and
non-Muslims are different from places which do not. This is important, because our instrumental
variables regression is essentially a latent average treatment effect (LATE) mainly for this subset
of districts. We find that districts which feature close elections are more populous, more urban
and have a higher population share of Muslims than places which do not feature close elections
(Table 5, panel A). This is not surprising, since Muslims are more likely to live in urban areas
which have a higher population. In other words, we are more likely to see close elections in
places where Muslims are more likely to contest elections, namely places where their population
share is higher. Interestingly, these places also have a significantly higher occurrence of HinduMuslim riots (Table 5, panel A). This means that focusing on places with close elections (where
our instrumental variables strategy is most effective) involves focusing on places with the
greatest incidence of religious violence i.e. our IV estimates are the LATE for a highly relevant
subset of districts. When using the full sample for our IV estimates, we will therefore explicitly
control for the presence of close elections, since such places are different from others. However,
14
the more precise instrumental variable estimates are obtained when restricting our focus to only
such places.
Our basic identification assumption is that, conditional on there being a close election,
places where Muslims win in close elections are similar to places where non-Muslims win the
close elections. We show that constituencies where Muslims win close elections are very similar
to those where non-Muslims win in terms of many political characteristics: the total number of
candidates, the share of Muslim candidates, and the vote share of the eventual winner (Table 5,
panel B). However, Muslim winners are significantly more likely to belong to the Congress
(INC) and less likely to belong to the BJP. This is similar to the overall party affiliation patterns
of Muslim candidates (Figure 3). However, it means that the presence of Muslim legislators
might be proxying for a general Congress policy agenda rather than a BJP one (this is akin to the
hypothesis of Muslim legislators having different preferences which leads them to join different
parties).
4.2 Instrumental Variables Estimates
We begin by discussing briefly the state-level instrumental variable results. In contrast to the
OLS regressions, our instrumental variable regressions show a consistent negative relationship
between the proportion of Muslim legislators and the incidence of Hindu-Muslim riots (Table 2,
columns 4-6). However, none of the estimated coefficients are statistically significant, suggesting
either that this relationship is not a strong one, or that the state level data are too aggregated and
noisy to estimate the relationship precisely. We proceed to conduct our analysis at the district
level.
We find some evidence that the presence of a Muslim legislator is associated with a
significant decline in the occurrence of Hindu-Muslim riots (Table 6). However, this result is not
universal. Instrumental variable estimates for the full sample do not show any significant
relationship, even after adding the occurrence of close elections as a control variable (Table 6,
columns 1 and 2). But when we restrict our focus to only the places where close elections occur,
we do see a very strong and significant relationship (Table 6, column 3). An exactly similar
pattern holds when we redefine close elections to those with a victory margin of 3% or less
instead of 5% (Table 6, columns 4-6). This suggests that Muslim legislators are likely to have a
big impact in certain kinds of places, namely places where there is significant electoral
15
competition between Muslims and non-Muslims. As documented in Table 5, these are places
which have a higher Muslim share and a higher occurrence of riots. We should note that these
instrumental variables estimates, if robust to further analysis, suggest potentially large effects of
minority political representation. The estimates in Table 6, Column 2 suggest that a one standard
deviation increase in the proportion of Muslim legislators could result in a 2.9 percentage point
decline in the probability of a Hindu-Muslim riots (i.e. a decline of 0.12 standard deviations) and
a 9% decline in the number of riots.
5. Conclusions and Further Research
This paper finds that raising the share of Muslim leaders in state assemblies in India results in
a substantial decline in the incidence of Hindu-Muslim conflict in the period 1980-2007.
However, this decline is observed only in certain places, namely those which feature a high
degree of political competition between Muslims and non-Muslims. This is consistent with
theories of political identity (eg. Besley and Coate, 1997), insofar as Muslims in India value
security more than Hindus (Mitra and Ray, 2010; Wilkinson, 2004). It also suggests a cause of
conflict- and hence a solution for the control of conflict- that has not been previously considered.
In work in progress we are extending the data back to 1960, testing robustness of these results,
investigating heterogeneity in this relationship and extending the analysis to look at public goods
other than security.
16
References
Beaman, Lori, Raghabendra Chattopadhyay, Esther Duflo, Rohini Pande and Petia Topalova.
