The Impact of Electronic Voting Machines on Electoral Frauds, Democracy, and Development

Transcription

1 The Impact of Electronic Voting Machines on Electoral Frauds, Democracy, and Development Debnath, Sisir Indian School of Business (ISB) Kapoor, Mudit Indian Statistical Institute (ISI), Delhi Ravi, Shamika Brookings Institution February 9, 2018 Abstract Free and fair elections are the cornerstone of democracy. One of the ways in which elections can be unfair is if there is electoral fraud, which was rampant throughout India till the late 1990s. To combat this problem, the Election Commission of India introduced electronic voting machines (EVMs). We exploit the phased roll-out of the EVMs in state assembly elections to study its impact on electoral fraud, democracy, and development. We find that the introduction of EVMs led to a significant decline in electoral fraud, increased voter turnout among more vulnerable populations, reduced the vote share of incumbent parties, and led to an increase in the provision of electricity, a key public good provided by the government. Keywords: Voting Technology, Electoral Fraud, Democracy, Development, Political Responsiveness JEL Codes: P16, H1 We would like to thank Abhiroop Mukhopadhyay, Sriniketh Nagavarapu, Sugato Dasgupta, Maitreesh Ghatak, Amit Seru, Tarun Jain, Shilpa Aggarwal, John Leahy, Leora Friedberg, Sheetal Sekhri, Rohini Somanathan, Subha Mani, Ravi Jagannathan, Kathleen Hagerty, David Matsa, Robert McDonald for their insightful suggestions. We thank Sandip Sukhtankar and Manasa Patnam for making the state assembly maps available to users. Comments from participants at the Indian School of Business summer camp, University of Virginia brown bag, and Finance seminar at Kellogg also helped to improve the paper substantially. Priyadarshi Amar, Arkodipto Sarkar, Mrithyunjayan Nilayamgode, and Puran Singh provided excellent research support. Existing errors are all ours. sisir debnath@isb.edu mkapoor@isid.ac.in sravi@brookings.edu 1

2 1 Introduction Everyone has the right to take part in the government of his country, directly or through freely chosen representatives. Everyone has the right of equal access to public service in his country. The will of the people shall be the basis of the authority of government; this will shall be expressed in periodic and genuine elections which shall be by universal and equal suffrage and shall be held by secret vote or by equivalent free voting procedures. -Article 21, Universal Declaration of Human Rights Free and fair election of political representatives is a cornerstone of democracy and a fundamental human right. 1 Voting procedure plays a significant role in the election of political representatives in a democracy, as it converts preferences of the voters into a political mandate, which in turn forms the basis for policymaking. Democracies vary in how votes get translated into political representatives - some have presidential systems, some parliamentary systems; within the United States, one chamber is elected proportional to population, the other not. But all democracies rely on votes being cast fairly. In practice, however, illegal efforts to shape electoral outcomes in a democracy are not uncommon (Lehoucq, 2003). 2 Electoral fraud undermines public trust in democratic institutions as elected leaders loose their legitimacy. Leaders elected through unfair means are more likely to divert resources away from the provision of public goods towards the perpetration of further electoral fraud exacerbating the lack of public trust. 3 India is the largest democracy in the world, with more than 800 million registered voters. 4 Electoral fraud has been one of the leading causes of concern in the country. Under the paper ballot system, politicians could hire criminal groups to capture polling booths and stuff ballot boxes with votes in their own favor (Verma, 2005). 5 To address such fraud and simplify the 1 UDHR is a historical document that sets out for the first time, fundamental human rights to be universally protected. 2 Kousser (1974) provides examples of the Democratic Party s use of fraud and violence to regain control of southern politics in the United States at the end of the nineteenth century. 3 Bueno de Mesquita et al. (1999). 4 In terms of the electorate size, India is more than four times the size of the second largest democracy, the Unites States of America. 5 Srinivas (1993) attributed the rise in violence in the early nineties to the politico-economic system and especially to booth capturing. Herstatt and Herstatt (2014) observes that Not only was the paper ballot system perceived as expensive and inefficient, it also had major security problems. One of the major problems is called booth capture. Often it happened that criminal groups, delegated by political parties, captured a polling station and literally stuffed the ballot box with large numbers of votes for the favoured candidate. 2

3 electoral procedure, the Election Commission of India (ECI) introduced electronic voting machines (EVMs) in the late 1990s. An important feature of the EVMs was that it could register only five votes per minute. This feature raised the cost of committing electoral fraud because polling booths would now have to be captured for a significantly longer period of time to rig the election. Besides reducing intentional fraud, the ECI also envisaged that EVMs would improve the efficiency of counting and tallying the electoral results, reducing the incidence of human error. 6 EVMs were introduced in India on an experimental basis in 1998 in a few assembly constituencies at the state level. They were then rolled out in phases for subsequent assembly elections throughout the country. Post-2001, EVMs replaced paper ballots for all state elections in India. The Constitution of India mandates that state assembly elections be held every five years. However, the election cycles of different states are staggered - not every state goes to polls in the same year. These staggered election cycles generate an exogenous variation in the introduction of EVMs between states. Additionally, there was substantial within state variation in the use of voting machines during the first few years of the new technology. For state assembly elections between 1998 and 2001, the ECI used voting machines in some constituencies of a state and paper ballots in others. We exploit these variations, both within and across states, to study the causal impact of EVMs on electoral outcomes, competitiveness of the elections, and measurable development outcomes. Using state assembly elections results between 1976 to 2007 and post-poll survey data, we provide strong evidence that introduction of EVMs led to a significant decline in electoral fraud in India. Under the paper ballot system, criminal groups captured many polling booths and stuffed ballot boxes with fake ballots resulting in an unusually high voter turnout. Using state assembly election data, we show that the introduction of EVMs led to a 3.5 percentage point decline in voter turnout. The decline was substantially larger in states prone to electoral fraud and electoral violence. An alternate explanation for our empirical findings could be that voters don t like using the EVMs, or the long waiting lines outside polling booths due to the limit on the number of votes per minute. To address these issues, we analyse post-poll survey data, which reveals that survey respondents were more likely to be able to cast their votes after introduction 6 The first attempt to abate electoral fraud in India was the introduction of the indelible ink in the 1962 parliamentary election to prevent multiple voting. The ink, still in use, is smeared on the nail and cuticle of the index finger of the right hand after voters have exercised their franchise. It cannot be washed away or erased for at least a week, sometimes up to 15 days. 3

4 of EVMs. This was specifically true for vulnerable categories of citizens such as illiterates, women, scheduled castes and tribes whose votes are typically easier to capture. Citizens also report less fear of violence or vote capturing at the polling booth after the introduction of EVMs. Additionally, we also find that EVMs led to a virtual elimination of rejected or error-ridden votes which were earlier significant in number. All of these patterns suggest that EVMs do not deter voters, but in fact encourage them. Therefore, a reduction in votes cast is a lower bound on the reduction in voter fraud. We also estimate the effect of EVMs on electoral competitiveness. More transparent elections with fewer rejected votes may change the distribution of actual voters, and therefore, the politicians in office. A more transparent election may also induce politicians to align with the preferences of the median voter (Meltzer and Richard, 1981) as predicted by citizen candidate models (Besley and Coate, 1997; Osborne and Slivinski, 1996). We cannot explore these channels separately, but instead, we focus on electoral competitiveness as we examine the effects of EVMs on vote share and chances of re-election for an incumbent party. Areas with the most electoral violence are likely those where violence and fraud are required to keep the incumbent in office, meaning that the incumbent is least likely to win a free and fair election. If EVMs reduce voter fraud, we would expect these areas to have the biggest change in incumbent vote share. We find that the introduction of EVMs leads to a significant decline in the vote share of the incumbent parties, an effect which is concentrated in the states prone to electoral violence. While there was a significantly positive impact of the EVMs on probability of re-election, our results show that this effect is concentrated in states with lower electoral violence. These results suggest that elections became more competitive and transparent, giving the incumbent an edge since they are able to affect the election through winning the votes of the citizens. Finally, we focus on whether the introduction of EVMs had any impact on provision of public goods. In a democracy, electoral goals often determine the distribution of discretionary grants and public goods. Arulampalam et al. (2009) find that swing states in India where the governing parties are aligned with the central government receive larger shares of federal grants. Public utility companies in India, often called State Electricity Boards, control the distribution of electricity within a state. These institutions are susceptible to political capture since the managers are accountable to elected officials in the state (Badiani and Jessoe, 2011). Since fair elections provide the electorate a means to improve the responsiveness of their elected officials 4

