Created
August 14, 2023 20:29
-
-
Save ygivenx/f1981f24406303d985446443f85e3e99 to your computer and use it in GitHub Desktop.
paper-democracy
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Democratic Backsliding in the World’s Largest Democracy | |
Sabyasachi Das* | |
Ashoka University | |
July 3, 2023 | |
Abstract | |
Democratic backsliding is a growing concern globally. This paper contributes to the discussion | |
by documenting irregular patterns in 2019 general election in India and identifying whether they are | |
due to electoral manipulation or precise control, i.e., incumbent party’s ability to precisely predict | |
and affect win margins through campaigning. I compile several new datasets and present evidence | |
that is consistent with electoral manipulation in closely contested constituencies and is less supportive of the precise control hypothesis. Manipulation appears to take the form of targeted electoral | |
discrimination against India’s largest minority group – Muslims, partly facilitated by weak monitoring by election observers. The results present a worrying development for the future of democracy. | |
JEL Codes: D72, D73, P00, Z12 | |
Keywords: Electoral fraud, precise control, democracy, economics of religion | |
*Das: Economics Department, Ashoka University, National Capital Region, India. Email: sabyasachi.das@ashoka.edu.in. | |
The author wishes to thank Lakshmi Iyer, Ajay Shenoy, Sam Asher, Milan Vaishnav, Amrita Dhillon, Sourav Bhattacharya, | |
Siddharth George, Rohit Lamba, Gaurav Chiplunkar, Raphael Susewind, Jonathan Lehne, Dev Patel, Neelanjan Sircar, Aaditya Dar, Vimal Balasubramaniam and Sugat Chaturvedi for valuable comments. Rakesh Kumar provided excellent research | |
assistance. The author is responsible for errors, if any. | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
I Introduction | |
Free and fair elections are cornerstone of a democracy. Yet in many democracies, the fairness of elections is increasingly in doubt. In 2020, for example, the Constitutional court of Malawi declared the | |
Presidential election to be fraudulent.1 The 2019 Presidential election result in Bolivia is also reported | |
to be have been manipulated (Escobari and Hoover 2020). In the US, one-third of voters believe that | |
Joe Biden won the 2020 Presidential election solely because of voter fraud, even though there is no | |
evidence favoring such a claim.2 Even before the election, the share of American voters reported to have | |
confidence in the honesty of elections has been declining for several years, and in 2019, stood at only | |
40%. | |
3 This figure is 50% for the entire world, according to the Gallup World Poll (2007-13). | |
The global erosion of trust in electoral institutions coincides with the autocratizing tendencies of | |
several democracies, known as democratic backsliding or deconsolidation (Waldner and Lust 2018, Foa | |
and Mounk 2016, 2017a,b). Freedom House 2021 report points out that global freedom deteriorated for | |
15 consecutive years, with 75 percent of the world living in a country that experienced deterioration in | |
2020. Democracy Report (2020) mentions: “For the first time since 2001, democracies are no longer | |
in the majority. [...] The countries that have autocratized the most over the last 10 years are Hungary, | |
Turkey, Poland, Serbia, Brazil and India.” While the overall pattern of democratic backsliding is based | |
primarily on subjective evaluation by experts, objective evidence on this trend is lacking (Little and | |
Meng 2023). | |
I contribute to this important debate by examining objective evidence of democratic backsliding | |
in the form of electoral manipulation in the world’s largest democracy – India. India is a somewhat | |
unusual case for electoral fraud as it stands out in terms of the public trust its election authority enjoys. | |
Two-third of its voters reported to have confidence in the honesty of elections in 2019, based on the | |
Gallup Poll survey. Moreover, the confidence in elections is rising in India at least since 2006.4 The | |
level of confidence is also higher than many democracies with strong institutions, such as Japan (57%), | |
France (57%), UK (61%) etc. The independence and institutional strength of the electoral authority | |
in charge of conducting elections, the Election Commission of India (ECI), is an important factor that | |
can potentially explain such high degree of confidence.5 This makes the ECI one of the most powerful | |
election management bodies in the world. | |
In the past few years, however, the credibility of the ECI has been called into question, with allegations of bias in scheduling of elections (Ramachandran 2022) and arbitrary deletion of names of registered Muslim voters (Malhotra 2019, Trivedi 2019, Naqvi 2022), both favoring the ruling party. The | |
recent democracy reports of the V-Dem Institute highlight that various indicators of democracy in India, | |
including the autonomy of the ECI, has been declining. Democracy Report (2021) have consequently | |
classified India as an “electoral autocracy”. As the V-Dem report points out, decline in the autonomy | |
of the ECI was one of the important factors contributing to the reclassification of India’s regime type. | |
Similarly, Freedom House has changed India’s status in 2021 from Free to Party Free (Repucci and | |
Slipowitz 2021). The Supreme Court of India, in a recent judgement in 2023, acknowledged the dangers | |
1 | |
https://www.nytimes.com/2020/02/03/world/africa/Malawi-president-election-fraud.html | |
2 | |
https://www.monmouth.edu/polling-institute/reports/MonmouthPoll_US_031721/ | |
3 | |
https://news.gallup.com/poll/285608/faith-elections-relatively-short-supply.aspx | |
4 | |
https://news.gallup.com/poll/248495/confidence-key-institutions-high-india-votes.aspx | |
5 | |
Section II provides a brief discussion on the independence of election authorities in India and the contextual details of | |
India’s general elections. | |
2 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
of a weak ECI and granted it significant autonomy and protection from executive overreach.6 | |
In light of these developments, I first document that the 2019 general election in India that reelected | |
the incumbent party shows significant irregularities in the election data – the density of the incumbent | |
party’s win margin variable exhibits a discontinuous jump at the threshold value of zero. It implies that in | |
constituencies that were closely contested between a candidate from the incumbent party and a rival, the | |
incumbent party (BJP) won disproportionately more of them than lost. This is known as the McCrary | |
test and is now a standard check for manipulation of running variable in the regression discontinuity | |
design (RDD) method used in analysis of political economy (Prakash et al. 2019, Nellis et al. 2016, | |
Bhalotra et al. 2014). I do not find similar discontinuities in the previous general elections for either | |
BJP or INC (Indian National Congress), the other major national party, as well as for state assembly | |
elections held simultaneously with the 2019 general election and those held subsequently. Moreover, | |
BJP’s disproportionate win of closely contested constituencies is primarily concentrated in states ruled | |
by the party at the time of election. | |
Failure of McCrary test however does not necessarily imply electoral fraud. If the incumbent party, | |
due to its superior electoral machinery, was able to accurately predict and affect win margins in closely | |
contested constituencies – a phenomenon known as precise control (Jeong and Shenoy 2020, Vogl 2014), | |
then it could also generate such patterns. The incumbent party in India may have been able to exercise | |
precise control in 2019 since it had significantly built up its organizational capacity in several states, | |
subsequent to its 2014 general election victory. It mobilized active party workers at the level of polling | |
stations who monitored and shaped voter attitudes, backed by centrally managed teams analyzing the | |
collected information and suggesting campaign strategies (Jha 2017). Precise control in this context, | |
therefore, if exercised, is likely to be facilitated by localized and targeted campaigning facilitated by | |
grassroots presence of the party organization. This can explain the patterns described above. In the | |
subsequent analysis I attempt to look for evidence that may distinguish between the two competing | |
hypotheses.7 | |
For my analysis, I put together several new datasets in addition to accessing the candidate level | |
general election results for 1977-2019 and state assembly election results for 2019-2021 from standard | |
sources. To examine precise control, I access the well-established post-poll survey – the National Election Survey (NES) of 2019 that gives micro data on election campaigning by political parties. To investigate election manipulation, I compile two different but official versions of constituency level Electronic | |
Voting Machine (EVM) turnout data (for 2019 general election) to directly measure data discrepancy. | |
The ECI initially released in its official website the “final” count of EVM votes polled for each Parliamentary Constituency (PC) for the first four out of seven phases of the 2019 elections (373 out of 543 | |
PCs). Subsequently, it released constituency wise number of votes counted in EVMs, which did not | |
match the initial numbers. When the media pointed out the discrepancy, the ECI removed the earlier | |
figures from its website. I access copies of the earlier turnout data to measure discrepancy. I also put | |
together the list of counting observers assigned to each constituency by the ECI to monitor counting | |
of votes in 2019. The data provides various characteristics of the counting observers such as the state | |
where they work, their cadre (i.e., whether they are part of the central or state bureaucracy), year of | |
joining service etc. Additionally, I compile polling station level election outcomes for the 2019 general | |
6 | |
https://www.thehindu.com/news/national/committee-of-pm-lop-cji-to-advice-on-appointment-of-electioncommissioners-supreme-court/article66570806.ece | |
7The excess mass of constituencies that BJP barely won could also arise purely due to chance. | |
3 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
election by scraping and parsing the scanned PDFs containing the data, available from the official websites of election authorities in individual states. To examine targeted voter suppression of Muslims as a | |
potential mechanism of manipulation, I compute electorate share of Muslims at the level of Assembly | |
Constituencies (ACs) using a 3 percent representative sample of voter lists and applying a highly accurate religion prediction algorithm on their names, and match it to polling stations and ACs.8 | |
. Section III | |
describes the datasets and their sources. | |
I use a new question added to the NES in 2019 to measure campaigning in the form of door-to-door | |
visits by BJP and other political parties in a representative sample of PCs to directly test for precise | |
control. I find that neither BJP nor any other party campaigned significantly harder in constituencies | |
that BJP barely won. Moreover, in BJP ruled states, campaigning by BJP does not exhibit statistically | |
significant discontinuity, while that for the other parties does. This makes precise control less likely to | |
be the primary mechanism. | |
Electoral manipulation, on the other hand, can take place at the stage of voter registration (registration manipulation) or at the time of voting or counting (turnout manipulation). To examine the mechanisms facilitating manipulation, I focus on Muslim voters who generally do not support BJP (Varshney | |
2019), and are easily identified in the voter list due to their culturally distanct names. Therefore, they | |
are potentially subject to both registration and turnout manipulation.9 | |
I consider two channels; first, | |
strategic deletion of Muslim names from the list of registered voters or electoral rolls (Lehne 2022). | |
Second, strategic suppression of Muslim votes at the time of voting (or counting) (Neggers 2018). I do | |
not consider the possibility of manipulation of EVMs themselves as a mechanism, as Purkayastha and | |
Sinha (2019) have pointed out that given its technology, it is hard to manipulate them at scale. | |
To test for registration manipulation, I compute growth rate of electorate (i.e., number of registered | |
voters) for each Parliamentary Constituency (PC) between 2014 and 2019. I show that the growth rate | |
falls discontinuously by 5 percentage points (compared to mean of 0.09) in PCs barely won by BJP, and | |
the fall is concentrated in PCs with higher share of Muslim electorate. To examine turnout manipulation, | |
I first examine the absolute difference between the two official versions of EVM turnout data. The | |
discrepancies could be due to administrative errors during counting of votes. However, the extent of | |
discrepancy, in that case, should not exhibit any discontinuous change with respect to the incumbent’s | |
win margin at its threshold value of zero. I however find that there is a large discontinuous increase in | |
the magnitude of data revision at the threshold. Consistent with previous results, the discontinuity is | |
concentrated in BJP ruled states. | |
I interpret the evidence on turnout discrepancy as indicative of manipulation done locally at the | |
polling stations, rather than resulting from aggregation fraud at the constituency level (Callen and Long | |
2015). It is unlikely that ECI would engage in direct tampering of turnout data ex-post. Moreover, | |
barring one case, the magnitude of data revision is smaller than BJP’s absolute margin of victory. I show | |
that polling station level election outcomes in the relevant PCs exhibit irregularities consistent with local | |
8Census data on religious composition of population is not ideal in this case, since the lowest level of geographic unit for | |
which such data is available is tehsil, which (a) does not always map to a single AC and (b) hard to map to polling stations, | |
since the map of geographic area covered by a polling station is not available and data on location of polling stations is also | |
error-prone (Hintson and Vaishnav 2021). Additionally, electorate share of Muslims is the ideal measure, which can differ | |
from their population share because of various reasons such as differential fertility and child survival rates etc., which could be | |
correlated with their support for BJP. | |
9Religious identity, especially the Hindu-Muslim divide, is a salient political cleavage in India (Bhalotra, Clots-Figueras, | |
Iyer, and Vecci 2021, Varshney 2003). Moreover, the salience of religion has heightened under BJP’s rule since 2014 (Khosla | |
and Vaishnav 2022). | |
4 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
manipulation. | |
To examine whether turnout manipulation was in part facilitated by weak monitoring of counting of | |
votes, I analyze the assignment of counting observers across PCs. I compute the fraction of counting | |
observers assigned in a PC who are from the State Civil Service (SCS), as opposed to the Indian Administrative Service (IAS).10 Since SCS officers are appointed by the state government, unlike the IAS | |
officers who are centrally appointed, they more likely to be politically pliable. I also compute the fraction of observers in a PC who are SCS and work in a BJP ruled state.11 I find that both fraction exhibits | |
large, positive and statistically significant discontinuity at the BJP win margin of zero. For the fraction | |
of SCS officers from BJP ruled states, the discontinuity is larger in magnitude in PCs of BJP ruled states, | |
while it is smaller and statistically insignificant for non-BJP ruled states. Additionally, in PCs won by | |
BJP, the fraction of counting observers who are SCS and come from BJP ruled states positively predicts | |
