Evidence from Close Elections in India∗ [Please ﬁnd the Latest Version Here]

Votes and Policies: Evidence from Close Elections in India∗ [Please ﬁnd the latest version here]

Sourav Sarkar†

December 7, 2018

Abstract

Electoral considerations affect government policies and economic outcomes in various ways. In this paper, I use a close election regression discontinuity design to study the development effects of political alignment between local legislative constituency representatives and state governments in India. I analyze policy and outcome variables from sources of non-proprietary data available annually at a legislative constituency level for the last decade. Constituencies with elected representatives aligned to the ruling party have less growth of visible long term fixed investment goods like new administrative headquarters and educational institutes. However, there is little evidence of aligned constituencies having less receipts and implementation of different government schemes or less growth in night-time luminosity. Together with previous findings of more economic growth due to less regulatory obstacle in aligned constituencies, my results can be rationalized by a theory in which the state government has different types of resources to transfer. The state government substitutes policies attributed more to the local constituency representatives with policies which are attributed primarily to the state government in constituencies whose representatives are not aligned to the ruling political party of the state.

∗I am grateful to my supervisors Abhijit Banerjee and Frank Schilbach for their invaluable guidance and support. I acknowledge the excellent research assistance of Sandeep Kumar for webscraping, geocoding and data preparation and Pranabes Dutta for his initiative in filing a RTI. Further, this paper has immensely benefited from advice, discussions and comments from Daniel Aronoff, David Atkin, Vivek Bhattacharya, Subhendu Bhowal, Surajit Das, Melissa Dell, Sugato Dasgupta, Dave Don- aldson, Leopoldo Fergusson, Chishio Furukawa, Emily Gallagher, Siddharth George, Ryan Hill, David Hughes, Namrata Kala, Ali Kakhbod, Gabriel Kreindler, Matt Lowe, Jacob Moscona, Benjamin Olken, Garima Sharma, Cory Smith, E. Somanathan and Samuel Young. I acknowledge financial help from the Shultz Fund, NUEPA for giving access to the raw data for school education and help from e-mail correspondence with Sam Asher and Paul Novosad. †Department of Economics, MIT. Email: [email protected]

1 1 Introduction

Maximizing social welfare without discrimination is a fundamental objective of most modern democracies built on regular elections. Elections are however mostly contests among political parties whose primary objective is electoral success. The two objectives may diverge significantly. The “baneful effects of the spirit of party” have been pointed out as early as 1796 in the US context (Washington and Hamilton[1796]). While the executive is supposed to serve the people, the members of the executive branch are mostly affiliated with a political party. With theories and evidence of economic voting (Dewan and Shepsle[2011], Healy and Malhotra[2013]) and the role of men and money in elections (Samuels[2002], Kapur and Vaishnav[2013], Vaishnav[2017]) it follows that the governmental policies have every reason to serve as tools for the gains of the political party and its electoral prospects, thereby affecting economic outcomes. Moreover, individual politicians, motivated by private gains, can also affect economic variables (Fisman, Schulz, and Vig[2014], Lehne, Shapiro, and Eynde[2018]) and this effect can be more prominent for politicians affiliated with the ruling political party (Fergusson, Harker, and Molina[2018]).

In this paper, I study the economic effects of a particular type of variation in political characteristic: alignment of the constituency’s representative to the ruling political party. The economic effect of political alignment in a locality is not conclusive, either theoretically or empirically. In their study on several economic outcomes in six African countries, Kramon and Posner[2013] show that while coethnics of the President may be better off in some outcomes, the reverse may hold in some other outcome variables. Although Brollo and Nannicini [2012] find more discretionary transfer of funds for infrastructure to aligned municipalities in Brazil, Bueno[2018] finds a simultaneous lower transfer of grants to non-governmental organizations in aligned municipalities. Callen, Gulzar, and Rezaee[2018] show that in the health sector of Punjab province in Pakistan, aligned regions have more resources but worse quality of health sector delivery. Since sections of the society who benefit from each type of policy are likely to be non-overlapping, there is need to study empirically different government policies and outcomes in the same setting and thereby derive indications for welfare implications of electoral politics in democracies.

I use a regression discontinuity design to identify the economic effect of alignment of the constituency’s representative to the ruling political party on various policies and outcome variables in India. I use this empirical strategy since alignment effects considered in alternative identification strategies like panel data regressions (Solé-Ollé and Sorribas- Navarro[2008]) or spatial discontinuity (Callen, Gulzar, and Rezaee[2018]) may also capture

2 some other attributes associated with vote share diﬀerence of the party, for instance party constituencies being strongholds for powerful leaders and factions within the party.1 Studying alignment eﬀect among close elections controls for these associations (Lee[2016]).

I analyze data of close elections for legislative assemblies in India from May 2008 to March 2018 (elections from May, 2008 are being conducted as per the latest delimitation of constituencies). Asher and Novosad[2017], using the same identification strategy for a previous decade, find more economic growth as indicated by higher satellite night lights, employment and stock prices in politically aligned constituencies in India.2 I use sources of non-proprietary data available annually at the legislative constituency level for the last decade (my data is spatially no coarser than a town/village and temporally no coarser than a year). I analyze the following variables: (a) long-term fixed investment goods such as new district headquarters, subdistrict headquarters, and new schools and higher educational institutions managed by the state government; (b) various rural employment, rural housing and road construction schemes that are implemented by the state government and total receipts and expenditure of schools; (c) satellite data on night lights. While the electorate is likely to attribute the decision on variables in category (a) almost entirely to the state government, survey data indicates that the electorate attribute variables in category (b) partly to the local politicians. Satellite night lights data is a combined outcome of different activities, both by public as well as private sector. According to Asher and Novosad[2017], the electorate attributes a part of such activities, like easing of bureaucratic regulations to local constituency representatives.

Following close elections, constituencies with elected representatives aligned to the ruling party have less growth of visible long term ﬁxed investment goods like new administrative headquarters and educational institutes. I estimate the index of growth of administrative headquarters to be 0.14 less in aligned constituencies as compared to the standard deviation of 0.72 among non-aligned constituencies and the index of growth of state government- managed educational institutes to be 0.12 less in aligned constituencies as compared to the standard deviation of 0.87 among non-aligned constituencies. Combining the above two indices, I estimate the index of visible long term ﬁxed investment goods in aligned con-

1Callen, Gulzar, and Rezaee[2018] does the analysis with both close election regression discontinuity as well as spatial discontinuity. However, their sample size is much smaller for the former analysis. 2Asher and Novosad[2017] use data for elections held from 1990-2012. Most of their data precedes my period of analysis. Except for satellite night lights (which they use from 1992-2008), their source of data is diﬀerent from what I use. They use the economic census (of 1990, 1998, and 2005) to measure employment growth and population census (of 1991 and 2001) to measure growth in public goods. They compile stock price data from several proprietary sources (2000-2013). Their data on mining permits are from the Bulletin of Mineral Information (2008-2013).

3 stituencies to be 0.2 less in aligned constituencies as compared to the standard deviation of 0.75 among non-aligned constituencies. My estimates are not statistically significant for subindices and for most variables pertaining to rural employment guarantee scheme, rural housing schemes, rural road construction scheme and total receipts and expenditure to schools. Exceptions are the total utilization of funds under central government sponsored housing schemes which I estimate to be 20.7% more in aligned constituencies and estimates for the total expenditure and completed length of roads under rural road construction scheme which are positive and statistically significant for specifications which include additional (parliamentary constituency-electoral period) fixed effects. Hence, there is little evidence of aligned constituencies having less of any variable pertaining to receipts and implementation of different government schemes. My estimates for the change in log luminosity using satellite night lights data are positive (although not statistically significant) and not statistically distinguishable from the estimates for the period 1992-2008 reported in Asher and Novosad [2017]. Also, during the earlier period of my analysis, Asher and Novosad[2017] find more mining permits being granted to aligned constituencies.

The above ﬁndings can be explained by a theory of substitution of various electoral strategies by the ruling party as pointed out by Bueno[2018] in the case of Brazil. This theory, inspired by the theory of heterogenous attribution (Arulampalam, Dasgupta, Dhillon, and Dutta[2009]), extends the predictions in Brollo and Nannicini[2012] and Asher and Novosad[2017] to point to the possibility of electoral strategies which are attributable to the party legislator being used more in aligned constituencies while electoral strategies which are mostly attributable to the state government are used more in the non-aligned ones.

My paper contributes to the literature on multiple strategies being used for electoral gains (List and Sturm[2006], Stokes, Dunning, Nazareno, and Brusco[2013], Kramon and Posner[2013]). In particular, my findings of less targeting of certain policies to aligned constituencies are an addition to the recent empirical literature on non-aligned constituencies/voters being favored for some goods and policies (Kramon and Posner[2013] for Africa, Bueno[2018] for Brazil, Callen, Gulzar, and Rezaee[2018] for Pakistan). This is in contrast to the findings in Brollo and Nannicini[2012], Migueis[2013], Bracco, Lockwood, Porcelli, and Redoano[2015], Asher and Novosad[2017]. My finding of a statistically significant impact for visible long-term fixed investments is also consistent with the theory of salience of visible goods in democracies (Mani and Mukand[2007]). Marx[2018] has found projects pertaining to visible goods being prone to electoral cycles in Africa.

My paper also contributes to the literature on redistributive politics and their implica-

4 tions in the Indian context (Khemani[2004], Arulampalam, Dasgupta, Dhillon, and Dutta [2009], Cole[2009], Asher and Novosad[2017]). Political scientists often view redistributive politics as dependent upon the level of development of a polity (Stokes, Dunning, Nazareno, and Brusco[2013], Weitz-Shapiro and Winters[2017]). Wilkinson[2007] predicted the possibility of different findings in redistributive politics in India for the period of my analysis due to increased transparency via media, etc. My results are indicative of political factors affecting some policies in India for the recent decade, although my findings of aligned constituencies getting less of some policies is novel in the Indian context.

The rest of the paper is organized as follows. Section2 discusses the institutional setting and lists the data sources, Section3 discusses the various policy variables used, the method of construction of index variables, my main empirical speciﬁcation and the various robustness checks that I use, Section4 discusses a model to guide my empirical results, Section5 checks for the validity of the empirical strategy, Section6 discusses the results, Section7 discusses possible alternative mechanisms, Section8 concludes.

2 Institutional Setting and Data

2.1 Institutional Setting

India is a Unitary Federation that is divided into 29 states (which comprise almost the entire area and population) and seven union territories. All 29 states and 2 Union Territories (the National Capital Territory of Delhi and Puducherry) have a parliamentary form of provincial government (henceforth I denote each of the 29 states and 2 Union Territories as states). Although there is a parliamentary form of government, known as the central government for the entire country, the Constitution of India vests signiﬁcant executive power to state governments, whose combined expenditure outstrips those of the central government.

Each state is divided into single-member Legislative Assembly constituencies for which elections are held with the winner decided by a first past the poll system. Elections are held regularly at five-year intervals (elections can be held before the end of a five-year term if no party or coalition is able to amass a majority). Elections are conducted by the Election Commission of India, a constitutionally established independent body. The party or coalition which secures the majority of Legislative Constituency seats forms the executive branch comprising of the Chief Minister of the state, the Cabinet and the Council

5 of Ministers.3 The electorate votes partly on the basis of the ruling political party and its chief ministerial candidate and partly on the basis of the legislative constituency candidate (as state election survey results in Lokniti-CSDS[2014b] show, although the importance of the former is generally higher than that of the latter, the importance attached to the legislative constituency candidate can be over 30%, as in the 2018 Karnataka Legislative Assembly election).

The number of Assembly Constituencies into which a state is divided varies markedly across states: the most populated state, Uttar Pradesh has 403 assembly constituencies while Pondicherry has only 30. Among the more populated states, the number of legislative constituencies is roughly proportional to the total population of the state according to the 1971 census. For states with lower populations, the population-constituency ratio is higher.

2.2 Choice of Data

Identification of potential alignment effects via a close election regression discontinuity re- quires a dataset at the electoral constituency level for each electoral period. Such data is not available for most of the variables since (a) the population census which reports data at the village level is decennial, while elections take place at five-year intervals; (b) annual data for most variables are available at the district level. Districts are on average eight times larger than the assembly constituencies.

I use policy variables and/or indicators of economic outcomes from all data sources which are non-proprietary and electronically available at a level temporally no coarser than a year and spatially no coarser than a village or town. My data includes creation of new administrative headquarters, variables pertaining to schools and higher educational institutes, variables from three government schemes which are implemented by the state government and satellite data on nighttime lights activity (I detail the reason for choosing the three

3Ten states (West Bengal, Uttar Pradesh, Kerala, Tamil Nadu, Uttarakhand, Karnataka, Andhra Pradesh, Telangana, Jharkhand and Maharashtra) have one additional nominated member of the Anglo- Indian community in the assembly. However, these additional legislators have hardly played a pivotal role in the formation of government in the case of a hung assembly. For details about the practical implications of an Anglo-Indian member, see Griﬃn[2018], Dailybite[2018], Ahmad[2018], and Basu[2018]. Also, seven states (Uttar Pradesh, Karnataka, Andhra Pradesh, Telangana, Bihar, Jammu & Kashmir and Maharashtra) have a bicameral legislature comprised of a Legislative Council in addition to the Legislative Assembly. The members of the Legislative Council are either indirectly elected or are nominated by the Governor of the state. The Council’s members have legislative powers and are also eligible to be members of the Council of Ministers which forms the executive branch. However, the members of the Legislative Council have hardly played a pivotal role in the determination of the party/coalition which forms the council of ministers.

6 government schemes later in this section).

I only use elections for periods after the 2008 delimitation since the non-proprietary shape ﬁle available for the previous period is less reliable.4 Moreover, the Delimitation Commission was formed in 2002. The economic implications of being politically aligned are diﬀerent if the constituencies for the subsequent elections vary. Further, many of the data sources that I use are available only for the last decade.

A caveat of the analysis is that I do not include the majority of government expenditures whose annual data is available at the district level at best. I am unable to use the wealth of information on major economic indicators available in the data of the National Sample Survey, different district level data on rural economic indicators compiled by ICRISAT- VDSA, data on health expenditures under national health mission, data on crime compiled by NCRB, etc. I also do not use data on major health investments since I could not find a list on date of establishment of hospitals and health centers in India analogous to that of education. Further, due to non-availability for most states, I do not use the data on the implementation of projects under the Local Area Development Funds of the Member of the Legislative Assemblies.5 Also, other important economic measures are also available from satellites, for instance measure of pollution. I do not report them since first, while an increased pollution-level may be harmful, decreased enforcement of anti-pollution norms can also give important benefits to some sections of the society (Holland[2016]) and second, the level of pollution may be positively associated with the level of economic activity.

2.3 Data Description

I describe the data below:

I Administrative Headquarters:

i New district headquarters: Each state in India is divided into districts for administrative convenience (the district administration is subordinate to the state administration). The administration of the district is centered at the district headquarters.

4The old assembly constituency shape files are available freely in Sukhtankar and Patnam[2018]. Although the shape files have been used previously (Vaishnav and Sircar[2012], Dhillon, Krishnan, Patnam, and Perroni[2016]), the website claims: “I make no claims whatsoever about the accuracy ...... Use at your own risk, and definitely do not expect the map to be anywhere close to accurate at the village level” 5Pranabes Dutta has filed an RTI for getting the data although even if available for most legislative assemblies the process is time-consuming. If available, this data may lead to important insights since these are projects under direct responsibility of the elected representatives.

7 After 1993, they are also seats of the District Councils (Zilla Parishad). The decision about formation of new districts is made by the government of the state. New districts have continuously been formed out of older districts (majority of these are splits of existing districts) and hence newer places have acquired the status of a district headquarter. 79 new districts have been formed in India since May 2008.6 Asher, Nagpal, and Novosad[2017] and Sarkar[2018] find economic benefits of being located close to a district headquarter in India. Hence, I perceive the allocation of a district headquarter to be a type of beneficial government policy. I obtained the data on the formation of new district headquarters in India from the respective district websites and the District Census Handbook of the Census of India.7 ii New subdistrict headquarters: Districts in India are further divided into subdistricts. Each subdistrict also has its headquarter. Analogous to districts and district headquarters, new subdistricts and subdistrict headquarters have continuously been formed in India. Although Bardhan and Mookherjee[2000] suggest the theoretical possibility of greater capture at lower levels of government, empirical evidence on it is scarce (for example Alatas, Banerjee, Hanna, Olken, Purnamasari, and Wai-Poi [2013] does not find evidence of capture in Indonesia). Hence, I assume that like district headquarters, the allocation of a subdistrict headquarter is a type of beneficial government policy. I obtained the data on the formation of new subdistrict headquarters in India from the Local Government Directory (LGD[2018]). The data mainly contain information on the formation of new subdistricts after the 2011 census of India. A total of about 550 new subdistricts were listed in the directory.8

II Education: Many studies have shown educational investments to have signiﬁcant economic returns (Psacharopoulos and Patrinos[2004]). Stasavage[2005] has argued for the electoral beneﬁts to be more for primary education than for higher education in African democracies. I use the data for both primary and secondary schools and for

6Electoral units of parliamentary elections and state legislative assembly elections are set by the De- limitation Commission of India and hence are unaffected by district formation by the state governments. The process of delimitation has been fairly neutral (see Jensenius[2013] for the earlier and Iyer and Reddy [2013] for the latest delimitation). Hence I do not analyze gerrymandering when analyzing the effect of state assembly elections in district formation. 7I take the date of formation to be the date of order wherever available. Otherwise, I take it to be the date from which the new district becomes effective. I obtained the location of district headquarters using google maps. 8I could not find the formation date of 1-2% subdistricts or locate the headquarters of 5% new subdistricts on google maps. I take the date of formation to be the date of order whenever available. Otherwise, I take it to be the date from which the new subdistrict becomes effective. I obtained the location of subdistrict headquarters using google maps. I could locate over 90% of the subdistrict headquarters that are listed in the census of 2011.

8 higher educational institutes.

i Schools: I obtained detailed data pertaining to schools in India on request from the website on School Report Cards (DISE[2018]). At the time of downloading, data was available up through academic year 2016-17. I construct the variables for the year of establishment of schools, total school receipts and expenditure from the general data ﬁle.9 The role of the state government in the functioning of schools is variable. The DISE data classiﬁes schools by the type of management: state government departments, local body, private aided, private unaided, central government, etc.10 ii Higher Educational Institutes: I downloaded data pertaining to higher education institutes from the website of the All India Survey of Higher Education (AISHE [2018]). I use the variable on the year of establishment of colleges, universities and standalone institutes. The majority of higher educational institutes are colleges and most of them are privately managed (in the main analysis, I use the data for only state government managed higher educational institutes).