5 by making them more accountable (Callen and Long, 2015), we study the impact of EVMs on the provision of electricity in a constituency. Using night-time lights data captured by satellites (luminosity) as a proxy for electricity provision, we find that use of EVMs significantly increased the provision of electricity in constituencies. The improvements were particularly striking in the fourth year after elections, that is, just before the next election. Our paper makes significant contributions to the growing literature on the effect of the voting process on electoral outcomes and development policy. For example, Fujiwara (2015) finds that EVMs reduced error-ridden and invalid votes in elections in Brazil. Paper ballots in Brazil required voters to write down the name of their preferred candidate. This requirement to write automatically led to a disenfranchisement of less educated people from the electoral process. The introduction of electronic voting increased government spending on health care, which disproportionately benefits the disenfranchised under paper ballots. Baland and Robinson (2008) examine the effect of a new secret ballot in Chile in 1958 on the voting behavior of laborers working for a landlord who serves as their patron. They find that before this reform, localities with more pervasive patron-client relationships tended to exhibit a much stronger support for the right-wing parties. After the reform, however, such difference disappeared. However, Card and Moretti (2007) shows that there was no significant effect of electronic voting on election outcomes in the closely contested 2004 US presidential elections. 7 Our paper also contributes to the important discussion on the relationship between democracy and growth. Acemoglu et al. (2014) finds strong evidence that democracy has a positive impact on long term growth. This result is in contrast to earlier work by Barro (1997) who finds that...democracy is not key to economic growth. Our paper complements Acemoglu et al. (2014) by showing the benefits of strengthening democracy in ways that their measure is too coarse to capture. For example, the adoption of EVMs to reduce electoral fraud by booth capturing over our period of interest can be interpreted as strengthening democratic institutions. However, if we were to use the measure of democracy used in Acemoglu et al. (2014), India would have maintained the same status of a free and democratic country during this period. 7 Ferraz and Finan (2011) study the impact of the electoral process on economic outcomes in Brazil. They use random audit reports of the local governments to construct measures of political corruption and combine this data with term-limits of the incumbent mayors. They find that mayors serving their first term with re-election incentives are significantly less corrupt compared to those facing a term limit. 5

6 2 Background 2.1 Political Institutions in India India is a federal parliamentary democratic republic with a dual polity system consisting of the Union Legislature at the national level and State Legislatures at the state level. The bicameral Union Legislature consists of the President of India, the Lower House (Lok Sabha or House of the People) representing the people of India, and the Upper House (Rajya Sabha or Council of States) representing the states of the Indian federation. Those elected or nominated to either house of the Parliament are referred to as Members of Parliament (or MPs). The states follow a similar structure where the Lower House is known as the Legislative Assembly (Vidhan Sabha), and the Upper House is called the Legislative Council (Vidhan Parishad). Those elected to the Legislative Assembly are referred to as Members of Legislative Assembly (or MLAs). Both the union and the states are divided into single-representative constituencies. Candidates compete in elections characterized by a first-past-the-post system, to represent the electoral districts in the legislature. The candidate with a plurality in a given constituency wins the seat. The party or coalition of parties with the majority of seats in the legislature has the first opportunity to form a government. Elections are scheduled quinquennially; although some states may have out-of-turn elections, mostly due to shifting of political alignments. In this paper, we focus primarily on the State Legislative Assembly election results as they exhibit more time-variation in the use of voting machines. 2.2 Electronic Voting Machines in India The objective of using electronic voting machines in India was to strengthen the electoral processes and to reduce the costs of conducting elections. Voting machines were used for the first time, as an experiment, in 1982 in North Paravur assembly by-election in the state of Kerala. Following the initial success, the ECI procured 150,000 machines in 1990 to use them on a national scale. However, the political parties were apprehensive about the security of the machines. A petition was filed questioning the statutory authority of the ECI to use EVMs. The Supreme Court ruled that voting machines could not be used without a necessary provision under the 6

7 law. 8 After the necessary amendments to the Constitution in December 1998, these machines were used in 16 selected constituencies in the state elections in Delhi, Madhya Pradesh, and Rajasthan. 9 These constituencies were were selected on the basis of their compact character and adequate infrastructure to manage the logistics for introducing EVMs. Availability of good road connectivity played a major role so that in the event of malfunctioning, these machines could be promptly replaced. The ECI publicized the usage of EVMs heavily to make sure the process of casting a vote using EVMs was well understood. Voting machines used in India can record a maximum of 3,840 votes. Since the number of registered voters in a polling station does not exceed 1500, the capacity of the machines is sufficient. These machines can accommodate a maximum of 64 candidates. Election officers, who are in charge of ten polling stations on an average, carry spare machines and are responsible for replacing faulty ones. In the event of a breakdown, votes recorded until the machine went out of order remain safe in the memory of the control unit, and it is not necessary to start the poll from the beginning. 10 These machines run on an ordinary 6-volt alkaline battery, and therefore, can be used in areas without electricity connections. All EVMs are provided with an end of poll button which, once pressed, renders them unable to record votes. The use of voting machines simplified the voting procedure and quickened the process of ascertaining results. It also reduced the cost of conducting elections as the ECI could avoid printing millions of ballots. Since EVMs can record only one response per voter, the possibility of rejected votes is virtually eliminated, as opposed to paper ballots where improper and multiple stamps could make a voter s choice unclear. The Goa legislative assembly election in June 1999 was held entirely with voting machines. For the parliamentary election held later in the same year, EVMs were used in parliamentary 45 constituencies out of 543, spread over 17 states covering 60 million voters. Among all the state assembly constituencies scheduled to hold elections in 1999 simultaneously with the parliamentary elections, only those that were within the confinement of these 45 parliamentary 8 To allay the concerns articulated by leaders of political parties about the machines security the ECI commissioned a study by an expert committee in The committee unanimously certified the machines as tamper-proof. A second committee was appointed by the commission in 2006 to evaluate the third-generation machines. In their report, the second committee also reiterated the belief that the machines were tamper-proof. However, some recent independent studies have raised several security issues with voting machines used in India (Wolchok et al., 2010). 9 The percentage of constituencies using EVMs in the states of Delhi, Madhya Pradesh, and Rajasthan were 9%, 2%, and 3%, respectively. 10 The rate of failure of voting machines is less than 0.5 percent. 7

8 constituencies used voting machines. For the state elections in the following year, 2000, once again, only the constituencies that were within the confinement of the same 45 parliamentary constituencies used EVMs. In February 2000, the Commission ordered the use of EVMs in 45 out of 90 Assembly seats in the state of Haryana. Table A.1 in the Appendix reports the fraction of constituencies in a state that used voting machines between Figure 2 plots the timeline of the introduction of EVMs in state assembly constituencies in India. 2.3 Prevention of Electoral Fraud using EVMs A serious concern with the use of paper ballots in India was booth capturing, wherein party loyalists would capture a polling booth by force and stuff the ballot box. 11 The EVMs were designed to discourage such fraud by limiting the rate of vote casting to five per minute. This feature did not rule out booth capturing completely but increased the time required to cast false ballots in large enough numbers, increasing the chances of the arrival of security forces. Second, with the voting machines, the presiding officer could press the close button to disable the device, preventing intruders from taking control of the polling booth. Third, under paper ballots the signature or thumb impression of a voter was recorded on the counterfoil of a ballot, which was not open to inspection except under the orders of a court, whereas under electronic voting, signatures or thumb impressions are maintained in a register which is open to inspection by the public or by anyone willing to file a petition to challenge election outcomes on the grounds of bogus voting. Lastly, paper ballots leave important discretionary decisions in the hands of officers who rule on the validity of unclear votes. The Election Commission has claimed that votes recorded in the machines are tamper proof, and that physical tampering of the devices is easily detectable. However, these claims were later contested by rigorous independent evaluations (Wolchok et al., 2010). 11 Booth capturing is carried out by hired criminals in a very organized way. First, booths are chosen carefully to minimize confrontation with the police and where citizen resistance is likely to be minimal. Second, those booths are targeted which are isolated or guarded by a small police force. Since a single booth is unlikely to affect the overall results, most candidates plan to target as many booths as possible. The candidates hire armed criminals, obtain vehicles and guns before the elections start, and spend considerable amount of money on these resources. (Verma, 2005) 8