the extent of turnout data discrepancy in the PC; in PCs that BJP lost, no such relationship holds. | |
I analyze polling station level election results to test for local manipulation. For each polling station, | |
I compute the vote share of BJP at that polling station relative to its vote share in the PC; I refer to this | |
as the relative BJP vote share. This makes comparison of polling stations across constituencies easier. | |
I show that within a constituency, relative BJP vote share typically hovers around one across polling | |
stations with different turnout, except in closely contested constituencies barely won by BJP in BJP ruled | |
states. In those constituencies, the relative vote share of BJP exhibits a large spike in polling stations with | |
high turnout. The pattern is replicated with a polling station level indicator of BJP’s vote share exceeding | |
95th percentile of its distribution.12 I compute the distribution of second digit in the polling station level | |
vote tallies of candidates to measure departure from Benford’s law at the polling station level. Benford’s | |
law (Benford 1938) specifies distribution of digits in naturally occurring numbers, and departures of | |
the observed distribution from Benford’s specification is often used as an indicator of manipulation. I | |
show that the departures from Benford’s law exhibit the same pattern. Additionally, I perform tests on | |
the shape of the BJP’s vote share density, proposed in more recent research on electoral fraud, and find | |
results consistent with fraud. Moreover, the spike in the relative BJP vote share mentioned above is | |
higher in PCs with larger discrepancy in turnout data. While the first couple of results are consistent | |
with both mechanisms, the rest of the results indicate manipulation. | |
Finally, manipulation in the form of targeted electoral discrimination against Muslim voters would | |
imply that within a PC barely won by BJP, high vote shares of the party should be concentrated in | |
areas with higher Muslim presence. On the other hand, if precise control is the appropriate explanation, | |
then we should expect the opposite, as the increase in BJP’s vote share in 2019 relative to 2014 came | |
primarily from Hindus, especially from its lower caste groups, while its support among Muslims was | |
low and constant across the two elections (Varshney 2019). | |
I match the data on AC level electorate share of Muslims (described above) to polling stations to test | |
the above hypothesis. ACs are smaller than PCs, and each PC contains about 7 ACs on average. The | |
matched data therefore provide us within PC variation in electorate share of Muslims across polling stations located in different ACs. I find that in PCs that BJP barely lost, its vote share is less likely to exceed | |
the 95th percentile in polling stations located in high Muslim share ACs within the PC. However, this | |
negative relationship gets significantly reduced in PCs barely won by the party; in those PCs, the likeli10There are typically multiple counting centers in a PC, each of which is assigned a counting observer. | |
11Observers are deployed in a state different from where they work. | |
12In those polling stations, BJP on average received 90% of votes cast. | |
5 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
hood of the event does not fall in ACs with higher Muslim share. This again supports the manipulation | |
hypothesis. | |
The paper is unable to comment on the overall extent of manipulation in the 2019 general election. | |
It focuses on closely contested constituencies as an empirical strategy to detect the presence of potential manipulation. Back of the envelope calculation shows that in PCs with BJP win margin less than | |
5%, BJP’s “excess” win is in about 11 PCs. Therefore, even if all the disproportionate wins of BJP in | |
closely contested PCs is due to manipulation, it likely would not have changed the government formation.13 Nonetheless, the results signify a worrying development for the future of democracy in India and | |
consequently, in the world at large. | |
This paper contributes to our understanding of democratic backsliding in consolidated democracies | |
using objective measures. Little and Meng (2023) argue that subjective evaluation of nature of democracy by experts may be subject to their biases. Therefore, claims about democratic backsliding need to | |
be grounded in more objective evidence. The authors examine objective measures of democracy and do | |
not find any evidence of systematic backsliding across democracies. They conclude, “[...] it may be the | |
case that major backsliding is occurring precisely in ways that elude objective measurement. However, | |
this is an extraordinary claim, which requires a stronger theoretical and empirical basis than has been | |
offered to date.” | |
Additionally, several studies examining democratic backsliding have focused on the “demand side” | |
issues, specifically, voters’ willingness to sacrifice democratic principles in the context of increased polarization (Braley et al. 2022, Fishkin et al. 2021, Graham and Svolik 2020), rise of populism (Martinelli | |
2016) etc., resulting in dismantling of check-and-balances (¸Sa¸smaz, Yagci, and Ziblatt 2022). The paper | |
shows that dilution of electoral integrity is also an important and “supply side” contributor to democratic | |
backsliding. Several consolidated or stable democracies, such as India14, Mexico15, Hungary (Scheppele | |
2022), have witnessed weakening of its electoral institutions in recent times. It is relevant to understand | |
whether and how this weakening contributes to democratic backsliding. There is little evidence in mature | |
democracies of direct electoral fraud typically observed in weaker democracies, such as ballot stuffing, | |
booth capturing or direct manipulation of data by election authorities. Incumbents in these countries are | |
likely to adopt subtler strategies, such as fragmentation of opposition (Arriola, Devaro, and Meng 2021) | |
or voter suppression (Manheim and Porter 2019) etc. My examination of the latest general election in | |
India adds to our understanding of this process. | |
Empirical analyses of electoral fraud have typically focused on weak democracies such as Afghanistan | |
(Callen and Long 2015), Ghana (Asunka et al. 2019), Nigeria (Onapajo and Uzodike 2014), Russia | |
(Enikolopov et al. 2013, Rundlett and Svolik 2016), Mexico during 1980s (Cantú 2019), nineteenth | |
century Germany (Ziblatt 2009), or local elections in robust democracies such as Japan (Fukumoto and | |
Horiuchi 2011). In these cases, the nature of fraud typically entails aggregation fraud, tampering of election documents at the polling station level etc. In case of India, my paper shows, manipulation took the | |
form of localized and targeted discrimination against a well-identified minority group, via manipulation | |
of voter registration as well as weaker monitoring of the election process. | |
Previous studies have employed several methods to detect electoral fraud – Cantoni and Pons (2020) | |
use sampled data on proven and suspected fraud cases, Asunka et al. (2019) and Enikolopov et al. | |
13BJP won 303 PCs and it needed 272 PCs to form the government. | |
14https://www.telegraphindia.com/india/election-commission-weak-kneed-say-former-officials/cid/1688448 | |
15https://www.bbc.com/news/world-latin-america-64742733 | |
6 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
(2013) examine effect of poll observers on incumbent vote share, Christensen and Schultz (2014) analyze turnout behavior of specific voting groups more likely to be targeted for frauds, James and Clark | |
(2020) conduct survey of polling station workers etc. My paper contributes methodologically by analyzing irregularities across polling stations and constituencies with different demographic composition | |
of minority voters and applying regression discontinuity and difference-in-discontinuity designs. Additionally, papers on electoral fraud typically employ one specific method to detect fraud. In contrast, | |
this paper employs a combination of methods to demonstrate consistent results. This is of particular | |
importance given that the nature of irregularities is more subtle, as one may expect in a consolidated | |
democracy, and hence, requires a deeper examination. | |
II Background and Context | |
Autonomy of the Election Authority in India: Election Commission of India is the central authority | |
in charge of conducting national (and state) elections in India. It was established in 1950. Several | |
scholars have highlighted the exemplary role played by the ECI in ensuring free and fair elections and | |
consequently, in the consolidation of India’s democracy, in spite of its challenging social, cultural and | |
economic environment. Banerjee (2017), for example, says: “In contrast to the usual inefficiencies of | |
Indian public institutions, the well-oiled machinery of the Election Commission stands out because of | |
its excellent performance in conducting elections on an unimaginably large scale.” (p 410) Sridharan | |
and Vaishnav (2017) point out: “What has emerged over the past six-and-a-half decades is an Election | |
Commission that has significant powers, far greater than what its counterparts in many democracies have | |
at their disposal. [...] According to a 1996 poll conducted by the Centre for the Study of Developing | |
Societies, the ECI was the most respected public institution in all of India with 62 per cent of respondents | |
favourably disposed. A 2008 study found that an even higher percentage – nearly 80 per cent – of Indians | |
surveyed expressed a high degree of trust in the Commission, second only to the army among state | |
institutions.” (p 419 of Kapur et al. (2018)) Multiple researchers have shown that redistricting in India | |
does not suffer from gerrymandering, a common phenomenon in the US, thanks to the independence of | |
the Delimitation Commission of India from any political interference (Kjelsrud et al. 2020, Nath et al. | |
2017, Iyer and Reddy 2013). Eggers et al. (2015) in their study of elections across a number of countries | |
find no evidence of manipulation of election results in India for the period 1977-2004. | |
Elections in India: India follows a Parliamentary system. The Parliament has 543 legislatures or | |
Members of Parliament (MPs), each of whom is elected from a Parliamentary Constituency (PC) using | |
the first-past-the-post rule. The national or general elections in India are conducted every 5 years, unless | |
there is an early dissolution of the government. There are several parties that field candidates in the | |
general elections. The two main national parties are the BJP (Bhartiya Janata Party) and the INC (Indian | |
National Congress). Apart from the national parties, there are several regional or state parties that are | |
important political actors in specific states. | |
The ECI also conducts the state elections in India. In a state election, voters from each Assembly | |
Constituency (AC) elect one representative (Member of Legislative Assembly) to the state legislature. | |
The size of the legislature in a state depends on its population. Taken together, there are roughly 4, 300 | |
ACs in India. An AC is always subsumed within a PC. The timing of state elections is not synchronized | |
7 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
with the general elections. During every general election, a subset of states has their state assembly | |
elections simultaneously with it. But the subset of states changes over time due to either early dissolution | |
of state government, or the central government or both (Balasubramaniam et al. 2021). In 2019, the states | |
of Andhra Pradesh, Orissa, Arunachal Pradesh and Sikkim had simultaneous general and state elections. | |
General Election in 2019: The most recent general election in India happened in 2019 that reelected | |
the incumbent coalition (the National Democratic Alliance or NDA), led by the BJP, to power. There are | |
two significant developments related to the 2019 general elections that are worth highlighting. First, the | |
incumbent party, BJP, had built up its grassroots organizational presence significantly in the lead up to | |
the 2019 elections, especially in certain states. The party deployed it efficiently during its 2019 election | |
campaign, as discussed in Jha (2017). The author describes that the party created polling booth level | |
committees who were in charge of connecting with voters enrolled in the booth, organizing membership | |
drives, collecting household level data on various social, demographic and economic indicators, along | |
with their political attitudes. The households were classified according to their intention to vote for | |
the party to decide the party’s campaign strategy. This localized campaign, in conjunction with the | |
allocation of abundant campaign resources, gave the party an edge over the other parties. | |
At the same time, there were reports of mass deletion of voter names of minority groups from | |
electoral rolls (Malhotra 2019, Trivedi 2019, Naqvi 2022). Since the incumbent party enjoys lower | |
electoral support among the minority groups, such deletions may provide an electoral advantage to the | |
party. Additionally, subsequent to the elections, the ECI released two “final” versions of the PC level | |
EVM turnout data that did not match (Agarwal 2019). ECI did not provide any accounting of the data | |
discrepancy.16 These reports raise fears about possible electoral manipulation during the 2019 elections. | |
III Data | |
Aggregate Election Results: I first access the candidate level Parliamentary election results from | |
1977-2019 and state assembly election results of 2019-2021. It is published by the Election Commission | |
of India, and is compiled and made public by the Trivedi Centre for Political Data (TCPD) at Ashoka | |
University (Bhogale et al. 2019).17 The data contain for each PC (AC, in case of state election) and | |
each election year, details of candidate names, their party affiliations, votes received by candidates, total | |
turnout and electorate size. | |
Two Versions of EVM Turnout Data: The Election Commission of India (ECI) initially published | |
“Final Voter Turnout” figures for the first four (out of seven) phases of the 2019 general election.18 | |
These figures reflect the PC wise number of votes polled in the Electronic Voting Machines (EVMs). | |
These numbers however do not match with the PC wise number of votes counted in the EVMs, as | |
available in the official website of the ECI. This is unusual as votes polled and votes counted in the | |
EVMs should be identical. The news media pointed out this discrepancy in the data, following which | |
16The Association for Democratic Reforms, an independent election watch body, has filed a petition in the Supreme Court | |
of India seeking reconciliation of the data: https://www.nationalheraldindia.com/india/adr-files-petition-in-supreme-court-onmismatch-in-evm-data. | |
17The data is publicly available from the TCPD’s website http://lokdhaba.ashoka.edu.in. | |
18For the rest of the PCs, it released the “estimated” turnout figures, and therefore, are not considered for analysis. | |
8 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
the ECI removed the “Final Voter Turnout” figures from its website. The PDF copies of the data are | |
publicly available here: https://www.scribd.com/docu ment/411811036/EC-s-votes-polled-data-Phase1. I digitize the data and match it against the (revised) official EVM turnout figures available in the | |
ECI’s website: https://eci.gov.in/files/file/10969-13-pc-wise-voters-turn-out/. | |
Counting Observers in 2019: The ECI appoints officials who are responsible for overseeing and | |
monitoring the counting of votes in each counting center. They are referred to as counting observers. | |
The counting observers have the power to stop the counting process or not declare the results if they | |
find breach of counting procedure or if they suspect that some form of fraud has taken place. They | |
