Do parties matter for ethnic violence? Evidence from India * Gareth Nellis † Yale University Michael Weaver ‡ Yale University Steven C. Rosenzweig § Yale University May 1, 2016 Abstract Ethnic group conflict is among the most serious threats facing young democracies. In this paper, we investigate whether the partisanship of incumbent politicians affects the incidence and severity of local ethnic violence. Using a novel application of the regression-discontinuity design, we show that as-if random victory by candidates rep- resenting India’s Congress party in close state assembly elections between 1962 and 2000 reduced Hindu-Muslim rioting. The effects are large. Simulations reveal that had Congress lost all close elections in this period, India would have experienced 11 percent more riots. Additional analyses suggest that Congress candidates’ dependence on local Muslim votes, as well as apprehensions about religious polarization of the electorate in the event of riots breaking out, are what drive the observed effect. Our findings shed new light on parties’ connection to ethnic conflict, the relevance of partisanship in developing states, and the puzzle of democratic consolidation in ethnically-divided societies. * We thank the anonymous reviewers, Steven Wilkinson, Tariq Thachil, and seminar partic- ipants at Yale, UC Berkeley, and the 2015 American Political Science Association conference for valuable comments. We are especially grateful to Francesca Jensenius for sharing election data with us. † Corresponding author. Ph.D. candidate, Yale University, Department of Political Science, New Haven, CT. Email: [email protected]‡ Ph.D. candidate, Yale University, Department of Political Science, New Haven, CT. Email: [email protected]§ Ph.D. candidate, Yale University, Department of Political Science, New Haven, CT. Email: [email protected]
66
Embed
Do parties matter for ethnic violence? Evidence …...Do parties matter for ethnic violence? Evidence from India Gareth Nellisy Yale University Michael Weaverz Yale University Steven
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Do parties matter for ethnic violence? Evidence fromIndia∗
Gareth Nellis†
Yale UniversityMichael Weaver‡
Yale UniversitySteven C. Rosenzweig§
Yale University
May 1, 2016
Abstract
Ethnic group conflict is among the most serious threats facing young democracies.In this paper, we investigate whether the partisanship of incumbent politicians affectsthe incidence and severity of local ethnic violence. Using a novel application of theregression-discontinuity design, we show that as-if random victory by candidates rep-resenting India’s Congress party in close state assembly elections between 1962 and2000 reduced Hindu-Muslim rioting. The effects are large. Simulations reveal that hadCongress lost all close elections in this period, India would have experienced 11 percentmore riots. Additional analyses suggest that Congress candidates’ dependence on localMuslim votes, as well as apprehensions about religious polarization of the electorate inthe event of riots breaking out, are what drive the observed effect. Our findings shed newlight on parties’ connection to ethnic conflict, the relevance of partisanship in developingstates, and the puzzle of democratic consolidation in ethnically-divided societies.
∗We thank the anonymous reviewers, Steven Wilkinson, Tariq Thachil, and seminar partic-ipants at Yale, UC Berkeley, and the 2015 American Political Science Association conferencefor valuable comments. We are especially grateful to Francesca Jensenius for sharing electiondata with us.†Corresponding author. Ph.D. candidate, Yale University, Department of Political Science,
New Haven, CT. Email: [email protected]‡Ph.D. candidate, Yale University, Department of Political Science, New Haven, CT. Email:
[email protected]§Ph.D. candidate, Yale University, Department of Political Science, New Haven, CT. Email:
Model SpecificationsDistrict fixed effects Yes Yes No NoDependent Variablet−1 No No Yes YesYear fixed effects Yes Yes Yes YesState time trends No Yes No Yes
Standard errors clustered at the district level reported in parentheses. Nis across 307 districts.
