The Value of Regulatory Discretion: Estimates from Environmental Inspections in India

Transcription

1 The Value of Regulatory Discretion: Estimates from Environmental Inspections in India Esther Duflo Michael Greenstone Rohini Pande Nicholas Ryan November 30, 2016 Abstract Should regulators be rule-bound, or does regulatory discretion, by allowing the use of private information, actually strengthen enforcement? In collaboration with the Gujarat (India) Pollution Control Board, we doubled inspection frequency for a random group of polluting industrial plants. The added inspections led to regulatory citations, but little improvement in environmental outcomes. We show evidence that the treatment inspections, which were allocated evenly across all plants, were poorly targeted relative to inspections allocated at the regulator s discretion in the status quo. To understand the experimental findings, we specify a structural model of regulator-plant interactions, where the regulator uses signals of plant emissions to target inspections, in order to maximize abatement, and plants abate only when threatened with regulatory penalties. The model results suggest that discretion is valuable; treatment inspections would have had a larger effect on pollution, had they been targeted using the regulator s private information. At the control level of inspections, removing discretion altogether, by mandating equal inspection rates for all plants, would have reduced plants pollution abatement by two-thirds. I Introduction Regulatory agencies enforce standards using imperfect information on the parties they regulate (Laffont and Tirole, 1993). This imperfection suggests a need for policy to be flexible, to allow regulators to use local information. Yet, flexibility lends discretion, which may lead to enforcement that reflects regulators personal objectives, rather than the social goals of regulation We thank Sanjiv Tyagi, R. G. Shah, and Hardik Shah for advice and support over the course of this project. We thank Pankaj Verma, Eric Dodge, Vipin Awatramani, Logan Clark, Yuanjian Li, Sam Norris, Nick Hagerty, Susanna Berkouwer and Kevin Rowe for excellent research assistance and numerous seminar participants for comments. We thank the Sustainability Science Program (SSP), the Harvard Environmental Economics Program, the Center for Energy and Environmental Policy Research (CEEPR), the International Initiative for Impact Evaluation (3ie), the International Growth Centre (IGC) and the National Science Foundation (NSF Award # ) for financial support. All views and errors are solely ours. MIT, eduflo@mit.edu University of Chicago, mgreenst@uchicago.edu Harvard, rohini_pande@harvard.edu Corresponding author. Yale, nicholas.ryan@yale.edu. 1

2 (Stigler, 1971; Leaver, 2009). This tension between rules and discretion underlies many debates on the design of regulatory agencies (Sitkin and Bies, 1994). 1 Environmental regulation in emerging economies is an important case in point. Countries like China and India report pollution that exceeds the highest levels ever recorded in rich countries, with dire health consequences (Chen et al., 2013; Greenstone et al., 2014). This pollution has often grown despite strict regulatory standards for emissions on the books. To see whether discretion may compromise, or help uphold, pollution standards, we collaborated with the Gujarat Pollution Control Board (GPCB), the agency in charge of enforcing environmental regulation in Gujarat, India, on a two-year field experiment that increased the frequency of environmental inspections for some industrial plants. The de jure environmental standards are strict maximum limits on the concentrations of pollutants that plants emit in air and water. There are also regulatory requirements for inspection frequency and national laws that build a criminal framework for punishment, giving regulators broad power to close down highly polluting plants. However, de facto, the GPCB does not have enough resources to inspect all plants at the required rate, and uses considerable discretion in deciding whom to inspect and penalize. In the control group, despite widespread noncompliance with pollution standards, 44% of plants are inspected less than once per year during the experiment. The experiment was run in a sample of 960 industrial plants, of which half were assigned to the treatment group of increased inspection frequency. 2 This inspection treatment had two features. First, it provided the resources necessary to bring all treatment plants up to at least the required minimum frequency of inspections. Second, it removed the regulator s discretion over these added inspections by allocating them randomly across all treatment plants. (The regulator carried on allocating its existing budget of inspections however it chose in both the treatment and control groups.) Aside from the added inspections and their allocation, the treatment did not alter pollution standards or the regulatory penalties for violations. The analysis is conducted with unusually rich data on the regulatory process, plant abatement 1 For example in India, the setting for this study, the central government in 2014 reformed labor inspections to be randomly assigned, deliberately to reduce discretion among inspectors (Indian Express, 2014). 2 The inspection treatment was cross-randomized with an audit reform experiment in the subset of 473 sample plants that were eligible for environmental audits (Duflo et al., 2013). In 1996, the High Court of Gujarat ordered GPCB to instate a third-party audit system wherein plants with the potential to be highly polluting must provide an annual audit report to GPCB. Duflo et al. (2013) evaluated a reform of this audit system and found that making third-party auditors more accountable to the regulator and less beholden to the plants they audit improves truth-telling and lowers pollution. 2