9 3 Data 3.1 Election Data The Election Commission of India (ECI) oversees, directs, and controls the Parliament and State Legislative Assembly elections. The results of these elections are published by the ECI in Statistical Reports of General Election to Lok Sabha (Parliament) and Vidhansabha (State Legislative Assembly). The format of these reports vary but overall they publish total number of electors and voters by gender, and the name, gender, party affiliation, and votes secured by each contestant at the constituency level. 12 We use these data to construct voter turnout, rejected votes, and winning margin. 13 Some of these reports provide information on postal ballots and re-poll orders at the constituency level. We use re-poll orders in 2004 Parliamentary elections to find heterogeneity in the main results. We restrict our analysis to the state assembly elections held between and the states where constituency boundaries did not change during this period, even though new states came into existence. 14 A total of 199 state elections were held during this period covering all 30 states and union territories and 4,119 assembly constituencies. 15 We use data from 160 of these elections in our estimation sample. 16 The dates of introduction of the EVMs are collected from ECI orders and several newspaper archives. 12 This data is available at the ECI s website for every general election held since Voter turnout is the ratio of total votes and electors; rejected vote is the difference between total votes and valid votes; winning margin is the difference between the fraction of votes of the winning and runners-up candidate. 14 As per the constitutional and legal provisions, the total number of legislative assembly seats in all states is to remain unaltered till the first census after Given the fixed number of constituencies and differential population growth across regions constituency boundaries are redrawn periodically by the Delimitation Commission to reduce the variation in constituency sizes. The representation from each state is not altered during this exercise. However, the number of reserved seats (by caste) in a state may change as per the most recent census. Till date the Delimitation Commission has been set up four times: 1952, 1963, 1973 and The recommendations of the third and fourth Delimitation Commission were implemented in the year 1976 and 2008, respectively. The first election after the latest round of delimitation was held in the state of Karnataka in Pondicherry and Delhi are the only Union Territories that elect a legislative assembly and are included in our analysis. 16 The 31 state elections that are missing are: Arunachal Pradesh (1978), Delhi (1977, 1983), Goa (1977, 1989, 1907), Jammu & Kashmir (1977, 1983, 1987), Manipur (1980, 1990, 1907), Mizoram (1978, 1984, 1987, 1989), Nagaland (1987), Meghalaya (1978, 1983, 1988), Pondicherry (1977, 1980, 1985, 1991, 1906), Sikkim (1979, 1985, 1989, 1904), and Tripura (1983, 1988). These elections are not included in our analysis mainly because the format of the reports for these elections are different than others. 9

10 3.2 Nighttime Lights We use annual satellite nighttime light images for the period and assembly constituency maps to construct a proxy measure of electricity provision. We obtain the satellite raster images from the National Aeronautics and Space Administration s (NASA) Defense Meteorological Satellite Programs Operational Linescan System (DMSP-OLS), a set of military weather satellites orbiting and recording high-resolution images of earth each night between 20:00 and 21:30 local time. 17 These images are available from 1992 onwards and are used to produce annual composites after dropping cloud cover, Aurora, solar glare (mainly near the poles), and fleeting lights such as forest fires and other noise. We use this series of images after masking the raster data for the geographic boundary of India. These pictures are scaled onto a geo-referenced 30 arc-second grid (approximately one sq. km.). Each pixel is encoded with a measure of its annual average brightness on a 6-bit scale from 0 to 63. Using state assembly constituency boundary maps we extract an annual time series of constituency level luminosity data between Figure 4 shows the nighttime lights images for the years 1992 (in Panel A) and 2007 (in Panel B) overlaid with the assembly boundaries. Political science literature has used nighttime lights extensively as a measure of the provision of electricity. Min et al. (2013) show that nighttime lights imagery can be used to approximate rural electrification in developing countries using DMSP-OLS and survey data from Mali and Senegal. Recent economic growth literature establishes a high degree of correlation between the traditional measures of growth and change in luminosity. Henderson et al. (2012) develop a statistical framework that uses nighttime lights to enhance existing income measures. Interpretation of evolving nighttime lights as growth requires some degree of interpolation. We mainly use nighttime lights as a measure of the provision of electricity which, in India, is a subject matter associated primarily with state-level policy-making. 3.3 Post Poll Surveys We use post-poll surveys conducted by the Centre for the Study of Developing Societies (CSDS), an autonomous organization, to corroborate our analysis using Election Commission data. A research program for studies on comparative democracy at the CSDS regularly conducts large- 17 These high-resolution images, captured at an altitude of 830 km above the earth, record concentrations of outdoor lights, fires, and gas flares at a resolution of 0.56 km and a smoothed resolution of 2.7 km. 10

11 scale scientific studies of political behavior, opinions, and attitudes of Indian electorate. Further, the center is considered a pioneer in large-scale pre-, post-, and exit-poll surveys for parliament and state legislative assembly elections. The first such study was conducted in the state of Kerala in 1965 but was not continued during the 1970s and 1980s. The study resumed in 1995 with a Bihar state assembly election and has covered most state elections since then. We focus on post-poll survey data for state legislative assembly elections during the period Of these elections, voting machines were used in 20 elections, 3 used both EVM and paper ballots, and 1 used only paper ballots. Post-poll surveys are conducted after the day of polling but before the declaration of results. We use the data on whether the eligible voter was able to vote, and whether she abstained from voting due to fear of violence at the polling station, vote capture, or the use of force. The surveys also collect demographic information such as age, gender, caste, and education. Some of the surveys also have questions on awareness of and opinion on EVMs. 4 Estimation Strategy As discussed earlier, our data is restricted to state assembly elections between covering 160 assembly elections and 3,827 constituencies. Each state has its own quinquennial electoral cycle and these cycles are not necessarily aligned. All assembly election after 2002 were held electronically, whereas all elections before 1998 used paper ballots. Therefore, time variation in use of electronic voting emerges due to a pre-determined state election schedule. For example, states that happen to hold an election in 2003 are likely to use EVMs before states that happen to hold an election in However, for the elections between 1998 and 2002, as reported in Table A.1, there was variation in the use of voting technology both within and across states; the ECI used EVMs in 45 out of 543 constituencies in the 1999 parliamentary (Lok Sabha) elections. Since these constituencies were chosen by the ECI on the basis of likelihood of success of EVMs, early adoption in the assembly constituencies inside the geographic confinement of the 45 parliamentary constituencies is likely endogenous. We have therefore dropped them from our main estimation sample. Our identification strategy is based on the quinquennial election cycle. The state elections 18 Assam, Bihar, Chattisgarh, Delhi, Haryana, Jammu & Kashmir, Jharkhand, Kerala, Madhya Pradesh, Meghalaya, Mizoram, Odisha, Punjab, Rajasthan, Tamilnadu, Uttarakhand, and West Bengal. 11

12 are held in different years primarily for historical reasons, which include the year in which the state was formed and the dissolution of state assemblies when incumbent governments lose a vote of confidence. Since EVMs were introduced across India post 2001, for each state the new voting technology got introduced in different years depending on their electoral cycle. For example, consider the states of Bihar and West Bengal and their election years between 1990 and In Bihar, the first elections in the 1990s was held in 1990, election two was in 1995 and election three was in 2000, whereas in West Bengal, the corresponding elections were held in 1991, 1996, and For our analysis, the identification strategy would be that in Bihar, for its three elections between 1990 and 2001, it did not have EVMs, whereas for West Bengal, elections one and two did not use EVMs, but election three in 2001 did. Therefore, in cycle three of the state elections, the treatment effects of the EVMs in West Bengal can be identified using the state of Bihar as a control. Our empirical specification below utilizes intra- and inter- state time variations to estimate the effects of electronic voting machines on outcome variables. Y apt = β 0 + β EV EV apt + τ t + α ap + π ap t + β x x ap(t 1) + ɛ apt (1) where the indices a, p and t represent assembly constituency, parliamentary constituency, and election year, respectively. EV apt is a binary variable taking the value one if assembly constituency a in parliamentary constituency p used electronic voting machines in year t. The coefficient β EV measures the average effect of electronic voting on outcomes such as voter turnout, rejected votes, or winning margin; this coefficient is our main parameter of interest. Assembly fixed effects (α ap ) control for time-invariant unobserved characteristics of constituencies that might have affected the use of voting machines and electoral fraud simultaneously. The year fixed effects (τ t ) control for national events that might have affected the use of voting machines and voter turnout and other electoral outcomes, such as national elections coinciding with state elections. Assembly constituency specific linear time trends (π ap t) capture the trends in assembly constituency characteristics that affect electoral outcomes due to demographic or preference changes and growth. The constituency and time varying characteristics (x ap(t 1) ) include variables that might influence election outcomes, such as the total number of electors, gender of the winning candidate and total number of candidates in the previous election. 12