35
Figure 1: Percentage of Muslim votes going to INC and BJS/BJP in Lok Sabha elections,1957–2009
1960 1970 1980 1990 2000 2010
020
4060
8010
0
Year
Vot
e S
hare ●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
● ●
● ●●●● ●●
●●
●
● ● ●
●● ● ●
●
●
●
●
●
●●
●
●
●
●
●●
●●
INC OverallINC among Muslims
BJP/BJS among Muslims
Notes: Data in green and yellow based on self-reports of Muslim voting found in varioussurveys conducted by the Centre for the Study of Developing Societies and newly compiledby the authors. The data are noisy due to the relatively small number of Muslim respon-dents in each survey. Data in gray shows the actual vote share received by the Congressin each election, according to the Election Commission of India. Lines represent Lowesssmoothing curves. In the post-Emergency election of 1977, the Bharatiya Jana Sangh(BJS) merged into the Janata Party, which accounts for the jump in Muslim support inthat year. The BJS was reconstituted as the BJP in 1980.
36
Figure 2: Instrumental variables estimates of the effect of CongSeatShare on riot outcomes
β: Effect of Congress Seatshare
Mod
el S
peci
ficat
ion
Saturated Control
Linear Control
−0.5 0.0 0.5
●
●
Pr(Any Riot)
Saturated Control
Linear Control
−4 −2 0 2 4
●
●
Log Days of Rioting
Saturated Control
Linear Control
●
●
Log Riot Casualties
Saturated Control
Linear Control
●
●
Log Riot Count
Notes: This presents coefficient estimates from IVLS regressions of logged or binary riotoutcomes on CongSeatShare, using the approach described in the Data and identificationsection. Bars represent 95% confidence intervals using robust standard errors clusteredat the district level. Saturated or Linear indicates how the control, CongCloseProp, isspecified in the model. N for all regressions is 2871, across 315 districts.
37
Figure 3: Difference in means estimates of the effect of CongSeatShare on riot outcomes
Mean
Gro
up
Control
Treated
Difference
−0.2 −0.1 0.0 0.1 0.2
●
Pr(Any Riot)
Control
Treated
Difference
−0.05 0.00 0.05
●
Days of Rioting
Control
Treated
Difference ●
Riot Casualties
Control
Treated
Difference ●
Riot Count
Notes: This presents the effects of Congress victory for a restricted sample of district-election cycle observations in which a single close election took place between a Congressand a non-Congress candidate. “Control” indicates Congress loss, and “Treated” indi-cates Congress victory. Bars represent 95% confidence intervals using robust standarderrors clustered at the district level. For riot count variables, we estimate differences oflog-transformed counts. For ease of interpreting these differences, the means and theirdifference were transformed back to their original (un-logged) scale by taking the expo-nent. This re-scaling explains why confidence intervals are asymmetric around the pointestimates. N is 644 across 263 districts.
38
Figure 4: Heterogeneous effects of CongSeatShare
β: Effect of Congress Seatshare
Con
ditio
n
YES
NO
−1.0 −0.5 0.0 0.5 1.0
●
●
High Muslim Pop.Any Riot
−1.0 −0.5 0.0 0.5 1.0
●
●
High Party FractionalizationAny Riot
YES
NO
−6 −4 −2 0 2 4 6
●
●
High Muslim Pop.Log Riot Count
−6 −4 −2 0 2 4 6
●
●
High Party FractionalizationLog Riot Count
Notes: This figure presents coefficient estimates from IVLS regressions of logged and bi-nary riot outcomes on CongSeatShare, using the approach described in the Data andidentification section. Bars represent 95% confidence intervals using robust standard er-rors clustered at the district level. N for high and low Muslim population are 1427 and1372, respectively. N for high and low party fractionalization are both 1397.
39
Online supplementary appendixA Identification
Table A1: χ2 Test of Out-comes in Close Elections
Observed Losses Wins
Percent 48.8 51.2Number (536) (563)
χ21 = 0.66, P = 0.42
Table A2: Correlation of Close Elec-tions over Time
Pearson correlation coefficients.N is 2556 for (T − 1), 2246 for(T − 2), and 1939 for (T − 3).
1
Fig
ure
A1:
Map
sillu
stra
ting
const
ruct
ion
ofri
ght-
han
d-s
ide
elec
tora
lva
riab
les
-0.4
6% +20.
