Criminally Accused Politicians in India: Economic Outcomes (Chapter 3)

7 Rough Calculation of the Effect on GDP In this section, we calculate a rough estimate of the impact of criminally accused politicians in terms of GDP loss. So far, due to the absence of local economic indicators, we use changes in the intensity of night-time lights as a proxy for economic activity; however, it is possible to estimate the impact in terms of GDP by using existing estimates of the elasticity of GDP with respect to night-time lights. These elasticities have been calculated in settings where data exist for both the intensity of night-time lights and local economic activity. In particular, we focus on Henderson et al. (2012) estimate for a global sample of low and medium income countries, 0.3, and Bickenbach et al.’s (2013) India-specific district-level estimate of elasticity of 0.107. Since there are no reliable and systematic figures for growth at the constituency level, the latter represents the most disaggregated estimate of the elasticity for India. These two elasticities allows us to calculate a rough upper and lower-bound estimate of the effect on growth rate of GDP. We present the “back of the envelop calculations” of GDP loss in Table 10. For the ease of calculation, we report the RD estimates for the three types of accusations in Panel A (i.e. estimates from column (1) of Table 2, and column (1) from Panel A and B of Table 3). In Panel B, we calculate the impact of electing a criminally accused candidate on the growth rate of GDP in column (1), for candidates accused of at least one financial charge in column (2), and finally for candidates accused of at least one serious charge in column (3). These estimates are calculated by multiplying the coefficients in Panel A with the relevant elasticities from the literature (i.e. the upper bound uses the elasticity of 0.3, while the lower-bound uses the elasticity of 0.107). Depending on the type of criminal accusation, we find estimates ranging from 2.3-6.5 percentage point lower GDP growth per year for our baseline estimates [column (1) in Panel A]. Lastly, in Panel C, we calculate what would the GDP loss look like under these different scenarios based on the estimates from Panel B. Using a conservative estimate of 6-percent GDP growth as a measure of average yearly constituency growth, this would imply that, on average, electing a criminally accused candidate would result in a 5.61 to 5.86-percent GDP growth per year (as compared to the 6-percent otherwise).42,43 These calculations are conservative since they are estimates of the 42 During this period, India experienced high growth rates ranging from 7.9 percent in 2003 to 9.8 percent in 2007. These growth rates are based on the estimated effects on Panel B. For instance, the upper bound estimate in column (1) is -6.5 percent. Consequently, the constituency would only experience 93.5 percent of the normal growth 43 yearly cost, such losses are likely to compound over the full 5 year term. These rough estimates of the cost in terms of GDP growth raise parallel questions in terms of the foregone poverty reduction and the effects on distribution. While the data do not exist to verify this, it is important to highlight that these are not just aggregate constituency-level costs; they are likely commensurate micro-level costs. 8 Robustness In this section, we perform further robustness checks using alternate functional forms, and alter- nate definitions of the dependent variable. We also re-estimate our main result after controlling for covariates in the RD regression (similar to Meyersson 2014), and finally examine the impact of top coding on our results. 