III Government Schemes Implemented by the State Government: I use data from three types of government schemes, almost all of which are central government sponsored schemes. A large portion of the funds for these schemes are borne by the central government although they are implemented by the state governments. The reason for choosing these variables is that unlike completely state-government-run schemes, the data for these are available for the entire country at a uniﬁed central portal. These are the Mahatma Gandhi National Rural Employment Guarantee Scheme (MGNREGS), the various rural housing schemes and the Pradhan Mantri Gram Sadak Yojana (PMGSY) which together constitute about 30% of the expenditure on centrally-sponsored schemes in 2018-19. The National Health Mission and National Education Mission constitute another 20% of the expenditure. Although I could not ﬁnd their annual expenditure at the legislative-assembly level on the website, many of the receipts and expenditures of the latter (from the universal literacy and secondary education programs SSA and RMSA respectively) are included in the total receipts and expenditure to schools found in the DISE data for schools which I have described. Each of the other centrally-sponsored schemes is no more than 7% of the total expenditure under centrally-sponsored schemes.

9The other files contain much detailed information pertaining to infrastructural facilities, teaching staff, enrollment and student’s performance. I do not use these variables in this paper. 10The GIS data for many of the schools are available in the website for school location mapping (School- GIS[2018]). For those which were not available, the postal codes of the schools were used to geolocate them. Postal code shape files for most of India are available in the postal department’s website (Postoffices[2018]) and the GIS files were webscraped. About 85% of the schools were matched to the assembly constituencies.

9 The state governments play a significant role in the implementation of some central government schemes also, notably the Public Distribution System (PDS) whose expenditure dwarfs those of schemes that I consider in this paper. I could not find data for either the PDS or the state government schemes (except those that are available for the number of houses under rural housing schemes, which I use) in a unified portal.

i MGNREGS: Started in 2006, MGNREGS was extended all over India in 2008. “Under the Act, any adult residing in rural areas who demands work has to be employed on local public works within 15 days. Failing that, an unemployment al- lowance is due” (Drèze and Khera[2017]). Studies have shown the efficacy of MGN- REGS in reducing poverty (Ravi and Engler[2015]; Bhattacharjee[2017], however, gives a contrary argument), increasing household consumption (Bose[2017]) and increasing rural wages (Carswell and De Neve[2014], Imbert and Papp[2015], Mu- ralidharan, Niehaus, and Sukhtankar[2017], Berg, Bhattacharyya, Rajasekhar, and Manjula[2018]). Electoral gains from MGNREGS in the case of Andhra Pradesh is documented in Maiorano[2014]. Creation of individual as well as community assets form an important part of MGNREGS.11 I webscraped the asset data from Bhuvan Geo-MGNREGA (Bhuvan- MGNREGA[2018]). 12 The webscraped data contains latitudes and longitudes along with the work financial year and the financial year of the work completion of the asset.13 There is a wide variation in the type of rural assets created under MGN- REGS, and this includes many of the works of PMGSY and some of the housing schemes I discuss subsequently. ii Housing Schemes: Many housing schemes to provide permanent houses to people are in operation in India. A large part of the governmental expenditure pertaining to housing schemes in rural regions happens under the Indira Awaas Yojana (IAY) which has been “restructured into Pradhan Mantri Awaas Yojana (PMAY-G) w.e.f. 1st April, 2016” (PMAY-G[2016]). 14 Financing for IAY and PMAY-G is shared between the central and the state governments although the implementation is done by the latter. While there are objective norms for the intra-state allotment of funds

11Since 2009, it is “mandatory for the states to ensure that at least 60% of MGNREGA related works undertaken in a district in terms of costs, is spent for creation of productive assets that are directly linked to agriculture and allied activities” (Mishra[2011]). 12Detailed data on many other variables on MGNREGS are available in (MGNREGA[2018]). I webscraped the former website since ﬁrst, the latter only contains data at the panchayat-level from 2011 and second, the former is available along with the latitudes and longitudes. 13At the time of webscraping, only about 10-20% of total assets were shown to be non-geotaggable or yet to be geotagged. 14Details of these housing schemes are mentioned in IAY[2013] and PMAY-G[2016].

10 like the data on the census, BPL households, and socio-economic and caste census, there is room for manipulation.15 There are also some state-sponsored housing schemes (example Biju Pucca Ghar Yojana in Orissa, the Lohia Gramin Awas Yojana in Uttar Pradesh, the Chief Minister BPL Awas Yojana in Rajasthan) and a few schemes for natural calamities like cyclones and earthquakes. I webscraped the data on both the physical progress as well as financial progress of various rural housing schemes (PMAY-G[2018]). The data on the physical progress reports information on the number of houses completed in each panchayat for each financial year from 2010 under each type of housing scheme (the two main sub- categories are all central schemes and all state schemes. The former is much greater in scale). The data on the financial progress reports data on the total utilization of funds under central schemes only for each financial year from 2010. The total allocation of funds is available only at the state level in the website. The level of utilization of funds is low in relation to total availability, hence indicating significant scope for variation in implementation due to state level politics.16 iii Pradhan Mantri Gram Sadak Yojana (PMGSY): Started in 2001, the PMGSY intends to link habitations in India by permanent roads. Several studies (e.g. Ag- garwal[2018], Adukia, Asher, and Novosad[2017]) find positive economic effects of this program (Asher and Novosad[2018], however, do not find an effect of PMGSY on several important outcomes). I webscraped the data for the physical progress of work under PMGSY (PMGSY [2018]) and geocoded the habitations using google. The webscraped data contains data on the year of sanction of each project and their details.

IV Satellite Nighttime Lights Data: I obtained these from two websites of the National Oceanic and Atmospheric Association (NOAA) (DMSP-OLS[2018], VIIRS[2018]). The dataset may not be comparable across time but should be usable with time ﬁxed eﬀects (Lowe[2014]). As per Elvidge, Baugh, Zhizhin, Hsu, and Ghosh[2017], the later series from VIIRS has not been corrected for cloud-free observations. Hence, for the VIIRS data, I take those pixels to be missing values if they are of zero value as well as if they contain zero cloud-free observations. Also the VIIRS dataset contains many outliers.

15Compared to MGNREGS, these schemes have received much less attention in academic studies and hence evidence documenting political manipulation is scarce (Verma, Gupta, and Birner[2017] who studies IAY in Bihar is an exception). However, as mentioned in IAY[2013] (section 4.2) and PMAY-G[2016], village local bodies have a signiﬁcant role in various stages of IAY and PMAY-G. 16I geo-coded the panchayats with the help of Google Maps and matched them with the help of names of the panchayats and districts.

11 x Hence, I take the following transformation of the pixels: y = |x| ∗ log(1 + |x|) if x 6= 0, y = 0 if x = 0.

V Data of State Legislative Assembly Elections: Detailed data on various state assembly elections in India are available on the website of the Election Commission of India (ECI[2018]). I obtained the dataset used in this paper in digital format from the Lok Dhaba interface of Ashoka University (TCPD[2018]). I use the dataset for regular state assembly elections conducted after the latest (2008) Delimitation.17 I list the details for the state elections in my period of analysis in Table A.1. The assembly constituency shapeﬁle used is from datameet (Datameet[2017]).

2.4 Policy Variables Used and Construction of Summary Indices

In this section I detail the policy variables I use. I follow Anderson[2008] to construct summary indices for groups.18 In what I mention henceforth, by weighted average I imply the weighted average of the demeaned variables as described in Anderson[2008].

I Policies involving growth of long-term investment goods which are taken by the state government:19 I believe that the electorate attributes these policies mostly to the state government (while the successful operation and beneﬁts from establishments like the administrative headquarters and educational institutes may involve eﬀorts of local politicians, it is unlikely that the electorate attribute the decision on establishment of such headquarters/institutions to the local constituency representative)

i Growth of Administrative Headquarters: I consider the growth of district headquarters and growth of subdistrict headquarters.20 I construct the index for administrative headquarters by taking a weighted average of growth in the number of district headquarters and growth in the number of subdistrict headquarters. ii Growth of Educational Institutions: I consider the growth of state-government- managed schools and the growth in the number of state-government-managed higher

17At the time of downloading the 2018 state assembly election of Karnataka had not yet been held. This does not impact the main result since no single political party could secure a majority in the 2018 Karnataka election. 18Anderson[2008]’s discussion of indices is based on O’Brien[1984] ∆xi 19By growth I mean , where ∆xi is the number of new good i while xi is the total number of good (1 + xi) i at the beginning of the electoral period. 20I take the existing subdistrict headquarters to be as listed in the 2011 census.

12 educational institutes.21 I construct the index for growth of educational institutes by taking a weighted average of growth in the number of state-government-managed schools and growth in the number of state-government-managed higher educational institutes.

I construct the index for goods involving long term ﬁxed investment and/or visible goods by taking the weighted average of the two indices for administrative headquarters and educational institutes.

II Government schemes implemented by the state government: The rural employment guarantee scheme, rural housing schemes and rural road construction scheme are implemented by the district administration whose officials function under the state government, although local political leaders and constituency representatives are also involved (see Gulzar and Pasquale[2017], Lehne, Shapiro, and Eynde[2018]). According to the survey results in Lokniti-CSDS[2014a], conditional on being a beneficiary, 14.7% of the respondents credited a local politician/party worker for a central government sponsored housing scheme while the figure is 16.1% for national employment guarantee scheme (this compares to 50.1% and 42.0% credit attribution to the state government for housing and employment schemes respectively). Schools (whose receipts and expenditure I consider) are managed by diverse bodies: while the largest number of schools are under the state government departments, a substantial percentage of schools are under private management, local government bodies or the central government. As a whole, I believe that the electorate attributes the receipts and implementation of the government schemes that I consider, at least partly, to the constituency representative.

i Rural Employment Program (MGNREGS): I consider log(1 + x) where x are total number of asset works under MGNREGS and total number of asset works completed under MGNREGS. I construct the index for rural employment program (MGN- REGS) by taking the weighted average of the above two variables. ii Housing Schemes: I consider log(1 + x) where x are the total number of houses built under diﬀerent rural housing schemes and total funds utilized under central

21I include schools which are managed by the department of education or the tribal/social welfare department, colleges and standalone institutes which are managed by the state government, and universities which are either state public universities or state open universities. Many educational institutes even when not managed by the state government are under its inﬂuence. For instance many schools are managed by local governments which work under the state government. Also many privately-managed institutes receive grants from the state government. In the Appendix, I report results for all types of schools and higher-education institutes.

13 government-sponsored housing schemes. I construct the index for housing schemes by taking the weighted average of the above two variables. iii Rural Road Construction Program (PMGSY): I consider log(1 + x) where x are the (population-weighted) total cost of the sanctioned schemes, (population-weighted) total expenditure till date of the sanctioned schemes, (population-weighted) total length of the road, (population-weighted) total length completed.22 I construct the index for rural road construction program by taking the weighted average of the above four indices. iv School Receipts and Expenditure: I take 100 ∗ log(1 + x), where x is the total school receipts and total school expenditures. I use the aggregate of all categories of receipts (expenditure) in the DISE data: the amount of school development grant receipt (expenditure), the amount of teacher learning material receipt (expenditure), school maintenance grants (expenditure) and funds (expenditure) from other sources. I construct the index for total school receipts and expenditures by taking a weighted average of total school receipts and total school expenditures.

I construct the index for government schemes by taking the weighted average of the above four indices for goverment schemes.

III Growth of Satellite Night Lights: I consider the difference in the logarithm of the average measure of satellite night lights at the end of the electoral period from that in the previous electoral period.23 Growth in satellite night lights can be both an outcome of direct increase in nighttime lights due to electrification efforts by the government (Min[2015]) as well as an indirect outcome of an increased economic activity. Asher and Novosad[2017] finds a significant increase of satellite night lights among aligned constituencies and attributes it to be reflective of higher economic growth due to less bureaucratic regulations in aligned constituencies. 22I use the population of a habitation as weights because many of the projects connect multiple habitations and hence each project can potentially span multiple assembly constituencies. Population of each habitation that a project covers is given in the data. 23I consider this measure to make it comparable with Asher and Novosad[2017]

14 3 Empirical Strategy

I run a close election regression discontinuity with the running variable to be Υip, where

r o vip − vip Υip = tot ∗ 100 (1) vip

r Here, ip is a state assembly constituency in a particular electoral period; vip is the total votes o polled by the candidate of the ruling party; vip is the total votes polled by the candidate tot securing the highest vote other than that of the ruling party; vip is the total votes polled in the assembly constituency in the last election. In other words, it is the vote difference of the ruling party candidate with the next-best other party candidate as a percentage of the total number of votes polled in the constituency in the last state assembly election. My preferred approach (I report the results using this approach) to objectively define a ruling party is to select the party which has secured a simple majority in the number of seats won on a first-past-the-poll basis. I leave out electoral periods in which no political party has secured a simple majority. I leave out constituencies in which the ruling party did not contest (these are mostly during the electoral periods when the party had a pre-poll alliance with another party).24 Apart from the benefit of objectivity, this approach leaves out coalition governments.25 An alternative is to include periods of coalition governments by identifying ruling coalitions and considering a party to be ruling if it is included in the ruling coalition. I report the results with ruling coalitions in appendixA. 26

In my main empirical speciﬁcation, I restrict the data to the relevant bandwidth and run the following regression:

0 0 1 2 Yip = β · Dip + γ1 · [f(|Υip|)] + γ2 · [g(|Υip|)] + αrp + αip + εip (2)

ip is a state-assembly constituency in a particular electoral period; Υip is the running vari-

24I assume that there is no major defection in the ruling party. This was mostly the case for our period (Andhra Pradesh between May 2009 to May 2014 and Arunachal Pradesh after May 2014 are exceptions). I do not include the state of Goa for the electoral period beginning 2012 since the majority party won by one seat and hence a ruling party legislator is also pivotal in determining the majority party. 25Rajani[2016] finds coalition politics to influence government spending in India. In a different context, Artés and Jurado[2018] finds a difference in government functioning between coalition and single party governments. 26I leave out the state of Bihar from my analysis since none of the coalition governments for either electoral period was stable. I do not include Delhi for the short period after December 2013 elections until elections in February 2015 and Jammu and Kashmir after December 2014 since the states were/are under an unstable coalition followed by President’s (Governor’s) rule.

15 able as defined in Equation1; Dip = 1(Υip > 0) is the indicator variable of whether the constituency’s representative is aligned to the ruling party; f(|Υip|) is a vector of polynomial m terms of |Υip|: (|Υip|, ...., |Υip| ) while g(|Υip|) is a vector of polynomial terms of |Υip| in- m 1 teracted with Dip: (|Υip| ∗ Dip, ...., |Υip| ∗ Dip), m ≥ 1; αrp are region-electoral period fixed 2 effects, αip are the fixed effects for the political party of the constituency representative; Yip is the dependent variable which is the particular economic outcome/policy. β is the effect of being aligned to the ruling party when compared to not being aligned to the ruling party 1 m 1 m 1 under close elections. γ1 = (γ1 , ..., γ1 ) and γ2 = (γ2 , ..., γ2 ) are parameter vectors. γ1 and 1 1 γ1 + γ2 are the increase in the policy variable Yip with respect to an increase in the absolute value of the vote share difference of the ruling party among the nonaligned constituencies 27 1 and aligned constituencies respectively when elections are reasonably close. γ2 is the additional relationship of the policy variable with respect to the absolute value of the vote share difference among the aligned constituencies as compared to nonaligned constituencies under close elections.

1 My preferred specification is to take αrp to be state-electoral period fixed effects and m = 1. My specification is a case of a stacked RD design since I consider many electoral periods for different states. There is no consensus on the optimal bandwidth choice in case of a stacked RD design (Bertanha[2017], Francis-Tan and Tannuri-Pianto[2018]). I report the results for bandwidths 0.5, 1, and 1.5 times the coverage error rate optimal bandwidth varying on both sides of the cut-off (suggested in Calonico, Cattaneo, and Farrell[2018], Calonico, Cattaneo, and Farrell[2018a] and Calonico, Cattaneo, and Farrell[2018b]) for m = 1 with the dependent variables partialled out with respect to the fixed effects (henceforth 0.5CCT, CCT, 1.5CCT bandwidths respectively).28 I also report the results with the CCT bandwidth 1 for m = 2 as well as for m = 1 taking αrp to be parliamentary constituency-electoral period fixed effects.29 I do not consider m > 2 since following Gelman and Imbens[2018], recent empirical works involving regression discontinuity design (e.g. Baskaran and Hessami[2018])

27 k k0 0 The effect of γ1 or γ2 for k, k > 1 is approximately zero under close elections. 28The bandwidth varies with both the dependent variable, as the value of m. I use the triangular kernel to select the coverage error rate optimal bandwidth varying on both sides of the cut-off. I use the STATA command rdbwselect (Calonico, Cattaneo, Farrell, and Titiunik[2017]). 29A parliamentary constituency is a regional unit from which a representative is elected to the Union parliament. While the total number of legislative assembly constituencies per each assembly constituency varies with states, on average there are about eight assembly constituencies per parliamentary constituency. Another alternative is to use district fixed effects. Each state is divided into districts. A district is a unit of state governmental administration and hence district fixed effects capture variations at the level of administrative capacity. Parliamentary constituencies on the other hand are political units, and parliamentary constituency fixed effects capture variations due to political factors of Union government elections. I do not add state assembly constituency fixed effects as in Bracco, Lockwood, Porcelli, and Redoano[2015] since the number of electoral periods with consecutive non-coalition governments is low and reducing the specification to close elections reduces the number of observations substantially.

16 have avoided using higher order polynomials. I cluster the standard errors at state (as in 2008) level (the number of clusters are no more than 23 which makes the critical value of the t-statistic higher than that of a standard normal distribution).

Many of the outcome variables are at an annual frequency and hence some of the data are for periods in which two electoral periods overlap. Also, some of the data is not available for the entire period. Hence, for each variable, I consider each electoral period to comprise x − x data from six years. I weight each year of the electoral period by 1 2 , where x and x 12 1 2 are the number of months of the electoral period and another electoral period respectively for which data are included in the year. Further, I weight each electoral period by the sum of the annual weights considering the weight to be zero for negative weights (by this method, I weigh electoral periods for which data are available for the entire period more than those for which data are only available partially). For each subindex, I take the weights to be a simple average of the weights of the variables which form the components of the subindex (analogously, for each index, I take the weights to be a simple average of the weights of the subindices which form the components of the index).

4 Theory and Empirical Implication

In this section, I use a simple model to spell out a theory which is consistent with the results that I discuss in Section6. Here, I consider a simple extension of the theory in Brollo and Nannicini[2012] from a single type of governmental policy to a possibility of diﬀerent type of policies which a government can allocate for electoral gains. Unlike Brollo and Nannicini [2012], whose theory predicts aligned constituencies to get atleast as much governmental policies as nonaligned constituencies, the model in this section predicts the possibility of aligned constituencies to get more of some type of policies and less of others (although not formally modeled, the theory and its predictions have been propounded in Bueno[2018]). I also propose some additional empirical implications of the observed relationship of policies with past electoral performance of the ruling political party.

There are two political parties, the ruling political party r, opposition political party o. I denote the set of constituencies as Ω. Each constituency i ∈ Ω has a parameter r o vi −vi r o Υi = r o ; vi and vi are the total votes polled by the candidate of the ruling party and the vi +vi opposition party respectively in the last assembly election. The vote share diﬀerence, Υi has a continuous distribution function of g(x) where g(x) > 0 for x ∈ [−1, 1].

