Do Criminally Accused Politicians Affect Economic Outcomes - Evidence from India Ch 2

4.2 Election Results and Affidavit Data We use the Election Commission of India (ECI) Statistical Reports on General Elections to State Legislative Assemblies for election data.24 These reports provide important information related to elections, such as the name and code of the constituency, candidates and their vote shares, electorate size (number of registered voters), number who turned out to vote (number of voters), the candidate’s gender and constituency type (whether reserved for SCs (STs) or non-reserved).25 For criminal accusations, we rely on the data from sown affidavits that have been collected and digitalized by the Election Watch, in collaboration with the Association for Democratic Reforms (ADR).26 The ADR data provide information on the number of criminal cases against each candidate, the charges associated with each criminal case, a classification of each accusation as serious or not, the asset and liabilities disclosures of each candidate and each candidate’s level of education. We consider all state elections held between 2004 and 2008. While the light data are available from 1992 onwards, we are limited by the data on affidavits, which became mandatory only after the Supreme Court order in 2003. Further, ADR data are available only for elections held after 2004. As a result, we have a sample of 20 states out of a total of 28 covering approximately 90 percent of India’s total electorate.27 Also, the constituency boundaries changed in 2008 meaning constituencies before and after delimitation are not comparable.28 Thus, between the Court order to file affidavits in 2003 and the redrawing of boundaries in 2008, we observe only 1 election per state. However, we utilize the night-lights data until 2012 for some states.29 Table A-1 reports the information on the 24 The reports are available at http://eci.nic.in/eci_main1/ElectionStatistics.aspx, accessed in May 2014. According to the Indian Constitution, certain seats are reserved for Scheduled Castes (SCs) and Scheduled Tribes (STs), the two historically disadvantaged minority groups. While registered voters from all social groups can vote, only an SC (ST) candidate may contest election from the seats reserved for SC (ST). 26 The ADR data is available for public use at www.myneta.info. 27 The included states are Arunachal Pradesh, Assam, Bihar, Goa, Gujarat, Haryana, Himachal Pradesh, Jharkhand, Kerala, Maharashtra, Manipur, Meghalaya, Nagaland, Odisha, Punjab, Tamil Nadu, Tripura, Uttar Pradesh, Uttarakhand and West Bengal. The states excluded from our analysis are Andhra Pradesh, Chhattisgarh, Jammu and Kashmir, Karnataka, Madhya Pradesh, Mizoram, Rajasthan and Sikkim. Note that these states are excluded from the analysis based on the pre-determined timing of their elections. Consequently, there is no reason to believe that there are any systematic differences between included and excluded states (particularly with respect to the growth of night lights or criminally accused candidates). 28 The boundaries for constituencies were fixed in 1976 until the Delimitation Act of 2002. This Act constituted a Delimitation Commission to redraw the constituency boundaries based on the 2001 census figures. Based on the delay in compiling the necessary data and in creating the new boundaries, the first election with redrawn boundaries was only held in Karnataka in 2008. Consequently, the period between 1976 and 2008 had fixed constituencies boundaries allowing for the comparison of satellite imagery across time. Once the new boundaries were implemented, it is not possible to make a comparison between the two periods. 29 The affidavits are available starting with the first election held after the Supreme Court order in 2003. For example, the order was first effective in 2004 in Arunachal Pradesh and the first election after the boundaries changed 25 number of assembly constituencies and information on year in which the elections were held in each state after the Supreme Court order from the ADR website. Our main variable of interest is the criminal accusations.30 A potential concern with accusations is that political rivals may file false cases to gain electoral advantages. Unfortunately, it is not possible to distinguish between “true” and “false” criminal accusations. Despite this limitation, these data have been widely used to measure criminal accusations (Aidt et al. 2015, Asher and Novosad 2014, Banerjee et al. 2014, Fisman et al. 2014). There is some evidence to suggest that false cases are not as frequent as might be believed. Looking at a sub-sample of states, Vaishnav (2011a) finds that accusations are unrelated to prior electoral performance (a proxy for popularity), incumbency, and the timing of elections. In addition, only cases filed 6 months prior to elections need to