Can Environmental Policy Reduce Infant Mortality? Evidence from the Ganga Pollution Cases

In many developing countries, environmental quality remains low and policies to improve it have been inconsistently effective. This paper conducts a case study of environmental policy in India, focusing on unprecedented Supreme Court rulings that targeted industrial pollution in the Ganga River. In a difference-in-differences framework, the rulings are found to have precipitated reductions in river pollution and one-month infant mortality, both of which persist for more than a decade. The authors then estimate a pollution-mortality dose-response function across twenty-nine rivers in the Ganga Basin, instrumenting for pollution with its upstream counterpart. The estimation reveals a significant external health burden of river pollution, not just in the district of measurement, but also on downstream communities. It further provides suggestive evidence that reducing pollution was an important driver behind declines in infant mortality observed after the rulings.

In many developing countries, environmental quality remains low and policies to improve it have been inconsistently effective. This paper conducts a case study of environmental policy in India, focusing on unprecedented Supreme Court rulings that targeted industrial pollution in the Ganga River. In a difference-in-differences framework, the rulings are found to have precipitated reductions in river pollution and one-month infant mortality, both of which persist for more than a decade. The authors then estimate a pollution-mortality dose-response function across twenty-nine rivers in the Ganga Basin, instrumenting for pollution with its upstream counterpart. The estimation reveals a significant external health burden of river pollution, not just in the district of measurement, but also on downstream communities. It further provides suggestive evidence that reducing pollution was an important driver behind declines in infant mortality observed after the rulings.

Introduction
Environmental economics and development economics are unified by the persistent puzzle of poor air and water quality in developing countries (Greenstone and Jack 2015). A growing literature provides evidence that pollution imposes a significant health burden (e.g. Jayachandran 2008, Ebenstein 2012, andBrainerd andMenon 2014), yet corresponding estimates of willingness-to-pay for environmental quality are surprisingly low (Kremer et al. 2011) and policies aimed at improving environmental quality have not reliably done so (Greenstone and Hanna 2014;Field, Glennerster, and Hussam 2011). While high levels of air and water pollution may be due in part to high marginal utility of consumption  and high marginal costs of pollution abatement (Davis 2008), they are also likely driven by rent-seeking behavior (Duflo et al. 2013 We investigate a landmark decision in India's environmental regulatory history: the Supreme Court case M.C. Mehta vs. Union of India, subsequently bifurcated and known as the "Ganga Pollution Cases." These cases represent India's first-ever environmental public-interest litigation, and their unprecedented rulingswhich mandated pollution cleanup by the tanning industry concentrated along the Ganga River in Kanpur, Uttar Pradeshmarked the rise of environmental activism by the Indian judiciary (Mehta 2009). Exploiting the quasi-random incidence of the litigation, we find that the ruling is strongly associated with reductions in both river pollution (as measured by biochemical oxygen demand [BOD]) and neonatal (one-month) mortality.
Our positive finding is an important data point because, in developing countries, there is very little evidence of environmental policies other than piped water provision being successful (Ravallion and Jalan 2003; Gamper-Rabindran, Khan, and Timmins 2010). Our reduced-form evidence highlights a rare instance of successful policy: judicial-led mandates primarily targeting industrial pollution. Our results suggest that the rulings improved several measures of water quality -BOD as well as calcium, sulfur, and chloride concentrations -and that mortality impacts occur primarily in the first month of life. Furthermore, the rulings' impacts on pollution and health persist over time for at least ten years as well as downstream into neighboring districts.
In the second phase of our analysis, we quantify the external infant health costs of river pollution in general and investigate the mechanisms of the Ganga Pollution Cases' impacts. To do so, we estimate a pollution-health dose-response function in the Ganga River Basin. Since pollution is potentially endogenous, we instrument for it using its upstream counterpart. This identification strategy has a strong logic: the decision to pollute upstream is orthogonal to downstream health inputs, but pollution reliably flows downstream nonetheless. The persistence of downstream impacts of the Supreme Court rulings, as well as recent research leveraging the upstream-downstream geographic relationship (Garg et al. 2017), provide further support for our instrumental variables (IV) strategy.
We first find that exceedance of India's water quality standards raises the risk of neonatal mortality by 10-14 percentage points, on average, in the IV framework; the analogous ordinary least squares (OLS) regression yields a statistical zero. Since our IV point estimates are identified by variation in upstream water quality, they also show that the mortality burden of river pollution persists in downstream districts. Reduced-form regression of downstream neonatal mortality on upstream water quality indicate that the downstream mortality effect is approximately twenty percent of the main, in-district effect. Prior research documents the flow of pollution into downstream water sources (e.g., Lipscomb and Mobarak 2017), but we believe we are the first to connect such spillovers to economicallymeaningful health impacts. Our work thus underscores the large social costs of continually poor environmental quality, both at the point of measurement and in nearby areas linked by surface water flow.
Next, we leverage the IV strategy to investigate the mechanisms at work. In general, pollution policy may affect health directly through improved environmental quality, but it may also do so through increased awareness and avoidance (Kremer et al. 2011; Graff-Zivin, Neidell, and Schlenker 2011). Reduced-form estimates of policy impacts aggregate the effects of all such channels. Our finding of simultaneous reductions in river pollution and neonatal mortality in the Kanpur region is thus a necessary, but not sufficient, condition for pollution being the primary channel of the verdicts' health impact.
Intuition suggests that if reduced water contamination is the only channel of mortality reductions, then two statements should hold empirically: there is no residual effect of policy on mortality after appropriately controlling for pollution; and the policy is a valid instrument for pollution. Testing the degree to which these statements hold is equivalent to testing the null hypothesis that the policy's mortality impacts are fully explained by its pollution impacts. Empirically, we formulate this as an overidentification test with two instruments for pollution: upstream pollution and the policy itself. This strategy is representative of a more general method of understanding policy impacts in which outcomes of a policy change are compared to the outcomes of orthogonal changes in structural parameters (e.g.,

Angelucci and Attanasio 2013).
Armed with a plausibly exogenous source of pollution variation, we confirm that the two aforementioned statements hold empirically. First, the policy variable is no longer a significant predictor of mortality once pollution is included alongside it as an explanatory variable. Second, our overidentification tests fail to reject the null hypothesis that the rulings are a valid instrument for river pollutioni.e., that water quality improvements are the only channel through which the verdicts affected neonatal mortality.
The remainder of this paper is organized as follows. Section 1 describes the Ganga Pollution Cases amidst the more general context of pollution and related policy in India. Section 2 describes the various sources of data, while sections 3 and 4 present our empirical strategy and results, respectively. Section 5 places our contribution within the environmental policy literature, especially with regard to the puzzle of persistently poor environmental quality in the face of large social costs of pollution.