2009. “Powerful Women: Does Exposure Reduce Bias?” Quarterly Journal of Economics
124(4): 1497-1540.
Bolhken, Anjali and Ernest Sergenti. 2010. “Economic growth and ethnic violence: An empirical
investigation of Hindu–Muslim riots in India.” Journal of Peace Research 47(5): 589-600.
bridge.
Chattopadhyay, Raghabendra and Esther Duflo. 2004. “Women as Policy Makers: Evidence
from a Randomized Policy Experiment in India.” Econometrica 72(5): 1409-1443.
Clots-Figueras, Irma. 2011a. “Women in Politics: Evidence from the Indian States.” Journal of
Public Economics 95: 664-690.
Clots-Figueras, Irma. 2011b. “Are Female Leaders Good for Education? Evidence from India.”
American Economic Journal: Applied Economics, Forthcoming.
Chua, A. (2003). World On Fire. How Exporting Free Market Democracy Breeds Ethnic Hatred
and Global Instability, New York: Anchor Books.
Collier, P. and A. Hoeffler (2004). “Greed and grievance in civil war: On economic causes of
civil war”, Oxford Economic Papers, 56 : 563-595
Gleditsch, Nils Petter, Peter Wallensteen, Mikael Eriksson, Margareta Sollenberg, and Håvard
Strand. 2002. “Armed Conflict 1946–2001: A New Dataset.” Journal of Peace Research, 39(5):
615–37.
Iyer, Lakshmi, Anandi Mani, Prachi Mishra and Petia Topalova. 2012. “The Power of Political
Voice: Women's Political Representation and Crime in India.” American Economic Journal:
Applied Economics, Forthcoming.
Jha, Saumitra. 2008. “Trade, institutions and religious tolerance: evidence from India.” Stanford
GSB Research Paper Series.
Mitra, Anirban and Debraj Ray. 2010. “Implications of an economic theory of conflict: HinduMuslim Violence in India.” NYU Working Paper, https://files.nyu.edu/am2840/public/hm.pdf,
accessed January 2012.
Ray, Debraj and Joan Esteban and Laura Mayoral. 2011. “Ethnicity and Conflict: An Empirical
Investigation.” American Economic Review, forthcoming.
Varshney, A. (2002): Ethnic Conflict and Civic Life: Hindus and Muslims in India. Yale
University Press, New Haven.
17
Varshney, A., and S. I. Wilkinson (2004): “Varshney Wilkinson Dataset on Hindu- Muslim
Violence in India, Version 2."
Wilkinson, S. I. (2004): Votes and violence: electoral competition and ethnic riots in India,
Cambridge studies in comparative politics. Cambridge University Press, Cam--(2005a): “Introduction," in (Wilkinson 2005b), chap. 1, pp. p.1{34.(ed.)
--(2005b): (ed.) Religious politics and communal violence in India. Oxford University Press,
New Delhi.
--(2005c): “Wilkinson dataset on Hindu-Muslim Violence in India, 1600-1950,"
18
Table 1: Summary statistics
#obs
Panel A: Electoral variables
Proportion of seats won by Muslims
Proportion of Muslim candidates
Proportion of seats with close elections between Muslims and non-Muslims (5% vote margin)
Mean
s.d.
Min
Max
464
464
0.063
0.071
0.060
0.055
0
0
0.29
0.27
464
0.025
0.023
0
0.11
464
0.015
0.014
0
0.07
464
0.012
0.013
0
0.07
464
0.007
0.008
0
0.05
Panel B: Incidence of Hindu-Muslim riots
Number of Hindu-Muslim riots
Log (0.1 + number of riots)
Dummy for occurrence of Hindu-Muslim riots
464
464
464
2.278
-0.646
0.502
6.273
1.773
0.501
0
-2.303
0
100
4.606
1
Part C: Other variables
Proportion of Muslim population
Proportion of urban population
Proportion male
Proportion of Scheduled Castes
Proportion of Scheduled Tribes
State election year
State GDP per capita
Effective number of parties
464
464
464
464
464
464
464
464
0.111
0.248
0.516
0.165
0.079
0.224
0.161
2.699
0.084
0.095
0.012
0.059
0.072
0.417
0.082
1.038
0.008
0.076
0.483
0.066
0
0
0.037
1.353
0.327
0.498
0.538
0.291
0.233
1
0.456
6.409
Proportion of seats with close elections between Muslims and non-Muslims (3% vote margin)
Proportion of seats won by Muslims in close elections against non-Muslims (5% vote margin)
Proportion of seats won by Muslims in close elections against non-Muslims (3% vote margin)
Summary statistics are for 16 major states over 1980-2008.