17 I denote the set of policies by Ψ. Each policy j ∈ Ψ has a parameter θj which has a distribution function of f(x) where f(x) ≥ 0, f(x) = 0 for x∈ / [0, 1]. θj measures the fraction of goodwill that gets reflected to the ruling party while (1−θj) measures the fraction of goodwill that gets reflected to the constituency representative (unlike Bracco, Lockwood, Porcelli, and Redoano[2015], I do not endogenise θj). Examples of policies with θj u 1 maybe allotment of new administrative headquarters, establishment of new educational institutes, etc. Implementation of different government schemes may have an intermediate value of θj while examples of policies of low values of θj maybe those involving easing of bureaucratic regulations like granting of mining permits (Asher and Novosad[2017]).

The state government, affiliated to the ruling political party, is responsible for policy- making. I do not model the possibility of local constituency representatives influencing policies (I briefly talk about the possibility of variation in the competence of legislators while discussing alternative mechanisms in Section7).

j The state government chooses an allocation x ≡ {xi }j∈Ψ,i∈Ω, x ≥ 0 to maximize the net return: Z Z j j U(x) = {u(θ , Υi) · v(xi ) + c(Xi)}djdi (3) i∈Ω j∈Ψ

R j 2 where (i) Xi ≡ j∈Ψ xi dj ∀i ∈ Ω; (ii) u : < 7−→ (0, ∞); v : [0, ∞) 7−→ (0, ∞); c : [0, ∞) 7−→ <; (iii) c0(0) ≥ 0; ∀y ≥ 0, v0(y) > 0, v00(y) < 0, c00(y) < 0; (iv) the maximization problem in equation3 has a unique solution.

Equation3 can be the outcome of various objectives of the ruling party which I do not model. Together with my subsequent assumptions on u : <2 7−→ (0, ∞); v : [0, ∞) 7−→ (0, ∞); c : [0, ∞) 7−→ <, equation3 nests Brollo and Nannicini[2012]’s model of the ruling party seeking to maximize a weighted average of electoral fortunes at federal and municipal level by targeting a single type of policy as a special case (the institutional structure in India is diﬀerent. Nevertheless, I can analogously think of the problem as a weighted average of social welfare and the future electoral success of the ruling party). As I discuss later in the section, equation3 can also incorporate various correlations of vote share of the ruling party in the past election with various other objectives of the ruling party, for instance the welfare of each factions among the ruling party (Persico, Pueblita, and Silverman[2011]).

Like Brollo and Nannicini[2012], the functional form in equation3 assumes separability j0 j00 0 00 of U(x) in any xi0 , xi00 , i 6= i . The separability assumption implies: (i) additional returns from a constituency to the ruling party does not aﬀect additional returns from another

18 constituency to the ruling party. This assumption holds if the electoral objective of the ruling party is to maximize the expected number of seats won; and (ii) additional policy to a constituency does not affect the cost of an additional policy to another constituency. This assumption may hold if the total resources to allocate is not fixed (state government in India, often have budget deficits). This assumption may also hold if, for example, the resource allotment for a policy is fixed for each constituency, although the implementation depends upon the level of effort (this may be plausible for several central government sponsored schemes like the rural road construction scheme: PMGSY).

Assumption 1. u(θ, Υ) is strictly increasing in θ ∀(θ, Υ) ∈

Assumption1 is motivated by the premise (e.g. in Arulampalam, Dasgupta, Dhillon, and Dutta[2009]) that among nonaligned constituencies, given Xi, returns from a policy to the ruling party increases with an increase in attribution to the state government vis-a-vis the constituency representative.

Assumption 2. u(θ, Υ) is continuous for ∀(θ, Υ) ∈

Assumption 3. (i) u(1, Υ) is continuous in Υ 0 0 (ii) (∀θ ∈ [0, 1), ∃(θ ) ∈ (0, ∞)) such that (∀δ > 0), ∼ [∀(Υ−, Υ+) ∈ (−δ, 0)X(0, δ), 0 0 0 u(θ , Υ−) − u(θ , Υ+) ≤ (θ )]

Assumption3 is motivated by the premise that: (i) if a policy is attributed completely to the state government, then given Xi, additional returns from the policy to the ruling party does not depend on the level of alignment; and (ii) (e.g. in Brollo and Nannicini [2012]) if a policy is partially attributed to the local constituency representative, then given

Xi, additional returns from the policy to the ruling party is more for aligned constituencies than for nonaligned constituencies under close elections. I discuss possible violations of assumption3 in Section7.

Assumption 4. 0 R 1 0 R δ ∀δ < 1, δ f(y)dy > 0 and ∃δ ∈ [0, 1) such that 0 f(y)dy > 0

Assumption4 implies that there exist some policies which are attributed wholly to the state government and also some policies which are not attributed wholly to the state government.

19 j Let {xî }j∈Ψ,i∈Ω be the optimal allocation. It is immediate from the above model that 0 0 j0 0 00 0 00 V j0 for any j ∈ Ψ, i ∈ Ω, xî0 > 0. Further, ∀(j ∈ Ψ, j ∈ Ψ, i ∈ Ω, i ∈ Ω), [(Υi0 = Υi00 ) (θ = j00 j0 j00 θ ) −→ (ˆxi0 =x î00 )]. Hence, the optimal allocation is a function: xˆ : SΥXSθ 7−→ (0, ∞), j where SΥ = {x : ∃i ∈ Ω s.t. Υi = x} and Sθ = {y : ∃j ∈ Ψ s.t. θ = y}. Total allocation to a ˆ constituency Xi is a function: X : SΥ 7−→ (0, ∞).

Lemma 1. 0 00 0 00 Suppose assumptions1 to3 hold. Then, ∃ δ0 > 0 such that (∀ j , j ∈ Ψ, j < j ) and

(∀(Υ−, Υ+) ∈ (−δ0, 0)X(0, δ0)): j00 j0 (i) xˆ(Υ−, θ ) > xˆ(Υ−, θ ) ˆ ˆ (ii) X(Υ+) > X(Υ−)

Proposition 1. 0 00 Suppose assumptions1 to4 hold. Then, ∃ δ1 > 0, p ∈ (0, 1), q ∈ [0, 1) such that (∀ j , j ∈ 0 00 Ψ, j < j ) and (∀(Υ−, Υ+) ∈ (−δ1, 0)X(0, δ1)): j00 j00 j00 (i)xˆ(Υ+, θ ) < xˆ(Υ−, θ ) if θ ∈ (p, 1] j0 j0 j0 (ii) xˆ(Υ+, θ ) > xˆ(Υ−, θ ) if θ = q

If a policy is attributed to the state government, then aligned constituencies get less of it under close elections. Further, there exists some policy which is not wholly attributed to the state government, such that aligned constituencies get more of it under close elections. Proposition1 has the following empirical implication:

Empirical Implication: β < 0 for policies which are attributed to the state government. In other words, if a policy is attributed to the state government, then aligned constituencies get less of it under close elections. However, there exists some policy which is not wholly attributed to the state government, such that β > 0. In other words, under close elections, aligned constituencies get more of some policies, which are attributed, at least partially, to the constituency representative.

I am however agnostic about the signs of γ1 and γ2 which can reﬂect correlations from various socio-economic factors, although there are suggestive implications from two types of relationships which have been generally recognized in the literature: ‘swing voter models’ and ‘loyal voter models’.30 ‘Swing voter models’ propose that when the ruling party’s objective is to maximize the expected number of assembly seats won in the next election, it is optimal for the party to target policies most to constituencies which are expected to be the most

30See Hirano, SnyderJr, and Ting[2009] for a short discussion of the related literature.

20 competitive in the next election. Since past competitiveness is a reasonable predictor of future competitiveness, if a policy is attributed wholly to the state government, then such a policy may be targeted most to the constituencies which had the lowest absolute vote share difference in the last assembly election. On the other hand, theories (e.g. Persico, Pueblita, and Silverman[2011]) and empirical evidence (e.g. Ansolabehere and Snyder[2006]) also indicate that government policies are often targeted to ruling party strongholds which are likely to have a positive relationship with past vote share difference of the ruling political party. Further, it is realistic in the Indian context for such a relationship to be a continuous function of the past vote share difference of the ruling party (Lee[2016]). Combining the above two theories with the theory in Arulampalam, Dasgupta, Dhillon, and Dutta[2009], I assert the following:31

Proposition 2. 0 00 0 00 ∃1 > 0, 2 ∈ (0, 1) such that (∀ j , j ∈ Ψ, j < j ) and (∀(Υ−, Υ+) ∈ (−1, 0)X(0, 1)), ˆ θ ∈ (2, 1) : xˆ(Υ , θj00 ) ∂ i ˆ j0 ∂X(Υi) xˆ(Υi, θ ) (i) |Υi=Υ− > 0 and |Υi=Υ− > 0. Hence, ∂Υi ∂Υi ˆ j00 j0 ∂xˆ(Υi, θ) j00 j0 ∂xˆ(Υi, θ ) ∂xˆ(Υi, θ ) |Υi=Υ− > 0 and since xˆ(Υ−, θ ) > xˆ(Υ−, θ ), |Υi=Υ− > |Υi=Υ− ∂Υi ∂Υi ∂Υi ˆ ˆ ∂xˆ(Υi, θ) ∂xˆ(Υi, θ) (ii) |Υi=Υ− > − |Υi=Υ+ ∂Υi ∂Υi

Empirical Implication: Under close elections, for policies which are attributed mostly to the state government, (i) γ1 < 0, or the relationship of policies with absolute vote share diﬀerence among nonaligned constituencies is negative; (ii) γ2 > 0, or the relationship of policies with absolute vote share diﬀerence among aligned constituencies is stronger than among nonaligned constituencies. The sign of the relationship of policies with

absolute vote share diﬀerence among aligned constituencies (γ1 + γ2) is however ambiguous.

Comparative Statics: The comparative statics also differ from the theory of single- type of policies in Brollo and Nannicini[2012]. An increase in the difference in benefit to the ruling party from an additional policy (partially attributed to the local constituency representative) between aligned and nonaligned constituencies can lead to nonaligned constituencies getting even more of policies which are attributed wholly to the state government.

31For proposition2 to hold in the model, I require additional assumptions on u : <2 7−→ (0, ∞); v : [0, ∞) 7−→ (0, ∞); c : [0, ∞) 7−→ <.

21 5 Validity of the RD Design

To show the validity of the regression equations2, I conduct a number of balance checks:

Testing for difference in density of the running variable at the cut-off: I implement the density tests proposed in Cattaneo, Jansson, and Ma[2017] to test for the null hypothesis of a difference in the density of the running variable (as I define in Equation1) on both sides of the cut-off. A large sample implication of a valid regression discontinuity design is continuity of the density of the running variable at the cut-off. Hence, a rejection of the null hypothesis may be indicative of a possible manipulation of the running variable, questioning the validity of a causal interpretation of the regression discontinuity estimates. The density plot in Figure1 does not reject the null of hypothesis of a difference in density of the running variable.

Balancedness of electoral characteristics of the constituency at baseline: A large sample implication of a valid regression discontinuity design is similarity of pre-treatment characteristics of both treatment and control groups on both sides of the cut-off. In order to show that my treatment and control constituencies do not differ in pre-treatment characteristics, I implement equation2 taking the dependent variable to be 100 ∗ log(1 + x) where x is: (a) The total electors in the assembly constituency in the last electoral period; (b) The total votes polled in the assembly constituency in the last electoral period; (c) The total rural population according to the 2011 census.32 In tables1,2 and3, none of the estimates of the coefficients testing for the difference in baseline characteristics for total number of electors, total votes polled and total rural population are statistically significant at the 10% level. Further, both the percentage difference in total electors or total votes polled are less than 1%. The aligned constituencies have 6.59% more total rural population than the nonaligned constituencies for my preferred specification with a linear polynomial and optimal CCT bandwidth, although the sign of the coefficients is sensitive to the choice of bandwidth, polynomial order or fixed effects.

32For this I geocoded the village directory of 2011 census using google maps.

22 6 Results

Graphical Evidence: I first graphically explore the effect of political alignment on policy and outcome variables with a regression discontinuity plot of the partialled out (with respect to state-period Fixed Effects and political party Fixed Effects) variable. The running variable (Υip) is the winning margin, in percentage, of the ruling party candidate (as I define in equation1). I restrict the plot to observations of Υip within the coverage error rate optimal bandwidth (varying on both sides of Υip = 0, selected using a triangular kernel) and fit observations with a linear polynomial of varying slope on both sides of Υip = 0. I report the regression discontinuity plot for the indices and subindices of long-term investment goods and government schemes as I describe in Section 2.4 and also for the change in log luminosity from satellite night-time lights data (Figures2 to 10). Since my preferred specification in equation2 includes fixed effects, the jump at the cut-off for the regression discontinuity plots differ from the estimates shown in the first row of column (1) in Tables4 to 12. Hence, the plots are only visual depictions of the observations around the threshold and not estimates of the difference in policy/outcome variables between aligned and nonaligned constituencies under close elections. At a margin of victory of zero, I observe a drop in the index for administrative headquarters, state government-managed educational institutes and the combined index for long-term investment goods, indicating that aligned constituencies have less growth in administrative headquarters and state government-managed educational institutes under close elections. There is however no significant difference in the plot between aligned and nonaligned constituencies for the various indices pertaining to government schemes implemented by the state government or growth in satellite night lights data.

Regression Results: Next, I describe the results from implementing the baseline spec- iﬁcation in equation2 for the outcome variables that I have described in Section 2.4.

I Index for growth of long-term investment goods which are taken by the state government: Under close elections, the estimate of the index for growth of long-term investment goods (including only state government-managed educational institutes) is 0.2 less in aligned constituencies under close elections for my preferred specification with a linear polynomial and optimal CCT bandwidth (Table6, first row, column (1)). The estimate is over one-fourth of the standard deviation of the index among nonaligned constituencies. The sign of the estimate is robust to use of alternative bandwidths (columns 2−3), quadratic specification (column (4)) or use of parliamentary-constituency-electoral period fixed ef-

23 fects (column (5)). The sign of the estimate of relationship of the index variable with the absolute value of vote share difference among nonaligned is constituencies is negative (Table6, second row) while the sign of the estimate for the difference in the relationship with the absolute value of vote share difference in aligned constituencies from nonaligned constituencies is positive (Table6, third row). The estimates are statistically significant for several specifications, including my preferred specification in column (1). The above estimates are consistent with a theory (Section4) of the electorate not attributing long-term investment goods which are taken by the state government to the local constituency representative. The estimates are also robust to the inclusion of educational institutes which are not directly managed by the state government (Table A.18) or when I consider electoral periods with coalition governments (Table A.21).33 The estimates are qualitatively similar for the subindices of administrative headquarters and state-government managed educational institutes, although the statistical significance is weaker.

i Growth of Administrative Headquarters: Under close elections, the estimate of the index for growth of administrative headquarters is 0.14 less in aligned constituencies under close elections for my preferred specification with a linear polynomial and optimal CCT bandwidth (Table4, first row, column (1)). The estimate is about one-fifth of the standard deviation of the index among nonaligned constituencies. The sign of the estimate is robust to use of alternative bandwidths (columns (2) − (3)), quadratic specification (column (4)) or use of parliamentary-constituency-electoral period fixed effects (column (5)). The sign of the estimate of relationship of the index variable with the absolute value of vote share difference among nonaligned is constituencies is negative (Table4, second row) while the sign of the estimate for the difference in the relationship with the absolute value of vote share difference in aligned constituencies from non-aligned constituencies is positive (Table4, third row). The estimates are statistically significant for several specifications, including my preferred specification in column (1). The above estimates are consistent with a theory (Section4) of the electorate not attributing decision to locate administrative headquarters to the local constituency representative. The estimates are qualitatively similar for the individual outcome variables: growth of district headquarters and growth of subdistrict headquarters, although the estimates are not statistically significant for most specifications (Ta-

33I do not include all educational institutes in my main speciﬁcation since the precise role played by the state government in the establishment of any type of educational institute is less clear.

24 bles A.3 and A.4). The t-statistic corresponding to the estimates in the first row for the growth of subdistrict headquarters in Table A.4 is however high for many of the specifications (for my preferred specification in column (1), the growth in subdistrict headquarters is 2.48% less in aligned constituencies when compared to nonaligned constituencies). ii Growth of Educational Institutions: Under close elections, the estimate of the index for growth of state government- managed educational institutes is 0.12 less in aligned constituencies under close elections for my preferred specification with a linear polynomial and optimal CCT bandwidth (Table4, first row, column (1)). The estimate is over 13% of the standard deviation of the index among nonaligned constituencies. The sign of the estimate is robust to use of alternative bandwidths (columns (2) − (3)), quadratic specification (column (4)) or use of parliamentary-constituency-electoral period fixed effects (column (5)), although not statistically significant at the 10% level in columns (2), (3) and (5). The sign of the estimate of relationship of the index variable with the absolute value of vote share difference among nonaligned is constituencies is negative (Table5, second row) while the sign of the estimate for the difference in the relationship with the absolute value of vote share difference in aligned constituencies from non-aligned constituencies is positive (Table5, third row) although not statistically significant for my preferred specification in column (1). The statistical significance of the estimates increase with the inclusion of educational institutes which are not directly managed by the state government (Table A.17). The estimates are not statistically significant for the individual outcome variables: growth of state-government-managed schools and growth of state-government-managed higher educational institutes (Tables A.5 and A.6).

My estimates with annual data on two categories of long-term investment goods document the possibility of aligned constituencies to get less of policies which are not attributed to the local constituency representative. This result complements the work in Asher and Novosad[2017] who does not find a statistically significant difference in growth of public infrastructure: roads, power, schools, irrigation and hospitals using decennial census data for the year 1991 and 2001.

II Government schemes implemented by the state government: Unlike the index for growth of long-term investment goods, there is little evidence of aligned constituencies to have less receipts or implementation of government schemes.