The Ganga Pollution Cases
In the aftermath of several decades of population and industrial growth, India's rivers are heavily pollutedparticularly in urban areas (Murty and Kumar 2011). Monitoring of biochemical oxygen demand (BOD) -a broad-based measure of organic water pollution-carried out by India's Central Pollution Control Board (CPCB) shows that, as of 2011, average readings at approximately 37 percent of all sampling stations did not meet the government's standard of acceptability for bathing set at 3 mg/l (CPCB 2017b). Waste from domestic consumption, industry, and agriculture all contribute to widespread water quality problems in surface and groundwater resources (Murty and Kumar 2011).
There have been efforts to improve Indian water quality, but there is little evidence that they have been successful. The most salient government effort to reduce river pollution is the National River Conservation Plan (NRCP), a federal top-down program targeting domestic pollution into India's surface waters. NRCP began in 1985 as the Ganga Action Plan but has expanded over 30 years to now cover 190 towns along 41 rivers across India. Its goal since 1987 has been to restore the Ganga River to the standard for outdoor bathing, as defined by India's "Designated Best Use" classification system. 1 The primary lever for achieving this goal has been the "interception, diversion, and treatment" of sewage (Madhya Pradesh Pollution Control Board 2017). To that end, 4,704 million-liters per day of sewage treatment capacity have been created since its inception (Ministry of Environment and Forests 2013). Nonetheless, popular media and non-governmental organizations have panned NRCP for reasons such as poor inter-agency cooperation, funding imbalances across sites, and an inability to keep pace with growing sewage loads (Suresh et al. 2007). Confirming public belief, Greenstone and Hanna (2014) find no discernible impact of NRCP on water quality levels.
The executive branch, however, is not the only source of environmental regulation in India; the Indian judiciary has, through the years, developed a reputation for environmental activism (Singh 2014). Article 21 of the Indian Constitution provides citizens with the "Right to Life", and much jurisprudence in recent years has centered on the protection of this constitutional right. This paper examines the first instance of Supreme Court involvement in issues of river pollution in India.
The story begins in the pilgrimage city of Haridwar along the Ganga River. The Ganga is India's longest river; in the Hindu religion, it is revered as a goddess. Upwards of 100,000 industrial operations (Mehta 2009) and 330 million people (CPCB 2009) reside in its basin (see Figure 1). In 1984, a matchstick tossed into the river by a smoker in Haridwar resulted in the river catching on fire for more than 30 hours, due to a toxic layer of chemicals produced by a pharmaceutical firm (Mehta 2009). In response to this event, environmental lawyer and social activist M.C. Mehta filed a writ petition in the Supreme Court of India charging that government authorities had not taken effective steps to prevent environmental pollution in the Ganga's waters. The scale of the case -the whole 2,500-km stretch of the river -proved to be intractable. The court requested that Mr. Mehta narrow his focus; he chose the city of Kanpur (Supreme Court of India 1985).
Kanpur is a city of 2.9 million people lying directly on the Ganga in the State of Uttar Pradesh).
For more than 100 years, it has been a major center for India's tannery industry. Of the 400 tanneries currently located in Kanpur (The Hindu 2016), most are concentrated in the neighborhood of Jajmau, which lay directly on the southern bank of the Ganga River. Leather processing is a highly polluting industry; the procedures for washing, liming, fleshing, tanning, splitting, and finishing involve a large number of chemicals (Cheremisinoff 2001). Tannery effluent is characterized by large amounts of organic material -which, when deposited in rivers, depletes dissolved oxygen levels and thus the overall health of the watershed -and heavy metals like chromium, which is a documented carcinogen (Stout et al. 2009). It is also known for its conspicuous reddish-brown color (Durai and Rajasimman 2011).
Mehta selected Kanpur despite not having been born in or lived in Kanpur. In interviews granted to our research team in June 2014, he explained that " [Kanpur] was in the middle of the Ganga Basin, the reddish color of the pollution made the pollution highly salient, and the city seemed representative of many other cities in the Ganga Basin." The court subsequently split the petition into two parts.
The first dealt with the tanneries of Kanpur and the second with the city government. These are now respectively called Mehta I and Mehta II in legislative digests, and are together known as the "Ganga Pollution Cases"the foundational water pollution litigation in the Indian court system. In October 1987, the Court invoked the Water Act and Environment (Protection) Act as well as Article 21 of the Indian Constitution to rule in Mr. Mehta's favor and order the tanneries of Jajmau to clean their wastewater within six months or shut down entirely. This was followed by a January 1988 judgment that required the Kanpur local municipal bodies to take several immediate measures to control water pollution: the relocation of 80,000 cattle housed in dairies or the safe removal of animal waste from these locations; the cleaning of the city's sewers; the building of larger sewer systems; the construction of public latrines; and an immediate ban on the disposal of corpses into the river. The court also required all schools to devote one hour each week to environmental education and awareness.
Of the 87 tanneries named in Mr. Mehta's petition, approximately 20 were shut down and at least 60 established primary treatment plants (PTPs). Moreover, several initiatives were undertaken in 1987 and 1988 to clean drains, expand the number of handpumps, and build latrines to improve sanitation systems in Jajmau (Alley 2002). Subsequent litigation in the Supreme Court over the past 25 years, and indeed many academic researchers of pollution in Kanpur, have argued that these projects were a failure and that newly established technologies were not appropriately maintained or used (Alley 2002;Singh 2007).

