Criminally Accused Politicians in India: Economic Outcomes (Chapter 1)

Do Criminally Accused Politicians Affect Economic Outcomes? Evidence from India Abstract We study the causal impact of electing criminally accused politicians to state legislative assemblies in India on the subsequent economic performance of their constituencies. Using data on the criminal background of candidates running for state assembly elections and a constituency-level measure of economic activity proxied by intensity of night-time lights, we employ a regression discontinuity design that controls for unobserved heterogeneity across constituencies and find 22-percentage point lower yearly growth in the intensity of night-time lights arising from the election of a criminally accused politician. These effects are driven by serious, financial and the number of criminal charges and appear to be concentrated in the less developed and more corrupt Indian states. Similar findings emerge for the provision of public goods using data on India’s major rural roads construction program. “They may protest the administrative machinery and thereby break the law, but they are seen as local heroes who are trying to help poor people by different means” - (NY Times, 2014) “Earlier politicians used criminals. Now the criminals themselves have entered politics” - (Associated Press, 2014) 1 Introduction Despite a history of widely contested and transparent elections, and the presence of a vibrant and open media, India elects an ever-increasing number of politicians facing criminal charges. The share of Members of Parliament, the national legislature, who face pending criminal charges has risen from 24 percent in 2004 to 34 percent in 2014 (NY Times 2014).1 While the election of criminally accused candidates to public office is concerning in any context, this is especially true for India. Large quantities of funds are distributed by the government through a wide variety of interventions and programs, which have been plagued by costly scandals with estimated losses in the hundreds of billions of dollars (Sukhtankar and Vaishnav 2015).2 This problem is exacerbated by a severely understaffed judiciary and police force, resulting in an extremely slow judicial system.3 Taken together, these realities create a context in which an influx of criminally accused politicians could be especially costly for an economy. In this paper, we examine the aggregate economic costs of electing criminally accused politicians to State Legislative Assemblies (Vidhan Sabhas) in India for elections held during the 2004 to 2008 period. It is generally accepted, both in the literature and in public debate, that the election of criminally accused candidate is likely to have substantial economic costs for a constituency. Despite this widespread belief of an adverse effect, no formal estimates exist. We take advantage of a Supreme Court of India order in 2003 which required all candidates seeking election to the Parliament of India 1 http://india.blogs.nytimes.com/2014/05/23/in-the-newly-elected-indian-parliament-worrying-trends/ ?_r=0. 2 Looking at a statutory wage increase for participants in India’s employment guarantee scheme (NREGS), Niehaus and Sukhtankar (2013) estimate marginal leakage of almost 100%. Similarly, Baskaran et al. (2015), Nagavarapu and Sekhri (2013), and Min and Golden (2014) find that the allocation of electricity is distorted by political incentives. 3 For instance, Suhkatankar and Vaishnav (2015) note that nearly 60 percent of police positions are unfilled in Uttar Pradesh. Nationwide, 20 to 30 percent of district, subordinate and High court seats are unfilled and close to a quarter of all cases have been pending for 5 years or more and there is a backlog of over 31 million cases. or to State Legislative Assemblies to disclose information on their criminal background in a sworn affidavit submitted to the Election Commission of India (ECI). The data from these affidavits not only allow us to identify criminally charged candidates, but also allow to differentiate candidates by the types and number of charges. We measure constituency-level economic activity using satellite data on the intensity of night-time lights, since no GDP data are available for the state assembly constituencies. Night-time lights data have been used as a robust proxy for economic activity and measures of micro-development (Bleakley and Lin 2012, Henderson et al. 2012, Hodler and Rashky 2014, Michalopoulos and Papaioannou 2013, 2014, Pinkovskiy 2013, Storeygard 2014) at both the national and sub-national levels and offer several advantages.4 First, unlike large household and economic surveys in India, night-time lights data are available at a highly disaggregated geographical level and can be aggregated to compile constituency-level measures. Second, these data are available annually, thus allowing for a more credible time series analysis. Third, night-time lights data are collected automatically by satellites and are therefore less prone to biases that may be present in administrative data. An important challenge in this setting is to account for the unobserved heterogeneity between constituencies that elect criminally accused