13 Identification of the effects of electronic voting that use inter-state time variation relies on comparison of two state elections at different time periods. Such comparisons may introduce bias in our estimates as state elections in different time periods might not be comparable. Likewise, identification exploiting the intra-state time variation assumes constituencies that used voting machines in a given state election are not systematically different from their counterparts using paper ballots. Since the variation in election technology is due only to election timing, which in turn is due to historical accidents as described earlier, we expect that any differences between these groups of constituencies will be due to random chance. Table 1 compares these groups on some observable characteristics and does show some significant differences. For example, on an average, there are 58,104 additional eligible voters for elections with voting machines compared to the elections held using paper ballots. The difference in the average number of voters between the two types of voting technologies is 36,364. If these differences only affect the dependent variable directly, then they do not bias our coefficient estimates, since they will get picked up by the assembly constituency specific time trends and election year fixed effects in our model. However, these time-varying unobserved differences may confound our estimates if they also determine the use of voting machines, thereby also having an indirect effect on the dependent variable. We address this issue by estimating the same model after restricting our estimation sample such that covariate balance is achieved between elections with voting machines and paper ballots. We discuss this issue in greater detail in Section Results 5.1 Voter Turnout Effect on Total Voters and Voter Turnout The effects of an electronic voting machine on the number of voters and voter turnout is theoretically ambiguous. Unlike paper ballots, Indian voting machines by default record only five votes per minute. As a result, corrupt politicians have to capture polling booths longer to cast false votes, increasing costs of fraud and the chances of detection. Therefore, electronic voting could reduce the total number of votes by reducing the number of fraudulent votes. Second, voters may turn out in greater numbers in constituencies where machines were used, either 13

14 because they enjoy using the machines more, because they think the elections are fairer, or because the Election Commission s heavy publicity for the machines also served as publicity for the elections themselves. We present the results of electronic voting on the number of voters and voter turnout in Table 3. Panel A in Table 3 reports the effects of EVM on the natural log of total, male and female voters for the period using the specification in equation 1. We find that electronic voting reduced the number of voters by 3.8 percent. 19 The effect varied by gender: male votes decreased by 4.8 percent (column 2), whereas female votes only decreased by 0.67 percent (column 3). Note that, specifications in Panel A control for a log of total-, male- and femaleelectors, respectively. Therefore, the estimated decline in voters is not on account of a spurious correlation between the introduction of voting machines and a simultaneous decline in registered voters. In Panel B, we report the coefficient on voter turnout, defined as the ratio of the number of voters to the number of electors. As Column (1) indicates the overall effects of electronic voting estimated at percentage points continues to be negative and statistically significant at one percent. Drop in male turnout was sharper compared to female turnout supporting the results in Panel B. Under the assumption that electronic voting affects voter turnout only through the reduction of electoral fraud, these estimates suggest that electronic voting machines might have reduced fraudulent votes. Given that the average winning margin for the period of (pre-evm period) was 15.8 percent, a drop in turnout by 3.51 percentage points could affect election outcomes substantially Alternative Explanations Another reason for the decline in votes could be the formation of long lines at polling booths due to the five votes per minute rule, which could discourage voters from exercising their franchise. Voters might also be averse to new technology, which would explain the estimated drop in voter turnout. We test these alternative explanations using the Lokniti post-poll surveys of eligible voters conducted by the Centre for the Study of Developing Societies (CSDS), which collected data on awareness and opinion about voting machines. Voters showed a clear preference for, rather than against, electronic voting percent of the respondents preferred electronic voting machines to paper ballots. Therefore, it is unlikely that the decline in total votes was 19 Since the dependent variable is in logarithmic scale the coefficient estimate of translates into (exp ) = -3.8 percent. 14

15 due to aversion to electronic voting. The survey also collected data on whether voters were able to cast their vote and if not, why. Table 4.A reports the effects of electronic voting on whether an eligible voter was able to cast her vote. The dependent variable takes the value one if she was able to vote and zero otherwise. All specifications reported in the table control for election year and assembly constituency fixed effects, and demographic characteristics of the respondent (age and gender). The results show that electronic voting increased a voter s ability to vote. As indicated in Column (1) the coefficient of Electronic Voting is (0.05 < p < 0.10). This result suggests that introduction of the machines only increased a voter s ability to cast her vote, eliminating concerns about the confounding factor of long lines or aversion to using new technology. In the rest of the columns, we interact Electronic Voting with indicators for several vulnerable groups of electors. The coefficient of the interaction between Electronic Voting and voters with below intermediate education in Column (2) on ability to vote is (p < 0.01), meaning that a voter who lacks an intermediate education was 6.8 percentage points more likely to be able to vote if elections were held using voting machines. Given the baseline average at 0.87 for voters with below intermediate education, the use of EVMs led to a 7.8 percent increase in the likelihood that a less educated voter will cast her vote. In the subsequent columns, we interact Electronic Voting with indicators for women, lower caste, senior citizen, and poorly educated female voters, respectively. All reported coefficients on the interactions are positive and highly significant suggesting the introduction of electronic voting increased the electoral participation of vulnerable groups. These results support our thesis that the drop in total cotes caused by electronic voting comes from a drop in fraudulent votes rather than a drop in legitimate votes. In Table 4.B, we explore the effects of electronic voting on the reasons behind voters inability to vote. The dependent variable measuring vote capture takes the value one if a respondent did not vote due to fear of violence, someone else cast her vote or was prevented from voting and zero otherwise. As in the previous table all specifications control for election years and assembly constituency fixed effects, and demographic characteristics of the respondent. We find that electronic voting had either zero or negative effect on the likelihood of vote capture. We estimate the overall effect of electronic voting to be negative, but statistically insignificant (Column 1). However, when we interact electronic voting with demographic characteristics to estimate the effects on vulnerable populations, we find significant negative effects (Columns 2 15

16 6). For example, voters without an intermediate education are 1.4% less likely to have their vote captured, which is significant at the ten percent level (Column 2). Similarly, for female, lower caste, senior citizen, and less educated female voters the estimated coefficients are measured at [p < 0.01], [0.01 < p < 0.05], [p > 0.1], and [p < 0.01], respectively. The results presented in Tables 4.A and 4.B taken together suggest that electronic voting improved vulnerable populations likelihood of participating in elections Effect on Rejected Votes Under the paper ballot system, voters applied a stamp against the election symbol of their preferred candidate to cast their vote. Unclear or multiple stamping could lead to rejection of the ballot. Unless a voter consciously chooses to waste her vote by marking the ballot in a confounding manner, rejected votes negate the will of the voters who failed to express their preference properly. Electronic voting could reduce the total number of rejected votes, since voters know when their choice is recorded, and so can avoid accidentally not choosing a candidate, and voters are also not allowed to choose more than one candidate, eliminating the possibility of choosing more than one. Both these features reduced the likelihood of a vote getting rejected with electronic voting machines. 20 Therefore, in the absence of electoral fraud, one would expect that with the introduction of EVMs the fraction of rejected ballots to decrease and valid ballots to increase. We analyze the effect of electronic voting on rejected and valid votes separately using the specification in equation 1. Table 5 reports the effect of voting machines on rejected and valid votes. All specifications control for election year fixed effects and parliamentary constituency specific linear time trends. We report the effect of electronic voting on turnout from Table 3 in Column 1 as [p < 0.01]. The effects of electronic voting on rejected and valid votes as a percentage of total electorates are reported in Columns 2 and 3, respectively. While the coefficient on Electronic Voting for rejected votes is negative at [p < 0.01] as expected, the effect on valid votes is estimated at -3.0 [p < 0.01]. These results suggest that reduced rejected votes induced by EVMs are not the principle reason behind the decline in turnout. It is the decline in valid votes that was primarily responsible for the reduced voter turnout. This counterintuitive result reinforces our claim that EVMs lead to a decline in electoral fraud. 20 See Table 1. 16

17 5.2 Heterogeneity in the Effects of EVMs In the previous section, we showed that electronic voting in India has led to a decline in the total number of votes, even though disadvantaged voters reported being more likely to be able to cast their vote. These results suggest that voting machines reduced electoral fraud. In this section, we present the differential effects of electronic voting for states where elections were more likely to be rigged. If the reduction in the total number of votes are biggest where elections were most likely to be rigged, that would again support our hypothesis that the decline in the total number of votes is due to a reduction in electoral fraud Re-poll Orders Since we cannot directly observe electoral fraud, we use re-poll orders issued by the Election Commission as a proxy. The Election Commission appoints observers in every constituency to monitor the elections to ensure free and fair voting. If these observers submit negative reports, the Commission may declare the results from a particular polling booth void and can issue orders for a re-poll. The re-poll orders are consistently available only for the 2004 parliamentary elections. Table A.3 ranks all states in India by per constituency re-poll orders. We create a dummy (Highest Re-poll States) for the four states of India with highest average re-poll orders. 21 As reported in Panel A of Table 6, the coefficients of the interaction between Electronic Voting and the Highest Re-poll States on the log of total, male, and female voters are at , , and , respectively. All of these estimates are significant at the one percent level. The main effect of Electronic Voting on the total number of votes recorded continues to be negative but its magnitude and significance decrease compared to the coefficients reported in Table 3. These results suggest that following the introduction of EVMs the drop in the number of voters and turnout in the four states with highest re-poll orders was almost triple the drop in the rest of the states. In the bottom panel of Table 6 we report that the use of EVM reduced voter turnout by additional 2.3 percent in the four legislative assemblies with highest re-poll orders (Column 1) while the main effects continue to be negative and significant. The heterogeneity in the effects 21 These states, Andhra Pradesh, Bihar, Jharkhand, and Orissa, are densely populated with 18.4 percent of the landmass and 21.8 percent of total population in India. 17

