Signature Redacted-- Accepted By

Essays in Development and Political Economy by

Mahvish Ifrah Shaukat

M.S., Stanford University (2009) M.A., Stanford University (2008) B.A., University of California, Davis (2006)

Submitted to the Department of Economics in partial fulfillment of the requirements for the degree of

Doctor of Philosophy in Economics

at the

MASSACHUSETTS INSTITUTE OF TECHNOLOGY

June 2019

@ 2019 Mahvish Shaukat. All rights reserved. The author hereby grants to Massachusetts Institute of Technology permission to reproduce and to distribute copies of this thesis document in whole or in part.

Signature redacted A uthor ...... I v Department of Economics 2019 Signature redactE May 15, Certified by...... Benjamin Olken Professor of Economics Signature redact ed Thesis Supervisor C ertified by ...... Esther Duflo Abdul Latif Jameel Professor of Poverty Alleviation and Development Economics Thesis Supervisor Signature redacted-- Accepted by...... Ricardo Caballero Ford International Professor of Economics Chairman, Departmental Committee on Graduate Studies

2019 UN 14 LIBRAMtES ARCHIVES

Essays in Development and Political Economy by Mahvish Ifrah Shaukat

Submitted to the Department of Economics on May 15, 2019, in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Economics

Abstract

This thesis studies three questions in economic development and political economy. Chapter 1 presents empirical evidence on the causal effect of close elections on political selection and performance in India. An extensive literature in political economy predicts electoral competition improves governance. Empirical work testing this theory is scarce, however, due in large part to the difficulty in identifying the causal impact of competition, which is itself an equilibrium outcome shaped by politician behavior. To address this identification challenge, this paper proposes a shift-share instrument that uses variation in aggregate party popularity over time ("shifts") and local party preferences across space ("initial shares") to predict the vote-margin gap in local elections. Using data on Indian state constituency elections, I find political parties respond strategically to competition by selecting high quality (high valence) candidates to run for office, where quality is proxied by lower criminality, higher education, and higher wealth. Once in office, however, these politicians perform no better in facilitating economic growth or providing public goods. Instead, I find parties make temporary, and likely inefficient, reallocations in state-run resources such as electricity in the run up to competitive elections. Together, the results suggest that while the anticipation of close elections may lead politicians to take costly actions prior to an election, electoral competition does not necessarily result in improved economic outcomes. Chapter 2 studies how the spatial allocation of public goods within cities depends on politicians and bureaucrats. In many cities around the world, the bureaucrats and politicians that comprise local governments manage non-overlapping administrative and political units. Us- ing geo-referenced data on public goods, property characteristics, and residents in two major cities in Pakistan, I use a spatial regression discontinuity design to estimate discontinuities in public goods provision and property values across these boundaries. These discontinuities are used to construct measures of political and administrative unit "value-added," and to quantify the variation across each type of unit. I find that the delimitation of political and administrative boundaries results in large variation in public goods quality and self- reported property valuations within cities. Political unit quality is correlated with a number of politician and constituency observables, suggesting that political factors may be important determinants of internal city structure. Chapter 3, coauthored with Adnan Q. Khan, Asim Khwaja, and Benjamin Olken, investi-

3 gates whether strengthening the link between local taxation and urban services can revitalize the social compact between citizen and state. A significant challenge to the provision of local public services in developing economies is the inability to raise adequate resources, especially through local taxation. In many countries, the social compact, whereby citizens agree to pay taxes to fund their desired services, is broken. A low willingness to pay taxes leads to low revenue collection, and prevents adequate service provision, which in turn reduces willingness to pay and can even lead to citizen disengagement from the state. We investigate whether strengthening the link between local collections and urban services can increase citizen's willingness to pay for services, improve service delivery, and enhance local politics. We test this in major urban centers in Punjab, Pakistan via several interventions - including eliciting citizen preferences for specific services when taxes are collected, earmarking revenue for specific services, and enabling local politicians - that credibly strengthen the link between tax collection and urban service provision. This paper presents the experimental design and reports first year impacts on tax payments. We find strengthening the link between taxes and services increases tax payments: taxpayers who failed to make a full payment prior to the interventions increase property tax payments after the interventions are introduced.

Thesis Supervisor: Benjamin Olken Title: Professor of Economics

Thesis Supervisor: Esther Duflo Title: Abdul Latif Jameel Professor of Poverty Alleviation and Development Economics

0111 - 111 11 1,11111111 1111 Acknowledgments

I am indebted to many people for their help, support, and guidance during the writing of this thesis. First and foremost, I would like to express my immense gratitude to my advisors, Ben Olken, Esther Duflo, and Abhijit Banerjee. I cannot find enough words to thank Ben. To say that he has been critical to every stage of my professional development would be an understatement. It was Ben's investment and confidence in me that paved the way to so many opportunities, including this PhD. I consider it one of my greatest fortunes to have him as my advisor. His invaluable insights, constant support, and boundless generosity improved every part of this thesis. Esther inspired me to study development economics, and I am incredibly lucky to have had her unwavering support and guidance, from my time at J-PAL as her RA through these years of graduate school. More than anyone else, she pushed me forward and challenged me to do better, all while believing in me and supporting me. I cannot thank her enough. Abhijit opened his door to me time and time again, listened to my (often ill-defined) ideas, and guided me through the research process. His sharp insights, enthusiasm for my research, and generosity with his time helped me overcome many roadblocks. In addition to my advisors, I have benefited tremendously from working with Adnan Q. Khan and Asim Khwaja. Asim and Adnan provided me my first opportunity to do field work, and I have learned a great deal from them about the challenges and rewards of collaborating with government partners. I hope to continue learning from them in the future. I have also had the privilege of working with and benefiting from the advice of many faculty at MIT, including Daron Acemoglu, David Atkin, Dave Donaldson, Frank Schilbach, and Tavneet Suri. I would also like to thank Emily Oster, who, along with Ben, introduced me to research in development economics, and Nicholas Hope, who encouraged me to take this path. One of the best parts of the PhD has been the incredible group of classmates that accompanied me on this journey. I would like to thank my development and political economy colleagues - Josh Dean, Chishio Furukawa, Nick Hagerty, Reshma Hussam, Donghee Jo, Gabriel Kreindler, Matt Lowe, Benjamin Marx, Rachael Meager, Mateo Montenegro, Arianna Ornaghi, Otis Reid, Cory Smith, Marco Tabellini, Roman Andr6s Zarate, and Ariel Zucker - for their friendship and support. Special thanks are due to my office mates, Sydnee Caldwell, Daniel Clark, Josh Dean, and Ariel Zucker, who made the day-to-day of the PhD much more enjoyable. Cara Nickolaus left us all too soon. I am grateful for having known her kindness and brilliance. I would also like to thank the Center for Economic Research in Pakistan for hosting

5 me throughout many summers of field work. Working with the "tax team" has been a wonderful experience, and a large part of this thesis would not have been possible without their tireless effort. I thank Osman ul Haq and Omer Qasim for coordinating our team, and facilitating nearly every aspect of my field research. I was fortunate that my time in Lahore often overlapped with that of several other PhD students, resulting in many friendships. I thank Niharika Singh in particular for being a continuous source of encouragement, first in Lahore and then again in Cambridge. My aunt and uncle, Shahla and Javed Iqbal, provided a home away from home for me during each and every one of these summers. I am immensely grateful to them for their love and kindness. I owe more than words can express to my family. My extended family - aunts, uncles, and cousins - have cheered my every step forward, and never allowed me to become discouraged by any step back. My brother, Shahruz, sister, Sonya, nieces, Imaal and Simal - each motivated and inspired me in countless ways. Finally, my father, Nadeem, and my mother, Nabila, gave me the unconditional, limitless love and support without which this thesis would never have been possible.

This thesis is dedicated to the memory of my grandfather, Dr. Abdul Latif.

6 Contents

1 Too Close to Call: Electoral Competition and Politician Behavior in India 11 1.1 Introduction ...... 11 1.1.1 Related Literature . . 16 1.2 Conceptual Framework .. . . 18 1.2.1 Campaign Strategy . . 19 1.2.2 Incumbent Performance 20 1.2.3 Testable Implications 21 1.3 Setting and Data ...... 21 1.3.1 Indian State Elections 21 1.3.2 Data ...... 23

1 A Pm irical Praimpwrk 25 1..1 MinSpcfiato...... 1.4.1 Main Specification ...... 25 1.4.2 Instrument for Electoral Competition ...... 27 1.4.3 Constructing the Instrument ...... 29 1.4.4 First Stage ...... 34 1.5 R esults ...... 35 1.5.1 Political Selection ...... 35 1.5.2 Resource Allocation ...... 43 1.5.3 Campaign Strategy: Additional Outcomes ...... 47 1.5.4 Incumbent Performance ...... 49

1.6 Conclusion ...... 52

2 Shaping Cities with Political and Administrative Boundaries 87

7 2.1 Introduction ...... 87 2.2 Empirical Setting ...... 91 2.3 Empirical Framework ...... 92 2.3.1 Fixed effects regression discontinuity design 93 2.3.2 Assessing magnitudes ...... 95 2.3.3 Correlates of political unit quality ...... 96 2.4 Data ...... 96 2.4.1 Description ...... 96 2.4.2 Summary Statistics and Balance ...... 99 2.5 R esults ...... 99 2.5.1 Estimating fixed effects across boundaries 100 2.5.2 Assessing magnitudes ...... 103 2.5.3 Correlates of political unit quality ...... 104 2.6 Concluding remarks ...... 107

3 Rebuilding the Social Compact: Urban Service Delivery and Property Taxes in Pakistan 129 3.1 Introduction ...... 129 3.2 Empirical Setting ...... 133 3.3 Conceptual Framework ...... 134 3.4 Experimental Design ...... 136 3.4.1 Sample ...... 136

3.4.2 Main Interventions ...... 137 3.5 Data and Empirical Framework ...... 140 3.5.1 Data ...... 140 3.5.2 Empirical Framework ...... 142 3.6 Implementation and First Year Re sults ...... 143 3.6.1 Balance Checks ...... 143 3.6.2 Tax morale at baseline ...... 143 3.6.3 Preference elicitation ...... 144

8 3.6.4 Service delivery ...... 145 3.6.5 First year impact on tax payments ...... 146 3.7 Concluding remarks ...... 147

9 10 Chapter 1

Too Close to Call: Electoral Competition and Politician Behavior in India*

1.1 Introduction

Electoral competition has long been theorized to improve the selection and performance of politicians. Much as market competition between firms can raise consumer welfare, so too can electoral competition raise voter welfare by creating incentives for political parties to select high quality candidates to run in elections, and for politicians to perform well once in office (see, for example, Stigler 1972, Wittman 1989, Wittman 1995).1 Empirical work testing this theory, however, is scarce, particularly for developing countries where competition may play an even greater role in disciplining politicians in the absence of strong institutions. This paper studies political selection and performance when an election is "too close to call" in advance. The setting is India, where a large body of work has demonstrated the

*I am very grateful to Abhijit Banerjee, Esther Duflo, and Ben Olken for their advice and generous support throughout this project. I thank Daron Acemoglu, Pablo Azar, Sydnee Caldwell, Josh Dean, Dave Donaldson, Chishio Furukawa, Nick Hagerty, Donghee Jo, Gabriel Kreindler, Matt Lowe, Monica Martinez- Bravo, Benjamin Marx, Mateo Montenegro, Wayne Sandholtz, Garima Sharma, Nabila Shaukat, Niharika Singh, Cory Smith, Tavneet Suri, Marco Tabellini, Roman Andres Zarate, Ariel Zucker and participants in the MIT Development Seminar for their many helpful comments and suggestions. Special thanks to John Firth and Nishith Prakash for generously sharing shape files and data, and to Shahan Shahid for helping collect data. All errors are my own. Electoral competition is often used to assess the strength of democratic institutions. For example, a core component of the Polity Score, which measures a country's level of democracy, is the competitiveness of elections.

11 first-order importance of political institutions in determining economic outcomes, 2 but where relatively little is known about the causal impact of electoral competition. A simple conceptual framework, shown in Figure 1, illustrates the possible channels electoral competition can affect politician behavior, and ultimately economic outcomes. First, political parties may respond strategically to a change in electoral competition prior to an election, selecting a better candidate, reallocating resources, or mobilizing voters to improve the odds of winning a competitive constituency. Second, if parties select a different type of candidate to run in competitive constituencies, the incumbent elected in a competitive election may perform better or worse than the incumbent elected in a lopsided election, with potential implications for economic outcomes. Finally, the anticipation of electoral competition in future elections may change performance incentives over the long-term, which may also determine economic outcomes. Identifying the causal impact of electoral competition empirically is challenging due to reverse causality. An incumbent politician's performance while in office affects her popularity, changing the level of competition she can expect to face in future elections. Similarly, the actions of political parties prior to an election determine how competitive an election will be. A party's decision to enter or exit a race, the selection of a well-qualified or charismatic candidate, an increase in campaign expenditure - all tilt the balance of a race in favor of one party or another. I address this identification challenge by developing a general framework to estimate the causal impact of electoral competition in first-past-the-post electoral systems. The framework is motivated by the observation that aggregate shifts to a party's popularity over time change the level of competition across constituencies non-monotonically depending on local party preferences. A positive shift in an incumbent party's popularity, for example, increases competition in constituencies where voters typically prefer the opposing party, but decreases competition in constituencies where voters typically lean towards the incumbent party. I use this observation to construct a shift-share instrument that leverages this non- monotonicity for identification. The instrument is constructed in two steps. In the first step,

2 See Besley and Burgess 2002, Pande 2003, Khemani 2004, Chattopadhyay and Duflo 2004, Cole 2009, etc. for examples.

12 I predict vote shares for political parties in each constituency and in each election year using a baseline measure of constituency party preferences ("shares") and a leave-out measure of the growth (or decline) in the aggregate popularity of each party across election years ("shifts"). This step is similar to the approach used to construct a typical shift-share instrument, which decomposes growth rates into aggregate and location-specific components, using the former to construct the instrument. In the second step, I use these predicted vote shares to predict electoral competition. Predicted electoral competition is then used to instrument for actual electoral competition. Because competition is a non-monotonic function of vote shares, I am able to control for any direct effects of party popularity, ensuring identification of the causal impact of competition between parties, rather than party preferences per se. The first part of my paper studies how political parties respond to electoral competition during a campaign. Before an election, a party must decide how to allocate candidates and resources across constituencies to maximize its chances of winning as many constituencies as possible.' I present a simple framework identifying how electoral competition may affect political selection and resource allocation. A party has greater incentive to place higher (observable) quality candidates in competitive constituencies, where voters are less likely to vote based on party loyalty, than in less competitive constituencies, where partisanship may be sufficient to deliver the party a win or a loss. A similar logic can be applied to resource allocation. There are higher electoral returns to placing resources in competitive constituencies than in biased constituencies, where either party loyalty on its own is sufficient to guarantee a win (since the bias is in favor of a party) or even a large influx of resources is unlikely to sway voters (since the bias is against a party). My results are largely consistent with this framework. Using data from Indian state elections from 1989 to 2007 and from 2008 to 2017,5 I first show that candidate turnover across elections is high, with incumbent and non-incumbent parties fielding new candidates across election cycles approximately 39% and 77% of the time, respectively. Electoral com-

3 1n Indian state elections, the party that wins the highest proportion of constituencies can form the state government. 4 This problem is a version of the Colonel Blotto game, which has no known general solution. A simple version of this game with homogenous locations was solved only recently in Roberson (2006). 5Because constituency boundaries were redrawn in 2008, I cannot link constituencies before and after this period. Some outcomes are only available for the earlier or later period.

13 petition increases the incumbent party's turnover rate, consistent with anecdotal evidence that incumbent parties remove sitting politicians unlikely to be re-elected to counter anti- incumbency bias in India. Electoral competition improves the candidate pool by leading to the selection of less corrupt candidates, as measured by their criminal records. In particular, a one standard deviation increase in competition, as measured by a 0.30 reduction in relative vote shares between the incumbent party and strongest opposition party, lowers the probability a candidate has a criminal record by 4.8 percentage points (18%). Electoral competition also affects the characteristics of elected politicians: an increase in competition lowers the probability an elected politician is corrupt by 7.8 percentage points (23%) and leads to a 33% increase in a politician's reported net assets at the time of election. That politician assets are higher in competitive constituencies may either indicate that these individuals are higher ability, with assets providing a proxy for a politician's private sector options (if they were previously not in office), or that politicians need to rely on personal wealth to finance campaigns in competitive constituencies. To explore this further, I analyze the effect of electoral competition on the allocation of campaign funds using official campaign finance and expenditure data.' Here, I find little evidence electoral competition affects campaign expenditure. Though most of my estimated effects are imprecise, they suggest if anything that an increase in electoral competition leads to a decrease in official spending. This result provides suggestive evidence that parties may rely on self-financing candidates to win competitive constituencies. In contrast to the results on campaign funds, I find electoral competition has a large and significant effect on the allocation of state-run resources prior to an election. A number of studies (Min and Golden, 2014; Baskaran et al. 2015) have shown Indian state governments induce electoral cycles in electricity provision. Using data on night lights, I find that electoral competition amplifies these cycles. A one standard deviation increase in anticipated competition leads to a 1.6 percentage point increase in a constituency's annual night lights growth in the year prior to an election. This effect is large in magnitude: for comparison, Asher and Novosad (2017) estimate a 4 percentage point increase in night lights

6 As discussed later in the paper, this official data is almost certainly an underestimate of true campaign costs.

14 growth for constituencies aligned with the state party over a five-year electoral term. Since it is unlikely that spatial and temporal shifts in electoral competition are coincident with resource needs, this reallocation is likely inefficient. Finally, I show that electoral competition also leads to the strategic entry and exit of parties in elections. Though the anticipated closeness of a race does not affect the total number of parties, it does change the type of party that chooses to enter an election. An increase in competition encourages the entry of major parties with a high anticipated likelihood of winning, and leads to the exit of small parties with a low anticipated likelihood of winning. In particular, a one standard deviation increase in the closeness of a race reduces the number of small parties by 11%. This suggests that small parties either strategically opt out of races where they are unlikely to garner even a small vote share, or that they are blocked from entering from major parties who fear losing crucial votes. The second part of my paper shows that despite changes in political selection and resource allocation prior to an election, electoral competition does not lead to a significant improvement in incumbent performance after an election. I measure incumbent performance using three economic indicators: employment growth, night lights growth, and public goods provision as measured by road building in a rural roads program. The incumbent of an Indian state assembly can plausibly influence all three indicators.7 I first test whether electoral competition has a selection effect on incumbent performance by changing the type of politician elected in competitive constituencies. Perhaps surprisingly, I do not find evidence that politicians elected in competitive constituencies perform better once in office. The estimated effect for each economic indicator is negative, and though statistically insignificant, I cannot rule out large negative effects on employment growth, night lights growth, and road building. Next, I test whether electoral competition influences incumbent performance through a moral hazard channel: an incumbent secure in her tenure may not be as motivated to exert effort as the incumbent who expects to face a high level of anticipated competition in the

7 Several studies show how politicians affect these outcomes: Asher and Novosad (2017), for example, show that politician alignment with the ruling party leads to significantly higher employment and increased night light growth in a constituency. Similarly, Lehne et al. (2016) and Prakash et al. (2016) provide strong evidence that politicians influence implementation of the rural roads program.

15 next election. Though here too, I cannot reject a null effect on any economic indicator, the confidence intervals do not rule out a large negative effect on employment growth, or a large positive effect on night lights growth. The effect on night lights is of similar magnitude to the increase in total light emissions observed in the year prior to an election, providing suggestive evidence that any increase in night lights is a result of actions taken close to an election, rather than throughout an incumbent's term in office. Taken together, the empirical analysis of incumbent performance suggests electoral competition may do little to improve economic outcomes and, if anything, may lead to worse outcomes. The negative (though insignificant) selection effects indicate that even though candidates selected to run in competitive constituencies are more electable, they are not necessarily better politicians once in office (at least when using observable economic indicators to measure performance). This implies the qualities that make a candidate electable

(e.g. wealth or, in the case of the incumbent party, being a "freshman" politician) are not necessarily correlated with the qualities that make a good politician. Similarly, the moral hazard results show that politicians only respond to anticipated electoral competition close to an election. This result is consistent with the large literature on electoral cycles (e.g. Khemani (2004), Cole (2009), etc.) that show politicians reallocate resources in response to reelection concerns, but only in election years. Whether this is because voters only pay attention to politician actions in this time period, or because it is difficult to forecast future electoral competition far in advance is unclear. In either case, however, the findings suggest that electoral competition has limited influence over incumbents during their full term in office.

1.1.1 Related Literature

This paper contributes to three strands of literature:

Electoral competition: A handful of empirical studies examine how electoral competition affects political selection and performance. On the selection side, De Paula and Scoppa (2011) and Galasso and Nannicini (2011) study Italian politicians and find that competition improves the quality of elected politicians. Similarly, Banerjee and Pande (2018) show that

16 stronger group identity in India worsened candidate quality. Banerjee et al. (2013) also show that reducing incumbency advantage in India motivated less politically experienced candidates to enter politics. On the performance side, the existing literature shows competition leads politicians to select less partisan, more growth oriented policies (Besley et al. (2010)), improve public goods provision (de Janvry et al. (2012), Arvate (2013), Acemoglu et al. (2014), Acemoglu et al. (2017), Kosec et al. (2015)), and reduce rent-seeking behavior (Svalerd and Vlachos (2009)). Nath (2015), on the other hand, finds that competition leads to worse performance by limiting an incumbent's ability to use dynamic incentives to motivate bureaucrats. Many of the studies above use institutional features to identify the effects of electoral competition.' In contrast, I identify the effects of competition using short-term changes to party popularity across election cycles. This variation is a direct measure of competition, rather than a combination of competition and institutional factors, which may have different implications for selection and performance. My identification strategy is closest to the handful of papers that exploit aggregate shifts in party popularity to instrument for electoral competition (Svaleryd and Vlachos (2009), Marx (2014), Kosec et al. (2015)). I develop a general framework building on this source of variation and using the non-monotonicity of competition in vote shares for identification.

Political selection: This paper is also related to a broader literature on how electoral institutions in general determine political selection (Caselli and Morelli (2004), Ferraz and Finan (2009), Keane and Merlo (2010), Dal B6 et al. (2017), etc.) - an important question given the large body of work demonstrating leaders play an important role in determining policy and economic performance. My empirical analysis shows how the anticipated likelihood of winning a race affects the quality of candidates and elected politicians.

Electoral cycles and swing districts: Finally, this paper contributes to the literature on political business cycles and the tactical redistribution of resources to competitive districts prior to an election (Baleiras and Costa (2004), Khemani (2004), Cole (2009), Min

8 For example, Acemoglu (2013) uses the number of ruling families in Sierra Leone, Besley et al. (2010) use the introduction of the Voting Rights Act in the United States, and Nath (2015) uses a change in constituency reservation status in India.

17 and Golden (2013), Baskaran et al. (2014)). Many of these papers rely on time variation in years to an election and cross-sectional variation in competition to determine whether politicians target resources to swing districts prior to an election. My paper contributes to this literature by using an instrumental variable to address the endogeneity concern that realized competition is likely a function of prior resource allocation. The remainder of the paper is organized as follows: Section 1.2 presents a simple conceptual framework identifying channels through which electoral competition can affect politician behavior. Section 1.3 provides some background of electoral competition in India and describes the data. Section 1.4 develops an empirical framework to identify the casual effect of electoral competition and presents the first stage results. Section 1.5 presents the results. Section 1.6 concludes.

1.2 Conceptual Framework

To understand how electoral competition affects politician behavior, I present a simple framework identifying possible channels of influence. My framework is adapted from Banerjee and Pande's (2018) model of political competition and candidate selection. For simplicity, suppose there are two parties, L and R. Constituencies vary in their degree of partisanship, which is parameterized by A E [-1, 1]. A high (positive) value of A implies a high degree of partisanship in favor of party L, while a low (negative) value of A implies a high degree of partisanship in favor of party R. Constituencies with A close to the median are compet-

itive. A voter selects party L if UL + AP >_ UR - AP, where UL (UR) denotes the utility a voter receives from party L's (R's) actions and P is a positive number capturing the tradeoff between a voter's valuation of party actions and a voter's partisanship. Electoral competition can affect politician behavior across two time periods: (1) prior to an election, when political parties are choosing a campaign strategy, and (2) after an election, when elected politicians are in office. Each period is considered below.

18 1.2.1 Campaign Strategy

Consider a political party's decision problem before an election. The party has a fixed pool of candidates and resources that must be allocated across constituencies, which vary in their degree of competitiveness. Selecting a better candidate or better resources relative to other parties can increase the odds of winning. The party's objective is to win as many constituencies as possible; winning the highest proportion of constituencies provides an opportunity to form a government at the state level. Parties determine their allocations simultaneously, and without knowledge of any competing party's strategy. The first problem is candidate allocation. Assume that candidates vary in valence and quality. A candidate's valence is a function of observable characteristics such as education, past experience, criminal background, etc. and is valued positively regardless of her ideology. For example, voters may favor candidates with more education than those with less. The former group of candidates would be considered "high valence"; the latter, "low valence". A candidate's quality, which determines her future performance in office, is unobserved but can potentially be inferred from valence. Voters determine who to vote for based on party loyalty and candidate type. A voter will select party L's candidate if

TL + AP > TR - AP, where TL (TR) is party L's (R's) candidate valence measure. How will parties allocate candidates? One possible equilibrium is for both parties to place their high-valence candidates in competitive constituencies. The intuition is straightforward: party L has little incentive to place its high-types in constituencies where voters are biased towards them. In these constituencies, party loyalty on its own can ensure a win. Party L also has little incentive to place its high-types in constituencies where voters are biased against them. Assuming high-types are limited, party L will not want to waste these candidates in constituencies where even a good candidate is unlikely to induce voters to switch from party R to L. In response, party R will also not want to waste its high-types in these constituencies, since it can win based on party loyalty alone. Both parties therefore allocate their high-type candidates to competitive constituencies.'

9 This argument relies on the assumption that even the best candidate from another party is unlikely to sway voters in highly partisan constituencies. A simple example can provide further intuition. Suppose there are three constituencies with A 1 = -1, A2 = 0, A3 = 1. Party L has three candidates ranked in order of type: AL > BL > CL. Similarly, Party R has three candidates ranked in order of type: AR > BR > CR.

19 Parties also have greater incentive to allocate resources to competitive constituencies. Competitive constituencies offer higher electoral returns than biased constituencies, where either party loyalty on its own is sufficient to guarantee a win (since the bias is in favor of a party) or even a large influx of resources is unlikely to sway voters (since the bias is against a party). 10

1.2.2 Incumbent Performance

After an election occurs, competition can affect incumbent performance through two channels: (1) selection, with competition changing the type of politician elected, or (2) moral hazard, with anticipated competition changing incentives. Selection occurs if parties allocate candidates strategically. Whether selection improves incumbent performance, however, is ambiguous. Voters choose who to vote for based on a candidate's electability, which may or may not be correlated with a candidate's quality. If electability and quality are positively correlated, than selection should improve subsequent politician performance. If, however, electability and quality are not positively correlated, selection may not affect performance at all. This can occur if voters are myopic or uninformed about what makes a "good" politician. Competition can also affect politician performance through a moral hazard channel. Incumbents expecting to face a highly competitive race in the future may exert greater effort to ensure they are re-elected. Moral hazard will occur only to the extent that politicians can anticipate the true level of competition they will face in an upcoming election. When the true level of competition is unknown until very close to an election, this moral hazard effect will be dampened.

Assume that CL AR - 2P, which implies that when a constituency is biased towards party L (A = 1), even party R's most electable candidate cannot beat party L's least electable candidate. Similarly, assume that CR> AL - 2P. Intuitively, voters in highly partisan constiutencies select only on party loyalty and ignore candidate type altogether. This implies that party L will place CL in the constituency with A = 1. Since party R is guaranteed to lose in this constituency, it will place BR, so as not to waste its best candidate AR. The best candidates from both parties, AR and AL will be placed in the most competitive constituency where A = 0. 1 0This argument assumes there is no trade-off between resources and candidates. In a richer model, parties may substitute between resources and candidates when making allocation decisions.

OF- 1.2.3 Testable Implications

This simple framework yields three testable implications:

1. Candidate selection: Parties will place high valence candidates in competitive constituencies.

2. Resource allocation: Parties will allocate more resources to competitive constituencies prior to an election.

3. Incumbent performance: Competition will improve incumbent performance through a selection effect if parties allocate high valence candidates to competitive constituencies and this type is positively correlated with quality. Competition will improve incumbment performance through a long-term moral hazard effect if politicians can anticipate future competition.

1.3 Setting and Data

1.3.1 Indian State Elections

My empirical analysis focuses on state elections in India. State elections are conducted in a first-past-the-post system, with the candidate obtaining a plurality of votes in an election winning the single-member constituency. The party that wins the highest proportion of constituencies within a state has the opportunity to form the state-level government, either on its own or by forming a coalition with other parties. Elections are held every five years," with different states holding elections in different years. Though the Indian National Congress Party dominated elections in India's early post-Independence years, several parties have become formidable challengers over the last few decades. These parties include the Bharatiya Janata Party (BJP), Congress' main opposition party at the national level, as well as several regional and caste-based parties that regularly win elections at the state-level. Figure 2A maps electoral competition over time and across regions in India. The average winning margin, which measures the difference in vote shares between the winner and runner-up candidate in an election, declined from between 15 to

"Elections can be called early, but most election terms last the full five years.