25 Under close elections, the estimate of the index for government schemes implemented by the state government is 0.02 (not statistically significant) higher in aligned constituencies for my preferred specification with a linear polynomial and optimal CCT bandwidth (Table 11, first row, column (1)). The estimate is little over 2% of the standard deviation of the index among nonaligned constituencies. The sign of the estimate varies with specifications. The estimate is positive and statistically significant at the 10% level for the specification with a linear polynomial and 1.5 times the optimal CCT bandwidth (column 3). None of the estimates are negative, except for the specification in column 2 which is −0.011 and not statistically significant. The sign of the estimates of relationship of the index variable with the absolute value of vote share difference in the second and third row of Table 11 is small, varies with specifications and is not statistically significant. The results remain qualitatively similar when I consider electoral periods with coalition governments (Table A.22). I detail the results for the individual indices and outcome variables below:

i Rural Employment Guarantee Scheme (MGNREGS): Under close elections, the estimate of the index for MGNREGS asset works is not statistically different among aligned and nonaligned constituencies. The estimate is 0.027 (not statistically significant) higher in aligned constituencies for my preferred specification with a linear polynomial and optimal CCT bandwidth (Table7, first row, column (1)). None of the estimates for the individual variables of number of asset works ongoing (Table A.7) or completed (Table A.8) is statistically significant. ii Housing Schemes: Under close elections, the estimate of the index for housing schemes is not statistically different among aligned and nonaligned constituencies. The estimate is 0.019 (not statistically significant) higher in aligned constituencies for my preferred specification with a linear polynomial and optimal CCT bandwidth (Table8, first row, column (1)). The estimate for the number of houses completed (Table A.9, first row) is not statistically significant. Aligned constituencies have more utilization of funds under central government sponsored housing schemes than nonaligned constituencies under close elections. The estimate for my preferred specification (Table A.10, first row, column (1)) is 20.7 and statistically significant at the 10% level. iii Rural Road Construction Program (PMGSY): Under close elections, the estimate of the index for rural road construction scheme is not statistically different among aligned and nonaligned constituencies. The estimate is 0.078 (not statistically significant) higher in aligned constituencies for my preferred

26 specification with a linear polynomial and optimal CCT bandwidth (Table8, first row, column (1)). However, the estimate is positive and statistically significant on use of parliamentary-constituency-electoral period fixed effects (column (5)). The estimates for the individual variables: total cost of projects (Table A.11, first row), total expenditure (Table A.12, first row), total length of roads (Table A.13, first row) and total length of roads completed (Table A.14, first row) is positive, although not statistically significant for my preferred specification. The estimate for the total expenditure on projects and total length of roads completed is statistically significant on use of parliamentary-constituency-electoral period fixed effects (column (5)). iv School Receipts and Expenditure: Under close elections, the estimate of the index for total school receipts and expenditures is not statistically different among aligned and nonaligned constituencies. The estimate is 0.0049 (not statistically significant) higher in aligned constituencies for my preferred specification with a linear polynomial and optimal CCT bandwidth (Table 10, first row, column (1)). None of the estimates for the individual variables of total school receipts (Table A.15) or total school expenditure (Table A.16) is statistically significant.

III Growth of Satellite Night Lights: There is little evidence of aligned constituencies to have less increase in log luminosity as indicated by satellite night lights data. Under close elections, the estimate of the change in log luminosity is 2.37 (not statistically significant) higher in aligned constituencies for my preferred specification with a linear polynomial and optimal CCT bandwidth (Table 11, first row, column (1)). The estimate is not statistically different from the estimates for an earlier period in Asher and Novosad[2017] who report a positive and statistically significant change in log luminosity during 1992-2012. The sign of the estimates is positive, although not statistically significant for other specifications (columns (2) − (5)). The estimate (not statistically significant) is more in magnitude when I consider electoral periods with coalition governments (Table A.23).

7 Alternative Mechanisms

In this section I discuss several alternative theories which can explain my result of aligned constituencies getting less of some policies than nonaligned constituencies under close elections.

27 7.1 Diﬀerences in future margin of victory based on alignment

My results can also be explained by a violation of assumption3 to allow for a higher return to the ruling party from allotting policies to nonaligned constituencies under close elections.

Formally, I consider an assumption diﬀerent from assumption3 in the model in Sec- tion4:

Assumption 5. For some m ∈ (0, 1), 0 0 0 (i) (∀θ ∈ (m, 1], ∃ (θ ) ∈ (0, ∞)) such that (∀δ > 0), ∼ [∀(Υ−, Υ+) ∈ (−δ, 0)X(0, δ), 0 0 0 0 [u(θ , Υ−) − u(θ , Υ+)] ≤ (θ )] (ii) u(m, Υ) is continuous in Υ 0 0 0 (iii) (∀θ ∈ [0, m), ∃ (θ ) ∈ (0, ∞)) such that (∀δ > 0), ∼ [∀(Υ−, Υ+) ∈ (−δ, 0)X(0, δ), 0 0 0 0 [u(θ , Υ+) − u(θ , Υ−)] ≤ (θ )]

Assumption5 is motivated by the premise that if a policy is mostly attributed to the state government, then given Xi, additional returns from the policy to the ruling party is more for nonaligned constituencies than for aligned constituencies under close elections. Analogously, I modify assumption4 to assumption6:

Assumption 6. R m1 R 1 ∃(0 < m1 < m < m2 < 1) such that f(y)dy > 0, f(y)dy > 0 0 m2 Proposition 3. 0 00 0 00 If assumptions1,2,5, and6 hold. Then, ∃ δ1 > 0, m ∈ (0, 1), m ∈ (0, 1), m < m such 0 00 0 00 that (∀ j , j ∈ Ψ, j < j ) and (∀(Υ−, Υ+) ∈ (−δ1, 0)X(0, δ1)): j00 j00 j00 00 (i)xˆ(Υ+, θ ) < xˆ(Υ−, θ ) if θ = m j0 j0 j0 0 (ii) xˆ(Υ+, θ ) > xˆ(Υ−, θ ) if θ = m

Assumption5 instead of assumption3 can hold if ex-ante identical constituencies are not ex-post identical from an electoral standpoint. This can arise if the future margin of victory of the ruling party in two constituencies diﬀer based upon the party alignment of the elected representative in the last assembly election.

If the objective of the ruling party is to maximize the expected number of seats won, then the ruling party should divert resources to constituencies which are expected to be competitive or where the margin of victory is expected to be low. Hence, a negative alignment

28 can be observed if nonaligned constituencies are expected to be more competitive in the next assembly election. I expect the actual margin of victory in the next election (although an outcome of policies), to be indicative of the degree of ex-ante competitiveness. For my period of analysis, Table A.19 shows that under close elections, the diﬀerence between aligned and nonaligned constituencies in the level of future electoral competitiveness is not statistically signiﬁcant.

Alternative objectives of the ruling party can predict a relationship of targeting of resources beyond absolute vote share diﬀerence. Snyder[1989] for instance, in his study of a two-party campaign ﬁnance game, also considers a possible objective of maximization of the probability of winning a majority of seats in the legislature. When the ruling party has a higher probability of winning a large majority of seats in the next election, it may not necessarily target swing constituencies since it may prefer to secure its chances of winning a majority of seats by diverting resources to some constituencies where the chances of winning are greater than that of the most competitive ones.34 Hence, when the ruling party is likely to win a majority of seats and has a higher probability of winning in nonaligned consistencies in the next assembly election, assumption5 instead of assumption3 may hold.

I expect the actual probability of re-election of the ruling party (although an outcome of policies), to be indicative of the degree of ex-ante probability of re-election. For my period of analysis, I find a significantly less probability of a ruling party to get reelected under close elections (Table A.20). This resonates with the findings of several studies (e.g. Uppal [2009]) which document a lower probability of reelection of incumbent legislators in Indian state assembly elections. However, the negative alignment effect persists for the growth of index of long-term investment goods when I exclude high majority electoral periods in states with prominent incumbency disadvantage (Table 13). I use Lee[2016] to identify states with prominent incumbency disadvantage in India.35 I consider high majority electoral periods to be those for which in the last assembly election, the ruling party had a share of total seats not below 62.585 (the median in my data).

34Even under this maximization principle, however, the ruling party has little reason to target more to constituencies where they are sure to win. 35Lee[2016] ﬁnds a strong and statistically signiﬁcant incumbency disadvantage in north, west and north- western states of India. These include the states of Bihar, Jharkhand, Chhattisgarh, Madhya Pradesh, Uttar Pradesh and Utttarakhand (north), Gujarat, Goa and Maharashtra, (west), Punjab, Haryana, Delhi, Rajasthan, Himachal Pradesh and Jammu and Kashmir (north-west).

29 7.2 Competence

Asher and Novosad[2017] argue that if success is correlated with the ability of a candidate for the economic growth of a constituency, then my estimate from Equation2 may be negative. This is because under close elections, elected candidates from the ruling party are less successful than average, while elected candidates from the non-ruling party are more successful than average (Caughey and Sekhon[2011], Ferraz and Finan[2011]). However, there are indications of goods which a non-ruling party legislator is unlikely to inﬂuence to also grow more in the non-ruling party constituency: for instance, administrative headquarters (tables A.3 and A.4) and state-government-managed educational institutes (tables A.5 and A.6). A legislative constituency representative can directly aﬀect the utilization of funds under the constituency local area development grants scheme. Availability of data on projects under the local area development scheme may allow me to better distinguish competence between elected candidates from ruling and non-ruling parties under close elections.

8 Conclusion

In this paper I find evidence of aligned constituencies having less growth in administrative headquarters and state-government managed educational institutes than nonaligned constituencies. However there is little evidence of aligned constituencies having less receipts and implementation of different government schemes or less growth in night-time luminosity. Together with previous findings of more growth in aligned constituencies, possibly due to easing of bureaucratic regulations, I can explain my findings by a theory where some policies are attributed partly to the local constituency representative while others are attributed mostly to the state government, with the state government giving the former more in aligned constituencies and substituting the former with the latter in nonaligned constituencies. Such a theory has several implications. First, this theory suggests a possibility of differential implications of political alignment among different groups of people. This is because sections of the society who benefit from each type of policy are likely to be non-overlapping. Second, this theory points to the possibility of an underestimate in the variation of targeting of individual policies if the observed variable is a mixture of different type of policies. Availability of data on projects under the local area development scheme, which a legislative constituency representative can directly affect, may allow me to better understand the mechanisms behind my results.

30 References

Anjali Adukia, Sam Asher, and Paul Novosad. Educational Investment Responses to Economic Opportunity: Evidence from Indian Road Construction. 2017. URL http: //www.dartmouth.edu/~novosad/adukia-asher-novosad-dise-roads.pdf.

Shilpa Aggarwal. Do rural roads create pathways out of poverty? Evidence from India. Journal of Development Economics, 133:375 – 395, 2018. ISSN 0304-3878. doi: https:// doi.org/10.1016/j.jdeveco.2018.01.004. URL http://www.sciencedirect.com/science/ article/pii/S0304387818300063.

Moinuddin Ahmad. If You Are Wondering What Is An ‘Anglo-Indian MLA’, Here’s What It’s All About. 2018. URL https://www.indiatimes.com/news/india/ if-you-are-wondering-what-is-an-anglo-indian-mla-here-s-what-it-s-all-about-345674. html.

AISHE. All India Survey of Higher Education. Ministry of Human Resource Development, 2018. URL http://aishe.nic.in/aishe/home.

Vivi Alatas, Abhijit Banerjee, Rema Hanna, Benjamin A. Olken, Ririn Purnamasari, and Matthew Wai-Poi. Does Elite Capture Matter? Local Elites and Targeted Welfare Pro- grams in Indonesia. NBER Working Paper, (18798), 2013. URL http://www.nber.org/ papers/w18798.

Michael L. Anderson. Multiple Inference and Gender Diﬀerences in the Eﬀects of Early Intervention: A Reevaluation of the Abecedarian, Perry Preschool, and Early Training Projects. Journal of the American Statistical Association, 103(484):1481–1495, 2008. doi: 10.1198/016214508000000841. URL https://doi.org/10.1198/016214508000000841.

Stephen Ansolabehere and James M. Snyder. Party Control of State Government and the Distribution of Public Expenditures. Scandinavian Journal of Economics, 108(4):547–569, 2006. ISSN 1467-9442. doi: 10.1111/j.1467-9442.2006.00470.x. URL http://dx.doi.org/ 10.1111/j.1467-9442.2006.00470.x.

Joaquín Artés and Ignacio Jurado. Government fragmentation and ﬁscal deﬁcits: a regression discontinuity approach. Public Choice, 175(3):367–391, Jun 2018. ISSN 1573-7101. doi: 10.1007/s11127-018-0548-y. URL https://doi.org/10.1007/s11127-018-0548-y.

Wiji Arulampalam, Sugato Dasgupta, Amrita Dhillon, and Bhaskar Dutta. Electoral goals and center-state transfers: A theoretical model and empirical evidence from In- dia. Journal of Development Economics, 88(1):103 – 119, 2009. ISSN 0304-3878. doi:

31 https://doi.org/10.1016/j.jdeveco.2008.01.001. URL http://www.sciencedirect.com/ science/article/pii/S0304387808000059.

Sam Asher and Paul Novosad. Politics and Local Economic Growth: Evidence from India. American Economic Journal: Applied Economics, 9(1):229–73, January 2017. doi: 10.1257/app.20150512. URL http://www.aeaweb.org/articles?id=10.1257/app. 20150512.

Sam Asher and Paul Novosad. Rural Roads and Local Economic Develop- ment. 2018. URL https://ramblinsamblog.files.wordpress.com/2018/02/ asher-novosad-rural-roads.pdf.

Sam Asher, Karan Nagpal, and Paul Novosad. The Cost of Distance: Geography and Governance in Rural India (working paper). 2017. URL http://www.cafral.org.in/ sfControl/content/Speech/214201834356PMKaranNagpalJMP.pdf.

Pranab K. Bardhan and Dilip Mookherjee. Capture and Governance at Local and National Levels. American Economic Review, 90(2):135–139, May 2000. doi: 10.1257/aer.90.2.135. URL http://www.aeaweb.org/articles?id=10.1257/aer.90.2.135.

Thushyanthan Baskaran and Zohal Hessami. Does the Election of a Female Leader Clear the Way for More Women in Politics? American Economic Journal: Economic Policy, 10(3):95–121, August 2018. doi: 10.1257/pol.20170045. URL http://www.aeaweb.org/ articles?id=10.1257/pol.20170045.

Indrani Basu. Karnataka Elections: As BJP Hunts For MLAs, Anglo-Indians Bow Out Of Floor Test. 2018. URL https://www.huffingtonpost.in/2018/05/18/ karnataka-elections-as-bjp-hunts-for-mlas-anglo-indians-bow-out-of-floor-test_ a_23437949/.

Erlend Berg, Sambit Bhattacharyya, D. Rajasekhar, and R. Manjula. Can public works increase equilibrium wages? Evidence from India’s National Rural Employment Guarantee. World Development, 103:239 – 254, 2018. ISSN 0305-750X. doi: https://doi.org/10.1016/ j.worlddev.2017.10.027. URL http://www.sciencedirect.com/science/article/pii/ S0305750X15311335.

Marinho Bertanha. Regression Discontinuity Design with Many Thresholds. 2017. URL https://ssrn.com/abstract=2712957.

Govind Bhattacharjee. MGNREGA as Distribution of Dole. Economic and Political Weekly, Vol. 52(Issue No. 25-26), 24 2017.

32 Bhuvan-MGNREGA. Ministry of Rural Development. 2018. URL https://bhuvan-app2. nrsc.gov.in/mgnrega/mgnrega_phase2.php.

Nayana Bose. Raising Consumption Through India’s National Rural Employment Guarantee Scheme. World Development, 96:245 – 263, 2017. ISSN 0305-750X. doi: https://doi.org/10. 1016/j.worlddev.2017.03.010. URL http://www.sciencedirect.com/science/article/ pii/S0305750X15304216.

Emanuele Bracco, Ben Lockwood, Francesco Porcelli, and Michela Redoano. Intergovern- mental grants as signals and the alignment eﬀect: Theory and evidence. Journal of Public Economics, 123:78 – 91, 2015. ISSN 0047-2727. doi: https://doi.org/10.1016/ j.jpubeco.2014.11.007. URL http://www.sciencedirect.com/science/article/pii/ S0047272714002291.

Fernanda Brollo and Tommaso Nannicini. Tying Your Enemy’s Hands in Close Races: The Politics of Federal Transfers in Brazil. American Political Science Review, 106(4):742–761, 2012. doi: 10.1017/S0003055412000433. URL https://www. cambridge.org/core/journals/american-political-science-review/article/ tying-your-enemys-hands-in-close-races-the-politics-of-federal-transfers-in-brazil/ 3262A98BFF83D8E2CED402039FA33C4E.

Natália S. Bueno. Bypassing the Enemy: Distributive Politics, Credit Claiming, and Non- state Organizations in Brazil. Comparative Political Studies, 51(3):304–340, 2018. doi: 10.1177/0010414017710255. URL https://doi.org/10.1177/0010414017710255.

Michael Callen, Saad Gulzar, and Arman Rezaee. Can Political Alignment be Costly. 2018. URL https://dl.dropboxusercontent.com/content_link/ NJpeGTgs2YNbzLaPg0I9xYRr4JPE0HRv043PfxRSvQKxcwlxg8CT9i9PGCLCBOOp/file.

S. Calonico, M. D. Cattaneo, and M. H. Farrell. Coverage Error Optimal Conﬁdence Inter- vals. ArXiv e-prints, August 2018a.

S. Calonico, M. D. Cattaneo, and M. H. Farrell. Optimal Bandwidth Choice for Robust Bias Corrected Inference in Regression Discontinuity Designs (Work- ing Paper). 2018b. URL http://www-personal.umich.edu/~cattaneo/papers/ Calonico-Cattaneo-Farrell_2018_CERRD.pdf.

Sebastian Calonico, Matias D. Cattaneo, Max H. Farrell, and Rocio Titiunik. rdrobust: Software for regression-discontinuity designs. Stata Journal, 17(2):372–404, June 2017. URL https://ideas.repec.org/a/tsj/stataj/v17y2017i2p372-404.html.

33 Sebastian Calonico, Matias D. Cattaneo, and Max H. Farrell. On the Eﬀect of Bias Esti- mation on Coverage Accuracy in Nonparametric Inference. Journal of the American Sta- tistical Association, 113(522):767–779, 2018. doi: 10.1080/01621459.2017.1285776. URL https://doi.org/10.1080/01621459.2017.1285776.

Grace Carswell and Geert De Neve. MGNREGA in Tamil Nadu: A Story of Success and Transformation? Journal of Agrarian Change, 14(4):564–585, 2014. ISSN 1471-0366. doi: 10.1111/joac.12054. URL http://dx.doi.org/10.1111/joac.12054.

M. D. Cattaneo, M. Jansson, and X. Ma. Simple Local Polynomial Density Estimators. working paper, University of Michigan, 2017. URL http://www-personal.umich.edu/ ~cattaneo/papers/Cattaneo-Jansson-Ma_2017_LocPolDensity.pdf.

Devin Caughey and Jasjeet S. Sekhon. Elections and the Regression Discontinuity Design: Lessons from Close U.S. House Races, 1942-2008. Political Analysis, 19(4):385–408, 2011. ISSN 10471987, 14764989. URL http://www.jstor.org/stable/41403727.

Shawn Cole. Fixing Market Failures or Fixing Elections? Agricultural Credit in India. American Economic Journal: Applied Economics, 1(1):219–50, January 2009. doi: 10. 1257/app.1.1.219. URL http://www.aeaweb.org/articles?id=10.1257/app.1.1.219.

Dailybite. Karnataka drama raises questions on practice of nominat- ing Anglo-Indian MLAs. 2018. URL https://www.dailyo.in/variety/ anglo-indians-legislative-assembly-karnataka-governor-vajubhai-vala/ story/1/24259.html.

Datameet. India Assembly Constituencies-shapeﬁles. Community Created Maps of India, 2017. URL http://projects.datameet.org/maps/assembly-constituencies/.