Table 2
Does the presence of Muslim legislators affect the incidence of Hindu-Muslim riots?
(1)
(2)
Panel A: Number of riots (negative binomial specification)
(3)
Proportion of Muslim legislators
-4.649
[3.025]
-0.038
[3.751]
-3.024
[4.818]
Observations
Log Likelihood
464
-849.87
464
-640.04
464
-634.44
(4)
(5)
(6)
Panel B: Log (0.1 + Number of riots)
Proportion of Muslim legislators
-0.096
[2.009]
1.013
[4.470]
-0.432
[4.955]
-4.482
[15.661]
-7.314
[7.004]
-10.174
[13.919]
-15.177
[11.498]
-10.174
[16.248]
-15.177
[11.224]
0
464
0.59
464
0.6
464
0.6
464
0.6
464
0.6
464
Proportion of close elections
between Muslims and non-Muslims
R-squared
Observations
Panel C: Whether any Hindu-Muslim riot occurred in the state (linear probability model)
Proportion of Muslim legislators
0.307
[0.561]
0.349
[1.417]
0.276
[1.533]
0.949
[4.524]
-1.697
[2.276]
-0.223
[4.032]
-4.284
[3.506]
-0.223
[4.821]
-4.284
[3.154]
0.48
464
OLS
0.48
464
IV (5%)
0.49
464
IV (3%)
0.49
464
IV (3%)
Proportion of close elections
between Muslims and non-Muslims
R-squared
Observations
Specification
0
0.47
464
464
OLS
OLS
Panel D: First stage
Proportion of Muslim legislators who
won in close elections against a non-Muslim
R-squared
Observations
State FE
Year FE
Controls
Clustering
Y
Y
stateelection
stateelection
Y
Y
Y
stateelection
0.634 *** 0.952 ***
[0.219]
[0.216]
0.97
0.97
464
464
Y
Y
Y
Y
Y
Y
stateelection state-election
0.952 ***
[0.229]
0.97
464
Y
Y
Y
state
Controls include demographic variables, state domestic product per capita and the effective number of political parties.
Demographic variables include proportion Muslim, proportion urban, proportion male, proportion of Scheduled Castes and
Scheduled Tribes and the proportion engaged in farming.
Standard errors in paranthesis. ***significant at 1%, ** significant at 5%, * significant at 10%
Table 3
District level relationship between Muslim legislators and Hindu-Muslim
riots (OLS)
(1)
(2)
Panel A: Log (0.1 + Number of riots)
Proportion of Muslim legislators
R-squared
Observations
(3)
(4)
0.389 **
[0.162]
-0.036
[0.118]
-0.01
[0.177]
-0.041
[0.212]
0.08
10220
0.25
10220
0.18
2067
0.2
1442
Panel B: Whether any Hindu-Muslim riot occurred in the state (linear probability model)
Proportion of Muslim legislators
R-squared
Observations
State*year fixed effects
District and year fixed effects
Clustering
0.151 ***
[0.058]
-0.006
[0.047]
0.01
[0.064]
0.006
[0.073]
0.07
10220
0.22
10220
0.17
2067
0.19
1442
Y
Y
district
district
Y
district
Y
district
Notes: Column (3) and (4) restrict the sample to observations where there was at least one close
election between a Muslim and a non-Muslim, with a vote margin <=5% and <=3%
respectively.