8.1 Sensitivity Analysis of RD Specification While earlier researchers emphasized the analysis of different bandwidths (Imbens and Lemieux 2008), recent studies broaden the focus to include alternate control functions (Dell 2010, Lee and Lemieux 2009, Meyersson 2014). We address this in Table A-4 which reports the RD effects for linear, quadratic, cubic, and quartic functions using the IK (h), CCT, h/2, and 2h bandwidth choices.44 Variations in the polynomial order in the control function are ordered by row and bandwidth choices by column in this table. By and large, we find that the RD estimates are negative and statistically significant and qualitatively similar to the effect estimated in Table 2. Statistical significance, however, is lost with larger bandwidth and/or polynomial order of the control function. 8.2 Alternate Dependent Variables Since night-time lights and their distribution can be measured in several ways, we explore two alternate definitions of the dependent variable: the intensity of night lights in levels [Log(N ight Lights)] which is 5.61 GDP growth per year. 44 In results we do not present here, we also re-calculate the RD estimates using alternate bandwidth choices, specifically 0.9h, 1.3h, 1.7h, 2.1h, where h(IK) = 6.35. We further repeat the same exercise for four additional bandwidths (i.e. 2.5, 5.0, 7.5, and 10.0) and find similar results. Both tables are available upon request. and the growth of night-time lights averaged over the election term of the candidate (Average Growth over the Election T erm). We present the estimates from RD effect for the two dependent variables in Table A-5.45 While we focus on the yearly growth of night-time lights (for better comparison across constituencies), both growth rates and levels are used when researchers talk about growth. Moreover, large percentage changes due to a small initial level can mask very small absolute changes. We therefore estimate the local linear RD regression using the IK bandwidth for night-time lights in levels and present the result in column (1). For our second measure, we consider the growth of night lights averaged over the election term of the candidate to address the potential effect of year to year volatility. We present this estimate in column (3). Results from Table A-5, columns (1)-(2) suggest that the point estimates remain statistically significant and negative using the alternate definitions of the dependent variable. 8.3 Controlling for Covariates In an RD framework, it is not necessary to control for various pre-determined covariates as the treatment is independent of these covariates at the cutoff. However, it is possible to directly control for the covariates and estimate the local linear RD regression. We present the RD results in Table A-6, where we do not control for baseline value of log night-lights in the year before election in Panel A and with the baseline value of log night-lights as a control in Panel B. We present the RD regression result with only year fixed effects in column (1), with state and year fixed effects in column (2) [similar to Table 2, column (1)], and finally in addition to state and year fixed effects we add constituency characteristics (i.e. growth of night lights in t-1, growth of night lights in t-2, electorate size, 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)) and candidate characteristics (for e.g. MLA’s incumbency status and gender, in the previous election) in column (2) . Overall, our results remain similar to our main findings in Table 2. This provides further reassurance about the validity of the RD design in our context. 45 The graphical illustration of the RD effect, and the validity of the balance test are available upon request. 8.4 Top Coding As previously noted, the night-time lights data are censored at 63. While this is not an issue in less developed areas, this could be an issue in the wealthiest and most populated areas where we cannot observe any changes above an intensity of 63. Although this is unlikely to be systematically correlated with the accusation status of elected candidates, we directly address this in Table A-7. In our RD sample, there are 63 fully lit constituencies. In Panel A, we drop any observations where the constituency-year pixel average is 63. In Panel B, we drop any constituency in which the average pixel intensity for any year is 63. In both cases, the results are both qualitatively and quantitatively similar to our main results. 