Data
To assess the relationship between policy, water quality, and health in Kanpur, we collect and combine three types of data: infant mortality, river pollution, and other variables to be used as controls. In most of our analysis, we restrict our sample geographic area to the Ganga Basin, which is defined as the area that drains the Ganga and all of its tributaries. We make this restriction because of the singularity of this area in the context of our analysis. The Ganga Basin is not only a much more densely populated region than anywhere else in India, but also a region in which water issues have received special government attention: the National River Conservation Plan (NRCP) focused exclusively on the Ganga Basin (including the Ganga, Yamuna, Damodar, Gomti, and Mahananda Rivers) throughout its first 10 years (1986)(1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995). Extending the analysis beyond this region might confound the effect of the Supreme Court ruling with the effect of broad scrutiny in the Ganga Basin during this time period.

Health data
We choose infant mortality as our primary health outcome. This choice conveys at least two significant statistical advantages. First, as noted by Chay and Greenstone (2003), infant survival is a shortterm measure; other measures of health stock, captured later in life, necessarily reflect the effect of many more health shocks experienced in an individual's lifetime. Second, complete birth histories are available in certain Indian demographic surveys, so we can construct long pseudo-panels of infant survival to maximize statistical power. In contrast, morbidity variables such as diarrhea incidence and low birth weight are only available cross-sectionally from the time of survey; too few of such surveys have been completed for us to construct a meaningful panel. Panel variation in infant mortality allows us to include detailed temporal and cross-sectional fixed effects in regression analyses, which account for time-varying shocks to national infant health and time-invariant characteristics of specific areas, respectively.
We focus primarily on deaths within the first month of life, i.e., neonatal mortality, although we also test for a policy impact on one-year mortality. The first month is the most vulnerable period of

Pollution data
To measure water quality, we use river pollution data collected under the auspices of India's National Water Quality Monitoring Program (NWMP). These data were originally gathered by Greenstone and Use of a broad-based indicator of pollution raises the issue that it is likely affected by a variety of factors, not just the activities of tanneries. During the Supreme Court hearings, lawyers representing the tanneries argued that other firms, a negligent Kanpur municipality, and other negligent municipalities were contributing to the pollution around the tanneries (Mehta 2009: 78-82). This argument, however, was eventually dismissed in light of a report from the CPCB which demonstrated that the specific pollutant readings in the river were not even theoretically capable of being generated by any entity other than the cluster of tanneries (Mehta 2009: 83-84). Field reports of Jajmau report no other major industry in the area (Alley 2002).
In most of our analysis, we parameterize BOD as a dummy variable equaling one if average BOD in a district-month exceeds the national standard of 3 mg/l for bathing in surface water. 2 This 3-mg/l threshold is also a stated goal of the National River Conservation Plan, and is consistent with the crit-ical values used by regulatory agencies in other countries (APHA 1992, ISO 1989. It is also consistent with the standards of proper measurement. According to the International Standards Organization (2003), the limit of detection for BOD is 3 mg/l; that is, BOD readings below three cannot be statistically distinguished from zero. These and other water quality standards in general highlight the nonlinear nature of the pollution-health relationship that has been documented in previous economics research on pollution and health (Chay andGreenstone 2003, Ebenstein 2012). For comparison, we also use the natural logarithm of BOD in certain analyses.
We also consider four other pollutants that shed additional light on the impacts of the Kanpur Supreme Court verdicts: calcium, sulfates, chlorides, and fecal coliforms (FCOLI). Calcium is the key component of lime, which is a standard ingredient used in the removal of hair and flesh and the splitting of the hide into its two primary layers. Sulfate and chloride ions, meanwhile, are the main components of the total dissolved solids (TDS) produced in tanning. 3 All three of these pollutants are measured in milligrams per liter (mg/l). Fecal coliforms (FCOLI) are an oft-used measure of domestic (as opposed to industrial) pollution, which was a major focus of the second verdict in the Ganga Pollution Cases. It is measured as the "most probable number" of coliform organisms per 100 milliliters of water (MPN/100 ml, reported in thousands). Other theoretically-relevant pollutants in our context are not recorded consistently in our time period. 4

Assignment of upstream pollution
Part of our analytical strategy relies on the measurement of pollution upstream of a given location.
Such measurement requires information on the precise location of pollution monitors. Unfortunately, latitude and longitude of monitors are incomplete and unreliable in our dataset. To circumvent this problem, we manually map each monitor according to the administrative descriptors provided (state, town, river) and an accompanying string description of location (e.g., "Sabarmati at Ahmedabad at V.N. Bridge"). With our monitors mapped, we trace the path of all rivers in our sample, from origin to last monitor downstream, and measure distances along the river between all pairs of neighboring monitors.
How far upstream should upstream pollution be measured? Two issues arise when answering this question. First, many water quality monitors in our dataset have more than one possible upstream counterpart. Second, there is no single "correct" distance at which to measure upstream pollution. 3 Chloride ions are highly soluble in water, so their concentration is insensitive to standard tannery waste treatment (UNIDO 2011). However, surface-water chloride levels may respond to other policy impacts on the tanning industry, such as reduction or cessation of tannery operations. 4 Total suspended solids (TSS) and total dissolved solids (TDS) are potential alternatives to BOD in our analysis, but the first of these is not recorded in large numbers in our data, and the second is not recorded in Kanpur prior to the verdicts. Chromium, perhaps the highest-profile pollutant in the tanning process, was not widely measured by the CPCB as of 2004.
The larger this distance, the weaker the upstream instrument will be as a result of pollution decay.
On the other hand, monitors that are closer together are more likely to be subject to common shocks such as rainfall, which would create spurious correlation.
We thus adopt a variety of definitions of "upstream", and check the robustness of our results to the definition used. To assign an upstream counterpart to a given pollution monitor, we use the following protocol: first, we follow the river upstream until it reaches a new district; then, we locate the nearest monitor along the river that falls within a distance range (in km) of [X, Y ] from the original monitor, where X ∈ {0, 20, 50, 75, 100} and Y ∈ {200, 300}. When a river splits upstream of a given monitor, so that there is an upstream monitor on each of two tributaries, we take the unweighted average pollution reading of these monitors as our upstream measure. 5 When there is no upstream monitor of any kind to be found, we use the river's origin as an upstream location (subject to the distance-range requirement) and assign the sample-wide minimum value of pollution as our upstream measure.