3 and pollution responses. On the regulatory process, we code 9,624 pieces of correspondence between the regulator and sample plants, which record their interactions over five years (from two years before the experiment through one year after). These documents include plant inspections, pollution readings, regulatory notices and penalties, as well as written responses from plants, such as documentation of the installation of abatement equipment. On plant behavior, we conducted an endline survey of pollution and abatement costs independently of the regulator, so that our outcome variables were not used in the regulatory process. The experimental results show that the added treatment inspections largely failed to reduce pollution. To unpack why, consider three links in the causal chain from inspections to emissions. First, the experiment was implemented, in the sense that inspection rates in the treatment group were twice those in the control, and treatment plants report higher perceived inspection rates, suggesting the scrutiny was felt. Second, treatment plants were more often found in violation of pollution standards and received more citations, which threaten punishment. Third, however, the higher level of scrutiny had little effect on actual punishment or plant abatement: treatment plants were not more likely to be penalized, and, while a few more plants that were already close to pollution standards complied in the treatment group, we cannot reject the null of no average treatment effect on plant pollution emissions. We argue that the last link of the chain broke because treatment inspections, allocated without discretion, were poorly targeted. While the treatment inspections found many violations, they did not detect more extreme violations, which are the kind that the regulator usually targets, and for which, in practice, it reserves the most costly penalties. We set out a model of regulator-plant interactions, with the aims of understanding how much regulation costs for plants and, from the side of the regulator, the value of discretion in targeting inspections. The model includes two stages. In the targeting stage, the regulator choses who to inspect, based on noisy signals of plants pollution, and plants decide whether to abate pollution given the threat of inspections. The second is a penalty stage where, after an inspection, penalties may be levied on non-compliant plants. The regulator s objective is to choose an inspection rule that maximizes abatement, subject to a budget constraint on total inspections, and taking behavior in the penalty stage into account. To keep the problem simple, we estimate regulatory action on penalties as a reduced-form policy function in the data. We 3

4 refer to this policy function henceforth as the regulatory machine, and hold it fixed in estimation and counterfactuals. This approach has the advantage of endogenizing inspections, which the experiment changed, while leaving untouched the penalty policy, which the experiment did not alter. 3 The plant s objective is to minimize total costs by trading off costly abatement actions against the risk of future inspections, and the penalties they may beget, which will depend on plant pollution. We estimate the model using the experimental data and the exogenous variation in inspection probability created by the experiment. Beginning with the penalty stage, the results show that regulatory penalties are costly when applied, but that the risk of penalties is remote. We estimate penalties in the model and find that a realized plant closure is estimated to cost a substantial US$ 50 thousand. 4 We then use these costs and the probability of penalties, given optimal decision-making by a plant in the penalty stage, to calculate the expected discounted value of regulation for a plant, at the time of an initial inspection. Given the low risk of penalties, the expected discounted value of an initial inspection to the plant is negative US$ 2 thousand. Even for plants found, on initial inspection, with a pollution reading of at least five times the regulatory standard, the expected value is negative US$ 6 thousand greater than for a plant with average pollution, but far smaller than the cost of certain punishment. A contribution of the model is to allow us to estimate costs of regulation that are basically never observed, like hassle, disruption and bribes. The model is estimated using the plant s revealed preference, in a dynamic problem, of when to install costly equipment or to wait and risk penalties. We observe the cost of equipment, and the model measures penalties and other unobserved costs of regulation as those costs that rationalize plants choices to install this equipment when faced with a risk of penalty. Thus, our cost estimates include direct abatement costs, formal penalties (such as the costs of having your plant closed down), as well as the hassle and bribes that may come with having your plant inspected or negotiating with the regulator. We find that these other kinds of regulatory cost are very important; after an inspection, about 3 We call this policy function the regulatory machine to emphasize that the regulatory objective is not modeled in the penalty stage the regulator just acts as in the data. A further argument for this approach is that, in reality, the regulatory machine may not be in the control of a single agent, like an official deciding how to target inspections. It is rather a complicated function of the behavior of many staff, actingunderlegalrulesand reputational concerns. 4 This cost is equivalent to about two months of mean plant profits, assuming a 10% profit margin, or four months of mean plant electricity bills. 4