18 of voting machines strengthens our earlier conclusion that drop in voter turnout reflects reduced electoral malpractices Close Elections in States with High Electoral Violence The estimates reported in Section 5.2 using the Lokniti data suggest before electronic voting machines were introduced a large fraction of vulnerable voters were unable to cast their vote due to fear of violence. Such concerns would be greater in the states where electoral violence is more likely and for elections that are close. Close elections are of particular interest because these are the elections that are more likely to have involved fraud, or which are likely to be worth indulging in fraud, given the benefits that can be gained at the margin. We proxy for electoral violence in the state with the spike in murder rates (number of murders per 1000 population) in election years before EVMs were introduced by calculating the average murder rates in election and non-election years in each state over the baseline period ( ), and order the states by the difference between the two. Political science literature often argues that elections with tiny winning margins are very similar and, therefore, are comparable to each other (Butler, 2009; Eggers and Hainmueller, 2009). To visualize the election outcomes for the close elections we plot voter turnout against the winning margin for the elections with and without EVMs in Figure 3. In Panel A, the sample is restricted to the states of Assam, Bihar, Jharkhand, Pondicherry, Punjab, Tripura, and Uttar Pradesh where our measure of electoral violence is the highest. The figure reveals that the average voter turnout is significantly lower in constituencies that used EVMs compared to their counterparts using paper ballots. We do not find any significant difference in voter turnout in Panel B, where the sample is restricted to the states with the lowest electoral violence. 22 We test this more rigorously in Table 7. All specifications reported in the table control for election year and state fixed effects, state specific time trends, gender of the winning candidate and the number of contestants in the previous election. Panel A suggests that, for constituencies with winning margin less than 3 percent, overall, male, and female turnout was lower for the constituencies that used EVMs (Column 1). The effects continue to remain significant (except for female voters) with a quadratic specification measured at (0.01 < p < 0.05), (0.01 < p < 0.05), and (p > 0.10) for overall, male, and female turnout, respectively 22 These states are Andhra Pradesh, Arunachal Pradesh, Karnataka, Kerala, Manipur, Meghalaya, Nagaland, and Sikkim. See Table A.2 for ranking of Indian states by electoral violence. 18

19 (Column 2). The relative magnitudes of these estimates are very similar to those of the effects of EVMs for the entire sample reported in Table 3 earlier. The effects of EVMs on voter turnout continues to persist when the sample is expanded to contain constituencies with winning margin less than 6 percent (Columns 3 and 4). All reported estimates confirm that voter turnout in close elections was significantly lower if elections were held using EVMs. Panel B reports the same estimates after restricting the data to the states of Assam, Bihar, Jharkhand, Pondicherry, Punjab, Tripura, and Uttar Pradesh where the spike in murder rate (murders per 1000 population) during election years was the highest. The effects on overall, male, and female turnout are each several times larger compared to Panel A, and they are all significant at one percent level irrespective of whether the sample is restricted to constituencies with less than 6 percent winning margin or those with less than 3 percent winning margin. Panel C reports the same estimates for the states with smaller election-year spikes in their murder rates. We estimate that the effects of EVMs in close elections for these states are largely non-existent. Systematic differences in the closely contested constituencies using EVMs and paper ballots may confound our results. To rule out such possibilities, we plot constituency characteristics, derived from the 2001 Census data, against winning margin by use of EVMs. 23 Figure 6 plots the predicted values (residuals after taking out the state fixed effects) of a local linear smoother estimated separately for constituencies that used EVM and paper ballots. None of the plotted characteristics reveal any systematic differences between constituencies that use EVMs and paper ballots at close elections. Based on these observations, it is reasonable to conclude that the differential effect observed on voter turnout across states with high and low electoral violence is highly correlated with voting technology Deviations in Voter Turnout While it is evident that electoral fraud can reduce the fairness and competitiveness of representative democracy, and that this directly detracts from the benefits inherent in this political institution, it is less obvious how such malpractices may be observed or measured. In fact, Alvarez et al. (2008) notes that electoral fraud does not even have a commonly accepted definition, particularly because it is largely contextual and subjective in nature. In particular, 23 These characteristics include percent workers engaged in cultivation, urban population, literacy rate (total, male and female), schedule caste, schedule tribes and female population as per the 2001 Census. 19

20 allegations of electoral fraud and abnormal socio-political behavior may provide an indication or evidence for the presence of electoral fraud, but political processes cannot be categorically judged as being either free and fair or otherwise, since the allegations may themselves be politically (tarnishing the opponent) or psychologically (being ungracious in defeat) motivated, and the evidence of suspicious activity may be confounded by several other unobservable factors. If we consider electoral fraud or manipulation, more generally, to be a process that distorts the political intentions of the voters during the electoral process through the use of undue influence, we run into the problem of being unable to observe the original intentions of the voters. The black box that is the secret ballot system not only protects the fairness of the electoral process, but also makes it nearly opaque to any form of auditing. The next best option, perhaps, is to look at the post-election statistics, try to identify any suspicious patterns, and predict the likelihood of electoral fraud. For example, Mebane and Kalinin (2009) uses the Benford s law, the observation that the digits in a number that arises from a large number of natural processes tends to be randomly distributed, whereas the digits of numbers those falsified by human agents tend not to be. This suggests that falsified vote counts may lead a forensic trail that may be observed through studying the distribution of particular digits in the tally. In cases where the fraud is expected to arise through the threat of violence or through ballot stuffing, studies such as Myagkov and Ordeshook (2008) suggest that the distribution of voter turnout may shed some light on electoral malpractice. Klimek et al. (2012) takes this analysis one step further and use a parametric model to quantify the impact of electoral fraud on electoral outcomes. In the Indian context, Singh (2017) also observes that the ECI is likely to allocate security personnel to polling booths based on their likelihood of booth capturing, which is measured using variables including abnormally high voter turnout and highly partisan vote distributions at a given polling booth. Furthermore, Quraishi (2014) notes that very high voter turnout is used by the Election Commission of India (ECI) to flag polling booths that might have been captured. The ideal constituency-level measure for abnormal voter turnout, and thereby for potential electoral fraud, is the proportion of polling booths in the constituency for which the deviation in voter turnout at the booth from the average turnout in the entire state is above a threshold level k (i.e. ). This measure could then be used to study the effect of the introduction of electronic voting on abnormal voter turnout at the constituency level. 20

21 We are constrained from using this ideal measure to identify the likelihood of electoral fraud in a constituency due to a lack of availability of booth-level data on voter turnout. Instead, we choose to measure abnormal turnout using an indicator variable that takes the value 1 when the deviation in average voter turnout in the constituency from the average turnout in the entire state is above a threshold level k and 0 otherwise (i.e. ). Given that this is an indicator variable that takes only the values 0 or 1 as determined ex post by the voter turnout in the constituencies, this may either be an overestimate or an underestimate of the likelihood of electoral fraud as given by the ideal measure. Consider, for example, a constituency whose average voter turnout exceeds that of the whole state by k + ɛ, where ɛ >). Our measure would flag this constituency as a potential source of fraud, whereas the ideal measure would calculate the probability of fraud as the mass of all polling booths with turnout strictly greater than the constituency average or at least k less than the state average (the latter of which is likely a negligible section of the left tail of the turnout distribution across polling booths in the constituency). Clearly, our measure is an overestimate, and the bias decreases as the constituency average turnout moves further away from the state average. Similarly, we can see that our measure is an underestimate for constituencies that are within the deviation threshold from the state average turnout, and that this bias decreases as the constituency average turnout approaches the state average. In other words, our measure of potential electoral fraud is increasingly biased around the threshold deviation in average turnout and exhibits a discontinuity in bias at the threshold, but is a fairly reasonable estimate elsewhere. We then use a Linear Probability Model in Table 8 to look at how the use of EVMs may affect the probability of seeing a given deviation. All specifications control for election-year and assembly-constituency fixed effects, average turnout in the assembly election, the gender of the winning candidate (t 1), the number of contestants (t 1), and assembly constituency specific time trends. We repeat this exercise for 3 values of this deviation, namely 5% (plausibly natural), 10%(improbable), and 15%(unlikely). Note that deviations at this level are seen in approximately 49%, 20%, and 7% respectively of all constituency-election pairs that used paper ballots. The results suggest that the introduction of EVMs does not significantly affect the likelihood of seeing a 5% deviation and only mildly affects the probability of seeing a 10% deviation, and then only in states with a high incidence of electoral violence. However, it does significantly reduce the probability of seeing very large deviations (15%), an effect 21

