How Does Electoral Competition Affect Bureaucratic Performance Part 2

Proposition 1: The unique optimal contract solving the second-best problem is characterized by: e∗2 = q q2∗ = (w + r − V ) (8) w+r−v r (9) 2 4 4 1 e∗1 = αR 7 [r + βV ∗ ] 7 q1∗ = αR 7 [r + βV ∗ ] 7 V ∗ solves: 8 9 2αR 7 [r + βV ] 7 − −7 1 3 − R(w + r − V ) 4 = 0 r 16 V 0∗ = 0 (10) (11) (12) (13) −2 7 where α = 8 . Proof: See Appendix.  3.3 Comparative Statics: Effect of Electoral Competition We now examine how the optimal level of effort implemented by the politician changes as level of competition changes. Case I: Pure re-election concerns, inability to use dynamic incentives In this case, the politician can only use static contracts each period. This means β(φ) = 0 and 2 4 e∗1 = αR(φ) 7 r 7 . As the level of competition increases, the marginal return to politicians from extra effort put into the project by the bureaucrat increases. This is because in competitive districts, the value of increase from a small vote share is much higher than in stronghold constituencies. Thus, re-election concerns suggest Rφ > 0. Differentiating the optimal effort level with respect to φ gives us: −5 4 de∗1 2 = αR 7 r 7 Rφ > 0 dφ 7 13 (14) This suggests as the level of competition increases, re-election concerns put pressure on the politician to induce a higher effort level form the bureaucrat. Now, starting from the optimal level of effort, what happens if the probability of winning goes to zero? Since the probability of winning being zero means lack of re-election concerns similar to that of low φ, 14 tells us that the optimal effort level should fall. Case II: Ability to use dynamic incentives, no re-election concerns This case corresponds to R(φ) = R. Now, whether the politician is able to use the promise of future payoffs to induce effort depends on β(φ). With high electoral competition, β(φ) ≈ 0 such that these politicians can only provide static contracts in each period. This reduces the 2 4 first period effort level to ecomptt = αR 7 r 7 . Dividing this expression with the expression for 1 the effort level that can be implemented by the recursive contract by politicians in strongholds, we get: ecomptt 1 estronghold 1 4 = r7 (r + βV ) 4 7 ⇒ <1 < estronghold ecomptt 1 1 This shows that as level of competition rises, politician’s ability to use the long-term contracts falls since the probability of them getting re-elected is low. This affects the optimal effort level that can be implemented. Now, what happens if the probability of winning falls to zero? In this case, both re-election concerns and dynamic incentives are shut down (β(0) = 0 and R(0)) ⇒ e∗1 → 0. Thus, starting from any effort level, exogenous decrease in winning probability should reduce optimal effort level. We can summarize the theoretical predictions in the following table: Table 1: Predictions for the Potential Mechanisms Mechanism 1 Re-election Concerns 2 Access to Dynamic Contracts δe δφ ∆e as Prob(win) → 0 (+) (-) (-) (-) Now, how do we interpret the two columns of table 1? The first thing to note is that extent of electoral competition is non-monotonic in the probability of winning as shown in the figure below. When the probability of winning is close to zero, the politician does not face any re-election concerns. Moreover, he does not have access to dynamic contracts. On 14 the other hand, when the probability of winning is close to one, the politicians can credibly make future promises to bureaucrats. However, the re-election concerns are weak at this end of the spectrum as well. Now, when the probability of winning is close to half, the pressure of electoral competition is the highest.12 The predictions in table 1 look at what the mechanisms have to say about change in optimal level of effort when we move away from probability of winning at half. Predictions in column (1) tell us how the level of effort changes when the probability of winning increases towards one. In column (2) however, we are looking at what the mechanisms have to say when we move away from half towards probability of winning being zero. Carrying out both the exercises can help us identify which of the mechanisms is dominant. 4 Data 4.1 Data Sources The main source of data on local public goods comes from the MPLAD database. We have data for each project that was recommended and sanctioned across all 392 constituencies in 12 major states in India over the period 1999-2009. These cover 72% of all constituencies in India and two Lok Sabha election terms: 1999-2004 and 2004-2009. 13 The details of the projects available to us include the type of the project (whether its a road project, drinking water project etc.); the cost incurred; the location; the day on which the project was recommended and the day on which it was sanctioned by the collector. For a subset of projects, 12 Strictly speaking, the probability of winning is half if there is a contest between two candidates. In India, there is a multi party system. However, in many cases there are two strong players and several smaller ones. Even if there are three strong players, for the incumbent, what matters is how far he is from the next best candidate. 