Other data
We include several types of variables as controls in many of our regression analyses. The main body of these controls consists of cross-sectional birth, mother, and child characteristics taken from RCH-2.
A pair of climate controls are created using monthly, gridded rainfall totals from the University of Delaware and air temperature averages from the Indian Meteorological Institute; we use these gridded averages to interpolate monthly rainfall and temperature values at each district (in mortality-only regressions) or monitor (in regressions involving pollution). Finally, we add time-varying measures of common effluent treatment plant (CETP) capacity and the incidence of the NRCP, which capture the non-uniform intensity of environmental cleanup and policy efforts within the Ganga Basin. 6

Empirical Strategy
Our aim is to determine whether or not the 1987 Supreme Court decision affected environmental quality or health outcomes. Our analysis has two broad parts. In the first part, we use difference-in-differences (DD)comparing outcomes in the affected area to those elsewhere in the Ganga Basin, before and after the verdictto estimate the impacts of Mehta vs. Union of India on infant mortality and river pollution. In the second, we employ instrumental variables (IV) to identify the direct relationship between river pollution and neonatal mortality. The IV strategy serves two purposes: first, it quantifies an important component of the external damages of river pollution; and second, it sheds some light on the mechanisms through which environmental regulation affects health.

Difference-in-differences
We specify the reduced-form impact of the Supreme Court verdict on neonatal mortality as follows: where M ortality idt is a dummy variable indicating whether a child i, born in district d, in year-month t, died within the first month of life. T dt captures the incidence of policythe Mehta vs. Union of India court decisions in our caseand takes a value of one in affected -or "treated"-districts after October 1987 and zero otherwise. X idt is a vector of individual, location-by-time characteristics, which includes district and year-month fixed effects. The verdict's impact on river pollution is specified analogously: where P ollution dt is a dummy for BOD greater than 3 mg/l in district d and year-month t.
From the pool of all Ganga Basin districts, we define our "treatment group" to comprise four districts: Kanpur Nagar, Unnao, Fatehpur, and Rae Bareli. This definition is justified by two facts.
First, these are the four districts whose water quality is most likely to be affected by the verdict.
Kanpur Nagar is the physical locus of the verdict; Unnao shares the same stretch of the Ganga River with Kanpur Nagar; and Fatehpur and Rae Bareli are the first districts downstream of these two, with borders approximately 28 and 44 kilometers from the Jajmau tannery cluster, respectively (in fact, the Jajmau tanneries are very near to Kanpur Nagar district's downstream border, so that the majority of Kanpur's population is not directly exposed to tannery pollution). Second, Unnao city contains its own cluster of tanneries, so inclusion of this district in the treatment group yields impacts that are net of any tanning output "leakage". Nonetheless, we also test the robustness of our results to the exclusion of one or more non-Kanpur districts from the treatment group. 7 The crux of our identification strategy is the assumption that Cov(T dt , idt ) = 0 -i.e., that the policy variable is uncorrelated with all unobserved predictors of neonatal mortality (and of pollution).
We argue that this is a plausible assumption because environmental public interest litigation had no To visually inspect the credibility of our identification assumption, we plot annual averages of both neonatal mortality and exceedance of the BOD threshold in the treatment and control groups over time; the annual averages smooth out otherwise noisy monthly time series. We then fit equations (1) and (2) to produce point estimates of policy-induced mean-shifts in these outcomes. In all cases, mortality regressions are run with child-month observations and pollution regressions are run with district-month observations.

Instrumental variables
Estimation of equation (1) yields the reduced-form impact of policy on neonatal mortality, but it provides no information on the mechanisms of that impact. Theoretically, the Supreme Court decision could have had an impact on mortality through three channels: (i) reduced water contamination, (ii) reduced exposure to contaminated water, and (iii) increased ability to cope with the consequences of pollution. In our context, it is likely that, in addition to reducing water contamination, the Supreme Court verdicts also drew attention to the river pollution problem; this could have driven households to avoid exposure through water-source switching or filtration, and to seek information on how to treat the In principle, knowledge of river pollution's direct impact on mortality can improve policy design.
First, in quantifying the health burden imposed by pollution, we gain a sense of the potential for health improvements through policies that reduce pollution. Second, we can use this "dose-response function" to translate estimated policy impacts on pollution (from equation 2) into a prediction of policy impacts on mortality exclusively through the reduced-contamination channel. If that prediction differs significantly from our reduced-form estimates of policy impacts on mortality (from equation 1), it is evidence that the policy operates at least in part through other channels. Note, moreover, that the reduced-contamination channel captures both the direct impact of the Supreme Court decision and any follow-on contamination-reducing policies communities could have undertaken in the aftermath of the verdicts.
Our primary obstacle in estimating a pollution-mortality dose-response function is the endogeneity of pollution. In general, pollution is correlated with other factors affecting mortality, such as urbanization (which brings access to health care facilities and education) and economic productivity (which raises incomes). To solve this problem, we instrument for pollution in a given location-time with its upstream counterpart, as defined in Section 2. The logic of this instrument leverages the unidirectional flow of rivers: upstream pollution reliably flows downstream (subject to some decay function of distance) to impart a negative health impact, and yet the determinants of this upstream pollution are plausibly orthogonal to downstream determinants of health. Our implicit assumption is that river flow is distinct from movements in economic, political and demographic variables, which likely take longer to diffuse from upstream to downstream.
We start by estimating the reduced-form analog of our IV strategy: where P ollution u dt is the pollution reading upstream of district d in year t. The primary parameter of interest in equation (3) is γ 3 , the impact of upstream pollution on downstream mortality. The point estimate of γ 3 gives us a first check on whether our IV strategy will successfully identify the dose-response function; it also provides a causal estimate of the downstream external damages created by pollution, hence extending the work of Lipscomb and Mobarak (2017) and Sigman (2002Sigman ( , 2005 on the presence of pollution spillovers. Moreover, we can test the credibility of our instrument with a falsification test: for every district in our pollution sample, we replace our first-choice "upstream district" with the nearest off-river district available in our data and re-estimate equation (3). Statistical insignificance of γ 3 in the falsification check is a necessary condition of instrument validity.
We then estimate our full two stage least squares (2SLS) model using, as a first stage, and, as a second stage, Coefficient γ 3 in equation (5)  Equations (4) and (5) also provide the basis for testing the mechanisms at work. We compare the estimated mortality burden of pollution with a single instrument (upstream pollution) to the estimated burden with two instruments (adding the Supreme Court verdict). If the policy affects health through non-pollution channels, then an overidentification test should reject the null hypothesis that the policy instrument is a valid instrument for pollution. This null will also be rejected if the effect of pollution on mortality is heterogeneous across (e.g.) locations or pollutants, so that each of the two instruments measures a different local average treatment effect. Failure to reject the null, on the other hand, provides suggestive evidence that pollution is the primary channel of policy-induced reductions in neonatal mortality. The overidentification test uses a C-statistic (see, e.g., Eichenbaum, Hansen, and Singleton 1988), also known as a difference-in-Sargan test statistic. It is equal to the difference of the two Sargan-Hansen J-statistics obtained from the regression using both T dt and P ollution u dt as instruments on the one hand and the regression using only P ollution u dt on the other hand. 8