have to report to the ECI if they take such actions. I access the list of counting observers for 2019 | |
general election from the official website of ECI. The data contain the names of officials assigned to | |
each PC, along with their ‘office state’, i.e., the state where they were currently working as a bureaucrat, | |
their ‘home state’, i.e., where they were born, whether they are an Indian Administrative Service (IAS) | |
officer or from the State Civil Service (SCS) and their year of joining the service. I am able to match | |
data for 539 PCs (out of 543) containing 1, 804 counting observers. | |
Polling Station Level Results: I put together polling station level election results for the 2019 general | |
election. Polling station level election records are available in each of the states’ Chief Electoral Officer’s | |
official website. The format of the data differs from state to state. While in one state the digitized data | |
is available, in most states the data come in the form of scanned PDFs containing the polling station | |
level results for each constituency. I scrape, digitize, clean and compile the results for 22 major states | |
of India covering more than 900, 000 polling stations. For each polling station, the data provide the PC | |
and AC it falls under, candidate-wise vote tallies (along with votes in favor of “None of the Above”) | |
and candidate’s party affiliation. This allows me to calculate the absolute turnout and vote share of BJP | |
at the polling station level. Except the state of Uttar Pradesh (UP), the data do not contain number of | |
electorates at the polling stations. Therefore, barring UP, it is not possible to calculate turnout rate at the | |
polling station level. | |
National Election Survey 2019: National Election Survey (NES) is a post-poll voter survey conducted by the Center for the Study of Developing Societies (CSDS). The surveys, conducted right after | |
every general election but before declaration of results, ask a representative sample of voters in a randomly selected sample of PCs questions about their political attitudes, knowledge and activities, among | |
other things. NES has been conducted regularly in India since 1990s and is a credible source of voter | |
preferences and political activities (Balasubramaniam et al. 2021, Banerjee et al. 2019, Thachil 2014). I | |
access the relevant sections of the NES 2019 data to examine the campaigning activities of parties. NES | |
2019 surveyed 24,236 voters across 208 PCs. | |
Muslim Electorate Share using Voter List: I create reliable estimates of Muslim electorate share | |
at the AC and PC level to examine electoral discrimination against Muslims as a potential source of | |
manipulation. For this, I use a 3 percent random sample of registered voters, representative at AC | |
level (about 25 million observations). I access this proprietary data from a private organization that has | |
compiled the full list of registered voters for the entire country using the electoral rolls published by | |
the ECI. The data uses electoral rolls published till 2018, and therefore, is not subject to any strategic | |
9 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
deletion that may have happened during the 2019 revision prior to the general election. I use a religion | |
prediction algorithm (with 97 percent accuracy) developed by Chaturvedi and Chaturvedi (2023) to | |
predict each voter’s religion from their name, which allows me to compute Muslim share in each AC.19 | |
Appendix Figure A1 plots the local polynomial relationship between vote share received by all Muslim | |
candidates running in an AC in a state assembly election (during 2008-2018) against the electorate | |
Muslim share and finds strong positive relationship. This indicates that my measure of AC level Muslim | |
share is reliable. | |
IV Irregularities in Aggregate Election Results | |
I first perform the McCrary test that checks for discontinuity in the distribution of win margin (of any | |
party) at the value of zero (Calonico et al. 2014, McCrary 2008) . The presence of discontinuity would | |
imply that there is disproportionately higher mass of closely contested electoral constituencies where | |
the party has barely won or lost, depending on whether the discontinuity is positive or negative. The | |
idea is that, if elections are fair, then conditional on an election being closely contested between party A | |
and any other party, the party A’s chance of winning would be close to 50%. This is because, whether | |
it ends up winning a really close election would effectively be random. I perform the McCrary tests by | |
computing the win margins for the two major national parties of India, namely BJP and INC. Any party | |
A’s win margin is defined as: | |
Party A win margin = (vote share of A - vote share of winner), if A loses | |
= (vote share of A - vote share of runner up), if A wins | |
BJP win margin therefore takes negative values in constituencies where it lost and positive values where | |
it won.20 When the variable takes values close to zero, it implies that BJP either lost or won the election | |
with a narrow margin. Similarly, a large negative (or positive) value would imply that BJP lost (or won) | |
that election with a large margin. Same is true for INC win margin. Figure 1a and 1b plot the densities | |
of BJP and INC win margins, respectively, for the 2019 general election. I observe a large discontinuous | |
jump in the density of BJP Win Margin just right of zero. This implies that conditional on a closely | |
contested election between a BJP candidate and another candidate, BJP was significantly more likely to | |
win that election than lose. I do not observe any discontinuity in the density of INC Win Margin.21 | |
Before commenting on the interpretation of this finding, I wish to point out that failure of McCrary | |
test in electoral context is rare, both in India as well as internationally. I perform the test for past general | |
elections of India using BJP and INC win margins. Table 1 reports the estimated discontinuities in the | |
densities of the two variables for all general elections going back to 1977. I find that the BJP win margin | |
in 2019 is the only case exhibiting statistically significant estimate of the discontinuity. To illustrate | |
this point more clearly, Appendix Table A1 reports the number and percentage of constituencies BJP | |
won and lost in constituencies with small absolute BJP win margins for the past 4 general elections. I | |
consider three narrow win margin bands - within 0.05, 0.03 and 0.02. For each band, the 2019 elections | |
19I thank Sugat Chaturvedi for implementing the algorithm in the data. | |
20This is a standard definition in this kind of exercise. Nellis et al. (2016), for example, use the same running variable | |
defined for INC to estimate the causal effect of electing an INC politician on violence. | |
21It implies that the disproportionately higher wins of BJP candidates were primarily against regional parties. | |
10 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
Figure 1—McCrary Tests Demonstrate Discontinuity in BJP Win Margin Distribution | |
(a) BJP Win Margin: All States (b) INC Win Margin: All States | |
(c) BJP Win Margin: BJP Ruled States (d) BJP Win Margin: Non-BJP Ruled States | |
show the most lop-sided share of win for BJP; the share of BJP victory is 69-74% in 2019, depending on | |
bandwidth. For each bandwidth, 2019 is the only year where the data rejects the null that the likelihood | |
of BJP victory is 0.5. In each case, the null hypothesis is rejected with with p-value less than 0.01, i.e., | |
there is less than 1% probability of observing the patterns with BJP’s true probability of victory in close | |
elections being 0.5. | |
Nellis, Weaver, Rosenzweig et al. (2016) find that INC win margin passes the McCrary test in state | |
assembly elections for the period 1962–2000. Uppal (2009) find the same using incumbent win margin | |
as the running variable for state elections during the period 1975–2003. Moreover, the only evidence of | |
failure of McCrary test that has been documented in a robust democracy, is in the context of elections in | |
the US (Jeong and Shenoy 2020, Vogl 2014, Caughey and Sekhon 2011). Eggers et al. (2015), however, | |
have shown that it is in fact an exception as the test works in a number of countries (including the US | |
and India) and for different time periods. Hence, the failure of the test in the 2019 general election in | |
India warrants notice and additional investigation. | |
Interpretation: While the result is consistent with possible manipulation of the election results in | |
favor of the BJP, the incumbent party, it is not the only interpretation. Alternatively, it could be that BJP, | |
being the incumbent, was able to exercise precise control over win margin, i.e., it was able to precisely | |
predict win margins, especially in constituencies where a close contest was expected, and was able to | |
11 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
Table 1—Estimates of the Discontinuity in the Density of BJP and Congress Win Margins | |
2019 2014 2009 2004 1999 1998 1996 1991 1989 1984 1980 1977 | |
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) | |
Estimate of discontinuity: BJP Win Margin 1.51** -0.24 -0.83 1.88 2.41 -0.79 -1.20 0.43 -0.01 | |
(0.75) (0.74) (1.15) (1.20) (1.63) (1.28) (0.94) (0.60) (1.24) | |
Estimate of discontinuity: INC Win Margin 0.78 -0.37 1.80 -1.02 -1.37 -0.24 0.66 -1.19 0.36 0.49 -0.54 0.89 | |
( 0.60) (0.73) (1.30) (1.03) (0.91) (1.01) (0.79) (0.81) (0.77) (0.86) (0.73) (0.71) | |
Notes: The table reports the estimates of the discontinuity in the density at the threshold value of zero for two running variables – BJP Win Margin (first row) and | |
INC/Congress Win Margin (second row). The year in each column refers to the general election year. Each estimate, therefore, comes from a separate test for a given | |
of running variable in a given general election year. The estimates are computed using the method proposed by Calonico et al. (2014). The robust standard errors are | |
reported in parentheses. *** p<0.01, ** p<0.05, * p<0.1 | |
affect it, thanks to its comparative advantage in electoral campaigning and greater access to resources. | |
Notice that it is not enough for BJP to predict the constituencies where it will face a close fight to | |
generate failure of McCrary test. In such a case, it would campaign harder in all the constituencies | |
expected to have a close contest, resulting in a uniform shift of its win margin to the right and hence, | |
no discontinuity would emerge at zero.22 The party would have to accurately predict whether they are | |
ahead or falling behind in the close contest. Hence, for precise control to be the explanation, BJP had | |
to accurately predict the sign as well as the magnitude of the win margin, to be able to target the set | |
of constituencies where it expects to lose in a close contest. Jeong and Shenoy (2020) have shown that | |
incumbent parties in US state legislative elections do exhibit behavior consistent with precise control | |
and it can explain their ability to consistently win majority of close races. While election prediction in | |
India is still not as sophisticated as in developed countries such as the US, it is possible that BJP, due to | |
its superior electoral machine, was able to precisely predict and affect win margins. | |
State Assembly Elections: Seven states had their state assembly elections in 2019, including four | |
states where the state elections were held concurrently with the general election.23 BJP was the incumbent party in the government in three of the seven states. I compute the BJP win margin for state election | |
results for BJP and non-BJP ruled states separately. I find that it does not exhibit failure of McCrary | |
test (Appendix Figures A2a and A2b). Same is true for state elections held in 2020 and 2021 (Appendix | |
Figures A2c and A2d). If precise control is the mechanism responsible for Figure 1a, then we should | |
expect it at work at state level elections as well, at least in 2019. I however do not find that. | |
BJP vs. Non-BJP Ruled States: I now perform the McCrary test for the 2019 general election in | |
two sub-samples of constituencies – those in states that were ruled by the BJP at the time of the 2019 | |
election and those in non-BJP ruled states.24 The two sub-samples have equal number of constituencies. | |
Figures 1c and 1d show the densities of BJP win margin for the two sub-samples respectively. I find | |
that for BJP ruled states, the density shows an even larger discontinuous jump to the right of threshold. | |
For non-BJP ruled states, the jump is muted. Appendix Table A2 reports the estimated discontinuities | |
in the densities for the two sub-samples of states separately. The estimate of the jump for BJP ruled | |
states is highly statistically significant (p-value = 0.007), while it is statistically insignificant (p-value | |
22Lee and Lemieux (2010) refer to this as imprecise control over the running variable and argue that the regression discontinuity design remains valid under imprecise control. | |
23The seven states are Arunachal Pradesh, Haryana, Maharashtra, Jharkhand, Andhra Pradesh, Odisha and Sikkim. | |
24In 2019, the BJP ruled states were Assam, Bihar, Goa, Gujarat, Haryana, Himachal Pradesh, Jharkhand, Maharashtra, | |
Manipur, Nagaland, Tripura, Uttar Pradesh, Uttarakhand. | |
12 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
Figure 2—McCrary Test for General Elections in 2014 and 2009 | |
(a) 2014 Election: BJP Ruled States (b) 2014 Election: Non-BJP Ruled States | |
(c) 2009 Election: BJP Ruled States (d) 2009 Election: Non-BJP Ruled States | |
= 0.84) for non-BJP ruled states. Therefore, the overall failure of McCrary test is primarily driven by | |
constituencies in the BJP ruled states. Those states, on the other hand, do not exhibit differential patterns | |
in the previous two general elections (Figure 2). | |
The results are consistent with both mechanisms. Having control over the state’s bureaucratic machinery can help a party target its manipulation efforts better, especially in a context where widespread | |
manipulation is hard to implement given the intense media attention during elections and vocal rival political parties.25 It is also consistent with precise control if being in power at the state government helps | |
in mobilizing party workers at the ground. Greater presence of party workers at the ground can generate | |
more precise information about a party’s expected vote share vis-a-vis the main rival party, which can | |
facilitate precise control.26 | |
Comparability of PCs that BJP Closely Won and Lost: Table 2 reports the estimates of discontinuity of various PC level electoral variables at the BJP win margin threshold of zero. It finds no systematic | |
25Even though bureaucrats (Indian Administrative Service officers) are employees of the central government, their appointment and promotion are influenced by state governments (Iyer and Mani 2012). During general election, they report to the | |
ECI, but the pool of officers available is shaped by the state government, making them pliable to the interests of the incumbent | |
party at the state. | |
26Among the states not ruled by BJP, several of them, such as Madhya Pradesh, Rajasthan, Karnataka, Orissa, have strong | |
presence of the party. Appendix Figure A3 shows the discontinuity for that subsample of states. It does not exhibit a differentially larger discontinuity than the one in Figure 1d. | |
13 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
differences between PCs that BJP barely lost and won. The variables examined are electorate size, | |
turnout rate27, number of candidates, reservation status for SC/STs, share of female candidates, share of | |
candidates switching political parties (i.e., turncoats), whether the incumbent is running in the election, | |
and whether BJP won the PC in the previous general election. The coefficient on BJP victory in 2014 | |