candidates and those that do not. For instance, criminally accused candidates may be more likely to run (and win) from certain constituencies than others. Therefore, constituencies that elect criminally accused candidates may not be comparable to those that elect non-accused candidates. Using a regression discontinuity design that credibly identifies the effect of electing criminally accused politicians by comparing constituencies which elect criminally accused with those which elect non-accused politicians in close elections, we find that electing a criminally accused politician has a large negative effect on economic activity in their constituency. On average, constituencies in which a criminally accused candidate barely won experienced roughly 22-percentage point lower yearly growth in the intensity of night-time lights than those constituencies which barely elected a non-accused candidate. Moreover, these effects are larger for politicians accused of serious or financial criminal charges. Using existing estimates from the literature on the elasticity of GDP growth to night-time lights growth, this is roughly equivalent to roughly 2.3-6.5 4 Pinkovskiy and Sala-i-Martin (2016) provides a literature review of the nighttime lights measure and propose a data-driven method to assess the relative quality of GDP per capita and survey means by comparing them to the evolution of satellite-recorded nighttime lights. Michalopoulos and Papaioannou (2013) cross-validate the satellite light density and regional development by examining the relationship between luminosity and economic performance using micro-level data from the Demographic and Health Surveys. percentage point lower GDP growth per year. We further find that the negative impact varies by state characteristics. In particular, the costs are more pronounced in states with high corruption levels, lower levels of development and plausibly weaker institutions (so-called ‘BIMAROU’ states). We find very similar effects using a proxy for public good provision: the roads built annually under the Pradhan Mantri Gram Sadak Yojana (PMGSY) program. The length (in kms) of roads built annually under the PMGSY is significantly lower in constituencies in which an accused politician barely won compared to constituencies in which a non-accused barely won (we discuss this in Section 6). Overall, these results highlight the high aggregate economic costs of electing lower quality politicians (i.e. criminally accused) and point to likely significant individual costs in foregone access to public services. Our paper contributes to several related literatures. Most narrowly, our findings contribute to the emerging literature on criminally accused politicians in India. Existing studies typically focus on the selection of these candidates; some examine the response of voters to information on criminal status or criminal charges and the potential mitigating effect of caste politics (Banerjee et al. 2014, Charchard 2014). Others examine the selection of these candidates by political parties (Aidt et al. 2012, Tiwari 2014, Vaishnav 2011a, 2011b, 2011c). However, this growing body of literature on preferences over politician type and how to fight the selection of corrupt politician implicitly assumes that such politicians are less desirable for their constituencies, although the actual effect is unknown. Our study provides the first evidence of the aggregate economic cost of electing criminally accused politicians.5 Since criminally accused politicians are not limited to India, this indicator may also be more broadly applicable (for e.g. Brazil and Pakistan).6,7 Additionally, we contribute to the discussion regarding whether criminal accusations are ‘too noisy’ a signal to be useful for research or policy. In part, this concern arises since political rivals may have the incentive and means to fabricate charges against opponents. Similarly, there are examples of political activists being charged while participating in democratic protests (Jaffrelot and Verniers 2014). We find that accusations can provide useful information regarding politicians as the economic costs vary with the type and the number of criminal charges. 5 Chemin (2012) shows that the election of accused politicians has distributional impacts by reducing the per capita expenditure of SC, ST, and OBC members. However, he does not investigate the overall effects. 6 Almost 200 legislators, or a third of Brazil’s Congress, are facing charges in trials overseen by the Supreme Federal Tribunal (see: http://www.law360.com/articles/457373/brazil-has-had-it-with-corporate-bribery). 