22 that is particularly pronounced in high-violence states. In particular, the magnitudes of the coefficients are reasonably large compared to the baseline averages, indicating that the effect is not merely statistically significant, but also economically so. This strengthens the argument that the introduction of EVMs reduced electoral fraud, particularly in constituencies where it was rampant. 5.3 Public Goods and Electoral Competition A transparent election may induce the incumbent politician to align themselves with the preferences of the median voter (Meltzer and Richard, 1981). Incumbent politicians often use fiscal policies to improve their chances of re-election. 24 Members of state legislative assemblies have less discretionary power and access to state funds as compared to members of the national parliament (Lok Sabha), but they routinely exploit state-controlled electricity distribution companies (discoms) to increase their chances of reelection. We test this indirectly by exploring the effects of electronic voting on the provision of electricity. We expect that, given the reduction in electoral fraud, politicians in constituencies implementing electronic voting have differentially stronger incentives to increase the provision of electricity in view of an upcoming election. Conversely, citizen candidate models predict that the consequences of a transparent election will manifest through the election of politicians who are closer to the median voter (Besley and Coate, 1997; Osborne and Slivinski, 1996). We explore this channel by investigating the differential impact of electronic voting on vote shares and reelection probability of the incumbent Provision of Electricity State-level corporations are a major player in the energy sector in India. They produce 41% of all electricity and are responsible for its transmission and distribution. Electricity is one of the key issues during state elections and the primary players in its provision are directly under state control. Using transmission loss data for a major state in India, Min and Golden (2014) show that transmission losses peak just before the state assembly elections. Politicians wield considerable power and exercise them to control and manipulate the provision of electricity to tilt election outcomes in their favor (Baskaran et al., 2014). We estimate the effect of electronic voting on electricity provision before elections, using annual constituency-level luminosity data 24 See Drazen (2001) and Alesina et al. (1997) for excellent reviews. 22

23 as a proxy for electricity provided. Due to data limitations, we restrict our analysis to the period between Table 9 reports the time-varying effects of electronic voting on luminosity. All specifications control for election year and assembly constituency fixed effects, assembly constituency-specific time trends, the gender of the winning candidate and the number of contestants in the previous election. The coefficient on Electronic Voting at means that luminosity is higher by 2.5 percent in constituencies that used electronic voting machines. The estimate is significant at one percent level. The coefficients on the interaction between Electronic Voting and the indicators for years since the last election indicate that the effect on log luminosity is much higher in the run-up to the next election. (Column 1) This pattern hints at an attempt by the ruling party to secure more votes at the next election through the provision of electricity. We repeat our analysis while controlling for the interactions between Electronic Voting, dummy variables for years since the last election, and an indicator for states that exhibit high levels of electoral violence 25 (Column 2). The coefficients on the triple interactions suggest that in addition to the increase in luminosity in the run-up to the next election, in the states with high electoral violence, luminosity increases immediately after an election with EVMs and then decreases over time compared to an election with paper ballots. The statistical significance of the coefficients on the interactions change, but the direction of the estimated coefficients remains the same even if we introduce the number of phases in which the election was held as an additional control (Column 3). The estimated coefficients reported in all specifications point out that provision of electricity in constituencies that used voting machines increases as the next election draws nearer Vote Shares and Reelection Weak institutions, lack of information, and political clientelism may lead to the capture of democracy by the political elite (Banerjee et al., 2010). Transparent elections or revealing information to the voters may reduce such inefficiencies. For example, disclosure of the local 25 These states are Assam, Bihar, Jharkhand, Pondicherry, Punjab, Tripura, and Uttar Pradesh. See Table A.2 for ranking of Indian states by electoral violence. 26 Elections in India are held quinquennially, but many of the states have had out-of-turn elections at some point. The date of election also largely depends on historical accident and logistical convenience, meaning that elections could happen any time during the year. These elections also tend to happen in phases, with some parts of the state going to polls days or even weeks before other parts do. 23

24 government s corrupt practices reduced the incumbent s likelihood of reelection in Brazil s municipalities (Ferraz and Finan, 2008). As argued in the previous sections, voting machines in India strengthened the electoral process, making voter fraud more difficult. Therefore, EVMs should improve political competitiveness reflected in the vote share of the incumbents and their reelection chances. We present the effects of electronic voting on vote shares and incumbent party s reelection in Table All specifications reported in Table 10 control for election year and assembly constituency fixed effects, assembly constituency-specific time trends, the gender of the winning candidate and the number of contestants in the previous election. Panel A reports the effects for all state elections. Incumbent party s vote share decreases by 5.34 percentage points with the introduction of voting machines, an estimate which is significant at the 1 percent level (Column 1). This drop represents a percent decrease from the baseline average vote share of 34.75%.. However, the effect on the likelihood of reelection of the incumbent is positive at 0.16, and significant at 1 percent (Column 2). While these results may seem contradictory at first, we believe that this may be the result of two conflicting channels through which electronic voting affects electoral outcomes. On the one hand, reduction in electoral fraud affects all large parties, effectively reducing their vote shares, since fraud is unlikely to significantly affect the votes received by independent candidates or candidates from small regional parties. On the other hand, apart from engaging in fraud, the incumbent has the additional ability to affect re-election by performing well, which he also has a larger incentive to do in light of more transparent elections under electronic voting. Under a first-past-the-post voting system, these channels are likely to lead to reduced vote shares to all major parties as well as increased re-election probabilities for the incumbent. In Panel B, we show how the effects of voting technology vary by levels of electoral violence. The coefficient on the interaction between the indicator for Electronic Voting and the High Electoral Violence States reported in the second row captures this effect. 28 The Incumbent 27 Ideally we want to explore the consequences of the change in voting technology on the incumbent candidate s reelection. Tracking a candidate s performance over several years of election is difficult on several accounts. First, candidate names are not spelled consistently in the election reports over time. Second, parties often field dummy candidates with matching names with their rival candidates. Finally, politicians in India often change their party affiliation making the tracking exercise difficult. Therefore, instead of incumbent candidates we focus on the effects of the change in technology on the incumbent parties. 28 As earlier, High Electoral Violence States takes the value one for the states of Assam, Bihar, Jharkhand, Pondicherry, Punjab, Tripura, and Uttar Pradesh. See Section for more details. 24

25 party s vote share in these states declined by an additional 4.34 percentage points with the introduction of voting machines with the estimate being significant at the five percent level (Column 1). The likelihood of the incumbent party s reelection in the highest re-poll order states increased by 4 percentage points compared to the rest of the states. This point estimate is, however, not significantly different from zero at even the ten percent level (Column 2). These results, combined with the increased provision of public goods, indicate that the introduction of voting machines significantly affected the vote shares of both the incumbent and the winning parties, particularly in states prone to electoral violence, and ended up increasing re-election probabilities across all states, possibly because the incumbent had the additional ability to affect re-election by performing well. 5.4 Robustness Our estimation strategy assumes that constituencies with electronic voting and paper ballots are comparable. However, the summary statistics of observables reported in Table 1 shows significant differences between them. As long as these observed or unobserved differences are time invariant, they will get absorbed by assembly constituency fixed effects. However, timevarying unobserved differences between elections held with electronic voting and paper ballots that are not captured by constituency-specific time trends may confound our estimates if they determine the use of voting machines and election outcomes simultaneously. We address this issue by estimating the same model (equation 1) after restricting our estimation sample to the region of common support on the predicted use of electronic voting. 29 To validate our results we also estimate the effects of voting machines in an instrumental variable framework by restricting the sample to the assembly constituencies that are part of the 45 parliamentary constituencies chosen for the use of electronic voting in the 1999 parliamentary elections, and exploiting the variation in the timing of assembly elections across these constituencies. As a placebo check, we examine whether the introduction of voting machines would have had any impact on the outcomes if we assign its use arbitrarily. Finally, we check for other security factors, in particular, whether greater policing is confounding our results. 29 This restriction is chosen so as to achieve covariate balance between elections held with voting machines and paper ballots. 25