31%
+1.9
7%
+3.9
4%-3
.68%
-6.5
1%
+0.1
4%-0.2
7%+7
.52%
+29.
07%
-29.
05%In
depe
nden
t var
iabl
e =
6/11
-0.4
6% +20.
31%
+1.9
7%
+3.9
4%-3
.68%
-6.5
1%
+0.1
4%-0.2
7%+7
.52%
+29.
07%
-29.
05%In
stru
men
tal v
aria
ble
= 1
/11
-0.4
6% +20.
31%
+1.9
7%
+3.9
4%-3
.68%
-6.5
1%
+0.1
4%-0.2
7%+7
.52%
+29.
07%
-29.
05%
Con
trol v
aria
ble
= 3/
11
Notes:
Map
dis
pla
ys
stat
eas
sem
bly
elec
tion
sin
Agr
aD
istr
ict,
Utt
arP
rad
esh
,in
1985
.N
um
ber
sre
pre
sent
the
Con
gres
sca
nd
idat
e’s
mar
gin
ofvic
tory
orlo
ssin
each
ML
Aco
nst
itu
ency
.T
he
num
ber
ofd
ark
shad
edar
eas
rep
rese
nts
the
nu
mer
ato
rof
the
vari
able
,w
hil
eth
ed
enom
inat
oris
the
tota
lnum
ber
ofM
LA
con
stit
uen
cies
inth
ed
istr
ict.
Th
issh
ows
ab
and
wid
thof
1%.
2
Figure A2: Balance Test: T-Test for difference in means between Congress wins and losses byless than 1 percent
Differences (Normalized)
Ele
ctio
n C
hara
cter
istic
s
Electors (1120)
ENP (1120)
INC Organization (940)
INC Organization (High) (940)
Invalid Missing (1120)
Invalid Votes (482)
% Muslim (1097)
Population (1046)
Population Density (972)
SC Seat (1120)
ST Seat (1120)
Total Votes (1120)
% Turnout (1120)
% Urban (1097)
% Women Candidates (1120)
−0.2 −0.1 0.0 0.1 0.2
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
Notes: Results from a two-sided t-test of the difference in means between closeelections won and lost by Congress candidates across several election character-istics. Confidence intervals are based on robust standard errors clustered at thedistrict level. N for each test is in parentheses next to the variable being tested.
3
Figure A3: Balance Test: Local Linear Regression Discontinuity
Differences (Normalized)
Ele
ctio
n C
hara
cter
istic
s
Electors (2937)
ENP (3324)
INC Organization (685)
INC Organization (High) (706)
Invalid Missing (3097)
Invalid Votes (1393)
% Muslim (2878)
Population (2297)
Population Density (2810)
SC Seat (2965)
ST Seat (2556)
Total Votes (3191)
% Turnout (2904)
% Urban (2795)
% Women Candidates (3097)
−0.4 −0.2 0.0 0.2 0.4
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
Notes: Results from a local linear regression to estimate differences at the discon-tinuity between Congress candidates winning and losing election across severalelection characteristics. Bandwidths are estimated using optimal bandwidth se-lection suggested by Imbens and Kalyanaraman. While bandwidths differ foreach outcome, they are all between two and three percent. Confidence intervalsare based on robust standard errors. N for each test is in parentheses next tothe variable being tested.
4
Figure A4: Randomization test—estimates showing that violence at time t−1 does not predictour instrument at time t
β: Effect of Previous Violence (normalized)
Ban
dwid
th
1%
2%
3%
−0.002 −0.001 0.000 0.001 0.002
●
●
●
Lagged Any Riot
1%
2%
3%
●
●
●
Lagged Riot Count
Notes: OLS regressions of the instrument (CongCloseWin) in election years t on violence measuresin the election cycle preceding the election (t−1). Regressions include CongCloseProp as a saturatedcontrol. Bars represent 95% confidence intervals based on robust standard errors clustered the districtlevel.