5 60% of the expected value of regulation to plants is due to expected future penalties, including closure, hassle costs, and the like, and the remainder from expected future abatement capital expenditures. The penalty estimates suggest targeting may be valuable. Since penalties, after an inspection, depend on the level of pollution an inspection finds, the regulator may induce more compliance by targeting inspections at dirtier plants. The experiment, by exogenously increasing inspection probability, provides a key source of identifying variation for the targeting stage, in which the firm decides whether to pre-emptively run abatement equipment to reduce future penalties. Turning to model results on targeting, we indeed find the regulator aggressively targets plants with higher pollution signals, even though its signals are weakly related to the true pollution level. For example, plants that the regulator expects to be the dirtiest are inspected about ten times per year, while nearly half of plants get less than one inspection annually. This type of aggressive targeting is optimal, despite the noisiness of the regulator s signal on plant emissions, because most plants have strictly positive costs of running their abatement equipment. If the regulator instead spread its inspections around evenly, all the plants will anticipate getting very few inspections, and none may find it worthwhile to abate. We use the model estimates to study counterfactual inspection regimes. We solve for the targeting rule that achieves the largest reduction in pollution emissions, varying the regulator s total inspection budget and the strength of its emissions signal, while holding the regulatory machine fixed. Additional inspections give the regulator more chances to find a violation and more reliable emissions signals allow better targeting. These counterfactuals are matched to the horizon of the experiment, and should be thought of as medium-run impacts, because they hold fixed the structure of penalties and, on the plant side, abatement capital. The simulations reveal that targeting inspections is crucial to inducing abatement. Relative to the status quo of targeting based on noisy signals of emissions, a policy that required the regulator to spread inspections evenly over all plants would cause total pollution abatement to fall by roughly two-thirds. Discretion is still valuable, but less so, as inspection levels rise. We estimate that treatment inspections would have induced fifeen percent more abatement than observed, if they had been allocated optimally using the regulator s information, instead of at random. Improving information makes discretion more powerful. If the regulator had a technology that 5

6 provided perfect information on emissions, and used this to target inspections and thus penalties, it could increase abatement 30% with a constant inspection budget. 5 By contrast, under uniform targeting there would be no change in policy or abatement with better monitoring technology. Such technologies are not a matter of science fiction: continuous emissions monitoring systems (CEMS) are used widely in the United States, and India has announced plans to roll out these devices in heavily polluting sectors (Central Pollution Control Board, 2013a, 2014a). The combination of a large-scale field experiment and structural estimation allows us to make several contributions. First, we believe this paper provides the first experimental evidence of the effect of inspection frequency on compliance with environmental regulations. 6 Second, our rich data on the regulatory process sheds light inside a previously black box, which holds the negotiations between the regulator and polluters. We use this data to structurally estimate a model of regulatory targeting and penalties and estimate the full costs of environmental regulation. Prior research in this spirit has been mainly theoretical, including influential work arguing that dynamic incentives, like those in our model, are necessary to explain why plants comply despite infrequent inspections (Harrington, 1988). Third, using the experimental variation in inspection frequency within the structural model allows us to value regulatory discretion, which has been recognized as important but very seldom empirically measured. 7 The rest of the paper runs as follows. Section II describes environmental regulation in India, 5 Of course, if information were truly perfect, inspections may become redundant altogether. In the present legal regime inspections are still necessary, even under perfect information, because only pollution samples taken in person during inspections have legal standing. 6 Past studies of regulatory inspections in the United States show that inspections reduce pollution significantly (Hanna and Oliva, 2010; Magat and Viscusi, 1990). All of these studies rely on observational data wherein dirtier plants are more likely to get inspections, and this endogeneity is a strong concern (See Shimshack (2014) for a recent survey.). Studies of regulatory efficacy in emerging economies are more mixed, with Tanaka (2013), for example, finding large reductions in pollution from a control policy in China and Greenstone and Hanna (2014) finding cuts in pollution in India from policies targeting air, but not water, pollution. On the costs of regulation, Greenstone et al. (2012) find U.S. regulations lower manufacturing productivity, and Ryan (2012), using a dynamic model, finds that the U.S. Clean Air Act Amendments raised entry costs in the cement industry. There are few studies of environmental regulation in developing countries, but negative productivity effects of rigid industrial regulation have been documented for India (Aghion et al., 2008; Besley and Burgess, 2004). 7 Environmental regulation is a classic setting for incentive regulation (Laffont and Tirole, 1993; Laffont, 1994; Boyer and Laffont, 1999). Limited regulatory capacity and limited commitment will typically cause the optimal regulatory policy in emerging economies to differ dramatically from settings with complete markets (Laffont, 2005; Estache and Wren-Lewis, 2009). Papers in organizational economics, such as Aghion and Tirole (1997) suggest that formal and real authority may then optimally diverge leading to substantial value for discretion. More broadly, our findings resonate with an emerging literature on effective policy design when state capacity is limited (Besley and Persson, 2010). For instance, consistent with this paper s findings Rasul and Rogger (2013) report significant gains from providing Nigerian bureaucrats autonomy in decision-making. Other papers note, to the contrary, that discretion in environmental regulation may be misused by bureaucrats and politicians (Burgess et al., 2012; Jia, 2014). 6