Results
We report summary statistics for each analysis variable in in the treatment group than in the control group (6 percent and 9 percent, respectively). These Ganga Basin rates are, in turn, larger than the analogous national rates of 5.3 percent and 7.6 percent in our data, respectively. 9 Meanwhile, average log-BOD is higher in the treatment group than in the control group (1.07 vs. 0.94), but the relationship flips when we summarize the frequency of exceeding the national standard of BOD<3 mg/l for "bathing class" water (0.34 in the treatment group vs. 0.39 in the control group). Moreover, there is no consistent relationship between treatment and control across all pollutants considered: high sulfate and chloride levels are more frequently observed in the Kanpur region, but high calcium levels are less frequently observed there, and high FCOLI levels are about equally likely to occur in each group. While Table 1 reveals several cross-sectional differences between the treatment and control, the inclusion of district fixed-effects controls for these differences.

Difference-in-differences: a graphical representation
We begin our analysis by graphically inspecting our difference-in-differences (DD) assumption of parallel trends in the treatment and control group. The results are depicted in Figures 3 and 4. In both figures, data points are annual averages of monthly observations in either the treatment group or the 8 In Appendix B, we formally derive this test from a sparse model of pollution, policy, and health. 9 According to the World Development Indicators (2017), the national infant mortality rate was 14.6 in 1971 and 5.78 in 2004; our measured sample-wide rate of 7.6 is closer to the latter number because births recorded in DLHS-2 are predominantly from more recent years. control group. Pre-ruling summary statistics for treatment and control are provided in Appendix Table   A1.  Using the entirety of the Ganga Basin in our control group allows us to smooth our counterfactual significantly. 11

The impact of Mehta vs. Union of India on infant health
To obtain a point estimate of the Supreme Court verdicts' impact on infant health, we estimate equation (1). Table 2 provides our first set of results from this estimation. In columns 1-3, we use the full 1971-2004 period, while in columns 4-6, we use the shorter 1980-1994 period. Within each of these triplets, the first column uses one-year (infant) mortality as the dependent variable, the second uses one-month (neonatal) mortality, and the third uses one-year mortality conditional on survival past the first month. In all columns, we multiply RCH-2 sampling weights by the inverse size of the relevant age cohort and use these as weights -for example, all observations from children at an age of one month are assigned a weight equal to (1/# of observations with age = one month). This weighting makes possible the translation of our point estimates, which come from monthly observations, into aggregate changes in mortality risk across the whole full first year of an infant's life. 12 The remainder of the columns in Table 2  We explore the timing and geography of the verdicts' impacts on neonatal health in Table 3, using the same child-month mortality panel as in Table 2. Columns 1-3 show results from neonatal mortality regressions using post-periods of different lengths, while column 4 shows results from simultaneous estimation of short-, medium-, and long-run policy impacts. Column 5 tests the robustness of the main effect to the use of all of India as a control group. All estimates are statistically significant at the five-percent level or lower. Magnitudes range from 2 to 3.5 percentage points in these columns, which imply a 16-29 percent drop in neonatal mortality rates as a result of the rulings. Pairwise F-tests reject the null hypothesis that the 1995-1999 (medium-run) coefficient is equal to either of the other two period-specific coefficients (p-value<0.01).
In column 6, we investigate the extent to which the mortality reduction is driven by reduced in utero exposure to pollution. To do so, we categorize treatment-group births by their degree of in utero exposure. All children born after the Supreme Court verdicts experience lower post-natal exposure.
However, children born in the first nine months after the Supreme Court verdicts (November 1987 to July 1988) experience only partial in utero reduction in exposure to contaminated water, while children born after July 1988 fully benefit from the effect of the verdicts. We define dummies for these two groups of children in the treatment region and estimate impacts for each group simultaneously.
The predictive effect of fully benefiting from the verdicts while in utero is a statistically-significant 2.9 percentage-point drop in neonatal mortality risk. Meanwhile, for children only partially "treated" by the Supreme Court rulings while in utero, we find a positive though statistically insignificant effect on mortality risk. This result is consistent with much existing evidence that links mother's health to the risks of neonatal mortality (Gluckman et al. 2008; Currie and Almond 2011). We note, however, that the difference between being partially and fully affected by the policy in utero and other temporal factors cannot be disentangled. That is, other changes in the treatment region after July 1988 could also explain the result -for instance, relatively weaker policy effectiveness in the initial year following the verdicts. Meanwhile, the lack of impacts beyond Allahabad is not surprising, because pollution decay and other inputs to water quality should attenuate the effect the rulings as we move our focus downstream.

The impact of Mehta vs. Union of India on river pollution
The Ganga Pollution Cases therefore seem to have had a beneficial effect on infant health. Did they do so by reducing river pollution? One necessary condition for this explanation to hold is that the verdict is associated with reduced pollution levels. We test this condition in our reduced-form DD framework, estimating equation 2. Panel A of Table 4 Table 3, column   7: the nearest group of districts downstream of the treatment group is associated with a large and significant average drop in exceedance of the water quality standard, and the point estimate monotonically drops in the next two groups downstream. The evidence thus implies that the rulings' impacts on both water quality and neonatal mortality persist some distance downstream of the cases' focus.
In columns 7 and 8, we use the logarithm of the BOD level as our outcome variable instead of the bathing class dummy. With this specification of BOD, and over the full time period (column 7), the rulings are associated with a marginally significant drop in log-BOD, but the association is smaller and insignificant in the short run. To investigate the divergent results between a dummy variable and the continuous logarithm form, we graphically examine the distribution of the logarithm of BOD by time and treatment status.