7 Criminality and politics in Pakistan (see: http://www.dawn.com/news/1200870). We also make a modest contribution to the literature on the quality of politicians which typically uses proxies such as education (Alcantara 2008, Besley et al. 2005, De Paola and Scoppa 2010, Martinez-Bravo 2014) and, more recently, personality (Callen et al. 2015). As our results demonstrate, whether or not a politician is criminally accused can have an important effect on the constituency level economic outcomes. Consequently, we believe that this may represent a preelection indicator of candidate quality. Finally, although our study focuses on India, it contributes to the broader understanding of the costs of electing lower quality politicians in clientelistic democracies. We find that the costs of electing criminally accused politicians are concentrated in states that are less developed and have higher levels of corruption. These findings are consistent with papers that suggest that low quality politicians (such as criminals) may provide targeted benefits to certain voters (Chandra 2004, Vaishnav 2011a, Banerjee and Pande 2007) based on certain salient factors, such as caste, ethnicity or class. Robinson and Verdier (2013) and van de Walle (2005) further suggest that less developed countries are associated with this type of clientelist or patronage politics. Bardhan and Mookherjee (2012) note that patronage democracies can lead to excessively short-term payoffs and a lack of long run investmetns, included the provision of public goods. In such democracies, criminally accused politicians may be especially detrimental for economic development and public goods delivery. The remainder of this paper is organized as follows. In Section 2 we provide the background on the elected representatives in India and discuss the corruption and criminality in Indian politics. Section 3 discusses the empirical strategy, followed by the data description and the validity of the regression discontinuity design in Section 4. We present the impact of electing criminally accused politicians on night-lights in Section 5. Section 6 estimates the impact of electing criminally accused on a measure public good–roads construction, and Section 7 calculates the effect on GDP. We present variety of robustness checks in Section 8, and conclude the paper in Section 9. 2 Background 2.1 Elected Representatives in India India is a federal republic with a parliamentary system of government. The Parliament of India, the national level legislature, consists of two Houses–an Upper House (also called the Rajya Sabha or the Council of States) and a Lower House (also called the Lok Sabha or the House of the People). Those elected or nominated to either House of the Parliament are referred to as Members of Parliament (or MPs). State assemblies in India follow a similar structure where the Upper House is called the Legislative Council (or Vidhan Parishad) and the Lower House is called the Legislative Assembly (or Vidhan Sabha). Those elected to the state Legislative Assembly are referred to as Members of the Legislative Assembly (or MLAs). Similar to the national level, the election system at the state level is a “first-past-the-post” system and constituencies are divided into single member constituencies. The term of each MLA is 5 years, although it is possible to have elections before the 5-year term mostly due to shifting of political alignments. The focus in this paper is on the members elected to the state legislative assemblies. The Indian Constitution grants elected representatives certain responsibilities. In particular, MLAs hold legislative, financial, and executive power. In addition to these constitutional powers, they also have control over the state bureaucracy, especially in promotions and job assignment/transfers (Asher and Novosad 2015, Iyer and Mani 2012, Krishnan and Somanathan 2013, Nath 2014, Sukhtankar and Vaishnav 2015)8 , which allows them to play a key role in the allocation of funds for various development projects, distribution of licenses, and facilitate access to governmental schemes through the bureaucratic machinery. Collaboration with or control of the bureaucracy also allows politicians to act as mediators between the private sector and the government, and to lobby political allies and business contacts to bring projects to their constituencies (Bussell 2012, Chopra 1996, Jensenius 2013). Finally, MLAs also have access to discretionary development funds through the Member of Legislative Assembly Constituency Development Scheme which they can spend on development projects within their constituencies. Therefore, elected representatives can 8 The nexus between politicians and bureaucrats, and in particular, the possibility of bribes involving job assignments/transfer of bureaucrats was recently alleged in a press conference on May 22, 2015 by the Chief Minister of Delhi (http://www.ndtv.com/video/player/news/ kejriwal-says-centre-has-betrayed-people-of-delhi-by-siding-with-lieutenant-governor/368367). both directly and indirectly affect economic activity in their constituencies. 