7 the experimental design and data collection. Section III gives reduced-form results on the effects of additional inspections on regulatory actions and plant pollution behavior. Section IV presents a model of regulatory targeting with dynamic penalties. Section V describes the empirical estimation of this model, and Section VI presents the structural estimation results on regulatory penalties and targeting. Section VII uses the model to simulate counterfactuals on the effect of regulatory discretion on pollution abatement. Section VIII concludes. II Context and Experimental Design This section describes the Indian regulatory framework, the experimental design and data, and presents summary statistics on the sample plants and their interactions with the regulator. A Context: Regulation of Industrial Pollution in India India has stringent command-and-control environmental regulations, which for industrial plants take the form of maximum allowable concentration standards for emissions in air and water. States may make their standards more strict than the national standards, but cannot relax them (Ministry of Environment and Forests, 1986). State Pollution Control Boards conduct basically all enforcement of environmental regulations. We describe the enforcement practices of our partner regulator, the Gujarat Pollution Control Board (GPCB), which are largely common with other Indian states. GPCB has 19 regional offices, each with several inspection teams and a Regional Officer in charge of assigning inspections. Plant inspections are undertaken by full-time engineer and scientist employees. During an inspection, the team observes the state of the plant and its environmental management and often, but not always, collects pollution samples for laboratory analysis. The inspection report and pollution readings from laboratory analysis enter the plant s file. Officers at the regional and head office review inspection and analysis reports, which include qualitative information, such as on waste management practices and plant housekeeping, the plant s history of violations and, of particular importance, pollution concentrations for air and water pollutants. The officers, not the field staff, then decide further action, including whether to inspect the plant again, warn the plant or impose penalties (on which see below), or to leave the plant alone. A plant inspection may occur for several reasons. First, regulations mandate routine inspec- 7

8 tion of plants in sectors with the highest pollution potential ( red category plants) every 90 days if they are large- or medium-scale and once per year if they are small-scale. 8 Routine inspections make up 35% of the total. Second, inspections occur when plants apply for an environmental consent, a license to operate a certain type of manufacturing, and for consent renewal every five years (30% of inspections). Third, the inspection may be to follow-up on a violation found in a prior inspection, or to verify that the plant has taken a specific action as it claimed (24% of inspections). The balance of 11% of inspections are in response to public complaints and for other miscellaneous reasons. Given these competing priorities and insufficient resources, inspection staff are spread thin. 9 Data from the year before the experiment shows that inspection rates for red-category plants of small-, medium- and large-scale occur roughly half as often as prescribed (once in 187, 224 and 534 days, respectively). A subset of red-category plants are also required to submit annual environmental audit reports to the regulator. Auditors conflict of interest has historically rendered these audits inaccurate, such that inspections remain the main way the regulator monitors plants (Duflo et al., 2013). While inspections are infrequent, plants found in violation of pollution standards can be harshly penalized. National environmental regulations empower the regulator to issue written directions for polluting plants to take abatement action or for the stoppage of water, electricity or other services to those plants (Ministry of Environment and Forests, 1986). The regulator applies these penalties often, including mandating that a plant install abatement equipment, mandating that a plant post a bond against future performance and ordering that a plant s water and electricity be disconnected, closing the plant down. Utility disconnections can remain in force until the plant has shown progress towards meeting environmental standards, often by installing abatement equipment. Because abatement capital is observable, it is used to demonstrate compliance, even when an initial pollutant violation could be remedied by operational changes. The duration of closure depends on the offense but is typically several weeks; in our data the 8 The GPCB follows a government classification for plants based on their reported scale of capital investment, with small-scale being investment less than INR 50m (US$ 1m), medium INR 50m to 100m (US$ 1m to 2m) and large above 100m (US$ 2m). Throughout the paper, we use an exchange rate of US$ 1 = INR 50. Prescribed inspection rates for these plants are comparable to those applied to large plants, by air pollution potential, in the United States (Hanna and Oliva, 2010). 9 Bhushan et al. (2009) note that the approved number of employees for many Boards has declined even as the number of plants the Boards were to regulate doubled or trebled. Of nominally approved staff positions, many Boards left a quarter or more unfilled due to financial or bureaucratic difficulties in hiring. 8

9 median duration of closure is 24 days. The regulator, in principle, can also take a violator to court for criminal sanction, but this is rare and does not occur in our data, because documenting violations to legal standards is burdensome and there are long delays in prosecution. In addition to formal penalties like closure, plants may incur other costs of regulation, such as disruptions to plant operations during inspections, or bribes to suppress inspection results or penalties. B Experimental Design Between 2009 (quarter 3) and 2011 (quarter 2), we worked with GPCB to raise inspection frequency for some plants within a sample of 960. The sample was drawn as follows. First, we identified the population of 3,455 red-category (i.e., high pollution potential) small- and mediumscale plants in three regions of Gujarat (Ahmedabad, Surat and Valsad). This population, which constitutes roughly 15% of the more than 20,000 regulated plants in Gujarat, is of special interest, as it is highly polluting but hard to regulate. From this population, we selected all 473 audit eligible plants in Ahmedabad and Surat. 10 Next, we randomly selected 488 plants from the remaining audit-ineligible population. Inspection treatment assignment was randomized within region by audit-treatment-status strata (treatment: 233 plants, control: 240 plants and non-audit-eligible: 488 plants). The treatment was thus cross-randomized and implemented concurrently with the audit reform treatment studied by Duflo et al. (2013). The 481 plants assigned to the inspection treatment were guaranteed at least one annual inspection. Specifically, in the first quarter the plant was assigned an initial inspection, after which it was randomly assigned on a quarterly basis to be inspected again with probability 0.66, conditional on strata. After four quarters this cycle started over. Thus the treatment assigned extra variation in inspection frequency within the treatment group of plants, up to a maximum of four inspections per year. Treatment inspections were conducted by regional GPCB teams consisting of an environmental engineer and scientist. In order to increase inspections without overburdening GPCB staff, we worked with GPCB to rehire three recently retired GPCB scientists (GPCB judged that scientists, who are the only staff able to take certain pollution samples, were the main constraint in 10 Audit-eligible plants are a kind of super-red category selected on three dimensions: what the plant produces, where it sends its effluent (i.e., wastewater), and the volume of that effluent. Most audit-eligible plants in our sample are eligible because they discharge more than 200 kl per day of effluent. Due to the requirements for audit eligibility, which were originally targeted at plants in Ahmedabad, few if any plants in Valsad are audit-eligible. 9