26 5.4.1 Covariate Balance Time variation in the introduction of voting machines in assembly elections was mainly determined by the state election cycles and the use of voting machines in selected constituencies in the 1999 parliamentary election. Therefore, our estimation strategy (equation 1) relies on comparing election outcomes of the same constituencies over time and cross-sectional comparison of constituencies using voting machines and paper ballots. One potential problem with this identification strategy is that constituencies using voting machines early might not have a proper counterfactual in the ones where they were only used later. The significant difference in the average characteristics by voting technology reported in Table 1 means that we cannot rule out this potential problem. Constituency fixed effects in our model eliminate concerns about time-invariant differences, but time-varying unobservable differences may confound our results. To address this issue, we implement the main specification after restricting the data to the region of common support. 30 Note that the differences in average characteristics for the restricted sample (Table A.4) are smaller in magnitude than those for the full sample(table 1) and statistically indistinguishable from zero. The idea behind this exercise is that, since we choose our constituencies to eliminate any differences in observable characteristics, one may expect any difference in unobservables to be also minimized, thereby reducing the impact of confounding factors, if any. The estimation results on voter turnout for the restricted sample is reported in Table A.5, which reports that the overall effect of electronic voting on voter turnout is measured at percentage points, and is significant at the ten percent level (Column 1). The drop in female turnout (-4.46) was sharper than the drop in male turnout (-3.44), once again suggesting that certain groups might be more vulnerable to vote capture. These results on the restricted sample are largely consistent, both in terms of direction and magnitude, with our earlier results on the full sample, and allay concerns regarding systematic unobservable differences between constituencies using different voting technology. The estimated coefficient for the restricted sample 30 Use of electronic voting is predicted using a linear probability model. The control variables include an indicator for use of voting machines in 1999 parliamentary elections, the number of eligible voters, number of candidates, urban population, literacy rate and total population by gender, fraction of the SC-ST population, the percentage of the population engaged in agriculture, and a linear time trend. Constituency level demographic characteristics are constructed using Primary Census Abstract, The choice of the common support between [0.3, 0.55] is determined after visually inspecting the distributions of predicted probability for EVM and paper ballot constituencies (See Figure 5). The results reported in Table A.5 are robust to other choices of intervals of common support or non-linear models for predicting the use of voting machines and are available upon request. 26

27 is larger in magnitude as compared to the estimated effect on the entire sample, reported in Table 3. This suggests that unobservable confounding factors, if any, led to an underestimation of the effect of electronic voting Instrumental Variable Estimates Our estimation strategy exploits within- and between-state variation in the time of the introduction of voting machines. Between-state variations were determined by state election cycles, making them plausibly exogenous, whereas characteristics of the constituencies might have determined variation in the use of voting machines within a state. If constituency fixed effects and constituency specific time trends fail to absorb such factors, then our model yields biased estimates. To address this concern, we create an instrument for the introduction of EVMs to isolate the exogenous variation. In the Lok Sabha elections of 1999, the ECI used EVMs in 45 out of 543 parliamentary constituencies. These constituencies were not randomly selected, and therefore the early use of EVMs in the assembly constituencies within them was not random. If these selection criteria were also correlated with a decline in total votes polled in state assembly elections in these constituencies, we may incorrectly attribute this decline to EVMs. Therefore, we would like to identify an instrument that predicts EVM use but not vote declines. First, to reduce the variations in infrastructural, demographic, and other factors that may have led to the selection of the 45 parliamentary constituencies by the ECI to use EVMs in the 1999 Lok Sabha elections 31, we restrict our sample to assembly constituencies within these 45 parliamentary constituencies, and to assembly elections until 2001, when only about half of these constituencies had had state elections using EVMs. Second, for these assembly elections, the introduction of voting machines was mainly determined by the coincidence of parliamentary and state elections. For the assembly constituencies within these 45 parliamentary constituencies and the states where state assembly and Lok Sabha elections were held together in 1999, the likelihood of using voting machines was much higher. We exploit this fact and use an indicator for the year 1999 as an instrument for electronic voting. The IV approach involves estimating a two-stage model as 31 Given that this was the first time these EVMs were being used, the ECI chose the 45 constituencies keeping in view their compact nature and the availability of necessary logistic so that in case of any malfunctioning of any machine in any booth could be promptly attended or even replaced, if necessary. 27

28 follows: First stage: EV apt = α 0 + α IV I year= τ a t + α a + α x x apt + ɛ apt (2) Second stage: Y apt = β 0 + β EV ÊV apt + τ a t + α a + β x x apt + ν apt (3) We report the two stage least squares (2SLS) estimation results in Table A.6. Panel A reports the first stage results. The coefficient on I year=1999 is 0.99 and it is significant at one percent level. The magnitude of the coefficient and the Kleibergen-Paap F-stat reported at the bottom rules out concerns about relevance and weakness of the instrument. Panel B reports the Two- Stage Least Squares estimate on overall, male, and female turnout, all of which decrease under electronic voting. The effect is highest for female voter turnout at -9.15%. These results are stronger than our previous results from the specification in equation 1, suggesting that previous estimates of the effect of EVM use on turnout were underestimates. However, the 2SLS-IV estimates should be interpreted as Local Average Treatment Effects (LATE) as opposed to the Average Treatment Effect on the Treated (ATT) presented in Table Placebo Year of Introduction of Voting Machines In this section we describe a falsification test to confirm the results reported in Table 3 represent the effects of electronic voting machines. We estimate our main specification in equation 1 after randomly assigning the voting technology either to paper ballots or electronic voting without changing the fractions of elections that use voting machines within each state. We expect the coefficient on the reassigned EVM variable to be, on average, statistically indistinguishable from zero. This process of reassignment and estimation is repeated 100 times, and the means and standard deviations of the estimated effects of shuffled electronic voting are reported in Column (2) of Table A.7. All estimated coefficients on electronic voting are very close to zero and are statistically insignificant. For reference, Column (1) displays the estimated effects as in Panel B of Table Policing and Elections Making elections more transparent is a continuous process. Apart from the introduction of voting machines, other major measures undertaken by the Election Commission include adoption 32 See Angrist and Pischke (2008) and Athey and Imbens (2006) for details. 28

29 of a model code of conduct (1971), annulment of election results in the event of booth capturing (1988), use of photo identity cards to identify voters (1993), and disclosure of assets and criminal antecedents by candidates (2003). Other recent security measures include verifiable paper trail, closed-circuit cameras, and identification of politically sensitive booths. The Election Commission does not have its staff for the administration of elections. Therefore, security personnel is requisitioned from the provincial and central government for policing. To prevent local political rivalries from influencing these officials, they are deputed away from their region of work. Moreover, state elections are conducted in several phases to maximize the use of security forces depending on the size of a state. To test whether factors such as greater policing might have coincided with the introduction of voting machines and confounded our results, we estimate our main empirical specification after controlling for the number of phases in which the election was conducted.the results are reported in Table A.8, where the estimated effect of the number of phases on the natural log of total voters is , significant at the one percent level (Panel A). This suggests that the marginal effect of an additional phase in an election is a decrease in the number of voters by 4.0 percentage points. Note that the inclusion of the number of phases as an additional control does not reduce the magnitude or significance of the effect of voting machines discussed earlier. On the contrary, the effect is slightly more negative than the estimates reported in Table 3. In Panel B, we report the effect of voting machines on voter turnout after controlling for the number of phases for each assembly election. The effect of voting machines is slightly more negative compared to the baseline results in Table 3 and continues to be significant at the one percent level. 6 Conclusion Free and fair elections are a cornerstone of democracy. In the 1990s, the Election Commission of India (ECI) introduced electronic voting machines (EVMs) to address electoral frauds and simplify the electoral procedure. Using electoral data from all the state assembly elections in India (largest democracy in the world) from 1976 to 2007, we analyse the effect of EVMs on electoral fraud, democracy, and select development outcomes. First, we looked at the impact of EVMs on electoral fraud. Though electoral fraud is difficult to detect, it is well documented in India. Prior to the introduction of EVMs, in many 29