13 These 12 major states are: Uttar Pradesh, Madhya Pradesh, Andhra Pradesh, West Bengal, Maharashtra, Gujarat, Kerala, Rajasthan, Tamil Nadu, Haryana, Chhattisgarh and Assam. 15 we also know the time lag between the day the project was sanctioned and the day on which the project actually got completed. We also know which implementing agency executed the project. In the MPLAD data, even though we know which village the project was recommended in, we do not know the identity of the administrative district the village belongs to. We only know which constituency it belongs to. Knowing the identity of the districts is important for us to know which bureaucrat interacted with which politician. In order to find out which district the project was implemented in, we match village and block names from the MPLAD database to the village names in Census data. Our first iteration used a string matching algorithm and in cases where we did not find unique matches, we manually matched the village names. Once we matched the village names across MPLAD data and Census data, it provided us with both district and constituency names for each village. In order to test whether the matching algorithm gave us correct matching, we use shape file data for villages from Census of India and shape files for constituency boundaries obtained from the Election Commission of India. Shape files contain the information on the boundaries of districts, constituencies, and villages. The observations appear as points on polygons. Points are the GPS coordinates and the polygons they form are the boundaries of the geographical unit under study. The following steps are followed in mapping projects from MPLAD data to districts: for each village, we find the centroid using the ArcGIS. We then take the centroid and figure out which polygon it belongs to in the district shape files and which polygon it belongs to in the electoral constituency shape files. This provides us with district-constituency pairs for each project. Using geospatial overlaps, we also obtain the extent to which the administrative and electoral boundaries overlap. Once we obtain the district name for each project in the MPLAD database, we then identify which bureaucrat served as the collector in that particular district in that particular month and year. This allows us to see the identity of the bureaucrat who approved each project. The information on bureaucrats is obtained from the Department of Personnel and Training (DoPT) in India. The DoPT keeps the records of all IAS officers in the form of Executive Record sheets. There is one ER sheet per bureaucrat. This provides us with information about their name, their cadre, their educational background as well as all assignments since the day they joined the civil services. We digitized these ERs in the form of a database and used it to figure out which bureaucrat was assigned as collector to which district over the 16 1999-2009 period. In any given district, over a ten year period, five to six bureaucrats to serve as collectors. So, knowing just We need to know exactly which bureaucrat served in any given know which bureaucrat sanctioned which project. This piece of our identification strategy. on average, we can expect the district is not enough. month. Only then can we information is essential for The election data is obtained from the Election Commission of India. We obtain data for all parliamentary elections from 1989-2009. The main variable of interest is the extent of political competition. We look at various measures of competition. The first one is victory margin which uses information on vote shares of the winner and runner up. The second measure is whether a constituency is a party stronghold or not. This measure looks at identity of the party that won previous consecutive elections. This data also provides us with the names of the winning candidates and their party affiliation. We can map the names across different electoral terms to see whether these politicians are rookies or have been serving for a long time. This allows us to control for politician’s experiences. We also use the party information to see if the MPs are from the same party as the party in power at the center. This proxies for political connection of the politicians. 4.2 Measuring Bureaucrat’s Performance The objective of this paper is to examine how bureaucrats’ performance is affected by electoral competition. For this exercise, we need to define a task assigned to bureaucrats and compare how they perform this task. In this paper we propose a new way of measuring bureaucratic performance: time taken to approve development projects. This is equal to the number of days between the date on which the politician recommended the project and the date on which the bureaucrat gave his stamp of approval. The main advantage of this measure is that this task is purely under the control of the bureaucrat: any delays in approval can be directly attributed to him. Alternative measures used in the literature are time taken to complete projects (i.e. time taken for construction to complete) and completion rates (i.e. proportion of projects that got completed)14 . While these measures are a good indicator of overall governance, we