5
Figure A5: Placebo test—estimates showing that the instrument cannot predict “pre-treatment” violence outcomes
β: Effect of Congress Seatshare
Ban
dwid
th
1%
2%
3%
−1.0 −0.5 0.0 0.5 1.0
●
●
●
Log Riot Count
1%
2%
3%
●
●
●
Pr(Any Riot)
Notes: Results from IVLS regressions of logged and binary riot outcomes at t− 1 (previous electioncycle) on CongSeatShare at time t. Bars represent 95% confidence intervals using robust standarderrors clustered at the district level.
6
Figure A6: Placebo test—estimates from the reduced-form maximum likelihood models show-ing that the instrument cannot predict “pre-treatment” violence outcomes
β: Effect of Congress Seatshare
Ban
dwid
th
1%
2%
3%
−4 −2 0 2 4
●
●
●
Any Riot
1%
2%
3%
●
●
●
Riot Count
Notes: This figure reproduces the previous figure using reduced-form negative binomial and probitregressions of riot outcomes at t − 1 (previous election cycle) on CongSeatShare at time t. Barsrepresent 95% confidence intervals based on robust standard errors clustered at the district level.
7
Figure A7: Randomization test—estimates showing that violence and INC performance attime t− 1 does not predict our instrument at time t
β: Effect of Pre−treatment Variables (normalized)
Pre
−tr
eatm
ent V
aria
bles
Lagged Riot Count
Lagged Any Riot
Lagged INC Seat %
Lagged INC District Voteshare
−0.001 0.000 0.001
●
●
●
●
Notes: OLS regressions of the instrument (CongCloseWin) in election years t on normalized violenceand INC performance measures in the election cycle preceding the election (t − 1). Regressionsinclude CongCloseProp as a saturated control. Bars represent 95% confidence intervals based onrobust standard errors clustered the district level. N is 2556 for all regressions.
8
B Data
Included states. The states included in the analysis are: Andhra Pradesh, Arunachal
Pradesh, Maharashtra, Orissa, Punjab, Rajasthan, Tamil Nadu, Uttar Pradesh, and West
Bengal. Other states—all of which, with the exception of Jammu and Kashmir, are extremely
small—were omitted because data were unavailable.
Creating the district panel. Our analyses required compiling a variety of data and ag-
gregating them to create a panel dataset for constant geographic units across time. This is
necessary as Indian administrative district boundaries have changed periodically. In 1961, for
example, there were 331 districts; by 2011 there were 640.
Changes to administrative district boundaries took two forms. The vast majority were
“simple” splits in which one district was cleanly divided into two or more districts. In other
cases, new districts were the result of “complex” splits: the new district’s territory was formed
out of multiple existing districts. Our raw annual data on riots are recorded using the district
boundaries as they existed at the time the riots took place. Our goal was to aggregate these
data back to 1961 districts.
We define the original unsplit districts as “parents” and the new districts as “children.”
To match parents to children, we used Appendix 1 to Table A-1 from the General Population
Tables (Part II-A) of the 2001, 1991, 1981, and 1971 Indian censuses. These tables record all
districts extant in the year of the census. When a new district has been created, the table
indicates the parent district or districts out of which it was carved. For each census round,
we identify the changes that took place.
Our dependent variables are count data. In the simplest case, district boundaries are
unchanged across census years. When the children districts are the result of a simple split,
aggregating backwards is straightforward: since there is only one parent district, we simply
sum up the counts of all its children. For complex splits, the procedure is more involved.
9
In such cases, we take a weighted sum of the counts from the children districts. Using the
Census tables, we calculate what proportion of the territory in a child district j came from
each parent district i and define this as the weight Wij. We compute some count variable X
for parent district i by taking the weighted portion of X from each child district j. That is,
we sum over the product of each Xj and Wij as follows:
Xi =
j∑1
XjWij (A)
More precisely, we use weights calculated from each census to bring districts back first from
2001 to 1991, then from 1991 to 1981, from 1981 to 1971, and finally from 1971 to 1961.35
We further had to map state legislative (MLA) constituencies onto the 1961 adminis-
trative districts in order to create our right-hand-side electoral variables: CongSeatShare,
CongCloseWin, and CongCloseProp. Like administrative districts, the boundaries of state
legislative constituencies changed over time. Throughout, however, these constituencies re-
mained perfect subsets of administrative districts.