be reported, therefore it is less likely the criminal charges are electorally motivated. In our baseline specification, we define a binary variable for whether or not a candidate is currently accused in any criminal case. Specifically, this variable takes a value of 1 if MLA faces any current criminal cases and 0 otherwise. Given the setup in a RD design, we only consider races in which, among the top-2 candidates, one is accused and the other is non-accused. This is because the RD design implicitly assumes that voters are faced with a choice between two types of candidates (accused and non-accused). In practice, the top-2 (or even all the candidates) might be of the same type. We therefore restrict the sample to constituencies in which the top-2 candidates represent each type (accused and non-accused). This restricts our sample from the full 2633 constituencies for which we have data to a smaller sample of 941 constituencies, which are observed annually during our sample period over two successive elections in each state, totaling over 3600 observations.31 While we are not the first study to use the data on criminal accusations, few studies, Vaishnav (2011a) being a notable exception, consider the type of charges. This is particularly important was held in 2009. So for Arunachal Pradesh, our (post-treatment) sample period is from 2004-2009. However, for Uttar Pradesh the first election after the order, took place in 2007 and the first election after the changes in the boundaries was held in 2012. As a result, our sample period for Uttar Pradesh is from 2007-2012. 30 It is not possible to examine convictions since there are only a handful of cases in which criminal proceedings lead to a conviction. Ideally, we would be able to use earlier accusations and examine whether candidates are subsequently convicted. However, the data on accusations are only recent and the Indian judicial process frequently takes years or even decades to resolve cases. According to Sukhtankar and Viashnav (2015), of the 76 MPs serving in the 15th Lok Sabha (2009 national elections) who faced ongoing criminal action, the case had been pending for an average 7 years. 31 Note that our results are robust to using the full sample. We choose this restricted sample since this is the implicit comparison in a RD design. According to Lehne et al. (2016) the median number of candidates per state assembly election is eight, however only the top-2 candidates are competitive: the third placed candidates average 7% of the vote, the fourth placed candidates average 3%, the fifth 1.6% and the rest less than 1%. since not all charges are of the same severity or relevance in assessing a candidate’s aptness for public office. We consider whether a specific charge associated with a criminal case is serious, and whether it is related to any financial wrongdoing. Since any definition of serious criminal charges is inherently arbitrary, we rely on the classification used by the ADR which is based on such factors as the maximum punishment under the law, their violent nature, and offenses under the Prevention of Corruption Act.32 Since ADR does not classify charges into financial and non-financial categories, we define a charge as financial if the corresponding IPC refers to a crime resulting in a loss to the public exchequer.33 Similar to the variable for any criminal case, we create binary variables for whether or not a candidate is accused of a serious or financial criminal charge. Our baseline definition of criminal accusations is whether a candidate has any criminal case against him/her. We refine our definition of criminally accused further by considering candidates who face multiple criminal cases. Insofar as there is a cost to framing false cases against politicians, we might expect that a higher number of cases might be a more reliable indicator of a politician’s true type. Additionally, this also provides a measure of the “intensity of the treatment”. Accordingly, we should expect the treatment effects to become larger as the treatment intensity increases. We consider two thresholds: a candidate is criminally accused if he/she has 2 or more cases or if he/she has 5 or more cases. Table 1-C reports descriptive statistics of the data on criminal accusations. Approximately 53percent of the winners in our sample report at least one criminal case, while 40-percent are facing at least one serious charge, and 19-percent at least one financial charge. Although some of the differences between the winners and the runners-up are statistically significant in the full sample in Table 1-B, these become insignificant when we look at the sample of close elections which is selected by the optimal bandwidth criterion. Table 1-B also reports the predetermined constituency and 32 ADR compiles detailed data on each candidate’s criminal cases and the type of charges framed in each