cannot use these to identify which bureaucrat is at fault if delays occur. Completing a project involves several individuals and in infrastructure projects, unobserved factors like the weather can be important reasons for delays. Since the execution process is more complex and has many 14 See Roggers (2014) 17 more unobservables compared to the sanctioning time, we prefer to use the later measure. If approvals take a long time, then we know that it is because the bureaucrat is not putting the effort to get things done. Being able to directly observe the actions of the bureaucrat gives us an advantage over other measures. Given that our measure is directly related to the decisions of a single individual, it is comforting to know that it is positively correlated to completion times. For a very small subset of projects in our database, we have information on the date at which the project construction began and the date at which the project construction ended. We use this additional information to construct a measure of delay in completion - this is the time between the date of sanctioning by the bureaucrat and the date at which the project construction was completed. Note that this measure of delay does not include time taken to sanction - it only looks at how long a project took to complete after it was sanctioned. Table 3 explores the relationship between time taken to sanction projects and time taken to complete projects. As we can see, there is a significant positive correlation between commencement and sanctioning time. This suggests that if bureaucrats perform better according to our measure, then then the projects get started faster. In column 2, we see that if bureaucrat gets project sanctioned within 10 days, then the the delay in commencement of projects reduces by more than 3 months. This is a very large magnitude. Moreover, if projects take more than 90 days to sanction, the then the project construction is delayed by about 7 months. Hence, the speed of sanctioning and speed of project construction move in the same direction15 . 5 Empirical Identification Strategy Consider the following model: y = µ0 + µ1 ∗ Competition +  (15) where y denotes the time taken to sanction a project. We are interested in establishing the causal effect of competition on time taken to sanction the project. Identifying the parameter µ1 requires two things: [1] there should be no reverse causality and [2] there should not be any other unobservables that affects our measure of competition. 15 Although time taken to sanction projects is a more direct way of measuring bureaucratic performance, it would be nice to replicate the tables with an additional measure. However, we have very little data on time taken to complete the projects, hence cannot carry out such an analysis. 18 In the literature, lagged victory margins are often used as measures of competition. This measure assumes that past victory margins are a good predictor of the future victory margins. Moreover, any deviations between the past and future victory margins are assumed to be random and independent to the outcome variable of interest. Now, in our setup, this assumption is violated if we believe that the bureaucrat’s performance between the two elections can affect the future victory margin. Moreover, Ravishankar (2009) shows that there is a strong anti-incumbency effect in Indian elections. Hence, past victory margins may not fully capture the re-election concerns faced by the incumbent. 5.1 Obtaining Exogenous Variation in Probability of Winning In order to obtain the causal effect of electoral competition on bureaucrats’ performance, we need to have a truly exogenous source of variation in political competition. To get such a variation, we take advantage of an information shock that occurred in the middle of the 2004-2009 legislative term. This shock exogenously changed some incumbents’ perceived probability of winning in the next elections to zero while not affecting other incumbents. Henceforth, we refer to the constituencies that are affected by the shock as the treated group while those that are not affected by this shock are the control group. We can then compare the sanctioning times of the treated and control groups before and after this shock. Since this event essentially causes an exogenous change in the electoral competition, the differences-in-difference strategy will give us a causal effect of change in political competition on the performance of the bureaucrat. The information shock we take advantage of is an outcome of the redistricting process that took place in India from 2002-2007. The Government of India froze the changing of electoral boundaries in 1976 and the freeze was supposed to be lifted after the 2001 census was carried out. A delimitation commission was set up in 2002 with the objective of redrawing the electoral boundaries such that the population shares across constituencies are equalized. The total number of constituencies were to remain unchanged. In addition to the equalization of population, the commission re-demarcated the SC/ST constituencies. The SC/ST constituencies are ‘reserved’ in the sense that only candidates who belong to the Scheduled Caste (SC) or Scheduled Tribe (ST) are allowed to contest. Those constituencies that are not reserved, allows any citizen to contest as a candidate. Article 330 of the Constitution of India says that the number of constituencies (or ‘seats’) reserved in any state should be proportional to the number of total seats within a state. Which seat is reserved depends on the relative population of SC/ST across constituencies. So, when the constituency borders are being redrawn, the politicians cannot predict whether their own 19 constituency will be reserved or not because the reservation status depends on how boundaries of other constituencies were re-drawn. This fact is important for the identification strategy. This is because politicians may attempt to influence the re-districting process in order to benefit from it. If the influence in re-districting is a function of political competition, then the concern is that this event may not be truly exogenous. But, since there is no threshold rule but reservation is based on relative population shares, the announcement of the reservation status acts as an exogenous information shock to the incumbents, even if the redistricting may have been endogenous. 16 The report of the delimitation commission of 2002 came in December 2007. This information shock is in the middle of the term 2004-2009. Figure 4 depicts the timing of the announcement. It shows that conditions for using difference-in-difference (DID) strategy are ripe. We now need to do the following: [1] clearly define the treatment group and [2] show how this exogenous information shock changed the perceived probability of winning for the treated group. A constituency can be of three types: [1] reserved for only candidates who belong to SC category, [2] reserved for ST category candidates and [3] GEN (General) category where any citizen can contest (including SC/ST candidates). Under 1972 delimitation, certain constituencies were reserved for SC or ST categories. The incumbents who took office in May 2004 were elected according to the reservation scheme as under 1972 directives. In 2007, the incumbents got to know the reservation status of their constituency for the 2009 election. All those constituencies where there was no change in the reservation status - are in the control group. The electoral districts where there were changes in the reservation status can be classified into two types of treatment groups: [1] REStoGEN - where the constituency was reserved under the 1972 delimitation but all types of candidates were allowed to contest under the 2002 delimitation. [2] GENtoRES - in this case the constituencies were open to all before and in 2007 it was announced that they will be reserved only for SC/ST candidates from 2009 elections. For incumbents who don’t belong to the SC/ST category in the GENtoRES constituencies (which is 87% of all candidates), the change in the reservation status means that they will not be allowed to contest again. Hence, their probability of winning in 2009 is deterministically zero. For the incumbents in REStoGEN, the probability of winning is not conceptually zero because they are, in principle, allowed to contest again. However, as we will show now, the 16 A paper by Iyer and Reddy (2013) provides further support for our identification strategy. They study the redistricting process in India and find that "the redistricting process does not appear to have been influenced by incumbent politicians to a great extent." 20 perceived probability of winning for these incumbents would have been very close to zero. In table 7, consider the group "Gen to Gen." There are 342 constituencies where all candidates could contest in 2004 as well as in 2009 elections. This group is a subset of the control group. Column (2) shows that out of all contestants in these constituencies, only 16% were SC/ST candidates. This means out of 4,477 candidates in these constituencies, only 716 were from reserved category. All these constituencies had at least one candidate who was from reserved category. Column (4) shows that out of these 342 constituencies, only 6 constituency had a winner who came from the reserved category. So, less than 2% of the constituencies open to all had a SC/ST leader. What do these statistics tell us? Consider an incumbent in the REStoGEN treatment group. In the middle of the term, he gets to know that the constituency he represents will be open to competition from other general category candidates. These are career politicians that have a fair amount of idea about what the chances of an SC/ST candidates are in winning an election in GEN category constituency. The above calculations tell us that in 2004, the only 2% of the reserved candidates are successful in winning a seat - this is a fairly low number. Hence, the increase in competition from general candidates is likely to take the perceived probability of winning for incumbents in REStoGEN category is close to zero. Moreover, there is another source of competition faced by these incumbents: intra-party competition. An incumbent of party A may be getting the party ticket to contest in the reserved category constituencies because he may be the best SC/ST candidate. But when the constituency gets de-reserved, this incumbent is up against the general category candidates to get the party ticket to contest in the first place. There are two facts that confirm this: [1] None of the incumbents in the REStoGEN treatment group got a ticket to contest again in 2009 and [2] ALL 2009 candidates in REStoGEN group belonged to the GEN category - none of the candidates were from SC/ST category! Hence once the information shock came in 2007, it is reasonable to assume that due to increase in both intra-party competition and competition from other contestants, the SC/ST incumbents’ perceived probability of winning in 2009 jumped very close to zero within the REStoGEN. 21 5.2 Isolating the Mechanisms While the natural experiment described above helps us determine the causal effect of competition on bureaucratic performance, it does not tell us which mechanism drives the results. This is because when the probability of winning goes to zero, both re-election concerns as well as access to dynamic contracts shut down. In order to see what happens when probability of winning goes to one, we use an alternative measure of extent of electoral pressures: whether a constituency is a party stronghold or not. We define stronghold as a dummy variable that takes value one if the same party won over all four elections between 1989-1998. Table 5 shows that the probability of a stronghold party winning again in 1999 is 65% and winning in 2004 is about 70%. These probabilities are significantly higher compared to non-stronghold constituencies. Note that by construction, our measure of stronghold compares those constituencies where parties have survived the anti-incumbency effect in the past with ones that did not survive it. Hence, the incumbents of stronghold constituencies face less electoral pressures compared to the non-strongholds. Now, how long bureaucrats take to sanction projects depends on their ability. If higher ability bureaucrats are assigned to strongholds, then the error term is correlated to the main regressor of interest. This gives rise to the following selection problem: if we observe a negative relationship between our competition measure and time taken to sanction, it could very well be driven by bureaucrat’s ability. To solve this problem, we exploit two features of our data: [a] panel structure and [b] boundary overlaps. The latter refers to the fact that the administrative and electoral boundaries do not perfectly overlap in India. A single district may have two or three Lok Sabha constituencies that overlap with it. Since the collector is the head bureaucrat in a district and the politician is the elected representative of the electoral constituencies, we have situations where one bureaucrat may deal with two or even three politicians. Figure 3 illustrates such a possibility of overlaps. In this example, the district overlaps with three constituencies: 1, 2, and 3. Each of these politicians face different levels of political competition in their constituencies. However, all three politicians have to depend on this bureaucrat for execution of projects within the district boundary depicted in black color. We can compare the average time taken by this bureaucrat for each of the three politicians. Since we keep the bureaucrat fixed, this controls for the unobserved ability. If we observe that the bureaucrat sanctions projects faster for the politician in the stronghold, then we can attribute the difference in performance to the difference in level of competition. Our discussion so far provides us with two models that we can use to isolate the mechanism 22 that drives the relationship between political competition on bureaucratic performance: yibpc = β0 + β1 ∗ P arty_Strongholdc + ibpc (16) yibpc = γ0 + γ1 ∗ (P rob(win) = 0)c ∗ P ost + γ2 ∗ (P rob(win) = 0)c + γ3 ∗ P ost + ibpc (17) where, yibpc denotes the time taken to sanction project i by bureaucrat b when paired with politician p in constituency c. The variable stronghold means that the same party had won all four elections prior to 1999 elections. Now, how do we interpret β1 and γ1 ? When the party of an incumbent has a stronghold over the constituency then it means that the probability he will win again is very high (close to 1). Hence, we can think of β1 as the effect on approval speed as probability of winning goes to 1. On the other hand, our exogenous shock works as a term limit on the incumbents. Therefore, γ1 looks at the effect on bureaucrat’s performance as the probability of winning goes to zero. As discussed under the theoretical framework, different mechanisms predict different signs on β1 and γ1 . The following table provides the predictions of the alternative channels: Table 2: Predictions for the Potential Mechanisms Mechanism 1 Re-election Concerns 2 Access to Dynamic Contracts β γ (+) (-) (+) (+) Note that our measure of bureaucratic performance is time taken to sanction a project which can be thought of as negative of e in the model. Hence, in column 2 of table 2, the signs are flipped compared to table 1, column 2. Moreover, since stronghold is negatively correlated to φ (the extent of competition), the signs in column 1 are the same in both tables. 23