We used the reports of the Delimitation Commission of India from 1961, 1971, and 1976 to
assign each legislative constituency at election time t to the administrative district to which it
belonged, also at time t. (After 1976, legislative districts were not redrawn until 2008, easing
the process for this period.) If these administrative districts had gone unchanged since 1961,
then no further work was needed—the MLA constituency was already matched to the correct
1961 district. If the constituency had ended up in child district produced by a simple split,
then we simply reallocated this seat to the original parent district. In cases where an MLA
constituency belonged to a child district produced via a complex split, we used tehsil and
village information contained within the Delimitation Reports, as well as district maps, to
manually assign the constituency to the correct 1961 district. In this manner, we were able
to accurately assign all MLA constituencies between 1962 and 2008 to 1961 administrative
35Equation A is easily generalized for simple changes. When the district remains unchanged,
i = j. When there is a simple split, each Wij = 1.
10
districts.
District Muslim population. To measure the proportion of the population in a district
that was Muslim, we used reports from the 1961, 1971, and 1981 censuses. These data included
the total population for a district and the total number of Muslims in a district. Applying
the same procedure for reconstructing 1961 district boundaries, we added up total population
and total Muslim population for 1961 districts. We thereby calculate the proportion of the
district that was the Muslim.
Congress state governments. We used secondary historical sources to compile a list of
all parties that formed state governments in India between 1961 and 2008. This list included
the party of the Chief Minister as well as any other parties in coalition governments. In our
analyses, we used this data to create a dummy variable indicating whether the Chief Minister
was from the Congress Party in a given state-year.
Riots. As mentioned in the paper, we use the Wilkinson-Varshney database of Hindu-
Muslim riots from 1950 to 1995. We append these with data collected by Mitra and Ray
(2014), bringing the panel to 2000. In cases where these data did not report the district in
which the riots occurred, we used the state and locality of the riot to find the district.
11
C Explanation of simulations
To simulate the expected count of riots for our entire sample (Figure D10), we estimate the
reduced-form equation of our instrumental variables design using negative binomial regression
and a 1% bandwidth. We then generate new copies of the data for several counter-factual
scenarios in which Congress won close elections with different probabilities: 0, 0.2, 0.4, 0.6,
0.8, and 1. Next, we generate 1,000 clustered bootstrap simulations. For each bootstrap
simulation, we estimate the vectors of coefficients, β, then calculate predicted values, y for the
actual data and for 250 counterfactual datasets for each victory probability,36 and transform
these into the expected counts by taking ey. Finally, for each scenario, we take the sum over
all observations in the data, giving us the expected number of riots. By using bootstrap
simulation, we are able to estimate the uncertainty of predictions from our model.
36This is necessary, since we have to randomly sample which close elections are won and
lost, given a probability of victory.
12
D Supplements to the main analysis
Table D1: Descriptive Statistics
Variables Mean SD N Min Max
Number of riots 0.348 1.509 2871 0 47.351Number of riot casualties 9.170 73.535 2871 0 2386.000Number of riot days 0.704 3.806 2871 0 88.351Any Riot 0.160 0.367 2871 0 1.000% Congress seats 0.436 0.319 2871 0 1.000% Congress close wins 0.018 0.052 2871 0 1.000% Congress close elections 0.036 0.072 2871 0 1.000Number of Seats 11.044 6.529 2871 1 55.000% Muslim 0.100 0.091 2799 0 0.614Congress chief minister 0.556 0.497 2863 0 1.000% Congress vote 0.362 0.133 2870 0 0.963% Turnout 0.572 0.119 2871 0 0.890
Table D2: First Stage F-Test
Bandwidth F-Statistic p
1% 55.50 0.00
13
Table D3: Correlates of Close Elections
Prop. Close Elections Any Close Election % Muslim % Urban
Prop. Close ElectionsAny Close Election 0.98% Muslim 0.05 0.07% Urban 0.04 0.05 0.18Population Density 0.12 0.15 0.41 0.11
Spearman rank correlation coefficients. N is 2787 for % Muslim and Urban and 2436 forPopulation Density.