2.2 Corruption and Criminality in Indian Politics Elected officials are widely reputed to be involved in corruption, mostly the graft and embezzlement of public funds (BBC News India 2012, India Today 2012). A recent paper by Sukhtankar and Vaishnav (2015) compiled an inventory of the biggest public corruption scandals uncovered after 2003, and found amounts totaling hundreds of billions of dollars.9 Fisman et al. (2014) utilize the asset disclosures of candidates for Indian state legislators and compare the asset growth of election winners versus runners-up to calculate the financial returns from holding public office relative to the private sector opportunities available to political candidates. They find that the estimated annual growth rate of the winners’ assets is 3-5 percent higher than those of the runners-up. Similarly, Bhavnani (2012) compares the change in winners’ and losers’ self-declared family assets in India’s two most recent state and national elections, and finds that the average election winner increased their assets by 4-6 percent a year.10 The issue of criminally accused candidates contesting elections in India is not new and has been debated at the highest level by the Election Commission of India, and the Indian Parliament. Even the Supreme Court of India showed serious concern about the increasing “criminalization” of politics, leading to its landmark judgment which made the criminal backgrounds of politicians public. In 2003, the Supreme Court of India required candidates seeking election to the Parliament or to a Legislative Assembly to file sworn affidavits detailing information on their professional and educational qualifications, their assets and liabilities and those of their immediate family, and their criminal convictions and charges. In particular, the affidavits require candidates to report prior convictions and any pending accusations for which the offence is punishable with imprisonment for two years or more, and in which a charge is framed or cognizance is taken by the Court of Law (that is, any criminal indictment). These charges are limited to those framed prior to the six-month period 9 Table 1 of Sukhtankar and Vaishnav (2015) estimates a mean scam “value” of Rs. 36,000 crore (about 5.6 billion USD), and a median of Rs. 12,000 crore (about 1.9 billion USD). 10 According to Banerjee et al. (2011(, in the case of Uttar Pradesh state legislators, the 287 elected MLAs in 2007 who ran for elections again in 2012 witnessed an increase in their average asset value from $220,613 to $658,804 over their 5 year term in office. At that time, the average annual salary of MLAs in Uttar Pradesh was approximately $12,000. The political affiliation was especially important as MLAs who belonged to the political party heading the state government (or the ruling party) saw their asset value increase by an average of $500,000. For opposition party members, this increase amounted to less than $300,000. preceding an election. Since candidates face penalties for lying on the sworn affidavits and since rival candidates (and the media) have incentives to verify information contained in these affidavits, any deliberate misreporting should be minimized.11 The Association for Democratic Reforms (ADR), an election watchdog, along with the National Election Watch have conducted so-called Election Watches for all state and federal elections since 2003 in India.12 Surprisingly, the percentage of MPs facing criminal charges has increased between the 2009 and 2014 elections for the national Parliament. The findings are similar for the state assembly elections. According to the ADR report, over 30 percent of the MLAs currently face criminal cases.13 3 Regression Discontinuity Design A key contribution of this paper is the identification of causal effect of electing criminally accused politicians to state assemblies in India on constituency level economic activity. The main challenge in doing this is that the victory of criminally accused politicians is not necessarily random; for example, criminally accused candidates may be more likely to run and win from certain constituencies than others in ways that are unobservable to us. As a result, average differences in economic activity between constituencies that elected an accused MLA and those that elected a non-accused MLA will result in a biased estimate of the effect of electing criminally accused candidates. A regression discontinuity (RD) design (Lee 2008, Imbens and Lemieux 2008) allows us to exploit a discontinuity in the treatment assignment to identify the causal effect of a treatment variable.14 In our setting, the assignment of treatment, whether a candidate is criminally accused or 11 These are sworn affidavits and there is a penalty for filing incorrect affidavit (e.g. disqualification, imprisonment for a term which may extend to six months, or a fine, or both). The affidavits can be accessed from the ECI’s website (http://eci.nic.in/eci/eci.html) and its website on candidate affidavit (http://eci.nic.in/eci_main1/ LinktoAffidavits.aspx). 12 An Election Watch comprises of background reports based on Criminal, Financial, Educational and Income Tax details of Candidates and Winners (MPs, MLAs and Ministers) who have contested Elections to State Assemblies, the Parliament and a few local bodies. 13 For example, in one of most populous and politically important state, Uttar Pradesh, 575 of the candidates for the 403 assembly seats had criminal backgrounds or faced criminal charges during the 2007 state legislative assembly elections. Out of these, 140 won the assembly seats. Unsurprisingly, following this success, an even greater number of criminally accused candidates (759) ran in the subsequent elections in 2012. Of these, 189 won seats in the state assembly (ADR, 2012a). 