21 20% in the 1950s to approximately 12% in the 2000s, an indicator of highly competitive elections.12 . Figure 2B plots the proportion of state assembly seats won in competitive or lopsided races. Here too, we see Indian elections have become increasingly competitive with the proportion of lopsided races falling from around 30% prior to 1990 to around 15% in the 2000s. The average state election fields 12 candidates, with approximately half representing parties and half running as independents. In most elections, however, the top two or three candidates capture 60 to 70% of the vote share, indicating the effective number of parties in an election is much lower. The top parties in a constituency vary by state. Figure 2C displays variation in vote shares in 2007 state elections in Gujarat, Punjab, and Uttar Pradesh. Congress was among the top three parties in each state, but its relative strength varied considerably: in Gujarat and Punjab, Congress captured 20 to 40% of the total vote share, but captured only 5% in Uttar Pradesh. The identity of the opposition party also varied by state: in. Gujarat, Congress faced the stiffest competition from BJP, but in Punjab, the Sikh regional party Shiromani Akali Dal (SAD) and in Uttar Pradesh, the caste-based Bahujan Samaj Party (BSP) formed the strongest opposition. Once elected, state governments command significant administrative and legislative power. State governments are responsible for the provision of public goods and services, overseeing credit, managing land rights, and collecting retail taxes (Asher and Novosad, 2017). In addition, state governments influence the management of state-run companies, either by direct intervention in company decisions or indirect political appointments. For example, state governments frequently intervene in the operation of state power utilities, selecting villages for electrification projects and determining the location and length of power outages (Baskaran et al., 2015). State governments also influence state-owned public sector banks by appointing board members to key committees that set credit levels (Cole 2009). Though the executive branch wields much of this power formally, elected members of state assembly constituencies (MLAs) play a significant informal role acting as intermediary between citizens and the state. Voters expect MLAs to help obtain goods and services

12For comparison, the average winning margin in the 2016 elections for the US House of Representatives was 37%.

22 (Asher and Novosad, 2017, Chopra, 1993) and assess MLAs accordingly on their ability to "get things done" (Jensenius, 2013). MLAs also exert control over the state bureaucracy via promotions and transfers (Iyer and Mani 2012, Sukhtankar and Vaishav 2015). Taken together, this formal and informal power over public goods, infrastructure, and bureaucrats suggests that state governments may have considerable influence on the economic prospects of a constituency. A growing body of empirical work has provided evidence in support of this hypothesis, demonstrating that the state government and MLAs influence employment and night light growth in a constituency (Asher and Novosad, 2017; Baskaran et al. 2015) and the provision of public goods (Lehne et al, 2016; Prakash et al, 2016).

1.3.2 Data

I conduct the empirical analysis on a panel of Indian state assembly constituencies from 1989 to 2007, and from 2008 to 2017.:3 To study the effects of electoral competition on politician behavior, I combine data from several sources, described in detail below:

Elections: Data on Indian state elections is assembled for each election held between 1989 and 2017 using the Election Commission of India's reported statistics. I collect data on each party that fielded a candidate in an election, party vote shares, and voter turnout. I also collect data on alliances between the party in control of the state government and other parties between 1989 and 2008 from Asher and Novosad (2017).

Candidate characteristics: Data on candidates is collected from the Election Commission and the Association for Democratic Reforms (ADR), a non-partisan organization dedicated to increasing transparency in electoral politics. The Election Commission reports the gender and names of all candidates in all election years, the latter allowing me to determine if a candidate has prior political experience or is a newcomer. ADR also reports a candidate's age, education level, declared assets and liabilities at the time of the election, and any criminal convictions against the candidate. This data, however, is available only from 2004 onwards.

13Because constituency boundaries were redrawn in a delimitation exercise in 2008, I cannot link constituencies before and after this period.

23 Official campaign financing and expenditure: The Election Commission of India publishes official campaign financing and expenditure data for all winning candidates of state assembly elections from 2008 onwards. Official spending is capped at Rs. 2.8 million (approximately $40, 000), but it is well-known that candidates spend in excess of that amount

(see Bussell et al. (2018)). This data therefore reflects how candidates finance campaigns officially and how this money is spent.

Night lights: I use satellite data on night lights to proxy for electricity supply and economic growth. The National Oceanic and Atmospheric Administration publishes gridded average annual night light data from 1992 onwards. Night lights are measured on a scale from 0 to 63, allowing me to compare night lights intensity across grids in the same year.1 4 I match grids to constituency polygons in each election year and compute the log of total night lights intensity. I estimate changes in night lights over a short, one year period and a longer term five year period. It is likely that changes in night lights over the shorter time horizon are driven primarily by changes in electricity supply, rather than economic growth. However, because I cannot conclude this definitively, I interpret night lights as a proxy for both electricity supply and economic growth.

Employment growth: Data on employment growth is obtained from the Economic Cen- suses conducted in 1990, 1998, and 2005. The Censuses track all firms from the manufac- turing and service sectors across India in repeated cross-sections. I use Asher and Novosad's (2017) Census data, which is merged to constituencies using location identifiers, and use their measure of employment growth: the change in log constituency-level employment from 1990 to 1998 and 1998 to 2005. This measure is an imperfect reflection of a politician's performance during her election term: the 7 year time frame spans two election terms so it captures a combination of the performance of a politician and her successor. Given the lack of high frequency, high resolution economic data, however, this measure allows for the best assessment of a politician's impact on economic performance in her constituency.

Public goods: Public goods are measured using construction data from the Pradhan Mantri

14Because satellite-specific factors such as sensor properties can affect how well night lights are measured, I can only compare grids in the same year, which are all measured using the same satellite technology. This is not a problem in my setting since my empirical specifications include year fixed effects.

24 Gram Sadak Yojana (PMGSY), a national rural roads plan to provide roads to unconnected villages. The program is implemented by bureaucrats, but there is strong evidence that politicians influence the allocation of road contracts (Lehne et al. 2016) and project completion (Prakash et al. 2016). Data is collected from 2000 onwards at the Census-block level. I use matched data from Prakash et al. (2016), which assigns a block to a constituency if at least 50% of villages in a block also belong to that particular constituency.

1.4 Empirical Framework

In this section, I present a general empirical framework to estimate the effects of electoral competition on politician behavior. First, I introduce the main specification (Section 1.4.1) and propose a shift-share instrument for electoral competition (Section 1.4.2). Next, I detail how the instrument is constructed and discuss the identifying assumptions (Section 1.4.3). Finally, I present the first stage results (Section 1.4.4).

1.4.1 Main Specification

The goal of this paper is to estimate the causal effect of electoral competition on politician behavior in Indian state elections. Since most outcomes are observed for constituencies across multiple years, the basic specification can be written as follows:

ycst = /Competitioncst + cac + 6 st + Ecst (1.1) where Ycst is an outcome for constituency c in state s and year t, Ccs, is a measure of electoral competition either anticipated in an upcoming election or realized in the past election, ac are constituency fixed effects and 6st are state-year fixed effects. Constituency fixed effects absorb time-invariant variation at the constituency level, while state-year fixed effects absorb variation across years within a state. The coefficient of interest, /, estimates the effect of competition on outcomes within the same constituency across years, compared to other constituencies in a given state-year. Standard errors are clustered at the constituency level.

25 Defining competition

I measure electoral competition in two ways: the winning margin and Herfindahl- Hirshman Index (HHI). I define competition as the difference in (realized or expected) vote shares between the incumbent and her opposition. In first-past-the-post electoral systems, where a party need only secure a plurality of votes to win, this difference reflects the level of competition an incumbent faces when trying to win an election. The closer an incumbent party's vote share to the opposition, the more competitive an election is for her. Using notation, we have:

Competitionct = 1 - |siest - soest| (1.2) where sicst is the vote share of the incumbent party and socst is the vote share of the opposition. In multi-party settings such as India where there may be more than one opposition party, socst is the vote share of the (realized or expected) strongest opposition, socst = max{scst, s2cst, . . , skcst}. There are two limitations to this measure. First, the measure is incumbent-centric, and ignores cases where a party other than that of the incumbent may be the relevant challenger to the opposition. Second, there are cases where the assumption that the incumbent party ignores the behavior of all other parties except the strongest opposition may be less likely to hold. For example, suppose in a three-party setting Constituency 1's vote shares for parties A, B and C are {0.50,0.48,0.02}, while Constituency 2's vote shares are {0.35, 0.33, 0.32}. Using my measure of competition for the incumbent party A, both constituencies would be assigned a value of 0.98, indicating highly competitive races. However, it is likely that party A faces a greater degree of competition in Constituency 2, where both opposition parties are strong challengers, than in Constituency 1, where only one opposition party poses a credible threat. To deal with these scenarios, I use an additional measure of competition: the dispersion of party vote shares. To construct this measure, I compute the

26 HHI for the incumbent and opposition parties. Using notation, we have:

k 2 Dispersionct = 1 - Z(spcst) (1.3) P=1

This measure is not incumbent-centric and increases as a party faces a credible threat from multiple opposition parties.1 5

1.4.2 Instrument for Electoral Competition

Identification challenge

The challenge in estimating equation (1) is that electoral competition is likely to be endogenous to outcomes of interest. In particular, unobserved time-varying constituency characteristics may be correlated with electoral competition and outcomes, leading to biased estimates of 0. Consider a specification estimating the effects of competition on candidate selection. A party's choice of candidate will change its expected vote share, shifting the level of competition. Suppose, for example, that a party fields a wealthy candidate who is able to finance vote buying exchanges prior to an election. If the party was otherwise not expected to do well in the race, this would increase the level of competition. If, on the other hand, the party was expected to do well, selecting a candidate who could take actions to increase the party's vote share even further would decrease the level of competition. In either case, estimates of the effect of competition on the selection of wealthy candidates would be biased (upwards in the first case, downwards in the second case). A similar concern holds when estimating the effects of competition on economic outcomes. Incumbents in constituencies with high economic growth or better public goods provision likely face less competition than those in constituencies with low economic growth or poor public goods provision. This reverse causation would bias estimates of 3 downwards. On the other hand, more parties may choose to field candidates in constituencies experiencing economic growth or better public goods. These constituencies are arguably easier to administer and offer higher potential rents. This would lead to greater competition, and

15If there are k parties, dispersion can range from 0 (vote shares are concentrated in one party), to 1 - (vote shares are dispersed equally among all k parties).

27 bias estimates of / upwards."

Motivation for instrument

To address these endogeneity problems, I propose a shift-share instrument that exploits changes to aggregate party popularity over time to predict electoral competition. To motivate the instrument, consider two types of voters: (1) partisan voters, who vote for a party based on ideological preferences, and (2) swing voters, who vote for a party based on performance. Swing voters observe a party's performance not only at the constituency level, but also at the state or national level. This latter source of variation can be used to construct an instrument. As an example, a state-level corruption scandal involving politicians from the Congress Party may compel swing voters to shift their support towards Congress' opposition. This shift in vote share will change the level of competition in a constituency, with the direction of the change depending on the policy preferences of its partisan voters. In constituencies with a large Congress base, the state-level corruption scandal will lead to an increase in competition by increasing the vote share of the opposition. In constituencies with a low Congress base, the corruption scandal will lead to a decrease in competition, since the opposition already has a secure vote share. Because changes in competition are driven by actions at the state rather than constituency level, they are not susceptible to the endogeneity concerns described above. The motivation for the instrument is shown graphically in Figure 3A. For simplicity, I illustrate a two-party setting, though the intuition carries over to the multi-party case. Suppose there are two parties, a left-wing party, L, and a right-wing party R. Constituency preferences vary from left (0 to 0.5) to right (0.5 to 1). Competition, as defined in equation (2), is plotted as a function of constituency policy preferences." Constituencies where policy preferences are close to the median are competitive. Constituencies where policy preferences

16Yet another potential source of bias is the correlation between changes in party vote shares and changes in constituency demographics such as income level or ethnicity. If voters select parties along ethnic lines (which is often the case in India), we may estimate / > 0 (alternatively, / < 0), not due to the effect of competition on politician performance, but because constituencies with more heterogenous (alternatively, homogenous) populations have better outcomes. 17We can also motivate the instrument by plotting dispersion, as defined in equation (3), as a function of constituency policy preferences. The functional form of electoral competition will change, but the intuition carries over.

28 are towards either extreme (left or right) are not competitive: the left-wing party is always elected in constituencies where preferences are towards the left, while the right-wing party is always elected in constituencies where preferences are towards the right. Suppose a state-level corruption scandal occurs, reducing the popularity of the right-wing party among swing voters. This negative shock, shown in Figure 3B, shifts the competition curve towards the right and changes the level of competition across constituencies non-monotonically. Constituencies near the median which were once competitive become less so as the left-wing party becomes relatively more attractive. On the other hand, constituencies which were once the right-wing party's stronghold become more competitive. In a multi-party setting such as India, the effects of state-level shifts in party popularity are similar, though more complicated since the composition of policy preferences can vary across constituencies. For example, some constituencies may be dominated by the left-wing and right-wing parties, while others may be dominated by the left-wing and more centrist parties. Aggregate changes in the right-wing party's popularity would affect these types of constituencies differently, with competition changing in constituencies choosing between the left-wing and right-wing parties, and remaining the same in constituencies choosing between the left-wing and centrist parties. The basic intuition, however - state-level changes in a party's popularity affect a constituency's level of competition - still holds.

1.4.3 Constructing the Instrument

Using the intuition presented in Figure 3, I construct a shift-share instrument, where "shifts" are changes to aggregate party popularity over time and "shares" are a measure of constituency policy preferences, to predict electoral competition. I construct the instrument in two steps. First, I predict party vote shares for each constituency and in each election year using the shifts and shares. Second, I use these predicted vote shares to predict electoral competition. Predicted electoral competition is used to instrument for actual electoral competition. Each step is described in detail below.

29 Predicting party vote shares

Party vote shares are predicted using state-level shifts in party popularity and constituency policy preferences. I compute the state-level shifts using a leave-out measure of the growth or decline in party support in a state across election years. By leaving out the change in party support in the constituency of interest, I ensure that changes in vote shares are not driven by location-specific factors such as the selection of a particular candidate or the actions of the incumbent. I use state-level growth rates, rather than an overall national growth rate, since party popularity in India is typically state-specific. The shifts are analogous to those used in the traditional shift-share framework, which decomposes growth rates into an aggregate component and location-specific component, using the former to construct the instrument. I compute constituency preferences using party vote shares in the baseline year, defined as the first election year in the sample period for each state. Baseline vote shares proxy for constituency partisanship, though they are a combination of both constituency partisanship and any shocks to party popularity in the baseline year.' 8 Denoting the baseline party vote share for each party p in each constituency c and state s as zp,, and the leave-out measure of state-level changes to party support from the baseline year to time t as g,-, the predicted vote share for party p is given by:

A = + g 1 (1.4)

The predicted party vote shares, spcst, will be close to the true party vote shares to the extent that baseline party support shifts along with state-level trends in party popularity. This is is likely to occur in constituencies with a large proportion of swing voters. It is unlikely to occur in constituencies with a large proportion of partisan voters loyal to a particular party and inelastic to state-level trends. To test if predicted party vote shares are close to actual party vote shares, I estimate

8 Ideally, constituency preferences would be measured by party registration or survey data on party affil- iation at baseline. Given that party registration data is unavailable, however, baseline vote shares provide the next best measure of constituency preferences.

30 the following specification in Table 1:

6 6 VoteShare,8,t = -yPredictedVoteSharepct+ ac + st + cst (1.5)

I predict party vote shares for four major parties in each state: the Indian National Congress Party (INC), the Bharatiya Janata Party (BJP), and the two largest parties in each state

(excluding INC and BJP). Together, these parties capture approximately 76% of the total vote share in each election year and represent the major players in most constituencies. 9

Before estimating the specification, I re-scale party vote shares to sum to 1. This ensures

that I capture the relative popularity of these four parties with respect to one another, rather

than to smaller parties or independent candidates. Each column in Table 1 reports results for a single party. The coefficient on predicted vote share for all parties is positive and statistically significant, suggesting that support for these parties is responsive to state-level trends.O A look at the data suggests the

"4errors" in prediction are largely of commission (I predict a positive vote share, when the

actual vote share is 0) or omission (I predict a 0 vote share, when the actual vote share is

positive). These are driven by idiosyncrasies in the baseline year. For example, if I do not

observe a party compete in a constituency in the baseline year, and I only observe negative shocks to the party's popularity in a state, I will always assign a 0 vote share to that party in that constituency - even if it fielded candidates in later elections. Aside from these types of errors, however, there is a strong relationship between predicted and actual vote shares.

Predicting electoral competition

I construct two instruments for electoral competition using the predicted vote shares described above. Using the measure for electoral competition defined in equation (2), predicted

9 1n principle, it is possible to extend the analysis to additional parties. However, to the extent that the popularity of smaller parties is driven by location-specific, rather than state-specific, factors, the predictions will be less powerful. 20 Note that the predicted vote shares are a combination of constituency-specific and state-year-specific variation. Predicted vote shares are not entirely collinear due to constituency-specific variation in aggregate shocks. The results in Table 1 are qualitatively similar when removing constituency fixed effects, state-year fixed effects, and both constituency and state-year fixed effects from the specification.

31 competition is given by:

PredictedCompetitiont = 1 -| icst - socst1 (1.6)

where sicst is the predicted vote share for the incumbent party and socst = max{slet, s2cst, . .. , Skcst} is the predicted vote share for the opposition party predicted to be the strongest. Using the measure for electoral competition defined in equation (3), predicted dispersion is given by:

k PredictedDispersioncst= 1 - Z(scst) 2 (1.7) p=1

Each predicted measure of electoral competition is used to instrument for actual predicted competition. The basic first stage equations are given by:

Competitionc t = 'PredictedCompetitionct + ac + 6 st + Ecst (1.8)

3 Dispersion,8 t =yPredictedDispersioncst+ ac + st + Ecst (1.9)

Identifying assumptions

The key identifying assumption for the instruments is that predicted competition or predicted dispersion are not correlated with outcomes other than through their effect on actual electoral competition, conditional on constituency and year fixed effects. There are two potential threats to this assumption: endogeneity of shares and the violation of the stable unit treatment value assumption (SUTVA). I address each potential threat below.

Shares: Shift-share instruments are a common empirical strategy in the trade and immigration literature." Several recent studies have focused on opening the "black box" of these instruments to clarify the assumptions needed for identification (Goldsmith-Pinkham et al (2018), Borusyak et al. (2018)). Goldsmith-Pinkham et al. (2018) argue that a shift-share instrument is asymptotically equivalent to using the shares themselves as instruments, im-

21Jaeger et al. (2018) note, "It is difficult to overstate the importance of this instrument for research on immigration. Few literatures rely so heavily on a single instrument or variants thereof."

32 plying that shares cannot be endogenous to outcomes.22 Unlike in the standard shift-share framework, the shifts in my setting affect predicted competition non-monotonically, with the same shift increasing competition in some constituencies while decreasing competition in others depending on the shares. In other words, the shares determine how predicted competition changes both in magnitude and direction in response to a shift. In the standard shift-share framework, by contrast, the instrument is the inner product of shares and shifts and the shares determine only the magnitude of the response to a shift, not the direction. 23 Since I am interested in identifying the effect of a non-monotonic function of shares (where the non-monotonicity is driven by the functional form of competition in vote shares), I can control for any effects driven by the shares themselves in my specifications. I estimate a modified version of equation (1) that controls directly for predicted vote shares:

k Ycst = 3Competitionct + j 0jsiCst + ac + 6 st + Ecst (1.10) i=1

The coefficients Oi control for the direct effects of the predicted vote shares for each party, while the coefficient of interest, 0, estimates the causal effect of competition on outcomes

Ycst -

SUTVA: A second identification concern is the potential violation of the stable unit treatment value assumption (SUTVA). Any shift in party vote shares treats all constituencies in a state, leaving no "pure control" group. To see this clearly, suppose the party in control of the state government prioritizes resource allocation to competitive constituencies. Assuming the level of resources is fixed, a shift in party vote shares that increases the number of competitive constituencies will lead to a reduction in resources in non-competitive constituencies. This concern can be addressed by interpreting 3 appropriately. 3 is the effect of experiencing a greater degree of competition on an outcome relative to other constituencies

22Borusyak et al. (2018) propose an alternative identification assumption relying on the exogeneity of shocks. However, their argument assumes the number of shocks grows in number with the sample - an assumption not applicable to my setting. 23See Appendix B for a simple example clarifying the difference between standard shift-share instruments, and the shift-share instrument used here.

33 in the same state-year that experienced a lower degree of competition. In other words, 3 estimates the effect of competition in the observed equilibrium allocation.

1.4.4 First Stage

Having presented the empirical framework for estimating the causal effect of electoral competition, I now turn to the first stage results. Table 2 estimates the relationship between predicted competition and actual competition (Panel A), and predicted dispersion and actual dispersion (Panel B). Each specification controls for constituency and year fixed effects. In Panel A, Column (1) regresses actual competition on predicted competition. In Columns (2) through (4), I add predicted party vote shares for BJP, Congress, and the top two parties in each state one-by-one. In Column (4), I estimate the preferred first stage equation controlling for all predicted party vote shares:

k Competitioncet = 'PredictedCompetitioncet + 3 6 jsicst + ac + 6st + Ecst (1.11) i=1

The results suggest a strong and robust relationship between actual and predicted competition. The coefficient for predicted competition is positive and statistically significant. Including the direct effects of predicted party vote shares does little to change the coefficient on predicted competition, which remains stable around 0.33. The F-statistic is well above conventional levels in all of the specifications. These results indicate the relationship between predicated and actual competition is driven by the non-monotonicity of predicted vote shares rather than direct policy preferences. In Panel B, I conduct a similar analysis regressing actual dispersion on predicted dispersion. Here again, there is a strong relationship between predicted and actual dispersion that holds even after controlling for the direct effects of predicted party vote shares. The F-statistic is lower than that for predicted competition, but still above conventional levels. In Figure 4, I present an example of how predicted and actual competition vary non-monotonically as a function of initial shares. To illustrate the non-monotonicity, I focus on competition between two parties, BJP and INC, in the state of Karnataka. Figure 4 plots predicted and actual competition as a function of INC vote shares across election

34 years. The blue dots show the actual level of competition between BJP and INC, while the red dots show the predicted level of competition between BJP and INC. Panel A displays the mechanical relationship between competition and INC vote shares in the baseline year. Panel B shows that competition shifts towards the right in 1994, as the BJP becomes a more competitive party vis-a-vis INC. Constituencies that were once INC strongholds become more competitive. In 1999, competition shifts back towards the left, indicating a positive shift for INC in Panel C. Finally, in Panel D, the BJP once again becomes competitive, threatening INC strongholds. Figure 4 also shows that the set of competitive constituencies in Karnataka changes across election years. This suggests a substantial share of constituencies respond to aggregate shifts in party popularity. In Appendix A, Table 1, I compute the standard deviation of competition and dispersion (actual and predicted) within the same constituency across election years to show this occurs in nearly all states: Column (2), for example, shows that the median standard deviation of predicted competition within constituencies is 0.12.24 For comparison, the corresponding standard deviation is 0.19 in Karnataka.

1.5 Results

This section presents the core empirical results on electoral competition and politician behavior. The first part of the analysis focuses on political parties' campaign strategy before an election, while the second part studies incumbent performance after an election.

1.5.1 Political Selection

I begin by presenting the effects of electoral competition on the selection of candidates (Section 1.5.1) and elected politicians (Section 1.5.1).

24Delhi and Tripura are exceptions, with low variation in predicted competition within constituencies. Constituencies in these states experience some variation in competition (the standard deviation of actual competition, shown in Column (1), is larger); but this variation is likely driven by factors other than state- level trends.

35 Candidates

Turnover: A necessary first step in examining how electoral competition affects candidate selection is to establish that candidates change over time. If political parties field the same candidates repeatedly across election cycles, then short-term changes in electoral competition will have little effect on the distribution of candidates across constituencies. I identify whether a candidate has changed across election cycles by conducting a fuzzy string match on candidate names. Average candidate turnover in Indian state elections is high: approximately 75% of candidates running for office are different than the candidate who ran on the party's ticket in the preceding election year. The turnover rate for candidates of the incumbent party is markedly lower in comparison: the incumbent party replaces a sitting politician only 39% of the time. Table 3 shows that electoral competition increases the incumbent party's turnover rate. I estimate equation (10) with candidate turnover as the dependent variable in four samples: (1) all parties' candidates, (2) the incumbent party's candidates, (3) the strongest opposition party's candidates, and (4) all remaining parties' candidates. To control for possible differences in political parties' pool of potential candidates, I also include party fixed effects in each specification. The results without party fixed effects, reported in Appendix Table 2, are largely similar qualitatively. Panel A presents estimates using the incumbent winning margin to measure electoral competition. Columns (1) and (2) report OLS and 2SLS estimates for the sample including all candidates. Both coefficients are negative and similar in magnitude, but the 2SLS point estimate is imprecise. Column (3) restricts the sample to the incumbent party. Though insignificant, the point estimate is positive and large in magnitude. Since the standard deviation of my measure of competition is approximately 0.3, the estimate suggests a one standard deviation increase in competitiveness leads to a 18.9 percentage point increase in the likelihood the sitting politician is replaced by the incumbent party (48% of the mean).

2 5 The corresponding estimate without party fixed effects is also positive, but smaller in magnitude. How- ever, the coefficient is still substantially larger than the effects on non-incumbent parties' turnover rate. The estimated effect using dispersion to measure electoral competition is large and significant, both with party fixed effects (at the 5% level) and without party fixed effects (at the 10% level).

36 In Panel B, I use the dispersion of vote shares to measure electoral competition. Column (3) shows that a one standard deviation increase in dispersion (approximately 0.2) increases the incumbent party's turnover rate by 36.6 percentage points - a doubling of the mean turnover rate of 39%. This estimate is significant at the 5% level. This result is consistent with anecdotal evidence that incumbent parties remove sitting politicians who are unlikely to be re-elected. 21 In settings such as India where there is significant anti- incumbency bias (Linden, 2004), replacing the incumbent politician may be a necessary strategy for the incumbent party to remain viable in competitive constituencies. I find no significant effect of electoral competition on non-incumbent parties' turnover rate. Average candidate turnover for these parties is fairly high (77% for the strongest opposition party and 92% for other parties). Columns (4) and (5) of Panel B report negative effects on turnover for non-incumbent parties, but these coefficients are imprecisely estimated. Though this result should be interpreted with caution (especially since I do not observe a similar effect in Panel A), it suggests that non-incumbent parties may have an incentive to remain with the candidate who ran in the last election. One possible reason may be name recognition: the candidate who ran in the previous election is likely recognizable and unlike the incumbent, cannot be judged by their performance in office. Taken together, these results suggest parties do field different candidates across Indian state election cycles on average, and that competition increases the incumbent party's turnover rate.

Criminal Record: I next examine whether electoral competition leads to the positive or negative selection of candidates in observable measures of valence. The first measure of valence is a candidate's criminal record at the time of declaring her candidacy. Data on candidates' background is available only for 2004 onwards. Since constituency boundaries were redrawn in 2008, I cannot compare constituencies before and after this time period. This means I observe only one election for most constituencies between 2004 and 2008, and only one election for most constituencies between 2008 and 2017.27

26See, for example, https://www.hindustantimes.com/india-news/bjp-to-conduct-survey-in-rajasthan-to- gauge-popularity-of-its-mlas/story-QeL2Tm3Oc2b4RsPwzXASHO.html. 27 Two actual election years are observed, but the baseline year - used to construct initial shares - cannot be used for analysis.

37 To deal with this issue, I run a modified version of equation (10) by stacking constituencies before and after delimitation and running a cross-sectional regression with additional controls. The 2SLS specification is as follows:

p Ycst = 3Competitionc t + 3 Cist + 6 st + 7p+ XtQ + Ecst (1.12) where -y are party fixed effects and X'cst is a vector of controls: the initial level of competition and initial vote shares in each constituency. The controls address the concern that unobserved variation across constituencies may bias estimates. 3 estimates the effect of electoral competition on candidate selection, conditional on a constituency's baseline level of competition, party preferences in the current and baseline year, and state-year fixed effects. Results are reported in Table 4A, and shown graphically in Figures 5A and 5B. Panel A of Table 4A shows a one standard deviation increase in electoral competition lowers the probability a candidate has a criminal record by 4.8 percentage points. Given that 27% of candidates report having a criminal record at the time of declaring their candidacy, this represents a substantial drop of 17.7%. Panel B shows the effects of electoral competition are even larger when measured by dispersion: a one standard deviation increase in the dispersion of party vote shares lowers the probability a candidate has a criminal record by 10.6 percentage points, or 40% of the mean. While the estimated effects are negative for incumbent and non-incumbent parties, the effects are more precisely estimated for the latter group (Columns (4) and (5)). This can be explained by the relatively higher candidate turnover rate for non-incumbent parties - non-incumbent parties frequently select new candidates in each election cycle; the incumbent party less so.

Education: Electoral competition also leads to the positive selection of candidates by educational attainment, with the effect concentrated in non-incumbent parties. I measure education on a 0 to 7 scale, ranging from no formal education (illiterate) to post-graduate education. Columns (1) and (2) of Table 4B report the OLS and 2SLS estimates for all candidates. While the OLS coefficient is insignificant, the 2SLS is positive and precisely estimated: a one standard increase in competition leads to a 0.13 category point increase in a

38 candidate's educational level (2.5% of the mean), while a one standard deviation increase in dispersion leads to a 0.47 category point increase (8.9% of the mean). The results are robust to different ways of coding education (e.g. years of schooling, secondary school completion versus incomplete or no secondary school). Columns (3)-(5) show the effects are driven by candidates of non-incumbent parties, rather than those of the incumbent party.

Net Assets: Next, I examine the effects of electoral competition on a candidate's net assets at the time of declaring candidacy. For non-incumbent candidates, net assets provide a proxy of a candidate's private sector outside options and can therefore be considered a measure of valence. Net assets are measured by the sum of reported movable and immovable assets minus any liabilities. The estimated effects of electoral competition, reported in Table 3C, are insignificant for all parties. Column (3) reports a negative coefficient for the incumbent party, while Column (4) reports a positive coefficient for the strongest opposition party. However, these estimates are too imprecise to distinguish. These results stand in contrast to those for criminality and education, where competition leads non-incumbent parties in particular to select higher valence candidates. Here, there is no significant effect on non-incumbent party candidates' net assets. One possible explanation is that education and criminality are both positively correlated with wealth: a candidate with at least a secondary school education reports 34% higher wealth than a candidate without secondary school education, while a candidate with a criminal record reports 36% higher wealth than a candidate without a criminal record. Since competition positively selects on the former and negatively on the latter, any effect on wealth could cancel out.

Gender and Age: Finally, I also test whether electoral competition leads to the selection of candidates that differ by gender or age. While these demographics are not necessarily linked to valence, it is worth exploring whether electoral competition can create space for traditionally less well-represented groups (e.g. women, younger individuals) to enter politics. Appendix Table 3A reports the effect of electoral competition on the gender of a candidate." I find little significant effect on a candidate's gender, except in the sample of opposition party candidates. Column (3), Panel A shows a one standard deviation increase in electoral

28Since data on gender is available in all election years, I can estimate the specification with constituency fixed effects (e.g. equation (10)).