9 Conclusions In this paper, we estimate the aggregate economic costs of electing criminally accused politicians at the constituency level using data on the intensity of night-time lights and the sworn affidavits of candidates on their criminal background. We find several important results. We find a large negative and causal impact as the yearly growth of the intensity of night lights is roughly 22-percentage point lower for constituencies that barely elect a criminally accused candidate as compared with those that do not. The estimated effect is not just statistically significant but it is also economically meaningful; this effect translates into roughly a 2.3-percentage point lower GDP growth per year [Column (1) from Panel B of Table 10]. While we only have aggregate constituency level outcomes, this forgone growth must also impact poverty reduction and other micro-level development outcomes. Using data on the specific charges elected representative (e.g. serious or financial criminal charge vs. any criminal charge) we find large variation in the impact of electing accused politicians on the intensity of night-time lights. In contrast, the election of candidates with only non-serious or nonfinancial charges has no impact on the subsequent growth. Additionally, this effect increases with the number of underlying criminal cases, which may serve as an indicator of ‘intensity of treatment’. This variation not only highlights the importance of the precise accusations but also more reliably captures the potential criminality of politicians. This possibility is further underlined by the gradual accumulation of the costs over time. We further examine how the impact of electing criminally accused candidates varies across state characteristics. We find that the costs are more pronounced in states with high corruption levels, lower levels of development and those which plausibly have weaker institutions (‘BIMAROU’ states). This suggests that the local context may mitigate the detrimental effect of electing lower quality politicians, such as accused politicians. While we cannot identify the precise channels, we believe that institutions can play an important role in constraining the actions of elected politicians. Additionally, in the spirit of Kremer’s ‘O-Ring’ theory (1993), local growth might depend on several complementary pieces including effective leadership. In certain contexts, such as less developed areas, there is no redundancy in leadership making the effects of inferior politicians more detrimental. Finally, using data on roads construction – a commonly used measure of public goods provision in developing countries – we find a large negative impact of criminally accused on the length of roads built under PMGSY. This negative effect varies by state characteristics and is largely driven by ‘BIMAROU’ and ‘Least Developed’ states. We interpret our key dependent variable as a measure of economic activity, but it may also proxy for access to electricity. Taken together, roads construction and access to electricity are arguably some of the most important public goods in India, thus the negative impact we estimate is likely to have important welfare consequences. Consequently, the election of accused politicians and, more generally, of lower quality politicians may have adverse effects along a variety of dimensions. More broadly, our results are consistent with the literature on patronage democracies. One manifestation of a patronage democracy is the election of politicians who are able and willing to provide targeted benefits (Burgess et al. 2015). These benefits could be targeted based on caste as in India (Chandra 2004), class, or ethnicity in other contexts. Therefore, instead of focusing on the overall outcomes (such as the delivery of public goods), voters focus on whether politicians can deliver targeted transfers to their specific group or caste. Not only are voters perhaps more likely to overlook accusations but these accusations