case. It reports the exact criminal charge(s) for each candidate as defined under the Indian Penal Code (IPC). The IPC is the main criminal code of India that covers all substantive aspects of criminal law. ADR defines serious criminal charges using eight criteria. They are: (1) Whether the maximum punishment for the offense committed is of five years or more, or; (2) Whether the offense is non-bailable, or; (3) Offenses pertaining to the electoral violation (IPC 171E or bribery), or; (4) Offense related to the loss to exchequer, or; (5) Offenses the nature of which are related to assault, murder, kidnap, rape, or; (6) Offenses that are mentioned in Representation of the People Act, or; (7) Offenses under Prevention of Corruption Act, or; (8) Offenses related to the Crimes against women. The following is a link to an online Appendix on ADR criteria for coding serious crimes:http://adrindia.org/content/ criteria-categorization-serious-criminal-cases. 33 This classification is based on consultations with several high level Indian Police Service officers and we classify the following IPCs as financial crimes: 171B, 171E, 230–262, 272, 273, 274, 275, 276, 378–420, and 466–489D. candidate characteristics for the previous year, thus only includes 1 data per included constituency. Further, Figure A-7 depicts the distribution of criminally accused MLAs across Indian States. 4.3 Validity of the RD Design There are two standard tests to show the validity of the RD design (Imbens and Lemieux 2008). The first is the McCrary (2008) density test for a discontinuity at the cutoff in the density of the forcing variable. In our context, this tests for whether criminally accused candidates disproportionately win close elections. For instance, criminally accused politicians might be able to manipulate elections and therefore be more likely to win close elections, thereby violating the identifying assumption that the treatment is randomly assigned. If this were the case, we would find a larger frequency of criminally accused candidates compared to non-accused candidates in the neighborhood of the cutoff. This would imply that the density of the margin of victory, the forcing variable, would show a discontinuity at the cutoff. Panels (a)-(b) of Figure 1 shows that the density of the victory margin above and below the cutoff is not statistically significant. The second test of the validity of the RD design is whether the observed characteristics of candidates and constituencies are continuous around the cutoff. That is, while the characteristics for criminally accused and non-accused candidates may be different over the entire sample, with the exception of the treatment, no other variable should be discontinuous around the cutoff. For instance, recent papers (Caughey and Sekhon 2011, Grimmer et al. 2011) have shown that in the context of U.S. elections, the incumbent party tends to have systematically greater chances of winning even when elections are close. However, Eggers et al. (2014) use data on 40,000 closely contested races in different electoral settings, including India, and do not find any systematic evidence of sorting or imbalance around the electoral thresholds. We formally check for continuity of various constituency characteristics in Figure 2 and candidate characteristics in Figure 3. The variable on the y-axis is net of state and year fixed effects. The dots in the scatter plot depict the averages over each successive 0.5% interval of the margin of victory. The curves are local linear regressions fit separately for positive and negative margins of victory using a triangular kernel and the optimal bandwidth calculator suggested by Imbens and Kalayanaraman (2012). The confidence intervals are the 95% confidence intervals plotted using standard errors that are clustered at the constituency level. In Panels (a)–(g) of Figure 2, we compare accused winners to non-accused winners on growth of night lights in the prior years (t-1 and t-2) and on several other constituency characteristics: the electorate size, the number of voters, whether a constituency was aligned with the ruling party in the state, in the previous election, and whether a constituency is reserved for Scheduled Caste (SC) or Scheduled Tribes (ST). Similarly, in Panels (a)–(c) of Figure 3, we perform the continuity checks for candidate characteristics: MLA’s incumbency status and gender, in the previous election. These results indicate that there are no statistically significant differences in the observed covariates around the cutoff. Therefore, the results from the McCrary test and the continuity of covariates strongly suggest that the assumptions underlying the RD design are valid in this setting and that the outcome of a close election is as good as random. 