is large and negative, though is noisily estimated. Therefore, in terms of various characteristics of PCs, | |
the ones that BJP barely lost vs won appear to be comparable. | |
Table 2—Comparability of PCs across BJP Win Margin Threshold | |
Electors Turnout #Candidates SC/ST Female Turncoat Incumbent BJP Won | |
rate reserved share share rerun in 2014 | |
(1) (2) (3) (4) (5) (6) (7) (8) | |
BJP Won 0.178 -0.029 -0.931 0.125 -0.004 -0.002 -0.009 -0.170 | |
(0.172) (0.036) (1.273) (0.154) (0.016) (0.010) (0.172) (0.181) | |
Mean Dep. var. 1.66 0.69 14.46 0.30 0.09 0.03 0.41 0.54 | |
Bandwidth (h | |
∗ | |
) 0.16 0.16 0.16 0.16 0.16 0.16 0.16 0.16 | |
Observations 189 189 189 189 189 189 189 189 | |
Notes: The table reports RDD estimates using BJP win margin on various PC level variables, such as electorate size in | |
millions (column 1), turnout share (column 2), number of candidates (column 3), Reservation status for SC/ST (column 4), | |
share of female candidates (column 5), share of candidates who switched parties (column 6), whether the incumbent is running (column 7) and whether BJP won the PC in 2014 (column 8). The bandwidth used is the optimal bandwidth used for | |
McCrary test in Figure 1a and Table A2. Standard errors are reported in the parentheses. *** p<0.01, ** p<0.05, * p<0.1 | |
V Evidence on Precise Control via Campaigning | |
BJP’s ability to exercise precise control is likely to arise from the party’s superior organizational strength | |
and the related election campaign strategy. It allowed the party to collect and utilize detailed and localized information about voters’ attitudes and voting intentions. This can result in precise control, as the | |
party can mobilize voters better than its opponent in closely contested elections, resulting in disproportionate wins. This is similar to Vogl (2014) who argues that in mayoral elections in southern US, Black | |
voters were better mobilized than White voters, resulting in disproportionate wins of Black candidates | |
in closely contested elections. Therefore, for precise control to be the primary explanation of Figure 1, I | |
expect the constituencies barely won by BJP to have significantly more campaigning by BJP relative to | |
other parties. It is well-established in the literature on political campaigning that a party’s relative campaigning, as opposed to its absolute campaigning activity, matters for its vote share (Bekkouche et al. | |
2022, Gerber 1998, Levitt 1994). Hence, the ideal test would estimate the discontinuity in the relative | |
campaigning by BJP at the BJP win margin value of zero. | |
There is no existing data on campaigning by political parties across constituencies in India. To | |
address this, I exploit a new question added to the post poll National Election Survey in 2019. In the | |
2019 round, the new question asks: “Did a candidate/party worker of the following parties come to your | |
house to ask for your vote in the last one month?” The survey listed the major political parties in each | |
state. The response to this question, therefore, reveals campaigning activity separately by individual | |
political parties in each of the sampled PCs.28 I use this question to define two dummy variables – home | |
27This is calculated using the revised turnout numbers as they are last version of data available with ECI. | |
28In previous rounds, the survey asked whether any candidate or party worker visited the respondent’s house for campaign14 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
Table 3—Campaigning by Political Parties in Closely Contested Elections | |
Home Visit by Party Worker/Candidate | |
Full Sample BJP Ruled States Non-BJP Ruled States | |
BJP Any Other BJP Any Other BJP Any Other | |
Party Party Party | |
(1) (2) (3) (4) (5) (6) | |
Panel A: BJP Win Margin ≤ 0.191 | |
BJP Won 0.03 0.00 0.05 0.34** 0.04 -0.09 | |
(0.10) (0.13) (0.17) (0.17) (0.15) (0.19) | |
Mean Dep. Var. 0.41 0.49 0.33 0.37 0.48 0.58 | |
Bandwidth (h | |
∗ | |
) 0.191 0.191 0.191 0.191 0.191 0.191 | |
Observation 8945 8945 3927 3927 5018 5018 | |
No. of PCs 76 76 32 32 44 44 | |
Panel B: BJP Win Margin ≤ 0.160 | |
BJP Won -0.01 -0.01 0.10 0.38** -0.01 -0.14 | |
(0.11) (0.14) (0.18) (0.17) (0.16) (0.20) | |
Mean Dep. Var. 0.41 0.49 0.33 0.39 0.47 0.57 | |
Bandwidth (h | |
∗ | |
) 0.160 0.160 0.160 0.160 0.160 0.160 | |
Observation 7897 7897 3297 3297 4600 4600 | |
No. of PCs 68 68 28 28 40 40 | |
Notes: The sample is individual level survey data from the National Election Survey (post | |
poll) 2019. The dependent variable in columns (1), (3), (5) is a dummy variable that takes | |
value one if a BJP party worker or candidate visited the house of the respondent to campaign | |
for general election and is zero otherwise. The dependent variable in columns (2), (4) and (6) | |
is also a dummy variable that indicates whether party worker or candidate from any other party | |
visited the house for campaigning. BJP Won is an indicator of whether BJP is the winner of | |
the Parliamentary Constituency (PC). Sample in Panel A consists of PCs with BJP win margin less than 0.191 – the optimal bandwidth calculated using the MSERD method proposed | |
by Calonico et al. (2014), while Panel B uses the sample of PCs with BJP win margin less | |
than 0.16 – the optimal bandwidth used for McCrary test in Figure 1a and Table A2. Columns | |
(1) and (2) use the full sample of PCs within the respective bandwidths. Columns (3) and (4) | |
restrict the sample to states ruled by BJP during 2019 general election. The last two columns | |
use the sample of non-BJP ruled states. Standard errors are clustered at the PC level and are | |
reported in the parentheses. *** p<0.01, ** p<0.05, * p<0.1 | |
visit by BJP, that takes value one if a BJP party worker or candidate visited the respondent’s house, and | |
home visit by any other party, that takes value one if any other party visited the house. The mean values | |
of the two variables are 0.38 (BJP) and 0.50 (any other party). Moreover, they are positively correlated | |
(r = 0.58), suggesting that parties tend to target similar set of “swing” voters. | |
I test whether likelihood of home visits by BJP and other parties increase discontinuously at BJP win | |
margin value of zero, and whether the increase for BJP is larger than other parties. Table 3 reports the | |
results. In Panel A, columns (1) and (2) report the RDD estimates for the two outcome variables using | |
the optimal bandwidth of 0.191. We find that both estimates are small in magnitude and statistically | |
insignificant. In columns (3) and (4), I restrict attention to BJP ruled states, as the failure of McCrary | |
test is concentrated in that sample. The coefficient for BJP home visits is 0.05, which is statistically | |
ing, i.e., it did not ask the question for each party separately. | |
15 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
insignificant, while that for any other party is 0.34, which is statistically significant at 5%. Therefore, in | |
this sample, the estimate for BJP is not larger than that of other parties. For non-BJP ruled states, both | |
estimates again are statistically insignificant. The results remain same if we use 0.16 as the bandwidth | |
(Panel B). Therefore, differentially greater campaigning by BJP relative to other parties cannot be the | |
primary reason for its disproportionate win in closely contested constituencies. | |
Social Media and Election Outcomes: BJP and other major parties extensively used social media | |
during the 2019 election campaign. While smartphone penetration in India is not widespread, social | |
media could potentially play a pivotal role in shaping voting behavior. There is no micro-data on social | |
media campaigning by political parties. The NES 2019, however, asks individuals about their social | |
media usage. For each social media platform, I define a dummy variable (“social media user”) that takes | |
value one if an individual uses that platform at least once every day and zero otherwise. Appendix Table | |
A3 shows how social media usage predicts voting in favor of BJP in the full sample. I find that only | |
Facebook users are statistically significantly more likely to vote in favor of BJP, while users of other | |
social media platforms either vote less for BJP or vote for other parties with equal likelihood.29 For each | |
PC p, I then run the following regression: | |
I(Voted for BJP)ip = αp + βpF b_userip + θ | |
′ | |
pXip + ϵip (1) | |
where the vector of controls Xip includes gender, age and caste categories. βp captures the propensity of | |
Facebook users in constituency p to vote differentially in favor of BJP. The estimate of βp therefore can | |
be interpreted as a proxy of BJP’s differential intensity of Facebook campaigning in the constituency. | |
I use βp as an outcome variable to test whether its value jumps discontinuously at BJP win margin of | |
zero. Appendix Table A4 column 1 reports the RDD coefficient. It is positive and statistically significant, | |
suggesting that constituencies barely won by BJP exhibited relatively more intense social media campaigning by the party. Columns 2 and 3 report the same coefficient when the RDD analysis is performed | |
on BJP ruled and non-BJP ruled states separately. The estimate in column 2 is small and statistically | |
insignificant, while that in column 3 is positive, comparable in magnitude to column 1 and statistically | |
significant. The evidence therefore cannot explain the patterns observed in the previous section that | |
failure of McCrary test is concentrated in the BJP ruled states. | |
VI Evidence on Manipulation | |
Manipulation of elections can take place at one of three stages of elections. First, at the time of voter | |
registration, in the form of targeted deletion of names of voters who are unlikely to vote for the incumbent party. I refer to it as registration manipulation. Second, at the time of voting, when polling officers | |
can strategically discriminate against registered voters, who are likely to vote against BJP. Finally, manipulation can take place at the time of counting of votes.30 Distinguishing between voting and counting | |
29This is consistent with the recent investigative media report that Facebook gave preferential rates to BJP for political ads | |
during 2019 campaign. See here: https://www.aljazeera.com/economy/2022/3/16/facebook-charged-bjp-lower-rates-for-indiapolls-ads-than-others. | |
30A fourth possibility is through manipulation of EVMs. However, some commentators have pointed about that widespread | |
manipulation of EVMs may be hard to achieve, given the technology (Purkayastha and Sinha 2019), making it an unlikely | |
mechanism. | |
16 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
manipulation is difficult. Barring one analysis that comments directly on counting manipulation, the | |
rest of the evidence are consistent with both voting and counting manipulation. Hence, I refer to both | |
as turnout manipulation. Section II above mentions media reports of potential registration and turnout | |
manipulations. In the sections below, I discuss evidence consistent with each of them. | |
VI.I Registration Manipulation | |
To examine the presence of this channel, I compute the growth in the number of electorate (i.e., number | |
of registered voters) in a PC between 2014 and 2019. For each PC p, I define: | |
Gp ≡ | |
Electoratep,2019 − Electoratep,2014 | |
Electoratep,2014 | |
If names were strategically deleted from the electoral rolls in an attempt to flip closely contested election | |
in favor of BJP, then we should expect the electorate growth rate to fall discontinuously at BJP win | |
margin value of zero. Moreover, if Muslims were the primary target of this strategic deletion, we expect | |
a greater fall in the electorate growth rate in PCs with higher Muslim electorate share. I implement the | |
regression discontinuity design on the full sample of PCs as well as on samples of PCs with Muslim | |
share greater and lower than the median of distribution of Muslim shares across PCs.31 | |
Table 4—Electorate Growth Rate Smaller in PCs Barely Won by BJP | |
Electorate Growth Rate (Gp) | |
Full High Low Full High Low | |
Sample Muslim Share Muslim Share Sample Muslim Share Muslim Share | |
(1) (2) (3) (4) (5) (6) | |
BJP Won -0.05*** -0.06** -0.02 -0.05*** -0.07*** -0.03 | |
(0.02) (0.02) (0.02) (0.01) (0.02) (0.02) | |
Mean Dep. Var. 0.10 0.10 0.09 0.09 0.09 0.09 | |
Observations 123 72 51 181 101 80 | |
Bandwidth (h | |
∗) 0.107 0.107 0.107 0.16 0.16 0.16 | |
Notes: The data is at Parliamentary Constituency (PC) level. The table reports the regression discontinuity design | |
estimate using BJP win margin as the running variable. The dependent variable in all the columns is the growth rate | |
in the PC electorate between 2014 and 2019. BJP Won is a dummy indicating whether BJP won the PC in 2019. | |
Columns (1) and (3) use the full sample of PCs. Columns (2) and (4) use PCs where the electorate share of Muslims | |
is higher than the median, and column (3) and (6) use PCs where the share is lower than median. Columns (1)-(3) use | |
the optimal bandwidth using the MSERD method specfied by Calonico et al. (2014), while columns (4)-(6) use the | |
optimal bandwidth calculated for McCrary test in Figure 1a. Robust standard errors are are reported in the parentheses. *** p<0.01, ** p<0.05, * p<0.1 | |
Table 4 shows the estimates for discontinuity for the three samples. Column (1) reports the RDD | |
estimate for the full sample using the optimal bandwidth 0.125. The estimated discontinuity is −0.05, | |
which is statistically significant at 1%. This implies that constituencies barely won by BJP had a 5 | |
percentage points smaller growth rate in electorate between 2014 and 2019 compared to PCs that it | |
barely lost. This is a large fall, given the mean growth rate of 0.09. Moreover, in PCs with higher | |
Muslim share, the estimated fall is 6 percentage points (column (2)), while it is 2 percentage points (and | |
statistically insignificant) in PCs with lower Muslim shares. The difference between the two estimates | |
31I calculate PC level Muslim electorate share by taking a weighted average of AC level Muslim shares using electorate | |
share of an AC as the weight. | |
17 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
in columns (2) and (3) is statistically significant at 10%. The result remains the same if we use the | |
bandwidth of 0.16. The result, therefore, is consistent with strategic deletion of Muslim names being an | |
important channel of manipulation. | |
Additionally, Appendix Table A5 reports the RDD coefficients for high and low Muslim share PCs | |
in the BJP ruled and non-BJP ruled states separately. Consistent with previous results, I find that the | |
only statistically significant coefficient is for the sample of high Muslim share PCs in BJP ruled states | |
(Column (1)). The Column (1) coefficient is also larger in magnitude than that low Muslim share in BJP | |
ruled states (Column (2)), though the difference is not statistically significant. The result for non-BJP | |
ruled states is also similar, though both the coefficients are noisily estimated.32 | |
VI.II Turnout Manipulation: EVM Turnout Data Discrepancy | |
I compile the two different EVM turnout figures described in Section III for the 373 PCs covered in | |
the first four phases of election. In 64% of PCs the turnout was revised up, and in the rest of the | |
cases it was revised down. I compute the absolute difference in vote tallies between the two reports. | |
While the median difference is 358, the 90th and 95th percentiles of the difference are 3302 and 7357, | |
respectively. The largest mismatch is of 57,747 votes in the Gautam Buddha Nagar constituency in Uttar | |
Pradesh. I define a dummy variable called “large” turnout discrepancy: it takes value one if the absolute | |
discrepancy is larger than the 95th percentile and zero otherwise. If the mismatch occurred due to some | |
administrative errors or glitches in the EVM, then we expect the “large” discrepancies to be randomly | |
spread across PCs with different BJP Win Margins. | |