might serve as a signal of the politician’s willingness to use the office to reward fellow-group members (Chauchard 2014, Wade 1985). If true, our results suggest three consequences. First, this can result in the election of criminally accused candidates, and therefore potentially explain the ever increasing number of accused politicians who are elected in India. Second, the election of lower quality candidates in patronage democracies leads to lower aggregate growth. Third, these effects, however, are mediated by the local context (especially the institutional and legal context). In the more developed, and less corrupt states in our sample, the effects of the accused politicians were lower, perhaps due to the strength of institutions. Although we study a particular context, lower quality politicians are believed to be pervasive in many developing countries. While the underlying cause is often context-specific and may range from caste-politics to tribal and ethnic voting, we believe that our analysis is suggestive for other contexts. Variables TABLE 1-A Descriptive Statistics of Dependent Variables RD Sample (IK bandwidth = 6.35) Overall Sample Accused Non-Accused Difference Accused Non-Accused Difference Growth of Night Lights in t+1 2.04 (1.73) 11.03 (0.04) 151,926.80 (4337.546) 2.43 (2.39) 11.01 (0.04) 157,719.70 (94841.109) No. of observations 1915 1701 Length of Roads built under PMGSY (kms) 20.37 (0.86) 20.48 (0.92) 560 535 Log of Night Lights in t+1 Night Lights average over Election Term No. of observations -0.39 -1.25 (2.90) (2.94) 0.02 10.82 (0.06) (0.07) -5,792.95 125,072.90 (6480.777) (5702.853) -0.10 (1.26) 1.79 (2.92) 11.02 (0.06) 148,250.30 (6663.625) 810 771 20.25 (1.21) 21.21 (1.38) 264 240 -3.03 (4.14) -0.20 (0.09) -23,177.42*** (8742.17) -0.97 (1.83) NOTES: The number of observations for the dependent variables are observed for each year of the election term. This is equivalent to 4 observations for each included constituency. Standard errors are given in parentheses. Asterisks denote significance levels (*=.10, **=.05, ***=.01) TABLE 1-B Descriptive Statistics of Pre-determined Characteristics Variables Growth of Night Lights in Previous Year Log of Night Lights in Previous Year Log Electorate Size in Previous Election Log Number Voted in Previous Election Turnout in Previous Election SC Constituency ST Constituency Ruling Party in Previous Election Incumbent in Previous Election Winner’s Gender in Previous Election Runner-up’s Gender in Previous Election No. of observations Length of Roads built under PMGSY (kms) in Previous Year No. of observations Accused 15.95 (3.37) 10.86 (0.07) 11.99 (0.02) 11.54 (0.02) 64.32 (0.47) 0.11 (0.01) 0.05 (0.01) 0.52 (0.02) 0.36 (0.02) 0.07 (0.01) 0.07 (0.01) Overall Sample Non-Accused Difference 30.81 -14.86** (6.71) (7.23) 10.83 0.04 (0.08) (0.11) 12.06 -0.07** (0.02) (0.03) 11.61 -0.07*** (0.02) (0.03) 64.42 -0.09 (0.54) (0.71) 0.13 -0.03 (0.02) (0.02) 0.05 0.00 (0.01) (0.01) 0.57 -0.05 (0.02) (0.03) 0.44 -0.08** (0.02) (0.03) 0.05 0.02 (0.01) (0.02) 0.05 0.02 (0.01) (0.02) 502 437 8.88 (0.56) 9.68 (0.73) 213 213 -0.80 (0.92) RD Sample (IK bandwidth = 6.35) Accused Non-Accused Difference 20.44 29.79 -9.35 (5.82) (8.64) (10.32) 10.67 10.87 -0.19 (0.11) (0.10) (0.15) 12.06 12.10 -0.04 (0.03) (0.03) (0.04) 11.60 11.64 -0.04 (0.03) (0.02) (0.04) 63.99 64.22 -0.23 (0.72) (0.81) (1.08) 0.14 0.09 0.05 (0.02) (0.02) (0.03) 0.03 0.03 0.00 (0.01) (0.01) (0.02) 0.52 0.51 0.02 (0.03) (0.04) (0.05) 0.30 0.39 -0.09* (0.03) (0.03) (0.05) 0.09 0.06 0.04 (0.02) (0.02) (0.03) 0.07 0.04 0.03 (0.02) (0.01) (0.02) 209 198 8.08 (0.65) 9.63 (0.97) 96 98 -1.55 (1.17) NOTES: The number of observations are from the previous election. This is equivalent to 1 observation per included constituency. Standard errors are given in parentheses. Asterisks denote significance levels (*=.10, **=.05, ***=.01) Variables TABLE 1-C Descriptive Statistics of Criminal Accusations RD Sample (IK bandwidth = 6.35) Overall Sample All Winners Runners-up Difference All Winners Runners-up