5 Criminally Accused Politicians and Economic Activity 5.1 Main Results We present the main results with a graphical illustration of the RD effect of electing criminally accused candidates in Figure 4 which plots the yearly growth of night-time lights against the margin of victory for the criminally accused candidates.34 The yearly growth of night-time lights is the residual from the regression of yearly growth of night-time lights on state and year dummies. The scatter plot shows the local averages of the residuals in each successive interval of 0.5 percent of margin of victory. The solid curves are plotted non-parametrically using a local linear regression with a triangular kernel and the optimal bandwidth criterion proposed by Imbens and Kalyanaraman (2012). Positive margins of victory indicate a constituency in which a criminally accused candidate won against a non-accused candidate (the runner-up), while a negative margin shows that she/he was the runner-up and that the winner was not criminally accused. The RD figure shows a sharp difference in the average yearly growth of night lights at the cutoff (M ARGIN = 0). The vertical difference between the red and blue lines reflects the estimated causal effect of electing a criminally accused candidate on the yearly growth of night-time lights. In particular, at the cutoff, there is a 34 The sample consists of elections in which, among the top-2 candidates, one is accused and the other is non-accused (see Section 4.2 for detailed discussion). statistically significant and negative effect of electing a criminally accused candidate. Table 2 reports the size of the estimated RD effects. In column (1) we use a local linear regression with the optimal bandwidth (h) criterion proposed by Imbens and Kalyanaraman (IK). This is our main specification and is analogous to Figure 4. We find a statistically significant negative effect of electing criminally accused candidates: the annual growth of night-time lights is approximately 22 percentage points lower in constituencies that barely elect a criminally accused candidate as compared to those constituencies that barely elected a candidate without accusations.35 For robustness, we also present these results using alternate bandwidths in columns (2)-(4). In column (2), we use a bandwidth calculated using the Calonico, Cattaneo and Titiunik (CCT) (2014) algorithm, while in columns (3)-(4) we halve and double the IK bandwidth, respectively. The results in column (2) are quantitatively identical to those in column (1). Halving the bandwidth in column (3) results in a slightly larger estimate, while doubling the bandwidth in column (4) leads to a smaller estimate. Overall, the results remain statistically significant and similar in magnitude. As discussed in sub-section 4.1, we use the intensity of night-time lights as a proxy for economic activity (see Pinkovskiy and Sala-i-Martin 2016 for an overview on the link between night-time lights data and GDP). It is, however, unclear what is driving the changes in night-time lights. For example, the growth in night-time lights could represent changes in the supply of electricity. Alternately, it could be attributable to changes in demand, or some combination of the two. Irrespective of whether these changes are supply or demand-side driven, economic activity and the intensity of night-time lights should be correlated (although certainly not one-to-one). While we currently interpret these results in terms of night-time lights intensity, we compute the corresponding decline in GDP using existing estimates of the elasticity of growth of night lights to the growth of GDP in Section 7. 5.2 Types and the Number of Charges In this subsection we begin by examining whether the costs of electing accused politicians vary based on the underlying charges. If the economic costs associated with electing criminal politicians are attributable to accused politicians, we would expect that both the type of charge and the number of underlying cases would matter. As noted in sub-section 4.2, we define the types of charges, first, 35 In Table A-2 we estimate Table 2 with year fixed effects (in Panel A) and with state-year fixed effects (in Panel B) and find the RD estimates to be identical. by their seriousness, and second, by whether the charges are related to any loss to public exchequer (financial crimes). We report the results by the type of charges in Table 3. In columns (1)-(4) of Panel A, we estimate the RD effect of electing candidates accused of financial charges on the yearly growth of night-time lights. In particular, we compare constituencies with a winner who is accused of at least 1 financial charge (and a loser who is not accused) to constituencies with a non-accused winner (and a loser who has at least 1 financial charge). Similarly, in columns (5)-(8), we examine the effect of electing candidates accused of only non-financial crimes where we compare constituencies with a winner who is accused of at