Figure 3—EVM Turnout Data Mismatch in Closely Contested Constituencies | |
(a) “Large” Data Discrepancy (b) Absolute Data Discrepancy | |
Figure 3a plots the relationship between the dummy variable and BJP win margin separately on | |
the two sides of the threshold value of zero. I find that the probability of “large” discrepancy jumps | |
significantly at zero. The estimate of the jump, using the method proposed by Calonico et al. (2014), is | |
0.26 (p-value = 0.008), implying that conditional on close election, the PCs that BJP barely won have 26 | |
percentage point larger likelihood of having a “large” mismatch than PCs that BJP barely lost. This is | |
a large effect considering the average value of the dummy variable, by construction, is 0.05. The result | |
implies that the sample of closely contested constituencies that were disproportionately won by BJP | |
32The p-value of the Column (3) coefficient is 0.103. | |
18 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
also has a disproportionately higher likelihood of “large” turnout revision. If the failure of McCrary test | |
demonstrated in the previous section involved manipulation of turnout figures, then we should expect this | |
pattern. Figure 3b plots the same graph directly using the absolute turnout discrepancy (in thousands) | |
and finds a similar pattern, though with a noisier estimate. The estimate of the discontinuity is 5.70 | |
(p-value = 0.09).33 Hence, the result is consistent with the manipulation hypothesis. Moreover, it is not | |
obvious why precise control would lead to larger turnout revisions by the ECI in the closely contested | |
constituencies won by the BJP. | |
Table 5—Heterogeneity in RDD Estimates of Turnout Discrepancy | |
BJP Ruled States Non-BJP Ruled States | |
(1) (2) | |
Panel A: “Large” Turnout Discrepancy | |
BJP Won 0.45** 0.16 | |
(0.19) (0.11) | |
Panel B: Absolute Turnout Discrepancy | |
BJP Won 15.54** -0.43 | |
(7.42) (1.02) | |
Bandwidth (h | |
∗ | |
) 0.153 0.153 | |
Notes: The table reports RDD estimates for two dependent variables – the dummy variable “Large” Turnout Discrepancy (Panel | |
A) which takes value one if the absolute discrepancy in turnout | |
data is larger than the 95th percentile, and the absolute turnout discrepancy, in thousands (Panel B). The running variable is BJP Win | |
Margin. The sample only includes the 373 constituencies for which | |
turnout discrepancy information is available. Column 1 has states | |
ruled by BJP in 2019, while column 2 has the rest of the states. | |
Optimal bandwidth for Panel A is calculated using the MSERD | |
method proposed by Calonico et al. (2014) and is maintained in | |
Panel B. *** p<0.01, ** p<0.05, * p<0.1 | |
BJP vs. Non-BJP Ruled States: Similar to the previous section, I test for heterogeneity across BJP | |
and non-BJP ruled states. Table 5 reports the RDD estimates for the two sub-samples for both outcome | |
variables. We observe in Panel A that the jump in the probability of “large” discrepancy is statistically | |
significant and large in magnitude for BJP ruled states, while it is statistically insignificant and smaller | |
in magnitude in non-BJP ruled states. The estimated jump in column 1 is 0.45 (and the value just to the | |
left of threshold is close to zero), i.e., the likelihood of “large” discrepancy in the PCs barely won by the | |
BJP is 9 times higher than what it would be under the random chance scenario. It is, on the other hand, | |
3 times higher for non-BJP ruled states. The results in Panel B are similar. While the estimated jump | |
in absolute discrepancy is more than 15, 000 votes (statistically significant at 5%) in BJP ruled states, | |
it is -430 (statistically insignificant) in non-BJP ruled states. The difference between the coefficients in | |
Panel A is not statistically significant, but in Panel B, it is significant at 5%. | |
33Appendix Figure A4 plots the same relationships using INC win margin and does not find discontinuity at the threshold. | |
19 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
Discontinuity in Turnout Difference: Appendix Figure A6 shows discontinuity in turnout difference | |
at the BJP win margin threshold of zero for the past general elections. Turnout difference for a PC in | |
a general election is the difference between its turnout rates in the current and previous elections. We | |
observe that the discontinuity is positive and statistically significant for 2019. The discontinuity estimate | |
is 0.023 (p-value = 0.05). For previous elections going back to 1989, the discontinuity estimates are | |
statistically insignificant. This is consistent with the result that turnout data discrepancy went up in PCs | |
barely won by BJP. | |
Interpretation: It would be inappropriate to treat the data revision as aggregation fraud (Callen and | |
Long 2015), i.e., one where higher level ECI officials directly engaged in turnout manipulation while | |
aggregating turnout data from polling station level results. Such acts by the ECI is unlikely. Moreover, | |
in all cases, barring one, the revision is not larger than the win margin. Rather, the revisions are likely | |
indicative of possible manipulations committed locally, at the polling stations. The local manipulations | |
could either be at the time of voting or counting. Most electoral frauds are decentralized in nature, as | |
Rundlett and Svolik (2016) point out. The analysis in the following section suggests that it was at least | |
partly facilitated by weaker monitoring during counting, while the next section provides evidence that | |
local manipulation may explain part of the observed turnout manipulation. | |
VI.III Counting Manipulation: Assignment of Counting Observers | |
I examine assignment of counting observers in PCs across the BJP win margin threshold. All counting | |
observers are assigned to a state different from their ‘office state’ and ‘home state’, as defined in Section | |
III. 36% of observers are from the SCS cadre. The SCS officers typically work in lower ranked positions | |
in a state bureaucracy, as compared to the IAS officers with same experience (Iyer and Mani 2012). | |
They are also more likely to be politically pliable by the state government, since they are appointed | |
by them, as opposed to the IAS officers who are appointed by the central government. I compute the | |
fraction of counting observers in a PC who come from the SCS cadre. About 50% of PCs have at least | |
one SCS observer assigned. Since I know the ‘office state’ of each observer, I also compute the fraction | |
of observers who are SCS and work in a BJP ruled state. The mean fraction is 0.13. | |
Table 6 columns (1) and (4) report the estimates of discontinuity in the two outcomes variables at the | |
BJP win margin threshold of zero. In both cases, we find that the RDD estimate is positive – 0.24 and | |
0.22, respectively. They are large in magnitude and statistically significant at 5%. Columns (2) and (5) | |
report the results for BJP ruled states and columns (3) and (6) for non-BJP ruled states.34 All coefficients | |
are positive and 3 out of the 4 coefficients are statistically significant; coefficients for BJP ruled states are | |
larger in magnitude. Specifically, for the fraction of observers who are SCS and come from BJP ruled | |
states, the coefficient is 0.37 and is statistically significant at 1% for BJP ruled states (column (5)) but | |
is 0.17 and statistically insignificant for non-BJP ruled states (column (6)). Appendix Figure A5 depicts | |
the RDD graphs for the four cases. The results indicate that more politically pliant counting observers | |
were assigned in PCs barely won by BJP, and the pattern is concentrated in BJP ruled states. | |
I regress the absolute data discrepancy in turnout and the indicator for “large” turnout discrepancy | |
computed in Section VI.II above on the fraction of counting observers who are SCS and from BJP ruled | |
states, whether BJP won the PC and their interaction. I include all PCs in the sample to check whether in | |
34Here BJP and non-BJP ruled states refer to the PCs where the observers were deployed. | |
20 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
Table 6—RDD Estimates of Characteristics of Counting Observers | |
SCS Observer SCS Observer from BJP States | |
All BJP Ruled Non-BJP All BJP Ruled Non-BJP | |
States States Ruled States States States Ruled States | |
(1) (2) (3) (4) (5) (6) | |
BJP Won 0.236** 0.335** 0.241* 0.224** 0.371*** 0.166 | |
(0.100) (0.131) (0.136) (0.104) (0.0739) (0.154) | |
Mean dep. var. 0.36 0.31 0.40 0.17 0.17 0.17 | |
Observations 188 83 105 188 83 105 | |
Bandwidth (h | |
∗ | |
) 0.16 0.16 0.16 0.16 0.16 0.16 | |
Notes: The table reports RDD estimates for two dependent variables – share of counting observers assigned to a PC who are State Civil Service (SCS) officers (columns (1)-(3)) and share of counting observers who are SCS and work in BJP ruled states (columns (4)-(6)). The running variable is BJP Win | |
Margin. The columns (1) and (4) include all PCs with BJP win margin within 0.16, columns (2) and (5) | |
include PCs with the same win margin and are in BJP ruled states, columns (3) and (6) include PCs with | |
the same win margin and are in non-BJP ruled states. Robust standard errors are reported in the parentheses. *** p<0.01, ** p<0.05, * p<0.1 | |
the full sample, assignment of politically pliant observers is correlated with data discrepancy. Appendix | |
Table A6 reports the results. I find that in PCs that BJP lost, the relationship between data discrepancy | |
and the fraction of SCS observers from BJP ruled states is negative. However, in PCs that BJP won, the | |
relationship turns positive for both outcomes. For the absolute discrepancy measure, the relationship is | |
statistically insignificant (p-value=0.104), while for the “large” discrepancy indicator, the relationship | |
is statistically significant at 5% (p-value = 0.041). This suggests that greater presence of politically | |
pliant counting observers in PCs barely won by BJP may have partly contributed towards discrepancy in | |
turnout data. | |
VI.IV Irregularities in Polling Station Outcomes | |
This section examines irregularities in polling station level election results for 2019 from 22 major states | |
of India.35 For each polling station, the data provide information on the total turnout and candidate wise | |
vote tallies. The data do not mention the number of electorates at the polling station level.36 If the | |
turnout discrepancy discussed above electorally benefited the incumbent party, then we should expect it | |
to be reflected in its vote share across polling stations. To examine this, I compute vote share of BJP | |
in each polling station in constituencies with a BJP candidate. To make polling station level BJP vote | |
shares comparable across PCs, I then compute the relative vote share of BJP in each polling station j in | |
each PC p: | |
Relative BJP vote sharejp = | |
BJP vote sharejp | |
BJP vote sharep | |
35The states are Andhra Pradesh, Assam, Bihar, Chhattisgarh, Delhi, Goa, Gujarat, Haryana, Himachal Pradesh, Jharkhand, Karnataka, Kerala, Madhya Pradesh, Maharashtra, Punjab, Rajasthan, Tamil Nadu, Telangana, Tripura, Uttar Pradesh, | |
Uttarakhand, West Bengal. | |
36The only exception is Uttar Pradesh; it releases the electorate size of polling stations along with vote tallies of parties in | |
the same dataset. | |
21 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
where BJP vote sharejp is BJP’s vote share in a polling station and BJP vote sharep is its vote share in | |
the entire constituency that the polling station belongs to. I use the revised turnout figures to calculate | |
vote shares, as that is the official data on turnout and are available for all PCs. Hence, relative BJP vote | |
share captures the party’s vote share in a polling station relative to its vote share in the constituency. | |
If its value is greater than one, then in the polling station, BJP’s vote share is higher than that in the | |
constituency. The average value of Relative BJP vote share should be approximately one. | |
Figure 4a plots, for the states ruled by BJP, the local polynomial relationship between the relative | |
BJP vote share and turnout at polling station level in closely contested constituencies, i.e., constituencies | |
where BJP’s win margin was less than 0.16 – the optimal bandwidth for the McCrary test performed in | |
Figure 1a. I estimate it separately for PCs where BJP won and lost, depicted by the solid and dashed | |
lines respectively. For comparison, Figure 4b plots the same graph for the states not ruled by BJP. Figure | |
4a shows that in both types of constituencies, the estimated relationship hovers around one for polling | |
stations with turnout 800 or below. In polling stations with higher turnout, relative BJP vote share spikes | |
in constituencies where BJP won, and falls in constituencies where BJP lost.37 In Figure 4b we do not | |
see such striking patterns. | |
I examine this directly by creating a dummy variable at the polling station level, called “high” | |
BJP vote share that takes value one if the BJP vote share in the polling station is higher than the 95th | |
percentile of the BJP vote share distribution in the entire sample. In these polling stations, BJP’s vote | |
share on average is 0.90. By construction, the average value of the dummy variable is 0.05. The dummy | |
variable essentially flags polling stations with “extreme” outcomes in favor of BJP. Figures 4c and 4d | |
plot the relationships for BJP ruled and non-BJP ruled states. The sample of PCs is same as before – | |
those with absolute BJP win margin within 0.16. We observe in Figure 4c that in both constituencies won | |
and lost by BJP, the average value of “high” BJP vote share is low in polling stations with turnout below | |
800. However, in larger turnout polling stations, the estimated relationships diverge. In constituencies | |
won by BJP, the average likelihood of “high” BJP vote share increases sharply, going beyond 0.4, while | |
the other graph remains flat. This is consistent with Figure 4a. Moreover, there is no such pattern | |
in the corresponding figure for non-BJP ruled states (Figure 4d). The pattern in Figure 4c is especially | |
noteworthy given the fact that I only consider closely contested constituencies for the estimation. Hence, | |
the rival party in these constituencies have received comparable vote share, which would make it less | |
likely for BJP to get “high” vote shares in any polling station. | |
Benford’a Law: I compute second digit distribution of absolute vote tallies across all candidates for | |
each polling station to check departures from the Benford’s law. Benford’s law specifies the distribution | |
of digits in different positions of naturally occurring numbers (Benford 1938, Raimi 1976). Manipulation of such numbers leads to a different distribution of digits, which allows analysts to detect the | |
manipulation (Hill et al. 1995). This method is used in a variety of contexts to detect fraud (Diekmann 2007, Nigrini 2012) – such as, income tax receipts, financial transactions, as well as elections. | |
In the context of election forensics, analysts usually focus on the distribution of second digits (Mebane | |
37The pattern for PCs lost by BJP is similar in constituencies that BJP lost with margin higher than 0.16 (Appendix Figure | |
A7). Therefore, in all the PCs that BJP lost, it got lower vote share in polling stations with high turnout. These polling stations | |
are likely to be located in urban centers. Since BJP’s primary support base is more urban than other parties, less support in | |
urban areas is a good indicator of its performance in a PC. | |
22 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
Figure 4—Distribution of BJP vote share across Polling Stations – Win margin ≤ 0.16 | |
(a) Relative BJP Vote Share: BJP ruled states (b) Relative BJP Vote Share: non-BJP ruled states | |
(c) “High” BJP Vote Share: BJP ruled states (d) “High” BJP Vote Share: non-BJP ruled states | |
(e) Benford distance: BJP ruled states (f) Benford distance: non-BJP ruled states | |