Standard errors in paranthesis. ***significant at 1%, ** significant at 5%, * significant at 10%
Table 4
Instrumental variable strategy: First stage
(1)
(2)
(3)
(4)
Dependent Variable: Proportion of Muslim legislators in district
(5)
(6)
Proportion of Muslim legislators who
won in close elections against a non-Muslim
0.870 *** 0.846 *** 0.867 *** 0.907 *** 0.854 *** 0.819 ***
[0.081]
[0.066]
[0.058]
[0.104]
[0.085]
[0.092]
Proportion of close elections between
0.278 *** -0.209 ***
[0.076]
[0.046]
Muslim and non-Muslim
R-squared
Observations
State*year fixed effects
District and year fixed effects
Vote margin for close elections
Clustering
0.33
10220
0.8
10220
Y
5%
district
Y
5%
district
0.262 ** -0.276 ***
[0.102]
[0.055]
0.38
2067
0.23
10220
Y
Y
5%
district
3%
district
0.78
10220
0.35
1442
Y
Y
3%
district
3%
district
Notes: Columns (3) and (6) restrict the sample to observations where there was at least one close election between a
Muslim and a non-Muslim, with vote margins as listed.
Standard errors in paranthesis. ***significant at 1%, ** significant at 5%, * significant at 10%
Table 5
Comparing Areas Where Muslims won in Close Elections to Other Areas
Panel A: Comparing Districts which had Close Elections between Muslims and non-Muslims
Districts which
had any close
Districts which had
elections
no close elections
between
between Muslims
Muslims and
and non-Muslims
non-Muslims
Demographic Characteristics from Census 2001
Total population
Fraction Muslim
Fraction urban
Fraction women
Fraction literate
Fraction of females literate
Number of Hindu-Muslim riots (1980-2007)
Dummy for any Hindu-Muslim riot (1980-2007)
Number of districts
2088816
0.054
0.214
0.484
0.549
0.449
0.050
0.035
172
3251808
0.163
0.253
0.484
0.534
0.439
0.141
0.079
193
Difference
1162992
0.109
0.039
0.000
-0.015
-0.010
0.091
0.045
***
***
**
***
***
Panel B: Comparing Constituencies where Muslims won in Close Elections to those in which non-Muslims Won
Sample: Constituencies where there was a close election between a Muslim and a non-Muslim (vote margin <=5%)
Non-Muslim won in
close election
Muslim won in
close election
356
13.640
0.322
0.082
0.366
0.024
0.177
0.320
0.087
355
12.727
0.358
0.111
0.371
0.024
0.349
0.031
0.085
Observations
Number of candidates
Fraction Muslim candidates
Fraction female candidates
Vote share of winner
Winner's vote margin
Winner is from INC
Winner is from BJP
Winner is from Communist parties (CPI/CPM)
***significant at 1%, ** significant at 5%, * significant at 10%
Difference
-0.914
0.036
0.029
0.005
0.001
0.172 ***
-0.289 ***
-0.003
Table 6
District level relationship between Muslim legislators and Hindu-Muslim riots (Instrumental
variable estimates)
(1)
(2)
Panel A: Log (0.1 + Number of riots)
Proportion of Muslim legislators
-0.277
[0.313]
0.017
[0.288]
0.491
[0.305]
0.07
10220
0.085
[0.142]
0.25
10220
(3)
(4)
(5)
(6)
-0.863 ***
[0.328]
-0.311
[0.435]
-0.345
[0.391]
-1.158 **
[0.583]
0.16
2067
0.555
[0.404]
0.07
10220
0.063
[0.204]
0.25
10220
0.17
1442
Proportion of close elections between
Muslim and non-Muslim
R-squared
Observations
Panel B: Whether any Hindu-Muslim riot occurred in the state (linear probability model)
Proportion of Muslim legislators
-0.074
[0.114]
0.044
[0.106]
Proportion of close elections between
Muslim and non-Muslim
0.172
[0.108]
0.022
[0.053]
R-squared
Observations
0.07
10220
0.21
10220
State*year fixed effects
District and year fixed effects
Vote margin for close elections
Clustering
Y
5%
district
Y
5%
district
-0.277 **
[0.116]
-0.07
[0.152]
-0.066
[0.144]
0.172
[0.136]
0.003
[0.075]
0.15
2067
0.06
10220
0.21
10220
Y
Y
5%
district
3%
district
-0.351 *
[0.205]
0.17
1442
Y
Y
3%
district
3%
district
Notes: Columns (2) and (4) restrict the sample to observations where there was at least one close election between a
Muslim and a non-Muslim, with vote margins as listed.
Standard errors in paranthesis. ***significant at 1%, ** significant at 5%, * significant at 10%