Difference Any case, proportion 0.50 (0.01) Any serious case, proportion 0.37 (0.01) Any financial case, proportion 0.18 (0.01) Multiple charges (>1), proportion 0.24 (0.01) Avg. number of serious charges 1.33 (0.08) Avg. number of financial charges 0.18 (0.01) No. of observations 1882 0.53 (0.02) 0.40 (0.02) 0.19 (0.01) 0.27 (0.01) 1.47 (0.11) 0.19 (0.01) 0.47 (0.02) 0.35 (0.02) 0.16 (0.01) 0.21 (0.01) 1.19 (0.12) 0.16 (0.01) 941 941 0.07*** (0.02) 0.06** (0.02) 0.03* (0.02) 0.06*** (0.02) 0.28* (0.16) 0.03* (0.02) 0.50 (0.02) 0.38 (0.02) 0.17 (0.01) 0.24 (0.02) 1.52 (0.16) 0.17 (0.01) 0.51 (0.02) 0.38 (0.02) 0.16 (0.02) 0.24 (0.02) 1.56 (0.21) 0.16 (0.02) 0.49 (0.02) 0.38 (0.02) 0.18 (0.02) 0.25 (0.02) 1.49 (0.25) 0.18 (0.02) 814 407 407 NOTES: Standard errors are given in parentheses. Asterisks denote significance levels (*=.10, **=.05, ***=.01) 0.03 (0.04) 0.00 (0.03) -0.02 (0.03) 0.00 (0.03) 0.08 (0.33) -0.02 (0.03) TABLE 2 Effect of Electing Criminally Accused Politicians on Growth of Night Lights Dependent Variable Growth of Night Lights (1) (2) (3) (4) Criminally Accused -21.74** -20.08** -23.95* -12.44** (8.83) (8.07) (13.28) (5.50) Bandwidth Size No. of observations State and Year Fixed Effects Bandwidth Type Polynomial order of control function 6.35 1,581 YES IK (h) 7.32 3.17 1,728 783 YES YES CCT h/2 Local Linear 12.70 2,543 YES 2h NOTES: Standard errors are clustered at the constituency level and given in parentheses. The dependent variable is the residual from the regression of year growth of night lights on state and year dummies. Criminally accused is a dummy variable that is 1 if a criminally accused candidate wins against a non-accused candidate and 0 if criminally accused candidate loses against a non-accused candidate. The RD estimates in columns (1)–(4) are on a local linear regression using a triangular kernel. Asterisks denote significance levels (*=.10, **=.05, ***=.01) TABLE 3 Effect of Electing Criminally Accused Politicians by Accusation Type Growth of Night Lights Dependent Variable Type of Accusation Financial Charge Non-Financial Charge (1) (2) (3) (4) (5) (6) (7) (8) PANEL A Local Linear Bandwidth Size No. of observations -39.39*** (14.50) 7.72 611 Type of Accusation Local Linear Bandwidth Size No. of observations State and Year Fixed Effects Bandwidth Type -36.00*** -38.62** -21.60** (13.78) (18.84) (9.74) 8.41 653 3.86 306 -9.51 (8.18) 15.44 8.97 958 1,332 PANEL B Serious Charge -9.71 -15.76 -3.01 (8.77) (14.13) (4.97) 8.19 1,249 4.48 724 17.94 1,976 Non-Serious Charge -33.19*** (11.75) -29.81*** (10.23) -29.39* (15.93) -18.67** (7.53) 8.46 (6.54) 7.93 (6.33) 5.96 (7.97) -0.18 (4.63) 5.49 1,070 YES IK (h) 6.78 1,257 YES CCT 2.75 506 YES h/2 10.99 1,729 YES 2h 7.38 422 YES IK (h) 7.78 426 YES CCT 3.69 226 YES h/2 14.75 707 YES 2h NOTES: Standard errors are clustered at the constituency level and given in parentheses. The dependent variable is the residual from the regression of state and year dummies. Results displayed in each panel-column come from a separate regression. In Panel A, columns (1)–(4) criminally accused is 1 for a winner who is accused of a financial crime and ran against a non-criminally accused loser; and 0 for a loser who is accused of a financial crime and ran against a non-criminally accused winner. In columns (5)–(8) criminally accused is 1 for a winner who is accused of a non-financial crime and ran against a non-criminally accused loser; and 0 for a loser who is accused of a non-financial crime and ran against a non-criminally accused winner. In Panel B, columns (1)–(4) criminally accused is 1 for a winner who is accused of a serious crime and ran against a non-criminally accused loser; and 0 for a loser who is accused of a serious crime and ran against a non-criminally accused winner. In columns (5)–(8) criminally accused is 1 for a winner who is accused of a non-serious crime and ran against a non-criminally accused loser; and 0 for a loser who is accused of a non-serious crime and ran against a non-criminally accused winner. The RD