2008a,b).38 However, treating a significant deviation from Benford’s distribution in a given constituency | |
as evidence of fraud (in that constituency) can lead to misleading conclusions, as researchers have shown | |
that even in cases without fraud, empirical distributions of second digits can deviate from Benford’s law | |
(Shikano and Mack 2011). This can happen due to a myriad of reasons as discussed by Mebane (2011). | |
I therefore do not test for deviations from Benford’s distribution in each PC individually. I argue that | |
presence (or absence) of patterns in deviations across PCs and polling stations is a better statistical test. | |
38There are other digit-based tests of electoral fraud, for example, examining distribution of last digits (Beber and Scacco | |
2012) etc. However, Benford’s law is the most widely used method in this context. | |
23 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
I compute the Euclidean distance between the second digit distribution of the vote tallies of candidates in each polling station and the “ideal” second digit distribution specified by Benford’s law. Figure | |
4e plots the Benford distance for each polling station against turnout for the same sample of PCs as Figure 4c. Figure 4f plots it for the sample of PCs used in Figure 4d. We observe that the Benford distance | |
typically falls with larger turnout, except for PCs won by BJP in BJP ruled states. In those PCs, the | |
Benford distance falls initially with turnout, but then rises for polling stations having turnout higher than | |
800. This is the same set of polling stations that exhibits irregular patterns, as discussed above. This | |
suggests that the “extreme” outcomes in favor of BJP observed in Figure 4c may have resulted from | |
some form of manipulation of vote tallies in that subset of polling stations. | |
Shape of vote share density: Some recent works on detection of electoral fraud examine the shape | |
of the density of vote share and turnout distributions using the booth (or precinct) level data. Rozenas | |
(2017), for example, test for presence of excess mass at coarse vote shares using a re-sampled kernel | |
density method. The method returns as output an estimated fraction of booths exhibiting fraudulent | |
results. I apply this method to the polling station level data from Uttar Pradesh (UP)– the only state | |
for which turnout rate as well as BJP’s vote share are available at polling stations, a requirement for | |
the analysis. Also, a significant number of PCs from UP are in the list of PCs with narrow win margin | |
(Appendix Table B1). I find that in the full sample, 0.13% booths are fraudulent. The share increases to | |
0.19% in PCs with BJP win margin less than 0.08 and won by BJP. The estimates are low but move in | |
the direction that is indicative of fraud. Additionally, for comparison, Rozenas (2017) analyze data from | |
Russian elections in 2011 and 2012 and find estimates of 0.94-0.97%. Klimek et al. (2012) argue that | |
the density of vote share (appropriately calculated and scaled) exhibits high kurtosis in case of fraud and | |
finds that Russian elections in 2011 and 2012 have kurtosis exceeding 10, while in other well-established | |
democracies in Europe, it is typically 5 or lower. In UP, the kurtosis is 29. Appendix Figure A8 shows | |
the density, which looks very similar in shape to those in Russian elections in 2011 and 2012 and unlike | |
those in other countries (Figure 2 in Klimek et al. (2012)). | |
Interpretation: The patterns observed in Figures 4a and 4c are consistent with both mechanisms. If | |
the incumbent party was able to accurately predict the win margins in closely contested constituencies, | |
and wished to affect them, it might be optimal for the party to target the larger polling stations, as they are | |
fewer in numbers and mostly located in urban areas, making voters easily accessible for campaigning. | |
This may result in high vote shares for the party in large turnout polling stations. Figure 4e and the | |
analysis of the shape of BJP’s vote share density, however, advance the manipulation hypothesis over | |
precise control. | |
VI.V Data Discrepancy and Irregularities at Polling Stations | |
To further distinguish between manipulation and precise control mechanisms, I utilize the fact that in | |
closely contested PCs barely won by BJP, the EVM data discrepancy is significantly larger. If discrepancies in turnout data and high vote share of BJP in large polling stations are both driven by manipulation, | |
then I should expect the two phenomena to be correlated, i.e., the irregular pattern in polling stations to | |
be primarily driven by constituencies with larger data revisions. However, if precise control is the explanation, then such patterns should be similar irrespective of whether data revision was large or small. | |
24 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
This is because, in that scenario, larger data revisions only reflect administrative errors during counting of votes, which should be uncorrelated with BJP’s ability to exercise precise control at the time of | |
elections. | |
To test this hypothesis formally, I estimate the difference-in-discontinuity specification (Grembi et al. | |
2016) specified below: | |
Rel. BJP vote sharejp = α1 + γ1BJP_W onp + γ2BJP_W onp × Dp (2) | |
+ β1BJP_M arginp + β2BJP_M arginp × BJP_W onp | |
+ Dp × {α2 + β3BJP_M arginp + β4BJP_M arginp × BJP_W onp}+ϵjp | |
where Dp is the absolute discrepancy in turnout data in PC p, measured in unit of 10,000 votes and the | |
rest are as defined before. γ1 measures the RDD estimate for relative BJP vote share in PCs without any | |
turnout discrepancy. γ2 measures the differential discontinuity in the relative BJP vote share in PCs with | |
additional discrepancy in 10,000 votes. Our coefficient of interest, therefore, is γ2. | |
Table 7—Discrepancy in Turnout Data and Irregularity in Election Results | |
Relative BJP vote share | |
All BJP ruled Non-BJP | |
states states ruled states | |
(1) (2) (3) | |
BJP Won -0.063** -0.105*** -0.094* | |
(0.030) (0.034) (0.048) | |
Absolute Turnout Discrepancy -0.234* -1.738 -0.174* | |
(0.137) (1.059) (0.092) | |
Absolute Turnout Discrepancy * BJP Won 0.236* 1.737 0.196 | |
(0.137) (1.060) (0.130) | |
Mean Dep. Var. 0.998 0.998 0.998 | |
Observations 183,275 82,921 100,354 | |
Bandwidth (h | |
∗ | |
) 0.160 0.160 0.160 | |
Notes: The data is at polling station level. The dependent variable in all columns is the | |
ratio of BJP vote share in a polling station and BJP vote share in the PC. Absolute Turnout | |
Discrepancy is the absolute mismatch (in unit of 10,000 votes) in EVM turnout data in | |
2019. Optimal bandwidth calculated for McCrary test in Figure 1a has been used in all | |
specifications. Standard errors are clustered at the constituency level and are reported in | |
the parentheses. *** p<0.01, ** p<0.05, * p<0.1 | |
Before discussing the results, I emphasize that even though the difference-in-discontinuity method is | |
used to estimate heterogeneity in causal effect of some treatment, that is not the appropriate interpretation | |
in this context. The estimation of equation (3) allows us to examine whether irregular outcomes in the | |
polling stations are positively correlated with extent of turnout discrepancy in PCs barely won by the | |
BJP. Table 7 reports the results for the full sample (column (1)), BJP ruled states (column (2)) and nonBJP ruled states (column (3)). We find that estimate of γ1 is negative and statistically significant in | |
all columns, i.e., PCs barely won by BJP with no data discrepancy exhibits a fall in relative BJP vote | |
share. However, estimate of γ2 in column (1) is positive and statistically significant at 10%. Moreover, | |
the magnitude of γ2 is about 4 times larger than γ1, suggesting that in PCs with discrepancy larger | |
25 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
than 2500 votes, the relative BJP vote share is higher in PCs won by BJP (relative to PCs lost by BJP). | |
Estimate of γ2 in column (2) is 17 times larger than γ1, while in column (3), it is twice as larger. The | |
coefficients in both columns however are noisily estimated. We therefore have weak and suggestive | |
evidence that turnout discrepancy (at the level of PCs) is positively correlated with irregular outcomes | |
at the polling stations in constituencies barely won by the BJP. | |
VI.VI Turnout Manipulation in High Muslim Share Areas | |
This section tests whether electoral discrimination of minorities, specifically Muslims, is a potential | |
source of turnout manipulation. Lehne (2022) shows, using individual voter level panel data on electoral | |
rolls from the state of Uttar Pradesh during 2012-2017, that in state assembly constituencies with BJP | |
incumbents (elected in 2012), Muslim voters have a significantly higher probability of being deleted | |
from the electoral rolls in 2017. Neggers (2018) shows using data from the state of Bihar, that polling | |
officers in charge of conducting election in a polling station exercise significant discretion in allowing | |
registered voters, specially from minority communities such as Muslims, to vote. Since Muslim names | |
are culturally distinct, Muslim voters are easily identified in the electoral roll. Therefore, they can | |
be subject to both strategic deletion (discussed above) and strategic discrimination. Moreover, such | |
exercises are easier in states controlled by the incumbent party, since the state government can influence | |
assignment of officials in charge of electoral roll revisions as well as polling officers. Hence, if fraud is | |
the appropriate explanation, I expect the polling station level irregularities to be concentrated in areas | |
within a PC that have high Muslim presence. | |
Precise control, on the other hand, would predict the opposite. This is because, the party historically | |
enjoyed minimal support among Muslim and consequently, spent significantly less effort in mobilizing | |
Muslim voters. Jha (2017), for example, points out that the party did not focus on areas with significant | |
Muslim presence, since it did not expect to get significant support from them. It instead directed its | |
efforts towards voters who could be converted to vote in favor of the party, especially those belonging | |
to lower castes among Hindus.39 This is consistent with Varshney (2019) who reports, using NES | |
data, that while support for the party increased substantially between 2014 and 2019, especially among | |
Scheduled Castes (SCs) and Other Backward Classes (OBCs) – two large disadvantaged caste groups | |
among Hindus, it remained constant among Muslims. In both elections, only 8 percent of Muslims are | |
reported to have voted for the BJP. Using the data on home visits by party workers, I also find that BJP | |
is significantly less likely to visit Muslim homes compared to non-Muslim homes, while other parties | |
are more likely to visit them (Appendix Table A7). Hence, if the polling station level irregularities are | |
due to exercise of precise control, we should expect it to be concentrated in areas within a PC that have | |
low Muslim share of the electorate. | |
Since Muslim electorate share at the polling stations is not known, I map each polling station to | |
the Assembly Constituency it falls under. Each AC is subsumed within a PC and each PC on average | |
contains about 7 ACs. The data on AC level Muslim electorate share (described in Section III) would | |
provide within-PC variation in Muslim electorate share across polling stations falling in different ACsegments. The final sample for this analysis contains more than 850, 000 polling stations mapped to 3098 | |
ACs (76% of all ACs) covering 475 PCs. The mean Muslim share in an AC is 0.14. However, there is | |
39In Uttar Pradesh, for example, during BJP’s state-wide membership drive it did not focus on the 13, 000 polling stations | |
with significant Muslim presence, since it did not receive any votes in those areas. | |
26 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
wide variation across ACs, with 5 | |
th percentile at 0.01 and 95th percentile being 0.43. Appendix Figure | |
A9 shows the distribution across all ACs. Appendix Table A8 regresses polling station level BJP vote | |
share on AC level Muslim share and finds a sizable and statistically significant negative relationship, both | |
across PCs as well as within a PC. Now, to shed light on the mechanisms discussed above I focus on close | |
election PCs, i.e., those with absolute BJP win margin within 0.16 and run the following specification: | |
Yjap = ϕp + γBJP_W onp ∗ Muslim_shareap + δMuslim_shareap | |
+ β1BJP_W inM arginp ∗ Muslim_shareap | |
+ β2BJP_W onp ∗ BJP_W inM arginp ∗ Muslim_shareap + ϵjap | |
Table 8—Polling Station Level Irregularities Concentrated in High Muslim Share ACs | |
BJP Rel. BJP BJP share | |
vote share vote share ≥ 95th pctile | |
(1) (2) (3) | |
Panel A: All states | |
Muslim Electorate Share in AC -0.745*** -1.811*** -0.194** | |
(0.087) (0.300) (0.085) | |
BJP Won * Muslim Electorate Share in AC 0.327** 0.818** 0.199* | |
(0.127) (0.369) (0.101) | |
Observations 280,391 280,391 280,391 | |
Panel B: BJP Ruled states | |
Muslim Electorate Share in AC -0.633*** -1.282*** -0.250** | |
(0.173) (0.373) (0.124) | |
BJP Won * Muslim Electorate Share in AC 0.367 0.695 0.269* | |
(0.224) (0.474) (0.147) | |
Observations 145,574 145,574 145,574 | |
PC Fixed Effect YES YES YES | |
Bandwidth (h | |
∗ | |
) 0.160 0.160 0.160 | |
Notes: The data is at polling station level. The dependent variables are polling station level | |
BJP vote share (column 1), relative BJP vote share (column 2) and a dummy variable that | |
takes value one if the BJP vote share in a polling station exceeds the 95th percentile (column | |
3). BJP Won is an indicator of whether BJP is the winner of the Parliamentary Constituency. | |
Muslim Electorate Share in AC is the share of Muslim voters in the Assembly Constituency in | |
which a polling station is located. The sample in Panel A is PCs from all states while that in | |
Panel B is BJP ruled states. Optimal bandwidth calculated for McCrary test in Figure 1a has | |
been used in all specifications. Standard errors are clustered at the PC level and are reported | |
in the parentheses. *** p<0.01, ** p<0.05, * p<0.1 | |
where j denotes polling station, a denotes AC and p denotes PC. Yjap is one of three outcome | |
variables – (i) vote share of BJP in a polling station, (ii) relative BJP vote share, as defined above, and | |
(iii) the indicator “high” BJP vote share defined above. ϕp is PC fixed effect and Muslim_shareap | |
is Muslim electorate share in AC a in PC p. The regression implements the difference-in-discontinuity | |
specification with PC fixed effects that subsume the running variable, the treatment BJP Won and their | |
interaction. It compares polling stations within a PC and checks if the BJP’s vote share is high or is more | |
27 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
like to exceed its 95th percentile in AC segments with higher Muslim share and whether this relationship | |
is different between PCs that BJP barely won and lost. δ estimates the relationship in PCs lost by BJP. | |
γ is the differential estimate for PCs won by BJP and is our coefficient of interest. Precise control | |
hypothesis implies γ < 0, while manipulation would imply γ > 0. | |
Table 8 reports the results. The three columns correspond to the three outcome variables mentioned | |
above. Panel A reports the results for the full sample, while Panel B reports it for the BJP ruled states. As | |
before, I restrict attention to PCs with BJP win margin within 0.16. In Panel A, the estimates of δ in all | |
the columns is negative and statistically significant at 1% or 5%. However, the estimates of γ are positive | |
and statistically significant at 10% or 5%. In Panel B, the estimates of δ are also positive, but in columns | |
(1) and (2) they are noisily estimated. The estimate for the dummy indicating “high” BJP vote share in | |