Mechanisms of policy transmission
Having identified a strong link between the Ganga Pollution Cases and the local health of both infants (Tables 2 and 3) and rivers (Table 4), we now turn to the questions of how water pollution affects infant health and what that means for Indian environmental policy. Our strategy for answering these questions relies on an estimation of a dose-response function for water pollution and neonatal mortality.
We use the merged sample of water pollution and neonatal mortality in this phase of analysis, which limits us to the 1986-2004 time period. To utilize our IV, we must further restrict our geographic focus to only those districts with non-missing upstream pollution measurements. Appendix Table A2 shows how the pool of available pollution monitors and their distance from upstream counterparts varies with our definition of upstream (which is detailed in Section 2). Appendix Table A3 summarizes the merged sample using our preferred upstream definition of [75 km, 200 km]. 13 Finally, Appendix Table  A4 reproduces the main results of our DiD analysis with the merged sample. Table 5 establishes a causal relationship between pollution and infant mortality. Column 1 shows the OLS relationship between BOD pollution (i.e., exceedance of the bathing-class standard) and neonatal mortality. It is a statistical zero, which may reflect the upward bias that would result from positive correlations between pollution, urbanization, and economic activity. Columns 2-5 replace the key independent variable with its upstream counterpart -the reduced-form analog of 2SLS with upstream BOD as the instrument. The specifications in these columns vary only by either sample time period or the definition of upstream. In stark contrast to the OLS result, the coefficient on upstream water quality is positive and significant across all four specifications. The interpretation of these coefficients is that exceeding the BOD standard for bathing-class water quality upstream increases the downstream likelihood of neonatal mortality by 2-3 percentage points -a significant spatial externality.
Our empirical strategy hinges on the assumption that upstream river pollution is exogenous once we condition on weather, policy, and district and year-month fixed effects. Implicit in this assumption is that these controls fully capture the endogenous location of households. To bolster this argument, we employ a placebo test that provides additional evidence supporting our exclusion restriction. In columns 6 and 7, each upstream district is replaced with the nearest off-river neighbor with nonmissing pollution data. If the resulting 'falsified' upstream pollution variable remains a significant predictor of downstream mortality, then we are likely capturing spurious, spatial correlation between a given district and its neighbors due to the regional nature of many health and environmental shocks.
This, however, is not the case: the falsified upstream pollution variable loses both its magnitude and significance, thereby suggesting that our true upstream variable is isolating quasi-random variation in pollution that originates upstream and yet flows downstream to other districts. The next step is to estimate the dose-response function and use it to test the importance of the pollution mechanism. The results of this step are documented in Table 6, which details the first and second stages of our 2SLS estimation and the associated over-identification tests. In columns 1-4, we employ a single instrument -upstream pollution, as before. Once again, the first three columns pertain to short-, medium-, and long-run periods of observation, while the fourth uses the full time period but parameterizes the policy variable as three period-specific dummies. The first-stage is strong in all four columns. 14 In the second stage, all specifications yield statistically significant point estimates of the pollution coefficient (in Panel B), which is approximately four to six times larger in magnitude than its upstream counterpart in Table 5. The implication is that surpassing the BOD standard of 3 mg/l raises the risk of neonatal mortality by 10-15 percentage points. lower bound in columns 4 and 5 of Table 5 as well as in columns 1-3 and 5-7 of Table A5. 14 The first stage is also very strong in IV models that replace BOD with sulfates, chlorides, or calcium. These results are available upon request.
The first four columns of Table 6 also provide some initial evidence on policy mechanisms. In these columns, the point estimates of the treatment coefficient in Panel B reveal whether the verdicts remain predictive of mortality impacts once pollution enters into the equation directly. In all cases, these point estimates are statistically insignificant, which strongly suggests that the verdicts acted on health primarily through the pollution channel. 15  In all cases, the 2-IV coefficient is smaller, but some of the deviation is a natural result of statistical noise. 16 The p-value measures the significance of the deviation, and in columns 5-8 it runs from 0.348 to 0.634. That is, in all cases we fail to reject the null hypothesis that pollution fully explains the verdicts' health impact. Appendix Table A5 shows that this finding is robust to the use of different definitions of upstream pollution and extending the control group to the entire remainder of India.