14 The seminal paper by Lee (2008) exploits a regression discontinuity design using electoral data. Studies using a similar design in the context of India and elsewhere include Asher and Novosad (2014), Bhalotra et al. (2013), Bhalotra and Clots-Figueras (2014), Broockman (2014), Clots-Figueras (2011, 2012), Fisman et al. (2014), and Uppal (2009). not (ACCU SED), is determined solely on the basis of a cutoff value, c=0, of the forcing variable, the victory margin (M ARGIN ). The treatment assignment follows a known deterministic rule, ACCU SED = 1 (M ARGIN ≥ 0), where 1(.) is the indicator function. The constituencies which fall below the cutoff (M ARGIN < 0), the control group (ACCU SED = 0), elect a non-accused candidate who won against an accused runner-up, and the victory margin in these elections is the difference between the vote shares of the accused runner-up and the non-accused winner. Constituencies that fall above the cutoff (M ARGIN ≥ 0), the treatment group (ACCU SED = 1), elect a criminally accused candidate who won against a non-accused runner-up, and the victory margin in these elections is the difference in vote shares of the accused winner and the non-accused runner-up. Therefore, at the victory margin of zero, the accusation status of a politician changes discontinuously from non-accused to criminally accused. Thus, as the victory margin becomes arbitrarily small (i.e. as we move closer to the cutoff), the outcome of an election is as good as random. As a result, constituencies that barely elected a non-accused politician in a close election serve as a valid counterfactual for constituencies that barely elect a criminally accused politician. We consider the following specification for estimating the RD treatment effect of electing a criminally accused candidate to state legislative assemblies relative to a non-accused candidate: GROW T Hi,s,t+1 = αs + βt+1 + γACCU SEDi,s,t + f (M ARGINi,s,t ) + µi,s,t+1 (1) ∀ M ARGINi,s,t ∈ (c − h, c + h) where GROW T Hi,s,t+1 is the yearly growth rate in night-time lights, our primary outcome of interest. This is measured as the difference in the natural log of night-time lights intensity for the constituency between the current and the previous period, for e.g. [Log(Yi,s,t+1 )-Log(Yi,s,t )], also widely accepted in the literature as a proxy for economic activity. We do not include the night light measure for the year of election as it could be driven by the previous candidate in our specification. αs is the state fixed effects and control for any time-invariant state characteristics. We follow Henderson et al. (2012) and Chen and Nordhaus (2011) by including year fixed effects βt+1 in our estimating equation to control for contemporaneous shocks affecting all units in a year, including any factors that may affect the overall brightness detected by a sensor in any given year. In addition, the year fixed effects also controls for any macroeconomic shocks or national policies that affected all states uniformly. In robustness tests, we also use year–state interaction to control for any state-level economic trends or policies. The variable ACCU SEDi,s,t is the treatment, M ARGINi,s,t is the forcing variable, and h is the neighborhood around the cutoff c=0, also referred to as the bandwidth. The control function f (M ARGINi,s,t ) is some continuous function, usually a n-order polynomial in the forcing variable on each side of c. Finally µi,s,t+1 is the error term. The coefficient of primary interest γ estimates the causal impact of electing criminally accused politicians to state assemblies in India on economic activity as proxied by the growth of night-time lights. The identification of this causal effect relies on fairly weak conditions on the conditional distribution of the error term µ, which is assumed to be a continuous function of the forcing variable (M ARGIN ). We estimate a local linear regression (Hahn, Todd, and Van der Klaauw 2001, Porter 2003, Imbens and Lemieux 2008) as it allows for a suitable bandwidth with a linear control function.15 Our preferred bandwidth specification follows the optimal bandwidth algorithm proposed by Imbens and Kalyanaraman (2012) [referred to as IK (h)]. As a robustness check, we also estimate the local linear regression using the optimal bandwidth proposed by Calonico, Cattaneo and Titiunik (2014), half the optimal bandwidth (h/2), and twice the optimal bandwidth (2h). Since growth in night-time lights is likely to be correlated over time within a constituency, the standard errors are clustered at the constituency level. 4 Data Description and Validity of the RD Design 4.1 Night-time Lights as a Measure of Economic Activity To study the costs associated with electing criminally accused candidates, we need a measure of economic activity at the state assembly constituency level, our unit of analysis. To the best of our knowledge, no such data exist in India.16 Large surveys, such as the National Sample Survey, the India Human Development Survey, and the Economic Census of Firms are only available at the 15 Different variations of equation (1) with different bandwidths and control function have been used in the literature. For example. Lee et al. (2004) use parametric regression-based higher order polynomials in the control function (second-order, third-order, and fourth-order polynomials), thus allowing all the observations to be used in the RD estimation. However, this method puts equal weight on observations far from the cutoff, which can be misleading (Gelman and Imbens 2014). 