10 conducting additional inspections). 11 These newly hired staff were in fact sometimes allocated to regular inspections, and regular staff often allocated to inspections assigned under the treatment, so that the teams were well-mixed. Using administrative data on staff assignments across all three regions, we observe that only one staff member, who was newly rehired, participated in treatment inspections only, whereas 32 staff members, mostly current employees, participated in both treatment and non-treatment inspections. Each morning, the inspection teams in each region assigned to do treatment inspections were randomly assigned a list of plants from the treatment group at which to conduct an initial routine inspection that day. This mimicked GPCB s practice of assigning teams to plants each day, except that for these teams the plant assignment, rather than being based on an official s discretion, was random. In all respects but targeting, control and treatment inspections were similar. Reports from treatment and control inspections were similarly used: the reports entered the same database, had samples analyzed by the same GPCB labs and had the same officials at the GPCB deciding whether or not to take action against plants. In Online Appendix Table B1, we test extensively for whether or not these follow-through actions were taken differently by treatment status, but find that treatment status did not influence actions, conditional on the results of an inspection. Towards the end of the inspection treatment, in the month prior to the endline survey, we also assigned, randomly and independently of the other treatments, some plants to receive a letter from GPCB reminding them of their obligations to meet emissions limits. This letter reiterated the terms of plants environmental consent, which in principle they already knew, but it may have also served to increase the salience of regulatory compliance. C Data The analysis uses two sources of data, an endline plant survey and GPCB administrative records. Independent agencies, mainly engineering departments of local universities, conducted the endline survey between April and July, 2011 (the experiment ended in June 2011), supervised by the research organization J-PAL South Asia. The survey collected pollution readings, expenditures for investment in, and operation of, pollution abatement equipment, and data on other aspects of plant operations. The GPCB issued letters that required plants to cooperate with the 11 GPCB observes a mandatory retirement age so many retired staff are willing to work. GPCB identified recently retired but interested staff in each region. 10

11 surveyors and stated truthfully that the results would not be used for regulatory matters. With this regulatory help, attrition was low and did not differ by treatment status; 12.9% of sample plants closed during the study and only 4.7% attrited for other reasons (See Online Appendix Tables B2 and B3). The second source of data, GPCB records of its own interactions with sample plants, allows us to examine the process of regulatory enforcement. Regulation, in practice, does not move directly from inspection to penalty but is mediated by formal, sometimes lengthy, bargaining between the regulator and regulated plants. The records in our data consist of internal regulatory documents, letters and actions sent to plants and the written responses that plants send back. Table 1 presents an overview of the types of documents in the data, grouped by the party that created a document and the actions that the documents represent, in our model of the regulatory machine and penalties. We observe 9,624 documents, of varying types, all of which are dated and many of which reference one another. The table maps what we observe in these documents into the regulatory and plant actions we will use in estimation. Table 1, Panel A describes the four actions of the regulator. When the regulator chooses Inspect, an inspection is documented in an inspection report, which describes GPGB s physical observations on visiting a plant. If, during an inspection, pollution samples are taken, then the inspection report is accompanied by an analysis report, giving the lab results on pollution concentrations for all samples. The action Warn covers several documents which threaten, but do not directly punish, the plant. These documents include letters urging the plant to clean up, citations for violations of pollution standards and closure notices that warn that the plant will be closed, absent some action to improve. The action Punish records only actual costly punishments, mainly plant closure. Plant closure is documented by a closure direction sent to the plant and/or a notice to the utility to cut off the plant s water or electricity. Finally, the regulator may, after aviolation, Accept that the plant is in compliance. Acceptance is documented by the regulator writing the plant to revoke a closure direction. If there is no such revocation, acceptance is also inferred, any time the regulator has not taken follow-up action within sixty days. Panel B of the table describes the actions of the plant and their underlying documentation. The plant has only two actions. We designate with the plant action Comply documents that record the installation of abatement equipment, typically with a certification or invoice from the 11