14
Figure D1: Hindu-Muslim riots in India by year, 1950–2000
Year
Num
ber
of R
iots
020
4060
8010
0
1950 1960 1970 1980 1990 2000
Notes: Data come from the Varshney-Wilkinson Dataset on Hindu-Muslim Violence inIndia and an extension of it to 2000 by Mitra and Ray (2014).
15
Fig
ure
D2:
Per
centa
geof
Musl
imvo
tes
goin
gto
vari
ous
par
ties
inIn
dia
nel
ecti
ons,
1957
–200
4
1957 1
1962 1
1962 2
1967 1
1967 2
1971 1
1971 2
1977 1
1977 2
1980 1
1980 2
1984 1
1991 1
1996 1
1998 1
1999 1
2004 1
0.00
0.25
0.50
0.75
17
13
134
13
16
11516
11213
14
15
13
13
16
13
136
16
11
136
13
13
15
10
19
13
18
12
12
14
12
14
Par
ties
Proportion of Respondents
Par
ties
(1)
Con
gres
s
(2)
Con
gres
s A
llies
(3)
CP
I/CP
M
(4)
Inde
pend
ents
(5)
Jana
ta D
al (
JD)
(6)
Jana
ta P
arty
(7)
Jan
San
gh
(8)
JD+
RJD
+S
P
(9)
JD+
SP
(10)
Lef
t Fro
nt
(11)
LO
K D
AL
(12)
Mus
lim L
eagu
e
(13)
Oth
ers
(14)
Sam
ajw
adi P
arty
(S
P)
(15)
Soc
ialis
ts
(16)
Sw
atan
tra
Notes:
Data
base
don
self
-rep
orts
ofM
usl
imvot
ing
fou
nd
inva
riou
ssu
rvey
sco
n-
du
cted
by
the
Cen
tre
for
the
Stu
dy
ofD
evel
opin
gS
oci
etie
san
dn
ewly
com
pil
edby
the
au
thors
.S
urv
eys
som
etim
esre
por
tM
usl
imvo
tesh
ares
for
form
al(i
.e.
pre
-el
ecti
on
)co
ali
tion
sof
par
ties
;th
isex
pla
ins
why
som
ep
arti
esin
the
lege
nd
are
gro
up
edto
get
her
.F
orel
ecti
onye
ars
wh
enw
eh
ave
two
surv
eys,
we
pro
vid
ese
pa-
rate
plo
tsfo
rea
chsu
rvey
.
16
Figure D3: Instrumental variables estimates of the effect of CongSeatShare on riot outcomes,multiple bandwidths
β: Effect of Congress Seatshare
Ban
dwid
th
1%
2%
3%
−0.5 0.0 0.5
●
●
●
Pr(Any Riot)
1%
2%
3%
−4 −2 0 2 4
●
●
●
Log Days of Rioting
1%
2%
3%
●
●
●
Log Riot Casualties
1%
2%
3%
●
●
●
Log Riot Count
Notes: This figure presents coefficient estimates from IVLS regressions of logged or binaryriot outcomes on CongSeatShare, using the approach described in the Data and Identifi-cation section, across multiple bandwidths. Bars represent 95% confidence intervals usingrobust standard errors clustered at the district level. Bandwidth refers to the margin ofvictory used to define a close election. N for all regressions is 2871, across 315 districts.The number of close elections for each bandwidth is 1099, 2212, and 3331 for 1%, 2%, and3%, respectively.