30 constituencies under the paper ballot system, polling booths were captured, and ballot boxes were stuffed with fake ballots which resulted in an unusually high voter turnout. Our estimates show that voter turnout and the number of rejected votes declines significantly with the use of EVMs. Moreover, independently collected post-poll survey data shows that the introduction of EVMs led to greater participation of the marginalized and vulnerable voters. These groups are also likely to report fewer instances of rigging or intimidation. We find evidence that the decline in voter turnout with the use of EVMs was more pronounced in states with a higher incidence of electoral violence and where the Election Commission was more likely to issue repoll orders. These results strongly suggest that introduction of EVMs reduced electoral fraud in India. We also find strong evidence that introduction of EVMs made the electoral process more competitive because it led to a decline in the winning margin and vote share of the winning candidate. We extend the analysis to check if the introduction of EVMs which reduced electoral fraud had any impact on measurable development outcomes. This is important because electoral goals often determine the distribution of discretionary grants and public goods in a democracy. Fair and competitive elections provide the electorate a means to improve the responsiveness of the elected officials by making them more accountable. We extend our study to the impact of EVMs on the provision of electricity within a constituency. We find that the constituencies using EVMs have better provision of electricity than their counterparts using paper ballots. The provision of electricity improves over time, and the effect is strongest for the year just before the subsequent election. The introduction of the machines may also have changed the composition of the voters as well as the characteristics of the elected politicians. It is difficult to identify the exact mechanisms or pathways given our data. This paper makes a significant contribution to the literature on democracy and development. We show that introduction of EVMs lead to strengthening of democratic institutions and better representation of marginalized and vulnerable segments of the society. We further show that stronger democratic institutions reduce electoral fraud and make elected officials more accountable. We find this in the improved provision of public goods such as electricity within constituencies with EVMs. 30

31 References Acemoglu, D., Naidu, S., Restrepo, P. and Robinson, J. A. (2014). Democracy Does Cause Growth. NBER Working Paper No Alesina, A., Roubini, N. and Cohen, G. (1997). Political Cycles and the Macroeconomy. Cambridge, MA: MIT Press. Alvarez, R. M., Hall, T. E. and Hyde, S. D. (eds.) (2008). Introduction: Studying Election Fraud, Brookings Institution Press, pp Angrist, J. D. and Pischke, J.-S. (2008). Mostly Harmless Econometrics: An Empiricist s Companion. Princeton University Press, 1st edn. Arulampalam, W., Dasgupta, S., Dhillon, A. and Dutta, B. (2009). Electoral Goals and Center-State Transfers: A Theoretical Model and Empirical Evidence from India. Journal of Development Economics, 88 (1), Athey, S. and Imbens, G. W. (2006). Identification and Inference in Nonlinear Difference-in- Differences Models. Econometrica, 74 (2), Badiani, R. and Jessoe, K. K. (2011). Elections at What Cost? The Impact of Electricity Subsidies on Groundwater Extraction and Agricultural Production. Working Paper. Baland, J. M. and Robinson, J. A. (2008). Land and Power: Theory and Evidence from Chile. American Economic Review, 98 (5), Banerjee, A., Kumar, S., Pande, R. and Su, F. (2010). Do informed voters make better choices? experimental evidence from urban india. Unpublished manuscript. povertyactionlab. org/node/2764. Barro, R. J. (1997). Getting it Right: Markets and Choices in a Free Society. The MIT Press. Baskaran, T., Min, B. and Uppal, Y. (2014). Election Cycles and Electricity Provision: Evidence from a Quasi-experiment with Indian Bye-elections. Working Paper. Besley, T. and Coate, S. (1997). An Economic Model of Representative Democracy. The Quarterly Journal of Economics, 112 (1), Bueno de Mesquita, B., Morrow, J. D., Siverson, R. M. and Smith, A. (1999). An Institutional Explanation of the Democratic Peace. American Political Science Review, 93 (4), Butler, D. M. (2009). A Regression Discontinuity Design Analysis of the Incumbency Advantage and Tenure in the US House. Electoral Studies, 28 (1), Callen, M. and Long, J. D. (2015). Institutional Corruption and Election Fraud: Evidence from a Field Experiment in Afghanistan. American Economic Review, 105 (1),

32 Card, D. and Moretti, E. (2007). Does Voting Technology Affect Election Ouctomes? Touch Screen Voting and the 2004 Presidential Election. The Review of Economics and Statistics., 4. Drazen, A. (2001). The Political Business Cycle After 25 Years. In NBER Macroeconomics Annual 2000, Volume 15, NBER Chapters, National Bureau of Economic Research, Inc, pp Eggers, A. c. and Hainmueller, J. (2009). MPs for Sale? Returns to Office in Postwar British Politics. American Political Science Review, 103, Ferraz, C. and Finan, F. (2008). Exposing Corrupt Politicians: The Effects of Brazil s Publicly Released Audits on Electoral Outcomes. The Quarterly Journal of Economics, 123(2), and (2011). Electoral Accountability and Corruption: Evidence from the Audits of Local Governments. American Economic Review, 101 (4), Fujiwara, T. (2015). Voting Technology, Political Responsiveness, and Infant Health: Evidence from Brazil. Econometrica, 83 (2), Henderson, J. V., Storeygard, A. and Weil, D. N. (2012). Measuring Economic Growth from Outer Space. American Economic Review, 102 (2), Herstatt, M. and Herstatt, C. (2014). India s Electronic Voting Machines (EVMs): Social construction of a frugal innovation. Working Papers 86, Hamburg University of Technology (TUHH), Institute for Technology and Innovation Management. Klimek, P., Yegorov, Y., Hanel, R. and Thurner, S. (2012). Statistical detection of systematic election irregularities. Proceedings of the National Academy of Sciences, 109 (41), Kousser, J. M. (1974). The shaping of southern politics: Suffrage restriction and the establishment of the one-party south, Yale University Press. Lehoucq, F. (2003). Electoral Fraud: Causes, Types, and Consequences. Annual Review of Political Science, 6 (1), Mebane, W. R. and Kalinin, K. (2009). Comparative election fraud detection. Meltzer, A. H. and Richard, S. F. (1981). A Rational Theory of the Size of Government. Journal of Political Economy, 89 (5), Min, B., Gaba, K. M., Sarr, O. F. and Agalassou, A. (2013). Detection of Rural Electrification in Africa using DMSP-OLS Night Lights Imagery. International Journal of Remote Sensing, 34 (22),

33 and Golden, M. (2014). Electoral Cycles in Electricity Losses in India. Energy Policy, 65, Myagkov, M. G. and Ordeshook, P. C. (2008). Russian Elections: An Oxymoron of Democracy. National Council for Eurasian and East European Research Seattle, WA. Osborne, M. J. and Slivinski, A. (1996). A Model of Political Competition with Citizen- Candidates. The Quarterly Journal of Economics, 111 (1), Quraishi, S. Y. (2014). An Undocumented Wonder: The Great Indian election. Rupa Publications. Singh, D. (2017). Safer elections and women turnout: Evidence from india. Srinivas, M. N. (1993). Changing Values in India Today. Economic and Political Weekly, 28 (19), Verma, A. (2005). Policing Elections in India. India Review, 4 (3-4), Wolchok, S., Wustrow, E., Halderman, J. A., Prasad, H. K., Kankipati, A., Sakhamuri, S. K., Yagati, V. and Gonggrijp, R. (2010). Security Analysis of India s Electronic Voting Machines. Proceedings of the 17th ACM Conference on Computer and Communications Security, pp

34 Figure 1: Paper Ballots and Electronic Voting Machines used in India. 34

35 Figure 2: Time-line of Introduction of Electronic Voting Machines in State Assembly Constituencies in India. Notes: The years of the introduction of electronic voting machines are obtained from the election commission s orders. 35

36 Figure 3: Turnout and Winning Margin by States: Local Polynomial Fit. 10 Panel A: States with High Electoral Violence Residual Turnout Winning Margin Panel B: States with Low Electoral Violence 5 Residual Turnout Winning Margin Electronic Voting Paper Ballots Notes: Data include results from state assembly elections between published by the Election Commission. In Panel A the data is restricted to the states of Assam, Bihar, Jharkhand, Pondicherry, Punjab, Tripura, and Uttar Pradesh, where electoral violence is the highest. In Panel B the data is restricted to the states of Andhra Pradesh, Arunachal Pradesh, Karnataka, Kerala, Manipur, Meghalaya, Nagaland, and Sikkim, where electoral violence is the lowest. The y-axis plots the residual turn-out after taking out assembly constituency fixed effects. 36

37 Figure 4: Nighttime Lights in India with Assembly Constituency Boundary. Panel A: 1992 Panel B: 2007 Notes: Satellite raster images are obtained from the National Aeronautics and Space Administration s (NASA) Defense Meteorological Satellite Programs Operational Linescan System (DMSP-OLS) is the first year and 2007 is the last year of our data. 37