12 vendor that did the work. Such installation is costly to the plant. The alternative for the plant is to Ignore the regulator. If the plant writes to the regulator to protest but does not demonstrate evidence of compliance, we also class this as Ignore, since protests are essentially cheap talk and involve no expenditures. 12 The action Ignore is also inferred from the absence of any plant response between regulatory actions. D The regulatory machine The regulator and plants interact in turn, responding to one another over time to resolve whether a plant should be punished or not based on the results of a given inspection. We group these actions into chains of related actions, using the characteristics of the underlying documents, and will refer to the regulator s policy function within a chain as the regulatory machine. There are three rules for linking actions in a chain. First, any document that explicitly references another document, and thus action, is linked to that prior action (accounting for 13% of links). So, if the regulator issues a citation referencing a particular inspection report, which found a violation, then the Warning action is linked to that prior Inspection. Second, actions concerning the same plant that follow one another within 30 days (or 60 days, for certain document types with longer lags) are linked. So, if the regulator issued a citation two weeks after an inspection, those two documents would be linked, even if the citation did not explicitly refer to the recent inspection. Third, within a chain and after the above links are formed, the regulator and the plant must alternate moves. We therefore impute the action Ignore for plants between consecutive regulatory moves (8,382 times), and the action Accept for the regulator, if the regulator leaves a plant alone without documenting this decision. Appendix A describes the data cleaning in more detail. The resulting chains of interactions between the regulator and plants give a picture of the probabilities of punishment and plant compliance. Table 2 summarizes the structure of the chained interactions between the regulator and plants across rounds. 13 Columns 1 through 6 give the frequency of actions of the regulator or plant in that round, from regulatory records, 12 For example, a plant where GPCB found high air pollution readings claimed in correspondence: At the time of visit our chilling plant accidentally failed to proper working, so chilling system of scrubber was not effective by simple water. Same time batch was under reaction and we were unable to stop our reaction at that time. Now it is working properly. That is, a piece of air pollution control equipment failed, causing pollution to be higher than normal during the visit. These types of explanations are common when plants are found well out of compliance. 13 Online Appendix Table B4 gives an example of an extended regulator-plant interaction. 12

13 and column 7 gives the total number of observations in that round. The stage begins with a regulatory inspection. The plant and the regulatory machine then alternate moves until the machine decides to Accept the plant s compliance. The rapidly descending numbers of observations in column 7 shows that the regulatory machine seldom penalizes plants: 87% of chains end after a single inspection, with the machine accepting in the third round. Over 900 chains continue beyond that stage, and a handful go on for a dozen rounds or more as violating plants are inspected and punished. If the machine does not immediately accept the state of the plant, then it is initially more likely to issue a warning: in round three, 9.5% of actions are warnings against 2.2% that are punishments. Thereafter, the machine is increasingly more likely to punish. Conditional on reaching a given round the probability of punishment rises monotonically with every round from 2.2% in round 3 to 18.1% in round 9 before turning downwards, in late rounds that are seldom observed. Initial plant compliance is low but rises monotonically, with probabilities of 0.4, 7.2, 8.9, 16.5 and 17.5 percent over the second through tenth rounds, before levelling off. Section IV C will model these interactions between plants and the regulatory machine. E Summary statistics and Randomization Check Table 3 presents sample means and a randomization check for the inspection treatment. Each row considers a separate plant characteristic; columns 1 and 2 report the means for treatment and control plants, respectively, using administrative data for the year prior to the experiment. Column 3 reports the coefficient 2 on the inspection treatment dummy T j from the following regression, where for each outcome Y j for plant j in region r, Y jr = r + 1 AuditSample j + 2 T j + j (1) where r are region effects and AuditSample j is a dummy for a plant belonging to the audit sample (i.e., being audit eligible, rather than being assigned to the audit treatment). 14 Standard deviations are in brackets and standard errors for the column 3 coefficient in parentheses. Panel A describes plant characteristics, which are well-balanced by treatment status. Forty- 14 Since only one region, Ahmedabad, contains both audit-eligible and audit-ineligible plants, this specification is equivalent to a full set of region eligible effects. 13

14 five percent of sample plants are in the textiles sector and about fifteen percent are chemical plants (mainly upstream dye or dye intermediate producers). The average plant generates nearly 200,000 liters of waste effluent per day. About half of sample plants have air emissions from a boiler and many are burning dirty fuels such as diesel or lignite in these boilers. Panel B reports recent plant-regulator interactions for the year prior to the study. The regulator cites plants for violating pollution standards often and follows through on those citations with costly penalties. The average plant was inspected about 1.2 times per year from January 2009 to the start of the experiment (July) in both the treatment and control groups. Despite limited contact with the regulator, many plants (18% in the control group) were cited for violations. Moreover, regulatory action did not stop with citations; about 20% of plants were mandated to install abatement equipment and 6% had their utilities disconnected. Across both panels we observe no statistically significant differences between treatment and control plants. III Results: Experimental Estimates This section examines how the inspection treatment affects regulatory actions, plant abatement costs and pollution emissions. A Regulatory Action Each row in Table 4 considers a different regulatory outcome. As before, columns 1 and 2 report the means for control and treatment plants, while column 3 reports the coefficient on the inspection treatment dummy from the estimation of equation (1). Panel A shows the rates of inspection in the treatment and control groups. Control plants were inspected an average of 1.40 times per year over the course of the experiment. Treatment plants were assigned to be inspected 2.12 more times per year and actually inspected an additional 1.71 times per year, more than doubling the annual rate of inspection, to 3.11 times. If we consider only initial inspections, that start a new chain of interactions with the regulator, the treatment increases these by 1.50 times per year. While inspection teams could well have induced crowd-in or crowd-out of regular GPCB inspections, we do not observe net substitution of GPCB s discretionary inspections towards, or 14

15 away from, treatment plants. 15 Panel B reports perceived inspection frequency. Though not officially told about increased inspection frequency, treatment plants recognized the change. When asked to recall how many inspections they received in a given year, both treatment and control plants overstated the number of inspections they got. Treatment plants, though, recalled being inspected a significant 0.71 times more than control plants in 2010, which is correct in sign but understates by 58% the actual difference in inspection rates. A placebo check shows there was no difference in perceived inspections for years prior to the experiment (e.g., the perceived difference in inspections recalled for 2008 was 0.13 with a standard error of 0.10). Figure 1, Panels A and B compare the distributions of annual inspection rates in the control and treatment groups, respectively. In these distributions, we include only the first inspection in a chain, not any follow-ups. The left side of the treatment distribution is shifted to the right; the mass bunched against zero in the control is instead in the range from 2-3 inspections per year. Only 5% of plants in the treatment have less than one annual inspection, a reduction of 34 percentage points from the control. The treatment shifts mass up the distribution, but the skinny right tails of the two distributions still look similar. Plants with this high rate of inspection are typically severe violators and, regardless of the treatment, are apparently in continual contact with the regulator. The additional treatment inspections led to more detected pollution violations and regulatory citations, which threaten action against plants. Table 4, Panel C examines the number of regulatory actions against sample plants. Moving down the table, the regulatory actions are ordered by increasing severity, from pollution readings, citations and warnings through to actions like utility disconnections that have a large cost to plants. Treatment plants are a significant 0.21 share more likely to have a pollution reading collected over the nearly two-year treatment, on a meager 0.38 base in the control. These readings lead directly to more treatment plants being found in violation of a standard (0.22 share) and a greater number of citations (0.20 share) for these violations. Given the treatment duration of about two years, this means about one in ten treatment plants receives an additional citation each year, more than doubling the citation rate 15 We test for crowd-out (the regulator shifting its own routine inspections away from the treatment group), and crowd-in (treatment inspections generating GPCB revisits to treatment plants). Online Appendix, Table B5 shows that these effects are small and statistically insignificant; hence the result in Table 4 that the net inspections were nearly equal to inspections assigned by the treatment. 15

16 in the control. Treatment plants see a statistically significant increase of 0.08 closure warnings, which formally threaten to close the plant unless remedial action is taken. Moving to more severe regulatory actions, with direct costs to plants, however, there is no significant evidence of a greater incidence of penalties in the treatment group. For example, closure directions are only insignificantly higher in treatment (coefficient 0.041, standard error 0.033). Other costly actions, such as the mandated installation of equipment or disconnection of utilities, are higher in the treatment but by very small and statistically insignificant amounts. The tapering off of treatment effects for more stringent punishments is not a mechanical effect, due to losing power to detect outcomes that are less frequent: closure warnings increase from 9.4% to 17% with the treatment, but closure directions increase only from 16% to 20%, for example. B Plant Abatement Costs, Pollution Emissions, and Regulatory Compliance This subsection examines whether the treatment increased abatement expenditures, reduced pollution and improved compliance. Table 5 presents estimates of treatment effects on abatement costs. Using descriptions of abatement expenditures in the endline survey, we separate abatement costs for capital and maintenance. 16 Abatement capital expenditures are observable by the regulator, and often mandated in response to a high pollution reading. Maintenance expenditures are not directly observable, but proxy for greater use of abatement equipment, and may thus be associated with lower pollution. More than half of control plants (0.567, column 1) install capital equipment, at an amortized cost of about US$ 2000 per year (column 2). However, the inspection treatment does not increase the probability that plants install capital equipment (column 1) or increase spending on that equipment (column 2). Maintenance expenditures are in some ways the mirror image of capital. Only 11% of plants report maintenance expenditures, with no significant treatmentinduced increase (column 3). We do, however, find that treatment plants increase the amount of maintenance expenditure by US$ 838 (standard error US$ 499, p-value < 0.10). This effect is large relative to the control level of maintenance expenditures, US$ 264 per year, but still small 16 We distinguish maintenance from capital costs by searching descriptions of expenditures for strings associated with maintenance, like maintain or change. See Appendix A for details. Capital costs are amortized into an annual flow of expenditures for comparison to maintenance costs. 16