least 1 crime but has no financial charge (and the runner up is not accused) to constituencies with a non-accused winner (and a runner up with at least 1 accusation but no financial accusation). We perform a similar analysis with serious charges in Panel B. We find consistent results: the type of charges matters. The coefficients for both financial and serious crimes are consistently significant and larger in magnitude (in absolute terms) than those estimated for any charge in column (1) of Table 2. In contrast, non-financial and non-serious charges are consistently statistically insignificant. In Table 4, we examine the effect of the number of criminal cases. A larger number of cases can be viewed as a “higher intensity treatment” and, insofar as there are costs to manufacturing and filing false charges, may be more likely to represent “true” accusations. In columns (1)-(4) we present the results of the impact of electing a candidate with two or more charges, while columns (5)-(8) present the results for candidates accused of five or more charges. Similar to the results with financial and serious charges, the number of criminal cases has a clear negative effect. The estimated coefficients are consistently negative and statistically significant and greater than the effect for any charge [column (1) in Table 2].36 Taken together, the results from Tables 3 and 4 demonstrate that the characteristics of the candidate, specifically the candidate’s accusation record, underlie the earlier results and that the costs increase with the severity of the charge. 36 We estimate similar models using the alternate thresholds of 2 and more financial (serious) charges, and 5 or more financial (serious) charges. The results are qualitatively similar to those in Table 4. The table is available upon request. 5.3 Timing of the Negative Effects We next examine whether the effect of electing accused politicians varies across the years of an accused politician’s term. That is, is the negative effect instantaneous (e.g. appear in the first year of the term) and then remains constant throughout the term or rather does it accumulate over time? In Table 5, we estimate the effect of elected an accused candidate separately for each year of the term. We also present the graphical illustration of the RD effect in Figure 5. The results show that the negative effect does not appear instantaneously; there are no effects in the years 1 or 2. In contrast, the estimated coefficients become statistically significant in years 3 and 4, with both the magnitude and significance of the estimated coefficient increasing over time. We interpret these findings to mean that the election of criminally accused politicians does not instantaneously result in lower economic activity and that the economic costs show up after a lag. For politicians to engage in corrupt behavior, they require collaboration with local bureaucrats (Iyer and Mani 2012). Consequently, a certain amount of time is necessary for corrupt politicians and bureaucrats to form a nexus and to engage in corrupt activity. For example, the effect of neglected public infrastructure, such as roads etc., may take some time to slow down economic activity. 5.4 Heterogeneous Impact by State Characteristics We now examine whether the effect of electing criminally accused politicians varies with state characteristics. First, we consider the so-called “BIMAROU” states of Bihar, Jharkhand, Odisha, Uttar Pradesh, and Uttarakhand (The acronym BIMAROU is formed using the first letters of the word ‘sick’: in Hindi). These states are widely considered to be lagging behind in terms of economic development and have been singled out for being corrupt and generally dysfunctional. Second, we consider the set of states that the Ministry of Finance classifies as being “Least Developed” and “Relatively Developed”.37 Finally, using measures of corruption created by Transparency International India’s (TII), drawn from the India Corruption Study of 2005, we classify states into “High-Corruption” and “Low-Corruption” states.38 37 The Ministry of Finance classifies the following states as being “Least Developed”: Arunachal Pradesh, Assam, Bihar, Jharkhand, Odisha and Uttar Pradesh and “Relatively Developed”: Goa, Haryana, Kerala, Maharashtra, Punjab, Tamil Nadu, and Uttarakhand. 