16 Although we could in principle use the Census of India to look at various measures of public goods, it is only available for the years 2001 and 2011, and therefore will not allow us to look at the annual change in outcome measures at the state assembly constituency level. district level. These cannot be used for three reasons. First, the number of constituencies varies across districts and there is no logical way to weight constituencies within districts.17 Second, even if a constituency level measure of economic activity could be derived, the above mentioned surveys are not available annually. Third, these surveys only capture measures of income and consumption at the district level. We use data on the intensity of night-time lights as a proxy for economic activity. Recent contributions of Henderson et al. (2012), Hodler and Rashky (2014), Michalopoulos and Papaioannou (2013, 2014), and Storeygard (2014) and previous work of Elvidge et al. (1997) finds the use of nighttime lights data as a useful proxy for development for regional analysis in countries with poor quality income data.18,19 Results from Doll et al. (2006) and Pinkovskiy (2013) shows that light density at night is a robust proxy of economic activity. More recently, Dhillon et al. (2016) use the National Election Survey from 2004, which surveys voters at the parliamentary constituency level, to examine the correlation of standard economic indicators with night-lights in India. They find the correlation of night-lights with wealth to be about 0.6, while that with income and education lies between 0.4 and 0.45.20 Overall, these studies find a strong within country relationship between GDP levels and night-time lights intensity and growth rates (see Pinkovskiy and Sala-i-Martin 2016). Night-time lights data offer several advantages, most importantly that the pixels can easily be aggregated to the constituency level and that the availability of annual frequency allows for more detailed temporal analysis. Moreover, whether through increased electrification or higher incomes levels, it should arguably be related to local economic activity. 17 Based on the Delimitation Order of 1976, the constituency boundaries remained fixed till 2008. As a result, there were 4,120 state assembly constituencies. According to the Delimitation Order of 2008, the number of Assembly constituencies are 4,033. Between the 2001 and 2011 Censuses, the number of districts increased from 593 to 640. 18 Henderson et al. (2012) shows that night-time lights can also identify short run fluctuations, including the Asian Financial Crisis in Indonesia between 1997 and 1998 and the Rwandan Genocide between 1993 and 1994. Thus satellite night-time lights data are a useful proxy for economic activity at temporal and geographic scales for which traditional data sets are of poor quality or are unavailable (Henderson et al. 2012). Additionally, prior research shows that the Defense Meteorological Satellite Program’s Operational Linescan System (DMSP-OLS) can reliably detect electrified villages in developing countries and that the night lights output is a useful proxy for electricity provision (Doll et al. 2006, Min et al. 2013, Baskaran et al. 2015). Recent papers have used night-time lights data to study growth of cities in sub-Saharan Africa (Storeygard 2014), production activity in blockaded Palestinian towns of the West Bank (Abrahams 2015, van der Weide et. al 2015), and urban form in China (Baum-Snow and Turner 2012) and India (Harari 2015). 19 Michalopoulos and Papaioannou (2014) show that light density is correlated strongly with proxies of public goods including as access to electrification, the presence of a sewage system, access to piped water, and education and development. Min (2008) finds a strong association between access to electricity, public-goods provision and night light intensity for low income countries. 20 Similar relationship holds at the district level in India (see Chaturvedi et al. 2011 and Bhandari and Roychowdhury 2011). The satellite data are collected by the National Aeronautics and Space Administration’s (NASA) Defense Meteorological Satellite Program’s Operational Linescan System (DMSP-OLS) via a set of military weather satellites that have been orbiting the earth since 1970.21 The result is a series of images covering the globe for each year from 1992 onwards (Elvidge et al. 1997, 2001). Images are scaled onto a geo-referenced 30 arc-second grid (approximately 1 km2 ). The key dependent variable is computed as the sum of all light values from all of the pixels within the boundaries of each state assembly constituencies as defined by the Election Commission of India.22 Thus, we aggregate nighttime lights output to the state level assembly constituency to estimate economic costs of electing criminally accused politicians. In the night-time lights data, each pixel is encoded with a measure of its annual average brightness on a 6-bit scale from 0 to 