17 in economic terms, given that sample plants spend on average US$ 12 thousand a month on electricity alone. Table 6 reports the results from regressions of pollution levels (column 1) and compliance (column 2) on treatment assignments for the inspection treatment, audit treatment and their interaction. Pollution is measured in standard deviations for each pollutant at the plant-bypollutant level and standard errors are clustered at the plant level. All specifications include region fixed effects and an indicator for whether a plant is audit-eligible. 17 The treatment reduced plant pollution emissions (column 1) by standard deviations (standard error standard deviations), but this is not statistically significant at conventional levels. This effect is about half the size of the statistically significant standard deviation reduction in pollution due to the audit treatment. 18 Column 2 repeats the column 1 specification, with the outcome variable changed to compliance, an indicator for whether a plant-by-pollutant reading is below the pollutant-specific standard. The inspection treatment marginally increased compliance with pollution standards: treatment plants are 3.7 percentage points (standard error 2.1 percentage points, p-value =0.087) more likely to comply, on a base of 61% compliant pollution readings in the control. (Multiple pollutants are observed in the survey and only 10% of plants are compliant on all pollution readings measured.) The audit treatment had a positive but smaller and statistically insignificant effect on compliance. 19 Compliance can increase without a large reduction in average pollution if plants near the standard were the most likely to respond to the inspection treatment. Figure 2, Panel A plots the coefficients on inspection treatment from regressions of indicators for a pollutant reading being in a given bin, relative to the regulatory standard, on treatment assignments (as in Table 6, column 17 Since only Ahmedabad includes firms that are both audit-eligible and ineligible, this specification is equivalent to using region-by-eligibility fixed effects. 18 The audit treatment effect on pollution reported in Duflo et al. (2013) was estimated in the inspection control group only and slightly larger. The audit-by-inspection interaction is large and positive, offsetting the reductions in pollution from the main effects of audits and inspections. The magnitude of the interaction is puzzling. One explanation is that audits fail to produce any additional information beyond what is uncovered in an inspection; indeed, the null that the sum of the interaction and audit coefficients is zero cannot be rejected (p-value =0.41). 19 To test the robustness of the inspection treatment effect on compliance, Online Appendix Table B7 reports placebo checks where compliance is coded to occur at various multiples of the real standard (2p, 5p, 10p). The effect of inspection treatment on compliance is only statistically significant at the true standard. Online Appendix Table B8 tests whether the letter treatment, of randomly sending letters reminding firms of their compliance obligations, had any effect on emissions and compliance. We cannot reject that the letter treatment had no effect on either treatment or compliance. 17

18 2, but with finer bins rather than a single dummy for compliance). Note that these endline survey pollution readings were collected by independent teams without knowledge of treatment status. The inspection treatment is estimated to reduce pollution readings just above the standard, in the range of (0.0,0.2] standard deviations, more than in any other bin, though this decrease is not statistically significant (p-value =0.17). However, it does significantly increase the number of readings just below the standard in [-0.2,0.0]. It appears that the treatment and its increased regulatory scrutiny shifted some plants that were modestly out of compliance with the de jure standard into compliance. 20 C Status Quo Targeting of Inspections Our structural analysis will argue that regulatory targeting is important to understand the results of the experiment. To support this argument, before laying out the model, we present three pieces of reduced-form evidence that inspection targeting does occur in the status quo. First, the regulator concentrates inspections on a relatively small share of plants. Second, treatment inspections found more polluting plants, relative to control group inspections, but no more highly polluting plants. Third, private information about plant pollution predicts later regulatory inspections. On the first point, recall that the distribution of inspection rates in the control is extremely skewed (Figure 1, Panel A). About 40% of plants are inspected less than once per year, though a few plants are visited nine times a year. The dispersion in inspection rates is striking, given that all sample plants belong, by design, to the highest pollution potential category. On the second point, Figure 3 shows the number of plants in the treatment and control groups with pollution readings that exceed the relevant pollutant-specific standard by varying factors (during the first year of the experiment, in order to minimize changes due to the treatment and focus on regulatory selection of plants). In the treatment group, the regulator takes many more pollution readings and finds roughly twice as many plants with pollution readings above the 20 The finding that inspections change compliance, albeit modestly, but not average pollution, is the reverse of our earlier finding that more truthful environmental audits reduce pollution from highly polluting plants only, with no effect on compliance (Duflo et al., 2013). Audits may plausibly differ from inspections because they focus on a set of audit-eligible firms with especially high pollution potential, and because an audit is made of three separate and detailed inspections during the year, offering very rich information. The inspection treatment is broader, but may provide information that is more closely substitutable with that from existing inspections, for plants that are already visited often. 18