17
Figure D4: Instrumental variables estimates of the effect of CongSeatShare on riot outcomes,with district and year fixed effects
β: Effect of Congress Seatshare
Mod
el S
peci
ficat
ion
District FEs
Year FEs
−0.5 0.0 0.5
●
●
Pr(Any Riot)
District FEs
Year FEs
−4 −2 0 2 4
●
●
Log Days of Rioting
District FEs
Year FEs
●
●
Log Riot Casualties
District FEs
Year FEs
●
●
Log Riot Count
This figure presents coefficient estimates from IVLS regressions of logged or binary riotoutcomes on CongSeatShare, using the approach described in the Data and Identificationsection, with district or year fixed effects. Bars represent 95% confidence intervals usingrobust standard errors clustered at the district level. N for all regressions is 2871, across315 districts.
18
Figure D5: Maximum likelihood estimates of reduced-form equation
β: Effect of Congress Seatshare
Mod
el S
peci
ficat
ion
Saturated Control
Linear Control
−4 −2 0 2 4
●
●
Any Riot
Saturated Control
Linear Control
−20 −10 0 10 20
●
●
Days of Rioting
Saturated Control
Linear Control
●
●
Riot Casualties
Saturated Control
Linear Control
●
●
Riot Count
This figure presents estimates from negative binomial (top three panels) and probit (bottompanel) regressions of the reduced-form equation. That is, unlogged riot outcomes regressedon CongCloseWin and CongCloseProp. Bars represent 95% confidence intervals derivedfrom robust standard errors clustered at the district level. The results are in line with thoseshown in Figure 2, although speficiations using a linear control are weaker and sometimesdrop out of conventional significance. N for all regressions is 2871, across 315 districts.
19
Figure D6: Instrumental variables estimates of the effect of CongSeatShare on riot outcomes,with standard errors clustered by state
β: Effect of Congress Seatshare
Mod
el S
peci
ficat
ion
Saturated Control
Linear Control
−1.0 −0.5 0.0 0.5 1.0
●
●
Pr(Any Riot)
Saturated Control
Linear Control
−5 0 5
●
●
Log Days of Rioting
Saturated Control
Linear Control
●
●
Log Riot Casualties
Saturated Control
Linear Control
●
●
Log Riot Count
This figure presents coefficient estimates from IVLS regressions of logged or binary riotoutcomes on CongSeatShare, using the approach described in the Data and Identificationsection. Bars represent 95% confidence intervals using robust standard errors clustered atthe state level using cluster bootstrapping and percentile confidence intervals. Because thebootstrapped distribution is asymmetric, the confidence intervals are asymmetric aroundthe point estimates. N for all regressions is 2871, across 315 districts.
20
Figure D7: Instrumental variables estimates of the effect of CongSeatShare on riot outcomes,with standard errors clustered by state-election year
β: Effect of Congress Seatshare
Mod
el S
peci
ficat
ion
Saturated Control
Linear Control
−0.5 0.0 0.5
●
●
Pr(Any Riot)
Saturated Control
Linear Control
−4 −2 0 2 4
●
●
Log Days of Rioting
Saturated Control
Linear Control
●
●
Log Riot Casualties
Saturated Control
Linear Control
●
●
Log Riot Count
This figure presents coefficient estimates from IVLS regressions of logged or binary riotoutcomes on CongSeatShare, using the approach described in the Data and Identificationsection. Bars represent 95% confidence intervals using robust standard errors clustered atthe state-election year level. N for all regressions is 2871, across 315 districts.
21
Figure D8: Proportion of elections contested closely by INC by election years.
Election Year
Pro
p. o
f Ele
ctio
ns
0.00
0.02
0.04
0.06
196219
6519
6719
6819
6919
7019
7119
7219
7419
7519
7719
7819
8019
8219
8319
8419
8519
8719
8919
9019
9119
9219
9319
9419
9519
9619
9719
9819
9920
00
Figure D9: Proportion of elections contested closely by INC by states.