Discussion
Our estimates of the Ganga Pollution Cases' sizable environmental and health impacts are important because they represent the first documented success in India's regulation of water quality. In addition, our findings are differentiated by the type of pollution studied. We provide the first quasi-experimental evidence of the association between biochemical oxygen demand and health. Prior work has focused on pollution from either domestic (Field, Glennerster, and Hussam 2011) or agricultural (Brainerd and Menon 2014) sources. Our own work links mortality to industrial pollution for the first time. 17 15 The three period-specific treatment coefficients are omitted from Panel B, column 4 for table conciseness but are similarly small and insignificant. 16 One back-of-the-envelope method of interpreting our results is to compare the product of our 'policy-on-pollution' and 'pollution-on-health' point estimates with our 'policy-on-health' point estimate. The relevant results must hold constant the time-period (1986)(1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004), geography (Ganga Basin) and upstream range (here, we choose [75,200]). The reduced form results for this specific sample are in Table A3 (column 1 and column 4) and Table A4 (column 3) respectively. The product of the two coefficients from the reduced form coefficients is 0.111 * (-0.381) = -0.042. This is larger than, but still in the range of, the policy-mortality coefficient of -0.034. We stress, however, that the overidentification test is the statistically correct way to interpret our policy mechanism results. 17 Our identified BOD impacts could, in principle, capture reductions in domestic pollution, but the absence of impacts on FCOLI -a more direct measure of domestic pollution -makes that explanation unlikely.
In addition, we have described and implemented a novel, replicable methodology for producing an unbiased estimate of the water pollution-health dose-response function leveraging the flow of rivers.
Our reduced-form estimate of the relationship between environmental quality and health mirrors those of Ebenstein (2012) and Brainerd and Menon (2014) and indicates that the costs of river pollution are significant. 18 The latter authors' findings are particularly consistent with our own results, in that they, too, specifically show neonatal mortality to be affected by water pollution in India. Given widespread breastfeeding practices in India, these results suggest that alternative modes of contamination other than drinking river water -such as bathing and person-to-person transmission -are at work (Cifuentes et al. 2000). Alternatively, pollution exposure in utero may explain our neonatal mortality estimates. While we have limited ability to investigate the timing of pollution exposure, our policy analysis suggests that infants whose gestation periods fall fully after the verdicts have a better probability of survival.
In spite of the harms of pollution, environmental quality remains low in many developing countries. One explanation is a high marginal utility of consumption (see Greenstone and Jack 2015), which could make the private opportunity cost of pollution reduction greater than the private benefit. In support of this hypothesis, Hanna and Oliva (2015a) show that when Indian households are randomly made wealthier (through a cash-and-livestock transfer), they choose to consume more energy but not cleaner energy. Relatedly, Greenstone and Hanna (2014) argue that, in India, air pollution policy has been more effective than water pollution policy because of relatively greater demand for air quality than for water quality. In our context, Kanpur Nagar district had BOD levels that placed it in the 63rd percentile of Ganga Basin water quality prior to the Supreme Court rulings. Higher pollution levels may, in principle, be associated with larger willingness-to-pay for environmental quality, but our qualitative investigations suggest that demand for water quality was quite low in Kanpur in the 1980s: there was no local movement to improve water quality at the time of Mehta's writ petition, but there was significant concern about the economic impacts of regulation in a city that relied heavily on the tanning industry for jobs. 19 18 Ebenstein (2012) shows that decreases in Chinese river water quality are associated with rises in adult deaths due to stomach cancer: a one-grade deterioration (from a six-grade scale) predicts a 10 percent rise in stomach-cancer mortality. Brainerd and Menon (2014), meanwhile, find that a 10 percent rise in agricultural pollution into India's rivers is associated with a 6 percent increase in neonatal mortality. 19 We conducted a qualitative research study in Kanpur in the winter of 2014. The study consisted of interviews of a large number of stakeholders and informants in the Ganga Pollution Cases. These included government officials, High marginal costs of abatement and political economic distortions are two other possible explanations for poor environmental quality in the developing world. Davis (2008) and Field, Glennerster, and Hussam (2011), for instance, identify environmental regulations whose designs had unintended consequences for air quality in Mexico and infant health in Bangladesh, respectively. Sigman (2002Sigman ( , 2005) and Lipscomb and Mobarak (2017), meanwhile, document greater water pollution upstream of administrative borders, where even a public agent may not fully internalize the social cost of poor water quality. Our IV analysis underscores this problem by documenting the existence of spatial spillovers in river pollution. It also connects this spatial externality directly to health outcomes for the first time, by showing that upstream pollution imparts a mortality burden on downstream communities.
River pollution is thus a collective action problem requiring central government regulation and/or inter-jurisdictional bargaining.
Why, then, did the Ganga Pollution Cases succeed where other Indian government action -such as the ambitious National River Conservation Plan (NRCP) -failed? One feature that distinguishes the verdicts from the NRCP is that the environmental policy produced by the former emanated from the judicial branch rather than the executive. Article 21 of the Indian Constitution provides citizens with the "Right to Life"; Mehta vs. Union of India is a watershed moment in Indian legal and environmental history because it was the first instance in which this constitutional 'lever' was used to drive environmental policy through the Indian Judiciary. It is plausible that decisions from the judiciary differ from those made by the executive in that they mandate agents to take specific -and verifiable -actions, rather than designing complex investment and incentive schemes that are arguably more difficult to implement (Davis 2008). Furthermore, the set of stakeholders empowered to monitor the execution of a judicial decision, which includes citizens, might reduce the scope for firms to cheat the government's monitoring system (Duflo et al. 2013

Conclusion
This paper provides empirical evidence that the Supreme Court decisions in Mehta vs. Union of India, which primarily targeted the tanning industry in Kanpur district, induced a drop in both river pollution and neonatal mortality. Our investigation of the mechanisms of policy indicates that the observed drop in pollution may have played an important role in explaining the identified mortality effect. In deriving owners of tanneries, two non-governmental organizations, several journalists, three professors at the Indian Institute of Technology, an operator of a Common Effluent Treatment Plant (CETP), elderly rickshaw pullers who worked in Jajmau in the 1980s, and taxi drivers. We examined official documents from the time that the Ganga Action Plan was implemented and progress reports in later years. For background on citizen activism, we are grateful to the Ecofriends organization, which was established in Kanpur in 1993. and conducting our tests of this pollution channel, we show how information about different potential mechanisms can be backed out from analysis even when data on all possible mechanisms are not available.
We believe our analysis represents an important contribution to the broader puzzle of continually poor environmental quality in developing countries. First, we have identified a precedent for successful water pollution policy in India. While one data point cannot identify the keys to broadly effective environmental policy, the attributes of the Ganga Pollution Cases and their context -including (but not limited to) judicial backing, industrial focus, and water quality improvements rather than behavior change -may inform the design of successful policy in an area where it has previously proven elusive.
Second, we have demonstrated that river pollution has a real, adverse impact on infant health in India. This is important because existing research predominantly focuses on air quality rather than water quality, and demand for the latter appears correspondingly lower than demand for the former.
Our results highlight the susceptibility of newborns in their first month of life, the harms of industrial pollution, and the association between mortality and the broad-based measure of biochemical oxygen demand. Methodologically, we show how the unidirectional flow of rivers can be leveraged for unbiased estimation of the water pollution-health dose-response function.
Finally, we have shown that the ultimate incidence of the costs of pollution is not limited to the origin of that pollution. Rather, water pollution flows downstream to other communities living along rivers, reducing not just water quality but also the likelihood of infant survival. This finding highlights the spatial externality inherent to pollution and underscores the need for inter-jurisdictional bargaining.
All in all, we provide several pieces of insight into the challenges of designing effective environmental policy in the context of a developing country.