estimates in columns (1)–(8) are on a local linear regression using a triangular kernel. Asterisks denote significance levels (*=.10, **=.05, ***=.01) TABLE 4 Effect of Electing Criminally Accused Politicians by Multiple Cases Growth of Night Lights Dependent Variable Multiple Cases ( >= 2 ) Multiple Cases ( >= 5 ) Type of Accusation (1) (2) (3) (4) (5) (6) (7) (8) Local Linear Bandwidth Size No. of observations State and Year Fixed Effects Bandwidth Type -30.91*** (10.36) 6.96 815 YES IK (h) -29.52*** -28.50** (10.08) (12.61) 7.39 842 YES CCT 3.48 413 YES h/2 -18.51** (7.23) -42.62** (20.16) -38.92** (19.28) -51.83** (26.05) -28.53** (14.28) 13.93 1,245 YES 2h 7.83 238 YES IK (h) 8.71 253 YES CCT 3.92 122 YES h/2 15.66 351 YES 2h NOTES: Standard errors are clustered at the constituency level and given in parentheses. The dependent variable is the residual from the regression of state and year dummies. Results displayed in each column come from a separate regression. In columns (1)–(4) criminally accused is 1 for a candidate who is accused of 2 or more criminal cases; and 0 otherwise. In columns (5)–(8), criminally accused is 1 for a candidate who is accused accused of 5 or more criminal cases. The RD estimates in columns (1)–(8) are on a local linear regression using a triangular kernel. Asterisks denote significance levels (*=.10, **=.05, ***=.01) Dependent Variable Criminally Accused TABLE 5 Effect of Electing Criminally Accused Politicians by Year in Power Log(Yist+1 ) − Log(Yist ) Log(Yist+2 ) − Log(Yist ) (2) (3) (4) (5) (6) (7) (8) (1) PANEL A -3.51 -3.36 -6.33 1.03 -10.44 -9.94 -8.78 -10.71 (7.91) (8.26) (10.30) (5.85) (10.41) (10.27) (14.30) (7.95) Bandwidth Size No. of observations Dependent Variable Criminally Accused Bandwidth Size No. of observations State and Year Fixed Effects Bandwidth Type Polynomial order of control function 9.71 552 7.85 4.86 19.42 7.3 7.85 3.65 14.59 473 304 809 442 501 234 695 Log(Yist+3 ) − Log(Yist ) Log(Yist+4 ) − Log(Yist ) PANEL B -57.67* -45.87* -46.1 -36.71* -101.05** -80.12** -104.09* -58.47** (30.77) (23.50) (31.07) (21.11) (41.36) (33.50) (60.03) (26.03) 4.69 289 YES IK (h) 7.85 438 YES CCT 2.34 144 YES h/2 9.38 5.61 528 331 YES YES 2h IK (h) Local Linear 7.85 423 YES CCT 2.8 157 YES h/2 11.22 548 YES 2h NOTES: Standard errors are clustered at the constituency level and given in parentheses. The dependent variable is the residual from the regression of state and year dummies. Criminally accused is a dummy variable that is 1 if a criminally accused candidate wins against a non-accused candidate and 0 if criminally accused candidate loses against a non-accused candidate. The RD estimates in columns (1)–(8) are based on a local linear regression using a triangular kernel. Asterisks denote significance levels (*=.10, **=.05, ***=.01) TABLE 6 Effect of Electing Criminally Accused Politicians by State Characteristics Dependent Variable Growth of Night Lights (1) (2) (3) (4) PANEL A: BIMAROU States Criminally Accused -45.76** -44.48** -39.22 -27.31** (18.17) (17.68) (24.82) (11.91) Bandwidth Size No. of observations Criminally Accused Bandwidth Size No. of observations Criminally Accused Bandwidth Size No. of observations State and Year Fixed Effects Bandwidth Type Polynomial order of control function 5.37 5.61 2.68 10.74 563 595 289 930 PANEL B: Least Developed States -49.46** -46.12** -43.74 -28.82** (20.19) (18.66) (27.73) (12.82) 5.06 5.71 2.53 10.13 535 595 277 905 PANEL C: High Corruption States -51.29** -52.37** -51.71 -30.23** (23.18) (23.92) (34.91) (14.76) 6.55 485 YES IK (h) 6.21 3.28 465 241 YES YES CCT h/2 Local Linear 13.11 776 YES 2h NOTES: Standard errors are clustered at the constituency level and given in parentheses. The dependent variable is the residual from the regression of state and year dummies. Results displayed in each column come year from a separate regression. Panel A includes the BIMAROU states: Bihar, Chattisgarh, Jharkhand, Orissa, Uttar Pradesh, and Uttarakhand; Panel B includes the Least Developed states as ranked by Ministry