Election Year
Pro
p. o
f Ele
ctio
ns
0.00
0.04
0.08
ANDHRA PRADESH
ARUNACHAL PRADESH
ASSAM
BIHAR
GUJARAT
HARYANA
HIMACHAL
PRADESH
KARNATAKA
KERALA
MADHYA
PRADESH
MADRAS
MAHARASHTRA
MYSORE
ORISSA
PUNJAB
RAJASTHAN
TAM
IL N
ADU
UTTAR P
RADESH
WEST B
ENGAL
22
Figure D10: Simulated difference in riots when Congress wins all close elections compared toits actual performance
−30
0−
100
010
020
030
040
0
0 0.2 0.4 0.6 0.8 1
Pr(INC victory) in close elections
Pre
dict
ed c
hang
e in
rio
ts
11768
22−19
−62−103
Notes: This figure plots the simulated predictions of how many fewer riots would haveoccurred if Congress had won close election with probabilities of 0, 0.2, 0.4, 0.6, 0.8,and 1, compared to its actual performance. The predictions are based on 1,000 clusteredbootstrapped replications of negative binomial regression estimates of the reduced-formequation from our instrumental variables design, using a 1% bandwidth (2 simulationsdrop due to failure for the model to converge). The figure reports the median change inriots in the middle of each simulated distribution. See Appendix C for further explanation.
23
Figure D11: Effects of CongSeatShare, by opposition party type
β: Effect of Congress Seatshare
INC
's O
ppon
ent i
n C
lose
Ele
ctio
ns
All Other Parties
Ethno−religious Parties
−1.0 −0.5 0.0 0.5 1.0
●
●
Pr(Any Riot)
All Other Parties
Ethno−religious Parties
−5 0 5
●
●
Log Days of Rioting
All Other Parties
Ethno−religious Parties
●
●
Log Riot Casualties
All Other Parties
Ethno−religious Parties
●
●
Log Riot Count
Notes: This figure presents coefficient estimates from IVLS regressions of logged or binaryriot outcomes on CongSeatShare, using the approach described in the Data and Identifica-tion section. Bars represent 95% confidence intervals using robust standard errors clusteredat the district level. CongCloseProp uses a saturated specification. “Ethnic parties” refersto the BJS/BJP and the Shiv Sena, parties which mobilize along the Hindu-Muslim ethnicdivide. “All Other Parties” refers to all parties other than the ethnic parties and the INC.N for all regressions is 2871, across 315 districts.
24
Figure D12: Effects of CongSeatShare, by time-period and state-government incumbency
β: Effect of Congress Seatshare
Con
ditio
n
YES
NO
−0.5 0.0 0.5
●
●
After 1989Any Riot
−0.5 0.0 0.5
●
●
INC GovernmentAny Riot
YES
NO
−4 −2 0 2 4
●
●
After 1989Log Riot Count
−4 −2 0 2 4
●
●
INC GovernmentLog Riot Count
Notes: This figure presents coefficient estimates from IVLS regressions of logged and bi-nary riot outcomes on CongSeatShare, using the approach described in the Data andidentification section. Bars represent 95% confidence intervals using robust standard er-rors clustered at the district level. N for Pre- and Post-1989 are 1953 and 918, respectively.N for Congress government and opposition are 1593 and 1270, respectively.
25
Figure D13: Reduced-form negative binomial estimates of heterogeneous effects
β: Effect of Congress Seatshare
Con
ditio
n
YES
NO
−5 0 5
●
●
High Muslim Pop.Any Riot
−5 0 5
●
●
High Party FractionalizationAny Riot
YES
NO
−10 −5 0 5 10
●
●
High Muslim Pop.Riot Count
−10 −5 0 5 10
●
●
High Party FractionalizationRiot Count
This figure presents estimates from negative binomial regressions of the reduced-formequation. That is, unlogged riot outcomes are regressed on CongCloseWin andCongCloseProp. Bars represent 95% confidence intervals derived from robust standarderrors clustered at the district level. The subgroups used to demonstrate the heterogeneouseffects are described in the main paper. N for high and low Muslim population are 1427and 1372, respectively. N for high and low party fractionalization are both 1397.