Torun Dewan and Kenneth A. Shepsle. Political Economy Models of Elections. Annual Review of Political Science, 14(1):311–330, 2011. doi: 10.1146/annurev.polisci.12.042507. 094704. URL https://doi.org/10.1146/annurev.polisci.12.042507.094704.

Amrita Dhillon, Pramila Krishnan, Manasa Patnam, and Carlo Perroni. Electoral Account- ability and the Natural Resource Curse: Theory and Evidence from India. CEPR Dis- cussion Paper No. DP11377, 2016. URL https://papers.ssrn.com/sol3/papers.cfm? abstract_id=2807788.

DISE. District Information System for Education. National University of Education Planning and Administration, 2018. URL http://schoolreportcards.in/SRC-New/.

34 DMSP-OLS. Version 4 DMSP-OLS Nighttime Lights Time Series. National Centers for En- vironmental Information, National Oceanic and Atmospheric Administration, 2018. URL https://ngdc.noaa.gov/eog/dmsp/downloadV4composites.html.

Jean Drèze and Reetika Khera. Recent Social Security Initiatives in India. World De- velopment, 98:555 – 572, 2017. ISSN 0305-750X. doi: https://doi.org/10.1016/j. worlddev.2017.05.035. URL http://www.sciencedirect.com/science/article/pii/ S0305750X17302097.

ECI. Election Results- Full Statistical Reports. Election Commission of India, 2018. URL https://eci.nic.in/eci_main1/ElectionStatistics.aspx.

Christopher D Elvidge, Kimberly Baugh, Mikhail Zhizhin, Feng Chi Hsu, and Tilottama Ghosh. VIIRS night-time lights. International Journal of Remote Sensing, 38(21):5860– 5879, 2017. doi: 10.1080/01431161.2017.1342050. URL https://doi.org/10.1080/ 01431161.2017.1342050.

Leopoldo Fergusson, Arturo Harker, and Carlos Molina. Do ghosts exist? Clientelistic networks and corruption in public education. 2018. URL https://www.dropbox.com/s/ qq6hys3jsxd557y/PaperDraft.pdf?dl=0".

Claudio Ferraz and Frederico Finan. Electoral Accountability and Corruption: Evidence from the Audits of Local Governments. American Economic Review, 101(4):1274–1311, June 2011. doi: 10.1257/aer.101.4.1274. URL http://www.aeaweb.org/articles?id= 10.1257/aer.101.4.1274.

Raymond Fisman, Florian Schulz, and Vikrant Vig. The Private Returns to Public Oﬃce. Journal of Political Economy, 122(4):806–862, 2014. doi: 10.1086/676334. URL https: //doi.org/10.1086/676334.

Andrew Francis-Tan and Maria Tannuri-Pianto. Black Movement: Using discontinuities in admissions to study the eﬀects of college quality and aﬃrmative action. Journal of Development Economics, 135:97 – 116, 2018. ISSN 0304-3878. doi: https://doi.org/10. 1016/j.jdeveco.2018.06.017. URL http://www.sciencedirect.com/science/article/ pii/S0304387818304863.

Andrew Gelman and Guido Imbens. Why High-Order Polynomials Should Not Be Used in Regression Discontinuity Designs. Journal of Business & Economic Statistics, 0(0):1–10, 2018. doi: 10.1080/07350015.2017.1366909. URL https://doi.org/10.1080/07350015. 2017.1366909.

35 Peter Griﬃn. The Hindu Explains: The Anglo Indian member and the numbers game. 2018. URL https://www.thehindu.com/news/national/karnataka/ the-hindu-explains-the-anglo-indian-member-and-the-numbers-game/ article23937744.ece.

Saad Gulzar and Benjamin J. Pasquale. Politicians, Bureaucrats, and Development: Ev- idence from India. American Political Science Review, 111(1):162–183, 2017. doi: 10.1017/S0003055416000502.

Andrew Healy and Neil Malhotra. Retrospective Voting Reconsidered. Annual Review of Political Science, 16(1):285–306, 2013. doi: 10.1146/annurev-polisci-032211-212920. URL https://doi.org/10.1146/annurev-polisci-032211-212920.

Shigeo Hirano, James M. SnyderJr, and Michael M. Ting. Distributive Politics with Pri- maries. The Journal of Politics, 71(4):1467–1480, 2009. doi: 10.1017/S0022381609990247. URL https://doi.org/10.1017/S0022381609990247.

Alisha C. Holland. Forbearance. American Political Science Review, 110(2):232–246, 2016. doi: 10.1017/S0003055416000083. URL https://www.cambridge.org/ core/journals/american-political-science-review/article/forbearance/ 3BE0D1D5085F962CE168D8891519AC60.

IAY. Indira Awaas Yojana(IAY): Guidelines. 2013. URL https://pmayg.nic.in/netiay/ IAY%20revised%20guidelines%20july%202013.pdf.

Clément Imbert and John Papp. Labor Market Eﬀects of Social Programs: Evidence from In- dia’s Employment Guarantee. American Economic Journal: Applied Economics, 7(2):233– 63, April 2015. doi: 10.1257/app.20130401. URL http://www.aeaweb.org/articles? id=10.1257/app.20130401.

Lakshmi Iyer and Maya Reddy. Redrawing the Lines: Did Political Incumbents Inﬂuence Electoral Redistricting in the World’s Largest Democracy? Harvard Business School Working Paper, No. 14-051, 2013. URL https://www.hbs.edu/faculty/Pages/item. aspx?num=46005.

Francesca Refsum Jensenius. Was the Delimitation Commission Unfair to Muslims? Studies in Indian Politics, 1(2):213–229, 2013. doi: 10.1177/2321023013511625. URL https: //doi.org/10.1177/2321023013511625.

36 Devesh Kapur and Milan Vaishnav. Quid Pro Quo: Builders, Politicians, and Election Finance in India - Working Paper 276. 2013. URL https://www.cgdev.org/sites/ default/files/Kapur_Vaishnav_election_finance_India-FINAL-0313.pdf.

Stuti Khemani. Political cycles in a developing economy: eﬀect of elections in the Indian States. Journal of Development Economics, 73(1):125 – 154, 2004. ISSN 0304-3878. doi: https://doi.org/10.1016/j.jdeveco.2003.01.002. URL http://www.sciencedirect.com/ science/article/pii/S0304387803001329.

Eric Kramon and Daniel N. Posner. Who Beneﬁts from Distributive Poli- tics? How the Outcome One Studies Aﬀects the Answer One Gets. Perspec- tives on Politics, 11(2):461–474, 2013. doi: 10.1017/S1537592713001035. URL https://www.cambridge.org/core/journals/perspectives-on-politics/article/ who-benefits-from-distributive-politics-how-the-outcome-one-studies-affects-the-answer-one-gets/ 7920F6966363B66413250514119EF2BF.

Alexander Lee. Incumbency, Parties, and Legislatures: Theory and Evidence from India. 2016. URL http://www.rochester.edu/college/faculty/alexander_lee/ wp-content/uploads/2016/09/incumbency5.pdf.

Jonathan Lehne, Jacob N. Shapiro, and Oliver Vanden Eynde. Building connections: Polit- ical corruption and road construction in India. Journal of Development Economics, 131: 62 – 78, 2018. ISSN 0304-3878. doi: https://doi.org/10.1016/j.jdeveco.2017.10.009. URL http://www.sciencedirect.com/science/article/pii/S0304387817300883.

LGD. Local Government Directory. 2018. URL https://lgdirectory.gov.in/.

John A. List and Daniel M. Sturm. How Elections Matter: Theory and Evidence from Environmental Policy. The Quarterly Journal of Economics, 121(5):1249–1281, 2006. URL https://doi.org/10.1093/qje/121.4.1249.

Lokniti-CSDS. National Election Studies-Postpoll (April-May 2014). 2014a. URL http:// www.lokniti.org/media/PDF-upload/1536130357_23397100_download_report.pdf.

Lokniti-CSDS. State Election Studies. 2014b. URL http://www.lokniti.org/ state-election-studies.

Matt Lowe. Night Lights and ArcGIS: A Brief Guide, January 2014, 2014. URL http://www.thehindu.com/thread/article17272322.ece/BINARY/Matt%20Lowe’s% 20guide%20on%20Night%20Light%20data%20analysis.

37 Diego Maiorano. The Politics of the Mahatma Gandhi National Rural Employment Guar- antee Act in Andhra Pradesh. World Development, 58:95 – 105, 2014. ISSN 0305-750X. doi: https://doi.org/10.1016/j.worlddev.2014.01.006. URL http://www.sciencedirect. com/science/article/pii/S0305750X14000072.

Anandi Mani and Sharun Mukand. Democracy, visibility and public good provision. Journal of Development Economics, 83(2):506 – 529, 2007. ISSN 0304-3878. doi: https://doi.org/10.1016/j.jdeveco.2005.06.008. URL http://www.sciencedirect.com/ science/article/pii/S0304387806000642. Papers from a Symposium: The Social Di- mensions of Microeconomic Behaviour in Low-Income Communities.

Benjamin Marx. Elections as Incentives: Project Completion and Visibility in African Poli- tics. 2018. URL https://economics.mit.edu/files/14240.

MGNREGA. MGNREGA Public Data Portal. Ministry of Rural Development and Evi- dence for Policy Design, 2018. URL http://nregarep2.nic.in/netnrega/dynamic2/ dynamicreport_new4.aspx.

Marco Migueis. The Eﬀect of Political Alignment on Transfers to Portuguese Municipalities. Economics & Politics, 25(1):110–133, 2013. ISSN 1468-0343. doi: 10.1111/ecpo.12005. URL http://dx.doi.org/10.1111/ecpo.12005.

Brian Min. Elections and Electricity in the Developing World. 2015.

Sushanta Mishra. Asset Creation under MGNREGA: A study in three districts in the state of Madhya Pradesh. 01 2011.

Karthik Muralidharan, Paul Niehaus, and Sandip Sukhtankar. General Equilibrium Eﬀects of (Improving) Public Employment Programs: Experimental Evidence from India. National Bureau of Economic Research (Working Paper), (23838), 2017. URL http://www.nber. org/papers/w23838.

Peter C. O’Brien. Procedures for Comparing Samples with Multiple Endpoints. Biometrics, 40(4):1079–1087, 1984. ISSN 0006341X, 15410420. URL http://www.jstor.org/stable/ 2531158.

Nicola Persico, José C. R. Pueblita, and Dan Silverman. Factions and Political Competition. Journal of Political Economy, 119(2):242–288, 2011. doi: 10.1086/660298. URL https: //doi.org/10.1086/660298.

38 PMAY-G. Framework for Implementation: Pradhan Mantri Awaas Yojana - Gramin. 2016. URL https://pmayg.nic.in/netiay/Uploaded/English_Book_Final.pdf.

PMAY-G. Pradhan Mantri Awaas Yojana-Gramin. Ministry of Rural Development, 2018. URL https://rhreporting.nic.in/netiay/newreport.aspx.

PMGSY. Pradhan Mantri Gram Sadak Yojana. Online Managing, Monitoring and Account- ability System, 2018. URL http://omms.nic.in//#.

Postoﬃces. Mapping of Post oﬃces. Ministry of Communications, 2018. URL http:// postalgis.nic.in/view/.

George Psacharopoulos and Harry Anthony Patrinos. Returns to investment in education: a further update. Education Economics, 12(2):111–134, 2004. doi: 10.1080/ 0964529042000239140. URL https://doi.org/10.1080/0964529042000239140.

Ishita Rajani. Are Public Funds Used to Maintain Ruling Coalitions? Evidence from India. pages 269–275, 2016. URL https://ssrn.com/abstract=3054872.

Shamika Ravi and Monika Engler. Workfare as an Eﬀective Way to Fight Poverty: The Case of India’s NREGS. World Development, 67:57 – 71, 2015. ISSN 0305-750X. doi: https://doi.org/10.1016/j.worlddev.2014.09.029. URL http://www.sciencedirect.com/ science/article/pii/S0305750X14002940.

David J. Samuels. Pork Barreling Is Not Credit Claiming or Advertising: Campaign Fi- nance and the Sources of the Personal Vote in Brazil. The Journal of Politics, 64(3):845– 863, 2002. doi: 10.1111/0022-3816.00149. URL https://doi.org/10.1111/0022-3816. 00149.

Sourav Sarkar. Formation of New Administrative Units: The Case of India. Ph.D. Thesis, MIT (in progress), 2018.

School-GIS. School Location Mapping, Department of School Education and Literacy. Min- istry of Human Resource Development, 2018. URL http://schoolgis.nic.in/.

James M. Snyder. Election Goals and the Allocation of Campaign Resources. Econometrica, 57(3):637–660, 1989. ISSN 00129682, 14680262. URL http://www.jstor.org/stable/ 1911056.

Albert Solé-Ollé and Pilar Sorribas-Navarro. The effects of partisan alignment on the allocation of intergovernmental transfers. Differences-in-differences estimates for Spain. Journal

39 of Public Economics, 92(12):2302 – 2319, 2008. ISSN 0047-2727. doi: https://doi.org/10. 1016/j.jpubeco.2007.06.014. URL http://www.sciencedirect.com/science/article/ pii/S0047272707000990. New Directions in Fiscal Federalism.

David Stasavage. Democracy and Education Spending in Africa. American Journal of Political Science, 49(2):343–358, 2005. ISSN 1540-5907. doi: 10.1111/j.0092-5853.2005. 00127.x. URL http://dx.doi.org/10.1111/j.0092-5853.2005.00127.x.

Susan C. Stokes, Thad Dunning, Marcelo Nazareno, and Valeria Brusco. Brokers, Voters, and Clientelism: The Puzzle of Distributive Politics. Cambridge Studies in Comparative Politics. Cambridge University Press, 2013. doi: 10.1017/CBO9781107324909.

Sandip Sukhtankar and Manasa Patnam. Indian Assembly Constituency Maps. 2018. URL http://www.dartmouth.edu/~sandip/data.html.

TCPD. Lok Dhaba. Trivedi Centre for Political Data, 2018. URL http://lokdhaba. ashoka.edu.in/LokDhaba-Shiny/.

Yogesh Uppal. The disadvantaged incumbents: estimating incumbency eﬀects in Indian state legislatures. Public Choice, 138(1):9–27, Jan 2009.

Milan Vaishnav. When crime pays : money and muscle in Indian politics. Yale University Press, 2017.

Milan Vaishnav and Neelanjan Sircar. Core or Swing? The Role of Electoral Context in Shaping Pork Barrel. 01 2012. URL https://nsircar.files.wordpress.com/2013/02/ vaishnav_sircar_03-12-12.pdf.

Rajiv Verma, Saurabh Gupta, and Regina Birner. Can Grassroots Mobilization of the Poorest Reduce Corruption? A Tale of Governance Reforms and Struggle against Petty Corruption in Bihar, India. Development and Change, 48(2):339–363, 2017. doi: 10.1111/ dech.12295. URL https://onlinelibrary.wiley.com/doi/abs/10.1111/dech.12295.

VIIRS. Version 1 VIIRS Day/Night Band Nighttime Lights. National Centers for Envi- ronmental Information, National Oceanic and Atmospheric Administration, 2018. URL https://ngdc.noaa.gov/eog/viirs/download_dnb_composites.html.

George Washington and Alexander Hamilton. George Washington’s Farewell Address. 1796. URL https://en.wikisource.org/wiki/Washington%27s_Farewell_Address.

40 Rebecca Weitz-Shapiro and Matthew S. Winters. Can Citizens Discern? Information Cred- ibility, Political Sophistication, and the Punishment of Corruption in Brazil. The Journal of Politics, 79(1):60–74, 2017. doi: 10.1086/687287. URL https://doi.org/10.1086/ 687287.

Steven I. Wilkinson. Explaining changing patterns of party–voter linkages in India (in Pa- trons, Clients and Policies: Patterns of Democratic Accountability and Political Competi- tion), page 110–140. Cambridge University Press, 2007.

41 9 Figures and Tables

Figure 1: Density Plots for Manipulation Test .04 .03 .02 Density .01 0 -30 -20 -10 0 10 20 Vote Share Difference

point estimate 90% C.I.

Notes: This figure shows the results of density tests proposed in Cattaneo, Jansson, and Ma((2017)) to test for the null hypothesis of a difference in the density of the running variable on both sides of the cut-off. The figure plots the estimated density function along with 90% confidence interval of the running variable, which is the winning margin (Υip), in percentage, of the ruling party candidate. The value of the manipulation test statistic (which asymptotically has a standard normal distribution under the null hypothesis of continuity of density at Υip = 0) is 0.390 (p-value 0.696).

42 Table 1: Balance Table: Total Electors

(1) (2) (3) (4) (5) Ruling -0.80 -0.31 -0.42 -0.60 0.38 (1.59) (2.34) (1.23) (1.87) (1.69)

|Vote Share Diﬀ| 0.047 0.42 0.16 0.39 0.14 (0.13) (0.49) (0.15) (0.45) (0.14)

|Vote Share Diﬀ|*Ruling 0.14 -0.46 -0.067 0.058 -0.14 (0.19) (0.59) (0.14) (0.56) (0.17)

Observations 2715 1471 3639 3289 3206 No. of Clusters 23 23 23 23 23

state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes

Interval (-8.51,11.89) (-4.26,5.94) (-12.77,17.83) (-11.16,15.28) (-10.03,15.66) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1

Control Mean 1202.53 1202.34 1204.15 1203.36 1201.94 Control s.d. 74.27 72.2 71.68 73.45 74.93 All regressions include party ﬁxed eﬀects; *p < 0.1, ** p < 0.05, *** p < 0.01

Notes: This table reports the estimates from the regression discontinuity design that relate the political alignment of the constituency representative to Yip, which is 100 ∗ log(1 + x) where x is the total number of electors in the last election. The first row gives the estimate of the treatment variable which is a dummy that is 1 if the margin of victory of the ruling party candidate in the last election (Υip) was positive. The second row gives the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip < 0. The third row gives the difference in the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip > 0 from constituencies with Υip < 0. I report results from local linear (columns (1) − (3)) and local quadratic (column (4)) regressions with state-electoral period fixed effects for various bandwidths: optimal CCT (columns (1), (4)), 0.5 times the optimal CCT (column (2)), 1.5 times the optimal CCT (column 3). Column (5) reports results from local linear regression with parliamentary constituency-electoral period fixed effects for CCT bandwidth. The rows entitled Control Mean and Control s.d. reports the mean and standard deviation of the dependent variable for constituencies with Υip < 0 respectively. Standard errors, clustered at state (as in 2008) level are in parenthesis.

43 Table 2: Balance Table: Total Votes Polled

(1) (2) (3) (4) (5) Ruling 0.62 0.31 -0.27 0.11 -0.55 (1.19) (1.87) (1.08) (1.52) (1.32)

|Vote Share Diﬀ| -0.0012 0.13 -0.090 0.26 0.072 (0.089) (0.18) (0.075) (0.23) (0.098)

|Vote Share Diﬀ|*Ruling -0.15 -0.17 0.0046 -0.10 -0.12 (0.12) (0.22) (0.093) (0.26) (0.13)

Observations 4299 2802 4803 4366 4383 No. of Clusters 23 23 23 23 23 state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes

Interval (-18.98,24.19) (-9.49,12.09) (-28.48,36.28) (-23.64,23.79) (-16.99,27.8) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1

Control Mean 1172.82 1171.33 1172.91 1172.84 1172.52 Control s.d. 66.69 70.09 66.59 66.54 67.29 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01 Notes: This table reports the estimates from the regression discontinuity design that relate the political alignment of the constituency representative to Yip, which is 100 ∗ log(1 + x) where x is the total number of votes polled in the last election. The first row gives the estimate of the treatment variable which is a dummy that is 1 if the margin of victory of the ruling party candidate in the last election (Υip) was positive. The second row gives the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip < 0. The third row gives the difference in the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip > 0 from constituencies with Υip < 0. Standard errors, clustered at state (as in 2008) level are in parenthesis.

44 Table 3: Balance Table: Total Rural Population

(1) (2) (3) (4) (5) Ruling 6.59 -5.69 10.7 -13.5 2.65 (18.4) (34.4) (17.7) (31.3) (21.2)

|Vote Share Diﬀ| -0.97 -2.05 -0.33 -4.39 -0.37 (1.92) (6.68) (1.70) (11.7) (3.08)

|Vote Share Diﬀ|*Ruling 1.21 2.50 0.00041 12.8 0.78 (1.84) (9.35) (1.97) (12.3) (3.85)

Observations 2548 1382 3454 2859 2393 No. of Clusters 23 23 23 23 23 state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes

Interval (-8.79,10.55) (-4.4,5.27) (-13.19,15.82) (-10.37,12.14) (-8.97,10.08) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1

Control Mean 1143.46 1150.76 1141.22 1141.64 1139.74 Control s.d. 287.78 280.86 288.58 289.53 293.56 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01 Notes: This table reports the estimates from the regression discontinuity design that relate the political alignment of the constituency representative to Yip, which is 100 ∗ log(1 + x) where x is the total rural population of the constituency according to 2011 census. The first row gives the estimate of the treatment variable which is a dummy that is 1 if the margin of victory of the ruling party candidate in the last election (Υip) was positive. The second row gives the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip < 0. The third row gives the difference in the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip > 0 from constituencies with Υip < 0. Standard errors, clustered at state (as in 2008) level are in parenthesis.

45 Figure 2: Index for Growth of Administrative Headquarters .2 .1 0 Growth of Administrative Headquarters -.1

-10 -5 0 5 10 Vote Share Difference

Notes: This figure shows an RD plot for the partialled out (with respect to state-period Fixed Effects and political party Fixed Effects) index variable (Yip) constructed by taking a weighted average (as per Anderson((2008))) of growth of district headquarters and growth of subdistrict headquarters. The running variable is the winning margin (Υip), in percentage, of the ruling party candidate. An observation to the right of the threshold has a representative from the ruling party. I restrict the plot for observations of Υip within the coverage error rate optimal bandwidth (varying on both sides of Υip = 0, selected using a triangular kernel). Each dot (with 90% confidence bands) is the local average of Yip in equally spaced bins of 4 on both sides of Υip = 0. The solid lines are from a linear polynomial fit (of varying slopes on both sides of Υip = 0) of the underlying observations.

46 Table 4: Growth of Administrative Headquarters

(1) (2) (3) (4) (5) Ruling -0.14∗∗ -0.22∗∗ -0.084∗ -0.19∗∗ -0.100∗ (0.059) (0.080) (0.048) (0.075) (0.057)

|Vote Share Diﬀ| -0.022∗∗∗ -0.016 -0.013∗∗∗ -0.041∗ -0.016 (0.0077) (0.030) (0.0043) (0.020) (0.0095)

|Vote Share Diﬀ|*Ruling 0.021∗ 0.039 0.011∗ 0.065∗∗ 0.016 (0.010) (0.032) (0.0064) (0.026) (0.011)

Observations 2534 1388 3405 3074 2262 No. of Clusters 23 23 23 23 23

state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes

Interval (-11.5,9.36) (-5.75,4.68) (-17.25,14.04) (-14.74,12.08) (-11.32,8.28) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1

Control Mean 0 .05 0 .01 .01 Control s.d. .72 .830 .73 .75 .73 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01 Notes: This table reports the estimates from the regression discontinuity design that relate the political alignment of the constituency representative to the index of growth administrative headquarters (Yip). The first row gives the estimate of the treatment variable which is a dummy that is 1 if the margin of victory of the ruling party candidate in the last election (Υip) was positive. The second row gives the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip < 0. The third row gives the difference in the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip > 0 from constituencies with Υip < 0. I report results from local linear (columns (1) − (3)) and local quadratic (column (4)) regressions with state-electoral period fixed effects for various bandwidths: optimal CCT (columns (1), (4)), 0.5 times the optimal CCT (column (2)), 1.5 times the optimal CCT (column 3). Column (5) reports results from local linear regression with parliamentary constituency-electoral period fixed effects for CCT bandwidth. The rows entitled Control Mean and Control s.d. reports the mean and standard deviation of the dependent variable for constituencies with Υip < 0 respectively. Standard errors, clustered at state (as in 2008) level are in parenthesis.

47 Figure 3: Index for Growth of Educational Institutes .2 .1 0 -.1 Growth of Educational Institutes -.2 -10 -5 0 5 Vote Share Difference

Notes: This figure shows an RD plot for the partialled out (with respect to state-period Fixed Effects and political party Fixed Effects) index variable (Yip) constructed by taking a weighted average (as per Anderson((2008))) of growth of state-government-managed schools and growth of state-government-managed higher educational institutes. The running variable is the winning margin (Υip), in percentage, of the ruling party candidate. An observation to the right of the threshold has a representative from the ruling party. I restrict the plot to observations of Υip within the coverage error rate optimal bandwidth (varying on both sides of Υip = 0, selected using a triangular kernel). Each dot (with 90% confidence bands) is the local average of Yip in equally spaced bins of 4 on both sides of Υip = 0. The solid lines are from a linear polynomial fit (of varying slopes on both sides of Υip = 0) of the underlying observations.

48 Table 5: Growth of Educational Institutes

(1) (2) (3) (4) (5) Ruling -0.12∗ -0.20 -0.12∗ -0.26 -0.15 (0.070) (0.12) (0.068) (0.16) (0.12)

|Vote Share Diﬀ| -0.024 -0.039 -0.0049∗ -0.060 -0.024 (0.019) (0.052) (0.0025) (0.056) (0.028)

|Vote Share Diﬀ|*Ruling 0.023 0.072 0.032∗∗ 0.087 0.036 (0.014) (0.056) (0.014) (0.063) (0.036)

Observations 1757 962 2362 2439 1823 No. of Clusters 23 23 23 23 23

state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes

Interval (-10.79,4.73) (-5.39,2.36) (-16.18,7.09) (-13.75,8.1) (-7.72,7.3) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1

Control Mean .02 .04 .02 .01 .03 Control s.d. .87 1.05 .88 .84 .96 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01 Notes: This table reports the estimates from the regression discontinuity design that relate the political alignment of the constituency representative to the index of growth educational institutes (Yip). The first row gives the estimate of the treatment variable which is a dummy that is 1 if the margin of victory of the ruling party candidate in the last election (Υip) was positive. The second row gives the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip < 0. The third row gives the difference in the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip > 0 from constituencies with Υip < 0. Standard errors, clustered at state (as in 2008) level are in parenthesis.

49 Figure 4: Index for Growth of Long Term Investment Goods .2 .1 0 -.1 Growth of Long-term Investment Goods -.2 -10 -5 0 5 Vote Share Difference

Notes: This figure shows an RD plot for the partialled out (with respect to state-period Fixed Effects and political party Fixed Effects) index variable (Yip) constructed by taking a weighted average (as per Anderson((2008))) of index for growth of administrative headquarters and index for growth of educational institutes. The running variable is the winning margin (Υip), in percentage, of the ruling party candidate. An observation to the right of the threshold has a representative from the ruling party. I restrict the plot to observations of Υip within the coverage error rate optimal bandwidth (varying on both sides of Υip = 0, selected using a triangular kernel). Each dot (with 90% confidence bands) is the local average of Yip in equally spaced bins of 4 on both sides of Υip = 0. The solid lines are from a linear polynomial fit (of varying slopes on both sides of Υip = 0) of the underlying observations.

50 Table 6: Growth of Long Term Investment Goods

(1) (2) (3) (4) (5) Ruling -0.20∗∗∗ -0.32∗∗∗ -0.15∗∗ -0.22∗∗ -0.16∗ (0.055) (0.11) (0.056) (0.081) (0.078)

|Vote Share Diﬀ| -0.025∗ -0.065 -0.014∗∗ -0.034∗ -0.019 (0.013) (0.044) (0.0058) (0.019) (0.019)

|Vote Share Diﬀ|*Ruling 0.036∗∗ 0.12∗∗ 0.032∗∗∗ 0.065∗∗ 0.033 (0.014) (0.053) (0.011) (0.029) (0.022)

Observations 1732 948 2350 2681 1774 No. of Clusters 23 23 23 23 23 state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes

Interval (-10.02,4.76) (-5.01,2.38) (-15.03,7.14) (-18.11,8.54) (-8.06,6.64) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1

Control Mean .02 .04 .01 .01 .04 Control s.d. .75 .84 .72 .73 .79 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01 Notes: This table reports the estimates from the regression discontinuity design that relate the political alignment of the constituency representative to the index of long-term investment goods (Yip). The first row gives the estimate of the treatment variable which is a dummy that is 1 if the margin of victory of the ruling party candidate in the last election (Υip) was positive. The second row gives the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip < 0. The third row gives the difference in the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip > 0 from constituencies with Υip < 0. Standard errors, clustered at state (as in 2008) level are in parenthesis.

51 Figure 5: Index of Rural Employment Scheme .15 .1 .05 0 -.05 Index of Rural Employment Scheme -.1 -10 -5 0 5 10 Vote Share Difference

Notes: This figure shows an RD plot for the partialled out (with respect to state-period Fixed Effects and political party Fixed Effects) index variable (Yip) constructed by taking a weighted average (as per Anderson((2008))) of 100 ∗ log(1 + x) where x are various variates relating to rural employment scheme. The running variable is the winning margin (Υip), in percentage, of the ruling party candidate. An observation to the right of the threshold has a representative from the ruling party. I restrict the plot to observations of Υip within the coverage error rate optimal bandwidth (varying on both sides of Υip = 0, selected using a triangular kernel). Each dot (with 90% confidence bands) is the local average of Yip in equally spaced bins of 4 on both sides of Υip = 0. The solid lines are from a linear polynomial fit (of varying slopes on both sides of Υip = 0) of the underlying observations.

52 Table 7: Index of Rural Employment Scheme

(1) (2) (3) (4) (5) Ruling 0.027 -0.038 0.017 -0.017 -0.017 (0.032) (0.057) (0.031) (0.056) (0.041)

|Vote Share Diﬀ| -0.0064 -0.0068 -0.0072 -0.018 0.0093 (0.0084) (0.016) (0.0050) (0.024) (0.010)

|Vote Share Diﬀ|*Ruling 0.0034 0.0034 0.0038 0.027 -0.0090 (0.0075) (0.018) (0.0051) (0.024) (0.0089)

Observations 2391 1275 3257 3601 2292 No. of Clusters 23 23 23 23 23

state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes

Interval (-7.96,10) (-3.98,5) (-11.93,15) (-9.92,19.87) (-5.68,11.48) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1

Control Mean .15 .15 .14 .15 .14 Control s.d. .98 .98 .98 .98 .98 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01 Notes: This table reports the estimates from the regression discontinuity design that relate the political alignment of the constituency representative to the index of rural employment scheme (Yip). The first row gives the estimate of the treatment variable which is a dummy that is 1 if the margin of victory of the ruling party candidate in the last election (Υip) was positive. The second row gives the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip < 0. The third row gives the difference in the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip > 0 from constituencies with Υip < 0. Standard errors, clustered at state (as in 2008) level are in parenthesis.

53 Figure 6: Index of Housing Scheme .1 .05 0 Index of Rural Housing Schemes -.05 -10 -5 0 5 10 Vote Share Difference

Notes: This figure shows an RD plot for the partialled out (with respect to state-period Fixed Effects and political party Fixed Effects) index variable (Yip) constructed by taking a weighted average (as per Anderson((2008))) of 100 ∗ log(1 + x) where x are various variates relating to implementation of various housing schemes. The running variable is the winning margin (Υip), in percentage, of the ruling party candidate. An observation to the right of the threshold has a representative from the ruling party. I restrict the plot to observations of Υip within the coverage error rate optimal bandwidth (varying on both sides of Υip = 0, selected using a triangular kernel). Each dot (with 90% confidence bands) is the local average of Yip in equally spaced bins of 4 on both sides of Υip = 0. The solid lines are from a linear polynomial fit (of varying slopes on both sides of Υip = 0) of the underlying observations.

54 Table 8: Index of housing schemes

(1) (2) (3) (4) (5) Ruling 0.019 0.043 0.035 0.021 0.015 (0.032) (0.052) (0.022) (0.030) (0.028)

|Vote Share Diﬀ| 0.0020 0.010 -0.0024 0.0031 0.0019 (0.0039) (0.016) (0.0026) (0.0083) (0.0033)

|Vote Share Diﬀ|*Ruling 0.0043 -0.0060 0.0013 0.0063 -0.00033 (0.0047) (0.021) (0.0033) (0.0094) (0.0038)

Observations 1963 1057 2651 2819 2268 No. of Clusters 21 21 21 21 21 state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes

Interval (-9.45,8.53) (-4.73,4.26) (-14.18,12.79) (-14.41,14.41) (-11.87,10.7) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1

Control Mean .05 .04 .05 .05 .05 Control s.d. .950 .96 .950 .950 .950 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01 Notes: This table reports the estimates from the regression discontinuity design that relate the political alignment of the constituency representative to the index of housing schemes (Yip). The first row gives the estimate of the treatment variable which is a dummy that is 1 if the margin of victory of the ruling party candidate in the last election (Υip) was positive. The second row gives the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip < 0. The third row gives the difference in the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip > 0 from constituencies with Υip < 0. Standard errors, clustered at state (as in 2008) level are in parenthesis.

55 Figure 7: Index of Rural Road Construction Schemes .1 .05 0 -.05 Index of Rural Road Construction Scheme -.1 -10 -5 0 5 10 Vote Share Difference

Notes: This figure shows an RD plot for the partialled out (with respect to state-period Fixed Effects and political party Fixed Effects) index variable (Yip) constructed by taking a weighted average (as per Anderson((2008))) of 100 ∗ log(1 + x) where x are various variates relating to sanction and implementation of rural road construction scheme. The running variable is the winning margin (Υip), in percentage, of the ruling party candidate. An observation to the right of the threshold has a representative from the ruling party. I restrict the plot to observations of Υip within the coverage error rate optimal bandwidth (varying on both sides of Υip = 0, selected using a triangular kernel). Each dot (with 90% confidence bands) is the local average of Yip in equally spaced bins of 4 on both sides of Υip = 0. The solid lines are from a linear polynomial fit (of varying slopes on both sides of Υip = 0) of the underlying observations.

56 Table 9: Index Rural Road Construction Schemes

(1) (2) (3) (4) (5) Ruling 0.078 0.092 0.042 0.082 0.12∗∗ (0.069) (0.099) (0.063) (0.088) (0.054)

|Vote Share Diﬀ| 0.0091 0.017 0.0042 0.018 0.015 (0.011) (0.020) (0.0085) (0.019) (0.0090)

|Vote Share Diﬀ|*Ruling -0.014 -0.021 -0.0048 -0.024 -0.021∗∗ (0.011) (0.021) (0.0091) (0.023) (0.0093)

Observations 2362 1271 3233 3381 2545 No. of Clusters 23 23 23 23 23

state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes

Interval (-8.58,9.56) (-4.29,4.78) (-12.88,14.34) (-13.23,15.5) (-7.04,12.41) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1

Control Mean .01 -.01 -.01 0 .01 Control s.d. .92 .92 .92 .92 .92 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01 Notes: This table reports the estimates from the regression discontinuity design that relate the political alignment of the constituency representative to the index of rural road construction scheme (Yip). The first row gives the estimate of the treatment variable which is a dummy that is 1 if the margin of victory of the ruling party candidate in the last election (Υip) was positive. The second row gives the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip < 0. The third row gives the difference in the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip > 0 from constituencies with Υip < 0. Standard errors, clustered at state (as in 2008) level are in parenthesis.

57 Figure 8: Index of School Receipts and Expenditures .05 0 -.05 Index of School Receipts and Expenditures -.1 -10 -5 0 5 10 15 Vote Share Difference

Notes: This figure shows an RD plot for the partialled out (with respect to state-period Fixed Effects and political party Fixed Effects) index variable (Yip) constructed by taking a weighted average (as per Anderson((2008))) of 100 ∗ log(1 + x), where x are total school receipts and total school expenditures. The running variable is the winning margin (Υip), in percentage, of the ruling party candidate. An observation to the right of the threshold has a representative from the ruling party. I restrict the plot to observations of Υip within the coverage error rate optimal bandwidth (varying on both sides of Υip = 0, selected using a triangular kernel). Each dot (with 90% confidence bands) is the local average of Yip in equally spaced bins of 4 on both sides of Υip = 0. The solid lines are from a linear polynomial fit (of varying slopes on both sides of Υip = 0) of the underlying observations.

58 Table 10: Index of School Receipts and Expenditures

(1) (2) (3) (4) (5) Ruling 0.0049 0.0062 0.036 0.00040 0.012 (0.024) (0.033) (0.023) (0.028) (0.027)

|Vote Share Diﬀ| -0.0044 -0.011 -0.00047 -0.011 -0.00062 (0.0034) (0.0083) (0.0024) (0.010) (0.0028)

|Vote Share Diﬀ|*Ruling 0.0028 0.0054 -0.0021 0.0076 -0.0024 (0.0037) (0.0091) (0.0035) (0.0098) (0.0041)

Observations 2862 1571 3721 3294 2154 No. of Clusters 23 23 23 23 23

state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes

Interval (-10.12,13.65) (-5.06,6.83) (-15.18,20.48) (-10.61,17.47) (-10.59,8.79) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1

Control Mean -.05 -.04 -.05 -.06 -.06 Control s.d. 1.03 1.02 1.04 1.04 1.05 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01 Notes: This table reports the estimates from the regression discontinuity design that relate the political alignment of the constituency representative to the index of total school receipts and expenditures (Yip). The first row gives the estimate of the treatment variable which is a dummy that is 1 if the margin of victory of the ruling party candidate in the last election (Υip) was positive. The second row gives the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip < 0. The third row gives the difference in the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip > 0 from constituencies with Υip < 0. Standard errors, clustered at state (as in 2008) level are in parenthesis.

59 Figure 9: Index for Government Schemes .1 .05 0 -.05 Index of Government Schemes -.1 -10 -5 0 5 10 Vote Share Difference

Notes: This figure shows an RD plot for the partialled out (with respect to state-period Fixed Effects and political party Fixed Effects) index variable (Yip) constructed by taking a weighted average (as per Anderson((2008))) of index for rural employment scheme, rural housing schemes, rural road construction scheme and school receipts and expenditure. The running variable is the winning margin (Υip), in percentage, of the ruling party candidate. An observation to the right of the threshold has a representative from the ruling party. I restrict the plot to observations of Υip within the coverage error rate optimal bandwidth (varying on both sides of Υip = 0, selected using a triangular kernel). Each dot (with 90% confidence bands) is the local average of Yip in equally spaced bins of 4 on both sides of Υip = 0. The solid lines are from a linear polynomial fit (of varying slopes on both sides of Υip = 0) of the underlying observations.

60 Table 11: Index of Government Schemes

(1) (2) (3) (4) (5) Ruling 0.020 -0.011 0.039∗ 0.038 0.018 (0.030) (0.048) (0.021) (0.044) (0.041)

|Vote Share Diﬀ| -0.0041 -0.0011 -0.0038 0.015 0.0096 (0.0055) (0.014) (0.0029) (0.022) (0.0076)

|Vote Share Diﬀ|*Ruling 0.0084 0.0054 0.0018 -0.0089 -0.0083 (0.0056) (0.015) (0.0040) (0.020) (0.0079)

Observations 2335 1261 3203 3459 2166 No. of Clusters 23 23 23 23 23 state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes

Interval (-8.95,9.23) (-4.47,4.61) (-13.42,13.84) (-11.44,17.18) (-6.61,10.05) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1

Control Mean .1 .09 .09 .09 .1 Control s.d. .95 .96 .96 .95 .95 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01 Notes: This table reports the estimates from the regression discontinuity design that relate the political alignment of the constituency representative to the index of government schemes (Yip). The first row gives the estimate of the treatment variable which is a dummy that is 1 if the margin of victory of the ruling party candidate in the last election (Υip) was positive. The second row gives the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip < 0. The third row gives the difference in the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip > 0 from constituencies with Υip < 0. Standard errors, clustered at state (as in 2008) level are in parenthesis.

61 Figure 10: Index for Growth of Satellite Night Lights 10 5 0 Change in log luminosity -5

-15 -10 -5 0 5 10 Vote Share Difference

Notes: This figure shows an RD plot for the partialled out (with respect to state-period Fixed Effects and political party Fixed Effects) variable for change in luminosity: 1 + Lightst [log ] ∗ 100, where xt is the average satellite night lights at the end of the 1 + Lightst−1 electoral period. The running variable is the winning margin (Υip), in percentage, of the ruling party candidate. An observation to the right of the threshold has a representative from the ruling party. I restrict the plot to observations of Υip within the coverage error rate optimal bandwidth (varying on both sides of Υip = 0, selected using a triangular kernel). Each dot (with 90% confidence bands) is the local average of Yip in equally spaced bins of 4 on both sides of Υip = 0. The solid lines are from a linear polynomial fit (of varying slopes on both sides of Υip = 0) of the underlying observations.

62 Table 12: Growth of Satellite Night Lights

(1) (2) (3) (4) (5) Ruling 2.37 1.32 0.034 2.27 3.56 (3.45) (5.24) (4.04) (4.89) (2.89)

|Vote Share Diﬀ| -0.057 -0.26 0.017 -0.43 0.41 (0.43) (1.05) (0.28) (1.52) (0.36)

|Vote Share Diﬀ|*Ruling -0.59 0.72 -0.15 -0.37 -0.43 (0.74) (1.83) (0.58) (1.75) (0.56)

Observations 2578 1420 3434 3286 2421 No. of Clusters 23 23 23 23 23 state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes

Interval (-12.33,9.31) (-6.17,4.66) (-18.5,13.97) (-10.5,15.59) (-9.35,10.09) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1

Control Mean -24.72 -24.27 -23.69 -23.4 -23.5 Control s.d. 95.34 96.21 95.11 94.79 94.86 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01 Notes: This table reports the estimates from the regression discontinuity design that relate the political alignment of the constituency representative to the change in log luminosity (Yip). The first row gives the estimate of the treatment variable which is a dummy that is 1 if the margin of victory of the ruling party candidate in the last election (Υip) was positive. The second row gives the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip < 0. The third row gives the difference in the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip > 0 from constituencies with Υip < 0. Standard errors, clustered at state (as in 2008) level are in parenthesis.

63 Table 13: Index of Long Term Investment Goods: Robustness to Alternative Ruling Party Objectives

(1) (2) (3) (4) (5) Ruling -0.22∗∗∗ -0.36∗∗ -0.16∗∗ -0.27∗∗ -0.19∗ (0.071) (0.14) (0.065) (0.11) (0.090) |Vote Share Diff| -0.026∗ -0.072 -0.015∗∗ -0.049 -0.017 (0.015) (0.052) (0.0067) (0.031) (0.024) |Vote Share Diff|*Ruling 0.047∗∗ 0.14∗∗ 0.036∗∗∗ 0.11∗∗ 0.039 (0.018) (0.062) (0.012) (0.040) (0.025) Observations 1371 742 1849 2043 1364 No. of Clusters 20 20 20 20 20 state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes Interval (-10.02,4.68) (-5.01,2.34) (-15.03,7.02) (-17.3,7.92) (-7.97,6.15) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1 Control Mean .04 .07 .03 .03 .07 Control s.d. .82 .92 .78 .79 .87 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01 Notes: This table reports the estimates from the regression discontinuity design that relate the political alignment of the constituency representative to the index of long-term investment goods (Yip). I exclude data from high majority electoral periods in states with incumbency disadvantage. The first row gives the estimate of the treatment variable which is a dummy that is 1 if the margin of victory of the ruling party candidate in the last election (Υip) was positive. The second row gives the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip < 0. The third row gives the difference in the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip > 0 from constituencies with Υip < 0. Standard errors, clustered at state (as in 2008) level are in parenthesis.

A Appendix: Tables and Figures

Table A.1: Description of Electoral Periods since May 2008

State Begin Month/Year End Month/Year Total Seats Majority Party Seat Share Jammu and Kashmir Dec,2008 Dec,2014 87 coalition Jammu and Kashmir Dec,2014 87 coalition Himachal Pradesh Dec,2012 Dec,2017 68 INC 52.94 Himachal Pradesh Dec,2017 68 BJP 64.71% Punjab March,2012 May,2017 117 coalition

64 Punjab March,2017 117 INC 65.81% Uttarakhand March,2012 May,2017 70 coalition Uttarakhand March,2017 70 BJP 81.42% Haryana Oct,2009 Oct, 2014 90 coalition Haryana Oct,2014 288 BJP 70% Delhi Dec,2008 Dec,2013 70 INC 61.4% Delhi Dec,2013 Feb,2015 70 President’s Rule Delhi Feb,2015 70 AAP 95.71% Rajasthan Dec,2008 Dec,2013 200 coalition Rajasthan Dec,2013 200 BJP 81.5% Uttar Pradesh March,2012 May,2017 403 SP 55.58% Uttar Pradesh March,2017 403 BJP 77.42% Bihar Nov,2010 Nov, 2015 243 unstable coalition Bihar Nov,2015 240 unstable coalition Sikkim May,2009 May, 2014 32 SDF 100% Sikkim May,2014 32 SDF 68.75% Arunachal Pradesh September,2009 May, 2014 60 INC 70% Arunachal Pradesh May,2014 60 INC (defection later) 70% Nagaland Feb,2013 March,2018 60 NPF 63.33% Nagaland March,2018 59 coalition Manipur March,2012 May,2017 60 INC 70% Manipur March,2017 60 coalition Mizoram Dec,2008 Dec,2013 40 INC 80% Mizoram Dec,2013 40 INC 85% Tripura Feb,2013 March,2018 60 CPI(M) 81.66% Tripura March,2018 59 BJP 59.32% Meghalaya Feb,2013 March,2018 60 coalition Meghalaya March,2018 59 coalition Assam May,2011 May,2016 126 INC 61.90% Assam May,2016 126 coalition West Bengal May,2011 May,2016 294 AITC 62.59% West Bengal May,2016 294 AITC 71.79% Jharkhand Dec,2009 Dec, 2014 81 coalition Jharkhand Dec,2014 81 coalition Odisha May,2009 May, 2014 147 BJD 70% Odisha May,2014 147 BJD 79.59% Chhattisgarh Dec,2008 Dec,2013 90 BJP 55.56% Chhattisgarh Dec,2013 90 BJP 54.44% Madhya Pradesh Dec,2008 Dec,2013 224 BJP 62.17% Madhya Pradesh Dec,2013 230 BJP 71.73% Gujarat Dec,2012 Dec,2017 182 BJP 63.19 Gujarat Dec,2017 182 BJP 54.4% Maharashtra Oct,2009 Oct, 2014 288 coalition

65 Maharashtra Oct,2014 288 coalition Andhra Pradesh May,2009 May, 2014 294 INC 53.06% Andhra Pradesh May,2014 175 TDP 79.59% Telangana May,2014 119 TRS 52.94% Karnataka May,2008 May, 2013 224 coalition Karnataka May,2013 May,2018 224 INC 54.46% Goa March,2012 May,2017 40 BJP 52.5% Goa March,2017 40 coalition Kerala May,2011 May,2016 140 coalition Kerala May,2016 140 coalition Tamil Nadu May,2011 May,2016 234 AIADMK 64.10% Tamil Nadu May,2016 232 AIADMK 57.76% Puducherry May,2011 May,2016 30 coalition Puducherry May,2016 30 coalition

Notes: The above table details the electoral periods since May, 2008 in India. Column 1 gives the state, column 2 gives the time of the election, column 3 gives the time of the next election, column 4 gives the total number of constituencies in the state, column 5 gives the ruling party (mentions details of coalition government/president’s rule if no party secured a majority), column 6 gives the percentage of seats secured by the ruling party if a single party secures a majority.

Table A.2: Duration (Months) of Governments: May,2008-March,2018

State Total Seats Non-Coalition Coalition Population (millions) Uttar Pradesh 403 72 - 199.81 West Bengal 294 70 - 91.34 Andhra Pradesh/Telangana 294 106 - 84.67 Maharashtra 287 - 91 112.837 Bihar 243 - 89 103.80 Tamil Nadu 234 70 - 72.19 Madhya Pradesh 224 112 - 72.60 Karnataka 224 58 60 61.13 Rajasthan 200 51 60 68.62 Gujarat 182 64 - 60.38 Odisha 147 106 - 41.95 Kerala 140 - 70 33.39 Assam 126 60 22 31.17 Punjab 117 12 60 27.70 Chhattisgarh 90 112 - 25.54 Haryana 90 41 60 25.35 Jharkhand 81 - 89 32.97 Jammu & Kashmir 87 - 112 12.55

66 Delhi 70 96 15 16.79 Uttarakhand 70 12 60 10.12 Himachal Pradesh 68 63 - 6.86 Tripura 60 61 - 3.67 Meghalaya 60 - 61 2.96 Manipur 60 72 12 2.72 Nagaland 60 60 - 1.98 Arunachal Pradesh 60 60 - 1.38 Goa 40 60 12 1.46 Mizoram 40 111 - 1.09 Sikkim 32 106 - 0.61 Pondicherry 30 - 70 1.2

Notes: The above table gives the details for each state that I include in the analysis. Column 1 gives the state, column 2 gives the total constituencies in the state, column 3 gives the number of months of non-coalition governments, column 4 gives the number of months of coalition governments, column 5 gives the total population of the state as of 2011 census.

67 Table A.3: Growth of District Headquarters

(1) (2) (3) (4) (5) Ruling -0.99 -2.47 -0.95 -2.38 -0.59 (1.22) (1.70) (1.12) (1.73) (1.62) |Vote Share Diff| -0.25∗∗ -0.33 -0.24∗∗ -0.45 -0.22∗ (0.11) (0.28) (0.084) (0.41) (0.12) |Vote Share Diff|*Ruling 0.19 0.60 0.17∗ 0.83 0.26 (0.14) (0.43) (0.091) (0.56) (0.25) Observations 3044 1735 3922 3137 2683 No. of Clusters 23 23 23 23 23 state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes Interval (-14.04,12.09) (-7.02,6.05) (-21.06,18.14) (-15.57,12.37) (-14.32,9.96) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1 Control Mean 1.45 1.99 1.24 1.38 1.43 Control s.d. 11.87 13.86 10.99 11.6 11.81 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01

Notes: This table reports the estimates from the regression discontinuity design that relate the political alignment of the constituency representative to the growth in district headquarters (Yip). The ﬁrst row gives the estimate of the treatment variable which is a dummy that is 1 if the margin of victory of the ruling party candidate in the last election (Υip) was positive. The second row gives the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip < 0. The third row gives the diﬀerence in the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip > 0 from constituencies with Υip < 0. Standard errors, clustered at state (as in 2008) level are in parenthesis.

Table A.4: Growth of Subdistrict Headquarters

(1) (2) (3) (4) (5) Ruling -2.48 -3.30 -0.99 -2.78∗ -3.19∗ (1.44) (1.95) (1.92) (1.53) (1.79) |Vote Share Diff| -0.21∗ -0.29 -0.087 -0.70 -0.21 (0.11) (0.51) (0.10) (0.58) (0.31) |Vote Share Diff|*Ruling 0.31 0.88 0.077 1.08 0.33 (0.22) (0.64) (0.21) (0.64) (0.41) Observations 2416 1316 3257 2998 2260 No. of Clusters 23 23 23 23 23 state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes Interval (-11.32,8.65) (-5.66,4.32) (-16.97,12.97) (-13.64,11.85) (-9.23,9.06) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1 Control Mean 2.63 3.14 2.86 2.74 2.77 Control s.d. 17.07 18.14 18.48 18.47 17.91 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01

Notes: This table reports the estimates from the regression discontinuity design that relate the political alignment of the constituency representative to the growth in subdistrict headquarters (Yip). The ﬁrst row gives the estimate of the treatment variable which is a dummy that is 1 if the margin of victory of the ruling party candidate in the last election (Υip) was positive. The second row gives the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip < 0. The third row gives the diﬀerence in the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip > 0 from constituencies with Υip < 0.

68 Table A.5: Growth of Schools

(1) (2) (3) (4) (5) Ruling -8.10 -10.8 -5.07 -14.2 -9.62 (6.74) (9.41) (4.02) (12.3) (8.78) |Vote Share Diff| -1.67 -4.61 -0.63 -5.00 -2.22 (1.39) (4.02) (0.51) (4.69) (2.20) |Vote Share Diff|*Ruling 2.13 4.07 1.03 5.79 2.72 (1.72) (3.61) (0.69) (4.98) (2.51) Observations 1734 927 2379 2215 2181 No. of Clusters 23 23 23 23 23 state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes Interval (-10.17,5.46) (-5.09,2.73) (-15.26,8.19) (-12.62,7.82) (-7.25,10.45) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1 Control Mean 8.85 10.11 8.24 8.79 9.75 Control s.d. 57.2 67.56 53.03 55.58 63.24 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01

Notes: This table reports the estimates from the regression discontinuity design that relate the political alignment of the constituency representative to the growth in state-government-managed schools (Yip). The ﬁrst row gives the estimate of the treatment variable which is a dummy that is 1 if the margin of victory of the ruling party candidate in the last election (Υip) was positive. The second row gives the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip < 0. The third row gives the diﬀerence in the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip > 0 from constituencies with Υip < 0.

Table A.6: Growth of Higher Educational Institutes

(1) (2) (3) (4) (5) Ruling -0.21 -0.22 -0.31 -0.30 -0.22 (0.14) (0.14) (0.30) (0.26) (0.20) |Vote Share Diff| -0.019∗∗ 0.019 -0.0076 -0.022 -0.020 (0.0085) (0.020) (0.0047) (0.023) (0.013) |Vote Share Diff|*Ruling 0.059∗ 0.066 0.081 0.068 0.060 (0.030) (0.087) (0.066) (0.077) (0.053) Observations 1886 1052 2488 2451 1780 No. of Clusters 23 23 23 23 23 state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes Interval (-12.43,4.99) (-6.21,2.49) (-18.64,7.48) (-14.84,7.89) (-10.61,5.61) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1 Control Mean .11 .17 .12 .12 .12 Control s.d. 1.26 1.56 1.29 1.34 1.32 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01

Notes: This table reports the estimates from the regression discontinuity design that relate the political alignment of the constituency representative to the growth in state-government-managed higher educational institutes (Yip). The ﬁrst row gives the estimate of the treatment variable which is a dummy that is 1 if the margin of victory of the ruling party candidate in the last election (Υip) was positive. The second row gives the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip < 0. The third row gives the diﬀerence in the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip > 0 from constituencies with Υip < 0.

69 Table A.7: Total Number of Works on Assets Under Rural Employment Scheme

(1) (2) (3) (4) (5) Ruling 4.52 -11.8 4.14 -5.34 -4.45 (10.6) (19.1) (10.6) (19.1) (14.9) |Vote Share Diff| -1.52 -1.89 -2.58 -6.25 3.24 (2.51) (5.25) (1.77) (7.91) (2.79) |Vote Share Diff|*Ruling 1.34 0.37 1.64 9.34 -2.92 (2.44) (6.06) (1.72) (7.89) (2.75) Observations 2403 1288 3276 3614 2279 No. of Clusters 23 23 23 23 23 state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes Interval (-7.87,10.16) (-3.93,5.08) (-11.8,15.24) (-9.93,20.04) (-5.520,11.49) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1 Control Mean 624.31 627.14 619.05 623.72 617.87 Control s.d. 345.07 345.28 346.1 344.19 348.55 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01

Table A.8: Total Number of Assets Completed Under Rural Employment Scheme

(1) (2) (3) (4) (5) Ruling 8.29 -16.7 7.30 -5.42 -3.59 (10.6) (17.4) (10.4) (18.0) (13.5) |Vote Share Diff| -1.98 -1.74 -1.93 -5.94 2.49 (2.67) (5.05) (1.42) (8.13) (3.76) |Vote Share Diff|*Ruling 1.01 2.01 0.51 8.68 -2.41 (2.38) (5.54) (1.61) (7.71) (3.22) Observations 2384 1277 3250 3582 2318 No. of Clusters 23 23 23 23 23 state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes Interval (-8.07,9.92) (-4.03,4.96) (-12.1,14.88) (-9.93,19.67) (-5.98,11.49) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1 Control Mean 664.54 666.93 659.69 662.04 663.03 Control s.d. 310.22 307.47 311.53 309.76 308.68 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01

70 Table A.9: Housing schemes: total number of houses built

(1) (2) (3) (4) (5) Ruling 11.2 27.1 18.1 19.4 19.7 (13.2) (19.1) (10.6) (12.4) (14.3) |Vote Share Diff| 0.27 4.75 -0.25 2.63 4.75∗∗ (1.80) (5.36) (0.94) (3.35) (2.14) |Vote Share Diff|*Ruling 1.06 -2.89 0.31 -0.56 -5.40∗∗ (1.82) (7.46) (1.15) (3.08) (2.16) Observations 1620 871 2218 2297 1499 No. of Clusters 21 21 21 21 19 state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes Interval (-9.33,7.71) (-4.66,3.85) (-13.99,11.56) (-15.02,12.18) (-6.62,8.92) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1 Control Mean 567.56 566.19 567.04 567.31 568.79 Control s.d. 297.67 299.23 298.19 297.97 297.09 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01

Table A.10: Housing schemes: total funds utilized

(1) (2) (3) (4) (5) Ruling 20.7∗ 27.9∗ 23.8∗∗ 17.6∗ 12.8 (11.8) (13.4) (9.29) (8.93) (13.0) |Vote Share Diff| 1.32 5.92 0.40 0.67 1.53 (1.41) (4.84) (0.74) (1.88) (1.20) |Vote Share Diff|*Ruling -1.05 -9.00 -1.07 1.00 -1.26 (1.72) (5.77) (0.89) (5.41) (1.23) Observations 1946 1073 2598 2207 1924 No. of Clusters 21 21 21 21 20 state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes Interval (-8.960,10.3) (-4.48,5.15) (-13.44,15.45) (-16.41,10.07) (-9,10.82) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1 Control Mean 629.82 622.14 628.63 630.07 629.92 Control s.d. 305.78 309.29 307.65 306.61 305.4 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01

71 Table A.11: Rural Road Construction (PMGSY): Total Cost of Project

(1) (2) (3) (4) (5) Ruling 46.2 14.4 19.7 33.9 59.1 (60.0) (83.9) (45.1) (67.7) (55.3) |Vote Share Diff| 4.76 -2.06 2.50 6.84 3.02 (7.48) (12.4) (4.77) (13.6) (6.56) |Vote Share Diff|*Ruling -7.01 -2.56 -2.14 -8.33 -6.33 (8.43) (17.7) (4.95) (17.4) (7.96) Observations 2352 1279 3228 3369 2464 No. of Clusters 23 23 23 23 23 state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes Interval (-9.19,9.25) (-4.59,4.63) (-13.78,13.88) (-14.01,15.1) (-7.92,11.2) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1 Control Mean 812.05 786.47 793.14 795 802.93 Control s.d. 686.7 693.4 687.96 687.67 688.32 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01

Notes: This table reports the estimates from the regression discontinuity design that relate the political alignment of the constituency representative to Yip, which is 100 ∗ log(1 + x), where x is the population-weighted (as I have discussed in Section 2.4) total cost on projects under the rural road construction scheme (PMGSY). The ﬁrst row gives the estimate of the treatment variable which is a dummy that is 1 if the margin of victory of the ruling party candidate in the last election (Υip) was positive. The second row gives the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip < 0. The third row gives the diﬀerence in the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip > 0 from constituencies with Υip < 0.

Table A.12: Rural Road Construction (PMGSY): Total Expenditure of Project

(1) (2) (3) (4) (5) Ruling 65.5 97.2 37.3 63.1 84.3∗∗∗ (41.1) (60.3) (39.1) (48.9) (27.0) |Vote Share Diff| 6.20 16.9 5.49 15.2 13.4∗∗ (9.55) (19.8) (5.18) (12.3) (6.26) |Vote Share Diff|*Ruling -12.4 -20.1 -6.56 -17.5 -17.8∗∗ (9.36) (15.3) (5.85) (13.4) (7.08) Observations 2374 1268 3241 3846 2565 No. of Clusters 23 23 23 23 23 state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes Interval (-7.83,9.99) (-3.91,5) (-11.74,14.99) (-12.78,20.98) (-7.24,12.38) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1 Control Mean 482.37 482.11 485.69 483.29 479.12 Control s.d. 650.06 650.23 650.03 649.24 649.35 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01

Notes: This table reports the estimates from the regression discontinuity design that relate the political alignment of the constituency representative to Yip, which is 100 ∗ log(1 + x), where x is the population-weighted (as I have discussed in Section 2.4) total expenditure on projects till date under the rural road construction scheme (PMGSY). The ﬁrst row gives the estimate of the treatment variable which is a dummy that is 1 if the margin of victory of the ruling party candidate in the last election (Υip) was positive. The second row gives the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip < 0. The third row gives the diﬀerence in the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip > 0 from constituencies with Υip < 0.

72 Table A.13: Rural Road Construction (PMGSY): Total Length of Roads

(1) (2) (3) (4) (5) Ruling 32.1 14.9 14.5 28.4 37.9 (40.9) (52.8) (30.7) (47.3) (36.8) |Vote Share Diff| 3.16 3.00 2.32 3.89 0.53 (5.40) (11.6) (3.40) (10.7) (5.01) |Vote Share Diff|*Ruling -4.08 -3.21 -1.91 -7.53 -2.45 (5.93) (10.3) (3.41) (12.4) (5.57) Observations 2386 1294 3262 3348 2523 No. of Clusters 23 23 23 23 23 state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes Interval (-9.15,9.470) (-4.57,4.73) (-13.72,14.2) (-13.63,15.04) (-8.02,11.51) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1 Control Mean 584.43 569.73 571.68 574.09 579.97 Control s.d. 505.05 509.9 505.71 505.41 506.29 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01

Notes: This table reports the estimates from the regression discontinuity design that relate the political alignment of the constituency representative to Yip, which is 100 ∗ log(1 + x), where x is the population-weighted (as I have discussed in Section 2.4) total length of roads date under the rural road construction scheme (PMGSY). The ﬁrst row gives the estimate of the treatment variable which is a dummy that is 1 if the margin of victory of the ruling party candidate in the last election (Υip) was positive. The second row gives the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip < 0. The third row gives the diﬀerence in the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip > 0 from constituencies with Υip < 0.

Table A.14: Rural Road Construction (PMGSY): Total Length of Roads Completed

(1) (2) (3) (4) (5) Ruling 32.5 25.1 16.4 25.5 55.9∗∗ (32.5) (38.3) (29.2) (37.4) (26.1) |Vote Share Diff| 2.03 5.81 2.60 2.70 5.62 (5.77) (9.82) (4.23) (9.68) (5.23) |Vote Share Diff|*Ruling -4.04 -7.46 -2.36 -6.66 -8.37∗ (5.38) (9.38) (4.20) (9.65) (4.80) Observations 2317 1244 3189 3217 2692 No. of Clusters 23 23 23 23 23 state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes Interval (-8.540,9.27) (-4.270,4.63) (-12.81,13.9) (-12.17,14.52) (-7.34,13.26) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1 Control Mean 523.22 510.99 514.6 516.14 515.72 Control s.d. 500.79 502.1 498.71 498.99 499.1 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01

Notes: This table reports the estimates from the regression discontinuity design that relate the political alignment of the constituency representative to Yip, which is 100 ∗ log(1 + x), where x is the population-weighted (as I have discussed in Section 2.4) total length of roads completed till date under the rural road construction scheme (PMGSY). The ﬁrst row gives the estimate of the treatment variable which is a dummy that is 1 if the margin of victory of the ruling party candidate in the last election (Υip) was positive. The second row gives the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip < 0. The third row gives the diﬀerence in the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip > 0 from constituencies with Υip < 0.

73 Table A.15: Total School Receipts

(1) (2) (3) (4) (5) Ruling 8.84 -1.53 10.5 -1.59 3.58 (10.8) (16.6) (10.5) (13.5) (14.1) |Vote Share Diff| -1.47 -6.30 -0.81 -7.10 -0.21 (1.17) (4.81) (1.09) (4.51) (1.17) |Vote Share Diff|*Ruling 0.19 3.63 -0.29 5.17 -0.53 (1.60) (5.66) (1.58) (4.59) (2.34) Observations 2537 1379 3390 3175 2131 No. of Clusters 23 23 23 23 23 state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes Interval (-8.9,12.05) (-4.45,6.02) (-13.34,18.07) (-10.05,17.34) (-9.47,9.44) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1 Control Mean 1425.05 1425.61 1422.18 1421.77 1425.84 Control s.d. 459.5 464.66 470.03 464.61 459.56 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01

Table A.16: Total School Expenditures

(1) (2) (3) (4) (5) Ruling 0.048 -0.089 0.069 -0.053 0.029 (0.10) (0.17) (0.097) (0.14) (0.14) |Vote Share Diff| -0.023 -0.073 -0.012 -0.073 0.00067 (0.013) (0.052) (0.012) (0.046) (0.014) |Vote Share Diff|*Ruling 0.017 0.059 0.0027 0.066 -0.0071 (0.014) (0.061) (0.015) (0.051) (0.023) Observations 2457 1328 3304 3061 2032 No. of Clusters 23 23 23 23 23 state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes Interval (-8.52,11.53) (-4.26,5.77) (-12.77,17.3) (-10,16.18) (-9.14,8.91) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1 Control Mean 14.43 14.38 14.34 14.37 14.41 Control s.d. 4.49 4.63 4.67 4.59 4.52 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01

74 Table A.17: Growth of Educational Institutes (including all educational institutes)

(1) (2) (3) (4) (5) Ruling -0.10∗ -0.094∗ -0.13∗ -0.16∗ -0.12∗ (0.051) (0.048) (0.071) (0.082) (0.064) |Vote Share Diff| -0.016∗ -0.015 -0.0086 -0.044∗ -0.029∗ (0.0092) (0.010) (0.0067) (0.025) (0.015) |Vote Share Diff|*Ruling 0.024∗∗ 0.029 0.027∗∗ 0.048∗ 0.028 (0.011) (0.018) (0.013) (0.028) (0.016) Observations 2153 1194 2864 2611 1656 No. of Clusters 23 23 23 23 23 state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes Interval (-12.78,6.57) (-6.39,3.28) (-19.18,9.85) (-16.16,8.69) (-7.83,6.07) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1 Control Mean .02 .05 .02 .02 .04 Control s.d. .81 .95 .77 .78 .88 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01

Notes: This table reports the estimates from the regression discontinuity design that relate the political alignment of the constituency representative to the index of growth educational institutes (Yip) considering all schools and higher educational institutes. The ﬁrst row gives the estimate of the treatment variable which is a dummy that is 1 if the margin of victory of the ruling party candidate in the last election (Υip) was positive. The second row gives the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip < 0. The third row gives the diﬀerence in the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip > 0 from constituencies with Υip < 0.

Table A.18: Growth of Long Term Investment Goods (including all educational institutes)

(1) (2) (3) (4) (5) Ruling -0.17∗∗∗ -0.24∗∗∗ -0.14∗∗∗ -0.18∗∗∗ -0.16∗∗∗ (0.054) (0.076) (0.044) (0.056) (0.053) |Vote Share Diff| -0.025∗∗∗ -0.040 -0.016∗∗∗ -0.036∗∗ -0.019 (0.0089) (0.034) (0.0051) (0.014) (0.013) |Vote Share Diff|*Ruling 0.037∗∗∗ 0.056 0.027∗∗∗ 0.051∗∗∗ 0.027∗ (0.012) (0.034) (0.0074) (0.014) (0.013) Observations 2084 1115 2839 3036 1816 No. of Clusters 23 23 23 23 23 state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes Interval (-10.55,6.75) (-5.27,3.37) (-15.82,10.12) (-20.42,10.62) (-8,6.92) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1 Control Mean .03 .07 .01 .01 .04 Control s.d. .74 .83 .72 .71 .79 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01

Notes: This table reports the estimates from the regression discontinuity design that relate the political alignment of the constituency representative to the index of growth of long-term investments (Yip) considering all schools and higher educational institutes. The ﬁrst row gives the estimate of the treatment variable which is a dummy that is 1 if the margin of victory of the ruling party candidate in the last election (Υip) was positive. The second row gives the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip < 0. The third row gives the diﬀerence in the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip > 0 from constituencies with Υip < 0.

75 Table A.19: Competitiveness of Next Election

(1) (2) (3) (4) (5) Ruling -0.80 -1.90 -0.065 0.099 -0.97 (1.63) (1.53) (1.55) (2.16) (1.53) |Vote Share Diff| -0.047 0.62 0.15 -0.23 -0.13 (0.34) (0.37) (0.12) (0.62) (0.33) |Vote Share Diff|*Ruling 0.0050 -0.37 -0.13 0.074 0.011 (0.43) (0.46) (0.20) (0.68) (0.42) Observations 1091 590 1465 1403 995 No. of Clusters 17 16 18 18 17 state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes Interval (-6.81,11.04) (-3.41,5.52) (-10.22,16.55) (-10.63,14.75) (-7.31,10.19) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1 Control Mean 12.08 11.61 12.32 12.51 11.8 Control s.d. 12.25 9.34 12.38 12.37 12.12 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01

Notes: This table reports the estimates from the regression discontinuity design that relate the political alignment of the constituency representative to Yip, which is |x| ∗ 100, where x is the vote share difference of the ruling party in the next assembly election. The first row gives the estimate of the treatment variable which is a dummy that is 1 if the margin of victory of the ruling party candidate in the last election (Υip) was positive. The second row gives the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip < 0. The third row gives the difference in the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip > 0 from constituencies with Υip < 0.

Table A.20: Reelection of Ruling Party Candidate

(1) (2) (3) (4) (5) Ruling -15.9∗∗ -15.1∗∗ -14.6∗∗ -14.9∗ -15.1∗∗ (6.38) (6.80) (5.47) (7.83) (6.36) |Vote Share Diff| -0.84 -0.83 -0.21 -0.99 -0.73 (0.66) (1.92) (0.31) (1.84) (0.72) |Vote Share Diff|*Ruling 2.06∗∗ 2.54 1.61∗∗∗ 1.90 1.97∗∗ (0.76) (1.92) (0.47) (1.88) (0.92) Observations 1436 816 1827 1813 1423 No. of Clusters 18 17 18 18 18 state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes Interval (-10.46,14.45) (-5.23,7.22) (-15.69,21.67) (-14.12,21.99) (-10.7,14.74) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1 Control Mean 35.76 35.99 35.07 35.09 35.73 Control s.d. 47.98 48.08 47.76 47.77 47.97 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01

Notes: This table reports the estimates from the regression discontinuity design that relate the political alignment of the constituency representative to Yip, which is |x| ∗ 100, where x is the indicator variable of whether the ruling party’s candidate gets elected in the next assembly election. The ﬁrst row gives the estimate of the treatment variable which is a dummy that is 1 if the margin of victory of the ruling party candidate in the last election (Υip) was positive. The second row gives the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip < 0. The third row gives the diﬀerence in the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip > 0 from constituencies with Υip < 0.

76 Table A.21: Growth of Long Term Investment Goods (including coalition governments)

(1) (2) (3) (4) (5) Ruling -0.20∗∗∗ -0.17∗ -0.100∗∗ -0.22∗∗ -0.16∗ (0.064) (0.086) (0.039) (0.085) (0.082) |Vote Share Diff| -0.023∗∗ -0.025 -0.0096∗∗ -0.036 -0.020 (0.0097) (0.032) (0.0036) (0.021) (0.015) |Vote Share Diff|*Ruling 0.054∗∗∗ 0.068 0.021∗∗∗ 0.078∗∗ 0.038∗ (0.016) (0.041) (0.0076) (0.034) (0.022) Observations 2686 1449 3631 3817 2599 No. of Clusters 29 29 29 29 29 state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes Interval (-11.19,5.31) (-5.60,2.65) (-16.79,7.96) (-16.15,9.08) (-9.32,6.51) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1 Control Mean .01 .06 .02 .01 .02 Control s.d. .75 .91 .77 .74 .79 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01

Notes: This table reports the estimates from the regression discontinuity design that relate the political alignment of the constituency representative to the index of long-term investment goods (Yip). I include data for coalition governments in this table and consider a ruling party to be one which is included in the ruling coalition. The ﬁrst row gives the estimate of the treatment variable which is a dummy that is 1 if the margin of victory of the ruling party candidate in the last election (Υip) was positive. The second row gives the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip < 0. The third row gives the diﬀerence in the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip > 0 from constituencies with Υip < 0.

Table A.22: Index of Government Schemes (including coalition governments)

(1) (2) (3) (4) (5) Ruling 0.011 -0.0018 0.025 0.019 -0.0080 (0.023) (0.036) (0.028) (0.031) (0.026) |Vote Share Diff| -0.0050 -0.0023 -0.0053∗ -0.0019 -0.0028 (0.0038) (0.0090) (0.0029) (0.0075) (0.0038) |Vote Share Diff|*Ruling 0.0020 -0.00057 0.0016 0.0029 0.0020 (0.0031) (0.0092) (0.0026) (0.0084) (0.0044) Observations 3532 1925 4727 4505 3738 No. of Clusters 28 28 28 28 28 state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes Interval (-10.87,10.17) (-5.44,5.09) (-16.31,15.26) (-15.54,14) (-10.71,11.94) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1 Control Mean .07 .07 .06 .06 .07 Control s.d. .94 .94 .93 .93 .94 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01

Notes: This table reports the estimates from the regression discontinuity design that relate the political alignment of the constituency representative to the index of government schemes (Yip). I include data for coalition governments in this table and consider a ruling party to be the party which is included in the ruling coalition. The ﬁrst row gives the estimate of the treatment variable which is a dummy that is 1 if the margin of victory of the ruling party candidate in the last election (Υip) was positive. The second row gives the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip < 0. The third row gives the diﬀerence in the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip > 0 from constituencies with Υip < 0.

77 Table A.23: Growth of Satellite Night Lights (including coalition governments)

(1) (2) (3) (4) (5) Ruling 5.11 2.09 1.76 4.78 3.34 (3.81) (5.01) (3.52) (4.43) (2.58) |Vote Share Diff| -0.37 -0.37 -0.43 -0.019 0.0091 (0.54) (1.12) (0.34) (1.37) (0.30) |Vote Share Diff|*Ruling -0.64 0.44 0.22 -0.70 -0.072 (0.66) (1.66) (0.50) (1.21) (0.40) Observations 3589 1976 4754 4896 3702 No. of Clusters 29 29 29 29 29 state-period F.E. yes yes yes yes no PC-period F.E. no no no no yes Interval (-12.49,9.20) (-6.24,4.60) (-18.73,13.81) (-14.64,16.17) (-12.07,10.61) Bandwidth type CCT 0.5*CCT 1.5*CCT CCT CCT Polynomial Order 1 1 1 2 1 Control Mean -45.23 -40.03 -46.12 -46.04 -44.78 Control s.d. 107.67 104.37 109.28 108.44 107.51 All regressions include party fixed effects; *p < 0.1, ** p < 0.05, *** p < 0.01

Notes: This table reports the estimates from the regression discontinuity design that relate the political alignment of the constituency representative to the change in log luminosity (Yip). I include data for coalition governments in this table and consider a ruling party to be the party which is included in the ruling coalition. The ﬁrst row gives the estimate of the treatment variable which is a dummy that is 1 if the margin of victory of the ruling party candidate in the last election (Υip) was positive. The second row gives the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip < 0. The third row gives the diﬀerence in the estimate of the relationship of Yip with the absolute value of Υip among constituencies with Υip > 0 from constituencies with Υip < 0.