a polling station (column (3)) is large in magnitude and statistically significant at 10%. The estimates of | |
γ and δ jointly indicate that the Muslim electorate share does not predict “extreme” outcomes in favor | |
BJP in PCs barely won by BJP, even though it strongly negatively predicts such outcomes in PCs barely | |
lost by the party. | |
Appendix Table A9 partitions the sample used in Panel B column (3) into polling stations with | |
turnout higher and lower than 800 and estimates the same specification. To allow comparison, column | |
(1) of Table A9 reports the same result as column (3) of Table 8. Columns (2) and (3) report the results for | |
the two sub-samples separately. We find that the the estimate of δ is large in magnitude and statistically | |
significant at 1% in column (2), while it is smaller in magnitude and statistically insignificant in column | |
(3). This is consistent with the graphs reported in Figure 4. | |
40 | |
VII Concluding Remarks | |
The paper documents irregularity in India’s 2019 general election data by showing that the incumbent | |
party’s win margin distribution exhibits excess mass at zero, while no such pattern exists either in previous general elections or in state elections held simultaneously and subsequently. This implies that the | |
incumbent party in 2019 won a disproportionate share of closely contested elections. Moreover, the pattern is concentrated in the states ruled by the incumbent party at that time. While the result is consistent | |
with electoral fraud or manipulation, the incumbent party’s superior ability to predict and affect win margin (i.e., precise control), owing to its significant advantage in electoral campaigning over other parties | |
can also explain it. To isolate the two mechanisms, I conduct a series of analyses to check for presence | |
of precise control and manipulation. I do not find that the incumbent party did greater door-to-door | |
campaigning than other parties in constituencies barely won by it. On the other hand, I find evidence | |
consistent with electoral manipulation at the stage of voter registration as well as at the time of voting | |
and counting (turnout manipulation). In both cases, the results point to strategic and targeted electoral | |
discrimination against Muslims, in the form of deletion of names from voter lists and suppression of | |
their votes during election, in part facilitated by weak monitoring by election observers. | |
The tests are, however, not proofs of fraud, nor does it suggest that manipulation was widespread. | |
Proving electoral manipulation in a robust democracy is a significantly harder task that would require | |
detailed investigation of electoral data in each constituency separately. In the 1960 Presidential election | |
in the US, for example, there was reporting of possible fraud in Illinois state that may have resulted in | |
40The result remains the same if I use turnout threshold of 700 or 600, instead of 800. | |
28 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
John F. Kennedy winning that state. Analysis of detailed data on recounting of votes from Cook county | |
showed patterns consistent with fraud, and yet, were not able to conclusively determine its magnitude | |
and whether it caused the result to flip in Kennedy’s favor (Kallina 1985). This case also highlights | |
that electoral fraud is often decentralized (Rundlett and Svolik 2016), as opposed to being implemented | |
centrally. Consequently, fraud may occur even in contexts where it would not have mattered for government formation. In 1960, Kennedy would have won the Presidential election even if he had lost Illinois. | |
Similarly, in my context, even if manipulation of election data drives all of the observed irregularities | |
in closely contested constituencies, the aggregate election outcomes in terms of government formation | |
would likely have remained unchanged. Appendix Table A10 reports the number of PCs with “excess” | |
BJP wins in closely contested PCs. It varies from 9-18, depending on the definition of a close contest; | |
the numbers are smaller than the lead of 31 PCs that BJP has over the threshold required to form government. Nonetheless, electoral fraud even in a single constituency would imply that such manipulations by | |
incumbent parties are possible. In view of the depletion of trust in electoral processes across the globe | |
and the exceptional integrity of India’s electoral institution in its past, the paper presents a worrying | |
development with potentially far-reaching consequences for the world’s largest democracy. | |
References | |
AGARWAL, P. (2019): “EVM Vote Count Mismatch In 370+ Seats and EC Refuses to Explain,” | |
The Quint, 31 May, 2019, https://www.thequint.com/news/india/lok–sabha–election–results–2019– | |
mismatch–in–votes–polled–and–counted–in–evm–on–multiple–seats. | |
ARRIOLA, L. R., J. DEVARO, AND A. MENG (2021): “Democratic subversion: Elite cooptation and | |
opposition fragmentation,” American Political Science Review, 115, 1358–1372. | |
ASUNKA, J., S. BRIERLEY, M. GOLDEN, E. KRAMON, AND G. OFOSU (2019): “Electoral fraud | |
or violence: The effect of observers on party manipulation strategies,” British Journal of Political | |
Science, 49, 129–151. | |
BALASUBRAMANIAM, V., A. Y. BHATIYA, AND S. DAS (2021): “Behavioral Voters in Synchronized | |
Elections: Evidence from India,” Available at SSRN 3636183. | |
BANERJEE, A., A. GETHIN, AND T. PIKETTY (2019): “Growing Cleavages in India?” Economic & | |
Political Weekly, 54, 35. | |
BANERJEE, M. (2017): “Vote,” South Asia: Journal of South Asian Studies, 40, 410–412. | |
BEBER, B. AND A. SCACCO (2012): “What the numbers say: A digit-based test for election fraud,” | |
Political analysis, 20, 211–234. | |
BEKKOUCHE, Y., J. CAGE, AND E. DEWITTE (2022): “The heterogeneous price of a vote: Evidence | |
from multiparty systems, 1993–2017,” Journal of Public Economics, 206, 104559. | |
BENFORD, F. (1938): “The law of anomalous numbers,” Proceedings of the American philosophical | |
society, 551–572. | |
BHALOTRA, S., I. CLOTS-FIGUERAS, G. CASSAN, AND L. IYER (2014): “Religion, politician identity | |
and development outcomes: Evidence from India,” Journal of Economic Behavior & Organization, | |
104, 4–17. | |
29 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
BHALOTRA, S., I. CLOTS-FIGUERAS, L. IYER, AND J. VECCI (2021): “Leader identity and coordination,” The Review of Economics and Statistics, 1–50. | |
BHOGALE, S., S. HANGAL, F. R. JENSENIUS, M. KUMAR, C. NARAYAN, B. U. NISSA, AND | |
G. VERNIERS (2019): TCPD-IED: TCPD Indian Elections Data v1, Trivedi Centre for Political | |
Data, Ashoka University. | |
BRALEY, A., G. S. LENZ, D. ADJODAH, H. RAHNAMA, AND A. PENTLAND (2022): “The Subversion | |
Dilemma: Why Voters Who Cherish Democracy Participate in Democratic Backsliding,” Aletheia. | |
CALLEN, M. AND J. D. LONG (2015): “Institutional corruption and election fraud: Evidence from a | |
field experiment in Afghanistan,” American Economic Review, 105, 354–81. | |
CALONICO, S., M. D. CATTANEO, AND R. TITIUNIK (2014): “Robust nonparametric confidence | |
intervals for regression-discontinuity designs,” Econometrica, 82, 2295–2326. | |
CANTONI, E. AND V. PONS (2020): “Strict ID Laws Don’t Stop Voters: Evidence from a US Nationwide Panel, 2008–2018,” Quarterly Journal of Economics. | |
CANTÚ, F. (2019): “The fingerprints of fraud: Evidence from Mexico’s 1988 presidential election,” | |
American Political Science Review, 113, 710–726. | |
CAUGHEY, D. AND J. S. SEKHON (2011): “Elections and the regression discontinuity design: Lessons | |
from close US house races, 1942–2008,” Political Analysis, 19, 385–408. | |
CHATURVEDI, R. AND S. CHATURVEDI (2023): “It’s All in the Name: A Character-Based Approach | |
to Infer Religion,” Political Analysis, 1–16. | |
CHRISTENSEN, R. AND T. J. SCHULTZ (2014): “Identifying election fraud using orphan and low | |
propensity voters,” American Politics Research, 42, 311–337. | |
DEMOCRACY REPORT, . (2020): Autocratization Surges – Resistance Grows, V-Dem Institute. | |
——— (2021): Autocratization Turns Viral, V-Dem Institute. | |
DIEKMANN, A. (2007): “Not the first digit! using benford’s law to detect fraudulent scientif ic data,” | |
Journal of Applied Statistics, 34, 321–329. | |
EGGERS, A. C., A. FOWLER, J. HAINMUELLER, A. B. HALL, AND J. M. SNYDER JR (2015): “On | |
the validity of the regression discontinuity design for estimating electoral effects: New evidence from | |
over 40,000 close races,” American Journal of Political Science, 59, 259–274. | |
ENIKOLOPOV, R., V. KOROVKIN, M. PETROVA, K. SONIN, AND A. ZAKHAROV (2013): “Field experiment estimate of electoral fraud in Russian parliamentary elections,” Proceedings of the National | |
Academy of Sciences, 110, 448–452. | |
ESCOBARI, D. AND G. A. HOOVER (2020): “Evo morales and electoral fraud in bolivia: a natural | |
experiment and discontinuity evidence,” Available at SSRN 3492928. | |
FISHKIN, J., A. SIU, L. DIAMOND, AND N. BRADBURN (2021): “Is deliberation an antidote to extreme partisan polarization? Reflections on “America in one room”,” American Political Science | |
Review, 115, 1464–1481. | |
FOA, R. S. AND Y. MOUNK (2016): “The danger of deconsolidation: The democratic disconnect,” | |
Journal of democracy, 27, 5–17. | |
——— (2017a): “The end of the consolidation paradigm,” Journal of Democracy. | |
30 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
——— (2017b): “The signs of deconsolidation,” Journal of democracy, 28, 5–15. | |
FUKUMOTO, K. AND Y. HORIUCHI (2011): “Making outsiders’ votes count: Detecting electoral fraud | |
through a natural experiment,” American Political Science Review, 105, 586–603. | |
GERBER, A. (1998): “Estimating the effect of campaign spending on senate election outcomes using | |
instrumental variables,” American Political science review, 92, 401–411. | |
GRAHAM, M. H. AND M. W. SVOLIK (2020): “Democracy in America? Partisanship, polarization, | |
and the robustness of support for democracy in the United States,” American Political Science Review, | |
114, 392–409. | |
GREMBI, V., T. NANNICINI, AND U. TROIANO (2016): “Do fiscal rules matter?” American Economic | |
Journal: Applied Economics, 1–30. | |
HILL, T. P. ET AL. (1995): “A statistical derivation of the significant-digit law,” Statistical science, 10, | |
354–363. | |
HINTSON, J. AND M. VAISHNAV (2021): “Who Rallies Around the Flag? Nationalist Parties, National | |
Security, and the 2019 Indian Election,” American Journal of Political Science. | |
IYER, L. AND A. MANI (2012): “Traveling agents: political change and bureaucratic turnover in India,” | |
Review of Economics and Statistics, 94, 723–739. | |
IYER, L. AND M. REDDY (2013): Redrawing the Lines: Did Political Incumbents Influence Electoral | |
Redistricting in the World’s Largest Democracy?, Harvard Business School. | |
JAMES, T. S. AND A. CLARK (2020): “Electoral integrity, voter fraud and voter ID in polling stations: | |
lessons from English local elections,” Policy Studies, 41, 190–209. | |
JEONG, D. AND A. SHENOY (2020): “Can the Party in Power Systematically Win a Majority in Close | |
Legislative Elections? Evidence from US State Assemblies,” Journal of Politics, (forthcoming). | |
JHA, P. (2017): How the BJP Wins: Inside India’s Greatest Election Machine, Juggernaut Books. | |
KALLINA, E. F. (1985): “Was the 1960 presidential election stolen? The case of Illinois,” Presidential | |
Studies Quarterly, 113–118. | |
KAPUR, D., P. B. MEHTA, AND M. VAISHNAV (2018): Rethinking public institutions in India, Oxford | |
University Press. | |
KHOSLA, M. AND M. VAISHNAV (2022): “India@ 75: Religion and Citizenship in India,” . | |
KJELSRUD, A., K. O. MOENE, AND L. VANDEWALLE (2020): “The political competition over life | |
and death: Evidence from infant mortality in India,” Tech. rep., Graduate Institute of International | |
and Development Studies Working Paper. | |
KLIMEK, P., Y. YEGOROV, R. HANEL, AND S. THURNER (2012): “Statistical detection of systematic | |
election irregularities,” Proceedings of the National Academy of Sciences, 109, 16469–16473. | |
LEE, D. S. AND T. LEMIEUX (2010): “Regression discontinuity designs in economics,” Journal of | |
economic literature, 48, 281–355. | |
LEHNE, J. (2022): “Incumbents, Minorities and Voter Purges: Evidence from 120 Million Voters’ | |
Registrations in India,” Unpublished Manuscript. | |
LEVITT, S. D. (1994): “Using repeat challengers to estimate the effect of campaign spending on election | |
outcomes in the US House,” Journal of Political Economy, 102, 777–798. | |
31 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
LITTLE, A. AND A. MENG (2023): “Subjective and Objective Measurement of Democratic Backsliding,” Available at SSRN 4327307. | |
MALHOTRA, A. (2019): “Allegations of mass voter exclusion cast shadow on India election,” AlJazeera, | |
30 April, 2019, https://www.aljazeera.com/news/2019/4/30/allegations–of–mass–voter–exclusion– | |
cast–shadow–on–india–election. | |
MANHEIM, L. M. AND E. G. PORTER (2019): “The Elephant in the Room: Intentional Voter Suppression,” The Supreme Court Review, 2018, 213–255. | |
MARTINELLI, A. (2016): “Populism and the crisis of representative democracy,” Populism on the Rise: | |
Democracies Under Challenge, 13–32. | |
MCCRARY, J. (2008): “Manipulation of the running variable in the regression discontinuity design: A | |
density test,” Journal of econometrics, 142, 698–714. | |
MEBANE, W. R. (2008a): “Election forensics: Outlier and digit tests in America and Russia,” in American Electoral Process conference, Center for the Study of Democratic Politics, Princeton University, | |
Citeseer. | |
——— (2008b): “Election forensics: The second-digit Benford’s law test and recent American presidential elections,” Election fraud: detecting and deterring electoral manipulation, 162–181. | |
——— (2011): “Comment on “Benford’s Law and the detection of election fraud”,” Political Analysis, | |
19, 269–272. | |
NAQVI, S. (2022): “The Vanished Minority Voter,” Deccan Herald, February 14, 2022. | |
NATH, A., D. MOOKHERJEE, ET AL. (2017): “Resource Transfers to Local Governments: Political | |
Manipulation and Voting Patterns in West Bengal,” in 2017 Meeting Papers, Society for Economic | |
Dynamics, 1266. | |
NEGGERS, Y. (2018): “Enfranchising your own? Experimental evidence on bureaucrat diversity and | |
election bias in India,” American Economic Review, 108, 1288–1321. | |
NELLIS, G., M. WEAVER, S. ROSENZWEIG, ET AL. (2016): “Do parties matter for ethnic violence? | |
Evidence from India,” Quarterly Journal of Political Science, 11, 249–277. | |
NIGRINI, M. J. (2012): Benford’s Law: Applications for forensic accounting, auditing, and fraud detection, vol. 586, John Wiley & Sons. | |
ONAPAJO, H. AND U. O. UZODIKE (2014): “Rigging through the courts: The judiciary and electoral | |
fraud in Nigeria,” Journal of African Elections, 13, 137–168. | |
PRAKASH, N., M. ROCKMORE, AND Y. UPPAL (2019): “Do criminally accused politicians affect | |
economic outcomes? Evidence from India,” Journal of Development Economics, 141, 102370. | |
PURKAYASTHA, P. AND B. SINHA (2019): “There is No Ghost in the Indian EVM,” Indian Policy | |
Forum. | |
RAIMI, R. A. (1976): “The first digit problem,” The American Mathematical Monthly, 83, 521–538. | |
RAMACHANDRAN, S. (2022): “India State Assembly Elections: Where Is the Level Playing Field?” | |
The Diplomat, 13 January, 2022, https://thediplomat.com/2022/01/india–state–assembly–elections– | |
where–is–the–level–playing–field/. | |
REPUCCI, S. AND A. SLIPOWITZ (2021): “Democracy under siege,” Freedom House. | |
32 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
ROZENAS, A. (2017): “Detecting election fraud from irregularities in vote-share distributions,” Political | |
Analysis, 25, 41–56. | |
RUNDLETT, A. AND M. W. SVOLIK (2016): “Deliver the vote! micromotives and macrobehavior in | |
electoral fraud,” American Political Science Review, 110, 180–197. | |
¸SA ¸SMAZ, A., A. H. YAGCI, AND D. ZIBLATT (2022): “How Voters Respond to Presidential Assaults | |
on Checks and Balances: Evidence from a Survey Experiment in Turkey,” Comparative Political | |