Control Districts
Notes: Both panels show kernel densities of residuals from a regression of ln(BOD) on CETP capacity, air temperature, total precipitation, and NRCP dummy and district and year-month fixed effects. Panel A shows residuals from treatment districts, while Panel B shows residuals from control districts. Both panels plot separate densities for before vs. after the Supreme Court verdicts.    1971-2004 1971-2004 1971-2004 1980-1994 1980-1994 1980-1994 Adj. R-Squared Notes: All results are based on estimation of equation (1). An observation is a child-month. In all columns, the dependent variable is a dummy for mortality. In columns 1 and 4, the sample includes all observations from the first twelve months of life and corresponds to infant mortality (IM). In columns 2 and 5, the sample includes only observations from the first month, corresponding to neonatal mortality (NM). Columns 3 and 6 include observations from months 2-12, which captures infant mortality conditional on neonatal survival (Cond. IM). All regressions include a set of controls (religion of the household head, caste of the household head, mother's age, mother's literacy, CETP capacity, air temperature, total precipitation, and NRCP dummy) and district and year-month fixed effects. Observations are weighted by the product of survey sampling weight and inverse size of the relevant age cohort (in months). Standard errors are clustered at the district level in parentheses. ***, **, and * indicate statistical significance at the 1, 5, and 10 percent levels, respectively.  1980-2004 1980-1999 1980-1994 1980-2004 1980-1994 1980-1994 1980-1994 Adj. R-Squared Notes: All results are based on estimation of equation (1). An observation is a child-month. The dependent variable in all columns is a binary variable equaling one if a child died in the first month of life. All regressions include a set of controls (religion of the household head, caste of the household head, mother's age, mother's literacy, CETP capacity, air temperature, total precipitation, and NRCP dummy) and district and year-month fixed effects. Observations are weighted by survey sampling weights. Standard errors are clustered at the district level in parentheses.
In columns 1-4, the endogenous variable (BOD > 3) is instrumented using its upstream analog ('1-IV'), while in columns 5-8, both upstream pollution and the policy variable ("1[Kanpur] X 1[Post-Verdict]") are used as instruments ('2-IV'). In 2-IV columns, the listed p-value corresponds to a C-statistic, which tests the null hypothesis that the endogenous variable is overidentified. All regressions include a set of controls (religion of the household head, caste of the household head, mother's age, mother's literacy, local CETP capacity, air temperature, total precipitation, and NRCP) as well as district and year-month fixed effects.
Observations are weighted by survey sampling weights. Standard errors are clustered at the district level in parentheses. ***, **, and * indicate statistical significance at the 1, 5, and 10 percent, respectively.       (1)

A Appendix tables and figures
(3)  1986-2004 1986-1999 1986-1994 1986-2004 1986-1999 1986-1994 Adj. R-Squared  (1), with a binary dependent variable equaling one if a child died in the first month of life. Columns 4-6 correspond to estimation of equation (2), with a binary dependent variable equaling one if district-average BOD exceeds 3 mg/l in a given month. All regressions include a set of controls (religion of the household head, caste of the household head, mother's age, mother's literacy, CETP capacity, air temperature, total precipitation, and NRCP dummy) and district and year-month fixed effects. Observations are wThe ed by survey sampling weights. Standard errors are clustered at the district level in parentheses. ***, **, and * indicate statistical significance at the 1, 5, and 10 percent levels, respectively. (3) (3) We derive a statistical test of the mechanisms of policy impact by leveraging two potential instruments for river pollution: (1) upstream river pollution, and (2) the policy itself. We begin by reprinting equation (1).
M ortality idt = a + bT dt + X idt γ + e idt (A.1) In this reduced-form model of infant mortality, identification rests on the assumption that Cov T dt , e idt = 0 -i.e., that the policy variable is uncorrelated with all unobserved predictors of neonatal mortality. If this zero-covariance assumption holds, b represents the net causal effect of policy on neonatal mortality, aggregated across all channels of impact.
To gauge the relative importance of the various channels, we parsimoniously model the structural determinants of mortality rates as follows: This equation is analogous to equation (5) in the main text, except that it partitions the space of unobserved risk factors into two: Z idt and ε idt . The former is a vector of all factors that are also correlated with environmental policy T dt . These include, but are not restricted to, individual awareness about river water contamination, changes in factor prices stemming from the implementation of environmental policy T dt , or any type of private or public interventions that might have been triggered by T dt . The latter captures the other risk factors of infant mortality and is, by construction, such that Cov T dt , ε idt |X idt = 0.
Our next step is to parameterize the relationship between policy and the the structural determinants of health (i.e., the right-hand side of equation A.2). We assume that Z idt and P ollution dt respond to environmental policy and other determinants in linear fashion. That is, and P ollution dt = α 2 + β 2 T dt +X idt γ 2 + ε 2 idt (A.4) We can then rewrite equation (A.2) by substituting for both Z idt and P ollution dt , so as to decompose the policy's effect on mortality into the different channels of impact: M ortality idt = α + βα 2 + α 1 δ + ββ 2 + β 1 δ T dt +X idt βγ 2 + γ + γ 1 δ + βε 2 idt + ε 1 idt δ + ε idt (A.5) The total impact of environmental policy T dt on infant mortality (equal to b in reduced form) is here given by ββ 2 + β 1 δ . The first term (ββ 2 ) measures the contribution of the pollution channel, and the second (β 1 δ) aggregates all other channels; the challenge is to estimate each of these terms separately.
We estimate β 2 from equation (2). The remaining components of the policy-mortality relationship are obtained from equation (A.2), by substituting the right-hand side of equation (A.3) in for the unobservable Z idt . This yields M ortality idt = α + α 1 δ + βP ollution dt +X idt γ + γ 1 δ + β 1 δ T dt + ε 1 idt δ + ε idt (A.6) which contains both β and the product β 1 δ. We estimate equation (A.6) using 2SLS, where we instrument for P ollution dt with its upstream analog, P ollution u dt . With unbiased estimates of both β and β 1 δ, we can then test H 0 : β 1 δ = 0, the null hypothesis that non-pollution channels of policy impact are inactive.
Note that under this null hypothesis, equation (A.6) can be rewritten M ortality idt = α + α 1 δ + βP ollution dt +X idt γ + γ 1 · δ + ε 1 idt δ + ε idt (A.7) so that T dt is additionally excluded from second-stage equation (A.7) and becomes another valid instrument for P ollution dt . One test of H 0 is therefore an over-identification test that assesses the orthogonality condition for T dt , as part of the larger set of instruments {T dt , P ollution u dt }. To implement such a test, we construct a C-statistic (see, e.g., Eichenbaum, Hansen, and Singleton 1988),which we describe further in the main text.