39 competition, as measured by the incumbent's winning margin, leads to a 3.9 percentage point increase in the likelihood the opposition party selects a female candidate. This is a sizable increase over the mean of 6%. However, the effect is no longer significant when measuring electoral competition by the dispersion of vote shares (Column (3), Panel B), and does not hold for non-incumbent parties other than the strongest opposition (Column (4)) . Appendix Table 3B reports the effect of electoral competition on the age of a candidate. Here too, I find little significant effect, except in the sample of non-incumbent parties (Column (5)), where the effects are significant but small in magnitude. For this sample, a one standard deviation increase in competition lowers the average candidate's age by about one year (from a mean age of 47 years) and a one standard deviation increase in the dispersion of vote shares lowers the average candidate's age by about 3 years. It is worth noting that the coefficients for all samples are negative, indicating candidates in competitive constituencies are younger, but the results are too noisy to reject the null.

Discussion: The results above, summarized in Figures 5A and 5B, provide evidence that electoral competition between political parties leads to the positive selection of candidates in Indian state elections. Candidates placed in competitive constituencies are less likely to have criminal records and are more educated than candidates placed in non-competitive constituencies. The underlying assumption of this analysis is that voters value a candidate's lack of criminal record and educational attainment, which is why they are selected to run in competitive constituencies. Of course, candidates without these attributes are still frequently elected (and re-elected) in India. One explanation is that voters actually prefer criminal or less (observably) qualified candidates because of their ethnic ties or the patronage they provide (Banerjee et al. 2014). Simply comparing the vote shares of high and low valence candidates cannot reveal voters' preferences since, as the above analysis shows, candidates are placed strategically across constituencies. The empirical evidence on voter preferences suggests that voters do not, in fact, prefer criminal or less qualified candidates. In a survey experiment asking voters in Uttar Pradesh to choose between hypothetical candidates with and without criminal records, for example, voters expressed strong preferences for the latter

40 (Banerjee et al. 2014). In another experiment in Delhi, though voters did not pay attention to candidates' criminal records, they were more likely to vote for non-incumbents with higher education (Banerjee et al. 2011).2' This suggests that low valence candidates may be elected not because of voters' preferences, but for other reasons such as a lack of viable competing candidates - an explanation consistent with the empirical analysis of electoral competition above.

Elected Politicians

The preceding analysis demonstrates that electoral competition improves the quality of the candidate pool. Whether electoral competition leads to the positive selection of elected politicians is arguably a more important question since these individuals determine policy once in office. The selection effects on elected politicians may differ from those on candidates if incumbent and non-incumbent party candidates differ on observables, and if certain types of candidates (e.g. more educated, wealthier, etc.) have a higher likelihood of winning. Table 5 estimates the effect of electoral competition on valence (as measured by criminality, educational attainment, and reported assets), and demographic characteristics (gender and age). The corresponding estimates are summarized in Figure 6. The results suggest that the politicians elected in competitive constituencies are less likely to have a criminal record and wealthier than those elected in non-competitive constituencies. Column (1) in Panel A shows a one standard deviation increase in competition decreases the likelihood an elected politician has a criminal record by 7.8 percentage points, or 23%. When using the dispersion of vote shares to measure competition in Panel B, the effect is still large and negative, but imprecisely estimated. Column (2) suggests a positive effect of competition on the educational attainment of educated politicians. These estimates are insignificant, however, and smaller in magnitude than the estimates for candidates. Finally, Column (3) shows elected politicians in competitive constituencies are much wealthier than those in non- competitive constituencies: a one standard deviation increase in competition leads to a 33%

2 9Dal B6 and Finan (2018) provides an overview of empirical evidence showing voters care about valence in a number of settings.

41 increase in reported net assets, while a one standard deviation increase in dispersion leads to a 57% increase. Table 5 also shows that electoral competition has little effect on the age or gender of Indian state politicians. The estimates in Columns (4) and (5) are small and insignificant, suggesting that at least in this setting, competition does not increase the representation of less well-represented groups in legislative assemblies.

Discussion: These results provide evidence that electoral competition also leads to the positive selection of elected politicians. This is seen most clearly in the decline in the likelihood an elected politician has a criminal record as a constituency becomes more competitive. The estimates for education are more muted, perhaps because electoral competition leads to the positive selection by educational attainment of only non-incumbent party candidates (see Table 4B). Since incumbent party and non-incumbent party candidates are almost equally likely to win as a constituency becomes more and more competitive, the selection effect on education is dampened.30 The strongest selection effect is in net assets: the politician elected in a competitive constituency is much wealthier than the politician elected in a non-competitive constituency. Given that there is no significant difference in the wealth of the candidate pool running in competitive and non-competitive constituencies, this implies that wealthier candidates are systematically more likely to win in competitive constituencies. How should we interpret these results? On the one hand, net assets provide a proxy for a politician's private sector outside options (if they were previously not in politics). Higher net assets may therefore signal higher ability, and imply that voters are more likely to select these higher ability individuals in competitive constituencies. On the other hand, these results may also imply that candidates need to rely on personal wealth to finance campaigns in competitive constituencies - the candidates without this source of financing are unable to win. Electoral competition, while improving the quality of elected politicians on some margins such as criminality, may therefore also act as a barrier to less wealthy candidates. I explore the effects of electoral

3 0This result also highlights the importance of looking at the selection effects on candidates and elected politicians separately - even if electoral competition improves the quality of the candidate pool, it may not necessarily improve the quality of elected politicians if the effects on the candidate pool are asymmetric.

42 competition on campaign financing in more detail in the next section.

1.5.2 Resource Allocation

Political parties may strategically allocate not only candidates, but also resources in response to short-term changes in electoral competition. This section examines how electoral competition affects the allocation of campaign funds (Section 1.5.2) and state-run resources (Section 1.5.2) prior to an election.

Campaign Funds

Though there is growing concern about the rising costs of election campaigns in India - in 2014, India was second only to the United States in campaign expenditure, spending an estimated $5 billion in total - few studies have examined how politicians fund their campaigns or how this money is spent due to a lack of reliable data. One notable exception is Bussell et al. (2018), who use politician self-reported survey data to study sources of campaign funding. The survey data suggests candidates for state assemblies are financed primarily through illicit sources (46% of total campaign costs), followed by individual donations (30%), personal income (12%), and party funds (8%). For this analysis, I use official financing and expenditure data submitted by elected politicians to the Election Commission. The Election Commission caps spending by state assembly candidates at Rs. 2.8 million (approximately $40,000), but even candidates vying for much lower stakes seats are known to spend well above this limit. The official data is therefore almost certainly an underestimate of true campaign costs. Still, the data provides a glimpse of how politicians fund campaigns officially and where this money is spent. Campaign data is only available from 2008 onwards, and only for elected politicians

(not all candidates). Since I observe one election in this time period for most constituencies (excluding the baseline election year), I estimate equation (11), which controls for the initial level of competition and initial vote shares in each constituency. Table 6A reports 2SLS estimates of the effect of electoral competition on official campaigning for three samples: (1) all elected politicians; (2) elected politicians of the in-

43 cumbent party; and (3) elected politicians of non-incumbent parties. The effects on total expenditure, presented in Columns (1)-(3), are negative, small in magnitude, and insignificant. The effects on funding from personal wealth are also negative, but large in magnitude. Though these effects are not significant, they suggest that, if anything, elected politicians in competitive constituencies spend less from their own pockets on campaigns - at least officially. Table 6A also shows that elected politicians in competitive constituencies receive less official funding from their parties. The estimated effect for politicians of the incumbent party, reported in Column (8), is significant at the 5% level, though only when electoral competition is measured using the dispersion of vote shares. The effect for politicians of non-incumbent parties is also negative, but not significant. Finally, Columns (10)-(12) estimate the effect of electoral competition on other sources of funding, defined as donations and gifts received from individuals or firms. Here, the point estimates are positive, but again are too noisy to reject the null. These estimates provide suggestive evidence that as constituencies become more competitive, politicians rely increasingly on non-party funds to officially finance their campaigns. Table 6B reports the corresponding 2SLS estimates for campaign expenditures. I analyze disaggregated expenditure for three categories: (1) meetings, print materials, and transportation, (2) electronic/ print media (including television ads), and (3) visits by party leaders. The bulk of campaign funding is spent on the first category. An increase in electoral competition significantly lowers campaign spending on this category (Column (1)), with the effect driven primarily by non-incumbent parties (Column (3)). Consistent with the estimates from Table 6A, I find that elected politicians of the incumbent party spend less on party visits: a one standard deviation in electoral competition as measured by the incumbent margin lowers official spending on party visits by 45%." While most estimated effects on campaign funds are imprecise, they provide suggestive evidence that elected politicians in competitive constituencies spend less officially on

31One possible explanation of this result is that incumbent parties may fear negative press when visiting constituencies where their candidate faces intense competition. A visit to such constituencies may generate negative press for the incumbent party if, for example, a campaign rally draws a small crowd. A visit to a stronghold constituency, by contrast, is likely to have the opposite effect: large crowds and positive press. The incentive to avoid "bad optics" may divert party funds to less competitive constituencies.

44 their campaigns than elected politicians in non-competitive constituencies. Whether spending using total funds - official and unofficial - is also lower in competitive constituencies is unclear. Two pieces of evidence suggest that this may not necessarily be the case. First, financing from donations and gifts is higher in competitive constituencies, suggesting these constituencies may also attract more unofficial donations. Second, as shown in the preceding section, elected politicians in competitive constituencies are much wealthier than those in non-competitive constituencies. This suggests politicians in competitive constituencies may draw on their personal wealth to finance their campaigns unofficially. Indeed, Sukhtankar and Vaishnav (2015) note that as costs of elections has surged in India, parties have become more reliant on self-financing candidates. Better data on campaign funds, perhaps using third-party assessments of visible displays of campaign expenditures, is needed to form a definite conclusion.

State Resources

Along with campaign funds, the party in control of the state government (henceforth "state party") can reallocate resources from state-run companies and institutions towards particular constituencies as part of its campaign strategy.3 2 To test how state resources are reallocated, I focus on electricity supply, which can be proxied using high-frequency and high-resolution satellite data on night lights. Though night lights can reflect both changes in economic activity and changes in electricity supply, I examine short-term changes in night lights in the year before and year after an election. These short-term changes are more likely to reflect changes in electricity supply, which can be reallocated immediately before and after an election, rather than changes in economic activity, which take a much longer time horizon to shift.

3 2See Khemani, 2004; Cole, 2009; Baskaran et al., 2015 for examples of political budget cycles in state-run resources including public goods, credit, and electricity supply, respectively.

45 I estimate the following 2SLS version of equation (10):

Ycst =/1Competitioncet + /2Competitioncet x Pre-Electiont P P + yPre-Electiont + c + (3 x Pre-Electiont+ acc + 6st + ecst i=1 i=1

The outcome variable of interest, Yest, is a measure of total nightlights in constituency. I restrict the sample to the years prior and after an election to estimate short-term changes in electricity supply. /1 captures how competition affects the yearly growth rate in night lights, while /1 + /2 captures the total effect of competition in the year prior to an election. Table 7 presents the results of this exercise using the yearly change in log total night lights as the outcome variable, and the incumbent margin as the measure of electoral competition.33 Columns (1) and (2) report OLS and 2SLS estimates of the equation above.

The OLS estimates are small in magnitude, 01 + 2= -0.006, and insignificant. In contrast, the 2SLS estimate of the total effect is large, 01 + 2= 0.053, and significant at the 5% level. The estimate implies a one standard deviation increase in electoral competition leads to a 1.6 percentage point increase in annual night lights growth. For comparison, Asher and Novosad (2017) estimate a 4 percentage point increase in night lights growth for constituencies aligned with the state party over a five-year electoral term. Baskaran et al. (2015) estimate a similar effect on night lights growth in special election years.34 More broadly, these results are in line with the literature on tactical reallocations motivated by reelection incentives (e.g. Khemani 2004, Cole 2009), which finds that politicians target resources towards swing districts in election years. To my knowledge, however, much of this literature relies on OLS to assess whether politicians target swing districts. The results in Table 7 highlight the importance of correcting for omitted variable bias in the OLS estimates. OLS may be severely downward biased, likely because constituencies that receive

33 In Appendix Table 4, I use an alternate measure: log of total light emissions. The results are qualitatively similar. 34 Baskaran et al. (2015) also find that the increase in night lights growth is concentrated in competitive constituencies by comparing constituencies where the margin of victory was less than 5% in the previous election to constituencies where the margin of victory was greater than 5%. In contrast to their study, I look at the impact of anticipated competition in the upcoming election year and use a continuous measure of competition.

46 large influxes of state resources prior to an election have less competitive races afterwards. In Columns (3) and (4), I test whether the reallocation in electricity supply is directed towards constituencies aligned with the ruling state party. Intuitively, the state party may care primarily about constituencies where its own incumbent in jeopardy, where a reallocation of resources can tilt electoral competition in its favor. It may care less about constituencies where an incumbent of another party is in jeopardy, since a reallocation of resources will likely be attributed to the other party rather than the state party. The OLS estimates are again much smaller in magnitude than the 2SLS estimates. Column (4) shows electricity supply is targeted towards constituencies aligned with the state party. Though the total effect is insignificant, the magnitude is large, /1 + /2 + /33 + /34= 0.046, (p-value = 0.182). While I cannot reject a similar effect for non-aligned constituencies, Column (4) provides suggestive evidence that state-run resources are primarily targeted towards competitive constituencies where the incumbent is aligned with the state ruling party. Overall, Table 7 demonstrates that electoral competition leads to a significant reallocation of state-run resources prior to an election towards swing constituencies. Since it is unlikely that spatial and temporal shifts in electoral competition are coincident with resource needs, this reallocation is likely inefficient.

1.5.3 Campaign Strategy: Additional Outcomes

Before turning to post-election outcomes, I examine the effect of electoral competition on two additional outcomes related to campaign strategy: party entry and exit decisions (Section 1.5.3), and voter mobilization (Section 1.5.3).

Party Entry and Exit

Table 8 shows that electoral competition increases the number of major parties with a high likelihood of winning in a race, and decreases the number of smaller parties with a low likelihood of winning in a race. Columns (1) and (2) report OLS and 2SLS estimates of the effect of competition on the total number of parties. While the OLS estimates are positive and significant, the 2SLS estimates are negative and imprecisely estimated. The

47 upward bias of the OLS estimate is likely due to reverse causality: an increase in the number of political parties increases the competitiveness of a race. The 2SLS estimate suggests that competition does not affect the overall number of parties. Even if the overall number of parties does not vary, competition may affect the type of party that chooses to enter an election. The 2SLS estimates show that major parties, defined as BJP and INC, and the top two regional parties in each state, are more likely to enter competitive elections. Column (4) shows that a one standard deviation increase in competitiveness, as measured by the difference in incumbent and strongest opposition party vote shares in Panel A, increases the number of major parties in a race by 15.5 percentage points (2.9% of the mean). Column (5) estimates the entry decision of major parties excluding the incumbent and strongest opposition. Reassuringly, the estimates are insignificant in Panel A, since the difference in incumbent and strongest opposition party vote shares is not the relevant competition-metric for these parties, and significant in Panel B, which uses the dispersion of all vote shares to measure competition. Finally, small parties are less likely to enter competitive elections. Column (8) indicates that a one standard deviation increase in competition reduces the number of small parties in a race by 24.1 percentage points (11.1% of the mean). A small party's decision to exit a race may be in anticipation of strategic voting. Voters in competitive races are more pivotal and hence have a greater incentive to allocate their votes strategically. Parties may act strategically themselves, opting out of races where they are unlikely to win.35 Alterna- tively, a small party's exit may be the result of vote-buying by major parties. These parties have an incentive to block entry of additional parties in competitive races to improve their chance of winning. Paying off a candidate of a small party so that she does not run can ensure more potential votes. In either case, the results of Table 8 suggest that parties choose which electoral races to enter strategically, opting in when they are likely to win and opting out when they are unlikely to.

35 This possibility has been discussed in the literature (Cox 1997, Fey 2007). Cox (1997), for example, argues that "(party) elite anticipation of strategic voting should lead to prudent withdrawals and hence a reduction in the number of competitors entering the field of battle."

48 Voter Mobilization

In selecting candidates or reallocating resources, a party may be attempting to either persuade undecided voters or mobilize new voters.36 Table 9 tests whether competition mobilizes voters by examining the effect of electoral competition on voter turnout. Columns (1) and (2) report the results from the OLS and 2SLS specifications. The estimated coefficients are small and precisely estimated, suggesting a null average effect on overall turnout. Panel A, Column (2) indicates that a one standard deviation increase in competition increases voter turnout by a negligible 0.06 percentage points (0.09% of the mean). The results for dispersion are qualitatively similar. This result stands in contrast to a number of studies that find a positive correlation between electoral competition and voter turnout in the United States (see, for example, Cox and Munger (1989)). Given that voter turnout is India is already relatively high at an average of 66%, it may be that most potential voters vote regardless of how competitive an election will be. This suggests that parties likely make strategic candidate and resource allocation decisions to persuade undecided voters in competitive constituencies, rather than increase turnout.

1.5.4 Incumbent Performance

I now consider the effects of electoral competition on incumbent performance in office. Com- petition can potentially affect incumbent performance through two main channels: (1) selection, with competition changing the type of politician elected, or (2) moral hazard, with anticipated competition changing incentives. Each of these effects is examined empirically below.

Selection

The preceding section on political selection shows that politicians elected in competitive constituencies differ on observables from politicians elected in non-competitive con- 36 1f "political leaders focus their efforts on the tight contests and forget about the cakewalks," voters in competitive constituencies are more likely to be mobilized by campaigns (Rosenstone and Hansen, 2003). Voter mobilization may also increase if competitive races attract greater media attention (Jackson (1996), for example, finds media coverage for competitive U.S. House races is greater than that for less competitive ones).

49 stituencies. Whether these observable (and any unobservable) differences have implications for governance, however, is unclear. To determine whether there is a selection effect on governance, I examine incumbent performance using three economic indicators: employment growth, change in night lights growth, and public goods provision as measured by new construction in the PMGSY rural roads program. As described in the data section, the empirical evidence suggests that the incumbent in an Indian state assembly constituency can have a significant impact on all three indicators. Nightlights growth and public goods provision are measured during an incumbent's full five-year term. Due to the unavailability of high frequency data, employment growth is measured over a seven-year time frame, therefore capturing a combination of an incumbent's performance and that of her successor. Table 10A, Panel A presents estimates of equation (10) for each economic indicator. Competition is measured by (1 minus the) winning margin at the time of an incumbent's election." I report both OLS and 2SLS estimates. The 2SLS estimates for each indicator are negative, but imprecisely estimated. Though I cannot reject the null of no effect for any indicator, the confidence intervals do not rule out large effect sizes. In particular, the upper bound of the negative effect size on each indicator from a one standard deviation increase in competition is 1.5 log points for employment growth; 7.9 percentage points for night lights growth; and 30.8% for length of all new roads.

Moral Hazard

The second channel through which competition can affect incumbent performance is moral hazard throughout the incumbent's term. An incumbent secure in her tenure may not be as motivated to exert effort when compared to the incumbent facing a high level of anticipated competition in the upcoming election. Table 10A, Panel B estimates the effect of anticipated competition using the expected winning margin between the incumbent and her strongest opponent in the upcoming election year. Here too, the results are imprecisely estimated. The moral hazard effect on employment growth is negative and similar in magnitude to the selection effect. I cannot rule out an effect size of a 1.7 log point decrease in

3 7 The first stage using the dispersion of vote shares to measure electoral competition is weaker in some of these samples due to the smaller number of observations.

50 employment growth from a one standard deviation increase in anticipated competition. In contrast to employment growth, the effect on night lights growth is positive. Column (4) shows the 95% confidence interval contains an effect size of 8.2 percentage points from a one standard deviation increase in anticipated competition. Finally, the moral hazard effect on public goods provision is large and negative (Columns (5) and (6)). However, I cannot infer the effect size from the 2SLS estimates because the first stage F-statistic is well below conventional levels in this restricted sample.

Discussion: These results, though imprecisely estimated, suggest that if anything electoral competition has a negative selection effect on incumbent performance and that moral hazard does little to improve incumbent performance. It is possible that incumbents react differently to anticipated competition depending on the level of competition at the time of their election. To check this, Table 10B estimates the effects of (1) electoral competition at the time of an incumbent's election and (2) anticipated electoral competition simultaneously. The results are largely consistent with Table 10A. Column (4) again shows that anticipated competition leads to a large, positive (though insignificant) increase in night lights growth. The night lights estimates here are of the same order of magnitude as those presented in Table 7, which estimates the increase in total lights emission in the year prior to an election.38 This provides suggestive evidence that any increase in night lights growth is taken close to an upcoming election, rather than throughout an incumbent's term. The selection results can be explained by two hypotheses. First, it may be that even though the candidates selected to run in competitive constituencies are more electable, they are not better politicians (at least when measured by observable economic indicators). The results on political selection suggest the politicians in competitive constituencies are wealthier - this wealth may be used during a campaign to ensure victory, but is not necessarily correlated with the ability to do well once in office. Second, elected politicians in competitive constituencies are less likely to have prior political experience (electoral competition increases the incumbent party's turnover rate, as shown in Table 3). It is possible that this inexperience leads to an initial negative effect on economic outcomes, but that eventually these politicians will perform better as they gain experience. 3 8 The point estimate in Table 10B is 0.058; in Table 7 it is 0.053.

51 There are also two stories to explain the moral hazard results. First, if voters are myopic, it would make sense for incumbents to save their effort until very close to an election when voters are more likely to pay attention and the electoral returns to effort are higher. One concern with this explanation, however, is that it assumes politicians can correctly forecast changes in competition. Given that my instrument relies on short-term changes in electoral competition, this may not be the case. A second story, therefore, is that politicians do respond to electoral competition when in office, but given uncertainty about future competition, they can only do so close to an election. Survey data on politicians' beliefs about anticipated competition could disentangle these explanations.

1.6 Conclusion

This paper estimates the effect of electoral competition on political selection and performance in Indian state elections. To identify the causal impact of electoral competition, I develop an empirical framework motivated by the observation that aggregate shifts to a party's popularity over time change the level of competition across constituencies non- monotonically depending on local party preferences. I use this observation to construct a shift-share instrument that leverages this non-monotonicity for identification. My main findings suggest that parties respond strategically to competition prior to an election. When a race becomes competitive, parties field candidates that are ostensibly higher quality, as indicated by their less corrupt background. Politicians in competitive constituencies are also wealthier. This result, combined with no significant effect of electoral competition on campaign funds, indicates that parties may rely on self-financing candidates to win competitive constituencies. I also find competitive constituencies experience a large, though temporary, increase in night light growth prior to an election, implying that the party in control of the state government reallocates state-run resources (e.g. electricity) strategically. Despite changes in campaign strategy prior to an election, however, I find no significant evidence that the politicians elected in competitive constituencies improve economic outcomes once in office, or improve performance in anticipation of future competition over the long-term - though the estimated effects have large confidence intervals. The findings in this paper open a number of avenues for future work. First, a

52 key contribution of this paper is the proposal of an identification strategy that can be easily used to study electoral competition in other countries with first-past-the-post electoral systems. Understanding how the effects of electoral competition vary across countries (developed versus developing countries, less autocratic versus more autocratic countries, etc.) is important for testing the external validity of the results presented here, and for developing richer models of electoral competition that take into account institutional structure. Second, the results in this paper suggest parties may make strategic trade-offs between candidates and resources when campaigning in competitive constituencies. Better data on campaign financing can shed light on this trade-off and the implications for political selection. Finally, this paper focuses on short-term changes to electoral competition. A long-term change to electoral competition - a result, perhaps, of gerrymandering or political polarization - may have very different consequences for political selection and performance and merits further investigation.

53 References

Acemoglu, Daron, Giuseppe De Feo, and Giacomo De Luca. 2017. "Weak States: Causes and Consequences of the Sicilian Mafia,". NBER Working Paper No. 24115.

Acemoglu, Daron, Tristan Reed, and James A. Robinson. 2014. "Chiefs: Economic Devel- opment and Elite Control of Civil Society in Sierra Leone", Journal of Political Economy, 122(2), 319-368.

Arvate, Paulo Roberto. 2013. "Electoral Competition and Local Government Responsive- ness in Brazil." World Development 43, 67-83.

Asher, Sam, and Paul Novosad. 2017. "Politics and Local Economic Growth: Evidence from India", American Economic Journal: Applied Economics 9 (1), 229-73.

Baleiras, Rui and Jose Costa. 2004. "To be or not to be in office again: an empirical test of a local political business cycle rationale", European Journal of Political Economy, 20(3), 665-671.

Banerjee, Abhijit, Esther Duflo, Clement Imbert, and Rohini Pande, 2017. "Entry, Exit and Candidate Selection: Evidence from India", Working Paper.

Banerjee, Abhijit and Rohini Pande. 2018. "Parochial Politics: Ethnic Preferences and Politician Corruption," Working Paper.

Baskaran, Thushyanthan and Brian Min. 2015. "Election cycles and electricity provision: Evidence from a quasi-experiment with Indian special elections," 126, 64-73.

Besley, Timothy, Torsten Persson, and Daniel M. Sturm. 2010. "Political Competition, Pol- icy and Growth: Theory and Evidence from the US", The Review of Economic Studies, 77(4), 1329-1352.

Borusyak, K., Hull, P., and Jaravel, X. 2018. "Quasi-experimental shift-share research de- signs," NBER Working Paper 24997.

54 Carrillo, Juan D. and Thomas Mariotti, 2001. "Electoral Competition and Politicians Turnover", European Economic Review 45, 1-25.

Caselli, F. and M. Morelli 2004. "Bad Politicians", Journal of Public Economics 88 (3- 4),759?782.

Cole, Shawn, 2009. "Fixing Market Failures or Fixing Elections", American Economic Jour- nal: Applied Economics, 1(1), 219-250.

Cox, Gary 1997. Making Votes Count. Cambridge.

Dal B6 , Ernesto and Frederico Finan. 2018. "Progress and Perspectives in the Study of Political Selection", NBER Working Paper 24783.

Dal B6 , Ernesto, Frederico Finan, Olle Folke, Torsten Persson, Johanna Rickne. 2017. "Who Becomes A Politician?", The Quarterly Journal of Economics, 134(4): 1877-1914.

De Paola, Maria and Vincenzo Scoppa. 2011. "Political competition and politician quality: evidence from Italian municipalities", Public Choice, 148(3), 547-559.

Ferraz, C. and F. Finan 2009. "Motivating politicians: The impacts of monetary incentives on quality and performance," NBER Working Papers 14906.

Fey, Mark. 2007, "Duverger's Law Without Strategic Voting", Mimeo. University of Rochester.

Galasso, Vincenzo and Tommaso Nannicini. 2011. "Competing on Good Politicians." Amer- ican Political Science Review,105(1), 79-99.

Goldsmith-Pinkham, Paul, Isaac Sorkin, and Henry Swift. 2018. "Bartik Instruments: What, When, Why, and How", Working paper, National Bureau of Economic Research.

Keane, M. P. and A. Merlo 2010. "Money, political ambition, and the career decisions of politicians," American Economic Journal: Microeconomics, 2(3), 186?215.

Khemani, Stuti. 2004. "Political Cycles in a Developing Economy: Effect of Elections in the Indian States", Journal of Development Economics 73, 125-154.

55 Kosec, Katrina, Hamza Haider, David J. Spielman, and Fatima Zaidi (2015). "The effects of

political competition on rural land: evidence from Pakistan," IFPRI Discussion Paper, 1441.

Lehne, Jonathan, Jacob N. Shapiro and Oliver Vanden Eynde. 2016. "Building Connections: Political Corruption and Road Construction in India." Working Paper.

Linden, Leigh, 2004. "Are Incumbents Really Advantaged? The Preference for Non- Incumbents in Indian National Elections." Working Paper.

Marx, Benjamin. 2014. "The Causal Effects of Close Elections". Mimeo, MIT.

Min, Brian and Miriam Golden, 2014. "Electoral cycles in electricity losses in India", Energy Policy 65, 619-625.

Nath, Anusha. 2015. "Bureaucrats and Politicians: How Does Electoral Competition Affect Bureaucratic Performance?? Institute for Economic Development (IED) Working Paper 269.

Paola, Maria de and Vincenzo Scoppa. 2011. "Political competition and politician quality: Evidence from Italian municipalities." Public Choice, 148 (3-4), 547-559.

Persson, Torsten and Guido Tabellini. 2000. Political Economics. MIT Press.

Roberson, Brian. 2006. "The Colonel Blotto game," Economic Theory 29: 1-24.

Stigler, George J. 1972. "Economic Competition and Political Competition", Public Choice, 13(1), 91-106.

Sukhtankar, Sandip and Milan Vaishnav. 2015. "Corruption in India: Bridging Research Evidence and Policy Options." Brookings-NCAER India Policy Forum 2014.

Svaleryd, Helena and Jonas Vlachos. 2009. "Political rents in a non-corrupt democracy", Journal of Public Economics, 93(3-4), 355-372.

Wittman, Donald. 1989. "Why Democracies Produce Efficient Results," Journal of Political Economy, 97(6): 1395?1424.

56 Wittman, Donald. 1995. The Myth of Democratic Failure: Why Political Institutions are Efficient (Chicago: University of Chicago Press).

57 Figure 1: Conceptual Framework

Campaign Strategy Incumbent Performance Short-term Moral Hazard Long-term Moral Hazard

Election, time t Election, time t+1

Incumbent Performance Selection

Note: This figure provides an overview of the conceptual framework linking electoral competition to politician selection and performance. In the campaign period, political parties allocate candidates and resources across constituencies to maximize the number of seats they win in the upcoming election. Strategic decisions taken during this time are motivated by "short-term moral hazard" concerns. In the post-election period, electoral competition may affect incumbent performance through two channels: (1) a "selection" channel, if competition changes the type of politician elected, or (2) a "long-term moral hazard" channel, if anticipated competition changes performance incentives over the long-term.

58 Figure 2A: Electoral Competition in Indian State Elections (1)

U, C4J -

0) 'U

1950 1960 1970 1980 1990 2000 2010 Year

- -- North India - East India ---- South India - Western India

Note: This figure plots the average winning margin between the winner and runner-up in Indian state elections by region across decades.

Figure 2B: Electoral Competition in Indian State Elections (2)

ILQ

O) UD .. . . 0 . . C 0 aN4 - 0 a.

'-1 1950 1960 1970 1980 1990 2000 2010 Year

0 Competitive (Victory Margin < 0.05) 0 Lopsided (Victory Margin >0.20)

Note: This figure plots the proportion of state assembly seats won in competitive races (victory margin < 0.05) and lopsided races (victory margin > 0.20) across decades.

59 Figure 2C: Vote Shares in Select States (2007)

-- O- BJP INC IND BJP INC IND SAD BJP BSP INC IND Gujarat Punjab Uttar Pradesh

Note: This figure plots vote shares for major parties in the 2007 state elections. BJP is the Bharatiya Janata Party; INC is the Indian National Congress; IND are independent candidates; SAD is the Shiromani Akali Dal party; and BSP is the Bahujan Samaj Party.

60 Figure 3A: Identification Strategy, Two Party Example

Competition = I - Vote-Margin Gap

0 .2 .4 .6 .8 1 Constituency Preferences (0 = Left, 1 = Right)

Note: This figure plots competition as a function of constituency policy preferences. Competition is defined (0.5 to as 1 minus the winning margin. Constituency policy preferences range from left (0 to 0.5) to right 1).

Figure 3B: Negative Aggregate Shock to Right-Wing Party

"'I I

I I c'~J I I

C4 - I I

0 . o .2 .4 .6 .8 1 Constituency Preferences (0 = Left, 1 = Right)

S Competition, After Shock 0 Competition, Before Shock

Note: This figure plots competition as a function of constituency policy preferences. Competition is defined to as 1 minus the winning margin. Constituency policy preferences range from left (0 to 0.5) to right (0.5 near the 1). The blue curve plots competition after a negative shock to the right-wing party. Constituencies median which were once competitive become less so as the left-wing party becomes relatively more attractive. Constituencies which were once the right-wing party's stronghold become more competitive.

61 Figure 4: Actual vs Predicted Competition in Karnataka

Panel A: 1989 PanelB: 1994

S :q 0 .O I .

0- 0

0 0 o .2 .4 .6I .8 1 0 .2 .4 .6 I .8 1 Initial Vote Shares, INC Initial Vote Shares, INC

Panel C: 1999 Panel D: 2004 * Go- I. 0 90V rCq eS 0 0 * ww.1 V:- 0 0- E S 8 0 001 0%1- 0 0 i i i I I I I 1 1 0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1 Initial Vote Shares, INC Initial Vote Shares, INC

0 Actual Competition, BJP vs INC 0 Predicted Competition, BJP vs INC

Note: This figure shows a scatterplot of actual and predicted competition between BJP and INC across election years in Karnataka. Competition is plotted as a function of initial vote shares for INC in the baseline year, 1989. The initial vote share is rescaled to indicate the level of support for INC in comparison to BJP. Panel A displays the mechanical relationship between competition and INC vote shares in the baseline year. Panel B shows that competition shifts towards the right in 1994, as the BJP becomes a more competitive party vis-a-vis INC. Panel C shows competition shifts back towards the left, indicating a positive shift for INC in 1999. Finally, Panel D shows the BJP once again becomes competitive, threatening INC strongholds.

62 Figure 5A: Candidate Selection, Demographics (All Candidates)

Panel A: Competition Panel B: Dispersion

Criminal Record - - - 0

Education ------p

Ln. Net Assets -

Female - -+

Age - p *i

0 I I I -.5 0 .5 -2 -1 o 1 2 3

Note: This figure presents estimates of the effect of competition and dispersion on candidate demographics. Competition is defined as 1 minus the expected margin between the incumbent and her strongest opponent. Dispersion is defined as 1 minus the HHI. Each row on the y-axis reports the treatment effect and 95% confidence interval based on robust standard errors. Each specification includes party fixed effects, baseline year competition controls, and state-year fixed effects. Sample period is 2004-2017. Constituencies in 2004-2007 and 2008-2017 are stacked since the 2008 delimitation prevents comparison across these periods. Dependent variables are normalized by their mean and standard deviations to allow for comparison across variables. Tables 4A-4C and Appendix Tables 3A-3B report corresponding coefficients.

63 Figure 5B: Candidate Selection, Demographics (Incumbent and Opposition Party Candidates)

Panel A: Competition Panel B: Dispersion

p Criminal Record - p

Education - p

Ln. Net Assets - -_

i-.-- Female - I o

- 0 i -- I Age - -4-_

-2 -1 0 1 2 3 -6 -4 -2 0 2 4

0 Incumbent Party * Opposition Party

Note: This figure presents estimates of the effect of competition and dispersion on candidate demographics separately for the candidate selected by the incumbent party and the (strongest) opposition party. Com- petition is defined as 1 minus the expected margin between the incumbent and her strongest opponent. Dispersion is defined as 1 minus the HHI. Each row on the y-axis reports the treatment effect and 95% confidence interval based on robust standard errors. Each specification includes party fixed effects, baseline year competition controls, and state-year fixed effects. Sample period is 2004-2017. Constituencies in 2004-2007 and 2008-2017 are stacked since the 2008 delimitation prevents comparison across these periods. Dependent variables are normalized by their mean and standard deviations to allow for comparison across variables. Tables 4A-4C and Appendix Tables 3A-3B report corresponding coefficients.

64 Figure 6: Winner Selection, Demographics (All Elected Politicians)

Panel A: Competition Panel B: Dispersion

Criminal Record1f -- +

Education - 1 0

Ln. Net Assets - p

Female - -- 0--

Age - 'p

-1 -.5 0 .5 1 4 -2 0 2 4

Note: This figure presents estimates of the effect of competition and dispersion on elected politician demographics. Competition is defined as 1 minus the expected margin between the incumbent and her strongest opponent. Dispersion is defined as 1 minus the HHI. Each row on the y-axis reports the treatment effect and 95% confidence interval based on robust standard errors. Each specification includes party fixed effects, baseline year competition controls, and state-year fixed effects. Sample period is 2004-2017. Constituencies in 2004-2007 and 2008-2017 are stacked since the 2008 delimitation prevents comparison across these periods. Dependent variables are normalized by their mean and standard deviations to allow for comparison across variables. Tables 5 reports corresponding coefficients.

65 Table 1: Predicted Vote Shares

Dep. Variable: Party Vote Share (1) (2) (3) (4) Predicted Vote Share, BJP 0.470*** (0.150) Predicted Vote Share, INC 0.600*** (0.085) Predicted Vote Share, State Party 1 0.888*** (0.058) Predicted Vote Share, State Party 2 0.732*** (0.113) Mean dep. var. 0.27 0.38 0.20 0.14 R2 0.78 0.74 0.81 0.79 Number of clusters 3013 3013 3013 3013 Obs 9425 9425 9425 9425 Notes: All regressions contain constituency and state-year fixed effects. Standard errors clustered by constituency. BJP is the Bharatiya Janata Party; INC is the Indian National Congress; and State Party 1 and 2 are the parties with the top vote shares (excluding INC and BJP) in each state. Party vote shares are re-scaled to sum to 1. Sample years are 1989-2008, excluding baseline election year. * p<0.10, ** p<0.05, *** p<0.01.

66 Table 2: First Stage

Panel A: Dep. Variable: Competition (1) (2) (3) (4) Predicted Competition 0.331*** 0.324*** 0.323*** 0.326*** (0.027) (0.027) (0.027) (0.027) Predicted Vote Share, BJP -0.239 -0.238 -0.049 (0.147) (0.145) (0.188) Predicted Vote Share, INC -0.291*** -0.114 (0.095) (0.167) Predicted Vote Share, State Party 1 0.205 (0.140) F-stat excluded instrument 149.3 141.5 142.9 147.1 R2 0.65 0.65 0.65 0.65 Number of clusters 3013 3013 3013 3013 Obs 9425 9425 9425 9425 Panel B: Dep. Variable: Dispersion (1) (2) (3) (4) Predicted Dispersion 0.183*** 0.165*** 0.163*** 0.171*** (0.035) (0.035) (0.035) (0.035) Predicted Vote Share, BJP -0.260*** -0.259*** -0.115 (0.088) (0.086) (0.103) Predicted Vote Share, INC -0.295*** -0.161* (0.061) (0.096) Predicted Vote Share, State Party 1 0.154* (0.079) F-stat excluded instrument 27.7 22.1 21.4 23.2 R2 0.68 0.68 0.68 0.68 Number of clusters 3013 3013 3013 3013 Obs 9425 9425 9425 9425 Notes: All regressions contain constituency and state-year fixed effects. Standard errors clustered by constituency. Competition is defined as 1 minus the winning margin. Dispersion is defined as 1 minus the HHI. BJP is the Bharatiya Janata Party; INC is the Indian National Congress; and State Party 1 and 2 are the parties with the top vote shares (excluding INC and BJP) in each state. Party vote shares are re-scaled to sum to 1. Sample years are 1989-2008, excluding baseline election year. * p<0.10, ** p<0.05, *** p<0.01.

67 Table 3: Candidate Turnover

Dep Variable: New Candidate All Parties Incumb. Party Opp. Party Other Parties (1) (2) (3) (4) (5) OLS 2SLS 2SLS 2SLS 2SLS Panel A: Competition -0.040*** -0.074 0.630 0.001 -0.009 (0.011) (0.054) (0.467) (0.109) (0.064)

Mean dep. var. 0.75 0.75 0.39 0.77 0.92 Controls Yes Yes Yes Yes Yes Party Fixed Effects Yes Yes Yes Yes Yes F-stat excluded instrument 221.28 18.06 214.74 152.00 Number of clusters 3013 3010 2370 2466 2826 Obs 30183 30180 6811 7246 15102 (1) (2) (3) (4) (5) OLS 2SLS 2SLS 2SLS 2SLS Panel B: Dispersion 0.006 0.371 1.829** -1.016 -0.373 (0.021) (0.284) (0.828) (0.759) (0.334)

Mean dep. var. 0.75 0.75 0.39 0.77 0.92 Controls Yes Yes Yes Yes Yes Party Fixed Effects Yes Yes Yes Yes Yes F-stat excluded instrument 29.37 20.12 15.15 17.91 Number of clusters 3013 3010 2370 2466 2826 Obs 30183 30180 6811 7246 15102 Notes: This table estimates the effect of electoral competition on whether a candidate running for office also ran in the previous election in the same constituency. Competition is defined as 1 minus the expected margin between the incumbent and her strongest opponent. Dispersion is defined as 1 minus the HHI. All regressions contain constituency and state-year fixed effects. Standard errors clustered by constituency. Vote shares rescaled to sum to 1.

68 Table 4A: Selection of Candidates, Criminal Record

Dep Variable: Candidate Has Criminal Record All Parties Incumb. Party Opp. Party Other Parties (1) (2) (3) (4) (5) OLS 2SLS 2SLS 2SLS 2SLS Panel A: Competition -0.019 -0.155*** -0.208 -0.242** -0.192** (0.018) (0.060) (0.253) (0.115) (0.086)

Mean dep. var. 0.27 0.27 0.31 0.30 0.23 Controls Yes Yes Yes Yes Yes Party Fixed Effects Yes Yes Yes Yes Yes F-stat excluded instrument 712.29 40.35 231.71 320.30 Obs 7879 7879 1852 2103 3971 (1) (2) (3) (4) (5) OLS 2SLS 2SLS 2SLS 2SLS Panel B: Dispersion -0.020 -0.532** -0.548 -1.328** -0.408 (0.032) (0.234) (0.459) (0.520) (0.309)

Mean dep. var. 0.27 0.27 0.31 0.30 0.23 Controls Yes Yes Yes Yes Yes Party Fixed Effects Yes Yes Yes Yes Yes F-stat excluded instrument 142.92 35.65 39.75 70.78 Obs 7879 7879 1852 2103 3971 Notes: This table estimates the effect of electoral competition on the criminal record of a candidate at the time of declaring candidacy. Crime is defined as 1 if the candidate has any criminal record, and 0 otherwise. Competition is defined as 1 minus the expected margin between the incumbent and her strongest opponent. Dispersion is defined as 1 minus the HHI. Specifications include party fixed effects, baseline year competition controls, and state-year fixed effects. Robust standard errors. Sample period is 2004-2017. Constituencies in 2004-2007 and 2008-2017 are stacked since the 2008 delimitation prevents comparison across these periods. Vote shares rescaled to sum to 1.

69 Table 4B: Selection of Candidates, Educational Attainment

Dep Variable: Candidate Education All Parties Incumb. Party Opp. Party Other Parties (1) (2) (3) (4) (5) OLS 2SLS 2SLS 2SLS 2SLS Panel A: Competition -0.041 0.429** -0.445 0.303 0.569* (0.061) (0.212) (0.836) (0.340) (0.339) Mean dep. var. 5.24 5.24 5.42 5.39 5.08 Controls Yes Yes Yes Yes Yes Party Fixed Effects Yes Yes Yes Yes Yes F-stat excluded instrument 675.61 34.46 219.24 304.61 Obs 7535 7535 1795 2017 3770 (1) (2) (3) (4) (5) OLS 2SLS 2SLS 2SLS 2SLS Panel B: Dispersion -0.091 2.353*** -0.832 2.428* 2.837** (0.107) (0.848) (1.337) (1.439) (1.324)

Mean dep. var. 5.24 5.24 5.42 5.39 5.08 Controls Yes Yes Yes Yes Yes Party Fixed Effects Yes Yes Yes Yes Yes F-stat excluded instrument 135.28 34.59 39.86 64.00 Obs 7535 7535 1795 2017 3770 Notes: This table estimates the effect of electoral competition on the educational attainment of a candidate at the time of declaring candidacy. Education is coded on a 0 to 7 scale, ranging from illiterate to post-graduate education. Competition is defined as 1 minus the expected margin between the incumbent and her strongest opponent. Dispersion is defined as 1 minus the HHI. Specifications include party fixed effects, baseline year competition controls, and state-year fixed effects. Robust standard errors. Sample period is 2004-2017. Constituencies in 2004-2007 and 2008-2017 are stacked since the 2008 delimitation prevents comparison across these periods. Vote shares rescaled to sum to 1.

70 Table 4C: Selection of Candidates, Net Assets

Dep Variable: Candidate Ln. Net Assets All Parties Incumb. Party Opp. Party Other Parties (1) (2) (3) (4) (5) OLS 2SLS 2SLS 2SLS 2SLS Panel A: Competition 0.026 0.347 -0.363 0.226 -0.321 (0.065) (0.226) (0.863) (0.366) (0.355) Mean dep. var. 15.73 15.73 16.21 16.19 15.20 Controls Yes Yes Yes Yes Yes Party Fixed Effects Yes Yes Yes Yes Yes F-stat excluded instrument 688.56 38.33 212.23 303.20 Obs 7290 7290 1805 2023 3506

(1) (2) (3) (4) (5) OLS 2SLS 2SLS 2SLS 2SLS Panel B: Dispersion 0.097 0.910 -1.518 1.344 -0.881 (0.112) (0.900) (1.593) (1.804) (1.276)

Mean dep. var. 15.73 15.73 16.21 16.19 15.20 Controls Yes Yes Yes Yes Yes Party Fixed Effects Yes Yes Yes Yes Yes F-stat excluded instrument 130.99 33.80 31.13 66.84 Obs 7290 7290 1805 2023 3506 Notes: This table estimates the effect of electoral competition on (log) net assets. A candidate's net assets are defined as assets net of liabilities at the time of declaring candidacy. Competition is defined as 1 minus the expected margin between the incumbent and her strongest opponent. Dispersion is defined as 1 minus the HHI. Specifications include party fixed effects, baseline year competition controls, and state-year fixed effects. Robust standard errors. Sample period is 2004-2017. Constituencies in 2004-2007 and 2008-2017 are stacked since the 2008 delimitation prevents comparison across these periods. Vote shares rescaled to sum to 1.

71 Table 5: Selection of Elected Politicians

Dependent Variable: Criminal Record Education Net Assets Age Gender (1) (2) (3) (4) (5) 2SLS 2SLS 2SLS 2SLS 2SLS Panel A: Competition -0.260** 0.273 1.089*** 1.950 0.013 (0.130) (0.371) (0.386) (3.057) (0.047)

Mean dep. var. 0.34 5.42 16.16 50.63 0.06 Controls Yes Yes Yes Yes Yes F-stat excluded instrument 171.52 168.20 161.20 171.52 226.70 Obs 2264 2207 2190 2264 7437 (1) (2) (3) (4) (5) 2SLS 2SLS 2SLS 2SLS 2SLS Panel B: Dispersion -0.684 1.559 2.872* 7.043 0.048 (0.506) (1.411) (1.677) (12.726) (0.202)

Mean dep. var. 0.34 5.42 16.16 50.63 0.06 Controls Yes Yes Yes Yes Yes F-stat excluded instrument 36.23 36.28 34.06 36.23 40.32 Obs 2264 2207 2190 2264 7437 Notes: This table estimates the effect of electoral competition on winner characteristics. Competition is defined as 1 minus the expected margin between the incumbent and her strongest opponent. Dispersion is defined as 1 minus the HHI. Specifications in Columns (1)-(4) include baseline year competition controls, and state-year fixed effects. Robust standard errors. Sample period is 2004-2017. Constituencies in 2004-2007 and 2008-2017 are stacked since the 2008 delimitation prevents comparison across these periods. Specification in Column (5) includes constituency and state-year fixed effects. Standard errors clustered by constituency. All vote shares rescaled to sum to 1.

72 Table 6A: Official Campaign Financing of Elected Politicians

Total Expenses Own Funds Party Funds Other Funds All Parties Incumb. Parties Opp. Parties All Parties Incumb. Parties Opp. Party All Parties Incumb. Parties Opp. Parties All Parties Incumb. Parties Opp. Parties (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Panel A: Competition -0.034 -0.112 0.134 -0.426 -0.607 -0.085 -0.649 -3.831 -0.881 0.257 0.848 0.827 (0.113) (0.200) (0.121) (0.938) (2.230) (0.976) (1.342) (2.588) (1.561) (1.363) (2.598) (1.534)

Mean dep. var. 13.94 13.93 13.95 5.31 5.10 5.42 5.21 4.03 5.82 7.03 6.14 7.49 Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes F-stat excluded instrument 268.99 88.21 205.70 268.99 88.21 205.70 268.99 88.21 205.70 268.99 88.21 205.70 Obs 1216 414 802 1216 414 802 1216 414 802 1216 414 802 (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Panel B: Dispersion -0.050 -0.338 0.577 -2.369 -6.724 -0.262 -5.589 -21.136** -6.149 1.289 1.734 5.462 (0.432) (0.682) (0.435) (2.747) (6.269) (3.197) (4.009) (9.087) (4.927) (4.210) (8.472) (4.928)

Mean dep. var. 13.94 13.93 13.95 5.31 5.10 5.42 5.21 4.03 5.82 7.03 6.14 7.49 Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes F-stat excluded instrument 79.26 25.23 53.83 79.26 25.23 53.83 79.26 25.23 53.83 79.26 25.23 53.83 Obs 1216 414 802 1216 414 802 1216 414 802 1216 414 802 Notes: This table reports 2SLS estimates of the effect of electoral competition on campaign financing for all winning candidates, winning candidates from the incumbent party, and winning candidates from non-incumbent parties. All dependent variables are in logs. Competition is defined as 1 minus the expected margin between the incumbent and her strongest opponent. Dispersion is defined as 1 minus the HHI Specifications include baseline year competition controls, and state-year fixed effects. Robust standard errors. Sample period is 2008-2017. Vote shares rescaled to sum to 1. Table 6B: Official Campaign Expenditure of Elected Politicians

Meetings, Materials, Transport Media Party Visits All Parties Incumb. Parties Opp. Parties All Parties Incumb. Parties Opp. Party All Parties Incumb. Parties Opp. Parties

(1) (2) (3) (4) (5) (6) (7) (8) (9) Panel A: Competition -2.434** 0.892 -1.462 1.157 1.766 1.668 -0.305 -1.489* 0.262 (1.009) (1.300) (1.217) (1.244) (2.358) (1.452) (0.359) (0.869) (0.256)

Mean dep. var. 10.19 13.29 9.09 7.52 7.98 7.29 0.07 0.10 0.05 Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes F-stat excluded instrument 307.37 88.21 215.19 268.99 88.21 205.70 268.99 88.21 205.70 Obs 1580 414 1166 1216 414 802 1216 414 802

(1) (2) (3) (4) (5) (6) (7) (8) (9) Panel B: Dispersion -9.935*** 4.477 -7.139** -0.212 4.585 0.154 -1.111 -4.379 0.607 -1 (2.836) (4.608) (3.525) (3.882) (7.796) (4.403) (1.235) (3.303) (0.578)

Mean dep. var. 10.19 13.29 9.09 7.52 7.98 7.29 0.07 0.10 0.05 Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes F-stat excluded instrument 108.21 25.23 63.25 79.26 25.23 53.83 79.26 25.23 53.83 Obs 1580 414 1166 1216 414 802 1216 414 802 Notes: This table reports 2SLS estimates of the effect of electoral competition on campaign expenditure for all winning candidates, winning candidates from the incumbent Darty , and winning candidates from non-incumbent parties. All dependent variables are in logs. Competition is defined as 1 minus the expected margin between the incumbent and her strongest opponent. Dispersion is defined as 1 minus the HHIL. Specifications include baseline year competition controls, and state-year fixed effects. Robust standard errors. Sample period is 2008-2017. Vote shares rescaled to sum to 1. Table 7: State Resource Allocation, Power Supply

Dep Var: Ln Lights, Annual Growth (1) (2) (3) (4) OLS 2SLS OLS 2SLS Competition 0.007 -0.094*** 0.019 -0.037 (0.008) (0.025) (0.013) (0.032) Competition x Pre-Election Year -0.013 0.147*** -0.032* 0.061 (0.011) (0.027) (0.019) (0.048) Competition x Aligned -0.019 -0.129*** (0.015) (0.044) Competition x Pre-Election Year x Aligned 0.032 0.151*** (0.023) (0.057)

Pre-Election Year x Aligned -0.029* -0.102*** (0.015) (0.034) Aligned 0.017* 0.084*** (0.010) (0.026) Pre-Election Year -0.200*** -0.264*** -0.187*** -0.210*** (0.016) (0.018) (0.019) (0.028) P-val, 41 + 32 = 0 0.447 0.026 0.386 0.579 P-val, 41 + 32 +13 + 4 =0 0.925 0.182 Mean dep. var. 0.002 0.002 0.001 0.001 Baseline dep. var Yes Yes Yes Yes Controls Yes Yes Yes Yes Number of clusters 2300 2300 2258 2258 Obs 9516 9516 9430 9430 Notes: This table estimates the effect of electoral competition on power supply measured by annual log growth in total light emissions before and after an election. Sample is restricted to the years preceding and following an election year. Aligned constituencies are those where the incumbent is a member of the party or coalition in control of the state government. Competition is defined as 1 minus the expected margin between the incumbent and her strongest opponent. All regressions contain constituency and state-year fixed effects. Standard errors clustered by constituency. Vote shares rescaled to sum to 1.

75 Table 8: Party Entry and Exit

Number of: Parties Major Parties Major Parties, Other Small Parties Independents (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) OLS 2SLS OLS 2SLS OLS 2SLS OLS 2SLS OLS 2SLS Panel A: Competition 0.298*** -0.280 0.844*** 0.518*** 0.151*** 0.046 -0.552*** -0.802** -1.201* -2.565 (0.083) (0.401) (0.045) (0.186) (0.049) (0.206) (0.072) (0.354) (0.716) (3.023)

Mean dep. var. 5.35 5.35 3.19 3.19 1.59 1.59 2.16 2.16 5.92 5.92 Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes F-stat excluded instrument 217.87 217.69 217.26 216.71 216.64 Number of clusters 3013 2879 3013 2873 3013 2879 3013 2879 3013 2879 Obs 9425 9291 9417 9277 9425 9291 9425 9291 9425 9291

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) OLS 2SLS OLS 2SLS OLS 2SLS OLS 2SLS OLS 2SLS Panel B: Dispersion 0.928*** -1.197 2.001*** 1.986*** 0.907*** 6.289*** -1.084*** -3.073* -1.223 3.805 (0.157) (1.827) (0.082) (0.757) (0.092) (1.184) (0.135) (1.600) (1.110) (9.469)

Mean dep. var. 5.35 5.35 3.19 3.19 1.59 1.59 2.16 2.16 5.92 5.92 Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes F-stat excluded instrument 34.90 37.94 34.71 34.18 34.22 Number of clusters 3013 2879 3013 2873 3013 2879 3013 2879 3013 2879 Obs 9425 9291 9417 9277 9425 9291 9425 9291 9425 9291 Notes: This table estimates the effect of electoral competition on party entry and exit. The dependent variables are the number of parties, major parties, major parties excluding the incumbent and opposition party, small parties and independent candidates. Major parties are defined as INC, BJP, and the top two parties in each state (exclude INC and BJP). Major parties - other are defined as the major parties excluding the incumbent party and the strongest opposition party. Small parties are defined as all remaining parties. All regressions contain constituency and state-year fixed effects. Competition is defined as 1 minus the expected margin between the incumbent and her strongest opponent. Dispersion is defined as 1 minus the HHI. Standard errors clustered by constituency. Vote shares rescaled to sum to 1. Table 9: Voter Mobilization

Dep. Variable: Voter Turnout (1) (2) OLS 2SLS Panel A: Competition 0.000 0.002 (0.003) (0.013)

Mean dep. var. 0.66 0.66 Controls Yes Yes F-stat excluded instrument 199.54 Number of clusters 2902 2795 Obs 9046 8939 (1) (2) OLS 2SLS Panel B: Dispersion -0.006 0.019 (0.005) (0.054)

Mean dep. var. 0.66 0.66 Controls Yes Yes F-stat excluded instrument 36.38 Number of clusters 2902 2795 Obs 9046 8939

Notes: This table estimates the effect of electoral competition on voter turnout. Competi- tion is defined as 1 minus the expected margin between the incumbent and her strongest opponent.

Dispersion is defined as 1 minus the HHL Sample period is 1989-2008, excluding baseline election year. All regressions contain constituency and state-year fixed effects. Standard errors clustered by constituency. Vote shares rescaled to sum to 1.

77 Table 1OA: Incumbent Performance, Economic Indicators

Ln. Employment Growth Ln. Lights Growth Ln. Total New Road Length (1) (2) (3) (4) (5) (6) OLS 2SLS OLS 2SLS OLS 2SLS Panel A: Competition 0.006 -0.012 -0.012 -0.109 -0.024 -0.451 (0.005) (0.020) (0.020) (0.079) (0.108) (0.293)

Mean dep. var. 0.03 0.03 0.04 0.04 3.10 3.10 Baseline dep. variable Yes Yes Yes Yes No No F-stat excluded instrument 56 111 107 Number of clusters 2271 1732 2187 2096 2134 1610

__4 Obs 4003 3464 5046 4955 3839 3315 Go (1) (2) (3) (4) (5) (6) OLS 2SLS OLS 2SLS OLS 2SLS Panel B: Anticipated Competition -0.005 -0.017 0.024 0.100 -0.976 -1.250 (0.013) (0.020) (0.022) (0.092) (1.717) (2.036)

Mean dep. var. 0.02 0.02 0.03 0.03 2.67 2.67 Baseline dep. variable Yes Yes Yes Yes No No F-stat excluded instrument 45 140 5 Number of clusters 2132 641 2112 1891 1701 130 Obs 2773 1282 4378 4157 1831 260 Notes: This table estimates the effect of electoral competition on incumbent performance using three economic indicators: employment growth, night lights growth, and public goods provision. Competition is defined as 1 minus the margin between the incumbent and her strongest opponent in the year the incumbent was elected. Anticipated competition is defined as 1 minus the expected margin between the incumbent and her strongest opponent in the upcoming election year. All regressions contain constituency and state-year fixed effects. Standard errors clustered by constituency. Vote shares rescaled to sum to 1. Table 10B: Incumbent Performance, Economic Indicators

Ln. Employment Growth Ln. Lights Growth Ln. Total New Road Length (1) (2) (3) (4) (5) OLS 2SLS OLS 2SLS OLS Panel C: Competition -0.004 -0.025 -0.006 -0.103 -0.001 (0.015) (0.023) (0.026) (0.110) (1.375) Anticipated Competition -0.007 -0.024 0.021 0.058 -0.910 (0.015) (0.021) (0.024) (0.112) (1.854) __4 P-val, /1 = #2 0.83 0.98 0.33 0.10 0.65 Mean dep. var. 0.02 0.02 0.03 0.03 2.67 Baseline dep. variable Yes Yes Yes Yes No F-stat excluded instruments 26 24 Number of clusters 2131 640 2111 1889 1700 Obs 2771 1280 4375 4153 1830 Notes: This table estimates the effect of electoral competition on incumbent performance using three economic indicators: employment growth, night lights growth, and public goods provision. Competition is defined as 1 minus the margin between the incumbent and her strongest opponent in the year the incumbent was elected. Anticipated competition is defined as 1 minus the expected margin between the incumbent and her strongest opponent in the upcoming election year. All regressions contain constituency and state-year fixed effects. Standard errors clustered by constituency. Vote shares rescaled to sum to 1. APPENDIX A

Appendix Table 1: Average Standard Deviation of Competition (within constituency, across years)

Competition Predicted Competition Dispersion Predicted Dispersion

(1) (2) (3) (4) Andhra Pradesh 0.21 0.08 0.09 0.02 Assam 0.20 0.11 0.10 0.05 Bihar 0.25 0.17 0.16 0.08 Delhi 0.11 0.01 0.04 0.01 Goa 0.23 0.12 0.13 0.04 Gujarat 0.18 0.04 0.06 0.01 Haryana 0.39 0.21 0.23 0.11 Himachal Pradesh 0.18 0.09 0.04 0.02 Karnataka 0.29 0.19 0.17 0.07 Kerala 0.18 0.29 0.09 0.17 Madhya Pradesh 0.11 0.07 0.04 0.01 Maharashtra 0.28 0.09 0.16 0.04 Manipur 0.36 0.20 0.22 0.08 Meghalaya 0.14 0.08 0.07 0.04 Orissa 0.28 0.15 0.10 0.08 Pondicherry 0.25 0.12 0.15 0.10 Punjab 0.18 0.09 0.06 0.02 Rajasthan 0.15 0.09 0.05 0.03 Sikkim 0.24 0.23 0.10 0.06 Tamil Nadu 0.30 0.16 0.16 0.04 Tripura 0.06 0.02 0.05 0.03 Uttar Pradesh 0.24 0.15 0.13 0.05 West Bengal 0.26 0.25 0.24 0.19 Notes: This table reports the (1) average standard deviation of competition (actual and predicted) and (2) average standard deviation of dispersion (actual and predicted) within a constituency across election years for each state. The table shows that on average, constituencies in most states experience considerable variation in competitiveness across election years.

80 Appendix Table 2: Candidate Thrnover (No Party Fixed Effects)

Dep Variable: New Candidate All Parties Incumb. Party Opp. Party Other Parties (1) (2) (3) (4) (5) OLS 2SLS 2SLS 2SLS 2SLS Panel A: Competition -0.064*** -0.066 0.229 0.014 0.025 (0.011) (0.054) (0.196) (0.109) (0.063)

Mean dep. var. 0.75 0.75 0.39 0.77 0.92 Controls Yes Yes Yes Yes Yes Party Fixed Effects No No No No No F-stat excluded instrument 219.43 84.59 228.48 156.43 Number of clusters 3013 3010 2370 2466 2826 Obs 30183 30180 6811 7246 15102 (1) (2) (3) (4) (5) OLS 2SLS 2SLS 2SLS 2SLS Panel B: Dispersion -0.048** 0.442 1.249* -1.080 -0.414 (0.021) (0.296) (0.723) (0.747) (0.342)

Mean dep. var. 0.75 0.75 0.39 0.77 0.92 Controls Yes Yes Yes Yes Yes Party Fixed Effects No No No No No F-stat excluded instrument 28.91 21.39 15.70 17.12 Number of clusters 3013 3010 2370 2466 2826 Obs 30183 30180 6811 7246 15102 Notes: This table estimates the effect of electoral competition on whether a candidate running for office also ran in the previous election in the same constituency. The specification is the same as in Table 3, but without party fixed effects. Competition is defined as 1 minus the expected margin between the incumbent and her strongest opponent. Dispersion is defined as 1 minus the HHI. All regressions contain constituency and state-year fixed effects. Standard errors clustered by constituency. Vote shares rescaled to sum to 1.

81 Appendix Table 3A: Candidate Demographics, Gender

Dep Variable: Female Candidate All Parties Incumb. Party Opp. Party Other Parties (1) (2) (3) (4) (5) OLS 2SLS 2SLS 2SLS 2SLS Panel A: Competition -0.007 0.012 0.182 0.129** -0.078 (0.007) (0.031) (0.193) (0.057) (0.048)

Mean dep. var. 0.06 0.06 0.06 0.06 0.05 Controls Yes Yes Yes Yes Yes Party Fixed Effects Yes Yes Yes Yes Yes F-stat excluded instrument 221.28 18.06 214.74 152.00 Number of clusters 3013 3010 2370 2466 2826 Obs 30183 30180 6811 7246 15102 (1) (2) (3) (4) (5) OLS 2SLS 2SLS 2SLS 2SLS Panel B: Dispersion -0.012 0.004 -0.205 0.112 0.065 (0.013) (0.149) (0.287) (0.395) (0.245)

Mean dep. var. 0.06 0.06 0.06 0.06 0.05 Controls Yes Yes Yes Yes Yes Party Fixed Effects Yes Yes Yes Yes Yes F-stat excluded instrument 29.37 20.12 15.15 17.91 Number of clusters 3013 3010 2370 2466 2826 Obs 30183 30180 6811 7246 15102 Notes: This table estimates the effect of electoral competition on the gender of a candidate. Competition is defined as 1 minus the expected margin between the incumbent and her strongest opponent. Dispersion is defined as 1 minus the HHI. All regressions contain constituency and state-year fixed effects. Standard errors clustered by constituency. Vote shares rescaled to sum to 1.

82 Appendix Table 3B: Candidate Demographics, Age

Dep Variable: Candidate Age All Parties Incumb. Party Opp. Party Other Parties (1) (2) (3) (4) (5) OLS 2SLS 2SLS 2SLS 2SLS Panel A: Competition 0.825* -1.431 -2.741 -1.363 -3.997* (0.441) (1.572) (6.855) (2.632) (2.381)

Mean dep. var. 48.39 48.39 50.89 49.50 46.61 Controls Yes Yes Yes Yes Yes Party Fixed Effects Yes Yes Yes Yes Yes F-stat excluded instrument 712.29 40.35 231.71 320.30 Obs 7879 7879 1852 2103 3971 (1) (2) (3) (4) (5) OLS 2SLS 2SLS 2SLS 2SLS Panel B: Dispersion 1.481* -9.214 -7.773 -5.087 -15.834* (0.793) (6.496) (13.281) (10.467) (9.461)

Mean dep. var. 48.39 48.39 50.89 49.50 46.61 Controls Yes Yes Yes Yes Yes Party Fixed Effects Yes Yes Yes Yes Yes F-stat excluded instrument 142.92 35.65 39.75 70.78 Obs 7879 7879 1852 2103 3971 Notes: This table estimates the effect of electoral competition on a candidate's age at the time of declaring candidacy. Competition is defined as 1 minus the expected margin between the incumbent and her strongest opponent. Specifications include party fixed effects, baseline year competition controls, and state-year fixed effects. Standard errors are clustered by constituency. Sample period is 2004-2017. Constituencies in 2004-2007 and 2008-2017 are stacked since the 2008 delimitation prevents comparison across these periods. Dispersion is defined as 1 minus the HHI. Vote shares rescaled to sum to 1.

83 Appendix Table 4: State Resource Allocation, Power Supply

Dep Var: Ln Lights (1) (2) (3) (4) OLS 2SLS OLS 2SLS Competition 0.025** 0.026 0.037** 0.025 (0.011) (0.043) (0.017) (0.045) Competition x Pre-Election Year -0.030*** 0.073*** -0.034*** 0.050 (0.008) (0.019) (0.013) (0.031) Competition x Aligned -0.020 -0.001 (0.017) (0.048) Competition x Pre-Election Year x Aligned 0.008 0.052 (0.015) (0.037)

Pre-Election Year x Aligned -0.015 -0.043* (0.010) (0.022) Aligned 0.021* 0.010 (0.012) (0.029) Pre-Election Year -0.108*** -0.149*** -0.103*** -0.131*** (0.011) (0.012) (0.012) (0.018)

P-val, #1 + /2 = 0 0.646 0.027 0.492 0.293 P-val, 01 + /2 + 33 + /4 = 0 0.434 0.027 Mean dep. var. 2.809 2.809 2.811 2.811 Baseline dep. var Yes Yes Yes Yes Controls Yes Yes Yes Yes Number of clusters 2300 2300 2258 2258 Obs 9516 9516 9430 9430 Notes: This table estimates the effect of electoral competition on power supply measured by the log of total light emissions before and after an election. Sample is restricted to the years preceding and following an election year. Aligned constituencies are those where the incumbent is a member of the party or coalition in control of the state government. Competition is defined as 1 minus the expected margin between the incumbent and her strongest opponent. All regressions contain constituency and state-year fixed effects. Standard errors clustered by constituency. Vote shares rescaled to sum to 1.

84 APPENDIX B

This section illustrates the difference between standard shift-share instruments, which are constructed by computing the inner product of shifts and shares, and the instrument used in this setting, which is constructed by computing a non-monotonic function of shifts and shares. Consider a simple example with two shares and one time period in a single state. Following the framework used in Goldsmith-Pinkham et al. (2018), the standard shift-share (Bartik) instrument is defined as:

Bc = z1cg + Z2c92

Since the shares, zic and Z2c, sum to 1, the instrument can be re-written as:

Bc = g2 + g C - gC)zic

The first stage equation regressing the endogenous regressor, xc, on the instrument is given by:

Xc = Bc + cc

=Y92c + (gj7c - g -c)zlc + Ec

This equation shows that using the shift-share instrument is numerically equivalent to using the first share, zic as an instrument. The only difference between the two approaches is that the shift-share instrument rescales the first stage coefficient by the difference in growth rates,

_c 1c g. In either case, however, the predicted regressor is the same. 91 -92 In contrast to the standard shift-share, my instrument for competition is given by:

PredictedCompetition = 1 - (z 1 c + g1c) (z2c + g2C)

Since the shares sum to 1, the instrument can be re-written as:

PredictedCompetition = 1 - |2z1c + g1 c - g2 -

The first stage equation regressing actual competition, Competitionc on the instrument, PredictedCompetitionc,

85 is given by:

Competitione = PredictedCompetition,+ E,

g2 - + C = 11 - 12zi + gi il

(2 - gi~' + g2 - 2zic)y + Ec, if zic > 0.5 - 2

(gi ( - g2 + 2zic)-y + cc, if zic < 0.5 - gg 2

Here, the shift-share instrument is not equivalent to using the shares as instruments. This equation shows that changes in the relative growth rates, g g , lead to changes in predicted competition depending on the share, zic.

86 Chapter 2

Shaping Cities with Political and Administrative Boundaries*

2.1 Introduction

Urbanization has increased dramatically over the last few decades, with more than half of the world's population now residing in cities (United Nations, 2014). Much of this urbanization has taken place in developing countries, where improvements in public goods provision have lagged behind rapid population growth.

The responsibility of providing local public goods and services typically falls to local governments. In many cities around the world, the bureaucrats and politicians that comprise these local governments manage non-overlapping administrative and political units. Bureaucrats collect and aggregate local revenue, and use this revenue to deliver services within their administrative units. Politicians, though usually not formally responsible for public goods provision, may monitor bureaucrats or use supplemental funds to influence what goods are delivered and where they are placed within their constituencies.

The division of cities into administrative and political units would suggest that a locality's assignment to a particular unit may be consequential for the extent and quality of public goods it

*For comments and suggestions that have improved this paper, I thank Daron Acemoglu, Abhijit Banerjee, Enrico Cantoni, Esther Duflo, Robert Gibbons, Nick Hagerty, Donghee Jo, Gabriel Kreindler, Matt Lowe, Benjamin Marx, Rachael Meager, Ben Olken, Arianna Ornaghi, Otis Reid, Frank Schilbach, Cory Smith, and Ariel Zucker. Special thanks to staff members at the MIT GIS Lab for their help in processing the GIS data. This paper uses data from Khan et al. 2016, and an ongoing randomized controlled trial joint with Adnan Q. Khan, Asim Khwaja, and Ben Olken. Funding support from the Samuel Tak Lee MIT Real Estate Entrepreneurship Laboratory is gratefully acknowledged.

87 receives, but there is surprisingly little empirical evidence on whether this is in fact the case. This paper studies how the delimitation of administrative and political units determines the allocation of public goods within cities, and whether these differences in allocations are reflected in property valuations. The empirical setting is Pakistan. Like many developing countries, Pakistan has experienced a wave of urbanization over the last few decades - nearly 40% of the population currently resides in cities. In these cities, administrators are responsible for providing local public goods such as street lights, roads, sanitation, and water. Politicians, in turn, monitor the administrators responsible for implementing services within their constituencies.

I compile a detailed, geo-referenced dataset on properties, residents, and public goods within two of the largest cities in the country, Lahore and Faisalabad. The administrative data is collected from property tax records that have been largely unchanged since 2002, when present-day political and administrative boundaries were drawn. The survey data records property and resident characteristics in 2013 and 2016, after these cities had changed dramatically. I also use an objective survey of public goods location and quality conducted in 2016. Together, the data provides a high- resolution view of internal city structure across time and space that is rarely available in a developing country context. Though the empirical setting is chosen primarily because of the availability of geo- referenced data on public goods and economic activity, key features of the local political structure

in Pakistan are similar to that in many other developing countries, suggesting the empirical results

may be applicable elsewhere.

I conduct three empirical exercises that, together, suggest that the assignment to political and administrative units affects the provision of public goods and property valuations in a locality. In

the first exercise, I test if political and administrative boundaries matter for public goods provision.

To do this, I use a spatial regression discontinuity framework to compare public goods provision in

localities in close proximity to each other, but across adjacent administrative or political units.

The challenge in this exercise is that multiple attributes of an administrative or politi-

cal unit vary across a boundary, any one of which may be correlated with outcomes of interest.

Politicians may differ in their ability to secure funds, for example, resulting in improvements in

public goods quality in some units, while others are left behind. Or conditional on politician type, electoral concerns may incentivize some politicians to improve public goods. Instead of focusing on

any single attribute that may vary across a boundary, I use a value-added framework to estimate

the total variation in outcomes of interest. To do this, I estimate fixed effects for each political and

administrative unit in a spatial regression discontinuity design (RDD). These fixed effects measure

88 the "value-added" in public goods quality from a locality's assignment to a particular political and administrative unit. Assuming these value-added effects are constant across space, this approach allows for a ranking of all political and administrative units based on discontinuous changes in public goods provision and property valuations across localities in close proximity to one another, but on opposite sides of a political or administrative boundary.

Implementing this procedure, I estimate significant discontinuities in public goods provision across political and administrative unit boundaries. I find that the quality and extent of services such as street lights and road repair vary significantly across political boundaries; while the quality of infrastructure such as dumpsters and public water taps vary significantly across administrative boundaries. I also estimate significant changes in self-reported property valuations across these boundaries. The changes in service provision across political units is particularly striking given that these units are not tied to a tax base to finance public goods. The results suggest that informal channels such as monitoring of bureaucrats may be critical for effective public goods provision within cities.

In the second exercise, I use an empirical Bayes procedure to quantify the estimated variation in public goods provision and property valuations across political and administrative units.

I find that changes in public goods provision are large across both political and administrative boundaries, though the relative magnitudes vary depending on the type of public good. I also estimate considerable variation in property valuations across political units in particular.

Finally, I correlate measures of political unit quality with observable characteristics of politicians and constituencies to understand which attributes explain the variation in public goods and property valuations. On the politician side, I find that a politician's tax payment is a strong, negative predictor of political unit quality. This correlation can be interpreted in either of two ways: either wealthier politicians (who pay more in taxes) tend to provide lower quality public goods within their constituencies, or less corrupt politicians (who would also pay more in taxes) tend to provide lower quality public goods. On the constituency side, higher overall turnout in recent elections is a strong, positive predictor of political unit quality. I find that this is not driven by own voter behavior: there is no variation in voter turnout at the boundary itself. This suggests that constituency-wide turnout may be correlated with the selection of better politicians, or, conditional on politician type, may result in better monitoring or incentivizing of politicians during their time in office.

This paper makes a number of contributions to the literature. First, to the best of my

89 knowledge, the consequences of political delimitation for public goods provision within cities has yet to be examined in the literature. Though there is a growing literature in political economy on the misallocation of public goods (e.g. Burgess et al., 2015, Kramon and Posner, 2014, Frank and

Rainer, 2012, Besley et al., 2004, Knight, 2002) and the consequences of delimitation (e.g. Hsiao, 2019, Lipscomb and Mobarak, 2017), these studies have largely focused on misallocation across regions. The welfare consequences of misallocation may be particularly large in the presence of externalities and spillovers within cities.

Empirical studies on internal city structure are in general scarce, both because of identification challenges and data limitations. A recent literature (e.g. Ahlfedt et al. 2015, Turner et al. 2014) has employed high-resolution data to study the distribution of land values within cities.

Though these studies consider governments to be important determinants of of the spatial allocation of economic activity, the focus is in large part on zoning or land use regulations as the key channels through which political actors wield influence. Less understood are the consequences of political boundaries - typically drawn with electoral considerations in mind (e.g. what boundaries ensure fairness, or conversely, maximizes the vote share of a party?), we know little about how these delimitations influence the allocation of public goods and, in turn, the growth and structure

of cities. As an initial step towards quantifying misallocation within cities, this paper shows that

political factors are in fact important determinants of the intra-city spatial allocation of local public

goods.

Second, I provide suggestive evidence that highly local public goods are capitalized into

(self-reported) property valuations. A number of papers have shown that housing prices reflect variation in access to goods and services (e.g. public schools in Black 1999; air quality in Chay

and Greenstone 2005 and Currie et al. 2015; street asphalting in Gonzalez-Navarro and Quintana-

Domeque 2016). Building on this work, I show that local public goods such as street lights and

roads are reflected in property values. In contrast to existing work, I show that sharp changes

in property values occur not only across administrative boundaries, which capture changes in the

financing of public goods, but also across political boundaries, which capture political factors that

determine access to goods and services. This result suggests that property values may also reflect

political access.

Methodologically, this paper develops a framework to estimate regression discontinuities

when there are changes in multiple features across a threshold. In particular, I show that the

value-added models traditionally used to estimate individual fixed effects in panel data settings

90 (e.g. Abowd et al. 1999, McCaffrey et al, 2003, Kane and Staiger 2008, Card et al. 2013, Kane et al. 2013, Finkelstein et al. 2016, Chetty and Hendren, 2018a, b.) can be applied in a spatial RDD context to quantify the total variation in outcomes generated by a threshold.

The remainder of the paper is organized as follows: Section 2.2 presents the setting. I describe the empirical framework in Section 2.3, and the data in Section 2.4. Sections 2.5 presents the main empirical results. Section 2.6 concludes.

2.2 Empirical Setting

The empirical setting is Lahore and Faisalabad, the two largest cities in Punjab, Pakistan with populations of 18 million and 4 million, respectively. As in many developing countries, these cities have experienced rapid urbanization over the last two decades.

Each city is divided into approximately 5 to 10 administrative units, known as the Town and Municipal Authorities (TMAs). These units are responsible for providing and managing local public goods such as street lights, water, sanitation, and road repair. Public goods are financed from property tax collected from within each TMA, and supplemented by non-tax revenue and provincial grants.' TMAs are staffed by career bureaucrats. By law, no political or elected person can hold a position in a TMA, but politicians may still influence staff composition by interfering with transfer or promotion decisions.

Each city is also divided into approximately 25-30 provincial and national assembly constituencies. These constituencies are served by MPAs and MNAs, respectively. In Lahore and

Faisalabad, the majority of these politicians have been members of the Pakistan Muslim League-

Noon (PML-N) party. Politicians serve for five years, and are re-elected frequently. Though the primary role of politicians is to participate in the provincial or national legislative assembly, they also influence public goods provision through formal and informal channels. Formally, MPAs and

MNAs receive discretionary funds from the federal and provincial governments. These constituency development funds (CDFs) can be spent on public goods (e.g. road repair, miscellaneous development schemes) within their jurisdiction. Though comprehensive data on these funds is largely unavailable, anecdotal evidence suggests these annual funds can reach up to Rs. 400 million (approximately $2.8 million) each year.

Informally, politicians monitor administrators in the TMAs. Politicians are known to

'Budgetary data for TMAs is not available.

91 select the location, type, and quality of public goods within their constituencies. In addition, politicians are often the first point of contact for citizens with municipal concerns and flag these concerns to the appropriate TMA staff. Because of this informal channel, citizens often fault politicians, rather than administrators, for poor public goods quality.

The present-day boundaries of administrative and political units were delimited around the same time in the early 2000s. TMAs were created in 2001 to decentralize municipal functions from city districts to smaller administrative units. 2 The boundaries of both national and provincial constituencies were last delimited in 2002 by the Pakistan Election Commission. Boundaries were redrawn to reflect 1998 census results, increasing the number of units in all provinces and affecting almost all existing boundaries. Though ostensibly guided by the 1998 census, the actual process of delimitation was opaque with Democracy Reporting International reporting that "The delimitation process [was] not formalized in administrative regulations, and the guidelines and policies for its conduct are not published." (2015).

Figure 1 displays administrative units, political units, and major roads in Lahore.3 As can be seen visually, boundaries are non-overlapping for the most part - through there are a small number of exceptions - and do not follow any of the major roads.

2.3 Empirical Framework

The empirical framework exploits variation in outcomes of localities located in close proximity to each other, but across adjacent administrative or political units. The idea is that these localities shared identical characteristics before a boundary was drawn, but because of their assignment to different units, now rely on a different bureaucratic or political structure for public goods provision.

The challenge with this approach is that multiple attributes of an administrative or political unit vary across a boundary, any one of which may be correlated with outcomes of interest.

The provision of public goods within a political unit, for example, may be influenced by politician identity (incumbents or those aligned with the ruling party may better able to secure funds and resources for their units); qualities of the political unit itself (larger units may be more difficult to monitor, leading to lower public goods quality); and finally, the residents who choose to reside within the political unit (resident preferences may change the type of public goods provided).

2 The TMAs have now been disbanded, with city-wide municipal corporations now responsible for providing public goods and services throughout a city. 3 Boundaries are similar in Faisalabad.

92 The standard approach in the literature is to focus on a single attribute that varies across a boundary. Instead, this paper proposes an alternative approach to estimate the total variation in an outcome of interest due to the boundary itself. The estimation framework proceeds in three steps.

First, I estimate a fixed effect for every administrative and political unit using a spatial regression discontinuity design (RDD). Second, using these estimated fixed effects, I apply an Empirical Bayes procedure to assess the magnitude of the variation in outcomes generated by each type of unit.

Finally, I correlate the fixed effects with observable characteristics of politicians and the units themselves to understand which attributes explain the variation in unit quality.4 Each step is described in detail below.

2.3.1 Fixed effects regression discontinuity design

I use a spatial RDD framework to estimate a set of fixed effects for political and administrative units. The specification of interest for estimating political unit fixed effects is given by:

6 Yijb = a3 + b + f(geographic location)ijb + -yk + C23j (2.1) where Yijb is an outcome of interest in grid i, political unit j, near political unit boundary b; aj

are political unit fixed effects for each unit j = 1,... , N in my data; 6b are political boundary fixed effects; and f(geographic location)ijb is a boundary-specific linear polynomial in latitude and longitude. Since political and administrative boundaries sometimes overlap, I also include administrative unit fixed effects -Yk. This ensures the variation in outcomes is due to changes in political

units, rather than administrative units. The specification is estimated for a range of bandwidths to establish robustness. 5 Spatial correlation in residuals is addressed by estimating Conley standard errors.

I estimate administrative unit fixed effects using the analogous specification:

Yikb = ak + 6b + f (geographic location)ikb + -)j + Eikb (2.2)

Here, Yikb is an outcome of interest in grid i, administrative unit k, near administrative unit boundary

b; ak are administrative unit fixed effects for each unit k = 1,.. . , N in my data; 6b are administrative

4Bureaucrat characteristics are not observed. 5There is no consensus on an optimal bandwidth for multi-dimensional RDDs (Dell and Olken, 2019).

93 boundary fixed effects; and f(geographic location)ikb is a boundary-specific linear polynomial in latitude and longitude.

These specifications rely on standard RDD assumptions, along with an additional constant treatment effects assumption to allow for a ranking of the fixed effects. The key RDD identifying assumption is that all factors other than assignment to unit vary smoothly across boundaries. This means the boundary can only capture the political or administrative unit treatment, rather than changes in any fundamentals that directly affect the intrinsic attractiveness of a locality. Boundaries cannot dived one quality of land from another. I provide support for this assumption by showing there are no discontinuities in locality characteristics across boundaries in 2002, when present-day units were delimited.

A second identifying assumption is that assignment to unit is the only discontinuous change across boundaries correlated with outcomes of interest. This assumption would be violated

if assignment to unit was correlated with other locality characteristics that determined which side of

a boundary had better public goods. Since the TMA, political and national assembly units are the

primary determinants of public goods quality during this time, this concern is largely addressed by

controlling for administrative (political) boundaries in the political (administrative) unit regressions.

RDD frameworks often rely on an additional assumption that there is no selective sorting

across boundaries. Instead of imposing this assumption, I consider sorting an outcome of interest

(e.g. as in Dell 2010), and explore the possibility that heterogenous residents sort into localities based on variation in political and administrative units. Under the RDD assumptions, Equations (1) and (2) generate a set of fixed effects mea-

suring the change in outcomes attributable to assignment to a particular political or administrate

unit. To make meaningful comparisons of these fixed effects, however, an additional assumption is

needed. I motivate this assumption with a stylized example. Figure 2 displays grids observed in a

city segmented into N = 5 political units. Suppose we are interested in estimating political unit

fixed effects, a1,... , a5. I define a connected set to be the set of political units that can be linked

via a boundary. In Figure 2, the connected set are all political units within a single city: al is

linked to a2 and a4; a2 is linked to a 3 ; and a3 is linked to a5. Within any connected set, we can

normalize a, = 0, and estimate all ai relative to a,. However, for this implicit ranking of units

to be valid, we need to make the additional assumption that treatment effects are constant across

space. This assumption requires each unit to have a constant effect in all locations relative to other

94 units in the connected set.6

The constant treatment effects assumption allows us to rank all units within a connected set - even if they share no border. In the stylized example, we can compare al to a5 by linking changes in outcomes moving from the border between a1 and a2; to the border between a2 and a; and finally to the border between a3 and a5 - so long as the assignment to unit treatment effect is constant anywhere within the unit.

2.3.2 Assessing magnitudes

The fixed effects generated from the RDD above allows us to assess the magnitude of variation in outcomes created by political and administrative units. Before we can do this, however, note that each fixed effect is estimated with noise:

63 = aj + ej (2.3)

It is therefore important to adjust the fixed effects to avoid overstating the true variance.

In a panel data setting, the noise can be reduced by an empirical Bayes procedure that reweights each fixed effect by a signal-to-noise ratio based on year-to-year variation in the estimate (e.g.

McCaffrey et al, 2003; Kane and Staiger 2008; Kane et al. 2013).

I apply the analogous procedure in this setting and reweight each fixed effect by a noise-to- signal ratio based on location-to-location variation in the estimate. This results in a set of posterior fixed effects:

Var(a) aj,adjusted =i Var(a) (2.4)

This procedures trades noisy, but unbiased estimates (dj), for more precise, but biased estimates (&j,adjusted), allowing us to recover the magnitude of variation created by the units. I apply this procedure to compare the variance of fixed effects across political units, with the variance of fixed effects across administrative units, to assess which is more consequential for outcomes.

6 The constant treatment effects assumption in this context is analogous to the one required for estimating value-added models using panel data (see, for example, Angrist 2017, Hull 2018).

95 2.3.3 Correlates of political unit quality

The fixed effects reflect the average causal effect of moving from one political or administrative unit to another. To understand what explains differences in "unit quality", I regress the posterior quality estimates on observable politician and unit characteristics such as politician alignment with the ruling party, or turnout in the preceding election year.

Since there are many potential covariates, running a simple OLS regression may result in overfitting. Instead, I use a regularization procedure to identify the most powerful correlations.

Ridge and Lasso procedures shrink regression coefficients by imposing a penalty on their size. I use

Ridge regressions, which do not shrink any of the coefficients to zero, and therefore perform better than Lasso when the point-mass of true zeros is small (Abadie and Kasy, 2017). This allows me to observe relative correlations for all covariates, rather than a restricted set that would be generated by Lasso. The larger the coefficient in a Ridge regression, the more important the variable in explaining the variation in posterior quality estimates (Hastie et al. 2009).

2.4 Data

2.4.1 Description

The empirical analysis requires geo-referenced data on properties, residents, and public goods within cities. I compile this data from administrative sources and surveys for Lahore and Faisalabad.

Together, the data provide a high-resolution view of internal city structure, first in 2002, when present-day political and administrative boundaries were drawn, and again more than a decade later, when these cities had changed dramatically. Below, I describe the data and sources used to construct the variables for analysis.

Properties: Data on properties is compiled from two cross-sectional surveys conducted in 2013 and

2016.7 Surveyors directly observed the characteristics of each property in randomly selected clusters in Lahore and Faisalabad. Each property was then matched (manually in 2013; and using a digital database in 2016) to an administrative property tax record. This tax record contains data from the tax department's own surveys. In addition to observing property characteristics such as land use, land area, covered area, number of floors, etc., the tax department assigned each property to one

7 The first survey was conducted in Khan et al. 2016, while the second for an ongoing randomized controlled trial.

96 of seven categories based on the property's location. High (low) category areas correspond to more

(less) attractive localities. For nearly 70% of properties observed in the 2013 survey, the property tax record contains data from the tax department's 2002 survey.8 The tax records therefore provide a view of city structure at the time of delimitation, and can be used to test for balance across boundaries.

I measure property value using the property owner's subjective assessment of her property's worth in 2013 and 2016. In contrast to the tax department's assessed valuation of a property, which is formulaic and based on property characteristics observed for the most part in 2002, the property owner's assessment captures variation in property values that may occur due to changes in property and locality characteristics over time. Figure Al displays surveyed properties in Lahore and Faisalabad.

Residents: The cross-sectional surveys also contain data on the residents of each sampled property.

To estimate whether there is sorting of residents across boundaries, I construct measures of year of residency, household size, and household expenditure.

Public goods: The analysis leverages a rich dataset of more than 18,000 geo-referenced public goods in Lahore and Faisalabad.9 These public goods - street lights, road repair, sanitation, and public water taps - are financed by local property taxes, and by grants from the provincial and federal governments. To collect this data, enumerators walked on every street in 500 localities within the two cities, tracking the quantity and quality of each type of public good on a smartphone application. GPS coordinates were taken for all observations of low-frequency goods (dumpsters, public taps) and every third observation of high-frequency goods (street lights, potholes, trash piles).

I construct measures of public goods quality by dividing each surveyed locality into 100 square meter grids. Each grid is assigned a measure of public goods quality by computing the number of observations of each type of public good within each grid. Streetlights are scaled by a point system, with a working street light assigned 10 points, and broken street lights assigned 5 points.

Political and administrative unit shape files: I match data on properties, residents, and

8 The tax department conducted province-wide property surveys in 2002 and 2013. New properties built between these years were assessed at the time of construction. Properties that experienced changes in form (e.g. covered area extended) or use (e.g. property converted from residential to commercial use) should in principle have been reassessed, but in practice this rarely occurred. 9 This data was collected for an ongoing randomized controlled trial.

97 public goods to political units using the Election Commission of Pakistan's maps of provincial constituencies. These maps are available in PDF form on the Election Commission's website, and have been manually converted to shape files by several teams.1 0 Though it is likely that there is some measurement error in these boundaries, so long as this error is not driven by unobservables correlated with outcomes, it will only add noise to the regression discontinuity estimates (e.g. Ozier 2018).

Elections: Data on national and provincial elections is assembled from three election years: 2002, 2008, and 2013. Though both Lahore and Faisalabad were dominated by the PML-N party from

2002 though 2013, there is considerable variation in turnout, PML-N vote share, and politician type across constituencies.

Politician characteristics: I use data published by the Electoral Commission of Pakistan (ECP) and Federal Board of Revenue (FBR) to construct a dataset of politician demographics and income tax returns. Politician demographics include age, gender, and education. Beginning in 2013, the

FBR published the total amount of income tax paid by all elected politicians at the end of every fiscal year. This data is used to construct two outcome variables: the amount of income tax paid and whether a politician filed taxes.1 1

Constituency development funds: Finally, I collect data on spending by MNA politicians within their constituencies using publicly available data on constituency development funds (CDFs) from

2008 to 2013.12 All elected politicians receive discretionary funds from the national or provincial government for local infrastructure projects within their constituencies. While CDFs are not the primary source -of funding for public goods, they are the only source of funding an elected politician has direct control over. By law all constituencies should receive a fixed amount, but actual amounts vary by constituency and can range anywhere from Rs. 1 million to Rs. 400 million (equivalent to approximately $7000 to $2.8 million). Funds are typically spent on development projects such as roads, electricity, health and water, though the project description is not always specified in the data.

101 use shape files distributed by Alhasan Systems. These shape files have been validated by those used by Callen et al. 2014. "Approximately 10-15% of politicians in each year do not file taxes at all. 1 2 Though MPA politicians also receive CDFs, data on allocated amounts or spending decisions has not been published systematically.

98 2.4.2 Summary Statistics and Balance

The empirical analysis requires that property characteristics are similar across political and administrative units at the time of delimitation. Table 1 estimates equations (1) and (2) using property tax records from 2002. The bandwidth is set to 250 meters. I report the F-statistic and corresponding p-value of the hypothesis test:

Ho : a, = a2 -- = aj = 0 (2.5)

To ensure there is no differential selection in the properties for which tax records were not updated since 2002, I start by testing for discontinuities in the year a tax record was last updated.

Table 1, Column 1 shows that there are no changes in whether a property's tax record was updated before or after 2002 across either type of political unit (provincial or national).

The variation in property tax record updates across administrative units is significant (p- value = 0.02). This suggests that the property tax department reassessed properties at differential rates across administrative boundaries. While this does not invalidate the identification strategy

(reassessments may have been driven by some administrative units changing faster than others), it does prevent the use of 2002 tax records to establish balance at baseline.

Since political units are balanced on reassessment rates, I use the 2002 tax records to check for variation in property characteristics around the time of delimitation. Columns (2) to

(4) checks for discontinuities in property tax liability, number of rooms, and the total area of the property. Out of the 6 comparisons made, 2 are significant at the 10% level, suggesting some imbalance at baseline. This suggests that at least in some localities within each city, there were existing differences in property characteristics when the boundaries were drawn. It is possible that some of the boundaries drawn in 2002 overlapped with existing boundaries, and so the observed variation in property characteristics stems from variation in pre-2002 political units. In ongoing work, I digitize the boundaries for these units and exclude any borders that overlap. In the results reported below, I control for these initial property characteristics where possible.

2.5 Results

This section presents the core empirical results on internal city structure. The first part of the analysis tests for discontinuities in public goods, property characteristics, and resident characteristics

99 across administrative and political boundaries. The second part assesses the magnitude of variation generated by each type of unit. The third part explores the correlates of political unit quality.

2.5.1 Estimating fixed effects across boundaries

I consider three sets of boundaries that may plausibly influence the extent and quality of public goods: provincial assembly boundaries; national assembly boundaries; and TMA boundaries.

In addition, I construct a set of placebo boundaries by shifting the provincial assembly boundaries away from their true location. These placebo boundaries provide supporting evidence that the results estimated for the true boundaries are not spurious.

Public goods: I start by estimating the fixed effects RDD specification for public goods provision.

Tables 2A and 2B and reports the F-statistic and corresponding p-value testing the null hypothesis

that all political or administrative unit fixed effects are equal to 0 (equation (2.5)). Table 2A ex-

plores service provision (street lights, potholes, broken manholes, and trash piles). The last column

constructs a service provision index by normalizing these variables by their means and standard

deviations and computing the average. Similarly, Table 2B reports estimates for infrastructure

provision (dumpsters, public water taps, water filtration taps) and a corresponding infrastructure

index.

Table 2A suggests there are discontinuities in service provision across provincial and na-

tional assembly constituencies in Panels A and B respectively. The null hypothesis is soundly

rejected for streetlights and road repair continuity across provincial assemblies (p-values are less

than 0.002); and for road repair continuity across national assemblies (p-value is less than 0.10).

Column (5) shows there are significant jumps in overall service provision across both types of po-

litical units. By contrast, Panel C shows there is only variation in street lights provision across

administrative units. The null hypothesis that the administrative unit fixed effects are equal to

zero for overall service provision cannot be rejected. Finally, Panel D shows there are no significant

discontinuities in any type of service provision across the placebo boundaries.

There are also significant discontinuities in infrastructure provision across political and

administrative units, though the results are more pronounced in the latter. Panel A shows that even

in close proximity to a boundary, the placement of dumpsters varies across provincial assemblies.

Panel C shows that administrative units vary in their provision of dumpsters and public water

100 taps."3 Overall infrastructure provision varies considerably across these boundaries - the p-value for the F-test is below 0.01. Panel D shows there is some significant variation in infrastructure across placebo boundaries, but there are no significant discontinuities in overall infrastructure provision when all of the infrastructure goods are combined into a single index, as shown in Column 5.

Together, these results provide evidence that assignment to political and administrative units are important determinants of even highly localized public goods. The discontinuities estimated across political units are particularly striking - these units are not linked to the property tax base which is ostensibly responsible for financing these public goods. That there is variation in service provision across these units suggests either that some politicians are securing funds from other resources (e.g. constituency development funds) to finance public goods, or that some politicians are influencing the administrators responsible for providing public goods by directing them to their own constituencies.

It is also worth highlighting that the results are consistent with the ruling party PML-

N's emphasis on road construction during this time period.14 The results in Table 2A shows that the condition of roads varies significantly across political - but not administrative - boundaries, suggesting assignment to a particular political unit determined where roads were built or repaired.

Property values: I next assess whether the variation in public goods provision estimated across political and administrative units is capitalized into property values. To do so, I estimate equations

1 and 2 with (the log of) self-reported property value assessments as the outcome variable.

Table 3 presents results for self-reported property values in the 2016 and 2013 surveys.

As a robustness check, I restrict the 2013 sample to properties whose tax records was not updated after 2002 and add additional control variables from the 2002 property records.

I estimate significant discontinuities in property values across provincial assemblies and administrative units in 2016. The null hypothesis for each F-test is rejected at the 5% level. By contrast, there is no significant variation in property values across localities that are on opposite sides of a national assembly boundary. Discontinuities across placebo boundaries are significant at the 10% level. It is possible that this result is driven by the low number of observations available

in the 2016 survey. If these observations are in close proximity to the original provincial assembly lines, the specification may pick up some of the variation generated across provincial assembly units

1 3That there is no corresponding variation in the location of water filtration plants may be due to their low number. Appendix Table Al shows that the average number of filtration plants per grid is 0.01. "See, for example, https://www.economist.com/asia/2017/01/19/pakistans-misguided-obsession-with- infrastructure.

101 rather than only the placebo provincial assembly units.

The results for property values in 2013 are less clear. I estimate no significant discontinuities for any set of boundaries in the specification without tax record controls. In Column (3), I restrict the sample to properties whose record was not updated after 2002 and use data on tax liability, number of floors and property area as controls. The control variable increase the precision of the fixed effects. Here, I estimate significant discontinuities across provincial assembly boundaries

(p-value < 0.10) and national assembly boundaries (p-value < 0.05). There are no discontinuities for the placebo boundaries with or without controls.

There are two key takeaways from Table 3. First, Column (1) indicates that variation in public goods may in fact be capitalized into property prices. Residents report significantly different property valuations across provincial assembly and administrative boundaries in 2016. Second, that these results are not entirely reflected in the 2013 survey suggest that these property valuations may not be stable and instead reflect only the contemporaneous allocation of public goods.l1 If the allocation of public goods varies over time due to changes in politicians or politician incentives, for example, then residents may report different property valuations depending on the state of their locality at a particular time.

Sorting: Given the variation in public goods across political and administrative units, residents may sort across boundaries according to preferences for public goods. Evidence of such sorting would be consistent with a Tiebout model of residents "voting with their feet" (Tiebout, 1956), though in contrast to the traditional Tiebout setup, this sorting would be driven by variation in political and administrative unit quality, rather than variation in tax rates.1 6

Table 4 tests for discontinuities in three resident observables: (1) the number of years the resident has lived at the property, (2) household size, and (3) per capita household expenditure.

The only consistent result across survey years is in household per capita expenditure across administrative boundaries: Columns (3) and (6) of Panel C provide suggestive evidence of sorting into different TMAs. The null hypothesis is rejected at the 5% level in 2016 and 2013. There are no corresponding differences in the number of years a resident has lived at the property or in household size across administrative boundaries.

There are no significant differences in resident characteristics across provincial assemblies.

5 ' The public goods survey was conducted around the same time as the 2016 property survey. 16 Property taxes are applied at the same rates throughout Punjab, with the rates determined by property assessments.

102 Panel B shows though localities differ in household per capita expenditure across national assembly boundaries in 2016 (Column (3)), I cannot reject the corresponding null hypothesis in 2013 (Column (6)).

2.5.2 Assessing magnitudes

The preceding analysis suggests that the delimitation of political and administrative units created significant variation in public goods provision and property valuations within cities. How large are the differences generated by assignment to a particular political or administrative unit? To answer this question, I analyze the variance of the estimated fixed effects. I start by implementing the empirical Bayes procedure outlined in Section 2.3.2 to avoid overstating the true variance. Figure

3 displays the distribution of provincial assembly fixed effects for the service provision index and the corresponding posteriors. As can be seen graphically, the posteriors have a tighter distribution around the mean than the initial estimates.

Next, I plot the distribution of posteriors for public goods and property valuations. Figure

4A shows the distributions for services; Figure 4B for infrastructure and property valuations. There is considerable variation in public goods provision across political and administrative units. The streetlights index, which combines measures of total number of streetlights and the proportion that are in working condition, varies across all three sets of boundaries. Table 5 shows that the standard deviation of the provincial assembly fixed effects is 0.08, while that for administrative units is 0.13.

I use a Brown-Forsythe test to compare variances across political and administrative units. The test for equality of variances is rejected when comparing provincial to administrative units (p-value =

0.0381), which suggests that there is greater variation in streetlights provision across administrative rather than political units.

In contrast, Figure 4B shows larger variation in the provision of dumpsters across political units. The standard deviation of provincial assembly posteriors is 0.0047, while that for administrative units is only 0.0009. As shown in Table 5, equality of variances is soundly rejected.

The service and infrastructure indices vary considerably more across administrative units than provincial assemblies. Note, however, that the variation for services across administrative units is estimated with noise, as shown in Table 2A. The variation for services across provincial assemblies is smaller in magnitude, but precisely estimated. The variation in infrastructure across administrative units is large and significant, as shown in Table 2B.

103 Finally, Figure 4B plots the distribution of posteriors estimated from discontinuities in property valuations across political and administrative boundaries. Here, there is considerable variation across provincial assemblies. Table 5 shows the standard deviation in provincial assembly posteriors for property valuations is 1.72, much larger than that for either national assemblies or administrative units.

Taken together, these results show that many of the estimated discontinuities across political and administrative units are not only significant (as shown in Tables 2 and 3), but also large in magnitude. Though the results across public goods vary, perhaps reflecting the ability of politicians or administrators to influence the provision of certain goods and services but not others, it is clear from Figures 4A and 4B that there is sizable variation in public goods quality and property valuations even across localities in close proximity to one another, but assigned to different units.

2.5.3 Correlates of political unit quality

What explains the variation in administrative and political unit quality? As discussed in Section

2.3.3, there may be any number of factors that can explain the variation in public goods and

property valuations observed across boundaries, from the qualities of the administrator or politician

overseeing a unit to the qualities of the unit itself. I explore potential channels by correlating the

estimates of political unit quality with a range of observables.1 7

Figure 5A displays coefficients from a Ridge regression of provincial assembly constituen-

cies on politician and constituency characteristics. The larger a coefficient in this regression, the

more important the variable as a predictor of political unit quality. Since the public goods and

second property survey were conducted in 2016, I collect data on the politicians elected to office in

the 2013 elections.

The top panel considers the correlates of political unit quality as measured by variations

in property valuations across boundaries. Politician demographics are moderately correlated with

political unit quality, with older politicians associated with higher quality. Since a politician's tax

liability is unobserved, the amount of tax paid by a politician can be considered a proxy for both

politician wealth and corruption. I find that the amount of tax paid by a politician is a negative

predictor of political unit quality. Incumbents tend to administer units with higher political unit

quality, as do members of the minority (non PML-N) parties. This latter result is somewhat

17Observable characteristics of administrative units, such as budgets and expenditures, are unavailable, preventing an analysis of the determinants of administrative unit quality in this context.

104 surprising given that these politicians likely had less access to resources to public goods from the

ruling party. The correlation between a political unit's minority party status and property valuations

may stem from other channels.

In the last three rows, I examine constituency characteristics from the 2013 elections.

Here, I find that the overall turnout in the 2013 elections is a strong predictor of political unit

quality. The magnitude of this coefficient is more than double than that for any of the other

coefficients.

The bottom panel considers the correlates of political unit quality as measured by vari-

ations in public goods provision directly. I focus on streetlights and potholes, which, as shown in

Table 2A, are discontinuous across provincial assembly boundaries. So that higher values of the

public goods measures reflect higher political unit quality, I show results for road quality, defined

as the absence of potholes. The results for these public goods measures of political unit quality are

largely consistent with the preceding analysis. The amount of tax paid by a politician is negatively

correlated with political unit quality as measured either by streetlights or road quality. Higher turnout is positively correlated with political unit quality, as is the winning margin obtained by the politician in the 2013 elections. Here, minority party status is negatively associated with political

unit quality, which supports the hypothesis that these politicians may be unable to access resources to provide public goods to their constituents.

Figure 5B shows the corresponding results for national assembly constituencies. The data

available for these politicians is somewhat different than that available for members of the provincial

assemblies. I do not have data on the age or education of these candidates. However, I do have expenditure data on the constituency development funds (CDFs) obtained by politicians in national

assemblies from 2008 to 2013. Though this data does not correspond exactly to when the public goods survey was conducted, to the extent that public goods provision during this period was still visible later on, CDF spending may still be correlated with political unit quality.

The top panel shows that the strongest predictor of national assembly quality is the winning margin of the politician elected in the 2013 elections. This result stand in contrast to the one estimated above, which showed a small, negative correlation between the winning margin and provincial assembly quality as measured by property valuations. I also find a negative correlation between turnout and national assembly quality, though this correlation is smaller in magnitude.

In the bottom panel, I regress quality as measured by public goods provision on politician and constituency observables. CDF spending is positively correlated with quality as measured

105 by either road quality or streetlights, suggesting that politician access to funds is an important determinant of political unit quality. The correlations with constituency characteristics such as turnout and winning margin differ depending on which measure of public goods is used, with road quality negatively correlated with turnout, for example, and streetlights positively correlated (albeit at a much smaller magnitude).

To understand the differences between the results for property valuations, streetlights and road quality in the national assembly constituencies, I estimate the correlation between posteriors measured by public goods and property valuations. Table A2 shows that while the posteriors for provincial assemblies are aligned, with higher property valuations associated with better public goods, the results are noisier for national assemblies. For example, national assemblies that are ranked higher in property valuations are ranked lower in roads quality. This is probably a result of the fact that the national assembly fixed effects using property valuations as an outcome measure are noisily estimated (cf. Table 3, Panel B), and their correlations with fixed effects using road quality as an outcome measure (which are significant at the 10% level as shown in Table 2A, Panel B) are not informative.

Discussion: The preceding analysis suggests a number of factors explain variation in political unit quality. I consider the provincial assembly estimates of political unit quality to be more informative, both because the spatial RDD estimates show significant variation in property valuations and public goods provision across these boundaries and because the quality estimate are positively correlated with one another, suggesting the estimates are picking up the same underlying variation in political units.

On the politician side, the amount of tax paid is a strong, negative predictor of political unit quality. This correlation can be interpreted in either of two ways: either wealthier politicians

(who pay more in taxes) tend to provide lower quality public goods within their constituencies, or less corrupt politicians (who would also pay more in taxes) tend to provide lower quality public goods. Absent direct data on politician assets, it is difficult to disentangle these two interpretations.

On the constituency side, higher overall turnout in recent elections is a strong, positive predictor of political unit quality. Appendix Table A3 shows that there is no variation in turnout at the boundary, which suggests that being part of a political unit that is more engaged - rather than own voter behavior - can result in better public goods and ultimately higher perceived property valuations. Higher overall turnout may lead to the selection of better politicians, or, conditional on

106 politician type, may lead to better monitoring or incentivizing of politicians during their time in office.

2.6 Concluding remarks

This paper highlights the role of political and administrative units in shaping internal city structure.

I use a spatial RDD design to estimate discontinuities in public goods provision across political and administrative unit boundaries. In Pakistan, these units are jointly responsible for providing local public goods such as street lights, roads, sanitation, and water - the political units for monitoring and mediating public goods provision, and the administrative units for actual delivery.

The assignment of a locality to a particular political or administrative unit results in large variation in public goods provision, even when comparing localities in close proximity to one another, but on opposite sides of a boundary. Variation in public goods may lead to to variation in (at least perceived) property valuations. Indeed, I find that political and administrative units ranked highly in many types of public goods quality are also assessed by residents to have higher property valuations. Political unit quality is correlated with a number of observables, including politician tax payments, an indicator of either wealth or low tax evasion, as well as constituency characteristics such as overall turnout in elections.

Given that key features of the local administrative and political structure in Pakistan are similar to that in many other developing countries, the results reported here likely generalize to other contexts. Further research is needed to understand the consequences of spatial distortions in public goods provision and property valuations created by intra-city units. The variation in public goods across political units in particular may have large welfare consequences, as the influence of politician or constituency-specific factors in determining where public goods are placed within cities may result in misallocation.

107 References

Abadie, Alberto and Maximilian Kasy (2017). "The Risk of Machine Learning," Working Paper.

Abowd, John M., Francis Kramarz, and David N. Margolis (1999). "High Wage Workers and High Wage Firms," Econometrica, 67 (2), 251-333.

Ahlfedt, G.M., Redding, S.J., Sturm, D.M. AND N. Wolf (2015), "The Economics of Density: Evidence from the Berlin Wall," Econometrica, 83(6): 2127-2189.

Angrist, Joshua, Peter Hull, Parag Pathak, and Christopher Walters, (2017). "Leveraging Lot-

teries for School Value-Added: Testing and Estimation," Quarterly Journal of Economics, 132, 871-919.

Besley, Timothy, Rohini Pande , and Vijayendra Rao (2004). "The Politics of Public Good Provi-

sion: Evidence from Indian Local Governments," Journal of European Economics Association, 2(2-3): 416-426.

Black, Sandra. (1999). "Do Better Schools Matter? Parental Valuation of Elementary Education." The Quarterly Journal of Economics, 114(2): 577-599.

Burgess, Robin, Remi Jedwab, Edward Miguel, Ameet Morjaria, and Gerard Padro i Miquel. (2015). "The Value of Democracy: Evidence from Road Building in Kenya." American Economic Review, 105(6): 1817-51.

Callen, Michael, Saad Gulzar, Ali Hasanain, and Yasir Khan, (2014). "The Political Economy of Public Employee Absence: Experimental Evidence from Pakistan," Working Paper.

Cantoni, Enrico, (2019). "A Precinct Too Far: Turnout and Voting Costs," American Economic Journal: Applied Economics, Forthcoming.

Card, David, Jorg Heining, and Patrick Kline, (2013). "Workplace Heterogeneity and the Rise of West German Wage Inequality," The Quarterly Journal of Economics, 128 (3), 967-1015.

Chay, Kenneth and Michael Greenstone (2005). "Does Air Quality Matter? Evidence from the Housing Market," Journal of Political Economy, 113(2): 376-424.

108 Chetty, Raj and Nathaniel Hendren, (2018a). "The Impacts of Neighborhoods on Intergenerational

Mobility I: Childhood Exposure Effects," Quarterly Journal of Economics. 133(3): 1107-1162.

Chetty, Raj and Nathaniel Hendren, (2018b). "The Impacts of Neighborhoods on Intergenerational

Mobility II: County-Level Estimates," Quarterly Journal of Economics, 133(3): 1163-1228.

Currie, Janet, Lucas Davis, Michael Greenstone, Reed Walker (2015). "Environmental Health

Risks and Housing Values: Evidence from 1,600 Toxic Plant Openings and Closings," American Economic Review, 105(2): 678-709.

Dell, Melissa (2010). "The Persistent Effects of Peru's Mining Mita," Econometrica 78(6): 1863- 1903.

Dell, Melissa and Benjamin Olken. (2019). "The Development Effects of the Extractive Colonial

Economy: The Dutch Cultivation System in Java", Review of Economic Studies, Forthcoming.

Democracy Reporting International (2015). "Electoral Delimitation in Pakistan: Formula for In-

equality." Briefing Paper No. 01 of 2015.

Finkelstein, Amy, Matthew Gentzkow, and Heidi Williams, (2016). "Sources of Geographic Vari-

ation in Health Care: Evidence from Patient Migration," The Quarterly Journal of Economics,

131(4), 1681-1726.

Frank, Raphael, and Ilia Rainer. (2012) "Does the Leader's Ethnicity Matter? Ethnic Favoritism,

Education, and Health in Sub-Saharan Africa." American Political Science Review, 106(2): 294-325.

Gonzalez-Navarro, Marco and Climent Quintana-Domeque (2016). "Paving Streets for the Poor:

Experimental Analysis of Infrastructure Effects," Review of Economics and Statistics 98(2):

254-267.

Hastie, Trevor, Tibshirani, Robert, and Friedman, Jerome, (2009). "The Elements of Statistical

Learning," Second Edition. Springer Series in Statistics.

Hull, Peter (2018). "Estimating Hospital Quality with Quasi-experimental Data," Working Paper.

Hsiao, Allan. (2019). "Misallocation and Infrastructure Investment: Evidence from Healthcare in

Indonesia." Working Paper.

109 Kane, Thomas, and Douglas Staiger, (2008). "Estimating Teacher Impacts on Stu- dent Achieve- ment: An Experimental Evaluation," NBER Working Paper 14607.

Kane, Thomas, Daniel McCaffrey, Trey Miller and Douglas Staiger, (2013). "Have We Identified Effective Teachers? Validating Measures of Effective Teaching Using Random Assignment," Seattle, WA: Bill and Melinda Gates Foundation.

Khan, A. Q., A. I. Khwaja, and B. A. Olken, (2016). "Tax Farming Redux: Experimental Evidence on Performance Pay for Tax Collectors." The Quarterly Journal of Economics, 131(1): 219-271.

Knight, Brian (2002). "Endogenous Federal Grants and Crowd-Out of State Government Spending: Theory and Evidence from the Federal Highway Aid Program." American Economic Review, 92, 71-92.

Kramon, Eric, and Daniel N. Posner. (2014) "Ethnic Favoritism in Primary Education in Kenya." Working Paper.

Lipscomb, Molly and Ahmed Mushfiq Mobarak. (2017). "Decentralization and Pollution Spillovers: Evidence from the Re-drawing of County Borders in Brazil." Review of Economic Studies, 84(1):464-502.

McCaffrey, Daniel F., J. R. Lockwood, Daniel M. Koretz, and Laura S. Hamilton, (2003). "Eval- uating Value-Added Models for Teacher Accountability," RAND, Report.

Oates, Wallace E. 1969. "The Effects of Property Taxes and Local Public Spending on Property Values: An Empirical Study of Tax Capitalization and the Tiebout Hypothesis." Journal of Political Economy, 77(6): 957-71.

Ozier, Owen. (2018). "The Impact of Secondary Schooling in Kenya: a Regression Discontinuity Analysis," Journal of Human Resources 53 (1): 157-188.

Tiebout, Charles. 1956. "A Pure Theory of Local Expenditures." Journal of Political Economy 64(5):416-24.

Turner, M. A., Haughwout, A. and van der Klaauw, W. (2014), "Land Use Regulation and Wel- fare." Econometrica, 82:1341-1403.

United Nations Population Division. (2014). World Urbanization Prospects.

110 Figure 1: Lahore Administrative and Political Boundaries

Major Roads - Administrative Units (TMAs)

- Political Units (Provincial Assenblies)

Note: This figure shows administrative boundaries (blue); political boundaries (black); and major roads (green) in Lahore.

111 Figure 2: Fixed Effects Spatial RDD

SmE *.s **. 11

U U

Note: This figure shows a stylized example of grids observed in a city segmented into N =5 political units. I define a connected set to be the set of political units that can be linked via a boundary. In this example, the connected set contains all political units within a single city: ai is linked to a2 and a4 ; a2 is linked to a3; and awis linked to ao. Within any connected set, we can normalize a1 = 0, and estimate all as relative to &i. Assuming constant treatment effects across locations, each cr3 estimated in equation (1) can be ranked with one another to form estimates of (relative) political unit quality.

112 Figure 3:

Provincial Assembly Estimates, Services Index 0-

LO-

-. 1 -.05 0 .05 .1

Estimates Posteriors

Note: This figure shows the distribution of estimated fixed effects (blue) and posteriors (red) for service provision variation across provincial assembly boundaries. Fixed effects are estimated using the spatial RDD design outlined in Section 3.1. Estimates are reweighted using an empirical Bayes procedure as described in Section 3.2. The empirical Bayes procedures "shrinks" the estimates towards the mean by a signal-to-noise ratio.

113 Figure 4A: Posterior Distributions

Streetlights Index Number of potholes

..,.8

8 _?;,C"l ·;;; C: ogQ) N

0 -.004 -. 002 0 .002 .004 -.0002 -.0001 0 .0001 .0002 1--- Province --- National --- Admin I 1--- Province --- National --- Adm in I

Service Index

-1 -.5 0 .5 1--- Province --- National --- Admin I

Note: This figure plots the kernel density of posteriors for provincial assemblies, national assemblies, and administrative units. In each figure, I first estimate a set of fixed effects across each type of boundary by estimating the spatial RDD described in Section 3.1 using the specified public good as an outcome variable. Then, I apply an empirical Bayes procedure to "shrink" the estimates towards the mean by a signal-to-noise ratio. This procedure ensures the variation in fixed effects is not overstated due to noise. The variance and standard deviations of these posteriors arc reported in Table 5 . Figure 4B: Posterior Distributions

Number of dumpsters Infrastructure Index g8

;:;, ~ 'iii C Q) 0

C\I~

-.0002 -.0001 0 .0001 .0002 -.5 0 .5 1--- Province --- National --- Admin I 1--- Province --- National --- Admin I

Property Worth (Ln)

(0 i'!- . 'iii C Q) osi:

0 -4 -2 0 2 4 j--- Province --- National --- Admin I

Note: This figure plots the kernel density of posteriors for provincial assemblies, national assemblies, and administrative units. In each figure, I first estimate a set of fixed effects across each type of boundary by estimating the spatial RDD described in Section 3.1 using the specified public good or property valuation as an outcome variable. Then, I apply an empirical Bayes procedure to "shrink" the estimates towards the mean by a signal-to-noise ratio. This procedure ensures the variation in fixed effects is not overstated due to noise. The variance and standard deviations of these posteriors- are reported in 'Table 5. Figure 5A: Correlates of political unit quality (Provincial Assemblies)

Age

Education I Amount tax paid

Incumbent

Minority party

Registered voters

Turnout (%)

Winning margin

-0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5

0 Property Valuation (2016)

Age

Educatio n

Amount tax paid

Incumbent

Minority party

Registered voters

Turnout (%)

Winning margin

-0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3

U Road Quality S Streetlights Index

Note: This figure shows the magnitudes of Ridge regression coefficients correlating provincial assembly quality and politician and constituency observables. Provincial assembly quality is measured by changes in property valuations (top panel) and public goods provision (bottom panel) across political boundaries. The politician observables refer to the politician in office from 2013 to 2018. The constituency observables refer to the 2013 elections.

116 Figure 5B: Correlates of political unit quality (National Assemblies)

CDF spending No

Amount tax paid

Minority party 0

Registered voters

Turnout (%)

Winning margin

-0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7

0 Property Valuation (2016)

CDF spending

Amount tax paid

Minority party II

Registered voters

Turnout (%) I

Winning margin U

-1.5 -1 -0.5 0 0.5 1 1.5

S Road Quality &Streetlights Index

Note: This figure shows the magnitudes of Ridge regression coefficients correlating national assembly quality and politician and constituency observables. National assembly quality is measured by changes in property valuations (top panel) and public goods provision (bottom panel) across political boundaries. CDF spending corresponds to the years 2008-2013. The politician observables refer to the politician in office from 2013 to 2018. The constituency observables refer to the 2013 elections.

117 Table 1: Balance Checks using 2002 Tax Records

Tax Record in 2002 Updated Tax Number of before 2002? Liability floors Area (1) (2) (3) (4) PanelA: Provincial F-statistic 1.21 1.4072* 0.5491 0.8943 P-value [0.2013] [0.0867] [0.9678] [0.618]

R2 0.5055 0.3526 0.2388 0.4888 Observations 965 846 846 839 Panel B: National F-statistic 1.0555 1.2425 1.5906* 0.5792 P-value [0.3946] [0.2306] [0.0666] [0.9002]

R2 0.4048 0.3819 0.266 0.1331 Observations 838 623 623 618 Panel C: Administrative F-statistic 2.1103** 1.3391 1.4072 0.3965 P-value [0.0178] [0.1995] [0.166] [0.9574]

R2 0.3358 0.3158 0.2045 0.1263 Observations 750 562 562 556 Notes: This table reportes F-statistics from the null hypothesis that all fixed effects estimated in equations (1) and (2) are equal to 0. P-values are reported in brackets. Standard errors are adjusted for spatial correlation using Conley (1999). Panel A reports p values for estimated discontinuities across provincial assembly boundaries; Panel B for discontinuities across national assembly boundaries; and Panel C for discontinuities across TMA boundaries. Each specification includes observations within 250 meters of the boundary.

118 Table 2A: Service provision

Broken Service Street Lights Potholes Broes Trashpiles Index manholesIne (1) (2) (3) (4) (5) Panel A: Provincial F-statistic 2.1242*** 1.8856*** 0.9828 0.7551 1.4835** P-value [0.0002] [0.0019] [0.4945] [0.8376] [0.0394]

R2 0.25 0.0495 0.0622 0.1139 0.1139 Observations 3596 3596 3596 3596 3596 Panel B: National F-statistic 1.065 1.5343* 0.8652 1.1351 1.7416** P-value [0.3837] [0.0793] [0.6104] [0.3155] [0.0336]

R2 0.1368 0.0901 0.0541 0.0453 0.067 Observations 2387 2387 2387 2387 2387 Panel C: Administrative F-statistic 7.7346*** 1.2391 1.4636 0.8124 0.7826 P-value [0] [0.2496] [0.1307] [0.638] [0.6691]

R2 0.1394 0.0955 0.071 0.0634 0.0481 Observations 2229 2229 2229 2229 2229 Panel D: Placebo F-statistic 1.1426 1.0353 0.2997 0.5955 0.1969 P-value [0.3178] [0.4139] [0.9922] [0.8594] [0.9991]

R2 0.1765 0.0892 0.0356 0.0673 0.0216 Observations 1502 1502 1502 1502 1502 Notes: This table reportes F-statistics from the null hypothesis that all fixed effects estimated in equations (1) and (2) are equal to 0. P-values are reported in brackets. Standard errors are adjusted for spatial correlation using Conley (1999). Panel A reports p values for estimated discontinuities across provincial assembly boundaries; Panel B for discontinuities across national assembly boundaries, Panel C for discontinuities across TMA boundaries, and Panel D for discontinuities across placebo boundaries. Placebo boundaries are generated by shifting the provincial assembly boundaries. Each specification includes observations within 250 meters of the boundary.

119 Table 2B: Infrastructure provision

Dumpsters Public water taps Filtration plants InfrastructureIndex

(1) (2) (3) (4) Panel A: Provincial F-statistic 1.5065** 0.5266 0.739 1.2881 P-value [0.0338] [0.987] [0.8562] [0.1289]

R2 0.0511 0.1192 0.0563 0.0484 Observations 3596 3596 3596 3596 Panel B: National F-statistic 0.5104 0.782 0.6493 0.86 P-value [0.9434] [0.7077] [0.8451] [0.6165]

R2 0.0667 0.0869 0.0787 0.0671 Observations 2387 2387 2387 2387 Panel C: Administrative F-statistic 1.9917** 2.0851*** 0.4534 3.7634*** P-value [0.0215] [0.0151] [0.9414] [0]

R2 0.0878 0.1036 0.0584 0.0481 Observations 2229 2229 2229 2229 Panel D: Placebo F-statistic 5.011*** 1.9916** 0.6559 0.1791 P-value [0] [0.0183] [0.8072] [0.9995]

R2 0.135 0.0756 0.0303 0.0216 Observations 1502 1502 1502 1502 Notes: This table reportes F-statistics from the null hypothesis that all fixed effects estimated in equations (1) and (2) are equal to 0. P-values are reported in brackets. Standard errors are adjusted for spatial correlation using Conley (1999). Panel A reports p values for estimated discontinuities across provincial assembly boundaries; Panel B for discontinuities across national assembly boundaries, Panel C for discontinuities across TMA boundaries, and Panel D for discontinuities across placebo boundaries. Placebo boundaries are generated by shifting the provincial assembly boundaries. Each specification includes observations within 250 meters of the boundary.

120 Table 3: Self-Reported Property Worth (Ln)

Survey 2016 Survey 2013

(1) (2) (3) Panel A: Provincial F-statistic 2.6266*** 0.8816 1.4129* P-value [0] [0.6612] [0.0867]

R2 0.4528 0.3914 0.6342 Observations 455 1659 614 Panel B: National F-statistic 0.6768 0.7004 1.7127** P-value [0.807] [0.8212] [0.0425]

R2 0.4784 0.3287 0.5101 Observations 372 1126 442 Panel C: Administrative F-statistic 1.9449** 1.0116 1.387 P-value [0.0394] [0.4337] [0.1773]

R2 0.4618 0.3899 0.6092 Observations 332 939 385 Panel D: Placebo F-statistic 1.8029* 1.0226 1.1533 P-value [0.0822] [0.428] [0.3284]

R2 0.4752 0.384 0.5951 Observations 172 527 223

Tax record 2002 controls X Notes: This table reportes F-statistics from the null hypothesis that all fixed effects estimated in equations (1) and (2) are equal to 0. P- values are reported in brackets. Standard errors are adjusted for spatial correlation using Conley (1999). Panel A reports p values for estimated discontinuities across provincial assembly boundaries; Panel B for discontinuities across national assembly boundaries, Panel C for discontinuities across TMA boundaries, and Panel D for discontinuities across placebo boundaries. Placebo boundaries are generated by shifting the provincial assembly boundaries. Each specification includes observations within 250 meters of the boundary.

121 Table 4: Resident Characteristics

Survey 2016 Survey 2013

Resident HH Size HH Exp. Resident HH Size HH Exp.

(1) (2) (3) (4) (5) (6) Panel A: Provincial F-statistic 1.3235 0.929 0.92 1.2441 1.266 1.1496 P-value [0.1263] [0.5749] [0.5884] [0.1619] [0.1438] [0.2572]

R2 0.3461 0.3085 0.4742 0.2492 0.243 0.3698 Observations 498 540 516 1675 1806 1791 Panel B: National F-statistic 0.9734 1.3295 2.1346*** 1.3659 0.814 0.9224 P-value [0.4833] [0.1761] [0.007] [0.1347] [0.6919] [0.5547]

R2 0.3007 0.3275 0.5129 0.3057 0.2166 0.3454 Observations 384 430 410 1144 1223 1208 Panel C: Administrative F-statistic 1.079 0.9883 1.9515** 1.3031 1.0253 2.1611** P-value [0.3781] [0.4533] [0.0379] [0.2172] [0.4213] [0.0145]

R2 0.3438 0.2604 0.3795 0.2743 0.1979 0.3419 Observations 347 393 386 992 1060 1058 Panel D: Placebo F-statistic 1.1466 0.5838 1.6256 2.6128*** 1.2652 1.1298 P-value [0.3358] [0.7902] [0.1219] [0.0016] [0.2304] [0.3309]

R2 0.3452 0.2426 0.5436 0.3114 0.2776 0.4148 Observations 193 206 195 544 570 565

Tax record 2002 controls Notes: This table reportes F-statistics from the null hypothesis that all fixed effects estimated in equations (1) and (2) are equal to 0. P-values are reported in brackets. Standard errors are adjusted for spatial correlation using Conley (1999). Panel A reports p values for estimated discontinuities across provincial assembly boundaries; Panel B for discontinuities across national assembly boundaries, Panel C for discontinuities across TMA boundaries, and Panel D for discontinuities across placebo boundaries. Placebo boundaries are generated by shifting the provincial assembly boundaries. Each specification includes observations within 250 meters of the boundary.

122 Table 5: Assessing magnitudes

Province National Administrative Variance FE 2.88 0.34 0.17 Std. Dev FE 1.70 0.59 0.42 15.72 Property Valuation Mean Dep. Var. 15.72 15.72 (Ln) Province vs. National 0.0003 Province vs. Admin. 0.0004 National vs. Admin. 0.3409 Variance FE 0.0070 0.0335 0.0164 Std. Dev FE 0.0836 0.1829 0.1280 Mean Dep. Var. 0.0013 0.0015 0.0014 Streetlights Index Provincevs. National 0.0000 Provincevs. Admin. 0.038 1 National vs. Admin. 0.0940 Variance FE 0.0000 0.0000 0.0000 Std. Dev FE 0.0047 0.0052 0.0009 Mean Dep. Var. 0.0000 0.0000 0.0000 Dumpsters Province vs. National 0.9691 Province vs. Admin. 0.0078 National vs. Admin. 0.0287 Variance FE 0.0000 0.0000 0.0000 Std. Dev FE 0.0037 0.0068 0.0044 Mean Dep. Var. 2.13E-05 3.27E-05 2.03E-05 Potholes Province vs. National 0.0040 Province vs. Admin. 0.9850 National vs. Admin. 0.0492 Variance FE 0.0000 0.0000 0.0000 Std. Dev FE 0.0008 0.0008 0.0002 Mean Dep. Var. 0.0000 0.0000 0.0000 Broken manholes Province vs. National 0.1516 Province vs. Admin. 0.0012 National vs. Admin. 0.2363 P-values from Brown-Forsythe test of equality of variances.

123 Appendix A

Figure Al: Surveyed properties

(a) Lahore 2013 Survey

(b) Lahore 2016 Survey

Note: This figure displays surveyed properties in 2013 and 2016 in Lahore. Political units are outlined in black; administrative units in blue. Major roads are displayed in green.

124 Table Al: Public goods summary statistics

PCilitical Jurisdiction Administrative Unit Mean Minimum Maximum Mean Minimum Maximum (1) (2) (3) (4) (5) (6) Street Lights 15.58 0.00 62.50 14.95 9.06 33.51 [13.89] [7.24] Potholes 0.14 0.00 0.50 0.15 0.00 0.40 [0.14] [0.14] Broken manholes 0.04 0.00 0.33 0.03 0.00 0.06 [0.07] [0.02] Trashpiles 0.18 0.00 1.03 0.16 0.01 0.46 [0.23] [0.14] Dumpsters 0.09 0.00 0.33 0.07 0.00 0.25 [0.09] [0.07] Filtration plants 0.01 0.00 0.05 0.01 0.00 0.04 [0.01] [0.02] Notes: Standard deviations in brackets.

125 Table A2: Correlation of property valuation and public goods posteriors

Province National Adminstrative

Property Worth (Ln)

Streetlights 0.0346 0.4809 0.3532 Potholes -0.272 0.3414 0.1825

Broken manholes -0.172 0.1043 -0.4534

Trash 0.149 0.1916 0.2672 Dumpsters 0.0374 0.0189 -0.1134 Taps 0.0557 -0.081 -0.3238

FiltrationPlants 0.0002 -0.5036 0.0144 Notes: This table reports correlations between posteriors that estimate changes in property worth with posteriors that estimate changes in public goods provision. The correlations are computed separately for Drovince. national. and adminsitrative units.

126 Table A3: Voting behavior near provincial assembly boundaries

Vote Vote PML-N (1) (2) F-statistic 1.2654 1.3294 P-value [0.1654] [0.1221]

R2 0.359 0.3411 Observations 543 516 Notes: This table reportes F-statistics from the null hypothesis that all fixed effects estimated in equation (1) are equal to 0. P-values are reported in brackets. Standard errors are adjusted for spatial correlation using Conley (1999). Each specification includes observations within 250 meters of the boundary.

127 128 Chapter 3

Rebuilding the Social Compact: Urban Service Delivery and Property Taxes in Pakistan

(Joint work with Adnan Q. Khan, Asim Khwaja, and Benjamin Olken)*

3.1 Introduction

The social compact between citizen and state - whereby a citizen pays taxes and receives public goods and services - is a critical link in the development process. This link is especially salient in the context of local governments, and a significant metric by which they are judged. However, if citizens perceive little benefit from their tax payments, or if local services are disconnected from local decision-making, the link between citizen and state can be broken. This can create a vicious cycle where citizens do not receive high quality services because resources are limited by low levels of local tax revenue. In turn, the low quality of services

*This project is the result of collaboration among many people. We thank Ali Abbas, Ismail Khan, Yanchuan Liu, Guillermo Palacios Diaz, and Fay Terrett for outstanding research assistance in Cambridge and Osman Haq, Khawaja Hussain, Ahmad Mustafa, Omer Qasim, Adeel Shafqat, Mehak Siddiquei, Shahan Shahid and the late Sadaqat Shah for outstanding research assistance in Lahore. We thank all the Secretaries, Director Generals, Directors, the Project Directors from the Punjab Department of Excise and Taxation, the Punjab Finance, Local Government and Community Development, Planning and Development departments and the Chief Secretary and Chief Minister's offices for their support over the many years of this project. This RCT was registered in the American Economic Association Registry for randomized control trials under trial number AEARCTR-0003270. Financial support for the project came from 3ie, the IGC, JPAL-GI, JPAL- USI, NSF, and STL. The views expressed here are those of the authors and do not necessarily reflect those of the many individuals or organizations acknowledged here.

129 leads to a low willingness to pay taxes, and a broader lack of trust in the state. Though policymakers in developing countries regularly express concern regarding the suboptimal equilibrium of low revenue generation and low public good provision, few studies have examined whether strengthening the link between taxation and service provision can increase citizens' willingness to pay taxes. Simply informing citizens of the tax-benefit link does not seem to increase compliance (Blumenthal, Christian, and Slemrod 2001; Scar- tascini 2015). Laboratory experiments show that eliciting and promising to follow taxpayer preferences on government spending, on the other hand, can increase tax compliance (Alm, Jackson, and McKee 1993; Lamberton, De Neve, and Norton 2014). Other studies have found correlations between survey results on tax morale and service provision (OECD 2013), or have linked ex-post tax compliance to public service provision (Gonzalez-Navarro and Quintana-Domeque 2014). But there is still little evidence of the link between taxes paid and services delivered in the real world. This paper provides what is to the best of our knowledge the first experimental evidence on this question. We partner with the Punjab, Pakistan provincial government to implement a series of interventions that strengthen the link between taxes and services in several ways - from simply eliciting taxpayers' preferences over local services, to earmarking a portion of tax revenues to be allocated to taxpayers' neighborhoods, to earmarking a portion of tax revenues to be allocated to taxpayers' neighborhoods according to their preferences. To the extent that citizens are reminded of the link between taxes and services or perceive public goods to more accurately reflect their preferences, they may experience a higher disutility from tax evasion, and tax compliance may increase. If so, these interventions could be a powerful policy reform and have positive implications over and above those on tax compliance by positively impacting citizens' views on and relationship with the state. We implement the interventions in a large-scale randomized controlled trial in La- hore and Faisalabad, the two largest cities in Punjab. We first construct a sample of 500 neighborhoods, comprising of 100 to 400 contiguous taxable properties. Each neighborhood is assigned to one of three interventions for two rounds: Local Allocation, Voice, or Voice- based Local Allocation. In Local Allocation neighborhoods, the local government commits to allocating a portion (35%) of property tax collected from a neighborhood to service provision in that same neighborhood. In the status quo, revenue is collected from larger administrative tax units and distributed at the city-level for services - so citizens currently have no sense of

130 how much of their taxes, if any, is spent on services within their locality, let alone how these services are chosen, or whether these services are the ones they desire. The Local Allocation intervention strengthens the geographic link between taxes paid and services provided. In Voice neighborhoods, citizens are asked to provide preferences on the types of local goods and services should be prioritized in their neighborhood. Currently there is no explicit process, aside from constituency politics, through which citizens can express their preferences on what spending should occur in their neighborhood. In the framework of the 2004 World Development Report, this approach - influencing policy through elections - is known as the "long route" to accountability. While it can be effective for broad policies, more direct approaches - known as the "short route" - can be more effective for building links between citizens and government. Evidence from other contexts suggests that allowing more direct participation in this process - i.e. "short route" approaches - can increase the perceived legitimacy of political decisions (Olken 2010). Citizen preferences in each Voice neighborhood are aggregated and shared with the local government in an attempt to improve the allocation of services. The intervention tests whether increasing citizen voice in the decision making process affects the type and quality of local public goods provided and, in turn, increases citizen willingness to pay for these services through greater tax compliance and tax morale. The Voice-based Local Allocation intervention combines preference elicitation and local allocation. Though more than 70% of local property taxes are designated for local goods and services in the status quo, simply eliciting preferences and providing them to the local government may not be sufficient to change taxpayer beliefs and attitudes if trust in the system is low. Similarly, simply earmarking funds for local allocation may be insufficient to change taxpayer behavior, as taxpayers may be uninformed about these efforts. In Voice- based Local Allocation neighborhoods, the local government allocates 35% of property tax revenue they receive to the specific goods and services requested by taxpayers. Citizens are informed of this earmarking, and the resultant service expenditures are indeed carried out in their locality. Since local politicians can play a critical role in building (or hindering) citizen trust, and in monitoring the provision of local services, we also examine how local politicians can facilitate the tax-service linkage and how that in turn affects citizens' attitudes towards the state and politics. We cross-randomize intervention neighborhoods into a Local Leader treat-

131 ment, in which we assign a local politician to coordinate taxpayer mobilization efforts in his constituency. These mobilization efforts aim to increase awareness of the scheme, encourage taxpayers to submit tax payments punctually and accurately, and remind taxpayers of the link (established by the scheme) between taxes paid and services delivered. We evaluate the impact of these interventions on a range of outcomes including tax payments, tax morale, public goods quality, and attitudes towards the state. Throughout the study, we collect monthly property-level tax data on tax assessments, tax payments, and the timing of tax payments. We supplement this data with baseline and endline surveys of a representative sample of properties in our sample, collecting detailed data on tax morale, perceptions of service quality, voting behavior, and engagement' with the state. To gain an objective measure of public goods provision, we measure the extent and quality of street lighting, roads, sanitation, and water in our neighborhoods before and after the interventions. Finally, we track the local politician response to the interventions in a survey of all politicians in our sample. We have completed the first round of interventions in nearly all neighborhoods. Cit- izens in Voice and Voice-based Local Allocation neighborhoods provided their preferences for services in their neighborhoods. These preferences were aggregated and shared with the local government. In Local Allocation and Voice-based Local Allocation neighborhoods, the local government earmarked 35% of tax revenue for service provision. Service delivery is nearing completion, with citizens in the Local Allocation group receiving some service financed by their tax revenue, and citizens in the Voice-based Local Allocation group receiving their preferred service financed by their tax revenue. We find the interventions led to a 7 percentage point increase in tax payments in the first fiscal year of implementation - even before services were delivered to all Local Allocation and Voice-based Local Allocation neighborhoods. The increase is largely concentrated in Local Allocation neighborhoods, though we estimate positive (but insignificant) effects on tax payments in all three interventions. Given these results, we expect to see larger effects on tax payments in the second fiscal year, once service delivery is complete and the perceived credibility of the interventions improve. The remainder of the paper is organized as follows: Section 3.2 provides a brief overview of the setting in which the study takes place. We present a simple model of tax compliance in Section 3.3. Section 3.4 outlines the experimental design. Section 3.5

132 describes the data and the empirical framework. Section 3.6 reports on the implementation of the interventions, and first year impacts on tax payments. Section 3.7 concludes.

3.2 Empirical Setting

The study takes place in Lahore and Faisalabad, the two most populous cities in Punjab with populations of 18 and 4 million, respectively. Like many developing countries, Pakistan has experienced a wave of urbanization over the last few decades with nearly 40% of the population currently residing in cities. Social and urban services, however, have not kept pace. Even compared to other countries in the region, Pakistan is an outlier in public goods quality (ADB 2014). Part of the reason for low public goods quality is insufficient finances. Public goods are financed by the Punjab Urban Immovable Property Tax (UIPT)', but revenue from this tax is abysmally low. This property tax accounts for only one tenth of one percent of Punjab's GDP (IGC 2015), which is roughly a fifth of the level of countries comparable to Pakistan (World Bank 2006). Many problems contribute to Punjab's low property tax revenue, including a narrow tax base, tax rates that do not reflect properties' true market value, tax evasion, corruption and poorly incentivized tax collectors (Khan et al. 2016). We work with the Punjab Excise and Taxation Department (E&T), which collects property tax, and the Punjab Local Government Department, which provides services. The E&T department levies property tax based on a formula that takes into account property and neighborhood characteristics. These include square footage of land and covered area, property use, occupation status, and locality quality ratings. Property tax is collected by tax inspectors, who are responsible for determining a property's tax liability and sending the annual tax bill to the property owner. Prior to 2016, property records were stored manually in separate offices; after a digitization campaign in 2016, the records are now stored in a secure online database that allows all tax inspectors to access historical and current property records for any assessed property in six major cities in Punjab. Though a property's tax liability is formulaic, it is based on the tax inspector's assessment of the property. This assessment is not verified by third parties, leaving considerable discretionary power to the tax inspector in determining the final liability. Collusion between taxpayers and tax inspectors is thought to be widespread, with tax inspectors misreporting

'Supplemental grants from the provincial and federal governments also provide funding for public goods.

133 property characteristics (or leaving out a property from the tax rolls entirely) to lower liabilities. Taxpayers can also evade taxes by simply paying less than the assessed amount. The difference between the tax liability and payment is added to an arrears account, and carried over to the subsequent fiscal year. The Local Government Department is responsible for using property tax revenue to provide local public goods in each city. These services include street lights, road repair, sanitation, waste removal, and water.2 Though 70% of property tax revenue is currently allocated for local public goods, our baseline survey shows most residents believe little if any of property tax is ultimately used for this purpose. Given the low quality of urban services, citizens in both Lahore and Faisalabad are increasingly choosing to "opt-out" of the social compact entirely by relying on non-state actors for service provision. In the last two decades, a significant proportion of the upper middle class and elite have moved to or formed private housing societies that charge residents fees to finance services within their neighborhoods. Others have chosen to remain where they are, but outsource certain services to private companies. In our surveys, we measure not only citizen perceptions of service quality, but also their reliance on state and non-state actors for service provision. For those relying on services provided by the Local Government Department, a citizen's first point of contact for municipal concerns is often a local politician. These local politicians are members of Union Councils, which are formally responsible for monitoring delivery of municipal services, maintenance of public areas, community mobilization and dispute resolution. Because these Councils have little to no development funds of their own, their main function is to intermediate between citizens and service providers. Current members of the Union Councils were elected in 2016 for a five-year term. Even though they are recently elected, a typical Union Council politician has been active in his or her neighborhood unofficially as a political broker, and is well-known in the community.

2In practice, sanitation and waste removal is outsourced to a separate department, the Waste Management and Sanitation Agency (WASA). We work with this agency directly to provide dumpsters and trash removal services in our neighborhoods.

134 3.3 Conceptual Framework

We present a simple conceptual framework to illustrate how strengthening the link between taxes and services can affect tax compliance. Suppose each taxpayer i faces a true tax liability

u(yi - Ti*) > (1 - r)u(yi - 7j) + 7Tru(yi - ri* - q)

Given reasonable audit probabilities and fines, however, this simple model is unable to explain observed rates of tax compliance, even in developing countries. A number of studies have highlighted other factors such as social sanctions (Luttmer and Singhal, 2014) and intrinsic motivation (Dwenger at al. 2016) to explain tax compliance. Here, we allow individuals' utility from paying taxies to be influenced by citizens' relationship to the state. This relationship may be influenced by the provision of desired public services. If the government is unable to provide high-quality services or if provided services do not match citizen preferences, then citizens' utility from paying taxes decreases, making them less willing to pay and more likely to evade. There are (at least) three factors that may be important to a taxpayer when con- sidering the social compact with the state:

1. Tax morale: Let qi capture tax morale, defined as a taxpayer's willingness to pay taxes to the state, even though the taxpayer's pecuniary payoff would be higher if she did not pay (Luttmer and Singhal 2014).

2. Type of public good: Let Ri(x) represent a taxpayer's ranking of a public good X, which can be financed from tax revenue. For example, a taxpayer may prefer her taxes are spent on repairing street lights in her neighborhood, which she considers to be more critical than trash collection. Ri(street lights) would therefore be higher than Ri (trash collection).

3. Amount of revenue spent on public good: Let -y(x) denote the (perceived) fraction of local revenue spent on the local public good. In each neighborhood of size n, the n government spends -y(x) ji of total local revenue on public good x in neighborhood i=1

135 We consider a two-period model. In the first period, the taxpayer chooses whether to pay taxes. In the second period, the taxpayer receives a public good financed from period 1 taxes. For simplicity, we assume a risk-neutral taxpayer, and allow the taxpayer to either pay the full tax liability i*, or pay 0. Taxpayer utility when paying and not paying the tax is given by:

Utility if payment: - r * + qi + 3Ri(x) 7(x) r

n Utility if no payment: 7r(-i - q) + 3Ri(x) [.(x) r

This implies the taxpayer will pay so long as:

-r1* - r(-* - q) + 0i + fRi(x)-y(x)T,* > 0

The willingness to pay taxes is increasing in tax morale qj, the preference ranking of the government's choice of public good Ri(x), and the proportion of revenue allocated to the taxpayer's own neighborhood. 3 Each of our interventions varies a parameter or set of parameters in this model. The Local Allocation intervention sets -y(x) to 35%, which is higher than the average fraction of local revenue taxpayers believe is spent on local public goods at baseline. The Voice intervention, which aggregates citizen preferences and transfers them to the local government, may increase tax morale by making citizens feel their voice is heard by the government. If the government chooses to act on citizen preferences, R(x) will also increase as the government selects a public good more aligned with citizen preferences. Finally, the Voice-based Local Allocation intervention sets -y(x) to 35% and ensures R(x) increases by mandating the selection of a public good that is ranked highly by citizens. Tax morale may increase as citizen preferences are collected, as in the Voice intervention, and in subsequent periods as public goods are delivered and citizens view the government delivering on its promise.

3 For simplicity, we ignore preferences for redistribution. In a more realistic model, taxpayers would have preferences over the proportion of revenue allocated not only to their own neighborhood, but to other neighborhoods as well.

136 3.4 Experimental Design

This section provides details on the sample construction (Section 3.4.1) and experimental design (Section 3.4.2).

3.4.1 Sample

The sample consists of 500 neighborhoods, comprising over 100,000 taxpayers in Lahore and Faisalabad. These neighborhoods were identified using geo-referenced property-level administrative data according to several key parameters. Neighborhoods were constructed to consist of around 100 to 400 contiguous taxable residential or commercial properties. This size ensures each neighborhood is small enough so that the incentive to free-ride is not too strong, citizen preferences can be aggregated, and the goods or services provided are utilized by most taxpayers in the neighborhood. However, the neighborhood size is not so small that the goods or services cannot be financed from local revenue. Neighborhoods also have a high density of taxable properties so that a large proportion of residents or shopkeepers in the neighborhood can potentially contribute revenue for service provision. Finally, neighborhoods are defined to be contiguous so there is some sense of social cohesion among taxpayers. This social cohesion may facilitate tax compliance by allowing taxpayers to encourage their neighbors to pay taxes. Figure 1 displays neighborhoods in Lahore and Faisalabad, color-coded by treatment assignment. Our sample of neighborhoods covers areas throughout each city.4 Ap- pendix Figure Al shows how neighborhoods were identified using geo-referenced property data.

3.4.2 Main Interventions

Neighborhoods were randomly allocated to one of three interventions: Local Allocation, Voice, or Voice-based Local Allocation. In addition, we cross-randomized neighborhoods into Local Leader interventions, where members of the Union Councils were encouraged to mobilize the community to enhance the tax-service link. Randomization was stratified by

4 One exception is the south-east portion of Lahore, which is excluded since this region consists primarily of private housing areas that are exempt from property tax and rely primarily on non-state actors for service provision.

137 income and property use (residential or commercial) to allow for the estimation of impact variation across sub-groups. We implemented the randomizations in public lotteries with representatives from the tax staff, local government, and Union councils present. Table 1 presents the experimental design. 350 neighborhoods were assigned to one of the three main interventions, while the remaining 150 neighborhoods formed the control group. In each intervention group, we randomly selected half of the neighborhoods to also receive the Local Leader intervention. The experiment is implemented for two rounds of preference elicitation and service delivery to trace dynamics as citizens see how their preferences are acted upon during the first year. Below, we describe each intervention and conclude with a discussion of citizen understanding of the interventions.

Local Allocation: In the status quo, revenue is collected from administrative tax units and transferred to local governments that allocate these to city-level services. However, there is no linkage between taxes paid and services received at a lower and likely more salient geographical unit: the neighborhood (a contiguous set of typically 100-400 households). To strengthen the link between taxes paid and services provided, local governments commit to allocate a portion (35%) of property tax collected from a Local Allocation neighborhood to that same neighborhood. Citizens are informed of this linkage during a door-to-door campaign from the tax authority, and through informational flyers, shown in Appendix Figure A2. In the door-to- door campaign, the tax authority conveyed intervention information via a smartphone app. We used GPS and random audio audits to verify that each citizen was contacted, and that the tax authority explained the intervention correctly. On average, the earmarked amount (35% of neighborhood-level property tax revenue) corresponds to Rs. 200,000 (approximately $1,500) and is sufficient to finance a range of services in varying quantities. We computed earmarked amounts using administrative property tax data (described in Section 3.5). The local governments could then select any service from a menu of options so long as the costs did not exceed the budgetary constraint. Service delivery takes place over a four to five month period. First, engineers survey each neighborhood to determine the type, quantity, and location of services to be delivered. Cost estimates are prepared to ensure services can be financed by the total amount of funds allocated to each neighborhood. Proposed services are submitted to the Mayor and relevant

138 government officials of each city for approval, after which contracts for service delivery are tendered. Service delivery is then implemented in each neighborhood. To ensure citizens are aware that services are delivered via the intervention (and not through other government initiatives), a poster or stencil painting is placed on each service. Each poster/stencil painting is color-coded according to the intervention (blue for Local Allocation services; orange for Voice services, and green for Voice-based Local Allocation services). Appendix Figure A3 shows an example of a blue logo visible on a Local Allocation trash can.

Voice: In the Voice intervention, tax staff solicit citizen preferences on which types of local goods and services should be prioritized in their neighborhood. The results of this preference elicitation is shared with the local government in an effort to improve the allocation of services. To collect preferences, tax staff visit each property in a Voice neighborhood and provide information about the intervention via a smartphone app. The tax staff then display a menu of services, and ask the citizen to select his or her top two choices. The preference elicitation screen on the smartphone app is displayed in Appendix Figure A4. We aggregate preferences by identifying the two services that are selected most often by taxpayers in each neighborhood. These preferences are conveyed to the local government, who can then choose whether or not to use these preferences when deciding spending allocations for the upcoming fiscal year. Citizens are also informed that the results of the preference elicitation exercise have been conveyed to the government.

Voice-based Local Allocation: This intervention implements the Local Allocation and Voice interventions in tandem. By both eliciting citizen preferences and requiring local governments to allocate a portion of property tax collected from a neighborhood to that same neighborhood in accordance with these preferences, it seeks to make the tax-services link even more salient and credible. Citizen preferences are collected as in the Voice intervention. The aggregated preferences are conveyed to the local government, but unlike in the Voice intervention, local governments are required to implement desired preferences. We compute the budget constraint for each neighborhood (35% of neighborhood-level property tax revenue) and ask the local government to survey each neighborhood to determine the amounts of each preferred

139 services that can be provided within the constraint. The process for service delivery is the same as in the Local Allocation intervention. Citizens are informed that the results of the preference elicitation exercise will be implemented and the expected timeline for service delivery. Appendix Figure A5 shows an information flyer for a neighborhood assigned to the Voice-based Local Allocation intervention. The flyer shows that citizens in this neighborhood selected new trashcans/dumpsters and streetlight repair as their preferred services. The allocation amount for these services is PKR 383,889 (approx. $3,000). Once services are delivered, a poster or stencil-painting links the service to the intervention, as in the Local Allocation intervention.

Local Leader: In order to understand whether the local political process can enhance the impact of strengthening the tax-service link, we cross-randomize an additional intervention that enables local politicians to directly support the effectiveness of the three schemes. The local politicians are members of Union Councils, local government bodies responsible for monitoring public services, dispute resolution, and for delivering certain municipal services. They are both the closest and most accessible political actor for the citizen, and, given their resources and knowledge, an effective intermediary between citizens and state. Indeed these local politicians are invariably the first political point of contact for the citizens' enabling them to better respond to the citizens' needs is therefore likely to be a key building block in rebuilding the citizens' faith in the state. The Local Leader intervention is cross-randomized with all three interventions: Local Allocation, Voice, and Voice-based Local Allocation. Local politicians selected for this intervention are allowed to intervene at different stages, depending on the treatment status of a neighborhood within their constituency: (1) In Voice and Voice-based Local Allocation neighborhoods, local politicians introduce the intervention to taxpayers during town hall meetings; (2) In Voice and Voice-based Local Allocation neighborhoods, local politicians monitor tax staff as they collect taxpayer preferences; (3) In Local Allocation and Voice-based Local Allocation neighborhoods, these politicians monitor and facilitate service delivery, using existing channels to pressure service providers and assisting providers in selecting service locations; (4) finally, in Local Allocation and Voice-based Local Allocation neighborhoods, local politicians hold public events to inaugurate new services and reinforce the link between taxes and services.

140 Knowledge and Credibility of the Interventions: To ensure citizens understood the intervention they were assigned to, the tax authority conveyed the key aspects of each intervention during the door-to-door visits and via informational flyers. Appendix Figure A5 displays the summary of each intervention conveyed to taxpayers. We test citizen understanding of the interventions at three points: First, we test citizen knowledge immediately after the tax staff have explains the intervention on the smartphone app. Second, we conduct random field and phone spot checks of citizens after the tax staff visit. Finally, we test citizen knowledge again in the endline survey. Though we did not survey citizens about the credibility of the interventions, qual- itative evidence from the field suggests that citizen views ranged from cautiously optimistic to skeptical that the local government would deliver public goods and services according to the experimental design. We believe credibility of the interventions will improve after the first round of service delivery is complete.

3.5 Data and Empirical Framework

3.5.1 Data

We collect detailed administrative and survey data on tax payments, public goods and services, and attitudes towards the state. Below, we describe key aspects of our data.

Administrative tax data: We collect administrative property-level data on tax assessments and tax payments. This data is available from FY2014 onwards on a monthly basis. For each property in our sample, we construct measures of tax assessed, tax paid, and timing of payment timings. The administrative data also contains detailed property characteristics such as property use (residential or commercial), ownership status (owned or rented), and property location (main or off road). In addition, we observe the property's valuation category which captures the quality of facilities and infrastructure in the property's locality. Each property is assigned a valuation category ranging from A to G. These property characteristics allow us to construct a rich set of property-level controls that increase the precision of our estimates.

Property survey data: We survey properties in our neighborhoods at baseline and at the end of the intervention period. Since it would be too costly to survey all 100 to 400

141 taxable properties within each neighborhood, we use a simple randomization strategy to select properties for the survey sample. In particular, we sample three GPS coordinates within each neighborhood and then survey three to four randomly chosen properties around that coordinate. This strategy ensures that surveyed properties form a representative picture of the typical property in each neighborhood. Each property in the baseline sample will be surveyed again at endline so as to create a panel and hence improve statistical precision in the analysis. The survey collects detailed data on usage of and perceptions of quality of relevant urban services, such as water, sanitation, waste removal, street maintenance, and lighting. We also obtain information on tax morale, voting behavior, and attitudes towards the government more generally. We match the property survey data to the administrative data using property identifiers. This allows us to geo-reference each observation in our survey sample.

Neighborhood survey: We supplement data on perceptions of public goods and services quality, with an objective assessment of public goods and services. The objective assessment is based on a street-level survey of each of our 500 neighborhoods. In each neighborhood, enumerators walked on every street tracking the quantity and quality of public goods and services on a smartphone application. Appendix Figure A6 displays the application interface, while Appendix Figure A7 displays geo-referenced public goods within a locality. GPS coordinates were taken for all observations of low-frequency goods (dumpsters, public taps) and every third observation of high-frequency goods (street lights, potholes, trash piles). This data allows us to directly observe the extent and quality of services provided in each neighborhood. By linking the GPS coordinates in the property survey to GPS coordinates in the neighborhood survey, we also use this data to measure heterogeneity in treatment effects by taxpayer proximity to urban services.

Preferences survey: The tax authority collects citizen preferences for local goods and services in Voice and Voice-based Local Allocation neighborhoods. Though this data is used as part of the interventions, it is also interesting in its own right, providing a high-resolution view of how preferences for goods and services vary across time and space. We also randomly select half of the citizens in this sample to provide an assessment of the current quality of services in their neighborhood.

142 Election data: Finally, we gather polling station data from the 2018 national elections to track potential impacts on voting behavior. This supplements the self-reported voting data collected in the property survey.

3.5.2 Empirical Framework

The empirical framework relies on the random assignment of neighborhoods to an intervention or control group. The basic specification for estimating the average treatment effect at the individual property level is given by:

Yinst /3Treatmentn 8 t + '-yi + Xinst + as + Cinst (3.1) where Yinst is an outcome of interest for property i in neighborhood n in stratum s at time t. When possible, we include the baseline level of the outcome variable, Yio. Xint is a set of baseline property-level characteristics, including the log of government assessed property worth, whether the taxpayer was identified as a defaulter at baseline, property use (residential or commercial), occupation status (owned or rented), location (main or off road), and valuation category (A through G). Since the randomization is stratified by income and property use, we include stratum fixed effects, as. Standard errors are clustered at the neighborhood level. To estimate the impact of separate sub-treatments, we estimate the following equation:

Yinst = i1Voicen8 t + 02Allocu8 t + /33VoiceAllocn8 t + TYfo + Xinst + as + Einst (3.2)

Our primary outcomes of interest are tax payments (assessed tax, paid tax, timing of payment), attitudes towards the government (including tax morale), and voter behavior. We also measure treatment impacts on (objective and subjective) measures of public goods quality. We analyze the impact of the treatments for four samples: (1) the full sample of taxpayers; (2) taxpayers who did not pay in full in the baseline year (FY2015 to 2016); (3)

143 taxpayers who did not pay at all in the baseline year; and (4) taxpayers who made a partial payment in the baseline year. We expect the interventions to have the greatest impact on tax payments in the latter three groups, where taxpayers can easily increase tax payments at their baseline assessment.5

3.6 Implementation and First Year Results

This section presents first year impacts of the interventions. In the first year, preferences were collected, aggregated, and shared with the local government in Voice and Voice-based Local Allocation neighborhoods. Service delivery is ongoing In Local Allocation and Voice-based Local Allocation neighborhoods.

3.6.1 Balance Checks

Table 2 reports balance checks on administrative and survey data at baseline. We compare control neighborhoods (Column 1) to all treatment neighborhoods (Column 2) and each treatment group separately (Columns 3 through 5). Each specification includes stratum fixed effects. Standard errors are clustered by neighborhood. The results show covariates are balanced across treatment groups at baseline: out of the 48 comparisons made (12 variables * 4 columns), 3 are significant at the 10 percent level.

3.6.2 Tax morale at baseline

We next present descriptive statistics on tax morale at baseline. Table 3 measures citizen attitudes towards the government (rows 1 through 3); beliefs about the importance of paying taxes (rows 4 through 5), and beliefs about tax compliance and the link between taxes and services (rows 6-8). Most citizens report feeling neutral on whether the government has helped them in the last year or uses tax revenue to provide services. However, more citizens disagree with the statement than agree with it, indicating relatively low levels of trust in the government to represent citizen interests or allocate tax revenue appropriately. Citizens feel strongly that it is important to pay taxes (though this response may be biased by experimenter demand

5 Tax payments can also increase via re-assessments. This is more costly, however, since the taxpayer would have to request a re-assessment from the tax authority.

144 effects), but only if those taxes are used to provide services, providing suggestive evidence that the link between taxes and services is critical for ensuring tax compliance. To elicit citizen beliefs on tax compliance, we asked what proportion of people in a citizen's neighborhood pay taxes, and what proportion of people in the country pay taxes. Neighbors are considered to be more tax compliant, with 33% of citizens reporting they believe 80-100% of people in their neighborhood pay taxes, but only 10% expressing the same sentiment about their countrymen. Finally, the majority of citizens believe only 0-20% of taxes are used to fund services. Only 5% believe the allocated proportion is near 70% - the actual proportion in the status quo.

3.6.3 Preference elicitation

Results from the preference elicitation exercise in Voice and Voice-based Local Al- location neighborhoods are reported in Figure 2. Each histogram shows the distribution of top-preferred services across the 250 neighborhoods by treatment status and geographic region. Strikingly, nearly half of all neighborhoods select street light repair as their top- preferred service. One potential concern with this result is that it may be driven by an ordering effect. Figure A3 shows that repair of street lights appears as the top listed preference on the smart phone app. In the second (ongoing) round of preference elicitation, we randomized the ordering of services. Reassuringly, we find that the distribution in this round is very similar, indicating the results reflect actual demand for street light repair, rather than ordering effects. The bottom panel shows there is some variation in preferences across region. For example, Lahore A, a relatively poorer and older part of Lahore, has higher demand for pot- hole repair, than either Lahore B - a more recently developed part of Lahore - or Faisalabad.

3.6.4 Service delivery

Services are delivered in Local Allocation and Voice-based Local Allocation neighborhoods.' In Local Allocation neighborhoods, services are selected by the local government department, while in Voice-based Local Allocation neighborhoods, services are selected by

6 We also monitor service delivery in Voice neighborhoods to see if they match citizen preferences. But this service delivery is optional, and not mandated by our interventions.

145 citizens. Figure 3 shows the distribution of services is similar in each group, through bureaucrats are more likely to opt for installation of new lampposts to address street lighting issues, while citizens are more likely to select street light repair. To test whether the selected services are more aligned with citizen preferences at baseline in Voice-based Local Allocation neighborhoods, we estimate the following specification:

Alignin, = /VoiceAllocn, + as + 6inst where Aligninst is a dummy equal to 1 if the delivered service matches the top (or top two) baseline preference of taxpayer i in neighborhood n and strata s and 0 otherwise. The sample is all taxpayers in the Local Allocation and Voice-based Local Allocation neighborhoods. Standard errors are clustered at the neighborhood level. We use two measures of citizen preferences at baseline: preferences for specified goods and services (street light repair, lamp posts, pot holes, road carpeting, trash removal, and dumpsters; and preferences for more general service categories (water, street lights, roads, and sanitation.) Table 4, Column 1 shows that only 17% of citizens in Local Allocation neighborhoods received or will receive a service that is aligned with their top specific preference. This average improves slightly to 27% when comparing services to the top two specific preferences in Column 2. The alignment does not improve in Voice-based Local Allocation neighborhoods. In Columns 3 and 4, we match services to baseline preferences for general service categories. Here, the alignment of services and preferences is 4.8 percentage points higher in Voice-based Local Allocation neighborhoods (34% increase). 7 This estimate is significant at the 5% level. The differential effect is larger when assessing the alignment of services to the top two general service categories: citizens in Voice-based Local Allocation neighborhoods are 7 percentage points more likely to receive services aligned with their preferences. We note, however, that the estimated effects for service alignment with the top preferred service category at baseline (Column (3)) and service alignment with the top two preferred service categories at baseline (Column (4)) are not statistically different.

7Note that the local allocation mean is slightly lower (14% vs 17%). This is because the general service categories contain water, which was not offered as a specific service in the interventions.

146 Even though we estimate significantly higher alignment in Voice-based Local Allo- cation neighborhoods, it is surprising that the average is not higher. One possible explanation is that preferences for service provision are not stable, and citizens reported different rank- ings in the baseline and preference survey. Another possible explanation is differences in the method of preference elicitation. In the baseline survey, we asked citizens to provide a full ranking of seven possible services. In the preference survey, we asked surveys to select their top two preferred services out of seven. It is possible that citizens changed their ranking because of these different prompts.

3.6.5 First year impact on tax payments

We now turn to first year impacts on administrative tax payments. The empirical analysis - including specifications, samples, primary outcomes, and selection of controls - follows a pre-analysis plan uploaded on the AEA RCT registry.' The tax payment data is total payments, which include current year tax payments and payments for arrears. We anticipate obtaining data separating current year payments and arrears; when we do so, we will conduct the analysis for current year payments and arrears separately. The current year tax payments are primary outcomes. All specifications use a set of property controls: the log of government assessed property worth, log of total covered area, whether the taxpayer defaulted on payments in the baseline year (FY2015-2016), whether the property is located on a main street, whether the property is rented or owned, the tax valuation category, and whether the property is residential or commercial. In the future, we may use a double-LASSO procedure a la Chernozhukov et al. to refine the control variables. Table 5 shows that tax payments increased in the first year of interventions. As described in Section 3.5.2, we analyze the impact of the treatment for four samples: (1) the full sample of taxpayers; (2) taxpayers who did not pay in full in the baseline year (FY2015-2016); (3) taxpayers who did not pay at all in the baseline year; and (4) taxpayers who made a partial payment in the baseline year. Specifications highlighted in bold indicate our primary outcome. Here, our primary outcome is the treatment effect on any payment, conditional on not paying in full in the baseline year. In the first year of interventions, taxpayers in this sample made payments that were on average 7 percentage points higher

8 The link for the AEA RCT registry is available here: http://www.socialscienceregistry.org/trials/3270/history/33596.

147 than the control group (12% increase). Tax payments are also higher in the first year of interventions conditional on not making any payment in the baseline year. Column (3) shows taxpayers in this sample made payments that were on average 8 percentage points higher than the control group (15% increase). Table 6 presents the same specifications by separating each treatment arm. In Column (2), we find that the effect on tax payments conditional on not paying in full in the baseline year is statistically significant in Local Allocation neighborhoods. In the first year of the intervention, taxpayers in these neighborhoods make tax payments as a proportion of the payable amount due 11 percentage points higher than the control mean. Column (3) shows that taxpayers in Local Allocation neighborhoods also make higher tax payments, conditional on not paying at all in the baseline year. For this group, payments are 14 percentage points higher (26% increase) in the first year of interventions. Though the tax payment increase is significant only in the local allocation treatment, all three treatments have a positive effect on payment amounts. Given that these tax payments were made before most Local Allocation and Voice-based Local Allocation neighborhoods received services, these results are promising. The results show taxpayers responded to the preference elicitation and information components of the treatments. The credibility of the interventions is likely to improve considerably after service delivery is complete. In addition, the precision of the estimates may improve once current year payments are examined separately.

3.7 Concluding remarks

This paper explores whether strengthening the link between taxes paid and services delivered can increase tax compliance, enhance tax morale, and improve service delivery, revitalizing the social compact between citizens and the state. We do so by introducing a series of interventions enabling citizens to better and more credibly express their preferences for types of local urban services and to link the provision of these services with the amount of local taxes paid. The first year impacts of the interventions on tax payments are promising, with taxpayers who failed to make a full payment in the baseline year increasing tax payments by about 7 percentage points in treated neighborhoods. Given that service delivery is ongoing, we anticipate larger effects after taxpayers have received services in their neighborhoods and

148 voiced their preferences again in anticipation for a second round of services. We also expect to capture a richer set of outcomes on tax morale and attitudes towards the state in the endline survey, providing novel empirical evidence on the role of citizen engagement in tax collection and state efficacy, the effect of service provision on tax morale, and the design of taxation and service provision systems.

149 References

Allingham, Michael G. and Agnar Sandmo, (1972). "Income tax evasion: A theoretical analysis". Journal of Public Economics, 1(3-4): 323-338.

Alm, James, Betty R. Jackson, and Michael McKee, (1993). "Fiscal Exchange, Collective Decision Institutions, and Tax Compliance." Journal of Economic Behavior and Organi- zation, 22(3): 285-303.

Asian Development Bank, (2014). Key Indicators for Asia and the Pacific 2014. Manila: Asian Development Bank.

Chetty, Sylvie, and Cecilia Pahlberg, (2014). "Networks and social capital." The Routledge Companion to International Entrepreneurship 117.

Dwenger, Nadja, Henrik Jacobsen Kleven, Imran Rasul, and Johannes Rincke (2014). "Ex- trinsic and Intrinsic Motivations for Tax Compliance: Evidence from a Field experiment in Germany." Working paper.

Gonzalez-Navarro, Marco, and Climent Quintana-Domeque (2014). "Local Public Goods and Property Tax Compliance: Evidence from Residential Street Pavement," Working paper, Economics Department, University of Toronto.

Khan, A. Q., A. I. Khwaja, and B. A. Olken, (2016). "Tax Farming Redux: Experimental Evidence on Performance Pay for Tax Collectors." The Quarterly Journal of Economics, 131(1): 219-271.

Khan, Adnan Q., Asim Ijaz Khwaja, and Benjamin A. Olken, (2019). "Making Moves Mat- ter: Experimental Evidence on Incentivizing Bureaucrats through Performance-Based Postings." American Economic Review, 109 (1): 237-70.

Kleven, Henrik Jacobsen, Martin B. Knudsen, Claus Thustrup Kreiner, Soren Pedersen, and Emmanuel Saez, (2011). "Unwilling or Unable to Cheat? Evidence from a Tax Audit Experiment in Denmark." Econometrica, 79(3): 651-692.

Lamberton, Cait Poynor, J. Emmanuel De Neve, and Michael I. Norton, (2014). "Elicit- ing Taxpayer Preferences Increases Tax Compliance." Working Paper 14-106, Harvard Business School.

150 Luttmer, Erzo F. P., and Monica Singhal, (2014)."Tax Morale,", Journal of Economic Per- spectives, 28(4): 149-68.

Pomeranz, Dina (2015). "No Taxation without Information: Deterrence and Self- Enforcement in the Value Added Tax," American Economic Review, 105(8): 2539:69.

151 Figure 1: Spatial Distribution of Sample Neighborhoods

1 4Mm I-U RIO ci a w A

tP$ U. U. Qi U

4' NK Treament A Control Local Allocation Voice Voice-based LocalAllocation

Note: This figure displays sample neighborhoods in Faisalabad (left panel) and Lahore (right panel). Treatment status is specified by color. Figure 2: Preferences for urban services

Top preference by treatment

0.48 0.45

0.20 C4 0.13 0.13 0.13 0.13 0.10 0.09 0.06 0.06 0.03

0 Voice Voice-based Local Allocation Repair street lights New lamp posts Repair potholes Road carpetting Trash removal New dumpsters

Top preference by city 0.49 i 0.47

0.41

0.31 26 VN- 0.21

0.11 0.11 04- .04

FAISALABAD LAHORE A LAHORE B

Repair street lights New lamp posts Repair potholes Road carpetting Trash removal New dumpsters

Note: These histograms display the distribution of top-preferred services across neighborhoods in the first round of preference collection. The top panel disaggregates preferences by treatment status; the bottom panel by geographic location. Lahore A is a relatively lower income part of Lahore; while Lahore B is a relatively higher income part of Lahore.

153 Figure 3: Service delivery distribution by treatment

CO -

C (DCY

______I I______I______Fix streetlight New streethight Potholes Road carpetting Trash Dumpsters m[:::= Local Allocation [I Voice-based Local Allocation

Note: This histogram displays the distribution of services selected by the local government (Local Allocation) and citizens (Voice-based Local Allocation).

154 Table 1: Experimental Design

Local Leader No Mobilization Mobilization Local Allocation 50 50 Voice 50 50 Voice-based Local Allocation 75 75 Control 150 0 Notes: This table shows the number of neighborhoods assigned to each of the three main interventions (Local Allocation, Voice, or Voice-based Local Allocation). In each intervention group, half of the neighborhoods were cross-randomized to receive Local Leader mobilization.

155 Table 2: Randomization Balance

Control Treatment Local Allocation Voice Voice-based Local Allocation

(1) (2) (3) (4) (5) Panel A: Adminstrative Tax Data Baseline tax payment (log) 8.334 0.031 0.004 0.045 0.044 [1.3111 (0.03) (0.04) (0.03) (0.03) Panel B: Property Characteristics (Baseline survey) Land area (sq. feet) 3230 -562 -532 -322 -766 [16641] (512) (572) (602) (501) Covered area (sq. feet) 3896 -1170 -1266 -1065 -1192 [24294 (761.09) (821.28) (810.02) (763.61) Number of floors 1.855 -0.036 -0.027 -0.072* -0.015 [0.6831 (0.03) (0.04) (0.04) (0.03) Rent property? 0.179 0.006 -0.006 0.028 0 [0.384] (0.02) (0.02) (0.02) (0.02) Panel C: Attitudes towards government (Baseline survey) Satisfaction with Union Council 2.465 0.021 0.038 0.125* -0.056 [0.94] (0.06) (0.07) (0.07) (0.07) Satisfaction with City/Local Govt 2.412 0.091 0.097 0.151* 0.05 [0.95] (0.06) (0.07) (0.08) (0.07) "Govt have helped me in last year." 2.755 -0.008 0.001 0.107 -0.086 [1.18] (0.08) (0.10) (0.10) (0.09) "Govt represent my interests." 2.492 0.016 0.064 0.079 -0.057 [1.11] (0.07) (0.09) (0.08) (0.08) "Govt use tax revenue to provide goods/services." 2.716 0.034 0.099 0.096 -0.051 [1.17] (0.07) (0.09) (0.09) (0.09) "It is important for citizens to pay taxes." 4.213 -0.072 -0.117 -0.028 -0.07 [0.90] (0.06) (0.08) (0.07) (0.07) How much can people like you affect what govt does? 2.338 0.067 0.133 0.043 0.043 [1.25] (0.09) (0.11) (0.11) (0.10) Notes: This table presents balance checks using administrative and survey data at baseline. Standard deviations are reported in brackets. Columns (2) - (5) report difference between control mean and treatment group with stratum fixed effects. Standard errors are clustered by neighborhood and reported in parentheses. The "satisfaction with Union Council" and "satisfaction with city/local govt" variables are coded 1 = dissatisfied, 4 = satisfied. The remaining variables in Panel C are coded from 1 = strongly disagree to 5 = strongly agree. Table 3: Baseline Tax Morale

(1) (2) (3) (4) (5) (6) Mean Strongly Disagree Somewhat Disagree Neutral Somewhat Agree Strongly Agree "Govt has helped me in last year." 2.75 22% 15% 35% 23% 5% [1.17] "Govt represents my interests." 2.5 25% 20% 35% 18% 1% [1.101 "Govt uses tax revenue to provide services for people like me." 2.74 21% 14% 39% 23% 4% [1.13] "It is impt for citizens to pay taxes." 4.16 3% 3% 13% 37% 44% f0.961 "People should only pay taxes if govt gives better services." 4.17 4% 4% 14% 26% 52% [1.07] Mean 0-20% 20-40% 40-60% 60-80% 80-100% What frac. in neighborhood pay taxes? 3.63 10% 10% 18% 27% 33% [1.31] c-fl What frac. in country pay taxes? 2.82 18% 25% 22% 23% 10% [1.26] What frac. of taxes spent on services? 1.7 55% 26% 14% 4% 1% [0.92] Notes: This table reports descriptive statistics on measures of tax morale. Column 1 reports the mean and standard deviation in brackets. The remaining columns report the proportion of respondents who selected each response. Data was collected prior to the interventions in the property tax survey described in Section 3.5.1. Table 4: Service Alignment with Baseline Preferences

Top Preference Top 2 Preference Top Category Top 2 Category (1) (2) (3) (4) Voice-based Local Allocation -0.010 0.018 0.048** 0.070* (0.028) (0.036) (0.024) (0.041) Local Allocation Mean 0.17 0.27 0.14 0.36 Obs 2440 2440 2440 2440 Notes: This table estimates the alignment of baseline preferences for services in Local Allocation and Voice-based Local Allocation neighborhoods. The sample is all taxpayers in these intervention neighborhoods in the baseline sample. Columns (1) and (2) compare services to baseline preferences for specified goods and services (street light repair, lamp posts, pot holes, road carpeting, trash removal, and dumpsters. Columns (3) and (4) compare services to baseline preferences for service categories (water, street lights, roads, and sanitation.) Each regression includes stratum fixed effects. Standard errors are clustered by neighborhood. Table 5: Treatment effect on payment as a proportion of payable amount (any treatment)

Payment as prop of payable amount (1) (2) (3) (4) FY2016-17 (Nov. 16 - June 17) Treatment 0.043 0.070** 0.081** -0.008 (0.026) (0.035) (0.038) (0.043) Baseline FY2015-16 -0.008** 0.045 0.028 (0.004) (0.038) (0.052) Baseline FY2014-15 -0.009*** -0.012*** -0.011*** -0.044*** (0.003) (0.003) (0.003) (0.011) Property characteristic covariates X X X X Did not pay in full in FY2015-2016 X Did not pay at all in FY2015-2016 X Made partial payment in FY2015-2016 X N 33986 21932 18662 3270 Mean of control group 0.648 0.590 0.536 0.909 Notes: OLS regressions of paid/payable ratios on treatment. Outcomes have been top-coded at 1%. The paid variable includes current and arrear payments. We will split the paid variable in future versions of this table. Treatment is a dummy showing whether the observation is in any of the three treatment arms. Property characteristics include log of government assessed property worth, log of total covered area, a dummy of defaulter in 2015, a dummy of location on the main street, a dummy of occupation status (rented or owned), a dummy of valuation category and dummies of residential and commercial use. The highlighted column is the treatment effect on payment as a proportion of the payable amount, conditional on not paying in full in the baseline year. All specifications have stratum fixed effects. Standard errors are clustered by neighborhood. * p<0.10, ** p<0.05, *** p<0.01

159 Table 6: Treatment effect on payment as a proportion of payable amount (treatment arms specified)

Payment as prop of payable amount (1) (2) (3) (4) FY2016-17 (Nov. 16 - June 17) Voice 0.038 0.067 0.068 0.037 (0.035) (0.046) (0.050) (0.059) Local Allocation 0.061* 0.109** 0.140** -0.066 (0.037) (0.050) (0.054) (0.055) Voice-based Local Allocation 0.033 0.047 0.053 0.004 (0.032) (0.042) (0.045) (0.051) Baseline FY2015-16 -0.008** 0.044 0.027 (0.004) (0.038) (0.052) Baseline FY2014-15 -0.009*** -0.012*** -0.010*** -0.044*** (0.003) (0.003) (0.003) (0.011) Property characteristic covariates X X X X Did not pay in full in FY2015-2016 X Did not pay at all in FY2015-2016 X Made partial payment in FY2015-2016 X N 33986 21932 18662 3270 Mean of control group 0.648 0.590 0.536 0.909 Notes: OLS regressions of paid/payable ratios on treatment. Outcomes have been top-coded at 1%. The paid variable includes current and arrear payments. The payable variable is defined using baseline FY15-16 values. We will split the paid variable in future versions of this table. Property characteristics include log of government assessed property worth, log of total covered area, a dummy of defaulter in 2015, a dummy of location on the main street, a dummy of occupation status (rented or owned), a dummy of valuation category and dummies of residential and commercial use. The highlighted column is the treatment effect on payment as a proportion of the payable amount, conditional on not paying in full in the baseline year. All specifications have stratum fixed effects. Standard errors are clustered by neighborhood. * p<0.10, ** p<0.05, *** p<0.01

160 Appendix: Supplemental Figures

Figure Al: Neighborhood Sample Construction

Note: This figure displays neighborhoods identified on Google Earth (left panel) and matched to geo- referenced property data (right panel). Each neighborhood is outlined in black. The geo-referenced property data displays taxable properties in green, and exempt properties in red. Properties are exempt if they are smaller than 5 marlas (about 125 square meters) or belong to widows, the disabled, retired federal and provincial government employees, or religious charitable institutions (Khan et al. 2016).

Figure A2: Information Flyers

Note: This figure displays color-coded flyers distributed to each property in intervention neighborhoods. The Local Allocation scheme is in blue; the Voice scheme in orange; and the Voice-based Local Allocation scheme in green. Each flyer conveys key components of the intervention and a logo.

161 .., A

Figure A3: Local Allocation Service Delivery

Note: This figure displays the blue Local Allocation logo on a trash can provided through the intervention. Color-coded posters (specific to each intervention) are displayed on or next to all provided services.

162 Figure A4: Preference Elicitation

Collect > Test Preference Collection Which of the following services would you want in your neighbourhood? Please select 2 D Repair of street lights Ii Install new lamp post D Repair of pot holes ? Carpeting of roads D Cleaning of gutters I Increase incidence of trash removal D New trashcan/dumpster

Note: This figure displays a screenshot of the Preference Elicitation page of the smaxtphone app. Citizens are asked to select their top two preferred services from a menu of options.

163 Figure A5: Voice-based Local Allocation Information Sheet

CITIZEN PREFERENCE 1: New trashcans/dumpsters CITIZEN PREFERENCE 2: Repair of street lights ALLOCATION AMOUNT: PKR 383,889

Note: This figure displays an information sheet informing citizens of the results of the preference elicitation exercise and the amount that will be allocated to preferred services. Citizens in Green Town III neighborhood selected new trashcans/dumpsters and streetlight repair as their preferred services. The allocation amount is PKR 383,889 (approx. $3,000)

164 Figure A6: Public goods data collection app

CoInsct~Fesupport p

Ckh on the pKcture Coroprondwiq to *t+at y-u We

Note: This figure displays the data collection app used to identify public goods within localities.

Figure A7: Geo-referenced public goods

- -*

Note: This figure displays public goods observed identified within a locality collected in the 2016 survey discussed in Section 3.4.1. Major roads are displayed in yellow. Surveyors walked on every street in 500 localities within Lahore and Faisalabad to identify street lights, pot holes, dumpsters, trash piles, and public water taps. Each dot displays a different public good.

165