Studies, 00104140211066216. | |
SCHEPPELE, K. L. (2022): “How viktor orbán wins,” Journal of Democracy, 33, 45–61. | |
SHIKANO, S. AND V. MACK (2011): “When Does the Second-Digit Benford’s Law-Test Signal an | |
Election Fraud?” Jahrbücher fur Nationalökonomie & Statistik, 231. | |
SRIDHARAN, E. AND M. VAISHNAV (2017): “Election commission of India,” Rethinking public institutions in India, 417–63. | |
THACHIL, T. (2014): “Elite parties and poor voters: Theory and evidence from India,” American Political Science Review, 108, 454–477. | |
TRIVEDI, D. (2019): “Missing Voters,” Frontline, 26 April, 2019, https://frontline.thehindu.com/cover– | |
story/article26781550.ece. | |
UPPAL, Y. (2009): “The disadvantaged incumbents: estimating incumbency effects in Indian state legislatures,” Public choice, 138, 9–27. | |
VARSHNEY, A. (2003): Ethnic conflict and civic life: Hindus and Muslims in India, Yale University | |
Press. | |
——— (2019): “Modi consolidates power: Electoral vibrancy, mounting liberal deficits,” Journal of | |
Democracy, 30, 63–77. | |
VOGL, T. S. (2014): “Race and the politics of close elections,” Journal of Public Economics, 109, | |
101–113. | |
WALDNER, D. AND E. LUST (2018): “Unwelcome change: Coming to terms with democratic backsliding,” Annual Review of Political Science, 21, 93–113. | |
ZIBLATT, D. (2009): “Shaping democratic practice and the causes of electoral fraud: The case of | |
nineteenth-century Germany,” American Political Science Review, 103, 1–21. | |
33 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
Appendix | |
A Additional Figures and Tables | |
Table A1—Number of Constituencies BJP Won vs Lost in Close Elections | |
2019 2014 2009 2004 | |
(1) (2) (3) (4) | |
Panel A: Absolute BJP Win Margin ≤ 0.05 | |
# (%) of Constituencies BJP Won 41 (69%) 29 (60%) 49 (51%) 47 (59%) | |
# (%) of Constituencies BJP Lost 18 (31%) 19 (40%) 48 (49%) 33 (41%) | |
Total # (%) of Constituencies 59 (100%) 48 (100%) 97 (100%) 80 (100%) | |
Panel B: Absolute BJP Win Margin ≤ 0.03 | |
# (%) of Constituencies BJP Won 28 (74%) 14 (58%) 29 (47%) 22 (56%) | |
# (%) of Constituencies BJP Lost 10 (26%) 10 (42%) 33 (53%) 17 (44%) | |
Total # (%) of Constituencies 38 (100%) 24 (100%) 62 (100%) 39 (100%) | |
Panel C: Absolute BJP Win Margin ≤ 0.02 | |
# (%) of Constituencies BJP Won 20 (74%) 10 (53%) 21 (46%) 18 (64%) | |
# (%) of Constituencies BJP Lost 7 (26%) 9 (47%) 25 (54%) 10 (36%) | |
Total # (%) of Constituencies 27 (100%) 19 (100%) 46 (100%) 28 (100%) | |
Notes: The table reports the number and percentage of constituencies that BJP won and lost and | |
total number of constituencies in 2019 (column 1), 2014 (column 2), 2009 (column 3) and 2004 | |
(column 4) general elections where the BJP’s absolute win margin was less than or equal to 0.05 | |
(Panel A), 0.03 (Panel B) and 0.02 (Panel C). | |
34 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
Table A2—Heterogeneity in Density Jump in BJP Win Margin | |
Panel A | |
BJP Ruled States Non-BJP Ruled States | |
(1) (2) | |
Density jump 3.14*** 0.21 | |
(1.15) (1.03) | |
Bandwidth (h | |
∗ | |
) 0.160 0.160 | |
Notes: The table reports the estimates in the difference in densities | |
of BJP win margin at the threshold value of zero using the McCrary | |
test. Column 1 reports it for the states ruled by BJP in 2019, while | |
column 2 reports for the rest of the states. Optimal bandwidth is | |
calculated using the MSERD method proposed by Calonico et al. | |
(2014). *** p<0.01, ** p<0.05, * p<0.1 | |
Table A3—Social Media Usage and Voting Behavior: NES 2019 | |
Voted for BJP | |
(1) (2) | |
Facebook user 0.035*** 0.027** | |
(0.013) (0.013) | |
Twitter user -0.081*** -0.078*** | |
(0.018) (0.017) | |
Whatsapp user 0.007 -0.008 | |
(0.013) (0.013) | |
Instagram user 0.020 0.017 | |
(0.014) (0.014) | |
YouTube user 0.001 -0.004 | |
(0.013) (0.013) | |
Constant 0.335*** 0.367*** | |
(0.004) (0.013) | |
Observations 22,037 22,037 | |
R-squared 0.002 0.007 | |
Notes: The sample is individual level survey data from the National Election Survey | |
(post poll) 2019. The dependent variable is | |
a dummy indicating whether the individual | |
reported to have voted for BJP in the 2019 | |
election. For any social media platform, an | |
individual is defined to be “user” of the platform if they use it at least once daily. Column | |
2 controls for the gender, age and dummies | |
for caste categories of individuals. Robust | |
standard errors are reported in the parentheses. *** p<0.01, ** p<0.05, * p<0.1 | |
35 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
Table A4—Discontinuity in βp Estimates in Closely Contested Constituencies | |
Dep. Var.: βp | |
Full BJP Ruled Non-BJP | |
Sample States Ruled States | |
(1) (2) (3) | |
BJP Won 0.24*** 0.09 0.27* | |
(0.09) (0.09) (0.16) | |
Bandwidth (h | |
∗ | |
) 0.16 0.16 0.16 | |
Observations 68 28 40 | |
Notes: The dependent variable is the estimate of the coefficient βp from running equation (1) for each PC. The table | |
reports the RDD estimates of a BJP victory using βp as the | |
outcome variable. Column 1 reports it for the full sample, | |
columns 2 reports it for the states ruled by BJP in 2019, and | |
column 3 reports for the rest of the states. Bandwidth used | |
is the optimal bandwidth used for McCrary test in Figure 1a. | |
*** p<0.01, ** p<0.05, * p<0.1 | |
Table A5—Electorate Growth Rate Smaller in PCs Barely Won by BJP | |
Electorate Growth Rate (Gc) | |
BJP Ruled States Non-BJP Ruled States | |
High Low High Low | |
Muslim Share Muslim Share Muslim Share Muslim Share | |
(1) (2) (3) (4) | |
BJP Won -0.06* -0.04 -0.04 -0.01 | |
(0.04) (0.03) (0.02) (0.02) | |
Mean Dep. Var. 0.11 0.10 0.08 0.07 | |
Observations 53 30 48 50 | |
Bandwidth (h | |
∗) 0.16 0.16 0.16 0.16 | |
Notes: The data is at Parliamentary Constituency (PC) level. The table reports the regression discontinuity design estimate using BJP win margin as the running variable. The dependent variable in all the columns is the growth rate in the PC electorate between 2014 and | |
2019. BJP Won is a dummy indicating whether BJP won the PC in 2019. Columns (1) and | |
(2) use the sample of PCs in BJP ruled states, while columns (3) and (4) use PCs in non-BJP | |
ruled states. Columns (1) and (3) use PCs where the electorate share of Muslims is higher | |
than the median, and column (2) and (4) use PCs where the share is lower than median. All | |
columns use the optimal bandwidth calculated for McCrary test in Figure 1a. Robust standard errors are are reported in the parentheses. *** p<0.01, ** p<0.05, * p<0.1 | |
36 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
Table A6—Association between Turnout Data Discrepancy and Counting Observer Characteristics | |
Absolute “Large” | |
Turnout Discrepancy Turnout Discrepancy | |
(1) (2) | |
BJP Won -475.1 -0.0269 | |
(623.9) (0.0229) | |
SCS Counting Observer from BJP Ruled State -463.2 -0.0471 | |
(1,009) (0.0663) | |
BJP Won * SCS Counting Observer from BJP Ruled State 4,373* 0.210** | |
(2,589) (0.104) | |
Constant 1,712*** 0.0544*** | |
(426.3) (0.0196) | |
H0 : β2 + β3 = 0 (p-value) 0.102 0.041 | |
Observations 370 370 | |
R-squared 0.012 0.016 | |
Notes: The data is at Parliamentary Constituency (PC) level. The dependent variable in column (1) is absolute discrepancy in turnout data and in column (2) a dummy variable that takes value one when the absolute discrepancy | |
exceeds 95th percentile of its distribution. BJP Won is a dummy indicator whether BJP won the PC in 2019. SCS | |
Counting Observer from BJP Ruled State is the fraction of counting observers assigned to a PC who are SCS cadre | |
and work in BJP ruled states. Robust standard errors are are reported in the parentheses. *** p<0.01, ** p<0.05, * | |
p<0.1 | |
Table A7—Campaigning among Muslim Voters by BJP and Other Parties | |
Home Visit by Party Worker/Candidate | |
from BJP from Any Other Party | |
(1) (2) (3) (4) (5) (6) | |
Muslim -0.14*** -0.12*** -0.12*** 0.05*** 0.04*** 0.03*** | |
(0.01) (0.01) (0.01) (0.01) (0.01) (0.01) | |
Mean Dep. Var. 0.38 0.38 0.38 0.50 0.50 0.50 | |
Fixed Effect State PC AC State PC AC | |
Observations 24,230 24,230 24,230 24,230 24,230 24,230 | |
No of PCs 208 208 208 208 208 208 | |
R-squared 0.14 0.24 0.31 0.23 0.32 0.39 | |
Notes: The sample is individual level survey data from the National Election Survey (post | |
poll) 2019. The dependent variable in columns (1)-(3) is a dummy variable that takes value | |
one if a BJP party worker or candidate visited the house of the respondent to campaign for | |
general election, and is zero otherwise. The dependent variable in columns (4)-(6) is also a | |
dummy variable that indicates whether party worker or candidate from any other party visited the house for campaigning. Muslim is dummy variable that takes value one if the survey | |
respondent is a Muslim. All columns control for the respondents’ age, age squared, gender | |
and education categories. Columns (1) and (4) have state fixed effects, (2) and (5) have PC | |
fixed effects, and (3) and (6) have AC fixed effects. Robust standard errors are reported in | |
the parentheses. *** p<0.01, ** p<0.05, * p<0.1 | |
37 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
Table A8—Correlation between AC Muslim Share and Polling Station Level BJP Vote Share | |
BJP vote share | |
(1) (2) (3) | |
Muslim Electorate Share in AC -0.37*** -0.31*** -0.43*** | |
(0.02) (0.02) (0.02) | |
Mean Dep. Var. 0.46 0.46 0.46 | |
Observations 674,253 674,253 674,253 | |
State Fixed Effect NO YES NO | |
PC Fixed Effect NO NO YES | |
Notes: The data is at polling station level. The dependent variable in | |
all the columns is BJP vote share in a polling station. Muslim Electorate | |
Share in AC is the share of Muslim voters in the Assembly Constituency | |
in which a polling station is located. Column (1) has no other controls, | |
while columns (2) and (3) have state and AC fixed effects respectively. | |
Standard errors are clustered at the AC level and are reported in the parentheses. *** p<0.01, ** p<0.05, * p<0.1 | |
38 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
Table A9—Polling Station Level Irregularities Concentrated in Polling Stations with High Muslim | |
Shares | |
BJP vote share ≥ 95th pctile | |
BJP Ruled P. S. Turnout P. S. Turnout | |
States ≥ 800 < 800 | |
(1) (2) (3) | |
Muslim Electorate Share in AC -0.250** -0.378*** -0.207 | |
(0.124) (0.073) (0.137) | |
BJP Won * Muslim Electorate Share in AC 0.269* 0.701*** 0.224 | |
(0.147) (0.174) (0.158) | |
Mean Dep. Var. 0.054 0.086 0.052 | |
Observations 145,574 8,310 137,263 | |
PC Fixed Effect YES YES YES | |
Bandwidth (h | |
∗ | |
) 0.16 0.16 0.16 | |
Notes: The data is at polling station level. The dependent variable in all the columns is a dummy | |
variable that takes value one if the BJP vote share in a polling station is larger than the 95th | |
percentile and zero otherwise. BJP Won is an indicator of whether BJP is the winner of the Parliamentary Constituency (PC). Muslim Electorate Share in AC is the share of Muslim electorate | |
in the Assembly Constituency in which a polling station is located. The sample in column (1) is | |
PCs in BJP ruled states with absolute BJP win margin within 0.16. Columns (2) and (3) samples | |
are partitions of the column (1) sample into polling stations with turnout higher than and less | |
than 800, respectively. All columns have PC fixed effect. The optimal bandwidth calculated for | |
McCrary test in Figure 1a has been used in all specifications. Standard errors are clustered at the | |
PC level and are reported in the parentheses. *** p<0.01, ** p<0.05, * p<0.1 | |
39 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
Table A10—Extent of “Excess” BJP Wins in Closely Contested PCs | |
BJP Win Margin | |
≤ 0.07 ≤ 0.05 ≤ 0.03 | |
(1) (2) (3) | |
# Close Election PCs 82 59 38 | |
# “Excess” BJP Wins 18 11 9 | |
Notes: The table reports the number of closely contested | |
Parliamentary Constituencies (PCs) in 2019 and the “excess” number of wins by BJP in those PCs relative to | |
the benchmark of 50% chance of winning. The three | |
columns use three different bandwidths to define a close | |
contest. Column (1) considers BJP win margin within | |
0.07 while columns (2) and (3) consider 0.05 and 0.03 | |
respectively. | |
40 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
Figure A1—Correlation between Muslim Electorate Share and Vote Share of Muslim Candidates | |
41 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
Figure A2—McCrary Test for State Assembly Elections | |
(a) 2019 Elections: BJP Ruled States (b) 2019 Elections: Non-BJP Ruled States | |
(c) 2020-21 Elections: BJP Ruled States (d) 2020-21 Elections: Non-BJP Ruled States | |
42 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
Figure A3—McCrary Test for Non-BJP Ruled States with Strong BJP Presence | |
Figure A4—EVM Turnout Data Discrepancy in Closely Contested Constituencies | |
(a) “Large” Data Discrepancy (b) Absolute Data Discrepancy | |
43 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
Figure A5—SCS Election Observers in Closely Contested Constituencies | |
(a) BJP Ruled States (b) Non-BJP Ruled States | |
(c) BJP Ruled States (d) Non-BJP Ruled States | |
44 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
Figure A6—Turnout Difference in Closely Contested PCs in past General Elections | |
(a) 2019 (b) 2014 | |
(c) 2009 (d) 2004 | |
(e) 1998 (f) 1996 | |
(g) 1991 (h) 1989 45 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
Figure A7—Differential Pattern only in Closely Contested Elections Won by BJP | |
46 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
Figure A8—BJP’s Vote Rate Density in Uttar Pradesh | |
47 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
Figure A9—Distribution of Muslim Electorate Share | |
48 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
B Constituency List | |
BJP Ruled State State Constituency BJP Won | |
(1) (2) (3) (4) | |
1 Assam KARIMGANJ 1 | |
1 Assam NOWGONG 0 | |
1 Bihar PATALIPUTRA 1 | |
1 Goa SOUTH GOA 0 | |
1 Haryana ROHTAK 1 | |
1 Jharkhand DUMKA 1 | |
1 Jharkhand KHUNTI 1 | |
1 Jharkhand LOHARDAGA 1 | |
1 Maharashtra CHANDRAPUR 0 | |
1 Maharashtra NANDED 1 | |
1 Manipur INNER MANIPUR 1 | |
1 Uttar Pradesh SAHARANPUR 0 | |
1 Uttar Pradesh MUZAFFARNAGAR 1 | |
1 Uttar Pradesh MEERUT 1 | |
1 Uttar Pradesh BAGHPAT 1 | |
1 Uttar Pradesh FIROZABAD 1 | |
1 Uttar Pradesh BADAUN 1 | |
1 Uttar Pradesh SULTANPUR 1 | |
1 Uttar Pradesh KANNAUJ 1 | |
1 Uttar Pradesh KAUSHAMBI 1 | |
1 Uttar Pradesh SHRAWASTI 0 | |
1 Uttar Pradesh BASTI 1 | |
1 Uttar Pradesh SANT KABIR NAGAR 1 | |
1 Uttar Pradesh BALLIA 1 | |
1 Uttar Pradesh MACHHLISHAHR 1 | |
1 Uttar Pradesh CHANDAULI 1 | |
1 Uttar Pradesh BHADOHI 1 | |
0 Andaman & Nicobar Islands ANDAMAN & NICOBAR ISLANDS 0 | |
0 Chhattisgarh RAIGARH 1 | |
0 Chhattisgarh KORBA 0 | |
0 Chhattisgarh BASTAR 0 | |
0 Chhattisgarh KANKER 1 | |
0 Dadra & Nagar Haveli DADRA AND NAGAR HAVELI 0 | |
0 Karnataka KOPPAL 1 | |
0 Karnataka BELLARY 1 | |
0 Karnataka TUMKUR 1 | |
49 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 | |
0 Karnataka CHAMARAJANAGAR 1 | |
0 Madhya Pradesh CHHINDWARA 0 | |
0 Odisha SAMBALPUR 1 | |
0 Odisha MAYURBHANJ 1 | |
0 Odisha BALASORE 1 | |
0 Odisha BHADRAK 0 | |
0 Odisha DHENKANAL 0 | |
0 Odisha BOLANGIR 1 | |
0 Odisha KALAHANDI 1 | |
0 Odisha NABARANGPUR 0 | |
0 Odisha PURI 0 | |
0 Odisha BHUBANESWAR 1 | |
0 Punjab HOSHIARPUR 1 | |
0 West Bengal COOCH BEHAR 1 | |
0 West Bengal RAIGANJ 1 | |
0 West Bengal BALURGHAT 1 | |
0 West Bengal MALDAHA DAKSHIN 0 | |
0 West Bengal KRISHNANAGAR 0 | |
0 West Bengal BARRACKPORE 1 | |
0 West Bengal DUM DUM 0 | |
0 West Bengal ARAMBAGH 0 | |
0 West Bengal JHARGRAM 1 | |
0 West Bengal BARDHAMAN DURGAPUR 1 | |
Notes: The table lists the 59 constituencies where the absolute BJP win margin was within 0.05. The first column | |
indicates whether the constituency belonged to a state ruled by the BJP during the 2019 general elections. The last | |
column indicates whether the BJP won that constituency in 2019. | |
50 | |
Electronic copy available at: https://ssrn.com/abstract=4512936 |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment