WPS8352 Policy Research Working Paper 8352 Decentralization and Redistribution Irrigation Reform in Pakistan’s Indus Basin Hanan G. Jacoby Ghazala Mansuri Freeha Fatima Development Research Group Poverty and Inequality Team February 2018 Policy Research Working Paper 8352 Abstract Does decentralizing the allocation of public resources bureaucratically managed, leading to substantial wealth reduce rent-seeking and improve equity? This paper stud- redistribution. The increase in water theft was greater ies a governance reform in Pakistan’s vast Indus Basin along channels with larger landowners situated upstream. irrigation system. Using canal discharge measurements These findings are consistent with a model in which decen- across all of Punjab province, the analysis finds that water tralization accentuates the political power of local elites theft increased on channels taken over by local farmer by shifting the arena in which water rights are contested. organizations compared with channels that remained This paper is a product of the Poverty and Inequality Team, Development Research Group. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The authors may be contacted at hjacoby@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Decentralization and Redistribution: Irrigation Reform in Pakistan’s Indus Basin∗ Hanan G. Jacoby† Ghazala Mansuri‡ Freeha Fatima§ Keywords: Rent-seeking, bureaucracy, elite capture, landownership inequality JEL codes: D73, P48, Q15, Q25 ∗ Funding for this project was provided by the World Bank’s Knowledge for Change Program, its Strategic Research Program, and its Research Support Budget. The project would not have been possible without the support of the Punjab Irrigation Department. The authors are particularly grateful to Mr. Habibullah Bodla at the PMIU for providing access to the discharge data and to Mr. Afzal Toor at PIDA for providing access to FO data. The team also thanks Mr. Saif Anjum (ex-Secretary Irrigation) for his consistent support. The views expressed herein are those of the authors, and do not necessarily reflect the opinions of the World Bank, its executive directors, or the countries they represent. † World Bank, 1818 H St NW, Washington DC 20433; e-mail: hjacoby@worldbank.org. ‡ World Bank; e-mail: gmansuri@worldbank.org. § World Bank; e-mail: ffatima@worldbank.org. 1 Introduction Perceived corruption and lack of accountability associated with top-down public service delivery has led to calls for greater decentralization in developing countries. Participatory or grass-roots governance, in which resource control resides with local elected bodies rather than with centralized bureaucracies, has gained currency among international agencies and donors (see, e.g., World Bank 2004), even though communal authority is by no means immune from rent-seeking in its various forms. A key empirical question is, therefore, whether the promise of local governance can be realized in practice and, if so, under what conditions. Yet, empirical investigation is hampered by lack of large-scale controlled experiments in decentralization combined with a paucity of objective data on behavior associated with rent- seeking.1 This paper takes advantage of a partial governance reform in the world’s largest canal irrigation system, that of Pakistan’s Indus Basin watershed. During the last decade, in an effort encouraged by the World Bank, the management of several large sub-systems in the Punjab was transferred from the provincial irrigation department to farmer organizations (FOs) organized at the channel level. We assess how this shift from bureaucratic to local control affected rent-seeking in the form of water theft along a channel. Effective management of large irrigation systems has proven elusive in both historical and contemporary experience (Meinzen-Dick, 2007).2 In the continuous gravity flow and rotation systems most common in Asia, volumetric pricing and widespread water trading– i.e., market-based allocation–faces daunting technical hurdles (Sampath 1992).3 Instead, 1 See Mansuri and Rao (2013) for a comprehensive and critical review of the evidence. A small literature looks at the impact of decentralization on corruption in cross-sections of countries (most recently, Fan et al. 2009). The challenges with cross-country analyses include heterogeneity in the nature of decentralization and, of course, reverse causation (See Bardhan and Mookherjee 2006b). 2 The celebrated writings of Elinor Ostrom take a slightly more optimistic view, arguing that self-governing institutions sometimes arise organically to solve collective action problems, at least in smaller-scale irrigation systems (Ostrom and Garner 1993). 3 Surface irrigation systems are distinct from other public utilities, such as piped water networks, in that property rights are vastly cheaper to enforce in the latter case; canal water thus has an important common property dimension (see Jacoby and Mansuri 2018). 1 irrigation bureaucracies have been established to operate and monitor centralized systems for the allocation of water as it makes its way down from the rivers and main canals to the network of distributaries, minor and sub-minor canals, and, finally, to the watercourse outlets, where it is delivered to individual farms. In such quota-based systems, users have a strong temptation to bribe local officials to “look the other way” as they use various means to illicitly enhance their water entitlement. Invariably, such water theft benefits farmers at the head of the channel, where water is first to arrive, at the expense of farmers at the tail (see Bromely et al. 1980, Wade 1982, and Chambers 1988). As even a cursory internet search reveals, canal water theft garners enormous media attention in Pakistan, where it is often portrayed as pitting large landlords at the head against multitudes of poor tail-enders. While decentralization strips authority from unelected irrigation department bureaucrats, farmer organizations may also be subject to capture by these upstream elites and, hence, may sanction as much (if not more) water theft than the irrigation department functionaries they replace. In this sense, irrigation reform and decentralization, more broadly, could merely change the venue for rent-seeking without ameliorating its underlying causes.4 To think about the implications of decentralizing irrigation management, we set out a simple model of water allocation, corruption, and rent-seeking along a canal system. Given the locational asymmetry, corruption and theft are concentrated at the head of a channel. However, theft induces rent-seeking by coalitions of gainers (farmers at the head) and losers (farmers at the tail), each with varying degrees of political influence. Under irrigation department control, lobbying effort is directed “over the head” of the local official involved in the corruption whereas, under decentralized control, it is directed toward the FO. The model has several 4 Rijsberman (2008) elaborates on the view that water-users associations are not a panacea. Punjab’s irrigation reform exemplifies what Meinzen-Dick (2007) terms “externally initiated programs...[with] top- down imposition of a rigid structure of user groups and uniform rules that would allow state agencies to recognize and interact with [them].” Such irrigation management transfers, according to her review of the evidence, have had mixed success. Vermillion (1997), considering much the same body of country case- studies, concludes that “the literature on irrigation management transfer does not yet allow analysts to draw strong conclusions about...impacts, either positive or negative.” p29. 2 empirical implications for the impact of decentralization and how this impact interacts with asymmetry in political influence. The centerpiece of our analysis of Pakistan’s irrigation reform is an administrative database maintained by the Punjab Irrigation Department and consisting of readings taken from wa- ter discharge gauges installed at the head and tail of each channel of the entire system. These data arguably provide an objective measure of water theft along a channel. More- over, water discharge data are available over the years 2006-2014, a period encompassing significant devolution of irrigation management to FOs. Importantly, we are also able to match villages along each irrigation channel back to unit record landownership data from recent Agricultural Censuses. This allows us to construct measures of differences in political power, proxied by landholdings, between head and tail villages and thus establish whether irrigation reform has had heterogeneous impacts along this dimension. In a companion paper (Jacoby and Mansuri 2018), we study the allocation of canal water in the presence of rent-seeking farmers and corruptible irrigation officials with career concerns. Using data from several hundred distributaries in Punjab that were not subject to irrigation reform, we find that, under bureaucratic control, the extent of water theft is substantially affected by the distribution of political power along a channel: where political influence is relatively concentrated at the head of a channel, water allocations are more favorable toward the head as reflected in both the canal discharge and land value differential between head and tail. In this paper, by contrast, we focus on how inequality interacts with decentralization. Although the literature recognizes that local governance is more likely to serve the interests of elites where economic and political power is more asymmetric, empirical support for this proposition remains thin (see Mansuri and Rao, 2013). We adopt two strategies for constructing a control group against which we compare the changes in canal water allocations following decentralization. Since FOs were phased-in starting in 2005, our first strategy is to look at variation in outcomes across distributaries 3 with early and late FO formation while controlling for channel-level fixed effects. In this ‘pipeline’ approach, FOs that became operational later (or not at all) serve as controls for those that became operational early. Our second, and ultimately preferable, strategy uses geographically matched controls drawn from adjacent administrative zones that provincial authorities had not (yet) directed to establish FOs. That is, control channels are chosen on the basis of being within a geographical buffer of given distance around a particular FO channel. In this case, and in contrast to the pipeline strategy, we compare changes in discharges over the same time period between neighboring FO and non-FO channels. To summarize our empirical results, we find strong evidence of an economically important decrease in the relative allocation of water to the tail of a channel once an FO becomes operational. Moreover, this decline is greater along FO channels where large landowners are a vis the tail. This latter finding implies that, more heavily concentrated at the head vis ` where power asymmetry is in the same direction as the inherent locational (head versus tail) asymmetry, decentralization leads to greater inequity in allocations. By contrast, where the power asymmetry and the inherent locational asymmetry work in opposition, the negative impact of decentralization is muted. This study contributes to a growing micro-empirical literature on the impact of decen- tralization. Alatas et al. (2013) and Beath et al. (2017) look for direct evidence of elite capture based on field experiments (in Indonesia and Afghanistan, respectively) in which the authority or accountability of extant local governments is randomly varied (see also Basurto et al., 2015, for a nonexperimental study along similar lines). None of these studies, however, compares local-level to top-down control, the pre-reform scenario considered in this paper. Moreover, as noted by Mookherjee (2015), empirical work to date focuses almost exclusively on intracommunity allocations, although the “effects of decentralization on intercommunity allocations are no less important,” precisely because much rent-seeking activity is under- taken by groups of actors in pursuit of their collective interests. Bardhan and Mookherjee 4 (2006c), using longitudinal data from West Bengal, find that district and/or state allocated development grants to villages are strongly negatively related to the percentage of low-caste poor households in the village. However, neither this study nor any other econometric anal- ysis of which we are aware addresses whether devolution of resource control can exacerbate conflict and inequality between communities. More broadly, this paper is related to, and the results consistent with, Acemoglu and Robinson (2008) who study the interaction between de jure and de facto political power: “A change in political institutions that modifies the distribution of de jure power,” they argue, “need not lead to a change in equilibrium economic institutions if it is associated with an offsetting change in the distribution of de facto political power.” Our question, in particular, is whether an institutional change (decentralization) reduces economic inequality or is instead thwarted by increased investments in political capture. We also follow Baland and Robinson (2008) and Anderson et al. (2015), among others, in associating land ownership in developing countries with political power or influence, although our mechanism (rent-seeking) is distinct from theirs (clientilism). The next section of the paper presents the institutional backdrop and data for our anal- ysis. Section 3 develops the model of corruption and rent-seeking along a canal system. Section 4 lays out the empirical methodologies and presents the main impacts of irrigation management reform in Punjab. Section 5 turns to the empirical analysis of power asymmetry along a channel and how it interacts with decentralization. Section 6 concludes the paper. 2 Context and Data 2.1 Indus Basin irrigation system The Indus Basin irrigation system, which accounts for 80% of Pakistan’s agricultural pro- duction, lies mostly in its most populous province, Punjab, wherein it encompasses 37,000 5 Figure 1: Indus Basin irrigation system in Punjab 6 Outlet discharge Authorized Branch canal Actual De-silted Distributary 0 Distance from head Minor canal Head Head outlets Tail outlets Watercourse Tail = PMIU discharge gauge Figure 2: System schematic with discharge gauges kilometers of canals and irrigates about 8.5 million hectacres (Figure 1). From the Indus, Jhelum, Chenab, Ravi, and Sutlej rivers issues a dense network of main canals, branch canals, distributaries, minors, and sub-minors, ultimately feeding 58,000 individual watercourses in Punjab alone (See Figure 2 for a schematic of the canal hierarchy.). Each watercourse outlet or mogha supplies irrigation to typically several dozen farmers according to a rotational system known as warabandi. Tracing its origins to British colonial rule and to the early development of irrigation in the Indus basin, the institution of warabandi (literally “fixed turns”) embodies a modified principle of equity: to each irrigator in propor- tion to his cultivated area. As discussed below, adherence to warabandi leads to an efficient allocation of canal water. At each level of the canal hierarchy in this continuous gravity-flow irrigation system, “authorized discharge” is allocated in proportion to cultivable command area (CCA). At the main canal level, irrigation department staff operate a series of gates regulating flow into the off-taking distributaries according to a rotational schedule. However, 7 since moghas are ungated, discharge into tertiary units, the watercourses, is determined by the width of the outlet; the greater the watercourse CCA, the greater the authorized outlet width (for a given canal discharge), and thus the greater the water in-take each week. Over the course of a week, proceeding from the head to the tail of the watercourse (see Figure 2), each farmer takes his pre-assigned turn at using the entire flow to irrigate his field, with the length of turn proportional to the size of the field. Although design discharge at any point along a channel accounts for seepage and con- veyance losses and is therefore a declining function of distance to the head (see Figure 2 inset), tail outlets should, in theory at least, receive their full water entitlement. In practice, however, discharge at the distributary head is often too low (Bandaragoda and Rehman 1995), or the canal is too silted up, for water to reach the tail outlets. Over-silting also re- sults in higher water levels at the channel head and, consequently, greater discharge at head outlets (Van Waijjen et al. 1997). Lack of canal maintenance, therefore, tends to favor head outlets, which may give rise to lobbying of the irrigation department by farmers at the tail outlets to increase maintenance and by those at the head outlets to suppress it. Although such manipulation is difficult to confirm,5 direct forms of water theft – i.e., tampering with outlets to increase width, siphoning off canal water with pipes, breaching of the canal banks, all supposedly undertaken with the connivance of irrigation officials – are pervasive in the Indus Basin (see, e.g., Rinaudo 2002; Rinaudo et al. 2000). 2.2 Irrigation management reform Formally launched with the passage of the Punjab Irrigation and Drainage Authority Act of 1997 by the provincial assembly, irrigation management reform in the Indus Basin, and specifically the devolution to water user’s associations, was strongly encouraged by the World 5 Yet, one apparently widespread practice having the same effect is placing large boulders or other ob- structions in the bed of a minor canal to increase flow at the head. 8 Bank.6 Administratively, Punjab’s irrigation system is divided into 17 circles. As part of the reform, Area Water Boards (AWBs) were established at the circle level with the responsibility of promoting the formation of FOs covering every water channel within the circle, with the FOs themselves tasked with the operations and management of distributaries and their off- taking channels. In particular, an FO is responsible for monitoring the rotational system to ensure equitable allocation along the distributary, for mediating and reporting water-related disputes among its irrigators, and for collecting water taxes to fund canal operations and maintenance. Five AWBs in what we will refer to as “FO circles” were initially directed to form FOs (see Figure 1). Subsequent roll-out to the remaining 12 circles has been indefinitely delayed due to concerns about FO performance. The formation of an FO involves the following steps: First, an outlet level chairman is elected by all landowners in each watercourse. Second, a secret ballot election is held at the level of the distributary (including off-taking channels), through which a nine-member management committee (president, vice-president, secretary, treasurer, and five executive members) is selected from among the outlet level chairmen. The management committee exercises all powers of the FO. Once elected, an FO does not start operations until its members are trained and it is registered with the AWB.7 Once operationalized, the FO membership remains in office for a tenure of three years, after which new elections are due. In practice, this electoral system has not functioned smoothly. Several incumbent FOs initiated legal action to remain in power and their tenures have been extended beyond the statutory 3-year term under court stays. Starting from the universe of 2,902 irrigation channels in the Punjab, dropping cases that either had zero discharge at the head throughout the 2006-14 period or in which the 6 See World Bank (1994). The Bank’s support was premised on the government instituting a package of reforms, only some of which were ultimately carried out. 7 In theory, FO members were to acquire formal training related to the daily operations and management of the system and be provided with ongoing institutional support. However, despite detailed rules and regulations to this effect, training and capacity building efforts stalled after the pilot phase in LCC East. 9 overseeing FO included a larger branch canal (3 FOs in all), leaves 2,860 channels. Of these, 1,007 are in FO circles, covered by 394 FOs, and 1,853 are in non-FO circles. The excess of channels over FOs in the former case reflects the fact that most distributaries have off- taking minors (and sub-minors) for which we also have discharge data. A distributary-level FO manages all of these minor canals as well. Appendix Table B.1 presents descriptive statistics for all channels by FO status of the circle. FO and non-FO circles look quite similar across design features, which include the number and location of outlets as well as position along parent channel (e.g., a minor canal’s “parent” is a distributary canal). Table 1 gives a timeline of FO operationalization in each of the circles where they have been formed. Between 2006 and 2014, FOs in LCC East and LCC West had completed one full tenure and started their second tenures, while FOs in Bahawalnagar, LBDC, and Derajat were in their first tenure. We do not have pre-reform data for LCC East because FOs there began their first tenure just prior to 2006. Also, because of delays in the election process, there was an interregnum between the two FO tenures in both LCC East and West. During this period, control of the channels reverted back to the irrigation department under a caretaker administration. Finally, note that legal action (court stays) extended the first tenure of 27 FOs (82 channels) in Bahawalnagar and extended the second tenure of 41 FOs (117 channels) in LCC East. An empirical concern addressed below is that these extensions may have occurred precisely in FOs where rent-seeking was intensifying. 2.3 Canal water discharge data Punjab Irrigation Department’s Program Monitoring and Implementation Unit (PMIU) has maintained daily records of authorized (designed) and actual discharge at the head and tail of each channel since 2006. Figure 2 illustrates the typical location of PMIU discharge gauges. Since tail discharge is measured at the last watercourse outlet of the channel, design discharge at the tail is never zero; all sanctioned outlets are entitled to off-take canal water. 10 Table 1: Timeline of FO Operationalization % channels in circle-year Circle No. channels (FOs) Authority 2006 2007 2008 2009 2010 2011 2012 2013 2014 LCC East 229 (84) PID (pre) 4 4 4 4 4 4 0 0 0 FO (1st) 96 96 96 0 0 0 4 4 0 PID (post) 0 0 0 96 96 24 12 12 32 FO (2nd) 0 0 0 0 0 73 84 85 68 LCC West 195 (73) PID (pre) 100 100 3 3 2 2 2 2 0 FO (1st) 0 0 97 97 98 1 1 0 2 PID (post) 0 0 0 0 0 98 98 98 24 FO (2nd) 0 0 0 0 0 0 0 0 74 Bahawalnagar 140 (67) PID (pre) 100 100 100 100 100 0 0 0 0 FO (1st) 0 0 0 0 0 100 100 100 59 PID (post) 0 0 0 0 0 0 0 0 41 FO (2nd) 0 0 0 0 0 0 0 0 0 LBDC 221 (52) PID (pre) 100 100 100 100 100 100 2 2 0 FO (1st) 0 0 0 0 0 0 98 98 100 PID (post) 0 0 0 0 0 0 0 0 0 FO (2nd) 0 0 0 0 0 0 0 0 0 Derajat 222 (120) PID (pre) 100 100 100 100 100 100 100 0 0 FO (1st) 0 0 0 0 0 0 0 100 100 PID (post) 0 0 0 0 0 0 0 0 0 FO (2nd) 0 0 0 0 0 0 0 0 0 Notes: Under the column heading “Authority” are PID (pre) = Punjab Irrigation Department pre-reform; FO (1st) = First tenure of Farmer Organization; PID (post) = Punjab Irrigation Department post-1st FO tenure; FO (2nd) = Second tenure of Farmer Organization. Thus, for example, in 2011, 167 (73% of 229) channels in LCC East were in their second FO tenure. We construct a version of the “delivery performance ratio” or DPR (see, e.g., Waijjen et al. 1997) for the economically most important kharif (summer) season, which runs from mid-April to mid-October. During rabi season, from November to March, 42% of channels in Punjab are dry. Letting d index days and t index year, define j d∈t Qjid DP Rit = (1) ¯ Qj d∈t id for j = H (ead), T (ail), where Qj ¯j id is daily discharge at position j of channel i and Qid is the corresponding authorized daily discharge. Figure 3 shows how head and tail DPRs vary across years for all channels in the 5 FO circles and 12 non-FO circles.8 Two key facts emerge: First, the Indus Basin irrigation 8 Even though a channel is in an FO circle, it may not actually come to be managed by an FO until as late as 2013 (see Table 1). 11 FO Circles Non-FO Circles 1 .8 .6 .4 .2 0 2006 2008 2010 2012 2014 2006 2008 2010 2012 2014 Actual/authorized discharge Head Tail Figure 3: Head and Tail DPRs by year for FO and non-FO channels system consistently under-provides surface water relative to its design parameters; i.e., the ratio of actual to authorized discharge is substantially less than one for the entire 9-year period. Second, the water shortfall is greater at the tail than at the head. To understand the greater water shortfall at the tail, recall that under the quota-based warbandi system each watercourse gets a share of flow into the channel determined by outlet width. Thus, if discharge measured at the head gauge of the channel is, say, 80% of authorized over the whole season, then each outlet along that channel, including the very last one where tail discharge is measured, would automatically receive 80% of its water entitlement or design flow. If, however, upstream outlets are enlarged or the canal is breached or silt is not removed in a timely manner – or, more benignly, the flow entering the channel is highly variable within the filling cycle – then relatively less water makes its way to the tail of the channel over the course of the season; for any given value of DP RH , there is a lower value of DP RT . 12 Defining tail shortage as H T T Sit = DP Rit − DP Rit , (2) we see in Figure 3 that average tail shortage across all channels and years is twice as large in FO circles (0.096) as in non-FO circles (0.048), which suggests that water theft may be more prevalent in FO circles. However, inferring anything about the causal impact of FOs is premature. Indeed, the pattern could reflect selection; i.e., reforms may have been initiated in areas where inequities in water allocation were more pervasive to begin with. 3 Conceptual Framework 3.1 Centralized bureaucracy versus decentralization Modeling public service delivery under alternative political institutions, such as centralized bureaucracy and local governance, using a common theoretical apparatus poses a distinct challenge. In perhaps the only other attempt to do so,9 Bardhan and Mookherjee (2006a) consider a bureaucratic hierarchy engaged in rent extraction under asymmetric information and compare it against a local elected government captured by elites. Decentralization “shifts control rights away from bribe extractors to those who respond to the interests of local users, owing to electoral pressures. However, they respond with a bias in favour of local elites” (p. 110). Bribe-taking is, consequently, replaced by biased fiscal transfers. In Bardhan and Mookherjee, the actions of bureaucrats are unconstrained by motivations associated with public service or career concerns (as discussed in, e.g., Dixit 2002). By contrast, Jacoby and Mansuri (2018) develop a model of bureaucratic canal water allocation in which corruption on the part of local irrigation department officials is constrained by a transfer threat coming from a higher level of the administration. The model is motivated by 9 Hoffmann et al. (2017) examine political allocations under centralized and decentralized structures when home constituencies are favored. However, there is no bureaucratic hierarchy in the model. 13 Rinaudo et al.’s (2000) observations, informed by extensive field-work in Pakistan’s Indus Basin, that “[i]nfluential farmers who are well connected to high-level administration officers or to local politicians. . .are able to put pressure on the local staff of the irrigation bureaucracy in charge of water distribution. . . [C]o-operative local staff. . .benefit from promotions and favourable postings. . .[I]rrigation agency staff regulates the competition between rent-seekers, and maintains the potential costs of tail-enders’ opposition under a threshold guaranteeing the stability of their position.” In this paper, we extend the model to cover the case of FO-managed channels. Under either form of management, water theft and corruption generates winners (farmers at the head of the channel) and losers (farmers at the downstream outlets) who receive less water than they are entitled to. Rent-seeking arises as these winners and losers lobby the powers- that-be to intercede on their behalf. To highlight the role of institutional structure, the only difference between centralized and decentralized systems lies in the incentive for corruption, which, in equilibrium, affects the incentive for rent-seeking. 3.2 Model preliminaries Assume a continuum of outlets along a channel indexed by n ∈ [0, N ], with n = 0 representing the first outlet at the head of the channel and n = N the last outlet at the tail of the channel. Suppose that each outlet has the same command area, normalized to one, and hence the same de jure endowment of water w0 . The de facto inflow of water to each outlet is given by the function w(n), which for the channel as a whole is constrained by N w(n)dn = N w0 . (3) 0 14 Agricultural output depends on water per acre cultivated, but with diminishing marginal product.10 The demand schedule for water D(w) is, therefore, downward sloping (D < 0 for ∀w). Suppose further that D(w0 ) > 0 and that surplus from off-take w is w s(w) = D(w)dw. (4) 0 So, the de jure allocation has a positive marginal value and confers a collective surplus or total value of s0 = s(w0 ) to farmers on the outlet. The efficient allocation of canal water along a channel maximizes N s(w(n))dn (5) 0 subject to (3), which requires that D(w(n)) be equal across outlets. The de jure allocation, with w(n) = w0 ∀ n, is thus efficient and deviations from equal per acre allocations, such as those discussed below, create deadweight losses.11 3.3 Theft and corruption Assume that canal water at each outlet is appropriated until its marginal value is zero subject to availability. Since water arrives first at the head of the channel, outlets at the head have first-mover advantage; some outlets at the tail must, therefore, get no water. Define outlet 10 Output, of course, also depends on purchased inputs such as seed and fertilizer, but to the extent that these are optimally chosen and that their prices do not vary along a channel, the presence of such complementary (to water) investments will not affect our analysis. 11 Chakravorty and Roumasset (1993) point out that equal per-acre allocation along a canal is not neces- sarily efficient once conveyance losses–i.e., water seepage into the channel itself–are taken into account. They show that, in this case, optimal inflow at each outlet should decline with distance to the head. Chakravorty and Roumasset’s simulations, however, indicate that these conveyance loss effects only become quantitatively relevant for outlets at a considerable distance from the head. With a median length of 7 kilometers, the channels that we consider are, in general, too short for conveyance losses to be consequential. Moreover, these simulations overstate the effect of canal seepage in our context by not accounting for the resulting aquifer recharge, which is recovered and used productively by farmers through groundwater pumping. 15 ˆ such that D(w off-take w ˆ by n ˆ ) = 0 and the ‘critical’ outlet n ˆwˆ = N w0 (using equation 3). Thus, all outlets n ∈ [0, n ˆ − w0 in excess of their legal entitlement and receive ˆ ] off-take w surplus s ˆ ), whereas all outlets n ∈ (ˆ ˆ = s(w n, N ] receive no water and get zero surplus. Now, consider the role of an authority, such as the irrigation department or an FO. While ˆ by fine-tuning the degree of outlet tampering the authority could, at some cost, set w < w ˆ − w0 is and other such violations, we assume instead that the amount of water theft w taken as a fait accompli (see Jacoby and Mansuri 2018 for an alternative justification of this assumption). However, once faced with an infraction, the official of the authority charges the farmers at the offending outlet a collective bribe of size b to overlook it (e.g., to not make a police report). How is the amount of this bribe set? A larger bribe, up to the maximum ˆ− s0 , yields higher income to the official, but there is a potential downside. willingness to pay s Before turning to the local official’s trade-off, we must first consider rent-seeking. 3.4 Rent-seeking Water theft creates groups of winners (head outlets) and losers (tail outlets), each of which lobbies the “powers-that-be” for its desired outcome. Define the head outlet coalition CH = ˆ ]} and the tail outlet coalition CT = {n|n ∈ (ˆ {n|n ∈ [0, n n, N ]}, where n ˆ is the last outlet that would receive water under the appropriation scenario described in the last subsection. CH and CT each try to sway the authority to, respectively, continue the water theft or to restore the de jure water allocation. As in Tullock (1980), we assume that the probability, P , of CH winning this contest depends on the effort level, ej , of both coalitions j = H, T as follows:12 ιH eH P = , (6) ιH eH + ιT eT 12 The linearity of each player’s effort in the probability function is a standard simplification in the literature on games of rent-seeking (see Nitzan 1994). 16 where the ιj represent the marginal influence of coalition j . When ιH = ιT , there is a power asymmetry along the channel; this is the sense in which intercommunity inequality matters for outcomes.13 Assuming a unitary marginal cost of effort,14 expected net surplus for CH is πH = P n ns0 − eH s − b) + (1 − P )ˆ ˆ (ˆ (7) ˆ s0 + P ∆H − eH , = n where ∆H = n s − s0 − b), and for CT is ˆ (ˆ ˆ )s0 − eT πT = (1 − P )(N − n (8) ˆ )s0 − P ∆T − eT = (N − n where ∆T = (N − n ˆ )s0 . Although we abstract here from free-riding on rent-seeking effort within each coalition, political influence ιj can be seen, in part, as a measure of the efficacy of collective action (as in Acemoglu and Robinson’s 2008 political contest model). The nature of rent-seeking activities may also differ between head and tail, given head outlets’ locational advantage. For instance, CT may stage protests or, rather, exercise “voice” (see Reinikka and Svensson 2004 for a model along these lines), whereas CH may engage in various more subtle forms of pressure and persuasion. We consider these different cases below. Suppose, now, that each coalition chooses its rent-seeking effort taking that of the other coalition as given. Given an interior solution, eT = ΩeH , where Ω = ∆T /∆H is the ratio of 13 Insofar as some of the rent-seeking effort translates into utility for the authority, there is an incentive for whoever is in charge to hold a lobbying contest with non-trivial win probabilites for each side. 14 This assumption, applied to lobbying effort by both head and tail coalitions, is innocuous. High (low) marginal influence ιj is equivalent to low (high) marginal cost of effort. 17 ˜ , where win-loss differentials. Thus, in the Nash equilibrium, P = P ˜ ( b) = ιH ∆H (b) P . (9) ιH ∆H (b) + ιT ∆T The equilibrium probability of maintaining corruption depends on each coalition’s net gains from winning the lobbying contest weighted by their marginal influence. We may write equation (9) more compactly as ˜ (b) = θ P . (10) θ + Ω(b) where θ = ιH /ιT is a parameter representing the relative influence of the head coalition a-vis the tail coalition. vis-` 3.5 Optimal bribe Bureaucracy: Bureaucracy is characterized by hierarchy; the official on the ground con- doning the water theft in exchange for a bribe is an agent of a higher level office. We assume that the local official is, in effect, paid an efficiency wage and thus has career concerns (see Jacoby and Mansuri 2018 and the citations therein). As long as he stays in his current ˆ b; otherwise, he receives his outside option, which we position he receives bribe income n normalize to zero. Whether the local official is retained depends on the pressure exerted on the irrigation department by the contending interests along the channel. If CH wins the lobbying contest, as described formally in the previous subsection, the official will be re- tained, whereas if CT wins, the official will be reassigned and replaced, at least temporarily, by direct irrigation department oversight. The local official chooses his bribe b for the channel to maximize expected income ˜ (b)ˆ VB (b) = P nb. (11) 18 Thus, the official faces a trade-off between greater bribe income, on the one hand, and a higher equilibrium probability of retaining his position, on the other. In particular, the higher the bribe, the less net surplus is available to head outlets and, hence, the less effort their coalition exerts to retain the official. Farmer Organizations: While FO officials are assumed to behave the same way as the local irrigation department officials, their objective function differs in a key respect. Under bureaucratic control, lobbying is directed upward, to the office with the authority to transfer a lower official. In a decentralized structure, farmers lobby the FO and at least part of this lobbying directly benefits FO officials. Decentralization thus breaks the separation between corruption and rent-seeking that prevails in the bureaucratic hierarchy. Suppose that the FO receives some utility u from rent-seeking effort. We may think of u as the perks of power or the value of political support to remain in power or to be reelected, or all of the above. Let us distinguish the two cases alluded to earlier. In the first case, an equal fraction of the equilibrium rent seeking efforts eH (b) and eT (b) provide utility to the FO; thus, u1 (b) = U (eT (b) + eH (b)). In the second case, the nature of rent-seeking on the part of CH is the same as above; CT , however, only exercises its voice option. Since protests provide no direct utility to the FO, u2 (b) = U (eH (b)). Depending on case c = 1, 2, the FO chooses b to maximize ˜ (b)ˆ VF (b) = P nb + uc (b), (12) VF can be seen to combine the local irrigation official’s objective VB with that of the higher- level department office, which we previously could ignore. Importantly, since uc < 0 — higher bribes, by curtailing valuable rent-seeking effort, make the FO worse off — the FO official has lower marginal corruption incentives than the irrigation official and, hence, charges a lower bribe. This result is formalized as lemma 1 in Appendix A. 19 3.6 Implications of decentralization Our outcome variable, tail shortage, is the expected difference in water available at the first ˜ (br )(w and last outlet of a channel. In terms of the model, T Sr = P ˜ (br ) (w0 − ˆ − 0) + 1 − P ˜ (br )w, w0 ) = P ˆ where r = B, F denote bureaucracy and FO, respectively. The model yields three results (see Appendix A for proofs): Proposition 1. T SF > T SB This result says that water theft increases after decentralization, or ∆T S = T SF − T SB > 0. Intuitively, the bribe amount falls under FO authority because, as noted, the marginal incentives for bribery are reduced. Water theft, however, is decreasing in the bribe amount, because higher bribes reduce head outlets’ surplus and, hence, their support for the status ˜ must fall). quo (so P Next, we have that decentralization increases water theft by more on channels along which head outlets are relatively powerful; i.e., along those with high θ: ∂ ∆T S Proposition 2. > 0. ∂θ Essentially, theft responds more, at the margin, to political influence when bribes are low (i.e., under FOs).15 Lastly, we have a symmetry result following directly from the definition θ = ιH /ιT : ∂ ∆T ∂ ∆T Proposition 3. =− . ∂ log ιH ∂ log ιT A 1 percent increase in head influence has an equivalent effect on the change in tail shortage as a 1 percent decrease in tail influence. In the remainder of the paper, we assess whether the experience of decentralization in the Indus Basin comports with these implications of our model. 15 Jacoby and Mansuri (2018) prove that ∂T S (bB )/∂θ > 0 so that, given Proposition 2, water theft is increasing in θ under both irrigation department and FO authority. 20 4 Main Impact of Irrigation Reform 4.1 Pipeline strategy Our regression model for tail shortage, exploiting the pipeline variation, is 1 2 T Sit = α1 τit + α2 τit + µi + δt + γc t + εit (13) k where the τit are indicators for whether channel i is in the midst of its first (k = 1) or second (k = 2) FO tenure during kharif season of year t. As noted, we control for channel fixed effects, µi , which sweep out permanent channel characteristics, such as those correlated with the likelihood of receiving an FO earlier rather than later. We also include year dummies δt and circle-specific time-trends, as represented by the penultimate term in equation (13). Difference-in-differences (or fixed effects) estimation of treatment effects is predicated on the parallel trends assumption, which is to say that, absent intervention, average outcomes would have evolved similarly for both treatment and control groups. Here, with the exception of LCC East, which has no pre-reform observations (see Table 1), we are able to estimate separate time trends, γc , for each circle.16 Thus, we directly control for differential pre- intervention time trends across the unit of policy choice (recall that Area Water Boards for the formation of FOs were established at the circle level). Nevertheless, our pipeline identification strategy maintains the assumption that intertemporal shocks to relative water availability at channel tails, the εit , are uncorrelated with FO operationalization — i.e., do not cause FOs to begin or end their tenures sooner or later. It is the threat posed by the possible failure of this assumption that motivates our second strategy below. 16 Our analysis of pre-trends in the three late-reforming FOs (Bahawalnagar, LBDC, and Derajat; see Table 1) is summarized in Appendix Figure B.1. While parallel trends between these FOs and all non-FO channels from 2006-2010 can be formally rejected, this is no longer the case when 2006 data are dropped. Below, therefore, we check our estimates for robustness to the removal of the 2006 observations. 21 Table 2: Baseline Treatment Effect Estimates – Pipeline Strategy (1) (2) (3) (4) (5) 1st FO tenure (α1 ) 0.0485*** 0.0474*** 0.0483*** 0.0427*** 0.0478*** (0.00716) (0.00797) (0.00724) (0.00742) (0.00719) 2nd FO tenure (α2 ) 0.00423 0.00438 0.00381 -0.00288 — (0.0101) (0.0101) (0.0100) (0.0111) p−values: H0 : γc = 0 0.000 0.000 0.000 0.000 0.000 H0 : α1 = α2 0.001 0.002 0.001 0.001 — R2 (within) 0.056 0.047 0.057 0.056 0.056 No. of observations 24983 23059 24791 22195 24146 No. of channels 2851 2626 2851 2848 2851 No. of FOs/distibutaries 1225 1101 1225 1223 1225 Notes: Robust standard errors in parentheses (*** p<0.01, ** p<0.05, * p<0.1), clustered on FO/distributary (distributary for non-FO channels). Dependent variable is tail shortage DP RH − DP RT . All specifications include channel fixed effects, year dummies, and circle-specific time trends (γc ). Guide to specifications: (1) All channel-years; (2) Drops LCC East; (3) Drops cases of FO tenure extended by court-stay; (4) Drops observations from 2006; (5) Drops all channel-years in second FO tenure. 4.2 Pipeline results Results for the pipeline approach using all channel-years (Table 2, column 1) indicate that the first FO tenure significantly increased tail shortage. The average treatment effect estimate of 0.0485 is 53% of average pre-reform tail shortage across all FO channels and years. Thus, the irrigation reform worsened discharge at the tail relative to the head by around half the original gap. The circle-specific time trends are strongly significant, net of overall year effects, as indicated in Table 2. However, we are not able to allow for a separate time trend for LCC East, as this FO circle has no pre-treatment observations. To ensure that LCC East is not driving our results as a consequence, we drop all observations from this FO from the estimation in column 2. Comparison to column 1 reveals that lack of pre-trends for LCC East is not a serious lacuna. The pipeline results imply no discernible effect of the second FO tenure. Although the standard error on the second tenure coefficient is somewhat larger than that on the first 22 tenure coefficient, lack of precision is clearly not the whole story since we can still strongly reject the equality restriction α1 = α2 . As noted, however, FOs had their second tenures in only two of the five FO circles (see Table 1); 84% of these observations are from LCC East, which, recall, was intended to be a showcase for the irrigation reform. Additionally, as also noted, a large proportion of second tenures in LCC East were extended by court stays. To check robustness against the concern that FO tenures were endogenously extended by legal action, we drop all such channel-year observations (whether in the first or second FO tenures) in column 3. The results are virtually unchanged. We also drop the first year of data (see column 4) in light of the fact that pre-trends with 2006 included are not strictly parallel (fn. 16). This also makes little difference, nor should it since we are already controlling for circle-specific time trends. Finally, in the last column of Table 1, we present the estimate for the first FO tenure treatment effect when all second tenure observations are dropped. 4.3 Spatial matching strategy Before setting out the regression model for use with spatially matched controls, we discuss our GIS buffer strategy. A buffer is a locus of GIS coordinates equidistant from each coordinate of an FO channel. Spatial matching consists in finding the set of channels from non-FO circles that lie entirely within a buffer of given radius. Figure 4 illustrates a 40 kilometer buffer for a channel in Bahawalnagar Circle along with one particular control channel, of which there are typically many.17 Compared to the pipeline strategy, spatial matching uses the same underlying channels (both FO and non-FO) but weights them differently. The choice of radius for the GIS buffer presents a trade-off. The smaller the radius, the more similar treatment and control channels are likely to be along unobserved dimensions (given spatial correlation in these unobservables). However, a smaller radius also implies a 17 There are no GIS shape files for circle borders, so we cannot match on the basis of distance to these administrative boundaries. 23 Figure 4: Example of 40 km buffer for spatial matching smaller likelihood of finding any channels lying both within the buffer and within an adjacent non-FO circle. A radius of 40 km, in particular, leads to a sample consisting of 302 FOs covering 747 channels, with 915 non-FO channels as controls (but each of these typically appearing in many buffers). Thus, the choice of 40 km radius implies a loss of 94 of our original 396 FOs in the sense that we do not have spatially matched controls for them. By contrast, moving to a 60 km buffer radius matches 348 FOs covering 883 channels (with 1,233 non-FO control channels). But shrinking the buffer radius down to 20 km nets only 130 FOs covering a mere 351 channels; since we believe that this is too few to constitute a useful sample, we do not pursue the 20 km buffer strategy. Indexing buffers by subscript b, our regression model becomes 1 2 T Sit = α1 τit + α2 τit + µi + γc t + φbt + ξit (14) 24 where φbt is a buffer-year fixed effect.18 In terms of the pipeline specification, we may think of εit = φbt + ξit with φbt as the spatially intra-correlated component of the intertemporal shock to relative water supply at channel tails. To understand the source of identifying variation in equation (14), we simplify the model to just two time periods, before and after reform. First-differencing over time for each channel in this case is equivalent to including channel fixed effects and yields 1 ∆T Sit = α1 ∆τit 2 + α2 ∆τit ˜b + ∆ξi , + γc + φ (15) ˜b is a buffer fixed effect. Thus, the average treatment effect of an FO tenure is iden- where φ tified off of within buffer variation in channel-level discharge differences (pre/post) between FO and non-FO channels that lie in adjacent FO and non-FO circles, respectively. In con- trast to the pipeline approach, the spatial matching estimator uses none of the variation in the timing of reform across FO circles since a given buffer can only contain channels from one FO circle. While the general time-pattern of tail shortage is absorbed in the buffer-year fixed ef- fects included in equation (14), circle-specific trends γc are estimable because channels from the same circle can appear in many different buffers. To allow for the possibility that tail shortages in FO circles and in adjacent non-FO circles were not on parallel trajectories prior to decentralization, we thus again control for circle-specific trends. Finally, let us emphasize that, insofar as the decomposition of the tail shortage shock εit into a spatially intra-correlated component φbt and a purely idiosyncratic component ξit is valid, the identi- fying assumptions are weaker in the spatial matching case than in the pipeline strategy; we only require that ξit be uncorrelated with changes in FO operationalization status. 18 Due to the high dimensionality of both the channel and buffer-year fixed effects, we must estimate equation (14) using an iterative technique (Guimaraes and Portugal, 2010). 25 Table 3: Baseline Treatment Effect Estimates – Spatial Matching 40 km buffer 60 km buffer (1) (2) (3) (4) (5) (6) 1st FO tenure (α1 ) 0.0465*** 0.0463*** 0.0500*** 0.0427*** 0.0424*** 0.0441*** (0.0089) (0.0089) (0.0087) (0.0087) (0.0088) (0.0086) 2nd FO tenure (α2 ) 0.0430*** 0.0429*** — 0.0280** 0.0278** — (0.0162) (0.0162) (0.0139) (0.0139) p−values: H0 : γc = 0 0.000 0.000 0.000 0.000 0.000 0.000 H0 : α1 = α2 0.845 0.851 — 0.367 0.371 — R2 (within) 0.598 0.598 0.599 0.567 0.567 0.567 Observations 223,109 222,970 222,508 674,296 674,121 673,517 Number of clusters 751 751 751 916 916 916 Number of FOs 302 302 302 348 348 348 Notes: Robust standard errors in parentheses (*** p<0.01, ** p<0.05, * p<0.1), clustered on FO/distributary (distributary for non-FO channels). Dependent variable is tail shortage DP RH − DP RT . All specifications include channel fixed effects, buffer-year fixed effects, and circle-specific time trends (co- efficients on which are γc ). Guide to specifications: (1,4) All channel-years; (2,5) Drops cases of FO tenure extended by court-stay; (3,6) Drops all channel-years in second FO tenure. 4.4 Spatial matching results The spatial matching strategy yields similar estimates of the first FO tenure treatment effect regardless of whether we adopt a 40 km (Table 3, column 1) or a 60 km (column 4) buffer radius. Relative to the pre-reform scenario in FO circles, the first tenure effects imply a 51% (40 km buffer) and 46% (60 km buffer) increase in tail shortage. The second FO tenure effects here are statistically significant and of similar magnitude to the first tenure effects, so that we cannot reject the equality restriction α1 = α2 . Moreover, none of these results depends on the inclusion of the potentially suspect observations involving court stays (cols. 2 and 6). Finally, Table 3 reports specifications that that drop channel-years in second FO tenures altogether (columns 3 and 6), which has little impact on the first tenure coefficient.19 19 Equality of circle-specific time-trends can be rejected in all specifications. Note, as well, that dropping observations from 2006, thereby rendering parallel the pre-trends in FO and non-FO circles (see fn. 16), does not appreciably affect our results (the estimate of α1 in specification (4) falls to 0.041 (0.009) and that of α2 falls to 0.022 (0.014)). 26 4.5 Discussion Two distinct panel data strategies have yielded broadly consistent findings: Irrigation reform in Punjab increased tail shortage initially (i.e., in the first FO tenure) by around 50% of the pre-reform baseline. Since, as we have argued, the efficient allocation of canal water involves zero tail shortage, decentralization had a social cost. In other words, the takeover by FOs could not both accentuate head-tail inequality in canal water and (through side-payments) lead to a Pareto improvement of welfare. While we cannot compute the social cost directly, outlet-level data on land values from Punjab (see Jacoby and Mansuri 2018) allow us to infer that the wealth redistribution was substantial; in particular, the reform increased the value of head-end land by about 9% relative to the value of tail-end land.20 Evidence on the second FO tenure, which is far less frequent in the data than the first tenure, is not as clear-cut. Using spatial matching, the estimated second tenure treatment effect is significantly different from zero and not significantly different from the first tenure effect, which is entirely plausible — there is no theoretical reason to suggest that these effects should differ. The pipeline strategy, by contrast, yields a more precise yet insignificant second tenure treatment effect. These divergent findings may indicate that the identifying assumption of the pipeline strategy is violated in the data; that water availability shocks are correlated with the timing of (second tenure) FO operationalization. Be that as it may, when we drop observations in second FO tenures, both the pipeline and spatial matching strategies yield virtually identical results. In the analysis to follow, therefore, we will rely on this restricted sample and, because it is more robust, the spatial matching strategy. 20 The data come from 3,922 outlets along 448 non-FO channels in Punjab. Within the same channel, land at a head outlet is valued at a 11.2% premium over land at a tail outlet (Jacoby and Mansuri 2018, Table 1). Moreover, average head DP R on these same channels is 0.052 higher than average tail DP R. Assuming, plausibly, that the entire head-tail land value differential is attributable to variation in canal water availability, a treatment effect of 0.043 is tantamount to a 11.2 × 0.043/0.052 = 9.3% increase in relative land values. 27 5 Role of Political Influence Under what conditions will decentralization produce more equitable allocations? Our the- oretical model formalizes the political process at the canal level as a rent-seeking contest between rival coalitions of irrigators. Asymmetry of political influence (θ) thus affects the outcome of irrigation reform. The empirical challenge is to measure the relative influence of outlets at the head versus those at the tail. In Pakistan, the natural proxy for political power is land ownership. Indeed, large landowners not only have more political clout but also a proportionally greater stake in the contest over water rights, and hence a greater incentive to deploy their clout.21 Despite active tenancy markets, land sales markets are relatively thin in Pakistan, with the bulk of ownership transferred through inheritance. As a consequence, the local distribution of land ownership can be seen as both stable and as largely independent of the distribution of farmer productivity or soil fertility (factors which are, at any rate, purged from our regression specifications using channel fixed effects). 5.1 Land ownership data and mouza matching We use data from four Agricultural Censuses (1980, 1990, 2000, and 2010) to characterize the distribution of landownership along Punjab’s irrigation channels. Since Pakistan carries out a “sample census,” about 13% of villages (mouzas ) are covered in any given round (and 9% of households), yielding roughly 3,500 villages per round with considerable overlap across rounds. Thus, between 1980 and 2010, nearly 7,700 unique villages appear in the Agricultural Census. Given the relative stability of the land ownership distribution over time, we treat the most recent observations on all of these villages equally for the purposes of constructing 21 Jacoby and Mansuri (2018) create an index of lobbying power that combines information on both individ- ual landownership and political/bureaucratic/hereditary office-holding, but the latter data are not available for the present sample of channels. 28 our aggregates.22 Irrigated villages from the census are matched to their corresponding canal outlets using village-outlet lists supplied by the Punjab Irrigation Department. Following irrigation department designation, head villages are defined as those that match to outlets on the upper 40% by length of a given channel; tail villages as those that match to outlets on the lower 20%. We compute land ownership statistics Lij by position j = H, T on channel i, such that Lij = G(ω1ij L1ij , ..., ω1ij LNij ij ), (16) where Nij is the number of census households matched to position j of channel i (we drop channel-positions with Nij < 20), the Lkij are the unit record landownership data, the ωkij are sample census population weights normalized to sum to one within a channel-position, and G : RNij → R1 is a statistic. Since theory is silent on the form of G, we experiment with several, varying the weight given to large landowners. Thus, while we use the (weighted) arithmetic mean (G is simply the summation operator in this case), we also try a version of the generalized mean G(x1 , ..., xM ) = ( xq m) 1/q , which puts greater weight on large values of xm insofar as q > 1, as well as the 75th and 90th percentile operators.23 Note that the Agricultural Census samples all types of households within each mouza, whether cultivating or not and whether they own land or not. Arguably, the population of cultivators and/or landowners is most relevant for the political-economy of irrigation. Since non-cultivating households without land should have little, if any, influence with the FO, or with the irrigation department for that matter, this population might reasonably be excluded 22 To the extent that land ownership data from the 1980s and 1990s are dated, they introduce measurement error biasing against finding (differential) treatment effects. 23 The Agricultural Census does not provide household landownership broken down by irrigated and rain- fed areas, even though it is the former type of land that is most germane to the lobbying effort along a channel. However, the Census does distinguish irrigated and rain-fed cultivated area at the household level. Therefore, we deflate household landownership by the ratio of cultivated area under irrigation (summed across household in a mouza ) to total cultivated area in the mouza. 29 in calculating channel position-level land statistics. On the other hand, including these non- farm households could have an important scaling function. For example, a community of 100 households each owning 100 acres is likely to have more influence than a community consisting of just a single 100-acre farm surrounded by 99 non-farm households; yet, mean landholdings across farm households is identical in these two communities, whereas mean landholdings across all households is indeed higher in the presumptively more powerful one. If the second community, instead, consisted of 100 farms of 100 acres and 100 non-farm households, the mean across all households would imply that the second is less powerful than the first when it is, in fact, equally powerful. In this case, using the means across farm households would (correctly) imply communities with identical lobbying influence (see Appendix Table B.2 for a visual guide to these examples). In short, for our purposes, there is no unambiguously valid choice of population over which to compute land ownership statistics. Prudence, therefore, dictates using both approaches. Because the four rounds of sample census data are not everywhere dense in villages, we are not able to match both head and tail mouzas for every channel. Our analysis of power asymmetry is, thus, based on fewer FOs than were present in the baseline samples. For the 60 km spatial matching sample, the number of FOs covered falls from 349 to 247, when land statistics are taken over only farm households, and to 252, when land statistics are taken over all census households.24 There is also a modest (positive) correlation between land ownership statistics at head and tail of the same channel (see Appendix Figure B.2), which is why these variables must be included together in the regressions. 24 Of the 247 FOs represented in the former case, 15% are in Bahawalnagar Circle, 29% are in Derajat, 17% are in LBDC, 21% are in LCC East, and 17% are in LCC West. The corresponding breakdown across all 396 FOs in Punjab is 17%, 30%, 13%, 21%, and 18%. In line with this similarity in composition, main treatment effects are very close to those in Table 3 when estimated on the smaller samples of channels with land data (results available upon request). 30 5.2 Heterogeneity results With these considerations, we now specify a mapping from the land distribution at position j of channel i, summarized by the statistic Lij , to political influence of the form ιij ∝ ¯ ij (mean exp(Lij ). If, for example, G is chosen to be the summation operator, then Lij = L ¯ iT , the mean difference ¯ iH − L landownership), and we would have log θi = log ιiH − log ιiT = L in landownership between head and tail. The augmented spatial matching specification (dropping second FO tenure effects) is 1 1 1 1 T Sit = α1 τit + δH τit LiH + δT τit LiT + τit Zi λ + µi + γc t + φbt + ξit , (17) where Zi is a vector of channel level characteristics—including its length, number of outlets, and position on parent channel—that might influence FO performance. Under the symmetry restriction δH = −δT (see Proposition 3), this regression is equivalent to interacting the 1 treatment dummy τit with log θi . Symmetry, recall, implies that head outlets obtain just as much additional influence over allocations at the margin from (say) one acre higher mean landownership at the head as from one acre lower mean landownership at the tail. Table 4 reports eight spatial matching specifications, crossing four versions of G with the two census populations. Even as the coefficients of interest in Table 4, the δj , vary in magnitude across specifications due to the different scaling of the Lij , a consistent pattern ˆT < 0 and we fail to reject the null of symmetry. That ˆH > 0 and δ emerges: In each case, δ this test has power is supported by the fact that we can reject (in all but one case) the joint null hypothesis that δH = δT = 0. Finally, in the restricted models (i.e., with δH = −δT ), we can strongly reject the null that relative political influence has no effect on water allocation along a channel. 31 Table 4: Influence Asymmetry and FO Performance — Spatial Matching Strategy Land Statistic: Mean Generalized mean (q = 1.5) 75th percentile 90th percentile (a) census farm hhs: τ 1 × LH (δH ) 0.00552** 0.00225** 0.00487** 0.00195* (0.00225) (0.00098) (0.00217) (0.00115) 1 τ × LT (δT ) -0.00512 -0.00174 -0.00489 -0.00175 (0.00330) (0.00113) (0.00310) (0.00124) τ 1 × (LH − LT ) 0.00534*** 0.00204*** 0.00488** 0.00185** (0.00204) (0.000733) (0.00199) (0.00093) p−values: H0 : λ = 0 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 H0 : δH = −δT 0.916 — 0.732 — 0.995 — 0.899 — H0 : δH = δT = 0 0.024 — 0.023 — 0.038 — 0.125 — R2 0.567 0.567 0.567 0.567 0.567 0.567 0.567 0.567 (b) all census hhs: 32 τ 1 × LH (δH ) 0.00975*** 0.00375** 0.00690** 0.00404*** (0.00343) (0.00158) (0.00279) (0.00140) 1 τ × LT (δT ) -0.0126*** -0.00426** -0.00882*** -0.00568*** (0.00470) (0.00167) (0.00327) (0.00206) τ 1 × (LH − LT ) 0.0109*** 0.00395*** 0.00781*** 0.00467*** (0.00301) (0.00120) (0.00230) (0.00121) p−values: H0 : λ = 0 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 H0 : δH = −δT 0.593 — 0.822 — 0.617 — 0.484 — H0 : δH = δT = 0 0.001 — 0.003 — 0.004 — 0.001 — R2 0.566 0.566 0.566 0.566 0.566 0.566 0.566 0.566 Notes: Robust standard errors in parentheses (*** p<0.01, ** p<0.05, * p<0.1), clustered on FO/distributary. Number of observations (clusters) [FOs]: 586,417 (892) [247] in panel (a); 602,425 (909) [252] in panel (b). All specifications use spatial matching with 60 km buffer. Dependent variable is tail shortage DP RH − DP RT . Channel-years in 2nd FO tenure dropped (τ 1 is indicator for first FO tenure). All specifications include channel fixed effects, circle-specific time trends, and buffer-year fixed effects. Lj denotes land ownership statistic at position j of channel computed over census farm households (panel (a)) or all census households (panel (b)). τ 1 is interacted with Z variables: log channel length, log number of outlets, whether channel is on head or middle of parent channel (tail omitted category), and whether channel is minor or sub-minor (distributary omitted category), the coefficients on which are denoted by λ. Census Farm Households All Census Households .15 .2 Marginal effect of FO tenure Marginal effect of FO tenure .1 .1 .05 0 0 -.1 -.05 -10 -5 0 5 10 -10 -5 0 5 10 Head mean land - Tail mean land Head mean land - Tail mean land .08 .1 .08 .06 .06 Fraction Fraction .04 .04 .02 .02 0 0 -10 -5 0 5 10 -10 -5 0 5 10 Head mean land - Tail mean land Head mean land - Tail mean land Figure 5: Marginal effect of 1st FO tenure by log θ Notes: Left and right upper panels show marginal effects of first FO tenure at different values of relative influence (log θ) based on, respectively, census farm households (panel (a), col. 2, Table 4) and all census households (panel (b), col. 2, Table 4). Lower panels show histograms of log θ for the corresponding specifications. Bars in upper panels denote 95% confidence intervals. Short-dashed vertical lines denote one standard deviation above and below the mean of log θ (long-dashed vertical line). We summarize our key finding in Figure 5, which plots the marginal effect of FO tenure based on the restricted estimates (column 2 of Table 4). In channels with relatively greater average landownership at the head than at the tail, the post-reform allocation of canal water worsened (disfavored the tail) to a significantly greater extent. In other words, decentral- ization not only appears to have aggravated rent-seeking, but to have aggravated it by more in channels along which political power asymmetry reinforces locational asymmetry (Propo- sition 2). Tail-end irrigators on channels along which large landowners most predominate ¯ iT is two standard deviations above ¯ iH − L at the head (specifically, on which log θi = L the mean) saw roughly a doubling of their relative shortage after reform, whereas tail-end 33 Table 5: Influence Asymmetry–Robustness to Channel-level Inequality Census Farm Households All Census Households Gini Top 5% share Landless prop. Gini Top 5% share Landless prop. 1 ¯T ) ¯H − L τ × (L 0.00540*** 0.00520*** 0.00534*** 0.0109*** 0.0108*** 0.0109*** (0.00202) (0.00201) (0.00205) (0.00299) (0.00297) (0.00295) τ 1 × inequality -0.0302 0.0471 -0.00525 0.142* 0.0557 0.123** (0.0893) (0.0569) (0.0659) (0.0861) (0.0524) (0.0490) R2 0.567 0.567 0.567 0.566 0.566 0.566 No. of Obs. 586,417 586,417 586,417 602,425 602,425 602,425 No. of clusters 892 892 892 909 909 909 No. of FOs 247 247 247 252 252 252 Notes: See notes to Table 4. Columns 1-3 should be compared to Table 4, panel (a), column 2; Columns 4-6 should be compared to Table 4, panel (b), column 2. irrigators on channels along which large landowners most predominate at the tail (log θi two standard deviations below the mean) suffered essentially no erosion in water allocation following reform. 5.3 Robustness to channel-level inequality A large literature on collective action in commons management highlights the importance of heterogeneity among users (e.g., Ostrom 1990; Baland and Platteau 1997), although the effect of inequality on outcomes is often theoretically ambiguous (Bardhan and Dayton- Johnson 2002). In the context of surface irrigation systems, Bardhan (2000) and Dayton- Johnson (2000) find that the landholdings Gini coefficient is negatively associated with co- operation in water allocation and channel maintenance. Our concern here is that, if FOs along channels with, say, greater wealth inequality produce less cooperative outcomes, and if overall channel-level land inequality is correlated with head-tail differences in landholdings, then our heterogeneity results in Table 5 may be spurious. To deal with this concern, we construct channel-level measures of land inequality—Gini coefficient, share of land owned by top 5%, and proportion of landless—using the Agricultural 34 Census and both reference populations (see subsection 5.1).25 These measures incorporate households from all census villages that match to outlets on the head, tail, and middle 40% of a given channel. In Table 5, we add these variables one-by-one to the specifications in ¯ T ) is virtually unaffected. ¯H − L column (2) of Table 4. In each case, the coefficient on τ 1 × (L 6 Conclusion How a shift in control from centralized bureaucracy to local government affects resource allocation has been empirical terra incognita up until now. It is worth reiterating why this is so: natural experiments in decentralization are extremely rare, rarer still in contexts where rent-seeking outcomes can be objectively measured. The devolution of irrigation management in Pakistan’s Indus Basin provides just such a felicitous combination. We have compared changes in water discharge along channels whose management was taken over by locally elected farmer organizations (FOs) to changes that occurred in channels that remained centrally managed. Water theft increased by more in the former case, leading to a large redistribution of wealth. That decentralization also increased water theft by more along channels with a greater preponderance of large landowners at the head suggests that investment in de facto political power (borrowing the terminology of Acemoglu and Robinson 2008) can sometimes more than offset changes in de jure political power brought about by institutional reform. Here, as our theoretical model indicates, decentralization shifts the lobbying arena from the upper-tier of the bureaucratic hierarchy to the communal governance structure, which leads to greater rent-seeking. While our evidence is not favorable to the decentralization effort in the Indus Basin inasmuch as it did not deliver on its promise of a more equitable (and efficient) distribution of canal water, it would be premature to throw out the reform baby with the bathwater. 25 For the calculation of the Gini coefficient, landholdings of landless households are set to 10−8 acres. 35 Successful decentralization would likely involve directly addressing power asymmetries along the irrigation system, such as by giving tail-enders exclusive control over FOs.26 In terms of the model, this policy would make it more likely that the efficient allocation is implemented; tail-enders would have every incentive to enforce the warabandi system, while head-end influence within the FO would be minimized. Regardless of the precise governance structure, however, continual support of the central government, both in setting and enforcing the rules of the game, is critical to effective local administration (see Mansuri and Rao 2013). 26 Merely establishing reservations whereby each FO must have a certain number of members or officers representing tail outlets may not be enough. In our data, a reasonable proportion of FO presidents (20%) and of the four-member FO management committee (18% on average) own land at the tail. In these cases, tail- enders’ interests are nominally represented in the FO. However, analysis similar to that in Table 4 (available from the authors upon request) reveals no significant difference in decentralization outcomes between FOs with and without tail representation. To be sure, caution must be exercised in interpreting these results as we do not understand why some FOs have officers with tail-holdings and others do not. Nevertheless, taken on their own terms, these findings do not support a partial reservation for tail-enders. 36 References [1] Acemoglu, D., & Robinson, J. (2008). Persistence of Power, Elites, and Institutions. American Economic Review, 98(1), 267-293. [2] Alatas, V., Banerjee, A., Hanna, R., Olken, B. A., Purnamasari, R., & Wai-Poi, M. (2013). Does elite capture matter? Local elites and targeted welfare programs in In- donesia (No. w18798). National Bureau of Economic Research. [3] Anderson, S., Francois, P., & Kotwal, A. (2015). Clientelism in Indian villages. American Economic Review, 105(6), 1780-1816. [4] Baland, J. M., & Platteau, J. P. (1997). Wealth inequality and efficiency in the commons Part I: the unregulated case. Oxford Economic Papers, 49(4), 451-482. [5] Baland, J. M., & Robinson, J. A. (2008). Land and power: Theory and evidence from Chile. American Economic Review, 98(5), 1737-1765. [6] Bardhan, P. (2000). Irrigation and cooperation: An empirical analysis of 48 irrigation communities in South India. Economic Development and cultural change, 48(4), 847- 865. [7] Bardhan, P., & Dayton-Johnson, J. (2002). Unequal irrigators: heterogeneity and com- mons management in large-scale multivariate research. The drama of the commons, 87-112. [8] Bardhan, P., & Mookherjee, D. (2006a). Decentralisation and accountability in infras- tructure delivery in developing countries. Economic Journal, 116(508), 101-127. [9] Bardhan, P., & Mookherjee, D. (2006b). Decentralization, corruption and government accountability. International handbook on the economics of corruption, 6, 161-188. 37 [10] Bardhan, P., & Mookherjee, D. (2006). Pro-poor targeting and accountability of local governments in West Bengal. Journal of development Economics, 79(2), 303-327. [11] Bandaragoda, D. J., & ur Rehman, S. (1995). Warabandi in Pakistan’s canal irrigation systems: Widening gap between theory and practice. No. 7. IWMI. [12] Basurto, P., Dupas, P., & Robinson, J. (2015). Decentralization and efficiency of subsidy targeting: Evidence from chiefs in rural Malawi. Unpublished, Stanford University. [13] Beath, A., Christia, F., & Enikolopov, R. (2017). Direct democracy and resource allo- cation: Experimental evidence from Afghanistan. Journal of Development Economics, 124, 199-213. [14] Bromley, D. W., Taylor, D. C., & Parker, D. E. (1980). Water reform and economic development: institutional aspects of water management in the developing countries. Economic Development and Cultural Change, 365-387. [15] Chakravorty, U., & Roumasset, J. (1991). Efficient spatial allocation of irrigation water. American Journal of Agricultural Economics, 73(1), 165-173. [16] Chambers, R. (1988). Managing canal irrigation: Practical analysis from south Asia. New Delhi: Oxford. [17] Dayton-Johnson, J. (2000). Determinants of collective action on the local commons: a model with evidence from Mexico. Journal of Development Economics, 62(1), 181-208. [18] Dixit, A. (2002). Incentives and Organizations in the Public Sector: An Interpretative Review. Journal of Human Resources, 37(4), 696-727. doi:10.2307/3069614 [19] Fan, C. S., Lin, C., & Treisman, D. (2009). Political decentralization and corruption: Evidence from around the world. Journal of Public Economics, 93(1), 14-34. 38 [20] Guimaraes, P. and P. Portugal (2010). A Simple Feasible Alternative Procedure to Estimate Models with High-Dimensional Fixed Effects, Stata Journal, 10(4), 628-649. [21] Hoffman, V., P. Jakiela, M. Kremer, and R. Sheely (2017). There is no place like home: Theory and evidence on decentralization and politician preferences. Unpublished. [22] Jacoby, H. & Mansuri, G. (2018). Governing the Commons? Water and Power in Pakistan’s Indus Basin. Policy Research Working Paper (forthcoming). World Bank, Washington DC. [23] Mansuri, G., and Rao, V. (2013). Localizing development: Does participation work?. World Bank Publications. [24] Mookherjee, D. (2015). Political decentralization. Annual Review of Economics 7, no. 1 (2015): 231-249. [25] Meinzen-Dick, R. (2007). Beyond panaceas in water institutions. Proceedings of the National Academy of Sciences, 104(39), 15200-15205. [26] Nitzan, S. (1994). Modelling rent-seeking contests. European Journal of Political Econ- omy 10, no. 1: 41-60. [27] Ostrom, E. (1990). Governing the Commons: The Evolution of Institutions for Collec- tive Action. New York: Cambridge University Press. [28] Ostrom, E. & Gardner, R. (1993). Coping with asymmetries in the commons: self- governing irrigation systems can work. Journal of Economic Perspectives, 7(4), 93-112. [29] Reinikka, R., & Svensson, J. (2004). Local capture: evidence from a central government transfer program in Uganda. Quarterly Journal of Economics, 119(2), 679-705. 39 [30] Rijsberman, F. R. (2008). Water for food: corruption in irrigation systems. Global Corruption Report 2008, P. 67-84. [31] Rinaudo, J. D., Strosser, P., & Thoyer, S. (2000). Distributing water or rents? Ex- amples from a public irrigation system in Pakistan. Canadian Journal of Development etudes du d´ Studies/Revue canadienne d’´ eveloppement, 21(1):113-139. [32] Rinaudo, J. D. (2002). Corruption and allocation of water: the case of public irrigation in Pakistan. Water Policy, 4(5), 405-422. [33] Sampath, R. K. (1992). Issues in irrigation pricing in developing countries. World De- velopment, 20(7), 967-977. [34] Tullock, G. (1980). Efficient rent-seeking. In Buchanan, J.; Tollison, R.; & Tullock, G., Toward a theory of the rent-seeking society, pp. 97-112. [35] Van Waijjen, E. G., Hart, W. W. H., Kuper, M. & Brouwer, R. (1997). Using a hydro- dynamic flow model to plan maintenance activities and improve irrigation water dis- tribution: application to the Fordwah distributary in Punjab, Pakistan. Irrigation and Drainage Systems 11, no. 4: 367-386. [36] Vermillion, D. L. (1997). Impacts of irrigation management transfer: A review of the evidence. Sri Lanka: International Irrigation Management Institute. [37] Wade, R. (1982). The system of administrative and political corruption: Canal irrigation in south India, J. of Development Studies, 18(3): 287-328. [38] World Bank (1994) Pakistan irrigation and drainage: Issues and options. Agricultural Operations Division, South Asia Region. Washington, DC: World Bank. [39] World Bank (2004) World Development Report 2004: Making Services Work for Poor People. Washington, DC. 40 Appendix A Proofs Lemma 1. bF < bB . Proof. The FOC for the bureaucratic official is VB (bB ) = 0 (the SOC VB < 0 is always satisfied) and for the FO official is VB (bF ) + uc (bF ) = 0, for cases c = 1, 2. Since equilibrium efforts are ej = P ˜ )∆j , we have eH + eT = P ˜ (1 − P ˜ (1 − P˜ )(∆H + ∆T ) = ∆H ∆T /(∆H + ∆T ) at the point of equal marginal influence, θ = 1. Thus, U (·)∆2 T ∆H u1 (bB ) = <0 (∆H + ∆T )2 ˆ . Likewise, with eH = ∆2 since ∆H (b) = −n 2 H ∆T /(∆H + ∆T ) , we have U (·)∆2 T ∆H ∆H u 2 ( bB ) = < 0. (∆H + ∆T )3 It follows that, in both cases 1 and 2, VB (bF ) > VB (bB ), which, by VB < 0, ⇒ bF < bB . ˜ (br )w, Proof of Proposition 1: We have T S (br ) = T Sr = P ˆ and Taylor expansion T S (bF ) ≈ T S (bB ) + T S (bB )(bF − bB ). ˜ (br ) < 0 ⇒ T S (bB ) < 0, lemma 1 ⇒ T SF > T SB . Since P Proof of Proposition 2: We have Taylor expansion T Sθ (bF ) ≈ T Sθ (bB ) + T Sθb (bB )(bF − bB ). ˜θb (bB )w where subscripts denote partial derivatives and T Sθb (bB ) = P ˆ . Next, ˜ ˜ ˜b bθ − Pθ . ˜θb = P + P P b b P˜ Evaluated at bB , the expression in square brackets is proportional to VB (bB ) and thus vanishes. Since P˜θ > 0, T Sθb (bB ) < 0, which, with lemma 1, proves the result. 41 B Additional Figures and Tables .06 .04 De-meaned tail shortage -.02 0 .02 -.04 2006 2007 2008 2009 2010 year FO (excluding LCC East & West) non-FO Figure B.1: Tail shortage pre-trends in FO and non-FO circles 42 Census Farm Households 80 80 tail gen. mean (q=1.5) 60 60 tail mean 40 40 20 20 0 0 0 20 40 60 0 20 40 60 head mean head gen. mean (q=1.5) 80 80 60 60 tail 75th pctile tail 90th pctile 40 40 20 20 0 0 0 20 40 60 0 20 40 60 head 75th pctile head 90th pctile All Census Households 50 50 tail gen. mean (q=1.5) 40 10 20 30 40 tail mean 20 30 10 0 0 0 10 20 30 40 50 0 10 20 30 40 50 head mean head gen. mean (q=1.5) 50 50 20 30 40 20 30 40 tail 75th pctile tail 90th pctile 10 10 0 0 0 10 20 30 40 50 0 10 20 30 40 50 head 75th pctile head 90th pctile Figure B.2: Land ownership statistics at FO channel head and tail 43 Table B.1: Descriptive Statistics for Sample Channels FO circles Non-FO circles N Number of outlets Position on parent N Number of outlets Position on parent Canal type (%) Head Middle Tail Head Middle Tail (%) Head Middle Tail Head Middle Tail Distributary 453 9.9 7.0 9.8 0.31 0.40 0.29 788 10.5 8.0 10.5 0.28 0.38 0.34 (45.0) (11.1) (8.3) (9.8) (0.46) (0.49) (0.45) (42.5) (10.7) (8.2) (9.5) (0.45) (0.49) (0.47 Minor 488 3.0 2.8 4.7 0.44 0.37 0.19 921 4.3 3.7 5.6 0.41 0.42 0.17 44 (48.5) (4.3) (3.5) (3.5) (0.50) (0.48) (0.39) (49.7) (5.1) (3.7) (4.3) (0.49) (0.49) (0.37 Sub-minor 66 1.8 1.8 3.7 0.53 0.38 0.09 144 2.8 2.6 4.1 0.45 0.40 0.15 (6.8) (2.4) (2.2) (2.5) (0.50) (0.49) (0.30) (8.0) (3.6) (2.7) (2.6) (0.50) (0.49) (0.35 Total 1007 6.0 4.7 7.0 0.39 0.38 0.23 1853 6.8 5.4 7.6 0.36 0.40 0.24 (100) (8.9) (6.6) (7.7) (0.49) (0.49) (0.42) (100) (8.5) (6.4) (7.4) (0.48) (0.49) (0.43) Notes: Figures are means or proportions (standard deviations in parentheses) unless otherwise noted. FO circles consist of the five Area Water Boards that formed FOs under the irrigation reform; non-FO circles consist of the 12 Area Water Boards that did not form FOs. Head, middle, and tail are defined as, respectively, the first 40%, second 40%, and last 20% of a channel by length. Parent refers to the canal from which a channel off-takes. Table B.2: Choice of Reference Sample: A Simple Example Most Highest mean landholdings Comparison powerful Farm hhs All hhs 1 vs. 2 1 = 1 1 vs. 3 = = 1 2 vs. 3 3 = 3 Community 1: 100 x 100 acre farm hhs Community 2: 1 x 100 acre farm hh + 99 x non-farm hhs Community 3: 100 x 100 acre farm hhs + 100 x non-farm hhs Notes: Most powerful community is based on the number of large landowners (equal sign denotes equally powerful). Mean landholdings refers to average across farm households (col. 3) or average across both farm and non-farm households (col. 4). 45