38 TII classifies the following states as being “High Corruption”: Tamil Nadu, Haryana, Jharkhand, Assam, and Bihar and “Low-Corruption”: Kerala, Himachal Pradesh, Gujarat, Maharashtra, Punjab, West Bengal, Odisha, and We present the results in Table 6. The RD estimate for BIMAROU states are reported in Panel (A), while Panel (B) reports the estimate for Least Developed states, and finally Panel (C) reports the estimate for High-Corruption states. These effects are negative and statistically significant. The size of the coefficients for BIMAROU and Least Developed states are roughly one and a half times larger than our main result [Table 2, column (1)] and slightly more than double for High-Corruption states [The coefficient is −21.74 for our baseline result (all states), −45.76 for BIMAROU states, −49.46 for Least Developed states, and −51.29 for High-Corruption states]. However, the results are statistically insignificant for Non-BIMAROU, Relatively Developed and Low Corruption states as shown in Table A-3. Since there is substantial overlap between the states in the BIMAROU, Least Developed and High Corruption classifications, it is difficult to isolate one particular factor. Rather, we see that the effects of electing accused politicians are largely confined to certain states. One common feature of these states is the relatively weaker institutions whether judicial, police or political. Anecdotal evidence suggests that in states with stronger institutions, the actions of criminal politicians are more constrained. In contrast, those states with lower quality institutions have reputations of lawless behavior and general impunity for politicians and bureaucrats. Thus, a criminally accused politician may be more able to compromise governance where institutions are weaker or less developed. 6 Criminally Accused Politicians and Public Good Our main results use night-time lights as a measure of economic activity. Using data from the PMGSY program, we examine an alternate measure of economic development and a proxy for public good provision: the length of roads built annually under PMGSY (in kms). We present the graphical illustration of the RD effect (and the balance test) of electing criminally accused candidates on the length of roads built annually in Figure 6 and the results from local linear regression by state characteristics and the types of accusation in Tables 7, 8 and 9. We report the results for all states in Panel A, BIMAROU states in Panel B, Least Developed states in Panel C, and High Corruption states in Panel D. We find that the length of roads constructed is lower in constituencies represented by criminally accused candidates (Panel A in Figure 6 and Table 7) as Uttar Pradesh. compared to constituencies that elected non-accused candidates. Similar to the night-time lights results (Table 6, Panel A and B), we find the negative impact varies by ‘BIMAROU’ and ‘Least Developed’ states, although the estimated coefficient no longer varies by the types of accusation. Overall, it is reassuring to find similar results using a completely different measure of economic development. Surprisingly, we do not find negative impact of electing criminally accused politicians on roads construction for the high corruption states; this may be due to the smaller number of road projects approved and constructed in some of the high corruption states.39 In addition, these states are less dependent than the BIMAROU or Least Developed states on federal transfer of funds. Lastly, we cannot rule out the fact that the nature of corruption might differ in high corruption states; for example, criminal politicians could be corrupt in ways that are harder to detect and not necessarily through the roads construction program. Road construction in India is widely believed to be plagued with rampant corruption and frequently involves manipulation of the tenders and the process of procurement, often leading to lower quality and unfinished road projects (WSJ 2012).40 ,41 By matching contractors’ and political can- didates’ last names in the PMGSY program, Lehne et al. (2016) finds a 63-percent increase in the share of roads allocated to contractors who share their last name when the politician gets elected. In addition, they show that the preferential allocation also increases the likelihood of missing roads using the 2011 census data. They advance two reasons to explain why a road would be listed as missing, both of which are indicative of corruption. Firstly, roads may be listed as completed without ever being built. Secondly, roads could be built with sub-standard materials leading to their complete or partial deterioration by the time of the 2011 census. These results complement our findings with the night-time lights and are consistent with the view that accused politicians are detrimental to economic development and public good delivery. 39 States that have recorded the highest road construction are the BIMAROU and Least Developed states, for e.g. Madhya Pradesh (63,548 km), followed by Rajasthan (58,462 km), Uttar Pradesh (45,905 km), Bihar (35,510 km) and Odisha (35,019 km). 40 http://blogs.wsj.com/indiarealtime/2012/05/04/road-building-still-tarred-with-corruption/. 41 A senior official reported that “road mafias” of contractors, engineers, the local police, civil servants, “and last but not least local politicians”, conspire to keep prices on road contracts above market rates (Singh 2005).