63. Thus, it is top-coded at 63 and censored at 0 (i.e. the brightest areas are not well measured and areas require some minimum level of light to be captured). Topcoding is a concern since we are unable to observe increases in lighting above 63. However, this should not be systematically related to close elections of criminally accused candidates. That said, we subsequently replicate our analysis with and without the top coded constituencies to verify the robustness. Another potential issue is blooming, which occurs when light from a brightly lit area spills over into neighboring areas. While this is a concern around large cities, our sample is predominately rural. However, for rural areas a related concern is whether night-time lights capture rural activity. As we discussed above, night-time lights also allows dimmer lights to be captured in rural, electrified areas. Lastly, the levels of light output are relative brightness values. Since there is no onboard radiance calibration on the satellite sensors, there is no way to convert the relative brightness values to an actual level of illumination. This complicates time series analysis because changes in observed brightness in different annual composites may be due to real changes in light output on the ground or due to technical factors related to gain levels or sensor properties. We follow Henderson et al. (2012) and Chen and Nordhaus (2011), and account for this limitation by including year fixed 21 The satellites record high resolution images of the entire earth each night typically between 8:30 and 10:00 pm local time. The images, captured at an altitude of 830 km above the earth, record concentrations of outdoor lights, fires, and gas flares at a fine resolution of 0.56 km and a smoothed resolution of 2.7 km. These images are used to produce annual composites during a calendar year, dropping images where lights are shrouded by cloud cover or overpowered by the aurora or solar glare, and removing ephemeral lights like fires, other temporary lighting phenomenon and noise. 22 We use GIS data on administrative boundaries of states and assembly constituencies to enable the aggregation within each state assembly constituencies. effects in our estimating equation to control for contemporaneous shocks affecting all units in a year, including any factors that may affect the overall brightness detected by a sensor in any given year. Finally, we utilize the data available on stable night lights that drop light values from pixels with unstable light signatures over time. Our key dependent variable is the yearly growth in night-time lights. This is the difference in the natural log of night-time lights intensity for the constituency between the periods (t + 1) and (t). As discussed earlier, this has been widely accepted in the literature as a proxy for economic activity. Another advantage of specifying the dependent variable in this form is that it allows us to roughly calculate the impact on GDP using estimates from the literature on the elasticity of GDP growth to night-time lights growth. We also present results using two alternate dependent variables: the natural log of night-time lights, and the growth of night-time lights averaged over the entire election term. These additional dependent variables are likely to be less sensitive when comparing growth of night-time lights over time. We report the descriptive statistics of the key dependent variables in Table 1-A. The dependent variables are observed for each year of the election term, thus we observe about 4 observations for each included state level constituency. Finally, we also use another outcome variable that is a proxy for public good provision at the constituency level: the length of roads built under the PMGSY program (in kms). PMGSY is a fully centrally sponsored rural roads construction program launched in 2000 that aims to provide all weather road connectivity in rural areas, and forms an integral part of the Government of India’s poverty reduction strategy. This program has been described as “unprecedented in its scale and scope” (Aggarwal 2015), connecting over 116,000 habitations with roads and another 23,000 currently under construction as of January 2016. Under PMGSY, about 360,000 kms of rural roads are being constructed with a projected investment of approximately US $14 billion for construction and US $9 billion for “upgradation” of existing tracks. The administrative records of length of roads built under PMGSY are publicly available in the Online Management and Monitoring System (OMMS). We match roads data to the state assembly constituency level data to estimate the impact of electing criminally accused politicians on this measure of public good in Section 6.23 23 The 2001 Census is the source for habitation-level data, which are collected by the PMGSY in order to determine the prioritization of roads.The PMGSY data are available at the census block level. Although there is no one-toone matching between census blocks and state assembly constituencies (for e.g., a block can span more than one constituency), we match a block to a specific constituency if at least 50 percent of the villages in the block fell in that constituency.