of Finance: Arunachal Pradesh, Assam, Bihar, Jharkhand, Odisha and Uttar Pradesh; and Panel C includes the High Corruption states as ranked by Transparency International India (TII) on index of corruption: Tamil Nadu, Haryana, Jharkhand, Assam, and Bihar. The RD estimates in columns (1)–(4) are on a local linear regression using a triangular kernel. Asterisks denote significance levels (*=.10, **=.05, ***=.01) TABLE 7 Effect of Electing Criminally Accused Politicians on Roads Dependent Variable Length of Roads Built (in kms) (1) (2) (3) (4) PANEL A: All States Criminally Accused -9.11** -9.04** -2.64 -5.63* (4.01) (3.99) (4.44) (3.05) Bandwidth Size No. of observations Criminally Accused Bandwidth Size No. of observations Criminally Accused Bandwidth Size No. of observations Criminally Accused Bandwidth Size No. of observations State and Year Fixed Effects Bandwidth Type Polynomial order of control function 6.75 6.87 3.38 13.5 532 535 255 818 PANEL B: BIMAROU States -16.17** -16.44** -10.95 -11.97** (6.98) (7.22) (7.53) (5.03) 7.92 7.35 3.96 15.83 290 263 151 412 PANEL C: Least Developed States -17.24** -17.25** -9.33 -13.05** (7.32) (7.34) (7.73) (5.33) 7.39 7.34 3.69 14.78 275 275 143 411 PANEL D: High Corruption States -0.72 1.1 0.47 4.09 (6.54) (7.67) (10.10) (4.60) 9.56 203 YES IK (h) 7.61 4.78 166 114 YES YES CCT h/2 Local Linear 19.11 284 YES 2h NOTES: Standard errors are clustered at the constituency level and given in parentheses. The dependent variable is the residual from the regression of state and year dummies. Results displayed in each column come year from a separate regression. Panel A includes all the states; Panel B includes the BIMAROU states: Bihar, Chattisgarh, Jharkhand, Orissa, Uttar Pradesh, and Uttarakhand; Panel C includes the Least Developed states as ranked by Ministry of Finance: Arunachal Pradesh, Assam, Bihar, Jharkhand, Odisha and Uttar Pradesh; and Panel D includes the High Corruption states as ranked by Transparency International India (TII) on index of corruption: Tamil Nadu, Haryana, Jharkhand, Assam, and Bihar. The RD estimates in columns (1)–(4) are on a local linear regression using a triangular kernel. Asterisks denote significance levels (*=.10, **=.05, ***=.01) TABLE 8 Effect of Electing Criminally Accused (Serious Charge) Politicians on Roads Dependent Variable Length of Roads Built (in kms) (1) (2) (3) (4) PANEL A: All States Serious Charge -8.85* -8.99* -3.74 -4.51 (4.97) (5.03) (5.50) (3.69) Bandwidth Size No. of observations Serious Charge Bandwidth Size No. of observations Serious Charge Bandwidth Size No. of observations Serious Charge Bandwidth Size No. of observations State and Year Fixed Effects Bandwidth Type Polynomial order of control function 6.99 6.73 3.5 13.99 412 408 204 618 PANEL B: BIMAROU States -16.44* -16.54* -5.57 -14.66** (9.00) (8.49) (7.33) (6.71) 5.69 6.94 2.84 11.38 191 211 95 289 PANEL C: Least Developed States -16.90** -16.90** -12.03 -11.59** (8.31) (8.54) (9.47) (5.82) 7.45 6.94 3.73 14.91 243 225 134 345 PANEL D: High Corruption States -1.74 -0.71 -3.79 0.23 (7.99) (8.49) (11.12) (5.35) 8.02 154 YES IK (h) 7.17 4.01 139 85 YES YES CCT h/2 Local Linear 16.03 217 YES 2h NOTES: Standard errors are clustered at the constituency level and given in parentheses. The dependent variable is the residual from the regression of state and year dummies. Results displayed in each column come year from a separate regression. Panel A includes all the states; Panel B includes the BIMAROU states: Bihar, Chattisgarh, Jharkhand, Orissa, Uttar Pradesh, and Uttarakhand; Panel C includes the Least Developed states as ranked by Ministry of Finance: Arunachal Pradesh, Assam, Bihar, Jharkhand, Odisha and Uttar Pradesh; and Panel D includes the High Corruption states as ranked by Transparency International India (TII) on index of corruption: Tamil Nadu, Haryana, Jharkhand, Assam, and Bihar. The RD estimates in columns (1)–(4) are on a local linear regression using a triangular kernel. Asterisks denote significance levels (*=.10, **=.05, ***=.01)